X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/ff6e181ae92fc6f1e89841290f461d1f2f9badd9..b226f5e54a60dc81db17b1260381d7dbfea3cdf1:/bsd/kern/kern_event.c diff --git a/bsd/kern/kern_event.c b/bsd/kern/kern_event.c index 870989b26..ecffc8db7 100644 --- a/bsd/kern/kern_event.c +++ b/bsd/kern/kern_event.c @@ -1,15 +1,20 @@ /* - * Copyright (c) 2000-2005 Apple Computer, Inc. All rights reserved. + * Copyright (c) 2000-2017 Apple Inc. All rights reserved. + * + * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ * - * @APPLE_LICENSE_HEADER_START@ - * * This file contains Original Code and/or Modifications of Original Code * as defined in and that are subject to the Apple Public Source License * Version 2.0 (the 'License'). You may not use this file except in - * compliance with the License. Please obtain a copy of the License at - * http://www.opensource.apple.com/apsl/ and read it before using this - * file. - * + * compliance with the License. The rights granted to you under the License + * may not be used to create, or enable the creation or redistribution of, + * unlawful or unlicensed copies of an Apple operating system, or to + * circumvent, violate, or enable the circumvention or violation of, any + * terms of an Apple operating system software license agreement. + * + * Please obtain a copy of the License at + * http://www.opensource.apple.com/apsl/ and read it before using this file. + * * The Original Code and all software distributed under the License are * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES, @@ -17,8 +22,8 @@ * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT. * Please see the License for the specific language governing rights and * limitations under the License. - * - * @APPLE_LICENSE_HEADER_END@ + * + * @APPLE_OSREFERENCE_LICENSE_HEADER_END@ * */ /*- @@ -50,6 +55,7 @@ * @(#)kern_event.c 1.0 (3/31/2000) */ #include +#include #include #include @@ -57,7 +63,7 @@ #include #include #include -#include +#include #include #include #include @@ -73,459 +79,1120 @@ #include #include #include +#include #include - -#include +#include +#include +#include +#include +#include +#include +#include + +#include #include +#include +#include #include #include +#include #include +#include #include +#include +#include +#include +#include +#include +#include #include +#include -extern void unix_syscall_return(int); +#include "net/net_str_id.h" -MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system"); +#include +#include -static inline void kqlock(struct kqueue *kq); -static inline void kqunlock(struct kqueue *kq); - -static int kqlock2knoteuse(struct kqueue *kq, struct knote *kn); -static int kqlock2knoteusewait(struct kqueue *kq, struct knote *kn); -static int kqlock2knotedrop(struct kqueue *kq, struct knote *kn); -static int knoteuse2kqlock(struct kqueue *kq, struct knote *kn); - -static void kqueue_wakeup(struct kqueue *kq); -static int kqueue_read(struct fileproc *fp, struct uio *uio, - kauth_cred_t cred, int flags, struct proc *p); -static int kqueue_write(struct fileproc *fp, struct uio *uio, - kauth_cred_t cred, int flags, struct proc *p); -static int kqueue_ioctl(struct fileproc *fp, u_long com, caddr_t data, - struct proc *p); -static int kqueue_select(struct fileproc *fp, int which, void *wql, - struct proc *p); -static int kqueue_close(struct fileglob *fp, struct proc *p); -static int kqueue_kqfilter(struct fileproc *fp, struct knote *kn, struct proc *p); -extern int kqueue_stat(struct fileproc *fp, struct stat *st, struct proc *p); - -static struct fileops kqueueops = { - kqueue_read, - kqueue_write, - kqueue_ioctl, - kqueue_select, - kqueue_close, - kqueue_kqfilter, - 0 -}; +#if CONFIG_MEMORYSTATUS +#include +#endif -static int kevent_copyin(user_addr_t *addrp, struct kevent *kevp, struct proc *p); -static int kevent_copyout(struct kevent *kevp, user_addr_t *addrp, struct proc *p); +extern thread_t port_name_to_thread(mach_port_name_t port_name); /* osfmk/kern/ipc_tt.h */ +extern mach_port_name_t ipc_entry_name_mask(mach_port_name_t name); /* osfmk/ipc/ipc_entry.h */ -static int kevent_callback(struct kqueue *kq, struct kevent *kevp, void *data); -static void kevent_continue(struct kqueue *kq, void *data, int error); -static void kevent_scan_continue(void *contp, wait_result_t wait_result); -static int kevent_process(struct kqueue *kq, kevent_callback_t callback, - void *data, int *countp, struct proc *p); -static void knote_put(struct knote *kn); -static int knote_fdpattach(struct knote *kn, struct filedesc *fdp, struct proc *p); -static void knote_drop(struct knote *kn, struct proc *p); -static void knote_activate(struct knote *kn); -static void knote_deactivate(struct knote *kn); -static void knote_enqueue(struct knote *kn); -static void knote_dequeue(struct knote *kn); -static struct knote *knote_alloc(void); -static void knote_free(struct knote *kn); -extern void knote_init(void); +#define KEV_EVTID(code) BSDDBG_CODE(DBG_BSD_KEVENT, (code)) -static int filt_fileattach(struct knote *kn); -static struct filterops file_filtops = - { 1, filt_fileattach, NULL, NULL }; +MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system"); -static void filt_kqdetach(struct knote *kn); -static int filt_kqueue(struct knote *kn, long hint); -static struct filterops kqread_filtops = - { 1, NULL, filt_kqdetach, filt_kqueue }; +#define KQ_EVENT NO_EVENT64 + +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); +static int kqueue_select(struct fileproc *fp, int which, void *wq_link_id, + 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); +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_kqfilter = kqueue_kqfilter, + .fo_drain = kqueue_drain, +}; -/* - * placeholder for not-yet-implemented filters - */ -static int filt_badattach(struct knote *kn); -static struct filterops bad_filtops = - { 0, filt_badattach, 0 , 0 }; +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 int kevent_register_wait_prepare(struct knote *kn, struct kevent_internal_s *kev); +static void kevent_register_wait_block(struct turnstile *ts, thread_t handoff_thread, + struct knote_lock_ctx *knlc, 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); +static inline void kqueue_release_last(struct proc *p, kqueue_t kqu); +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); +static int kqueue_queue_empty(struct kqueue *kq, kq_index_t qos_index); + +static struct kqtailq *kqueue_get_suppressed_queue(kqueue_t kq, struct knote *kn); +static void kqueue_threadreq_initiate(struct kqueue *kq, struct kqrequest *kqr, kq_index_t qos, int flags); + +static void kqworkq_update_override(struct kqworkq *kqwq, struct knote *kn, kq_index_t qos); +static void kqworkq_unbind(proc_t p, struct kqrequest *kqr); +static thread_qos_t kqworkq_unbind_locked(struct kqworkq *kqwq, struct kqrequest *kqr, thread_t thread); +static struct kqrequest *kqworkq_get_request(struct kqworkq *kqwq, kq_index_t qos_index); + +static void kqworkloop_update_override(struct kqworkloop *kqwl, kq_index_t override_index); +static void kqworkloop_unbind(proc_t p, struct kqworkloop *kwql); +static thread_qos_t kqworkloop_unbind_locked(struct kqworkloop *kwql, thread_t thread); +static kq_index_t kqworkloop_owner_override(struct kqworkloop *kqwl); +enum { + KQWL_UTQ_NONE, + /* + * The wakeup qos is the qos of QUEUED knotes. + * + * 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 + * 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. + * This override is cleared when the knote suppressed list becomes empty. + */ + KQWL_UTQ_UPDATE_WAKEUP_OVERRIDE, + KQWL_UTQ_RESET_WAKEUP_OVERRIDE, + /* + * 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_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 kqworkloop_end_processing(struct kqworkloop *kqwl, int flags, int kevent_flags); + +static int knote_process(struct knote *kn, kevent_callback_t callback, void *callback_data, + struct filt_process_s *process_data); + +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_internal_s *kev, bool is_fd, struct proc *p); + +static void knote_drop(struct kqueue *kq, struct knote *kn, struct knote_lock_ctx *knlc); +static struct knote *knote_alloc(void); +static void knote_free(struct knote *kn); + +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 bool knote_should_apply_qos_override(struct kqueue *kq, struct knote *kn, + int result, thread_qos_t *qos_out); +static void knote_apply_qos_override(struct knote *kn, kq_index_t qos_index); +static void knote_adjust_qos(struct kqueue *kq, struct knote *kn, int result); +static void knote_reset_priority(struct knote *kn, pthread_priority_t pp); +static kq_index_t knote_get_qos_override_index(struct knote *kn); +static void knote_set_qos_overcommit(struct knote *kn); + +static zone_t knote_zone; +static zone_t kqfile_zone; +static zone_t kqworkq_zone; +static zone_t kqworkloop_zone; +#if DEVELOPMENT || DEBUG +#define KEVENT_PANIC_ON_WORKLOOP_OWNERSHIP_LEAK (1U << 0) +#define KEVENT_PANIC_ON_NON_ENQUEUED_PROCESS (1U << 1) +#define KEVENT_PANIC_BOOT_ARG_INITIALIZED (1U << 31) + +#define KEVENT_PANIC_DEFAULT_VALUE (0) +static uint32_t +kevent_debug_flags(void) +{ + static uint32_t flags = KEVENT_PANIC_DEFAULT_VALUE; -static int filt_procattach(struct knote *kn); -static void filt_procdetach(struct knote *kn); -static int filt_proc(struct knote *kn, long hint); + if ((flags & KEVENT_PANIC_BOOT_ARG_INITIALIZED) == 0) { + uint32_t value = 0; + if (!PE_parse_boot_argn("kevent_debug", &value, sizeof(value))) { + value = KEVENT_PANIC_DEFAULT_VALUE; + } + value |= KEVENT_PANIC_BOOT_ARG_INITIALIZED; + os_atomic_store(&flags, value, relaxed); + } + return flags; +} +#endif -static struct filterops proc_filtops = - { 0, filt_procattach, filt_procdetach, filt_proc }; +#define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask)) -extern struct filterops fs_filtops; +/* placeholder for not-yet-implemented filters */ +static int filt_badattach(struct knote *kn, struct kevent_internal_s *kev); +static int filt_badevent(struct knote *kn, long hint); +SECURITY_READ_ONLY_EARLY(static struct filterops) bad_filtops = { + .f_attach = filt_badattach, +}; -extern struct filterops sig_filtops; +#if CONFIG_MEMORYSTATUS +extern const struct filterops memorystatus_filtops; +#endif /* CONFIG_MEMORYSTATUS */ +extern const struct filterops fs_filtops; +extern const struct filterops sig_filtops; +extern const struct filterops machport_filtops; +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; +extern const struct filterops soexcept_filtops; +extern const struct filterops spec_filtops; +extern const struct filterops bpfread_filtops; +extern const struct filterops necp_fd_rfiltops; +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; +/* + * + * Rules for adding new filters to the system: + * Public filters: + * - 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 + * 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. + */ +SECURITY_READ_ONLY_EARLY(static struct filterops *) 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, + [~EVFILT_VM] = &bad_filtops, + [~EVFILT_SOCK] = &file_filtops, +#if CONFIG_MEMORYSTATUS + [~EVFILT_MEMORYSTATUS] = &memorystatus_filtops, +#else + [~EVFILT_MEMORYSTATUS] = &bad_filtops, +#endif + [~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_PTMX] = &ptmx_kqops, +}; -/* Timer filter */ -static int filt_timercompute(struct knote *kn, uint64_t *abs_time); -static void filt_timerexpire(void *knx, void *param1); -static int filt_timerattach(struct knote *kn); -static void filt_timerdetach(struct knote *kn); -static int filt_timer(struct knote *kn, long hint); +/* waitq prepost callback */ +void waitq_set__CALLING_PREPOST_HOOK__(void *kq_hook, void *knote_hook, int qos); -static struct filterops timer_filtops = - { 0, filt_timerattach, filt_timerdetach, filt_timer }; +static inline struct kqworkloop * +kqr_kqworkloop(struct kqrequest *kqr) +{ + if (kqr->kqr_state & KQR_WORKLOOP) { + return __container_of(kqr, struct kqworkloop, kqwl_request); + } + return NULL; +} -/* to avoid arming timers that fire quicker than we can handle */ -static uint64_t filt_timerfloor = 0; +static inline kqueue_t +kqr_kqueue(proc_t p, struct kqrequest *kqr) +{ + kqueue_t kqu; + if (kqr->kqr_state & KQR_WORKLOOP) { + kqu.kqwl = kqr_kqworkloop(kqr); + } else { + kqu.kqwq = (struct kqworkq *)p->p_fd->fd_wqkqueue; + assert(kqr >= kqu.kqwq->kqwq_request && + kqr < kqu.kqwq->kqwq_request + KQWQ_NBUCKETS); + } + return kqu; +} -static lck_mtx_t _filt_timerlock; -static void filt_timerlock(void); -static void filt_timerunlock(void); +static inline boolean_t +is_workqueue_thread(thread_t thread) +{ + return (thread_get_tag(thread) & THREAD_TAG_WORKQUEUE); +} /* - * Sentinel marker for a thread scanning through the list of - * active knotes. + * kqueue/note lock implementations + * + * The kqueue lock guards the kq state, the state of its queues, + * 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 + * be guarded by the kq lock - as it needs to be taken when we + * already have the waitq set lock held (during the waitq hook + * callback). It might be better to use the waitq lock itself + * for this, but the IRQ requirements make that difficult). + * + * Knote flags, filter flags, and associated data are protected + * by the underlying object lock - and are only ever looked at + * by calling the filter to get a [consistent] snapshot of that + * data. */ -static struct filterops threadmarker_filtops = - { 0, filt_badattach, 0, 0 }; +static lck_grp_attr_t *kq_lck_grp_attr; +static lck_grp_t *kq_lck_grp; +static lck_attr_t *kq_lck_attr; -static zone_t knote_zone; +static inline void +kqlock(kqueue_t kqu) +{ + lck_spin_lock(&kqu.kq->kq_lock); +} -#define KN_HASHSIZE 64 /* XXX should be tunable */ -#define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask)) +static inline void +kqlock_held(__assert_only kqueue_t kqu) +{ + LCK_SPIN_ASSERT(&kqu.kq->kq_lock, LCK_ASSERT_OWNED); +} -#if 0 -extern struct filterops aio_filtops; -#endif +static inline void +kqunlock(kqueue_t kqu) +{ + lck_spin_unlock(&kqu.kq->kq_lock); +} -/* - * Table for for all system-defined filters. - */ -static struct filterops *sysfilt_ops[] = { - &file_filtops, /* EVFILT_READ */ - &file_filtops, /* EVFILT_WRITE */ -#if 0 - &aio_filtops, /* EVFILT_AIO */ -#else - &bad_filtops, /* EVFILT_AIO */ -#endif - &file_filtops, /* EVFILT_VNODE */ - &proc_filtops, /* EVFILT_PROC */ - &sig_filtops, /* EVFILT_SIGNAL */ - &timer_filtops, /* EVFILT_TIMER */ - &bad_filtops, /* EVFILT_MACHPORT */ - &fs_filtops /* EVFILT_FS */ -}; +static inline void +kq_req_lock(kqueue_t kqu) +{ + assert(kqu.kq->kq_state & (KQ_WORKLOOP | KQ_WORKQ)); + lck_spin_lock(&kqu.kq->kq_reqlock); +} -/* - * kqueue/note lock attributes and implementations - * - * kqueues have locks, while knotes have use counts - * Most of the knote state is guarded by the object lock. - * the knote "inuse" count and status use the kqueue lock. - */ -lck_grp_attr_t * kq_lck_grp_attr; -lck_grp_t * kq_lck_grp; -lck_attr_t * kq_lck_attr; +static inline void +kq_req_unlock(kqueue_t kqu) +{ + assert(kqu.kq->kq_state & (KQ_WORKLOOP | KQ_WORKQ)); + lck_spin_unlock(&kqu.kq->kq_reqlock); +} + +static inline void +kq_req_held(__assert_only kqueue_t kqu) +{ + assert(kqu.kq->kq_state & (KQ_WORKLOOP | KQ_WORKQ)); + LCK_SPIN_ASSERT(&kqu.kq->kq_reqlock, LCK_ASSERT_OWNED); +} static inline void -kqlock(struct kqueue *kq) +knhash_lock(proc_t p) { - lck_spin_lock(&kq->kq_lock); + lck_mtx_lock(&p->p_fd->fd_knhashlock); } static inline void -kqunlock(struct kqueue *kq) +knhash_unlock(proc_t p) { - lck_spin_unlock(&kq->kq_lock); + lck_mtx_unlock(&p->p_fd->fd_knhashlock); } -/* - * Convert a kq lock to a knote use referece. +#pragma mark knote locks + +/* + * Enum used by the knote_lock_* functions. * - * If the knote is being dropped, we can't get - * a use reference, so just return with it - * still locked. - * - * - kq locked at entry - * - unlock on exit if we get the use reference + * KNOTE_KQ_LOCK_ALWAYS + * The function will always return with the kq lock held. + * + * KNOTE_KQ_UNLOCK_ON_SUCCESS + * The function will return with the kq lock held if it was successful + * (knote_lock() is the only function that can fail). + * + * KNOTE_KQ_UNLOCK_ON_FAILURE + * The function will return with the kq lock held if it was unsuccessful + * (knote_lock() is the only function that can fail). + * + * KNOTE_KQ_UNLOCK: + * The function returns with the kq unlocked. */ -static int -kqlock2knoteuse(struct kqueue *kq, struct knote *kn) +#define KNOTE_KQ_LOCK_ALWAYS 0x0 +#define KNOTE_KQ_LOCK_ON_SUCCESS 0x1 +#define KNOTE_KQ_LOCK_ON_FAILURE 0x2 +#define KNOTE_KQ_UNLOCK 0x3 + +#if DEBUG || DEVELOPMENT +__attribute__((noinline, not_tail_called, disable_tail_calls)) +void knote_lock_ctx_chk(struct knote_lock_ctx *knlc) { - if (kn->kn_status & KN_DROPPING) - return 0; - kn->kn_inuse++; + /* evil hackery to make sure no one forgets to unlock */ + assert(knlc->knlc_state == KNOTE_LOCK_CTX_UNLOCKED); +} +#endif + +static struct knote_lock_ctx * +knote_lock_ctx_find(struct kqueue *kq, struct knote *kn) +{ + struct knote_lock_ctx *ctx; + LIST_FOREACH(ctx, &kq->kq_knlocks, knlc_le) { + if (ctx->knlc_knote == kn) return ctx; + } + panic("knote lock context not found: %p", kn); + __builtin_trap(); +} + +/* slowpath of knote_lock() */ +__attribute__((noinline)) +static bool __result_use_check +knote_lock_slow(struct kqueue *kq, struct knote *kn, + struct knote_lock_ctx *knlc, int kqlocking) +{ + kqlock_held(kq); + + struct knote_lock_ctx *owner_lc = knote_lock_ctx_find(kq, kn); + thread_t owner_thread = owner_lc->knlc_thread; + +#if DEBUG || DEVELOPMENT + knlc->knlc_state = KNOTE_LOCK_CTX_WAITING; +#endif + + thread_reference(owner_thread); + TAILQ_INSERT_TAIL(&owner_lc->knlc_head, knlc, knlc_tqe); + assert_wait(&kn->kn_status, THREAD_UNINT | THREAD_WAIT_NOREPORT); kqunlock(kq); - return 1; - } -/* - * Convert a kq lock to a knote use referece. + if (thread_handoff_deallocate(owner_thread) == THREAD_RESTART) { + if (kqlocking == KNOTE_KQ_LOCK_ALWAYS || + kqlocking == KNOTE_KQ_LOCK_ON_FAILURE) { + kqlock(kq); + } +#if DEBUG || DEVELOPMENT + assert(knlc->knlc_state == KNOTE_LOCK_CTX_WAITING); + knlc->knlc_state = KNOTE_LOCK_CTX_UNLOCKED; +#endif + return false; + } +#if DEBUG || DEVELOPMENT + assert(knlc->knlc_state == KNOTE_LOCK_CTX_LOCKED); +#endif + if (kqlocking == KNOTE_KQ_LOCK_ALWAYS || + kqlocking == KNOTE_KQ_LOCK_ON_SUCCESS) { + kqlock(kq); + } + return true; +} + +/* + * Attempts to take the "knote" lock. * - * If the knote is being dropped, we can't get - * a use reference, so just return with it - * still locked. - * - * - kq locked at entry - * - kq always unlocked on exit + * Called with the kqueue lock held. + * + * Returns true if the knote lock is acquired, false if it has been dropped */ -static int -kqlock2knoteusewait(struct kqueue *kq, struct knote *kn) +static bool __result_use_check +knote_lock(struct kqueue *kq, struct knote *kn, struct knote_lock_ctx *knlc, + int kqlocking) { - if (!kqlock2knoteuse(kq, kn)) { - kn->kn_status |= KN_DROPWAIT; - assert_wait(&kn->kn_status, THREAD_UNINT); + kqlock_held(kq); + +#if DEBUG || DEVELOPMENT + assert(knlc->knlc_state == KNOTE_LOCK_CTX_UNLOCKED); +#endif + knlc->knlc_knote = kn; + knlc->knlc_thread = current_thread(); + TAILQ_INIT(&knlc->knlc_head); + + if (__improbable(kn->kn_status & KN_LOCKED)) { + return knote_lock_slow(kq, kn, knlc, kqlocking); + } + + /* + * 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(&kq->kq_knlocks, knlc, knlc_le); + 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(kq); - thread_block(THREAD_CONTINUE_NULL); - return 0; } - return 1; - } + return true; +} -/* - * Convert from a knote use reference back to kq lock. +/* + * Unlocks a knote successfully locked with knote_lock(). * - * Drop a use reference and wake any waiters if - * this is the last one. + * Called with the kqueue lock held. * - * The exit return indicates if the knote is - * still alive - but the kqueue lock is taken - * unconditionally. + * Returns with the kqueue lock held according to KNOTE_KQ_* flags */ -static int -knoteuse2kqlock(struct kqueue *kq, struct knote *kn) +static void +knote_unlock(struct kqueue *kq, struct knote *kn, + struct knote_lock_ctx *knlc, int flags) { - kqlock(kq); - if ((--kn->kn_inuse == 0) && - (kn->kn_status & KN_USEWAIT)) { - kn->kn_status &= ~KN_USEWAIT; - thread_wakeup(&kn->kn_inuse); - } - return ((kn->kn_status & KN_DROPPING) == 0); - } - -/* - * Convert a kq lock to a knote drop referece. - * - * 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. - * - * - kq locked at entry - * - always unlocked on exit. - * - caller can't hold any locks that would prevent - * the other dropper from completing. + kqlock_held(kq); + + assert(knlc->knlc_knote == kn); + assert(kn->kn_status & KN_LOCKED); +#if DEBUG || DEVELOPMENT + assert(knlc->knlc_state == KNOTE_LOCK_CTX_LOCKED); +#endif + + struct knote_lock_ctx *next_owner_lc = TAILQ_FIRST(&knlc->knlc_head); + + LIST_REMOVE(knlc, knlc_le); + + if (next_owner_lc) { + assert(next_owner_lc->knlc_knote == kn); + TAILQ_REMOVE(&knlc->knlc_head, next_owner_lc, knlc_tqe); + + assert(TAILQ_EMPTY(&next_owner_lc->knlc_head)); + TAILQ_CONCAT(&next_owner_lc->knlc_head, &knlc->knlc_head, knlc_tqe); + LIST_INSERT_HEAD(&kq->kq_knlocks, next_owner_lc, knlc_le); +#if DEBUG || DEVELOPMENT + next_owner_lc->knlc_state = KNOTE_LOCK_CTX_LOCKED; +#endif + } else { + kn->kn_status &= ~KN_LOCKED; + } + if (kn->kn_inuse == 0) { + /* + * No f_event() in flight anymore, we can leave QoS "Merge" mode + * + * See knote_should_apply_qos_override() + */ + kn->kn_status &= ~KN_MERGE_QOS; + } + if (flags & KNOTE_KQ_UNLOCK) { + kqunlock(kq); + } + if (next_owner_lc) { + thread_wakeup_thread(&kn->kn_status, next_owner_lc->knlc_thread); + } +#if DEBUG || DEVELOPMENT + knlc->knlc_state = KNOTE_LOCK_CTX_UNLOCKED; +#endif +} + +/* + * Aborts all waiters for a knote lock, and unlock the knote. + * + * Called with the kqueue lock held. + * + * Returns with the kqueue lock held according to KNOTE_KQ_* flags */ -static int -kqlock2knotedrop(struct kqueue *kq, struct knote *kn) +static void +knote_unlock_cancel(struct kqueue *kq, struct knote *kn, + struct knote_lock_ctx *knlc, int kqlocking) { + kqlock_held(kq); - if ((kn->kn_status & KN_DROPPING) == 0) { - kn->kn_status |= KN_DROPPING; - if (kn->kn_inuse > 0) { - kn->kn_status |= KN_USEWAIT; - assert_wait(&kn->kn_inuse, THREAD_UNINT); - kqunlock(kq); - thread_block(THREAD_CONTINUE_NULL); - } else - kqunlock(kq); - return 1; - } else { - kn->kn_status |= KN_DROPWAIT; - assert_wait(&kn->kn_status, THREAD_UNINT); + assert(knlc->knlc_knote == kn); + assert(kn->kn_status & KN_LOCKED); + assert(kn->kn_status & KN_DROPPING); + + LIST_REMOVE(knlc, knlc_le); + kn->kn_status &= ~KN_LOCKED; + + if (kqlocking == KNOTE_KQ_UNLOCK || + kqlocking == KNOTE_KQ_LOCK_ON_FAILURE) { kqunlock(kq); - thread_block(THREAD_CONTINUE_NULL); - return 0; } + if (!TAILQ_EMPTY(&knlc->knlc_head)) { + thread_wakeup_with_result(&kn->kn_status, THREAD_RESTART); + } +#if DEBUG || DEVELOPMENT + knlc->knlc_state = KNOTE_LOCK_CTX_UNLOCKED; +#endif } - -/* - * Release a knote use count reference. + +/* + * Call the f_event hook of a given filter. + * + * Takes a use count to protect against concurrent drops. */ static void -knote_put(struct knote *kn) +knote_call_filter_event(struct kqueue *kq, struct knote *kn, long hint) { - struct kqueue *kq = kn->kn_kq; + int result, dropping = 0; + kqlock_held(kq); + + if (kn->kn_status & (KN_DROPPING | KN_VANISHED)) + return; + + kn->kn_inuse++; + kqunlock(kq); + result = filter_call(knote_fops(kn), f_event(kn, hint)); kqlock(kq); - if ((--kn->kn_inuse == 0) && - (kn->kn_status & KN_USEWAIT)) { - kn->kn_status &= ~KN_USEWAIT; - thread_wakeup(&kn->kn_inuse); + + dropping = (kn->kn_status & KN_DROPPING); + + if (!dropping && (result & FILTER_ACTIVE)) { + if (result & FILTER_ADJUST_EVENT_QOS_BIT) + knote_adjust_qos(kq, kn, result); + knote_activate(kn); } - kqunlock(kq); - } + if (--kn->kn_inuse == 0) { + if ((kn->kn_status & KN_LOCKED) == 0) { + /* + * We're the last f_event() call and there's no other f_* call in + * flight, we can leave QoS "Merge" mode. + * + * See knote_should_apply_qos_override() + */ + kn->kn_status &= ~KN_MERGE_QOS; + } + if (dropping) { + waitq_wakeup64_all((struct waitq *)&kq->kq_wqs, + CAST_EVENT64_T(&kn->kn_inuse), + THREAD_AWAKENED, WAITQ_ALL_PRIORITIES); + } + } +} + +/* + * 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_filter_events(struct kqueue *kq, struct knote *kn) +{ + wait_result_t wr = THREAD_NOT_WAITING; + + kqlock_held(kq); + + assert(kn->kn_status & KN_DROPPING); + + if (kn->kn_inuse) { + wr = waitq_assert_wait64((struct waitq *)&kq->kq_wqs, + CAST_EVENT64_T(&kn->kn_inuse), + THREAD_UNINT | THREAD_WAIT_NOREPORT, TIMEOUT_WAIT_FOREVER); + } + kqunlock(kq); + if (wr == THREAD_WAITING) { + thread_block(THREAD_CONTINUE_NULL); + } +} +#pragma mark file_filtops static int -filt_fileattach(struct knote *kn) +filt_fileattach(struct knote *kn, struct kevent_internal_s *kev) { - - return (fo_kqfilter(kn->kn_fp, kn, current_proc())); + return fo_kqfilter(kn->kn_fp, kn, kev, vfs_context_current()); } -#define f_flag f_fglob->fg_flag -#define f_type