X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/5ba3f43ea354af8ad55bea84372a2bc834d8757c..refs/heads/master:/bsd/kern/kern_event.c diff --git a/bsd/kern/kern_event.c b/bsd/kern/kern_event.c index f64bef436..f3277df75 100644 --- a/bsd/kern/kern_event.c +++ b/bsd/kern/kern_event.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2000-2017 Apple Inc. All rights reserved. + * Copyright (c) 2000-2020 Apple Inc. All rights reserved. * * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ * @@ -55,7 +55,7 @@ * @(#)kern_event.c 1.0 (3/31/2000) */ #include -#include +#include #include #include @@ -75,6 +75,7 @@ #include #include #include +#include // SYS_* constants #include #include #include @@ -85,8 +86,8 @@ #include #include #include -#include -#include +#include +#include #include #include @@ -102,8 +103,10 @@ #include #include +#include +#include +#include #include -#include #include "net/net_str_id.h" @@ -114,102 +117,72 @@ #include #endif -extern thread_t port_name_to_thread(mach_port_name_t port_name); /* osfmk/kern/ipc_tt.h */ +#if DEVELOPMENT || DEBUG +#define KEVENT_PANIC_ON_WORKLOOP_OWNERSHIP_LEAK (1U << 0) +#define KEVENT_PANIC_ON_NON_ENQUEUED_PROCESS (1U << 1) +TUNABLE(uint32_t, kevent_debug_flags, "kevent_debug", 0); +#endif + +static LCK_GRP_DECLARE(kq_lck_grp, "kqueue"); +SECURITY_READ_ONLY_EARLY(vm_packing_params_t) kn_kq_packing_params = + VM_PACKING_PARAMS(KNOTE_KQ_PACKED); + extern mach_port_name_t ipc_entry_name_mask(mach_port_name_t name); /* osfmk/ipc/ipc_entry.h */ +extern int cansignal(struct proc *, kauth_cred_t, struct proc *, int); /* bsd/kern/kern_sig.c */ #define KEV_EVTID(code) BSDDBG_CODE(DBG_BSD_KEVENT, (code)) /* - * JMM - this typedef needs to be unified with pthread_priority_t - * and mach_msg_priority_t. It also needs to be the same type - * everywhere. + * If you need accounting for KM_KQUEUE consider using + * KALLOC_HEAP_DEFINE to define a zone view. */ -typedef int32_t qos_t; - -MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system"); - -#define KQ_EVENT NO_EVENT64 - -#define KNUSE_NONE 0x0 -#define KNUSE_STEAL_DROP 0x1 -#define KNUSE_BOOST 0x2 -static int kqlock2knoteuse(struct kqueue *kq, struct knote *kn, int flags); -static int kqlock2knotedrop(struct kqueue *kq, struct knote *kn); -static int kqlock2knotedetach(struct kqueue *kq, struct knote *kn, int flags); -static int knoteuse2kqlock(struct kqueue *kq, struct knote *kn, int flags); - -static int kqueue_read(struct fileproc *fp, struct uio *uio, - int flags, vfs_context_t ctx); -static int kqueue_write(struct fileproc *fp, struct uio *uio, - int flags, vfs_context_t ctx); -static int kqueue_ioctl(struct fileproc *fp, u_long com, caddr_t data, - vfs_context_t ctx); +#define KM_KQUEUE KHEAP_DEFAULT + +#define KQ_EVENT NO_EVENT64 + static int kqueue_select(struct fileproc *fp, int which, void *wq_link_id, - vfs_context_t ctx); + vfs_context_t ctx); static int kqueue_close(struct fileglob *fg, vfs_context_t ctx); static int kqueue_kqfilter(struct fileproc *fp, struct knote *kn, - struct kevent_internal_s *kev, vfs_context_t ctx); + struct kevent_qos_s *kev); static int kqueue_drain(struct fileproc *fp, vfs_context_t ctx); static const struct fileops kqueueops = { - .fo_type = DTYPE_KQUEUE, - .fo_read = kqueue_read, - .fo_write = kqueue_write, - .fo_ioctl = kqueue_ioctl, - .fo_select = kqueue_select, - .fo_close = kqueue_close, + .fo_type = DTYPE_KQUEUE, + .fo_read = fo_no_read, + .fo_write = fo_no_write, + .fo_ioctl = fo_no_ioctl, + .fo_select = kqueue_select, + .fo_close = kqueue_close, + .fo_drain = kqueue_drain, .fo_kqfilter = kqueue_kqfilter, - .fo_drain = kqueue_drain, }; -static void kevent_put_kq(struct proc *p, kqueue_id_t id, struct fileproc *fp, struct kqueue *kq); -static int kevent_internal(struct proc *p, - kqueue_id_t id, kqueue_id_t *id_out, - user_addr_t changelist, int nchanges, - user_addr_t eventlist, int nevents, - user_addr_t data_out, uint64_t data_available, - unsigned int flags, user_addr_t utimeout, - kqueue_continue_t continuation, - int32_t *retval); -static int kevent_copyin(user_addr_t *addrp, struct kevent_internal_s *kevp, - struct proc *p, unsigned int flags); -static int kevent_copyout(struct kevent_internal_s *kevp, user_addr_t *addrp, - struct proc *p, unsigned int flags); -char * kevent_description(struct kevent_internal_s *kevp, char *s, size_t n); - -static void kqueue_interrupt(struct kqueue *kq); -static int kevent_callback(struct kqueue *kq, struct kevent_internal_s *kevp, - void *data); -static void kevent_continue(struct kqueue *kq, void *data, int error); -static void kqueue_scan_continue(void *contp, wait_result_t wait_result); -static int kqueue_process(struct kqueue *kq, kevent_callback_t callback, void *callback_data, - struct filt_process_s *process_data, int *countp, struct proc *p); -static struct kqtailq *kqueue_get_base_queue(struct kqueue *kq, kq_index_t qos_index); -static struct kqtailq *kqueue_get_high_queue(struct kqueue *kq, kq_index_t qos_index); -static int kqueue_queue_empty(struct kqueue *kq, kq_index_t qos_index); - -static struct kqtailq *kqueue_get_suppressed_queue(struct kqueue *kq, kq_index_t qos_index); - -static void kqworkq_request_thread(struct kqworkq *kqwq, kq_index_t qos_index); -static void kqworkq_request_help(struct kqworkq *kqwq, kq_index_t qos_index); -static void kqworkq_update_override(struct kqworkq *kqwq, kq_index_t qos_index, kq_index_t override_index); -static void kqworkq_bind_thread_impl(struct kqworkq *kqwq, kq_index_t qos_index, thread_t thread, unsigned int flags); -static void kqworkq_unbind_thread(struct kqworkq *kqwq, kq_index_t qos_index, thread_t thread, unsigned int flags); -static struct kqrequest *kqworkq_get_request(struct kqworkq *kqwq, kq_index_t qos_index); +static inline int kevent_modern_copyout(struct kevent_qos_s *, user_addr_t *); +static int kevent_register_wait_prepare(struct knote *kn, struct kevent_qos_s *kev, int result); +static void kevent_register_wait_block(struct turnstile *ts, thread_t handoff_thread, + thread_continue_t cont, struct _kevent_register *cont_args) __dead2; +static void kevent_register_wait_return(struct _kevent_register *cont_args) __dead2; +static void kevent_register_wait_cleanup(struct knote *kn); -enum { - KQWL_UO_NONE = 0, - KQWL_UO_OLD_OVERRIDE_IS_SYNC_UI = 0x1, - KQWL_UO_NEW_OVERRIDE_IS_SYNC_UI = 0x2, - KQWL_UO_UPDATE_SUPPRESS_SYNC_COUNTERS = 0x4, - KQWL_UO_UPDATE_OVERRIDE_LAZY = 0x8 -}; +static struct kqtailq *kqueue_get_suppressed_queue(kqueue_t kq, struct knote *kn); +static void kqueue_threadreq_initiate(struct kqueue *kq, workq_threadreq_t, kq_index_t qos, int flags); -static void kqworkloop_update_override(struct kqworkloop *kqwl, kq_index_t qos_index, kq_index_t override_index, uint32_t flags); -static void kqworkloop_bind_thread_impl(struct kqworkloop *kqwl, thread_t thread, unsigned int flags); -static void kqworkloop_unbind_thread(struct kqworkloop *kqwl, thread_t thread, unsigned int flags); -static inline kq_index_t kqworkloop_combined_qos(struct kqworkloop *kqwl, boolean_t *); -static void kqworkloop_update_suppress_sync_count(struct kqrequest *kqr, uint32_t flags); +static void kqworkq_unbind(proc_t p, workq_threadreq_t); +static thread_qos_t kqworkq_unbind_locked(struct kqworkq *kqwq, workq_threadreq_t, thread_t thread); +static workq_threadreq_t kqworkq_get_request(struct kqworkq *kqwq, kq_index_t qos_index); + +static void kqworkloop_unbind(struct kqworkloop *kwql); + +enum kqwl_unbind_locked_mode { + KQWL_OVERRIDE_DROP_IMMEDIATELY, + KQWL_OVERRIDE_DROP_DELAYED, +}; +static void kqworkloop_unbind_locked(struct kqworkloop *kwql, thread_t thread, + enum kqwl_unbind_locked_mode how); +static void kqworkloop_unbind_delayed_override_drop(thread_t thread); +static kq_index_t kqworkloop_override(struct kqworkloop *kqwl); +static void kqworkloop_set_overcommit(struct kqworkloop *kqwl); enum { KQWL_UTQ_NONE, /* @@ -217,12 +190,14 @@ enum { * * This QoS is accounted for with the events override in the * kqr_override_index field. It is raised each time a new knote is queued at - * a given QoS. The kqr_wakeup_indexes field is a superset of the non empty + * a given QoS. The kqwl_wakeup_indexes field is a superset of the non empty * knote buckets and is recomputed after each event delivery. */ KQWL_UTQ_UPDATE_WAKEUP_QOS, KQWL_UTQ_UPDATE_STAYACTIVE_QOS, KQWL_UTQ_RECOMPUTE_WAKEUP_QOS, + KQWL_UTQ_UNBINDING, /* attempt to rebind */ + KQWL_UTQ_PARKING, /* * The wakeup override is for suppressed knotes that have fired again at * a higher QoS than the one for which they are suppressed already. @@ -231,157 +206,74 @@ enum { KQWL_UTQ_UPDATE_WAKEUP_OVERRIDE, KQWL_UTQ_RESET_WAKEUP_OVERRIDE, /* - * The async QoS is the maximum QoS of an event enqueued on this workloop in + * The QoS is the maximum QoS of an event enqueued on this workloop in * userland. It is copied from the only EVFILT_WORKLOOP knote with * a NOTE_WL_THREAD_REQUEST bit set allowed on this workloop. If there is no * such knote, this QoS is 0. */ - KQWL_UTQ_SET_ASYNC_QOS, - /* - * The sync waiters QoS is the maximum QoS of any thread blocked on an - * EVFILT_WORKLOOP knote marked with the NOTE_WL_SYNC_WAIT bit. - * If there is no such knote, this QoS is 0. - */ - KQWL_UTQ_SET_SYNC_WAITERS_QOS, + KQWL_UTQ_SET_QOS_INDEX, KQWL_UTQ_REDRIVE_EVENTS, }; static void kqworkloop_update_threads_qos(struct kqworkloop *kqwl, int op, kq_index_t qos); -static void kqworkloop_request_help(struct kqworkloop *kqwl, kq_index_t qos_index); - -static int knote_process(struct knote *kn, kevent_callback_t callback, void *callback_data, - struct filt_process_s *process_data, struct proc *p); -#if 0 -static void knote_put(struct knote *kn); -#endif +static int kqworkloop_end_processing(struct kqworkloop *kqwl, int flags, int kevent_flags); -static int kq_add_knote(struct kqueue *kq, struct knote *kn, - struct kevent_internal_s *kev, struct proc *p, int *knoteuse_flags); -static struct knote *kq_find_knote_and_kq_lock(struct kqueue *kq, struct kevent_internal_s *kev, bool is_fd, struct proc *p); -static void kq_remove_knote(struct kqueue *kq, struct knote *kn, struct proc *p, kn_status_t *kn_status, uint16_t *kq_state); - -static void knote_drop(struct knote *kn, struct proc *p); static struct knote *knote_alloc(void); static void knote_free(struct knote *kn); +static int kq_add_knote(struct kqueue *kq, struct knote *kn, + struct knote_lock_ctx *knlc, struct proc *p); +static struct knote *kq_find_knote_and_kq_lock(struct kqueue *kq, + struct kevent_qos_s *kev, bool is_fd, struct proc *p); + +static void knote_activate(kqueue_t kqu, struct knote *kn, int result); +static void knote_dequeue(kqueue_t kqu, struct knote *kn); + +static void knote_apply_touch(kqueue_t kqu, struct knote *kn, + struct kevent_qos_s *kev, int result); +static void knote_suppress(kqueue_t kqu, struct knote *kn); +static void knote_unsuppress(kqueue_t kqu, struct knote *kn); +static void knote_drop(kqueue_t kqu, struct knote *kn, struct knote_lock_ctx *knlc); + +// both these functions may dequeue the knote and it is up to the caller +// to enqueue the knote back +static void knote_adjust_qos(struct kqueue *kq, struct knote *kn, int result); +static void knote_reset_priority(kqueue_t kqu, struct knote *kn, pthread_priority_t pp); + +static ZONE_DECLARE(knote_zone, "knote zone", + sizeof(struct knote), ZC_CACHING | ZC_ZFREE_CLEARMEM); +static ZONE_DECLARE(kqfile_zone, "kqueue file zone", + sizeof(struct kqfile), ZC_ZFREE_CLEARMEM); +static ZONE_DECLARE(kqworkq_zone, "kqueue workq zone", + sizeof(struct kqworkq), ZC_ZFREE_CLEARMEM); +static ZONE_DECLARE(kqworkloop_zone, "kqueue workloop zone", + sizeof(struct kqworkloop), ZC_CACHING | ZC_ZFREE_CLEARMEM); + +#define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask)) + +static int filt_no_attach(struct knote *kn, struct kevent_qos_s *kev); +static void filt_no_detach(struct knote *kn); +static int filt_bad_event(struct knote *kn, long hint); +static int filt_bad_touch(struct knote *kn, struct kevent_qos_s *kev); +static int filt_bad_process(struct knote *kn, struct kevent_qos_s *kev); -static void knote_activate(struct knote *kn); -static void knote_deactivate(struct knote *kn); - -static void knote_enable(struct knote *kn); -static void knote_disable(struct knote *kn); - -static int knote_enqueue(struct knote *kn); -static void knote_dequeue(struct knote *kn); - -static void knote_suppress(struct knote *kn); -static void knote_unsuppress(struct knote *kn); -static void knote_wakeup(struct knote *kn); - -static kq_index_t knote_get_queue_index(struct knote *kn); -static struct kqtailq *knote_get_queue(struct knote *kn); -static kq_index_t knote_get_req_index(struct knote *kn); -static kq_index_t knote_get_qos_index(struct knote *kn); -static void knote_set_qos_index(struct knote *kn, kq_index_t qos_index); -static kq_index_t knote_get_qos_override_index(struct knote *kn); -static kq_index_t knote_get_sync_qos_override_index(struct knote *kn); -static void knote_set_qos_override_index(struct knote *kn, kq_index_t qos_index, boolean_t override_is_sync); -static void knote_set_qos_overcommit(struct knote *kn); - -static int filt_fileattach(struct knote *kn, struct kevent_internal_s *kev); -SECURITY_READ_ONLY_EARLY(static struct filterops) file_filtops = { - .f_isfd = 1, - .f_attach = filt_fileattach, -}; - -static void filt_kqdetach(struct knote *kn); -static int filt_kqueue(struct knote *kn, long hint); -static int filt_kqtouch(struct knote *kn, struct kevent_internal_s *kev); -static int filt_kqprocess(struct knote *kn, struct filt_process_s *data, struct kevent_internal_s *kev); -SECURITY_READ_ONLY_EARLY(static struct filterops) kqread_filtops = { - .f_isfd = 1, - .f_detach = filt_kqdetach, - .f_event = filt_kqueue, - .f_touch = filt_kqtouch, - .f_process = filt_kqprocess, -}; - -/* placeholder for not-yet-implemented filters */ -static int filt_badattach(struct knote *kn, struct kevent_internal_s *kev); SECURITY_READ_ONLY_EARLY(static struct filterops) bad_filtops = { - .f_attach = filt_badattach, -}; - -static int filt_procattach(struct knote *kn, struct kevent_internal_s *kev); -static void filt_procdetach(struct knote *kn); -static int filt_proc(struct knote *kn, long hint); -static int filt_proctouch(struct knote *kn, struct kevent_internal_s *kev); -static int filt_procprocess(struct knote *kn, struct filt_process_s *data, struct kevent_internal_s *kev); -SECURITY_READ_ONLY_EARLY(static struct filterops) proc_filtops = { - .f_attach = filt_procattach, - .f_detach = filt_procdetach, - .f_event = filt_proc, - .f_touch = filt_proctouch, - .f_process = filt_procprocess, + .f_attach = filt_no_attach, + .f_detach = filt_no_detach, + .f_event = filt_bad_event, + .f_touch = filt_bad_touch, + .f_process = filt_bad_process, }; #if CONFIG_MEMORYSTATUS extern const struct filterops memorystatus_filtops; #endif /* CONFIG_MEMORYSTATUS */ - extern const struct filterops fs_filtops; - extern const struct filterops sig_filtops; - -static zone_t knote_zone; -static zone_t kqfile_zone; -static zone_t kqworkq_zone; -static zone_t kqworkloop_zone; - -#define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask)) - -/* Mach portset filter */ extern const struct filterops machport_filtops; - -/* User filter */ -static int filt_userattach(struct knote *kn, struct kevent_internal_s *kev); -static void filt_userdetach(struct knote *kn); -static int filt_user(struct knote *kn, long hint); -static int filt_usertouch(struct knote *kn, struct kevent_internal_s *kev); -static int filt_userprocess(struct knote *kn, struct filt_process_s *data, struct kevent_internal_s *kev); -SECURITY_READ_ONLY_EARLY(static struct filterops) user_filtops = { - .f_attach = filt_userattach, - .f_detach = filt_userdetach, - .f_event = filt_user, - .f_touch = filt_usertouch, - .f_process = filt_userprocess, -}; - -static lck_spin_t _filt_userlock; -static void filt_userlock(void); -static void filt_userunlock(void); - -/* Workloop filter */ -static bool filt_wlneeds_boost(struct kevent_internal_s *kev); -static int filt_wlattach(struct knote *kn, struct kevent_internal_s *kev); -static int filt_wlpost_attach(struct knote *kn, struct kevent_internal_s *kev); -static void filt_wldetach(struct knote *kn); -static int filt_wlevent(struct knote *kn, long hint); -static int filt_wltouch(struct knote *kn, struct kevent_internal_s *kev); -static int filt_wldrop_and_unlock(struct knote *kn, struct kevent_internal_s *kev); -static int filt_wlprocess(struct knote *kn, struct filt_process_s *data, struct kevent_internal_s *kev); -SECURITY_READ_ONLY_EARLY(static struct filterops) workloop_filtops = { - .f_needs_boost = filt_wlneeds_boost, - .f_attach = filt_wlattach, - .f_post_attach = filt_wlpost_attach, - .f_detach = filt_wldetach, - .f_event = filt_wlevent, - .f_touch = filt_wltouch, - .f_drop_and_unlock = filt_wldrop_and_unlock, - .f_process = filt_wlprocess, -}; - +extern const struct filterops pipe_nfiltops; extern const struct filterops pipe_rfiltops; extern const struct filterops pipe_wfiltops; extern const struct filterops ptsd_kqops; +extern const struct filterops ptmx_kqops; extern const struct filterops soread_filtops; extern const struct filterops sowrite_filtops; extern const struct filterops sock_filtops; @@ -393,7 +285,12 @@ extern const struct filterops fsevent_filtops; extern const struct filterops vnode_filtops; extern const struct filterops tty_filtops; +const static struct filterops file_filtops; +const static struct filterops kqread_filtops; +const static struct filterops proc_filtops; const static struct filterops timer_filtops; +const static struct filterops user_filtops; +const static struct filterops workloop_filtops; /* * @@ -402,169 +299,124 @@ const static struct filterops timer_filtops; * - Add a new "EVFILT_" option value to bsd/sys/event.h (typically a negative value) * in the exported section of the header * - Update the EVFILT_SYSCOUNT value to reflect the new addition - * - Add a filterops to the sysfilt_ops array. Public filters should be added at the end + * - Add a filterops to the sysfilt_ops array. Public filters should be added at the end * of the Public Filters section in the array. * Private filters: * - Add a new "EVFILT_" value to bsd/sys/event.h (typically a positive value) * in the XNU_KERNEL_PRIVATE section of the header * - Update the EVFILTID_MAX value to reflect the new addition - * - Add a filterops to the sysfilt_ops. Private filters should be added at the end of - * the Private filters section of the array. + * - Add a filterops to the sysfilt_ops. Private filters should be added at the end of + * the Private filters section of the array. */ -SECURITY_READ_ONLY_EARLY(static struct filterops *) sysfilt_ops[EVFILTID_MAX] = { +static_assert(EVFILTID_MAX < UINT8_MAX, "kn_filtid expects this to be true"); +static const struct filterops * const sysfilt_ops[EVFILTID_MAX] = { /* Public Filters */ - [~EVFILT_READ] = &file_filtops, - [~EVFILT_WRITE] = &file_filtops, - [~EVFILT_AIO] = &bad_filtops, - [~EVFILT_VNODE] = &file_filtops, - [~EVFILT_PROC] = &proc_filtops, - [~EVFILT_SIGNAL] = &sig_filtops, - [~EVFILT_TIMER] = &timer_filtops, - [~EVFILT_MACHPORT] = &machport_filtops, - [~EVFILT_FS] = &fs_filtops, - [~EVFILT_USER] = &user_filtops, - &bad_filtops, - &bad_filtops, - [~EVFILT_SOCK] = &file_filtops, + [~EVFILT_READ] = &file_filtops, + [~EVFILT_WRITE] = &file_filtops, + [~EVFILT_AIO] = &bad_filtops, + [~EVFILT_VNODE] = &file_filtops, + [~EVFILT_PROC] = &proc_filtops, + [~EVFILT_SIGNAL] = &sig_filtops, + [~EVFILT_TIMER] = &timer_filtops, + [~EVFILT_MACHPORT] = &machport_filtops, + [~EVFILT_FS] = &fs_filtops, + [~EVFILT_USER] = &user_filtops, + [~EVFILT_UNUSED_11] = &bad_filtops, + [~EVFILT_VM] = &bad_filtops, + [~EVFILT_SOCK] = &file_filtops, #if CONFIG_MEMORYSTATUS - [~EVFILT_MEMORYSTATUS] = &memorystatus_filtops, + [~EVFILT_MEMORYSTATUS] = &memorystatus_filtops, #else - [~EVFILT_MEMORYSTATUS] = &bad_filtops, + [~EVFILT_MEMORYSTATUS] = &bad_filtops, #endif - [~EVFILT_EXCEPT] = &file_filtops, - + [~EVFILT_EXCEPT] = &file_filtops, [~EVFILT_WORKLOOP] = &workloop_filtops, /* Private filters */ - [EVFILTID_KQREAD] = &kqread_filtops, - [EVFILTID_PIPE_R] = &pipe_rfiltops, - [EVFILTID_PIPE_W] = &pipe_wfiltops, - [EVFILTID_PTSD] = &ptsd_kqops, - [EVFILTID_SOREAD] = &soread_filtops, - [EVFILTID_SOWRITE] = &sowrite_filtops, - [EVFILTID_SCK] = &sock_filtops, - [EVFILTID_SOEXCEPT] = &soexcept_filtops, - [EVFILTID_SPEC] = &spec_filtops, - [EVFILTID_BPFREAD] = &bpfread_filtops, - [EVFILTID_NECP_FD] = &necp_fd_rfiltops, - [EVFILTID_FSEVENT] = &fsevent_filtops, - [EVFILTID_VN] = &vnode_filtops, - [EVFILTID_TTY] = &tty_filtops + [EVFILTID_KQREAD] = &kqread_filtops, + [EVFILTID_PIPE_N] = &pipe_nfiltops, + [EVFILTID_PIPE_R] = &pipe_rfiltops, + [EVFILTID_PIPE_W] = &pipe_wfiltops, + [EVFILTID_PTSD] = &ptsd_kqops, + [EVFILTID_SOREAD] = &soread_filtops, + [EVFILTID_SOWRITE] = &sowrite_filtops, + [EVFILTID_SCK] = &sock_filtops, + [EVFILTID_SOEXCEPT] = &soexcept_filtops, + [EVFILTID_SPEC] = &spec_filtops, + [EVFILTID_BPFREAD] = &bpfread_filtops, + [EVFILTID_NECP_FD] = &necp_fd_rfiltops, + [EVFILTID_FSEVENT] = &fsevent_filtops, + [EVFILTID_VN] = &vnode_filtops, + [EVFILTID_TTY] = &tty_filtops, + [EVFILTID_PTMX] = &ptmx_kqops, + + /* fake filter for detached knotes, keep last */ + [EVFILTID_DETACHED] = &bad_filtops, }; /* waitq prepost callback */ -void waitq_set__CALLING_PREPOST_HOOK__(void *kq_hook, void *knote_hook, int qos); +void waitq_set__CALLING_PREPOST_HOOK__(waitq_set_prepost_hook_t *kq_hook); -#ifndef _PTHREAD_PRIORITY_EVENT_MANAGER_FLAG -#define _PTHREAD_PRIORITY_EVENT_MANAGER_FLAG 0x02000000 /* pthread event manager bit */ -#endif -#ifndef _PTHREAD_PRIORITY_OVERCOMMIT_FLAG -#define _PTHREAD_PRIORITY_OVERCOMMIT_FLAG 0x80000000 /* request overcommit threads */ -#endif -#ifndef _PTHREAD_PRIORITY_QOS_CLASS_MASK -#define _PTHREAD_PRIORITY_QOS_CLASS_MASK 0x003fff00 /* QoS class mask */ -#endif -#ifndef _PTHREAD_PRIORITY_QOS_CLASS_SHIFT_32 -#define _PTHREAD_PRIORITY_QOS_CLASS_SHIFT_32 8 -#endif - -static inline __kdebug_only -uintptr_t -kqr_thread_id(struct kqrequest *kqr) +static inline bool +kqr_thread_bound(workq_threadreq_t kqr) { - return (uintptr_t)thread_tid(kqr->kqr_thread); + return kqr->tr_state == WORKQ_TR_STATE_BOUND; } -static inline -boolean_t is_workqueue_thread(thread_t thread) +static inline bool +kqr_thread_requested_pending(workq_threadreq_t kqr) { - return (thread_get_tag(thread) & THREAD_TAG_WORKQUEUE); + workq_tr_state_t tr_state = kqr->tr_state; + return tr_state > WORKQ_TR_STATE_IDLE && tr_state < WORKQ_TR_STATE_BOUND; } -static inline -void knote_canonicalize_kevent_qos(struct knote *kn) +static inline bool +kqr_thread_requested(workq_threadreq_t kqr) { - struct kqueue *kq = knote_get_kq(kn); - unsigned long canonical; - - if ((kq->kq_state & (KQ_WORKQ | KQ_WORKLOOP)) == 0) - return; - - /* preserve manager and overcommit flags in this case */ - canonical = pthread_priority_canonicalize(kn->kn_qos, FALSE); - kn->kn_qos = (qos_t)canonical; + return kqr->tr_state != WORKQ_TR_STATE_IDLE; } -static inline -kq_index_t qos_index_from_qos(struct knote *kn, qos_t qos, boolean_t propagation) +static inline thread_t +kqr_thread_fast(workq_threadreq_t kqr) { - struct kqueue *kq = knote_get_kq(kn); - kq_index_t qos_index; - unsigned long flags = 0; - - if ((kq->kq_state & (KQ_WORKQ | KQ_WORKLOOP)) == 0) - return QOS_INDEX_KQFILE; - - qos_index = (kq_index_t)thread_qos_from_pthread_priority( - (unsigned long)qos, &flags); - - if (kq->kq_state & KQ_WORKQ) { - /* workq kqueues support requesting a manager thread (non-propagation) */ - if (!propagation && (flags & _PTHREAD_PRIORITY_EVENT_MANAGER_FLAG)) - return KQWQ_QOS_MANAGER; - } - - return qos_index; + assert(kqr_thread_bound(kqr)); + return kqr->tr_thread; } -static inline -qos_t qos_from_qos_index(kq_index_t qos_index) +static inline thread_t +kqr_thread(workq_threadreq_t kqr) { - /* should only happen for KQ_WORKQ */ - if (qos_index == KQWQ_QOS_MANAGER) - return _PTHREAD_PRIORITY_EVENT_MANAGER_FLAG; - - if (qos_index == 0) - return THREAD_QOS_UNSPECIFIED; - - /* Should have support from pthread kext support */ - return (1 << (qos_index - 1 + - _PTHREAD_PRIORITY_QOS_CLASS_SHIFT_32)); + return kqr_thread_bound(kqr) ? kqr->tr_thread : THREAD_NULL; } -/* kqr lock must be held */ -static inline -unsigned long pthread_priority_for_kqrequest( - struct kqrequest *kqr, - kq_index_t qos_index) +static inline struct kqworkloop * +kqr_kqworkloop(workq_threadreq_t kqr) { - unsigned long priority = qos_from_qos_index(qos_index); - if (kqr->kqr_state & KQR_THOVERCOMMIT) { - priority |= _PTHREAD_PRIORITY_OVERCOMMIT_FLAG; + if (kqr->tr_flags & WORKQ_TR_FLAG_WORKLOOP) { + return __container_of(kqr, struct kqworkloop, kqwl_request); } - return priority; + return NULL; } -static inline -kq_index_t qos_index_for_servicer(int qos_class, thread_t thread, int flags) +static inline kqueue_t +kqr_kqueue(proc_t p, workq_threadreq_t kqr) { -#pragma unused(thread) - kq_index_t qos_index; - - if (flags & KEVENT_FLAG_WORKQ_MANAGER) - return KQWQ_QOS_MANAGER; - - qos_index = (kq_index_t)qos_class; - assert(qos_index > 0 && qos_index < KQWQ_QOS_MANAGER); - - return qos_index; + kqueue_t kqu; + if (kqr->tr_flags & WORKQ_TR_FLAG_WORKLOOP) { + kqu.kqwl = kqr_kqworkloop(kqr); + } else { + kqu.kqwq = p->p_fd->fd_wqkqueue; + assert(kqr >= kqu.kqwq->kqwq_request && + kqr < kqu.kqwq->kqwq_request + KQWQ_NBUCKETS); + } + return kqu; } /* * kqueue/note lock implementations * * The kqueue lock guards the kq state, the state of its queues, - * and the kqueue-aware status and use counts of individual knotes. + * and the kqueue-aware status and locks of individual knotes. * * The kqueue workq lock is used to protect state guarding the * interaction of the kqueue with the workq. This state cannot @@ -578,380 +430,609 @@ kq_index_t qos_index_for_servicer(int qos_class, thread_t thread, int flags) * by calling the filter to get a [consistent] snapshot of that * data. */ -lck_grp_attr_t * kq_lck_grp_attr; -lck_grp_t * kq_lck_grp; -lck_attr_t * kq_lck_attr; static inline void -kqlock(struct kqueue *kq) +kqlock(kqueue_t kqu) { - lck_spin_lock(&kq->kq_lock); + lck_spin_lock(&kqu.kq->kq_lock); } static inline void -kqlock_held(__assert_only struct kqueue *kq) +kqlock_held(__assert_only kqueue_t kqu) { - LCK_SPIN_ASSERT(&kq->kq_lock, LCK_ASSERT_OWNED); + LCK_SPIN_ASSERT(&kqu.kq->kq_lock, LCK_ASSERT_OWNED); } static inline void -kqunlock(struct kqueue *kq) +kqunlock(kqueue_t kqu) { - lck_spin_unlock(&kq->kq_lock); + lck_spin_unlock(&kqu.kq->kq_lock); } static inline void -knhash_lock(proc_t p) +knhash_lock(struct filedesc *fdp) { - lck_mtx_lock(&p->p_fd->fd_knhashlock); + lck_mtx_lock(&fdp->fd_knhashlock); } static inline void -knhash_unlock(proc_t p) +knhash_unlock(struct filedesc *fdp) { - lck_mtx_unlock(&p->p_fd->fd_knhashlock); + lck_mtx_unlock(&fdp->fd_knhashlock); } - -/* - * Convert a kq lock to a knote use referece. - * - * If the knote is being dropped, or has - * vanished, we can't get a use reference. - * Just return with it still locked. - * - * - kq locked at entry - * - unlock on exit if we get the use reference - */ -static int -kqlock2knoteuse(struct kqueue *kq, struct knote *kn, int flags) +/* wait event for knote locks */ +static inline event_t +knote_lock_wev(struct knote *kn) { - if (kn->kn_status & (KN_DROPPING | KN_VANISHED)) - return (0); + return (event_t)(&kn->kn_hook); +} - assert(kn->kn_status & KN_ATTACHED); - kn->kn_inuse++; - if (flags & KNUSE_BOOST) { - set_thread_rwlock_boost(); - } - kqunlock(kq); - return (1); +/* wait event for kevent_register_wait_* */ +static inline event64_t +knote_filt_wev64(struct knote *kn) +{ + /* kdp_workloop_sync_wait_find_owner knows about this */ + return CAST_EVENT64_T(kn); } -/* - * - kq locked at entry - * - kq unlocked at exit - */ -__disable_tail_calls -static wait_result_t -knoteusewait(struct kqueue *kq, struct knote *kn) -{ - kn->kn_status |= KN_USEWAIT; - waitq_assert_wait64((struct waitq *)&kq->kq_wqs, - CAST_EVENT64_T(&kn->kn_status), - THREAD_UNINT, TIMEOUT_WAIT_FOREVER); - kqunlock(kq); - return thread_block(THREAD_CONTINUE_NULL); +/* wait event for knote_post/knote_drop */ +static inline event64_t +knote_post_wev64(struct knote *kn) +{ + return CAST_EVENT64_T(&kn->kn_kevent); } -static bool -knoteuse_needs_boost(struct knote *kn, struct kevent_internal_s *kev) +/*! + * @function knote_has_qos + * + * @brief + * Whether the knote has a regular QoS. + * + * @discussion + * kn_qos_override is: + * - 0 on kqfiles + * - THREAD_QOS_LAST for special buckets (stayactive, manager) + * + * Other values mean the knote participates to QoS propagation. + */ +static inline bool +knote_has_qos(struct knote *kn) { - if (knote_fops(kn)->f_needs_boost) { - return knote_fops(kn)->f_needs_boost(kev); - } - return false; + return kn->kn_qos_override > 0 && kn->kn_qos_override < THREAD_QOS_LAST; } +#pragma mark knote locks + /* - * Convert from a knote use reference back to kq lock. + * Enum used by the knote_lock_* functions. * - * Drop a use reference and wake any waiters if - * this is the last one. + * KNOTE_KQ_LOCK_ALWAYS + * The function will always return with the kq lock held. * - * If someone is trying to drop the knote, but the - * caller has events they must deliver, take - * responsibility for the drop later - and wake the - * other attempted dropper in a manner that informs - * him of the transfer of responsibility. + * KNOTE_KQ_LOCK_ON_SUCCESS + * The function will return with the kq lock held if it was successful + * (knote_lock() is the only function that can fail). * - * The exit return indicates if the knote is still alive - * (or if not, the other dropper has been given the green - * light to drop it). + * KNOTE_KQ_LOCK_ON_FAILURE + * The function will return with the kq lock held if it was unsuccessful + * (knote_lock() is the only function that can fail). * - * The kqueue lock is re-taken unconditionally. + * KNOTE_KQ_UNLOCK: + * The function returns with the kq unlocked. */ -static int -knoteuse2kqlock(struct kqueue *kq, struct knote *kn, int flags) -{ - int dropped = 0; - int steal_drop = (flags & KNUSE_STEAL_DROP); +enum kqlocking { + KNOTE_KQ_LOCK_ALWAYS, + KNOTE_KQ_LOCK_ON_SUCCESS, + KNOTE_KQ_LOCK_ON_FAILURE, + KNOTE_KQ_UNLOCK, +}; - kqlock(kq); - if (flags & KNUSE_BOOST) { - clear_thread_rwlock_boost(); +static struct knote_lock_ctx * +knote_lock_ctx_find(kqueue_t kqu, struct knote *kn) +{ + struct knote_lock_ctx *ctx; + LIST_FOREACH(ctx, &kqu.kq->kq_knlocks, knlc_link) { + if (ctx->knlc_knote == kn) { + return ctx; + } } + panic("knote lock context not found: %p", kn); + __builtin_trap(); +} - if (--kn->kn_inuse == 0) { - - if ((kn->kn_status & KN_ATTACHING) != 0) { - kn->kn_status &= ~KN_ATTACHING; - } +/* slowpath of knote_lock() */ +__attribute__((noinline)) +static bool __result_use_check +knote_lock_slow(kqueue_t kqu, struct knote *kn, + struct knote_lock_ctx *knlc, int kqlocking) +{ + struct knote_lock_ctx *owner_lc; + struct uthread *uth = current_uthread(); + wait_result_t wr; - if ((kn->kn_status & KN_USEWAIT) != 0) { - wait_result_t result; + kqlock_held(kqu); - /* If we need to, try and steal the drop */ - if (kn->kn_status & KN_DROPPING) { - if (steal_drop && !(kn->kn_status & KN_STOLENDROP)) { - kn->kn_status |= KN_STOLENDROP; - } else { - dropped = 1; - } - } + owner_lc = knote_lock_ctx_find(kqu, kn); +#if DEBUG || DEVELOPMENT + knlc->knlc_state = KNOTE_LOCK_CTX_WAITING; +#endif + owner_lc->knlc_waiters++; - /* wakeup indicating if ANY USE stole the drop */ - result = (kn->kn_status & KN_STOLENDROP) ? - THREAD_RESTART : THREAD_AWAKENED; + /* + * Make our lock context visible to knote_unlock() + */ + uth->uu_knlock = knlc; - kn->kn_status &= ~KN_USEWAIT; - waitq_wakeup64_all((struct waitq *)&kq->kq_wqs, - CAST_EVENT64_T(&kn->kn_status), - result, - WAITQ_ALL_PRIORITIES); - } else { - /* should have seen use-wait if dropping with use refs */ - assert((kn->kn_status & (KN_DROPPING|KN_STOLENDROP)) == 0); - } + wr = lck_spin_sleep_with_inheritor(&kqu.kq->kq_lock, LCK_SLEEP_UNLOCK, + knote_lock_wev(kn), owner_lc->knlc_thread, + THREAD_UNINT | THREAD_WAIT_NOREPORT, TIMEOUT_WAIT_FOREVER); - } else if (kn->kn_status & KN_DROPPING) { - /* not the last ref but want to steal a drop if present */ - if (steal_drop && ((kn->kn_status & KN_STOLENDROP) == 0)) { - kn->kn_status |= KN_STOLENDROP; + if (wr == THREAD_RESTART) { + /* + * We haven't been woken up by knote_unlock() but knote_unlock_cancel. + * We need to cleanup the state since no one did. + */ + uth->uu_knlock = NULL; +#if DEBUG || DEVELOPMENT + assert(knlc->knlc_state == KNOTE_LOCK_CTX_WAITING); + knlc->knlc_state = KNOTE_LOCK_CTX_UNLOCKED; +#endif - /* but we now have to wait to be the last ref */ - knoteusewait(kq, kn); - kqlock(kq); - } else { - dropped = 1; + if (kqlocking == KNOTE_KQ_LOCK_ALWAYS || + kqlocking == KNOTE_KQ_LOCK_ON_FAILURE) { + kqlock(kqu); + } + return false; + } else { + if (kqlocking == KNOTE_KQ_LOCK_ALWAYS || + kqlocking == KNOTE_KQ_LOCK_ON_SUCCESS) { + kqlock(kqu); +#if DEBUG || DEVELOPMENT + /* + * This state is set under the lock so we can't + * really assert this unless we hold the lock. + */ + assert(knlc->knlc_state == KNOTE_LOCK_CTX_LOCKED); +#endif } + return true; } - - return (!dropped); } /* - * Convert a kq lock to a knote use reference - * (for the purpose of detaching AND vanishing it). - * - * If the knote is being dropped, we can't get - * a detach reference, so wait for the knote to - * finish dropping before returning. + * Attempts to take the "knote" lock. * - * If the knote is being used for other purposes, - * we cannot detach it until those uses are done - * as well. Again, just wait for them to finish - * (caller will start over at lookup). + * Called with the kqueue lock held. * - * - kq locked at entry - * - unlocked on exit + * Returns true if the knote lock is acquired, false if it has been dropped */ -static int -kqlock2knotedetach(struct kqueue *kq, struct knote *kn, int flags) +static bool __result_use_check +knote_lock(kqueue_t kqu, struct knote *kn, struct knote_lock_ctx *knlc, + enum kqlocking kqlocking) { - if ((kn->kn_status & KN_DROPPING) || kn->kn_inuse) { - /* have to wait for dropper or current uses to go away */ - knoteusewait(kq, kn); - return (0); + kqlock_held(kqu); + +#if DEBUG || DEVELOPMENT + assert(knlc->knlc_state == KNOTE_LOCK_CTX_UNLOCKED); +#endif + knlc->knlc_knote = kn; + knlc->knlc_thread = current_thread(); + knlc->knlc_waiters = 0; + + if (__improbable(kn->kn_status & KN_LOCKED)) { + return knote_lock_slow(kqu, kn, knlc, kqlocking); } - assert((kn->kn_status & KN_VANISHED) == 0); - assert(kn->kn_status & KN_ATTACHED); - kn->kn_status &= ~KN_ATTACHED; - kn->kn_status |= KN_VANISHED; - if (flags & KNUSE_BOOST) { - clear_thread_rwlock_boost(); + + /* + * When the knote will be dropped, the knote lock is taken before + * KN_DROPPING is set, and then the knote will be removed from any + * hash table that references it before the lock is canceled. + */ + assert((kn->kn_status & KN_DROPPING) == 0); + LIST_INSERT_HEAD(&kqu.kq->kq_knlocks, knlc, knlc_link); + kn->kn_status |= KN_LOCKED; +#if DEBUG || DEVELOPMENT + knlc->knlc_state = KNOTE_LOCK_CTX_LOCKED; +#endif + + if (kqlocking == KNOTE_KQ_UNLOCK || + kqlocking == KNOTE_KQ_LOCK_ON_FAILURE) { + kqunlock(kqu); } - kn->kn_inuse++; - kqunlock(kq); - return (1); + return true; } /* - * Convert a kq lock to a knote drop reference. + * Unlocks a knote successfully locked with knote_lock(). * - * If the knote is in use, wait for the use count - * to subside. We first mark our intention to drop - * it - keeping other users from "piling on." - * If we are too late, we have to wait for the - * other drop to complete. + * Called with the kqueue lock held. * - * - kq locked at entry - * - always unlocked on exit. - * - caller can't hold any locks that would prevent - * the other dropper from completing. + * Returns with the kqueue lock held according to KNOTE_KQ_* mode. */ -static int -kqlock2knotedrop(struct kqueue *kq, struct knote *kn) +static void +knote_unlock(kqueue_t kqu, struct knote *kn, + struct knote_lock_ctx *knlc, enum kqlocking kqlocking) { - int oktodrop; - wait_result_t result; + kqlock_held(kqu); - oktodrop = ((kn->kn_status & (KN_DROPPING | KN_ATTACHING)) == 0); - /* if another thread is attaching, they will become the dropping thread */ - kn->kn_status |= KN_DROPPING; - knote_unsuppress(kn); - knote_dequeue(kn); - if (oktodrop) { - if (kn->kn_inuse == 0) { - kqunlock(kq); - return (oktodrop); - } + assert(knlc->knlc_knote == kn); + assert(kn->kn_status & KN_LOCKED); +#if DEBUG || DEVELOPMENT + assert(knlc->knlc_state == KNOTE_LOCK_CTX_LOCKED); +#endif + + LIST_REMOVE(knlc, knlc_link); + + if (knlc->knlc_waiters) { + thread_t thread = THREAD_NULL; + + wakeup_one_with_inheritor(knote_lock_wev(kn), THREAD_AWAKENED, + LCK_WAKE_DEFAULT, &thread); + + /* + * knote_lock_slow() publishes the lock context of waiters + * in uthread::uu_knlock. + * + * Reach out and make this context the new owner. + */ + struct uthread *ut = get_bsdthread_info(thread); + struct knote_lock_ctx *next_owner_lc = ut->uu_knlock; + + assert(next_owner_lc->knlc_knote == kn); + next_owner_lc->knlc_waiters = knlc->knlc_waiters - 1; + LIST_INSERT_HEAD(&kqu.kq->kq_knlocks, next_owner_lc, knlc_link); +#if DEBUG || DEVELOPMENT + next_owner_lc->knlc_state = KNOTE_LOCK_CTX_LOCKED; +#endif + ut->uu_knlock = NULL; + thread_deallocate_safe(thread); + } else { + kn->kn_status &= ~KN_LOCKED; + } + + if ((kn->kn_status & KN_MERGE_QOS) && !(kn->kn_status & KN_POSTING)) { + /* + * No f_event() in flight anymore, we can leave QoS "Merge" mode + * + * See knote_adjust_qos() + */ + kn->kn_status &= ~KN_MERGE_QOS; + } + if (kqlocking == KNOTE_KQ_UNLOCK) { + kqunlock(kqu); } - result = knoteusewait(kq, kn); - /* THREAD_RESTART == another thread stole the knote drop */ - return (result == THREAD_AWAKENED); +#if DEBUG || DEVELOPMENT + knlc->knlc_state = KNOTE_LOCK_CTX_UNLOCKED; +#endif } -#if 0 /* - * Release a knote use count reference. + * Aborts all waiters for a knote lock, and unlock the knote. + * + * Called with the kqueue lock held. + * + * Returns with the kqueue unlocked. */ static void -knote_put(struct knote *kn) +knote_unlock_cancel(struct kqueue *kq, struct knote *kn, + struct knote_lock_ctx *knlc) { - struct kqueue *kq = knote_get_kq(kn); + kqlock_held(kq); - kqlock(kq); - if (--kn->kn_inuse == 0) { - if ((kn->kn_status & KN_USEWAIT) != 0) { - kn->kn_status &= ~KN_USEWAIT; - waitq_wakeup64_all((struct waitq *)&kq->kq_wqs, - CAST_EVENT64_T(&kn->kn_status), - THREAD_AWAKENED, - WAITQ_ALL_PRIORITIES); - } - } + assert(knlc->knlc_knote == kn); + assert(kn->kn_status & KN_LOCKED); + assert(kn->kn_status & KN_DROPPING); + + LIST_REMOVE(knlc, knlc_link); + kn->kn_status &= ~KN_LOCKED; kqunlock(kq); -} -#endif -static int -filt_fileattach(struct knote *kn, struct kevent_internal_s *kev) -{ - return (fo_kqfilter(kn->kn_fp, kn, kev, vfs_context_current())); + if (knlc->knlc_waiters) { + wakeup_all_with_inheritor(knote_lock_wev(kn), THREAD_RESTART); + } +#if DEBUG || DEVELOPMENT + knlc->knlc_state = KNOTE_LOCK_CTX_UNLOCKED; +#endif } -#define f_flag f_fglob->fg_flag -#define f_msgcount f_fglob->fg_msgcount -#define f_cred f_fglob->fg_cred -#define f_ops f_fglob->fg_ops -#define f_offset f_fglob->fg_offset -#define f_data f_fglob->fg_data - +/* + * Call the f_event hook of a given filter. + * + * Takes a use count to protect against concurrent drops. + */ static void -filt_kqdetach(struct knote *kn) +knote_post(struct knote *kn, long hint) { - struct kqfile *kqf = (struct kqfile *)kn->kn_fp->f_data; - struct kqueue *kq = &kqf->kqf_kqueue; + struct kqueue *kq = knote_get_kq(kn); + int dropping, result; kqlock(kq); - KNOTE_DETACH(&kqf->kqf_sel.si_note, kn); - kqunlock(kq); -} -/*ARGSUSED*/ -static int -filt_kqueue(struct knote *kn, __unused long hint) -{ - struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; - int count; + if (__improbable(kn->kn_status & (KN_DROPPING | KN_VANISHED))) { + return kqunlock(kq); + } - count = kq->kq_count; - return (count > 0); -} + if (__improbable(kn->kn_status & KN_POSTING)) { + panic("KNOTE() called concurrently on knote %p", kn); + } -static int -filt_kqtouch(struct knote *kn, struct kevent_internal_s *kev) -{ -#pragma unused(kev) - struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; - int res; + kn->kn_status |= KN_POSTING; + kqunlock(kq); + result = filter_call(knote_fops(kn), f_event(kn, hint)); kqlock(kq); - kn->kn_data = kq->kq_count; - if ((kn->kn_status & KN_UDATA_SPECIFIC) == 0) - kn->kn_udata = kev->udata; - res = (kn->kn_data > 0); - kqunlock(kq); + dropping = (kn->kn_status & KN_DROPPING); - return res; -} + if (!dropping && (result & FILTER_ACTIVE)) { + knote_activate(kq, kn, result); + } -static int -filt_kqprocess(struct knote *kn, struct filt_process_s *data, struct kevent_internal_s *kev) -{ -#pragma unused(data) - struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; - int res; + if ((kn->kn_status & KN_LOCKED) == 0) { + /* + * There's no other f_* call in flight, we can leave QoS "Merge" mode. + * + * See knote_adjust_qos() + */ + kn->kn_status &= ~(KN_POSTING | KN_MERGE_QOS); + } else { + kn->kn_status &= ~KN_POSTING; + } - kqlock(kq); - kn->kn_data = kq->kq_count; - res = (kn->kn_data > 0); - if (res) { - *kev = kn->kn_kevent; - if (kn->kn_flags & EV_CLEAR) - kn->kn_data = 0; + if (__improbable(dropping)) { + waitq_wakeup64_all((struct waitq *)&kq->kq_wqs, knote_post_wev64(kn), + THREAD_AWAKENED, WAITQ_ALL_PRIORITIES); } - kqunlock(kq); - return res; + kqunlock(kq); } -#pragma mark EVFILT_PROC - -static int -filt_procattach(struct knote *kn, __unused struct kevent_internal_s *kev) +/* + * Called by knote_drop() to wait for the last f_event() caller to be done. + * + * - kq locked at entry + * - kq unlocked at exit + */ +static void +knote_wait_for_post(struct kqueue *kq, struct knote *kn) { - struct proc *p; + wait_result_t wr = THREAD_NOT_WAITING; - assert(PID_MAX < NOTE_PDATAMASK); + kqlock_held(kq); - if ((kn->kn_sfflags & (NOTE_TRACK | NOTE_TRACKERR | NOTE_CHILD)) != 0) { - kn->kn_flags = EV_ERROR; - kn->kn_data = ENOTSUP; - return 0; - } + assert(kn->kn_status & KN_DROPPING); - p = proc_find(kn->kn_id); - if (p == NULL) { - kn->kn_flags = EV_ERROR; - kn->kn_data = ESRCH; - return 0; + if (kn->kn_status & KN_POSTING) { + wr = waitq_assert_wait64((struct waitq *)&kq->kq_wqs, + knote_post_wev64(kn), THREAD_UNINT | THREAD_WAIT_NOREPORT, + TIMEOUT_WAIT_FOREVER); + } + kqunlock(kq); + if (wr == THREAD_WAITING) { + thread_block(THREAD_CONTINUE_NULL); + } +} + +#pragma mark knote helpers for filters + +OS_ALWAYS_INLINE +void +knote_set_error(struct knote *kn, int error) +{ + kn->kn_flags |= EV_ERROR; + kn->kn_sdata = error; +} + +OS_ALWAYS_INLINE +int64_t +knote_low_watermark(const struct knote *kn) +{ + return (kn->kn_sfflags & NOTE_LOWAT) ? kn->kn_sdata : 1; +} + +/*! + * @function knote_fill_kevent_with_sdata + * + * @brief + * Fills in a kevent from the current content of a knote. + * + * @discussion + * This is meant to be called from filter's f_event hooks. + * The kevent data is filled with kn->kn_sdata. + * + * kn->kn_fflags is cleared if kn->kn_flags has EV_CLEAR set. + * + * Using knote_fill_kevent is typically preferred. + */ +OS_ALWAYS_INLINE +void +knote_fill_kevent_with_sdata(struct knote *kn, struct kevent_qos_s *kev) +{ +#define knote_assert_aliases(name1, offs1, name2) \ + static_assert(offsetof(struct kevent_qos_s, name1) + offs1 == \ + offsetof(struct kevent_internal_s, name2), \ + "kevent_qos_s::" #name1 " and kevent_internal_s::" #name2 "need to alias") + /* + * All the code makes assumptions on these aliasing, + * so make sure we fail the build if we ever ever ever break them. + */ + knote_assert_aliases(ident, 0, kei_ident); +#ifdef __LITTLE_ENDIAN__ + knote_assert_aliases(filter, 0, kei_filter); // non trivial overlap + knote_assert_aliases(filter, 1, kei_filtid); // non trivial overlap +#else + knote_assert_aliases(filter, 0, kei_filtid); // non trivial overlap + knote_assert_aliases(filter, 1, kei_filter); // non trivial overlap +#endif + knote_assert_aliases(flags, 0, kei_flags); + knote_assert_aliases(qos, 0, kei_qos); + knote_assert_aliases(udata, 0, kei_udata); + knote_assert_aliases(fflags, 0, kei_fflags); + knote_assert_aliases(xflags, 0, kei_sfflags); // non trivial overlap + knote_assert_aliases(data, 0, kei_sdata); // non trivial overlap + knote_assert_aliases(ext, 0, kei_ext); +#undef knote_assert_aliases + + /* + * Fix the differences between kevent_qos_s and kevent_internal_s: + * - xflags is where kn_sfflags lives, we need to zero it + * - fixup the high bits of `filter` where kn_filtid lives + */ + *kev = *(struct kevent_qos_s *)&kn->kn_kevent; + kev->xflags = 0; + kev->filter |= 0xff00; + if (kn->kn_flags & EV_CLEAR) { + kn->kn_fflags = 0; + } +} + +/*! + * @function knote_fill_kevent + * + * @brief + * Fills in a kevent from the current content of a knote. + * + * @discussion + * This is meant to be called from filter's f_event hooks. + * The kevent data is filled with the passed in data. + * + * kn->kn_fflags is cleared if kn->kn_flags has EV_CLEAR set. + */ +OS_ALWAYS_INLINE +void +knote_fill_kevent(struct knote *kn, struct kevent_qos_s *kev, int64_t data) +{ + knote_fill_kevent_with_sdata(kn, kev); + kev->filter = kn->kn_filter; + kev->data = data; +} + + +#pragma mark file_filtops + +static int +filt_fileattach(struct knote *kn, struct kevent_qos_s *kev) +{ + return fo_kqfilter(kn->kn_fp, kn, kev); +} + +SECURITY_READ_ONLY_EARLY(static struct filterops) file_filtops = { + .f_isfd = 1, + .f_attach = filt_fileattach, +}; + +#pragma mark kqread_filtops + +#define f_flag fp_glob->fg_flag +#define f_ops fp_glob->fg_ops +#define f_data fp_glob->fg_data +#define f_lflags fp_glob->fg_lflags + +static void +filt_kqdetach(struct knote *kn) +{ + struct kqfile *kqf = (struct kqfile *)kn->kn_fp->f_data; + struct kqueue *kq = &kqf->kqf_kqueue; + + kqlock(kq); + KNOTE_DETACH(&kqf->kqf_sel.si_note, kn); + kqunlock(kq); +} + +static int +filt_kqueue(struct knote *kn, __unused long hint) +{ + struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; + + return kq->kq_count > 0; +} + +static int +filt_kqtouch(struct knote *kn, struct kevent_qos_s *kev) +{ +#pragma unused(kev) + struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; + int res; + + kqlock(kq); + res = (kq->kq_count > 0); + kqunlock(kq); + + return res; +} + +static int +filt_kqprocess(struct knote *kn, struct kevent_qos_s *kev) +{ + struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; + int res = 0; + + kqlock(kq); + if (kq->kq_count) { + knote_fill_kevent(kn, kev, kq->kq_count); + res = 1; + } + kqunlock(kq); + + return res; +} + +SECURITY_READ_ONLY_EARLY(static struct filterops) kqread_filtops = { + .f_isfd = 1, + .f_detach = filt_kqdetach, + .f_event = filt_kqueue, + .f_touch = filt_kqtouch, + .f_process = filt_kqprocess, +}; + +#pragma mark proc_filtops + +static int +filt_procattach(struct knote *kn, __unused struct kevent_qos_s *kev) +{ + struct proc *p; + + assert(PID_MAX < NOTE_PDATAMASK); + + if ((kn->kn_sfflags & (NOTE_TRACK | NOTE_TRACKERR | NOTE_CHILD)) != 0) { + knote_set_error(kn, ENOTSUP); + return 0; + } + + p = proc_find((int)kn->kn_id); + if (p == NULL) { + knote_set_error(kn, ESRCH); + return 0; } - const int NoteExitStatusBits = NOTE_EXIT | NOTE_EXITSTATUS; + const uint32_t NoteExitStatusBits = NOTE_EXIT | NOTE_EXITSTATUS; - if ((kn->kn_sfflags & NoteExitStatusBits) == NoteExitStatusBits) + if ((kn->kn_sfflags & NoteExitStatusBits) == NoteExitStatusBits) { do { pid_t selfpid = proc_selfpid(); - if (p->p_ppid == selfpid) - break; /* parent => ok */ - + if (p->p_ppid == selfpid) { + break; /* parent => ok */ + } if ((p->p_lflag & P_LTRACED) != 0 && - (p->p_oppid == selfpid)) - break; /* parent-in-waiting => ok */ - + (p->p_oppid == selfpid)) { + break; /* parent-in-waiting => ok */ + } + if (cansignal(current_proc(), kauth_cred_get(), p, SIGKILL)) { + break; /* allowed to signal => ok */ + } proc_rele(p); - kn->kn_flags = EV_ERROR; - kn->kn_data = EACCES; + knote_set_error(kn, EACCES); return 0; } while (0); + } - proc_klist_lock(); + kn->kn_proc = p; + kn->kn_flags |= EV_CLEAR; /* automatically set */ + kn->kn_sdata = 0; /* incoming data is ignored */ - kn->kn_ptr.p_proc = p; /* store the proc handle */ + proc_klist_lock(); KNOTE_ATTACH(&p->p_klist, kn); @@ -963,7 +1044,7 @@ filt_procattach(struct knote *kn, __unused struct kevent_internal_s *kev) * only captures edge-triggered events after this point * so it can't already be fired. */ - return (0); + return 0; } @@ -979,9 +1060,9 @@ filt_procdetach(struct knote *kn) proc_klist_lock(); - p = kn->kn_ptr.p_proc; + p = kn->kn_proc; if (p != PROC_NULL) { - kn->kn_ptr.p_proc = PROC_NULL; + kn->kn_proc = PROC_NULL; KNOTE_DETACH(&p->p_klist, kn); } @@ -989,7 +1070,7 @@ filt_procdetach(struct knote *kn) } static int -filt_proc(struct knote *kn, long hint) +filt_procevent(struct knote *kn, long hint) { u_int event; @@ -998,7 +1079,7 @@ filt_proc(struct knote *kn, long hint) /* * Note: a lot of bits in hint may be obtained from the knote * To free some of those bits, see Freeing up - * bits in hint for filt_proc + * bits in hint for filt_procevent * * mask off extra data */ @@ -1013,20 +1094,21 @@ filt_proc(struct knote *kn, long hint) * parent and these knotes re-fired. */ if (event & NOTE_EXIT) { - if ((kn->kn_ptr.p_proc->p_oppid != 0) - && (knote_get_kq(kn)->kq_p->p_pid != kn->kn_ptr.p_proc->p_ppid)) { + if ((kn->kn_proc->p_oppid != 0) + && (knote_get_kq(kn)->kq_p->p_pid != kn->kn_proc->p_ppid)) { /* * This knote is not for the current ptrace(2) parent, ignore. */ return 0; } - } + } /* * if the user is interested in this event, record it. */ - if (kn->kn_sfflags & event) + if (kn->kn_sfflags & event) { kn->kn_fflags |= event; + } #pragma clang diagnostic push #pragma clang diagnostic ignored "-Wdeprecated-declarations" @@ -1038,62 +1120,62 @@ filt_proc(struct knote *kn, long hint) /* * The kernel has a wrapper in place that returns the same data - * as is collected here, in kn_data. Any changes to how + * as is collected here, in kn_hook32. Any changes to how * NOTE_EXITSTATUS and NOTE_EXIT_DETAIL are collected * should also be reflected in the proc_pidnoteexit() wrapper. */ if (event == NOTE_EXIT) { - kn->kn_data = 0; + kn->kn_hook32 = 0; if ((kn->kn_sfflags & NOTE_EXITSTATUS) != 0) { kn->kn_fflags |= NOTE_EXITSTATUS; - kn->kn_data |= (hint & NOTE_PDATAMASK); + kn->kn_hook32 |= (hint & NOTE_PDATAMASK); } if ((kn->kn_sfflags & NOTE_EXIT_DETAIL) != 0) { kn->kn_fflags |= NOTE_EXIT_DETAIL; - if ((kn->kn_ptr.p_proc->p_lflag & - P_LTERM_DECRYPTFAIL) != 0) { - kn->kn_data |= NOTE_EXIT_DECRYPTFAIL; + if ((kn->kn_proc->p_lflag & + P_LTERM_DECRYPTFAIL) != 0) { + kn->kn_hook32 |= NOTE_EXIT_DECRYPTFAIL; } - if ((kn->kn_ptr.p_proc->p_lflag & - P_LTERM_JETSAM) != 0) { - kn->kn_data |= NOTE_EXIT_MEMORY; - switch (kn->kn_ptr.p_proc->p_lflag & P_JETSAM_MASK) { + if ((kn->kn_proc->p_lflag & + P_LTERM_JETSAM) != 0) { + kn->kn_hook32 |= NOTE_EXIT_MEMORY; + switch (kn->kn_proc->p_lflag & P_JETSAM_MASK) { case P_JETSAM_VMPAGESHORTAGE: - kn->kn_data |= NOTE_EXIT_MEMORY_VMPAGESHORTAGE; + kn->kn_hook32 |= NOTE_EXIT_MEMORY_VMPAGESHORTAGE; break; case P_JETSAM_VMTHRASHING: - kn->kn_data |= NOTE_EXIT_MEMORY_VMTHRASHING; + kn->kn_hook32 |= NOTE_EXIT_MEMORY_VMTHRASHING; break; case P_JETSAM_FCTHRASHING: - kn->kn_data |= NOTE_EXIT_MEMORY_FCTHRASHING; + kn->kn_hook32 |= NOTE_EXIT_MEMORY_FCTHRASHING; break; case P_JETSAM_VNODE: - kn->kn_data |= NOTE_EXIT_MEMORY_VNODE; + kn->kn_hook32 |= NOTE_EXIT_MEMORY_VNODE; break; case P_JETSAM_HIWAT: - kn->kn_data |= NOTE_EXIT_MEMORY_HIWAT; + kn->kn_hook32 |= NOTE_EXIT_MEMORY_HIWAT; break; case P_JETSAM_PID: - kn->kn_data |= NOTE_EXIT_MEMORY_PID; + kn->kn_hook32 |= NOTE_EXIT_MEMORY_PID; break; case P_JETSAM_IDLEEXIT: - kn->kn_data |= NOTE_EXIT_MEMORY_IDLE; + kn->kn_hook32 |= NOTE_EXIT_MEMORY_IDLE; break; } } - if ((kn->kn_ptr.p_proc->p_csflags & - CS_KILLED) != 0) { - kn->kn_data |= NOTE_EXIT_CSERROR; + if ((kn->kn_proc->p_csflags & + CS_KILLED) != 0) { + kn->kn_hook32 |= NOTE_EXIT_CSERROR; } } } /* if we have any matching state, activate the knote */ - return (kn->kn_fflags != 0); + return kn->kn_fflags != 0; } static int -filt_proctouch(struct knote *kn, struct kevent_internal_s *kev) +filt_proctouch(struct knote *kn, struct kevent_qos_s *kev) { int res; @@ -1101,8 +1183,6 @@ filt_proctouch(struct knote *kn, struct kevent_internal_s *kev) /* accept new filter flags and mask off output events no long interesting */ kn->kn_sfflags = kev->fflags; - if ((kn->kn_status & KN_UDATA_SPECIFIC) == 0) - kn->kn_udata = kev->udata; /* restrict the current results to the (smaller?) set of new interest */ /* @@ -1119,52 +1199,77 @@ filt_proctouch(struct knote *kn, struct kevent_internal_s *kev) } static int -filt_procprocess(struct knote *kn, struct filt_process_s *data, struct kevent_internal_s *kev) +filt_procprocess(struct knote *kn, struct kevent_qos_s *kev) { -#pragma unused(data) - int res; + int res = 0; proc_klist_lock(); - res = (kn->kn_fflags != 0); - if (res) { - *kev = kn->kn_kevent; - kn->kn_flags |= EV_CLEAR; /* automatically set */ - kn->kn_fflags = 0; - kn->kn_data = 0; + if (kn->kn_fflags) { + knote_fill_kevent(kn, kev, kn->kn_hook32); + kn->kn_hook32 = 0; + res = 1; } proc_klist_unlock(); return res; } +SECURITY_READ_ONLY_EARLY(static struct filterops) proc_filtops = { + .f_attach = filt_procattach, + .f_detach = filt_procdetach, + .f_event = filt_procevent, + .f_touch = filt_proctouch, + .f_process = filt_procprocess, +}; -#pragma mark EVFILT_TIMER +#pragma mark timer_filtops +struct filt_timer_params { + uint64_t deadline; /* deadline in abs/cont time + * (or 0 if NOTE_ABSOLUTE and deadline is in past) */ + uint64_t leeway; /* leeway in abstime, or 0 if none */ + uint64_t interval; /* interval in abstime or 0 if non-repeating timer */ +}; /* * Values stored in the knote at rest (using Mach absolute time units) * - * kn->kn_hook where the thread_call object is stored + * kn->kn_thcall where the thread_call object is stored * kn->kn_ext[0] next deadline or 0 if immediate expiration * kn->kn_ext[1] leeway value * kn->kn_sdata interval timer: the interval * absolute/deadline timer: 0 - * kn->kn_data fire count + * kn->kn_hook32 timer state (with gencount) + * + * TIMER_IDLE: + * The timer has either never been scheduled or been cancelled. + * It is safe to schedule a new one in this state. + * + * TIMER_ARMED: + * The timer has been scheduled + * + * TIMER_FIRED + * The timer has fired and an event needs to be delivered. + * When in this state, the callout may still be running. + * + * TIMER_IMMEDIATE + * The timer has fired at registration time, and the callout was never + * dispatched. */ +#define TIMER_IDLE 0x0 +#define TIMER_ARMED 0x1 +#define TIMER_FIRED 0x2 +#define TIMER_IMMEDIATE 0x3 +#define TIMER_STATE_MASK 0x3 +#define TIMER_GEN_INC 0x4 -static lck_mtx_t _filt_timerlock; - -static void filt_timerlock(void) { lck_mtx_lock(&_filt_timerlock); } -static void filt_timerunlock(void) { lck_mtx_unlock(&_filt_timerlock); } - -static inline void filt_timer_assert_locked(void) +static void +filt_timer_set_params(struct knote *kn, struct filt_timer_params *params) { - LCK_MTX_ASSERT(&_filt_timerlock, LCK_MTX_ASSERT_OWNED); + kn->kn_ext[0] = params->deadline; + kn->kn_ext[1] = params->leeway; + kn->kn_sdata = params->interval; } -/* state flags stored in kn_hookid */ -#define TIMER_RUNNING 0x1 -#define TIMER_CANCELWAIT 0x2 - /* * filt_timervalidate - process data from user * @@ -1175,20 +1280,21 @@ static inline void filt_timer_assert_locked(void) * kn_sfflags style of timer, unit of measurement * * Output: - * kn_sdata either interval in abstime or 0 if non-repeating timer - * ext[0] fire deadline in abs/cont time - * (or 0 if NOTE_ABSOLUTE and deadline is in past) + * struct filter_timer_params to apply to the filter with + * filt_timer_set_params when changes are ready to be commited. * * Returns: * EINVAL Invalid user data parameters + * ERANGE Various overflows with the parameters * * Called with timer filter lock held. */ static int -filt_timervalidate(struct knote *kn) +filt_timervalidate(const struct kevent_qos_s *kev, + struct filt_timer_params *params) { /* - * There are 4 knobs that need to be chosen for a timer registration: + * There are 5 knobs that need to be chosen for a timer registration: * * A) Units of time (what is the time duration of the specified number) * Absolute and interval take: @@ -1218,13 +1324,11 @@ filt_timervalidate(struct knote *kn) * expires when mach_continuous_time() is > the passed in value. */ - filt_timer_assert_locked(); - uint64_t multiplier; boolean_t use_abstime = FALSE; - switch (kn->kn_sfflags & (NOTE_SECONDS|NOTE_USECONDS|NOTE_NSECONDS|NOTE_MACHTIME)) { + switch (kev->fflags & (NOTE_SECONDS | NOTE_USECONDS | NOTE_NSECONDS | NOTE_MACHTIME)) { case NOTE_SECONDS: multiplier = NSEC_PER_SEC; break; @@ -1242,36 +1346,40 @@ filt_timervalidate(struct knote *kn) multiplier = NSEC_PER_SEC / 1000; break; default: - return (EINVAL); + return EINVAL; } /* transform the leeway in kn_ext[1] to same time scale */ - if (kn->kn_sfflags & NOTE_LEEWAY) { + if (kev->fflags & NOTE_LEEWAY) { uint64_t leeway_abs; if (use_abstime) { - leeway_abs = (uint64_t)kn->kn_ext[1]; - } else { + leeway_abs = (uint64_t)kev->ext[1]; + } else { uint64_t leeway_ns; - if (os_mul_overflow((uint64_t)kn->kn_ext[1], multiplier, &leeway_ns)) - return (ERANGE); + if (os_mul_overflow((uint64_t)kev->ext[1], multiplier, &leeway_ns)) { + return ERANGE; + } nanoseconds_to_absolutetime(leeway_ns, &leeway_abs); } - kn->kn_ext[1] = leeway_abs; + params->leeway = leeway_abs; + } else { + params->leeway = 0; } - if (kn->kn_sfflags & NOTE_ABSOLUTE) { + if (kev->fflags & NOTE_ABSOLUTE) { uint64_t deadline_abs; if (use_abstime) { - deadline_abs = (uint64_t)kn->kn_sdata; + deadline_abs = (uint64_t)kev->data; } else { uint64_t calendar_deadline_ns; - if (os_mul_overflow((uint64_t)kn->kn_sdata, multiplier, &calendar_deadline_ns)) - return (ERANGE); + if (os_mul_overflow((uint64_t)kev->data, multiplier, &calendar_deadline_ns)) { + return ERANGE; + } /* calendar_deadline_ns is in nanoseconds since the epoch */ @@ -1304,20 +1412,21 @@ filt_timervalidate(struct knote *kn) * it does not change the calendar timebase. */ - if (kn->kn_sfflags & NOTE_MACH_CONTINUOUS_TIME) + if (kev->fflags & NOTE_MACH_CONTINUOUS_TIME) { clock_continuoustime_interval_to_deadline(interval_abs, - &deadline_abs); - else + &deadline_abs); + } else { clock_absolutetime_interval_to_deadline(interval_abs, - &deadline_abs); + &deadline_abs); + } } else { deadline_abs = 0; /* cause immediate expiration */ } } - kn->kn_ext[0] = deadline_abs; - kn->kn_sdata = 0; /* NOTE_ABSOLUTE is non-repeating */ - } else if (kn->kn_sdata < 0) { + params->deadline = deadline_abs; + params->interval = 0; /* NOTE_ABSOLUTE is non-repeating */ + } else if (kev->data < 0) { /* * Negative interval timers fire immediately, once. * @@ -1331,226 +1440,208 @@ filt_timervalidate(struct knote *kn) * We now skip the power-wasting hot spin phase and go straight to the idle phase. */ - kn->kn_sdata = 0; /* non-repeating */ - kn->kn_ext[0] = 0; /* expire immediately */ + params->deadline = 0; /* expire immediately */ + params->interval = 0; /* non-repeating */ } else { uint64_t interval_abs = 0; if (use_abstime) { - interval_abs = (uint64_t)kn->kn_sdata; + interval_abs = (uint64_t)kev->data; } else { uint64_t interval_ns; - if (os_mul_overflow((uint64_t)kn->kn_sdata, multiplier, &interval_ns)) - return (ERANGE); + if (os_mul_overflow((uint64_t)kev->data, multiplier, &interval_ns)) { + return ERANGE; + } nanoseconds_to_absolutetime(interval_ns, &interval_abs); } uint64_t deadline = 0; - if (kn->kn_sfflags & NOTE_MACH_CONTINUOUS_TIME) + if (kev->fflags & NOTE_MACH_CONTINUOUS_TIME) { clock_continuoustime_interval_to_deadline(interval_abs, &deadline); - else + } else { clock_absolutetime_interval_to_deadline(interval_abs, &deadline); + } - kn->kn_sdata = interval_abs; /* default to a repeating timer */ - kn->kn_ext[0] = deadline; + params->deadline = deadline; + params->interval = interval_abs; } - return (0); + return 0; } - - - /* * filt_timerexpire - the timer callout routine - * - * Just propagate the timer event into the knote - * filter routine (by going through the knote - * synchronization point). Pass a hint to - * indicate this is a real event, not just a - * query from above. */ static void -filt_timerexpire(void *knx, __unused void *spare) +filt_timerexpire(void *knx, void *state_on_arm) { - struct klist timer_list; struct knote *kn = knx; - filt_timerlock(); - - kn->kn_hookid &= ~TIMER_RUNNING; - - /* no "object" for timers, so fake a list */ - SLIST_INIT(&timer_list); - SLIST_INSERT_HEAD(&timer_list, kn, kn_selnext); - - KNOTE(&timer_list, 1); + uint32_t state = (uint32_t)(uintptr_t)state_on_arm; + uint32_t fired_state = state ^ TIMER_ARMED ^ TIMER_FIRED; - /* if someone is waiting for timer to pop */ - if (kn->kn_hookid & TIMER_CANCELWAIT) { + if (os_atomic_cmpxchg(&kn->kn_hook32, state, fired_state, relaxed)) { + // our f_event always would say FILTER_ACTIVE, + // so be leaner and just do it. struct kqueue *kq = knote_get_kq(kn); - waitq_wakeup64_all((struct waitq *)&kq->kq_wqs, - CAST_EVENT64_T(&kn->kn_hook), - THREAD_AWAKENED, - WAITQ_ALL_PRIORITIES); - - kn->kn_hookid &= ~TIMER_CANCELWAIT; + kqlock(kq); + knote_activate(kq, kn, FILTER_ACTIVE); + kqunlock(kq); + } else { + /* + * The timer has been reprogrammed or canceled since it was armed, + * and this is a late firing for the timer, just ignore it. + */ } - - filt_timerunlock(); } /* - * Cancel a running timer (or wait for the pop). - * Timer filter lock is held. - * May drop and retake the timer filter lock. + * Does this deadline needs a timer armed for it, or has it expired? */ -static void -filt_timercancel(struct knote *kn) +static bool +filt_timer_is_ready(struct knote *kn) { - filt_timer_assert_locked(); - - assert((kn->kn_hookid & TIMER_CANCELWAIT) == 0); - - /* if no timer, then we're good */ - if ((kn->kn_hookid & TIMER_RUNNING) == 0) - return; - - thread_call_t callout = (thread_call_t)kn->kn_hook; + uint64_t now, deadline = kn->kn_ext[0]; - /* cancel the callout if we can */ - if (thread_call_cancel(callout)) { - kn->kn_hookid &= ~TIMER_RUNNING; - return; + if (deadline == 0) { + return true; } - /* cancel failed, we have to wait for the in-flight expire routine */ - - kn->kn_hookid |= TIMER_CANCELWAIT; - - struct kqueue *kq = knote_get_kq(kn); - - waitq_assert_wait64((struct waitq *)&kq->kq_wqs, - CAST_EVENT64_T(&kn->kn_hook), - THREAD_UNINT, TIMEOUT_WAIT_FOREVER); - - filt_timerunlock(); - thread_block(THREAD_CONTINUE_NULL); - filt_timerlock(); - - assert((kn->kn_hookid & TIMER_CANCELWAIT) == 0); - assert((kn->kn_hookid & TIMER_RUNNING) == 0); + if (kn->kn_sfflags & NOTE_MACH_CONTINUOUS_TIME) { + now = mach_continuous_time(); + } else { + now = mach_absolute_time(); + } + return deadline <= now; } +/* + * Arm a timer + * + * It is the responsibility of the caller to make sure the timer call + * has completed or been cancelled properly prior to arming it. + */ static void filt_timerarm(struct knote *kn) { - filt_timer_assert_locked(); - - assert((kn->kn_hookid & TIMER_RUNNING) == 0); - - thread_call_t callout = (thread_call_t)kn->kn_hook; - uint64_t deadline = kn->kn_ext[0]; uint64_t leeway = kn->kn_ext[1]; + uint32_t state; int filter_flags = kn->kn_sfflags; unsigned int timer_flags = 0; - if (filter_flags & NOTE_CRITICAL) + if (filter_flags & NOTE_CRITICAL) { timer_flags |= THREAD_CALL_DELAY_USER_CRITICAL; - else if (filter_flags & NOTE_BACKGROUND) + } else if (filter_flags & NOTE_BACKGROUND) { timer_flags |= THREAD_CALL_DELAY_USER_BACKGROUND; - else + } else { timer_flags |= THREAD_CALL_DELAY_USER_NORMAL; + } - if (filter_flags & NOTE_LEEWAY) + if (filter_flags & NOTE_LEEWAY) { timer_flags |= THREAD_CALL_DELAY_LEEWAY; + } - if (filter_flags & NOTE_MACH_CONTINUOUS_TIME) + if (filter_flags & NOTE_MACH_CONTINUOUS_TIME) { timer_flags |= THREAD_CALL_CONTINUOUS; + } - thread_call_enter_delayed_with_leeway(callout, NULL, - deadline, leeway, - timer_flags); + /* + * Move to ARMED. + * + * We increase the gencount, and setup the thread call with this expected + * state. It means that if there was a previous generation of the timer in + * flight that needs to be ignored, then 3 things are possible: + * + * - the timer fires first, filt_timerexpire() and sets the state to FIRED + * but we clobber it with ARMED and a new gencount. The knote will still + * be activated, but filt_timerprocess() which is serialized with this + * call will not see the FIRED bit set and will not deliver an event. + * + * - this code runs first, but filt_timerexpire() comes second. Because it + * knows an old gencount, it will debounce and not activate the knote. + * + * - filt_timerexpire() wasn't in flight yet, and thread_call_enter below + * will just cancel it properly. + * + * This is important as userspace expects to never be woken up for past + * timers after filt_timertouch ran. + */ + state = os_atomic_load(&kn->kn_hook32, relaxed); + state &= ~TIMER_STATE_MASK; + state += TIMER_GEN_INC + TIMER_ARMED; + os_atomic_store(&kn->kn_hook32, state, relaxed); - kn->kn_hookid |= TIMER_RUNNING; + thread_call_enter_delayed_with_leeway(kn->kn_thcall, + (void *)(uintptr_t)state, deadline, leeway, timer_flags); } /* - * Does this knote need a timer armed for it, or should it be ready immediately? + * Mark a timer as "already fired" when it is being reprogrammed + * + * If there is a timer in flight, this will do a best effort at canceling it, + * but will not wait. If the thread call was in flight, having set the + * TIMER_IMMEDIATE bit will debounce a filt_timerexpire() racing with this + * cancelation. */ -static boolean_t -filt_timer_is_ready(struct knote *kn) +static void +filt_timerfire_immediate(struct knote *kn) { - uint64_t now; - - if (kn->kn_sfflags & NOTE_MACH_CONTINUOUS_TIME) - now = mach_continuous_time(); - else - now = mach_absolute_time(); + uint32_t state; - uint64_t deadline = kn->kn_ext[0]; + static_assert(TIMER_IMMEDIATE == TIMER_STATE_MASK, + "validate that this atomic or will transition to IMMEDIATE"); + state = os_atomic_or_orig(&kn->kn_hook32, TIMER_IMMEDIATE, relaxed); - if (deadline < now) - return TRUE; - else - return FALSE; + if ((state & TIMER_STATE_MASK) == TIMER_ARMED) { + thread_call_cancel(kn->kn_thcall); + } } /* * Allocate a thread call for the knote's lifetime, and kick off the timer. */ static int -filt_timerattach(struct knote *kn, __unused struct kevent_internal_s *kev) +filt_timerattach(struct knote *kn, struct kevent_qos_s *kev) { thread_call_t callout; + struct filt_timer_params params; int error; - callout = thread_call_allocate_with_options(filt_timerexpire, - (thread_call_param_t)kn, THREAD_CALL_PRIORITY_HIGH, - THREAD_CALL_OPTIONS_ONCE); - - if (NULL == callout) { - kn->kn_flags = EV_ERROR; - kn->kn_data = ENOMEM; + if ((error = filt_timervalidate(kev, ¶ms)) != 0) { + knote_set_error(kn, error); return 0; } - filt_timerlock(); - - if ((error = filt_timervalidate(kn)) != 0) { - kn->kn_flags = EV_ERROR; - kn->kn_data = error; - filt_timerunlock(); + callout = thread_call_allocate_with_options(filt_timerexpire, + (thread_call_param_t)kn, THREAD_CALL_PRIORITY_HIGH, + THREAD_CALL_OPTIONS_ONCE); - __assert_only boolean_t freed = thread_call_free(callout); - assert(freed); + if (NULL == callout) { + knote_set_error(kn, ENOMEM); return 0; } - kn->kn_hook = (void*)callout; - kn->kn_hookid = 0; + filt_timer_set_params(kn, ¶ms); + kn->kn_thcall = callout; kn->kn_flags |= EV_CLEAR; + os_atomic_store(&kn->kn_hook32, TIMER_IDLE, relaxed); /* NOTE_ABSOLUTE implies EV_ONESHOT */ - if (kn->kn_sfflags & NOTE_ABSOLUTE) + if (kn->kn_sfflags & NOTE_ABSOLUTE) { kn->kn_flags |= EV_ONESHOT; + } - boolean_t timer_ready = FALSE; - - if ((timer_ready = filt_timer_is_ready(kn))) { - /* cause immediate expiration */ - kn->kn_data = 1; + if (filt_timer_is_ready(kn)) { + os_atomic_store(&kn->kn_hook32, TIMER_IMMEDIATE, relaxed); + return FILTER_ACTIVE; } else { filt_timerarm(kn); + return 0; } - - filt_timerunlock(); - - return timer_ready; } /* @@ -1559,34 +1650,17 @@ filt_timerattach(struct knote *kn, __unused struct kevent_internal_s *kev) static void filt_timerdetach(struct knote *kn) { - thread_call_t callout; + __assert_only boolean_t freed; - filt_timerlock(); - - callout = (thread_call_t)kn->kn_hook; - filt_timercancel(kn); - - filt_timerunlock(); - - __assert_only boolean_t freed = thread_call_free(callout); + /* + * Unconditionally cancel to make sure there can't be any filt_timerexpire() + * running anymore. + */ + thread_call_cancel_wait(kn->kn_thcall); + freed = thread_call_free(kn->kn_thcall); assert(freed); } -/* - * filt_timerevent - post events to a timer knote - * - * Called in the context of filt_timerexpire with - * the filt_timerlock held - */ -static int -filt_timerevent(struct knote *kn, __unused long hint) -{ - filt_timer_assert_locked(); - - kn->kn_data = 1; - return (1); -} - /* * filt_timertouch - update timer knote with new user input * @@ -1595,54 +1669,35 @@ filt_timerevent(struct knote *kn, __unused long hint) * pops have gone off (in kn_data). */ static int -filt_timertouch( - struct knote *kn, - struct kevent_internal_s *kev) +filt_timertouch(struct knote *kn, struct kevent_qos_s *kev) { + struct filt_timer_params params; + uint32_t changed_flags = (kn->kn_sfflags ^ kev->fflags); int error; - filt_timerlock(); - - /* - * cancel current call - drops and retakes lock - * TODO: not safe against concurrent touches? - */ - filt_timercancel(kn); + if (changed_flags & NOTE_ABSOLUTE) { + kev->flags |= EV_ERROR; + kev->data = EINVAL; + return 0; + } - /* clear if the timer had previously fired, the user no longer wants to see it */ - kn->kn_data = 0; + if ((error = filt_timervalidate(kev, ¶ms)) != 0) { + kev->flags |= EV_ERROR; + kev->data = error; + return 0; + } /* capture the new values used to compute deadline */ - kn->kn_sdata = kev->data; + filt_timer_set_params(kn, ¶ms); kn->kn_sfflags = kev->fflags; - kn->kn_ext[0] = kev->ext[0]; - kn->kn_ext[1] = kev->ext[1]; - if ((kn->kn_status & KN_UDATA_SPECIFIC) == 0) - kn->kn_udata = kev->udata; - - /* recalculate deadline */ - error = filt_timervalidate(kn); - if (error) { - /* no way to report error, so mark it in the knote */ - kn->kn_flags |= EV_ERROR; - kn->kn_data = error; - filt_timerunlock(); - return 1; + if (filt_timer_is_ready(kn)) { + filt_timerfire_immediate(kn); + return FILTER_ACTIVE | FILTER_UPDATE_REQ_QOS; + } else { + filt_timerarm(kn); + return FILTER_UPDATE_REQ_QOS; } - - boolean_t timer_ready = FALSE; - - if ((timer_ready = filt_timer_is_ready(kn))) { - /* cause immediate expiration */ - kn->kn_data = 1; - } else { - filt_timerarm(kn); - } - - filt_timerunlock(); - - return timer_ready; } /* @@ -1653,29 +1708,45 @@ filt_timertouch( * counters for the next time. */ static int -filt_timerprocess( - struct knote *kn, - __unused struct filt_process_s *data, - struct kevent_internal_s *kev) +filt_timerprocess(struct knote *kn, struct kevent_qos_s *kev) { - filt_timerlock(); + uint32_t state = os_atomic_load(&kn->kn_hook32, relaxed); - if (kn->kn_data == 0 || (kn->kn_hookid & TIMER_CANCELWAIT)) { + /* + * filt_timerprocess is serialized with any filter routine except for + * filt_timerexpire which atomically does a TIMER_ARMED -> TIMER_FIRED + * transition, and on success, activates the knote. + * + * Hence, we don't need atomic modifications of the state, only to peek at + * whether we see any of the "FIRED" state, and if we do, it is safe to + * do simple state machine transitions. + */ + switch (state & TIMER_STATE_MASK) { + case TIMER_IDLE: + case TIMER_ARMED: /* - * kn_data = 0: - * The timer hasn't yet fired, so there's nothing to deliver - * TIMER_CANCELWAIT: - * touch is in the middle of canceling the timer, - * so don't deliver or re-arm anything - * * This can happen if a touch resets a timer that had fired * without being processed */ - filt_timerunlock(); return 0; } - if (kn->kn_sdata != 0 && ((kn->kn_flags & EV_ERROR) == 0)) { + os_atomic_store(&kn->kn_hook32, state & ~TIMER_STATE_MASK, relaxed); + + /* + * Copy out the interesting kevent state, + * but don't leak out the raw time calculations. + * + * TODO: potential enhancements - tell the user about: + * - deadline to which this timer thought it was expiring + * - return kn_sfflags in the fflags field so the client can know + * under what flags the timer fired + */ + knote_fill_kevent(kn, kev, 1); + kev->ext[0] = 0; + /* kev->ext[1] = 0; JMM - shouldn't we hide this too? */ + + if (kn->kn_sdata != 0) { /* * This is a 'repeating' timer, so we have to emit * how many intervals expired between the arm @@ -1685,15 +1756,13 @@ filt_timerprocess( * this could easily be done in the client... */ - /* The timer better have had expired... */ - assert((kn->kn_hookid & TIMER_RUNNING) == 0); - uint64_t now; - if (kn->kn_sfflags & NOTE_MACH_CONTINUOUS_TIME) + if (kn->kn_sfflags & NOTE_MACH_CONTINUOUS_TIME) { now = mach_continuous_time(); - else + } else { now = mach_absolute_time(); + } uint64_t first_deadline = kn->kn_ext[0]; uint64_t interval_abs = kn->kn_sdata; @@ -1711,18 +1780,11 @@ filt_timerprocess( * and be in repeating mode, so therefore it must have been * more than 'interval' time since the attach or last * successful touch. - * - * An unsuccessful touch would: - * disarm the timer - * clear kn_data - * clear kn_sdata - * set EV_ERROR - * all of which will prevent this code from running. */ assert(num_fired > 0); /* report how many intervals have elapsed to the user */ - kn->kn_data = (int64_t) num_fired; + kev->data = (int64_t)num_fired; /* We only need to re-arm the timer if it's not about to be destroyed */ if ((kn->kn_flags & EV_ONESHOT) == 0) { @@ -1733,93 +1795,44 @@ filt_timerprocess( kn->kn_ext[0] = new_deadline; + /* + * This can't shortcut setting up the thread call, because + * knote_process deactivates EV_CLEAR knotes unconditionnally. + */ filt_timerarm(kn); } } - /* - * Copy out the interesting kevent state, - * but don't leak out the raw time calculations. - * - * TODO: potential enhancements - tell the user about: - * - deadline to which this timer thought it was expiring - * - return kn_sfflags in the fflags field so the client can know - * under what flags the timer fired - */ - *kev = kn->kn_kevent; - kev->ext[0] = 0; - /* kev->ext[1] = 0; JMM - shouldn't we hide this too? */ - - /* we have delivered the event, reset the timer pop count */ - kn->kn_data = 0; - - filt_timerunlock(); - return 1; + return FILTER_ACTIVE; } SECURITY_READ_ONLY_EARLY(static struct filterops) timer_filtops = { + .f_extended_codes = true, .f_attach = filt_timerattach, .f_detach = filt_timerdetach, - .f_event = filt_timerevent, + .f_event = filt_bad_event, .f_touch = filt_timertouch, .f_process = filt_timerprocess, }; - -#pragma mark EVFILT_USER - - -static void -filt_userlock(void) -{ - lck_spin_lock(&_filt_userlock); -} - -static void -filt_userunlock(void) -{ - lck_spin_unlock(&_filt_userlock); -} +#pragma mark user_filtops static int -filt_userattach(struct knote *kn, __unused struct kevent_internal_s *kev) +filt_userattach(struct knote *kn, __unused struct kevent_qos_s *kev) { - /* EVFILT_USER knotes are not attached to anything in the kernel */ - /* Cant discover this knote until after attach - so no lock needed */ - kn->kn_hook = NULL; if (kn->kn_sfflags & NOTE_TRIGGER) { - kn->kn_hookid = 1; + kn->kn_hook32 = FILTER_ACTIVE; } else { - kn->kn_hookid = 0; + kn->kn_hook32 = 0; } - return (kn->kn_hookid); -} - -static void -filt_userdetach(__unused struct knote *kn) -{ - /* EVFILT_USER knotes are not attached to anything in the kernel */ -} - -static int -filt_user( - __unused struct knote *kn, - __unused long hint) -{ - panic("filt_user"); - return 0; + return kn->kn_hook32; } static int -filt_usertouch( - struct knote *kn, - struct kevent_internal_s *kev) +filt_usertouch(struct knote *kn, struct kevent_qos_s *kev) { uint32_t ffctrl; int fflags; - int active; - - filt_userlock(); ffctrl = kev->fflags & NOTE_FFCTRLMASK; fflags = kev->fflags & NOTE_FFLAGSMASK; @@ -1838,177 +1851,242 @@ filt_usertouch( } kn->kn_sdata = kev->data; - if ((kn->kn_status & KN_UDATA_SPECIFIC) == 0) - kn->kn_udata = kev->udata; - if (kev->fflags & NOTE_TRIGGER) { - kn->kn_hookid = 1; + kn->kn_hook32 = FILTER_ACTIVE; } - active = kn->kn_hookid; - - filt_userunlock(); - - return (active); + return (int)kn->kn_hook32; } static int -filt_userprocess( - struct knote *kn, - __unused struct filt_process_s *data, - struct kevent_internal_s *kev) +filt_userprocess(struct knote *kn, struct kevent_qos_s *kev) { - filt_userlock(); - - if (kn->kn_hookid == 0) { - filt_userunlock(); - return 0; - } + int result = (int)kn->kn_hook32; - *kev = kn->kn_kevent; - kev->fflags = (volatile UInt32)kn->kn_sfflags; - kev->data = kn->kn_sdata; - if (kn->kn_flags & EV_CLEAR) { - kn->kn_hookid = 0; - kn->kn_data = 0; - kn->kn_fflags = 0; + if (result) { + /* EVFILT_USER returns the data that was passed in */ + knote_fill_kevent_with_sdata(kn, kev); + kev->fflags = kn->kn_sfflags; + if (kn->kn_flags & EV_CLEAR) { + /* knote_fill_kevent cleared kn_fflags */ + kn->kn_hook32 = 0; + } } - filt_userunlock(); - return 1; + return result; } -#pragma mark EVFILT_WORKLOOP +SECURITY_READ_ONLY_EARLY(static struct filterops) user_filtops = { + .f_extended_codes = true, + .f_attach = filt_userattach, + .f_detach = filt_no_detach, + .f_event = filt_bad_event, + .f_touch = filt_usertouch, + .f_process = filt_userprocess, +}; -#if DEBUG || DEVELOPMENT -/* - * see src/queue_internal.h in libdispatch - */ -#define DISPATCH_QUEUE_ENQUEUED 0x1ull -#endif +#pragma mark workloop_filtops + +#define EPREEMPTDISABLED (-1) static inline void filt_wllock(struct kqworkloop *kqwl) { - lck_mtx_lock(&kqwl->kqwl_statelock); + lck_spin_lock(&kqwl->kqwl_statelock); } static inline void filt_wlunlock(struct kqworkloop *kqwl) { - lck_mtx_unlock(&kqwl->kqwl_statelock); -} - -static inline void -filt_wlheld(__assert_only struct kqworkloop *kqwl) -{ - LCK_MTX_ASSERT(&kqwl->kqwl_statelock, LCK_MTX_ASSERT_OWNED); + lck_spin_unlock(&kqwl->kqwl_statelock); } -#define WL_OWNER_SUSPENDED ((thread_t)(~0ull)) /* special owner when suspended */ - +/* + * Returns true when the interlock for the turnstile is the workqueue lock + * + * When this is the case, all turnstiles operations are delegated + * to the workqueue subsystem. + * + * This is required because kqueue_threadreq_bind_prepost only holds the + * workqueue lock but needs to move the inheritor from the workloop turnstile + * away from the creator thread, so that this now fulfilled request cannot be + * picked anymore by other threads. + */ static inline bool -filt_wlowner_is_valid(thread_t owner) +filt_wlturnstile_interlock_is_workq(struct kqworkloop *kqwl) { - return owner != THREAD_NULL && owner != WL_OWNER_SUSPENDED; + return kqr_thread_requested_pending(&kqwl->kqwl_request); } -static inline bool -filt_wlshould_end_ownership(struct kqworkloop *kqwl, - struct kevent_internal_s *kev, int error) +static void +filt_wlupdate_inheritor(struct kqworkloop *kqwl, struct turnstile *ts, + turnstile_update_flags_t flags) { - thread_t owner = kqwl->kqwl_owner; - return (error == 0 || error == ESTALE) && - (kev->fflags & NOTE_WL_END_OWNERSHIP) && - (owner == current_thread() || owner == WL_OWNER_SUSPENDED); -} + turnstile_inheritor_t inheritor = TURNSTILE_INHERITOR_NULL; + workq_threadreq_t kqr = &kqwl->kqwl_request; -static inline bool -filt_wlshould_update_ownership(struct kevent_internal_s *kev, int error) -{ - return error == 0 && (kev->fflags & NOTE_WL_DISCOVER_OWNER) && - kev->ext[EV_EXTIDX_WL_ADDR]; -} + /* + * binding to the workq should always happen through + * workq_kern_threadreq_update_inheritor() + */ + assert(!filt_wlturnstile_interlock_is_workq(kqwl)); -static inline bool -filt_wlshould_set_async_qos(struct kevent_internal_s *kev, int error, - kq_index_t async_qos) -{ - if (error != 0) { - return false; - } - if (async_qos != THREAD_QOS_UNSPECIFIED) { - return true; - } - if ((kev->fflags & NOTE_WL_THREAD_REQUEST) && (kev->flags & EV_DELETE)) { - /* see filt_wlprocess() */ - return true; + if ((inheritor = kqwl->kqwl_owner)) { + flags |= TURNSTILE_INHERITOR_THREAD; + } else if ((inheritor = kqr_thread(kqr))) { + flags |= TURNSTILE_INHERITOR_THREAD; } - return false; + + turnstile_update_inheritor(ts, inheritor, flags); } +#define EVFILT_WORKLOOP_EFAULT_RETRY_COUNT 100 +#define FILT_WLATTACH 0 +#define FILT_WLTOUCH 1 +#define FILT_WLDROP 2 + __result_use_check static int -filt_wlupdateowner(struct kqworkloop *kqwl, struct kevent_internal_s *kev, - int error, kq_index_t async_qos) +filt_wlupdate(struct kqworkloop *kqwl, struct knote *kn, + struct kevent_qos_s *kev, kq_index_t qos_index, int op) { - struct kqrequest *kqr = &kqwl->kqwl_request; + user_addr_t uaddr = CAST_USER_ADDR_T(kev->ext[EV_EXTIDX_WL_ADDR]); + workq_threadreq_t kqr = &kqwl->kqwl_request; thread_t cur_owner, new_owner, extra_thread_ref = THREAD_NULL; kq_index_t cur_override = THREAD_QOS_UNSPECIFIED; - kq_index_t old_owner_override = THREAD_QOS_UNSPECIFIED; - boolean_t ipc_override_is_sync = false; - boolean_t old_owner_override_is_sync = false; - int action = KQWL_UTQ_NONE; + int efault_retry = EVFILT_WORKLOOP_EFAULT_RETRY_COUNT; + int action = KQWL_UTQ_NONE, error = 0; + bool wl_inheritor_updated = false, needs_wake = false; + uint64_t kdata = kev->ext[EV_EXTIDX_WL_VALUE]; + uint64_t mask = kev->ext[EV_EXTIDX_WL_MASK]; + uint64_t udata = 0; + struct turnstile *ts = TURNSTILE_NULL; + + filt_wllock(kqwl); - filt_wlheld(kqwl); +again: + new_owner = cur_owner = kqwl->kqwl_owner; /* - * The owner is only changed under both the filt_wllock and the - * kqwl_req_lock. Looking at it with either one held is fine. + * Phase 1: + * + * If asked, load the uint64 value at the user provided address and compare + * it against the passed in mask and expected value. + * + * If NOTE_WL_DISCOVER_OWNER is specified, translate the loaded name as + * a thread reference. + * + * If NOTE_WL_END_OWNERSHIP is specified and the currently known owner is + * the current thread, then end ownership. + * + * Lastly decide whether we need to perform a QoS update. */ - cur_owner = kqwl->kqwl_owner; - if (filt_wlshould_end_ownership(kqwl, kev, error)) { - new_owner = THREAD_NULL; - } else if (filt_wlshould_update_ownership(kev, error)) { + if (uaddr) { /* - * Decipher the owner port name, and translate accordingly. - * The low 2 bits were borrowed for other flags, so mask them off. + * Until exists, + * disabling preemption copyin forces any + * vm_fault we encounter to fail. */ - uint64_t udata = kev->ext[EV_EXTIDX_WL_VALUE]; - mach_port_name_t new_owner_name = (mach_port_name_t)udata & ~0x3; - if (new_owner_name != MACH_PORT_NULL) { - new_owner_name = ipc_entry_name_mask(new_owner_name); - } - - if (MACH_PORT_VALID(new_owner_name)) { - new_owner = port_name_to_thread(new_owner_name); - if (new_owner == THREAD_NULL) - return EOWNERDEAD; - extra_thread_ref = new_owner; - } else if (new_owner_name == MACH_PORT_DEAD) { - new_owner = WL_OWNER_SUSPENDED; - } else { + error = copyin_atomic64(uaddr, &udata); + + /* + * If we get EFAULT, drop locks, and retry. + * If we still get an error report it, + * else assume the memory has been faulted + * and attempt to copyin under lock again. + */ + switch (error) { + case 0: + break; + case EFAULT: + if (efault_retry-- > 0) { + filt_wlunlock(kqwl); + error = copyin_atomic64(uaddr, &udata); + filt_wllock(kqwl); + if (error == 0) { + goto again; + } + } + OS_FALLTHROUGH; + default: + goto out; + } + + /* Update state as copied in. */ + kev->ext[EV_EXTIDX_WL_VALUE] = udata; + + if ((udata & mask) != (kdata & mask)) { + error = ESTALE; + } else if (kev->fflags & NOTE_WL_DISCOVER_OWNER) { /* - * We never want to learn a new owner that is NULL. - * Ownership should be ended with END_OWNERSHIP. + * Decipher the owner port name, and translate accordingly. + * The low 2 bits were borrowed for other flags, so mask them off. + * + * Then attempt translation to a thread reference or fail. */ - new_owner = cur_owner; + mach_port_name_t name = (mach_port_name_t)udata & ~0x3; + if (name != MACH_PORT_NULL) { + name = ipc_entry_name_mask(name); + extra_thread_ref = port_name_to_thread(name, + PORT_TO_THREAD_IN_CURRENT_TASK); + if (extra_thread_ref == THREAD_NULL) { + error = EOWNERDEAD; + goto out; + } + new_owner = extra_thread_ref; + } } - } else { - new_owner = cur_owner; } - if (filt_wlshould_set_async_qos(kev, error, async_qos)) { - action = KQWL_UTQ_SET_ASYNC_QOS; + if ((kev->fflags & NOTE_WL_END_OWNERSHIP) && new_owner == current_thread()) { + new_owner = THREAD_NULL; + } + + if (error == 0) { + if ((kev->fflags & NOTE_WL_THREAD_REQUEST) && (kev->flags & EV_DELETE)) { + action = KQWL_UTQ_SET_QOS_INDEX; + } else if (qos_index && kqr->tr_kq_qos_index != qos_index) { + action = KQWL_UTQ_SET_QOS_INDEX; + } + + if (op == FILT_WLTOUCH) { + /* + * Save off any additional fflags/data we just accepted + * But only keep the last round of "update" bits we acted on which helps + * debugging a lot. + */ + kn->kn_sfflags &= ~NOTE_WL_UPDATES_MASK; + kn->kn_sfflags |= kev->fflags; + if (kev->fflags & NOTE_WL_SYNC_WAKE) { + needs_wake = (kn->kn_thread != THREAD_NULL); + } + } else if (op == FILT_WLDROP) { + if ((kn->kn_sfflags & (NOTE_WL_SYNC_WAIT | NOTE_WL_SYNC_WAKE)) == + NOTE_WL_SYNC_WAIT) { + /* + * When deleting a SYNC_WAIT knote that hasn't been woken up + * explicitly, issue a wake up. + */ + kn->kn_sfflags |= NOTE_WL_SYNC_WAKE; + needs_wake = (kn->kn_thread != THREAD_NULL); + } + } } - if (cur_owner == new_owner && action == KQWL_UTQ_NONE) { + + /* + * Phase 2: + * + * Commit ownership and QoS changes if any, possibly wake up waiters + */ + + if (cur_owner == new_owner && action == KQWL_UTQ_NONE && !needs_wake) { goto out; } - kqwl_req_lock(kqwl); + kqlock(kqwl); /* If already tracked as servicer, don't track as owner */ - if ((kqr->kqr_state & KQR_BOUND) && new_owner == kqr->kqr_thread) { - kqwl->kqwl_owner = new_owner = THREAD_NULL; + if (new_owner == kqr_thread(kqr)) { + new_owner = THREAD_NULL; } if (cur_owner != new_owner) { @@ -2017,31 +2095,20 @@ filt_wlupdateowner(struct kqworkloop *kqwl, struct kevent_internal_s *kev, /* we just transfered this ref to kqwl_owner */ extra_thread_ref = THREAD_NULL; } - cur_override = kqworkloop_combined_qos(kqwl, &ipc_override_is_sync); - old_owner_override = kqr->kqr_dsync_owner_qos; - old_owner_override_is_sync = kqr->kqr_owner_override_is_sync; + cur_override = kqworkloop_override(kqwl); - if (filt_wlowner_is_valid(new_owner)) { + if (new_owner) { /* override it before we drop the old */ if (cur_override != THREAD_QOS_UNSPECIFIED) { - thread_add_ipc_override(new_owner, cur_override); - } - if (ipc_override_is_sync) { - thread_add_sync_ipc_override(new_owner); + thread_add_kevent_override(new_owner, cur_override); } - /* Update the kqr to indicate that owner has sync ipc override */ - kqr->kqr_dsync_owner_qos = cur_override; - kqr->kqr_owner_override_is_sync = ipc_override_is_sync; - thread_starts_owning_workloop(new_owner); - if ((kqr->kqr_state & (KQR_THREQUESTED | KQR_BOUND)) == KQR_THREQUESTED) { + if (kqr_thread_requested_pending(kqr)) { if (action == KQWL_UTQ_NONE) { action = KQWL_UTQ_REDRIVE_EVENTS; } } - } else if (new_owner == THREAD_NULL) { - kqr->kqr_dsync_owner_qos = THREAD_QOS_UNSPECIFIED; - kqr->kqr_owner_override_is_sync = false; - if ((kqr->kqr_state & (KQR_THREQUESTED | KQR_WAKEUP)) == KQR_WAKEUP) { + } else { + if (!kqr_thread_requested(kqr) && kqr->tr_kq_wakeup) { if (action == KQWL_UTQ_NONE) { action = KQWL_UTQ_REDRIVE_EVENTS; } @@ -2050,72 +2117,98 @@ filt_wlupdateowner(struct kqworkloop *kqwl, struct kevent_internal_s *kev, } if (action != KQWL_UTQ_NONE) { - kqworkloop_update_threads_qos(kqwl, action, async_qos); + kqworkloop_update_threads_qos(kqwl, action, qos_index); } - kqwl_req_unlock(kqwl); - - /* Now that we are unlocked, drop the override and ref on old owner */ - if (new_owner != cur_owner && filt_wlowner_is_valid(cur_owner)) { - if (old_owner_override != THREAD_QOS_UNSPECIFIED) { - thread_drop_ipc_override(cur_owner); + ts = kqwl->kqwl_turnstile; + if (cur_owner != new_owner && ts) { + if (action == KQWL_UTQ_REDRIVE_EVENTS) { + /* + * Note that when action is KQWL_UTQ_REDRIVE_EVENTS, + * the code went through workq_kern_threadreq_initiate() + * and the workqueue has set the inheritor already + */ + assert(filt_wlturnstile_interlock_is_workq(kqwl)); + } else if (filt_wlturnstile_interlock_is_workq(kqwl)) { + workq_kern_threadreq_lock(kqwl->kqwl_p); + workq_kern_threadreq_update_inheritor(kqwl->kqwl_p, kqr, new_owner, + ts, TURNSTILE_IMMEDIATE_UPDATE); + workq_kern_threadreq_unlock(kqwl->kqwl_p); + if (!filt_wlturnstile_interlock_is_workq(kqwl)) { + /* + * If the workq is no longer the interlock, then + * workq_kern_threadreq_update_inheritor() has finished a bind + * and we need to fallback to the regular path. + */ + filt_wlupdate_inheritor(kqwl, ts, TURNSTILE_IMMEDIATE_UPDATE); + } + wl_inheritor_updated = true; + } else { + filt_wlupdate_inheritor(kqwl, ts, TURNSTILE_IMMEDIATE_UPDATE); + wl_inheritor_updated = true; } - if (old_owner_override_is_sync) { - thread_drop_sync_ipc_override(cur_owner); + + /* + * We need a turnstile reference because we are dropping the interlock + * and the caller has not called turnstile_prepare. + */ + if (wl_inheritor_updated) { + turnstile_reference(ts); } - thread_ends_owning_workloop(cur_owner); - thread_deallocate(cur_owner); } -out: - if (extra_thread_ref) { - thread_deallocate(extra_thread_ref); + if (needs_wake && ts) { + waitq_wakeup64_thread(&ts->ts_waitq, knote_filt_wev64(kn), + kn->kn_thread, THREAD_AWAKENED); + if (op == FILT_WLATTACH || op == FILT_WLTOUCH) { + disable_preemption(); + error = EPREEMPTDISABLED; + } } - return error; -} -static int -filt_wldebounce( - struct kqworkloop *kqwl, - struct kevent_internal_s *kev, - int default_result) -{ - user_addr_t addr = CAST_USER_ADDR_T(kev->ext[EV_EXTIDX_WL_ADDR]); - uint64_t udata; - int error; + kqunlock(kqwl); - /* we must have the workloop state mutex held */ - filt_wlheld(kqwl); +out: + /* + * Phase 3: + * + * Unlock and cleanup various lingering references and things. + */ + filt_wlunlock(kqwl); - /* Do we have a debounce address to work with? */ - if (addr) { - uint64_t kdata = kev->ext[EV_EXTIDX_WL_VALUE]; - uint64_t mask = kev->ext[EV_EXTIDX_WL_MASK]; +#if CONFIG_WORKLOOP_DEBUG + KQWL_HISTORY_WRITE_ENTRY(kqwl, { + .updater = current_thread(), + .servicer = kqr_thread(kqr), /* Note: racy */ + .old_owner = cur_owner, + .new_owner = new_owner, - error = copyin_word(addr, &udata, sizeof(udata)); - if (error) { - return error; - } + .kev_ident = kev->ident, + .error = (int16_t)error, + .kev_flags = kev->flags, + .kev_fflags = kev->fflags, - /* update state as copied in */ - kev->ext[EV_EXTIDX_WL_VALUE] = udata; + .kev_mask = mask, + .kev_value = kdata, + .in_value = udata, + }); +#endif // CONFIG_WORKLOOP_DEBUG - /* If the masked bits don't match, reject it as stale */ - if ((udata & mask) != (kdata & mask)) { - return ESTALE; - } + if (wl_inheritor_updated) { + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_NOT_HELD); + turnstile_deallocate_safe(ts); + } -#if DEBUG || DEVELOPMENT - if ((kev->fflags & NOTE_WL_THREAD_REQUEST) && !(kev->flags & EV_DELETE)) { - if ((udata & DISPATCH_QUEUE_ENQUEUED) == 0) { - panic("kevent: workloop %#016llx is not enqueued " - "(kev:%p dq_state:%#016llx)", kev->udata, kev, udata); - } + if (cur_owner && new_owner != cur_owner) { + if (cur_override != THREAD_QOS_UNSPECIFIED) { + thread_drop_kevent_override(cur_owner); } -#endif + thread_deallocate_safe(cur_owner); } - - return default_result; + if (extra_thread_ref) { + thread_deallocate_safe(extra_thread_ref); + } + return error; } /* @@ -2126,119 +2219,127 @@ filt_wldebounce( * - data is set to the error if any */ static inline void -filt_wlremember_last_update( - __assert_only struct kqworkloop *kqwl, - struct knote *kn, - struct kevent_internal_s *kev, - int error) +filt_wlremember_last_update(struct knote *kn, struct kevent_qos_s *kev, + int error) { - filt_wlheld(kqwl); kn->kn_fflags = kev->fflags; - kn->kn_data = error; + kn->kn_sdata = error; memcpy(kn->kn_ext, kev->ext, sizeof(kev->ext)); } -/* - * Return which operations on EVFILT_WORKLOOP need to be protected against - * knoteusewait() causing priority inversions. - */ -static bool -filt_wlneeds_boost(struct kevent_internal_s *kev) -{ - if (kev == NULL) { - /* - * this is an f_process() usecount, and it can cause a drop to wait - */ - return true; +static int +filt_wlupdate_sync_ipc(struct kqworkloop *kqwl, struct knote *kn, + struct kevent_qos_s *kev, int op) +{ + user_addr_t uaddr = (user_addr_t) kev->ext[EV_EXTIDX_WL_ADDR]; + uint64_t kdata = kev->ext[EV_EXTIDX_WL_VALUE]; + uint64_t mask = kev->ext[EV_EXTIDX_WL_MASK]; + uint64_t udata = 0; + int efault_retry = EVFILT_WORKLOOP_EFAULT_RETRY_COUNT; + int error = 0; + + if (op == FILT_WLATTACH) { + (void)kqueue_alloc_turnstile(&kqwl->kqwl_kqueue); + } else if (uaddr == 0) { + return 0; } - if (kev->fflags & NOTE_WL_THREAD_REQUEST) { + + filt_wllock(kqwl); + +again: + + /* + * Do the debounce thing, the lock serializing the state is the knote lock. + */ + if (uaddr) { /* - * All operations on thread requests may starve drops or re-attach of - * the same knote, all of them need boosts. None of what we do under - * thread-request usecount holds blocks anyway. + * Until exists, + * disabling preemption copyin forces any + * vm_fault we encounter to fail. */ - return true; - } - if (kev->fflags & NOTE_WL_SYNC_WAIT) { + error = copyin_atomic64(uaddr, &udata); + /* - * this may call filt_wlwait() and we don't want to hold any boost when - * woken up, this would cause background threads contending on - * dispatch_sync() to wake up at 64 and be preempted immediately when - * this drops. + * If we get EFAULT, drop locks, and retry. + * If we still get an error report it, + * else assume the memory has been faulted + * and attempt to copyin under lock again. */ - return false; + switch (error) { + case 0: + break; + case EFAULT: + if (efault_retry-- > 0) { + filt_wlunlock(kqwl); + error = copyin_atomic64(uaddr, &udata); + filt_wllock(kqwl); + if (error == 0) { + goto again; + } + } + OS_FALLTHROUGH; + default: + goto out; + } + + kev->ext[EV_EXTIDX_WL_VALUE] = udata; + kn->kn_ext[EV_EXTIDX_WL_VALUE] = udata; + + if ((udata & mask) != (kdata & mask)) { + error = ESTALE; + goto out; + } } - /* - * SYNC_WAIT knotes when deleted don't need to be rushed, there's no - * detach/reattach race with these ever. In addition to this, when the - * SYNC_WAIT knote is dropped, the caller is no longer receiving the - * workloop overrides if any, and we'd rather schedule other threads than - * him, he's not possibly stalling anything anymore. - */ - return (kev->flags & EV_DELETE) == 0; + if (op == FILT_WLATTACH) { + error = filt_wlattach_sync_ipc(kn); + if (error == 0) { + disable_preemption(); + error = EPREEMPTDISABLED; + } + } + +out: + filt_wlunlock(kqwl); + return error; } static int -filt_wlattach(struct knote *kn, struct kevent_internal_s *kev) +filt_wlattach(struct knote *kn, struct kevent_qos_s *kev) { struct kqueue *kq = knote_get_kq(kn); struct kqworkloop *kqwl = (struct kqworkloop *)kq; - int error = 0; + int error = 0, result = 0; kq_index_t qos_index = 0; - if ((kq->kq_state & KQ_WORKLOOP) == 0) { + if (__improbable((kq->kq_state & KQ_WORKLOOP) == 0)) { error = ENOTSUP; goto out; } -#if DEVELOPMENT || DEBUG - if (kev->ident == 0 && kev->udata == 0 && kev->fflags == 0) { - struct kqrequest *kqr = &kqwl->kqwl_request; - - kqwl_req_lock(kqwl); - kev->fflags = 0; - if (kqr->kqr_dsync_waiters) { - kev->fflags |= NOTE_WL_SYNC_WAIT; - } - if (kqr->kqr_qos_index) { - kev->fflags |= NOTE_WL_THREAD_REQUEST; - } - if (kqwl->kqwl_owner == WL_OWNER_SUSPENDED) { - kev->ext[0] = ~0ull; - } else { - kev->ext[0] = thread_tid(kqwl->kqwl_owner); - } - kev->ext[1] = thread_tid(kqwl->kqwl_request.kqr_thread); - kev->ext[2] = thread_owned_workloops_count(current_thread()); - kev->ext[3] = kn->kn_kevent.ext[3]; - kqwl_req_unlock(kqwl); - error = EBUSY; - goto out; - } -#endif - - /* Some simple validation */ - int command = (kn->kn_sfflags & NOTE_WL_COMMANDS_MASK); + uint32_t command = (kn->kn_sfflags & NOTE_WL_COMMANDS_MASK); switch (command) { case NOTE_WL_THREAD_REQUEST: if (kn->kn_id != kqwl->kqwl_dynamicid) { error = EINVAL; goto out; } - qos_index = qos_index_from_qos(kn, kn->kn_qos, FALSE); - if (qos_index < THREAD_QOS_MAINTENANCE || - qos_index > THREAD_QOS_USER_INTERACTIVE) { + qos_index = _pthread_priority_thread_qos(kn->kn_qos); + if (qos_index == THREAD_QOS_UNSPECIFIED) { error = ERANGE; goto out; } + if (kqwl->kqwl_request.tr_kq_qos_index) { + /* + * There already is a thread request, and well, you're only allowed + * one per workloop, so fail the attach. + */ + error = EALREADY; + goto out; + } break; case NOTE_WL_SYNC_WAIT: case NOTE_WL_SYNC_WAKE: - if (kq->kq_state & KQ_NO_WQ_THREAD) { - error = ENOTSUP; - goto out; - } if (kn->kn_id == kqwl->kqwl_dynamicid) { error = EINVAL; goto out; @@ -2252,800 +2353,588 @@ filt_wlattach(struct knote *kn, struct kevent_internal_s *kev) goto out; } break; + + case NOTE_WL_SYNC_IPC: + if ((kn->kn_flags & EV_DISABLE) == 0) { + error = EINVAL; + goto out; + } + if (kn->kn_sfflags & (NOTE_WL_UPDATE_QOS | NOTE_WL_DISCOVER_OWNER)) { + error = EINVAL; + goto out; + } + break; default: error = EINVAL; goto out; } - filt_wllock(kqwl); - kn->kn_hook = NULL; - - if (command == NOTE_WL_THREAD_REQUEST && kqwl->kqwl_request.kqr_qos_index) { - /* - * There already is a thread request, and well, you're only allowed - * one per workloop, so fail the attach. - * - * Note: kqr_qos_index is always set with the wllock held, so we - * don't need to take the kqr lock. - */ - error = EALREADY; + if (command == NOTE_WL_SYNC_IPC) { + error = filt_wlupdate_sync_ipc(kqwl, kn, kev, FILT_WLATTACH); } else { - /* Make sure user and kernel are in agreement on important state */ - error = filt_wldebounce(kqwl, kev, 0); + error = filt_wlupdate(kqwl, kn, kev, qos_index, FILT_WLATTACH); } - error = filt_wlupdateowner(kqwl, kev, error, qos_index); - filt_wlunlock(kqwl); + if (error == EPREEMPTDISABLED) { + error = 0; + result = FILTER_THREADREQ_NODEFEER; + } out: if (error) { - kn->kn_flags |= EV_ERROR; /* If userland wants ESTALE to be hidden, fail the attach anyway */ if (error == ESTALE && (kn->kn_sfflags & NOTE_WL_IGNORE_ESTALE)) { error = 0; } - kn->kn_data = error; - return 0; + knote_set_error(kn, error); + return result; + } + if (command == NOTE_WL_SYNC_WAIT) { + return kevent_register_wait_prepare(kn, kev, result); } - /* Just attaching the thread request successfully will fire it */ - return command == NOTE_WL_THREAD_REQUEST; + if (command == NOTE_WL_THREAD_REQUEST) { + /* + * Thread Request knotes need an explicit touch to be active again, + * so delivering an event needs to also consume it. + */ + kn->kn_flags |= EV_CLEAR; + return result | FILTER_ACTIVE; + } + return result; } -__attribute__((noinline,not_tail_called)) -static int -filt_wlwait(struct kqworkloop *kqwl, - struct knote *kn, - struct kevent_internal_s *kev) +static void __dead2 +filt_wlwait_continue(void *parameter, wait_result_t wr) { - filt_wlheld(kqwl); - assert((kn->kn_sfflags & NOTE_WL_SYNC_WAKE) == 0); - - /* - * Hint to the wakeup side that this thread is waiting. Also used by - * stackshot for waitinfo. - */ - kn->kn_hook = current_thread(); + struct _kevent_register *cont_args = parameter; + struct kqworkloop *kqwl = cont_args->kqwl; - thread_set_pending_block_hint(current_thread(), kThreadWaitWorkloopSyncWait); + kqlock(kqwl); + if (filt_wlturnstile_interlock_is_workq(kqwl)) { + workq_kern_threadreq_lock(kqwl->kqwl_p); + turnstile_complete((uintptr_t)kqwl, &kqwl->kqwl_turnstile, NULL, TURNSTILE_WORKLOOPS); + workq_kern_threadreq_unlock(kqwl->kqwl_p); + } else { + turnstile_complete((uintptr_t)kqwl, &kqwl->kqwl_turnstile, NULL, TURNSTILE_WORKLOOPS); + } + kqunlock(kqwl); - wait_result_t wr = assert_wait(kn, THREAD_ABORTSAFE); + turnstile_cleanup(); - if (wr == THREAD_WAITING) { - kq_index_t qos_index = qos_index_from_qos(kn, kev->qos, TRUE); - struct kqrequest *kqr = &kqwl->kqwl_request; + if (wr == THREAD_INTERRUPTED) { + cont_args->kev.flags |= EV_ERROR; + cont_args->kev.data = EINTR; + } else if (wr != THREAD_AWAKENED) { + panic("Unexpected wait result: %d", wr); + } - thread_t thread_to_handoff = THREAD_NULL; /* holds +1 thread ref */ + kevent_register_wait_return(cont_args); +} - thread_t kqwl_owner = kqwl->kqwl_owner; - if (filt_wlowner_is_valid(kqwl_owner)) { - thread_reference(kqwl_owner); - thread_to_handoff = kqwl_owner; - } +/* + * Called with the workloop mutex held, most of the time never returns as it + * calls filt_wlwait_continue through a continuation. + */ +static void __dead2 +filt_wlpost_register_wait(struct uthread *uth, struct knote *kn, + struct _kevent_register *cont_args) +{ + struct kqworkloop *kqwl = cont_args->kqwl; + workq_threadreq_t kqr = &kqwl->kqwl_request; + struct turnstile *ts; + bool workq_locked = false; - kqwl_req_lock(kqwl); + kqlock_held(kqwl); - if (qos_index) { - assert(kqr->kqr_dsync_waiters < UINT16_MAX); - kqr->kqr_dsync_waiters++; - if (qos_index > kqr->kqr_dsync_waiters_qos) { - kqworkloop_update_threads_qos(kqwl, - KQWL_UTQ_SET_SYNC_WAITERS_QOS, qos_index); - } - } + if (filt_wlturnstile_interlock_is_workq(kqwl)) { + workq_kern_threadreq_lock(kqwl->kqwl_p); + workq_locked = true; + } - if ((kqr->kqr_state & KQR_BOUND) && thread_to_handoff == THREAD_NULL) { - assert(kqr->kqr_thread != THREAD_NULL); - thread_t servicer = kqr->kqr_thread; + ts = turnstile_prepare((uintptr_t)kqwl, &kqwl->kqwl_turnstile, + TURNSTILE_NULL, TURNSTILE_WORKLOOPS); - thread_reference(servicer); - thread_to_handoff = servicer; + if (workq_locked) { + workq_kern_threadreq_update_inheritor(kqwl->kqwl_p, + &kqwl->kqwl_request, kqwl->kqwl_owner, ts, + TURNSTILE_DELAYED_UPDATE); + if (!filt_wlturnstile_interlock_is_workq(kqwl)) { + /* + * if the interlock is no longer the workqueue lock, + * then we don't need to hold it anymore. + */ + workq_kern_threadreq_unlock(kqwl->kqwl_p); + workq_locked = false; } + } + if (!workq_locked) { + /* + * If the interlock is the workloop's, then it's our responsibility to + * call update_inheritor, so just do it. + */ + filt_wlupdate_inheritor(kqwl, ts, TURNSTILE_DELAYED_UPDATE); + } - kqwl_req_unlock(kqwl); - - filt_wlunlock(kqwl); - - /* TODO: use continuation based blocking */ - - /* consume a refcount on thread_to_handoff, then thread_block() */ - wr = thread_handoff(thread_to_handoff); - thread_to_handoff = THREAD_NULL; - - filt_wllock(kqwl); - - /* clear waiting state (only one waiting thread - so no race) */ - assert(kn->kn_hook == current_thread()); + thread_set_pending_block_hint(uth->uu_thread, kThreadWaitWorkloopSyncWait); + waitq_assert_wait64(&ts->ts_waitq, knote_filt_wev64(kn), + THREAD_ABORTSAFE, TIMEOUT_WAIT_FOREVER); - if (qos_index) { - kqwl_req_lock(kqwl); - assert(kqr->kqr_dsync_waiters > 0); - if (--kqr->kqr_dsync_waiters == 0) { - assert(kqr->kqr_dsync_waiters_qos); - kqworkloop_update_threads_qos(kqwl, - KQWL_UTQ_SET_SYNC_WAITERS_QOS, 0); - } - kqwl_req_unlock(kqwl); - } + if (workq_locked) { + workq_kern_threadreq_unlock(kqwl->kqwl_p); } - kn->kn_hook = NULL; - - switch (wr) { - case THREAD_AWAKENED: - return 0; - case THREAD_INTERRUPTED: - return EINTR; - case THREAD_RESTART: - return ECANCELED; - default: - panic("filt_wlattach: unexpected wait result %d", wr); - return EINVAL; + thread_t thread = kqwl->kqwl_owner ?: kqr_thread(kqr); + if (thread) { + thread_reference(thread); } + + kevent_register_wait_block(ts, thread, filt_wlwait_continue, cont_args); } /* called in stackshot context to report the thread responsible for blocking this thread */ void kdp_workloop_sync_wait_find_owner(__assert_only thread_t thread, - event64_t event, - thread_waitinfo_t *waitinfo) + event64_t event, thread_waitinfo_t *waitinfo) { - struct knote *kn = (struct knote*) event; - assert(kdp_is_in_zone(kn, "knote zone")); + extern zone_t thread_zone; + struct knote *kn = (struct knote *)event; - assert(kn->kn_hook == thread); + zone_require(knote_zone, kn); + + assert(kn->kn_thread == thread); struct kqueue *kq = knote_get_kq(kn); - assert(kdp_is_in_zone(kq, "kqueue workloop zone")); + + zone_require(kqworkloop_zone, kq); assert(kq->kq_state & KQ_WORKLOOP); struct kqworkloop *kqwl = (struct kqworkloop *)kq; - struct kqrequest *kqr = &kqwl->kqwl_request; + workq_threadreq_t kqr = &kqwl->kqwl_request; thread_t kqwl_owner = kqwl->kqwl_owner; - thread_t servicer = kqr->kqr_thread; - - if (kqwl_owner == WL_OWNER_SUSPENDED) { - waitinfo->owner = STACKSHOT_WAITOWNER_SUSPENDED; - } else if (kqwl_owner != THREAD_NULL) { - assert(kdp_is_in_zone(kqwl_owner, "threads")); + if (kqwl_owner != THREAD_NULL) { + zone_require(thread_zone, kqwl_owner); waitinfo->owner = thread_tid(kqwl->kqwl_owner); - } else if (servicer != THREAD_NULL) { - assert(kdp_is_in_zone(servicer, "threads")); - - waitinfo->owner = thread_tid(servicer); - } else if (kqr->kqr_state & KQR_THREQUESTED) { + } else if (kqr_thread_requested_pending(kqr)) { waitinfo->owner = STACKSHOT_WAITOWNER_THREQUESTED; + } else if (kqr->tr_state >= WORKQ_TR_STATE_BINDING) { + zone_require(thread_zone, kqr->tr_thread); + waitinfo->owner = thread_tid(kqr->tr_thread); } else { waitinfo->owner = 0; } waitinfo->context = kqwl->kqwl_dynamicid; - - return; -} - -/* - * Takes kqueue locked, returns locked, may drop in the middle and/or block for a while - */ -static int -filt_wlpost_attach(struct knote *kn, struct kevent_internal_s *kev) -{ - struct kqueue *kq = knote_get_kq(kn); - struct kqworkloop *kqwl = (struct kqworkloop *)kq; - int error = 0; - - if (kev->fflags & NOTE_WL_SYNC_WAIT) { - if (kqlock2knoteuse(kq, kn, KNUSE_NONE)) { - filt_wllock(kqwl); - /* if the wake has already preposted, don't wait */ - if ((kn->kn_sfflags & NOTE_WL_SYNC_WAKE) == 0) - error = filt_wlwait(kqwl, kn, kev); - filt_wlunlock(kqwl); - knoteuse2kqlock(kq, kn, KNUSE_NONE); - } - } - return error; } static void -filt_wldetach(__assert_only struct knote *kn) +filt_wldetach(struct knote *kn) { - assert(knote_get_kq(kn)->kq_state & KQ_WORKLOOP); - - /* - * Thread requests have nothing to detach. - * Sync waiters should have been aborted out - * and drop their refs before we could drop/ - * detach their knotes. - */ - assert(kn->kn_hook == NULL); + if (kn->kn_sfflags & NOTE_WL_SYNC_IPC) { + filt_wldetach_sync_ipc(kn); + } else if (kn->kn_thread) { + kevent_register_wait_cleanup(kn); + } } static int -filt_wlevent( - __unused struct knote *kn, - __unused long hint) +filt_wlvalidate_kev_flags(struct knote *kn, struct kevent_qos_s *kev, + thread_qos_t *qos_index) { - panic("filt_wlevent"); - return 0; -} + uint32_t new_commands = kev->fflags & NOTE_WL_COMMANDS_MASK; + uint32_t sav_commands = kn->kn_sfflags & NOTE_WL_COMMANDS_MASK; -static int -filt_wlvalidate_kev_flags(struct knote *kn, struct kevent_internal_s *kev) -{ - int new_commands = kev->fflags & NOTE_WL_COMMANDS_MASK; - int sav_commands = kn->kn_sfflags & NOTE_WL_COMMANDS_MASK; - int error = 0; + if ((kev->fflags & NOTE_WL_DISCOVER_OWNER) && (kev->flags & EV_DELETE)) { + return EINVAL; + } + if (kev->fflags & NOTE_WL_UPDATE_QOS) { + if (kev->flags & EV_DELETE) { + return EINVAL; + } + if (sav_commands != NOTE_WL_THREAD_REQUEST) { + return EINVAL; + } + if (!(*qos_index = _pthread_priority_thread_qos(kev->qos))) { + return ERANGE; + } + } switch (new_commands) { case NOTE_WL_THREAD_REQUEST: /* thread requests can only update themselves */ - if (sav_commands != new_commands) - error = EINVAL; + if (sav_commands != NOTE_WL_THREAD_REQUEST) { + return EINVAL; + } break; case NOTE_WL_SYNC_WAIT: - if (kev->fflags & NOTE_WL_END_OWNERSHIP) - error = EINVAL; - /* FALLTHROUGH */ + if (kev->fflags & NOTE_WL_END_OWNERSHIP) { + return EINVAL; + } + goto sync_checks; + case NOTE_WL_SYNC_WAKE: - /* waits and wakes can update themselves or their counterparts */ - if (!(sav_commands & (NOTE_WL_SYNC_WAIT | NOTE_WL_SYNC_WAKE))) - error = EINVAL; - if (kev->fflags & NOTE_WL_UPDATE_QOS) - error = EINVAL; - if ((kev->flags & (EV_ENABLE | EV_DELETE)) == EV_ENABLE) - error = EINVAL; - if (kev->flags & EV_DELETE) { - /* - * Really this is not supported: there is absolutely no reason - * whatsoever to want to fail the drop of a NOTE_WL_SYNC_WAIT knote. - */ - if (kev->ext[EV_EXTIDX_WL_ADDR] && kev->ext[EV_EXTIDX_WL_MASK]) { - error = EINVAL; - } +sync_checks: + if (!(sav_commands & (NOTE_WL_SYNC_WAIT | NOTE_WL_SYNC_WAKE))) { + return EINVAL; + } + if ((kev->flags & (EV_ENABLE | EV_DELETE)) == EV_ENABLE) { + return EINVAL; + } + break; + + case NOTE_WL_SYNC_IPC: + if (sav_commands != NOTE_WL_SYNC_IPC) { + return EINVAL; + } + if ((kev->flags & (EV_ENABLE | EV_DELETE)) == EV_ENABLE) { + return EINVAL; } break; default: - error = EINVAL; - } - if ((kev->flags & EV_DELETE) && (kev->fflags & NOTE_WL_DISCOVER_OWNER)) { - error = EINVAL; + return EINVAL; } - return error; + return 0; } static int -filt_wltouch( - struct knote *kn, - struct kevent_internal_s *kev) +filt_wltouch(struct knote *kn, struct kevent_qos_s *kev) { - struct kqueue *kq = knote_get_kq(kn); - int error = 0; - struct kqworkloop *kqwl; - - assert(kq->kq_state & KQ_WORKLOOP); - kqwl = (struct kqworkloop *)kq; + struct kqworkloop *kqwl = (struct kqworkloop *)knote_get_kq(kn); + thread_qos_t qos_index = THREAD_QOS_UNSPECIFIED; + int result = 0; - error = filt_wlvalidate_kev_flags(kn, kev); + int error = filt_wlvalidate_kev_flags(kn, kev, &qos_index); if (error) { goto out; } - filt_wllock(kqwl); - - /* Make sure user and kernel are in agreement on important state */ - error = filt_wldebounce(kqwl, kev, 0); - if (error) { - error = filt_wlupdateowner(kqwl, kev, error, 0); - goto out_unlock; + uint32_t command = kev->fflags & NOTE_WL_COMMANDS_MASK; + if (command == NOTE_WL_SYNC_IPC) { + error = filt_wlupdate_sync_ipc(kqwl, kn, kev, FILT_WLTOUCH); + } else { + error = filt_wlupdate(kqwl, kn, kev, qos_index, FILT_WLTOUCH); + filt_wlremember_last_update(kn, kev, error); + } + if (error == EPREEMPTDISABLED) { + error = 0; + result = FILTER_THREADREQ_NODEFEER; } - int new_command = kev->fflags & NOTE_WL_COMMANDS_MASK; - switch (new_command) { - case NOTE_WL_THREAD_REQUEST: - assert(kqwl->kqwl_request.kqr_qos_index != THREAD_QOS_UNSPECIFIED); - break; - - case NOTE_WL_SYNC_WAIT: - /* - * we need to allow waiting several times on the same knote because - * of EINTR. If it's already woken though, it won't block. - */ - break; - - case NOTE_WL_SYNC_WAKE: - if (kn->kn_sfflags & NOTE_WL_SYNC_WAKE) { - /* disallow waking the same knote twice */ - error = EALREADY; - goto out_unlock; - } - if (kn->kn_hook) { - thread_wakeup_thread((event_t)kn, (thread_t)kn->kn_hook); +out: + if (error) { + if (error == ESTALE && (kev->fflags & NOTE_WL_IGNORE_ESTALE)) { + /* If userland wants ESTALE to be hidden, do not activate */ + return result; } - break; - - default: - error = EINVAL; - goto out_unlock; + kev->flags |= EV_ERROR; + kev->data = error; + return result; } - - /* - * Save off any additional fflags/data we just accepted - * But only keep the last round of "update" bits we acted on which helps - * debugging a lot. - */ - kn->kn_sfflags &= ~NOTE_WL_UPDATES_MASK; - kn->kn_sfflags |= kev->fflags; - kn->kn_sdata = kev->data; - - kq_index_t qos_index = THREAD_QOS_UNSPECIFIED; - - if (kev->fflags & NOTE_WL_UPDATE_QOS) { - qos_t qos = pthread_priority_canonicalize(kev->qos, FALSE); - - if (kn->kn_qos != qos) { - qos_index = qos_index_from_qos(kn, qos, FALSE); - if (qos_index == THREAD_QOS_UNSPECIFIED) { - error = ERANGE; - goto out_unlock; - } - kqlock(kq); - if (kn->kn_status & KN_QUEUED) { - knote_dequeue(kn); - knote_set_qos_index(kn, qos_index); - knote_enqueue(kn); - knote_wakeup(kn); - } else { - knote_set_qos_index(kn, qos_index); - } - kn->kn_qos = qos; - kqunlock(kq); + if (command == NOTE_WL_SYNC_WAIT && !