X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/39236c6e673c41db228275375ab7fdb0f837b292..ecc0ceb4089d506a0b8d16686a95817b331af9cb:/bsd/kern/kern_memorystatus.c diff --git a/bsd/kern/kern_memorystatus.c b/bsd/kern/kern_memorystatus.c index 852037af4..22f7edbd3 100644 --- a/bsd/kern/kern_memorystatus.c +++ b/bsd/kern/kern_memorystatus.c @@ -31,16 +31,18 @@ #include #include #include -#include +#include #include #include #include #include +#include #include #include #include #include #include +#include #include #include #include @@ -60,6 +62,38 @@ #include +#if CONFIG_JETSAM +/* For logging clarity */ +static const char *jetsam_kill_cause_name[] = { + "" , + "jettisoned" , /* kMemorystatusKilled */ + "highwater" , /* kMemorystatusKilledHiwat */ + "vnode-limit" , /* kMemorystatusKilledVnodes */ + "vm-pageshortage" , /* kMemorystatusKilledVMPageShortage */ + "vm-thrashing" , /* kMemorystatusKilledVMThrashing */ + "fc-thrashing" , /* kMemorystatusKilledFCThrashing */ + "per-process-limit" , /* kMemorystatusKilledPerProcessLimit */ + "diagnostic" , /* kMemorystatusKilledDiagnostic */ + "idle-exit" , /* kMemorystatusKilledIdleExit */ +}; + +/* Does cause indicate vm or fc thrashing? */ +static boolean_t +is_thrashing(unsigned cause) +{ + switch (cause) { + case kMemorystatusKilledVMThrashing: + case kMemorystatusKilledFCThrashing: + return TRUE; + default: + return FALSE; + } +} + +/* Callback into vm_compressor.c to signal that thrashing has been mitigated. */ +extern void vm_thrashing_jetsam_done(void); +#endif + /* These are very verbose printfs(), enable with * MEMORYSTATUS_DEBUG_LOG */ @@ -72,6 +106,70 @@ do { \ #define MEMORYSTATUS_DEBUG(cond, format, ...) #endif +/* + * Active / Inactive limit support + * proc list must be locked + * + * The SET_*** macros are used to initialize a limit + * for the first time. + * + * The CACHE_*** macros are use to cache the limit that will + * soon be in effect down in the ledgers. + */ + +#define SET_ACTIVE_LIMITS_LOCKED(p, limit, is_fatal) \ +MACRO_BEGIN \ +(p)->p_memstat_memlimit_active = (limit); \ + (p)->p_memstat_state &= ~P_MEMSTAT_MEMLIMIT_ACTIVE_EXC_TRIGGERED; \ + if (is_fatal) { \ + (p)->p_memstat_state |= P_MEMSTAT_MEMLIMIT_ACTIVE_FATAL; \ + } else { \ + (p)->p_memstat_state &= ~P_MEMSTAT_MEMLIMIT_ACTIVE_FATAL; \ + } \ +MACRO_END + +#define SET_INACTIVE_LIMITS_LOCKED(p, limit, is_fatal) \ +MACRO_BEGIN \ +(p)->p_memstat_memlimit_inactive = (limit); \ + (p)->p_memstat_state &= ~P_MEMSTAT_MEMLIMIT_INACTIVE_EXC_TRIGGERED; \ + if (is_fatal) { \ + (p)->p_memstat_state |= P_MEMSTAT_MEMLIMIT_INACTIVE_FATAL; \ + } else { \ + (p)->p_memstat_state &= ~P_MEMSTAT_MEMLIMIT_INACTIVE_FATAL; \ + } \ +MACRO_END + +#define CACHE_ACTIVE_LIMITS_LOCKED(p, trigger_exception) \ +MACRO_BEGIN \ +(p)->p_memstat_memlimit = (p)->p_memstat_memlimit_active; \ + if ((p)->p_memstat_state & P_MEMSTAT_MEMLIMIT_ACTIVE_FATAL) { \ + (p)->p_memstat_state |= P_MEMSTAT_FATAL_MEMLIMIT; \ + } else { \ + (p)->p_memstat_state &= ~P_MEMSTAT_FATAL_MEMLIMIT; \ + } \ + if ((p)->p_memstat_state & P_MEMSTAT_MEMLIMIT_ACTIVE_EXC_TRIGGERED) { \ + trigger_exception = FALSE; \ + } else { \ + trigger_exception = TRUE; \ + } \ +MACRO_END + +#define CACHE_INACTIVE_LIMITS_LOCKED(p, trigger_exception) \ +MACRO_BEGIN \ +(p)->p_memstat_memlimit = (p)->p_memstat_memlimit_inactive; \ + if ((p)->p_memstat_state & P_MEMSTAT_MEMLIMIT_INACTIVE_FATAL) { \ + (p)->p_memstat_state |= P_MEMSTAT_FATAL_MEMLIMIT; \ + } else { \ + (p)->p_memstat_state &= ~P_MEMSTAT_FATAL_MEMLIMIT; \ + } \ + if ((p)->p_memstat_state & P_MEMSTAT_MEMLIMIT_INACTIVE_EXC_TRIGGERED) { \ + trigger_exception = FALSE; \ + } else { \ + trigger_exception = TRUE; \ + } \ +MACRO_END + + /* General tunables */ unsigned long delta_percentage = 5; @@ -105,8 +203,9 @@ struct filterops memorystatus_filtops = { }; enum { - kMemorystatusNoPressure = 1, - kMemorystatusPressure = 2 + kMemorystatusNoPressure = 0x1, + kMemorystatusPressure = 0x2, + kMemorystatusLowSwap = 0x4 }; /* Idle guard handling */ @@ -120,11 +219,15 @@ static void memorystatus_schedule_idle_demotion_locked(proc_t p, boolean_t set_s static void memorystatus_invalidate_idle_demotion_locked(proc_t p, boolean_t clean_state); static void memorystatus_reschedule_idle_demotion_locked(void); -static void memorystatus_update_priority_locked(proc_t p, int priority); +static void memorystatus_update_priority_locked(proc_t p, int priority, boolean_t head_insert); + +boolean_t is_knote_registered_modify_task_pressure_bits(struct knote*, int, task_t, vm_pressure_level_t, vm_pressure_level_t); +void memorystatus_send_low_swap_note(void); int memorystatus_wakeup = 0; unsigned int memorystatus_level = 0; +unsigned int memorystatus_early_boot_level = 0; static int memorystatus_list_count = 0; @@ -141,9 +244,9 @@ uint64_t memstat_idle_demotion_deadline = 0; static unsigned int memorystatus_dirty_count = 0; -#if !CONFIG_JETSAM -static boolean_t kill_idle_exit = FALSE; -#endif +#if CONFIG_JETSAM +SYSCTL_INT(_kern, OID_AUTO, max_task_pmem, CTLFLAG_RD|CTLFLAG_LOCKED|CTLFLAG_MASKED, &max_task_footprint_mb, 0, ""); +#endif // CONFIG_JETSAM int @@ -169,53 +272,101 @@ static void memorystatus_thread(void *param __unused, wait_result_t wr __unused) #if CONFIG_JETSAM +static int memorystatus_cmd_set_jetsam_memory_limit(pid_t pid, int32_t high_water_mark, __unused int32_t *retval, boolean_t is_fatal_limit); + +static int memorystatus_cmd_set_memlimit_properties(pid_t pid, user_addr_t buffer, size_t buffer_size, __unused int32_t *retval); + +static int memorystatus_set_memlimit_properties(pid_t pid, memorystatus_memlimit_properties_t *entry); + +static int memorystatus_cmd_get_memlimit_properties(pid_t pid, user_addr_t buffer, size_t buffer_size, __unused int32_t *retval); + +static boolean_t proc_jetsam_state_is_active_locked(proc_t); + +int proc_get_memstat_priority(proc_t, boolean_t); + /* Kill processes exceeding their limit either under memory pressure (1), or as soon as possible (0) */ #define LEGACY_HIWATER 1 -static int memorystatus_highwater_enabled = 1; +static boolean_t memorystatus_idle_snapshot = 0; -extern unsigned int vm_page_free_count; -extern unsigned int vm_page_active_count; -extern unsigned