f_fglob->fg_type -#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 +SECURITY_READ_ONLY_EARLY(static struct filterops) file_filtops = { + .f_isfd = 1, + .f_attach = filt_fileattach, +}; + +#pragma mark kqread_filtops + +#define f_flag f_fglob->fg_flag +#define f_ops f_fglob->fg_ops +#define f_data f_fglob->fg_data static void filt_kqdetach(struct knote *kn) { - struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; + struct kqfile *kqf = (struct kqfile *)kn->kn_fp->f_data; + struct kqueue *kq = &kqf->kqf_kqueue; kqlock(kq); - KNOTE_DETACH(&kq->kq_sel.si_note, kn); + 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; + return (kq->kq_count > 0); +} + +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; + + kqlock(kq); + kn->kn_data = kq->kq_count; + res = (kn->kn_data > 0); + + kqunlock(kq); + + return res; +} + +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; + + kqlock(kq); kn->kn_data = kq->kq_count; - return (kn->kn_data > 0); + res = (kn->kn_data > 0); + if (res) { + *kev = kn->kn_kevent; + if (kn->kn_flags & EV_CLEAR) + kn->kn_data = 0; + } + 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) +filt_procattach(struct knote *kn, __unused struct kevent_internal_s *kev) { struct proc *p; - int funnel_state; - - funnel_state = thread_funnel_set(kernel_flock, TRUE); - p = pfind(kn->kn_id); + 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(kn->kn_id); if (p == NULL) { - thread_funnel_set(kernel_flock, funnel_state); - return (ESRCH); + knote_set_error(kn, ESRCH); + return 0; } - kn->kn_flags |= EV_CLEAR; /* automatically set */ + const int NoteExitStatusBits = NOTE_EXIT | NOTE_EXITSTATUS; - /* - * internal flag indicating registration done by kernel - */ - if (kn->kn_flags & EV_FLAG1) { - kn->kn_data = (int)kn->kn_sdata; /* ppid */ - kn->kn_fflags = NOTE_CHILD; - kn->kn_flags &= ~EV_FLAG1; - } + if ((kn->kn_sfflags & NoteExitStatusBits) == NoteExitStatusBits) + do { + pid_t selfpid = proc_selfpid(); + + if (p->p_ppid == selfpid) + break; /* parent => ok */ + + if ((p->p_lflag & P_LTRACED) != 0 && + (p->p_oppid == selfpid)) + break; /* parent-in-waiting => ok */ + + proc_rele(p); + knote_set_error(kn, EACCES); + return 0; + } while (0); + + proc_klist_lock(); + + kn->kn_ptr.p_proc = p; /* store the proc handle */ - /* XXX lock the proc here while adding to the list? */ KNOTE_ATTACH(&p->p_klist, kn); - thread_funnel_set(kernel_flock, funnel_state); + proc_klist_unlock(); + + proc_rele(p); + /* + * only captures edge-triggered events after this point + * so it can't already be fired. + */ return (0); } + /* * The knote may be attached to a different process, which may exit, - * leaving nothing for the knote to be attached to. So when the process - * exits, the knote is marked as DETACHED and also flagged as ONESHOT so - * it will be deleted when read out. However, as part of the knote deletion, - * this routine is called, so a check is needed to avoid actually performing - * a detach, because the original process does not exist any more. + * leaving nothing for the knote to be attached to. In that case, + * the pointer to the process will have already been nulled out. */ static void filt_procdetach(struct knote *kn) { struct proc *p; - int funnel_state; - funnel_state = thread_funnel_set(kernel_flock, TRUE); - p = pfind(kn->kn_id); + proc_klist_lock(); - if (p != (struct proc *)NULL) + p = kn->kn_ptr.p_proc; + if (p != PROC_NULL) { + kn->kn_ptr.p_proc = PROC_NULL; KNOTE_DETACH(&p->p_klist, kn); + } - thread_funnel_set(kernel_flock, funnel_state); + proc_klist_unlock(); } static int filt_proc(struct knote *kn, long hint) { u_int event; - int funnel_state; - funnel_state = thread_funnel_set(kernel_flock, TRUE); + /* ALWAYS CALLED WITH proc_klist_lock */ /* + * 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 + * * mask off extra data */ event = (u_int)hint & NOTE_PCTRLMASK; + /* + * termination lifecycle events can happen while a debugger + * has reparented a process, in which case notifications + * should be quashed except to the tracing parent. When + * the debugger reaps the child (either via wait4(2) or + * process exit), the child will be reparented to the original + * 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)) { + /* + * 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) kn->kn_fflags |= event; - /* - * process is gone, so flag the event as finished. - */ - if (event == NOTE_EXIT) { - kn->kn_flags |= (EV_EOF | EV_ONESHOT); - thread_funnel_set(kernel_flock, funnel_state); - return (1); +#pragma clang diagnostic push +#pragma clang diagnostic ignored "-Wdeprecated-declarations" + if ((event == NOTE_REAP) || ((event == NOTE_EXIT) && !(kn->kn_sfflags & NOTE_REAP))) { + kn->kn_flags |= (EV_EOF | EV_ONESHOT); } +#pragma clang diagnostic pop + /* - * process forked, and user wants to track the new process, - * so attach a new knote to it, and immediately report an - * event with the parent's pid. + * The kernel has a wrapper in place that returns the same data + * as is collected here, in kn_data. Any changes to how + * NOTE_EXITSTATUS and NOTE_EXIT_DETAIL are collected + * should also be reflected in the proc_pidnoteexit() wrapper. */ - if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) { - struct kevent kev; - int error; - - /* - * register knote with new process. - */ - kev.ident = hint & NOTE_PDATAMASK; /* pid */ - kev.filter = kn->kn_filter; - kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1; - kev.fflags = kn->kn_sfflags; - kev.data = kn->kn_id; /* parent */ - kev.udata = kn->kn_kevent.udata; /* preserve udata */ - error = kevent_register(kn->kn_kq, &kev, NULL); - if (error) - kn->kn_fflags |= NOTE_TRACKERR; - } - event = kn->kn_fflags; - thread_funnel_set(kernel_flock, funnel_state); + if (event == NOTE_EXIT) { + kn->kn_data = 0; + if ((kn->kn_sfflags & NOTE_EXITSTATUS) != 0) { + kn->kn_fflags |= NOTE_EXITSTATUS; + kn->kn_data |= (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_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) { + case P_JETSAM_VMPAGESHORTAGE: + kn->kn_data |= NOTE_EXIT_MEMORY_VMPAGESHORTAGE; + break; + case P_JETSAM_VMTHRASHING: + kn->kn_data |= NOTE_EXIT_MEMORY_VMTHRASHING; + break; + case P_JETSAM_FCTHRASHING: + kn->kn_data |= NOTE_EXIT_MEMORY_FCTHRASHING; + break; + case P_JETSAM_VNODE: + kn->kn_data |= NOTE_EXIT_MEMORY_VNODE; + break; + case P_JETSAM_HIWAT: + kn->kn_data |= NOTE_EXIT_MEMORY_HIWAT; + break; + case P_JETSAM_PID: + kn->kn_data |= NOTE_EXIT_MEMORY_PID; + break; + case P_JETSAM_IDLEEXIT: + kn->kn_data |= NOTE_EXIT_MEMORY_IDLE; + break; + } + } + if ((kn->kn_ptr.p_proc->p_csflags & + CS_KILLED) != 0) { + kn->kn_data |= NOTE_EXIT_CSERROR; + } + } + } + + /* if we have any matching state, activate the knote */ + return (kn->kn_fflags != 0); +} + +static int +filt_proctouch(struct knote *kn, struct kevent_internal_s *kev) +{ + int res; + + proc_klist_lock(); + + /* accept new filter flags and mask off output events no long interesting */ + kn->kn_sfflags = kev->fflags; + + /* restrict the current results to the (smaller?) set of new interest */ + /* + * For compatibility with previous implementations, we leave kn_fflags + * as they were before. + */ + //kn->kn_fflags &= kn->kn_sfflags; + + res = (kn->kn_fflags != 0); + + proc_klist_unlock(); + + return res; +} + +static int +filt_procprocess(struct knote *kn, struct filt_process_s *data, struct kevent_internal_s *kev) +{ +#pragma unused(data) + int res; + + 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; + } + 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_proc, + .f_touch = filt_proctouch, + .f_process = filt_procprocess, +}; + +#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_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_hookid timer state + * + * 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 - return (event != 0); +static void +filt_timer_set_params(struct knote *kn, struct filt_timer_params *params) +{ + kn->kn_ext[0] = params->deadline; + kn->kn_ext[1] = params->leeway; + kn->kn_sdata = params->interval; } /* - * filt_timercompute - compute absolute timeout + * filt_timervalidate - process data from user + * + * Sets up the deadline, interval, and leeway from the provided user data * - * The saved-data field in the knote contains the - * time value. The saved filter-flags indicates - * the unit of measurement. + * Input: + * kn_sdata timer deadline or interval time + * kn_sfflags style of timer, unit of measurement * - * If the timeout is not absolute, adjust it for - * the current time. + * Output: + * 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_timercompute(struct knote *kn, uint64_t *abs_time) +filt_timervalidate(const struct kevent_internal_s *kev, + struct filt_timer_params *params) { + /* + * 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: + * NOTE_SECONDS, NOTE_USECONDS, NOTE_NSECONDS, NOTE_MACHTIME + * Defaults to milliseconds if not specified + * + * B) Clock epoch (what is the zero point of the specified number) + * For interval, there is none + * For absolute, defaults to the gettimeofday/calendar epoch + * With NOTE_MACHTIME, uses mach_absolute_time() + * With NOTE_MACHTIME and NOTE_MACH_CONTINUOUS_TIME, uses mach_continuous_time() + * + * C) The knote's behavior on delivery + * Interval timer causes the knote to arm for the next interval unless one-shot is set + * Absolute is a forced one-shot timer which deletes on delivery + * TODO: Add a way for absolute to be not forced one-shot + * + * D) Whether the time duration is relative to now or absolute + * Interval fires at now + duration when it is set up + * Absolute fires at now + difference between now walltime and passed in walltime + * With NOTE_MACHTIME it fires at an absolute MAT or MCT. + * + * E) Whether the timer continues to tick across sleep + * By default all three do not. + * For interval and absolute, NOTE_MACH_CONTINUOUS_TIME causes them to tick across sleep + * With NOTE_ABSOLUTE | NOTE_MACHTIME | NOTE_MACH_CONTINUOUS_TIME: + * expires when mach_continuous_time() is > the passed in value. + */ + uint64_t multiplier; - uint64_t raw; - switch (kn->kn_sfflags & (NOTE_SECONDS|NOTE_USECONDS|NOTE_NSECONDS)) { + boolean_t use_abstime = FALSE; + + switch (kev->fflags & (NOTE_SECONDS|NOTE_USECONDS|NOTE_NSECONDS|NOTE_MACHTIME)) { case NOTE_SECONDS: multiplier = NSEC_PER_SEC; break; @@ -535,1742 +1202,7734 @@ filt_timercompute(struct knote *kn, uint64_t *abs_time) case NOTE_NSECONDS: multiplier = 1; break; + case NOTE_MACHTIME: + multiplier = 0; + use_abstime = TRUE; + break; case 0: /* milliseconds (default) */ multiplier = NSEC_PER_SEC / 1000; break; default: - return EINVAL; - } - nanoseconds_to_absolutetime((uint64_t)kn->kn_sdata * multiplier, &raw); - if (raw <= filt_timerfloor) { - *abs_time = 0; - return 0; + return (EINVAL); } - if ((kn->kn_sfflags & NOTE_ABSOLUTE) == NOTE_ABSOLUTE) { - uint32_t seconds, nanoseconds; - uint64_t now; - clock_get_calendar_nanotime(&seconds, &nanoseconds); - nanoseconds_to_absolutetime((uint64_t)seconds * NSEC_PER_SEC + nanoseconds, - &now); - if (now >= raw + filt_timerfloor) { - *abs_time = 0; - return 0; - } - raw -= now; - } - clock_absolutetime_interval_to_deadline(raw, abs_time); - return 0; -} + /* transform the leeway in kn_ext[1] to same time scale */ + if (kev->fflags & NOTE_LEEWAY) { + uint64_t leeway_abs; -/* - * 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) -{ - struct klist timer_list; - struct knote *kn = knx; + if (use_abstime) { + leeway_abs = (uint64_t)kev->ext[1]; + } else { + uint64_t leeway_ns; + if (os_mul_overflow((uint64_t)kev->ext[1], multiplier, &leeway_ns)) + return (ERANGE); - /* no "object" for timers, so fake a list */ - SLIST_INIT(&timer_list); - SLIST_INSERT_HEAD(&timer_list, kn, kn_selnext); - KNOTE(&timer_list, 1); -} + nanoseconds_to_absolutetime(leeway_ns, &leeway_abs); + } -/* - * data contains amount of time to sleep, in milliseconds, - * or a pointer to a timespec structure. - */ -static int -filt_timerattach(struct knote *kn) -{ - thread_call_t callout; - uint64_t deadline; - int error; + params->leeway = leeway_abs; + } else { + params->leeway = 0; + } - error = filt_timercompute(kn, &deadline); - if (error) - return (error); + if (kev->fflags & NOTE_ABSOLUTE) { + uint64_t deadline_abs; - if (deadline) { - callout = thread_call_allocate(filt_timerexpire, kn); - if (NULL == callout) - return (ENOMEM); - } else { - /* handle as immediate */ - kn->kn_sdata = 0; - callout = NULL; - } + if (use_abstime) { + deadline_abs = (uint64_t)kev->data; + } else { + uint64_t calendar_deadline_ns; - filt_timerlock(); - kn->kn_hook = (caddr_t)callout; + if (os_mul_overflow((uint64_t)kev->data, multiplier, &calendar_deadline_ns)) + return (ERANGE); - /* absolute=EV_ONESHOT */ - if (kn->kn_sfflags & NOTE_ABSOLUTE) - kn->kn_flags |= EV_ONESHOT; + /* calendar_deadline_ns is in nanoseconds since the epoch */ - if (deadline) { - /* all others - if not faking immediate */ - kn->kn_flags |= EV_CLEAR; - thread_call_enter_delayed(callout, deadline); - kn->kn_hookid = 0; - } else { - /* fake immediate */ - kn->kn_hookid = 1; - } - filt_timerunlock(); - return (0); -} + clock_sec_t seconds; + clock_nsec_t nanoseconds; -static void -filt_timerdetach(struct knote *kn) -{ - thread_call_t callout; + /* + * Note that the conversion through wall-time is only done once. + * + * If the relationship between MAT and gettimeofday changes, + * the underlying timer does not update. + * + * TODO: build a wall-time denominated timer_call queue + * and a flag to request DTRTing with wall-time timers + */ + clock_get_calendar_nanotime(&seconds, &nanoseconds); - filt_timerlock(); - callout = (thread_call_t)kn->kn_hook; - if (callout != NULL) { - boolean_t cancelled; - - /* cancel the callout if we can */ - cancelled = thread_call_cancel(callout); - if (cancelled) { - /* got it, just free it */ - kn->kn_hook = NULL; - filt_timerunlock(); - thread_call_free(callout); - return; - } - /* we have to wait for the expire routine. */ - kn->kn_hookid = -1; /* we are detaching */ - assert_wait(&kn->kn_hook, THREAD_UNINT); - filt_timerunlock(); - thread_block(THREAD_CONTINUE_NULL); - assert(kn->kn_hook == NULL); - return; - } - /* nothing to do */ - filt_timerunlock(); -} + uint64_t calendar_now_ns = (uint64_t)seconds * NSEC_PER_SEC + nanoseconds; + /* if deadline is in the future */ + if (calendar_now_ns < calendar_deadline_ns) { + uint64_t interval_ns = calendar_deadline_ns - calendar_now_ns; + uint64_t interval_abs; + nanoseconds_to_absolutetime(interval_ns, &interval_abs); -static int -filt_timer(struct knote *kn, __unused long hint) -{ - int result; - - if (hint) { - /* real timer pop */ - thread_call_t callout; - boolean_t detaching; - - filt_timerlock(); - - kn->kn_data++; - - detaching = (kn->kn_hookid < 0); - callout = (thread_call_t)kn->kn_hook; - - if (!detaching && (kn->kn_flags & EV_ONESHOT) == 0) { - uint64_t deadline; - int error; - - /* user input data may have changed - deal */ - error = filt_timercompute(kn, &deadline); - if (error) { - kn->kn_flags |= EV_ERROR; - kn->kn_data = error; - } else if (deadline == 0) { - /* revert to fake immediate */ - kn->kn_flags &= ~EV_CLEAR; - kn->kn_sdata = 0; - kn->kn_hookid = 1; + /* + * Note that the NOTE_MACH_CONTINUOUS_TIME flag here only + * causes the timer to keep ticking across sleep, but + * it does not change the calendar timebase. + */ + + if (kev->fflags & NOTE_MACH_CONTINUOUS_TIME) + clock_continuoustime_interval_to_deadline(interval_abs, + &deadline_abs); + else + clock_absolutetime_interval_to_deadline(interval_abs, + &deadline_abs); } else { - /* keep the callout and re-arm */ - thread_call_enter_delayed(callout, deadline); - filt_timerunlock(); - return 1; + deadline_abs = 0; /* cause immediate expiration */ } } - kn->kn_hook = NULL; - filt_timerunlock(); - thread_call_free(callout); - - /* if someone is waiting for timer to pop */ - if (detaching) - thread_wakeup(&kn->kn_hook); - return 1; - } + params->deadline = deadline_abs; + params->interval = 0; /* NOTE_ABSOLUTE is non-repeating */ + } else if (kev->data < 0) { + /* + * Negative interval timers fire immediately, once. + * + * Ideally a negative interval would be an error, but certain clients + * pass negative values on accident, and expect an event back. + * + * In the old implementation the timer would repeat with no delay + * N times until mach_absolute_time() + (N * interval) underflowed, + * then it would wait ~forever by accidentally arming a timer for the far future. + * + * We now skip the power-wasting hot spin phase and go straight to the idle phase. + */ - /* user-query */ - filt_timerlock(); + params->deadline = 0; /* expire immediately */ + params->interval = 0; /* non-repeating */ + } else { + uint64_t interval_abs = 0; - /* change fake timer to real if needed */ - while (kn->kn_hookid > 0 && kn->kn_sdata > 0) { - int error; + if (use_abstime) { + interval_abs = (uint64_t)kev->data; + } else { + uint64_t interval_ns; + if (os_mul_overflow((uint64_t)kev->data, multiplier, &interval_ns)) + return (ERANGE); - /* update the fake timer (make real) */ - kn->kn_hookid = 0; - kn->kn_data = 0; - filt_timerunlock(); - error = filt_timerattach(kn); - filt_timerlock(); - if (error) { - kn->kn_flags |= EV_ERROR; - kn->kn_data = error; - filt_timerunlock(); - return 1; + nanoseconds_to_absolutetime(interval_ns, &interval_abs); } - } - /* if still fake, pretend it fired */ - if (kn->kn_hookid > 0) - kn->kn_data = 1; + uint64_t deadline = 0; - result = (kn->kn_data != 0); - filt_timerunlock(); - return result; + if (kev->fflags & NOTE_MACH_CONTINUOUS_TIME) + clock_continuoustime_interval_to_deadline(interval_abs, &deadline); + else + clock_absolutetime_interval_to_deadline(interval_abs, &deadline); + + params->deadline = deadline; + params->interval = interval_abs; + } + + return (0); } +/* + * filt_timerexpire - the timer callout routine + */ static void -filt_timerlock(void) +filt_timerexpire(void *knx, __unused void *spare) { - lck_mtx_lock(&_filt_timerlock); + struct knote *kn = knx; + int v; + + if (os_atomic_cmpxchgv(&kn->kn_hookid, TIMER_ARMED, TIMER_FIRED, + &v, relaxed)) { + // our f_event always would say FILTER_ACTIVE, + // so be leaner and just do it. + struct kqueue *kq = knote_get_kq(kn); + kqlock(kq); + knote_activate(kn); + kqunlock(kq); + } else { + /* + * From TIMER_ARMED, the only allowed transition are: + * - to TIMER_FIRED through the timer callout just above + * - to TIMER_IDLE due to filt_timercancel() which will wait for the + * timer callout (and any possible invocation of filt_timerexpire) to + * have finished before the state is changed again. + */ + assert(v == TIMER_IDLE); + } } static void -filt_timerunlock(void) +filt_timercancel(struct knote *kn) { - lck_mtx_unlock(&_filt_timerlock); + if (os_atomic_xchg(&kn->kn_hookid, TIMER_IDLE, relaxed) == TIMER_ARMED) { + /* cancel the thread call and wait for any filt_timerexpire in flight */ + thread_call_cancel_wait((thread_call_t)kn->kn_hook); + } } /* - * JMM - placeholder for not-yet-implemented filters - */ -static int -filt_badattach(__unused struct knote *kn) -{ - return(ENOTSUP); -} - - -struct kqueue * -kqueue_alloc(struct proc *p) + * Does this deadline needs a timer armed for it, or has it expired? + */ +static bool +filt_timer_is_ready(struct knote *kn) { - struct filedesc *fdp = p->p_fd; - struct kqueue *kq; + uint64_t now, deadline = kn->kn_ext[0]; - MALLOC_ZONE(kq, struct kqueue *, sizeof(struct kqueue), M_KQUEUE, M_WAITOK); - if (kq != NULL) { - bzero(kq, sizeof(struct kqueue)); - lck_spin_init(&kq->kq_lock, kq_lck_grp, kq_lck_attr); - TAILQ_INIT(&kq->kq_head); - TAILQ_INIT(&kq->kq_inprocess); - kq->kq_fdp = fdp; + if (deadline == 0) { + return true; } - 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); + if (kn->kn_sfflags & NOTE_MACH_CONTINUOUS_TIME) { + now = mach_continuous_time(); + } else { + now = mach_absolute_time(); } - - return kq; + return deadline <= now; } - /* - * kqueue_dealloc - detach all knotes from a kqueue and free it + * Arm a timer * - * 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. + * It is the responsibility of the caller to make sure the timer call + * has completed or been cancelled properly prior to arming it. */ -void -kqueue_dealloc(struct kqueue *kq, struct proc *p) +static void +filt_timerarm(struct knote *kn) { - struct filedesc *fdp = p->p_fd; - struct knote *kn; - int i; + uint64_t deadline = kn->kn_ext[0]; + uint64_t leeway = kn->kn_ext[1]; - proc_fdlock(p); - for (i = 0; i < fdp->fd_knlistsize; i++) { - kn = SLIST_FIRST(&fdp->fd_knlist[i]); - while (kn != NULL) { - if (kq == kn->kn_kq) { - kqlock(kq); - proc_fdunlock(p); - /* drop it ourselves or wait */ - if (kqlock2knotedrop(kq, kn)) { - kn->kn_fop->f_detach(kn); - knote_drop(kn, p); - } - proc_fdlock(p); - /* start over at beginning of list */ - kn = SLIST_FIRST(&fdp->fd_knlist[i]); - continue; - } - kn = SLIST_NEXT(kn, kn_link); - } - } - if (fdp->fd_knhashmask != 0) { - for (i = 0; i < (int)fdp->fd_knhashmask + 1; i++) { - kn = SLIST_FIRST(&fdp->fd_knhash[i]); - while (kn != NULL) { - if (kq == kn->kn_kq) { - kqlock(kq); - proc_fdunlock(p); - /* drop it ourselves or wait */ - if (kqlock2knotedrop(kq, kn)) { - kn->kn_fop->f_detach(kn); - knote_drop(kn, p); - } - proc_fdlock(p); - /* start over at beginning of list */ - kn = SLIST_FIRST(&fdp->fd_knhash[i]); - continue; - } - kn = SLIST_NEXT(kn, kn_link); - } - } - } - proc_fdunlock(p); - lck_spin_destroy(&kq->kq_lock, kq_lck_grp); - FREE_ZONE(kq, sizeof(struct kqueue), M_KQUEUE); + int filter_flags = kn->kn_sfflags; + unsigned int timer_flags = 0; + + assert(os_atomic_load(&kn->kn_hookid, relaxed) == TIMER_IDLE); + + if (filter_flags & NOTE_CRITICAL) + timer_flags |= THREAD_CALL_DELAY_USER_CRITICAL; + else if (filter_flags & NOTE_BACKGROUND) + timer_flags |= THREAD_CALL_DELAY_USER_BACKGROUND; + else + timer_flags |= THREAD_CALL_DELAY_USER_NORMAL; + + if (filter_flags & NOTE_LEEWAY) + timer_flags |= THREAD_CALL_DELAY_LEEWAY; + + if (filter_flags & NOTE_MACH_CONTINUOUS_TIME) + timer_flags |= THREAD_CALL_CONTINUOUS; + + os_atomic_store(&kn->kn_hookid, TIMER_ARMED, relaxed); + thread_call_enter_delayed_with_leeway((thread_call_t)kn->kn_hook, NULL, + deadline, leeway, timer_flags); } -int -kqueue(struct proc *p, __unused struct kqueue_args *uap, register_t *retval) +/* + * Allocate a thread call for the knote's lifetime, and kick off the timer. + */ +static int +filt_timerattach(struct knote *kn, struct kevent_internal_s *kev) { - struct kqueue *kq; - struct fileproc *fp; - int fd, error; + thread_call_t callout; + struct filt_timer_params params; + int error; - error = falloc(p, &fp, &fd); - if (error) { - return (error); + if ((error = filt_timervalidate(kev, ¶ms)) != 0) { + knote_set_error(kn, error); + return 0; } - kq = kqueue_alloc(p); - if (kq == NULL) { - fp_free(p, fd, fp); - return (ENOMEM); + callout = thread_call_allocate_with_options(filt_timerexpire, + (thread_call_param_t)kn, THREAD_CALL_PRIORITY_HIGH, + THREAD_CALL_OPTIONS_ONCE); + + if (NULL == callout) { + knote_set_error(kn, ENOMEM); + return 0; } - fp->f_flag = FREAD | FWRITE; - fp->f_type = DTYPE_KQUEUE; - fp->f_ops = &kqueueops; - fp->f_data = (caddr_t)kq; + filt_timer_set_params(kn, ¶ms); + kn->kn_hook = callout; + kn->kn_flags |= EV_CLEAR; + os_atomic_store(&kn->kn_hookid, TIMER_IDLE, relaxed); - proc_fdlock(p); - *fdflags(p, fd) &= ~UF_RESERVED; - fp_drop(p, fd, fp, 1); - proc_fdunlock(p); + /* NOTE_ABSOLUTE implies EV_ONESHOT */ + if (kn->kn_sfflags & NOTE_ABSOLUTE) + kn->kn_flags |= EV_ONESHOT; - *retval = fd; - return (error); + if (filt_timer_is_ready(kn)) { + os_atomic_store(&kn->kn_hookid, TIMER_IMMEDIATE, relaxed); + return FILTER_ACTIVE; + } else { + filt_timerarm(kn); + return 0; + } } -int -kqueue_portset_np(__unused struct proc *p, - __unused struct kqueue_portset_np_args *uap, - __unused register_t *retval) +/* + * Shut down the timer if it's running, and free the callout. + */ +static void +filt_timerdetach(struct knote *kn) { - /* JMM - Placeholder for now */ - return (ENOTSUP); -} + __assert_only boolean_t freed; -int -kqueue_from_portset_np(__unused struct proc *p, - __unused struct kqueue_from_portset_np_args *uap, - __unused register_t *retval) -{ - /* JMM - Placeholder for now */ - return (ENOTSUP); + /* + * Unconditionally cancel to make sure there can't be any filt_timerexpire() + * running anymore. + */ + thread_call_cancel_wait((thread_call_t)kn->kn_hook); + freed = thread_call_free((thread_call_t)kn->kn_hook); + assert(freed); } +/* + * filt_timertouch - update timer knote with new user input + * + * Cancel and restart the timer based on new user data. When + * the user picks up a knote, clear the count of how many timer + * pops have gone off (in kn_data). + */ static int -kevent_copyin(user_addr_t *addrp, struct kevent *kevp, struct proc *p) +filt_timertouch(struct knote *kn, struct kevent_internal_s *kev) { - int advance; + struct filt_timer_params params; + uint32_t changed_flags = (kn->kn_sfflags ^ kev->fflags); int error; - if (IS_64BIT_PROCESS(p)) { - struct user_kevent kev64; + if (changed_flags & NOTE_ABSOLUTE) { + kev->flags |= EV_ERROR; + kev->data = EINVAL; + return 0; + } - advance = sizeof(kev64); - error = copyin(*addrp, (caddr_t)&kev64, advance); - if (error) - return error; - kevp->ident = CAST_DOWN(uintptr_t, kev64.ident); - kevp->filter = kev64.filter; - kevp->flags = kev64.flags; - kevp->fflags = kev64.fflags; - kevp->data = CAST_DOWN(intptr_t, kev64.data); - kevp->udata = kev64.udata; + if ((error = filt_timervalidate(kev, ¶ms)) != 0) { + kev->flags |= EV_ERROR; + kev->data = error; + return 0; + } + + /* capture the new values used to compute deadline */ + filt_timercancel(kn); + filt_timer_set_params(kn, ¶ms); + kn->kn_sfflags = kev->fflags; + + if (filt_timer_is_ready(kn)) { + os_atomic_store(&kn->kn_hookid, TIMER_IMMEDIATE, relaxed); + return FILTER_ACTIVE | FILTER_UPDATE_REQ_QOS; } else { - /* - * compensate for legacy in-kernel kevent layout - * where the udata field is alredy 64-bit. - */ - advance = sizeof(*kevp) + sizeof(void *) - sizeof(user_addr_t); - error = copyin(*addrp, (caddr_t)kevp, advance); + filt_timerarm(kn); + return FILTER_UPDATE_REQ_QOS; } - if (!error) - *addrp += advance; - return error; } +/* + * filt_timerprocess - query state of knote and snapshot event data + * + * Determine if the timer has fired in the past, snapshot the state + * of the kevent for returning to user-space, and clear pending event + * counters for the next time. + */ static int -kevent_copyout(struct kevent *kevp, user_addr_t *addrp, struct proc *p) +filt_timerprocess( + struct knote *kn, + __unused struct filt_process_s *data, + struct kevent_internal_s *kev) { - int advance; - int error; + /* + * 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 (os_atomic_load(&kn->kn_hookid, relaxed)) { + case TIMER_IDLE: + case TIMER_ARMED: + /* + * This can happen if a touch resets a timer that had fired + * without being processed + */ + return 0; + } - if (IS_64BIT_PROCESS(p)) { - struct user_kevent kev64; + os_atomic_store(&kn->kn_hookid, TIMER_IDLE, relaxed); - kev64.ident = (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; - advance = sizeof(kev64); - error = copyout((caddr_t)&kev64, *addrp, advance); + /* + * 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? */ + + if (kn->kn_sdata == 0) { + kev->data = 1; } else { /* - * compensate for legacy in-kernel kevent layout - * where the udata field is alredy 64-bit. + * This is a 'repeating' timer, so we have to emit + * how many intervals expired between the arm + * and the process. + * + * A very strange style of interface, because + * this could easily be done in the client... */ - advance = sizeof(*kevp) + sizeof(void *) - sizeof(user_addr_t); - error = copyout((caddr_t)kevp, *addrp, advance); - } - if (!error) - *addrp += advance; - return error; -} -/* - * kevent_continue - continue a kevent syscall after blocking - * - * assume we inherit a use count on the kq fileglob. - */ + uint64_t now; -static void -kevent_continue(__unused struct kqueue *kq, void *data, int error) -{ - struct _kevent *cont_args; - struct fileproc *fp; - register_t *retval; - int noutputs; - int fd; - struct proc *p = current_proc(); + if (kn->kn_sfflags & NOTE_MACH_CONTINUOUS_TIME) + now = mach_continuous_time(); + else + now = mach_absolute_time(); - cont_args = (struct _kevent *)data; - noutputs = cont_args->eventout; - retval = cont_args->retval; - fd = cont_args->fd; - fp = cont_args->fp; + uint64_t first_deadline = kn->kn_ext[0]; + uint64_t interval_abs = kn->kn_sdata; + uint64_t orig_arm_time = first_deadline - interval_abs; - fp_drop(p, fd, fp, 0); + assert(now > orig_arm_time); + assert(now > first_deadline); - /* 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); -} + uint64_t elapsed = now - orig_arm_time; -/* - * kevent - [syscall] register and wait for kernel events - * - */ + uint64_t num_fired = elapsed / interval_abs; -int -kevent(struct proc *p, struct kevent_args *uap, register_t *retval) -{ - user_addr_t changelist = uap->changelist; - user_addr_t ueventlist = uap->eventlist; - int nchanges = uap->nchanges; - int nevents = uap->nevents; - int fd = uap->fd; + /* + * To reach this code, we must have seen the timer pop + * and be in repeating mode, so therefore it must have been + * more than 'interval' time since the attach or last + * successful touch. + */ + assert(num_fired > 0); - struct _kevent *cont_args; - uthread_t ut; - struct kqueue *kq; - struct fileproc *fp; - struct kevent kev; - int error, noutputs; - struct timeval atv; + /* report how many intervals have elapsed to the user */ + kev->data = (int64_t)num_fired; - /* convert timeout to absolute - if we have one */ - if (uap->timeout != USER_ADDR_NULL) { - struct timeval rtv; - if ( IS_64BIT_PROCESS(p) ) { - struct user_timespec ts; - error = copyin( uap->timeout, &ts, sizeof(ts) ); - if ((ts.tv_sec & 0xFFFFFFFF00000000ull) != 0) - error = EINVAL; - else - TIMESPEC_TO_TIMEVAL(&rtv, &ts); - } else { - struct timespec ts; - error = copyin( uap->timeout, &ts, sizeof(ts) ); - TIMESPEC_TO_TIMEVAL(&rtv, &ts); + /* We only need to re-arm the timer if it's not about to be destroyed */ + if ((kn->kn_flags & EV_ONESHOT) == 0) { + /* fire at the end of the next interval */ + uint64_t new_deadline = first_deadline + num_fired * interval_abs; + + assert(new_deadline > now); + + 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); } - if (error) - return error; - if (itimerfix(&rtv)) - return EINVAL; - getmicrouptime(&atv); - timevaladd(&atv, &rtv); - } else { - atv.tv_sec = 0; - atv.tv_usec = 0; } - /* get a usecount for the kq itself */ - if ((error = fp_getfkq(p, fd, &fp, &kq)) != 0) - return(error); + return FILTER_ACTIVE; +} - /* register all the change requests the user provided... */ - noutputs = 0; - while (nchanges > 0 && error == 0) { - error = kevent_copyin(&changelist, &kev, p); - if (error) - break; - - kev.flags &= ~EV_SYSFLAGS; - error = kevent_register(kq, &kev, p); - if (error && nevents > 0) { - kev.flags = EV_ERROR; - kev.data = error; - error = kevent_copyout(&kev, &ueventlist, p); - if (error == 0) { - nevents--; - noutputs++; - } - } - nchanges--; +SECURITY_READ_ONLY_EARLY(static struct filterops) timer_filtops = { + .f_extended_codes = true, + .f_attach = filt_timerattach, + .f_detach = filt_timerdetach, + .f_event = filt_badevent, + .f_touch = filt_timertouch, + .f_process = filt_timerprocess, +}; + +#pragma mark user_filtops + +static int +filt_userattach(struct knote *kn, __unused struct kevent_internal_s *kev) +{ + if (kn->kn_sfflags & NOTE_TRIGGER) { + kn->kn_hookid = FILTER_ACTIVE; + } else { + kn->kn_hookid = 0; } + return (kn->kn_hookid); +} - /* store the continuation/completion data in the uthread */ - ut = (uthread_t)get_bsdthread_info(current_thread()); - cont_args = (struct _kevent *)&ut->uu_state.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; - - if (nevents > 0 && noutputs == 0 && error == 0) - error = kevent_scan(kq, kevent_callback, - kevent_continue, cont_args, - &atv, p); - kevent_continue(kq, cont_args, error); - /* NOTREACHED */ - return error; +static void +filt_userdetach(__unused struct knote *kn) +{ + /* EVFILT_USER knotes are not attached to anything in the kernel */ } +static int +filt_usertouch(struct knote *kn, struct kevent_internal_s *kev) +{ + uint32_t ffctrl; + int fflags; -/* - * kevent_callback - callback for each individual event - * - * called with nothing locked - * caller holds a reference on the kqueue - */ + ffctrl = kev->fflags & NOTE_FFCTRLMASK; + fflags = kev->fflags & NOTE_FFLAGSMASK; + switch (ffctrl) { + case NOTE_FFNOP: + break; + case NOTE_FFAND: + kn->kn_sfflags &= fflags; + break; + case NOTE_FFOR: + kn->kn_sfflags |= fflags; + break; + case NOTE_FFCOPY: + kn->kn_sfflags = fflags; + break; + } + kn->kn_sdata = kev->data; + + if (kev->fflags & NOTE_TRIGGER) { + kn->kn_hookid = FILTER_ACTIVE; + } + return (int)kn->kn_hookid; +} static int -kevent_callback(__unused struct kqueue *kq, struct kevent *kevp, void *data) +filt_userprocess( + struct knote *kn, + __unused struct filt_process_s *data, + struct kevent_internal_s *kev) { - struct _kevent *cont_args; - int error; + int result = (int)kn->kn_hookid; + + if (result) { + *kev = kn->kn_kevent; + kev->fflags = 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; + } + } - cont_args = (struct _kevent *)data; - assert(cont_args->eventout < cont_arg->eventcount); + return result; +} - /* - * Copy out the appropriate amount of event data for this user. - */ - error = kevent_copyout(kevp, &cont_args->eventlist, current_proc()); +SECURITY_READ_ONLY_EARLY(static struct filterops) user_filtops = { + .f_extended_codes = true, + .f_attach = filt_userattach, + .f_detach = filt_userdetach, + .f_event = filt_badevent, + .f_touch = filt_usertouch, + .f_process = filt_userprocess, +}; - /* - * 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; +#pragma mark workloop_filtops + +static inline void +filt_wllock(struct kqworkloop *kqwl) +{ + lck_mtx_lock(&kqwl->kqwl_statelock); +} + +static inline void +filt_wlunlock(struct kqworkloop *kqwl) +{ + lck_mtx_unlock(&kqwl->kqwl_statelock); } /* - * kevent_register - add a new event to a kqueue - * - * Creates a mapping between the event source and - * the kqueue via a knote data structure. + * Returns true when the interlock for the turnstile is the workqueue lock * - * Because many/most the event sources are file - * descriptor related, the knote is linked off - * the filedescriptor table for quick access. + * When this is the case, all turnstiles operations are delegated + * to the workqueue subsystem. * - * called with nothing locked - * caller holds a reference on the kqueue + * 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_wlturnstile_interlock_is_workq(struct kqworkloop *kqwl) +{ + struct kqrequest *kqr = &kqwl->kqwl_request; + return (kqr->kqr_state & KQR_THREQUESTED) && + (kqr->kqr_thread == THREAD_NULL); +} -int -kevent_register(struct kqueue *kq, struct kevent *kev, struct proc *p) +static void +filt_wlupdate_inheritor(struct kqworkloop *kqwl, struct turnstile *ts, + turnstile_update_flags_t flags) { - struct filedesc *fdp = kq->kq_fdp; - struct filterops *fops; - struct fileproc *fp = NULL; - struct knote *kn = NULL; - int error = 0; + turnstile_inheritor_t inheritor = TURNSTILE_INHERITOR_NULL; + struct kqrequest *kqr = &kqwl->kqwl_request; - if (kev->filter < 0) { - if (kev->filter + EVFILT_SYSCOUNT < 0) - return (EINVAL); - fops = sysfilt_ops[~kev->filter]; /* to 0-base index */ - } else { + /* + * binding to the workq should always happen through + * workq_kern_threadreq_update_inheritor() + */ + assert(!filt_wlturnstile_interlock_is_workq(kqwl)); + + if ((inheritor = kqwl->kqwl_owner)) { + flags |= TURNSTILE_INHERITOR_THREAD; + } else if ((inheritor = kqr->kqr_thread)) { + flags |= TURNSTILE_INHERITOR_THREAD; + } + + turnstile_update_inheritor(ts, inheritor, flags); +} + +#define FILT_WLATTACH 0 +#define FILT_WLTOUCH 1 +#define FILT_WLDROP 2 + +__result_use_check +static int +filt_wlupdate(struct kqworkloop *kqwl, struct knote *kn, + struct kevent_internal_s *kev, kq_index_t qos_index, int op) +{ + user_addr_t uaddr = CAST_USER_ADDR_T(kev->ext[EV_EXTIDX_WL_ADDR]); + struct kqrequest *kqr = &kqwl->kqwl_request; + thread_t cur_owner, new_owner, extra_thread_ref = THREAD_NULL; + kq_index_t cur_owner_override = THREAD_QOS_UNSPECIFIED; + int action = KQWL_UTQ_NONE, error = 0; + bool needs_wake = false, needs_wllock = false; + uint64_t kdata = kev->ext[EV_EXTIDX_WL_VALUE]; + uint64_t mask = kev->ext[EV_EXTIDX_WL_MASK]; + uint64_t udata = 0; + + if (kev->fflags & (NOTE_WL_END_OWNERSHIP | NOTE_WL_DISCOVER_OWNER)) { /* - * XXX - * filter attach routine is responsible for insuring that - * the identifier can be attached to it. + * If we're maybe going to change the kqwl_owner, + * then we need to hold the filt_wllock(). */ - printf("unknown filter: %d\n", kev->filter); - return (EINVAL); + needs_wllock = true; + } else if (kqr->kqr_thread == current_thread()) { + /* + * Servicer updates need to be serialized with + * any ownership change too, as the kqr_thread value influences the + * outcome of handling NOTE_WL_DISCOVER_OWNER. + */ + needs_wllock = true; } - /* this iocount needs to be dropped if it is not registered */ - if (fops->f_isfd && (error = fp_lookup(p, kev->ident, &fp, 0)) != 0) - return(error); - - restart: - proc_fdlock(p); - if (fops->f_isfd) { - /* fd-based knotes are linked off the fd table */ - if (kev->ident < (u_int)fdp->fd_knlistsize) { - SLIST_FOREACH(kn, &fdp->fd_knlist[kev->ident], kn_link) - if (kq == kn->kn_kq && - kev->filter == kn->kn_filter) - break; - } + if (needs_wllock) { + filt_wllock(kqwl); + /* + * The kqwl owner is set under both the req and filter lock, + * meaning it's fine to look at it under any. + */ + new_owner = cur_owner = kqwl->kqwl_owner; } else { - /* hash non-fd knotes here too */ - if (fdp->fd_knhashmask != 0) { - struct klist *list; - - list = &fdp->fd_knhash[ - KN_HASH((u_long)kev->ident, fdp->fd_knhashmask)]; - SLIST_FOREACH(kn, list, kn_link) - if (kev->ident == kn->kn_id && - kq == kn->kn_kq && - kev->filter == kn->kn_filter) - break; - } + new_owner = cur_owner = THREAD_NULL; } /* - * kn now contains the matching knote, or NULL if no match + * 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. */ - if (kn == NULL) { - if ((kev->flags & (EV_ADD|EV_DELETE)) == EV_ADD) { - kn = knote_alloc(); - if (kn == NULL) { - proc_fdunlock(p); - error = ENOMEM; - goto done; - } - kn->kn_fp = fp; - kn->kn_kq = kq; - kn->kn_tq = &kq->kq_head; - kn->kn_fop = fops; - kn->kn_sfflags = kev->fflags; - kn->kn_sdata = kev->data; - kev->fflags = 0; - kev->data = 0; - kn->kn_kevent = *kev; - kn->kn_inuse = 1; /* for f_attach() */ - kn->kn_status = 0; - - /* before anyone can find it */ - if (kev->flags & EV_DISABLE) - kn->kn_status |= KN_DISABLED; - - error = knote_fdpattach(kn, fdp, p); - proc_fdunlock(p); + if (uaddr) { + error = copyin_word(uaddr, &udata, sizeof(udata)); + if (error) { + goto out; + } - if (error) { - knote_free(kn); - goto done; - } + /* 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) { /* - * apply reference count to knote structure, and - * do not release it at the end of this routine. + * 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. */ - fp = NULL; + 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); + if (extra_thread_ref == THREAD_NULL) { + error = EOWNERDEAD; + goto out; + } + new_owner = extra_thread_ref; + } + } + } + + 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->kqr_qos_index != qos_index) { + action = KQWL_UTQ_SET_QOS_INDEX; + } + + if (op == FILT_WLTOUCH) { /* - * If the attach fails here, we can drop it knowing - * that nobody else has a reference to the knote. + * 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. */ - if ((error = fops->f_attach(kn)) != 0) { - knote_drop(kn, p); - goto done; + kn->kn_sfflags &= ~NOTE_WL_UPDATES_MASK; + kn->kn_sfflags |= kev->fflags; + kn->kn_sdata = kev->data; + if (kev->fflags & NOTE_WL_SYNC_WAKE) { + needs_wake = (kn->kn_hook != THREAD_NULL); } - } else { - proc_fdunlock(p); - error = ENOENT; - goto done; - } - } else { - /* existing knote - get kqueue lock */ - kqlock(kq); - proc_fdunlock(p); - - if (kev->flags & EV_DELETE) { - knote_dequeue(kn); - kn->kn_status |= KN_DISABLED; - if (kqlock2knotedrop(kq, kn)) { - kn->kn_fop->f_detach(kn); - knote_drop(kn, p); + } 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_hook != THREAD_NULL); } - goto done; - } - - /* update status flags for existing knote */ - if (kev->flags & EV_DISABLE) { - knote_dequeue(kn); - kn->kn_status |= KN_DISABLED; - } else if (kev->flags & EV_ENABLE) { - kn->kn_status &= ~KN_DISABLED; - if (kn->kn_status & KN_ACTIVE) - knote_enqueue(kn); } + } - /* - * If somebody is in the middle of dropping this - * knote - go find/insert a new one. But we have - * wait for this one to go away first. - */ - if (!kqlock2knoteusewait(kq, kn)) - /* kqueue unlocked */ - goto restart; + /* + * Phase 2: + * + * Commit ownership and QoS changes if any, possibly wake up waiters + */ - /* - * The user may change some filter values after the - * initial EV_ADD, but doing so will not reset any - * filter which have already been triggered. - */ - kn->kn_sfflags = kev->fflags; - kn->kn_sdata = kev->data; - kn->kn_kevent.udata = kev->udata; + if (cur_owner == new_owner && action == KQWL_UTQ_NONE && !needs_wake) { + goto out; } - - /* still have use ref on knote */ - if (kn->kn_fop->f_event(kn, 0)) { - if (knoteuse2kqlock(kq, kn)) - knote_activate(kn); - kqunlock(kq); - } else { - knote_put(kn); - } - -done: - if (fp != NULL) - fp_drop(p, kev->ident, fp, 0); - return (error); -} - -/* - * kevent_process - process the triggered events in a kqueue - * - * 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 a reference on the kqueue. - * kqueue locked on entry and exit - but may be dropped - */ -static int -kevent_process(struct kqueue *kq, - kevent_callback_t callback, - void *data, - int *countp, - struct proc *p) -{ - struct knote *kn; - struct kevent kev; - int nevents; - int error; + kq_req_lock(kqwl); - restart: - if (kq->kq_count == 0) { - *countp = 0; - return 0; + /* If already tracked as servicer, don't track as owner */ + if (new_owner == kqr->kqr_thread) { + new_owner = THREAD_NULL; } - /* if someone else is processing the queue, wait */ - if (!TAILQ_EMPTY(&kq->kq_inprocess)) { - assert_wait(&kq->kq_inprocess, THREAD_UNINT); - kq->kq_state |= KQ_PROCWAIT; - kqunlock(kq); - thread_block(THREAD_CONTINUE_NULL); - kqlock(kq); - goto restart; - } + if (cur_owner != new_owner) { + kqwl->kqwl_owner = new_owner; + if (new_owner == extra_thread_ref) { + /* we just transfered this ref to kqwl_owner */ + extra_thread_ref = THREAD_NULL; + } + cur_owner_override = kqworkloop_owner_override(kqwl); - error = 0; - nevents = 0; - while (error == 0 && - (kn = TAILQ_FIRST(&kq->kq_head)) != NULL) { + if (cur_owner) { + thread_ends_owning_workloop(cur_owner); + } - /* - * move knote to the processed queue. - * this is also protected by the kq lock. - */ - assert(kn->kn_tq == &kq->kq_head); - TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe); - kn->kn_tq = &kq->kq_inprocess; - TAILQ_INSERT_TAIL(&kq->kq_inprocess, kn, kn_tqe); + if (new_owner) { + /* override it before we drop the old */ + if (cur_owner_override != THREAD_QOS_UNSPECIFIED) { + thread_add_ipc_override(new_owner, cur_owner_override); + } + thread_starts_owning_workloop(new_owner); + if ((kqr->kqr_state & KQR_THREQUESTED) && !kqr->kqr_thread) { + if (action == KQWL_UTQ_NONE) { + action = KQWL_UTQ_REDRIVE_EVENTS; + } + } + } else { + if ((kqr->kqr_state & (KQR_THREQUESTED | KQR_WAKEUP)) == KQR_WAKEUP) { + if (action == KQWL_UTQ_NONE) { + action = KQWL_UTQ_REDRIVE_EVENTS; + } + } + } + } - /* - * Non-EV_ONESHOT events must be re-validated. - * - * Convert our lock to a use-count and call the event's - * filter routine to update. - * - * If the event is dropping (or no longer valid), we - * already have it off the active queue, so just - * finish the job of deactivating it. - */ - if ((kn->kn_flags & EV_ONESHOT) == 0) { - int result; + struct turnstile *ts = kqwl->kqwl_turnstile; + bool wl_inheritor_updated = false; - if (kqlock2knoteuse(kq, kn)) { - - /* call the filter with just a ref */ - result = kn->kn_fop->f_event(kn, 0); + if (action != KQWL_UTQ_NONE) { + kqworkloop_update_threads_qos(kqwl, action, qos_index); + } - if (!knoteuse2kqlock(kq, kn) || result == 0) { - knote_deactivate(kn); - continue; - } - } else { - knote_deactivate(kn); - continue; + 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; } /* - * Got a valid triggered knote with the kqueue - * still locked. Snapshot the data, and determine - * how to dispatch the knote for future events. + * We need a turnstile reference because we are dropping the interlock + * and the caller has not called turnstile_prepare. */ - kev = kn->kn_kevent; - - /* now what happens to it? */ - if (kn->kn_flags & EV_ONESHOT) { - knote_deactivate(kn); - if (kqlock2knotedrop(kq, kn)) { - kn->kn_fop->f_detach(kn); - knote_drop(kn, p); - } - } else if (kn->kn_flags & EV_CLEAR) { - knote_deactivate(kn); - kn->kn_data = 0; - kn->kn_fflags = 0; - kqunlock(kq); - } else { - /* - * leave on in-process queue. We'll - * move all the remaining ones back - * the kq queue and wakeup any - * waiters when we are done. - */ - kqunlock(kq); + if (wl_inheritor_updated) { + turnstile_reference(ts); } + } - /* callback to handle each event as we find it */ - error = (callback)(kq, &kev, data); - nevents++; + if (needs_wake && ts) { + waitq_wakeup64_thread(&ts->ts_waitq, CAST_EVENT64_T((event_t)kn), + (thread_t)kn->kn_hook, THREAD_AWAKENED); + } - kqlock(kq); + kq_req_unlock(kqwl); + + if (wl_inheritor_updated) { + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_NOT_HELD); + turnstile_deallocate(ts); } +out: /* - * With the kqueue still locked, move any knotes - * remaining on the in-process queue back to the - * kq's queue and wake up any waiters. + * Phase 3: + * + * Unlock and cleanup various lingering references and things. */ - while ((kn = TAILQ_FIRST(&kq->kq_inprocess)) != NULL) { - assert(kn->kn_tq == &kq->kq_inprocess); - TAILQ_REMOVE(&kq->kq_inprocess, kn, kn_tqe); - kn->kn_tq = &kq->kq_head; - TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe); + if (needs_wllock) { + filt_wlunlock(kqwl); } - if (kq->kq_state & KQ_PROCWAIT) { - kq->kq_state &= ~KQ_PROCWAIT; - thread_wakeup(&kq->kq_inprocess); + +#if CONFIG_WORKLOOP_DEBUG + KQWL_HISTORY_WRITE_ENTRY(kqwl, { + .updater = current_thread(), + .servicer = kqr->kqr_thread, /* Note: racy */ + .old_owner = cur_owner, + .new_owner = new_owner, + + .kev_ident = kev->ident, + .error = (int16_t)error, + .kev_flags = kev->flags, + .kev_fflags = kev->fflags, + + .kev_mask = mask, + .kev_value = kdata, + .in_value = udata, + }); +#endif // CONFIG_WORKLOOP_DEBUG + + if (cur_owner && new_owner != cur_owner) { + if (cur_owner_override != THREAD_QOS_UNSPECIFIED) { + thread_drop_ipc_override(cur_owner); + } + thread_deallocate(cur_owner); } - *countp = nevents; + if (extra_thread_ref) { + thread_deallocate(extra_thread_ref); + } return error; } +/* + * Remembers the last updated that came in from userspace for debugging reasons. + * - fflags is mirrored from the userspace kevent + * - ext[i, i != VALUE] is mirrored from the userspace kevent + * - ext[VALUE] is set to what the kernel loaded atomically + * - data is set to the error if any + */ +static inline void +filt_wlremember_last_update(struct knote *kn, struct kevent_internal_s *kev, + int error) +{ + kn->kn_fflags = kev->fflags; + kn->kn_data = error; + memcpy(kn->kn_ext, kev->ext, sizeof(kev->ext)); +} -static void -kevent_scan_continue(void *data, wait_result_t wait_result) +static int +filt_wlattach(struct knote *kn, struct kevent_internal_s *kev) { - uthread_t ut = (uthread_t)get_bsdthread_info(current_thread()); - struct _kevent_scan * cont_args = &ut->uu_state.ss_kevent_scan; - struct kqueue *kq = (struct kqueue *)data; - int error; - int count; + struct kqueue *kq = knote_get_kq(kn); + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + int error = 0; + kq_index_t qos_index = 0; - /* convert the (previous) wait_result to a proper error */ - switch (wait_result) { - case THREAD_AWAKENED: - kqlock(kq); - error = kevent_process(kq, cont_args->call, cont_args, &count, current_proc()); - if (error == 0 && count == 0) { - assert_wait_deadline(kq, THREAD_ABORTSAFE, cont_args->deadline); - kq->kq_state |= KQ_SLEEP; - kqunlock(kq); - thread_block_parameter(kevent_scan_continue, kq); - /* NOTREACHED */ + if ((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; + + kq_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; + } + 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]; + kq_req_unlock(kqwl); + error = EBUSY; + goto out; + } +#endif + + int 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 = _pthread_priority_thread_qos(kn->kn_qos); + if (qos_index == THREAD_QOS_UNSPECIFIED) { + error = ERANGE; + goto out; + } + if (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. + */ + error = EALREADY; + goto out; } - kqunlock(kq); - break; - case THREAD_TIMED_OUT: - error = EWOULDBLOCK; break; - case THREAD_INTERRUPTED: - error = EINTR; + case NOTE_WL_SYNC_WAIT: + case NOTE_WL_SYNC_WAKE: + if (kn->kn_id == kqwl->kqwl_dynamicid) { + error = EINVAL; + goto out; + } + if ((kn->kn_flags & EV_DISABLE) == 0) { + error = EINVAL; + goto out; + } + if (kn->kn_sfflags & NOTE_WL_END_OWNERSHIP) { + error = EINVAL; + goto out; + } break; default: - panic("kevent_scan_cont() - invalid wait_result (%d)", wait_result); - error = 0; + error = EINVAL; + goto out; } - - /* call the continuation with the results */ - assert(cont_args->cont != NULL); - (cont_args->cont)(kq, cont_args->data, error); + + error = filt_wlupdate(kqwl, kn, kev, qos_index, FILT_WLATTACH); + +out: + if (error) { + /* If userland wants ESTALE to be hidden, fail the attach anyway */ + if (error == ESTALE && (kn->kn_sfflags & NOTE_WL_IGNORE_ESTALE)) { + error = 0; + } + knote_set_error(kn, error); + return 0; + } + if (command == NOTE_WL_SYNC_WAIT) { + return kevent_register_wait_prepare(kn, kev); + } + /* Just attaching the thread request successfully will fire it */ + 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 FILTER_ACTIVE; + } + return 