(kn->kn_sfflags & NOTE_WL_SYNC_WAKE)) { + return kevent_register_wait_prepare(kn, kev, result); + } + /* Just touching the thread request successfully will fire it */ + if (command == NOTE_WL_THREAD_REQUEST) { + if (kev->fflags & NOTE_WL_UPDATE_QOS) { + result |= FILTER_UPDATE_REQ_QOS; } + result |= FILTER_ACTIVE; } + return result; +} + +static bool +filt_wlallow_drop(struct knote *kn, struct kevent_qos_s *kev) +{ + struct kqworkloop *kqwl = (struct kqworkloop *)knote_get_kq(kn); - error = filt_wlupdateowner(kqwl, kev, 0, qos_index); + int error = filt_wlvalidate_kev_flags(kn, kev, NULL); if (error) { - goto out_unlock; + goto out; } - if (new_command == NOTE_WL_SYNC_WAIT) { - /* if the wake has already preposted, don't wait */ - if ((kn->kn_sfflags & NOTE_WL_SYNC_WAKE) == 0) - error = filt_wlwait(kqwl, kn, kev); + uint32_t command = (kev->fflags & NOTE_WL_COMMANDS_MASK); + if (command == NOTE_WL_SYNC_IPC) { + error = filt_wlupdate_sync_ipc(kqwl, kn, kev, FILT_WLDROP); + } else { + error = filt_wlupdate(kqwl, kn, kev, 0, FILT_WLDROP); + filt_wlremember_last_update(kn, kev, error); } + assert(error != EPREEMPTDISABLED); -out_unlock: - filt_wlremember_last_update(kqwl, kn, kev, error); - filt_wlunlock(kqwl); out: if (error) { if (error == ESTALE && (kev->fflags & NOTE_WL_IGNORE_ESTALE)) { - /* If userland wants ESTALE to be hidden, do not activate */ - return 0; + return false; } kev->flags |= EV_ERROR; kev->data = error; - return 0; + return false; } - /* Just touching the thread request successfully will fire it */ - return new_command == NOTE_WL_THREAD_REQUEST; + return true; } static int -filt_wldrop_and_unlock( - struct knote *kn, - struct kevent_internal_s *kev) +filt_wlprocess(struct knote *kn, struct kevent_qos_s *kev) { - struct kqueue *kq = knote_get_kq(kn); - struct kqworkloop *kqwl = (struct kqworkloop *)kq; - int error = 0, knoteuse_flags = KNUSE_NONE; + struct kqworkloop *kqwl = (struct kqworkloop *)knote_get_kq(kn); + int rc = 0; - kqlock_held(kq); + assert(kn->kn_sfflags & NOTE_WL_THREAD_REQUEST); - assert(kev->flags & EV_DELETE); - assert(kq->kq_state & KQ_WORKLOOP); + kqlock(kqwl); - error = filt_wlvalidate_kev_flags(kn, kev); - if (error) { - goto out; + if (kqwl->kqwl_owner) { + /* + * userspace sometimes due to events being + * delivered but not triggering a drain session can cause a process + * of the thread request knote. + * + * When that happens, the automatic deactivation due to process + * would swallow the event, so we have to activate the knote again. + */ + knote_activate(kqwl, kn, FILTER_ACTIVE); + } else { +#if DEBUG || DEVELOPMENT + if (kevent_debug_flags & KEVENT_PANIC_ON_NON_ENQUEUED_PROCESS) { + /* + * see src/queue_internal.h in libdispatch + */ +#define DISPATCH_QUEUE_ENQUEUED 0x1ull + user_addr_t addr = CAST_USER_ADDR_T(kn->kn_ext[EV_EXTIDX_WL_ADDR]); + task_t t = current_task(); + uint64_t val; + if (addr && task_is_active(t) && !task_is_halting(t) && + copyin_atomic64(addr, &val) == 0 && + val && (val & DISPATCH_QUEUE_ENQUEUED) == 0 && + (val >> 48) != 0xdead && (val >> 48) != 0 && (val >> 48) != 0xffff) { + panic("kevent: workloop %#016llx is not enqueued " + "(kn:%p dq_state:%#016llx kev.dq_state:%#016llx)", + kn->kn_udata, kn, val, kn->kn_ext[EV_EXTIDX_WL_VALUE]); + } + } +#endif + knote_fill_kevent(kn, kev, 0); + kev->fflags = kn->kn_sfflags; + rc |= FILTER_ACTIVE; } - if (kn->kn_sfflags & NOTE_WL_THREAD_REQUEST) { - knoteuse_flags |= KNUSE_BOOST; - } + kqunlock(kqwl); - /* take a usecount to allow taking the filt_wllock */ - if (!kqlock2knoteuse(kq, kn, knoteuse_flags)) { - /* knote is being dropped already */ - error = EINPROGRESS; - goto out; + if (rc & FILTER_ACTIVE) { + workq_thread_set_max_qos(kqwl->kqwl_p, &kqwl->kqwl_request); } + return rc; +} - filt_wllock(kqwl); - - /* - * Make sure user and kernel are in agreement on important state - * - * Userland will modify bits to cause this to fail for the touch / drop - * race case (when a drop for a thread request quiescing comes in late after - * the workloop has been woken up again). - */ - error = filt_wldebounce(kqwl, kev, 0); +SECURITY_READ_ONLY_EARLY(static struct filterops) workloop_filtops = { + .f_extended_codes = true, + .f_attach = filt_wlattach, + .f_detach = filt_wldetach, + .f_event = filt_bad_event, + .f_touch = filt_wltouch, + .f_process = filt_wlprocess, + .f_allow_drop = filt_wlallow_drop, + .f_post_register_wait = filt_wlpost_register_wait, +}; - if (!knoteuse2kqlock(kq, kn, knoteuse_flags)) { - /* knote is no longer alive */ - error = EINPROGRESS; - goto out_unlock; - } +#pragma mark - kqueues allocation and deallocation - if (!error && (kn->kn_sfflags & NOTE_WL_THREAD_REQUEST) && kn->kn_inuse) { - /* - * There is a concurrent drop or touch happening, we can't resolve this, - * userland has to redrive. - * - * The race we're worried about here is the following: - * - * f_touch | f_drop_and_unlock - * ------------------------+-------------------------------------------- - * | kqlock() - * | kqlock2knoteuse() - * | filt_wllock() - * | debounces successfully - * kqlock() | - * kqlock2knoteuse | - * filt_wllock() | - * | knoteuse2kqlock() - * | filt_wlunlock() - * | kqlock2knotedrop() - * debounces successfully | - * filt_wlunlock() | - * caller WAKES f_drop | - * | performs drop, but f_touch should have won - * - * So if the usecount is not 0 here, we need to wait for it to drop and - * redrive the whole logic (including looking up the knote again). - */ - filt_wlunlock(kqwl); - knoteusewait(kq, kn); - return ERESTART; - } +/*! + * @enum kqworkloop_dealloc_flags_t + * + * @brief + * Flags that alter kqworkloop_dealloc() behavior. + * + * @const KQWL_DEALLOC_NONE + * Convenient name for "no flags". + * + * @const KQWL_DEALLOC_SKIP_HASH_REMOVE + * Do not remove the workloop fromt he hash table. + * This is used for process tear-down codepaths as the workloops have been + * removed by the caller already. + */ +OS_OPTIONS(kqworkloop_dealloc_flags, unsigned, + KQWL_DEALLOC_NONE = 0x0000, + KQWL_DEALLOC_SKIP_HASH_REMOVE = 0x0001, + ); - /* - * If error is 0 this will set kqr_qos_index to THREAD_QOS_UNSPECIFIED - * - * If error is 0 or ESTALE this may drop ownership and cause a thread - * request redrive, however the kqlock is held which prevents f_process() to - * run until we did the drop for real. - */ - error = filt_wlupdateowner(kqwl, kev, error, 0); - if (error) { - goto out_unlock; - } +static void +kqworkloop_dealloc(struct kqworkloop *, kqworkloop_dealloc_flags_t, uint32_t); - if ((kn->kn_sfflags & (NOTE_WL_SYNC_WAIT | NOTE_WL_SYNC_WAKE)) == - NOTE_WL_SYNC_WAIT) { - /* - * When deleting a SYNC_WAIT knote that hasn't been woken up - * explicitly, issue a wake up. - */ - kn->kn_sfflags |= NOTE_WL_SYNC_WAKE; - if (kn->kn_hook) { - thread_wakeup_thread((event_t)kn, (thread_t)kn->kn_hook); - } +OS_NOINLINE OS_COLD OS_NORETURN +static void +kqworkloop_retain_panic(struct kqworkloop *kqwl, uint32_t previous) +{ + if (previous == 0) { + panic("kq(%p) resurrection", kqwl); + } else { + panic("kq(%p) retain overflow", kqwl); } +} -out_unlock: - filt_wlremember_last_update(kqwl, kn, kev, error); - filt_wlunlock(kqwl); +OS_NOINLINE OS_COLD OS_NORETURN +static void +kqworkloop_release_panic(struct kqworkloop *kqwl) +{ + panic("kq(%p) over-release", kqwl); +} -out: - if (error == 0) { - /* If nothing failed, do the regular knote drop. */ - if (kqlock2knotedrop(kq, kn)) { - knote_drop(kn, current_proc()); - } else { - error = EINPROGRESS; +OS_ALWAYS_INLINE +static inline bool +kqworkloop_try_retain(struct kqworkloop *kqwl) +{ + uint32_t old_ref, new_ref; + os_atomic_rmw_loop(&kqwl->kqwl_retains, old_ref, new_ref, relaxed, { + if (__improbable(old_ref == 0)) { + os_atomic_rmw_loop_give_up(return false); } - } else { - kqunlock(kq); - } - if (error == ESTALE && (kev->fflags & NOTE_WL_IGNORE_ESTALE)) { - error = 0; - } - if (error == EINPROGRESS) { - /* - * filt_wlprocess() makes sure that no event can be delivered for - * NOTE_WL_THREAD_REQUEST knotes once a drop is happening, and - * NOTE_WL_SYNC_* knotes are never fired. - * - * It means that EINPROGRESS is about a state that userland cannot - * observe for this filter (an event being delivered concurrently from - * a drop), so silence the error. - */ - error = 0; - } - return error; + if (__improbable(old_ref >= KQ_WORKLOOP_RETAINS_MAX)) { + kqworkloop_retain_panic(kqwl, old_ref); + } + new_ref = old_ref + 1; + }); + return true; } -static int -filt_wlprocess( - struct knote *kn, - __unused struct filt_process_s *data, - struct kevent_internal_s *kev) +OS_ALWAYS_INLINE +static inline void +kqworkloop_retain(struct kqworkloop *kqwl) { - struct kqueue *kq = knote_get_kq(kn); - struct kqworkloop *kqwl = (struct kqworkloop *)kq; - struct kqrequest *kqr = &kqwl->kqwl_request; - int rc = 0; - - assert(kq->kq_state & KQ_WORKLOOP); - - /* only thread requests should get here */ - assert(kn->kn_sfflags & NOTE_WL_THREAD_REQUEST); - if (kn->kn_sfflags & NOTE_WL_THREAD_REQUEST) { - filt_wllock(kqwl); - assert(kqr->kqr_qos_index != THREAD_QOS_UNSPECIFIED); - if (kqwl->kqwl_owner) { - /* - * userspace sometimes due to events being - * delivered but not triggering a drain session can cause a process - * of the thread request knote. - * - * When that happens, the automatic deactivation due to process - * would swallow the event, so we have to activate the knote again. - */ - kqlock(kq); - knote_activate(kn); - kqunlock(kq); - } else if (kqr->kqr_qos_index) { -#if DEBUG || DEVELOPMENT - user_addr_t addr = CAST_USER_ADDR_T(kn->kn_ext[EV_EXTIDX_WL_ADDR]); - task_t t = current_task(); - uint64_t val; - if (addr && task_is_active(t) && !task_is_halting(t) && - copyin_word(addr, &val, sizeof(val)) == 0 && - val && (val & DISPATCH_QUEUE_ENQUEUED) == 0) { - panic("kevent: workloop %#016llx is not enqueued " - "(kn:%p dq_state:%#016llx kev.dq_state:%#016llx)", - kn->kn_udata, kn, val, - kn->kn_ext[EV_EXTIDX_WL_VALUE]); - } -#endif - *kev = kn->kn_kevent; - kev->fflags = kn->kn_sfflags; - kev->data = kn->kn_sdata; - kev->qos = kn->kn_qos; - rc = 1; - } - filt_wlunlock(kqwl); + uint32_t previous = os_atomic_inc_orig(&kqwl->kqwl_retains, relaxed); + if (__improbable(previous == 0 || previous >= KQ_WORKLOOP_RETAINS_MAX)) { + kqworkloop_retain_panic(kqwl, previous); } - return rc; } -#pragma mark kevent / knotes - -/* - * JMM - placeholder for not-yet-implemented filters - */ -static int -filt_badattach(__unused struct knote *kn, __unused struct kevent_internal_s *kev) +OS_ALWAYS_INLINE +static inline void +kqueue_retain(kqueue_t kqu) { - kn->kn_flags |= EV_ERROR; - kn->kn_data = ENOTSUP; - return 0; + if (kqu.kq->kq_state & KQ_DYNAMIC) { + kqworkloop_retain(kqu.kqwl); + } } -struct kqueue * -kqueue_alloc(struct proc *p, unsigned int flags) +OS_ALWAYS_INLINE +static inline void +kqworkloop_release_live(struct kqworkloop *kqwl) { - struct filedesc *fdp = p->p_fd; - struct kqueue *kq = NULL; - int policy; - void *hook = NULL; - uint64_t kq_addr_offset; - - if (flags & KEVENT_FLAG_WORKQ) { - struct kqworkq *kqwq; - int i; - - kqwq = (struct kqworkq *)zalloc(kqworkq_zone); - if (kqwq == NULL) - return NULL; - - kq = &kqwq->kqwq_kqueue; - bzero(kqwq, sizeof (struct kqworkq)); - - kqwq->kqwq_state = KQ_WORKQ; - - for (i = 0; i < KQWQ_NBUCKETS; i++) { - TAILQ_INIT(&kq->kq_queue[i]); - } - for (i = 0; i < KQWQ_NQOS; i++) { - kqwq->kqwq_request[i].kqr_qos_index = i; - } - - lck_spin_init(&kqwq->kqwq_reqlock, kq_lck_grp, kq_lck_attr); - policy = SYNC_POLICY_FIFO; - hook = (void *)kqwq; - - } else if (flags & KEVENT_FLAG_WORKLOOP) { - struct kqworkloop *kqwl; - int i; - - kqwl = (struct kqworkloop *)zalloc(kqworkloop_zone); - if (kqwl == NULL) - return NULL; - - bzero(kqwl, sizeof (struct kqworkloop)); - - kqwl->kqwl_state = KQ_WORKLOOP | KQ_DYNAMIC; - kqwl->kqwl_retains = 1; /* donate a retain to creator */ - - kq = &kqwl->kqwl_kqueue; - for (i = 0; i < KQWL_NBUCKETS; i++) { - TAILQ_INIT(&kq->kq_queue[i]); - } - TAILQ_INIT(&kqwl->kqwl_request.kqr_suppressed); - - lck_spin_init(&kqwl->kqwl_reqlock, kq_lck_grp, kq_lck_attr); - lck_mtx_init(&kqwl->kqwl_statelock, kq_lck_grp, kq_lck_attr); - - policy = SYNC_POLICY_FIFO; - if (flags & KEVENT_FLAG_WORKLOOP_NO_WQ_THREAD) { - policy |= SYNC_POLICY_PREPOST; - kq->kq_state |= KQ_NO_WQ_THREAD; - } else { - hook = (void *)kqwl; - } - - } else { - struct kqfile *kqf; - - kqf = (struct kqfile *)zalloc(kqfile_zone); - if (kqf == NULL) - return NULL; - - kq = &kqf->kqf_kqueue; - bzero(kqf, sizeof (struct kqfile)); - TAILQ_INIT(&kq->kq_queue[0]); - TAILQ_INIT(&kqf->kqf_suppressed); - - policy = SYNC_POLICY_FIFO | SYNC_POLICY_PREPOST; + uint32_t refs = os_atomic_dec_orig(&kqwl->kqwl_retains, relaxed); + if (__improbable(refs <= 1)) { + kqworkloop_release_panic(kqwl); } +} - waitq_set_init(&kq->kq_wqs, policy, NULL, hook); - lck_spin_init(&kq->kq_lock, kq_lck_grp, kq_lck_attr); - kq->kq_p = p; - - if (fdp->fd_knlistsize < 0) { - proc_fdlock(p); - if (fdp->fd_knlistsize < 0) - fdp->fd_knlistsize = 0; /* this process has had a kq */ - proc_fdunlock(p); +OS_ALWAYS_INLINE +static inline void +kqueue_release_live(kqueue_t kqu) +{ + if (kqu.kq->kq_state & KQ_DYNAMIC) { + kqworkloop_release_live(kqu.kqwl); } - - kq_addr_offset = ((uintptr_t)kq - (uintptr_t)VM_MIN_KERNEL_AND_KEXT_ADDRESS); - /* Assert that the address can be pointer compacted for use with knote */ - assert(kq_addr_offset < (uint64_t)(1ull << KNOTE_KQ_BITSIZE)); - return (kq); } -/* - * knotes_dealloc - detach all knotes for the process and drop them - * - * Called with proc_fdlock held. - * Returns with it locked. - * May drop it temporarily. - * Process is in such a state that it will not try to allocate - * any more knotes during this process (stopped for exit or exec). - */ -void -knotes_dealloc(proc_t p) +OS_ALWAYS_INLINE +static inline void +kqworkloop_release(struct kqworkloop *kqwl) { - struct filedesc *fdp = p->p_fd; - struct kqueue *kq; - struct knote *kn; - struct klist *kn_hash = NULL; - int i; + uint32_t refs = os_atomic_dec_orig(&kqwl->kqwl_retains, relaxed); - /* Close all the fd-indexed knotes up front */ - if (fdp->fd_knlistsize > 0) { - for (i = 0; i < fdp->fd_knlistsize; i++) { - while ((kn = SLIST_FIRST(&fdp->fd_knlist[i])) != NULL) { - kq = knote_get_kq(kn); - kqlock(kq); - proc_fdunlock(p); - /* drop it ourselves or wait */ - if (kqlock2knotedrop(kq, kn)) { - knote_drop(kn, p); - } - proc_fdlock(p); - } - } - /* free the table */ - FREE(fdp->fd_knlist, M_KQUEUE); - fdp->fd_knlist = NULL; + if (__improbable(refs <= 1)) { + kqworkloop_dealloc(kqwl, KQWL_DEALLOC_NONE, refs - 1); } - fdp->fd_knlistsize = -1; - - knhash_lock(p); - proc_fdunlock(p); +} - /* Clean out all the hashed knotes as well */ - if (fdp->fd_knhashmask != 0) { - for (i = 0; i <= (int)fdp->fd_knhashmask; i++) { - while ((kn = SLIST_FIRST(&fdp->fd_knhash[i])) != NULL) { - kq = knote_get_kq(kn); - kqlock(kq); - knhash_unlock(p); - /* drop it ourselves or wait */ - if (kqlock2knotedrop(kq, kn)) { - knote_drop(kn, p); - } - knhash_lock(p); - } - } - kn_hash = fdp->fd_knhash; - fdp->fd_knhashmask = 0; - fdp->fd_knhash = NULL; +OS_ALWAYS_INLINE +static inline void +kqueue_release(kqueue_t kqu) +{ + if (kqu.kq->kq_state & KQ_DYNAMIC) { + kqworkloop_release(kqu.kqwl); } +} - knhash_unlock(p); +/*! + * @function kqueue_destroy + * + * @brief + * Common part to all kqueue dealloc functions. + */ +OS_NOINLINE +static void +kqueue_destroy(kqueue_t kqu, zone_t zone) +{ + /* + * waitq_set_deinit() remove the KQ's waitq set from + * any select sets to which it may belong. + * + * The order of these deinits matter: before waitq_set_deinit() returns, + * waitq_set__CALLING_PREPOST_HOOK__ may be called and it will take the + * kq_lock. + */ + waitq_set_deinit(&kqu.kq->kq_wqs); + lck_spin_destroy(&kqu.kq->kq_lock, &kq_lck_grp); - /* free the kn_hash table */ - if (kn_hash) - FREE(kn_hash, M_KQUEUE); + zfree(zone, kqu.kq); +} - proc_fdlock(p); +/*! + * @function kqueue_init + * + * @brief + * Common part to all kqueue alloc functions. + */ +static kqueue_t +kqueue_init(kqueue_t kqu, waitq_set_prepost_hook_t *hook, int policy) +{ + waitq_set_init(&kqu.kq->kq_wqs, policy, NULL, hook); + lck_spin_init(&kqu.kq->kq_lock, &kq_lck_grp, LCK_ATTR_NULL); + return kqu; } +#pragma mark kqfile allocation and deallocation -/* - * kqueue_dealloc - detach all knotes from a kqueue and free it - * - * We walk each list looking for knotes referencing this - * this kqueue. If we find one, we try to drop it. But - * if we fail to get a drop reference, that will wait - * until it is dropped. So, we can just restart again - * safe in the assumption that the list will eventually - * not contain any more references to this kqueue (either - * we dropped them all, or someone else did). - * - * Assumes no new events are being added to the kqueue. - * Nothing locked on entry or exit. - * - * Workloop kqueues cant get here unless all the knotes - * are already gone and all requested threads have come - * and gone (cancelled or arrived). +/*! + * @function kqueue_dealloc + * + * @brief + * Detach all knotes from a kqfile and free it. + * + * @discussion + * We walk each list looking for knotes referencing this + * this kqueue. If we find one, we try to drop it. But + * if we fail to get a drop reference, that will wait + * until it is dropped. So, we can just restart again + * safe in the assumption that the list will eventually + * not contain any more references to this kqueue (either + * we dropped them all, or someone else did). + * + * Assumes no new events are being added to the kqueue. + * Nothing locked on entry or exit. */ void kqueue_dealloc(struct kqueue *kq) { - struct proc *p; - struct filedesc *fdp; + KNOTE_LOCK_CTX(knlc); + struct proc *p = kq->kq_p; + struct filedesc *fdp = p->p_fd; struct knote *kn; - int i; - - if (kq == NULL) - return; - p = kq->kq_p; - fdp = p->p_fd; + assert(kq && (kq->kq_state & (KQ_WORKLOOP | KQ_WORKQ)) == 0); proc_fdlock(p); - for (i = 0; i < fdp->fd_knlistsize; i++) { + for (int i = 0; i < fdp->fd_knlistsize; i++) { kn = SLIST_FIRST(&fdp->fd_knlist[i]); while (kn != NULL) { if (kq == knote_get_kq(kn)) { - assert((kq->kq_state & KQ_WORKLOOP) == 0); kqlock(kq); proc_fdunlock(p); - /* drop it ourselves or wait */ - if (kqlock2knotedrop(kq, kn)) { - knote_drop(kn, p); + if (knote_lock(kq, kn, &knlc, KNOTE_KQ_LOCK_ON_SUCCESS)) { + knote_drop(kq, kn, &knlc); } proc_fdlock(p); /* start over at beginning of list */ @@ -3055,22 +2944,21 @@ kqueue_dealloc(struct kqueue *kq) kn = SLIST_NEXT(kn, kn_link); } } - knhash_lock(p); + + knhash_lock(fdp); proc_fdunlock(p); if (fdp->fd_knhashmask != 0) { - for (i = 0; i < (int)fdp->fd_knhashmask + 1; i++) { + for (int i = 0; i < (int)fdp->fd_knhashmask + 1; i++) { kn = SLIST_FIRST(&fdp->fd_knhash[i]); while (kn != NULL) { if (kq == knote_get_kq(kn)) { - assert((kq->kq_state & KQ_WORKLOOP) == 0); kqlock(kq); - knhash_unlock(p); - /* drop it ourselves or wait */ - if (kqlock2knotedrop(kq, kn)) { - knote_drop(kn, p); + knhash_unlock(fdp); + if (knote_lock(kq, kn, &knlc, KNOTE_KQ_LOCK_ON_SUCCESS)) { + knote_drop(kq, kn, &knlc); } - knhash_lock(p); + knhash_lock(fdp); /* start over at beginning of list */ kn = SLIST_FIRST(&fdp->fd_knhash[i]); continue; @@ -3079,5515 +2967,5093 @@ kqueue_dealloc(struct kqueue *kq) } } } - knhash_unlock(p); - - if (kq->kq_state & KQ_WORKLOOP) { - struct kqworkloop *kqwl = (struct kqworkloop *)kq; - struct kqrequest *kqr = &kqwl->kqwl_request; - thread_t cur_owner = kqwl->kqwl_owner; - - assert(TAILQ_EMPTY(&kqwl->kqwl_request.kqr_suppressed)); - if (filt_wlowner_is_valid(cur_owner)) { - /* - * If the kqueue had an owner that prevented the thread request to - * go through, then no unbind happened, and we may have lingering - * overrides to drop. - */ - if (kqr->kqr_dsync_owner_qos != THREAD_QOS_UNSPECIFIED) { - thread_drop_ipc_override(cur_owner); - kqr->kqr_dsync_owner_qos = THREAD_QOS_UNSPECIFIED; - } - - if (kqr->kqr_owner_override_is_sync) { - thread_drop_sync_ipc_override(cur_owner); - kqr->kqr_owner_override_is_sync = 0; - } - thread_ends_owning_workloop(cur_owner); - thread_deallocate(cur_owner); - kqwl->kqwl_owner = THREAD_NULL; - } - } - - /* - * waitq_set_deinit() remove the KQ's waitq set from - * any select sets to which it may belong. - */ - waitq_set_deinit(&kq->kq_wqs); - lck_spin_destroy(&kq->kq_lock, kq_lck_grp); - - if (kq->kq_state & KQ_WORKQ) { - struct kqworkq *kqwq = (struct kqworkq *)kq; - - lck_spin_destroy(&kqwq->kqwq_reqlock, kq_lck_grp); - zfree(kqworkq_zone, kqwq); - } else if (kq->kq_state & KQ_WORKLOOP) { - struct kqworkloop *kqwl = (struct kqworkloop *)kq; - - assert(kqwl->kqwl_retains == 0); - lck_spin_destroy(&kqwl->kqwl_reqlock, kq_lck_grp); - lck_mtx_destroy(&kqwl->kqwl_statelock, kq_lck_grp); - zfree(kqworkloop_zone, kqwl); - } else { - struct kqfile *kqf = (struct kqfile *)kq; - - zfree(kqfile_zone, kqf); - } -} + knhash_unlock(fdp); -static inline void -kqueue_retain(struct kqueue *kq) -{ - struct kqworkloop *kqwl = (struct kqworkloop *)kq; - uint32_t previous; - - if ((kq->kq_state & KQ_DYNAMIC) == 0) - return; - - previous = OSIncrementAtomic(&kqwl->kqwl_retains); - if (previous == KQ_WORKLOOP_RETAINS_MAX) - panic("kq(%p) retain overflow", kq); - - if (previous == 0) - panic("kq(%p) resurrection", kq); + kqueue_destroy(kq, kqfile_zone); } -#define KQUEUE_CANT_BE_LAST_REF 0 -#define KQUEUE_MIGHT_BE_LAST_REF 1 - -static inline int -kqueue_release(struct kqueue *kq, __assert_only int possibly_last) +/*! + * @function kqueue_alloc + * + * @brief + * Allocate a kqfile. + */ +struct kqueue * +kqueue_alloc(struct proc *p) { - struct kqworkloop *kqwl = (struct kqworkloop *)kq; + struct kqfile *kqf; - if ((kq->kq_state & KQ_DYNAMIC) == 0) { - return 0; - } + /* + * kqfiles are created with kqueue() so we need to wait for + * the first kevent syscall to know which bit among + * KQ_KEV_{32,64,QOS} will be set in kqf_state + */ + kqf = zalloc_flags(kqfile_zone, Z_WAITOK | Z_ZERO); + kqf->kqf_p = p; + TAILQ_INIT_AFTER_BZERO(&kqf->kqf_queue); + TAILQ_INIT_AFTER_BZERO(&kqf->kqf_suppressed); - assert(kq->kq_state & KQ_WORKLOOP); /* for now */ - uint32_t refs = OSDecrementAtomic(&kqwl->kqwl_retains); - if (__improbable(refs == 0)) { - panic("kq(%p) over-release", kq); - } - if (refs == 1) { - assert(possibly_last); - } - return refs == 1; + return kqueue_init(kqf, NULL, SYNC_POLICY_FIFO | SYNC_POLICY_PREPOST).kq; } +/*! + * @function kqueue_internal + * + * @brief + * Core implementation for kqueue and guarded_kqueue_np() + */ int -kqueue_body(struct proc *p, fp_allocfn_t fp_zalloc, void *cra, int32_t *retval) +kqueue_internal(struct proc *p, fp_allocfn_t fp_zalloc, void *cra, int32_t *retval) { struct kqueue *kq; struct fileproc *fp; int fd, error; - error = falloc_withalloc(p, - &fp, &fd, vfs_context_current(), fp_zalloc, cra); + error = falloc_withalloc(p, &fp, &fd, vfs_context_current(), fp_zalloc, cra); if (error) { - return (error); + return error; } - kq = kqueue_alloc(p, 0); + kq = kqueue_alloc(p); if (kq == NULL) { fp_free(p, fd, fp); - return (ENOMEM); + return ENOMEM; } fp->f_flag = FREAD | FWRITE; fp->f_ops = &kqueueops; fp->f_data = kq; + fp->f_lflags |= FG_CONFINED; proc_fdlock(p); - *fdflags(p, fd) |= UF_EXCLOSE; + *fdflags(p, fd) |= UF_EXCLOSE | UF_FORKCLOSE; procfdtbl_releasefd(p, fd, NULL); fp_drop(p, fd, fp, 1); proc_fdunlock(p); *retval = fd; - return (error); + return error; } +/*! + * @function kqueue + * + * @brief + * The kqueue syscall. + */ int kqueue(struct proc *p, __unused struct kqueue_args *uap, int32_t *retval) { - return (kqueue_body(p, fileproc_alloc_init, NULL, retval)); + return kqueue_internal(p, fileproc_alloc_init, NULL, retval); } -static int -kevent_copyin(user_addr_t *addrp, struct kevent_internal_s *kevp, struct proc *p, - unsigned int flags) +#pragma mark kqworkq allocation and deallocation + +/*! + * @function kqworkq_dealloc + * + * @brief + * Deallocates a workqueue kqueue. + * + * @discussion + * This only happens at process death, or for races with concurrent + * kevent_get_kqwq calls, hence we don't have to care about knotes referencing + * this kqueue, either there are none, or someone else took care of them. + */ +void +kqworkq_dealloc(struct kqworkq *kqwq) { - int advance; - int error; + kqueue_destroy(kqwq, kqworkq_zone); +} - if (flags & KEVENT_FLAG_LEGACY32) { - bzero(kevp, sizeof (*kevp)); - - if (IS_64BIT_PROCESS(p)) { - struct user64_kevent kev64; - - advance = sizeof (kev64); - error = copyin(*addrp, (caddr_t)&kev64, advance); - if (error) - return (error); - kevp->ident = kev64.ident; - kevp->filter = kev64.filter; - kevp->flags = kev64.flags; - kevp->udata = kev64.udata; - kevp->fflags = kev64.fflags; - kevp->data = kev64.data; - } else { - struct user32_kevent kev32; - - advance = sizeof (kev32); - error = copyin(*addrp, (caddr_t)&kev32, advance); - if (error) - return (error); - kevp->ident = (uintptr_t)kev32.ident; - kevp->filter = kev32.filter; - kevp->flags = kev32.flags; - kevp->udata = CAST_USER_ADDR_T(kev32.udata); - kevp->fflags = kev32.fflags; - kevp->data = (intptr_t)kev32.data; - } - } else if (flags & KEVENT_FLAG_LEGACY64) { - struct kevent64_s kev64; +/*! + * @function kqworkq_alloc + * + * @brief + * Allocates a workqueue kqueue. + * + * @discussion + * This is the slow path of kevent_get_kqwq. + * This takes care of making sure procs have a single workq kqueue. + */ +OS_NOINLINE +static struct kqworkq * +kqworkq_alloc(struct proc *p, unsigned int flags) +{ + struct kqworkq *kqwq, *tmp; + + kqwq = zalloc_flags(kqworkq_zone, Z_WAITOK | Z_ZERO); - bzero(kevp, sizeof (*kevp)); - - advance = sizeof (struct kevent64_s); - error = copyin(*addrp, (caddr_t)&kev64, advance); - if (error) - return(error); - kevp->ident = kev64.ident; - kevp->filter = kev64.filter; - kevp->flags = kev64.flags; - kevp->udata = kev64.udata; - kevp->fflags = kev64.fflags; - kevp->data = kev64.data; - kevp->ext[0] = kev64.ext[0]; - kevp->ext[1] = kev64.ext[1]; - + assert((flags & KEVENT_FLAG_LEGACY32) == 0); + if (flags & KEVENT_FLAG_LEGACY64) { + kqwq->kqwq_state = KQ_WORKQ | KQ_KEV64; } else { - struct kevent_qos_s kevqos; + kqwq->kqwq_state = KQ_WORKQ | KQ_KEV_QOS; + } + kqwq->kqwq_p = p; - bzero(kevp, sizeof (*kevp)); + for (int i = 0; i < KQWQ_NBUCKETS; i++) { + TAILQ_INIT_AFTER_BZERO(&kqwq->kqwq_queue[i]); + TAILQ_INIT_AFTER_BZERO(&kqwq->kqwq_suppressed[i]); + } + for (int i = 0; i < KQWQ_NBUCKETS; i++) { + /* + * Because of how the bucketized system works, we mix overcommit + * sources with not overcommit: each time we move a knote from + * one bucket to the next due to overrides, we'd had to track + * overcommitness, and it's really not worth it in the workloop + * enabled world that track this faithfully. + * + * Incidentally, this behaves like the original manager-based + * kqwq where event delivery always happened (hence is + * "overcommit") + */ + kqwq->kqwq_request[i].tr_state = WORKQ_TR_STATE_IDLE; + kqwq->kqwq_request[i].tr_flags = WORKQ_TR_FLAG_KEVENT; + if (i != KQWQ_QOS_MANAGER) { + kqwq->kqwq_request[i].tr_flags |= WORKQ_TR_FLAG_OVERCOMMIT; + } + kqwq->kqwq_request[i].tr_kq_qos_index = (kq_index_t)i; + } - advance = sizeof (struct kevent_qos_s); - error = copyin(*addrp, (caddr_t)&kevqos, advance); - if (error) - return error; - kevp->ident = kevqos.ident; - kevp->filter = kevqos.filter; - kevp->flags = kevqos.flags; - kevp->qos = kevqos.qos; -// kevp->xflags = kevqos.xflags; - kevp->udata = kevqos.udata; - kevp->fflags = kevqos.fflags; - kevp->data = kevqos.data; - kevp->ext[0] = kevqos.ext[0]; - kevp->ext[1] = kevqos.ext[1]; - kevp->ext[2] = kevqos.ext[2]; - kevp->ext[3] = kevqos.ext[3]; - } - if (!error) - *addrp += advance; - return (error); -} + kqueue_init(kqwq, &kqwq->kqwq_waitq_hook, SYNC_POLICY_FIFO); -static int -kevent_copyout(struct kevent_internal_s *kevp, user_addr_t *addrp, struct proc *p, - unsigned int flags) -{ - user_addr_t addr = *addrp; - int advance; - int error; + if (!os_atomic_cmpxchgv(&p->p_fd->fd_wqkqueue, NULL, kqwq, &tmp, release)) { + kqworkq_dealloc(kqwq); + return tmp; + } - /* - * fully initialize the differnt output event structure - * types from the internal kevent (and some universal - * defaults for fields not represented in the internal - * form). - */ - if (flags & KEVENT_FLAG_LEGACY32) { - assert((flags & KEVENT_FLAG_STACK_EVENTS) == 0); + return kqwq; +} - if (IS_64BIT_PROCESS(p)) { - struct user64_kevent kev64; +#pragma mark kqworkloop allocation and deallocation - advance = sizeof (kev64); - bzero(&kev64, advance); - - /* - * deal with the special case of a user-supplied - * value of (uintptr_t)-1. - */ - kev64.ident = (kevp->ident == (uintptr_t)-1) ? - (uint64_t)-1LL : (uint64_t)kevp->ident; - - kev64.filter = kevp->filter; - kev64.flags = kevp->flags; - kev64.fflags = kevp->fflags; - kev64.data = (int64_t) kevp->data; - kev64.udata = kevp->udata; - error = copyout((caddr_t)&kev64, addr, advance); - } else { - struct user32_kevent kev32; - - advance = sizeof (kev32); - bzero(&kev32, advance); - kev32.ident = (uint32_t)kevp->ident; - kev32.filter = kevp->filter; - kev32.flags = kevp->flags; - kev32.fflags = kevp->fflags; - kev32.data = (int32_t)kevp->data; - kev32.udata = kevp->udata; - error = copyout((caddr_t)&kev32, addr, advance); - } - } else if (flags & KEVENT_FLAG_LEGACY64) { - struct kevent64_s kev64; +#define KQ_HASH(val, mask) (((val) ^ (val >> 8)) & (mask)) +#define CONFIG_KQ_HASHSIZE CONFIG_KN_HASHSIZE - advance = sizeof (struct kevent64_s); - if (flags & KEVENT_FLAG_STACK_EVENTS) { - addr -= advance; - } - bzero(&kev64, advance); - kev64.ident = kevp->ident; - kev64.filter = kevp->filter; - kev64.flags = kevp->flags; - kev64.fflags = kevp->fflags; - kev64.data = (int64_t) kevp->data; - kev64.udata = kevp->udata; - kev64.ext[0] = kevp->ext[0]; - kev64.ext[1] = kevp->ext[1]; - error = copyout((caddr_t)&kev64, addr, advance); - } else { - struct kevent_qos_s kevqos; - - advance = sizeof (struct kevent_qos_s); - if (flags & KEVENT_FLAG_STACK_EVENTS) { - addr -= advance; - } - bzero(&kevqos, advance); - kevqos.ident = kevp->ident; - kevqos.filter = kevp->filter; - kevqos.flags = kevp->flags; - kevqos.qos = kevp->qos; - kevqos.udata = kevp->udata; - kevqos.fflags = kevp->fflags; - kevqos.xflags = 0; - kevqos.data = (int64_t) kevp->data; - kevqos.ext[0] = kevp->ext[0]; - kevqos.ext[1] = kevp->ext[1]; - kevqos.ext[2] = kevp->ext[2]; - kevqos.ext[3] = kevp->ext[3]; - error = copyout((caddr_t)&kevqos, addr, advance); - } - if (!error) { - if (flags & KEVENT_FLAG_STACK_EVENTS) - *addrp = addr; - else - *addrp = addr + advance; - } - return (error); +OS_ALWAYS_INLINE +static inline void +kqhash_lock(struct filedesc *fdp) +{ + lck_mtx_lock_spin_always(&fdp->fd_kqhashlock); } -static int -kevent_get_data_size(struct proc *p, - uint64_t data_available, - unsigned int flags, - user_size_t *residp) +OS_ALWAYS_INLINE +static inline void +kqhash_unlock(struct filedesc *fdp) { - user_size_t resid; - int error = 0; + lck_mtx_unlock(&fdp->fd_kqhashlock); +} - if (data_available != USER_ADDR_NULL) { - if (flags & KEVENT_FLAG_KERNEL) { - resid = *(user_size_t *)(uintptr_t)data_available; - } else if (IS_64BIT_PROCESS(p)) { - user64_size_t usize; - error = copyin((user_addr_t)data_available, &usize, sizeof(usize)); - resid = (user_size_t)usize; - } else { - user32_size_t usize; - error = copyin((user_addr_t)data_available, &usize, sizeof(usize)); - resid = (user_size_t)usize; - } - if (error) - return(error); - } else { - resid = 0; - } - *residp = resid; - return 0; +OS_ALWAYS_INLINE +static inline void +kqworkloop_hash_insert_locked(struct filedesc *fdp, kqueue_id_t id, + struct kqworkloop *kqwl) +{ + struct kqwllist *list = &fdp->fd_kqhash[KQ_HASH(id, fdp->fd_kqhashmask)]; + LIST_INSERT_HEAD(list, kqwl, kqwl_hashlink); } -static int -kevent_put_data_size(struct proc *p, - uint64_t data_available, - unsigned int flags, - user_size_t resid) +OS_ALWAYS_INLINE +static inline struct kqworkloop * +kqworkloop_hash_lookup_locked(struct filedesc *fdp, kqueue_id_t id) { - int error = 0; + struct kqwllist *list = &fdp->fd_kqhash[KQ_HASH(id, fdp->fd_kqhashmask)]; + struct kqworkloop *kqwl; - if (data_available) { - if (flags & KEVENT_FLAG_KERNEL) { - *(user_size_t *)(uintptr_t)data_available = resid; - } else if (IS_64BIT_PROCESS(p)) { - user64_size_t usize = (user64_size_t)resid; - error = copyout(&usize, (user_addr_t)data_available, sizeof(usize)); - } else { - user32_size_t usize = (user32_size_t)resid; - error = copyout(&usize, (user_addr_t)data_available, sizeof(usize)); + LIST_FOREACH(kqwl, list, kqwl_hashlink) { + if (kqwl->kqwl_dynamicid == id) { + return kqwl; } } - return error; + return NULL; } -/* - * kevent_continue - continue a kevent syscall after blocking - * - * assume we inherit a use count on the kq fileglob. - */ +static struct kqworkloop * +kqworkloop_hash_lookup_and_retain(struct filedesc *fdp, kqueue_id_t kq_id) +{ + struct kqworkloop *kqwl = NULL; + + kqhash_lock(fdp); + if (__probable(fdp->fd_kqhash)) { + kqwl = kqworkloop_hash_lookup_locked(fdp, kq_id); + if (kqwl && !kqworkloop_try_retain(kqwl)) { + kqwl = NULL; + } + } + kqhash_unlock(fdp); + return kqwl; +} -__attribute__((noreturn)) +OS_NOINLINE static void -kevent_continue(__unused struct kqueue *kq, void *data, int error) +kqworkloop_hash_init(struct filedesc *fdp) { - struct _kevent *cont_args; - struct fileproc *fp; - uint64_t data_available; - user_size_t data_size; - user_size_t data_resid; - unsigned int flags; - int32_t *retval; - int noutputs; - int fd; - struct proc *p = current_proc(); - - cont_args = (struct _kevent *)data; - data_available = cont_args->data_available; - flags = cont_args->process_data.fp_flags; - data_size = cont_args->process_data.fp_data_size; - data_resid = cont_args->process_data.fp_data_resid; - noutputs = cont_args->eventout; - retval = cont_args->retval; - fd = cont_args->fd; - fp = cont_args->fp; - - kevent_put_kq(p, fd, fp, kq); - - /* don't abandon other output just because of residual copyout failures */ - if (error == 0 && data_available && data_resid != data_size) { - (void)kevent_put_data_size(p, data_available, flags, data_resid); - } + struct kqwllist *alloc_hash; + u_long alloc_mask; - /* don't restart after signals... */ - if (error == ERESTART) - error = EINTR; - else if (error == EWOULDBLOCK) - error = 0; - if (error == 0) - *retval = noutputs; - unix_syscall_return(error); + kqhash_unlock(fdp); + alloc_hash = hashinit(CONFIG_KQ_HASHSIZE, M_KQUEUE, &alloc_mask); + kqhash_lock(fdp); + + /* See if we won the race */ + if (__probable(fdp->fd_kqhashmask == 0)) { + fdp->fd_kqhash = alloc_hash; + fdp->fd_kqhashmask = alloc_mask; + } else { + kqhash_unlock(fdp); + hashdestroy(alloc_hash, M_KQUEUE, alloc_mask); + kqhash_lock(fdp); + } } -/* - * kevent - [syscall] register and wait for kernel events +/*! + * @function kqworkloop_dealloc * + * @brief + * Deallocates a workloop kqueue. + * + * @discussion + * Knotes hold references on the workloop, so we can't really reach this + * function unless all of these are already gone. + * + * Nothing locked on entry or exit. + * + * @param flags + * Unless KQWL_DEALLOC_SKIP_HASH_REMOVE is set, the workloop is removed + * from its hash table. + * + * @param current_ref + * This function is also called to undo a kqworkloop_alloc in case of + * allocation races, expected_ref is the current refcount that is expected + * on the workloop object, usually 0, and 1 when a dealloc race is resolved. */ -int -kevent(struct proc *p, struct kevent_args *uap, int32_t *retval) +static void +kqworkloop_dealloc(struct kqworkloop *kqwl, kqworkloop_dealloc_flags_t flags, + uint32_t current_ref) { - unsigned int flags = KEVENT_FLAG_LEGACY32; + thread_t cur_owner; - return kevent_internal(p, - (kqueue_id_t)uap->fd, NULL, - uap->changelist, uap->nchanges, - uap->eventlist, uap->nevents, - 0ULL, 0ULL, - flags, - uap->timeout, - kevent_continue, - retval); -} + if (__improbable(current_ref > 1)) { + kqworkloop_release_panic(kqwl); + } + assert(kqwl->kqwl_retains == current_ref); -int -kevent64(struct proc *p, struct kevent64_args *uap, int32_t *retval) -{ - unsigned int flags; + /* pair with kqunlock() and other kq locks */ + os_atomic_thread_fence(acquire); + + cur_owner = kqwl->kqwl_owner; + if (cur_owner) { + if (kqworkloop_override(kqwl) != THREAD_QOS_UNSPECIFIED) { + thread_drop_kevent_override(cur_owner); + } + thread_deallocate(cur_owner); + kqwl->kqwl_owner = THREAD_NULL; + } - /* restrict to user flags and set legacy64 */ - flags = uap->flags & KEVENT_FLAG_USER; - flags |= KEVENT_FLAG_LEGACY64; + if (kqwl->kqwl_state & KQ_HAS_TURNSTILE) { + struct turnstile *ts; + turnstile_complete((uintptr_t)kqwl, &kqwl->kqwl_turnstile, + &ts, TURNSTILE_WORKLOOPS); + turnstile_cleanup(); + turnstile_deallocate(ts); + } - return kevent_internal(p, - (kqueue_id_t)uap->fd, NULL, - uap->changelist, uap->nchanges, - uap->eventlist, uap->nevents, - 0ULL, 0ULL, - flags, - uap->timeout, - kevent_continue, - retval); -} + if ((flags & KQWL_DEALLOC_SKIP_HASH_REMOVE) == 0) { + struct filedesc *fdp = kqwl->kqwl_p->p_fd; -int -kevent_qos(struct proc *p, struct kevent_qos_args *uap, int32_t *retval) -{ - /* restrict to user flags */ - uap->flags &= KEVENT_FLAG_USER; - - return kevent_internal(p, - (kqueue_id_t)uap->fd, NULL, - uap->changelist, uap->nchanges, - uap->eventlist, uap->nevents, - uap->data_out, (uint64_t)uap->data_available, - uap->flags, - 0ULL, - kevent_continue, - retval); -} - -int -kevent_qos_internal(struct proc *p, int fd, - user_addr_t changelist, int nchanges, - user_addr_t eventlist, int nevents, - user_addr_t data_out, user_size_t *data_available, - unsigned int flags, - int32_t *retval) -{ - return kevent_internal(p, - (kqueue_id_t)fd, NULL, - changelist, nchanges, - eventlist, nevents, - data_out, (uint64_t)data_available, - (flags | KEVENT_FLAG_KERNEL), - 0ULL, - NULL, - retval); + kqhash_lock(fdp); + LIST_REMOVE(kqwl, kqwl_hashlink); + kqhash_unlock(fdp); + } + + assert(TAILQ_EMPTY(&kqwl->kqwl_suppressed)); + assert(kqwl->kqwl_owner == THREAD_NULL); + assert(kqwl->kqwl_turnstile == TURNSTILE_NULL); + + lck_spin_destroy(&kqwl->kqwl_statelock, &kq_lck_grp); + kqueue_destroy(kqwl, kqworkloop_zone); } -int -kevent_id(struct proc *p, struct kevent_id_args *uap, int32_t *retval) +/*! + * @function kqworkloop_alloc + * + * @brief + * Allocates a workloop kqueue. + */ +static void +kqworkloop_init(struct kqworkloop *kqwl, proc_t p, + kqueue_id_t id, workq_threadreq_param_t *trp) { - /* restrict to user flags */ - uap->flags &= KEVENT_FLAG_USER; + kqwl->kqwl_state = KQ_WORKLOOP | KQ_DYNAMIC | KQ_KEV_QOS; + kqwl->kqwl_retains = 1; /* donate a retain to creator */ + kqwl->kqwl_dynamicid = id; + kqwl->kqwl_p = p; + if (trp) { + kqwl->kqwl_params = trp->trp_value; + } + + workq_tr_flags_t tr_flags = WORKQ_TR_FLAG_WORKLOOP; + if (trp) { + if (trp->trp_flags & TRP_PRIORITY) { + tr_flags |= WORKQ_TR_FLAG_WL_OUTSIDE_QOS; + } + if (trp->trp_flags) { + tr_flags |= WORKQ_TR_FLAG_WL_PARAMS; + } + } + kqwl->kqwl_request.tr_state = WORKQ_TR_STATE_IDLE; + kqwl->kqwl_request.tr_flags = tr_flags; + + for (int i = 0; i < KQWL_NBUCKETS; i++) { + TAILQ_INIT_AFTER_BZERO(&kqwl->kqwl_queue[i]); + } + TAILQ_INIT_AFTER_BZERO(&kqwl->kqwl_suppressed); + + lck_spin_init(&kqwl->kqwl_statelock, &kq_lck_grp, LCK_ATTR_NULL); - return kevent_internal(p, - (kqueue_id_t)uap->id, NULL, - uap->changelist, uap->nchanges, - uap->eventlist, uap->nevents, - uap->data_out, (uint64_t)uap->data_available, - (uap->flags | KEVENT_FLAG_DYNAMIC_KQUEUE), - 0ULL, - kevent_continue, - retval); + kqueue_init(kqwl, &kqwl->kqwl_waitq_hook, SYNC_POLICY_FIFO); } -int -kevent_id_internal(struct proc *p, kqueue_id_t *id, - user_addr_t changelist, int nchanges, - user_addr_t eventlist, int nevents, - user_addr_t data_out, user_size_t *data_available, - unsigned int flags, - int32_t *retval) -{ - return kevent_internal(p, - *id, id, - changelist, nchanges, - eventlist, nevents, - data_out, (uint64_t)data_available, - (flags | KEVENT_FLAG_KERNEL | KEVENT_FLAG_DYNAMIC_KQUEUE), - 0ULL, - NULL, - retval); -} - +/*! + * @function kqworkloop_get_or_create + * + * @brief + * Wrapper around kqworkloop_alloc that handles the uniquing of workloops. + * + * @returns + * 0: success + * EINVAL: invalid parameters + * EEXIST: KEVENT_FLAG_DYNAMIC_KQ_MUST_NOT_EXIST is set and a collision exists. + * ENOENT: KEVENT_FLAG_DYNAMIC_KQ_MUST_EXIST is set and the entry wasn't found. + * ENOMEM: allocation failed + */ static int -kevent_get_timeout(struct proc *p, - user_addr_t utimeout, - unsigned int flags, - struct timeval *atvp) +kqworkloop_get_or_create(struct proc *p, kqueue_id_t id, + workq_threadreq_param_t *trp, unsigned int flags, struct kqworkloop **kqwlp) { - struct timeval atv; + struct filedesc *fdp = p->p_fd; + struct kqworkloop *alloc_kqwl = NULL; + struct kqworkloop *kqwl = NULL; int error = 0; - if (flags & KEVENT_FLAG_IMMEDIATE) { - getmicrouptime(&atv); - } else if (utimeout != USER_ADDR_NULL) { - struct timeval rtv; - if (flags & KEVENT_FLAG_KERNEL) { - struct timespec *tsp = (struct timespec *)utimeout; - TIMESPEC_TO_TIMEVAL(&rtv, tsp); - } else if (IS_64BIT_PROCESS(p)) { - struct user64_timespec ts; - error = copyin(utimeout, &ts, sizeof(ts)); - if ((ts.tv_sec & 0xFFFFFFFF00000000ull) != 0) - error = EINVAL; - else - TIMESPEC_TO_TIMEVAL(&rtv, &ts); - } else { - struct user32_timespec ts; - error = copyin(utimeout, &ts, sizeof(ts)); - TIMESPEC_TO_TIMEVAL(&rtv, &ts); - } - if (error) - return (error); - if (itimerfix(&rtv)) - return (EINVAL); - getmicrouptime(&atv); - timevaladd(&atv, &rtv); - } else { - /* wait forever value */ - atv.tv_sec = 0; - atv.tv_usec = 0; + assert(!trp || (flags & KEVENT_FLAG_DYNAMIC_KQ_MUST_NOT_EXIST)); + + if (id == 0 || id == (kqueue_id_t)-1) { + return EINVAL; } - *atvp = atv; - return 0; -} -static int -kevent_set_kq_mode(struct kqueue *kq, unsigned int flags) -{ - /* each kq should only be used for events of one type */ - kqlock(kq); - if (kq->kq_state & (KQ_KEV32 | KQ_KEV64 | KQ_KEV_QOS)) { - if (flags & KEVENT_FLAG_LEGACY32) { - if ((kq->kq_state & KQ_KEV32) == 0) { - kqunlock(kq); - return EINVAL; + for (;;) { + kqhash_lock(fdp); + if (__improbable(fdp->fd_kqhash == NULL)) { + kqworkloop_hash_init(fdp); + } + + kqwl = kqworkloop_hash_lookup_locked(fdp, id); + if (kqwl) { + if (__improbable(flags & KEVENT_FLAG_DYNAMIC_KQ_MUST_NOT_EXIST)) { + /* + * If MUST_NOT_EXIST was passed, even if we would have failed + * the try_retain, it could have gone the other way, and + * userspace can't tell. Let'em fix their race. + */ + error = EEXIST; + break; + } + + if (__probable(kqworkloop_try_retain(kqwl))) { + /* + * This is a valid live workloop ! + */ + *kqwlp = kqwl; + error = 0; + break; } - } else if (kq->kq_state & KQ_KEV32) { - kqunlock(kq); - return EINVAL; } - } else if (flags & KEVENT_FLAG_LEGACY32) { - kq->kq_state |= KQ_KEV32; - } else if (flags & KEVENT_FLAG_LEGACY64) { - kq->kq_state |= KQ_KEV64; - } else { - kq->kq_state |= KQ_KEV_QOS; + + if (__improbable(flags & KEVENT_FLAG_DYNAMIC_KQ_MUST_EXIST)) { + error = ENOENT; + break; + } + + /* + * We didn't find what we were looking for. + * + * If this is the second time we reach this point (alloc_kqwl != NULL), + * then we're done. + * + * If this is the first time we reach this point (alloc_kqwl == NULL), + * then try to allocate one without blocking. + */ + if (__probable(alloc_kqwl == NULL)) { + alloc_kqwl = zalloc_flags(kqworkloop_zone, Z_NOWAIT | Z_ZERO); + } + if (__probable(alloc_kqwl)) { + kqworkloop_init(alloc_kqwl, p, id, trp); + kqworkloop_hash_insert_locked(fdp, id, alloc_kqwl); + kqhash_unlock(fdp); + *kqwlp = alloc_kqwl; + return 0; + } + + /* + * We have to block to allocate a workloop, drop the lock, + * allocate one, but then we need to retry lookups as someone + * else could race with us. + */ + kqhash_unlock(fdp); + + alloc_kqwl = zalloc_flags(kqworkloop_zone, Z_WAITOK | Z_ZERO); } - kqunlock(kq); - return 0; + + kqhash_unlock(fdp); + + if (__improbable(alloc_kqwl)) { + zfree(kqworkloop_zone, alloc_kqwl); + } + + return error; } -#define KQ_HASH(val, mask) (((val) ^ (val >> 8)) & (mask)) -#define CONFIG_KQ_HASHSIZE CONFIG_KN_HASHSIZE +#pragma mark - knotes -static inline void -kqhash_lock(proc_t p) +static int +filt_no_attach(struct knote *kn, __unused struct kevent_qos_s *kev) { - lck_mtx_lock_spin_always(&p->p_fd->fd_kqhashlock); + knote_set_error(kn, ENOTSUP); + return 0; } -static inline void -kqhash_lock_held(__assert_only proc_t p) +static void +filt_no_detach(__unused struct knote *kn) { - LCK_MTX_ASSERT(&p->p_fd->fd_kqhashlock, LCK_MTX_ASSERT_OWNED); } -static inline void -kqhash_unlock(proc_t p) +static int __dead2 +filt_bad_event(struct knote *kn, long hint) { - lck_mtx_unlock(&p->p_fd->fd_kqhashlock); + panic("%s[%d](%p, %ld)", __func__, kn->kn_filter, kn, hint); } -static void -kqueue_hash_init_if_needed(proc_t p) +static int __dead2 +filt_bad_touch(struct knote *kn, struct kevent_qos_s *kev) { - struct filedesc *fdp = p->p_fd; - - kqhash_lock_held(p); - - if (__improbable(fdp->fd_kqhash == NULL)) { - struct kqlist *alloc_hash; - u_long alloc_mask; - - kqhash_unlock(p); - alloc_hash = hashinit(CONFIG_KQ_HASHSIZE, M_KQUEUE, &alloc_mask); - kqhash_lock(p); + panic("%s[%d](%p, %p)", __func__, kn->kn_filter, kn, kev); +} - /* See if we won the race */ - if (fdp->fd_kqhashmask == 0) { - fdp->fd_kqhash = alloc_hash; - fdp->fd_kqhashmask = alloc_mask; - } else { - kqhash_unlock(p); - FREE(alloc_hash, M_KQUEUE); - kqhash_lock(p); - } - } +static int __dead2 +filt_bad_process(struct knote *kn, struct kevent_qos_s *kev) +{ + panic("%s[%d](%p, %p)", __func__, kn->kn_filter, kn, kev); } /* - * Called with the kqhash_lock() held + * knotes_dealloc - detach all knotes for the process and drop them + * + * Called with proc_fdlock held. + * Returns with it locked. + * May drop it temporarily. + * Process is in such a state that it will not try to allocate + * any more knotes during this process (stopped for exit or exec). */ -static void -kqueue_hash_insert( - struct proc *p, - kqueue_id_t id, - struct kqueue *kq) +void +knotes_dealloc(proc_t p) { - struct kqworkloop *kqwl = (struct kqworkloop *)kq; struct filedesc *fdp = p->p_fd; - struct kqlist *list; + struct kqueue *kq; + struct knote *kn; + struct klist *kn_hash = NULL; + u_long kn_hashmask; + int i; + + /* Close all the fd-indexed knotes up front */ + if (fdp->fd_knlistsize > 0) { + for (i = 0; i < fdp->fd_knlistsize; i++) { + while ((kn = SLIST_FIRST(&fdp->fd_knlist[i])) != NULL) { + kq = knote_get_kq(kn); + kqlock(kq); + proc_fdunlock(p); + knote_drop(kq, kn, NULL); + proc_fdlock(p); + } + } + /* free the table */ + kheap_free(KM_KQUEUE, fdp->fd_knlist, + fdp->fd_knlistsize * sizeof(struct klist *)); + } + fdp->fd_knlistsize = 0; - /* should hold the kq hash lock */ - kqhash_lock_held(p); + knhash_lock(fdp); + proc_fdunlock(p); - if ((kq->kq_state & KQ_DYNAMIC) == 0) { - assert(kq->kq_state & KQ_DYNAMIC); - return; + /* Clean out all the hashed knotes as well */ + if (fdp->fd_knhashmask != 0) { + for (i = 0; i <= (int)fdp->fd_knhashmask; i++) { + while ((kn = SLIST_FIRST(&fdp->fd_knhash[i])) != NULL) { + kq = knote_get_kq(kn); + kqlock(kq); + knhash_unlock(fdp); + knote_drop(kq, kn, NULL); + knhash_lock(fdp); + } + } + kn_hash = fdp->fd_knhash; + kn_hashmask = fdp->fd_knhashmask; + fdp->fd_knhashmask = 0; + fdp->fd_knhash = NULL; } - /* only dynamically allocate workloop kqs for now */ - assert(kq->kq_state & KQ_WORKLOOP); - assert(fdp->fd_kqhash); + knhash_unlock(fdp); - kqwl->kqwl_dynamicid = id; + if (kn_hash) { + hashdestroy(kn_hash, M_KQUEUE, kn_hashmask); + } - list = &fdp->fd_kqhash[KQ_HASH(id, fdp->fd_kqhashmask)]; - SLIST_INSERT_HEAD(list, kqwl, kqwl_hashlink); + proc_fdlock(p); } -/* Called with kqhash_lock held */ -static void -kqueue_hash_remove( - struct proc *p, - struct kqueue *kq) +/* + * kqworkloops_dealloc - rebalance retains on kqworkloops created with + * scheduling parameters + * + * Called with proc_fdlock held. + * Returns with it locked. + * Process is in such a state that it will not try to allocate + * any more knotes during this process (stopped for exit or exec). + */ +void +kqworkloops_dealloc(proc_t p) { - struct kqworkloop *kqwl = (struct kqworkloop *)kq; struct filedesc *fdp = p->p_fd; - struct kqlist *list; + struct kqworkloop *kqwl, *kqwln; + struct kqwllist tofree; - /* should hold the kq hash lock */ - kqhash_lock_held(p); - - if ((kq->kq_state & KQ_DYNAMIC) == 0) { - assert(kq->kq_state & KQ_DYNAMIC); + if (!(fdp->fd_flags & FD_WORKLOOP)) { return; } - assert(kq->kq_state & KQ_WORKLOOP); /* for now */ - list = &fdp->fd_kqhash[KQ_HASH(kqwl->kqwl_dynamicid, fdp->fd_kqhashmask)]; - SLIST_REMOVE(list, kqwl, kqworkloop, kqwl_hashlink); -} -/* Called with kqhash_lock held */ -static struct kqueue * -kqueue_hash_lookup(struct proc *p, kqueue_id_t id) -{ - struct filedesc *fdp = p->p_fd; - struct kqlist *list; - struct kqworkloop *kqwl; - - /* should hold the kq hash lock */ - kqhash_lock_held(p); + kqhash_lock(fdp); - if (fdp->fd_kqhashmask == 0) return NULL; + if (fdp->fd_kqhashmask == 0) { + kqhash_unlock(fdp); + return; + } - list = &fdp->fd_kqhash[KQ_HASH(id, fdp->fd_kqhashmask)]; - SLIST_FOREACH(kqwl, list, kqwl_hashlink) { - if (kqwl->kqwl_dynamicid == id) { - struct kqueue *kq = (struct kqueue *)kqwl; + LIST_INIT(&tofree); - assert(kq->kq_state & KQ_DYNAMIC); - assert(kq->kq_state & KQ_WORKLOOP); /* for now */ - return kq; + for (size_t i = 0; i <= fdp->fd_kqhashmask; i++) { + LIST_FOREACH_SAFE(kqwl, &fdp->fd_kqhash[i], kqwl_hashlink, kqwln) { + /* + * kqworkloops that have scheduling parameters have an + * implicit retain from kqueue_workloop_ctl that needs + * to be balanced on process exit. + */ + assert(kqwl->kqwl_params); + LIST_REMOVE(kqwl, kqwl_hashlink); + LIST_INSERT_HEAD(&tofree, kqwl, kqwl_hashlink); } } - return NULL; + + kqhash_unlock(fdp); + + LIST_FOREACH_SAFE(kqwl, &tofree, kqwl_hashlink, kqwln) { + kqworkloop_dealloc(kqwl, KQWL_DEALLOC_SKIP_HASH_REMOVE, 1); + } } -static inline void -kqueue_release_last(struct proc *p, struct kqueue *kq) -{ - if (kq->kq_state & KQ_DYNAMIC) { - kqhash_lock(p); - if (kqueue_release(kq, KQUEUE_MIGHT_BE_LAST_REF)) { - kqueue_hash_remove(p, kq); - kqhash_unlock(p); - kqueue_dealloc(kq); - } else { - kqhash_unlock(p); +static int +kevent_register_validate_priority(struct kqueue *kq, struct knote *kn, + struct kevent_qos_s *kev) +{ + /* We don't care about the priority of a disabled or deleted knote */ + if (kev->flags & (EV_DISABLE | EV_DELETE)) { + return 0; + } + + if (kq->kq_state & KQ_WORKLOOP) { + /* + * Workloops need valid priorities with a QOS (excluding manager) for + * any enabled knote. + * + * When it is pre-existing, just make sure it has a valid QoS as + * kevent_register() will not use the incoming priority (filters who do + * have the responsibility to validate it again, see filt_wltouch). + * + * If the knote is being made, validate the incoming priority. + */ + if (!_pthread_priority_thread_qos(kn ? kn->kn_qos : kev->qos)) { + return ERANGE; } } + + return 0; } -static struct kqueue * -kevent_get_bound_kq(__assert_only struct proc *p, thread_t thread, - unsigned int kev_flags, unsigned int kq_flags) +/* + * Prepare a filter for waiting after register. + * + * The f_post_register_wait hook will be called later by kevent_register() + * and should call kevent_register_wait_block() + */ +static int +kevent_register_wait_prepare(struct knote *kn, struct kevent_qos_s *kev, int rc) { - struct kqueue *kq; - struct uthread *ut = get_bsdthread_info(thread); + thread_t thread = current_thread(); - assert(p == get_bsdthreadtask_info(thread)); + assert(knote_fops(kn)->f_extended_codes); - if (!(ut->uu_kqueue_flags & kev_flags)) - return NULL; + if (kn->kn_thread == NULL) { + thread_reference(thread); + kn->kn_thread = thread; + } else if (kn->kn_thread != thread) { + /* + * kn_thread may be set from a previous aborted wait + * However, it has to be from the same thread. + */ + kev->flags |= EV_ERROR; + kev->data = EXDEV; + return 0; + } - kq = ut->uu_kqueue_bound; - if (!kq) - return NULL; + return FILTER_REGISTER_WAIT | rc; +} - if (!