int vm_page_inactive_count; -extern unsigned int vm_page_throttled_count; -extern unsigned int vm_page_purgeable_count; -extern unsigned int vm_page_wire_count; +static int memorystatus_highwater_enabled = 1; /* Update the cached memlimit data. This should be removed. */ unsigned int memorystatus_delta = 0; -static unsigned int memorystatus_available_pages = (unsigned int)-1; -static unsigned int memorystatus_available_pages_pressure = 0; -static unsigned int memorystatus_available_pages_critical = 0; static unsigned int memorystatus_available_pages_critical_base = 0; -static unsigned int memorystatus_last_foreground_pressure_pages = (unsigned int)-1; -#if !LATENCY_JETSAM +//static unsigned int memorystatus_last_foreground_pressure_pages = (unsigned int)-1; static unsigned int memorystatus_available_pages_critical_idle_offset = 0; -#endif + +/* Jetsam Loop Detection */ +static boolean_t memorystatus_jld_enabled = TRUE; /* Enables jetsam loop detection on all devices */ +static uint32_t memorystatus_jld_eval_period_msecs = 0; /* Init pass sets this based on device memory size */ +static int memorystatus_jld_eval_aggressive_count = 3; /* Raise the priority max after 'n' aggressive loops */ +static int memorystatus_jld_eval_aggressive_priority_band_max = 15; /* Kill aggressively up through this band */ + +/* + * A FG app can request that the aggressive jetsam mechanism display some leniency in the FG band. This 'lenient' mode is described as: + * --- if aggressive jetsam kills an app in the FG band and gets back >=AGGRESSIVE_JETSAM_LENIENT_MODE_THRESHOLD memory, it will stop the aggressive march further into and up the jetsam bands. + * + * RESTRICTIONS: + * - Such a request is respected/acknowledged only once while that 'requesting' app is in the FG band i.e. if aggressive jetsam was + * needed and the 'lenient' mode was deployed then that's it for this special mode while the app is in the FG band. + * + * - If the app is still in the FG band and aggressive jetsam is needed again, there will be no stop-and-check the next time around. + * + * - Also, the transition of the 'requesting' app away from the FG band will void this special behavior. + */ + +#define AGGRESSIVE_JETSAM_LENIENT_MODE_THRESHOLD 25 +boolean_t memorystatus_aggressive_jetsam_lenient_allowed = FALSE; +boolean_t memorystatus_aggressive_jetsam_lenient = FALSE; + +#if DEVELOPMENT || DEBUG +/* + * Jetsam Loop Detection tunables. + */ + +SYSCTL_UINT(_kern, OID_AUTO, memorystatus_jld_eval_period_msecs, CTLFLAG_RW|CTLFLAG_LOCKED, &memorystatus_jld_eval_period_msecs, 0, ""); +SYSCTL_UINT(_kern, OID_AUTO, memorystatus_jld_eval_aggressive_count, CTLFLAG_RW|CTLFLAG_LOCKED, &memorystatus_jld_eval_aggressive_count, 0, ""); +SYSCTL_UINT(_kern, OID_AUTO, memorystatus_jld_eval_aggressive_priority_band_max, CTLFLAG_RW|CTLFLAG_LOCKED, &memorystatus_jld_eval_aggressive_priority_band_max, 0, ""); +#endif /* DEVELOPMENT || DEBUG */ #if DEVELOPMENT || DEBUG static unsigned int memorystatus_jetsam_panic_debug = 0; static unsigned int memorystatus_jetsam_policy = kPolicyDefault; static unsigned int memorystatus_jetsam_policy_offset_pages_diagnostic = 0; +static unsigned int memorystatus_debug_dump_this_bucket = 0; #endif -static boolean_t kill_under_pressure = FALSE; +static unsigned int memorystatus_thread_wasted_wakeup = 0; +static uint32_t kill_under_pressure_cause = 0; + +/* + * default jetsam snapshot support + */ static memorystatus_jetsam_snapshot_t *memorystatus_jetsam_snapshot; #define memorystatus_jetsam_snapshot_list memorystatus_jetsam_snapshot->entries - static unsigned int memorystatus_jetsam_snapshot_count = 0; static unsigned int memorystatus_jetsam_snapshot_max = 0; +static uint64_t memorystatus_jetsam_snapshot_last_timestamp = 0; +static uint64_t memorystatus_jetsam_snapshot_timeout = 0; +#define JETSAM_SNAPSHOT_TIMEOUT_SECS 30 + +/* + * snapshot support for memstats collected at boot. + */ +static memorystatus_jetsam_snapshot_t memorystatus_at_boot_snapshot; static void memorystatus_clear_errors(void); -static void memorystatus_get_task_page_counts(task_t task, uint32_t *footprint, uint32_t *max_footprint); -static int memorystatus_send_note(int event_code, void *data, size_t data_length); +static void memorystatus_get_task_page_counts(task_t task, uint32_t *footprint, uint32_t *max_footprint, uint32_t *max_footprint_lifetime, uint32_t *purgeable_pages); static uint32_t memorystatus_build_state(proc_t p); static void memorystatus_update_levels_locked(boolean_t critical_only); -static boolean_t memorystatus_issue_pressure_kevent(boolean_t pressured); +//static boolean_t memorystatus_issue_pressure_kevent(boolean_t pressured); static boolean_t memorystatus_kill_specific_process(pid_t victim_pid, uint32_t cause); -static boolean_t memorystatus_kill_top_process(boolean_t any, uint32_t cause, int32_t *priority, uint32_t *errors); +static boolean_t memorystatus_kill_top_process(boolean_t any, boolean_t sort_flag, uint32_t cause, int32_t *priority, uint32_t *errors); +static boolean_t memorystatus_kill_top_process_aggressive(boolean_t any, uint32_t cause, int aggr_count, int32_t priority_max, uint32_t *errors); #if LEGACY_HIWATER static boolean_t memorystatus_kill_hiwat_proc(uint32_t *errors); #endif @@ -223,18 +374,60 @@ static boolean_t memorystatus_kill_hiwat_proc(uint32_t *errors); static boolean_t memorystatus_kill_process_async(pid_t victim_pid, uint32_t cause); static boolean_t memorystatus_kill_process_sync(pid_t victim_pid, uint32_t cause); +/* Priority Band Sorting Routines */ +static int memorystatus_sort_bucket(unsigned int bucket_index, int sort_order); +static int memorystatus_sort_by_largest_coalition_locked(unsigned int bucket_index, int coal_sort_order); +static void memorystatus_sort_by_largest_process_locked(unsigned int bucket_index); +static int memorystatus_move_list_locked(unsigned int bucket_index, pid_t *pid_list, int list_sz); + +/* qsort routines */ +typedef int (*cmpfunc_t)(const void *a, const void *b); +extern void qsort(void *a, size_t n, size_t es, cmpfunc_t cmp); +static int memstat_asc_cmp(const void *a, const void *b); + #endif /* CONFIG_JETSAM */ /* VM pressure */ +extern unsigned int vm_page_free_count; +extern