0; } +static void __dead2 +filt_wlwait_continue(void *parameter, wait_result_t wr) +{ + struct _kevent_register *cont_args = parameter; + struct kqworkloop *kqwl = (struct kqworkloop *)cont_args->kq; + struct kqrequest *kqr = &kqwl->kqwl_request; + + kq_req_lock(kqwl); + kqr->kqr_dsync_waiters--; + if (filt_wlturnstile_interlock_is_workq(kqwl)) { + workq_kern_threadreq_lock(kqwl->kqwl_p); + turnstile_complete((uintptr_t)kqwl, &kqwl->kqwl_turnstile, NULL); + workq_kern_threadreq_unlock(kqwl->kqwl_p); + } else { + turnstile_complete((uintptr_t)kqwl, &kqwl->kqwl_turnstile, NULL); + } + kq_req_unlock(kqwl); + + turnstile_cleanup(); + + 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); + } + + kevent_register_wait_return(cont_args); +} /* - * kevent_scan - scan and wait for events in a kqueue - * - * Process the triggered events in a kqueue. - * - * 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. + * Called with the workloop mutex held, most of the time never returns as it + * calls filt_wlwait_continue through a continuation. */ - -int -kevent_scan(struct kqueue *kq, - kevent_callback_t callback, - kevent_continue_t continuation, - void *data, - struct timeval *atvp, - struct proc *p) +static void __dead2 +filt_wlpost_register_wait(struct uthread *uth, struct knote_lock_ctx *knlc, + struct _kevent_register *cont_args) { - thread_continue_t cont = THREAD_CONTINUE_NULL; - uint64_t deadline; - int error; - int first; + struct kqworkloop *kqwl = (struct kqworkloop *)cont_args->kq; + struct kqrequest *kqr = &kqwl->kqwl_request; + struct turnstile *ts; + bool workq_locked = false; - assert(callback != NULL); + kq_req_lock(kqwl); - first = 1; - for (;;) { - wait_result_t wait_result; - int count; + kqr->kqr_dsync_waiters++; - /* - * Make a pass through the kq to find events already - * triggered. - */ - kqlock(kq); - error = kevent_process(kq, callback, data, &count, p); - if (error || count) - break; /* lock still held */ + if (filt_wlturnstile_interlock_is_workq(kqwl)) { + workq_kern_threadreq_lock(kqwl->kqwl_p); + workq_locked = true; + } - /* 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) { - uint32_t seconds, nanoseconds; - uint64_t now; - - clock_get_uptime(&now); - nanoseconds_to_absolutetime((uint64_t)atvp->tv_sec * NSEC_PER_SEC + - atvp->tv_usec * 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 */ - } + ts = turnstile_prepare((uintptr_t)kqwl, &kqwl->kqwl_turnstile, + TURNSTILE_NULL, TURNSTILE_WORKLOOPS); - if (continuation) { - uthread_t ut = (uthread_t)get_bsdthread_info(current_thread()); - struct _kevent_scan *cont_args = &ut->uu_state.ss_kevent_scan; - - cont_args->call = callback; - cont_args->cont = continuation; - cont_args->deadline = deadline; - cont_args->data = data; - cont = kevent_scan_continue; - } + 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); + } - /* go ahead and wait */ - assert_wait_deadline(kq, THREAD_ABORTSAFE, deadline); - kq->kq_state |= KQ_SLEEP; - kqunlock(kq); - wait_result = thread_block_parameter(cont, kq); - /* NOTREACHED if (continuation != NULL) */ + thread_set_pending_block_hint(uth->uu_thread, kThreadWaitWorkloopSyncWait); + waitq_assert_wait64(&ts->ts_waitq, CAST_EVENT64_T(cont_args->knote), + THREAD_ABORTSAFE, TIMEOUT_WAIT_FOREVER); - switch (wait_result) { - case THREAD_AWAKENED: - continue; - case THREAD_TIMED_OUT: - return EWOULDBLOCK; - case THREAD_INTERRUPTED: - return EINTR; - default: - panic("kevent_scan - bad wait_result (%d)", - wait_result); - error = 0; - } + if (workq_locked) { + workq_kern_threadreq_unlock(kqwl->kqwl_p); } - kqunlock(kq); - return error; -} + thread_t thread = kqwl->kqwl_owner ?: kqr->kqr_thread; + if (thread) { + thread_reference(thread); + } + kq_req_unlock(kqwl); -/* - * XXX - * This could be expanded to call kqueue_scan, if desired. - */ -/*ARGSUSED*/ -static int -kqueue_read(__unused struct fileproc *fp, - __unused struct uio *uio, - __unused kauth_cred_t cred, - __unused int flags, - __unused struct proc *p) -{ - return (ENXIO); + kevent_register_wait_block(ts, thread, knlc, filt_wlwait_continue, cont_args); } -/*ARGSUSED*/ -static int -kqueue_write(__unused struct fileproc *fp, - __unused struct uio *uio, - __unused kauth_cred_t cred, - __unused int flags, - __unused struct proc *p) +/* 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) { - return (ENXIO); + struct knote *kn = (struct knote *)event; + assert(kdp_is_in_zone(kn, "knote zone")); + + assert(kn->kn_hook == thread); + + struct kqueue *kq = knote_get_kq(kn); + assert(kdp_is_in_zone(kq, "kqueue workloop zone")); + assert(kq->kq_state & KQ_WORKLOOP); + + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + struct kqrequest *kqr = &kqwl->kqwl_request; + + thread_t kqwl_owner = kqwl->kqwl_owner; + thread_t servicer = kqr->kqr_thread; + + if (kqwl_owner != THREAD_NULL) { + assert(kdp_is_in_zone(kqwl_owner, "threads")); + + 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) { + waitinfo->owner = STACKSHOT_WAITOWNER_THREQUESTED; + } else { + waitinfo->owner = 0; + } + + waitinfo->context = kqwl->kqwl_dynamicid; } -/*ARGSUSED*/ -static int -kqueue_ioctl(__unused struct fileproc *fp, - __unused u_long com, - __unused caddr_t data, - __unused struct proc *p) +static void +filt_wldetach(__assert_only struct knote *kn) { - return (ENOTTY); + assert(knote_get_kq(kn)->kq_state & KQ_WORKLOOP); + if (kn->kn_hook) { + kevent_register_wait_cleanup(kn); + } } -/*ARGSUSED*/ static int -kqueue_select(struct fileproc *fp, int which, void *wql, struct proc *p) +filt_wlvalidate_kev_flags(struct knote *kn, struct kevent_internal_s *kev, + thread_qos_t *qos_index) { - struct kqueue *kq = (struct kqueue *)fp->f_data; - int retnum = 0; + int new_commands = kev->fflags & NOTE_WL_COMMANDS_MASK; + int sav_commands = kn->kn_sfflags & NOTE_WL_COMMANDS_MASK; - if (which == FREAD) { - kqlock(kq); - if (kq->kq_count) { - retnum = 1; - } else { - selrecord(p, &kq->kq_sel, wql); - kq->kq_state |= KQ_SEL; + 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; } - kqunlock(kq); } - return (retnum); -} -/* - * kqueue_close - - */ -/*ARGSUSED*/ -static int -kqueue_close(struct fileglob *fg, struct proc *p) -{ - struct kqueue *kq = (struct kqueue *)fg->fg_data; + switch (new_commands) { + case NOTE_WL_THREAD_REQUEST: + /* thread requests can only update themselves */ + if (sav_commands != NOTE_WL_THREAD_REQUEST) + return EINVAL; + break; - kqueue_dealloc(kq, p); - fg->fg_data = NULL; - return (0); + case NOTE_WL_SYNC_WAIT: + if (kev->fflags & NOTE_WL_END_OWNERSHIP) + return EINVAL; + goto sync_checks; + + case NOTE_WL_SYNC_WAKE: + 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; + + default: + return EINVAL; + } + return 0; } -/*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 proc *p) +filt_wltouch(struct knote *kn, struct kevent_internal_s *kev) { - struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; + struct kqworkloop *kqwl = (struct kqworkloop *)knote_get_kq(kn); + thread_qos_t qos_index = THREAD_QOS_UNSPECIFIED; + + int error = filt_wlvalidate_kev_flags(kn, kev, &qos_index); + if (error) { + goto out; + } - if (kn->kn_filter != EVFILT_READ) - return (1); + error = filt_wlupdate(kqwl, kn, kev, qos_index, FILT_WLTOUCH); + filt_wlremember_last_update(kn, kev, error); + if (error) { + goto out; + } - kn->kn_fop = &kqread_filtops; - kqlock(kq); - KNOTE_ATTACH(&kq->kq_sel.si_note, kn); - kqunlock(kq); - return (0); +out: + if (error) { + if (error == ESTALE && (kev->fflags & NOTE_WL_IGNORE_ESTALE)) { + /* If userland wants ESTALE to be hidden, do not activate */ + return 0; + } + kev->flags |= EV_ERROR; + kev->data = error; + return 0; + } + int command = kev->fflags & NOTE_WL_COMMANDS_MASK; + if (command == NOTE_WL_SYNC_WAIT && !(kn->kn_sfflags & NOTE_WL_SYNC_WAKE)) { + return kevent_register_wait_prepare(kn, kev); + } + /* Just touching the thread request successfully will fire it */ + if (command == NOTE_WL_THREAD_REQUEST) { + if (kev->fflags & NOTE_WL_UPDATE_QOS) { + return FILTER_ACTIVE | FILTER_UPDATE_REQ_QOS; + } + return FILTER_ACTIVE; + } + return 0; } -/*ARGSUSED*/ -int -kqueue_stat(struct fileproc *fp, struct stat *st, __unused struct proc *p) +static bool +filt_wlallow_drop(struct knote *kn, struct kevent_internal_s *kev) { - struct kqueue *kq = (struct kqueue *)fp->f_data; + struct kqworkloop *kqwl = (struct kqworkloop *)knote_get_kq(kn); - bzero((void *)st, sizeof(*st)); - st->st_size = kq->kq_count; - st->st_blksize = sizeof(struct kevent); - st->st_mode = S_IFIFO; - return (0); + int error = filt_wlvalidate_kev_flags(kn, kev, NULL); + if (error) { + goto out; + } + + error = filt_wlupdate(kqwl, kn, kev, 0, FILT_WLDROP); + filt_wlremember_last_update(kn, kev, error); + if (error) { + goto out; + } + +out: + if (error) { + if (error == ESTALE && (kev->fflags & NOTE_WL_IGNORE_ESTALE)) { + return false; + } + kev->flags |= EV_ERROR; + kev->data = error; + return false; + } + return true; } -/* - * Called with the kqueue locked - */ -static void -kqueue_wakeup(struct kqueue *kq) +static int +filt_wlprocess( + struct knote *kn, + __unused struct filt_process_s *data, + struct kevent_internal_s *kev) { + struct kqworkloop *kqwl = (struct kqworkloop *)knote_get_kq(kn); + int rc = 0; + + assert(kn->kn_sfflags & NOTE_WL_THREAD_REQUEST); - if (kq->kq_state & KQ_SLEEP) { - kq->kq_state &= ~KQ_SLEEP; - thread_wakeup(kq); + filt_wllock(kqwl); + + 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(kqwl); + knote_activate(kn); + kqunlock(kqwl); + } 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_word(addr, &val, sizeof(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 + *kev = kn->kn_kevent; + kev->fflags = kn->kn_sfflags; + kev->data = kn->kn_sdata; + kev->qos = kn->kn_qos; + rc |= FILTER_ACTIVE; } - if (kq->kq_state & KQ_SEL) { - kq->kq_state &= ~KQ_SEL; - selwakeup(&kq->kq_sel); + + filt_wlunlock(kqwl); + + if (rc & FILTER_ACTIVE) { + workq_thread_set_max_qos(kqwl->kqwl_p, &kqwl->kqwl_request); } - KNOTE(&kq->kq_sel.si_note, 0); + return rc; } -void -klist_init(struct klist *list) -{ - SLIST_INIT(list); -} +SECURITY_READ_ONLY_EARLY(static struct filterops) workloop_filtops = { + .f_extended_codes = true, + .f_attach = filt_wlattach, + .f_detach = filt_wldetach, + .f_event = filt_badevent, + .f_touch = filt_wltouch, + .f_process = filt_wlprocess, + .f_allow_drop = filt_wlallow_drop, + .f_post_register_wait = filt_wlpost_register_wait, +}; +#pragma mark kevent / knotes /* - * 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 hind) and not deadlock itself. - * - * The object lock should also hold off pending - * detach/drop operations. But we'll prevent it here - * too - just in case. + * JMM - placeholder for not-yet-implemented filters */ -void -knote(struct klist *list, long hint) +static int +filt_badevent(struct knote *kn, long hint) { - struct knote *kn; + panic("%s[%d](%p, %ld)", __func__, kn->kn_filter, kn, hint); + return 0; +} - SLIST_FOREACH(kn, list, kn_selnext) { - struct kqueue *kq = kn->kn_kq; +static int +filt_badattach(__unused struct knote *kn, __unused struct kevent_internal_s *kev) +{ + knote_set_error(kn, ENOTSUP); + return 0; +} - kqlock(kq); - if (kqlock2knoteuse(kq, kn)) { - int result; +struct kqueue * +kqueue_alloc(struct proc *p, unsigned int flags) +{ + struct filedesc *fdp = p->p_fd; + struct kqueue *kq = NULL; + int policy; + void *hook = NULL; - /* call the event with only a use count */ - result = kn->kn_fop->f_event(kn, hint); + if (flags & KEVENT_FLAG_WORKQ) { + struct kqworkq *kqwq; + int i; - /* if its not going away and triggered */ - if (knoteuse2kqlock(kq, kn) && result) - knote_activate(kn); - /* lock held again */ + 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(&kqwq->kqwq_queue[i]); } - kqunlock(kq); + for (i = 0; i < KQWQ_NBUCKETS; i++) { + if (i != KQWQ_QOS_MANAGER) { + /* + * 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].kqr_state |= KQR_THOVERCOMMIT; + } + kqwq->kqwq_request[i].kqr_qos_index = i; + TAILQ_INIT(&kqwq->kqwq_request[i].kqr_suppressed); + } + + 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 */ + kqwl->kqwl_request.kqr_state = KQR_WORKLOOP; + + kq = &kqwl->kqwl_kqueue; + for (i = 0; i < KQWL_NBUCKETS; i++) { + TAILQ_INIT(&kqwl->kqwl_queue[i]); + } + TAILQ_INIT(&kqwl->kqwl_request.kqr_suppressed); + + lck_mtx_init(&kqwl->kqwl_statelock, kq_lck_grp, kq_lck_attr); + + policy = SYNC_POLICY_FIFO; + 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(&kqf->kqf_queue); + TAILQ_INIT(&kqf->kqf_suppressed); + + policy = SYNC_POLICY_FIFO | SYNC_POLICY_PREPOST; + } + + waitq_set_init(&kq->kq_wqs, policy, NULL, hook); + lck_spin_init(&kq->kq_lock, kq_lck_grp, kq_lck_attr); + lck_spin_init(&kq->kq_reqlock, 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); } + + return (kq); } /* - * 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. + * 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). */ -int -knote_attach(struct klist *list, struct knote *kn) +void +knotes_dealloc(proc_t p) { - int ret = SLIST_EMPTY(list); - SLIST_INSERT_HEAD(list, kn, kn_selnext); - return ret; + struct filedesc *fdp = p->p_fd; + struct kqueue *kq; + struct knote *kn; + struct klist *kn_hash = NULL; + 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 */ + FREE(fdp->fd_knlist, M_KQUEUE); + fdp->fd_knlist = NULL; + } + 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); + knote_drop(kq, kn, NULL); + knhash_lock(p); + } + } + kn_hash = fdp->fd_knhash; + fdp->fd_knhashmask = 0; + fdp->fd_knhash = NULL; + } + + knhash_unlock(p); + + /* free the kn_hash table */ + if (kn_hash) + FREE(kn_hash, M_KQUEUE); + + proc_fdlock(p); } /* - * 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. + * kqworkloop_invalidate + * + * Invalidate ownership of a workloop. + * + * This is meant to be used so that any remnant of overrides and ownership + * information is dropped before a kqworkloop can no longer be found in the + * global hash table and have ghost workloop ownership left over. + * + * Possibly returns a thread to deallocate in a safe context. */ -int -knote_detach(struct klist *list, struct knote *kn) +static thread_t +kqworkloop_invalidate(struct kqworkloop *kqwl) { - SLIST_REMOVE(list, kn, knote, kn_selnext); - return SLIST_EMPTY(list); + thread_t cur_owner = kqwl->kqwl_owner; + + assert(TAILQ_EMPTY(&kqwl->kqwl_request.kqr_suppressed)); + if (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 (kqworkloop_owner_override(kqwl) != THREAD_QOS_UNSPECIFIED) { + thread_drop_ipc_override(cur_owner); + } + thread_ends_owning_workloop(cur_owner); + kqwl->kqwl_owner = THREAD_NULL; + } + + return cur_owner; } /* - * remove all knotes referencing a specified fd + * kqueue_dealloc - detach all knotes from a kqueue and free it * - * Essentially an inlined knote_remove & knote_drop - * when we know for sure that the thing is a file - * - * Entered with the proc_fd lock already held. - * It returns the same way, but may drop it temporarily. + * 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). */ void -knote_fdclose(struct proc *p, int fd) +kqueue_dealloc(struct kqueue *kq) { - struct filedesc *fdp = p->p_fd; - struct klist *list; + struct proc *p; + struct filedesc *fdp; struct knote *kn; + int i; + + if (kq == NULL) + return; - list = &fdp->fd_knlist[fd]; - while ((kn = SLIST_FIRST(list)) != NULL) { - struct kqueue *kq = kn->kn_kq; + p = kq->kq_p; + fdp = p->p_fd; - kqlock(kq); + /* + * Workloops are refcounted by their knotes, so there's no point + * spending a lot of time under these locks just to deallocate one. + */ + if ((kq->kq_state & KQ_WORKLOOP) == 0) { + KNOTE_LOCK_CTX(knlc); + + proc_fdlock(p); + for (i = 0; i < fdp->fd_knlistsize; i++) { + kn = SLIST_FIRST(&fdp->fd_knlist[i]); + while (kn != NULL) { + if (kq == knote_get_kq(kn)) { + kqlock(kq); + proc_fdunlock(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 */ + kn = SLIST_FIRST(&fdp->fd_knlist[i]); + continue; + } + kn = SLIST_NEXT(kn, kn_link); + } + } + + knhash_lock(p); proc_fdunlock(p); - /* - * Convert the lock to a drop ref. - * If we get it, go ahead and drop it. - * Otherwise, we waited for it to - * be dropped by the other guy, so - * it is safe to move on in the list. - */ - if (kqlock2knotedrop(kq, kn)) { - kn->kn_fop->f_detach(kn); - knote_drop(kn, p); + if (fdp->fd_knhashmask != 0) { + for (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)) { + kqlock(kq); + knhash_unlock(p); + if (knote_lock(kq, kn, &knlc, KNOTE_KQ_LOCK_ON_SUCCESS)) { + knote_drop(kq, kn, &knlc); + } + knhash_lock(p); + /* start over at beginning of list */ + kn = SLIST_FIRST(&fdp->fd_knhash[i]); + continue; + } + kn = SLIST_NEXT(kn, kn_link); + } + } } - - proc_fdlock(p); + knhash_unlock(p); + } + + if (kq->kq_state & KQ_WORKLOOP) { + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + thread_t cur_owner = kqworkloop_invalidate(kqwl); + + if (cur_owner) thread_deallocate(cur_owner); + + if (kqwl->kqwl_request.kqr_state & KQR_ALLOCATED_TURNSTILE) { + struct turnstile *ts; + turnstile_complete((uintptr_t)kqwl, &kqwl->kqwl_turnstile, &ts); + turnstile_cleanup(); + turnstile_deallocate(ts); + } else { + assert(kqwl->kqwl_turnstile == NULL); + } + } - /* the fd tables may have changed - start over */ - list = &fdp->fd_knlist[fd]; + /* + * 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); + lck_spin_destroy(&kq->kq_reqlock, kq_lck_grp); + + if (kq->kq_state & KQ_WORKQ) { + zfree(kqworkq_zone, (struct kqworkq *)kq); + } else if (kq->kq_state & KQ_WORKLOOP) { + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + + assert(kqwl->kqwl_retains == 0); + lck_mtx_destroy(&kqwl->kqwl_statelock, kq_lck_grp); + zfree(kqworkloop_zone, kqwl); + } else { + zfree(kqfile_zone, (struct kqfile *)kq); } } -/* proc_fdlock held on entry (and exit) */ -static int -knote_fdpattach(struct knote *kn, struct filedesc *fdp, __unused struct proc *p) +static inline void +kqueue_retain(struct kqueue *kq) { - struct klist *list = NULL; + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + uint32_t previous; - if (! kn->kn_fop->f_isfd) { - if (fdp->fd_knhashmask == 0) - fdp->fd_knhash = hashinit(KN_HASHSIZE, M_KQUEUE, - &fdp->fd_knhashmask); - list = &fdp->fd_knhash[KN_HASH(kn->kn_id, fdp->fd_knhashmask)]; - } else { - if ((u_int)fdp->fd_knlistsize <= kn->kn_id) { - u_int size = 0; + if ((kq->kq_state & KQ_DYNAMIC) == 0) + return; - /* have to grow the fd_knlist */ - size = fdp->fd_knlistsize; - while (size <= kn->kn_id) - size += KQEXTENT; - MALLOC(list, struct klist *, - size * sizeof(struct klist *), M_KQUEUE, M_WAITOK); - if (list == NULL) - return (ENOMEM); - - 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; + 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); +} + +#define KQUEUE_CANT_BE_LAST_REF 0 +#define KQUEUE_MIGHT_BE_LAST_REF 1 + +static inline int +kqueue_release(kqueue_t kqu, __assert_only int possibly_last) +{ + if ((kqu.kq->kq_state & KQ_DYNAMIC) == 0) { + return 0; + } + + assert(kqu.kq->kq_state & KQ_WORKLOOP); /* for now */ + uint32_t refs = OSDecrementAtomic(&kqu.kqwl->kqwl_retains); + if (__improbable(refs == 0)) { + panic("kq(%p) over-release", kqu.kq); + } + if (refs == 1) { + assert(possibly_last); + } + return refs == 1; +} + +int +kqueue_body(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); + if (error) { + return (error); + } + + kq = kqueue_alloc(p, 0); + if (kq == NULL) { + fp_free(p, fd, fp); + return (ENOMEM); + } + + fp->f_flag = FREAD | FWRITE; + fp->f_ops = &kqueueops; + fp->f_data = kq; + + proc_fdlock(p); + *fdflags(p, fd) |= UF_EXCLOSE; + procfdtbl_releasefd(p, fd, NULL); + fp_drop(p, fd, fp, 1); + proc_fdunlock(p); + + *retval = fd; + return (error); +} + +int +kqueue(struct proc *p, __unused struct kqueue_args *uap, int32_t *retval) +{ + return (kqueue_body(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) +{ + int advance; + int error; + + 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; + + 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]; + + } else { + struct kevent_qos_s kevqos; + + bzero(kevp, sizeof (*kevp)); + + 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); +} + +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; + + /* + * 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); + + if (IS_64BIT_PROCESS(p)) { + struct user64_kevent kev64; + + 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; + + 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); +} + +static int +kevent_get_data_size( + struct proc *p, + uint64_t data_available, + unsigned int flags, + user_size_t *residp) +{ + user_size_t resid; + int error = 0; + + 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; +} + +static int +kevent_put_data_size( + struct proc *p, + uint64_t data_available, + unsigned int flags, + user_size_t resid) +{ + int error = 0; + + 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)); + } + } + return error; +} + +/* + * kevent_continue - continue a kevent syscall after blocking + * + * assume we inherit a use count on the kq fileglob. + */ +__attribute__((noreturn)) +static void +kevent_continue(__unused struct kqueue *kq, void *data, int error) +{ + 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); + } + + /* 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); +} + +/* + * kevent - [syscall] register and wait for kernel events + * + */ +int +kevent(struct proc *p, struct kevent_args *uap, int32_t *retval) +{ + unsigned int flags = KEVENT_FLAG_LEGACY32; + + 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); +} + +int +kevent64(struct proc *p, struct kevent64_args *uap, int32_t *retval) +{ + unsigned int flags; + + /* restrict to user flags and set legacy64 */ + flags = uap->flags & KEVENT_FLAG_USER; + flags |= KEVENT_FLAG_LEGACY64; + + 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); +} + +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); +} + +int +kevent_id(struct proc *p, struct kevent_id_args *uap, int32_t *retval) +{ + /* restrict to user flags */ + uap->flags &= KEVENT_FLAG_USER; + + 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); +} + +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); +} + +static int +kevent_get_timeout(struct proc *p, + user_addr_t utimeout, + unsigned int flags, + struct timeval *atvp) +{ + struct timeval atv; + 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; + } + *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; + } + } 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; + } + kqunlock(kq); + return 0; +} + +#define KQ_HASH(val, mask) (((val) ^ (val >> 8)) & (mask)) +#define CONFIG_KQ_HASHSIZE CONFIG_KN_HASHSIZE + +static inline void +kqhash_lock(proc_t p) +{ + lck_mtx_lock_spin_always(&p->p_fd->fd_kqhashlock); +} + +static inline void +kqhash_lock_held(__assert_only proc_t p) +{ + LCK_MTX_ASSERT(&p->p_fd->fd_kqhashlock, LCK_MTX_ASSERT_OWNED); +} + +static inline void +kqhash_unlock(proc_t p) +{ + lck_mtx_unlock(&p->p_fd->fd_kqhashlock); +} + +static void +kqueue_hash_init_if_needed(proc_t p) +{ + 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); + + /* 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); + } + } +} + +/* + * Called with the kqhash_lock() held + */ +static void +kqueue_hash_insert( + struct proc *p, + kqueue_id_t id, + struct kqueue *kq) +{ + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + struct filedesc *fdp = p->p_fd; + struct kqlist *list; + + /* should hold the kq hash lock */ + kqhash_lock_held(p); + + if ((kq->kq_state & KQ_DYNAMIC) == 0) { + assert(kq->kq_state & KQ_DYNAMIC); + return; + } + + /* only dynamically allocate workloop kqs for now */ + assert(kq->kq_state & KQ_WORKLOOP); + assert(fdp->fd_kqhash); + + kqwl->kqwl_dynamicid = id; + + list = &fdp->fd_kqhash[KQ_HASH(id, fdp->fd_kqhashmask)]; + SLIST_INSERT_HEAD(list, kqwl, kqwl_hashlink); +} + +/* Called with kqhash_lock held */ +static void +kqueue_hash_remove( + struct proc *p, + struct kqueue *kq) +{ + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + struct filedesc *fdp = p->p_fd; + struct kqlist *list; + + /* should hold the kq hash lock */ + kqhash_lock_held(p); + + if ((kq->kq_state & KQ_DYNAMIC) == 0) { + assert(kq->kq_state & KQ_DYNAMIC); + 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); + + if (fdp->fd_kqhashmask == 0) return NULL; + + 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; + + assert(kq->kq_state & KQ_DYNAMIC); + assert(kq->kq_state & KQ_WORKLOOP); /* for now */ + return kq; + } + } + return NULL; +} + +static inline void +kqueue_release_last(struct proc *p, kqueue_t kqu) +{ + struct kqueue *kq = kqu.kq; + if (kq->kq_state & KQ_DYNAMIC) { + kqhash_lock(p); + if (kqueue_release(kq, KQUEUE_MIGHT_BE_LAST_REF)) { + thread_t cur_owner = kqworkloop_invalidate(kqu.kqwl); + kqueue_hash_remove(p, kq); + kqhash_unlock(p); + if (cur_owner) thread_deallocate(cur_owner); + kqueue_dealloc(kq); + } else { + kqhash_unlock(p); + } + } +} + +/* + * 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 filedesc *fdp = p->p_fd; + struct kqlist *list; + struct kqworkloop *kqwl, *kqwln; + struct kqlist tofree; + int i; + + if (!