(kq->kq_state & kq_flags)) - return NULL; +/* + * Cleanup a kevent_register_wait_prepare() effect for threads that have been + * aborted instead of properly woken up with thread_wakeup_thread(). + */ +static void +kevent_register_wait_cleanup(struct knote *kn) +{ + thread_t thread = kn->kn_thread; + kn->kn_thread = NULL; + thread_deallocate(thread); +} - return kq; +/* + * Must be called at the end of a f_post_register_wait call from a filter. + */ +static void +kevent_register_wait_block(struct turnstile *ts, thread_t thread, + thread_continue_t cont, struct _kevent_register *cont_args) +{ + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_HELD); + kqunlock(cont_args->kqwl); + cont_args->handoff_thread = thread; + thread_handoff_parameter(thread, cont, cont_args, THREAD_HANDOFF_NONE); } -static int -kevent_get_kq(struct proc *p, kqueue_id_t id, unsigned int flags, struct fileproc **fpp, int *fdp, struct kqueue **kqp) +/* + * Called by Filters using a f_post_register_wait to return from their wait. + */ +static void +kevent_register_wait_return(struct _kevent_register *cont_args) { - struct filedesc *descp = p->p_fd; - struct fileproc *fp = NULL; - struct kqueue *kq; - int fd = 0; + struct kqworkloop *kqwl = cont_args->kqwl; + struct kevent_qos_s *kev = &cont_args->kev; int error = 0; - /* Was the workloop flag passed? Then it is for sure only a workloop */ - if (flags & KEVENT_FLAG_DYNAMIC_KQUEUE) { - assert(flags & KEVENT_FLAG_WORKLOOP); - if (id == (kqueue_id_t)-1 && - (flags & KEVENT_FLAG_KERNEL) && - (flags & KEVENT_FLAG_WORKLOOP)) { - - assert(is_workqueue_thread(current_thread())); + if (cont_args->handoff_thread) { + thread_deallocate(cont_args->handoff_thread); + } - /* - * when kevent_id_internal is called from within the - * kernel, and the passed 'id' value is '-1' then we - * look for the currently bound workloop kq. - * - * Until pthread kext avoids calling in to kevent_id_internal - * for threads whose fulfill is canceled, calling in unbound - * can't be fatal. - */ - kq = kevent_get_bound_kq(p, current_thread(), - KEVENT_FLAG_WORKLOOP, KQ_WORKLOOP); - if (kq) { - kqueue_retain(kq); - } else { - struct uthread *ut = get_bsdthread_info(current_thread()); + if (kev->flags & (EV_ERROR | EV_RECEIPT)) { + if ((kev->flags & EV_ERROR) == 0) { + kev->flags |= EV_ERROR; + kev->data = 0; + } + error = kevent_modern_copyout(kev, &cont_args->ueventlist); + if (error == 0) { + cont_args->eventout++; + } + } - /* If thread is unbound due to cancel, just return an error */ - if (ut->uu_kqueue_flags == KEVENT_FLAG_WORKLOOP_CANCELED) { - ut->uu_kqueue_flags = 0; - error = ECANCELED; - } else { - panic("Unbound thread called kevent_internal with id=-1" - " uu_kqueue_flags:0x%x, uu_kqueue_bound:%p", - ut->uu_kqueue_flags, ut->uu_kqueue_bound); - } - } + kqworkloop_release(kqwl); + if (error == 0) { + *(int32_t *)¤t_uthread()->uu_rval = cont_args->eventout; + } + unix_syscall_return(error); +} - *fpp = NULL; - *fdp = 0; - *kqp = kq; - return error; - } +/* + * kevent_register - add a new event to a kqueue + * + * Creates a mapping between the event source and + * the kqueue via a knote data structure. + * + * Because many/most the event sources are file + * descriptor related, the knote is linked off + * the filedescriptor table for quick access. + * + * called with nothing locked + * caller holds a reference on the kqueue + */ - /* try shortcut on kq lookup for bound threads */ - kq = kevent_get_bound_kq(p, current_thread(), KEVENT_FLAG_WORKLOOP, KQ_WORKLOOP); - if (kq != NULL && ((struct kqworkloop *)kq)->kqwl_dynamicid == id) { +int +kevent_register(struct kqueue *kq, struct kevent_qos_s *kev, + struct knote **kn_out) +{ + struct proc *p = kq->kq_p; + const struct filterops *fops; + struct knote *kn = NULL; + int result = 0, error = 0; + unsigned short kev_flags = kev->flags; + KNOTE_LOCK_CTX(knlc); - if (flags & KEVENT_FLAG_DYNAMIC_KQ_MUST_NOT_EXIST) { - error = EEXIST; - kq = NULL; - goto out; - } + if (__probable(kev->filter < 0 && kev->filter + EVFILT_SYSCOUNT >= 0)) { + fops = sysfilt_ops[~kev->filter]; /* to 0-base index */ + } else { + error = EINVAL; + goto out; + } - /* retain a reference while working with this kq. */ - assert(kq->kq_state & KQ_DYNAMIC); - kqueue_retain(kq); - error = 0; - goto out; + /* restrict EV_VANISHED to adding udata-specific dispatch kevents */ + if (__improbable((kev->flags & EV_VANISHED) && + (kev->flags & (EV_ADD | EV_DISPATCH2)) != (EV_ADD | EV_DISPATCH2))) { + error = EINVAL; + goto out; + } + + /* Simplify the flags - delete and disable overrule */ + if (kev->flags & EV_DELETE) { + kev->flags &= ~EV_ADD; + } + if (kev->flags & EV_DISABLE) { + kev->flags &= ~EV_ENABLE; + } + + if (kq->kq_state & KQ_WORKLOOP) { + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWL_REGISTER), + ((struct kqworkloop *)kq)->kqwl_dynamicid, + kev->udata, kev->flags, kev->filter); + } else if (kq->kq_state & KQ_WORKQ) { + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWQ_REGISTER), + 0, kev->udata, kev->flags, kev->filter); + } else { + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQ_REGISTER), + VM_KERNEL_UNSLIDE_OR_PERM(kq), + kev->udata, kev->flags, kev->filter); + } + +restart: + /* find the matching knote from the fd tables/hashes */ + kn = kq_find_knote_and_kq_lock(kq, kev, fops->f_isfd, p); + error = kevent_register_validate_priority(kq, kn, kev); + result = 0; + if (error) { + goto out; + } + + if (kn == NULL && (kev->flags & EV_ADD) == 0) { + /* + * No knote found, EV_ADD wasn't specified + */ + + if ((kev_flags & EV_ADD) && (kev_flags & EV_DELETE) && + (kq->kq_state & KQ_WORKLOOP)) { + /* + * For workloops, understand EV_ADD|EV_DELETE as a "soft" delete + * that doesn't care about ENOENT, so just pretend the deletion + * happened. + */ + } else { + error = ENOENT; } + goto out; + } else if (kn == NULL) { + /* + * No knote found, need to attach a new one (attach) + */ - /* look for the kq on the hash table */ - kqhash_lock(p); - kq = kqueue_hash_lookup(p, id); - if (kq == NULL) { - kqhash_unlock(p); + struct fileproc *knote_fp = NULL; - if (flags & KEVENT_FLAG_DYNAMIC_KQ_MUST_EXIST) { - error = ENOENT; + /* grab a file reference for the new knote */ + if (fops->f_isfd) { + if ((error = fp_lookup(p, (int)kev->ident, &knote_fp, 0)) != 0) { goto out; } + } - struct kqueue *alloc_kq; - alloc_kq = kqueue_alloc(p, flags); - if (alloc_kq) { - kqhash_lock(p); - kqueue_hash_init_if_needed(p); - kq = kqueue_hash_lookup(p, id); - if (kq == NULL) { - /* insert our new one */ - kq = alloc_kq; - kqueue_hash_insert(p, id, kq); - kqhash_unlock(p); - } else { - /* lost race, retain existing workloop */ - kqueue_retain(kq); - kqhash_unlock(p); - kqueue_release(alloc_kq, KQUEUE_MIGHT_BE_LAST_REF); - kqueue_dealloc(alloc_kq); - } - } else { - error = ENOMEM; - goto out; + kn = knote_alloc(); + if (kn == NULL) { + error = ENOMEM; + if (knote_fp != NULL) { + fp_drop(p, (int)kev->ident, knote_fp, 0); } - } else { + goto out; + } - if (flags & KEVENT_FLAG_DYNAMIC_KQ_MUST_NOT_EXIST) { - kqhash_unlock(p); - kq = NULL; - error = EEXIST; - goto out; - } + kn->kn_fp = knote_fp; + kn->kn_is_fd = fops->f_isfd; + kn->kn_kq_packed = VM_PACK_POINTER((vm_offset_t)kq, KNOTE_KQ_PACKED); + kn->kn_status = 0; - /* retain a reference while working with this kq. */ - assert(kq->kq_state & KQ_DYNAMIC); - kqueue_retain(kq); - kqhash_unlock(p); + /* was vanish support requested */ + if (kev->flags & EV_VANISHED) { + kev->flags &= ~EV_VANISHED; + kn->kn_status |= KN_REQVANISH; } - - } else if (flags & KEVENT_FLAG_WORKQ) { - /* must already exist for bound threads. */ - if (flags & KEVENT_FLAG_KERNEL) { - assert(descp->fd_wqkqueue != NULL); + + /* snapshot matching/dispatching protocol flags into knote */ + if (kev->flags & EV_DISABLE) { + kn->kn_status |= KN_DISABLED; } /* - * use the private kq associated with the proc workq. - * Just being a thread within the process (and not - * being the exit/exec thread) is enough to hold a - * reference on this special kq. + * copy the kevent state into knote + * protocol is that fflags and data + * are saved off, and cleared before + * calling the attach routine. + * + * - kn->kn_sfflags aliases with kev->xflags + * - kn->kn_sdata aliases with kev->data + * - kn->kn_filter is the top 8 bits of kev->filter */ - kq = descp->fd_wqkqueue; - if (kq == NULL) { - struct kqueue *alloc_kq = kqueue_alloc(p, KEVENT_FLAG_WORKQ); - if (alloc_kq == NULL) - return ENOMEM; - - knhash_lock(p); - if (descp->fd_wqkqueue == NULL) { - kq = descp->fd_wqkqueue = alloc_kq; - knhash_unlock(p); - } else { - knhash_unlock(p); - kq = descp->fd_wqkqueue; - kqueue_dealloc(alloc_kq); + kn->kn_kevent = *(struct kevent_internal_s *)kev; + kn->kn_sfflags = kev->fflags; + kn->kn_filtid = (uint8_t)~kev->filter; + kn->kn_fflags = 0; + knote_reset_priority(kq, kn, kev->qos); + + /* Add the knote for lookup thru the fd table */ + error = kq_add_knote(kq, kn, &knlc, p); + if (error) { + knote_free(kn); + if (knote_fp != NULL) { + fp_drop(p, (int)kev->ident, knote_fp, 0); } + + if (error == ERESTART) { + goto restart; + } + goto out; } - } else { - /* get a usecount for the kq itself */ - fd = (int)id; - if ((error = fp_getfkq(p, fd, &fp, &kq)) != 0) - return (error); - } - if ((error = kevent_set_kq_mode(kq, flags)) != 0) { - /* drop the usecount */ - if (fp != NULL) - fp_drop(p, fd, fp, 0); - return error; - } -out: - *fpp = fp; - *fdp = fd; - *kqp = kq; - - return error; -} + /* fp reference count now applies to knote */ -static void -kevent_put_kq( - struct proc *p, - kqueue_id_t id, - struct fileproc *fp, - struct kqueue *kq) -{ - kqueue_release_last(p, kq); - if (fp != NULL) { - assert((kq->kq_state & KQ_WORKQ) == 0); - fp_drop(p, (int)id, fp, 0); - } -} + /* + * we can't use filter_call() because f_attach can change the filter ops + * for a filter that supports f_extended_codes, so we need to reload + * knote_fops() and not use `fops`. + */ + result = fops->f_attach(kn, kev); + if (result && !knote_fops(kn)->f_extended_codes) { + result = FILTER_ACTIVE; + } -static uint64_t -kevent_workloop_serial_no_copyin(proc_t p, uint64_t workloop_id) -{ - uint64_t serial_no = 0; - user_addr_t addr; - int rc; + kqlock(kq); - if (workloop_id == 0 || p->p_dispatchqueue_serialno_offset == 0) { - return 0; - } - addr = (user_addr_t)(workloop_id + p->p_dispatchqueue_serialno_offset); + if (result & FILTER_THREADREQ_NODEFEER) { + enable_preemption(); + } - if (proc_is64bit(p)) { - rc = copyin(addr, (caddr_t)&serial_no, sizeof(serial_no)); - } else { - uint32_t serial_no32 = 0; - rc = copyin(addr, (caddr_t)&serial_no32, sizeof(serial_no32)); - serial_no = serial_no32; - } - return rc == 0 ? serial_no : 0; -} + if (kn->kn_flags & EV_ERROR) { + /* + * Failed to attach correctly, so drop. + */ + kn->kn_filtid = EVFILTID_DETACHED; + error = (int)kn->kn_sdata; + knote_drop(kq, kn, &knlc); + result = 0; + goto out; + } -int -kevent_exit_on_workloop_ownership_leak(thread_t thread) -{ - proc_t p = current_proc(); - struct filedesc *fdp = p->p_fd; - kqueue_id_t workloop_id = 0; - os_reason_t reason; - mach_vm_address_t addr; - uint32_t reason_size; + /* + * end "attaching" phase - now just attached + * + * Mark the thread request overcommit, if appropos + * + * If the attach routine indicated that an + * event is already fired, activate the knote. + */ + if ((kn->kn_qos & _PTHREAD_PRIORITY_OVERCOMMIT_FLAG) && + (kq->kq_state & KQ_WORKLOOP)) { + kqworkloop_set_overcommit((struct kqworkloop *)kq); + } + } else if (!knote_lock(kq, kn, &knlc, KNOTE_KQ_LOCK_ON_SUCCESS)) { + /* + * The knote was dropped while we were waiting for the lock, + * we need to re-evaluate entirely + */ - kqhash_lock(p); - if (fdp->fd_kqhashmask > 0) { - for (uint32_t i = 0; i < fdp->fd_kqhashmask + 1; i++) { - struct kqworkloop *kqwl; + goto restart; + } else if (kev->flags & EV_DELETE) { + /* + * Deletion of a knote (drop) + * + * If the filter wants to filter drop events, let it do so. + * + * defer-delete: when trying to delete a disabled EV_DISPATCH2 knote, + * we must wait for the knote to be re-enabled (unless it is being + * re-enabled atomically here). + */ - SLIST_FOREACH(kqwl, &fdp->fd_kqhash[i], kqwl_hashlink) { - struct kqueue *kq = &kqwl->kqwl_kqueue; - if ((kq->kq_state & KQ_DYNAMIC) && kqwl->kqwl_owner == thread) { - workloop_id = kqwl->kqwl_dynamicid; - break; - } + if (knote_fops(kn)->f_allow_drop) { + bool drop; + + kqunlock(kq); + drop = knote_fops(kn)->f_allow_drop(kn, kev); + kqlock(kq); + + if (!drop) { + goto out_unlock; } } - } - kqhash_unlock(p); - assert(workloop_id); - reason = os_reason_create(OS_REASON_LIBSYSTEM, - OS_REASON_LIBSYSTEM_CODE_WORKLOOP_OWNERSHIP_LEAK); - if (reason == OS_REASON_NULL) { - goto out; - } + if ((kev->flags & EV_ENABLE) == 0 && + (kn->kn_flags & EV_DISPATCH2) == EV_DISPATCH2 && + (kn->kn_status & KN_DISABLED) != 0) { + kn->kn_status |= KN_DEFERDELETE; + error = EINPROGRESS; + goto out_unlock; + } - reason->osr_flags |= OS_REASON_FLAG_GENERATE_CRASH_REPORT; - reason_size = 2 * sizeof(uint64_t); - reason_size = kcdata_estimate_required_buffer_size(2, reason_size); - if (os_reason_alloc_buffer(reason, reason_size) != 0) { + knote_drop(kq, kn, &knlc); goto out; - } + } else { + /* + * Regular update of a knote (touch) + * + * Call touch routine to notify filter of changes in filter values + * (and to re-determine if any events are fired). + * + * If the knote is in defer-delete, avoid calling the filter touch + * routine (it has delivered its last event already). + * + * If the touch routine had no failure, + * apply the requested side effects to the knote. + */ - struct kcdata_descriptor *kcd = &reason->osr_kcd_descriptor; + if (kn->kn_status & (KN_DEFERDELETE | KN_VANISHED)) { + if (kev->flags & EV_ENABLE) { + result = FILTER_ACTIVE; + } + } else { + kqunlock(kq); + result = filter_call(knote_fops(kn), f_touch(kn, kev)); + kqlock(kq); + if (result & FILTER_THREADREQ_NODEFEER) { + enable_preemption(); + } + } + + if (kev->flags & EV_ERROR) { + result = 0; + goto out_unlock; + } - if (kcdata_get_memory_addr(kcd, EXIT_REASON_WORKLOOP_ID, - sizeof(workloop_id), &addr) == KERN_SUCCESS) { - kcdata_memcpy(kcd, addr, &workloop_id, sizeof(workloop_id)); + if ((kn->kn_flags & EV_UDATA_SPECIFIC) == 0 && + kn->kn_udata != kev->udata) { + // this allows klist_copy_udata() not to take locks + os_atomic_store_wide(&kn->kn_udata, kev->udata, relaxed); + } + if ((kev->flags & EV_DISABLE) && !(kn->kn_status & KN_DISABLED)) { + kn->kn_status |= KN_DISABLED; + knote_dequeue(kq, kn); + } } - uint64_t serial_no = kevent_workloop_serial_no_copyin(p, workloop_id); - if (serial_no && kcdata_get_memory_addr(kcd, EXIT_REASON_DISPATCH_QUEUE_NO, - sizeof(serial_no), &addr) == KERN_SUCCESS) { - kcdata_memcpy(kcd, addr, &serial_no, sizeof(serial_no)); + /* accept new kevent state */ + knote_apply_touch(kq, kn, kev, result); + +out_unlock: + /* + * When the filter asked for a post-register wait, + * we leave the kqueue locked for kevent_register() + * to call the filter's f_post_register_wait hook. + */ + if (result & FILTER_REGISTER_WAIT) { + knote_unlock(kq, kn, &knlc, KNOTE_KQ_LOCK_ALWAYS); + *kn_out = kn; + } else { + knote_unlock(kq, kn, &knlc, KNOTE_KQ_UNLOCK); } out: -#if DEVELOPMENT || DEBUG - psignal_try_thread_with_reason(p, thread, SIGABRT, reason); - return 0; -#else - return exit_with_reason(p, W_EXITCODE(0, SIGKILL), (int *)NULL, - FALSE, FALSE, 0, reason); -#endif + /* output local errors through the kevent */ + if (error) { + kev->flags |= EV_ERROR; + kev->data = error; + } + return result; } - +/* + * knote_process - process a triggered event + * + * Validate that it is really still a triggered event + * by calling the filter routines (if necessary). Hold + * a use reference on the knote to avoid it being detached. + * + * If it is still considered triggered, we will have taken + * a copy of the state under the filter lock. We use that + * snapshot to dispatch the knote for future processing (or + * not, if this was a lost event). + * + * Our caller assures us that nobody else can be processing + * events from this knote during the whole operation. But + * others can be touching or posting events to the knote + * interspersed with our processing it. + * + * caller holds a reference on the kqueue. + * kqueue locked on entry and exit - but may be dropped + */ static int -kevent_servicer_detach_preflight(thread_t thread, unsigned int flags, struct kqueue *kq) +knote_process(struct knote *kn, kevent_ctx_t kectx, + kevent_callback_t callback) { + struct kevent_qos_s kev; + struct kqueue *kq = knote_get_kq(kn); + KNOTE_LOCK_CTX(knlc); + int result = FILTER_ACTIVE; int error = 0; - struct kqworkloop *kqwl; - struct uthread *ut; - struct kqrequest *kqr; - - if (!(flags & KEVENT_FLAG_WORKLOOP) || !(kq->kq_state & KQ_WORKLOOP)) - return EINVAL; - - /* only kq created with KEVENT_FLAG_WORKLOOP_NO_WQ_THREAD from userspace can have attached threads */ - if (!(kq->kq_state & KQ_NO_WQ_THREAD)) - return EINVAL; - - /* allow detach only on not wq threads */ - if (is_workqueue_thread(thread)) - return EINVAL; - - /* check that the current thread is bound to the requested wq */ - ut = get_bsdthread_info(thread); - if (ut->uu_kqueue_bound != kq) - return EINVAL; + bool drop = false; - kqwl = (struct kqworkloop *)kq; - kqwl_req_lock(kqwl); - kqr = &kqwl->kqwl_request; + /* + * Must be active or stayactive + * Must be queued and not disabled/suppressed or dropping + */ + assert(kn->kn_status & KN_QUEUED); + assert(kn->kn_status & (KN_ACTIVE | KN_STAYACTIVE)); + assert(!(kn->kn_status & (KN_DISABLED | KN_SUPPRESSED | KN_DROPPING))); - /* check that the wq is bound to the thread */ - if ((kqr->kqr_state & KQR_BOUND) == 0 || (kqr->kqr_thread != thread)) - error = EINVAL; + if (kq->kq_state & KQ_WORKLOOP) { + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWL_PROCESS), + ((struct kqworkloop *)kq)->kqwl_dynamicid, + kn->kn_udata, kn->kn_status | (kn->kn_id << 32), + kn->kn_filtid); + } else if (kq->kq_state & KQ_WORKQ) { + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWQ_PROCESS), + 0, kn->kn_udata, kn->kn_status | (kn->kn_id << 32), + kn->kn_filtid); + } else { + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQ_PROCESS), + VM_KERNEL_UNSLIDE_OR_PERM(kq), kn->kn_udata, + kn->kn_status | (kn->kn_id << 32), kn->kn_filtid); + } - kqwl_req_unlock(kqwl); + if (!knote_lock(kq, kn, &knlc, KNOTE_KQ_LOCK_ALWAYS)) { + /* + * When the knote is dropping or has dropped, + * then there's nothing we want to process. + */ + return EJUSTRETURN; + } - return error; -} + /* + * While waiting for the knote lock, we may have dropped the kq lock. + * and a touch may have disabled and dequeued the knote. + */ + if (!(kn->kn_status & KN_QUEUED)) { + knote_unlock(kq, kn, &knlc, KNOTE_KQ_LOCK_ALWAYS); + return EJUSTRETURN; + } -static void -kevent_servicer_detach_thread(struct proc *p, kqueue_id_t id, thread_t thread, - unsigned int flags, struct kqueue *kq) -{ - struct kqworkloop *kqwl; - struct uthread *ut; + /* + * For deferred-drop or vanished events, we just create a fake + * event to acknowledge end-of-life. Otherwise, we call the + * filter's process routine to snapshot the kevent state under + * the filter's locking protocol. + * + * suppress knotes to avoid returning the same event multiple times in + * a single call. + */ + knote_suppress(kq, kn); - assert((flags & KEVENT_FLAG_WORKLOOP) && (kq->kq_state & KQ_WORKLOOP)); + if (kn->kn_status & (KN_DEFERDELETE | KN_VANISHED)) { + uint16_t kev_flags = EV_DISPATCH2 | EV_ONESHOT; + if (kn->kn_status & KN_DEFERDELETE) { + kev_flags |= EV_DELETE; + } else { + kev_flags |= EV_VANISHED; + } - /* allow detach only on not wqthreads threads */ - assert(!is_workqueue_thread(thread)); + /* create fake event */ + kev = (struct kevent_qos_s){ + .filter = kn->kn_filter, + .ident = kn->kn_id, + .flags = kev_flags, + .udata = kn->kn_udata, + }; + } else { + kqunlock(kq); + kev = (struct kevent_qos_s) { }; + result = filter_call(knote_fops(kn), f_process(kn, &kev)); + kqlock(kq); + } - /* only kq created with KEVENT_FLAG_WORKLOOP_NO_WQ_THREAD from userspace can have attached threads */ - assert(kq->kq_state & KQ_NO_WQ_THREAD); + /* + * Determine how to dispatch the knote for future event handling. + * not-fired: just return (do not callout, leave deactivated). + * One-shot: If dispatch2, enter deferred-delete mode (unless this is + * is the deferred delete event delivery itself). Otherwise, + * drop it. + * Dispatch: don't clear state, just mark it disabled. + * Cleared: just leave it deactivated. + * Others: re-activate as there may be more events to handle. + * This will not wake up more handlers right now, but + * at the completion of handling events it may trigger + * more handler threads (TODO: optimize based on more than + * just this one event being detected by the filter). + */ + if ((result & FILTER_ACTIVE) == 0) { + if ((kn->kn_status & (KN_ACTIVE | KN_STAYACTIVE)) == 0) { + /* + * Stay active knotes should not be unsuppressed or we'd create an + * infinite loop. + * + * Some knotes (like EVFILT_WORKLOOP) can be reactivated from + * within f_process() but that doesn't necessarily make them + * ready to process, so we should leave them be. + * + * For other knotes, since we will not return an event, + * there's no point keeping the knote suppressed. + */ + knote_unsuppress(kq, kn); + } + knote_unlock(kq, kn, &knlc, KNOTE_KQ_LOCK_ALWAYS); + return EJUSTRETURN; + } - /* check that the current thread is bound to the requested kq */ - ut = get_bsdthread_info(thread); - assert(ut->uu_kqueue_bound == kq); + if (result & FILTER_ADJUST_EVENT_QOS_BIT) { + knote_adjust_qos(kq, kn, result); + } + kev.qos = _pthread_priority_combine(kn->kn_qos, kn->kn_qos_override); - kqwl = (struct kqworkloop *)kq; + if (kev.flags & EV_ONESHOT) { + if ((kn->kn_flags & EV_DISPATCH2) == EV_DISPATCH2 && + (kn->kn_status & KN_DEFERDELETE) == 0) { + /* defer dropping non-delete oneshot dispatch2 events */ + kn->kn_status |= KN_DEFERDELETE | KN_DISABLED; + } else { + drop = true; + } + } else if (kn->kn_flags & EV_DISPATCH) { + /* disable all dispatch knotes */ + kn->kn_status |= KN_DISABLED; + } else if ((kn->kn_flags & EV_CLEAR) == 0) { + /* re-activate in case there are more events */ + knote_activate(kq, kn, FILTER_ACTIVE); + } - kqlock(kq); - - /* unbind the thread. - * unbind itself checks if still processing and ends it. + /* + * callback to handle each event as we find it. + * If we have to detach and drop the knote, do + * it while we have the kq unlocked. */ - kqworkloop_unbind_thread(kqwl, thread, flags); - - kqunlock(kq); + if (drop) { + knote_drop(kq, kn, &knlc); + } else { + knote_unlock(kq, kn, &knlc, KNOTE_KQ_UNLOCK); + } - kevent_put_kq(p, id, NULL, kq); + if (kev.flags & EV_VANISHED) { + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KNOTE_VANISHED), + kev.ident, kn->kn_udata, kn->kn_status | (kn->kn_id << 32), + kn->kn_filtid); + } - return; + error = (callback)(&kev, kectx); + kqlock(kq); + return error; } +/* + * Returns -1 if the kqueue was unbound and processing should not happen + */ +#define KQWQAE_BEGIN_PROCESSING 1 +#define KQWQAE_END_PROCESSING 2 +#define KQWQAE_UNBIND 3 static int -kevent_servicer_attach_thread(thread_t thread, unsigned int flags, struct kqueue *kq) +kqworkq_acknowledge_events(struct kqworkq *kqwq, workq_threadreq_t kqr, + int kevent_flags, int kqwqae_op) { - int error = 0; - struct kqworkloop *kqwl; - struct uthread *ut; - struct kqrequest *kqr; + thread_qos_t old_override = THREAD_QOS_UNSPECIFIED; + thread_t thread = kqr_thread_fast(kqr); + struct knote *kn; + int rc = 0; + bool unbind; + struct kqtailq *suppressq = &kqwq->kqwq_suppressed[kqr->tr_kq_qos_index]; - if (!(flags & KEVENT_FLAG_WORKLOOP) || !(kq->kq_state & KQ_WORKLOOP)) - return EINVAL; + kqlock_held(&kqwq->kqwq_kqueue); - /* only kq created with KEVENT_FLAG_WORKLOOP_NO_WQ_THREAD from userspace can have attached threads*/ - if (!(kq->kq_state & KQ_NO_WQ_THREAD)) - return EINVAL; + if (!TAILQ_EMPTY(suppressq)) { + /* + * Return suppressed knotes to their original state. + * For workq kqueues, suppressed ones that are still + * truly active (not just forced into the queue) will + * set flags we check below to see if anything got + * woken up. + */ + while ((kn = TAILQ_FIRST(suppressq)) != NULL) { + assert(kn->kn_status & KN_SUPPRESSED); + knote_unsuppress(kqwq, kn); + } + } - /* allow attach only on not wqthreads */ - if (is_workqueue_thread(thread)) - return EINVAL; +#if DEBUG || DEVELOPMENT + thread_t self = current_thread(); + struct uthread *ut = get_bsdthread_info(self); - /* check that the thread is not already bound */ - ut = get_bsdthread_info(thread); - if (ut->uu_kqueue_bound != NULL) - return EINVAL; + assert(thread == self); + assert(ut->uu_kqr_bound == kqr); +#endif // DEBUG || DEVELOPMENT - assert(ut->uu_kqueue_flags == 0); + if (kqwqae_op == KQWQAE_UNBIND) { + unbind = true; + } else if ((kevent_flags & KEVENT_FLAG_PARKING) == 0) { + unbind = false; + } else { + unbind = !kqr->tr_kq_wakeup; + } + if (unbind) { + old_override = kqworkq_unbind_locked(kqwq, kqr, thread); + rc = -1; + /* + * request a new thread if we didn't process the whole queue or real events + * have happened (not just putting stay-active events back). + */ + if (kqr->tr_kq_wakeup) { + kqueue_threadreq_initiate(&kqwq->kqwq_kqueue, kqr, + kqr->tr_kq_qos_index, 0); + } + } - kqlock(kq); - kqwl = (struct kqworkloop *)kq; - kqwl_req_lock(kqwl); - kqr = &kqwl->kqwl_request; + if (rc == 0) { + /* + * Reset wakeup bit to notice events firing while we are processing, + * as we cannot rely on the bucket queue emptiness because of stay + * active knotes. + */ + kqr->tr_kq_wakeup = false; + } - /* check that the kqueue is not already bound */ - if (kqr->kqr_state & (KQR_BOUND | KQR_THREQUESTED | KQR_DRAIN)) { - error = EINVAL; - goto out; + if (old_override) { + thread_drop_kevent_override(thread); } - assert(kqr->kqr_thread == NULL); - assert((kqr->kqr_state & KQR_PROCESSING) == 0); + return rc; +} - kqr->kqr_state |= KQR_THREQUESTED; - kqr->kqr_qos_index = THREAD_QOS_UNSPECIFIED; - kqr->kqr_override_index = THREAD_QOS_UNSPECIFIED; - kqr->kqr_dsync_owner_qos = THREAD_QOS_UNSPECIFIED; - kqr->kqr_owner_override_is_sync = 0; +/* + * Return 0 to indicate that processing should proceed, + * -1 if there is nothing to process. + * + * Called with kqueue locked and returns the same way, + * but may drop lock temporarily. + */ +static int +kqworkq_begin_processing(struct kqworkq *kqwq, workq_threadreq_t kqr, + int kevent_flags) +{ + int rc = 0; - kqworkloop_bind_thread_impl(kqwl, thread, KEVENT_FLAG_WORKLOOP); + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWQ_PROCESS_BEGIN) | DBG_FUNC_START, + 0, kqr->tr_kq_qos_index); - /* get a ref on the wlkq on behalf of the attached thread */ - kqueue_retain(kq); + rc = kqworkq_acknowledge_events(kqwq, kqr, kevent_flags, + KQWQAE_BEGIN_PROCESSING); -out: - kqwl_req_unlock(kqwl); - kqunlock(kq); + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWQ_PROCESS_BEGIN) | DBG_FUNC_END, + thread_tid(kqr_thread(kqr)), kqr->tr_kq_wakeup); - return error; + return rc; } -static inline -boolean_t kevent_args_requesting_events(unsigned int flags, int nevents) +static thread_qos_t +kqworkloop_acknowledge_events(struct kqworkloop *kqwl) { - return (!(flags & KEVENT_FLAG_ERROR_EVENTS) && nevents > 0); -} + kq_index_t qos = THREAD_QOS_UNSPECIFIED; + struct knote *kn, *tmp; -static int -kevent_internal(struct proc *p, - kqueue_id_t id, kqueue_id_t *id_out, - user_addr_t changelist, int nchanges, - user_addr_t ueventlist, int nevents, - user_addr_t data_out, uint64_t data_available, - unsigned int flags, - user_addr_t utimeout, - kqueue_continue_t continuation, - int32_t *retval) -{ - struct _kevent *cont_args; - uthread_t ut; - struct kqueue *kq; - struct fileproc *fp = NULL; - int fd = 0; - struct kevent_internal_s kev; - int error, noutputs; - struct timeval atv; - user_size_t data_size; - user_size_t data_resid; - thread_t thread = current_thread(); + kqlock_held(kqwl); - /* Don't allow user-space threads to process output events from the workq kqs */ - if (((flags & (KEVENT_FLAG_WORKQ | KEVENT_FLAG_KERNEL)) == KEVENT_FLAG_WORKQ) && - kevent_args_requesting_events(flags, nevents)) - return EINVAL; + TAILQ_FOREACH_SAFE(kn, &kqwl->kqwl_suppressed, kn_tqe, tmp) { + /* + * If a knote that can adjust QoS is disabled because of the automatic + * behavior of EV_DISPATCH, the knotes should stay suppressed so that + * further overrides keep pushing. + */ + if (knote_fops(kn)->f_adjusts_qos && (kn->kn_status & KN_DISABLED) && + (kn->kn_status & (KN_STAYACTIVE | KN_DROPPING)) == 0 && + (kn->kn_flags & (EV_DISPATCH | EV_DISABLE)) == EV_DISPATCH) { + qos = MAX(qos, kn->kn_qos_override); + continue; + } + knote_unsuppress(kqwl, kn); + } - /* restrict dynamic kqueue allocation to workloops (for now) */ - if ((flags & (KEVENT_FLAG_DYNAMIC_KQUEUE | KEVENT_FLAG_WORKLOOP)) == KEVENT_FLAG_DYNAMIC_KQUEUE) - return EINVAL; + return qos; +} - if (flags & (KEVENT_FLAG_WORKLOOP_SERVICER_ATTACH | KEVENT_FLAG_WORKLOOP_SERVICER_DETACH | - KEVENT_FLAG_DYNAMIC_KQ_MUST_EXIST | KEVENT_FLAG_DYNAMIC_KQ_MUST_NOT_EXIST | KEVENT_FLAG_WORKLOOP_NO_WQ_THREAD)) { +static int +kqworkloop_begin_processing(struct kqworkloop *kqwl, unsigned int kevent_flags) +{ + workq_threadreq_t kqr = &kqwl->kqwl_request; + struct kqueue *kq = &kqwl->kqwl_kqueue; + thread_qos_t qos_override; + thread_t thread = kqr_thread_fast(kqr); + int rc = 0, op = KQWL_UTQ_NONE; - /* allowed only on workloops when calling kevent_id from user-space */ - if (!(flags & KEVENT_FLAG_WORKLOOP) || (flags & KEVENT_FLAG_KERNEL) || !(flags & KEVENT_FLAG_DYNAMIC_KQUEUE)) - return EINVAL; + kqlock_held(kq); - /* cannot attach and detach simultaneously*/ - if ((flags & KEVENT_FLAG_WORKLOOP_SERVICER_ATTACH) && (flags & KEVENT_FLAG_WORKLOOP_SERVICER_DETACH)) - return EINVAL; + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWL_PROCESS_BEGIN) | DBG_FUNC_START, + kqwl->kqwl_dynamicid, 0, 0); - /* cannot ask for events and detach */ - if ((flags & KEVENT_FLAG_WORKLOOP_SERVICER_DETACH) && kevent_args_requesting_events(flags, nevents)) - return EINVAL; + /* nobody else should still be processing */ + assert((kq->kq_state & KQ_PROCESSING) == 0); - } + kq->kq_state |= KQ_PROCESSING; - /* prepare to deal with stack-wise allocation of out events */ - if (flags & KEVENT_FLAG_STACK_EVENTS) { - int scale = ((flags & KEVENT_FLAG_LEGACY32) ? - (IS_64BIT_PROCESS(p) ? sizeof(struct user64_kevent) : - sizeof(struct user32_kevent)) : - ((flags & KEVENT_FLAG_LEGACY64) ? sizeof(struct kevent64_s) : - sizeof(struct kevent_qos_s))); - ueventlist += nevents * scale; + if (!TAILQ_EMPTY(&kqwl->kqwl_suppressed)) { + op = KQWL_UTQ_RESET_WAKEUP_OVERRIDE; } - /* convert timeout to absolute - if we have one (and not immediate) */ - error = kevent_get_timeout(p, utimeout, flags, &atv); - if (error) - return error; - - /* copyin initial value of data residual from data_available */ - error = kevent_get_data_size(p, data_available, flags, &data_size); - if (error) - return error; - - /* get the kq we are going to be working on */ - error = kevent_get_kq(p, id, flags, &fp, &fd, &kq); - if (error) - return error; - - /* only bound threads can receive events on workloops */ - if ((flags & KEVENT_FLAG_WORKLOOP) && kevent_args_requesting_events(flags, nevents)) { - ut = (uthread_t)get_bsdthread_info(thread); - if (ut->uu_kqueue_bound != kq) { - error = EXDEV; - goto out; + if (kevent_flags & KEVENT_FLAG_PARKING) { + /* + * When "parking" we want to process events and if no events are found + * unbind. + * + * However, non overcommit threads sometimes park even when they have + * more work so that the pool can narrow. For these, we need to unbind + * early, so that calling kqworkloop_update_threads_qos() can ask the + * workqueue subsystem whether the thread should park despite having + * pending events. + */ + if (kqr->tr_flags & WORKQ_TR_FLAG_OVERCOMMIT) { + op = KQWL_UTQ_PARKING; + } else { + op = KQWL_UTQ_UNBINDING; } - + } + if (op == KQWL_UTQ_NONE) { + goto done; } - /* attach the current thread if necessary */ - if (flags & KEVENT_FLAG_WORKLOOP_SERVICER_ATTACH) { - error = kevent_servicer_attach_thread(thread, flags, kq); - if (error) - goto out; + qos_override = kqworkloop_acknowledge_events(kqwl); + + if (op == KQWL_UTQ_UNBINDING) { + kqworkloop_unbind_locked(kqwl, thread, KQWL_OVERRIDE_DROP_IMMEDIATELY); + kqworkloop_release_live(kqwl); } - else { - /* before processing events and committing to the system call, return an error if the thread cannot be detached when requested */ - if (flags & KEVENT_FLAG_WORKLOOP_SERVICER_DETACH) { - error = kevent_servicer_detach_preflight(thread, flags, kq); - if (error) - goto out; + kqworkloop_update_threads_qos(kqwl, op, qos_override); + if (op == KQWL_UTQ_PARKING) { + if (!TAILQ_EMPTY(&kqwl->kqwl_queue[KQWL_BUCKET_STAYACTIVE])) { + /* + * We cannot trust tr_kq_wakeup when looking at stay active knotes. + * We need to process once, and kqworkloop_end_processing will + * handle the unbind. + */ + } else if (!kqr->tr_kq_wakeup || kqwl->kqwl_owner) { + kqworkloop_unbind_locked(kqwl, thread, KQWL_OVERRIDE_DROP_DELAYED); + kqworkloop_release_live(kqwl); + rc = -1; + } + } else if (op == KQWL_UTQ_UNBINDING) { + if (kqr_thread(kqr) == thread) { + /* + * The thread request fired again, passed the admission check and + * got bound to the current thread again. + */ + } else { + rc = -1; } } - if (id_out && kq && (flags & KEVENT_FLAG_WORKLOOP)) { - assert(kq->kq_state & KQ_WORKLOOP); - struct kqworkloop *kqwl; - kqwl = (struct kqworkloop *)kq; - *id_out = kqwl->kqwl_dynamicid; + if (rc == 0) { + /* + * Reset wakeup bit to notice stay active events firing while we are + * processing, as we cannot rely on the stayactive bucket emptiness. + */ + kqwl->kqwl_wakeup_indexes &= ~KQWL_STAYACTIVE_FIRED_BIT; + } else { + kq->kq_state &= ~KQ_PROCESSING; } - /* register all the change requests the user provided... */ - noutputs = 0; - while (nchanges > 0 && error == 0) { - error = kevent_copyin(&changelist, &kev, p, flags); - if (error) - break; + if (rc == -1) { + kqworkloop_unbind_delayed_override_drop(thread); + } - /* Make sure user doesn't pass in any system flags */ - kev.flags &= ~EV_SYSFLAGS; +done: + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWL_PROCESS_BEGIN) | DBG_FUNC_END, + kqwl->kqwl_dynamicid, 0, 0); - kevent_register(kq, &kev, p); + return rc; +} - if (nevents > 0 && - ((kev.flags & EV_ERROR) || (kev.flags & EV_RECEIPT))) { - if (kev.flags & EV_RECEIPT) { - kev.flags |= EV_ERROR; - kev.data = 0; - } - error = kevent_copyout(&kev, &ueventlist, p, flags); - if (error == 0) { - nevents--; - noutputs++; - } - } else if (kev.flags & EV_ERROR) { - error = kev.data; +/* + * Return 0 to indicate that processing should proceed, + * -1 if there is nothing to process. + * EBADF if the kqueue is draining + * + * Called with kqueue locked and returns the same way, + * but may drop lock temporarily. + * May block. + */ +static int +kqfile_begin_processing(struct kqfile *kq) +{ + struct kqtailq *suppressq; + + kqlock_held(kq); + + assert((kq->kqf_state & (KQ_WORKQ | KQ_WORKLOOP)) == 0); + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQ_PROCESS_BEGIN) | DBG_FUNC_START, + VM_KERNEL_UNSLIDE_OR_PERM(kq), 0); + + /* wait to become the exclusive processing thread */ + for (;;) { + if (kq->kqf_state & KQ_DRAIN) { + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQ_PROCESS_BEGIN) | DBG_FUNC_END, + VM_KERNEL_UNSLIDE_OR_PERM(kq), 2); + return EBADF; } - nchanges--; - } - /* short-circuit the scan if we only want error events */ - if (flags & KEVENT_FLAG_ERROR_EVENTS) - nevents = 0; - - /* process pending events */ - if (nevents > 0 && noutputs == 0 && error == 0) { - /* store the continuation/completion data in the uthread */ - ut = (uthread_t)get_bsdthread_info(thread); - cont_args = &ut->uu_kevent.ss_kevent; - cont_args->fp = fp; - cont_args->fd = fd; - cont_args->retval = retval; - cont_args->eventlist = ueventlist; - cont_args->eventcount = nevents; - cont_args->eventout = noutputs; - cont_args->data_available = data_available; - cont_args->process_data.fp_fd = (int)id; - cont_args->process_data.fp_flags = flags; - cont_args->process_data.fp_data_out = data_out; - cont_args->process_data.fp_data_size = data_size; - cont_args->process_data.fp_data_resid = data_size; - - error = kqueue_scan(kq, kevent_callback, - continuation, cont_args, - &cont_args->process_data, - &atv, p); - - /* process remaining outputs */ - noutputs = cont_args->eventout; - data_resid = cont_args->process_data.fp_data_resid; - - /* copyout residual data size value (if it needs to be copied out) */ - /* don't abandon other output just because of residual copyout failures */ - if (error == 0 && data_available && data_resid != data_size) { - (void)kevent_put_data_size(p, data_available, flags, data_resid); + if ((kq->kqf_state & KQ_PROCESSING) == 0) { + break; } + + /* if someone else is processing the queue, wait */ + kq->kqf_state |= KQ_PROCWAIT; + suppressq = &kq->kqf_suppressed; + waitq_assert_wait64((struct waitq *)&kq->kqf_wqs, + CAST_EVENT64_T(suppressq), THREAD_UNINT | THREAD_WAIT_NOREPORT, + TIMEOUT_WAIT_FOREVER); + + kqunlock(kq); + thread_block(THREAD_CONTINUE_NULL); + kqlock(kq); } - /* detach the current thread if necessary */ - if (flags & KEVENT_FLAG_WORKLOOP_SERVICER_DETACH) { - assert(fp == NULL); - kevent_servicer_detach_thread(p, id, thread, flags, kq); + /* Nobody else processing */ + + /* clear pre-posts and KQ_WAKEUP now, in case we bail early */ + waitq_set_clear_preposts(&kq->kqf_wqs); + kq->kqf_state &= ~KQ_WAKEUP; + + /* anything left to process? */ + if (TAILQ_EMPTY(&kq->kqf_queue)) { + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQ_PROCESS_BEGIN) | DBG_FUNC_END, + VM_KERNEL_UNSLIDE_OR_PERM(kq), 1); + return -1; } -out: - kevent_put_kq(p, id, fp, kq); + /* convert to processing mode */ + kq->kqf_state |= KQ_PROCESSING; - /* don't restart after signals... */ - if (error == ERESTART) - error = EINTR; - else if (error == EWOULDBLOCK) - error = 0; - if (error == 0) - *retval = noutputs; - return (error); -} + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQ_PROCESS_BEGIN) | DBG_FUNC_END, + VM_KERNEL_UNSLIDE_OR_PERM(kq)); + return 0; +} /* - * kevent_callback - callback for each individual event + * Try to end the processing, only called when a workq thread is attempting to + * park (KEVENT_FLAG_PARKING is set). * - * called with nothing locked - * caller holds a reference on the kqueue + * When returning -1, the kqworkq is setup again so that it is ready to be + * processed. */ static int -kevent_callback(__unused struct kqueue *kq, struct kevent_internal_s *kevp, - void *data) +kqworkq_end_processing(struct kqworkq *kqwq, workq_threadreq_t kqr, + int kevent_flags) { - struct _kevent *cont_args; - int error; - - cont_args = (struct _kevent *)data; - assert(cont_args->eventout < cont_args->eventcount); + if (!TAILQ_EMPTY(&kqwq->kqwq_queue[kqr->tr_kq_qos_index])) { + /* remember we didn't process everything */ + kqr->tr_kq_wakeup = true; + } - /* - * Copy out the appropriate amount of event data for this user. - */ - error = kevent_copyout(kevp, &cont_args->eventlist, current_proc(), - cont_args->process_data.fp_flags); + if (kevent_flags & KEVENT_FLAG_PARKING) { + /* + * if acknowledge events "succeeds" it means there are events, + * which is a failure condition for end_processing. + */ + int rc = kqworkq_acknowledge_events(kqwq, kqr, kevent_flags, + KQWQAE_END_PROCESSING); + if (rc == 0) { + return -1; + } + } - /* - * If there isn't space for additional events, return - * a harmless error to stop the processing here - */ - if (error == 0 && ++cont_args->eventout == cont_args->eventcount) - error = EWOULDBLOCK; - return (error); + return 0; } /* - * kevent_description - format a description of a kevent for diagnostic output + * Try to end the processing, only called when a workq thread is attempting to + * park (KEVENT_FLAG_PARKING is set). + * + * When returning -1, the kqworkq is setup again so that it is ready to be + * processed (as if kqworkloop_begin_processing had just been called). * - * called with a 256-byte string buffer + * If successful and KEVENT_FLAG_PARKING was set in the kevent_flags, + * the kqworkloop is unbound from its servicer as a side effect. */ - -char * -kevent_description(struct kevent_internal_s *kevp, char *s, size_t n) +static int +kqworkloop_end_processing(struct kqworkloop *kqwl, int flags, int kevent_flags) { - snprintf(s, n, - "kevent=" - "{.ident=%#llx, .filter=%d, .flags=%#x, .udata=%#llx, .fflags=%#x, .data=%#llx, .ext[0]=%#llx, .ext[1]=%#llx}", - kevp->ident, - kevp->filter, - kevp->flags, - kevp->udata, - kevp->fflags, - kevp->data, - kevp->ext[0], - kevp->ext[1] ); + struct kqueue *kq = &kqwl->kqwl_kqueue; + workq_threadreq_t kqr = &kqwl->kqwl_request; + thread_qos_t qos_override; + thread_t thread = kqr_thread_fast(kqr); + int rc = 0; - return (s); -} + kqlock_held(kq); -/* - * kevent_register - add a new event to a kqueue - * - * Creates a mapping between the event source and - * the kqueue via a knote data structure. - * - * Because many/most the event sources are file - * descriptor related, the knote is linked off - * the filedescriptor table for quick access. - * - * called with nothing locked - * caller holds a reference on the kqueue - */ + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWL_PROCESS_END) | DBG_FUNC_START, + kqwl->kqwl_dynamicid, 0, 0); -void -kevent_register(struct kqueue *kq, struct kevent_internal_s *kev, - __unused struct proc *ctxp) -{ - struct proc *p = kq->kq_p; - const struct filterops *fops; - struct knote *kn = NULL; - int result = 0; - int error = 0; - unsigned short kev_flags = kev->flags; - int knoteuse_flags = KNUSE_NONE; + if (flags & KQ_PROCESSING) { + assert(kq->kq_state & KQ_PROCESSING); - if (kev->filter < 0) { - if (kev->filter + EVFILT_SYSCOUNT < 0) { - error = EINVAL; - goto out; + /* + * If we still have queued stayactive knotes, remember we didn't finish + * processing all of them. This should be extremely rare and would + * require to have a lot of them registered and fired. + */ + if (!TAILQ_EMPTY(&kqwl->kqwl_queue[KQWL_BUCKET_STAYACTIVE])) { + kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_UPDATE_WAKEUP_QOS, + KQWL_BUCKET_STAYACTIVE); } - fops = sysfilt_ops[~kev->filter]; /* to 0-base index */ - } else { - error = EINVAL; - goto out; - } - /* restrict EV_VANISHED to adding udata-specific dispatch kevents */ - if ((kev->flags & EV_VANISHED) && - (kev->flags & (EV_ADD | EV_DISPATCH2)) != (EV_ADD | EV_DISPATCH2)) { - error = EINVAL; - goto out; + /* + * When KEVENT_FLAG_PARKING is set, we need to attempt an unbind while + * still under the lock. + * + * So we do everything kqworkloop_unbind() would do, but because we're + * inside kqueue_process(), if the workloop actually received events + * while our locks were dropped, we have the opportunity to fail the end + * processing and loop again. + * + * This avoids going through the process-wide workqueue lock hence + * scales better. + */ + if (kevent_flags & KEVENT_FLAG_PARKING) { + qos_override = kqworkloop_acknowledge_events(kqwl); + } } - /* Simplify the flags - delete and disable overrule */ - if (kev->flags & EV_DELETE) - kev->flags &= ~EV_ADD; - if (kev->flags & EV_DISABLE) - kev->flags &= ~EV_ENABLE; - - if (kq->kq_state & KQ_WORKLOOP) { - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_REGISTER), - ((struct kqworkloop *)kq)->kqwl_dynamicid, - kev->udata, kev->flags, kev->filter); - } else if (kq->kq_state & KQ_WORKQ) { - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWQ_REGISTER), - 0, kev->udata, kev->flags, kev->filter); + if (kevent_flags & KEVENT_FLAG_PARKING) { + kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_PARKING, qos_override); + if (kqr->tr_kq_wakeup && !kqwl->kqwl_owner) { + /* + * Reset wakeup bit to notice stay active events firing while we are + * processing, as we cannot rely on the stayactive bucket emptiness. + */ + kqwl->kqwl_wakeup_indexes &= ~KQWL_STAYACTIVE_FIRED_BIT; + rc = -1; + } else { + kqworkloop_unbind_locked(kqwl, thread, KQWL_OVERRIDE_DROP_DELAYED); + kqworkloop_release_live(kqwl); + kq->kq_state &= ~flags; + } } else { - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQ_REGISTER), - VM_KERNEL_UNSLIDE_OR_PERM(kq), - kev->udata, kev->flags, kev->filter); + kq->kq_state &= ~flags; + kq->kq_state |= KQ_R2K_ARMED; + kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_RECOMPUTE_WAKEUP_QOS, 0); } -restart: - - /* find the matching knote from the fd tables/hashes */ - kn = kq_find_knote_and_kq_lock(kq, kev, fops->f_isfd, p); + if ((kevent_flags & KEVENT_FLAG_PARKING) && rc == 0) { + kqworkloop_unbind_delayed_override_drop(thread); + } - if (kn == NULL) { - if (kev->flags & EV_ADD) { - struct fileproc *knote_fp = NULL; - - /* grab a file reference for the new knote */ - if (fops->f_isfd) { - if ((error = fp_lookup(p, kev->ident, &knote_fp, 0)) != 0) { - goto out; - } - } - - kn = knote_alloc(); - if (kn == NULL) { - error = ENOMEM; - if (knote_fp != NULL) - fp_drop(p, kev->ident, knote_fp, 0); - goto out; - } - - kn->kn_fp = knote_fp; - knote_set_kq(kn, kq); - kqueue_retain(kq); /* retain a kq ref */ - kn->kn_filtid = ~kev->filter; - kn->kn_inuse = 1; /* for f_attach() */ - kn->kn_status = KN_ATTACHING | KN_ATTACHED; - - /* was vanish support requested */ - if (kev->flags & EV_VANISHED) { - kev->flags &= ~EV_VANISHED; - kn->kn_status |= KN_REQVANISH; - } - - /* snapshot matching/dispatching protcol flags into knote */ - if (kev->flags & EV_DISPATCH) - kn->kn_status |= KN_DISPATCH; - if (kev->flags & EV_UDATA_SPECIFIC) - kn->kn_status |= KN_UDATA_SPECIFIC; - - /* - * copy the kevent state into knote - * protocol is that fflags and data - * are saved off, and cleared before - * calling the attach routine. - */ - kn->kn_kevent = *kev; - kn->kn_sfflags = kev->fflags; - kn->kn_sdata = kev->data; - kn->kn_fflags = 0; - kn->kn_data = 0; - - /* invoke pthread kext to convert kevent qos to thread qos */ - knote_canonicalize_kevent_qos(kn); - knote_set_qos_index(kn, qos_index_from_qos(kn, kn->kn_qos, FALSE)); - - /* before anyone can find it */ - if (kev->flags & EV_DISABLE) { - /* - * do this before anyone can find it, - * this can't call knote_disable() because it expects having - * the kqlock held - */ - kn->kn_status |= KN_DISABLED; - } - - /* Add the knote for lookup thru the fd table */ - error = kq_add_knote(kq, kn, kev, p, &knoteuse_flags); - if (error) { - (void)kqueue_release(kq, KQUEUE_CANT_BE_LAST_REF); - knote_free(kn); - if (knote_fp != NULL) - fp_drop(p, kev->ident, knote_fp, 0); - - if (error == ERESTART) { - error = 0; - goto restart; - } - goto out; - } - - /* fp reference count now applies to knote */ - /* rwlock boost is now held */ - - /* call filter attach routine */ - result = fops->f_attach(kn, kev); - - /* - * Trade knote use count for kq lock. - * Cannot be dropped because we held - * KN_ATTACHING throughout. - */ - knoteuse2kqlock(kq, kn, KNUSE_STEAL_DROP | knoteuse_flags); - - if (kn->kn_flags & EV_ERROR) { - /* - * Failed to attach correctly, so drop. - * All other possible users/droppers - * have deferred to us. Save the error - * to return to our caller. - */ - kn->kn_status &= ~KN_ATTACHED; - kn->kn_status |= KN_DROPPING; - error = kn->kn_data; - kqunlock(kq); - knote_drop(kn, p); - goto out; - } - - /* end "attaching" phase - now just attached */ - kn->kn_status &= ~KN_ATTACHING; - - if (kn->kn_status & KN_DROPPING) { - /* - * Attach succeeded, but someone else - * deferred their drop - now we have - * to do it for them. - */ - kqunlock(kq); - knote_drop(kn, p); - goto out; - } - - /* Mark the thread request overcommit - if appropos */ - knote_set_qos_overcommit(kn); - - /* - * If the attach routine indicated that an - * event is already fired, activate the knote. - */ - if (result) - knote_activate(kn); - - if (knote_fops(kn)->f_post_attach) { - error = knote_fops(kn)->f_post_attach(kn, kev); - if (error) { - kqunlock(kq); - goto out; - } - } - - } else { - if ((kev_flags & (EV_ADD | EV_DELETE)) == (EV_ADD | EV_DELETE) && - (kq->kq_state & KQ_WORKLOOP)) { - /* - * For workloops, understand EV_ADD|EV_DELETE as a "soft" delete - * that doesn't care about ENOENT, so just pretend the deletion - * happened. - */ - } else { - error = ENOENT; - } - goto out; - } - - } else { - /* existing knote: kqueue lock already taken by kq_find_knote_and_kq_lock */ - - if ((kn->kn_status & (KN_DROPPING | KN_ATTACHING)) != 0) { - /* - * The knote is not in a stable state, wait for that - * transition to complete and then redrive the lookup. - */ - knoteusewait(kq, kn); - goto restart; - } - - if (kev->flags & EV_DELETE) { - - /* - * If attempting to delete a disabled dispatch2 knote, - * we must wait for the knote to be re-enabled (unless - * it is being re-enabled atomically here). - */ - if ((kev->flags & EV_ENABLE) == 0 && - (kn->kn_status & (KN_DISPATCH2 | KN_DISABLED)) == - (KN_DISPATCH2 | KN_DISABLED)) { - kn->kn_status |= KN_DEFERDELETE; - kqunlock(kq); - error = EINPROGRESS; - } else if (knote_fops(kn)->f_drop_and_unlock) { - /* - * The filter has requested to handle EV_DELETE events - * - * ERESTART means the kevent has to be re-evaluated - */ - error = knote_fops(kn)->f_drop_and_unlock(kn, kev); - if (error == ERESTART) { - error = 0; - goto restart; - } - } else if (kqlock2knotedrop(kq, kn)) { - /* standard/default EV_DELETE path */ - knote_drop(kn, p); - } else { - /* - * The kqueue is unlocked, it's not being - * dropped, and kqlock2knotedrop returned 0: - * this means that someone stole the drop of - * the knote from us. - */ - error = EINPROGRESS; - } - goto out; - } - - /* - * If we are re-enabling a deferred-delete knote, - * just enable it now and avoid calling the - * filter touch routine (it has delivered its - * last event already). - */ - if ((kev->flags & EV_ENABLE) && - (kn->kn_status & KN_DEFERDELETE)) { - assert(kn->kn_status & KN_DISABLED); - knote_activate(kn); - knote_enable(kn); - kqunlock(kq); - goto out; - } - - /* - * If we are disabling, do it before unlocking and - * calling the touch routine (so no processing can - * see the new kevent state before the disable is - * applied). - */ - if (kev->flags & EV_DISABLE) - knote_disable(kn); - - /* - * Convert the kqlock to a use reference on the - * knote so we can call the filter touch routine. - */ - if (knoteuse_needs_boost(kn, kev)) { - knoteuse_flags |= KNUSE_BOOST; - } - if (kqlock2knoteuse(kq, kn, knoteuse_flags)) { - /* - * Call touch routine to notify filter of changes - * in filter values (and to re-determine if any - * events are fired). - */ - result = knote_fops(kn)->f_touch(kn, kev); - - /* Get the kq lock back (don't defer droppers). */ - if (!knoteuse2kqlock(kq, kn, knoteuse_flags)) { - kqunlock(kq); - goto out; - } - - /* Handle errors during touch routine */ - if (kev->flags & EV_ERROR) { - error = kev->data; - kqunlock(kq); - goto out; - } - - /* Activate it if the touch routine said to */ - if (result) - knote_activate(kn); - } - - /* Enable the knote if called for */ - if (kev->flags & EV_ENABLE) - knote_enable(kn); - - } - - /* still have kqlock held and knote is valid */ - kqunlock(kq); - -out: - /* output local errors through the kevent */ - if (error) { - kev->flags |= EV_ERROR; - kev->data = error; - } -} + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWL_PROCESS_END) | DBG_FUNC_END, + kqwl->kqwl_dynamicid, 0, 0); + return rc; +} /* - * knote_process - process a triggered event - * - * Validate that it is really still a triggered event - * by calling the filter routines (if necessary). Hold - * a use reference on the knote to avoid it being detached. - * - * If it is still considered triggered, we will have taken - * a copy of the state under the filter lock. We use that - * snapshot to dispatch the knote for future processing (or - * not, if this was a lost event). - * - * Our caller assures us that nobody else can be processing - * events from this knote during the whole operation. But - * others can be touching or posting events to the knote - * interspersed with our processing it. + * Called with kqueue lock held. * - * caller holds a reference on the kqueue. - * kqueue locked on entry and exit - but may be dropped + * 0: no more events + * -1: has more events + * EBADF: kqueue is in draining mode */ static int -knote_process(struct knote *kn, - kevent_callback_t callback, - void *callback_data, - struct filt_process_s *process_data, - struct proc *p) +kqfile_end_processing(struct kqfile *kq) { - struct kevent_internal_s kev; - struct kqueue *kq = knote_get_kq(kn); - int result = 0; - int error = 0; + struct kqtailq *suppressq = &kq->kqf_suppressed; + struct knote *kn; + int procwait; - bzero(&kev, sizeof(kev)); + kqlock_held(kq); - /* - * Must be active or stayactive - * Must be queued and not disabled/suppressed - */ - assert(kn->kn_status & KN_QUEUED); - assert(kn->kn_status & (KN_ACTIVE|KN_STAYACTIVE)); - assert(!