unsigned int vm_page_active_count; +extern unsigned int vm_page_inactive_count; +extern unsigned int vm_page_throttled_count; +extern unsigned int vm_page_purgeable_count; +extern unsigned int vm_page_wire_count; + #if VM_PRESSURE_EVENTS #include "vm_pressure.h" -extern boolean_t memorystatus_warn_process(pid_t pid); +extern boolean_t memorystatus_warn_process(pid_t pid, boolean_t critical); vm_pressure_level_t memorystatus_vm_pressure_level = kVMPressureNormal; +#if CONFIG_MEMORYSTATUS +unsigned int memorystatus_available_pages = (unsigned int)-1; +unsigned int memorystatus_available_pages_pressure = 0; +unsigned int memorystatus_available_pages_critical = 0; +unsigned int memorystatus_frozen_count = 0; +unsigned int memorystatus_suspended_count = 0; + +/* + * We use this flag to signal if we have any HWM offenders + * on the system. This way we can reduce the number of wakeups + * of the memorystatus_thread when the system is between the + * "pressure" and "critical" threshold. + * + * The (re-)setting of this variable is done without any locks + * or synchronization simply because it is not possible (currently) + * to keep track of HWM offenders that drop down below their memory + * limit and/or exit. So, we choose to burn a couple of wasted wakeups + * by allowing the unguarded modification of this variable. + */ +boolean_t memorystatus_hwm_candidates = 0; + +static int memorystatus_send_note(int event_code, void *data, size_t data_length); +#endif /* CONFIG_MEMORYSTATUS */ + #endif /* VM_PRESSURE_EVENTS */ /* Freeze */ @@ -244,6 +437,10 @@ vm_pressure_level_t memorystatus_vm_pressure_level = kVMPressureNormal; boolean_t memorystatus_freeze_enabled = FALSE; int memorystatus_freeze_wakeup = 0; +lck_grp_attr_t *freezer_lck_grp_attr; +lck_grp_t *freezer_lck_grp; +static lck_mtx_t freezer_mutex; + static inline boolean_t memorystatus_can_freeze_processes(void); static boolean_t memorystatus_can_freeze(boolean_t *memorystatus_freeze_swap_low); @@ -252,13 +449,13 @@ static void memorystatus_freeze_thread(void *param __unused, wait_result_t wr __ /* Thresholds */ static unsigned int memorystatus_freeze_threshold = 0; -static unsigned int memorystatus_freeze_pages_min = FREEZE_PAGES_MIN; -static unsigned int memorystatus_freeze_pages_max = FREEZE_PAGES_MAX; - -static unsigned int memorystatus_frozen_count = 0; +static unsigned int memorystatus_freeze_pages_min = 0; +static unsigned int memorystatus_freeze_pages_max = 0; static unsigned int memorystatus_freeze_suspended_threshold = FREEZE_SUSPENDED_THRESHOLD_DEFAULT; +static unsigned int memorystatus_freeze_daily_mb_max = FREEZE_DAILY_MB_MAX_DEFAULT; + /* Stats */ static uint64_t memorystatus_freeze_count = 0; static uint64_t memorystatus_freeze_pageouts = 0; @@ -271,17 +468,105 @@ static throttle_interval_t throttle_intervals[] = { static uint64_t memorystatus_freeze_throttle_count = 0; -static unsigned int memorystatus_suspended_count = 0; static unsigned int memorystatus_suspended_footprint_total = 0; +extern uint64_t vm_swap_get_free_space(void); + +static boolean_t memorystatus_freeze_update_throttle(); + #endif /* CONFIG_FREEZE */ /* Debug */ +extern struct knote *vm_find_knote_from_pid(pid_t, struct klist *); + #if DEVELOPMENT || DEBUG #if CONFIG_JETSAM +static void +memorystatus_debug_dump_bucket_locked (unsigned int bucket_index) +{ + proc_t p = NULL; + uint32_t pages = 0; + uint32_t pages_in_mb = 0; + unsigned int b = bucket_index; + boolean_t traverse_all_buckets = FALSE; + + if (bucket_index >= MEMSTAT_BUCKET_COUNT) { + traverse_all_buckets = TRUE; + b = 0; + } else { + traverse_all_buckets = FALSE; + b = bucket_index; + } + + /* + * Missing from this dump is the value actually + * stored in the ledger... also, format could be better. + */ + printf("memorystatus_debug_dump ***START***\n"); + printf("bucket [pid] [pages/pages-mb] state [EP / RP] dirty deadline [C-limit / A-limit / IA-limit] name\n"); + p = memorystatus_get_first_proc_locked(&b, traverse_all_buckets); + while (p) { + memorystatus_get_task_page_counts(p->task, &pages, NULL, NULL, NULL); + pages_in_mb = (pages * 4096) /1024 / 1024; + printf("%d [%d] [%d/%dMB] 0x%x [%d / %d] 0x%x %lld [%d%s / %d%s / %d%s] %s\n", + b, p->p_pid, pages, pages_in_mb, + p->p_memstat_state, p->p_memstat_effectivepriority, p->p_memstat_requestedpriority, p->p_memstat_dirty, p->p_memstat_idledeadline, + p->p_memstat_memlimit, + (p->p_memstat_state & P_MEMSTAT_FATAL_MEMLIMIT ? "F " : "NF"), + p->p_memstat_memlimit_active, + (p->p_memstat_state & P_MEMSTAT_MEMLIMIT_ACTIVE_FATAL ? "F " : "NF"), + p->p_memstat_memlimit_inactive, + (p->p_memstat_state & P_MEMSTAT_MEMLIMIT_INACTIVE_FATAL ? "F " : "NF"), + (p->p_comm ? p->p_comm : "unknown")); + p = memorystatus_get_next_proc_locked(&b, p, traverse_all_buckets); + } + printf("memorystatus_debug_dump ***END***\n"); +} + +static int +sysctl_memorystatus_debug_dump_bucket SYSCTL_HANDLER_ARGS +{ +#pragma unused(oidp, arg2) + int bucket_index = 0; + int error; + error = SYSCTL_OUT(req, arg1, sizeof(int)); + if (error || !req->newptr) { + return (error); + } + error = SYSCTL_IN(req, &bucket_index, sizeof(int)); + if (error || !req->newptr) { + return (error); + } + if (bucket_index >= MEMSTAT_BUCKET_COUNT) { + /* + * All jetsam buckets will be dumped. + */ + } else { + /* + * Only a single bucket will be dumped. + */ + } + + proc_list_lock(); + memorystatus_debug_dump_bucket_locked(bucket_index); + proc_list_unlock(); + memorystatus_debug_dump_this_bucket = bucket_index; + return (error); +} + +/* + * Debug aid to look at jetsam buckets and proc jetsam fields. + * Use this sysctl to act on a particular jetsam bucket. + * Writing the sysctl triggers the dump. + * Usage: sysctl kern.memorystatus_debug_dump_this_bucket= + */ + +SYSCTL_PROC(_kern, OID_AUTO, memorystatus_debug_dump_this_bucket, CTLTYPE_INT|CTLFLAG_RW|CTLFLAG_LOCKED, &memorystatus_debug_dump_this_bucket, 0, sysctl_memorystatus_debug_dump_bucket, "I", ""); + + /* Debug aid to aid determination of limit */ static int @@ -291,7 +576,6 @@ sysctl_memorystatus_highwater_enable SYSCTL_HANDLER_ARGS proc_t p; unsigned int b = 0; int error, enable = 0; - int32_t memlimit; error = SYSCTL_OUT(req, arg1, sizeof(int)); if (error || !req->newptr) { @@ -311,17 +595,35 @@ sysctl_memorystatus_highwater_enable SYSCTL_HANDLER_ARGS p = memorystatus_get_first_proc_locked(&b, TRUE); while (p) { + boolean_t trigger_exception; + if (enable) { - if ((p->p_memstat_state & P_MEMSTAT_MEMLIMIT_BACKGROUND) && (p->p_memstat_effectivepriority >= JETSAM_PRIORITY_FOREGROUND)) { - memlimit = -1; + /* + * No need to consider P_MEMSTAT_MEMLIMIT_BACKGROUND anymore. + * Background limits are described via the inactive limit slots. + */ + + if (proc_jetsam_state_is_active_locked(p) == TRUE) { + CACHE_ACTIVE_LIMITS_LOCKED(p, trigger_exception); } else { - memlimit = p->p_memstat_memlimit; + CACHE_INACTIVE_LIMITS_LOCKED(p, trigger_exception); } + } else { - memlimit = -1; + /* + * Disabling limits does not touch the stored variants. + * Set the cached limit fields to system_wide defaults. + */ + p->p_memstat_memlimit = -1; + p->p_memstat_state |= P_MEMSTAT_FATAL_MEMLIMIT; + trigger_exception = TRUE; } - task_set_phys_footprint_limit_internal(p->task, (memlimit > 0) ? memlimit : -1, NULL, TRUE); - + + /* + * Enforce the cached limit by writing to the ledger. + */ + task_set_phys_footprint_limit_internal(p->task, (p->p_memstat_memlimit > 0) ? p->p_memstat_memlimit: -1, NULL, trigger_exception); + p = memorystatus_get_next_proc_locked(&b, p, TRUE); } @@ -330,16 +632,17 @@ sysctl_memorystatus_highwater_enable SYSCTL_HANDLER_ARGS proc_list_unlock(); return 0; + } +SYSCTL_INT(_kern, OID_AUTO, memorystatus_idle_snapshot, CTLFLAG_RW|CTLFLAG_LOCKED, &memorystatus_idle_snapshot, 0, ""); + SYSCTL_PROC(_kern, OID_AUTO, memorystatus_highwater_enabled, CTLTYPE_INT|CTLFLAG_RW|CTLFLAG_LOCKED, &memorystatus_highwater_enabled, 0, sysctl_memorystatus_highwater_enable, "I", ""); SYSCTL_UINT(_kern, OID_AUTO, memorystatus_available_pages, CTLFLAG_RD|CTLFLAG_LOCKED, &memorystatus_available_pages, 0, ""); SYSCTL_UINT(_kern, OID_AUTO, memorystatus_available_pages_critical, CTLFLAG_RD|CTLFLAG_LOCKED, &memorystatus_available_pages_critical, 0, ""); SYSCTL_UINT(_kern, OID_AUTO, memorystatus_available_pages_critical_base, CTLFLAG_RW|CTLFLAG_LOCKED, &memorystatus_available_pages_critical_base, 0, ""); -#if !LATENCY_JETSAM SYSCTL_UINT(_kern, OID_AUTO, memorystatus_available_pages_critical_idle_offset, CTLFLAG_RW|CTLFLAG_LOCKED, &memorystatus_available_pages_critical_idle_offset, 0, ""); -#endif /* Diagnostic code */ @@ -418,34 +721,64 @@ SYSCTL_UINT(_kern, OID_AUTO, memorystatus_jetsam_policy_offset_pages_diagnostic, SYSCTL_UINT(_kern, OID_AUTO, memorystatus_available_pages_pressure, CTLFLAG_RW|CTLFLAG_LOCKED, &memorystatus_available_pages_pressure, 0, ""); -static int -sysctl_memorystatus_vm_pressure_level SYSCTL_HANDLER_ARGS -{ -#pragma unused(arg1, arg2, oidp) - int error = 0; - - error = priv_check_cred(kauth_cred_get(), PRIV_VM_PRESSURE, 0); - if (error) - return (error); - - return SYSCTL_OUT(req, &memorystatus_vm_pressure_level, sizeof(memorystatus_vm_pressure_level)); -} -SYSCTL_PROC(_kern, OID_AUTO, memorystatus_vm_pressure_level, CTLTYPE_INT|CTLFLAG_RD|CTLFLAG_LOCKED|CTLFLAG_MASKED, - 0, 0, &sysctl_memorystatus_vm_pressure_level, "I", ""); +/* + * This routine is used for targeted notifications + * regardless of system memory pressure. + * "memnote" is the current user. + */ static int sysctl_memorystatus_vm_pressure_send SYSCTL_HANDLER_ARGS { #pragma unused(arg1, arg2) - int error, pid = 0; + int error = 0, pid = 0; + int ret = 0; + struct knote *kn = NULL; + boolean_t found_knote = FALSE; error = sysctl_handle_int(oidp, &pid, 0, req); if (error || !req->newptr) return (error); - return vm_dispatch_pressure_note_to_pid(pid, FALSE); + /* + * We inspect 3 lists here for targeted notifications: + * - memorystatus_klist + * - vm_pressure_klist + * - vm_pressure_dormant_klist + * + * The vm_pressure_* lists are tied to the old VM_PRESSURE + * notification mechanism. We intend to stop using that + * mechanism and, in turn, get rid of the 2 lists and + * vm_dispatch_pressure_note_to_pid() too. + */ + + memorystatus_klist_lock(); + + SLIST_FOREACH(kn, &memorystatus_klist, kn_selnext) { + proc_t knote_proc = kn->kn_kq->kq_p; + pid_t knote_pid = knote_proc->p_pid; + + if (knote_pid == pid) { + /* + * Forcibly send this pid a "warning" memory pressure notification. + */ + kn->kn_fflags = NOTE_MEMORYSTATUS_PRESSURE_WARN; + found_knote = TRUE; + } + } + + if (found_knote) { + KNOTE(&memorystatus_klist, 0); + ret = 0; + } else { + ret = vm_dispatch_pressure_note_to_pid(pid, FALSE); + } + + memorystatus_klist_unlock(); + + return ret; } SYSCTL_PROC(_kern, OID_AUTO, memorystatus_vm_pressure_send, CTLTYPE_INT|CTLFLAG_WR|CTLFLAG_LOCKED|CTLFLAG_MASKED, @@ -455,10 +788,10 @@ SYSCTL_PROC(_kern, OID_AUTO, memorystatus_vm_pressure_send, CTLTYPE_INT|CTLFLAG_ #endif /* CONFIG_JETSAM */ -#endif /* DEVELOPMENT || DEBUG */ - #if CONFIG_FREEZE +SYSCTL_UINT(_kern, OID_AUTO, memorystatus_freeze_daily_mb_max, CTLFLAG_RW|CTLFLAG_LOCKED, &memorystatus_freeze_daily_mb_max, 0, ""); + SYSCTL_UINT(_kern, OID_AUTO, memorystatus_freeze_threshold, CTLFLAG_RW|CTLFLAG_LOCKED, &memorystatus_freeze_threshold, 0, ""); SYSCTL_UINT(_kern, OID_AUTO, memorystatus_freeze_pages_min, CTLFLAG_RW|CTLFLAG_LOCKED, &memorystatus_freeze_pages_min, 0, ""); @@ -473,41 +806,74 @@ boolean_t memorystatus_freeze_throttle_enabled = TRUE; SYSCTL_UINT(_kern, OID_AUTO, memorystatus_freeze_throttle_enabled, CTLFLAG_RW|CTLFLAG_LOCKED, &memorystatus_freeze_throttle_enabled, 0, ""); /* - * Enabled via: Enable the sysctl_memorystatus_freeze/thaw sysctls on Release KC - * - * TODO: Manual trigger of freeze and thaw for dev / debug kernels only. - * Disable/restrict the sysctl_memorystatus_freeze/thaw sysctls on Release KC + * Manual trigger of freeze and thaw for dev / debug kernels only. */ static int sysctl_memorystatus_freeze SYSCTL_HANDLER_ARGS { #pragma unused(arg1, arg2) - int error, pid = 0; proc_t p; + if (memorystatus_freeze_enabled == FALSE) { + return ENOTSUP; + } + error = sysctl_handle_int(oidp, &pid, 0, req); if (error || !req->newptr) return (error); + if (pid == 2) { + vm_pageout_anonymous_pages(); + + return 0; + } + + lck_mtx_lock(&freezer_mutex); + p = proc_find(pid); if (p != NULL) { uint32_t