(fdp->fd_flags & FD_WORKLOOP)) { + return; + } + + SLIST_INIT(&tofree); + + kqhash_lock(p); + assert(fdp->fd_kqhashmask != 0); + + for (i = 0; i <= (int)fdp->fd_kqhashmask; i++) { + list = &fdp->fd_kqhash[i]; + SLIST_FOREACH_SAFE(kqwl, list, 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); + SLIST_REMOVE(list, kqwl, kqworkloop, kqwl_hashlink); + SLIST_INSERT_HEAD(&tofree, kqwl, kqwl_hashlink); + } + } + + kqhash_unlock(p); + + SLIST_FOREACH_SAFE(kqwl, &tofree, kqwl_hashlink, kqwln) { + struct kqueue *kq = (struct kqueue *)kqwl; + __assert_only bool released; + released = kqueue_release(kq, KQUEUE_MIGHT_BE_LAST_REF); + assert(released); + kqueue_dealloc(kq); + } +} + +static struct kqueue * +kevent_get_bound_kqworkloop(thread_t thread) +{ + struct uthread *ut = get_bsdthread_info(thread); + struct kqrequest *kqr = ut->uu_kqr_bound; + + return kqr ? (struct kqueue *)kqr_kqworkloop(kqr) : NULL; +} + +static int +kevent_get_kq(struct proc *p, kqueue_id_t id, workq_threadreq_param_t *trp, + unsigned int flags, struct fileproc **fpp, int *fdp, + struct kqueue **kqp) +{ + struct filedesc *descp = p->p_fd; + struct fileproc *fp = NULL; + struct kqueue *kq = NULL; + int fd = 0; + int error = 0; + thread_t th = current_thread(); + + assert(!trp || (flags & KEVENT_FLAG_WORKLOOP)); + + /* Was the workloop flag passed? Then it is for sure only a workloop */ + if (flags & KEVENT_FLAG_DYNAMIC_KQUEUE) { + assert(flags & KEVENT_FLAG_WORKLOOP); + assert(!trp || (flags & KEVENT_FLAG_DYNAMIC_KQ_MUST_NOT_EXIST)); + kq = kevent_get_bound_kqworkloop(th); + + /* + * 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. + */ + if (id == (kqueue_id_t)-1 && + (flags & KEVENT_FLAG_KERNEL) && + (flags & KEVENT_FLAG_WORKLOOP)) { + + if (!is_workqueue_thread(th) || !kq) { + return EINVAL; + } + + kqueue_retain(kq); + goto out; + } + + if (id == 0 || id == (kqueue_id_t)-1) { + return EINVAL; + } + + /* try shortcut on kq lookup for bound threads */ + if (kq != NULL && ((struct kqworkloop *)kq)->kqwl_dynamicid == id) { + + if (flags & KEVENT_FLAG_DYNAMIC_KQ_MUST_NOT_EXIST) { + return EEXIST; + } + + /* retain a reference while working with this kq. */ + assert(kq->kq_state & KQ_DYNAMIC); + kqueue_retain(kq); + goto out; + } + + /* look for the kq on the hash table */ + kqhash_lock(p); + kq = kqueue_hash_lookup(p, id); + if (kq == NULL) { + kqhash_unlock(p); + + if (flags & KEVENT_FLAG_DYNAMIC_KQ_MUST_EXIST) { + return ENOENT; + } + + struct kqueue *alloc_kq; + alloc_kq = kqueue_alloc(p, flags); + if (!alloc_kq) { + return ENOMEM; + } + + 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; + if (trp) { + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + kqwl->kqwl_params = trp->trp_value; + } + kqueue_hash_insert(p, id, kq); + kqhash_unlock(p); + } else if (flags & KEVENT_FLAG_DYNAMIC_KQ_MUST_NOT_EXIST) { + /* lost race and caller wants an error */ + kqhash_unlock(p); + kqueue_release(alloc_kq, KQUEUE_MIGHT_BE_LAST_REF); + kqueue_dealloc(alloc_kq); + return EEXIST; + } 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 { + + if (flags & KEVENT_FLAG_DYNAMIC_KQ_MUST_NOT_EXIST) { + kqhash_unlock(p); + return EEXIST; + } + + /* retain a reference while working with this kq. */ + assert(kq->kq_state & KQ_DYNAMIC); + kqueue_retain(kq); + kqhash_unlock(p); + } + + } else if (flags & KEVENT_FLAG_WORKQ) { + /* must already exist for bound threads. */ + if (flags & KEVENT_FLAG_KERNEL) { + assert(descp->fd_wqkqueue != NULL); + } + + /* + * 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. + */ + 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); + } + } + } 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; +} + +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); + } +} + +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; + + if (workloop_id == 0 || p->p_dispatchqueue_serialno_offset == 0) { + return 0; + } + addr = (user_addr_t)(workloop_id + p->p_dispatchqueue_serialno_offset); + + 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; +} + +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 = OS_REASON_NULL; + mach_vm_address_t addr; + uint32_t reason_size; + + kqhash_lock(p); + 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) { + struct kqueue *kq = &kqwl->kqwl_kqueue; + if ((kq->kq_state & KQ_DYNAMIC) && kqwl->kqwl_owner == thread) { + workloop_id = kqwl->kqwl_dynamicid; + break; + } + } + } + } + kqhash_unlock(p); + + reason = os_reason_create(OS_REASON_LIBSYSTEM, + OS_REASON_LIBSYSTEM_CODE_WORKLOOP_OWNERSHIP_LEAK); + if (reason == OS_REASON_NULL) { + goto out; + } + + 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) { + goto out; + } + + if (workloop_id) { + struct kcdata_descriptor *kcd = &reason->osr_kcd_descriptor; + + 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)); + } + + 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)); + } + } +out: +#if DEVELOPMENT || DEBUG + if (kevent_debug_flags() & KEVENT_PANIC_ON_WORKLOOP_OWNERSHIP_LEAK) { + panic("thread %p in task %p is leaked workloop 0x%016llx ownership", + thread, p->task, workloop_id); + } + 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 +} + +static inline boolean_t +kevent_args_requesting_events(unsigned int flags, int nevents) +{ + return (!(flags & KEVENT_FLAG_ERROR_EVENTS) && nevents > 0); +} + +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) +{ + uthread_t ut; + struct kqueue *kq; + struct fileproc *fp = NULL; + int fd = 0; + struct kevent_internal_s kev; + int error, noutputs, register_rc; + bool needs_end_processing = false; + struct timeval atv; + user_size_t data_size; + user_size_t data_resid; + thread_t thread = current_thread(); + KNOTE_LOCK_CTX(knlc); + + /* 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; + + if (flags & KEVENT_FLAG_PARKING) { + if (!kevent_args_requesting_events(flags, nevents) || id != (kqueue_id_t)-1) + return EINVAL; + } + + /* restrict dynamic kqueue allocation to workloops (for now) */ + if ((flags & (KEVENT_FLAG_DYNAMIC_KQUEUE | KEVENT_FLAG_WORKLOOP)) == KEVENT_FLAG_DYNAMIC_KQUEUE) + return EINVAL; + + if ((flags & (KEVENT_FLAG_WORKLOOP)) && (flags & (KEVENT_FLAG_WORKQ))) + return EINVAL; + + if (flags & (KEVENT_FLAG_DYNAMIC_KQ_MUST_EXIST | KEVENT_FLAG_DYNAMIC_KQ_MUST_NOT_EXIST)) { + + /* 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; + } + + /* 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; + } + + /* 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, NULL, flags, &fp, &fd, &kq); +#if CONFIG_WORKLOOP_DEBUG + ut = (uthread_t)get_bsdthread_info(thread); + UU_KEVENT_HISTORY_WRITE_ENTRY(ut, { + .uu_kqid = id, + .uu_kq = error ? NULL : kq, + .uu_error = error, + .uu_nchanges = nchanges, + .uu_nevents = nevents, + .uu_flags = flags, + }); +#endif // CONFIG_WORKLOOP_DEBUG + if (error) + return error; + + /* only bound threads can receive events on workloops */ + if (flags & KEVENT_FLAG_WORKLOOP) { + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + struct kqrequest *kqr = &kqwl->kqwl_request; + + assert(kq->kq_state & KQ_WORKLOOP); + + if (kevent_args_requesting_events(flags, nevents)) { + if (kq != kevent_get_bound_kqworkloop(thread)) { + error = EXDEV; + goto out; + } + + kq_req_lock(kqwl); + /* + * 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. + */ + kqr->kqr_state &= ~KQR_R2K_NOTIF_ARMED; + needs_end_processing = true; + kq_req_unlock(kq); + } + + if (id_out) { + *id_out = kqwl->kqwl_dynamicid; + } + + } + + /* 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; + + /* Make sure user doesn't pass in any system flags */ + kev.flags &= ~EV_SYSFLAGS; + + register_rc = kevent_register(kq, &kev, &knlc); + if (register_rc & FILTER_REGISTER_WAIT) { + kqlock_held(kq); + + // 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 && nevents >= 1 && (flags & KEVENT_FLAG_ERROR_EVENTS)) { + struct _kevent_register *cont_args; + /* store the continuation/completion data in the uthread */ + ut = (uthread_t)get_bsdthread_info(thread); + cont_args = &ut->uu_save.uus_kevent_register; + cont_args->kev = kev; + cont_args->kq = kq; + cont_args->fp = fp; + cont_args->fd = fd; + cont_args->ueventlist = ueventlist; + cont_args->flags = flags; + cont_args->retval = retval; + cont_args->eventcount = nevents; + cont_args->eventout = noutputs; + knote_fops(cont_args->knote)->f_post_register_wait(ut, &knlc, cont_args); + panic("f_post_register_wait returned (kev: %p)", &kev); + } + + kev.flags |= EV_ERROR; + kev.data = ENOTSUP; + knote_unlock(kq, knlc.knlc_knote, &knlc, KNOTE_KQ_UNLOCK); + } + + // keep in sync with kevent_register_wait_return() + if (nevents > 0 && (kev.flags & (EV_ERROR|EV_RECEIPT))) { + if ((kev.flags & EV_ERROR) == 0) { + 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; + } + 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) { + struct _kevent *cont_args; + /* store the continuation/completion data in the uthread */ + ut = (uthread_t)get_bsdthread_info(thread); + cont_args = &ut->uu_save.uus_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; + + /* + * kqworkloop_end_processing() will happen at the end of kqueue_scan() + */ + needs_end_processing = false; + + 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); + } + } + +out: + if (__improbable(needs_end_processing)) { + /* + * If we didn't through kqworkloop_end_processing(), + * we need to do it here. + */ + kqlock(kq); + kqworkloop_end_processing((struct kqworkloop *)kq, 0, 0); + kqunlock(kq); + } + kevent_put_kq(p, id, fp, kq); + + /* don't restart after signals... */ + if (error == ERESTART) + error = EINTR; + else if (error == EWOULDBLOCK) + error = 0; + if (error == 0) + *retval = noutputs; + return (error); +} + + +/* + * kevent_callback - callback for each individual event + * + * called with nothing locked + * caller holds a reference on the kqueue + */ +static int +kevent_callback(__unused struct kqueue *kq, struct kevent_internal_s *kevp, + void *data) +{ + struct _kevent *cont_args; + int error; + + cont_args = (struct _kevent *)data; + assert(cont_args->eventout < cont_args->eventcount); + + /* + * 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 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); +} + +/* + * kevent_description - format a description of a kevent for diagnostic output + * + * called with a 256-byte string buffer + */ + +char * +kevent_description(struct kevent_internal_s *kevp, char *s, size_t n) +{ + 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] ); + + return (s); +} + +static int +kevent_register_validate_priority(struct kqueue *kq, struct knote *kn, + struct kevent_internal_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; +} + +/* + * 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_internal_s *kev) +{ + thread_t thread = current_thread(); + struct uthread *uth = get_bsdthread_info(thread); + + assert(knote_fops(kn)->f_extended_codes); + + if (kn->kn_hook == NULL) { + thread_reference(thread); + kn->kn_hook = thread; + } else if (kn->kn_hook != thread) { + /* + * kn_hook 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; + } + + uth->uu_save.uus_kevent_register.knote = kn; + return FILTER_REGISTER_WAIT; +} + +/* + * 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_hook; + kn->kn_hook = NULL; + thread_deallocate(thread); +} + +/* + * 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, + struct knote_lock_ctx *knlc, thread_continue_t cont, + struct _kevent_register *cont_args) +{ + knote_unlock(cont_args->kq, cont_args->knote, knlc, KNOTE_KQ_UNLOCK); + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_NOT_HELD); + cont_args->handoff_thread = thread; + thread_handoff_parameter(thread, cont, cont_args); +} + +/* + * 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 kqueue *kq = cont_args->kq; + proc_t p = kq->kq_p; + struct kevent_internal_s *kev = &cont_args->kev; + int error = 0; + + if (cont_args->handoff_thread) { + thread_deallocate(cont_args->handoff_thread); + } + + if (kev->flags & (EV_ERROR|EV_RECEIPT)) { + if ((kev->flags & EV_ERROR) == 0) { + kev->flags |= EV_ERROR; + kev->data = 0; + } + error = kevent_copyout(kev, &cont_args->ueventlist, p, cont_args->flags); + if (error == 0) cont_args->eventout++; + } + + kevent_put_kq(p, cont_args->fd, cont_args->fp, kq); + if (error == 0) { + *cont_args->retval = cont_args->eventout; + } + unix_syscall_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 + */ + +int +kevent_register(struct kqueue *kq, struct kevent_internal_s *kev, + struct knote_lock_ctx *knlc) +{ + 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; + + if (kev->filter < 0) { + if (kev->filter + EVFILT_SYSCOUNT < 0) { + error = EINVAL; + goto out; + } + 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; + } + + /* 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); + } else { + KDBG_FILTERED(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) + */ + + 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; + kn->kn_kq_packed = (intptr_t)(struct kqueue *)kq; + kqueue_retain(kq); /* retain a kq ref */ + kn->kn_filtid = ~kev->filter; + 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; + if (kev->flags & EV_DISABLE) + kn->kn_status |= KN_DISABLED; + + /* + * 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; + knote_reset_priority(kn, kev->qos); + + /* Add the knote for lookup thru the fd table */ + error = kq_add_knote(kq, kn, knlc, p); + 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) { + goto restart; + } + goto out; + } + + /* fp reference count now applies to knote */ + + /* + * 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; + } + + kqlock(kq); + + if (kn->kn_flags & EV_ERROR) { + /* + * Failed to attach correctly, so drop. + */ + kn->kn_status &= ~(KN_ATTACHED | KN_ATTACHING); + error = kn->kn_data; + knote_drop(kq, kn, knlc); + result = 0; + goto out; + } + + /* + * 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. + */ + kn->kn_status &= ~KN_ATTACHING; + knote_set_qos_overcommit(kn); + + if (result & FILTER_ACTIVE) { + if (result & FILTER_ADJUST_EVENT_QOS_BIT) + knote_adjust_qos(kq, kn, result); + knote_activate(kn); + } + + } 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 + */ + + 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). + */ + + 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; + } + + if ((kev->flags & EV_ENABLE) == 0 && + (kn->kn_status & (KN_DISPATCH2 | KN_DISABLED)) == + (KN_DISPATCH2 | KN_DISABLED)) { + kn->kn_status |= KN_DEFERDELETE; + error = EINPROGRESS; + goto out_unlock; + } + + 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. + */ + + 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 (kev->flags & EV_ERROR) { + result = 0; + } else { + /* accept new kevent state */ + if ((kn->kn_status & KN_UDATA_SPECIFIC) == 0) + kn->kn_udata = kev->udata; + if (kev->flags & EV_DISABLE) + knote_disable(kn); + if (result & (FILTER_UPDATE_REQ_QOS | FILTER_ADJUST_EVENT_QOS_BIT)) + knote_dequeue(kn); + if ((result & FILTER_UPDATE_REQ_QOS) && + kev->qos && kev->qos != kn->kn_qos) { + knote_reset_priority(kn, kev->qos); + } + if (result & FILTER_ACTIVE) { + thread_qos_t qos; + if (result & FILTER_ADJUST_EVENT_QOS_BIT) { + if (knote_should_apply_qos_override(kq, kn, result, &qos)) { + knote_apply_qos_override(kn, qos); + } + } + knote_activate(kn); + } + if (result & (FILTER_UPDATE_REQ_QOS | FILTER_ADJUST_EVENT_QOS_BIT)) { + if (knote_enqueue(kn) && (kn->kn_status & KN_ACTIVE)) { + knote_wakeup(kn); + } + } + if (kev->flags & EV_ENABLE) + knote_enable(kn); + } + } + +out_unlock: + if ((result & FILTER_REGISTER_WAIT) == 0) { + /* + * When the filter asked for a post-register wait, + * we leave the knote and kqueue locked for kevent_register() + * to call the filter's f_post_register_wait hook. + */ + knote_unlock(kq, kn, knlc, KNOTE_KQ_UNLOCK); + } + +out: + /* 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 +knote_process(struct knote *kn, + kevent_callback_t callback, + void *callback_data, + struct filt_process_s *process_data) +{ + struct kevent_internal_s kev; + struct kqueue *kq = knote_get_kq(kn); + KNOTE_LOCK_CTX(knlc); + int result = FILTER_ACTIVE; + int error = 0; + bool drop = false; + + bzero(&kev, sizeof(kev)); + + /* + * 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))); + + 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); + } + + if ((kn->kn_status & KN_DROPPING) || + !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; + } + + /* + * 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(kn); + + if (kn->kn_status & (KN_DEFERDELETE | KN_VANISHED)) { + /* create fake event */ + kev.filter = kn->kn_filter; + kev.ident = kn->kn_id; + kev.flags = (kn->kn_status & KN_DEFERDELETE) ? EV_DELETE : EV_VANISHED; + kev.flags |= (EV_DISPATCH2 | EV_ONESHOT); + kev.udata = kn->kn_udata; + } else { + /* deactivate - so new activations indicate a wakeup */ + knote_deactivate(kn); + + kqunlock(kq); + result = filter_call(knote_fops(kn), f_process(kn, process_data, &kev)); + kqlock(kq); + } + + /* + * 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(kn); + } + knote_unlock(kq, kn, &knlc, KNOTE_KQ_LOCK_ALWAYS); + return EJUSTRETURN; + } + + 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); + + if (kev.flags & EV_ONESHOT) { + 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); + } else { + drop = true; + } + } 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); + } + + /* + * 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 (drop) { + knote_drop(kq, kn, &knlc); + } else { + knote_unlock(kq, kn, &knlc, KNOTE_KQ_UNLOCK); + } + + if (kev.flags & EV_VANISHED) { + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KNOTE_VANISHED), + kev.ident, kn->kn_udata, kn->kn_status | (kn->kn_id << 32), + kn->kn_filtid); + } + + error = (callback)(kq, &kev, callback_data); + 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 +kqworkq_acknowledge_events(struct kqworkq *kqwq, struct kqrequest *kqr, + int kevent_flags, int kqwqae_op) +{ + thread_qos_t old_override = THREAD_QOS_UNSPECIFIED; + thread_t thread = kqr->kqr_thread; + struct knote *kn; + int rc = 0; + bool seen_stayactive = false, unbind; + + kqlock_held(&kqwq->kqwq_kqueue); + + if (!TAILQ_EMPTY(&kqr->kqr_suppressed)) { + /* + * 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(&kqr->kqr_suppressed)) != NULL) { + assert(kn->kn_status & KN_SUPPRESSED); + knote_unsuppress(kn); + if (kn->kn_status & KN_STAYACTIVE) { + seen_stayactive = true; + } + } + } + + kq_req_lock(kqwq); + +#if DEBUG || DEVELOPMENT + thread_t self = current_thread(); + struct uthread *ut = get_bsdthread_info(self); + + assert(kqr->kqr_state & KQR_THREQUESTED); + assert(kqr->kqr_thread == self); + assert(ut->uu_kqr_bound == kqr); +#endif // DEBUG || DEVELOPMENT + + if (kqwqae_op == KQWQAE_UNBIND) { + unbind = true; + } else if ((kevent_flags & KEVENT_FLAG_PARKING) == 0) { + unbind = false; + } else if (kqwqae_op == KQWQAE_BEGIN_PROCESSING && seen_stayactive) { + /* + * When we unsuppress stayactive knotes, for the kind that are hooked + * through select, we need to process once before we can assert there's + * no event pending. Hence we can't unbind during BEGIN PROCESSING. + */ + unbind = false; + } else { + unbind = ((kqr->kqr_state & KQR_WAKEUP) == 0); + } + 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->kqr_state & KQR_WAKEUP) { + kqueue_threadreq_initiate(&kqwq->kqwq_kqueue, kqr, + kqr->kqr_qos_index, 0); + } + } + + 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->kqr_state &= ~KQR_WAKEUP; + } + + kq_req_unlock(kqwq); + + if (old_override) { + thread_drop_ipc_override(thread); + } + + return rc; +} + +/* + * 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, struct kqrequest *kqr, + int kevent_flags) +{ + int rc = 0; + + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWQ_PROCESS_BEGIN) | DBG_FUNC_START, + 0, kqr->kqr_qos_index); + + rc = kqworkq_acknowledge_events(kqwq, kqr, kevent_flags, + KQWQAE_BEGIN_PROCESSING); + + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWQ_PROCESS_BEGIN) | DBG_FUNC_END, + thread_tid(kqr->kqr_thread), kqr->kqr_state); + + return rc; +} + +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 thread_qos_t +kqworkloop_acknowledge_events(struct kqworkloop *kqwl) +{ + struct kqrequest *kqr = &kqwl->kqwl_request; + kq_index_t qos = THREAD_QOS_UNSPECIFIED; + 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) { + qos = MAX(qos, knote_get_qos_override_index(kn)); + continue; + } + knote_unsuppress(kn); + } + + return qos; +} + +static int +kqworkloop_begin_processing(struct kqworkloop *kqwl, unsigned int kevent_flags) +{ + struct kqrequest *kqr = &kqwl->kqwl_request; + struct kqueue *kq = &kqwl->kqwl_kqueue; + thread_qos_t old_override = THREAD_QOS_UNSPECIFIED, qos_override; + thread_t thread = kqr->kqr_thread; + int rc = 0, op = KQWL_UTQ_NONE; + + kqlock_held(kq); + + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_PROCESS_BEGIN) | DBG_FUNC_START, + kqwl->kqwl_dynamicid, 0, 0); + + /* nobody else should still be processing */ + assert((kq->kq_state & KQ_PROCESSING) == 0); + + kq->kq_state |= KQ_PROCESSING; + + if (!TAILQ_EMPTY(&kqr->kqr_suppressed)) { + op = KQWL_UTQ_RESET_WAKEUP_OVERRIDE; + } + + 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->kqr_state & KQR_THOVERCOMMIT) { + op = KQWL_UTQ_PARKING; + } else { + op = KQWL_UTQ_UNBINDING; + } + } + if (op == KQWL_UTQ_NONE) { + goto done; + } + + qos_override = kqworkloop_acknowledge_events(kqwl); + + kq_req_lock(kqwl); + + if (op == KQWL_UTQ_UNBINDING) { + old_override = kqworkloop_unbind_locked(kqwl, thread); + (void)kqueue_release(kqwl, KQUEUE_CANT_BE_LAST_REF); + } + 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 KQR_WAKEUP when looking at stay active knotes. + * We need to process once, and kqworkloop_end_processing will + * handle the unbind. + */ + } else if ((kqr->kqr_state & KQR_WAKEUP) == 0 || kqwl->kqwl_owner) { + old_override = kqworkloop_unbind_locked(kqwl, thread); + (void)kqueue_release(kqwl, KQUEUE_CANT_BE_LAST_REF); + rc = -1; + } + } else if (op == KQWL_UTQ_UNBINDING) { + if (kqr->kqr_thread == thread) { + /* + * The thread request fired again, passed the admission check and + * got bound to the current thread again. + */ + } else { + rc = -1; + } + } + + 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. + */ + kqr->kqr_wakeup_indexes &= ~KQWL_STAYACTIVE_FIRED_BIT; + } else { + kq->kq_state &= ~KQ_PROCESSING; + } + + kq_req_unlock(kqwl); + + if (old_override) { + thread_drop_ipc_override(thread); + } + +done: + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_PROCESS_BEGIN) | DBG_FUNC_END, + kqwl->kqwl_dynamicid, 0, 0); + + return rc; +} + +/* + * 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 +kqfile_begin_processing(struct kqueue *kq) +{ + struct kqtailq *suppressq; + + kqlock_held(kq); + + assert((kq->kq_state & (KQ_WORKQ | KQ_WORKLOOP)) == 0); + KDBG_FILTERED(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->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 ((kq->kq_state & KQ_PROCESSING) == 0) + break; + + /* if someone else is processing the queue, wait */ + kq->kq_state |= KQ_PROCWAIT; + suppressq = kqueue_get_suppressed_queue(kq, NULL); + waitq_assert_wait64((struct waitq *)&kq->kq_wqs, + CAST_EVENT64_T(suppressq), THREAD_UNINT | THREAD_WAIT_NOREPORT, + 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_KQFILE)) { + 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; +} + +/* + * 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. + */ +static int +kqworkq_end_processing(struct kqworkq *kqwq, struct kqrequest *kqr, + int kevent_flags) +{ + if (!kqueue_queue_empty(&kqwq->kqwq_kqueue, kqr->kqr_qos_index)) { + /* remember we didn't process everything */ + kq_req_lock(kqwq); + kqr->kqr_state |= KQR_WAKEUP; + kq_req_unlock(kqwq); + } + + 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; + } + } + + return 0; +} + +/* + * 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). + * + * If successful and KEVENT_FLAG_PARKING was set in the kevent_flags, + * the kqworkloop is unbound from its servicer as a side effect. + */ +static int +kqworkloop_end_processing(struct kqworkloop *kqwl, int flags, int kevent_flags) +{ + struct kqueue *kq = &kqwl->kqwl_kqueue; + struct kqrequest *kqr = &kqwl->kqwl_request; + thread_qos_t old_override = THREAD_QOS_UNSPECIFIED, qos_override; + thread_t thread = kqr->kqr_thread; + int rc = 0; + + kqlock_held(kq); + + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_PROCESS_END) | DBG_FUNC_START, + kqwl->kqwl_dynamicid, 0, 0); + + if (flags & KQ_PROCESSING) { + assert(kq->kq_state & KQ_PROCESSING); + + /* + * 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])) { + kq_req_lock(kqwl); + kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_UPDATE_WAKEUP_QOS, + KQWL_BUCKET_STAYACTIVE); + kq_req_unlock(kqwl); + } + + /* + * 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); + } + } + + kq_req_lock(kqwl); + + if (kevent_flags & KEVENT_FLAG_PARKING) { + kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_PARKING, qos_override); + if ((kqr->kqr_state & KQR_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. + */ + kqr->kqr_wakeup_indexes &= ~KQWL_STAYACTIVE_FIRED_BIT; + rc = -1; + } else { + old_override = kqworkloop_unbind_locked(kqwl, thread); + (void)kqueue_release(kqwl, KQUEUE_CANT_BE_LAST_REF); + kq->kq_state &= ~flags; + } + } else { + kq->kq_state &= ~flags; + kqr->kqr_state |= KQR_R2K_NOTIF_ARMED; + kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_RECOMPUTE_WAKEUP_QOS, 0); + } + + kq_req_unlock(kqwl); + + if (old_override) { + thread_drop_ipc_override(thread); + } + + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_PROCESS_END) | DBG_FUNC_END, + kqwl->kqwl_dynamicid, 0, 0); + + return rc; +} + +/* + * Called with kqueue lock held. + */ +static void +kqfile_end_processing(struct kqueue *kq) +{ + struct knote *kn; + struct kqtailq *suppressq; + int procwait; + + kqlock_held(kq); + + assert((kq->kq_state & (KQ_WORKQ|KQ_WORKLOOP)) == 0); + + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQ_PROCESS_END), + VM_KERNEL_UNSLIDE_OR_PERM(kq), 0); + + /* + * Return suppressed knotes to their original state. + */ + suppressq = kqueue_get_suppressed_queue(kq, NULL); + while ((kn = TAILQ_FIRST(suppressq)) != NULL) { + assert(kn->kn_status & KN_SUPPRESSED); + knote_unsuppress(kn); + } + + 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); + } +} + +static int +kqueue_workloop_ctl_internal(proc_t p, uintptr_t cmd, uint64_t __unused options, + struct kqueue_workloop_params *params, int *retval) +{ + int error = 0; + int fd; + struct fileproc *fp; + struct kqueue *kq; + struct kqworkloop *kqwl; + struct filedesc *fdp = p->p_fd; + workq_threadreq_param_t trp = { }; + + switch (cmd) { + case KQ_WORKLOOP_CREATE: + if (!params->kqwlp_flags) { + error = EINVAL; + break; + } + + if ((params->kqwlp_flags & KQ_WORKLOOP_CREATE_SCHED_PRI) && + (params->kqwlp_sched_pri < 1 || + params->kqwlp_sched_pri > 63 /* MAXPRI_USER */)) { + error = EINVAL; + break; + } + + if ((params->kqwlp_flags & KQ_WORKLOOP_CREATE_SCHED_POL) && + invalid_policy(params->kqwlp_sched_pol)) { + error = EINVAL; + break; + } + + 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 (params->kqwlp_flags & KQ_WORKLOOP_CREATE_SCHED_PRI) { + trp.trp_flags |= TRP_PRIORITY; + trp.trp_pri = params->kqwlp_sched_pri; + } + if (params->kqwlp_flags & KQ_WORKLOOP_CREATE_SCHED_POL) { + trp.trp_flags |= TRP_POLICY; + trp.trp_pol = 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; + } + + error = kevent_get_kq(p, params->kqwlp_id, &trp, + KEVENT_FLAG_DYNAMIC_KQUEUE | KEVENT_FLAG_WORKLOOP | + KEVENT_FLAG_DYNAMIC_KQ_MUST_NOT_EXIST , &fp, &fd, &kq); + if (error) { + break; + } + + 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 = kevent_get_kq(p, params->kqwlp_id, NULL, + KEVENT_FLAG_DYNAMIC_KQUEUE | KEVENT_FLAG_WORKLOOP | + KEVENT_FLAG_DYNAMIC_KQ_MUST_EXIST , &fp, &fd, &kq); + if (error) { + break; + } + kqlock(kq); + kqwl = (struct kqworkloop *)kq; + trp.trp_value = kqwl->kqwl_params; + if (trp.trp_flags && !