(kn->kn_status & (KN_DISABLED|KN_SUPPRESSED|KN_DROPPING))); + assert((kq->kqf_state & (KQ_WORKQ | KQ_WORKLOOP)) == 0); - if (kq->kq_state & KQ_WORKLOOP) { - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_PROCESS), - ((struct kqworkloop *)kq)->kqwl_dynamicid, - kn->kn_udata, kn->kn_status | (kn->kn_id << 32), - kn->kn_filtid); - } else if (kq->kq_state & KQ_WORKQ) { - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWQ_PROCESS), - 0, kn->kn_udata, kn->kn_status | (kn->kn_id << 32), - kn->kn_filtid); - } else { - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQ_PROCESS), - VM_KERNEL_UNSLIDE_OR_PERM(kq), kn->kn_udata, - kn->kn_status | (kn->kn_id << 32), kn->kn_filtid); - } + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQ_PROCESS_END), + VM_KERNEL_UNSLIDE_OR_PERM(kq), 0); /* - * For deferred-drop or vanished events, we just create a fake - * event to acknowledge end-of-life. Otherwise, we call the - * filter's process routine to snapshot the kevent state under - * the filter's locking protocol. + * Return suppressed knotes to their original state. */ - if (kn->kn_status & (KN_DEFERDELETE | KN_VANISHED)) { - /* create fake event */ - kev.filter = kn->kn_filter; - kev.ident = kn->kn_id; - kev.qos = kn->kn_qos; - kev.flags = (kn->kn_status & KN_DEFERDELETE) ? - EV_DELETE : EV_VANISHED; - kev.flags |= (EV_DISPATCH2 | EV_ONESHOT); - kev.udata = kn->kn_udata; - result = 1; - - knote_suppress(kn); - } else { - int flags = KNUSE_NONE; - /* deactivate - so new activations indicate a wakeup */ - knote_deactivate(kn); - - /* suppress knotes to avoid returning the same event multiple times in a single call. */ - knote_suppress(kn); - - if (knoteuse_needs_boost(kn, NULL)) { - flags |= KNUSE_BOOST; - } - /* convert lock to a knote use reference */ - if (!kqlock2knoteuse(kq, kn, flags)) - panic("dropping knote found on queue\n"); - - /* call out to the filter to process with just a ref */ - result = knote_fops(kn)->f_process(kn, process_data, &kev); - if (result) flags |= KNUSE_STEAL_DROP; - - /* - * convert our reference back to a lock. accept drop - * responsibility from others if we've committed to - * delivering event data. - */ - if (!knoteuse2kqlock(kq, kn, flags)) { - /* knote dropped */ - kn = NULL; - } + while ((kn = TAILQ_FIRST(suppressq)) != NULL) { + assert(kn->kn_status & KN_SUPPRESSED); + knote_unsuppress(kq, kn); } - if (kn != NULL) { - /* - * Determine how to dispatch the knote for future event handling. - * not-fired: just return (do not callout, leave deactivated). - * One-shot: If dispatch2, enter deferred-delete mode (unless this is - * is the deferred delete event delivery itself). Otherwise, - * drop it. - * stolendrop:We took responsibility for someone else's drop attempt. - * treat this just like one-shot and prepare to turn it back - * into a deferred delete if required. - * Dispatch: don't clear state, just mark it disabled. - * Cleared: just leave it deactivated. - * Others: re-activate as there may be more events to handle. - * This will not wake up more handlers right now, but - * at the completion of handling events it may trigger - * more handler threads (TODO: optimize based on more than - * just this one event being detected by the filter). - */ + procwait = (kq->kqf_state & KQ_PROCWAIT); + kq->kqf_state &= ~(KQ_PROCESSING | KQ_PROCWAIT); - if (result == 0) - return (EJUSTRETURN); - - if ((kev.flags & EV_ONESHOT) || (kn->kn_status & KN_STOLENDROP)) { - if ((kn->kn_status & (KN_DISPATCH2 | KN_DEFERDELETE)) == KN_DISPATCH2) { - /* defer dropping non-delete oneshot dispatch2 events */ - kn->kn_status |= KN_DEFERDELETE; - knote_disable(kn); - - /* if we took over another's drop clear those flags here */ - if (kn->kn_status & KN_STOLENDROP) { - assert(kn->kn_status & KN_DROPPING); - /* - * the knote will be dropped when the - * deferred deletion occurs - */ - kn->kn_status &= ~(KN_DROPPING|KN_STOLENDROP); - } - } else if (kn->kn_status & KN_STOLENDROP) { - /* We now own the drop of the knote. */ - assert(kn->kn_status & KN_DROPPING); - knote_unsuppress(kn); - kqunlock(kq); - knote_drop(kn, p); - kqlock(kq); - } else if (kqlock2knotedrop(kq, kn)) { - /* just EV_ONESHOT, _not_ DISPATCH2 */ - knote_drop(kn, p); - kqlock(kq); - } - } else if (kn->kn_status & KN_DISPATCH) { - /* disable all dispatch knotes */ - knote_disable(kn); - } else if ((kev.flags & EV_CLEAR) == 0) { - /* re-activate in case there are more events */ - knote_activate(kn); - } + if (procwait) { + /* first wake up any thread already waiting to process */ + waitq_wakeup64_all((struct waitq *)&kq->kqf_wqs, + CAST_EVENT64_T(suppressq), THREAD_AWAKENED, WAITQ_ALL_PRIORITIES); } - /* - * callback to handle each event as we find it. - * If we have to detach and drop the knote, do - * it while we have the kq unlocked. - */ - if (result) { - kqunlock(kq); - error = (callback)(kq, &kev, callback_data); - kqlock(kq); + if (kq->kqf_state & KQ_DRAIN) { + return EBADF; } - return (error); + return (kq->kqf_state & KQ_WAKEUP) ? -1 : 0; } - -/* - * Return 0 to indicate that processing should proceed, - * -1 if there is nothing to process. - * - * Called with kqueue locked and returns the same way, - * but may drop lock temporarily. - */ static int -kqworkq_begin_processing(struct kqworkq *kqwq, kq_index_t qos_index, int flags) +kqueue_workloop_ctl_internal(proc_t p, uintptr_t cmd, uint64_t __unused options, + struct kqueue_workloop_params *params, int *retval) { - struct kqrequest *kqr; - thread_t self = current_thread(); - __assert_only struct uthread *ut = get_bsdthread_info(self); - - assert(kqwq->kqwq_state & KQ_WORKQ); - assert(qos_index < KQWQ_NQOS); - - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWQ_PROCESS_BEGIN) | DBG_FUNC_START, - flags, qos_index); - - kqwq_req_lock(kqwq); - - kqr = kqworkq_get_request(kqwq, qos_index); - - /* manager skips buckets that haven't asked for its help */ - if (flags & KEVENT_FLAG_WORKQ_MANAGER) { + int error = 0; + struct kqworkloop *kqwl; + struct filedesc *fdp = p->p_fd; + workq_threadreq_param_t trp = { }; - /* If nothing for manager to do, just return */ - if ((kqr->kqr_state & KQWQ_THMANAGER) == 0) { - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWQ_PROCESS_BEGIN) | DBG_FUNC_END, - 0, kqr->kqr_state); - kqwq_req_unlock(kqwq); - return -1; + switch (cmd) { + case KQ_WORKLOOP_CREATE: + if (!params->kqwlp_flags) { + error = EINVAL; + break; } - /* bind manager thread from this time on */ - kqworkq_bind_thread_impl(kqwq, qos_index, self, flags); - - } else { - /* We should already be bound to this kqueue */ - assert(kqr->kqr_state & KQR_BOUND); - assert(kqr->kqr_thread == self); - assert(ut->uu_kqueue_bound == (struct kqueue *)kqwq); - assert(ut->uu_kqueue_qos_index == qos_index); - assert((ut->uu_kqueue_flags & flags) == ut->uu_kqueue_flags); - } - - /* - * we should have been requested to be here - * and nobody else should still be processing - */ - assert(kqr->kqr_state & KQR_WAKEUP); - assert(kqr->kqr_state & KQR_THREQUESTED); - assert((kqr->kqr_state & KQR_PROCESSING) == 0); - - /* reset wakeup trigger to catch new events after we start processing */ - kqr->kqr_state &= ~KQR_WAKEUP; - - /* convert to processing mode */ - kqr->kqr_state |= KQR_PROCESSING; - - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWQ_PROCESS_BEGIN) | DBG_FUNC_END, - kqr_thread_id(kqr), kqr->kqr_state); - - kqwq_req_unlock(kqwq); - return 0; -} - -static inline bool -kqworkloop_is_processing_on_current_thread(struct kqworkloop *kqwl) -{ - struct kqueue *kq = &kqwl->kqwl_kqueue; - - kqlock_held(kq); - if (kq->kq_state & KQ_PROCESSING) { - /* - * KQ_PROCESSING is unset with the kqlock held, and the kqr thread is - * never modified while KQ_PROCESSING is set, meaning that peeking at - * its value is safe from this context. - */ - return kqwl->kqwl_request.kqr_thread == current_thread(); - } - return false; -} - -static void -kqworkloop_acknowledge_events(struct kqworkloop *kqwl, boolean_t clear_ipc_override) -{ - struct kqrequest *kqr = &kqwl->kqwl_request; - struct knote *kn, *tmp; - - kqlock_held(&kqwl->kqwl_kqueue); - - TAILQ_FOREACH_SAFE(kn, &kqr->kqr_suppressed, kn_tqe, tmp) { - /* - * If a knote that can adjust QoS is disabled because of the automatic - * behavior of EV_DISPATCH, the knotes should stay suppressed so that - * further overrides keep pushing. - */ - if (knote_fops(kn)->f_adjusts_qos && (kn->kn_status & KN_DISABLED) && - (kn->kn_status & (KN_STAYACTIVE | KN_DROPPING)) == 0 && - (kn->kn_flags & (EV_DISPATCH | EV_DISABLE)) == EV_DISPATCH) { - /* - * When called from unbind, clear the sync ipc override on the knote - * for events which are delivered. - */ - if (clear_ipc_override) { - knote_adjust_sync_qos(kn, THREAD_QOS_UNSPECIFIED, FALSE); - } - continue; + if ((params->kqwlp_flags & KQ_WORKLOOP_CREATE_SCHED_PRI) && + (params->kqwlp_sched_pri < 1 || + params->kqwlp_sched_pri > 63 /* MAXPRI_USER */)) { + error = EINVAL; + break; } - knote_unsuppress(kn); - } -} - -static int -kqworkloop_begin_processing(struct kqworkloop *kqwl, - __assert_only unsigned int flags) -{ - struct kqrequest *kqr = &kqwl->kqwl_request; - struct kqueue *kq = &kqwl->kqwl_kqueue; - - kqlock_held(kq); - - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_PROCESS_BEGIN) | DBG_FUNC_START, - kqwl->kqwl_dynamicid, flags, 0); - - kqwl_req_lock(kqwl); - - /* nobody else should still be processing */ - assert((kqr->kqr_state & KQR_PROCESSING) == 0); - assert((kq->kq_state & KQ_PROCESSING) == 0); - - kqr->kqr_state |= KQR_PROCESSING | KQR_R2K_NOTIF_ARMED; - kq->kq_state |= KQ_PROCESSING; - - kqwl_req_unlock(kqwl); - - kqworkloop_acknowledge_events(kqwl, FALSE); - - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_PROCESS_BEGIN) | DBG_FUNC_END, - kqwl->kqwl_dynamicid, flags, 0); - - return 0; -} - -/* - * Return 0 to indicate that processing should proceed, - * -1 if there is nothing to process. - * - * Called with kqueue locked and returns the same way, - * but may drop lock temporarily. - * May block. - */ -static int -kqueue_begin_processing(struct kqueue *kq, kq_index_t qos_index, unsigned int flags) -{ - struct kqtailq *suppressq; - - kqlock_held(kq); - - if (kq->kq_state & KQ_WORKQ) { - return kqworkq_begin_processing((struct kqworkq *)kq, qos_index, flags); - } else if (kq->kq_state & KQ_WORKLOOP) { - return kqworkloop_begin_processing((struct kqworkloop*)kq, flags); - } - - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQ_PROCESS_BEGIN) | DBG_FUNC_START, - VM_KERNEL_UNSLIDE_OR_PERM(kq), flags); - assert(qos_index == QOS_INDEX_KQFILE); - - /* wait to become the exclusive processing thread */ - for (;;) { - if (kq->kq_state & KQ_DRAIN) { - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQ_PROCESS_BEGIN) | DBG_FUNC_END, - VM_KERNEL_UNSLIDE_OR_PERM(kq), 2); - return -1; + if ((params->kqwlp_flags & KQ_WORKLOOP_CREATE_SCHED_POL) && + invalid_policy(params->kqwlp_sched_pol)) { + error = EINVAL; + break; } - if ((kq->kq_state & KQ_PROCESSING) == 0) + if ((params->kqwlp_flags & KQ_WORKLOOP_CREATE_CPU_PERCENT) && + (params->kqwlp_cpu_percent <= 0 || + params->kqwlp_cpu_percent > 100 || + params->kqwlp_cpu_refillms <= 0 || + params->kqwlp_cpu_refillms > 0x00ffffff)) { + error = EINVAL; break; - - /* if someone else is processing the queue, wait */ - kq->kq_state |= KQ_PROCWAIT; - suppressq = kqueue_get_suppressed_queue(kq, qos_index); - waitq_assert_wait64((struct waitq *)&kq->kq_wqs, - CAST_EVENT64_T(suppressq), - THREAD_UNINT, TIMEOUT_WAIT_FOREVER); - - kqunlock(kq); - thread_block(THREAD_CONTINUE_NULL); - kqlock(kq); - } - - /* Nobody else processing */ - - /* clear pre-posts and KQ_WAKEUP now, in case we bail early */ - waitq_set_clear_preposts(&kq->kq_wqs); - kq->kq_state &= ~KQ_WAKEUP; - - /* anything left to process? */ - if (kqueue_queue_empty(kq, qos_index)) { - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQ_PROCESS_BEGIN) | DBG_FUNC_END, - VM_KERNEL_UNSLIDE_OR_PERM(kq), 1); - return -1; - } - - /* convert to processing mode */ - kq->kq_state |= KQ_PROCESSING; - - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQ_PROCESS_BEGIN) | DBG_FUNC_END, - VM_KERNEL_UNSLIDE_OR_PERM(kq)); - - return 0; -} - -/* - * kqworkq_end_processing - Complete the processing of a workq kqueue - * - * We may have to request new threads. - * This can happen there are no waiting processing threads and: - * - there were active events we never got to (count > 0) - * - we pended waitq hook callouts during processing - * - we pended wakeups while processing (or unsuppressing) - * - * Called with kqueue lock held. - */ -static void -kqworkq_end_processing(struct kqworkq *kqwq, kq_index_t qos_index, int flags) -{ -#pragma unused(flags) - - struct kqueue *kq = &kqwq->kqwq_kqueue; - struct kqtailq *suppressq = kqueue_get_suppressed_queue(kq, qos_index); - - thread_t self = current_thread(); - struct uthread *ut = get_bsdthread_info(self); - struct knote *kn; - struct kqrequest *kqr; - thread_t thread; - - assert(kqwq->kqwq_state & KQ_WORKQ); - assert(qos_index < KQWQ_NQOS); - - /* Are we really bound to this kqueue? */ - if (ut->uu_kqueue_bound != kq) { - assert(ut->uu_kqueue_bound == kq); - return; - } - - kqr = kqworkq_get_request(kqwq, qos_index); - - kqwq_req_lock(kqwq); - - /* Do we claim to be manager? */ - if (flags & KEVENT_FLAG_WORKQ_MANAGER) { - - /* bail if not bound that way */ - if (ut->uu_kqueue_qos_index != KQWQ_QOS_MANAGER || - (ut->uu_kqueue_flags & KEVENT_FLAG_WORKQ_MANAGER) == 0) { - assert(ut->uu_kqueue_qos_index == KQWQ_QOS_MANAGER); - assert(ut->uu_kqueue_flags & KEVENT_FLAG_WORKQ_MANAGER); - kqwq_req_unlock(kqwq); - return; } - /* bail if this request wasn't already getting manager help */ - if ((kqr->kqr_state & KQWQ_THMANAGER) == 0 || - (kqr->kqr_state & KQR_PROCESSING) == 0) { - kqwq_req_unlock(kqwq); - return; + if (params->kqwlp_flags & KQ_WORKLOOP_CREATE_SCHED_PRI) { + trp.trp_flags |= TRP_PRIORITY; + trp.trp_pri = (uint8_t)params->kqwlp_sched_pri; } - } else { - if (ut->uu_kqueue_qos_index != qos_index || - (ut->uu_kqueue_flags & KEVENT_FLAG_WORKQ_MANAGER)) { - assert(ut->uu_kqueue_qos_index == qos_index); - assert((ut->uu_kqueue_flags & KEVENT_FLAG_WORKQ_MANAGER) == 0); - kqwq_req_unlock(kqwq); - return; + if (params->kqwlp_flags & KQ_WORKLOOP_CREATE_SCHED_POL) { + trp.trp_flags |= TRP_POLICY; + trp.trp_pol = (uint8_t)params->kqwlp_sched_pol; + } + if (params->kqwlp_flags & KQ_WORKLOOP_CREATE_CPU_PERCENT) { + trp.trp_flags |= TRP_CPUPERCENT; + trp.trp_cpupercent = (uint8_t)params->kqwlp_cpu_percent; + trp.trp_refillms = params->kqwlp_cpu_refillms; } - } - - assert(kqr->kqr_state & KQR_BOUND); - thread = kqr->kqr_thread; - assert(thread == self); - - assert(kqr->kqr_state & KQR_PROCESSING); - - /* If we didn't drain the whole queue, re-mark a wakeup being needed */ - if (!kqueue_queue_empty(kq, qos_index)) - kqr->kqr_state |= KQR_WAKEUP; - - kqwq_req_unlock(kqwq); - - /* - * Return suppressed knotes to their original state. - * For workq kqueues, suppressed ones that are still - * truly active (not just forced into the queue) will - * set flags we check below to see if anything got - * woken up. - */ - while ((kn = TAILQ_FIRST(suppressq)) != NULL) { - assert(kn->kn_status & KN_SUPPRESSED); - knote_unsuppress(kn); - } - - kqwq_req_lock(kqwq); - - /* Indicate that we are done processing this request */ - kqr->kqr_state &= ~KQR_PROCESSING; - /* - * Drop our association with this one request and its - * override on us. - */ - kqworkq_unbind_thread(kqwq, qos_index, thread, flags); + error = kqworkloop_get_or_create(p, params->kqwlp_id, &trp, + KEVENT_FLAG_DYNAMIC_KQUEUE | KEVENT_FLAG_WORKLOOP | + KEVENT_FLAG_DYNAMIC_KQ_MUST_NOT_EXIST, &kqwl); + if (error) { + break; + } - /* - * request a new thread if we didn't process the whole - * queue or real events have happened (not just putting - * stay-active events back). - */ - if (kqr->kqr_state & KQR_WAKEUP) { - if (kqueue_queue_empty(kq, qos_index)) { - kqr->kqr_state &= ~KQR_WAKEUP; + if (!(fdp->fd_flags & FD_WORKLOOP)) { + /* FD_WORKLOOP indicates we've ever created a workloop + * via this syscall but its only ever added to a process, never + * removed. + */ + proc_fdlock(p); + fdp->fd_flags |= FD_WORKLOOP; + proc_fdunlock(p); + } + break; + case KQ_WORKLOOP_DESTROY: + error = kqworkloop_get_or_create(p, params->kqwlp_id, NULL, + KEVENT_FLAG_DYNAMIC_KQUEUE | KEVENT_FLAG_WORKLOOP | + KEVENT_FLAG_DYNAMIC_KQ_MUST_EXIST, &kqwl); + if (error) { + break; + } + kqlock(kqwl); + trp.trp_value = kqwl->kqwl_params; + if (trp.trp_flags && !(trp.trp_flags & TRP_RELEASED)) { + trp.trp_flags |= TRP_RELEASED; + kqwl->kqwl_params = trp.trp_value; + kqworkloop_release_live(kqwl); } else { - kqworkq_request_thread(kqwq, qos_index); + error = EINVAL; } + kqunlock(kqwl); + kqworkloop_release(kqwl); + break; } - kqwq_req_unlock(kqwq); + *retval = 0; + return error; } -static void -kqworkloop_end_processing(struct kqworkloop *kqwl, int nevents, - unsigned int flags) +int +kqueue_workloop_ctl(proc_t p, struct kqueue_workloop_ctl_args *uap, int *retval) { - struct kqrequest *kqr = &kqwl->kqwl_request; - struct kqueue *kq = &kqwl->kqwl_kqueue; - - kqlock_held(kq); - - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_PROCESS_END) | DBG_FUNC_START, - kqwl->kqwl_dynamicid, flags, 0); - - if ((kq->kq_state & KQ_NO_WQ_THREAD) && nevents == 0 && - (flags & KEVENT_FLAG_IMMEDIATE) == 0) { - /* - * We may soon block, but have returned no - * kevents that need to be kept supressed for overriding purposes. - * - * It is hence safe to acknowledge events and unsuppress everything, so - * that if we block we can observe all events firing. - */ - kqworkloop_acknowledge_events(kqwl, TRUE); + struct kqueue_workloop_params params = { + .kqwlp_id = 0, + }; + if (uap->sz < sizeof(params.kqwlp_version)) { + return EINVAL; } - kqwl_req_lock(kqwl); - - assert(kqr->kqr_state & KQR_PROCESSING); - assert(kq->kq_state & KQ_PROCESSING); - - kq->kq_state &= ~KQ_PROCESSING; - kqr->kqr_state &= ~KQR_PROCESSING; - kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_RECOMPUTE_WAKEUP_QOS, 0); - - kqwl_req_unlock(kqwl); - - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_PROCESS_END) | DBG_FUNC_END, - kqwl->kqwl_dynamicid, flags, 0); -} - -/* - * Called with kqueue lock held. - */ -static void -kqueue_end_processing(struct kqueue *kq, kq_index_t qos_index, - int nevents, unsigned int flags) -{ - struct knote *kn; - struct kqtailq *suppressq; - int procwait; - - kqlock_held(kq); - - assert((kq->kq_state & KQ_WORKQ) == 0); - - if (kq->kq_state & KQ_WORKLOOP) { - return kqworkloop_end_processing((struct kqworkloop *)kq, nevents, flags); + size_t copyin_sz = MIN(sizeof(params), uap->sz); + int rv = copyin(uap->addr, ¶ms, copyin_sz); + if (rv) { + return rv; } - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQ_PROCESS_END), - VM_KERNEL_UNSLIDE_OR_PERM(kq), flags); - - assert(qos_index == QOS_INDEX_KQFILE); - - /* - * Return suppressed knotes to their original state. - */ - suppressq = kqueue_get_suppressed_queue(kq, qos_index); - while ((kn = TAILQ_FIRST(suppressq)) != NULL) { - assert(kn->kn_status & KN_SUPPRESSED); - knote_unsuppress(kn); + if (params.kqwlp_version != (int)uap->sz) { + return EINVAL; } - procwait = (kq->kq_state & KQ_PROCWAIT); - kq->kq_state &= ~(KQ_PROCESSING | KQ_PROCWAIT); - - if (procwait) { - /* first wake up any thread already waiting to process */ - waitq_wakeup64_all((struct waitq *)&kq->kq_wqs, - CAST_EVENT64_T(suppressq), - THREAD_AWAKENED, - WAITQ_ALL_PRIORITIES); - } + return kqueue_workloop_ctl_internal(p, uap->cmd, uap->options, ¶ms, + retval); } -/* - * kqwq_internal_bind - bind thread to processing workq kqueue - * - * Determines if the provided thread will be responsible for - * servicing the particular QoS class index specified in the - * parameters. Once the binding is done, any overrides that may - * be associated with the cooresponding events can be applied. - * - * This should be called as soon as the thread identity is known, - * preferably while still at high priority during creation. - * - * - caller holds a reference on the process (and workq kq) - * - the thread MUST call kevent_qos_internal after being bound - * or the bucket of events may never be delivered. - * - Nothing locked - * (unless this is a synchronous bind, then the request is locked) - */ +/*ARGSUSED*/ static int -kqworkq_internal_bind( - struct proc *p, - kq_index_t qos_index, - thread_t thread, - unsigned int flags) +kqueue_select(struct fileproc *fp, int which, void *wq_link_id, + __unused vfs_context_t ctx) { - struct kqueue *kq; - struct kqworkq *kqwq; - struct kqrequest *kqr; - struct uthread *ut = get_bsdthread_info(thread); - - /* If no process workq, can't be our thread. */ - kq = p->p_fd->fd_wqkqueue; + struct kqfile *kq = (struct kqfile *)fp->f_data; + struct kqtailq *suppressq = &kq->kqf_suppressed; + struct kqtailq *queue = &kq->kqf_queue; + struct knote *kn; + int retnum = 0; - if (kq == NULL) + if (which != FREAD) { return 0; - - assert(kq->kq_state & KQ_WORKQ); - kqwq = (struct kqworkq *)kq; - - /* - * No need to bind the manager thread to any specific - * bucket, but still claim the thread. - */ - if (qos_index == KQWQ_QOS_MANAGER) { - assert(ut->uu_kqueue_bound == NULL); - assert(flags & KEVENT_FLAG_WORKQ_MANAGER); - ut->uu_kqueue_bound = kq; - ut->uu_kqueue_qos_index = qos_index; - ut->uu_kqueue_flags = flags; - - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWQ_BIND), - thread_tid(thread), flags, qos_index); - - return 1; } - /* - * If this is a synchronous bind callback, the request - * lock is already held, so just do the bind. - */ - if (flags & KEVENT_FLAG_SYNCHRONOUS_BIND) { - kqwq_req_held(kqwq); - /* strip out synchronout bind flag */ - flags &= ~KEVENT_FLAG_SYNCHRONOUS_BIND; - kqworkq_bind_thread_impl(kqwq, qos_index, thread, flags); - return 1; - } + kqlock(kq); + + assert((kq->kqf_state & KQ_WORKQ) == 0); /* - * check the request that corresponds to our qos_index - * to see if there is an outstanding request. + * If this is the first pass, link the wait queue associated with the + * the kqueue onto the wait queue set for the select(). Normally we + * use selrecord() for this, but it uses the wait queue within the + * selinfo structure and we need to use the main one for the kqueue to + * catch events from KN_STAYQUEUED sources. So we do the linkage manually. + * (The select() call will unlink them when it ends). */ - kqr = kqworkq_get_request(kqwq, qos_index); - assert(kqr->kqr_qos_index == qos_index); - kqwq_req_lock(kqwq); + if (wq_link_id != NULL) { + thread_t cur_act = current_thread(); + struct uthread * ut = get_bsdthread_info(cur_act); - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWQ_BIND), - thread_tid(thread), flags, qos_index, kqr->kqr_state); + kq->kqf_state |= KQ_SEL; + waitq_link((struct waitq *)&kq->kqf_wqs, ut->uu_wqset, + WAITQ_SHOULD_LOCK, (uint64_t *)wq_link_id); - if ((kqr->kqr_state & KQR_THREQUESTED) && - (kqr->kqr_state & KQR_PROCESSING) == 0) { + /* always consume the reserved link object */ + waitq_link_release(*(uint64_t *)wq_link_id); + *(uint64_t *)wq_link_id = 0; - if ((kqr->kqr_state & KQR_BOUND) && - thread == kqr->kqr_thread) { - /* duplicate bind - claim the thread */ - assert(ut->uu_kqueue_bound == kq); - assert(ut->uu_kqueue_qos_index == qos_index); - kqwq_req_unlock(kqwq); - return 1; - } - if ((kqr->kqr_state & (KQR_BOUND | KQWQ_THMANAGER)) == 0) { - /* ours to bind to */ - kqworkq_bind_thread_impl(kqwq, qos_index, thread, flags); - kqwq_req_unlock(kqwq); - return 1; - } + /* + * selprocess() is expecting that we send it back the waitq + * that was just added to the thread's waitq set. In order + * to not change the selrecord() API (which is exported to + * kexts), we pass this value back through the + * void *wq_link_id pointer we were passed. We need to use + * memcpy here because the pointer may not be properly aligned + * on 32-bit systems. + */ + void *wqptr = &kq->kqf_wqs; + memcpy(wq_link_id, (void *)&wqptr, sizeof(void *)); } - kqwq_req_unlock(kqwq); - return 0; -} -static void -kqworkloop_bind_thread_impl(struct kqworkloop *kqwl, - thread_t thread, - __assert_only unsigned int flags) -{ - assert(flags & KEVENT_FLAG_WORKLOOP); - - /* the request object must be locked */ - kqwl_req_held(kqwl); - - struct kqrequest *kqr = &kqwl->kqwl_request; - struct uthread *ut = get_bsdthread_info(thread); - boolean_t ipc_override_is_sync; - kq_index_t qos_index = kqworkloop_combined_qos(kqwl, &ipc_override_is_sync); - - /* nobody else bound so finally bind (as a workloop) */ - assert(kqr->kqr_state & KQR_THREQUESTED); - assert((kqr->kqr_state & (KQR_BOUND | KQR_PROCESSING)) == 0); - assert(thread != kqwl->kqwl_owner); - - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_BIND), - kqwl->kqwl_dynamicid, (uintptr_t)thread_tid(thread), - qos_index, - (uintptr_t)(((uintptr_t)kqr->kqr_override_index << 16) | - (((uintptr_t)kqr->kqr_state) << 8) | - ((uintptr_t)ipc_override_is_sync))); - - kqr->kqr_state |= KQR_BOUND | KQR_R2K_NOTIF_ARMED; - kqr->kqr_thread = thread; - - /* bind the workloop to the uthread */ - ut->uu_kqueue_bound = (struct kqueue *)kqwl; - ut->uu_kqueue_flags = flags; - ut->uu_kqueue_qos_index = qos_index; - assert(ut->uu_kqueue_override_is_sync == 0); - ut->uu_kqueue_override_is_sync = ipc_override_is_sync; - if (qos_index) { - thread_add_ipc_override(thread, qos_index); - } - if (ipc_override_is_sync) { - thread_add_sync_ipc_override(thread); + if (kqfile_begin_processing(kq) == -1) { + kqunlock(kq); + return 0; } -} -/* - * workloop_fulfill_threadreq - bind thread to processing workloop - * - * The provided thread will be responsible for delivering events - * associated with the given kqrequest. Bind it and get ready for - * the thread to eventually arrive. - * - * If WORKLOOP_FULFILL_THREADREQ_SYNC is specified, the callback - * within the context of the pthread_functions->workq_threadreq - * callout. In this case, the request structure is already locked. - */ -int -workloop_fulfill_threadreq(struct proc *p, - workq_threadreq_t req, - thread_t thread, - int flags) -{ - int sync = (flags & WORKLOOP_FULFILL_THREADREQ_SYNC); - int cancel = (flags & WORKLOOP_FULFILL_THREADREQ_CANCEL); - struct kqrequest *kqr; - struct kqworkloop *kqwl; + if (!TAILQ_EMPTY(queue)) { + /* + * there is something queued - but it might be a + * KN_STAYACTIVE knote, which may or may not have + * any events pending. Otherwise, we have to walk + * the list of knotes to see, and peek at the + * (non-vanished) stay-active ones to be really sure. + */ + while ((kn = (struct knote *)TAILQ_FIRST(queue)) != NULL) { + if (kn->kn_status & KN_ACTIVE) { + retnum = 1; + goto out; + } + assert(kn->kn_status & KN_STAYACTIVE); + knote_suppress(kq, kn); + } - kqwl = (struct kqworkloop *)((uintptr_t)req - - offsetof(struct kqworkloop, kqwl_request) - - offsetof(struct kqrequest, kqr_req)); - kqr = &kqwl->kqwl_request; + /* + * There were no regular events on the queue, so take + * a deeper look at the stay-queued ones we suppressed. + */ + while ((kn = (struct knote *)TAILQ_FIRST(suppressq)) != NULL) { + KNOTE_LOCK_CTX(knlc); + int result = 0; - /* validate we're looking at something valid */ - if (kqwl->kqwl_p != p || - (kqwl->kqwl_state & KQ_WORKLOOP) == 0) { - assert(kqwl->kqwl_p == p); - assert(kqwl->kqwl_state & KQ_WORKLOOP); - return EINVAL; - } - - if (!sync) - kqwl_req_lock(kqwl); + /* If didn't vanish while suppressed - peek at it */ + if ((kn->kn_status & KN_DROPPING) || !knote_lock(kq, kn, &knlc, + KNOTE_KQ_LOCK_ON_FAILURE)) { + continue; + } - /* Should be a pending request */ - if ((kqr->kqr_state & KQR_BOUND) || - (kqr->kqr_state & KQR_THREQUESTED) == 0) { + result = filter_call(knote_fops(kn), f_peek(kn)); - assert((kqr->kqr_state & KQR_BOUND) == 0); - assert(kqr->kqr_state & KQR_THREQUESTED); - if (!sync) - kqwl_req_unlock(kqwl); - return EINPROGRESS; - } + kqlock(kq); + knote_unlock(kq, kn, &knlc, KNOTE_KQ_LOCK_ALWAYS); - assert((kqr->kqr_state & KQR_DRAIN) == 0); + /* unsuppress it */ + knote_unsuppress(kq, kn); - /* - * Is it a cancel indication from pthread. - * If so, we must be exiting/exec'ing. Forget - * our pending request. - */ - if (cancel) { - kqr->kqr_state &= ~KQR_THREQUESTED; - kqr->kqr_state |= KQR_DRAIN; - } else { - /* do the actual bind? */ - kqworkloop_bind_thread_impl(kqwl, thread, KEVENT_FLAG_WORKLOOP); + /* has data or it has to report a vanish */ + if (result & FILTER_ACTIVE) { + retnum = 1; + goto out; + } + } } - if (!sync) - kqwl_req_unlock(kqwl); +out: + kqfile_end_processing(kq); + kqunlock(kq); + return retnum; +} - if (cancel) - kqueue_release_last(p, &kqwl->kqwl_kqueue); /* may dealloc kq */ +/* + * kqueue_close - + */ +/*ARGSUSED*/ +static int +kqueue_close(struct fileglob *fg, __unused vfs_context_t ctx) +{ + struct kqfile *kqf = (struct kqfile *)fg->fg_data; + assert((kqf->kqf_state & KQ_WORKQ) == 0); + kqueue_dealloc(&kqf->kqf_kqueue); + fg->fg_data = NULL; return 0; } - /* - * kevent_qos_internal_bind - bind thread to processing kqueue - * - * Indicates that the provided thread will be responsible for - * servicing the particular QoS class index specified in the - * parameters. Once the binding is done, any overrides that may - * be associated with the cooresponding events can be applied. - * - * This should be called as soon as the thread identity is known, - * preferably while still at high priority during creation. - * - * - caller holds a reference on the kqueue. - * - the thread MUST call kevent_qos_internal after being bound - * or the bucket of events may never be delivered. - * - Nothing locked (may take mutex or block). + * Max depth of the nested kq path that can be created. + * Note that this has to be less than the size of kq_level + * to avoid wrapping around and mislabeling the level. */ +#define MAX_NESTED_KQ 1000 -int -kevent_qos_internal_bind( - struct proc *p, - int qos_class, - thread_t thread, - unsigned int flags) +/*ARGSUSED*/ +/* + * The callers has taken a use-count reference on this kqueue and will donate it + * to the kqueue we are being added to. This keeps the kqueue from closing until + * that relationship is torn down. + */ +static int +kqueue_kqfilter(struct fileproc *fp, struct knote *kn, + __unused struct kevent_qos_s *kev) { - kq_index_t qos_index; - - assert(flags & KEVENT_FLAG_WORKQ); - - if (thread == THREAD_NULL || (flags & KEVENT_FLAG_WORKQ) == 0) { - return EINVAL; - } + struct kqfile *kqf = (struct kqfile *)fp->f_data; + struct kqueue *kq = &kqf->kqf_kqueue; + struct kqueue *parentkq = knote_get_kq(kn); - /* get the qos index we're going to service */ - qos_index = qos_index_for_servicer(qos_class, thread, flags); + assert((kqf->kqf_state & KQ_WORKQ) == 0); - if (kqworkq_internal_bind(p, qos_index, thread, flags)) + if (parentkq == kq || kn->kn_filter != EVFILT_READ) { + knote_set_error(kn, EINVAL); return 0; + } - return EINPROGRESS; -} - - -static void -kqworkloop_internal_unbind( - struct proc *p, - thread_t thread, - unsigned int flags) -{ - struct kqueue *kq; - struct kqworkloop *kqwl; - struct uthread *ut = get_bsdthread_info(thread); - - assert(ut->uu_kqueue_bound != NULL); - kq = ut->uu_kqueue_bound; - assert(kq->kq_state & KQ_WORKLOOP); - kqwl = (struct kqworkloop *)kq; + /* + * We have to avoid creating a cycle when nesting kqueues + * inside another. Rather than trying to walk the whole + * potential DAG of nested kqueues, we just use a simple + * ceiling protocol. When a kqueue is inserted into another, + * we check that the (future) parent is not already nested + * into another kqueue at a lower level than the potenial + * child (because it could indicate a cycle). If that test + * passes, we just mark the nesting levels accordingly. + * + * Only up to MAX_NESTED_KQ can be nested. + * + * Note: kqworkq and kqworkloop cannot be nested and have reused their + * kq_level field, so ignore these as parent. + */ - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_UNBIND), - kqwl->kqwl_dynamicid, (uintptr_t)thread_tid(thread), - flags, 0); + kqlock(parentkq); - if (!(kq->kq_state & KQ_NO_WQ_THREAD)) { - assert(is_workqueue_thread(thread)); + if ((parentkq->kq_state & (KQ_WORKQ | KQ_WORKLOOP)) == 0) { + if (parentkq->kq_level > 0 && + parentkq->kq_level < kq->kq_level) { + kqunlock(parentkq); + knote_set_error(kn, EINVAL); + return 0; + } - kqlock(kq); - kqworkloop_unbind_thread(kqwl, thread, flags); - kqunlock(kq); + /* set parent level appropriately */ + uint16_t plevel = (parentkq->kq_level == 0)? 2: parentkq->kq_level; + if (plevel < kq->kq_level + 1) { + if (kq->kq_level + 1 > MAX_NESTED_KQ) { + kqunlock(parentkq); + knote_set_error(kn, EINVAL); + return 0; + } + plevel = kq->kq_level + 1; + } - /* If last reference, dealloc the workloop kq */ - kqueue_release_last(p, kq); - } else { - assert(!is_workqueue_thread(thread)); - kevent_servicer_detach_thread(p, kqwl->kqwl_dynamicid, thread, flags, kq); + parentkq->kq_level = plevel; } -} -static void -kqworkq_internal_unbind( - struct proc *p, - kq_index_t qos_index, - thread_t thread, - unsigned int flags) -{ - struct kqueue *kq; - struct kqworkq *kqwq; - struct uthread *ut; - kq_index_t end_index; - - assert(thread == current_thread()); - ut = get_bsdthread_info(thread); - - kq = p->p_fd->fd_wqkqueue; - assert(kq->kq_state & KQ_WORKQ); - assert(ut->uu_kqueue_bound == kq); + kqunlock(parentkq); - kqwq = (struct kqworkq *)kq; - - /* end servicing any requests we might own */ - end_index = (qos_index == KQWQ_QOS_MANAGER) ? - 0 : qos_index; + kn->kn_filtid = EVFILTID_KQREAD; kqlock(kq); + KNOTE_ATTACH(&kqf->kqf_sel.si_note, kn); + /* indicate nesting in child, if needed */ + if (kq->kq_level == 0) { + kq->kq_level = 1; + } - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWQ_UNBIND), - (uintptr_t)thread_tid(thread), flags, qos_index); - - do { - kqworkq_end_processing(kqwq, qos_index, flags); - } while (qos_index-- > end_index); - - ut->uu_kqueue_bound = NULL; - ut->uu_kqueue_qos_index = 0; - ut->uu_kqueue_flags = 0; - + int count = kq->kq_count; kqunlock(kq); + return count > 0; } /* - * kevent_qos_internal_unbind - unbind thread from processing kqueue - * - * End processing the per-QoS bucket of events and allow other threads - * to be requested for future servicing. - * - * caller holds a reference on the kqueue. - * thread is the current thread. + * kqueue_drain - called when kq is closed */ - -int -kevent_qos_internal_unbind( - struct proc *p, - int qos_class, - thread_t thread, - unsigned int flags) +/*ARGSUSED*/ +static int +kqueue_drain(struct fileproc *fp, __unused vfs_context_t ctx) { -#pragma unused(qos_class) + struct kqfile *kqf = (struct kqfile *)fp->fp_glob->fg_data; - struct uthread *ut; - struct kqueue *kq; - unsigned int bound_flags; - bool check_flags; + assert((kqf->kqf_state & KQ_WORKQ) == 0); - ut = get_bsdthread_info(thread); - if (ut->uu_kqueue_bound == NULL) { - /* early out if we are already unbound */ - assert(ut->uu_kqueue_flags == 0); - assert(ut->uu_kqueue_qos_index == 0); - assert(ut->uu_kqueue_override_is_sync == 0); - return EALREADY; + kqlock(kqf); + kqf->kqf_state |= KQ_DRAIN; + + /* wakeup sleeping threads */ + if ((kqf->kqf_state & (KQ_SLEEP | KQ_SEL)) != 0) { + kqf->kqf_state &= ~(KQ_SLEEP | KQ_SEL); + (void)waitq_wakeup64_all((struct waitq *)&kqf->kqf_wqs, + KQ_EVENT, + THREAD_RESTART, + WAITQ_ALL_PRIORITIES); } - assert(flags & (KEVENT_FLAG_WORKQ | KEVENT_FLAG_WORKLOOP)); - assert(thread == current_thread()); + /* wakeup threads waiting their turn to process */ + if (kqf->kqf_state & KQ_PROCWAIT) { + assert(kqf->kqf_state & KQ_PROCESSING); - check_flags = flags & KEVENT_FLAG_UNBIND_CHECK_FLAGS; + kqf->kqf_state &= ~KQ_PROCWAIT; + (void)waitq_wakeup64_all((struct waitq *)&kqf->kqf_wqs, + CAST_EVENT64_T(&kqf->kqf_suppressed), + THREAD_RESTART, WAITQ_ALL_PRIORITIES); + } - /* Get the kqueue we started with */ - kq = ut->uu_kqueue_bound; - assert(kq != NULL); - assert(kq->kq_state & (KQ_WORKQ | KQ_WORKLOOP)); + kqunlock(kqf); + return 0; +} - /* get flags and QoS parameters we started with */ - bound_flags = ut->uu_kqueue_flags; +/*ARGSUSED*/ +int +kqueue_stat(struct kqueue *kq, void *ub, int isstat64, proc_t p) +{ + assert((kq->kq_state & KQ_WORKQ) == 0); - /* Unbind from the class of workq */ - if (kq->kq_state & KQ_WORKQ) { - if (check_flags && !(flags & KEVENT_FLAG_WORKQ)) { - return EINVAL; - } + kqlock(kq); + if (isstat64 != 0) { + struct stat64 *sb64 = (struct stat64 *)ub; - kqworkq_internal_unbind(p, ut->uu_kqueue_qos_index, thread, bound_flags); - } else { - if (check_flags && !(flags & KEVENT_FLAG_WORKLOOP)) { - return EINVAL; + bzero((void *)sb64, sizeof(*sb64)); + sb64->st_size = kq->kq_count; + if (kq->kq_state & KQ_KEV_QOS) { + sb64->st_blksize = sizeof(struct kevent_qos_s); + } else if (kq->kq_state & KQ_KEV64) { + sb64->st_blksize = sizeof(struct kevent64_s); + } else if (IS_64BIT_PROCESS(p)) { + sb64->st_blksize = sizeof(struct user64_kevent); + } else { + sb64->st_blksize = sizeof(struct user32_kevent); } + sb64->st_mode = S_IFIFO; + } else { + struct stat *sb = (struct stat *)ub; - kqworkloop_internal_unbind(p, thread, bound_flags); + bzero((void *)sb, sizeof(*sb)); + sb->st_size = kq->kq_count; + if (kq->kq_state & KQ_KEV_QOS) { + sb->st_blksize = sizeof(struct kevent_qos_s); + } else if (kq->kq_state & KQ_KEV64) { + sb->st_blksize = sizeof(struct kevent64_s); + } else if (IS_64BIT_PROCESS(p)) { + sb->st_blksize = sizeof(struct user64_kevent); + } else { + sb->st_blksize = sizeof(struct user32_kevent); + } + sb->st_mode = S_IFIFO; } - + kqunlock(kq); return 0; } +static inline bool +kqueue_threadreq_can_use_ast(struct kqueue *kq) +{ + if (current_proc() == kq->kq_p) { + /* + * Setting an AST from a non BSD syscall is unsafe: mach_msg_trap() can + * do combined send/receive and in the case of self-IPC, the AST may bet + * set on a thread that will not return to userspace and needs the + * thread the AST would create to unblock itself. + * + * At this time, we really want to target: + * + * - kevent variants that can cause thread creations, and dispatch + * really only uses kevent_qos and kevent_id, + * + * - workq_kernreturn (directly about thread creations) + * + * - bsdthread_ctl which is used for qos changes and has direct impact + * on the creator thread scheduling decisions. + */ + switch (current_uthread()->syscall_code) { + case SYS_kevent_qos: + case SYS_kevent_id: + case SYS_workq_kernreturn: + case SYS_bsdthread_ctl: + return true; + } + } + return false; +} + /* - * kqueue_process - process the triggered events in a kqueue + * Interact with the pthread kext to request a servicing there at a specific QoS + * level. * - * Walk the queued knotes and validate that they are - * really still triggered events by calling the filter - * routines (if necessary). Hold a use reference on - * the knote to avoid it being detached. For each event - * that is still considered triggered, invoke the - * callback routine provided. + * - Caller holds the workq request lock * - * caller holds a reference on the kqueue. - * kqueue locked on entry and exit - but may be dropped - * kqueue list locked (held for duration of call) + * - May be called with the kqueue's wait queue set locked, + * so cannot do anything that could recurse on that. */ - -static int -kqueue_process(struct kqueue *kq, - kevent_callback_t callback, - void *callback_data, - struct filt_process_s *process_data, - int *countp, - struct proc *p) +static void +kqueue_threadreq_initiate(struct kqueue *kq, workq_threadreq_t kqr, + kq_index_t qos, int flags) { - unsigned int flags = process_data ? process_data->fp_flags : 0; - struct uthread *ut = get_bsdthread_info(current_thread()); - kq_index_t start_index, end_index, i; - struct knote *kn; - int nevents = 0; - int error = 0; + assert(kqr->tr_kq_wakeup); + assert(kqr_thread(kqr) == THREAD_NULL); + assert(!kqr_thread_requested(kqr)); + struct turnstile *ts = TURNSTILE_NULL; - /* - * Based on the mode of the kqueue and the bound QoS of the servicer, - * determine the range of thread requests that need checking - */ - if (kq->kq_state & KQ_WORKQ) { - if (flags & KEVENT_FLAG_WORKQ_MANAGER) { - start_index = KQWQ_QOS_MANAGER; - } else if (ut->uu_kqueue_bound != kq) { - return EJUSTRETURN; - } else { - start_index = ut->uu_kqueue_qos_index; - } + if (workq_is_exiting(kq->kq_p)) { + return; + } - /* manager services every request in a workq kqueue */ - assert(start_index > 0 && start_index <= KQWQ_QOS_MANAGER); - end_index = (start_index == KQWQ_QOS_MANAGER) ? 0 : start_index; + kqlock_held(kq); - } else if (kq->kq_state & KQ_WORKLOOP) { - if (ut->uu_kqueue_bound != kq) - return EJUSTRETURN; + if (kq->kq_state & KQ_WORKLOOP) { + __assert_only struct kqworkloop *kqwl = (struct kqworkloop *)kq; - /* - * Single request servicing - * we want to deliver all events, regardless of the QOS - */ - start_index = end_index = THREAD_QOS_UNSPECIFIED; + assert(kqwl->kqwl_owner == THREAD_NULL); + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWL_THREQUEST), + kqwl->kqwl_dynamicid, 0, qos, kqr->tr_kq_wakeup); + ts = kqwl->kqwl_turnstile; + /* Add a thread request reference on the kqueue. */ + kqworkloop_retain(kqwl); } else { - start_index = end_index = QOS_INDEX_KQFILE; + assert(kq->kq_state & KQ_WORKQ); + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWQ_THREQUEST), + -1, 0, qos, kqr->tr_kq_wakeup); } - - i = start_index; - - do { - if (kqueue_begin_processing(kq, i, flags) == -1) { - *countp = 0; - /* Nothing to process */ - continue; - } + /* + * New-style thread request supported. + * Provide the pthread kext a pointer to a workq_threadreq_s structure for + * its use until a corresponding kqueue_threadreq_bind callback. + */ + if (kqueue_threadreq_can_use_ast(kq)) { + flags |= WORKQ_THREADREQ_SET_AST_ON_FAILURE; + } + if (qos == KQWQ_QOS_MANAGER) { + qos = WORKQ_THREAD_QOS_MANAGER; + } + if (!workq_kern_threadreq_initiate(kq->kq_p, kqr, ts, qos, flags)) { /* - * loop through the enqueued knotes associated with this request, - * processing each one. Each request may have several queues - * of knotes to process (depending on the type of kqueue) so we - * have to loop through all the queues as long as we have additional - * space. + * Process is shutting down or exec'ing. + * All the kqueues are going to be cleaned up + * soon. Forget we even asked for a thread - + * and make sure we don't ask for more. */ - error = 0; - - struct kqtailq *base_queue = kqueue_get_base_queue(kq, i); - struct kqtailq *queue = kqueue_get_high_queue(kq, i); - do { - while (error == 0 && (kn = TAILQ_FIRST(queue)) != NULL) { - error = knote_process(kn, callback, callback_data, process_data, p); - if (error == EJUSTRETURN) { - error = 0; - } else { - nevents++; - } - /* error is EWOULDBLOCK when the out event array is full */ - } - } while (error == 0 && queue-- > base_queue); + kq->kq_state &= ~KQ_R2K_ARMED; + kqueue_release_live(kq); + } +} - if ((kq->kq_state & KQ_WORKQ) == 0) { - kqueue_end_processing(kq, i, nevents, flags); - } +/* + * kqueue_threadreq_bind_prepost - prepost the bind to kevent + * + * This is used when kqueue_threadreq_bind may cause a lock inversion. + */ +__attribute__((always_inline)) +void +kqueue_threadreq_bind_prepost(struct proc *p __unused, workq_threadreq_t kqr, + struct uthread *ut) +{ + ut->uu_kqr_bound = kqr; + kqr->tr_thread = ut->uu_thread; + kqr->tr_state = WORKQ_TR_STATE_BINDING; +} - if (error == EWOULDBLOCK) { - /* break out if no more space for additional events */ - error = 0; - break; - } - } while (i-- > end_index); +/* + * kqueue_threadreq_bind_commit - commit a bind prepost + * + * The workq code has to commit any binding prepost before the thread has + * a chance to come back to userspace (and do kevent syscalls) or be aborted. + */ +void +kqueue_threadreq_bind_commit(struct proc *p, thread_t thread) +{ + struct uthread *ut = get_bsdthread_info(thread); + workq_threadreq_t kqr = ut->uu_kqr_bound; + kqueue_t kqu = kqr_kqueue(p, kqr); - *countp = nevents; - return (error); + kqlock(kqu); + if (kqr->tr_state == WORKQ_TR_STATE_BINDING) { + kqueue_threadreq_bind(p, kqr, thread, 0); + } + kqunlock(kqu); } static void -kqueue_scan_continue(void *data, wait_result_t wait_result) +kqueue_threadreq_modify(kqueue_t kqu, workq_threadreq_t kqr, kq_index_t qos, + workq_kern_threadreq_flags_t flags) { - thread_t self = current_thread(); - uthread_t ut = (uthread_t)get_bsdthread_info(self); - struct _kqueue_scan * cont_args = &ut->uu_kevent.ss_kqueue_scan; - struct kqueue *kq = (struct kqueue *)data; - struct filt_process_s *process_data = cont_args->process_data; - int error; - int count; - - /* convert the (previous) wait_result to a proper error */ - switch (wait_result) { - case THREAD_AWAKENED: { - kqlock(kq); - retry: - error = kqueue_process(kq, cont_args->call, cont_args->data, - process_data, &count, current_proc()); - if (error == 0 && count == 0) { - if (kq->kq_state & KQ_DRAIN) { - kqunlock(kq); - goto drain; - } + assert(kqr_thread_requested_pending(kqr)); - if (kq->kq_state & KQ_WAKEUP) - goto retry; + kqlock_held(kqu); - waitq_assert_wait64((struct waitq *)&kq->kq_wqs, - KQ_EVENT, THREAD_ABORTSAFE, - cont_args->deadline); - kq->kq_state |= KQ_SLEEP; - kqunlock(kq); - thread_block_parameter(kqueue_scan_continue, kq); - /* NOTREACHED */ - } - kqunlock(kq); - } break; - case THREAD_TIMED_OUT: - error = EWOULDBLOCK; - break; - case THREAD_INTERRUPTED: - error = EINTR; - break; - case THREAD_RESTART: - drain: - error = EBADF; - break; - default: - panic("%s: - invalid wait_result (%d)", __func__, - wait_result); - error = 0; + if (kqueue_threadreq_can_use_ast(kqu.kq)) { + flags |= WORKQ_THREADREQ_SET_AST_ON_FAILURE; } - - /* call the continuation with the results */ - assert(cont_args->cont != NULL); - (cont_args->cont)(kq, cont_args->data, error); + workq_kern_threadreq_modify(kqu.kq->kq_p, kqr, qos, flags); } - /* - * kqueue_scan - scan and wait for events in a kqueue - * - * Process the triggered events in a kqueue. + * kqueue_threadreq_bind - bind thread to processing kqrequest * - * If there are no events triggered arrange to - * wait for them. If the caller provided a - * continuation routine, then kevent_scan will - * also. - * - * The callback routine must be valid. - * The caller must hold a use-count reference on the kq. + * The provided thread will be responsible for delivering events + * associated with the given kqrequest. Bind it and get ready for + * the thread to eventually arrive. */ +void +kqueue_threadreq_bind(struct proc *p, workq_threadreq_t kqr, thread_t thread, + unsigned int flags) +{ + kqueue_t kqu = kqr_kqueue(p, kqr); + struct uthread *ut = get_bsdthread_info(thread); -int -kqueue_scan(struct kqueue *kq, - kevent_callback_t callback, - kqueue_continue_t continuation, - void *callback_data, - struct filt_process_s *process_data, - struct timeval *atvp, - struct proc *p) -{ - thread_continue_t cont = THREAD_CONTINUE_NULL; - unsigned int flags; - uint64_t deadline; - int error; - int first; - int fd; - - assert(callback != NULL); + kqlock_held(kqu); - /* - * Determine which QoS index we are servicing - */ - flags = (process_data) ? process_data->fp_flags : 0; - fd = (process_data) ? process_data->fp_fd : -1; + assert(ut->uu_kqueue_override == 0); - first = 1; - for (;;) { - wait_result_t wait_result; - int count; + if (kqr->tr_state == WORKQ_TR_STATE_BINDING) { + assert(ut->uu_kqr_bound == kqr); + assert(kqr->tr_thread == thread); + } else { + assert(kqr_thread_requested_pending(kqr)); + assert(kqr->tr_thread == THREAD_NULL); + assert(ut->uu_kqr_bound == NULL); + ut->uu_kqr_bound = kqr; + kqr->tr_thread = thread; + } - /* - * Make a pass through the kq to find events already - * triggered. - */ - kqlock(kq); - error = kqueue_process(kq, callback, callback_data, - process_data, &count, p); - if (error || count) - break; /* lock still held */ - - /* looks like we have to consider blocking */ - if (first) { - first = 0; - /* convert the timeout to a deadline once */ - if (atvp->tv_sec || atvp->tv_usec) { - uint64_t now; - - clock_get_uptime(&now); - nanoseconds_to_absolutetime((uint64_t)atvp->tv_sec * NSEC_PER_SEC + - atvp->tv_usec * (long)NSEC_PER_USEC, - &deadline); - if (now >= deadline) { - /* non-blocking call */ - error = EWOULDBLOCK; - break; /* lock still held */ - } - deadline -= now; - clock_absolutetime_interval_to_deadline(deadline, &deadline); - } else { - deadline = 0; /* block forever */ - } + kqr->tr_state = WORKQ_TR_STATE_BOUND; - if (continuation) { - uthread_t ut = (uthread_t)get_bsdthread_info(current_thread()); - struct _kqueue_scan *cont_args = &ut->uu_kevent.ss_kqueue_scan; + if (kqu.kq->kq_state & KQ_WORKLOOP) { + struct turnstile *ts = kqu.kqwl->kqwl_turnstile; - cont_args->call = callback; - cont_args->cont = continuation; - cont_args->deadline = deadline; - cont_args->data = callback_data; - cont_args->process_data = process_data; - cont = kqueue_scan_continue; + if (__improbable(thread == kqu.kqwl->kqwl_owner)) { + /* + * shows that asserting here is not ok. + * + * This is not supposed to happen for correct use of the interface, + * but it is sadly possible for userspace (with the help of memory + * corruption, such as over-release of a dispatch queue) to make + * the creator thread the "owner" of a workloop. + * + * Once that happens, and that creator thread picks up the same + * workloop as a servicer, we trip this codepath. We need to fixup + * the state to forget about this thread being the owner, as the + * entire workloop state machine expects servicers to never be + * owners and everything would basically go downhill from here. + */ + kqu.kqwl->kqwl_owner = THREAD_NULL; + if (kqworkloop_override(kqu.kqwl)) { + thread_drop_kevent_override(thread); } } - if (kq->kq_state & KQ_DRAIN) { - kqunlock(kq); - return EBADF; - } - - /* If awakened during processing, try again */ - if (kq->kq_state & KQ_WAKEUP) { - kqunlock(kq); - continue; + if (ts && (flags & KQUEUE_THREADERQ_BIND_NO_INHERITOR_UPDATE) == 0) { + /* + * Past this point, the interlock is the kq req lock again, + * so we can fix the inheritor for good. + */ + filt_wlupdate_inheritor(kqu.kqwl, ts, TURNSTILE_IMMEDIATE_UPDATE); + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_HELD); } - /* go ahead and wait */ - waitq_assert_wait64_leeway((struct waitq *)&kq->kq_wqs, - KQ_EVENT, THREAD_ABORTSAFE, - TIMEOUT_URGENCY_USER_NORMAL, - deadline, TIMEOUT_NO_LEEWAY); - kq->kq_state |= KQ_SLEEP; - kqunlock(kq); - wait_result = thread_block_parameter(cont, kq); - /* NOTREACHED if (continuation != NULL) */ + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWL_BIND), kqu.kqwl->kqwl_dynamicid, + thread_tid(thread), kqr->tr_kq_qos_index, + (kqr->tr_kq_override_index << 16) | kqr->tr_kq_wakeup); - switch (wait_result) { - case THREAD_AWAKENED: - continue; - case THREAD_TIMED_OUT: - return EWOULDBLOCK; - case THREAD_INTERRUPTED: - return EINTR; - case THREAD_RESTART: - return EBADF; - default: - panic("%s: - bad wait_result (%d)", __func__, - wait_result); - error = 0; + ut->uu_kqueue_override = kqr->tr_kq_override_index; + if (kqr->tr_kq_override_index) { + thread_add_servicer_override(thread, kqr->tr_kq_override_index); } + } else { + assert(kqr->tr_kq_override_index == 0); + + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWQ_BIND), -1, + thread_tid(thread), kqr->tr_kq_qos_index, + (kqr->tr_kq_override_index << 16) | kqr->tr_kq_wakeup); } - kqunlock(kq); - return (error); } - /* - * XXX - * This could be expanded to call kqueue_scan, if desired. + * kqueue_threadreq_cancel - abort a pending thread request + * + * Called when exiting/exec'ing. Forget our pending request. */ -/*ARGSUSED*/ -static int -kqueue_read(__unused struct fileproc *fp, - __unused struct uio *uio, - __unused int flags, - __unused vfs_context_t ctx) +void +kqueue_threadreq_cancel(struct proc *p, workq_threadreq_t kqr) { - return (ENXIO); + kqueue_release(kqr_kqueue(p, kqr)); } -/*ARGSUSED*/ -static int -kqueue_write(__unused struct fileproc *fp, - __unused struct uio *uio, - __unused int flags, - __unused vfs_context_t ctx) +workq_threadreq_param_t +kqueue_threadreq_workloop_param(workq_threadreq_t kqr) { - return (ENXIO); + struct kqworkloop *kqwl; + workq_threadreq_param_t trp; + + assert(kqr->tr_flags & WORKQ_TR_FLAG_WORKLOOP); + kqwl = __container_of(kqr, struct kqworkloop, kqwl_request); + trp.trp_value = kqwl->kqwl_params; + return trp; } -/*ARGSUSED*/ -static int -kqueue_ioctl(__unused struct fileproc *fp, - __unused u_long com, - __unused caddr_t data, - __unused vfs_context_t ctx) +/* + * kqueue_threadreq_unbind - unbind thread from processing kqueue + * + * End processing the per-QoS bucket of events and allow other threads + * to be requested for future servicing. + * + * caller holds a reference on the kqueue. + */ +void +kqueue_threadreq_unbind(struct proc *p, workq_threadreq_t kqr) { - return (ENOTTY); + if (kqr->tr_flags & WORKQ_TR_FLAG_WORKLOOP) { + kqworkloop_unbind(kqr_kqworkloop(kqr)); + } else { + kqworkq_unbind(p, kqr); + } } -/*ARGSUSED*/ -static int -kqueue_select(struct fileproc *fp, int which, void *wq_link_id, - __unused vfs_context_t ctx) +/* + * If we aren't already busy processing events [for this QoS], + * request workq thread support as appropriate. + * + * TBD - for now, we don't segregate out processing by QoS. + * + * - May be called with the kqueue's wait queue set locked, + * so cannot do anything that could recurse on that. + */ +static void +kqworkq_wakeup(struct kqworkq *kqwq, kq_index_t qos_index) { - struct kqueue *kq = (struct kqueue *)fp->f_data; - struct kqtailq *queue; - struct kqtailq *suppressq; - struct knote *kn; - int retnum = 0; - - if (which != FREAD) - return (0); + workq_threadreq_t kqr = kqworkq_get_request(kqwq, qos_index); - kqlock(kq); - - assert((kq->kq_state & KQ_WORKQ) == 0); + /* convert to thread qos value */ + assert(qos_index < KQWQ_NBUCKETS); - /* - * If this is the first pass, link the wait queue associated with the - * the kqueue onto the wait queue set for the select(). Normally we - * use selrecord() for this, but it uses the wait queue within the - * selinfo structure and we need to use the main one for the kqueue to - * catch events from KN_STAYQUEUED sources. So we do the linkage manually. - * (The select() call will unlink them when it ends). - */ - if (wq_link_id != NULL) { - thread_t cur_act = current_thread(); - struct uthread * ut = get_bsdthread_info(cur_act); + if (!kqr->tr_kq_wakeup) { + kqr->tr_kq_wakeup = true; + if (!kqr_thread_requested(kqr)) { + kqueue_threadreq_initiate(&kqwq->kqwq_kqueue, kqr, qos_index, 0); + } + } +} - kq->kq_state |= KQ_SEL; - waitq_link((struct waitq *)&kq->kq_wqs, ut->uu_wqset, - WAITQ_SHOULD_LOCK, (uint64_t *)wq_link_id); +/* + * This represent the asynchronous QoS a given workloop contributes, + * hence is the max of the current active knotes (override index) + * and the workloop max qos (userspace async qos). + */ +static kq_index_t +kqworkloop_override(struct kqworkloop *kqwl) +{ + workq_threadreq_t kqr = &kqwl->kqwl_request; + return MAX(kqr->tr_kq_qos_index, kqr->tr_kq_override_index); +} - /* always consume the reserved link object */ - waitq_link_release(*(uint64_t *)wq_link_id); - *(uint64_t *)wq_link_id = 0; +static inline void +kqworkloop_request_fire_r2k_notification(struct kqworkloop *kqwl) +{ + workq_threadreq_t kqr = &kqwl->kqwl_request; - /* - * selprocess() is expecting that we send it back the waitq - * that was just added to the thread's waitq set. In order - * to not change the selrecord() API (which is exported to - * kexts), we pass this value back through the - * void *wq_link_id pointer we were passed. We need to use - * memcpy here because the pointer may not be properly aligned - * on 32-bit systems. - */ - void *wqptr = &kq->kq_wqs; - memcpy(wq_link_id, (void *)&wqptr, sizeof(void *)); - } + kqlock_held(kqwl); - if (kqueue_begin_processing(kq, QOS_INDEX_KQFILE, 0) == -1) { - kqunlock(kq); - return (0); + if (kqwl->kqwl_state & KQ_R2K_ARMED) { + kqwl->kqwl_state &= ~KQ_R2K_ARMED; + act_set_astkevent(kqr_thread_fast(kqr), AST_KEVENT_RETURN_TO_KERNEL); } +} - queue = kqueue_get_base_queue(kq, QOS_INDEX_KQFILE); - if (!TAILQ_EMPTY(queue)) { - /* - * there is something queued - but it might be a - * KN_STAYACTIVE knote, which may or may not have - * any events pending. Otherwise, we have to walk - * the list of knotes to see, and peek at the - * (non-vanished) stay-active ones to be really sure. - */ - while ((kn = (struct knote *)TAILQ_FIRST(queue)) != NULL) { - if (kn->kn_status & KN_ACTIVE) { - retnum = 1; - goto out; - } - assert(kn->kn_status & KN_STAYACTIVE); - knote_suppress(kn); - } +static void +kqworkloop_update_threads_qos(struct kqworkloop *kqwl, int op, kq_index_t qos) +{ + workq_threadreq_t kqr = &kqwl->kqwl_request; + struct kqueue *kq = &kqwl->kqwl_kqueue; + kq_index_t old_override = kqworkloop_override(kqwl); + kq_index_t i; - /* - * There were no regular events on the queue, so take - * a deeper look at the stay-queued ones we suppressed. - */ - suppressq = kqueue_get_suppressed_queue(kq, QOS_INDEX_KQFILE); - while ((kn = (struct knote *)TAILQ_FIRST(suppressq)) != NULL) { - unsigned peek = 1; + kqlock_held(kqwl); - assert(!knoteuse_needs_boost(kn, NULL)); + switch (op) { + case KQWL_UTQ_UPDATE_WAKEUP_QOS: + if (qos == KQWL_BUCKET_STAYACTIVE) { + /* + * the KQWL_BUCKET_STAYACTIVE is not a QoS bucket, we only remember + * a high watermark (kqwl_stayactive_qos) of any stay active knote + * that was ever registered with this workloop. + * + * When waitq_set__CALLING_PREPOST_HOOK__() wakes up any stay active + * knote, we use this high-watermark as a wakeup-index, and also set + * the magic KQWL_BUCKET_STAYACTIVE bit to make sure we remember + * there is at least one stay active knote fired until the next full + * processing of this bucket. + */ + kqwl->kqwl_wakeup_indexes |= KQWL_STAYACTIVE_FIRED_BIT; + qos = kqwl->kqwl_stayactive_qos; + assert(qos); + } + if (kqwl->kqwl_wakeup_indexes & (1 << qos)) { + assert(kqr->tr_kq_wakeup); + break; + } - /* If didn't vanish while suppressed - peek at it */ - if (kqlock2knoteuse(kq, kn, KNUSE_NONE)) { - peek = knote_fops(kn)->f_peek(kn); + kqwl->kqwl_wakeup_indexes |= (1 << qos); + kqr->tr_kq_wakeup = true; + kqworkloop_request_fire_r2k_notification(kqwl); + goto recompute; - /* if it dropped while getting lock - move on */ - if (!knoteuse2kqlock(kq, kn, KNUSE_NONE)) - continue; + case KQWL_UTQ_UPDATE_STAYACTIVE_QOS: + assert(qos); + if (kqwl->kqwl_stayactive_qos < qos) { + kqwl->kqwl_stayactive_qos = qos; + if (kqwl->kqwl_wakeup_indexes & KQWL_STAYACTIVE_FIRED_BIT) { + assert(kqr->tr_kq_wakeup); + kqwl->kqwl_wakeup_indexes |= (1 << qos); + goto recompute; } + } + break; - /* unsuppress it */ - knote_unsuppress(kn); - - /* has data or it has to report a vanish */ - if (peek > 0) { - retnum = 1; - goto out; + case KQWL_UTQ_PARKING: + case KQWL_UTQ_UNBINDING: + kqr->tr_kq_override_index = qos; + OS_FALLTHROUGH; + case KQWL_UTQ_RECOMPUTE_WAKEUP_QOS: + if (op == KQWL_UTQ_RECOMPUTE_WAKEUP_QOS) { + assert(qos == THREAD_QOS_UNSPECIFIED); + } + i = KQWL_BUCKET_STAYACTIVE; + if (TAILQ_EMPTY(&kqwl->kqwl_suppressed)) { + kqr->tr_kq_override_index = THREAD_QOS_UNSPECIFIED; + } + if (!TAILQ_EMPTY(&kqwl->kqwl_queue[i]) && + (kqwl->kqwl_wakeup_indexes & KQWL_STAYACTIVE_FIRED_BIT)) { + /* + * If the KQWL_STAYACTIVE_FIRED_BIT is set, it means a stay active + * knote may have fired, so we need to merge in kqwl_stayactive_qos. + * + * Unlike other buckets, this one is never empty but could be idle. + */ + kqwl->kqwl_wakeup_indexes &= KQWL_STAYACTIVE_FIRED_BIT; + kqwl->kqwl_wakeup_indexes |= (1 << kqwl->kqwl_stayactive_qos); + } else { + kqwl->kqwl_wakeup_indexes = 0; + } + for (i = THREAD_QOS_UNSPECIFIED + 1; i < KQWL_BUCKET_STAYACTIVE; i++) { + if (!TAILQ_EMPTY(&kqwl->kqwl_queue[i])) { + kqwl->kqwl_wakeup_indexes |= (1 << i); } } - } - -out: - kqueue_end_processing(kq, QOS_INDEX_KQFILE, retnum, 0); - kqunlock(kq); - return (retnum); -} + if (kqwl->kqwl_wakeup_indexes) { + kqr->tr_kq_wakeup = true; + kqworkloop_request_fire_r2k_notification(kqwl); + } else { + kqr->tr_kq_wakeup = false; + } + goto recompute; -/* - * kqueue_close - - */ -/*ARGSUSED*/ -static int -kqueue_close(struct fileglob *fg, __unused vfs_context_t ctx) -{ - struct kqfile *kqf = (struct kqfile *)fg->fg_data; + case KQWL_UTQ_RESET_WAKEUP_OVERRIDE: + kqr->tr_kq_override_index = qos; + goto recompute; - assert((kqf->kqf_state & KQ_WORKQ) == 0); - kqueue_dealloc(&kqf->kqf_kqueue); - fg->fg_data = NULL; - return (0); -} + case KQWL_UTQ_UPDATE_WAKEUP_OVERRIDE: +recompute: + /* + * When modifying the wakeup QoS or the override QoS, we always need to + * maintain our invariant that kqr_override_index is at least as large + * as the highest QoS for which an event is fired. + * + * However this override index can be larger when there is an overriden + * suppressed knote pushing on the kqueue. + */ + if (kqwl->kqwl_wakeup_indexes > (1 << qos)) { + qos = (uint8_t)(fls(kqwl->kqwl_wakeup_indexes) - 1); /* fls is 1-based */ + } + if (kqr->tr_kq_override_index < qos) { + kqr->tr_kq_override_index = qos; + } + break; -/*ARGSUSED*/ -/* - * The callers has taken a use-count reference on this kqueue and will donate it - * to the kqueue we are being added to. This keeps the kqueue from closing until - * that relationship is torn down. - */ -static int -kqueue_kqfilter(__unused struct fileproc *fp, struct knote *kn, - __unused struct kevent_internal_s *kev, __unused vfs_context_t ctx) -{ - struct kqfile *kqf = (struct kqfile *)kn->kn_fp->f_data; - struct kqueue *kq = &kqf->kqf_kqueue; - struct kqueue *parentkq = knote_get_kq(kn); + case KQWL_UTQ_REDRIVE_EVENTS: + break; - assert((kqf->kqf_state & KQ_WORKQ) == 0); + case KQWL_UTQ_SET_QOS_INDEX: + kqr->tr_kq_qos_index = qos; + break; - if (parentkq == kq || - kn->kn_filter != EVFILT_READ) { - kn->kn_flags = EV_ERROR; - kn->kn_data = EINVAL; - return 0; + default: + panic("unknown kqwl thread qos update operation: %d", op); } + thread_t kqwl_owner = kqwl->kqwl_owner; + thread_t servicer = kqr_thread(kqr); + boolean_t qos_changed = FALSE; + kq_index_t new_override = kqworkloop_override(kqwl); + /* - * We have to avoid creating a cycle when nesting kqueues - * inside another. Rather than trying to walk the whole - * potential DAG of nested kqueues, we just use a simple - * ceiling protocol. When a kqueue is inserted into another, - * we check that the (future) parent is not already nested - * into another kqueue at a lower level than the potenial - * child (because it could indicate a cycle). If that test - * passes, we just mark the nesting levels accordingly. + * Apply the diffs to the owner if applicable */ + if (kqwl_owner) { +#if 0 + /* JMM - need new trace hooks for owner overrides */ + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWL_THADJUST), + kqwl->kqwl_dynamicid, thread_tid(kqwl_owner), kqr->tr_kq_qos_index, + (kqr->tr_kq_override_index << 16) | kqr->tr_kq_wakeup); +#endif + if (new_override == old_override) { + // nothing to do + } else if (old_override == THREAD_QOS_UNSPECIFIED) { + thread_add_kevent_override(kqwl_owner, new_override); + } else if (new_override == THREAD_QOS_UNSPECIFIED) { + thread_drop_kevent_override(kqwl_owner); + } else { /* old_override != new_override */ + thread_update_kevent_override(kqwl_owner, new_override); + } + } - kqlock(parentkq); - if (parentkq->kq_level > 0 && - parentkq->kq_level < kq->kq_level) - { - kqunlock(parentkq); - kn->kn_flags = EV_ERROR; - kn->kn_data = EINVAL; - return 0; - } else { - /* set parent level appropriately */ - if (parentkq->kq_level == 0) - parentkq->kq_level = 2; - if (parentkq->kq_level < kq->kq_level + 1) - parentkq->kq_level = kq->kq_level + 1; - kqunlock(parentkq); + /* + * apply the diffs to the servicer + */ + if (!kqr_thread_requested(kqr)) { + /* + * No servicer, nor thread-request + * + * Make a new thread request, unless there is an owner (or the workloop + * is suspended in userland) or if there is no asynchronous work in the + * first place. + */ - kn->kn_filtid = EVFILTID_KQREAD; - kqlock(kq); - KNOTE_ATTACH(&kqf->kqf_sel.si_note, kn); - /* indicate nesting in child, if needed */ - if (kq->kq_level == 0) - kq->kq_level = 1; + if (kqwl_owner == NULL && kqr->tr_kq_wakeup) { + int initiate_flags = 0; + if (op == KQWL_UTQ_UNBINDING) { + initiate_flags = WORKQ_THREADREQ_ATTEMPT_REBIND; + } + kqueue_threadreq_initiate(kq, kqr, new_override, initiate_flags); + } + } else if (servicer) { + /* + * Servicer in flight + * + * Just apply the diff to the servicer + */ + struct uthread *ut = get_bsdthread_info(servicer); + if (ut->uu_kqueue_override != new_override) { + if (ut->uu_kqueue_override == THREAD_QOS_UNSPECIFIED) { + thread_add_servicer_override(servicer, new_override); + } else if (new_override == THREAD_QOS_UNSPECIFIED) { + thread_drop_servicer_override(servicer); + } else { /* ut->uu_kqueue_override != new_override */ + thread_update_servicer_override(servicer, new_override); + } + ut->uu_kqueue_override = new_override; + qos_changed = TRUE; + } + } else if (new_override == THREAD_QOS_UNSPECIFIED) { + /* + * No events to deliver anymore. + * + * However canceling with turnstiles is challenging, so the fact that + * the request isn't useful will be discovered by the servicer himself + * later on. + */ + } else if (old_override != new_override) { + /* + * Request is in flight + * + * Apply the diff to the thread request + */ + kqueue_threadreq_modify(kq, kqr, new_override, WORKQ_THREADREQ_NONE); + qos_changed = TRUE; + } - int count = kq->kq_count; - kqunlock(kq); - return (count > 0); + if (qos_changed) { + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWL_THADJUST), kqwl->kqwl_dynamicid, + thread_tid(servicer), kqr->tr_kq_qos_index, + (kqr->tr_kq_override_index << 16) | kqr->tr_kq_wakeup); } } -/* - * kqueue_drain - called when kq is closed - */ -/*ARGSUSED*/ -static int -kqueue_drain(struct fileproc *fp, __unused vfs_context_t ctx) +static void +kqworkloop_wakeup(struct kqworkloop *kqwl, kq_index_t qos) { - struct kqueue *kq = (struct kqueue *)fp->f_fglob->fg_data; - - assert((kq->kq_state & KQ_WORKQ) == 0); + if ((kqwl->kqwl_state & KQ_PROCESSING) && + kqr_thread(&kqwl->kqwl_request) == current_thread()) { + /* + * kqworkloop_end_processing() will perform the required QoS + * computations when it unsets the processing mode. + */ + return; + } - kqlock(kq); - kq->kq_state |= KQ_DRAIN; - kqueue_interrupt(kq); - kqunlock(kq); - return (0); + kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_UPDATE_WAKEUP_QOS, qos); } -/*ARGSUSED*/ -int -kqueue_stat(struct kqueue *kq, void *ub, int isstat64, proc_t p) +static struct kqtailq * +kqueue_get_suppressed_queue(kqueue_t kq, struct knote *kn) { - assert((kq->kq_state & KQ_WORKQ) == 0); - - kqlock(kq); - if (isstat64 != 0) { - struct stat64 *sb64 = (struct stat64 *)ub; - - bzero((void *)sb64, sizeof(*sb64)); - sb64->st_size = kq->kq_count; - if (kq->kq_state & KQ_KEV_QOS) - sb64->st_blksize = sizeof(struct kevent_qos_s); - else if (kq->kq_state & KQ_KEV64) - sb64->st_blksize = sizeof(struct kevent64_s); - else if (IS_64BIT_PROCESS(p)) - sb64->st_blksize = sizeof(struct user64_kevent); - else - sb64->st_blksize = sizeof(struct user32_kevent); - sb64->st_mode = S_IFIFO; + if (kq.kq->kq_state & KQ_WORKLOOP) { + return &kq.kqwl->kqwl_suppressed; + } else if (kq.kq->kq_state & KQ_WORKQ) { + return &kq.kqwq->kqwq_suppressed[kn->kn_qos_index]; } else { - struct stat *sb = (struct stat *)ub; - - bzero((void *)sb, sizeof(*sb)); - sb->st_size = kq->kq_count; - if (kq->kq_state & KQ_KEV_QOS) - sb->st_blksize = sizeof(struct kevent_qos_s); - else if (kq->kq_state & KQ_KEV64) - sb->st_blksize = sizeof(struct kevent64_s); - else if (IS_64BIT_PROCESS(p)) - sb->st_blksize = sizeof(struct user64_kevent); - else - sb->st_blksize = sizeof(struct user32_kevent); - sb->st_mode = S_IFIFO; + return &kq.kqf->kqf_suppressed; } - kqunlock(kq); - return (0); } -/* - * Interact with the pthread kext to request a servicing there. - * Eventually, this will request threads at specific QoS levels. - * For now, it only requests a dispatch-manager-QoS thread, and - * only one-at-a-time. - * - * - Caller holds the workq request lock - * - * - May be called with the kqueue's wait queue set locked, - * so cannot do anything that could recurse on that. - */ -static void -kqworkq_request_thread( - struct kqworkq *kqwq, - kq_index_t qos_index) +struct turnstile * +kqueue_alloc_turnstile(kqueue_t kqu) { - struct kqrequest *kqr; - - assert(kqwq->kqwq_state & KQ_WORKQ); - assert(qos_index < KQWQ_NQOS); + struct kqworkloop *kqwl = kqu.kqwl; + kq_state_t kq_state; - kqr = kqworkq_get_request(kqwq, qos_index); + kq_state = os_atomic_load(&kqu.kq->kq_state, dependency); + if (kq_state & KQ_HAS_TURNSTILE) { + /* force a dependency to pair with the atomic or with release below */ + return os_atomic_load_with_dependency_on(&kqwl->kqwl_turnstile, + (uintptr_t)kq_state); + } - assert(kqr->kqr_state & KQR_WAKEUP); + if (!(kq_state & KQ_WORKLOOP)) { + return TURNSTILE_NULL; + } - /* - * If we have already requested a thread, and it hasn't - * started processing yet, there's no use hammering away - * on the pthread kext. - */ - if (kqr->kqr_state & KQR_THREQUESTED) - return; + struct turnstile *ts = turnstile_alloc(), *free_ts = TURNSTILE_NULL; + bool workq_locked = false; - assert((kqr->kqr_state & KQR_BOUND) == 0); + kqlock(kqu); - /* request additional workq threads if appropriate */ - if (pthread_functions != NULL && - pthread_functions->workq_reqthreads != NULL) { - unsigned int flags = KEVENT_FLAG_WORKQ; - unsigned long priority; - thread_t wqthread; + if (filt_wlturnstile_interlock_is_workq(kqwl)) { + workq_locked = true; + workq_kern_threadreq_lock(kqwl->kqwl_p); + } - /* Compute the appropriate pthread priority */ - priority = qos_from_qos_index(qos_index); + if (kqwl->kqwl_state & KQ_HAS_TURNSTILE) { + free_ts = ts; + ts = kqwl->kqwl_turnstile; + } else { + ts = turnstile_prepare((uintptr_t)kqwl, &kqwl->kqwl_turnstile, + ts, TURNSTILE_WORKLOOPS); -#if 0 - /* JMM - for now remain compatible with old invocations */ - /* set the over-commit flag on the request if needed */ - if (kqr->kqr_state & KQR_THOVERCOMMIT) - priority |= _PTHREAD_PRIORITY_OVERCOMMIT_FLAG; -#endif /* 0 */ - - /* Compute a priority based on qos_index. */ - struct workq_reqthreads_req_s request = { - .priority = priority, - .count = 1 - }; + /* release-barrier to pair with the unlocked load of kqwl_turnstile above */ + os_atomic_or(&kqwl->kqwl_state, KQ_HAS_TURNSTILE, release); - /* mark that we are making a request */ - kqr->kqr_state |= KQR_THREQUESTED; - if (qos_index == KQWQ_QOS_MANAGER) - kqr->kqr_state |= KQWQ_THMANAGER; + if (filt_wlturnstile_interlock_is_workq(kqwl)) { + workq_kern_threadreq_update_inheritor(kqwl->kqwl_p, + &kqwl->kqwl_request, kqwl->kqwl_owner, + ts, TURNSTILE_IMMEDIATE_UPDATE); + /* + * The workq may no longer be the interlock after this. + * In which case the inheritor wasn't updated. + */ + } + if (!filt_wlturnstile_interlock_is_workq(kqwl)) { + filt_wlupdate_inheritor(kqwl, ts, TURNSTILE_IMMEDIATE_UPDATE); + } + } - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWQ_THREQUEST), - 0, qos_index, - (((uintptr_t)kqr->kqr_override_index << 8) | - (uintptr_t)kqr->kqr_state)); - wqthread = (*pthread_functions->workq_reqthreads)(kqwq->kqwq_p, 1, &request); + if (workq_locked) { + workq_kern_threadreq_unlock(kqwl->kqwl_p); + } - /* We've been switched to the emergency/manager thread */ - if (wqthread == (thread_t)-1) { - assert(qos_index != KQWQ_QOS_MANAGER); - kqr->kqr_state |= KQWQ_THMANAGER; - return; - } + kqunlock(kqu); - /* - * bind the returned thread identity - * This goes away when we switch to synchronous callback - * binding from the pthread kext. - */ - if (wqthread != NULL) { - kqworkq_bind_thread_impl(kqwq, qos_index, wqthread, flags); - } + if (free_ts) { + turnstile_deallocate(free_ts); + } else { + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_NOT_HELD); } + return ts; } -/* - * If we aren't already busy processing events [for this QoS], - * request workq thread support as appropriate. - * - * TBD - for now, we don't segregate out processing by QoS. - * - * - May be called with the kqueue's wait queue set locked, - * so cannot do anything that could recurse on that. - */ -static void -kqworkq_request_help( - struct kqworkq *kqwq, - kq_index_t qos_index) +__attribute__((always_inline)) +struct turnstile * +kqueue_turnstile(kqueue_t kqu) { - struct kqrequest *kqr; - - /* convert to thread qos value */ - assert(qos_index < KQWQ_NQOS); - - kqwq_req_lock(kqwq); - kqr = kqworkq_get_request(kqwq, qos_index); - - if ((kqr->kqr_state & KQR_WAKEUP) == 0) { - /* Indicate that we needed help from this request */ - kqr->kqr_state |= KQR_WAKEUP; + kq_state_t kq_state = os_atomic_load(&kqu.kq->kq_state, relaxed); + if (kq_state & KQ_WORKLOOP) { + return os_atomic_load(&kqu.kqwl->kqwl_turnstile, relaxed); + } + return TURNSTILE_NULL; +} - /* Go assure a thread request has been made */ - kqworkq_request_thread(kqwq, qos_index); +__attribute__((always_inline)) +struct turnstile * +kqueue_threadreq_get_turnstile(workq_threadreq_t kqr) +{ + struct kqworkloop *kqwl = kqr_kqworkloop(kqr); + if (kqwl) { + return os_atomic_load(&kqwl->kqwl_turnstile, relaxed); } - kqwq_req_unlock(kqwq); + return TURNSTILE_NULL; } static void -kqworkloop_threadreq_impl(struct kqworkloop *kqwl, kq_index_t qos_index) +kqworkloop_set_overcommit(struct kqworkloop *kqwl) { - struct kqrequest *kqr = &kqwl->kqwl_request; - unsigned long pri = pthread_priority_for_kqrequest(kqr, qos_index); - int op, ret; - - assert((kqr->kqr_state & (KQR_THREQUESTED | KQR_BOUND)) == KQR_THREQUESTED); + workq_threadreq_t kqr = &kqwl->kqwl_request; /* - * New-style thread request supported. Provide - * the pthread kext a pointer to a workq_threadreq_s - * structure for its use until a corresponding - * workloop_fulfill_threqreq callback. + * This test is racy, but since we never remove this bit, + * it allows us to avoid taking a lock. */ - if (current_proc() == kqwl->kqwl_kqueue.kq_p) { - op = WORKQ_THREADREQ_WORKLOOP_NO_THREAD_CALL; - } else { - op = WORKQ_THREADREQ_WORKLOOP; + if (kqr->tr_flags & WORKQ_TR_FLAG_OVERCOMMIT) { + return; } -again: - ret = (*pthread_functions->workq_threadreq)(kqwl->kqwl_p, &kqr->kqr_req, - WORKQ_THREADREQ_WORKLOOP, pri, 0); - switch (ret) { - case ENOTSUP: - assert(op == WORKQ_THREADREQ_WORKLOOP_NO_THREAD_CALL); - op = WORKQ_THREADREQ_WORKLOOP; - goto again; - - case ECANCELED: - case EINVAL: - /* - * Process is shutting down or exec'ing. - * All the kqueues are going to be cleaned up - * soon. Forget we even asked for a thread - - * and make sure we don't ask for more. - */ - kqueue_release((struct kqueue *)kqwl, KQUEUE_CANT_BE_LAST_REF); - kqr->kqr_state &= ~KQR_THREQUESTED; - kqr->kqr_state |= KQR_DRAIN; - break; - case EAGAIN: - assert(op == WORKQ_THREADREQ_WORKLOOP_NO_THREAD_CALL); - act_set_astkevent(current_thread(), AST_KEVENT_REDRIVE_THREADREQ); - break; + kqlock_held(kqwl); - default: - assert(ret == 0); + if (kqr_thread_requested_pending(kqr)) { + kqueue_threadreq_modify(kqwl, kqr, kqr->tr_qos, + WORKQ_THREADREQ_MAKE_OVERCOMMIT); + } else { + kqr->tr_flags |= WORKQ_TR_FLAG_OVERCOMMIT; } } static void -kqworkloop_threadreq_modify(struct kqworkloop *kqwl, kq_index_t qos_index) +kqworkq_update_override(struct kqworkq *kqwq, struct knote *kn, + kq_index_t override_index) { - struct kqrequest *kqr = &kqwl->kqwl_request; - unsigned long pri = pthread_priority_for_kqrequest(kqr, qos_index); - int ret, op = WORKQ_THREADREQ_CHANGE_PRI_NO_THREAD_CALL; - - assert((kqr->kqr_state & (KQR_THREQUESTED | KQR_BOUND)) == KQR_THREQUESTED); + workq_threadreq_t kqr; + kq_index_t old_override_index; + kq_index_t queue_index = kn->kn_qos_index; - if (current_proc() == kqwl->kqwl_kqueue.kq_p) { - op = WORKQ_THREADREQ_CHANGE_PRI_NO_THREAD_CALL; - } else { - op = WORKQ_THREADREQ_CHANGE_PRI; + if (override_index <= queue_index) { + return; } -again: - ret = (*pthread_functions->workq_threadreq_modify)(kqwl->kqwl_p, - &kqr->kqr_req, op, pri, 0); - switch (ret) { - case ENOTSUP: - assert(op == WORKQ_THREADREQ_CHANGE_PRI_NO_THREAD_CALL); - op = WORKQ_THREADREQ_CHANGE_PRI; - goto again; - - case EAGAIN: - assert(op == WORKQ_THREADREQ_WORKLOOP_NO_THREAD_CALL); - act_set_astkevent(current_thread(), AST_KEVENT_REDRIVE_THREADREQ); - break; - case ECANCELED: - case EINVAL: - case 0: - break; + kqr = kqworkq_get_request(kqwq, queue_index); - default: - assert(ret == 0); + kqlock_held(kqwq); + + old_override_index = kqr->tr_kq_override_index; + if (override_index > MAX(kqr->tr_kq_qos_index, old_override_index)) { + thread_t servicer = kqr_thread(kqr); + kqr->tr_kq_override_index = override_index; + + /* apply the override to [incoming?] servicing thread */ + if (servicer) { + if (old_override_index) { + thread_update_kevent_override(servicer, override_index); + } else { + thread_add_kevent_override(servicer, override_index); + } + } } } -/* - * Interact with the pthread kext to request a servicing thread. - * This will request a single thread at the highest QoS level - * for which there is work (whether that was the requested QoS - * for an event or an override applied to a lower-QoS request). - * - * - Caller holds the workloop request lock - * - * - May be called with the kqueue's wait queue set locked, - * so cannot do anything that could recurse on that. - */ static void -kqworkloop_request_thread(struct kqworkloop *kqwl, kq_index_t qos_index) +kqueue_update_override(kqueue_t kqu, struct knote *kn, thread_qos_t qos) { - struct kqrequest *kqr; - - assert(kqwl->kqwl_state & KQ_WORKLOOP); - - kqr = &kqwl->kqwl_request; + if (kqu.kq->kq_state & KQ_WORKLOOP) { + kqworkloop_update_threads_qos(kqu.kqwl, KQWL_UTQ_UPDATE_WAKEUP_OVERRIDE, + qos); + } else { + kqworkq_update_override(kqu.kqwq, kn, qos); + } +} - assert(kqwl->kqwl_owner == THREAD_NULL); - assert((kqr->kqr_state & KQR_BOUND) == 0); - assert((kqr->kqr_state & KQR_THREQUESTED) == 0); - assert(!(kqwl->kqwl_kqueue.kq_state & KQ_NO_WQ_THREAD)); +static void +kqworkloop_unbind_locked(struct kqworkloop *kqwl, thread_t thread, + enum kqwl_unbind_locked_mode how) +{ + struct uthread *ut = get_bsdthread_info(thread); + workq_threadreq_t kqr = &kqwl->kqwl_request; - /* If we're draining thread requests, just bail */ - if (kqr->kqr_state & KQR_DRAIN) - return; + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWL_UNBIND), kqwl->kqwl_dynamicid, + thread_tid(thread), 0, 0); - if (pthread_functions != NULL && - pthread_functions->workq_threadreq != NULL) { - /* - * set request state flags, etc... before calling pthread - * This assures they are set before a possible synchronous - * callback to workloop_fulfill_threadreq(). - */ - kqr->kqr_state |= KQR_THREQUESTED; + kqlock_held(kqwl); - /* Add a thread request reference on the kqueue. */ - kqueue_retain((struct kqueue *)kqwl); + assert(ut->uu_kqr_bound == kqr); + ut->uu_kqr_bound = NULL; + if (how == KQWL_OVERRIDE_DROP_IMMEDIATELY && + ut->uu_kqueue_override != THREAD_QOS_UNSPECIFIED) { + thread_drop_servicer_override(thread); + ut->uu_kqueue_override = THREAD_QOS_UNSPECIFIED; + } - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_THREQUEST), - kqwl->kqwl_dynamicid, - 0, qos_index, kqr->kqr_state); - kqworkloop_threadreq_impl(kqwl, qos_index); - } else { - panic("kqworkloop_request_thread"); - return; + if (kqwl->kqwl_owner == NULL && kqwl->kqwl_turnstile) { + turnstile_update_inheritor(kqwl->kqwl_turnstile, + TURNSTILE_INHERITOR_NULL, TURNSTILE_IMMEDIATE_UPDATE); + turnstile_update_inheritor_complete(kqwl->kqwl_turnstile, + TURNSTILE_INTERLOCK_HELD); } + + kqr->tr_thread = THREAD_NULL; + kqr->tr_state = WORKQ_TR_STATE_IDLE; + kqwl->kqwl_state &= ~KQ_R2K_ARMED; } static void -kqworkloop_update_sync_override_state(struct kqworkloop *kqwl, boolean_t sync_ipc_override) +kqworkloop_unbind_delayed_override_drop(thread_t thread) { - struct kqrequest *kqr = &kqwl->kqwl_request; - kqwl_req_lock(kqwl); - kqr->kqr_has_sync_override = sync_ipc_override; - kqwl_req_unlock(kqwl); - + struct uthread *ut = get_bsdthread_info(thread); + assert(ut->uu_kqr_bound == NULL); + if (ut->uu_kqueue_override != THREAD_QOS_UNSPECIFIED) { + thread_drop_servicer_override(thread); + ut->uu_kqueue_override = THREAD_QOS_UNSPECIFIED; + } } -static inline kq_index_t -kqworkloop_combined_qos(struct kqworkloop *kqwl, boolean_t *ipc_override_is_sync) +/* + * kqworkloop_unbind - Unbind the servicer thread of a workloop kqueue + * + * It will acknowledge events, and possibly request a new thread if: + * - there were active events left + * - we pended waitq hook callouts during processing + * - we pended wakeups while processing (or unsuppressing) + * + * Called with kqueue lock held. + */ +static void +kqworkloop_unbind(struct kqworkloop *kqwl) { - struct kqrequest *kqr = &kqwl->kqwl_request; - kq_index_t override; + struct kqueue *kq = &kqwl->kqwl_kqueue; + workq_threadreq_t kqr = &kqwl->kqwl_request; + thread_t thread = kqr_thread_fast(kqr); + int op = KQWL_UTQ_PARKING; + kq_index_t qos_override = THREAD_QOS_UNSPECIFIED; + + assert(thread == current_thread()); - *ipc_override_is_sync = FALSE; - override = MAX(MAX(kqr->kqr_qos_index, kqr->kqr_override_index), - kqr->kqr_dsync_waiters_qos); + kqlock(kqwl); - if (kqr->kqr_sync_suppress_count > 0 || kqr->kqr_has_sync_override) { - *ipc_override_is_sync = TRUE; - override = THREAD_QOS_USER_INTERACTIVE; + /* + * Forcing the KQ_PROCESSING flag allows for QoS updates because of + * unsuppressing knotes not to be applied until the eventual call to + * kqworkloop_update_threads_qos() below. + */ + assert((kq->kq_state & KQ_PROCESSING) == 0); + if (!TAILQ_EMPTY(&kqwl->kqwl_suppressed)) { + kq->kq_state |= KQ_PROCESSING; + qos_override = kqworkloop_acknowledge_events(kqwl); + kq->kq_state &= ~KQ_PROCESSING; } - return override; -} -static inline void -kqworkloop_request_fire_r2k_notification(struct kqworkloop *kqwl) -{ - struct kqrequest *kqr = &kqwl->kqwl_request; + kqworkloop_unbind_locked(kqwl, thread, KQWL_OVERRIDE_DROP_DELAYED); + kqworkloop_update_threads_qos(kqwl, op, qos_override); - kqwl_req_held(kqwl); + kqunlock(kqwl); - if (kqr->kqr_state & KQR_R2K_NOTIF_ARMED) { - assert(kqr->kqr_state & KQR_BOUND); - assert(kqr->kqr_thread); + /* + * Drop the override on the current thread last, after the call to + * kqworkloop_update_threads_qos above. + */ + kqworkloop_unbind_delayed_override_drop(thread); - kqr->kqr_state &= ~KQR_R2K_NOTIF_ARMED; - act_set_astkevent(kqr->kqr_thread, AST_KEVENT_RETURN_TO_KERNEL); - } + /* If last reference, dealloc the workloop kq */ + kqworkloop_release(kqwl); } -static void -kqworkloop_update_threads_qos(struct kqworkloop *kqwl, int op, kq_index_t qos) +static thread_qos_t +kqworkq_unbind_locked(struct kqworkq *kqwq, + workq_threadreq_t kqr, thread_t thread) { - const uint8_t KQWL_STAYACTIVE_FIRED_BIT = (1 << 0); - - struct kqrequest *kqr = &kqwl->kqwl_request; - boolean_t old_ipc_override_is_sync = FALSE; - kq_index_t old_qos = kqworkloop_combined_qos(kqwl, &old_ipc_override_is_sync); - struct kqueue *kq = &kqwl->kqwl_kqueue; - bool static_thread = (kq->kq_state & KQ_NO_WQ_THREAD); - kq_index_t i; + struct uthread *ut = get_bsdthread_info(thread); + kq_index_t old_override = kqr->tr_kq_override_index; - /* must hold the kqr lock */ - kqwl_req_held(kqwl); + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KQWQ_UNBIND), -1, + thread_tid(kqr_thread(kqr)), kqr->tr_kq_qos_index, 0); - switch (op) { - case KQWL_UTQ_UPDATE_WAKEUP_QOS: - if (qos == KQWL_BUCKET_STAYACTIVE) { - /* - * the KQWL_BUCKET_STAYACTIVE is not a QoS bucket, we only remember - * a high watermark (kqr_stayactive_qos) of any stay active knote - * that was ever registered with this workloop. - * - * When waitq_set__CALLING_PREPOST_HOOK__() wakes up any stay active - * knote, we use this high-watermark as a wakeup-index, and also set - * the magic KQWL_BUCKET_STAYACTIVE bit to make sure we remember - * there is at least one stay active knote fired until the next full - * processing of this bucket. - */ - kqr->kqr_wakeup_indexes |= KQWL_STAYACTIVE_FIRED_BIT; - qos = kqr->kqr_stayactive_qos; - assert(qos); - assert(!static_thread); - } - if (kqr->kqr_wakeup_indexes & (1 << qos)) { - assert(kqr->kqr_state & KQR_WAKEUP); - break; - } + kqlock_held(kqwq); - kqr->kqr_wakeup_indexes |= (1 << qos); - kqr->kqr_state |= KQR_WAKEUP; - kqworkloop_request_fire_r2k_notification(kqwl); - goto recompute_async; + assert(ut->uu_kqr_bound == kqr); + ut->uu_kqr_bound = NULL; + kqr->tr_thread = THREAD_NULL; + kqr->tr_state = WORKQ_TR_STATE_IDLE; + kqr->tr_kq_override_index = THREAD_QOS_UNSPECIFIED; + kqwq->kqwq_state &= ~KQ_R2K_ARMED; - case KQWL_UTQ_UPDATE_STAYACTIVE_QOS: - assert(qos); - if (kqr->kqr_stayactive_qos < qos) { - kqr->kqr_stayactive_qos = qos; - if (kqr->kqr_wakeup_indexes & KQWL_STAYACTIVE_FIRED_BIT) { - assert(kqr->kqr_state & KQR_WAKEUP); - kqr->kqr_wakeup_indexes |= (1 << qos); - goto recompute_async; - } - } - break; + return old_override; +} - case KQWL_UTQ_RECOMPUTE_WAKEUP_QOS: - kqlock_held(kq); // to look at kq_queues - kqr->kqr_has_sync_override = FALSE; - i = KQWL_BUCKET_STAYACTIVE; - if (TAILQ_EMPTY(&kqr->kqr_suppressed)) { - kqr->kqr_override_index = THREAD_QOS_UNSPECIFIED; - } - if (!TAILQ_EMPTY(&kq->kq_queue[i]) && - (kqr->kqr_wakeup_indexes & KQWL_STAYACTIVE_FIRED_BIT)) { - /* - * If the KQWL_STAYACTIVE_FIRED_BIT is set, it means a stay active - * knote may have fired, so we need to merge in kqr_stayactive_qos. - * - * Unlike other buckets, this one is never empty but could be idle. - */ - kqr->kqr_wakeup_indexes &= KQWL_STAYACTIVE_FIRED_BIT; - kqr->kqr_wakeup_indexes |= (1 << kqr->kqr_stayactive_qos); - } else { - kqr->kqr_wakeup_indexes = 0; - } - for (i = THREAD_QOS_UNSPECIFIED + 1; i < KQWL_BUCKET_STAYACTIVE; i++) { - if (!TAILQ_EMPTY(&kq->kq_queue[i])) { - kqr->kqr_wakeup_indexes |= (1 << i); - struct knote *kn = TAILQ_FIRST(&kqwl->kqwl_kqueue.kq_queue[i]); - if (i == THREAD_QOS_USER_INTERACTIVE && - kn->kn_qos_override_is_sync) { - kqr->kqr_has_sync_override = TRUE; - } - } - } - if (kqr->kqr_wakeup_indexes) { - kqr->kqr_state |= KQR_WAKEUP; - kqworkloop_request_fire_r2k_notification(kqwl); +/* + * kqworkq_unbind - unbind of a workq kqueue from a thread + * + * We may have to request new threads. + * This can happen there are no waiting processing threads and: + * - there were active events we never got to (count > 0) + * - we pended waitq hook callouts during processing + * - we pended wakeups while processing (or unsuppressing) + */ +static void +kqworkq_unbind(proc_t p, workq_threadreq_t kqr) +{ + struct kqworkq *kqwq = (struct kqworkq *)p->p_fd->fd_wqkqueue; + __assert_only int rc; + + kqlock(kqwq); + rc = kqworkq_acknowledge_events(kqwq, kqr, 0, KQWQAE_UNBIND); + assert(rc == -1); + kqunlock(kqwq); +} + +workq_threadreq_t +kqworkq_get_request(struct kqworkq *kqwq, kq_index_t qos_index) +{ + assert(qos_index < KQWQ_NBUCKETS); + return &kqwq->kqwq_request[qos_index]; +} + +static void +knote_reset_priority(kqueue_t kqu, struct knote *kn, pthread_priority_t pp) +{ + kq_index_t qos = _pthread_priority_thread_qos(pp); + + if (kqu.kq->kq_state & KQ_WORKLOOP) { + assert((pp & _PTHREAD_PRIORITY_EVENT_MANAGER_FLAG) == 0); + pp = _pthread_priority_normalize(pp); + } else if (kqu.kq->kq_state & KQ_WORKQ) { + if (qos == THREAD_QOS_UNSPECIFIED) { + /* On workqueues, outside of QoS means MANAGER */ + qos = KQWQ_QOS_MANAGER; + pp = _PTHREAD_PRIORITY_EVENT_MANAGER_FLAG; } else { - kqr->kqr_state &= ~KQR_WAKEUP; + pp = _pthread_priority_normalize(pp); } - assert(qos == THREAD_QOS_UNSPECIFIED); - goto recompute_async; + } else { + pp = _pthread_unspecified_priority(); + qos = THREAD_QOS_UNSPECIFIED; + } - case KQWL_UTQ_RESET_WAKEUP_OVERRIDE: - kqr->kqr_override_index = THREAD_QOS_UNSPECIFIED; - assert(qos == THREAD_QOS_UNSPECIFIED); - goto recompute_async; + kn->kn_qos = (int32_t)pp; - case KQWL_UTQ_UPDATE_WAKEUP_OVERRIDE: - recompute_async: - /* - * When modifying the wakeup QoS or the async override QoS, we always - * need to maintain our invariant that kqr_override_index is at least as - * large as the highest QoS for which an event is fired. - * - * However this override index can be larger when there is an overriden - * suppressed knote pushing on the kqueue. - */ - if (kqr->kqr_wakeup_indexes > (1 << qos)) { - qos = fls(kqr->kqr_wakeup_indexes) - 1; /* fls is 1-based */ - } - if (kqr->kqr_override_index < qos) { - kqr->kqr_override_index = qos; - } - break; + if ((kn->kn_status & KN_MERGE_QOS) == 0 || qos > kn->kn_qos_override) { + /* Never lower QoS when in "Merge" mode */ + kn->kn_qos_override = qos; + } - case KQWL_UTQ_REDRIVE_EVENTS: - break; + /* only adjust in-use qos index when not suppressed */ + if (kn->kn_status & KN_SUPPRESSED) { + kqueue_update_override(kqu, kn, qos); + } else if (kn->kn_qos_index != qos) { + knote_dequeue(kqu, kn); + kn->kn_qos_index = qos; + } +} - case KQWL_UTQ_SET_ASYNC_QOS: - filt_wlheld(kqwl); - kqr->kqr_qos_index = qos; - break; +static void +knote_adjust_qos(struct kqueue *kq, struct knote *kn, int result) +{ + thread_qos_t qos_index = (result >> FILTER_ADJUST_EVENT_QOS_SHIFT) & 7; - case KQWL_UTQ_SET_SYNC_WAITERS_QOS: - filt_wlheld(kqwl); - kqr->kqr_dsync_waiters_qos = qos; - break; + kqlock_held(kq); - default: - panic("unknown kqwl thread qos update operation: %d", op); - } + assert(result & FILTER_ADJUST_EVENT_QOS_BIT); + assert(qos_index < THREAD_QOS_LAST); - boolean_t new_ipc_override_is_sync = FALSE; - kq_index_t new_qos = kqworkloop_combined_qos(kqwl, &new_ipc_override_is_sync); - thread_t kqwl_owner = kqwl->kqwl_owner; - thread_t servicer = kqr->kqr_thread; - __assert_only int ret; + /* + * Early exit for knotes that should not change QoS + */ + if (__improbable(!knote_fops(kn)->f_adjusts_qos)) { + panic("filter %d cannot change QoS", kn->kn_filtid); + } else if (__improbable(!knote_has_qos(kn))) { + return; + } /* - * Apply the diffs to the owner if applicable + * knotes with the FALLBACK flag will only use their registration QoS if the + * incoming event has no QoS, else, the registration QoS acts as a floor. */ - if (filt_wlowner_is_valid(kqwl_owner)) { -#if 0 - /* JMM - need new trace hooks for owner overrides */ - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_THADJUST), - kqwl->kqwl_dynamicid, - (kqr->kqr_state & KQR_BOUND) ? thread_tid(kqwl_owner) : 0, - (kqr->kqr_qos_index << 8) | new_qos, - (kqr->kqr_override_index << 8) | kqr->kqr_state); -#endif - if (new_qos == kqr->kqr_dsync_owner_qos) { - // nothing to do - } else if (kqr->kqr_dsync_owner_qos == THREAD_QOS_UNSPECIFIED) { - thread_add_ipc_override(kqwl_owner, new_qos); - } else if (new_qos == THREAD_QOS_UNSPECIFIED) { - thread_drop_ipc_override(kqwl_owner); - } else /* kqr->kqr_dsync_owner_qos != new_qos */ { - thread_update_ipc_override(kqwl_owner, new_qos); + thread_qos_t req_qos = _pthread_priority_thread_qos_fast(kn->kn_qos); + if (kn->kn_qos & _PTHREAD_PRIORITY_FALLBACK_FLAG) { + if (qos_index == THREAD_QOS_UNSPECIFIED) { + qos_index = req_qos; } - kqr->kqr_dsync_owner_qos = new_qos; - - if (new_ipc_override_is_sync && - !kqr->kqr_owner_override_is_sync) { - thread_add_sync_ipc_override(kqwl_owner); - } else if (!new_ipc_override_is_sync && - kqr->kqr_owner_override_is_sync) { - thread_drop_sync_ipc_override(kqwl_owner); + } else { + if (qos_index < req_qos) { + qos_index = req_qos; } - kqr->kqr_owner_override_is_sync = new_ipc_override_is_sync; + } + if ((kn->kn_status & KN_MERGE_QOS) && (qos_index < kn->kn_qos_override)) { + /* Never lower QoS when in "Merge" mode */ + return; } - /* - * apply the diffs to the servicer - */ - if (static_thread) { + if ((kn->kn_status & KN_LOCKED) && (kn->kn_status & KN_POSTING)) { /* - * Statically bound thread + * When we're trying to update the QoS override and that both an + * f_event() and other f_* calls are running concurrently, any of these + * in flight calls may want to perform overrides that aren't properly + * serialized with each other. + * + * The first update that observes this racy situation enters a "Merge" + * mode which causes subsequent override requests to saturate the + * override instead of replacing its value. * - * These threads don't participates in QoS overrides today, just wakeup - * the thread blocked on this kqueue if a new event arrived. + * This mode is left when knote_unlock() or knote_post() + * observe that no other f_* routine is in flight. */ + kn->kn_status |= KN_MERGE_QOS; + } - switch (op) { - case KQWL_UTQ_UPDATE_WAKEUP_QOS: - case KQWL_UTQ_UPDATE_STAYACTIVE_QOS: - case KQWL_UTQ_RECOMPUTE_WAKEUP_QOS: - break; + /* + * Now apply the override if it changed. + */ - case KQWL_UTQ_RESET_WAKEUP_OVERRIDE: - case KQWL_UTQ_UPDATE_WAKEUP_OVERRIDE: - case KQWL_UTQ_REDRIVE_EVENTS: - case KQWL_UTQ_SET_ASYNC_QOS: - case KQWL_UTQ_SET_SYNC_WAITERS_QOS: - panic("should never be called"); - break; - } + if (kn->kn_qos_override == qos_index) { + return; + } - kqlock_held(kq); + kn->kn_qos_override = qos_index; - if ((kqr->kqr_state & KQR_BOUND) && (kqr->kqr_state & KQR_WAKEUP)) { - assert(servicer && !is_workqueue_thread(servicer)); - if (kq->kq_state & (KQ_SLEEP | KQ_SEL)) { - kq->kq_state &= ~(KQ_SLEEP | KQ_SEL); - waitq_wakeup64_all((struct waitq *)&kq->kq_wqs, KQ_EVENT, - THREAD_AWAKENED, WAITQ_ALL_PRIORITIES); - } - } - } else if ((kqr->kqr_state & KQR_THREQUESTED) == 0) { + if (kn->kn_status & KN_SUPPRESSED) { /* - * No servicer, nor thread-request + * For suppressed events, the kn_qos_index field cannot be touched as it + * allows us to know on which supress queue the knote is for a kqworkq. * - * Make a new thread request, unless there is an owner (or the workloop - * is suspended in userland) or if there is no asynchronous work in the - * first place. + * Also, there's no natural push applied on the kqueues when this field + * changes anyway. We hence need to apply manual overrides in this case, + * which will be cleared when the events are later acknowledged. */ + kqueue_update_override(kq, kn, qos_index); + } else if (kn->kn_qos_index != qos_index) { + knote_dequeue(kq, kn); + kn->kn_qos_index = qos_index; + } +} - if (kqwl_owner == THREAD_NULL && (kqr->kqr_state & KQR_WAKEUP)) { - kqworkloop_request_thread(kqwl, new_qos); - } - } else if ((kqr->kqr_state & KQR_BOUND) == 0 && - (kqwl_owner || (kqr->kqr_state & KQR_WAKEUP) == 0)) { - /* - * No servicer, thread request in flight we want to cancel - * - * We just got rid of the last knote of the kqueue or noticed an owner - * with a thread request still in flight, take it back. - */ - ret = (*pthread_functions->workq_threadreq_modify)(kqwl->kqwl_p, - &kqr->kqr_req, WORKQ_THREADREQ_CANCEL, 0, 0); - if (ret == 0) { - kqr->kqr_state &= ~KQR_THREQUESTED; - kqueue_release(kq, KQUEUE_CANT_BE_LAST_REF); - } - } else { - boolean_t qos_changed = FALSE; +/* + * Called back from waitq code when no threads waiting and the hook was set. + * + * Preemption is disabled - minimal work can be done in this context!!! + */ +void +waitq_set__CALLING_PREPOST_HOOK__(waitq_set_prepost_hook_t *kq_hook) +{ + kqueue_t kqu; - /* - * Servicer or request is in flight - * - * Just apply the diff to the servicer or the thread request - */ - if (kqr->kqr_state & KQR_BOUND) { - servicer = kqr->kqr_thread; - struct uthread *ut = get_bsdthread_info(servicer); - if (ut->uu_kqueue_qos_index != new_qos) { - if (ut->uu_kqueue_qos_index == THREAD_QOS_UNSPECIFIED) { - thread_add_ipc_override(servicer, new_qos); - } else if (new_qos == THREAD_QOS_UNSPECIFIED) { - thread_drop_ipc_override(servicer); - } else /* ut->uu_kqueue_qos_index != new_qos */ { - thread_update_ipc_override(servicer, new_qos); - } - ut->uu_kqueue_qos_index = new_qos; - qos_changed = TRUE; - } + kqu.kq = __container_of(kq_hook, struct kqueue, kq_waitq_hook); + assert(kqu.kq->kq_state & (KQ_WORKQ | KQ_WORKLOOP)); - if (new_ipc_override_is_sync != ut->uu_kqueue_override_is_sync) { - if (new_ipc_override_is_sync && - !ut->uu_kqueue_override_is_sync) { - thread_add_sync_ipc_override(servicer); - } else if (!new_ipc_override_is_sync && - ut->uu_kqueue_override_is_sync) { - thread_drop_sync_ipc_override(servicer); - } - ut->uu_kqueue_override_is_sync = new_ipc_override_is_sync; - qos_changed = TRUE; - } - } else if (old_qos != new_qos) { - assert(new_qos); - kqworkloop_threadreq_modify(kqwl, new_qos); - qos_changed = TRUE; - } - if (qos_changed) { - servicer = kqr->kqr_thread; - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_THADJUST), - kqwl->kqwl_dynamicid, - (kqr->kqr_state & KQR_BOUND) ? thread_tid(servicer) : 0, - (kqr->kqr_qos_index << 16) | (new_qos << 8) | new_ipc_override_is_sync, - (kqr->kqr_override_index << 8) | kqr->kqr_state); + kqlock(kqu); + + if (kqu.kq->kq_count > 0) { + if (kqu.kq->kq_state & KQ_WORKLOOP) { + kqworkloop_wakeup(kqu.kqwl, KQWL_BUCKET_STAYACTIVE); + } else { + kqworkq_wakeup(kqu.kqwq, KQWQ_QOS_MANAGER); } } + + kqunlock(kqu); } -static void -kqworkloop_request_help(struct kqworkloop *kqwl, kq_index_t qos_index) +void +klist_init(struct klist *list) { - /* convert to thread qos value */ - assert(qos_index < KQWL_NBUCKETS); - - kqwl_req_lock(kqwl); - kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_UPDATE_WAKEUP_QOS, qos_index); - kqwl_req_unlock(kqwl); + SLIST_INIT(list); } + /* - * These arrays described the low and high qindexes for a given qos_index. - * The values come from the chart in (must stay in sync). + * Query/Post each knote in the object's list + * + * The object lock protects the list. It is assumed + * that the filter/event routine for the object can + * determine that the object is already locked (via + * the hint) and not deadlock itself. + * + * The object lock should also hold off pending + * detach/drop operations. */ -static kq_index_t _kqwq_base_index[KQWQ_NQOS] = {0, 0, 6, 11, 15, 18, 20, 21}; -static kq_index_t _kqwq_high_index[KQWQ_NQOS] = {0, 5, 10, 14, 17, 19, 20, 21}; - -static struct kqtailq * -kqueue_get_base_queue(struct kqueue *kq, kq_index_t qos_index) +void +knote(struct klist *list, long hint) { - if (kq->kq_state & KQ_WORKQ) { - assert(qos_index < KQWQ_NQOS); - return &kq->kq_queue[_kqwq_base_index[qos_index]]; - } else if (kq->kq_state & KQ_WORKLOOP) { - assert(qos_index < KQWL_NBUCKETS); - return &kq->kq_queue[qos_index]; - } else { - assert(qos_index == QOS_INDEX_KQFILE); - return &kq->kq_queue[QOS_INDEX_KQFILE]; + struct knote *kn; + + SLIST_FOREACH(kn, list, kn_selnext) { + knote_post(kn, hint); } } -static struct kqtailq * -kqueue_get_high_queue(struct kqueue *kq, kq_index_t qos_index) +/* + * attach a knote to the specified list. Return true if this is the first entry. + * The list is protected by whatever lock the object it is associated with uses. + */ +int +knote_attach(struct klist *list, struct knote *kn) { - if (kq->kq_state & KQ_WORKQ) { - assert(qos_index < KQWQ_NQOS); - return &kq->kq_queue[_kqwq_high_index[qos_index]]; - } else if (kq->kq_state & KQ_WORKLOOP) { - assert(qos_index < KQWL_NBUCKETS); - return &kq->kq_queue[KQWL_BUCKET_STAYACTIVE]; - } else { - assert(qos_index == QOS_INDEX_KQFILE); - return &kq->kq_queue[QOS_INDEX_KQFILE]; - } + int ret = SLIST_EMPTY(list); + SLIST_INSERT_HEAD(list, kn, kn_selnext); + return ret; } -static int -kqueue_queue_empty(struct kqueue *kq, kq_index_t qos_index) +/* + * detach a knote from the specified list. Return true if that was the last entry. + * The list is protected by whatever lock the object it is associated with uses. + */ +int +knote_detach(struct klist *list, struct knote *kn) { - struct kqtailq *base_queue = kqueue_get_base_queue(kq, qos_index); - struct kqtailq *queue = kqueue_get_high_queue(kq, qos_index); - - do { - if (!TAILQ_EMPTY(queue)) - return 0; - } while (queue-- > base_queue); - return 1; + SLIST_REMOVE(list, kn, knote, kn_selnext); + return SLIST_EMPTY(list); } -static struct kqtailq * -kqueue_get_suppressed_queue(struct kqueue *kq, kq_index_t qos_index) +/* + * knote_vanish - Indicate that the source has vanished + * + * If the knote has requested EV_VANISHED delivery, + * arrange for that. Otherwise, deliver a NOTE_REVOKE + * event for backward compatibility. + * + * The knote is marked as having vanished, but is not + * actually detached from the source in this instance. + * The actual detach is deferred until the knote drop. + * + * Our caller already has the object lock held. Calling + * the detach routine would try to take that lock + * recursively - which likely is not supported. + */ +void +knote_vanish(struct klist *list, bool make_active) { - struct kqtailq *res; - struct kqrequest *kqr; - - if (kq->kq_state & KQ_WORKQ) { - struct kqworkq *kqwq = (struct kqworkq *)kq; + struct knote *kn; + struct knote *kn_next; - kqr = kqworkq_get_request(kqwq, qos_index); - res = &kqr->kqr_suppressed; - } else if (kq->kq_state & KQ_WORKLOOP) { - struct kqworkloop *kqwl = (struct kqworkloop *)kq; + SLIST_FOREACH_SAFE(kn, list, kn_selnext, kn_next) { + struct kqueue *kq = knote_get_kq(kn); - kqr = &kqwl->kqwl_request; - res = &kqr->kqr_suppressed; - } else { - struct kqfile *kqf = (struct kqfile *)kq; - res = &kqf->kqf_suppressed; + kqlock(kq); + if (__probable(kn->kn_status & KN_REQVANISH)) { + /* + * If EV_VANISH supported - prepare to deliver one + */ + kn->kn_status |= KN_VANISHED; + } else { + /* + * Handle the legacy way to indicate that the port/portset was + * deallocated or left the current Mach portspace (modern technique + * is with an EV_VANISHED protocol). + * + * Deliver an EV_EOF event for these changes (hopefully it will get + * delivered before the port name recycles to the same generation + * count and someone tries to re-register a kevent for it or the + * events are udata-specific - avoiding a conflict). + */ + kn->kn_flags |= EV_EOF | EV_ONESHOT; + } + if (make_active) { + knote_activate(kq, kn, FILTER_ACTIVE); + } + kqunlock(kq); } - return res; } -static kq_index_t -knote_get_queue_index(struct knote *kn) +/* + * Force a lazy allocation of the waitqset link + * of the kq_wqs associated with the kn + * if it wasn't already allocated. + * + * This allows knote_link_waitq to never block + * if reserved_link is not NULL. + */ +void +knote_link_waitqset_lazy_alloc(struct knote *kn) { - kq_index_t override_index = knote_get_qos_override_index(kn); - kq_index_t qos_index = knote_get_qos_index(kn); struct kqueue *kq = knote_get_kq(kn); - kq_index_t res; - - if (kq->kq_state & KQ_WORKQ) { - res = _kqwq_base_index[qos_index]; - if (override_index > qos_index) - res += override_index - qos_index; - assert(res <= _kqwq_high_index[qos_index]); - } else if (kq->kq_state & KQ_WORKLOOP) { - res = MAX(override_index, qos_index); - assert(res < KQWL_NBUCKETS); - } else { - assert(qos_index == QOS_INDEX_KQFILE); - assert(override_index == QOS_INDEX_KQFILE); - res = QOS_INDEX_KQFILE; - } - return res; -} - -static struct kqtailq * -knote_get_queue(struct knote *kn) -{ - kq_index_t qindex = knote_get_queue_index(kn); - - return &(knote_get_kq(kn))->kq_queue[qindex]; -} - -static kq_index_t -knote_get_req_index(struct knote *kn) -{ - return kn->kn_req_index; -} - -static kq_index_t -knote_get_qos_index(struct knote *kn) -{ - return kn->kn_qos_index; + waitq_set_lazy_init_link(&kq->kq_wqs); } -static void -knote_set_qos_index(struct knote *kn, kq_index_t qos_index) +/* + * Check if a lazy allocation for the waitqset link + * of the kq_wqs is needed. + */ +boolean_t +knote_link_waitqset_should_lazy_alloc(struct knote *kn) { struct kqueue *kq = knote_get_kq(kn); - - assert(qos_index < KQWQ_NQOS); - assert((kn->kn_status & KN_QUEUED) == 0); - - if (kq->kq_state & KQ_WORKQ) { - assert(qos_index > THREAD_QOS_UNSPECIFIED); - } else if (kq->kq_state & KQ_WORKLOOP) { - /* XXX this policy decision shouldn't be here */ - if (qos_index == THREAD_QOS_UNSPECIFIED) - qos_index = THREAD_QOS_LEGACY; - } else - qos_index = QOS_INDEX_KQFILE; - - /* always set requested */ - kn->kn_req_index = qos_index; - - /* only adjust in-use qos index when not suppressed */ - if ((kn->kn_status & KN_SUPPRESSED) == 0) - kn->kn_qos_index = qos_index; + return waitq_set_should_lazy_init_link(&kq->kq_wqs); } -static void -knote_set_qos_overcommit(struct knote *kn) +/* + * For a given knote, link a provided wait queue directly with the kqueue. + * Wakeups will happen via recursive wait queue support. But nothing will move + * the knote to the active list at wakeup (nothing calls knote()). Instead, + * we permanently enqueue them here. + * + * kqueue and knote references are held by caller. + * waitq locked by caller. + * + * caller provides the wait queue link structure and insures that the kq->kq_wqs + * is linked by previously calling knote_link_waitqset_lazy_alloc. + */ +int +knote_link_waitq(struct knote *kn, struct waitq *wq, uint64_t *reserved_link) { struct kqueue *kq = knote_get_kq(kn); - struct kqrequest *kqr; - - /* turn overcommit on for the appropriate thread request? */ - if (kn->kn_qos & _PTHREAD_PRIORITY_OVERCOMMIT_FLAG) { - if (kq->kq_state & KQ_WORKQ) { - kq_index_t qos_index = knote_get_qos_index(kn); - struct kqworkq *kqwq = (struct kqworkq *)kq; - - kqr = kqworkq_get_request(kqwq, qos_index); - - kqwq_req_lock(kqwq); - kqr->kqr_state |= KQR_THOVERCOMMIT; - kqwq_req_unlock(kqwq); - } else if (kq->kq_state & KQ_WORKLOOP) { - struct kqworkloop *kqwl = (struct kqworkloop *)kq; - - kqr = &kqwl->kqwl_request; + kern_return_t kr; - kqwl_req_lock(kqwl); - kqr->kqr_state |= KQR_THOVERCOMMIT; - kqwl_req_unlock(kqwl); - } + kr = waitq_link(wq, &kq->kq_wqs, WAITQ_ALREADY_LOCKED, reserved_link); + if (kr == KERN_SUCCESS) { + knote_markstayactive(kn); + return 0; + } else { + return EINVAL; } } -static kq_index_t -knote_get_qos_override_index(struct knote *kn) -{ - return kn->kn_qos_override; -} - -static void -knote_set_qos_override_index(struct knote *kn, kq_index_t override_index, - boolean_t override_is_sync) +/* + * Unlink the provided wait queue from the kqueue associated with a knote. + * Also remove it from the magic list of directly attached knotes. + * + * Note that the unlink may have already happened from the other side, so + * ignore any failures to unlink and just remove it from the kqueue list. + * + * On success, caller is responsible for the link structure + */ +int +knote_unlink_waitq(struct knote *kn, struct waitq *wq) { struct kqueue *kq = knote_get_kq(kn); - kq_index_t qos_index = knote_get_qos_index(kn); - kq_index_t old_override_index = knote_get_qos_override_index(kn); - boolean_t old_override_is_sync = kn->kn_qos_override_is_sync; - uint32_t flags = 0; + kern_return_t kr; - assert((kn->kn_status & KN_QUEUED) == 0); + kr = waitq_unlink(wq, &kq->kq_wqs); + knote_clearstayactive(kn); + return (kr != KERN_SUCCESS) ? EINVAL : 0; +} - if (override_index == KQWQ_QOS_MANAGER) { - assert(qos_index == KQWQ_QOS_MANAGER); - } else { - assert(override_index < KQWQ_QOS_MANAGER); - } +/* + * remove all knotes referencing a specified fd + * + * Entered with the proc_fd lock already held. + * It returns the same way, but may drop it temporarily. + */ +void +knote_fdclose(struct proc *p, int fd) +{ + struct klist *list; + struct knote *kn; + KNOTE_LOCK_CTX(knlc); - kn->kn_qos_override = override_index; - kn->kn_qos_override_is_sync = override_is_sync; +restart: + list = &p->p_fd->fd_knlist[fd]; + SLIST_FOREACH(kn, list, kn_link) { + struct kqueue *kq = knote_get_kq(kn); - /* - * If this is a workq/workloop kqueue, apply the override to the - * servicing thread. - */ - if (kq->kq_state & KQ_WORKQ) { - struct kqworkq *kqwq = (struct kqworkq *)kq; + kqlock(kq); - assert(qos_index > THREAD_QOS_UNSPECIFIED); - kqworkq_update_override(kqwq, qos_index, override_index); - } else if (kq->kq_state & KQ_WORKLOOP) { - struct kqworkloop *kqwl = (struct kqworkloop *)kq; + if (kq->kq_p != p) { + panic("%s: proc mismatch (kq->kq_p=%p != p=%p)", + __func__, kq->kq_p, p); + } + + /* + * If the knote supports EV_VANISHED delivery, + * transition it to vanished mode (or skip over + * it if already vanished). + */ + if (kn->kn_status & KN_VANISHED) { + kqunlock(kq); + continue; + } - if ((kn->kn_status & KN_SUPPRESSED) == KN_SUPPRESSED) { - flags = flags | KQWL_UO_UPDATE_SUPPRESS_SYNC_COUNTERS; + proc_fdunlock(p); + if (!knote_lock(kq, kn, &knlc, KNOTE_KQ_LOCK_ON_SUCCESS)) { + /* the knote was dropped by someone, nothing to do */ + } else if (kn->kn_status & KN_REQVANISH) { + kn->kn_status |= KN_VANISHED; - if (override_index == THREAD_QOS_USER_INTERACTIVE - && override_is_sync) { - flags = flags | KQWL_UO_NEW_OVERRIDE_IS_SYNC_UI; + kqunlock(kq); + knote_fops(kn)->f_detach(kn); + if (kn->kn_is_fd) { + fp_drop(p, (int)kn->kn_id, kn->kn_fp, 0); } + kn->kn_filtid = EVFILTID_DETACHED; + kqlock(kq); - if (old_override_index == THREAD_QOS_USER_INTERACTIVE - && old_override_is_sync) { - flags = flags | KQWL_UO_OLD_OVERRIDE_IS_SYNC_UI; - } + knote_activate(kq, kn, FILTER_ACTIVE); + knote_unlock(kq, kn, &knlc, KNOTE_KQ_UNLOCK); + } else { + knote_drop(kq, kn, &knlc); } - assert(qos_index > THREAD_QOS_UNSPECIFIED); - kqworkloop_update_override(kqwl, qos_index, override_index, flags); + proc_fdlock(p); + goto restart; } } -static kq_index_t -knote_get_sync_qos_override_index(struct knote *kn) -{ - return kn->kn_qos_sync_override; -} - -static void -kqworkq_update_override(struct kqworkq *kqwq, kq_index_t qos_index, kq_index_t override_index) +/* + * knote_fdfind - lookup a knote in the fd table for process + * + * If the filter is file-based, lookup based on fd index. + * Otherwise use a hash based on the ident. + * + * Matching is based on kq, filter, and ident. Optionally, + * it may also be based on the udata field in the kevent - + * allowing multiple event registration for the file object + * per kqueue. + * + * fd_knhashlock or fdlock held on entry (and exit) + */ +static struct knote * +knote_fdfind(struct kqueue *kq, + const struct kevent_internal_s *kev, + bool is_fd, + struct proc *p) { - struct kqrequest *kqr; - kq_index_t old_override_index; + struct filedesc *fdp = p->p_fd; + struct klist *list = NULL; + struct knote *kn = NULL; - if (override_index <= qos_index) { - return; + /* + * determine where to look for the knote + */ + if (is_fd) { + /* fd-based knotes are linked off the fd table */ + if (kev->kei_ident < (u_int)fdp->fd_knlistsize) { + list = &fdp->fd_knlist[kev->kei_ident]; + } + } else if (fdp->fd_knhashmask != 0) { + /* hash non-fd knotes here too */ + list = &fdp->fd_knhash[KN_HASH((u_long)kev->kei_ident, fdp->fd_knhashmask)]; } - kqr = kqworkq_get_request(kqwq, qos_index); - - kqwq_req_lock(kqwq); - old_override_index = kqr->kqr_override_index; - if (override_index > MAX(kqr->kqr_qos_index, old_override_index)) { - kqr->kqr_override_index = override_index; - - /* apply the override to [incoming?] servicing thread */ - if (kqr->kqr_state & KQR_BOUND) { - thread_t wqthread = kqr->kqr_thread; - - /* only apply if non-manager */ - assert(wqthread); - if ((kqr->kqr_state & KQWQ_THMANAGER) == 0) { - if (old_override_index) - thread_update_ipc_override(wqthread, override_index); - else - thread_add_ipc_override(wqthread, override_index); + /* + * scan the selected list looking for a match + */ + if (list != NULL) { + SLIST_FOREACH(kn, list, kn_link) { + if (kq == knote_get_kq(kn) && + kev->kei_ident == kn->kn_id && + kev->kei_filter == kn->kn_filter) { + if (kev->kei_flags & EV_UDATA_SPECIFIC) { + if ((kn->kn_flags & EV_UDATA_SPECIFIC) && + kev->kei_udata == kn->kn_udata) { + break; /* matching udata-specific knote */ + } + } else if ((kn->kn_flags & EV_UDATA_SPECIFIC) == 0) { + break; /* matching non-udata-specific knote */ + } } } } - kqwq_req_unlock(kqwq); + return kn; } -/* called with the kqworkq lock held */ -static void -kqworkq_bind_thread_impl( - struct kqworkq *kqwq, - kq_index_t qos_index, - thread_t thread, - unsigned int flags) +/* + * kq_add_knote- Add knote to the fd table for process + * while checking for duplicates. + * + * All file-based filters associate a list of knotes by file + * descriptor index. All other filters hash the knote by ident. + * + * May have to grow the table of knote lists to cover the + * file descriptor index presented. + * + * fd_knhashlock and fdlock unheld on entry (and exit). + * + * Takes a rwlock boost if inserting the knote is successful. + */ +static int +kq_add_knote(struct kqueue *kq, struct knote *kn, struct knote_lock_ctx *knlc, + struct proc *p) { - /* request lock must be held */ - kqwq_req_held(kqwq); - - struct kqrequest *kqr = kqworkq_get_request(kqwq, qos_index); - assert(kqr->kqr_state & KQR_THREQUESTED); - - if (qos_index == KQWQ_QOS_MANAGER) - flags |= KEVENT_FLAG_WORKQ_MANAGER; + struct filedesc *fdp = p->p_fd; + struct klist *list = NULL; + int ret = 0; + bool is_fd = kn->kn_is_fd; + uint64_t nofile = proc_limitgetcur(p, RLIMIT_NOFILE, TRUE); - struct uthread *ut = get_bsdthread_info(thread); + if (is_fd) { + proc_fdlock(p); + } else { + knhash_lock(fdp); + } - /* - * If this is a manager, and the manager request bit is - * not set, assure no other thread is bound. If the bit - * is set, make sure the old thread is us (or not set). - */ - if (flags & KEVENT_FLAG_WORKQ_MANAGER) { - if ((kqr->kqr_state & KQR_BOUND) == 0) { - kqr->kqr_state |= (KQR_BOUND | KQWQ_THMANAGER); - TAILQ_INIT(&kqr->kqr_suppressed); - kqr->kqr_thread = thread; - ut->uu_kqueue_bound = (struct kqueue *)kqwq; - ut->uu_kqueue_qos_index = KQWQ_QOS_MANAGER; - ut->uu_kqueue_flags = (KEVENT_FLAG_WORKQ | - KEVENT_FLAG_WORKQ_MANAGER); - } else { - assert(kqr->kqr_state & KQR_BOUND); - assert(thread == kqr->kqr_thread); - assert(ut->uu_kqueue_bound == (struct kqueue *)kqwq); - assert(ut->uu_kqueue_qos_index == KQWQ_QOS_MANAGER); - assert(ut->uu_kqueue_flags & KEVENT_FLAG_WORKQ_MANAGER); - } - return; + if (knote_fdfind(kq, &kn->kn_kevent, is_fd, p) != NULL) { + /* found an existing knote: we can't add this one */ + ret = ERESTART; + goto out_locked; } - /* Just a normal one-queue servicing thread */ - assert(kqr->kqr_state & KQR_THREQUESTED); - assert(kqr->kqr_qos_index == qos_index); + /* knote was not found: add it now */ + if (!is_fd) { + if (fdp->fd_knhashmask == 0) { + u_long size = 0; - if ((kqr->kqr_state & KQR_BOUND) == 0) { - kqr->kqr_state |= KQR_BOUND; - TAILQ_INIT(&kqr->kqr_suppressed); - kqr->kqr_thread = thread; + list = hashinit(CONFIG_KN_HASHSIZE, M_KQUEUE, &size); + if (list == NULL) { + ret = ENOMEM; + goto out_locked; + } - /* apply an ipc QoS override if one is