purgeable, wired, clean, dirty; boolean_t shared; uint32_t max_pages = 0; - if (DEFAULT_FREEZER_IS_ACTIVE || DEFAULT_FREEZER_COMPRESSED_PAGER_IS_ACTIVE) { - max_pages = MIN(default_pager_swap_pages_free(), memorystatus_freeze_pages_max); + if (DEFAULT_FREEZER_IS_ACTIVE || DEFAULT_FREEZER_COMPRESSED_PAGER_IS_SWAPBACKED) { + + unsigned int avail_swap_space = 0; /* in pages. */ + + if (DEFAULT_FREEZER_IS_ACTIVE) { + /* + * Freezer backed by default pager and swap file(s). + */ + avail_swap_space = default_pager_swap_pages_free(); + } else { + /* + * Freezer backed by the compressor and swap file(s) + * while will hold compressed data. + */ + avail_swap_space = vm_swap_get_free_space() / PAGE_SIZE_64; + } + + max_pages = MIN(avail_swap_space, memorystatus_freeze_pages_max); + } else { + /* + * We only have the compressor without any swap. + */ max_pages = UINT32_MAX - 1; } + error = task_freeze(p->task, &purgeable, &wired, &clean, &dirty, max_pages, &shared, FALSE); proc_rele(p); if (error) error = EIO; + + lck_mtx_unlock(&freezer_mutex); return error; } + + lck_mtx_unlock(&freezer_mutex); return EINVAL; } @@ -522,6 +888,10 @@ sysctl_memorystatus_available_pages_thaw SYSCTL_HANDLER_ARGS int error, pid = 0; proc_t p; + if (memorystatus_freeze_enabled == FALSE) { + return ENOTSUP; + } + error = sysctl_handle_int(oidp, &pid, 0, req); if (error || !req->newptr) return (error); @@ -544,11 +914,126 @@ SYSCTL_PROC(_kern, OID_AUTO, memorystatus_thaw, CTLTYPE_INT|CTLFLAG_WR|CTLFLAG_L #endif /* CONFIG_FREEZE */ +#endif /* DEVELOPMENT || DEBUG */ + extern kern_return_t kernel_thread_start_priority(thread_continue_t continuation, void *parameter, integer_t priority, thread_t *new_thread); +#if CONFIG_JETSAM +/* + * Picks the sorting routine for a given jetsam priority band. + * + * Input: + * bucket_index - jetsam priority band to be sorted. + * sort_order - JETSAM_SORT_xxx from kern_memorystatus.h + * Currently sort_order is only meaningful when handling + * coalitions. + * + * Return: + * 0 on success + * non-0 on failure + */ +static int memorystatus_sort_bucket(unsigned int bucket_index, int sort_order) +{ + int coal_sort_order; + + /* + * Verify the jetsam priority + */ + if (bucket_index >= MEMSTAT_BUCKET_COUNT) { + return(EINVAL); + } + +#if DEVELOPMENT || DEBUG + if (sort_order == JETSAM_SORT_DEFAULT) { + coal_sort_order = COALITION_SORT_DEFAULT; + } else { + coal_sort_order = sort_order; /* only used for testing scenarios */ + } +#else + /* Verify default */ + if (sort_order == JETSAM_SORT_DEFAULT) { + coal_sort_order = COALITION_SORT_DEFAULT; + } else { + return(EINVAL); + } +#endif + + proc_list_lock(); + switch (bucket_index) { + case JETSAM_PRIORITY_FOREGROUND: + if (memorystatus_sort_by_largest_coalition_locked(bucket_index, coal_sort_order) == 0) { + /* + * Fall back to per process sorting when zero coalitions are found. + */ + memorystatus_sort_by_largest_process_locked(bucket_index); + } + break; + default: + memorystatus_sort_by_largest_process_locked(bucket_index); + break; + } + proc_list_unlock(); + + return(0); +} + +/* + * Sort processes by size for a single jetsam bucket. + */ + +static void memorystatus_sort_by_largest_process_locked(unsigned int bucket_index) +{ + proc_t p = NULL, insert_after_proc = NULL, max_proc = NULL; + proc_t next_p = NULL, prev_max_proc = NULL; + uint32_t pages = 0, max_pages = 0; + memstat_bucket_t *current_bucket; + + if (bucket_index >= MEMSTAT_BUCKET_COUNT) { + return; + } + + current_bucket = &memstat_bucket[bucket_index]; + + p = TAILQ_FIRST(¤t_bucket->list); + + while (p) { + memorystatus_get_task_page_counts(p->task, &pages, NULL, NULL, NULL); + max_pages = pages; + max_proc = p; + prev_max_proc = p; + + while ((next_p = TAILQ_NEXT(p, p_memstat_list)) != NULL) { + /* traversing list until we find next largest process */ + p=next_p; + memorystatus_get_task_page_counts(p->task, &pages, NULL, NULL, NULL); + if (pages > max_pages) { + max_pages = pages; + max_proc = p; + } + } + + if (prev_max_proc != max_proc) { + /* found a larger process, place it in the list */ + TAILQ_REMOVE(¤t_bucket->list, max_proc, p_memstat_list); + if (insert_after_proc == NULL) { + TAILQ_INSERT_HEAD(¤t_bucket->list, max_proc, p_memstat_list); + } else { + TAILQ_INSERT_AFTER(¤t_bucket->list, insert_after_proc, max_proc, p_memstat_list); + } + prev_max_proc = max_proc; + } + + insert_after_proc = max_proc; + + p = TAILQ_NEXT(max_proc, p_memstat_list); + } +} + +#endif /* CONFIG_JETSAM */ + static proc_t memorystatus_get_first_proc_locked(unsigned int *bucket_index, boolean_t search) { memstat_bucket_t *current_bucket; proc_t next_p; @@ -593,6 +1078,11 @@ memorystatus_init(void) kern_return_t result; int i; +#if CONFIG_FREEZE + memorystatus_freeze_pages_min = FREEZE_PAGES_MIN; + memorystatus_freeze_pages_max = FREEZE_PAGES_MAX; +#endif + nanoseconds_to_absolutetime((uint64_t)DEFERRED_IDLE_EXIT_TIME_SECS * NSEC_PER_SEC, &memorystatus_idle_delay_time); /* Init buckets */ @@ -616,11 +1106,11 @@ memorystatus_init(void) assert(freeze_threshold_percentage < 100); #if CONFIG_JETSAM + /* device tree can request to take snapshots for idle-exit kills by default */ + PE_get_default("kern.jetsam_idle_snapshot", &memorystatus_idle_snapshot, sizeof(memorystatus_idle_snapshot)); + memorystatus_delta = delta_percentage * atop_64(max_mem) / 100; -#if !LATENCY_JETSAM memorystatus_available_pages_critical_idle_offset = idle_offset_percentage * atop_64(max_mem) / 100; -#endif - memorystatus_available_pages_critical_base = (critical_threshold_percentage / delta_percentage) * memorystatus_delta; memorystatus_jetsam_snapshot_max = maxproc; @@ -631,8 +1121,21 @@ memorystatus_init(void) panic("Could not allocate memorystatus_jetsam_snapshot"); } + nanoseconds_to_absolutetime((uint64_t)JETSAM_SNAPSHOT_TIMEOUT_SECS * NSEC_PER_SEC, &memorystatus_jetsam_snapshot_timeout); + + memset(&memorystatus_at_boot_snapshot, 0, sizeof(memorystatus_jetsam_snapshot_t)); + /* No contention at this point */ memorystatus_update_levels_locked(FALSE); + + /* Jetsam Loop Detection */ + if (max_mem <= (512 * 1024 * 1024)) { + /* 512 MB devices */ + memorystatus_jld_eval_period_msecs = 8000; /* 8000 msecs == 8 second