(trp.trp_flags & TRP_RELEASED)) { + trp.trp_flags |= TRP_RELEASED; + kqueue_release(kq, KQUEUE_CANT_BE_LAST_REF); + } else { + error = EINVAL; + } + kqunlock(kq); + kqueue_release_last(p, kq); + break; + } + *retval = 0; + return error; +} + +int +kqueue_workloop_ctl(proc_t p, struct kqueue_workloop_ctl_args *uap, int *retval) +{ + struct kqueue_workloop_params params = { + .kqwlp_id = 0, + }; + if (uap->sz < sizeof(params.kqwlp_version)) { + return EINVAL; + } + + size_t copyin_sz = MIN(sizeof(params), uap->sz); + int rv = copyin(uap->addr, ¶ms, copyin_sz); + if (rv) { + return rv; + } + + if (params.kqwlp_version != (int)uap->sz) { + return EINVAL; + } + + return kqueue_workloop_ctl_internal(p, uap->cmd, uap->options, ¶ms, + retval); +} + +/* + * kqueue_process - process the triggered events in a kqueue + * + * 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) + */ +static int +kqueue_process(struct kqueue *kq, + kevent_callback_t callback, + void *callback_data, + struct filt_process_s *process_data, + int *countp) +{ + struct uthread *ut = get_bsdthread_info(current_thread()); + struct kqrequest *kqr = ut->uu_kqr_bound; + struct knote *kn; + unsigned int flags = process_data ? process_data->fp_flags : 0; + int nevents = 0, error = 0, rc = 0; + struct kqtailq *base_queue, *queue; + kqueue_t kqu = { .kq = kq }; +#if DEBUG || DEVELOPMENT + int retries = 64; +#endif + + if (kq->kq_state & KQ_WORKQ) { + if (kqr == NULL || (kqr->kqr_state & KQR_WORKLOOP)) { + return EJUSTRETURN; + } + rc = kqworkq_begin_processing(kqu.kqwq, kqr, flags); + } else if (kq->kq_state & KQ_WORKLOOP) { + if (ut->uu_kqr_bound != &kqu.kqwl->kqwl_request) { + return EJUSTRETURN; + } + rc = kqworkloop_begin_processing(kqu.kqwl, flags); + } else { + rc = kqfile_begin_processing(kq); + } + + if (rc == -1) { + /* Nothing to process */ + *countp = 0; + return 0; + } + + /* + * 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_again: + if (kq->kq_state & KQ_WORKQ) { + base_queue = queue = &kqu.kqwq->kqwq_queue[kqr->kqr_qos_index]; + } else if (kq->kq_state & KQ_WORKLOOP) { + base_queue = &kqu.kqwl->kqwl_queue[0]; + queue = &kqu.kqwl->kqwl_queue[KQWL_NBUCKETS - 1]; + } else { + base_queue = queue = &kq->kq_queue[QOS_INDEX_KQFILE]; + } + + do { + while (error == 0 && (kn = TAILQ_FIRST(queue)) != NULL) { + error = knote_process(kn, callback, callback_data, process_data); + if (error == EJUSTRETURN) { + error = 0; + } else { + nevents++; + } + /* error is EWOULDBLOCK when the out event array is full */ + } + + if (error == EWOULDBLOCK) { + /* break out if no more space for additional events */ + error = 0; + break; + } + } while (queue-- > base_queue); + + *countp = nevents; + + /* + * If KEVENT_FLAG_PARKING is set, and no kevents have been returned, + * we want to unbind the kqrequest from the thread. + * + * 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 returned events however, end processing never fails. + */ + if (error || nevents) flags &= ~KEVENT_FLAG_PARKING; + if (kq->kq_state & KQ_WORKQ) { + rc = kqworkq_end_processing(kqu.kqwq, kqr, flags); + } else if (kq->kq_state & KQ_WORKLOOP) { + rc = kqworkloop_end_processing(kqu.kqwl, KQ_PROCESSING, flags); + } else { + kqfile_end_processing(kq); + rc = 0; + } + if (rc == -1) { + assert(flags & KEVENT_FLAG_PARKING); +#if DEBUG || DEVELOPMENT + if (retries-- == 0) { + panic("kevent: way too many knote_process retries, kq: %p (0x%02x)", + kq, kq->kq_state); + } +#endif + goto process_again; + } + return error; +} + +static void +kqueue_scan_continue(void *data, wait_result_t wait_result) +{ + thread_t self = current_thread(); + uthread_t ut = (uthread_t)get_bsdthread_info(self); + struct _kqueue_scan * cont_args = &ut->uu_save.uus_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); + if (error == 0 && count == 0) { + if (kq->kq_state & KQ_DRAIN) { + kqunlock(kq); + goto drain; + } + + if (kq->kq_state & KQ_WAKEUP) + goto retry; + + 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; + } + + /* call the continuation with the results */ + assert(cont_args->cont != NULL); + (cont_args->cont)(kq, cont_args->data, error); +} + + +/* + * kqueue_scan - scan and wait for events in a kqueue + * + * Process the triggered events in a kqueue. + * + * 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. + */ +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, + __unused 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); + + /* + * Determine which QoS index we are servicing + */ + flags = (process_data) ? process_data->fp_flags : 0; + fd = (process_data) ? process_data->fp_fd : -1; + + first = 1; + for (;;) { + wait_result_t wait_result; + int count; + + /* + * Make a pass through the kq to find events already + * triggered. + */ + kqlock(kq); + error = kqueue_process(kq, callback, callback_data, + process_data, &count); + 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 */ + } + + if (continuation) { + uthread_t ut = (uthread_t)get_bsdthread_info(current_thread()); + struct _kqueue_scan *cont_args = &ut->uu_save.uus_kqueue_scan; + + 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 (kq->kq_state & KQ_DRAIN) { + kqunlock(kq); + return EBADF; + } + + /* If awakened during processing, try again */ + if (kq->kq_state & KQ_WAKEUP) { + kqunlock(kq); + continue; + } + + /* 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) */ + + 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; + } + } + kqunlock(kq); + return (error); +} + + +/* + * XXX + * This could be expanded to call kqueue_scan, if desired. + */ +/*ARGSUSED*/ +static int +kqueue_read(__unused struct fileproc *fp, + __unused struct uio *uio, + __unused int flags, + __unused vfs_context_t ctx) +{ + return (ENXIO); +} + +/*ARGSUSED*/ +static int +kqueue_write(__unused struct fileproc *fp, + __unused struct uio *uio, + __unused int flags, + __unused vfs_context_t ctx) +{ + return (ENXIO); +} + +/*ARGSUSED*/ +static int +kqueue_ioctl(__unused struct fileproc *fp, + __unused u_long com, + __unused caddr_t data, + __unused vfs_context_t ctx) +{ + return (ENOTTY); +} + +/*ARGSUSED*/ +static int +kqueue_select(struct fileproc *fp, int which, void *wq_link_id, + __unused vfs_context_t ctx) +{ + 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); + + kqlock(kq); + + assert((kq->kq_state & KQ_WORKQ) == 0); + + /* + * 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); + + kq->kq_state |= KQ_SEL; + waitq_link((struct waitq *)&kq->kq_wqs, ut->uu_wqset, + WAITQ_SHOULD_LOCK, (uint64_t *)wq_link_id); + + /* always consume the reserved link object */ + waitq_link_release(*(uint64_t *)wq_link_id); + *(uint64_t *)wq_link_id = 0; + + /* + * 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 *)); + } + + if (kqfile_begin_processing(kq) == -1) { + kqunlock(kq); + return (0); + } + + queue = &kq->kq_queue[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); + } + + /* + * 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, NULL); + while ((kn = (struct knote *)TAILQ_FIRST(suppressq)) != NULL) { + KNOTE_LOCK_CTX(knlc); + int result = 0; + + /* 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; + } + + result = filter_call(knote_fops(kn), f_peek(kn)); + + kqlock(kq); + knote_unlock(kq, kn, &knlc, KNOTE_KQ_LOCK_ALWAYS); + + /* unsuppress it */ + knote_unsuppress(kn); + + /* has data or it has to report a vanish */ + if (result & FILTER_ACTIVE) { + retnum = 1; + goto out; + } + } + } + +out: + kqfile_end_processing(kq); + kqunlock(kq); + return (retnum); +} + +/* + * 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); +} + +/* + * 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 + +/*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); + uint16_t plevel = 0; + + assert((kqf->kqf_state & KQ_WORKQ) == 0); + + if (parentkq == kq || kn->kn_filter != EVFILT_READ) { + knote_set_error(kn, EINVAL); + return 0; + } + + /* + * 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. + */ + + kqlock(parentkq); + if (parentkq->kq_level > 0 && + parentkq->kq_level < kq->kq_level) + { + kqunlock(parentkq); + knote_set_error(kn, EINVAL); + return 0; + } else { + /* set parent level appropriately */ + 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; + } + + parentkq->kq_level = plevel; + kqunlock(parentkq); + + 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; + + int count = kq->kq_count; + kqunlock(kq); + return (count > 0); + } +} + +/* + * kqueue_drain - called when kq is closed + */ +/*ARGSUSED*/ +static int +kqueue_drain(struct fileproc *fp, __unused vfs_context_t ctx) +{ + struct kqueue *kq = (struct kqueue *)fp->f_fglob->fg_data; + + assert((kq->kq_state & KQ_WORKQ) == 0); + + kqlock(kq); + kq->kq_state |= KQ_DRAIN; + kqueue_interrupt(kq); + kqunlock(kq); + return (0); +} + +/*ARGSUSED*/ +int +kqueue_stat(struct kqueue *kq, void *ub, int isstat64, proc_t p) +{ + 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; + } 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; + } + kqunlock(kq); + return (0); +} + +/* + * Interact with the pthread kext to request a servicing there at a specific QoS + * level. + * + * - 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 +kqueue_threadreq_initiate(struct kqueue *kq, struct kqrequest *kqr, + kq_index_t qos, int flags) +{ + assert(kqr->kqr_state & KQR_WAKEUP); + assert(kqr->kqr_thread == THREAD_NULL); + assert((kqr->kqr_state & KQR_THREQUESTED) == 0); + struct turnstile *ts = TURNSTILE_NULL; + + if (workq_is_exiting(kq->kq_p)) { + return; + } + + /* Add a thread request reference on the kqueue. */ + kqueue_retain(kq); + + kq_req_held(kq); + + if (kq->kq_state & KQ_WORKLOOP) { + __assert_only struct kqworkloop *kqwl = (struct kqworkloop *)kq; + + assert(kqwl->kqwl_owner == THREAD_NULL); + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_THREQUEST), + kqwl->kqwl_dynamicid, 0, qos, kqr->kqr_state); + ts = kqwl->kqwl_turnstile; + } else { + assert(kq->kq_state & KQ_WORKQ); + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWQ_THREQUEST), + -1, 0, qos, kqr->kqr_state); + } + + kqr->kqr_state |= KQR_THREQUESTED; + + /* + * 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 0 // 45129862 + if ((kq->kq_state & KQ_WORKLOOP) && current_proc() == kq->kq_p) { + flags |= WORKQ_THREADREQ_SET_AST_ON_FAILURE; + } +#endif + if (qos == KQWQ_QOS_MANAGER) { + qos = WORKQ_THREAD_QOS_MANAGER; + } + if (!workq_kern_threadreq_initiate(kq->kq_p, kqr, ts, qos, flags)) { + /* + * 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. + */ + kqr->kqr_state &= ~(KQR_THREQUESTED | KQR_R2K_NOTIF_ARMED); + kqueue_release(kq, KQUEUE_CANT_BE_LAST_REF); + } +} + +/* + * kqueue_threadreq_bind_prepost - prepost the bind to kevent + * + * This is used when kqueue_threadreq_bind may cause a lock inversion. + */ +void +kqueue_threadreq_bind_prepost(struct proc *p __unused, workq_threadreq_t req, + thread_t thread) +{ + struct kqrequest *kqr = __container_of(req, struct kqrequest, kqr_req); + struct uthread *ut = get_bsdthread_info(thread); + + req->tr_binding_thread = thread; + ut->uu_kqr_bound = kqr; + req->tr_state = TR_STATE_BINDING; + + struct kqworkloop *kqwl = kqr_kqworkloop(kqr); + if (kqwl && kqwl->kqwl_turnstile) { + struct turnstile *ts = kqwl->kqwl_turnstile; + /* + * While a thread request is in flight, the workqueue + * is the interlock for the turnstile and can update the inheritor. + */ + turnstile_update_inheritor(ts, thread, TURNSTILE_IMMEDIATE_UPDATE | + TURNSTILE_INHERITOR_THREAD); + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_HELD); + } +} + +/* + * 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); + struct kqrequest *kqr = ut->uu_kqr_bound; + kqueue_t kqu = kqr_kqueue(p, kqr); + + kq_req_lock(kqu); + if (kqr->kqr_req.tr_state == TR_STATE_BINDING) { + kqueue_threadreq_bind(p, &kqr->kqr_req, thread, 0); + } + kq_req_unlock(kqu); +} + +static void +kqueue_threadreq_modify(struct kqueue *kq, struct kqrequest *kqr, kq_index_t qos) +{ + assert(kqr->kqr_state & KQR_THREQUESTED); + assert(kqr->kqr_thread == THREAD_NULL); + + kq_req_held(kq); + + int flags = 0; +#if 0 // 45129862 + if ((kq->kq_state & KQ_WORKLOOP) && kq->kq_p == current_proc()) { + flags |= WORKQ_THREADREQ_SET_AST_ON_FAILURE; + } +#endif + workq_kern_threadreq_modify(kq->kq_p, kqr, qos, flags); +} + +/* + * kqueue_threadreq_bind - bind thread to processing kqrequest + * + * 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 req, thread_t thread, + unsigned int flags) +{ + struct kqrequest *kqr = __container_of(req, struct kqrequest, kqr_req); + kqueue_t kqu = kqr_kqueue(p, kqr); + struct uthread *ut = get_bsdthread_info(thread); + + kq_req_held(kqu); + + assert(kqr->kqr_state & KQR_THREQUESTED); + assert(kqr->kqr_thread == THREAD_NULL); + assert(ut->uu_kqueue_override == 0); + + if (kqr->kqr_req.tr_state == TR_STATE_BINDING) { + assert(ut->uu_kqr_bound == kqr); + assert(kqr->kqr_req.tr_binding_thread == thread); + kqr->kqr_req.tr_state = TR_STATE_IDLE; + kqr->kqr_req.tr_binding_thread = NULL; + } else { + assert(ut->uu_kqr_bound == NULL); + } + + ut->uu_kqr_bound = kqr; + kqr->kqr_thread = thread; + + if (kqu.kq->kq_state & KQ_WORKLOOP) { + struct turnstile *ts = kqu.kqwl->kqwl_turnstile; + + 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_owner_override(kqu.kqwl)) { + thread_drop_ipc_override(thread); + } + thread_ends_owning_workloop(thread); + } + + 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); + } + + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_BIND), kqu.kqwl->kqwl_dynamicid, + thread_tid(thread), kqr->kqr_qos_index, + (kqr->kqr_override_index << 16) | kqr->kqr_state); + + ut->uu_kqueue_override = kqr->kqr_override_index; + if (kqr->kqr_override_index) { + thread_add_ipc_override(thread, kqr->kqr_override_index); + } + } else { + assert(kqr->kqr_override_index == 0); + + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWQ_BIND), -1, + thread_tid(thread), kqr->kqr_qos_index, + (kqr->kqr_override_index << 16) | kqr->kqr_state); + } +} + +/* + * kqueue_threadreq_cancel - abort a pending thread request + * + * Called when exiting/exec'ing. Forget our pending request. + */ +void +kqueue_threadreq_cancel(struct proc *p, workq_threadreq_t req) +{ + struct kqrequest *kqr = __container_of(req, struct kqrequest, kqr_req); + kqueue_t kqu = kqr_kqueue(p, kqr); + + kq_req_lock(kqu); + + assert(kqr->kqr_thread == THREAD_NULL); + assert(kqr->kqr_state & KQR_THREQUESTED); + kqr->kqr_state &= ~(KQR_THREQUESTED | KQR_R2K_NOTIF_ARMED); + + kq_req_unlock(kqu); + + kqueue_release_last(p, kqu); /* may dealloc kqu */ +} + +workq_threadreq_param_t +kqueue_threadreq_workloop_param(workq_threadreq_t req) +{ + struct kqrequest *kqr = __container_of(req, struct kqrequest, kqr_req); + struct kqworkloop *kqwl; + workq_threadreq_param_t trp; + + assert(kqr->kqr_state & KQR_WORKLOOP); + kqwl = __container_of(kqr, struct kqworkloop, kqwl_request); + trp.trp_value = kqwl->kqwl_params; + return trp; +} + +/* + * 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, struct kqrequest *kqr) +{ + if (kqr->kqr_state & KQR_WORKLOOP) { + kqworkloop_unbind(p, kqr_kqworkloop(kqr)); + } else { + kqworkq_unbind(p, kqr); + } +} + +/* + * 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) +{ + struct kqrequest *kqr; + + /* convert to thread qos value */ + assert(qos_index < KQWQ_NBUCKETS); + + kq_req_lock(kqwq); + kqr = kqworkq_get_request(kqwq, qos_index); + + if ((kqr->kqr_state & KQR_WAKEUP) == 0) { + kqr->kqr_state |= KQR_WAKEUP; + if ((kqr->kqr_state & KQR_THREQUESTED) == 0) { + kqueue_threadreq_initiate(&kqwq->kqwq_kqueue, kqr, qos_index, 0); + } + } + kq_req_unlock(kqwq); +} + +static kq_index_t +kqworkloop_owner_override(struct kqworkloop *kqwl) +{ + struct kqrequest *kqr = &kqwl->kqwl_request; + return MAX(kqr->kqr_qos_index, kqr->kqr_override_index); +} + +static inline void +kqworkloop_request_fire_r2k_notification(struct kqworkloop *kqwl) +{ + struct kqrequest *kqr = &kqwl->kqwl_request; + + kq_req_held(kqwl); + + if (kqr->kqr_state & KQR_R2K_NOTIF_ARMED) { + assert(kqr->kqr_thread); + kqr->kqr_state &= ~KQR_R2K_NOTIF_ARMED; + act_set_astkevent(kqr->kqr_thread, AST_KEVENT_RETURN_TO_KERNEL); + } +} + +static void +kqworkloop_update_threads_qos(struct kqworkloop *kqwl, int op, kq_index_t qos) +{ + struct kqrequest *kqr = &kqwl->kqwl_request; + struct kqueue *kq = &kqwl->kqwl_kqueue; + kq_index_t old_owner_override = kqworkloop_owner_override(kqwl); + kq_index_t i; + + /* must hold the kqr lock */ + kq_req_held(kqwl); + + 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); + } + if (kqr->kqr_wakeup_indexes & (1 << qos)) { + assert(kqr->kqr_state & KQR_WAKEUP); + break; + } + + kqr->kqr_wakeup_indexes |= (1 << qos); + kqr->kqr_state |= KQR_WAKEUP; + kqworkloop_request_fire_r2k_notification(kqwl); + goto recompute; + + 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; + } + } + break; + + case KQWL_UTQ_PARKING: + case KQWL_UTQ_UNBINDING: + kqr->kqr_override_index = qos; + /* FALLTHROUGH */ + case KQWL_UTQ_RECOMPUTE_WAKEUP_QOS: + if (op == KQWL_UTQ_RECOMPUTE_WAKEUP_QOS) { + assert(qos == THREAD_QOS_UNSPECIFIED); + } + kqlock_held(kqwl); // to look at kq_queues + i = KQWL_BUCKET_STAYACTIVE; + if (TAILQ_EMPTY(&kqr->kqr_suppressed)) { + kqr->kqr_override_index = THREAD_QOS_UNSPECIFIED; + } + if (!TAILQ_EMPTY(&kqwl->kqwl_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(&kqwl->kqwl_queue[i])) { + kqr->kqr_wakeup_indexes |= (1 << i); + } + } + if (kqr->kqr_wakeup_indexes) { + kqr->kqr_state |= KQR_WAKEUP; + kqworkloop_request_fire_r2k_notification(kqwl); + } else { + kqr->kqr_state &= ~KQR_WAKEUP; + } + goto recompute; + + case KQWL_UTQ_RESET_WAKEUP_OVERRIDE: + kqr->kqr_override_index = qos; + goto recompute; + + 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 (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; + + case KQWL_UTQ_REDRIVE_EVENTS: + break; + + case KQWL_UTQ_SET_QOS_INDEX: + kqr->kqr_qos_index = qos; + break; + + default: + panic("unknown kqwl thread qos update operation: %d", op); + } + + thread_t kqwl_owner = kqwl->kqwl_owner; + thread_t servicer = kqr->kqr_thread; + boolean_t qos_changed = FALSE; + kq_index_t new_owner_override = kqworkloop_owner_override(kqwl); + + /* + * Apply the diffs to the owner if applicable + */ + if (kqwl_owner) { +#if 0 + /* JMM - need new trace hooks for owner overrides */ + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_THADJUST), + kqwl->kqwl_dynamicid, thread_tid(kqwl_owner), kqr->kqr_qos_index, + (kqr->kqr_override_index << 16) | kqr->kqr_state); +#endif + if (new_owner_override == old_owner_override) { + // nothing to do + } else if (old_owner_override == THREAD_QOS_UNSPECIFIED) { + thread_add_ipc_override(kqwl_owner, new_owner_override); + } else if (new_owner_override == THREAD_QOS_UNSPECIFIED) { + thread_drop_ipc_override(kqwl_owner); + } else /* old_owner_override != new_owner_override */ { + thread_update_ipc_override(kqwl_owner, new_owner_override); + } + } + + /* + * apply the diffs to the servicer + */ + if ((kqr->kqr_state & KQR_THREQUESTED) == 0) { + /* + * 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. + */ + + if (kqwl_owner == NULL && (kqr->kqr_state & KQR_WAKEUP)) { + int initiate_flags = 0; + if (op == KQWL_UTQ_UNBINDING) { + initiate_flags = WORKQ_THREADREQ_ATTEMPT_REBIND; + } + kqueue_threadreq_initiate(kq, kqr, new_owner_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 != kqr->kqr_override_index) { + if (ut->uu_kqueue_override == THREAD_QOS_UNSPECIFIED) { + thread_add_ipc_override(servicer, kqr->kqr_override_index); + } else if (kqr->kqr_override_index == THREAD_QOS_UNSPECIFIED) { + thread_drop_ipc_override(servicer); + } else /* ut->uu_kqueue_override != kqr->kqr_override_index */ { + thread_update_ipc_override(servicer, kqr->kqr_override_index); + } + ut->uu_kqueue_override = kqr->kqr_override_index; + qos_changed = TRUE; + } + } else if (new_owner_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_owner_override != new_owner_override) { + /* + * Request is in flight + * + * Apply the diff to the thread request + */ + kqueue_threadreq_modify(kq, kqr, new_owner_override); + qos_changed = TRUE; + } + + if (qos_changed) { + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_THADJUST), kqwl->kqwl_dynamicid, + thread_tid(kqr->kqr_thread), kqr->kqr_qos_index, + (kqr->kqr_override_index << 16) | kqr->kqr_state); + } +} + +static void +kqworkloop_request_help(struct kqworkloop *kqwl, kq_index_t qos_index) +{ + /* convert to thread qos value */ + assert(qos_index < KQWL_NBUCKETS); + + kq_req_lock(kqwl); + kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_UPDATE_WAKEUP_QOS, qos_index); + kq_req_unlock(kqwl); +} + +static struct kqtailq * +kqueue_get_queue(struct kqueue *kq, kq_index_t qos_index) +{ + if (kq->kq_state & KQ_WORKQ) { + assert(qos_index < KQWQ_NBUCKETS); + } else if (kq->kq_state & KQ_WORKLOOP) { + assert(qos_index < KQWL_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 &kq->kq_queue[qos_index]; +} + +static int +kqueue_queue_empty(struct kqueue *kq, kq_index_t qos_index) +{ + return TAILQ_EMPTY(kqueue_get_queue(kq, qos_index)); +} + +static struct kqtailq * +kqueue_get_suppressed_queue(kqueue_t kq, struct knote *kn) +{ + if (kq.kq->kq_state & KQ_WORKQ) { + return &kqworkq_get_request(kq.kqwq, kn->kn_qos_index)->kqr_suppressed; + } else if (kq.kq->kq_state & KQ_WORKLOOP) { + return &kq.kqwl->kqwl_request.kqr_suppressed; + } else { + return &kq.kqf->kqf_suppressed; + } +} + +static struct turnstile * +kqueue_get_turnstile(kqueue_t kqu, bool can_alloc) +{ + uint8_t kqr_state; + + if ((kqu.kq->kq_state & KQ_WORKLOOP) == 0) { + return TURNSTILE_NULL; + } + + kqr_state = os_atomic_load(&kqu.kqwl->kqwl_request.kqr_state, relaxed); + if (kqr_state & KQR_ALLOCATED_TURNSTILE) { + /* force a dependency to pair with the atomic or with release below */ + return os_atomic_load_with_dependency_on(&kqu.kqwl->kqwl_turnstile, + kqr_state); + } + + if (!can_alloc) { + return TURNSTILE_NULL; + } + + struct turnstile *ts = turnstile_alloc(), *free_ts = TURNSTILE_NULL; + + kq_req_lock(kqu); + if (filt_wlturnstile_interlock_is_workq(kqu.kqwl)) { + workq_kern_threadreq_lock(kqu.kqwl->kqwl_p); + } + + if (kqu.kqwl->kqwl_request.kqr_state & KQR_ALLOCATED_TURNSTILE) { + free_ts = ts; + ts = kqu.kqwl->kqwl_turnstile; + } else { + ts = turnstile_prepare((uintptr_t)kqu.kqwl, &kqu.kqwl->kqwl_turnstile, + ts, TURNSTILE_WORKLOOPS); + + /* release-barrier to pair with the unlocked load of kqwl_turnstile above */ + os_atomic_or(&kqu.kqwl->kqwl_request.kqr_state, + KQR_ALLOCATED_TURNSTILE, release); + } + + if (filt_wlturnstile_interlock_is_workq(kqu.kqwl)) { + workq_kern_threadreq_unlock(kqu.kqwl->kqwl_p); + } + kq_req_unlock(kqu.kqwl); + + if (free_ts) { + turnstile_deallocate(free_ts); + } + return