needed */ - if (kqr->kqr_override_index) { - assert(kqr->kqr_qos_index); - assert(kqr->kqr_override_index > kqr->kqr_qos_index); - assert(thread_get_ipc_override(thread) == THREAD_QOS_UNSPECIFIED); - thread_add_ipc_override(thread, kqr->kqr_override_index); + fdp->fd_knhash = list; + fdp->fd_knhashmask = size; } - /* indicate that we are processing in the uthread */ - ut->uu_kqueue_bound = (struct kqueue *)kqwq; - ut->uu_kqueue_qos_index = qos_index; - ut->uu_kqueue_flags = flags; + list = &fdp->fd_knhash[KN_HASH(kn->kn_id, fdp->fd_knhashmask)]; + SLIST_INSERT_HEAD(list, kn, kn_link); + ret = 0; + goto out_locked; } else { - /* - * probably syncronously bound AND post-request bound - * this logic can go away when we get rid of post-request bind - */ - assert(kqr->kqr_state & KQR_BOUND); - assert(thread == kqr->kqr_thread); - assert(ut->uu_kqueue_bound == (struct kqueue *)kqwq); - assert(ut->uu_kqueue_qos_index == qos_index); - assert((ut->uu_kqueue_flags & flags) == flags); - } -} + /* knote is fd based */ -static void -kqworkloop_update_override( - struct kqworkloop *kqwl, - kq_index_t qos_index, - kq_index_t override_index, - uint32_t flags) -{ - struct kqrequest *kqr = &kqwl->kqwl_request; + if ((u_int)fdp->fd_knlistsize <= kn->kn_id) { + u_int size = 0; + + /* Make sure that fd stays below current process's soft limit AND system allowed per-process limits */ + if (kn->kn_id >= (uint64_t) nofile + || kn->kn_id >= (uint64_t)maxfilesperproc) { + ret = EINVAL; + goto out_locked; + } + /* have to grow the fd_knlist */ + size = fdp->fd_knlistsize; + while (size <= kn->kn_id) { + size += KQEXTENT; + } - kqwl_req_lock(kqwl); + if (size >= (UINT_MAX / sizeof(struct klist *))) { + ret = EINVAL; + goto out_locked; + } - /* Do not override on attached threads */ - if (kqr->kqr_state & KQR_BOUND) { - assert(kqr->kqr_thread); + list = kheap_alloc(KM_KQUEUE, size * sizeof(struct klist *), + Z_WAITOK); + if (list == NULL) { + ret = ENOMEM; + goto out_locked; + } - if (kqwl->kqwl_kqueue.kq_state & KQ_NO_WQ_THREAD) { - kqwl_req_unlock(kqwl); - assert(!is_workqueue_thread(kqr->kqr_thread)); - return; + bcopy((caddr_t)fdp->fd_knlist, (caddr_t)list, + fdp->fd_knlistsize * sizeof(struct klist *)); + bzero((caddr_t)list + + fdp->fd_knlistsize * sizeof(struct klist *), + (size - fdp->fd_knlistsize) * sizeof(struct klist *)); + kheap_free(KM_KQUEUE, fdp->fd_knlist, + fdp->fd_knlistsize * sizeof(struct klist *)); + fdp->fd_knlist = list; + fdp->fd_knlistsize = size; } - } - /* Update sync ipc counts on kqr for suppressed knotes */ - if (flags & KQWL_UO_UPDATE_SUPPRESS_SYNC_COUNTERS) { - kqworkloop_update_suppress_sync_count(kqr, flags); + list = &fdp->fd_knlist[kn->kn_id]; + SLIST_INSERT_HEAD(list, kn, kn_link); + ret = 0; + goto out_locked; } - if ((flags & KQWL_UO_UPDATE_OVERRIDE_LAZY) == 0) { - kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_UPDATE_WAKEUP_OVERRIDE, - MAX(qos_index, override_index)); +out_locked: + if (ret == 0) { + kqlock(kq); + assert((kn->kn_status & KN_LOCKED) == 0); + (void)knote_lock(kq, kn, knlc, KNOTE_KQ_UNLOCK); + kqueue_retain(kq); /* retain a kq ref */ } - kqwl_req_unlock(kqwl); -} - -static void -kqworkloop_update_suppress_sync_count( - struct kqrequest *kqr, - uint32_t flags) -{ - if (flags & KQWL_UO_NEW_OVERRIDE_IS_SYNC_UI) { - kqr->kqr_sync_suppress_count++; + if (is_fd) { + proc_fdunlock(p); + } else { + knhash_unlock(fdp); } - if (flags & KQWL_UO_OLD_OVERRIDE_IS_SYNC_UI) { - assert(kqr->kqr_sync_suppress_count > 0); - kqr->kqr_sync_suppress_count--; - } + return ret; } /* - * kqworkloop_unbind_thread - Unbind the servicer thread of a workloop kqueue - * - * It will end the processing phase in case it was still processing: + * kq_remove_knote - remove a knote from the fd table for process * - * We may have to request a new thread for not KQ_NO_WQ_THREAD workloop. - * This can happen if : - * - there were active events at or above our QoS we never got to (count > 0) - * - we pended waitq hook callouts during processing - * - we pended wakeups while processing (or unsuppressing) + * If the filter is file-based, remove based on fd index. + * Otherwise remove from the hash based on the ident. * - * Called with kqueue lock held. + * fd_knhashlock and fdlock unheld on entry (and exit). */ - static void -kqworkloop_unbind_thread( - struct kqworkloop *kqwl, - thread_t thread, - __unused unsigned int flags) +kq_remove_knote(struct kqueue *kq, struct knote *kn, struct proc *p, + struct knote_lock_ctx *knlc) { - struct kqueue *kq = &kqwl->kqwl_kqueue; - struct kqrequest *kqr = &kqwl->kqwl_request; - - kqlock_held(kq); + struct filedesc *fdp = p->p_fd; + struct klist *list = NULL; + uint16_t kq_state; + bool is_fd = kn->kn_is_fd; - assert((kq->kq_state & KQ_PROCESSING) == 0); - if (kq->kq_state & KQ_PROCESSING) { - return; + if (is_fd) { + proc_fdlock(p); + } else { + knhash_lock(fdp); } - /* - * Forcing the KQ_PROCESSING flag allows for QoS updates because of - * unsuppressing knotes not to be applied until the eventual call to - * kqworkloop_update_threads_qos() below. - */ - kq->kq_state |= KQ_PROCESSING; - kqworkloop_acknowledge_events(kqwl, TRUE); - kq->kq_state &= ~KQ_PROCESSING; - - kqwl_req_lock(kqwl); - - /* deal with extraneous unbinds in release kernels */ - assert((kqr->kqr_state & (KQR_BOUND | KQR_PROCESSING)) == KQR_BOUND); - if ((kqr->kqr_state & (KQR_BOUND | KQR_PROCESSING)) != KQR_BOUND) { - kqwl_req_unlock(kqwl); - return; + if (is_fd) { + assert((u_int)fdp->fd_knlistsize > kn->kn_id); + list = &fdp->fd_knlist[kn->kn_id]; + } else { + list = &fdp->fd_knhash[KN_HASH(kn->kn_id, fdp->fd_knhashmask)]; } + SLIST_REMOVE(list, kn, knote, kn_link); - assert(thread == current_thread()); - assert(kqr->kqr_thread == thread); - if (kqr->kqr_thread != thread) { - kqwl_req_unlock(kqwl); - return; + kqlock(kq); + kq_state = kq->kq_state; + if (knlc) { + knote_unlock_cancel(kq, kn, knlc); + } else { + kqunlock(kq); } - - struct uthread *ut = get_bsdthread_info(thread); - kq_index_t old_qos_index = ut->uu_kqueue_qos_index; - boolean_t ipc_override_is_sync = ut->uu_kqueue_override_is_sync; - ut->uu_kqueue_bound = NULL; - ut->uu_kqueue_qos_index = 0; - ut->uu_kqueue_override_is_sync = 0; - ut->uu_kqueue_flags = 0; - - /* unbind the servicer thread, drop overrides */ - kqr->kqr_thread = NULL; - kqr->kqr_state &= ~(KQR_BOUND | KQR_THREQUESTED | KQR_R2K_NOTIF_ARMED); - kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_RECOMPUTE_WAKEUP_QOS, 0); - - kqwl_req_unlock(kqwl); - - /* - * Drop the override on the current thread last, after the call to - * kqworkloop_update_threads_qos above. - */ - if (old_qos_index) { - thread_drop_ipc_override(thread); + if (is_fd) { + proc_fdunlock(p); + } else { + knhash_unlock(fdp); } - if (ipc_override_is_sync) { - thread_drop_sync_ipc_override(thread); + + if (kq_state & KQ_DYNAMIC) { + kqworkloop_release((struct kqworkloop *)kq); } } -/* called with the kqworkq lock held */ -static void -kqworkq_unbind_thread( - struct kqworkq *kqwq, - kq_index_t qos_index, - thread_t thread, - __unused unsigned int flags) -{ - struct kqrequest *kqr = kqworkq_get_request(kqwq, qos_index); - kq_index_t override_index = 0; - - /* request lock must be held */ - kqwq_req_held(kqwq); +/* + * kq_find_knote_and_kq_lock - lookup a knote in the fd table for process + * and, if the knote is found, acquires the kqlock while holding the fd table lock/spinlock. + * + * fd_knhashlock or fdlock unheld on entry (and exit) + */ - assert(thread == current_thread()); +static struct knote * +kq_find_knote_and_kq_lock(struct kqueue *kq, struct kevent_qos_s *kev, + bool is_fd, struct proc *p) +{ + struct filedesc *fdp = p->p_fd; + struct knote *kn; - if ((kqr->kqr_state & KQR_BOUND) == 0) { - assert(kqr->kqr_state & KQR_BOUND); - return; + if (is_fd) { + proc_fdlock(p); + } else { + knhash_lock(fdp); } - assert(kqr->kqr_thread == thread); - assert(TAILQ_EMPTY(&kqr->kqr_suppressed)); - - /* - * If there is an override, drop it from the current thread - * and then we are free to recompute (a potentially lower) - * minimum override to apply to the next thread request. + /* + * Temporary horrible hack: + * this cast is gross and will go away in a future change. + * It is OK to do because we don't look at xflags/s_fflags, + * and that when we cast down the kev this way, + * the truncated filter field works. */ - if (kqr->kqr_override_index) { - struct kqtailq *base_queue = kqueue_get_base_queue(&kqwq->kqwq_kqueue, qos_index); - struct kqtailq *queue = kqueue_get_high_queue(&kqwq->kqwq_kqueue, qos_index); - - /* if not bound to a manager thread, drop the current ipc override */ - if ((kqr->kqr_state & KQWQ_THMANAGER) == 0) { - thread_drop_ipc_override(thread); - } + kn = knote_fdfind(kq, (struct kevent_internal_s *)kev, is_fd, p); - /* recompute the new override */ - do { - if (!TAILQ_EMPTY(queue)) { - override_index = queue - base_queue + qos_index; - break; - } - } while (queue-- > base_queue); + if (kn) { + kqlock(kq); + assert(knote_get_kq(kn) == kq); } - /* Mark it unbound */ - kqr->kqr_thread = NULL; - kqr->kqr_state &= ~(KQR_BOUND | KQR_THREQUESTED | KQWQ_THMANAGER); - - /* apply the new override */ - if (override_index > kqr->kqr_qos_index) { - kqr->kqr_override_index = override_index; + if (is_fd) { + proc_fdunlock(p); } else { - kqr->kqr_override_index = THREAD_QOS_UNSPECIFIED; + knhash_unlock(fdp); } -} -struct kqrequest * -kqworkq_get_request(struct kqworkq *kqwq, kq_index_t qos_index) -{ - assert(qos_index < KQWQ_NQOS); - return &kqwq->kqwq_request[qos_index]; + return kn; } -void -knote_adjust_qos(struct knote *kn, qos_t new_qos, qos_t new_override, kq_index_t sync_override_index) +__attribute__((noinline)) +static void +kqfile_wakeup(struct kqfile *kqf, __unused kq_index_t qos) { - struct kqueue *kq = knote_get_kq(kn); - boolean_t override_is_sync = FALSE; + /* flag wakeups during processing */ + if (kqf->kqf_state & KQ_PROCESSING) { + kqf->kqf_state |= KQ_WAKEUP; + } - if (kq->kq_state & (KQ_WORKQ | KQ_WORKLOOP)) { - kq_index_t new_qos_index; - kq_index_t new_override_index; - kq_index_t servicer_qos_index; + /* wakeup a thread waiting on this queue */ + if (kqf->kqf_state & (KQ_SLEEP | KQ_SEL)) { + kqf->kqf_state &= ~(KQ_SLEEP | KQ_SEL); + waitq_wakeup64_all((struct waitq *)&kqf->kqf_wqs, KQ_EVENT, + THREAD_AWAKENED, WAITQ_ALL_PRIORITIES); + } - new_qos_index = qos_index_from_qos(kn, new_qos, FALSE); - new_override_index = qos_index_from_qos(kn, new_override, TRUE); + /* wakeup other kqueues/select sets we're inside */ + KNOTE(&kqf->kqf_sel.si_note, 0); +} - /* make sure the servicer qos acts as a floor */ - servicer_qos_index = qos_index_from_qos(kn, kn->kn_qos, FALSE); - if (servicer_qos_index > new_qos_index) - new_qos_index = servicer_qos_index; - if (servicer_qos_index > new_override_index) - new_override_index = servicer_qos_index; - if (sync_override_index >= new_override_index) { - new_override_index = sync_override_index; - override_is_sync = TRUE; - } +static struct kqtailq * +knote_get_tailq(kqueue_t kqu, struct knote *kn) +{ + kq_index_t qos_index = kn->kn_qos_index; - kqlock(kq); - if (new_qos_index != knote_get_req_index(kn) || - new_override_index != knote_get_qos_override_index(kn) || - override_is_sync != kn->kn_qos_override_is_sync) { - if (kn->kn_status & KN_QUEUED) { - knote_dequeue(kn); - knote_set_qos_index(kn, new_qos_index); - knote_set_qos_override_index(kn, new_override_index, override_is_sync); - knote_enqueue(kn); - knote_wakeup(kn); - } else { - knote_set_qos_index(kn, new_qos_index); - knote_set_qos_override_index(kn, new_override_index, override_is_sync); - } - } - kqunlock(kq); + if (kqu.kq->kq_state & KQ_WORKLOOP) { + assert(qos_index < KQWL_NBUCKETS); + } else if (kqu.kq->kq_state & KQ_WORKQ) { + assert(qos_index < KQWQ_NBUCKETS); + } else { + assert(qos_index == QOS_INDEX_KQFILE); } + static_assert(offsetof(struct kqueue, kq_queue) == sizeof(struct kqueue), + "struct kqueue::kq_queue must be exactly at the end"); + return &kqu.kq->kq_queue[qos_index]; } -void -knote_adjust_sync_qos(struct knote *kn, kq_index_t sync_qos, boolean_t lock_kq) +static void +knote_enqueue(kqueue_t kqu, struct knote *kn, kn_status_t wakeup_mask) { - struct kqueue *kq = knote_get_kq(kn); - kq_index_t old_sync_override; - kq_index_t qos_index = knote_get_qos_index(kn); - uint32_t flags = 0; + kqlock_held(kqu); - /* Tracking only happens for UI qos */ - if (sync_qos != THREAD_QOS_USER_INTERACTIVE && - sync_qos != THREAD_QOS_UNSPECIFIED) { + if ((kn->kn_status & (KN_ACTIVE | KN_STAYACTIVE)) == 0) { return; } - if (lock_kq) - kqlock(kq); - - if (kq->kq_state & KQ_WORKLOOP) { - struct kqworkloop *kqwl = (struct kqworkloop *)kq; - - old_sync_override = knote_get_sync_qos_override_index(kn); - if (old_sync_override != sync_qos) { - kn->kn_qos_sync_override = sync_qos; + if (kn->kn_status & (KN_DISABLED | KN_SUPPRESSED | KN_DROPPING)) { + return; + } - /* update sync ipc counters for suppressed knotes */ - if ((kn->kn_status & KN_SUPPRESSED) == KN_SUPPRESSED) { - flags = flags | KQWL_UO_UPDATE_SUPPRESS_SYNC_COUNTERS; + if ((kn->kn_status & KN_QUEUED) == 0) { + struct kqtailq *queue = knote_get_tailq(kqu, kn); - /* Do not recalculate kqwl override, it would be done later */ - flags = flags | KQWL_UO_UPDATE_OVERRIDE_LAZY; + TAILQ_INSERT_TAIL(queue, kn, kn_tqe); + kn->kn_status |= KN_QUEUED; + kqu.kq->kq_count++; + } else if ((kn->kn_status & KN_STAYACTIVE) == 0) { + return; + } - if (sync_qos == THREAD_QOS_USER_INTERACTIVE) { - flags = flags | KQWL_UO_NEW_OVERRIDE_IS_SYNC_UI; - } + if (kn->kn_status & wakeup_mask) { + if (kqu.kq->kq_state & KQ_WORKLOOP) { + kqworkloop_wakeup(kqu.kqwl, kn->kn_qos_index); + } else if (kqu.kq->kq_state & KQ_WORKQ) { + kqworkq_wakeup(kqu.kqwq, kn->kn_qos_index); + } else { + kqfile_wakeup(kqu.kqf, kn->kn_qos_index); + } + } +} - if (old_sync_override == THREAD_QOS_USER_INTERACTIVE) { - flags = flags | KQWL_UO_OLD_OVERRIDE_IS_SYNC_UI; - } +__attribute__((always_inline)) +static inline void +knote_dequeue(kqueue_t kqu, struct knote *kn) +{ + if (kn->kn_status & KN_QUEUED) { + struct kqtailq *queue = knote_get_tailq(kqu, kn); - kqworkloop_update_override(kqwl, qos_index, sync_qos, - flags); - } + // attaching the knote calls knote_reset_priority() without + // the kqlock which is fine, so we can't call kqlock_held() + // if we're not queued. + kqlock_held(kqu); - } + TAILQ_REMOVE(queue, kn, kn_tqe); + kn->kn_status &= ~KN_QUEUED; + kqu.kq->kq_count--; } - if (lock_kq) - kqunlock(kq); } +/* called with kqueue lock held */ static void -knote_wakeup(struct knote *kn) +knote_suppress(kqueue_t kqu, struct knote *kn) { - struct kqueue *kq = knote_get_kq(kn); - kq_index_t qos_index = knote_get_qos_index(kn); + struct kqtailq *suppressq; - kqlock_held(kq); + kqlock_held(kqu); - if (kq->kq_state & KQ_WORKQ) { - /* request a servicing thread */ - struct kqworkq *kqwq = (struct kqworkq *)kq; + assert((kn->kn_status & KN_SUPPRESSED) == 0); + assert(kn->kn_status & KN_QUEUED); - kqworkq_request_help(kqwq, qos_index); + knote_dequeue(kqu, kn); + /* deactivate - so new activations indicate a wakeup */ + kn->kn_status &= ~KN_ACTIVE; + kn->kn_status |= KN_SUPPRESSED; + suppressq = kqueue_get_suppressed_queue(kqu, kn); + TAILQ_INSERT_TAIL(suppressq, kn, kn_tqe); +} - } else if (kq->kq_state & KQ_WORKLOOP) { - /* request a servicing thread */ - struct kqworkloop *kqwl = (struct kqworkloop *)kq; +__attribute__((always_inline)) +static inline void +knote_unsuppress_noqueue(kqueue_t kqu, struct knote *kn) +{ + struct kqtailq *suppressq; - if (kqworkloop_is_processing_on_current_thread(kqwl)) { - /* - * kqworkloop_end_processing() will perform the required QoS - * computations when it unsets the processing mode. - */ - return; - } - kqworkloop_request_help(kqwl, qos_index); - } else { - struct kqfile *kqf = (struct kqfile *)kq; + kqlock_held(kqu); - /* flag wakeups during processing */ - if (kq->kq_state & KQ_PROCESSING) - kq->kq_state |= KQ_WAKEUP; + assert(kn->kn_status & KN_SUPPRESSED); - /* wakeup a thread waiting on this queue */ - if (kq->kq_state & (KQ_SLEEP | KQ_SEL)) { - kq->kq_state &= ~(KQ_SLEEP | KQ_SEL); - waitq_wakeup64_all((struct waitq *)&kq->kq_wqs, - KQ_EVENT, - THREAD_AWAKENED, - WAITQ_ALL_PRIORITIES); - } + kn->kn_status &= ~KN_SUPPRESSED; + suppressq = kqueue_get_suppressed_queue(kqu, kn); + TAILQ_REMOVE(suppressq, kn, kn_tqe); - /* wakeup other kqueues/select sets we're inside */ - KNOTE(&kqf->kqf_sel.si_note, 0); + /* + * If the knote is no longer active, reset its push, + * and resynchronize kn_qos_index with kn_qos_override + * for knotes with a real qos. + */ + if ((kn->kn_status & KN_ACTIVE) == 0 && knote_has_qos(kn)) { + kn->kn_qos_override = _pthread_priority_thread_qos_fast(kn->kn_qos); } + kn->kn_qos_index = kn->kn_qos_override; } -/* - * Called with the kqueue locked - */ +/* called with kqueue lock held */ static void -kqueue_interrupt(struct kqueue *kq) +knote_unsuppress(kqueue_t kqu, struct knote *kn) { - assert((kq->kq_state & KQ_WORKQ) == 0); + if (kn->kn_status & KN_SUPPRESSED) { + knote_unsuppress_noqueue(kqu, kn); - /* wakeup sleeping threads */ - if ((kq->kq_state & (KQ_SLEEP | KQ_SEL)) != 0) { - kq->kq_state &= ~(KQ_SLEEP | KQ_SEL); - (void)waitq_wakeup64_all((struct waitq *)&kq->kq_wqs, - KQ_EVENT, - THREAD_RESTART, - WAITQ_ALL_PRIORITIES); + /* don't wakeup if unsuppressing just a stay-active knote */ + knote_enqueue(kqu, kn, KN_ACTIVE); } +} - /* wakeup threads waiting their turn to process */ - if (kq->kq_state & KQ_PROCWAIT) { - struct kqtailq *suppressq; +__attribute__((always_inline)) +static inline void +knote_mark_active(struct knote *kn) +{ + if ((kn->kn_status & KN_ACTIVE) == 0) { + KDBG_DEBUG(KEV_EVTID(BSD_KEVENT_KNOTE_ACTIVATE), + kn->kn_udata, kn->kn_status | (kn->kn_id << 32), + kn->kn_filtid); + } - assert(kq->kq_state & KQ_PROCESSING); + kn->kn_status |= KN_ACTIVE; +} - kq->kq_state &= ~KQ_PROCWAIT; - suppressq = kqueue_get_suppressed_queue(kq, QOS_INDEX_KQFILE); - (void)waitq_wakeup64_all((struct waitq *)&kq->kq_wqs, - CAST_EVENT64_T(suppressq), - THREAD_RESTART, - WAITQ_ALL_PRIORITIES); +/* called with kqueue lock held */ +static void +knote_activate(kqueue_t kqu, struct knote *kn, int result) +{ + assert(result & FILTER_ACTIVE); + if (result & FILTER_ADJUST_EVENT_QOS_BIT) { + // may dequeue the knote + knote_adjust_qos(kqu.kq, kn, result); } + knote_mark_active(kn); + knote_enqueue(kqu, kn, KN_ACTIVE | KN_STAYACTIVE); } /* - * Called back from waitq code when no threads waiting and the hook was set. - * - * Interrupts are likely disabled and spin locks are held - minimal work - * can be done in this context!!! - * - * JMM - in the future, this will try to determine which knotes match the - * wait queue wakeup and apply these wakeups against those knotes themselves. - * For now, all the events dispatched this way are dispatch-manager handled, - * so hard-code that for now. + * This function applies changes requested by f_attach or f_touch for + * a given filter. It proceeds in a carefully chosen order to help + * every single transition do the minimal amount of work possible. */ -void -waitq_set__CALLING_PREPOST_HOOK__(void *kq_hook, void *knote_hook, int qos) +static void +knote_apply_touch(kqueue_t kqu, struct knote *kn, struct kevent_qos_s *kev, + int result) { -#pragma unused(knote_hook, qos) - - struct kqueue *kq = (struct kqueue *)kq_hook; + kn_status_t wakeup_mask = KN_ACTIVE; - if (kq->kq_state & KQ_WORKQ) { - struct kqworkq *kqwq = (struct kqworkq *)kq; + if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) { + /* + * When a stayactive knote is reenabled, we may have missed wakeups + * while it was disabled, so we need to poll it. To do so, ask + * knote_enqueue() below to reenqueue it. + */ + wakeup_mask |= KN_STAYACTIVE; + kn->kn_status &= ~KN_DISABLED; - kqworkq_request_help(kqwq, KQWQ_QOS_MANAGER); + /* + * it is possible for userland to have knotes registered for a given + * workloop `wl_orig` but really handled on another workloop `wl_new`. + * + * In that case, rearming will happen from the servicer thread of + * `wl_new` which if `wl_orig` is no longer being serviced, would cause + * this knote to stay suppressed forever if we only relied on + * kqworkloop_acknowledge_events to be called by `wl_orig`. + * + * However if we see the KQ_PROCESSING bit on `wl_orig` set, we can't + * unsuppress because that would mess with the processing phase of + * `wl_orig`, however it also means kqworkloop_acknowledge_events() + * will be called. + */ + if (__improbable(kn->kn_status & KN_SUPPRESSED)) { + if ((kqu.kq->kq_state & KQ_PROCESSING) == 0) { + knote_unsuppress_noqueue(kqu, kn); + } + } + } - } else if (kq->kq_state & KQ_WORKLOOP) { - struct kqworkloop *kqwl = (struct kqworkloop *)kq; + if ((result & FILTER_UPDATE_REQ_QOS) && kev->qos && kev->qos != kn->kn_qos) { + // may dequeue the knote + knote_reset_priority(kqu, kn, kev->qos); + } - kqworkloop_request_help(kqwl, KQWL_BUCKET_STAYACTIVE); + /* + * When we unsuppress above, or because of knote_reset_priority(), + * the knote may have been dequeued, we need to restore the invariant + * that if the knote is active it needs to be queued now that + * we're done applying changes. + */ + if (result & FILTER_ACTIVE) { + knote_activate(kqu, kn, result); + } else { + knote_enqueue(kqu, kn, wakeup_mask); } -} -void -klist_init(struct klist *list) -{ - SLIST_INIT(list); + if ((result & FILTER_THREADREQ_NODEFEER) && + act_clear_astkevent(current_thread(), AST_KEVENT_REDRIVE_THREADREQ)) { + workq_kern_threadreq_redrive(kqu.kq->kq_p, WORKQ_THREADREQ_NONE); + } } - /* - * Query/Post each knote in the object's list + * knote_drop - disconnect and drop the knote * - * The object lock protects the list. It is assumed - * that the filter/event routine for the object can - * determine that the object is already locked (via - * the hint) and not deadlock itself. + * Called with the kqueue locked, returns with the kqueue unlocked. * - * The object lock should also hold off pending - * detach/drop operations. But we'll prevent it here - * too (by taking a use reference) - just in case. + * If a knote locking context is passed, it is canceled. + * + * The knote may have already been detached from + * (or not yet attached to) its source object. */ -void -knote(struct klist *list, long hint) +static void +knote_drop(struct kqueue *kq, struct knote *kn, struct knote_lock_ctx *knlc) { - struct knote *kn; - - SLIST_FOREACH(kn, list, kn_selnext) { - struct kqueue *kq = knote_get_kq(kn); + struct proc *p = kq->kq_p; - kqlock(kq); + kqlock_held(kq); - assert(!knoteuse_needs_boost(kn, NULL)); + assert((kn->kn_status & KN_DROPPING) == 0); + if (knlc == NULL) { + assert((kn->kn_status & KN_LOCKED) == 0); + } + kn->kn_status |= KN_DROPPING; - /* If we can get a use reference - deliver event */ - if (kqlock2knoteuse(kq, kn, KNUSE_NONE)) { - int result; + if (kn->kn_status & KN_SUPPRESSED) { + knote_unsuppress_noqueue(kq, kn); + } else { + knote_dequeue(kq, kn); + } + knote_wait_for_post(kq, kn); - /* call the event with only a use count */ - result = knote_fops(kn)->f_event(kn, hint); + knote_fops(kn)->f_detach(kn); - /* if its not going away and triggered */ - if (knoteuse2kqlock(kq, kn, KNUSE_NONE) && result) - knote_activate(kn); - /* kq lock held */ - } - kqunlock(kq); + /* kq may be freed when kq_remove_knote() returns */ + kq_remove_knote(kq, kn, p, knlc); + if (kn->kn_is_fd && ((kn->kn_status & KN_VANISHED) == 0)) { + fp_drop(p, (int)kn->kn_id, kn->kn_fp, 0); } + + knote_free(kn); } -/* - * attach a knote to the specified list. Return true if this is the first entry. - * The list is protected by whatever lock the object it is associated with uses. - */ -int -knote_attach(struct klist *list, struct knote *kn) +void +knote_init(void) { - int ret = SLIST_EMPTY(list); - SLIST_INSERT_HEAD(list, kn, kn_selnext); - return (ret); +#if CONFIG_MEMORYSTATUS + /* Initialize the memorystatus list lock */ + memorystatus_kevent_init(&kq_lck_grp, LCK_ATTR_NULL); +#endif } +SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL); -/* - * detach a knote from the specified list. Return true if that was the last entry. - * The list is protected by whatever lock the object it is associated with uses. - */ -int -knote_detach(struct klist *list, struct knote *kn) +const struct filterops * +knote_fops(struct knote *kn) { - SLIST_REMOVE(list, kn, knote, kn_selnext); - return (SLIST_EMPTY(list)); + return sysfilt_ops[kn->kn_filtid]; } -/* - * knote_vanish - Indicate that the source has vanished - * - * If the knote has requested EV_VANISHED delivery, - * arrange for that. Otherwise, deliver a NOTE_REVOKE - * event for backward compatibility. - * - * The knote is marked as having vanished, but is not - * actually detached from the source in this instance. - * The actual detach is deferred until the knote drop. - * - * Our caller already has the object lock held. Calling - * the detach routine would try to take that lock - * recursively - which likely is not supported. - */ -void -knote_vanish(struct klist *list) +static struct knote * +knote_alloc(void) { - struct knote *kn; - struct knote *kn_next; - - SLIST_FOREACH_SAFE(kn, list, kn_selnext, kn_next) { - struct kqueue *kq = knote_get_kq(kn); - int result; - - kqlock(kq); - - assert(!knoteuse_needs_boost(kn, NULL)); - - if ((kn->kn_status & KN_DROPPING) == 0) { - /* If EV_VANISH supported - prepare to deliver one */ - if (kn->kn_status & KN_REQVANISH) { - kn->kn_status |= KN_VANISHED; - knote_activate(kn); - - } else if (kqlock2knoteuse(kq, kn, KNUSE_NONE)) { - /* call the event with only a use count */ - result = knote_fops(kn)->f_event(kn, NOTE_REVOKE); - - /* if its not going away and triggered */ - if (knoteuse2kqlock(kq, kn, KNUSE_NONE) && result) - knote_activate(kn); - /* lock held again */ - } - } - kqunlock(kq); - } + return zalloc_flags(knote_zone, Z_WAITOK | Z_ZERO); } -/* - * For a given knote, link a provided wait queue directly with the kqueue. - * Wakeups will happen via recursive wait queue support. But nothing will move - * the knote to the active list at wakeup (nothing calls knote()). Instead, - * we permanently enqueue them here. - * - * kqueue and knote references are held by caller. - * waitq locked by caller. - * - * caller provides the wait queue link structure. - */ -int -knote_link_waitq(struct knote *kn, struct waitq *wq, uint64_t *reserved_link) +static void +knote_free(struct knote *kn) { - struct kqueue *kq = knote_get_kq(kn); - kern_return_t kr; + assert((kn->kn_status & (KN_LOCKED | KN_POSTING)) == 0); + zfree(knote_zone, kn); +} - kr = waitq_link(wq, &kq->kq_wqs, WAITQ_ALREADY_LOCKED, reserved_link); - if (kr == KERN_SUCCESS) { - knote_markstayactive(kn); - return (0); - } else { - return (EINVAL); - } +#pragma mark - syscalls: kevent, kevent64, kevent_qos, kevent_id + +kevent_ctx_t +kevent_get_context(thread_t thread) +{ + uthread_t ut = get_bsdthread_info(thread); + return &ut->uu_save.uus_kevent; } -/* - * Unlink the provided wait queue from the kqueue associated with a knote. - * Also remove it from the magic list of directly attached knotes. - * - * Note that the unlink may have already happened from the other side, so - * ignore any failures to unlink and just remove it from the kqueue list. - * - * On success, caller is responsible for the link structure - */ -int -knote_unlink_waitq(struct knote *kn, struct waitq *wq) +static inline bool +kevent_args_requesting_events(unsigned int flags, int nevents) { - struct kqueue *kq = knote_get_kq(kn); - kern_return_t kr; + return !(flags & KEVENT_FLAG_ERROR_EVENTS) && nevents > 0; +} - kr = waitq_unlink(wq, &kq->kq_wqs); - knote_clearstayactive(kn); - return ((kr != KERN_SUCCESS) ? EINVAL : 0); +static inline int +kevent_adjust_flags_for_proc(proc_t p, int flags) +{ + __builtin_assume(p); + return flags | (IS_64BIT_PROCESS(p) ? KEVENT_FLAG_PROC64 : 0); } -/* - * remove all knotes referencing a specified fd +/*! + * @function kevent_get_kqfile * - * Essentially an inlined knote_remove & knote_drop - * when we know for sure that the thing is a file + * @brief + * Lookup a kqfile by fd. * - * Entered with the proc_fd lock already held. - * It returns the same way, but may drop it temporarily. + * @discussion + * Callers: kevent, kevent64, kevent_qos + * + * This is not assumed to be a fastpath (kqfile interfaces are legacy) */ -void -knote_fdclose(struct proc *p, int fd, int force) +OS_NOINLINE +static int +kevent_get_kqfile(struct proc *p, int fd, int flags, + struct fileproc **fpp, struct kqueue **kqp) { - struct klist *list; - struct knote *kn; + int error = 0; + struct kqueue *kq; -restart: - list = &p->p_fd->fd_knlist[fd]; - SLIST_FOREACH(kn, list, kn_link) { - struct kqueue *kq = knote_get_kq(kn); + error = fp_get_ftype(p, fd, DTYPE_KQUEUE, EBADF, fpp); + if (__improbable(error)) { + return error; + } + kq = (struct kqueue *)(*fpp)->f_data; + uint16_t kq_state = os_atomic_load(&kq->kq_state, relaxed); + if (__improbable((kq_state & (KQ_KEV32 | KQ_KEV64 | KQ_KEV_QOS)) == 0)) { kqlock(kq); - - if (kq->kq_p != p) - panic("%s: proc mismatch (kq->kq_p=%p != p=%p)", - __func__, kq->kq_p, p); - - /* - * If the knote supports EV_VANISHED delivery, - * transition it to vanished mode (or skip over - * it if already vanished). - */ - if (!force && (kn->kn_status & KN_REQVANISH)) { - - if ((kn->kn_status & KN_VANISHED) == 0) { - proc_fdunlock(p); - - assert(!knoteuse_needs_boost(kn, NULL)); - - /* get detach reference (also marks vanished) */ - if (kqlock2knotedetach(kq, kn, KNUSE_NONE)) { - /* detach knote and drop fp use reference */ - knote_fops(kn)->f_detach(kn); - if (knote_fops(kn)->f_isfd) - fp_drop(p, kn->kn_id, kn->kn_fp, 0); - - /* activate it if it's still in existence */ - if (knoteuse2kqlock(kq, kn, KNUSE_NONE)) { - knote_activate(kn); - } - kqunlock(kq); - } - proc_fdlock(p); - goto restart; + kq_state = kq->kq_state; + if (!(kq_state & (KQ_KEV32 | KQ_KEV64 | KQ_KEV_QOS))) { + if (flags & KEVENT_FLAG_LEGACY32) { + kq_state |= KQ_KEV32; + } else if (flags & KEVENT_FLAG_LEGACY64) { + kq_state |= KQ_KEV64; } else { - kqunlock(kq); - continue; + kq_state |= KQ_KEV_QOS; } + kq->kq_state = kq_state; } + kqunlock(kq); + } - proc_fdunlock(p); - - /* - * Convert the kq lock to a drop ref. - * If we get it, go ahead and drop it. - * Otherwise, we waited for the blocking - * condition to complete. Either way, - * we dropped the fdlock so start over. - */ - if (kqlock2knotedrop(kq, kn)) { - knote_drop(kn, p); - } - - proc_fdlock(p); - goto restart; + /* + * kqfiles can't be used through the legacy kevent() + * and other interfaces at the same time. + */ + if (__improbable((bool)(flags & KEVENT_FLAG_LEGACY32) != + (bool)(kq_state & KQ_KEV32))) { + fp_drop(p, fd, *fpp, 0); + return EINVAL; } + + *kqp = kq; + return 0; } -/* - * knote_fdfind - lookup a knote in the fd table for process - * - * If the filter is file-based, lookup based on fd index. - * Otherwise use a hash based on the ident. +/*! + * @function kevent_get_kqwq * - * Matching is based on kq, filter, and ident. Optionally, - * it may also be based on the udata field in the kevent - - * allowing multiple event registration for the file object - * per kqueue. + * @brief + * Lookup or create the process kqwq (faspath). * - * fd_knhashlock or fdlock held on entry (and exit) + * @discussion + * Callers: kevent64, kevent_qos */ -static struct knote * -knote_fdfind(struct kqueue *kq, - struct kevent_internal_s *kev, - bool is_fd, - struct proc *p) +OS_ALWAYS_INLINE +static int +kevent_get_kqwq(proc_t p, int flags, int nevents, struct kqueue **kqp) { - struct filedesc *fdp = p->p_fd; - struct klist *list = NULL; - struct knote *kn = NULL; + struct kqworkq *kqwq = p->p_fd->fd_wqkqueue; - /* - * determine where to look for the knote - */ - if (is_fd) { - /* fd-based knotes are linked off the fd table */ - if (kev->ident < (u_int)fdp->fd_knlistsize) { - list = &fdp->fd_knlist[kev->ident]; + if (__improbable(kevent_args_requesting_events(flags, nevents))) { + return EINVAL; + } + if (__improbable(kqwq == NULL)) { + kqwq = kqworkq_alloc(p, flags); + if (__improbable(kqwq == NULL)) { + return ENOMEM; } - } else if (fdp->fd_knhashmask != 0) { - /* hash non-fd knotes here too */ - list = &fdp->fd_knhash[KN_HASH((u_long)kev->ident, fdp->fd_knhashmask)]; } - /* - * scan the selected list looking for a match - */ - if (list != NULL) { - SLIST_FOREACH(kn, list, kn_link) { - if (kq == knote_get_kq(kn) && - kev->ident == kn->kn_id && - kev->filter == kn->kn_filter) { - if (kev->flags & EV_UDATA_SPECIFIC) { - if ((kn->kn_status & KN_UDATA_SPECIFIC) && - kev->udata == kn->kn_udata) { - break; /* matching udata-specific knote */ - } - } else if ((kn->kn_status & KN_UDATA_SPECIFIC) == 0) { - break; /* matching non-udata-specific knote */ - } - } + *kqp = &kqwq->kqwq_kqueue; + return 0; +} + +#pragma mark kevent copyio + +/*! + * @function kevent_get_data_size + * + * @brief + * Copies in the extra data size from user-space. + */ +static int +kevent_get_data_size(int flags, user_addr_t data_avail, user_addr_t data_out, + kevent_ctx_t kectx) +{ + if (!data_avail || !data_out) { + kectx->kec_data_size = 0; + kectx->kec_data_resid = 0; + } else if (flags & KEVENT_FLAG_PROC64) { + user64_size_t usize = 0; + int error = copyin((user_addr_t)data_avail, &usize, sizeof(usize)); + if (__improbable(error)) { + return error; + } + kectx->kec_data_resid = kectx->kec_data_size = (user_size_t)usize; + } else { + user32_size_t usize = 0; + int error = copyin((user_addr_t)data_avail, &usize, sizeof(usize)); + if (__improbable(error)) { + return error; } + kectx->kec_data_avail = data_avail; + kectx->kec_data_resid = kectx->kec_data_size = (user_size_t)usize; } - return kn; + kectx->kec_data_out = data_out; + kectx->kec_data_avail = data_avail; + return 0; } -/* - * kq_add_knote- Add knote to the fd table for process - * while checking for duplicates. - * - * All file-based filters associate a list of knotes by file - * descriptor index. All other filters hash the knote by ident. - * - * May have to grow the table of knote lists to cover the - * file descriptor index presented. - * - * fd_knhashlock and fdlock unheld on entry (and exit). +/*! + * @function kevent_put_data_size * - * Takes a rwlock boost if inserting the knote is successful. + * @brief + * Copies out the residual data size to user-space if any has been used. */ static int -kq_add_knote(struct kqueue *kq, struct knote *kn, - struct kevent_internal_s *kev, - struct proc *p, int *knoteuse_flags) +kevent_put_data_size(unsigned int flags, kevent_ctx_t kectx) { - struct filedesc *fdp = p->p_fd; - struct klist *list = NULL; - int ret = 0; - bool is_fd = knote_fops(kn)->f_isfd; - - if (is_fd) - proc_fdlock(p); - else - knhash_lock(p); - - if (knote_fdfind(kq, kev, is_fd, p) != NULL) { - /* found an existing knote: we can't add this one */ - ret = ERESTART; - goto out_locked; + if (kectx->kec_data_resid == kectx->kec_data_size) { + return 0; + } + if (flags & KEVENT_FLAG_KERNEL) { + *(user_size_t *)(uintptr_t)kectx->kec_data_avail = kectx->kec_data_resid; + return 0; } + if (flags & KEVENT_FLAG_PROC64) { + user64_size_t usize = (user64_size_t)kectx->kec_data_resid; + return copyout(&usize, (user_addr_t)kectx->kec_data_avail, sizeof(usize)); + } else { + user32_size_t usize = (user32_size_t)kectx->kec_data_resid; + return copyout(&usize, (user_addr_t)kectx->kec_data_avail, sizeof(usize)); + } +} - /* knote was not found: add it now */ - if (!is_fd) { - if (fdp->fd_knhashmask == 0) { - u_long size = 0; +/*! + * @function kevent_legacy_copyin + * + * @brief + * Handles the copyin of a kevent/kevent64 event. + */ +static int +kevent_legacy_copyin(user_addr_t *addrp, struct kevent_qos_s *kevp, unsigned int flags) +{ + int error; - list = hashinit(CONFIG_KN_HASHSIZE, M_KQUEUE, - &size); - if (list == NULL) { - ret = ENOMEM; - goto out_locked; - } + assert((flags & (KEVENT_FLAG_LEGACY32 | KEVENT_FLAG_LEGACY64)) != 0); - fdp->fd_knhash = list; - fdp->fd_knhashmask = size; - } + if (flags & KEVENT_FLAG_LEGACY64) { + struct kevent64_s kev64; - list = &fdp->fd_knhash[KN_HASH(kn->kn_id, fdp->fd_knhashmask)]; - SLIST_INSERT_HEAD(list, kn, kn_link); - ret = 0; - goto out_locked; + error = copyin(*addrp, (caddr_t)&kev64, sizeof(kev64)); + if (__improbable(error)) { + return error; + } + *addrp += sizeof(kev64); + *kevp = (struct kevent_qos_s){ + .ident = kev64.ident, + .filter = kev64.filter, + /* Make sure user doesn't pass in any system flags */ + .flags = kev64.flags & ~EV_SYSFLAGS, + .udata = kev64.udata, + .fflags = kev64.fflags, + .data = kev64.data, + .ext[0] = kev64.ext[0], + .ext[1] = kev64.ext[1], + }; + } else if (flags & KEVENT_FLAG_PROC64) { + struct user64_kevent kev64; + error = copyin(*addrp, (caddr_t)&kev64, sizeof(kev64)); + if (__improbable(error)) { + return error; + } + *addrp += sizeof(kev64); + *kevp = (struct kevent_qos_s){ + .ident = kev64.ident, + .filter = kev64.filter, + /* Make sure user doesn't pass in any system flags */ + .flags = kev64.flags & ~EV_SYSFLAGS, + .udata = kev64.udata, + .fflags = kev64.fflags, + .data = kev64.data, + }; } else { - /* knote is fd based */ - - if ((u_int)fdp->fd_knlistsize <= kn->kn_id) { - u_int size = 0; - - if (kn->kn_id >= (uint64_t)p->p_rlimit[RLIMIT_NOFILE].rlim_cur - || kn->kn_id >= (uint64_t)maxfiles) { - ret = EINVAL; - goto out_locked; - } - /* have to grow the fd_knlist */ - size = fdp->fd_knlistsize; - while (size <= kn->kn_id) - size += KQEXTENT; - - if (size >= (UINT_MAX/sizeof(struct klist *))) { - ret = EINVAL; - goto out_locked; - } - - MALLOC(list, struct klist *, - size * sizeof(struct klist *), M_KQUEUE, M_WAITOK); - if (list == NULL) { - ret = ENOMEM; - goto out_locked; - } + struct user32_kevent kev32; - bcopy((caddr_t)fdp->fd_knlist, (caddr_t)list, - fdp->fd_knlistsize * sizeof(struct klist *)); - bzero((caddr_t)list + - fdp->fd_knlistsize * sizeof(struct klist *), - (size - fdp->fd_knlistsize) * sizeof(struct klist *)); - FREE(fdp->fd_knlist, M_KQUEUE); - fdp->fd_knlist = list; - fdp->fd_knlistsize = size; + error = copyin(*addrp, (caddr_t)&kev32, sizeof(kev32)); + if (__improbable(error)) { + return error; } + *addrp += sizeof(kev32); + *kevp = (struct kevent_qos_s){ + .ident = (uintptr_t)kev32.ident, + .filter = kev32.filter, + /* Make sure user doesn't pass in any system flags */ + .flags = kev32.flags & ~EV_SYSFLAGS, + .udata = CAST_USER_ADDR_T(kev32.udata), + .fflags = kev32.fflags, + .data = (intptr_t)kev32.data, + }; + } - list = &fdp->fd_knlist[kn->kn_id]; - SLIST_INSERT_HEAD(list, kn, kn_link); - ret = 0; - goto out_locked; + return 0; +} +/*! + * @function kevent_modern_copyin + * + * @brief + * Handles the copyin of a kevent_qos/kevent_id event. + */ +static int +kevent_modern_copyin(user_addr_t *addrp, struct kevent_qos_s *kevp) +{ + int error = copyin(*addrp, (caddr_t)kevp, sizeof(struct kevent_qos_s)); + if (__probable(!error)) { + /* Make sure user doesn't pass in any system flags */ + *addrp += sizeof(struct kevent_qos_s); + kevp->flags &= ~EV_SYSFLAGS; } + return error; +} -out_locked: - if (ret == 0 && knoteuse_needs_boost(kn, kev)) { - set_thread_rwlock_boost(); - *knoteuse_flags = KNUSE_BOOST; +/*! + * @function kevent_legacy_copyout + * + * @brief + * Handles the copyout of a kevent/kevent64 event. + */ +static int +kevent_legacy_copyout(struct kevent_qos_s *kevp, user_addr_t *addrp, unsigned int flags) +{ + int advance; + int error; + + assert((flags & (KEVENT_FLAG_LEGACY32 | KEVENT_FLAG_LEGACY64)) != 0); + + /* + * fully initialize the differnt output event structure + * types from the internal kevent (and some universal + * defaults for fields not represented in the internal + * form). + * + * Note: these structures have no padding hence the C99 + * initializers below do not leak kernel info. + */ + if (flags & KEVENT_FLAG_LEGACY64) { + struct kevent64_s kev64 = { + .ident = kevp->ident, + .filter = kevp->filter, + .flags = kevp->flags, + .fflags = kevp->fflags, + .data = (int64_t)kevp->data, + .udata = kevp->udata, + .ext[0] = kevp->ext[0], + .ext[1] = kevp->ext[1], + }; + advance = sizeof(struct kevent64_s); + error = copyout((caddr_t)&kev64, *addrp, advance); + } else if (flags & KEVENT_FLAG_PROC64) { + /* + * deal with the special case of a user-supplied + * value of (uintptr_t)-1. + */ + uint64_t ident = (kevp->ident == (uintptr_t)-1) ? + (uint64_t)-1LL : (uint64_t)kevp->ident; + struct user64_kevent kev64 = { + .ident = ident, + .filter = kevp->filter, + .flags = kevp->flags, + .fflags = kevp->fflags, + .data = (int64_t) kevp->data, + .udata = (user_addr_t) kevp->udata, + }; + advance = sizeof(kev64); + error = copyout((caddr_t)&kev64, *addrp, advance); } else { - *knoteuse_flags = KNUSE_NONE; + struct user32_kevent kev32 = { + .ident = (uint32_t)kevp->ident, + .filter = kevp->filter, + .flags = kevp->flags, + .fflags = kevp->fflags, + .data = (int32_t)kevp->data, + .udata = (uint32_t)kevp->udata, + }; + advance = sizeof(kev32); + error = copyout((caddr_t)&kev32, *addrp, advance); } - if (is_fd) - proc_fdunlock(p); - else - knhash_unlock(p); + if (__probable(!error)) { + *addrp += advance; + } + return error; +} - return ret; +/*! + * @function kevent_modern_copyout + * + * @brief + * Handles the copyout of a kevent_qos/kevent_id event. + */ +OS_ALWAYS_INLINE +static inline int +kevent_modern_copyout(struct kevent_qos_s *kevp, user_addr_t *addrp) +{ + int error = copyout((caddr_t)kevp, *addrp, sizeof(struct kevent_qos_s)); + if (__probable(!error)) { + *addrp += sizeof(struct kevent_qos_s); + } + return error; } -/* - * kq_remove_knote - remove a knote from the fd table for process - * and copy kn_status an kq_state while holding kqlock and - * fd table locks. +#pragma mark kevent core implementation + +/*! + * @function kevent_callback_inline * - * If the filter is file-based, remove based on fd index. - * Otherwise remove from the hash based on the ident. + * @brief + * Callback for each individual event * - * fd_knhashlock and fdlock unheld on entry (and exit). + * @discussion + * This is meant to be inlined in kevent_modern_callback and + * kevent_legacy_callback. */ -static void -kq_remove_knote(struct kqueue *kq, struct knote *kn, struct proc *p, - kn_status_t *kn_status, uint16_t *kq_state) +OS_ALWAYS_INLINE +static inline int +kevent_callback_inline(struct kevent_qos_s *kevp, kevent_ctx_t kectx, bool legacy) { - struct filedesc *fdp = p->p_fd; - struct klist *list = NULL; - bool is_fd; - - is_fd = knote_fops(kn)->f_isfd; + int error; - if (is_fd) - proc_fdlock(p); - else - knhash_lock(p); + assert(kectx->kec_process_noutputs < kectx->kec_process_nevents); - if (is_fd) { - assert ((u_int)fdp->fd_knlistsize > kn->kn_id); - list = &fdp->fd_knlist[kn->kn_id]; + /* + * Copy out the appropriate amount of event data for this user. + */ + if (legacy) { + error = kevent_legacy_copyout(kevp, &kectx->kec_process_eventlist, + kectx->kec_process_flags); } else { - list = &fdp->fd_knhash[KN_HASH(kn->kn_id, fdp->fd_knhashmask)]; + error = kevent_modern_copyout(kevp, &kectx->kec_process_eventlist); } - SLIST_REMOVE(list, kn, knote, kn_link); - - kqlock(kq); - *kn_status = kn->kn_status; - *kq_state = kq->kq_state; - kqunlock(kq); - if (is_fd) - proc_fdunlock(p); - else - knhash_unlock(p); + /* + * If there isn't space for additional events, return + * a harmless error to stop the processing here + */ + if (error == 0 && ++kectx->kec_process_noutputs == kectx->kec_process_nevents) { + error = EWOULDBLOCK; + } + return error; } -/* - * kq_find_knote_and_kq_lock - lookup a knote in the fd table for process - * and, if the knote is found, acquires the kqlock while holding the fd table lock/spinlock. +/*! + * @function kevent_modern_callback * - * fd_knhashlock or fdlock unheld on entry (and exit) + * @brief + * Callback for each individual modern event. + * + * @discussion + * This callback handles kevent_qos/kevent_id events. */ - -static struct knote * -kq_find_knote_and_kq_lock(struct kqueue *kq, - struct kevent_internal_s *kev, - bool is_fd, - struct proc *p) +static int +kevent_modern_callback(struct kevent_qos_s *kevp, kevent_ctx_t kectx) { - struct knote * ret; + return kevent_callback_inline(kevp, kectx, /*legacy*/ false); +} - if (is_fd) - proc_fdlock(p); - else - knhash_lock(p); +/*! + * @function kevent_legacy_callback + * + * @brief + * Callback for each individual legacy event. + * + * @discussion + * This callback handles kevent/kevent64 events. + */ +static int +kevent_legacy_callback(struct kevent_qos_s *kevp, kevent_ctx_t kectx) +{ + return kevent_callback_inline(kevp, kectx, /*legacy*/ true); +} - ret = knote_fdfind(kq, kev, is_fd, p); +/*! + * @function kevent_cleanup + * + * @brief + * Handles the cleanup returning from a kevent call. + * + * @discussion + * kevent entry points will take a reference on workloops, + * and a usecount on the fileglob of kqfiles. + * + * This function undoes this on the exit paths of kevents. + * + * @returns + * The error to return to userspace. + */ +static int +kevent_cleanup(kqueue_t kqu, int flags, int error, kevent_ctx_t kectx) +{ + // poll should not call any codepath leading to this + assert((flags & KEVENT_FLAG_POLL) == 0); - if (ret) { - kqlock(kq); + if (flags & KEVENT_FLAG_WORKLOOP) { + kqworkloop_release(kqu.kqwl); + } else if (flags & KEVENT_FLAG_WORKQ) { + /* nothing held */ + } else { + fp_drop(kqu.kqf->kqf_p, kectx->kec_fd, kectx->kec_fp, 0); } - if (is_fd) - proc_fdunlock(p); - else - knhash_unlock(p); + /* don't restart after signals... */ + if (error == ERESTART) { + error = EINTR; + } else if (error == 0) { + /* don't abandon other output just because of residual copyout failures */ + (void)kevent_put_data_size(flags, kectx); + } - return ret; + if (flags & KEVENT_FLAG_PARKING) { + thread_t th = current_thread(); + struct uthread *uth = get_bsdthread_info(th); + if (uth->uu_kqr_bound) { + thread_unfreeze_base_pri(th); + } + } + return error; } -/* - * knote_drop - disconnect and drop the knote + +/*! + * @function kqueue_process * - * Called with the kqueue unlocked and holding a - * "drop reference" on the knote in question. - * This reference is most often aquired thru a call - * to kqlock2knotedrop(). But it can also be acquired - * through stealing a drop reference via a call to - * knoteuse2knotedrop() or during the initial attach - * of the knote. + * @brief + * Process the triggered events in a kqueue. * - * The knote may have already been detached from - * (or not yet attached to) its source object. + * @discussion + * Walk the queued knotes and validate that they are really still triggered + * events by calling the filter routines (if necessary). + * + * For each event that is still considered triggered, invoke the callback + * routine provided. + * + * caller holds a reference on the kqueue. + * kqueue locked on entry and exit - but may be dropped + * kqueue list locked (held for duration of call) + * + * This is only called by kqueue_scan() so that the compiler can inline it. + * + * @returns + * - 0: no event was returned, no other error occured + * - EBADF: the kqueue is being destroyed (KQ_DRAIN is set) + * - EWOULDBLOCK: (not an error) events have been found and we should return + * - EFAULT: copyout failed + * - filter specific errors */ -static void -knote_drop(struct knote *kn, __unused struct proc *ctxp) +static int +kqueue_process(kqueue_t kqu, int flags, kevent_ctx_t kectx, + kevent_callback_t callback) { - struct kqueue *kq = knote_get_kq(kn); - struct proc *p = kq->kq_p; - kn_status_t kn_status; - uint16_t kq_state; + workq_threadreq_t kqr = current_uthread()->uu_kqr_bound; + struct knote *kn; + int error = 0, rc = 0; + struct kqtailq *base_queue, *queue; +#if DEBUG || DEVELOPMENT + int retries = 64; +#endif + uint16_t kq_type = (kqu.kq->kq_state & (KQ_WORKQ | KQ_WORKLOOP)); - /* If we are attached, disconnect from the source first */ - if (kn->kn_status & KN_ATTACHED) { - knote_fops(kn)->f_detach(kn); + if (kq_type & KQ_WORKQ) { + rc = kqworkq_begin_processing(kqu.kqwq, kqr, flags); + } else if (kq_type & KQ_WORKLOOP) { + rc = kqworkloop_begin_processing(kqu.kqwl, flags); + } else { +kqfile_retry: + rc = kqfile_begin_processing(kqu.kqf); + if (rc == EBADF) { + return EBADF; + } } - /* Remove the source from the appropriate hash */ - kq_remove_knote(kq, kn, p, &kn_status, &kq_state); + if (rc == -1) { + /* Nothing to process */ + return 0; + } /* - * If a kqueue_dealloc is happening in parallel for the kq - * pointed by the knote the kq could be aready deallocated - * at this point. - * Do not access the kq after the kq_remove_knote if it is - * not a KQ_DYNAMIC. + * loop through the enqueued knotes associated with this request, + * processing each one. Each request may have several queues + * of knotes to process (depending on the type of kqueue) so we + * have to loop through all the queues as long as we have additional + * space. */ - /* determine if anyone needs to know about the drop */ - assert((kn_status & (KN_DROPPING | KN_SUPPRESSED | KN_QUEUED)) == KN_DROPPING); +process_again: + if (kq_type & KQ_WORKQ) { + base_queue = queue = &kqu.kqwq->kqwq_queue[kqr->tr_kq_qos_index]; + } else if (kq_type & KQ_WORKLOOP) { + base_queue = &kqu.kqwl->kqwl_queue[0]; + queue = &kqu.kqwl->kqwl_queue[KQWL_NBUCKETS - 1]; + } else { + base_queue = queue = &kqu.kqf->kqf_queue; + } + + do { + while ((kn = TAILQ_FIRST(queue)) != NULL) { + error = knote_process(kn, kectx, callback); + if (error == EJUSTRETURN) { + error = 0; + } else if (__improbable(error)) { + /* error is EWOULDBLOCK when the out event array is full */ + goto stop_processing; + } + } + } while (queue-- > base_queue); + + if (kectx->kec_process_noutputs) { + /* callers will transform this into no error */ + error = EWOULDBLOCK; + } +stop_processing: /* - * If KN_USEWAIT is set, some other thread was trying to drop the kn. - * Or it was in kqueue_dealloc, so the kqueue_dealloc did not happen - * because that thread was waiting on this wake, or it was a drop happening - * because of a kevent_register that takes a reference on the kq, and therefore - * the kq cannot be deallocated in parallel. + * If KEVENT_FLAG_PARKING is set, and no kevents have been returned, + * we want to unbind the kqrequest from the thread. * - * It is safe to access kq->kq_wqs if needswakeup is set. + * However, because the kq locks are dropped several times during process, + * new knotes may have fired again, in which case, we want to fail the end + * processing and process again, until it converges. + * + * If we have an error or returned events, end processing never fails. */ - if (kn_status & KN_USEWAIT) - waitq_wakeup64_all((struct waitq *)&kq->kq_wqs, - CAST_EVENT64_T(&kn->kn_status), - THREAD_RESTART, - WAITQ_ALL_PRIORITIES); + if (error) { + flags &= ~KEVENT_FLAG_PARKING; + } + if (kq_type & KQ_WORKQ) { + rc = kqworkq_end_processing(kqu.kqwq, kqr, flags); + } else if (kq_type & KQ_WORKLOOP) { + rc = kqworkloop_end_processing(kqu.kqwl, KQ_PROCESSING, flags); + } else { + rc = kqfile_end_processing(kqu.kqf); + } - if (knote_fops(kn)->f_isfd && ((kn->kn_status & KN_VANISHED) == 0)) - fp_drop(p, kn->kn_id, kn->kn_fp, 0); + if (__probable(error)) { + return error; + } - knote_free(kn); + if (__probable(rc >= 0)) { + assert(rc == 0 || rc == EBADF); + return rc; + } - /* - * release reference on dynamic kq (and free if last). - * Will only be last if this is from fdfree, etc... - * because otherwise processing thread has reference. - */ - if (kq_state & KQ_DYNAMIC) - kqueue_release_last(p, kq); +#if DEBUG || DEVELOPMENT + if (retries-- == 0) { + panic("kevent: way too many knote_process retries, kq: %p (0x%04x)", + kqu.kq, kqu.kq->kq_state); + } +#endif + if (kq_type & (KQ_WORKQ | KQ_WORKLOOP)) { + assert(flags & KEVENT_FLAG_PARKING); + goto process_again; + } else { + goto kqfile_retry; + } } -/* called with kqueue lock held */ +/*! + * @function kqueue_scan_continue + * + * @brief + * The continuation used by kqueue_scan for kevent entry points. + * + * @discussion + * Assumes we inherit a use/ref count on the kq or its fileglob. + * + * This is called by kqueue_scan if neither KEVENT_FLAG_POLL nor + * KEVENT_FLAG_KERNEL was set, and the caller had to wait. + */ +OS_NORETURN OS_NOINLINE static void -knote_activate(struct knote *kn) +kqueue_scan_continue(void *data, wait_result_t wait_result) { - if (kn->kn_status & KN_ACTIVE) - return; + uthread_t ut = current_uthread(); + kevent_ctx_t kectx = &ut->uu_save.uus_kevent; + int error = 0, flags = kectx->kec_process_flags; + struct kqueue *kq = data; - KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KNOTE_ACTIVATE), - kn->kn_udata, kn->kn_status | (kn->kn_id << 32), - kn->kn_filtid); + /* + * only kevent variants call in here, so we know the callback is + * kevent_legacy_callback or kevent_modern_callback. + */ + assert((flags & (KEVENT_FLAG_POLL | KEVENT_FLAG_KERNEL)) == 0); - kn->kn_status |= KN_ACTIVE; - if (knote_enqueue(kn)) - knote_wakeup(kn); -} + switch (wait_result) { + case THREAD_AWAKENED: + if (__improbable(flags & (KEVENT_FLAG_LEGACY32 | KEVENT_FLAG_LEGACY64))) { + error = kqueue_scan(kq, flags, kectx, kevent_legacy_callback); + } else { + error = kqueue_scan(kq, flags, kectx, kevent_modern_callback); + } + break; + case THREAD_TIMED_OUT: + error = 0; + break; + case THREAD_INTERRUPTED: + error = EINTR; + break; + case THREAD_RESTART: + error = EBADF; + break; + default: + panic("%s: - invalid wait_result (%d)", __func__, wait_result); + } -/* called with kqueue lock held */ -static void -knote_deactivate(struct knote *kn) -{ - kn->kn_status &= ~KN_ACTIVE; - if ((kn->kn_status & KN_STAYACTIVE) == 0) - knote_dequeue(kn); + + error = kevent_cleanup(kq, flags, error, kectx); + *(int32_t *)&ut->uu_rval = kectx->kec_process_noutputs; + unix_syscall_return(error); } -/* called with kqueue lock held */ -static void -knote_enable(struct knote *kn) +/*! + * @function kqueue_scan + * + * @brief + * Scan and wait for events in a kqueue (used by poll & kevent). + * + * @discussion + * Process the triggered events in a kqueue. + * + * If there are no events triggered arrange to wait for them: + * - unless KEVENT_FLAG_IMMEDIATE is set in kectx->kec_process_flags + * - possibly until kectx->kec_deadline expires + * + * When it waits, and that neither KEVENT_FLAG_POLL nor KEVENT_FLAG_KERNEL + * are set, then it will wait in the kqueue_scan_continue continuation. + * + * poll() will block in place, and KEVENT_FLAG_KERNEL calls + * all pass KEVENT_FLAG_IMMEDIATE and will not wait. + * + * @param kq + * The kqueue being scanned. + * + * @param flags + * The KEVENT_FLAG_* flags for this call. + * + * @param kectx + * The context used for this scan. + * The uthread_t::uu_save.uus_kevent storage is used for this purpose. + * + * @param callback + * The callback to be called on events sucessfully processed. + * (Either kevent_legacy_callback, kevent_modern_callback or poll_callback) + */ +int +kqueue_scan(struct kqueue *kq, int flags, kevent_ctx_t kectx, + kevent_callback_t callback) { - if ((kn->kn_status & KN_DISABLED) == 0) - return; - - kn->kn_status &= ~KN_DISABLED; + int error; - if (kn->kn_status & KN_SUPPRESSED) { - /* Clear the sync qos on the knote */ - knote_adjust_sync_qos(kn, THREAD_QOS_UNSPECIFIED, FALSE); + for (;;) { + kqlock(kq); + error = kqueue_process(kq, flags, kectx, callback); /* - * it is possible for userland to have knotes registered for a given - * workloop `wl_orig` but really handled on another workloop `wl_new`. - * - * In that case, rearming will happen from the servicer thread of - * `wl_new` which if `wl_orig` is no longer being serviced, would cause - * this knote to stay suppressed forever if we only relied on - * kqworkloop_acknowledge_events to be called by `wl_orig`. - * - * However if we see the KQ_PROCESSING bit on `wl_orig` set, we can't - * unsuppress because that would mess with the processing phase of - * `wl_orig`, however it also means kqworkloop_acknowledge_events() - * will be called. + * If we got an error, events returned (EWOULDBLOCK) + * or blocking was disallowed (KEVENT_FLAG_IMMEDIATE), + * just return. */ - struct kqueue *kq = knote_get_kq(kn); - if ((kq->kq_state & KQ_PROCESSING) == 0) { - knote_unsuppress(kn); + if (__probable(error || (flags & KEVENT_FLAG_IMMEDIATE))) { + kqunlock(kq); + return error == EWOULDBLOCK ? 