window */ + } else { + /* 1GB and larger devices */ + memorystatus_jld_eval_period_msecs = 6000; /* 6000 msecs == 6 second window */ + } #endif #if CONFIG_FREEZE @@ -651,10 +1154,29 @@ memorystatus_init(void) extern void vm_wake_compactor_swapper(void); +/* + * The jetsam no frills kill call + * Return: 0 on success + * error code on failure (EINVAL...) + */ +static int +jetsam_do_kill(proc_t p, int jetsam_flags) { + int error = 0; + error = exit1_internal(p, W_EXITCODE(0, SIGKILL), (int *)NULL, FALSE, FALSE, jetsam_flags); + return(error); +} + +/* + * Wrapper for processes exiting with memorystatus details + */ static boolean_t memorystatus_do_kill(proc_t p, uint32_t cause) { - int retval = 0; + int error = 0; + __unused pid_t victim_pid = p->p_pid; + + KERNEL_DEBUG_CONSTANT( (BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_DO_KILL)) | DBG_FUNC_START, + victim_pid, cause, vm_page_free_count, 0, 0); #if CONFIG_JETSAM && (DEVELOPMENT || DEBUG) if (memorystatus_jetsam_panic_debug & (1 << cause)) { @@ -669,16 +1191,18 @@ memorystatus_do_kill(proc_t p, uint32_t cause) { case kMemorystatusKilledVnodes: jetsam_flags |= P_JETSAM_VNODE; break; case kMemorystatusKilledVMPageShortage: jetsam_flags |= P_JETSAM_VMPAGESHORTAGE; break; case kMemorystatusKilledVMThrashing: jetsam_flags |= P_JETSAM_VMTHRASHING; break; + case kMemorystatusKilledFCThrashing: jetsam_flags |= P_JETSAM_FCTHRASHING; break; case kMemorystatusKilledPerProcessLimit: jetsam_flags |= P_JETSAM_PID; break; case kMemorystatusKilledIdleExit: jetsam_flags |= P_JETSAM_IDLEEXIT; break; } - retval = exit1_internal(p, W_EXITCODE(0, SIGKILL), (int *)NULL, FALSE, FALSE, jetsam_flags); + error = jetsam_do_kill(p, jetsam_flags); - if (COMPRESSED_PAGER_IS_ACTIVE || DEFAULT_FREEZER_COMPRESSED_PAGER_IS_ACTIVE) { - vm_wake_compactor_swapper(); - } - - return (retval == 0); + KERNEL_DEBUG_CONSTANT( (BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_DO_KILL)) | DBG_FUNC_END, + victim_pid, cause, vm_page_free_count, error, 0); + + vm_wake_compactor_swapper(); + + return (error == 0); } /* @@ -721,11 +1245,12 @@ memorystatus_perform_idle_demotion(__unused void *spare1, __unused void *spare2) if (current_time >= p->p_memstat_idledeadline) { #if DEBUG || DEVELOPMENT if (!(p->p_memstat_dirty & P_DIRTY_MARKED)) { - printf("memorystatus_perform_idle_demotion: moving process %d to idle band, but never dirtied (0x%x)!\n", p->p_pid, p->p_memstat_dirty); + printf("memorystatus_perform_idle_demotion: moving process %d [%s] to idle band, but never dirtied (0x%x)!\n", + p->p_pid, (p->p_comm ? p->p_comm : "(unknown)"), p->p_memstat_dirty); } #endif memorystatus_invalidate_idle_demotion_locked(p, TRUE); - memorystatus_update_priority_locked(p, JETSAM_PRIORITY_IDLE); + memorystatus_update_priority_locked(p, JETSAM_PRIORITY_IDLE, false); // The prior process has moved out of the demotion bucket, so grab the new head and continue p = TAILQ_FIRST(&demotion_bucket->list); @@ -746,35 +1271,53 @@ memorystatus_perform_idle_demotion(__unused void *spare1, __unused void *spare2) static void memorystatus_schedule_idle_demotion_locked(proc_t p, boolean_t set_state) { - MEMORYSTATUS_DEBUG(1, "memorystatus_schedule_idle_demotion_locked: scheduling demotion to idle band for process %d (dirty:0x%x, set_state %d, demotions %d).\n", + boolean_t present_in_deferred_bucket = FALSE; + + if (p->p_memstat_effectivepriority == JETSAM_PRIORITY_IDLE_DEFERRED) { + present_in_deferred_bucket = TRUE; + } + + MEMORYSTATUS_DEBUG(1, "memorystatus_schedule_idle_demotion_locked: scheduling demotion to idle band for pid %d (dirty:0x%x, set_state %d, demotions %d).\n", p->p_pid, p->p_memstat_dirty, set_state, memorystatus_scheduled_idle_demotions); - assert((p->p_memstat_dirty & (P_DIRTY_IDLE_EXIT_ENABLED|P_DIRTY_DEFER_IN_PROGRESS)) == (P_DIRTY_IDLE_EXIT_ENABLED|P_DIRTY_DEFER_IN_PROGRESS)); + assert((p->p_memstat_dirty & P_DIRTY_IDLE_EXIT_ENABLED) == P_DIRTY_IDLE_EXIT_ENABLED); if (set_state) { assert(p->p_memstat_idledeadline == 0); + p->p_memstat_dirty |= P_DIRTY_DEFER_IN_PROGRESS; p->p_memstat_idledeadline = mach_absolute_time() + memorystatus_idle_delay_time; } - assert(p->p_memstat_idledeadline); + assert(p->p_memstat_idledeadline); - memorystatus_scheduled_idle_demotions++; + if (present_in_deferred_bucket == FALSE) { + memorystatus_scheduled_idle_demotions++; + } } static void memorystatus_invalidate_idle_demotion_locked(proc_t p, boolean_t clear_state) { - MEMORYSTATUS_DEBUG(1, "memorystatus_invalidate_idle_demotion(): invalidating demotion to idle band for process %d (clear_state %d, demotions %d).\n", + boolean_t present_in_deferred_bucket = FALSE; + + if (p->p_memstat_effectivepriority == JETSAM_PRIORITY_IDLE_DEFERRED) { + present_in_deferred_bucket = TRUE; + assert(p->p_memstat_idledeadline); + } + + MEMORYSTATUS_DEBUG(1, "memorystatus_invalidate_idle_demotion(): invalidating demotion to idle band for pid %d (clear_state %d, demotions %d).\n", p->p_pid, clear_state, memorystatus_scheduled_idle_demotions); - assert(p->p_memstat_idledeadline); if (clear_state) { p->p_memstat_idledeadline = 0; p->p_memstat_dirty &= ~P_DIRTY_DEFER_IN_PROGRESS; } - memorystatus_scheduled_idle_demotions--; + if (present_in_deferred_bucket == TRUE) { + memorystatus_scheduled_idle_demotions--; + } + assert(memorystatus_scheduled_idle_demotions >= 0); } @@ -791,11 +1334,12 @@ memorystatus_reschedule_idle_demotion_locked(void) { proc_t p; demotion_bucket = &memstat_bucket[JETSAM_PRIORITY_IDLE_DEFERRED]; p = TAILQ_FIRST(&demotion_bucket->list); - assert(p && p->p_memstat_idledeadline); - if (memstat_idle_demotion_deadline != p->p_memstat_idledeadline){ - thread_call_enter_delayed(memorystatus_idle_demotion_call, p->p_memstat_idledeadline); - memstat_idle_demotion_deadline = p->p_memstat_idledeadline; + assert(p && p->p_memstat_idledeadline); + + if (memstat_idle_demotion_deadline != p->p_memstat_idledeadline){ + thread_call_enter_delayed(memorystatus_idle_demotion_call, p->p_memstat_idledeadline); + memstat_idle_demotion_deadline = p->p_memstat_idledeadline; } } } @@ -809,7 +1353,7 @@ memorystatus_add(proc_t p, boolean_t locked) { memstat_bucket_t *bucket; - MEMORYSTATUS_DEBUG(1, "memorystatus_list_add(): adding process %d with priority %d.