ts; +} + +struct turnstile * +kqueue_turnstile(struct kqueue *kq) +{ + return kqueue_get_turnstile(kq, false); +} + +struct turnstile * +kqueue_alloc_turnstile(struct kqueue *kq) +{ + return kqueue_get_turnstile(kq, true); +} + +static struct kqtailq * +knote_get_queue(struct knote *kn) +{ + return kqueue_get_queue(knote_get_kq(kn), kn->kn_qos_index); +} + +static void +knote_reset_priority(struct knote *kn, pthread_priority_t pp) +{ + struct kqueue *kq = knote_get_kq(kn); + kq_index_t qos = _pthread_priority_thread_qos(pp); + + assert((kn->kn_status & KN_QUEUED) == 0); + + if (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 { + pp = _pthread_priority_normalize(pp); + } + } else if (kq->kq_state & KQ_WORKLOOP) { + assert((pp & _PTHREAD_PRIORITY_EVENT_MANAGER_FLAG) == 0); + pp = _pthread_priority_normalize(pp); + } else { + pp = _pthread_unspecified_priority(); + qos = THREAD_QOS_UNSPECIFIED; + } + + kn->kn_qos = pp; + kn->kn_req_index = qos; + + 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; + } + + /* only adjust in-use qos index when not suppressed */ + if ((kn->kn_status & KN_SUPPRESSED) == 0) { + kn->kn_qos_index = qos; + } else if (kq->kq_state & KQ_WORKQ) { + kqworkq_update_override((struct kqworkq *)kq, kn, qos); + } else if (kq->kq_state & KQ_WORKLOOP) { + kqworkloop_update_override((struct kqworkloop *)kq, qos); + } +} + +static void +knote_set_qos_overcommit(struct knote *kn) +{ + struct kqueue *kq = knote_get_kq(kn); + + /* turn overcommit on for the appropriate thread request? */ + if ((kn->kn_qos & _PTHREAD_PRIORITY_OVERCOMMIT_FLAG) && + (kq->kq_state & KQ_WORKLOOP)) { + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + struct kqrequest *kqr = &kqwl->kqwl_request; + + /* + * This test is racy, but since we never remove this bit, + * it allows us to avoid taking a lock. + */ + if (kqr->kqr_state & KQR_THOVERCOMMIT) { + return; + } + + kq_req_lock(kqwl); + kqr->kqr_state |= KQR_THOVERCOMMIT; + if (!kqr->kqr_thread && (kqr->kqr_state & KQR_THREQUESTED)) { + kqueue_threadreq_modify(kq, kqr, kqr->kqr_req.tr_qos); + } + kq_req_unlock(kqwl); + } +} + +static kq_index_t +knote_get_qos_override_index(struct knote *kn) +{ + return kn->kn_qos_override; +} + +static void +kqworkq_update_override(struct kqworkq *kqwq, struct knote *kn, + kq_index_t override_index) +{ + struct kqrequest *kqr; + kq_index_t old_override_index; + kq_index_t queue_index = kn->kn_qos_index; + + if (override_index <= queue_index) { + return; + } + + kqr = kqworkq_get_request(kqwq, queue_index); + + kq_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_thread) { + if (old_override_index) + thread_update_ipc_override(kqr->kqr_thread, override_index); + else + thread_add_ipc_override(kqr->kqr_thread, override_index); + } + } + kq_req_unlock(kqwq); +} + +static void +kqworkloop_update_override(struct kqworkloop *kqwl, kq_index_t override_index) +{ + kq_req_lock(kqwl); + kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_UPDATE_WAKEUP_OVERRIDE, + override_index); + kq_req_unlock(kqwl); +} + +static thread_qos_t +kqworkloop_unbind_locked(struct kqworkloop *kqwl, thread_t thread) +{ + struct uthread *ut = get_bsdthread_info(thread); + struct kqrequest *kqr = &kqwl->kqwl_request; + kq_index_t ipc_override = ut->uu_kqueue_override; + + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWL_UNBIND), kqwl->kqwl_dynamicid, + thread_tid(thread), 0, 0); + + kq_req_held(kqwl); + assert(ut->uu_kqr_bound == kqr); + ut->uu_kqr_bound = NULL; + ut->uu_kqueue_override = THREAD_QOS_UNSPECIFIED; + + 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->kqr_thread = NULL; + kqr->kqr_state &= ~(KQR_THREQUESTED | KQR_R2K_NOTIF_ARMED); + return ipc_override; +} + +/* + * 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(proc_t p, struct kqworkloop *kqwl) +{ + struct kqueue *kq = &kqwl->kqwl_kqueue; + struct kqrequest *kqr = &kqwl->kqwl_request; + thread_t thread = kqr->kqr_thread; + int op = KQWL_UTQ_PARKING; + kq_index_t ipc_override, qos_override = THREAD_QOS_UNSPECIFIED; + + assert(thread == current_thread()); + + kqlock(kqwl); + + /* + * 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(&kqr->kqr_suppressed)) { + kq->kq_state |= KQ_PROCESSING; + qos_override = kqworkloop_acknowledge_events(kqwl); + kq->kq_state &= ~KQ_PROCESSING; + } + + kq_req_lock(kqwl); + + ipc_override = kqworkloop_unbind_locked(kqwl, thread); + kqworkloop_update_threads_qos(kqwl, op, qos_override); + + kq_req_unlock(kqwl); + + kqunlock(kqwl); + + /* + * Drop the override on the current thread last, after the call to + * kqworkloop_update_threads_qos above. + */ + if (ipc_override) { + thread_drop_ipc_override(thread); + } + + /* If last reference, dealloc the workloop kq */ + kqueue_release_last(p, kqwl); +} + +static thread_qos_t +kqworkq_unbind_locked(__assert_only struct kqworkq *kqwq, + struct kqrequest *kqr, thread_t thread) +{ + struct uthread *ut = get_bsdthread_info(thread); + kq_index_t old_override = kqr->kqr_override_index; + + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KQWQ_UNBIND), -1, + thread_tid(kqr->kqr_thread), kqr->kqr_qos_index, 0); + + kq_req_held(kqwq); + assert(ut->uu_kqr_bound == kqr); + ut->uu_kqr_bound = NULL; + kqr->kqr_thread = NULL; + kqr->kqr_state &= ~(KQR_THREQUESTED | KQR_R2K_NOTIF_ARMED); + kqr->kqr_override_index = THREAD_QOS_UNSPECIFIED; + + return old_override; +} + +/* + * 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, struct kqrequest *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); +} + +struct kqrequest * +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_apply_qos_override(struct knote *kn, kq_index_t qos_index) +{ + assert((kn->kn_status & KN_QUEUED) == 0); + + kn->kn_qos_override = qos_index; + + if (kn->kn_status & KN_SUPPRESSED) { + struct kqueue *kq = knote_get_kq(kn); + /* + * 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. + * + * 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. + */ + if (kq->kq_state & KQ_WORKQ) { + kqworkq_update_override((struct kqworkq *)kq, kn, qos_index); + } else { + kqworkloop_update_override((struct kqworkloop *)kq, qos_index); + } + } else { + kn->kn_qos_index = qos_index; + } +} + +static bool +knote_should_apply_qos_override(struct kqueue *kq, struct knote *kn, int result, + thread_qos_t *qos_out) +{ + thread_qos_t qos_index = (result >> FILTER_ADJUST_EVENT_QOS_SHIFT) & 7; + + kqlock_held(kq); + + assert(result & FILTER_ADJUST_EVENT_QOS_BIT); + assert(qos_index < THREAD_QOS_LAST); + + /* + * Early exit for knotes that should not change QoS + * + * It is safe to test kn_req_index against MANAGER / STAYACTIVE because + * knotes with such kn_req_index values never change for their entire + * lifetime. + */ + if (__improbable(!knote_fops(kn)->f_adjusts_qos)) { + panic("filter %d cannot change QoS", kn->kn_filtid); + } else if (kq->kq_state & KQ_WORKLOOP) { + if (kn->kn_req_index == KQWL_BUCKET_STAYACTIVE) { + return false; + } + } else if (kq->kq_state & KQ_WORKQ) { + if (kn->kn_req_index == KQWQ_QOS_MANAGER) { + return false; + } + } else { + return false; + } + + /* + * 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 (kn->kn_qos & _PTHREAD_PRIORITY_FALLBACK_FLAG) { + if (qos_index == THREAD_QOS_UNSPECIFIED) + qos_index = kn->kn_req_index; + } else { + if (qos_index < kn->kn_req_index) + qos_index = kn->kn_req_index; + } + if ((kn->kn_status & KN_MERGE_QOS) && (qos_index < kn->kn_qos_override)) { + /* Never lower QoS when in "Merge" mode */ + return false; + } + + if ((kn->kn_status & KN_LOCKED) && kn->kn_inuse) { + /* + * 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. + * + * This mode is left when knote_unlock() or knote_call_filter_event() + * observe that no other f_* routine is in flight. + */ + kn->kn_status |= KN_MERGE_QOS; + } + + if (kn->kn_qos_override == qos_index) { + return false; + } + + *qos_out = qos_index; + return true; +} + +static void +knote_adjust_qos(struct kqueue *kq, struct knote *kn, int result) +{ + thread_qos_t qos; + if (knote_should_apply_qos_override(kq, kn, result, &qos)) { + knote_dequeue(kn); + knote_apply_qos_override(kn, qos); + if (knote_enqueue(kn) && (kn->kn_status & KN_ACTIVE)) { + knote_wakeup(kn); + } + } +} + +static void +knote_wakeup(struct knote *kn) +{ + struct kqueue *kq = knote_get_kq(kn); + + kqlock_held(kq); + + if (kq->kq_state & KQ_WORKQ) { + struct kqworkq *kqwq = (struct kqworkq *)kq; + + kqworkq_request_help(kqwq, kn->kn_qos_index); + } else if (kq->kq_state & KQ_WORKLOOP) { + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + + /* + * kqworkloop_end_processing() will perform the required QoS + * computations when it unsets the processing mode. + */ + if (!kqworkloop_is_processing_on_current_thread(kqwl)) { + kqworkloop_request_help(kqwl, kn->kn_qos_index); + } + } else { + struct kqfile *kqf = (struct kqfile *)kq; + + /* flag wakeups during processing */ + if (kq->kq_state & KQ_PROCESSING) + kq->kq_state |= KQ_WAKEUP; + + /* 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); + } + + /* wakeup other kqueues/select sets we're inside */ + KNOTE(&kqf->kqf_sel.si_note, 0); + } +} + +/* + * Called with the kqueue locked + */ +static void +kqueue_interrupt(struct kqueue *kq) +{ + assert((kq->kq_state & KQ_WORKQ) == 0); + + /* 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); + } + + /* wakeup threads waiting their turn to process */ + if (kq->kq_state & KQ_PROCWAIT) { + struct kqtailq *suppressq; + + assert(kq->kq_state & KQ_PROCESSING); + + kq->kq_state &= ~KQ_PROCWAIT; + suppressq = kqueue_get_suppressed_queue(kq, NULL); + (void)waitq_wakeup64_all((struct waitq *)&kq->kq_wqs, + CAST_EVENT64_T(suppressq), + THREAD_RESTART, + WAITQ_ALL_PRIORITIES); + } +} + +/* + * 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. + */ +void +waitq_set__CALLING_PREPOST_HOOK__(void *kq_hook, void *knote_hook, int qos) +{ +#pragma unused(knote_hook, qos) + + struct kqueue *kq = (struct kqueue *)kq_hook; + + if (kq->kq_state & KQ_WORKQ) { + struct kqworkq *kqwq = (struct kqworkq *)kq; + + kqworkq_request_help(kqwq, KQWQ_QOS_MANAGER); + } else if (kq->kq_state & KQ_WORKLOOP) { + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + + kqworkloop_request_help(kqwl, KQWL_BUCKET_STAYACTIVE); + } +} + +void +klist_init(struct klist *list) +{ + SLIST_INIT(list); +} + + +/* + * 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. + */ +void +knote(struct klist *list, long hint) +{ + struct knote *kn; + + SLIST_FOREACH(kn, list, kn_selnext) { + struct kqueue *kq = knote_get_kq(kn); + kqlock(kq); + knote_call_filter_event(kq, kn, hint); + kqunlock(kq); + } +} + +/* + * 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) +{ + int ret = SLIST_EMPTY(list); + SLIST_INSERT_HEAD(list, kn, kn_selnext); + return (ret); +} + +/* + * 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) +{ + SLIST_REMOVE(list, kn, knote, kn_selnext); + return (SLIST_EMPTY(list)); +} + +/* + * 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) +{ + struct knote *kn; + struct knote *kn_next; + + SLIST_FOREACH_SAFE(kn, list, kn_selnext, kn_next) { + struct kqueue *kq = knote_get_kq(kn); + + kqlock(kq); + if (kn->kn_status & KN_REQVANISH) { + /* If EV_VANISH supported - prepare to deliver one */ + kn->kn_status |= KN_VANISHED; + knote_activate(kn); + } else { + knote_call_filter_event(kq, kn, NOTE_REVOKE); + } + kqunlock(kq); + } +} + +/* + * 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) +{ + struct kqueue *kq = knote_get_kq(kn); + waitq_set_lazy_init_link(&kq->kq_wqs); +} + +/* + * 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); + return waitq_set_should_lazy_init_link(&kq->kq_wqs); +} + +/* + * 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); + kern_return_t kr; + + kr = waitq_link(wq, &kq->kq_wqs, WAITQ_ALREADY_LOCKED, reserved_link); + if (kr == KERN_SUCCESS) { + knote_markstayactive(kn); + return (0); + } else { + return (EINVAL); + } +} + +/* + * 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); + kern_return_t kr; + + kr = waitq_unlink(wq, &kq->kq_wqs); + knote_clearstayactive(kn); + return ((kr != KERN_SUCCESS) ? EINVAL : 0); +} + +/* + * 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); + +restart: + list = &p->p_fd->fd_knlist[fd]; + SLIST_FOREACH(kn, list, kn_link) { + struct kqueue *kq = knote_get_kq(kn); + + 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 (kn->kn_status & KN_VANISHED) { + kqunlock(kq); + continue; + } + + 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; + kn->kn_status &= ~KN_ATTACHED; + + kqunlock(kq); + knote_fops(kn)->f_detach(kn); + if (knote_fops(kn)->f_isfd) + fp_drop(p, kn->kn_id, kn->kn_fp, 0); + kqlock(kq); + + knote_activate(kn); + knote_unlock(kq, kn, &knlc, KNOTE_KQ_UNLOCK); + } else { + knote_drop(kq, kn, &knlc); + } + + proc_fdlock(p); + goto restart; + } +} + +/* + * 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, + struct kevent_internal_s *kev, + bool is_fd, + struct proc *p) +{ + struct filedesc *fdp = p->p_fd; + struct klist *list = NULL; + struct knote *kn = NULL; + + /* + * 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]; + } + } 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 */ + } + } + } + } + return kn; +} + +/* + * 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) +{ + 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, &kn->kn_kevent, is_fd, p) != NULL) { + /* found an existing knote: we can't add this one */ + ret = ERESTART; + goto out_locked; + } + + /* knote was not found: add it now */ + if (!is_fd) { + if (fdp->fd_knhashmask == 0) { + u_long size = 0; + + list = hashinit(CONFIG_KN_HASHSIZE, M_KQUEUE, &size); + if (list == NULL) { + ret = ENOMEM; + goto out_locked; + } + + fdp->fd_knhash = list; + fdp->fd_knhashmask = size; + } + + 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 { + /* 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; + } + + 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; + } + + list = &fdp->fd_knlist[kn->kn_id]; + SLIST_INSERT_HEAD(list, kn, kn_link); + ret = 0; + goto out_locked; + + } + +out_locked: + if (ret == 0) { + kqlock(kq); + assert((kn->kn_status & KN_LOCKED) == 0); + (void)knote_lock(kq, kn, knlc, KNOTE_KQ_UNLOCK); + } + if (is_fd) + proc_fdunlock(p); + else + knhash_unlock(p); + + return ret; +} + +/* + * kq_remove_knote - remove a knote from the fd table for process + * + * If the filter is file-based, remove based on fd index. + * Otherwise remove from the hash based on the ident. + * + * fd_knhashlock and fdlock unheld on entry (and exit). + */ +static void +kq_remove_knote(struct kqueue *kq, struct knote *kn, struct proc *p, + struct knote_lock_ctx *knlc) +{ + struct filedesc *fdp = p->p_fd; + struct klist *list = NULL; + uint16_t kq_state; + bool is_fd; + + is_fd = knote_fops(kn)->f_isfd; + + if (is_fd) + proc_fdlock(p); + else + knhash_lock(p); + + 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); + + kqlock(kq); + kq_state = kq->kq_state; + if (knlc) { + knote_unlock_cancel(kq, kn, knlc, KNOTE_KQ_UNLOCK); + } else { + kqunlock(kq); + } + if (is_fd) + proc_fdunlock(p); + else + knhash_unlock(p); + + if (kq_state & KQ_DYNAMIC) + kqueue_release_last(p, kq); +} + +/* + * 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) + */ + +static struct knote * +kq_find_knote_and_kq_lock(struct kqueue *kq, struct kevent_internal_s *kev, + bool is_fd, struct proc *p) +{ + struct knote * ret; + + if (is_fd) + proc_fdlock(p); + else + knhash_lock(p); + + ret = knote_fdfind(kq, kev, is_fd, p); + + if (ret) { + kqlock(kq); + } + + if (is_fd) + proc_fdunlock(p); + else + knhash_unlock(p); + + return ret; +} +/* + * knote_drop - disconnect and drop the knote + * + * Called with the kqueue locked, returns with the kqueue unlocked. + * + * 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. + */ +static void +knote_drop(struct kqueue *kq, struct knote *kn, struct knote_lock_ctx *knlc) +{ + struct proc *p = kq->kq_p; + + kqlock_held(kq); + + assert((kn->kn_status & KN_DROPPING) == 0); + if (knlc == NULL) { + assert((kn->kn_status & KN_LOCKED) == 0); + } + kn->kn_status |= KN_DROPPING; + + knote_unsuppress(kn); + knote_dequeue(kn); + knote_wait_for_filter_events(kq, kn); + + /* If we are attached, disconnect from the source first */ + if (kn->kn_status & KN_ATTACHED) { + knote_fops(kn)->f_detach(kn); + } + + /* kq may be freed when kq_remove_knote() returns */ + kq_remove_knote(kq, kn, p, knlc); + if (knote_fops(kn)->f_isfd && ((kn->kn_status & KN_VANISHED) == 0)) + fp_drop(p, kn->kn_id, kn->kn_fp, 0); + + knote_free(kn); +} + +/* called with kqueue lock held */ +static void +knote_activate(struct knote *kn) +{ + if (kn->kn_status & KN_ACTIVE) + return; + + KDBG_FILTERED(KEV_EVTID(BSD_KEVENT_KNOTE_ACTIVATE), + kn->kn_udata, kn->kn_status | (kn->kn_id << 32), + kn->kn_filtid); + + kn->kn_status |= KN_ACTIVE; + if (knote_enqueue(kn)) + knote_wakeup(kn); +} + +/* 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); +} + +/* called with kqueue lock held */ +static void +knote_enable(struct knote *kn) +{ + if ((kn->kn_status & KN_DISABLED) == 0) + return; + + kn->kn_status &= ~KN_DISABLED; + + if (kn->kn_status & KN_SUPPRESSED) { + /* + * 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. + */ + struct kqueue *kq = knote_get_kq(kn); + if ((kq->kq_state & KQ_PROCESSING) == 0) { + knote_unsuppress(kn); + } + } else if (knote_enqueue(kn)) { + knote_wakeup(kn); + } +} + +/* called with kqueue lock held */ +static void +knote_disable(struct knote *kn) +{ + if (kn->kn_status & KN_DISABLED) + return; + + kn->kn_status |= KN_DISABLED; + knote_dequeue(kn); +} + +/* called with kqueue lock held */ +static void +knote_suppress(struct knote *kn) +{ + struct kqtailq *suppressq; + struct kqueue *kq = knote_get_kq(kn); + + kqlock_held(kq); + + if (kn->kn_status & KN_SUPPRESSED) + return; + + knote_dequeue(kn); + kn->kn_status |= KN_SUPPRESSED; + suppressq = kqueue_get_suppressed_queue(kq, kn); + TAILQ_INSERT_TAIL(suppressq, kn, kn_tqe); +} + +/* called with kqueue lock held */ +static void +knote_unsuppress(struct knote *kn) +{ + struct kqtailq *suppressq; + struct kqueue *kq = knote_get_kq(kn); + + kqlock_held(kq); + + if ((kn->kn_status & KN_SUPPRESSED) == 0) + return; + + kn->kn_status &= ~KN_SUPPRESSED; + suppressq = kqueue_get_suppressed_queue(kq, kn); + TAILQ_REMOVE(suppressq, kn, kn_tqe); + + /* + * If the knote is no longer active, reset its push, + * and resynchronize kn_qos_index with kn_qos_override + */ + if ((kn->kn_status & KN_ACTIVE) == 0) { + kn->kn_qos_override = kn->kn_req_index; + } + kn->kn_qos_index = kn->kn_qos_override; + + /* don't wakeup if unsuppressing just a stay-active knote */ + if (knote_enqueue(kn) && (kn->kn_status & KN_ACTIVE)) { + knote_wakeup(kn); + } + + if ((kq->kq_state & KQ_WORKLOOP) && TAILQ_EMPTY(suppressq)) { + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + + if (kqworkloop_is_processing_on_current_thread(kqwl)) { + /* + * kqworkloop_end_processing() or kqworkloop_begin_processing() + * will perform the required QoS computations when it unsets the + * processing mode. + */ + } else { + kq_req_lock(kqwl); + kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_RESET_WAKEUP_OVERRIDE, 0); + kq_req_unlock(kqwl); + } + } +} + +/* called with kqueue lock held */ +static int +knote_enqueue(struct knote *kn) +{ + 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); + + kqlock_held(kq); + TAILQ_INSERT_TAIL(queue, kn, kn_tqe); + kn->kn_status |= KN_QUEUED; + kq->kq_count++; + return 1; + } + return ((kn->kn_status & KN_STAYACTIVE) != 0); +} + + +/* called with kqueue lock held */ +static void +knote_dequeue(struct knote *kn) +{ + struct kqueue *kq = knote_get_kq(kn); + struct kqtailq *queue; + + kqlock_held(kq); + + if ((kn->kn_status & KN_QUEUED) == 0) + return; + + queue = knote_get_queue(kn); + TAILQ_REMOVE(queue, kn, kn_tqe); + kn->kn_status &= ~KN_QUEUED; + kq->kq_count--; +} + +void +knote_init(void) +{ + knote_zone = zinit(sizeof(struct knote), 8192*sizeof(struct knote), + 8192, "knote zone"); + + kqfile_zone = zinit(sizeof(struct kqfile), 8192*sizeof(struct kqfile), + 8192, "kqueue file zone"); + + kqworkq_zone = zinit(sizeof(struct kqworkq), 8192*sizeof(struct kqworkq), + 8192, "kqueue workq zone"); + + kqworkloop_zone = zinit(sizeof(struct kqworkloop), 8192*sizeof(struct kqworkloop), + 8192, "kqueue workloop zone"); + + /* allocate kq lock group attribute and group */ + kq_lck_grp_attr = lck_grp_attr_alloc_init(); + + kq_lck_grp = lck_grp_alloc_init("kqueue", kq_lck_grp_attr); + + /* Allocate kq lock attribute */ + kq_lck_attr = lck_attr_alloc_init(); + +#if CONFIG_MEMORYSTATUS + /* Initialize the memorystatus list lock */ + memorystatus_kevent_init(kq_lck_grp, kq_lck_attr); +#endif +} +SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL) + +const struct filterops * +knote_fops(struct knote *kn) +{ + return sysfilt_ops[kn->kn_filtid]; +} + +static struct knote * +knote_alloc(void) +{ + struct knote *kn = ((struct knote *)zalloc(knote_zone)); + bzero(kn, sizeof(struct knote)); + return kn; +} + +static void +knote_free(struct knote *kn) +{ + assert(kn->kn_inuse == 0); + assert((kn->kn_status & KN_LOCKED) == 0); + zfree(knote_zone, kn); +} + +#if SOCKETS +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#ifndef ROUNDUP64 +#define ROUNDUP64(x) P2ROUNDUP((x), sizeof (u_int64_t)) +#endif + +#ifndef ADVANCE64 +#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 int kev_attach(struct socket *so, int proto, struct proc *p); +static int kev_detach(struct socket *so); +static int kev_control(struct socket *so, u_long cmd, caddr_t data, + struct ifnet *ifp, struct proc *p); +static lck_mtx_t * event_getlock(struct socket *, int); +static int event_lock(struct socket *, int, void *); +static int event_unlock(struct socket *, int, void *); + +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, +}; + +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, +} +}; + +__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"); + +struct kevtstat kevtstat; +SYSCTL_PROC(_net_systm_kevt, OID_AUTO, stats, + CTLTYPE_STRUCT | CTLFLAG_RD | CTLFLAG_LOCKED, 0, 0, + kevt_getstat, "S,kevtstat", ""); + +SYSCTL_PROC(_net_systm_kevt, OID_AUTO, pcblist, + CTLTYPE_STRUCT | CTLFLAG_RD | CTLFLAG_LOCKED, 0, 0, + kevt_pcblist, "S,xkevtpcb", ""); + +static lck_mtx_t * +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) + panic("%s: so=%p usecount=%d lrh= %s\n", __func__, + so, so->so_usecount, solockhistory_nr(so)); + /* NOTREACHED */ + } else { + panic("%s: so=%p NULL NO so_pcb %s\n", __func__, + so, solockhistory_nr(so)); + /* NOTREACHED */ + } + return (&ev_pcb->evp_mtx); +} + +static int +event_lock(struct socket *so, int refcount, void *lr) +{ + void *lr_saved; + + if (lr == NULL) + lr_saved = __builtin_return_address(0); + else + lr_saved = lr; + + if (so->so_pcb != NULL) { + lck_mtx_lock(&((struct kern_event_pcb *)so->so_pcb)->evp_mtx); + } else { + panic("%s: so=%p NO PCB! lr=%p lrh= %s\n", __func__, + so, lr_saved, solockhistory_nr(so)); + /* NOTREACHED */ + } + + if (so->so_usecount < 0) { + panic("%s: so=%p so_pcb=%p lr=%p ref=%d lrh= %s\n", __func__, + so, so->so_pcb, lr_saved, so->so_usecount, + solockhistory_nr(so)); + /* NOTREACHED */ + } + + 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); +} + +static int +event_unlock(struct socket *so, int refcount, void *lr) +{ + void *lr_saved; + lck_mtx_t *mutex_held; + + if (lr == NULL) + lr_saved = __builtin_return_address(0); + else + lr_saved = lr; + + if (refcount) { + so->so_usecount--; + } + if (so->so_usecount < 0) { + panic("%s: so=%p usecount=%d lrh= %s\n", __func__, + so, so->so_usecount, solockhistory_nr(so)); + /* NOTREACHED */ + } + if (so->so_pcb == NULL) { + panic("%s: so=%p NO PCB usecount=%d lr=%p lrh= %s\n", __func__, + so, so->so_usecount, (void *)lr_saved, + solockhistory_nr(so)); + /* NOTREACHED */ + } + mutex_held = (&((struct kern_event_pcb *)so->so_pcb)->evp_mtx); + + 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; + + if (so->so_usecount == 0) { + VERIFY(so->so_flags & SOF_PCBCLEARING); + event_sofreelastref(so); + } else { + lck_mtx_unlock(mutex_held); + } + + return (0); +} + +static int +event_sofreelastref(struct socket *so) +{ + struct kern_event_pcb *ev_pcb = (struct kern_event_pcb *)so->so_pcb; + + LCK_MTX_ASSERT(&(ev_pcb->evp_mtx), LCK_MTX_ASSERT_OWNED); + + so->so_pcb = NULL; + + /* + * Disable upcall in the event another thread is in kev_post_msg() + * appending record to the receive socket buffer, since sbwakeup() + * may release the socket lock otherwise. + */ + so->so_rcv.sb_flags &= ~SB_UPCALL; + so->so_snd.sb_flags &= ~SB_UPCALL; + so->so_event = sonullevent; + lck_mtx_unlock(&(ev_pcb->evp_mtx)); + + LCK_MTX_ASSERT(&(ev_pcb->evp_mtx), LCK_MTX_ASSERT_NOTOWNED); + lck_rw_lock_exclusive(kev_rwlock); + LIST_REMOVE(ev_pcb, evp_link); + kevtstat.kes_pcbcount--; + kevtstat.kes_gencnt++; + lck_rw_done(kev_rwlock); + kev_delete(ev_pcb); + + sofreelastref(so, 1); + return (0); +} + +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; + +/* + * Install the protosw's for the NKE manager. Invoked at extension load time + */ +void +kern_event_init(struct domain *dp) +{ + struct protosw *pr; + int i; + + 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++) + 