0 : error; + } + + waitq_assert_wait64_leeway((struct waitq *)&kq->kq_wqs, + KQ_EVENT, THREAD_ABORTSAFE, TIMEOUT_URGENCY_USER_NORMAL, + kectx->kec_deadline, TIMEOUT_NO_LEEWAY); + kq->kq_state |= KQ_SLEEP; + + kqunlock(kq); + + if (__probable((flags & (KEVENT_FLAG_POLL | KEVENT_FLAG_KERNEL)) == 0)) { + thread_block_parameter(kqueue_scan_continue, kq); + __builtin_unreachable(); + } + + wait_result_t wr = thread_block(THREAD_CONTINUE_NULL); + switch (wr) { + case THREAD_AWAKENED: + break; + case THREAD_TIMED_OUT: + return 0; + case THREAD_INTERRUPTED: + return EINTR; + case THREAD_RESTART: + return EBADF; + default: + panic("%s: - bad wait_result (%d)", __func__, wr); } - } else if (knote_enqueue(kn)) { - knote_wakeup(kn); } } -/* called with kqueue lock held */ -static void -knote_disable(struct knote *kn) +/*! + * @function kevent_internal + * + * @brief + * Common kevent code. + * + * @discussion + * Needs to be inlined to specialize for legacy or modern and + * eliminate dead code. + * + * This is the core logic of kevent entry points, that will: + * - register kevents + * - optionally scan the kqueue for events + * + * The caller is giving kevent_internal a reference on the kqueue + * or its fileproc that needs to be cleaned up by kevent_cleanup(). + */ +OS_ALWAYS_INLINE +static inline int +kevent_internal(kqueue_t kqu, + user_addr_t changelist, int nchanges, + user_addr_t ueventlist, int nevents, + int flags, kevent_ctx_t kectx, int32_t *retval, + bool legacy) { - if (kn->kn_status & KN_DISABLED) - return; + int error = 0, noutputs = 0, register_rc; - kn->kn_status |= KN_DISABLED; - knote_dequeue(kn); -} + /* only bound threads can receive events on workloops */ + if (!legacy && (flags & KEVENT_FLAG_WORKLOOP)) { +#if CONFIG_WORKLOOP_DEBUG + UU_KEVENT_HISTORY_WRITE_ENTRY(current_uthread(), { + .uu_kqid = kqu.kqwl->kqwl_dynamicid, + .uu_kq = error ? NULL : kqu.kq, + .uu_error = error, + .uu_nchanges = nchanges, + .uu_nevents = nevents, + .uu_flags = flags, + }); +#endif // CONFIG_WORKLOOP_DEBUG -/* called with kqueue lock held */ -static void -knote_suppress(struct knote *kn) -{ - struct kqtailq *suppressq; - struct kqueue *kq = knote_get_kq(kn); + if (flags & KEVENT_FLAG_KERNEL) { + /* see kevent_workq_internal */ + error = copyout(&kqu.kqwl->kqwl_dynamicid, + ueventlist - sizeof(kqueue_id_t), sizeof(kqueue_id_t)); + kectx->kec_data_resid -= sizeof(kqueue_id_t); + if (__improbable(error)) { + goto out; + } + } - kqlock_held(kq); + if (kevent_args_requesting_events(flags, nevents)) { + /* + * Disable the R2K notification while doing a register, if the + * caller wants events too, we don't want the AST to be set if we + * will process these events soon. + */ + kqlock(kqu); + kqu.kq->kq_state &= ~KQ_R2K_ARMED; + kqunlock(kqu); + flags |= KEVENT_FLAG_NEEDS_END_PROCESSING; + } + } - if (kn->kn_status & KN_SUPPRESSED) - return; + /* register all the change requests the user provided... */ + while (nchanges > 0 && error == 0) { + struct kevent_qos_s kev; + struct knote *kn = NULL; - knote_dequeue(kn); - kn->kn_status |= KN_SUPPRESSED; - suppressq = kqueue_get_suppressed_queue(kq, knote_get_qos_index(kn)); - TAILQ_INSERT_TAIL(suppressq, kn, kn_tqe); + if (legacy) { + error = kevent_legacy_copyin(&changelist, &kev, flags); + } else { + error = kevent_modern_copyin(&changelist, &kev); + } + if (error) { + break; + } - if ((kq->kq_state & KQ_WORKLOOP) && - knote_get_qos_override_index(kn) == THREAD_QOS_USER_INTERACTIVE && - kn->kn_qos_override_is_sync) { - struct kqworkloop *kqwl = (struct kqworkloop *)kq; - /* update the sync qos override counter for suppressed knotes */ - kqworkloop_update_override(kqwl, knote_get_qos_index(kn), - knote_get_qos_override_index(kn), - (KQWL_UO_UPDATE_SUPPRESS_SYNC_COUNTERS | KQWL_UO_NEW_OVERRIDE_IS_SYNC_UI)); + register_rc = kevent_register(kqu.kq, &kev, &kn); + if (__improbable(!legacy && (register_rc & FILTER_REGISTER_WAIT))) { + thread_t thread = current_thread(); + + kqlock_held(kqu); + + if (act_clear_astkevent(thread, AST_KEVENT_REDRIVE_THREADREQ)) { + workq_kern_threadreq_redrive(kqu.kq->kq_p, WORKQ_THREADREQ_NONE); + } + + // f_post_register_wait is meant to call a continuation and not to + // return, which is why we don't support FILTER_REGISTER_WAIT if + // KEVENT_FLAG_ERROR_EVENTS is not passed, or if the event that + // waits isn't the last. + // + // It is implementable, but not used by any userspace code at the + // moment, so for now return ENOTSUP if someone tries to do it. + if (nchanges == 1 && noutputs < nevents && + (flags & KEVENT_FLAG_KERNEL) == 0 && + (flags & KEVENT_FLAG_PARKING) == 0 && + (flags & KEVENT_FLAG_ERROR_EVENTS) && + (flags & KEVENT_FLAG_WORKLOOP)) { + uthread_t ut = get_bsdthread_info(thread); + + /* + * store the continuation/completion data in the uthread + * + * Note: the kectx aliases with this, + * and is destroyed in the process. + */ + ut->uu_save.uus_kevent_register = (struct _kevent_register){ + .kev = kev, + .kqwl = kqu.kqwl, + .eventout = noutputs, + .ueventlist = ueventlist, + }; + knote_fops(kn)->f_post_register_wait(ut, kn, + &ut->uu_save.uus_kevent_register); + __builtin_unreachable(); + } + kqunlock(kqu); + + kev.flags |= EV_ERROR; + kev.data = ENOTSUP; + } else { + assert((register_rc & FILTER_REGISTER_WAIT) == 0); + } + + // keep in sync with kevent_register_wait_return() + if (noutputs < nevents && (kev.flags & (EV_ERROR | EV_RECEIPT))) { + if ((kev.flags & EV_ERROR) == 0) { + kev.flags |= EV_ERROR; + kev.data = 0; + } + if (legacy) { + error = kevent_legacy_copyout(&kev, &ueventlist, flags); + } else { + error = kevent_modern_copyout(&kev, &ueventlist); + } + if (error == 0) { + noutputs++; + } + } else if (kev.flags & EV_ERROR) { + error = (int)kev.data; + } + nchanges--; } -} -/* called with kqueue lock held */ -static void -knote_unsuppress(struct knote *kn) -{ - struct kqtailq *suppressq; - struct kqueue *kq = knote_get_kq(kn); + if ((flags & KEVENT_FLAG_ERROR_EVENTS) == 0 && + nevents > 0 && noutputs == 0 && error == 0) { + kectx->kec_process_flags = flags; + kectx->kec_process_nevents = nevents; + kectx->kec_process_noutputs = 0; + kectx->kec_process_eventlist = ueventlist; + + if (legacy) { + error = kqueue_scan(kqu.kq, flags, kectx, kevent_legacy_callback); + } else { + error = kqueue_scan(kqu.kq, flags, kectx, kevent_modern_callback); + } + + noutputs = kectx->kec_process_noutputs; + } else if (!legacy && (flags & KEVENT_FLAG_NEEDS_END_PROCESSING)) { + /* + * If we didn't through kqworkloop_end_processing(), + * we need to do it here. + * + * kqueue_scan will call kqworkloop_end_processing(), + * so we only need to do it if we didn't scan. + */ + kqlock(kqu); + kqworkloop_end_processing(kqu.kqwl, 0, 0); + kqunlock(kqu); + } - kqlock_held(kq); + *retval = noutputs; +out: + return kevent_cleanup(kqu.kq, flags, error, kectx); +} - if ((kn->kn_status & KN_SUPPRESSED) == 0) - return; +#pragma mark modern syscalls: kevent_qos, kevent_id, kevent_workq_internal - /* Clear the sync qos on the knote */ - knote_adjust_sync_qos(kn, THREAD_QOS_UNSPECIFIED, FALSE); +/*! + * @function kevent_modern_internal + * + * @brief + * The backend of the kevent_id and kevent_workq_internal entry points. + * + * @discussion + * Needs to be inline due to the number of arguments. + */ +OS_NOINLINE +static int +kevent_modern_internal(kqueue_t kqu, + user_addr_t changelist, int nchanges, + user_addr_t ueventlist, int nevents, + int flags, kevent_ctx_t kectx, int32_t *retval) +{ + return kevent_internal(kqu.kq, changelist, nchanges, + ueventlist, nevents, flags, kectx, retval, /*legacy*/ false); +} - kn->kn_status &= ~KN_SUPPRESSED; - suppressq = kqueue_get_suppressed_queue(kq, knote_get_qos_index(kn)); - TAILQ_REMOVE(suppressq, kn, kn_tqe); +/*! + * @function kevent_id + * + * @brief + * The kevent_id() syscall. + */ +int +kevent_id(struct proc *p, struct kevent_id_args *uap, int32_t *retval) +{ + int error, flags = uap->flags & KEVENT_FLAG_USER; + uthread_t uth = current_uthread(); + workq_threadreq_t kqr = uth->uu_kqr_bound; + kevent_ctx_t kectx = &uth->uu_save.uus_kevent; + kqueue_t kqu; - /* udate in-use qos to equal requested qos */ - kn->kn_qos_index = kn->kn_req_index; + flags = kevent_adjust_flags_for_proc(p, flags); + flags |= KEVENT_FLAG_DYNAMIC_KQUEUE; - /* don't wakeup if unsuppressing just a stay-active knote */ - if (knote_enqueue(kn) && (kn->kn_status & KN_ACTIVE)) { - knote_wakeup(kn); + if (__improbable((flags & (KEVENT_FLAG_WORKQ | KEVENT_FLAG_WORKLOOP)) != + KEVENT_FLAG_WORKLOOP)) { + return EINVAL; } - if ((kq->kq_state & KQ_WORKLOOP) && !(kq->kq_state & KQ_NO_WQ_THREAD) && - knote_get_qos_override_index(kn) == THREAD_QOS_USER_INTERACTIVE && - kn->kn_qos_override_is_sync) { - struct kqworkloop *kqwl = (struct kqworkloop *)kq; - - /* update the sync qos override counter for suppressed knotes */ - kqworkloop_update_override(kqwl, knote_get_qos_index(kn), - knote_get_qos_override_index(kn), - (KQWL_UO_UPDATE_SUPPRESS_SYNC_COUNTERS | KQWL_UO_OLD_OVERRIDE_IS_SYNC_UI)); + error = kevent_get_data_size(flags, uap->data_available, uap->data_out, kectx); + if (__improbable(error)) { + return error; } - if (TAILQ_EMPTY(suppressq) && (kq->kq_state & KQ_WORKLOOP) && - !(kq->kq_state & KQ_NO_WQ_THREAD)) { - struct kqworkloop *kqwl = (struct kqworkloop *)kq; - if (kqworkloop_is_processing_on_current_thread(kqwl)) { - /* - * kqworkloop_end_processing() will perform the required QoS - * computations when it unsets the processing mode. - */ - } else { - kqwl_req_lock(kqwl); - kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_RESET_WAKEUP_OVERRIDE, 0); - kqwl_req_unlock(kqwl); + kectx->kec_deadline = 0; + kectx->kec_fp = NULL; + kectx->kec_fd = -1; + /* the kec_process_* fields are filled if kqueue_scann is called only */ + + /* + * Get the kq we are going to be working on + * As a fastpath, look at the currently bound workloop. + */ + kqu.kqwl = kqr ? kqr_kqworkloop(kqr) : NULL; + if (kqu.kqwl && kqu.kqwl->kqwl_dynamicid == uap->id) { + if (__improbable(flags & KEVENT_FLAG_DYNAMIC_KQ_MUST_NOT_EXIST)) { + return EEXIST; + } + kqworkloop_retain(kqu.kqwl); + } else if (__improbable(kevent_args_requesting_events(flags, uap->nevents))) { + return EXDEV; + } else { + error = kqworkloop_get_or_create(p, uap->id, NULL, flags, &kqu.kqwl); + if (__improbable(error)) { + return error; } } -} -/* called with kqueue lock held */ -static void -knote_update_sync_override_state(struct knote *kn) -{ - struct kqtailq *queue = knote_get_queue(kn); - struct kqueue *kq = knote_get_kq(kn); + return kevent_modern_internal(kqu, uap->changelist, uap->nchanges, + uap->eventlist, uap->nevents, flags, kectx, retval); +} - if (!(kq->kq_state & KQ_WORKLOOP) || - knote_get_queue_index(kn) != THREAD_QOS_USER_INTERACTIVE) - return; +/**! + * @function kevent_workq_internal + * + * @discussion + * This function is exported for the sake of the workqueue subsystem. + * + * It is called in two ways: + * - when a thread is about to go to userspace to ask for pending event + * - when a thread is returning from userspace with events back + * + * the workqueue subsystem will only use the following flags: + * - KEVENT_FLAG_STACK_DATA (always) + * - KEVENT_FLAG_IMMEDIATE (always) + * - KEVENT_FLAG_PARKING (depending on whether it is going to or returning from + * userspace). + * + * It implicitly acts on the bound kqueue, and for the case of workloops + * will copyout the kqueue ID before anything else. + * + * + * Pthread will have setup the various arguments to fit this stack layout: + * + * +-------....----+--------------+-----------+--------------------+ + * | user stack | data avail | nevents | pthread_self() | + * +-------....----+--------------+-----------+--------------------+ + * ^ ^ + * data_out eventlist + * + * When a workloop is used, the workloop ID is copied out right before + * the eventlist and is taken from the data buffer. + * + * @warning + * This function is carefuly tailored to not make any call except the final tail + * call into kevent_modern_internal. (LTO inlines current_uthread()). + * + * This function is performance sensitive due to the workq subsystem. + */ +int +kevent_workq_internal(struct proc *p, + user_addr_t changelist, int nchanges, + user_addr_t eventlist, int nevents, + user_addr_t data_out, user_size_t *data_available, + unsigned int flags, int32_t *retval) +{ + uthread_t uth = current_uthread(); + workq_threadreq_t kqr = uth->uu_kqr_bound; + kevent_ctx_t kectx = &uth->uu_save.uus_kevent; + kqueue_t kqu; + + assert(flags == (KEVENT_FLAG_STACK_DATA | KEVENT_FLAG_IMMEDIATE) || + flags == (KEVENT_FLAG_STACK_DATA | KEVENT_FLAG_IMMEDIATE | KEVENT_FLAG_PARKING)); + + kectx->kec_data_out = data_out; + kectx->kec_data_avail = (uint64_t)data_available; + kectx->kec_data_size = *data_available; + kectx->kec_data_resid = *data_available; + kectx->kec_deadline = 0; + kectx->kec_fp = NULL; + kectx->kec_fd = -1; + /* the kec_process_* fields are filled if kqueue_scann is called only */ + + flags = kevent_adjust_flags_for_proc(p, flags); + + if (kqr->tr_flags & WORKQ_TR_FLAG_WORKLOOP) { + kqu.kqwl = __container_of(kqr, struct kqworkloop, kqwl_request); + kqworkloop_retain(kqu.kqwl); + + flags |= KEVENT_FLAG_WORKLOOP | KEVENT_FLAG_DYNAMIC_KQUEUE | + KEVENT_FLAG_KERNEL; + } else { + kqu.kqwq = p->p_fd->fd_wqkqueue; - /* Update the sync ipc state on workloop */ - struct kqworkloop *kqwl = (struct kqworkloop *)kq; - boolean_t sync_ipc_override = FALSE; - if (!TAILQ_EMPTY(queue)) { - struct knote *kn_head = TAILQ_FIRST(queue); - if (kn_head->kn_qos_override_is_sync) - sync_ipc_override = TRUE; + flags |= KEVENT_FLAG_WORKQ | KEVENT_FLAG_KERNEL; } - kqworkloop_update_sync_override_state(kqwl, sync_ipc_override); + + return kevent_modern_internal(kqu, changelist, nchanges, + eventlist, nevents, flags, kectx, retval); } -/* called with kqueue lock held */ -static int -knote_enqueue(struct knote *kn) +/*! + * @function kevent_qos + * + * @brief + * The kevent_qos() syscall. + */ +int +kevent_qos(struct proc *p, struct kevent_qos_args *uap, int32_t *retval) { - if ((kn->kn_status & (KN_ACTIVE | KN_STAYACTIVE)) == 0 || - (kn->kn_status & (KN_DISABLED | KN_SUPPRESSED | KN_DROPPING))) - return 0; - - if ((kn->kn_status & KN_QUEUED) == 0) { - struct kqtailq *queue = knote_get_queue(kn); - struct kqueue *kq = knote_get_kq(kn); + uthread_t uth = current_uthread(); + kevent_ctx_t kectx = &uth->uu_save.uus_kevent; + int error, flags = uap->flags & KEVENT_FLAG_USER; + struct kqueue *kq; - kqlock_held(kq); - /* insert at head for sync ipc waiters */ - if (kn->kn_qos_override_is_sync) { - TAILQ_INSERT_HEAD(queue, kn, kn_tqe); - } else { - TAILQ_INSERT_TAIL(queue, kn, kn_tqe); - } - kn->kn_status |= KN_QUEUED; - kq->kq_count++; - knote_update_sync_override_state(kn); - return 1; + if (__improbable(flags & KEVENT_ID_FLAG_USER)) { + return EINVAL; } - return ((kn->kn_status & KN_STAYACTIVE) != 0); -} + flags = kevent_adjust_flags_for_proc(p, flags); -/* called with kqueue lock held */ -static void -knote_dequeue(struct knote *kn) -{ - struct kqueue *kq = knote_get_kq(kn); - struct kqtailq *queue; + error = kevent_get_data_size(flags, uap->data_available, uap->data_out, kectx); + if (__improbable(error)) { + return error; + } - kqlock_held(kq); + kectx->kec_deadline = 0; + kectx->kec_fp = NULL; + kectx->kec_fd = uap->fd; + /* the kec_process_* fields are filled if kqueue_scann is called only */ - if ((kn->kn_status & KN_QUEUED) == 0) - return; + /* get the kq we are going to be working on */ + if (__probable(flags & KEVENT_FLAG_WORKQ)) { + error = kevent_get_kqwq(p, flags, uap->nevents, &kq); + } else { + error = kevent_get_kqfile(p, uap->fd, flags, &kectx->kec_fp, &kq); + } + if (__improbable(error)) { + return error; + } - queue = knote_get_queue(kn); - TAILQ_REMOVE(queue, kn, kn_tqe); - kn->kn_status &= ~KN_QUEUED; - kq->kq_count--; - knote_update_sync_override_state(kn); + return kevent_modern_internal(kq, uap->changelist, uap->nchanges, + uap->eventlist, uap->nevents, flags, kectx, retval); } -void -knote_init(void) +#pragma mark legacy syscalls: kevent, kevent64 + +/*! + * @function kevent_legacy_get_deadline + * + * @brief + * Compute the deadline for the legacy kevent syscalls. + * + * @discussion + * This is not necessary if KEVENT_FLAG_IMMEDIATE is specified, + * as this takes precedence over the deadline. + * + * This function will fail if utimeout is USER_ADDR_NULL + * (the caller should check). + */ +static int +kevent_legacy_get_deadline(int flags, user_addr_t utimeout, uint64_t *deadline) { - knote_zone = zinit(sizeof(struct knote), 8192*sizeof(struct knote), - 8192, "knote zone"); + struct timespec ts; - kqfile_zone = zinit(sizeof(struct kqfile), 8192*sizeof(struct kqfile), - 8192, "kqueue file zone"); + if (flags & KEVENT_FLAG_PROC64) { + struct user64_timespec ts64; + int error = copyin(utimeout, &ts64, sizeof(ts64)); + if (__improbable(error)) { + return error; + } + ts.tv_sec = (unsigned long)ts64.tv_sec; + ts.tv_nsec = (long)ts64.tv_nsec; + } else { + struct user32_timespec ts32; + int error = copyin(utimeout, &ts32, sizeof(ts32)); + if (__improbable(error)) { + return error; + } + ts.tv_sec = ts32.tv_sec; + ts.tv_nsec = ts32.tv_nsec; + } + if (!timespec_is_valid(&ts)) { + return EINVAL; + } - kqworkq_zone = zinit(sizeof(struct kqworkq), 8192*sizeof(struct kqworkq), - 8192, "kqueue workq zone"); + clock_absolutetime_interval_to_deadline(tstoabstime(&ts), deadline); + return 0; +} - kqworkloop_zone = zinit(sizeof(struct kqworkloop), 8192*sizeof(struct kqworkloop), - 8192, "kqueue workloop zone"); +/*! + * @function kevent_legacy_internal + * + * @brief + * The core implementation for kevent and kevent64 + */ +OS_NOINLINE +static int +kevent_legacy_internal(struct proc *p, struct kevent64_args *uap, + int32_t *retval, int flags) +{ + uthread_t uth = current_uthread(); + kevent_ctx_t kectx = &uth->uu_save.uus_kevent; + struct kqueue *kq; + int error; - /* allocate kq lock group attribute and group */ - kq_lck_grp_attr = lck_grp_attr_alloc_init(); + if (__improbable(uap->flags & KEVENT_ID_FLAG_USER)) { + return EINVAL; + } - kq_lck_grp = lck_grp_alloc_init("kqueue", kq_lck_grp_attr); + flags = kevent_adjust_flags_for_proc(p, flags); - /* Allocate kq lock attribute */ - kq_lck_attr = lck_attr_alloc_init(); + kectx->kec_data_out = 0; + kectx->kec_data_avail = 0; + kectx->kec_data_size = 0; + kectx->kec_data_resid = 0; + kectx->kec_deadline = 0; + kectx->kec_fp = NULL; + kectx->kec_fd = uap->fd; + /* the kec_process_* fields are filled if kqueue_scann is called only */ - /* Initialize the timer filter lock */ - lck_mtx_init(&_filt_timerlock, kq_lck_grp, kq_lck_attr); + /* convert timeout to absolute - if we have one (and not immediate) */ + if (__improbable(uap->timeout && !(flags & KEVENT_FLAG_IMMEDIATE))) { + error = kevent_legacy_get_deadline(flags, uap->timeout, + &kectx->kec_deadline); + if (__improbable(error)) { + return error; + } + } - /* Initialize the user filter lock */ - lck_spin_init(&_filt_userlock, kq_lck_grp, kq_lck_attr); + /* get the kq we are going to be working on */ + if (flags & KEVENT_FLAG_WORKQ) { + error = kevent_get_kqwq(p, flags, uap->nevents, &kq); + } else { + error = kevent_get_kqfile(p, uap->fd, flags, &kectx->kec_fp, &kq); + } + if (__improbable(error)) { + return error; + } -#if CONFIG_MEMORYSTATUS - /* Initialize the memorystatus list lock */ - memorystatus_kevent_init(kq_lck_grp, kq_lck_attr); -#endif + return kevent_internal(kq, uap->changelist, uap->nchanges, + uap->eventlist, uap->nevents, flags, kectx, retval, + /*legacy*/ true); } -SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL) -const struct filterops * -knote_fops(struct knote *kn) +/*! + * @function kevent + * + * @brief + * The legacy kevent() syscall. + */ +int +kevent(struct proc *p, struct kevent_args *uap, int32_t *retval) { - return sysfilt_ops[kn->kn_filtid]; -} + struct kevent64_args args = { + .fd = uap->fd, + .changelist = uap->changelist, + .nchanges = uap->nchanges, + .eventlist = uap->eventlist, + .nevents = uap->nevents, + .timeout = uap->timeout, + }; -static struct knote * -knote_alloc(void) -{ - struct knote *kn; - kn = ((struct knote *)zalloc(knote_zone)); - *kn = (struct knote) { .kn_qos_override = 0, .kn_qos_sync_override = 0, .kn_qos_override_is_sync = 0 }; - return kn; + return kevent_legacy_internal(p, &args, retval, KEVENT_FLAG_LEGACY32); } -static void -knote_free(struct knote *kn) +/*! + * @function kevent64 + * + * @brief + * The legacy kevent64() syscall. + */ +int +kevent64(struct proc *p, struct kevent64_args *uap, int32_t *retval) { - zfree(knote_zone, kn); + int flags = (uap->flags & KEVENT_FLAG_USER) | KEVENT_FLAG_LEGACY64; + return kevent_legacy_internal(p, uap, retval, flags); } +#pragma mark - socket interface + #if SOCKETS #include #include @@ -8600,18 +8066,15 @@ knote_free(struct knote *kn) #include #ifndef ROUNDUP64 -#define ROUNDUP64(x) P2ROUNDUP((x), sizeof (u_int64_t)) +#define ROUNDUP64(x) P2ROUNDUP((x), sizeof (u_int64_t)) #endif #ifndef ADVANCE64 -#define ADVANCE64(p, n) (void*)((char *)(p) + ROUNDUP64(n)) +#define ADVANCE64(p, n) (void*)((char *)(p) + ROUNDUP64(n)) #endif -static lck_grp_attr_t *kev_lck_grp_attr; -static lck_attr_t *kev_lck_attr; -static lck_grp_t *kev_lck_grp; -static decl_lck_rw_data(,kev_lck_data); -static lck_rw_t *kev_rwlock = &kev_lck_data; +static LCK_GRP_DECLARE(kev_lck_grp, "Kernel Event Protocol"); +static LCK_RW_DECLARE(kev_rwlock, &kev_lck_grp); static int kev_attach(struct socket *so, int proto, struct proc *p); static int kev_detach(struct socket *so); @@ -8625,29 +8088,29 @@ static int event_sofreelastref(struct socket *); static void kev_delete(struct kern_event_pcb *); static struct pr_usrreqs event_usrreqs = { - .pru_attach = kev_attach, - .pru_control = kev_control, - .pru_detach = kev_detach, - .pru_soreceive = soreceive, + .pru_attach = kev_attach, + .pru_control = kev_control, + .pru_detach = kev_detach, + .pru_soreceive = soreceive, }; static struct protosw eventsw[] = { -{ - .pr_type = SOCK_RAW, - .pr_protocol = SYSPROTO_EVENT, - .pr_flags = PR_ATOMIC, - .pr_usrreqs = &event_usrreqs, - .pr_lock = event_lock, - .pr_unlock = event_unlock, - .pr_getlock = event_getlock, -} + { + .pr_type = SOCK_RAW, + .pr_protocol = SYSPROTO_EVENT, + .pr_flags = PR_ATOMIC, + .pr_usrreqs = &event_usrreqs, + .pr_lock = event_lock, + .pr_unlock = event_unlock, + .pr_getlock = event_getlock, + } }; __private_extern__ int kevt_getstat SYSCTL_HANDLER_ARGS; __private_extern__ int kevt_pcblist SYSCTL_HANDLER_ARGS; SYSCTL_NODE(_net_systm, OID_AUTO, kevt, - CTLFLAG_RW|CTLFLAG_LOCKED, 0, "Kernel event family"); + CTLFLAG_RW | CTLFLAG_LOCKED, 0, "Kernel event family"); struct kevtstat kevtstat; SYSCTL_PROC(_net_systm_kevt, OID_AUTO, stats, @@ -8655,8 +8118,8 @@ SYSCTL_PROC(_net_systm_kevt, OID_AUTO, stats, kevt_getstat, "S,kevtstat", ""); SYSCTL_PROC(_net_systm_kevt, OID_AUTO, pcblist, - CTLTYPE_STRUCT | CTLFLAG_RD | CTLFLAG_LOCKED, 0, 0, - kevt_pcblist, "S,xkevtpcb", ""); + CTLTYPE_STRUCT | CTLFLAG_RD | CTLFLAG_LOCKED, 0, 0, + kevt_pcblist, "S,xkevtpcb", ""); static lck_mtx_t * event_getlock(struct socket *so, int flags) @@ -8664,17 +8127,18 @@ event_getlock(struct socket *so, int flags) #pragma unused(flags) struct kern_event_pcb *ev_pcb = (struct kern_event_pcb *)so->so_pcb; - if (so->so_pcb != NULL) { - if (so->so_usecount < 0) + if (so->so_pcb != NULL) { + if (so->so_usecount < 0) { panic("%s: so=%p usecount=%d lrh= %s\n", __func__, so, so->so_usecount, solockhistory_nr(so)); - /* NOTREACHED */ + } + /* NOTREACHED */ } else { panic("%s: so=%p NULL NO so_pcb %s\n", __func__, so, solockhistory_nr(so)); /* NOTREACHED */ } - return (&ev_pcb->evp_mtx); + return &ev_pcb->evp_mtx; } static int @@ -8682,14 +8146,15 @@ event_lock(struct socket *so, int refcount, void *lr) { void *lr_saved; - if (lr == NULL) + if (lr == NULL) { lr_saved = __builtin_return_address(0); - else + } else { lr_saved = lr; + } if (so->so_pcb != NULL) { lck_mtx_lock(&((struct kern_event_pcb *)so->so_pcb)->evp_mtx); - } else { + } else { panic("%s: so=%p NO PCB! lr=%p lrh= %s\n", __func__, so, lr_saved, solockhistory_nr(so)); /* NOTREACHED */ @@ -8702,12 +8167,13 @@ event_lock(struct socket *so, int refcount, void *lr) /* NOTREACHED */ } - if (refcount) + if (refcount) { so->so_usecount++; + } so->lock_lr[so->next_lock_lr] = lr_saved; - so->next_lock_lr = (so->next_lock_lr+1) % SO_LCKDBG_MAX; - return (0); + so->next_lock_lr = (so->next_lock_lr + 1) % SO_LCKDBG_MAX; + return 0; } static int @@ -8716,10 +8182,11 @@ event_unlock(struct socket *so, int refcount, void *lr) void *lr_saved; lck_mtx_t *mutex_held; - if (lr == NULL) + if (lr == NULL) { lr_saved = __builtin_return_address(0); - else + } else { lr_saved = lr; + } if (refcount) { so->so_usecount--; @@ -8739,7 +8206,7 @@ event_unlock(struct socket *so, int refcount, void *lr) LCK_MTX_ASSERT(mutex_held, LCK_MTX_ASSERT_OWNED); so->unlock_lr[so->next_unlock_lr] = lr_saved; - so->next_unlock_lr = (so->next_unlock_lr+1) % SO_LCKDBG_MAX; + so->next_unlock_lr = (so->next_unlock_lr + 1) % SO_LCKDBG_MAX; if (so->so_usecount == 0) { VERIFY(so->so_flags & SOF_PCBCLEARING); @@ -8748,7 +8215,7 @@ event_unlock(struct socket *so, int refcount, void *lr) lck_mtx_unlock(mutex_held); } - return (0); + return 0; } static int @@ -8771,27 +8238,26 @@ event_sofreelastref(struct socket *so) lck_mtx_unlock(&(ev_pcb->evp_mtx)); LCK_MTX_ASSERT(&(ev_pcb->evp_mtx), LCK_MTX_ASSERT_NOTOWNED); - lck_rw_lock_exclusive(kev_rwlock); + lck_rw_lock_exclusive(&kev_rwlock); LIST_REMOVE(ev_pcb, evp_link); kevtstat.kes_pcbcount--; kevtstat.kes_gencnt++; - lck_rw_done(kev_rwlock); + lck_rw_done(&kev_rwlock); kev_delete(ev_pcb); sofreelastref(so, 1); - return (0); + return 0; } -static int event_proto_count = (sizeof (eventsw) / sizeof (struct protosw)); +static int event_proto_count = (sizeof(eventsw) / sizeof(struct protosw)); static struct kern_event_head kern_event_head; static u_int32_t static_event_id = 0; -#define EVPCB_ZONE_MAX 65536 -#define EVPCB_ZONE_NAME "kerneventpcb" -static struct zone *ev_pcb_zone; +static ZONE_DECLARE(ev_pcb_zone, "kerneventpcb", + sizeof(struct kern_event_pcb), ZC_ZFREE_CLEARMEM); /* * Install the protosw's for the NKE manager. Invoked at extension load time @@ -8805,42 +8271,9 @@ kern_event_init(struct domain *dp) VERIFY(!(dp->dom_flags & DOM_INITIALIZED)); VERIFY(dp == systemdomain); - kev_lck_grp_attr = lck_grp_attr_alloc_init(); - if (kev_lck_grp_attr == NULL) { - panic("%s: lck_grp_attr_alloc_init failed\n", __func__); - /* NOTREACHED */ - } - - kev_lck_grp = lck_grp_alloc_init("Kernel Event Protocol", - kev_lck_grp_attr); - if (kev_lck_grp == NULL) { - panic("%s: lck_grp_alloc_init failed\n", __func__); - /* NOTREACHED */ - } - - kev_lck_attr = lck_attr_alloc_init(); - if (kev_lck_attr == NULL) { - panic("%s: lck_attr_alloc_init failed\n", __func__); - /* NOTREACHED */ - } - - lck_rw_init(kev_rwlock, kev_lck_grp, kev_lck_attr); - if (kev_rwlock == NULL) { - panic("%s: lck_mtx_alloc_init failed\n", __func__); - /* NOTREACHED */ - } - - for (i = 0, pr = &eventsw[0]; i < event_proto_count; i++, pr++) + for (i = 0, pr = &eventsw[0]; i < event_proto_count; i++, pr++) { net_add_proto(pr, dp, 1); - - ev_pcb_zone = zinit(sizeof(struct kern_event_pcb), - EVPCB_ZONE_MAX * sizeof(struct kern_event_pcb), 0, EVPCB_ZONE_NAME); - if (ev_pcb_zone == NULL) { - panic("%s: failed allocating ev_pcb_zone", __func__); - /* NOTREACHED */ } - zone_change(ev_pcb_zone, Z_EXPAND, TRUE); - zone_change(ev_pcb_zone, Z_CALLERACCT, TRUE); } static int @@ -8850,33 +8283,31 @@ kev_attach(struct socket *so, __unused int proto, __unused struct proc *p) struct kern_event_pcb *ev_pcb; error = soreserve(so, KEV_SNDSPACE, KEV_RECVSPACE); - if (error != 0) - return (error); - - if ((ev_pcb = (struct kern_event_pcb *)zalloc(ev_pcb_zone)) == NULL) { - return (ENOBUFS); + if (error != 0) { + return error; } - bzero(ev_pcb, sizeof(struct kern_event_pcb)); - lck_mtx_init(&ev_pcb->evp_mtx, kev_lck_grp, kev_lck_attr); + + ev_pcb = zalloc_flags(ev_pcb_zone, Z_WAITOK | Z_ZERO); + lck_mtx_init(&ev_pcb->evp_mtx, &kev_lck_grp, LCK_ATTR_NULL); ev_pcb->evp_socket = so; ev_pcb->evp_vendor_code_filter = 0xffffffff; so->so_pcb = (caddr_t) ev_pcb; - lck_rw_lock_exclusive(kev_rwlock); + lck_rw_lock_exclusive(&kev_rwlock); LIST_INSERT_HEAD(&kern_event_head, ev_pcb, evp_link); kevtstat.kes_pcbcount++; kevtstat.kes_gencnt++; - lck_rw_done(kev_rwlock); + lck_rw_done(&kev_rwlock); - return (error); + return error; } static void kev_delete(struct kern_event_pcb *ev_pcb) { VERIFY(ev_pcb != NULL); - lck_mtx_destroy(&ev_pcb->evp_mtx, kev_lck_grp); + lck_mtx_destroy(&ev_pcb->evp_mtx, &kev_lck_grp); zfree(ev_pcb_zone, ev_pcb); } @@ -8890,22 +8321,23 @@ kev_detach(struct socket *so) so->so_flags |= SOF_PCBCLEARING; } - return (0); + return 0; } /* * For now, kev_vendor_code and mbuf_tags use the same * mechanism. */ -errno_t kev_vendor_code_find( - const char *string, - u_int32_t *out_vendor_code) +errno_t +kev_vendor_code_find( + const char *string, + u_int32_t *out_vendor_code) { if (strlen(string) >= KEV_VENDOR_CODE_MAX_STR_LEN) { - return (EINVAL); + return EINVAL; } - return (net_str_id_find_internal(string, out_vendor_code, - NSI_VENDOR_CODE, 1)); + return net_str_id_find_internal(string, out_vendor_code, + NSI_VENDOR_CODE, 1); } errno_t @@ -8915,19 +8347,20 @@ kev_msg_post(struct kev_msg *event_msg) net_str_id_first_last(&min_vendor, &max_vendor, NSI_VENDOR_CODE); - if (event_msg == NULL) - return (EINVAL); + if (event_msg == NULL) { + return EINVAL; + } - /* + /* * Limit third parties to posting events for registered vendor codes * only */ if (event_msg->vendor_code < min_vendor || event_msg->vendor_code > max_vendor) { - OSIncrementAtomic64((SInt64 *)&kevtstat.kes_badvendor); - return (EINVAL); + os_atomic_inc(&kevtstat.kes_badvendor, relaxed); + return EINVAL; } - return (kev_post_msg(event_msg)); + return kev_post_msg(event_msg); } int @@ -8944,28 +8377,30 @@ kev_post_msg(struct kev_msg *event_msg) total_size = KEV_MSG_HEADER_SIZE; for (i = 0; i < 5; i++) { - if (event_msg->dv[i].data_length == 0) + if (event_msg->dv[i].data_length == 0) { break; + } total_size += event_msg->dv[i].data_length; } if (total_size > MLEN) { - OSIncrementAtomic64((SInt64 *)&kevtstat.kes_toobig); - return (EMSGSIZE); + os_atomic_inc(&kevtstat.kes_toobig, relaxed); + return EMSGSIZE; } m = m_get(M_WAIT, MT_DATA); if (m == 0) { - OSIncrementAtomic64((SInt64 *)&kevtstat.kes_nomem); - return (ENOMEM); + os_atomic_inc(&kevtstat.kes_nomem, relaxed); + return ENOMEM; } ev = mtod(m, struct kern_event_msg *); total_size = KEV_MSG_HEADER_SIZE; tmp = (char *) &ev->event_data[0]; for (i = 0; i < 5; i++) { - if (event_msg->dv[i].data_length == 0) + if (event_msg->dv[i].data_length == 0) { break; + } total_size += event_msg->dv[i].data_length; bcopy(event_msg->dv[i].data_ptr, tmp, @@ -8981,7 +8416,7 @@ kev_post_msg(struct kev_msg *event_msg) ev->event_code = event_msg->event_code; m->m_len = total_size; - lck_rw_lock_shared(kev_rwlock); + lck_rw_lock_shared(&kev_rwlock); for (ev_pcb = LIST_FIRST(&kern_event_head); ev_pcb; ev_pcb = LIST_NEXT(ev_pcb, evp_link)) { @@ -9014,11 +8449,11 @@ kev_post_msg(struct kev_msg *event_msg) m2 = m_copym(m, 0, m->m_len, M_WAIT); if (m2 == 0) { - OSIncrementAtomic64((SInt64 *)&kevtstat.kes_nomem); + os_atomic_inc(&kevtstat.kes_nomem, relaxed); m_free(m); lck_mtx_unlock(&ev_pcb->evp_mtx); - lck_rw_done(kev_rwlock); - return (ENOMEM); + lck_rw_done(&kev_rwlock); + return ENOMEM; } if (sbappendrecord(&ev_pcb->evp_socket->so_rcv, m2)) { /* @@ -9029,16 +8464,16 @@ kev_post_msg(struct kev_msg *event_msg) 1, m->m_len, MBUF_TC_BE); sorwakeup(ev_pcb->evp_socket); - OSIncrementAtomic64((SInt64 *)&kevtstat.kes_posted); + os_atomic_inc(&kevtstat.kes_posted, relaxed); } else { - OSIncrementAtomic64((SInt64 *)&kevtstat.kes_fullsock); + os_atomic_inc(&kevtstat.kes_fullsock, relaxed); } lck_mtx_unlock(&ev_pcb->evp_mtx); } m_free(m); - lck_rw_done(kev_rwlock); + lck_rw_done(&kev_rwlock); - return (0); + return 0; } static int @@ -9054,32 +8489,32 @@ kev_control(struct socket *so, u_int32_t *id_value = (u_int32_t *) data; switch (cmd) { - case SIOCGKEVID: - *id_value = static_event_id; - break; - case SIOCSKEVFILT: - ev_pcb = (struct kern_event_pcb *) so->so_pcb; - ev_pcb->evp_vendor_code_filter = kev_req->vendor_code; - ev_pcb->evp_class_filter = kev_req->kev_class; - ev_pcb->evp_subclass_filter = kev_req->kev_subclass; - break; - case SIOCGKEVFILT: - ev_pcb = (struct kern_event_pcb *) so->so_pcb; - kev_req->vendor_code = ev_pcb->evp_vendor_code_filter; - kev_req->kev_class = ev_pcb->evp_class_filter; - kev_req->kev_subclass = ev_pcb->evp_subclass_filter; - break; - case SIOCGKEVVENDOR: - kev_vendor = (struct kev_vendor_code *)data; - /* Make sure string is NULL terminated */ - kev_vendor->vendor_string[KEV_VENDOR_CODE_MAX_STR_LEN-1] = 0; - return (net_str_id_find_internal(kev_vendor->vendor_string, - &kev_vendor->vendor_code, NSI_VENDOR_CODE, 0)); - default: - return (ENOTSUP); + case SIOCGKEVID: + *id_value = static_event_id; + break; + case SIOCSKEVFILT: + ev_pcb = (struct kern_event_pcb *) so->so_pcb; + ev_pcb->evp_vendor_code_filter = kev_req->vendor_code; + ev_pcb->evp_class_filter = kev_req->kev_class; + ev_pcb->evp_subclass_filter = kev_req->kev_subclass; + break; + case SIOCGKEVFILT: + ev_pcb = (struct kern_event_pcb *) so->so_pcb; + kev_req->vendor_code = ev_pcb->evp_vendor_code_filter; + kev_req->kev_class = ev_pcb->evp_class_filter; + kev_req->kev_subclass = ev_pcb->evp_subclass_filter; + break; + case SIOCGKEVVENDOR: + kev_vendor = (struct kev_vendor_code *)data; + /* Make sure string is NULL terminated */ + kev_vendor->vendor_string[KEV_VENDOR_CODE_MAX_STR_LEN - 1] = 0; + return net_str_id_find_internal(kev_vendor->vendor_string, + &kev_vendor->vendor_code, NSI_VENDOR_CODE, 0); + default: + return ENOTSUP; } - return (0); + return 0; } int @@ -9088,7 +8523,7 @@ kevt_getstat SYSCTL_HANDLER_ARGS #pragma unused(oidp, arg1, arg2) int error = 0; - lck_rw_lock_shared(kev_rwlock); + lck_rw_lock_shared(&kev_rwlock); if (req->newptr != USER_ADDR_NULL) { error = EPERM; @@ -9102,9 +8537,9 @@ kevt_getstat SYSCTL_HANDLER_ARGS error = SYSCTL_OUT(req, &kevtstat, MIN(sizeof(struct kevtstat), req->oldlen)); done: - lck_rw_done(kev_rwlock); + lck_rw_done(&kev_rwlock); - return (error); + return error; } __private_extern__ int @@ -9112,37 +8547,38 @@ kevt_pcblist SYSCTL_HANDLER_ARGS { #pragma unused(oidp, arg1, arg2) int error = 0; - int n, i; + uint64_t n, i; struct xsystmgen xsg; void *buf = NULL; - size_t item_size = ROUNDUP64(sizeof (struct xkevtpcb)) + - ROUNDUP64(sizeof (struct xsocket_n)) + - 2 * ROUNDUP64(sizeof (struct xsockbuf_n)) + - ROUNDUP64(sizeof (struct xsockstat_n)); + size_t item_size = ROUNDUP64(sizeof(struct xkevtpcb)) + + ROUNDUP64(sizeof(struct xsocket_n)) + + 2 * ROUNDUP64(sizeof(struct xsockbuf_n)) + + ROUNDUP64(sizeof(struct xsockstat_n)); struct kern_event_pcb *ev_pcb; - buf = _MALLOC(item_size, M_TEMP, M_WAITOK | M_ZERO); - if (buf == NULL) - return (ENOMEM); + buf = kheap_alloc(KHEAP_TEMP, item_size, Z_WAITOK | Z_ZERO); + if (buf == NULL) { + return ENOMEM; + } - lck_rw_lock_shared(kev_rwlock); + lck_rw_lock_shared(&kev_rwlock); n = kevtstat.kes_pcbcount; if (req->oldptr == USER_ADDR_NULL) { - req->oldidx = (n + n/8) * item_size; + req->oldidx = (size_t) ((n + n / 8) * item_size); goto done; } if (req->newptr != USER_ADDR_NULL) { error = EPERM; goto done; } - bzero(&xsg, sizeof (xsg)); - xsg.xg_len = sizeof (xsg); + bzero(&xsg, sizeof(xsg)); + xsg.xg_len = sizeof(xsg); xsg.xg_count = n; xsg.xg_gen = kevtstat.kes_gencnt; xsg.xg_sogen = so_gencnt; - error = SYSCTL_OUT(req, &xsg, sizeof (xsg)); + error = SYSCTL_OUT(req, &xsg, sizeof(xsg)); if (error) { goto done; } @@ -9159,13 +8595,13 @@ kevt_pcblist SYSCTL_HANDLER_ARGS i++, ev_pcb = LIST_NEXT(ev_pcb, evp_link)) { struct xkevtpcb *xk = (struct xkevtpcb *)buf; struct xsocket_n *xso = (struct xsocket_n *) - ADVANCE64(xk, sizeof (*xk)); + ADVANCE64(xk, sizeof(*xk)); struct xsockbuf_n *xsbrcv = (struct xsockbuf_n *) - ADVANCE64(xso, sizeof (*xso)); + ADVANCE64(xso, sizeof(*xso)); struct xsockbuf_n *xsbsnd = (struct xsockbuf_n *) - ADVANCE64(xsbrcv, sizeof (*xsbrcv)); + ADVANCE64(xsbrcv, sizeof(*xsbrcv)); struct xsockstat_n *xsostats = (struct xsockstat_n *) - ADVANCE64(xsbsnd, sizeof (*xsbsnd)); + ADVANCE64(xsbsnd, sizeof(*xsbsnd)); bzero(buf, item_size); @@ -9180,9 +8616,9 @@ kevt_pcblist SYSCTL_HANDLER_ARGS sotoxsocket_n(ev_pcb->evp_socket, xso); sbtoxsockbuf_n(ev_pcb->evp_socket ? - &ev_pcb->evp_socket->so_rcv : NULL, xsbrcv); + &ev_pcb->evp_socket->so_rcv : NULL, xsbrcv); sbtoxsockbuf_n(ev_pcb->evp_socket ? - &ev_pcb->evp_socket->so_snd : NULL, xsbsnd); + &ev_pcb->evp_socket->so_snd : NULL, xsbsnd); sbtoxsockstat_n(ev_pcb->evp_socket, xsostats); lck_mtx_unlock(&ev_pcb->evp_mtx); @@ -9198,21 +8634,22 @@ kevt_pcblist SYSCTL_HANDLER_ARGS * while we were processing this request, and it * might be necessary to retry. */ - bzero(&xsg, sizeof (xsg)); - xsg.xg_len = sizeof (xsg); + bzero(&xsg, sizeof(xsg)); + xsg.xg_len = sizeof(xsg); xsg.xg_count = n; xsg.xg_gen = kevtstat.kes_gencnt; xsg.xg_sogen = so_gencnt; - error = SYSCTL_OUT(req, &xsg, sizeof (xsg)); + error = SYSCTL_OUT(req, &xsg, sizeof(xsg)); if (error) { goto done; } } done: - lck_rw_done(kev_rwlock); + lck_rw_done(&kev_rwlock); - return (error); + kheap_free(KHEAP_TEMP, buf, item_size); + return error; } #endif /* SOCKETS */ @@ -9226,57 +8663,69 @@ fill_kqueueinfo(struct kqueue *kq, struct kqueue_info * kinfo) st = &kinfo->kq_stat; st->vst_size = kq->kq_count; - if (kq->kq_state & KQ_KEV_QOS) + if (kq->kq_state & KQ_KEV_QOS) { st->vst_blksize = sizeof(struct kevent_qos_s); - else if (kq->kq_state & KQ_KEV64) + } else if (kq->kq_state & KQ_KEV64) { st->vst_blksize = sizeof(struct kevent64_s); - else + } else { st->vst_blksize = sizeof(struct kevent); + } st->vst_mode = S_IFIFO; st->vst_ino = (kq->kq_state & KQ_DYNAMIC) ? - ((struct kqworkloop *)kq)->kqwl_dynamicid : 0; + ((struct kqworkloop *)kq)->kqwl_dynamicid : 0; /* flags exported to libproc as PROC_KQUEUE_* (sys/proc_info.h) */ #define PROC_KQUEUE_MASK (KQ_SEL|KQ_SLEEP|KQ_KEV32|KQ_KEV64|KQ_KEV_QOS|KQ_WORKQ|KQ_WORKLOOP) kinfo->kq_state = kq->kq_state & PROC_KQUEUE_MASK; - return (0); + return 0; } static int -fill_kqueue_dyninfo(struct kqueue *kq, struct kqueue_dyninfo *kqdi) +fill_kqueue_dyninfo(struct kqworkloop *kqwl, struct kqueue_dyninfo *kqdi) { - struct kqworkloop *kqwl = (struct kqworkloop *)kq; - struct kqrequest *kqr = &kqwl->kqwl_request; + workq_threadreq_t kqr = &kqwl->kqwl_request; + workq_threadreq_param_t trp = {}; int err; - if ((kq->kq_state & KQ_WORKLOOP) == 0) { + if ((kqwl->kqwl_state & KQ_WORKLOOP) == 0) { return EINVAL; } - if ((err = fill_kqueueinfo(kq, &kqdi->kqdi_info))) { + if ((err = fill_kqueueinfo(&kqwl->kqwl_kqueue, &kqdi->kqdi_info))) { return err; } - kqwl_req_lock(kqwl); + kqlock(kqwl); + + kqdi->kqdi_servicer = thread_tid(kqr_thread(kqr)); + kqdi->kqdi_owner = thread_tid(kqwl->kqwl_owner); + kqdi->kqdi_request_state = kqr->tr_state; + kqdi->kqdi_async_qos = kqr->tr_kq_qos_index; + kqdi->kqdi_events_qos = kqr->tr_kq_override_index; + kqdi->kqdi_sync_waiters = 0; + kqdi->kqdi_sync_waiter_qos = 0; - if (kqr->kqr_thread) { - kqdi->kqdi_servicer = thread_tid(kqr->kqr_thread); + trp.trp_value = kqwl->kqwl_params; + if (trp.trp_flags & TRP_PRIORITY) { + kqdi->kqdi_pri = trp.trp_pri; + } else { + kqdi->kqdi_pri = 0; } - if (kqwl->kqwl_owner == WL_OWNER_SUSPENDED) { - kqdi->kqdi_owner = ~0ull; + if (trp.trp_flags & TRP_POLICY) { + kqdi->kqdi_pol = trp.trp_pol; } else { - kqdi->kqdi_owner = thread_tid(kqwl->kqwl_owner); + kqdi->kqdi_pol = 0; } - kqdi->kqdi_request_state = kqr->kqr_state; - kqdi->kqdi_async_qos = kqr->kqr_qos_index; - kqdi->kqdi_events_qos = kqr->kqr_override_index; - kqdi->kqdi_sync_waiters = kqr->kqr_dsync_waiters; - kqdi->kqdi_sync_waiter_qos = kqr->kqr_dsync_waiters_qos; + if (trp.trp_flags & TRP_CPUPERCENT) { + kqdi->kqdi_cpupercent = trp.trp_cpupercent; + } else { + kqdi->kqdi_cpupercent = 0; + } - kqwl_req_unlock(kqwl); + kqunlock(kqwl); return 0; } @@ -9286,6 +8735,7 @@ void knote_markstayactive(struct knote *kn) { struct kqueue *kq = knote_get_kq(kn); + kq_index_t qos; kqlock(kq); kn->kn_status |= KN_STAYACTIVE; @@ -9294,62 +8744,63 @@ knote_markstayactive(struct knote *kn) * Making a knote stay active is a property of the knote that must be * established before it is fully attached. */ - assert(kn->kn_status & KN_ATTACHING); + assert((kn->kn_status & (KN_QUEUED | KN_SUPPRESSED)) == 0); /* handle all stayactive knotes on the (appropriate) manager */ - if (kq->kq_state & KQ_WORKQ) { - knote_set_qos_index(kn, KQWQ_QOS_MANAGER); - } else if (kq->kq_state & KQ_WORKLOOP) { + if (kq->kq_state & KQ_WORKLOOP) { struct kqworkloop *kqwl = (struct kqworkloop *)kq; - kqwl_req_lock(kqwl); - assert(kn->kn_req_index && kn->kn_req_index < THREAD_QOS_LAST); - kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_UPDATE_STAYACTIVE_QOS, - kn->kn_req_index); - kqwl_req_unlock(kqwl); - knote_set_qos_index(kn, KQWL_BUCKET_STAYACTIVE); + + qos = _pthread_priority_thread_qos(kn->kn_qos); + assert(qos && qos < THREAD_QOS_LAST); + kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_UPDATE_STAYACTIVE_QOS, qos); + qos = KQWL_BUCKET_STAYACTIVE; + } else if (kq->kq_state & KQ_WORKQ) { + qos = KQWQ_QOS_MANAGER; + } else { + qos = THREAD_QOS_UNSPECIFIED; } - knote_activate(kn); + kn->kn_qos_override = qos; + kn->kn_qos_index = qos; + + knote_activate(kq, kn, FILTER_ACTIVE); kqunlock(kq); } void knote_clearstayactive(struct knote *kn) { - kqlock(knote_get_kq(kn)); - kn->kn_status &= ~KN_STAYACTIVE; - knote_deactivate(kn); - kqunlock(knote_get_kq(kn)); + struct kqueue *kq = knote_get_kq(kn); + kqlock(kq); + kn->kn_status &= ~(KN_STAYACTIVE | KN_ACTIVE); + knote_dequeue(kq, kn); + kqunlock(kq); } static unsigned long kevent_extinfo_emit(struct kqueue *kq, struct knote *kn, struct kevent_extinfo *buf, - unsigned long buflen, unsigned long nknotes) + unsigned long buflen, unsigned long nknotes) { for (; kn; kn = SLIST_NEXT(kn, kn_link)) { if (kq == knote_get_kq(kn)) { if (nknotes < buflen) { struct kevent_extinfo *info = &buf[nknotes]; - struct kevent_internal_s *kevp = &kn->kn_kevent; kqlock(kq); - info->kqext_kev = (struct kevent_qos_s){ - .ident = kevp->ident, - .filter = kevp->filter, - .flags = kevp->flags, - .fflags = kevp->fflags, - .data = (int64_t)kevp->data, - .udata = kevp->udata, - .ext[0] = kevp->ext[0], - .ext[1] = kevp->ext[1], - .ext[2] = kevp->ext[2], - .ext[3] = kevp->ext[3], - .qos = kn->kn_req_index, - }; - info->kqext_sdata = kn->kn_sdata; - info->kqext_status = kn->kn_status; - info->kqext_sfflags = kn->kn_sfflags; + info->kqext_kev = *(struct kevent_qos_s *)&kn->kn_kevent; + if (knote_has_qos(kn)) { + info->kqext_kev.qos = + _pthread_priority_thread_qos_fast(kn->kn_qos); + } else { + info->kqext_kev.qos = kn->kn_qos_override; + } + info->kqext_kev.filter |= 0xff00; /* sign extend filter */ + info->kqext_kev.xflags = 0; /* this is where sfflags lives */ + info->kqext_kev.data = 0; /* this is where sdata lives */ + info->kqext_sdata = kn->kn_sdata; + info->kqext_status = kn->kn_status; + info->kqext_sfflags = kn->kn_sfflags; kqunlock(kq); } @@ -9364,7 +8815,7 @@ kevent_extinfo_emit(struct kqueue *kq, struct knote *kn, struct kevent_extinfo * int kevent_copyout_proc_dynkqids(void *proc, user_addr_t ubuf, uint32_t ubufsize, - int32_t *nkqueues_out) + int32_t *nkqueues_out) { proc_t p = (proc_t)proc; struct filedesc *fdp = p->p_fd; @@ -9381,24 +8832,27 @@ kevent_copyout_proc_dynkqids(void *proc, user_addr_t ubuf, uint32_t ubufsize, goto out; } - buflen = min(ubuflen, PROC_PIDDYNKQUEUES_MAX); + buflen = MIN(ubuflen, PROC_PIDDYNKQUEUES_MAX); if (ubuflen != 0) { if (os_mul_overflow(sizeof(kqueue_id_t), buflen, &bufsize)) { err = ERANGE; goto out; } - kq_ids = kalloc(bufsize); - assert(kq_ids != NULL); + kq_ids = kheap_alloc(KHEAP_TEMP, bufsize, Z_WAITOK | Z_ZERO); + if (!kq_ids) { + err = ENOMEM; + goto out; + } } - kqhash_lock(p); + kqhash_lock(fdp); if (fdp->fd_kqhashmask > 0) { for (uint32_t i = 0; i < fdp->fd_kqhashmask + 1; i++) { struct kqworkloop *kqwl; - SLIST_FOREACH(kqwl, &fdp->fd_kqhash[i], kqwl_hashlink) { + LIST_FOREACH(kqwl, &fdp->fd_kqhash[i], kqwl_hashlink) { /* report the number of kqueues, even if they don't all fit */ if (nkqueues < buflen) { kq_ids[nkqueues] = kqwl->kqwl_dynamicid; @@ -9408,11 +8862,11 @@ kevent_copyout_proc_dynkqids(void *proc, user_addr_t ubuf, uint32_t ubufsize, } } - kqhash_unlock(p); + kqhash_unlock(fdp); if (kq_ids) { size_t copysize; - if (os_mul_overflow(sizeof(kqueue_id_t), min(ubuflen, nkqueues), ©size)) { + if (os_mul_overflow(sizeof(kqueue_id_t), MIN(buflen, nkqueues), ©size)) { err = ERANGE; goto out; } @@ -9423,7 +8877,7 @@ kevent_copyout_proc_dynkqids(void *proc, user_addr_t ubuf, uint32_t ubufsize, out: if (kq_ids) { - kfree(kq_ids, bufsize); + kheap_free(KHEAP_TEMP, kq_ids, bufsize); } if (!err) { @@ -9434,10 +8888,10 @@ out: int kevent_copyout_dynkqinfo(void *proc, kqueue_id_t kq_id, user_addr_t ubuf, - uint32_t ubufsize, int32_t *size_out) + uint32_t ubufsize, int32_t *size_out) { proc_t p = (proc_t)proc; - struct kqueue *kq; + struct kqworkloop *kqwl; int err = 0; struct kqueue_dyninfo kqdi = { }; @@ -9447,14 +8901,10 @@ kevent_copyout_dynkqinfo(void *proc, kqueue_id_t kq_id, user_addr_t ubuf, return ENOBUFS; } - kqhash_lock(p); - kq = kqueue_hash_lookup(p, kq_id); - if (!kq) { - kqhash_unlock(p); + kqwl = kqworkloop_hash_lookup_and_retain(p->p_fd, kq_id); + if (!kqwl) { return ESRCH; } - kqueue_retain(kq); - kqhash_unlock(p); /* * backward compatibility: allow the argument to this call to only be @@ -9462,45 +8912,39 @@ kevent_copyout_dynkqinfo(void *proc, kqueue_id_t kq_id, user_addr_t ubuf, */ if (ubufsize >= sizeof(struct kqueue_dyninfo)) { ubufsize = sizeof(struct kqueue_dyninfo); - err = fill_kqueue_dyninfo(kq, &kqdi); + err = fill_kqueue_dyninfo(kqwl, &kqdi); } else { ubufsize = sizeof(struct kqueue_info); - err = fill_kqueueinfo(kq, &kqdi.kqdi_info); + err = fill_kqueueinfo(&kqwl->kqwl_kqueue, &kqdi.kqdi_info); } if (err == 0 && (err = copyout(&kqdi, ubuf, ubufsize)) == 0) { *size_out = ubufsize; } - kqueue_release_last(p, kq); + kqworkloop_release(kqwl); return err; } int kevent_copyout_dynkqextinfo(void *proc, kqueue_id_t kq_id, user_addr_t ubuf, - uint32_t ubufsize, int32_t *nknotes_out) + uint32_t ubufsize, int32_t *nknotes_out) { proc_t p = (proc_t)proc; - struct kqueue *kq; + struct kqworkloop *kqwl; int err; - assert(p != NULL); - - kqhash_lock(p); - kq = kqueue_hash_lookup(p, kq_id); - if (!kq) { - kqhash_unlock(p); + kqwl = kqworkloop_hash_lookup_and_retain(p->p_fd, kq_id); + if (!kqwl) { return ESRCH; } - kqueue_retain(kq); - kqhash_unlock(p); - err = pid_kqueue_extinfo(p, kq, ubuf, ubufsize, nknotes_out); - kqueue_release_last(p, kq); + err = pid_kqueue_extinfo(p, &kqwl->kqwl_kqueue, ubuf, ubufsize, nknotes_out); + kqworkloop_release(kqwl); return err; } int pid_kqueue_extinfo(proc_t p, struct kqueue *kq, user_addr_t ubuf, - uint32_t bufsize, int32_t *retval) + uint32_t bufsize, int32_t *retval) { struct knote *kn; int i; @@ -9511,14 +8955,14 @@ pid_kqueue_extinfo(proc_t p, struct kqueue *kq, user_addr_t ubuf, struct kevent_extinfo *kqext = NULL; /* arbitrary upper limit to cap kernel memory usage, copyout size, etc. */ - buflen = min(buflen, PROC_PIDFDKQUEUE_KNOTES_MAX); + buflen = MIN(buflen, PROC_PIDFDKQUEUE_KNOTES_MAX); - kqext = kalloc(buflen * sizeof(struct kevent_extinfo)); + kqext = kheap_alloc(KHEAP_TEMP, + buflen * sizeof(struct kevent_extinfo), Z_WAITOK | Z_ZERO); if (kqext == NULL) { err = ENOMEM; goto out; } - bzero(kqext, buflen * sizeof(struct kevent_extinfo)); proc_fdlock(p); for (i = 0; i < fdp->fd_knlistsize; i++) { @@ -9529,41 +8973,37 @@ pid_kqueue_extinfo(proc_t p, struct kqueue *kq, user_addr_t ubuf, if (fdp->fd_knhashmask != 0) { for (i = 0; i < (int)fdp->fd_knhashmask + 1; i++) { - kqhash_lock(p); + knhash_lock(fdp); kn = SLIST_FIRST(&fdp->fd_knhash[i]); nknotes = kevent_extinfo_emit(kq, kn, kqext, buflen, nknotes); - kqhash_unlock(p); + knhash_unlock(fdp); } } - assert(bufsize >= sizeof(struct kevent_extinfo) * min(buflen, nknotes)); - err = copyout(kqext, ubuf, sizeof(struct kevent_extinfo) * min(buflen, nknotes)); + assert(bufsize >= sizeof(struct kevent_extinfo) * MIN(buflen, nknotes)); + err = copyout(kqext, ubuf, sizeof(struct kevent_extinfo) * MIN(buflen, nknotes)); - out: - if (kqext) { - kfree(kqext, buflen * sizeof(struct kevent_extinfo)); - kqext = NULL; - } +out: + kheap_free(KHEAP_TEMP, kqext, buflen * sizeof(struct kevent_extinfo)); if (!err) { - *retval = min(nknotes, PROC_PIDFDKQUEUE_KNOTES_MAX); + *retval = (int32_t)MIN(nknotes, PROC_PIDFDKQUEUE_KNOTES_MAX); } return err; } static unsigned int klist_copy_udata(struct klist *list, uint64_t *buf, - unsigned int buflen, unsigned int nknotes) + unsigned int buflen, unsigned int nknotes) { - struct kevent_internal_s *kev; struct knote *kn; SLIST_FOREACH(kn, list, kn_link) { if (nknotes < buflen) { - struct kqueue *kq = knote_get_kq(kn); - kqlock(kq); - kev = &(kn->kn_kevent); - buf[nknotes] = kev->udata; - kqunlock(kq); + /* + * kevent_register will always set kn_udata atomically + * so that we don't have to take any kqlock here. + */ + buf[nknotes] = os_atomic_load_wide(&kn->kn_udata, relaxed); } /* we return total number of knotes, which may be more than requested */ nknotes++; @@ -9572,28 +9012,14 @@ klist_copy_udata(struct klist *list, uint64_t *buf, return nknotes; } -static unsigned int -kqlist_copy_dynamicids(__assert_only proc_t p, struct kqlist *list, - uint64_t *buf, unsigned int buflen, unsigned int nids) -{ - kqhash_lock_held(p); - struct kqworkloop *kqwl; - SLIST_FOREACH(kqwl, list, kqwl_hashlink) { - if (nids < buflen) { - buf[nids] = kqwl->kqwl_dynamicid; - } - nids++; - } - return nids; -} - int -kevent_proc_copy_uptrs(void *proc, uint64_t *buf, int bufsize) +kevent_proc_copy_uptrs(void *proc, uint64_t *buf, uint32_t bufsize) { proc_t p = (proc_t)proc; struct filedesc *fdp = p->p_fd; unsigned int nuptrs = 0; - unsigned long buflen = bufsize / sizeof(uint64_t); + unsigned int buflen = bufsize / sizeof(uint64_t); + struct kqworkloop *kqwl; if (buflen > 0) { assert(buf != NULL); @@ -9603,35 +9029,32 @@ kevent_proc_copy_uptrs(void *proc, uint64_t *buf, int bufsize) for (int i = 0; i < fdp->fd_knlistsize; i++) { nuptrs = klist_copy_udata(&fdp->fd_knlist[i], buf, buflen, nuptrs); } - knhash_lock(p); proc_fdunlock(p); + + knhash_lock(fdp); if (fdp->fd_knhashmask != 0) { - for (int i = 0; i < (int)fdp->fd_knhashmask + 1; i++) { + for (size_t i = 0; i < fdp->fd_knhashmask + 1; i++) { nuptrs = klist_copy_udata(&fdp->fd_knhash[i], buf, buflen, nuptrs); } } - knhash_unlock(p); + knhash_unlock(fdp); - kqhash_lock(p); + kqhash_lock(fdp); if (fdp->fd_kqhashmask != 0) { - for (int i = 0; i < (int)fdp->fd_kqhashmask + 1; i++) { - nuptrs = kqlist_copy_dynamicids(p, &fdp->fd_kqhash[i], buf, buflen, - nuptrs); + for (size_t i = 0; i < fdp->fd_kqhashmask + 1; i++) { + LIST_FOREACH(kqwl, &fdp->fd_kqhash[i], kqwl_hashlink) { + if (nuptrs < buflen) { + buf[nuptrs] = kqwl->kqwl_dynamicid; + } + nuptrs++; + } } } - kqhash_unlock(p); + kqhash_unlock(fdp); return (int)nuptrs; } -static void -kevent_redrive_proc_thread_request(proc_t p) -{ - __assert_only int ret; - ret = (*pthread_functions->workq_threadreq)(p, NULL, WORKQ_THREADREQ_REDRIVE, 0, 0); - assert(ret == 0 || ret == ECANCELED); -} - static void kevent_set_return_to_kernel_user_tsd(proc_t p, thread_t thread) { @@ -9642,12 +9065,8 @@ kevent_set_return_to_kernel_user_tsd(proc_t p, thread_t thread) uint64_t ast_flags64 = 0; struct uthread *ut = get_bsdthread_info(thread); - if (ut->uu_kqueue_bound != NULL) { - if (ut->uu_kqueue_flags & KEVENT_FLAG_WORKLOOP) { - ast_flags64 |= R2K_WORKLOOP_PENDING_EVENTS; - } else if (ut->uu_kqueue_flags & KEVENT_FLAG_WORKQ) { - ast_flags64 |= R2K_WORKQ_PENDING_EVENTS; - } + if (ut->uu_kqr_bound != NULL) { + ast_flags64 |= R2K_WORKLOOP_PENDING_EVENTS; } if (ast_flags64 == 0) { @@ -9665,10 +9084,10 @@ kevent_set_return_to_kernel_user_tsd(proc_t p, thread_t thread) } if (copyout((proc_is_64bit ? (void *)&ast_flags64 : (void *)&ast_flags32), - (user_addr_t)ast_addr, - user_addr_size) != 0) { + (user_addr_t)ast_addr, + user_addr_size) != 0) { printf("pid %d (tid:%llu): copyout of return_to_kernel ast flags failed with " - "ast_addr = %llu\n", p->p_pid, thread_tid(current_thread()), ast_addr); + "ast_addr = %llu\n", p->p_pid, thread_tid(current_thread()), ast_addr); } } @@ -9678,7 +9097,7 @@ kevent_ast(thread_t thread, uint16_t bits) proc_t p = current_proc(); if (bits & AST_KEVENT_REDRIVE_THREADREQ) { - kevent_redrive_proc_thread_request(p); + workq_kern_threadreq_redrive(p, WORKQ_THREADREQ_CAN_CREATE_THREADS); } if (bits & AST_KEVENT_RETURN_TO_KERNEL) { kevent_set_return_to_kernel_user_tsd(p, thread); @@ -9695,8 +9114,6 @@ kevent_sysctl SYSCTL_HANDLER_ARGS #pragma unused(oidp, arg2) uintptr_t type = (uintptr_t)arg1; uint64_t bound_id = 0; - struct uthread *ut; - struct kqueue *kq; if (type != KEVENT_SYSCTL_BOUND_ID) { return EINVAL; @@ -9706,16 +9123,16 @@ kevent_sysctl SYSCTL_HANDLER_ARGS return EINVAL; } - ut = get_bsdthread_info(current_thread()); + struct uthread *ut = get_bsdthread_info(current_thread()); if (!ut) { return EFAULT; } - kq = ut->uu_kqueue_bound; - if (kq) { - if (kq->kq_state & KQ_WORKLOOP) { - bound_id = ((struct kqworkloop *)kq)->kqwl_dynamicid; - } else if (kq->kq_state & KQ_WORKQ) { + workq_threadreq_t kqr = ut->uu_kqr_bound; + if (kqr) { + if (kqr->tr_flags & WORKQ_TR_FLAG_WORKLOOP) { + bound_id = kqr_kqworkloop(kqr)->kqwl_dynamicid; + } else { bound_id = -1; } } @@ -9724,12 +9141,12 @@ kevent_sysctl SYSCTL_HANDLER_ARGS } SYSCTL_NODE(_kern, OID_AUTO, kevent, CTLFLAG_RW | CTLFLAG_LOCKED, 0, - "kevent information"); + "kevent information"); SYSCTL_PROC(_kern_kevent, OID_AUTO, bound_id, - CTLTYPE_QUAD | CTLFLAG_RD | CTLFLAG_LOCKED | CTLFLAG_MASKED, - (void *)KEVENT_SYSCTL_BOUND_ID, - sizeof(kqueue_id_t), kevent_sysctl, "Q", - "get the ID of the bound kqueue"); + CTLTYPE_QUAD | CTLFLAG_RD | CTLFLAG_LOCKED | CTLFLAG_MASKED, + (void *)KEVENT_SYSCTL_BOUND_ID, + sizeof(kqueue_id_t), kevent_sysctl, "Q", + "get the ID of the bound kqueue"); #endif /* DEVELOPMENT || DEBUG */