\n", p->pid, priority); + MEMORYSTATUS_DEBUG(1, "memorystatus_list_add(): adding pid %d with priority %d.\n", p->p_pid, p->p_memstat_effectivepriority); if (!locked) { proc_list_lock(); @@ -822,6 +1366,10 @@ memorystatus_add(proc_t p, boolean_t locked) bucket = &memstat_bucket[p->p_memstat_effectivepriority]; + if (p->p_memstat_effectivepriority == JETSAM_PRIORITY_IDLE_DEFERRED) { + assert(bucket->count == memorystatus_scheduled_idle_demotions); + } + TAILQ_INSERT_TAIL(&bucket->list, p, p_memstat_list); bucket->count++; @@ -837,8 +1385,16 @@ exit: return 0; } +/* + * Description: + * Moves a process from one jetsam bucket to another. + * which changes the LRU position of the process. + * + * Monitors transition between buckets and if necessary + * will update cached memory limits accordingly. + */ static void -memorystatus_update_priority_locked(proc_t p, int priority) +memorystatus_update_priority_locked(proc_t p, int priority, boolean_t head_insert) { memstat_bucket_t *old_bucket, *new_bucket; @@ -849,42 +1405,151 @@ memorystatus_update_priority_locked(proc_t p, int priority) return; } - MEMORYSTATUS_DEBUG(1, "memorystatus_update_priority_locked(): setting process %d to priority %d\n", p->p_pid, priority); + MEMORYSTATUS_DEBUG(1, "memorystatus_update_priority_locked(): setting pid %d to priority %d, inserting at %s\n", + p->p_pid, priority, head_insert ? "head" : "tail"); old_bucket = &memstat_bucket[p->p_memstat_effectivepriority]; + if (p->p_memstat_effectivepriority == JETSAM_PRIORITY_IDLE_DEFERRED) { + assert(old_bucket->count == (memorystatus_scheduled_idle_demotions + 1)); + } + TAILQ_REMOVE(&old_bucket->list, p, p_memstat_list); old_bucket->count--; new_bucket = &memstat_bucket[priority]; - TAILQ_INSERT_TAIL(&new_bucket->list, p, p_memstat_list); + if (head_insert) + TAILQ_INSERT_HEAD(&new_bucket->list, p, p_memstat_list); + else + TAILQ_INSERT_TAIL(&new_bucket->list, p, p_memstat_list); new_bucket->count++; - + #if CONFIG_JETSAM - if (memorystatus_highwater_enabled && (p->p_memstat_state & P_MEMSTAT_MEMLIMIT_BACKGROUND)) { - if (((priority >= JETSAM_PRIORITY_FOREGROUND) && (p->p_memstat_effectivepriority < JETSAM_PRIORITY_FOREGROUND)) || - ((priority < JETSAM_PRIORITY_FOREGROUND) && (p->p_memstat_effectivepriority >= JETSAM_PRIORITY_FOREGROUND))) { - int32_t memlimit = (priority >= JETSAM_PRIORITY_FOREGROUND) ? -1 : p->p_memstat_memlimit; - task_set_phys_footprint_limit_internal(p->task, (memlimit > 0) ? memlimit : -1, NULL, TRUE); + if (memorystatus_highwater_enabled) { + boolean_t trigger_exception; + + /* + * If cached limit data is updated, then the limits + * will be enforced by writing to the ledgers. + */ + boolean_t ledger_update_needed = TRUE; + + /* + * No need to consider P_MEMSTAT_MEMLIMIT_BACKGROUND anymore. + * Background limits are described via the inactive limit slots. + * + * Here, we must update the cached memory limit if the task + * is transitioning between: + * active <--> inactive + * FG <--> BG + * but: + * dirty <--> clean is ignored + * + * We bypass processes that have opted into dirty tracking because + * a move between buckets does not imply a transition between the + * dirty <--> clean state. + * Setting limits on processes opted into dirty tracking is handled + * in memorystatus_dirty_set() where the transition is very clear. + */ + + if (p->p_memstat_dirty & P_DIRTY_TRACK) { + + ledger_update_needed = FALSE; + + } else if ((priority >= JETSAM_PRIORITY_FOREGROUND) && (p->p_memstat_effectivepriority < JETSAM_PRIORITY_FOREGROUND)) { + /* + * inactive --> active + * BG --> FG + * assign active state + */ + CACHE_ACTIVE_LIMITS_LOCKED(p, trigger_exception); + + } else if ((priority < JETSAM_PRIORITY_FOREGROUND) && (p->p_memstat_effectivepriority >= JETSAM_PRIORITY_FOREGROUND)) { + /* + * active --> inactive + * FG --> BG + * assign inactive state + */ + CACHE_INACTIVE_LIMITS_LOCKED(p, trigger_exception); + } else { + /* + * The transition between jetsam priority buckets apparently did + * not affect active/inactive state. + * This is not unusual... especially during startup when + * processes are getting established in their respective bands. + */ + ledger_update_needed = FALSE; + } + + /* + * Enforce the new limits by writing to the ledger + */ + if (ledger_update_needed) { + task_set_phys_footprint_limit_internal(p->task, (p->p_memstat_memlimit > 0) ? p->p_memstat_memlimit : -1, NULL, trigger_exception); + + MEMORYSTATUS_DEBUG(3, "memorystatus_update_priority_locked: new limit on pid %d (%dMB %s) priority old --> new (%d --> %d) dirty?=0x%x %s\n", + p->p_pid, (p->p_memstat_memlimit > 0 ? p->p_memstat_memlimit : -1), + (p->p_memstat_state & P_MEMSTAT_FATAL_MEMLIMIT ? "F " : "NF"), p->p_memstat_effectivepriority, priority, p->p_memstat_dirty, + (p->p_memstat_dirty ? ((p->p_memstat_dirty & P_DIRTY) ? "isdirty" : "isclean") : "")); } } -#endif + +#endif /* CONFIG_JETSAM */ p->p_memstat_effectivepriority = priority; memorystatus_check_levels_locked(); } +/* + * + * Description: Update the jetsam priority and memory limit attributes for a given process. + * + * Parameters: + * p init this process's jetsam information. + * priority The jetsam priority band + * user_data user specific data, unused by the kernel + * effective guards against race if process's update already occurred + * update_memlimit When true we know this is the init step via the posix_spawn path. + * + * memlimit_active Value in megabytes; The monitored footprint level while the + * process is active. Exceeding it may result in termination + * based on it's associated fatal flag. + * + * memlimit_active_is_fatal When a process is active and exceeds its memory footprint, + * this describes whether or not it should be immediately fatal. + * + * memlimit_inactive Value in megabytes; The monitored footprint level while the + * process is inactive. Exceeding it may result in termination + * based on it's associated fatal flag. + * + * memlimit_inactive_is_fatal When a process is inactive and exceeds its memory footprint, + * this describes whether or not it should be immediatly fatal. + * + * memlimit_background This process has a high-water-mark while in the