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 +kev_attach(struct socket *so, __unused int proto, __unused struct proc *p) +{ + int error = 0; + 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); + } + bzero(ev_pcb, sizeof(struct kern_event_pcb)); + lck_mtx_init(&ev_pcb->evp_mtx, kev_lck_grp, kev_lck_attr); + + 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); + LIST_INSERT_HEAD(&kern_event_head, ev_pcb, evp_link); + kevtstat.kes_pcbcount++; + kevtstat.kes_gencnt++; + lck_rw_done(kev_rwlock); + + 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); + zfree(ev_pcb_zone, ev_pcb); +} + +static int +kev_detach(struct socket *so) +{ + struct kern_event_pcb *ev_pcb = (struct kern_event_pcb *) so->so_pcb; + + if (ev_pcb != NULL) { + soisdisconnected(so); + so->so_flags |= SOF_PCBCLEARING; + } + + 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) +{ + if (strlen(string) >= KEV_VENDOR_CODE_MAX_STR_LEN) { + return (EINVAL); + } + return (net_str_id_find_internal(string, out_vendor_code, + NSI_VENDOR_CODE, 1)); +} + +errno_t +kev_msg_post(struct kev_msg *event_msg) +{ + mbuf_tag_id_t min_vendor, max_vendor; + + net_str_id_first_last(&min_vendor, &max_vendor, NSI_VENDOR_CODE); + + 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); + } + return (kev_post_msg(event_msg)); +} + +int +kev_post_msg(struct kev_msg *event_msg) +{ + struct mbuf *m, *m2; + struct kern_event_pcb *ev_pcb; + struct kern_event_msg *ev; + char *tmp; + u_int32_t total_size; + int i; + + /* Verify the message is small enough to fit in one mbuf w/o cluster */ + total_size = KEV_MSG_HEADER_SIZE; + + for (i = 0; i < 5; i++) { + 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); + } + + m = m_get(M_WAIT, MT_DATA); + if (m == 0) { + OSIncrementAtomic64((SInt64 *)&kevtstat.kes_nomem); + 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) + break; + + total_size += event_msg->dv[i].data_length; + bcopy(event_msg->dv[i].data_ptr, tmp, + event_msg->dv[i].data_length); + tmp += event_msg->dv[i].data_length; + } + + ev->id = ++static_event_id; + ev->total_size = total_size; + ev->vendor_code = event_msg->vendor_code; + ev->kev_class = event_msg->kev_class; + ev->kev_subclass = event_msg->kev_subclass; + ev->event_code = event_msg->event_code; + + m->m_len = total_size; + lck_rw_lock_shared(kev_rwlock); + for (ev_pcb = LIST_FIRST(&kern_event_head); + ev_pcb; + ev_pcb = LIST_NEXT(ev_pcb, evp_link)) { + lck_mtx_lock(&ev_pcb->evp_mtx); + if (ev_pcb->evp_socket->so_pcb == NULL) { + lck_mtx_unlock(&ev_pcb->evp_mtx); + continue; + } + if (ev_pcb->evp_vendor_code_filter != KEV_ANY_VENDOR) { + if (ev_pcb->evp_vendor_code_filter != ev->vendor_code) { + lck_mtx_unlock(&ev_pcb->evp_mtx); + continue; + } + + if (ev_pcb->evp_class_filter != KEV_ANY_CLASS) { + if (ev_pcb->evp_class_filter != ev->kev_class) { + lck_mtx_unlock(&ev_pcb->evp_mtx); + continue; + } + + if ((ev_pcb->evp_subclass_filter != + KEV_ANY_SUBCLASS) && + (ev_pcb->evp_subclass_filter != + ev->kev_subclass)) { + lck_mtx_unlock(&ev_pcb->evp_mtx); + continue; + } + } + } + + m2 = m_copym(m, 0, m->m_len, M_WAIT); + if (m2 == 0) { + OSIncrementAtomic64((SInt64 *)&kevtstat.kes_nomem); + m_free(m); + lck_mtx_unlock(&ev_pcb->evp_mtx); + lck_rw_done(kev_rwlock); + return (ENOMEM); + } + if (sbappendrecord(&ev_pcb->evp_socket->so_rcv, m2)) { + /* + * We use "m" for the socket stats as it would be + * unsafe to use "m2" + */ + so_inc_recv_data_stat(ev_pcb->evp_socket, + 1, m->m_len, MBUF_TC_BE); + + sorwakeup(ev_pcb->evp_socket); + OSIncrementAtomic64((SInt64 *)&kevtstat.kes_posted); + } else { + OSIncrementAtomic64((SInt64 *)&kevtstat.kes_fullsock); + } + lck_mtx_unlock(&ev_pcb->evp_mtx); + } + m_free(m); + lck_rw_done(kev_rwlock); + + return (0); +} + +static int +kev_control(struct socket *so, + u_long cmd, + caddr_t data, + __unused struct ifnet *ifp, + __unused struct proc *p) +{ + struct kev_request *kev_req = (struct kev_request *) data; + struct kern_event_pcb *ev_pcb; + struct kev_vendor_code *kev_vendor; + 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); + } + + return (0); +} + +int +kevt_getstat SYSCTL_HANDLER_ARGS +{ +#pragma unused(oidp, arg1, arg2) + int error = 0; + + lck_rw_lock_shared(kev_rwlock); + + if (req->newptr != USER_ADDR_NULL) { + error = EPERM; + goto done; + } + if (req->oldptr == USER_ADDR_NULL) { + req->oldidx = sizeof(struct kevtstat); + goto done; + } + + error = SYSCTL_OUT(req, &kevtstat, + MIN(sizeof(struct kevtstat), req->oldlen)); +done: + lck_rw_done(kev_rwlock); + + return (error); +} + +__private_extern__ int +kevt_pcblist SYSCTL_HANDLER_ARGS +{ +#pragma unused(oidp, arg1, arg2) + int error = 0; + int 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)); + struct kern_event_pcb *ev_pcb; + + buf = _MALLOC(item_size, M_TEMP, M_WAITOK | M_ZERO); + if (buf == NULL) + return (ENOMEM); + + lck_rw_lock_shared(kev_rwlock); + + n = kevtstat.kes_pcbcount; + + if (req->oldptr == USER_ADDR_NULL) { + req->oldidx = (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); + xsg.xg_count = n; + xsg.xg_gen = kevtstat.kes_gencnt; + xsg.xg_sogen = so_gencnt; + error = SYSCTL_OUT(req, &xsg, sizeof (xsg)); + if (error) { + goto done; + } + /* + * We are done if there is no pcb + */ + if (n == 0) { + goto done; + } + + i = 0; + for (i = 0, ev_pcb = LIST_FIRST(&kern_event_head); + i < n && ev_pcb != NULL; + 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)); + struct xsockbuf_n *xsbrcv = (struct xsockbuf_n *) + ADVANCE64(xso, sizeof (*xso)); + struct xsockbuf_n *xsbsnd = (struct xsockbuf_n *) + ADVANCE64(xsbrcv, sizeof (*xsbrcv)); + struct xsockstat_n *xsostats = (struct xsockstat_n *) + ADVANCE64(xsbsnd, sizeof (*xsbsnd)); + + bzero(buf, item_size); + + lck_mtx_lock(&ev_pcb->evp_mtx); + + xk->kep_len = sizeof(struct xkevtpcb); + xk->kep_kind = XSO_EVT; + xk->kep_evtpcb = (uint64_t)VM_KERNEL_ADDRPERM(ev_pcb); + xk->kep_vendor_code_filter = ev_pcb->evp_vendor_code_filter; + xk->kep_class_filter = ev_pcb->evp_class_filter; + xk->kep_subclass_filter = ev_pcb->evp_subclass_filter; + + sotoxsocket_n(ev_pcb->evp_socket, xso); + sbtoxsockbuf_n(ev_pcb->evp_socket ? + &ev_pcb->evp_socket->so_rcv : NULL, xsbrcv); + sbtoxsockbuf_n(ev_pcb->evp_socket ? + &ev_pcb->evp_socket->so_snd : NULL, xsbsnd); + sbtoxsockstat_n(ev_pcb->evp_socket, xsostats); + + lck_mtx_unlock(&ev_pcb->evp_mtx); + + error = SYSCTL_OUT(req, buf, item_size); + } + + if (error == 0) { + /* + * Give the user an updated idea of our state. + * If the generation differs from what we told + * her before, she knows that something happened + * while we were processing this request, and it + * might be necessary to retry. + */ + 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)); + if (error) { + goto done; } - list = &fdp->fd_knlist[kn->kn_id]; } - SLIST_INSERT_HEAD(list, kn, kn_link); - return (0); + +done: + lck_rw_done(kev_rwlock); + + return (error); } +#endif /* SOCKETS */ -/* - * should be called at spl == 0, since we don't want to hold spl - * while calling fdrop and free. - */ -static void -knote_drop(struct knote *kn, struct proc *p) +int +fill_kqueueinfo(struct kqueue *kq, struct kqueue_info * kinfo) { - struct filedesc *fdp = p->p_fd; - struct kqueue *kq = kn->kn_kq; - struct klist *list; + struct vinfo_stat * st; - proc_fdlock(p); - if (kn->kn_fop->f_isfd) - list = &fdp->fd_knlist[kn->kn_id]; - else - list = &fdp->fd_knhash[KN_HASH(kn->kn_id, fdp->fd_knhashmask)]; + st = &kinfo->kq_stat; - SLIST_REMOVE(list, kn, knote, kn_link); - kqlock(kq); - knote_dequeue(kn); - if (kn->kn_status & KN_DROPWAIT) - thread_wakeup(&kn->kn_status); - kqunlock(kq); - proc_fdunlock(p); + st->vst_size = kq->kq_count; + if (kq->kq_state & KQ_KEV_QOS) + st->vst_blksize = sizeof(struct kevent_qos_s); + else if (kq->kq_state & KQ_KEV64) + st->vst_blksize = sizeof(struct kevent64_s); + 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; - if (kn->kn_fop->f_isfd) - fp_drop(p, kn->kn_id, kn->kn_fp, 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; - knote_free(kn); + return (0); } -/* called with kqueue lock held */ -static void -knote_activate(struct knote *kn) +static int +fill_kqueue_dyninfo(struct kqueue *kq, struct kqueue_dyninfo *kqdi) { - struct kqueue *kq = kn->kn_kq; + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + struct kqrequest *kqr = &kqwl->kqwl_request; + workq_threadreq_param_t trp = {}; + int err; - kn->kn_status |= KN_ACTIVE; - knote_enqueue(kn); - kqueue_wakeup(kq); - } + if ((kq->kq_state & KQ_WORKLOOP) == 0) { + return EINVAL; + } -/* called with kqueue lock held */ -static void -knote_deactivate(struct knote *kn) -{ - kn->kn_status &= ~KN_ACTIVE; - knote_dequeue(kn); -} + if ((err = fill_kqueueinfo(kq, &kqdi->kqdi_info))) { + return err; + } -/* called with kqueue lock held */ -static void -knote_enqueue(struct knote *kn) -{ - struct kqueue *kq = kn->kn_kq; + kq_req_lock(kqwl); - if ((kn->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) { - struct kqtailq *tq = kn->kn_tq; + kqdi->kqdi_servicer = thread_tid(kqr->kqr_thread); + kqdi->kqdi_owner = thread_tid(kqwl->kqwl_owner); + 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 = 0; - TAILQ_INSERT_TAIL(tq, kn, kn_tqe); - kn->kn_status |= KN_QUEUED; - kq->kq_count++; - } -} + trp.trp_value = kqwl->kqwl_params; + if (trp.trp_flags & TRP_PRIORITY) + kqdi->kqdi_pri = trp.trp_pri; + else + kqdi->kqdi_pri = 0; -/* called with kqueue lock held */ -static void -knote_dequeue(struct knote *kn) -{ - struct kqueue *kq = kn->kn_kq; + if (trp.trp_flags & TRP_POLICY) + kqdi->kqdi_pol = trp.trp_pol; + else + kqdi->kqdi_pol = 0; - assert((kn->kn_status & KN_DISABLED) == 0); - if ((kn->kn_status & KN_QUEUED) == KN_QUEUED) { - struct kqtailq *tq = kn->kn_tq; + if (trp.trp_flags & TRP_CPUPERCENT) + kqdi->kqdi_cpupercent = trp.trp_cpupercent; + else + kqdi->kqdi_cpupercent = 0; - TAILQ_REMOVE(tq, kn, kn_tqe); - kn->kn_tq = &kq->kq_head; - kn->kn_status &= ~KN_QUEUED; - kq->kq_count--; - } + kq_req_unlock(kqwl); + + return 0; } + void -knote_init(void) +knote_markstayactive(struct knote *kn) { - knote_zone = zinit(sizeof(struct knote), 8192*sizeof(struct knote), 8192, "knote zone"); + struct kqueue *kq = knote_get_kq(kn); + kq_index_t qos; - /* allocate kq lock group attribute and group */ - kq_lck_grp_attr= lck_grp_attr_alloc_init(); - lck_grp_attr_setstat(kq_lck_grp_attr); + kqlock(kq); + kn->kn_status |= KN_STAYACTIVE; - kq_lck_grp = lck_grp_alloc_init("kqueue", kq_lck_grp_attr); + /* + * 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) { + qos = KQWQ_QOS_MANAGER; + } else if (kq->kq_state & KQ_WORKLOOP) { + struct kqworkloop *kqwl = (struct kqworkloop *)kq; + + qos = _pthread_priority_thread_qos(kn->kn_qos); + assert(qos && qos < THREAD_QOS_LAST); + kq_req_lock(kq); + kqworkloop_update_threads_qos(kqwl, KQWL_UTQ_UPDATE_STAYACTIVE_QOS, qos); + kq_req_unlock(kq); + qos = KQWL_BUCKET_STAYACTIVE; + } else { + qos = THREAD_QOS_UNSPECIFIED; + } - /* Allocate kq lock attribute */ - kq_lck_attr = lck_attr_alloc_init(); - lck_attr_setdefault(kq_lck_attr); + kn->kn_req_index = qos; + kn->kn_qos_override = qos; + kn->kn_qos_index = qos; - /* Initialize the timer filter lock */ - lck_mtx_init(&_filt_timerlock, kq_lck_grp, kq_lck_attr); + knote_activate(kn); + kqunlock(kq); } -SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL) -static struct knote * -knote_alloc(void) +void +knote_clearstayactive(struct knote *kn) { - return ((struct knote *)zalloc(knote_zone)); + kqlock(knote_get_kq(kn)); + kn->kn_status &= ~KN_STAYACTIVE; + knote_deactivate(kn); + kqunlock(knote_get_kq(kn)); } -static void -knote_free(struct knote *kn) +static unsigned long +kevent_extinfo_emit(struct kqueue *kq, struct knote *kn, struct kevent_extinfo *buf, + unsigned long buflen, unsigned long nknotes) { - zfree(knote_zone, kn); + 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; + + kqunlock(kq); + } + + /* we return total number of knotes, which may be more than requested */ + nknotes++; + } + } + + return nknotes; } -#include -#include -#include -#include -#include -#include -#include -#include -#include +int +kevent_copyout_proc_dynkqids(void *proc, user_addr_t ubuf, uint32_t ubufsize, + int32_t *nkqueues_out) +{ + proc_t p = (proc_t)proc; + struct filedesc *fdp = p->p_fd; + unsigned int nkqueues = 0; + unsigned long ubuflen = ubufsize / sizeof(kqueue_id_t); + size_t buflen, bufsize; + kqueue_id_t *kq_ids = NULL; + int err = 0; + assert(p != NULL); -static int kev_attach(struct socket *so, int proto, struct proc *p); -static int kev_detach(struct socket *so); -static int kev_control(struct socket *so, u_long cmd, caddr_t data, struct ifnet *ifp, struct proc *p); - -struct pr_usrreqs event_usrreqs = { - pru_abort_notsupp, pru_accept_notsupp, kev_attach, pru_bind_notsupp, pru_connect_notsupp, - pru_connect2_notsupp, kev_control, kev_detach, pru_disconnect_notsupp, - pru_listen_notsupp, pru_peeraddr_notsupp, pru_rcvd_notsupp, pru_rcvoob_notsupp, - pru_send_notsupp, pru_sense_null, pru_shutdown_notsupp, pru_sockaddr_notsupp, - pru_sosend_notsupp, soreceive, pru_sopoll_notsupp -}; + if (ubuf == USER_ADDR_NULL && ubufsize != 0) { + err = EINVAL; + goto out; + } -struct protosw eventsw[] = { - { - SOCK_RAW, &systemdomain, SYSPROTO_EVENT, PR_ATOMIC, - 0, 0, 0, 0, - 0, - 0, 0, 0, 0, -#if __APPLE__ - 0, -#endif - &event_usrreqs, - 0, 0, 0, -#if __APPLE__ - {0, 0}, 0, {0} -#endif - } -}; + buflen = min(ubuflen, PROC_PIDDYNKQUEUES_MAX); -static -struct kern_event_head kern_event_head; + if (ubuflen != 0) { + if (os_mul_overflow(sizeof(kqueue_id_t), buflen, &bufsize)) { + err = ERANGE; + goto out; + } + kq_ids = kalloc(bufsize); + if (!kq_ids) { + err = ENOMEM; + goto out; + } + bzero(kq_ids, bufsize); + } -static u_long static_event_id = 0; -struct domain *sysdom = &systemdomain; + kqhash_lock(p); + + 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) { + /* report the number of kqueues, even if they don't all fit */ + if (nkqueues < buflen) { + kq_ids[nkqueues] = kqwl->kqwl_dynamicid; + } + nkqueues++; + } + } + } + + kqhash_unlock(p); + + if (kq_ids) { + size_t copysize; + if (os_mul_overflow(sizeof(kqueue_id_t), min(buflen, nkqueues), ©size)) { + err = ERANGE; + goto out; + } + + assert(ubufsize >= copysize); + err = copyout(kq_ids, ubuf, copysize); + } + +out: + if (kq_ids) { + kfree(kq_ids, bufsize); + } + + if (!err) { + *nkqueues_out = (int)min(nkqueues, PROC_PIDDYNKQUEUES_MAX); + } + return err; +} -static lck_grp_t *evt_mtx_grp; -static lck_attr_t *evt_mtx_attr; -static lck_grp_attr_t *evt_mtx_grp_attr; -lck_mtx_t *evt_mutex; -/* - * Install the protosw's for the NKE manager. Invoked at - * extension load time - */ int -kern_event_init(void) +kevent_copyout_dynkqinfo(void *proc, kqueue_id_t kq_id, user_addr_t ubuf, + uint32_t ubufsize, int32_t *size_out) { - int retval; + proc_t p = (proc_t)proc; + struct kqueue *kq; + int err = 0; + struct kqueue_dyninfo kqdi = { }; - if ((retval = net_add_proto(eventsw, &systemdomain)) != 0) { - log(LOG_WARNING, "Can't install kernel events protocol (%d)\n", retval); - return(retval); + assert(p != NULL); + + if (ubufsize < sizeof(struct kqueue_info)) { + return ENOBUFS; } - - /* - * allocate lock group attribute and group for kern event - */ - evt_mtx_grp_attr = lck_grp_attr_alloc_init(); - evt_mtx_grp = lck_grp_alloc_init("eventlist", evt_mtx_grp_attr); - + kqhash_lock(p); + kq = kqueue_hash_lookup(p, kq_id); + if (!kq) { + kqhash_unlock(p); + return ESRCH; + } + kqueue_retain(kq); + kqhash_unlock(p); + /* - * allocate the lock attribute for mutexes + * backward compatibility: allow the argument to this call to only be + * a struct kqueue_info */ - evt_mtx_attr = lck_attr_alloc_init(); - lck_attr_setdefault(evt_mtx_attr); - evt_mutex = lck_mtx_alloc_init(evt_mtx_grp, evt_mtx_attr); - if (evt_mutex == NULL) - return (ENOMEM); - - return(KERN_SUCCESS); + if (ubufsize >= sizeof(struct kqueue_dyninfo)) { + ubufsize = sizeof(struct kqueue_dyninfo); + err = fill_kqueue_dyninfo(kq, &kqdi); + } else { + ubufsize = sizeof(struct kqueue_info); + err = fill_kqueueinfo(kq, &kqdi.kqdi_info); + } + if (err == 0 && (err = copyout(&kqdi, ubuf, ubufsize)) == 0) { + *size_out = ubufsize; + } + kqueue_release_last(p, kq); + return err; } -static int -kev_attach(struct socket *so, __unused int proto, __unused struct proc *p) +int +kevent_copyout_dynkqextinfo(void *proc, kqueue_id_t kq_id, user_addr_t ubuf, + uint32_t ubufsize, int32_t *nknotes_out) { - int error; - struct kern_event_pcb *ev_pcb; + proc_t p = (proc_t)proc; + struct kqueue *kq; + int err; - error = soreserve(so, KEV_SNDSPACE, KEV_RECVSPACE); - if (error) - return error; + assert(p != NULL); - MALLOC(ev_pcb, struct kern_event_pcb *, sizeof(struct kern_event_pcb), M_PCB, M_WAITOK); - if (ev_pcb == 0) - return ENOBUFS; + kqhash_lock(p); + kq = kqueue_hash_lookup(p, kq_id); + if (!kq) { + kqhash_unlock(p); + return ESRCH; + } + kqueue_retain(kq); + kqhash_unlock(p); - ev_pcb->ev_socket = so; - ev_pcb->vendor_code_filter = 0xffffffff; + err = pid_kqueue_extinfo(p, kq, ubuf, ubufsize, nknotes_out); + kqueue_release_last(p, kq); + return err; +} - so->so_pcb = (caddr_t) ev_pcb; - lck_mtx_lock(evt_mutex); - LIST_INSERT_HEAD(&kern_event_head, ev_pcb, ev_link); - lck_mtx_unlock(evt_mutex); +int +pid_kqueue_extinfo(proc_t p, struct kqueue *kq, user_addr_t ubuf, + uint32_t bufsize, int32_t *retval) +{ + struct knote *kn; + int i; + int err = 0; + struct filedesc *fdp = p->p_fd; + unsigned long nknotes = 0; + unsigned long buflen = bufsize / sizeof(struct kevent_extinfo); + struct kevent_extinfo *kqext = NULL; - return 0; -} + /* arbitrary upper limit to cap kernel memory usage, copyout size, etc. */ + buflen = min(buflen, PROC_PIDFDKQUEUE_KNOTES_MAX); + kqext = kalloc(buflen * sizeof(struct kevent_extinfo)); + if (kqext == NULL) { + err = ENOMEM; + goto out; + } + bzero(kqext, buflen * sizeof(struct kevent_extinfo)); -static int -kev_detach(struct socket *so) -{ - struct kern_event_pcb *ev_pcb = (struct kern_event_pcb *) so->so_pcb; + proc_fdlock(p); + for (i = 0; i < fdp->fd_knlistsize; i++) { + kn = SLIST_FIRST(&fdp->fd_knlist[i]); + nknotes = kevent_extinfo_emit(kq, kn, kqext, buflen, nknotes); + } + proc_fdunlock(p); - if (ev_pcb != 0) { - lck_mtx_lock(evt_mutex); - LIST_REMOVE(ev_pcb, ev_link); - lck_mtx_unlock(evt_mutex); - FREE(ev_pcb, M_PCB); - so->so_pcb = 0; - so->so_flags |= SOF_PCBCLEARING; - } + if (fdp->fd_knhashmask != 0) { + for (i = 0; i < (int)fdp->fd_knhashmask + 1; i++) { + kqhash_lock(p); + kn = SLIST_FIRST(&fdp->fd_knhash[i]); + nknotes = kevent_extinfo_emit(kq, kn, kqext, buflen, nknotes); + kqhash_unlock(p); + } + } + + 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; + } - return 0; + if (!err) { + *retval = min(nknotes, PROC_PIDFDKQUEUE_KNOTES_MAX); + } + return err; } -/* - * For now, kev_vender_code and mbuf_tags use the same - * mechanism. - */ -extern errno_t mbuf_tag_id_find_internal(const char *string, u_long *out_id, - int create); +static unsigned int +klist_copy_udata(struct klist *list, uint64_t *buf, + 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); + } + /* we return total number of knotes, which may be more than requested */ + nknotes++; + } -errno_t kev_vendor_code_find( - const char *string, - u_long *out_vender_code) + 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) { - if (strlen(string) >= KEV_VENDOR_CODE_MAX_STR_LEN) { - return EINVAL; + kqhash_lock_held(p); + struct kqworkloop *kqwl; + SLIST_FOREACH(kqwl, list, kqwl_hashlink) { + if (nids < buflen) { + buf[nids] = kqwl->kqwl_dynamicid; + } + nids++; } - return mbuf_tag_id_find_internal(string, out_vender_code, 1); + return nids; } -extern void mbuf_tag_id_first_last(u_long *first, u_long *last); +int +kevent_proc_copy_uptrs(void *proc, uint64_t *buf, int bufsize) +{ + proc_t p = (proc_t)proc; + struct filedesc *fdp = p->p_fd; + unsigned int nuptrs = 0; + unsigned long buflen = bufsize / sizeof(uint64_t); + + if (buflen > 0) { + assert(buf != NULL); + } + + proc_fdlock(p); + 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); + if (fdp->fd_knhashmask != 0) { + for (int i = 0; i < (int)fdp->fd_knhashmask + 1; i++) { + nuptrs = klist_copy_udata(&fdp->fd_knhash[i], buf, buflen, nuptrs); + } + } + knhash_unlock(p); + + kqhash_lock(p); + 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); + } + } + kqhash_unlock(p); + + return (int)nuptrs; +} -errno_t kev_msg_post(struct kev_msg *event_msg) +static void +kevent_set_return_to_kernel_user_tsd(proc_t p, thread_t thread) { - u_long min_vendor, max_vendor; - - mbuf_tag_id_first_last(&min_vendor, &max_vendor); - - 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) - { - return EINVAL; + uint64_t ast_addr; + bool proc_is_64bit = !!(p->p_flag & P_LP64); + size_t user_addr_size = proc_is_64bit ? 8 : 4; + uint32_t ast_flags32 = 0; + uint64_t ast_flags64 = 0; + struct uthread *ut = get_bsdthread_info(thread); + + if (ut->uu_kqr_bound != NULL) { + ast_flags64 |= R2K_WORKLOOP_PENDING_EVENTS; + } + + if (ast_flags64 == 0) { + return; + } + + if (!(p->p_flag & P_LP64)) { + ast_flags32 = (uint32_t)ast_flags64; + assert(ast_flags64 < 0x100000000ull); + } + + ast_addr = thread_rettokern_addr(thread); + if (ast_addr == 0) { + return; + } + + if (copyout((proc_is_64bit ? (void *)&ast_flags64 : (void *)&ast_flags32), + (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); } - - return kev_post_msg(event_msg); } - -int kev_post_msg(struct kev_msg *event_msg) +void +kevent_ast(thread_t thread, uint16_t bits) { - struct mbuf *m, *m2; - struct kern_event_pcb *ev_pcb; - struct kern_event_msg *ev; - char *tmp; - unsigned long total_size; - int i; + proc_t p = current_proc(); - /* Verify the message is small enough to fit in one mbuf w/o cluster */ - total_size = KEV_MSG_HEADER_SIZE; - - for (i = 0; i < 5; i++) { - if (event_msg->dv[i].data_length == 0) - break; - total_size += event_msg->dv[i].data_length; + if (bits & AST_KEVENT_REDRIVE_THREADREQ) { + 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); } - - if (total_size > MLEN) { - return EMSGSIZE; - } - - m = m_get(M_DONTWAIT, MT_DATA); - if (m == 0) - return ENOBUFS; - - 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) - break; - - total_size += event_msg->dv[i].data_length; - bcopy(event_msg->dv[i].data_ptr, tmp, - event_msg->dv[i].data_length); - tmp += event_msg->dv[i].data_length; - } - - ev->id = ++static_event_id; - ev->total_size = total_size; - ev->vendor_code = event_msg->vendor_code; - ev->kev_class = event_msg->kev_class; - ev->kev_subclass = event_msg->kev_subclass; - ev->event_code = event_msg->event_code; - - m->m_len = total_size; - lck_mtx_lock(evt_mutex); - for (ev_pcb = LIST_FIRST(&kern_event_head); - ev_pcb; - ev_pcb = LIST_NEXT(ev_pcb, ev_link)) { - - if (ev_pcb->vendor_code_filter != KEV_ANY_VENDOR) { - if (ev_pcb->vendor_code_filter != ev->vendor_code) - continue; - - if (ev_pcb->class_filter != KEV_ANY_CLASS) { - if (ev_pcb->class_filter != ev->kev_class) - continue; - - if ((ev_pcb->subclass_filter != KEV_ANY_SUBCLASS) && - (ev_pcb->subclass_filter != ev->kev_subclass)) - continue; - } - } - - m2 = m_copym(m, 0, m->m_len, M_NOWAIT); - if (m2 == 0) { - m_free(m); - lck_mtx_unlock(evt_mutex); - return ENOBUFS; - } - socket_lock(ev_pcb->ev_socket, 1); - if (sbappendrecord(&ev_pcb->ev_socket->so_rcv, m2)) - sorwakeup(ev_pcb->ev_socket); - socket_unlock(ev_pcb->ev_socket, 1); - } - - m_free(m); - lck_mtx_unlock(evt_mutex); - return 0; } +#if DEVELOPMENT || DEBUG + +#define KEVENT_SYSCTL_BOUND_ID 1 + static int -kev_control(struct socket *so, - u_long cmd, - caddr_t data, - __unused struct ifnet *ifp, - __unused struct proc *p) +kevent_sysctl SYSCTL_HANDLER_ARGS { - struct kev_request *kev_req = (struct kev_request *) data; - struct kern_event_pcb *ev_pcb; - struct kev_vendor_code *kev_vendor; - u_long *id_value = (u_long *) data; - - - switch (cmd) { - - case SIOCGKEVID: - *id_value = static_event_id; - break; - - case SIOCSKEVFILT: - ev_pcb = (struct kern_event_pcb *) so->so_pcb; - ev_pcb->vendor_code_filter = kev_req->vendor_code; - ev_pcb->class_filter = kev_req->kev_class; - ev_pcb->subclass_filter = kev_req->kev_subclass; - break; - - case SIOCGKEVFILT: - ev_pcb = (struct kern_event_pcb *) so->so_pcb; - kev_req->vendor_code = ev_pcb->vendor_code_filter; - kev_req->kev_class = ev_pcb->class_filter; - kev_req->kev_subclass = ev_pcb->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 mbuf_tag_id_find_internal(kev_vendor->vendor_string, - &kev_vendor->vendor_code, 0); - - default: - return ENOTSUP; +#pragma unused(oidp, arg2) + uintptr_t type = (uintptr_t)arg1; + uint64_t bound_id = 0; + + if (type != KEVENT_SYSCTL_BOUND_ID) { + return EINVAL; + } + + if (req->newptr) { + return EINVAL; + } + + struct uthread *ut = get_bsdthread_info(current_thread()); + if (!ut) { + return EFAULT; + } + + struct kqrequest *kqr = ut->uu_kqr_bound; + if (kqr) { + if (kqr->kqr_state & KQR_WORKLOOP) { + bound_id = kqr_kqworkloop(kqr)->kqwl_dynamicid; + } else { + bound_id = -1; + } } - - return 0; -} + return sysctl_io_number(req, bound_id, sizeof(bound_id), NULL, NULL); +} +SYSCTL_NODE(_kern, OID_AUTO, kevent, CTLFLAG_RW | CTLFLAG_LOCKED, 0, + "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"); +#endif /* DEVELOPMENT || DEBUG */