background. + * No longer meaningful. Background limits are described via + * the inactive slots. Flag is ignored. + * + * + * Returns: 0 Success + * non-0 Failure + */ + int -memorystatus_update(proc_t p, int priority, uint64_t user_data, boolean_t effective, boolean_t update_memlimit, int32_t memlimit, boolean_t memlimit_background) +memorystatus_update(proc_t p, int priority, uint64_t user_data, boolean_t effective, boolean_t update_memlimit, + int32_t memlimit_active, boolean_t memlimit_active_is_fatal, + int32_t memlimit_inactive, boolean_t memlimit_inactive_is_fatal, + __unused boolean_t memlimit_background) { int ret; + boolean_t head_insert = false; #if !CONFIG_JETSAM -#pragma unused(update_memlimit, memlimit, memlimit_background) -#endif +#pragma unused(update_memlimit, memlimit_active, memlimit_inactive) +#pragma unused(memlimit_active_is_fatal, memlimit_inactive_is_fatal) +#endif /* !CONFIG_JETSAM */ + + MEMORYSTATUS_DEBUG(1, "memorystatus_update: changing pid %d: priority %d, user_data 0x%llx\n", p->p_pid, priority, user_data); - MEMORYSTATUS_DEBUG(1, "memorystatus_update: changing process %d: priority %d, user_data 0x%llx\n", p->p_pid, priority, user_data); - KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_UPDATE) | DBG_FUNC_START, p->p_pid, priority, user_data, effective, 0); if (priority == -1) { @@ -893,12 +1558,16 @@ memorystatus_update(proc_t p, int priority, uint64_t user_data, boolean_t effect } else if (priority == JETSAM_PRIORITY_IDLE_DEFERRED) { /* JETSAM_PRIORITY_IDLE_DEFERRED is reserved for internal use; if requested, adjust to JETSAM_PRIORITY_IDLE. */ priority = JETSAM_PRIORITY_IDLE; + } else if (priority == JETSAM_PRIORITY_IDLE_HEAD) { + /* JETSAM_PRIORITY_IDLE_HEAD inserts at the head of the idle queue */ + priority = JETSAM_PRIORITY_IDLE; + head_insert = TRUE; } else if ((priority < 0) || (priority >= MEMSTAT_BUCKET_COUNT)) { /* Sanity check */ ret = EINVAL; goto out; } - + proc_list_lock(); assert(!(p->p_memstat_state & P_MEMSTAT_INTERNAL)); @@ -906,7 +1575,16 @@ memorystatus_update(proc_t p, int priority, uint64_t user_data, boolean_t effect if (effective && (p->p_memstat_state & P_MEMSTAT_PRIORITYUPDATED)) { ret = EALREADY; proc_list_unlock(); - MEMORYSTATUS_DEBUG(1, "memorystatus_update: effective change specified for pid %d, but change already occurred.\n", pid); + MEMORYSTATUS_DEBUG(1, "memorystatus_update: effective change specified for pid %d, but change already occurred.\n", p->p_pid); + goto out; + } + + if ((p->p_memstat_state & P_MEMSTAT_TERMINATED) || ((p->p_listflag & P_LIST_EXITED) != 0)) { + /* + * This could happen when a process calling posix_spawn() is exiting on the jetsam thread. + */ + ret = EBUSY; + proc_list_unlock(); goto out; } @@ -916,20 +1594,118 @@ memorystatus_update(proc_t p, int priority, uint64_t user_data, boolean_t effect #if CONFIG_JETSAM if (update_memlimit) { - p->p_memstat_memlimit = memlimit; + boolean_t trigger_exception; + + /* + * Posix_spawn'd processes come through this path to instantiate ledger limits. + * Forked processes do not come through this path, so no ledger limits exist. + * (That's why forked processes can consume unlimited memory.) + */ + + MEMORYSTATUS_DEBUG(3, "memorystatus_update(enter): pid %d, priority %d, dirty=0x%x, Active(%dMB %s), Inactive(%dMB, %s)\n", + p->p_pid, priority, p->p_memstat_dirty, + memlimit_active, (memlimit_active_is_fatal ? "F " : "NF"), + memlimit_inactive, (memlimit_inactive_is_fatal ? "F " : "NF")); + if (memlimit_background) { - /* Will be set as priority is updated */ - p->p_memstat_state |= P_MEMSTAT_MEMLIMIT_BACKGROUND; - } else { - /* Otherwise, apply now */ - if (memorystatus_highwater_enabled) { - task_set_phys_footprint_limit_internal(p->task, (memlimit > 0) ? memlimit : -1, NULL, TRUE); + + /* + * With 2-level HWM support, we no longer honor P_MEMSTAT_MEMLIMIT_BACKGROUND. + * Background limits are described via the inactive limit slots. + */ + + // p->p_memstat_state |= P_MEMSTAT_MEMLIMIT_BACKGROUND; + +#if DEVELOPMENT || DEBUG + printf("memorystatus_update: WARNING %s[%d] set unused flag P_MEMSTAT_MEMLIMIT_BACKGROUND [A==%dMB %s] [IA==%dMB %s]\n", + (p->p_comm ? p->p_comm : "unknown"), p->p_pid, + memlimit_active, (memlimit_active_is_fatal ? "F " : "NF"), + memlimit_inactive, (memlimit_inactive_is_fatal ? "F " : "NF")); +#endif /* DEVELOPMENT || DEBUG */ + } + + if (memlimit_active <= 0) { + /* + * This process will have a system_wide task limit when active. + * System_wide task limit is always fatal. + * It's quite common to see non-fatal flag passed in here. + * It's not an error, we just ignore it. + */ + + /* + * For backward compatibility with some unexplained launchd behavior, + * we allow a zero sized limit. But we still enforce system_wide limit + * when written to the ledgers. + */ + + if (memlimit_active < 0) { + memlimit_active = -1; /* enforces system_wide task limit */ } + memlimit_active_is_fatal = TRUE; + } + + if (memlimit_inactive <= 0) { + /* + * This process will have a system_wide task limit when inactive. + * System_wide task limit is always fatal. + */ + + memlimit_inactive = -1; + memlimit_inactive_is_fatal = TRUE; + } + + /* + * Initialize the active limit variants for this process. + */ + SET_ACTIVE_LIMITS_LOCKED(p, memlimit_active, memlimit_active_is_fatal); + + /* + * Initialize the inactive limit variants for this process. + */ + SET_INACTIVE_LIMITS_LOCKED(p, memlimit_inactive, memlimit_inactive_is_fatal); + + /* + * Initialize the cached limits for target process. + * When the target process is dirty tracked, it's typically + * in a clean state. Non dirty tracked processes are + * typically active (Foreground or above). + * But just in case, we don't make assumptions... + */ + + if (proc_jetsam_state_is_active_locked(p) == TRUE) { + CACHE_ACTIVE_LIMITS_LOCKED(p, trigger_exception); + } else { + CACHE_INACTIVE_LIMITS_LOCKED(p, trigger_exception); + } + + /* + * Enforce the cached limit by writing to the ledger. + */ + if (memorystatus_highwater_enabled) { + /* apply now */ + assert(trigger_exception == TRUE); + task_set_phys_footprint_limit_internal(p->task, ((p->p_memstat_memlimit > 0) ? p->p_memstat_memlimit : -1), NULL, trigger_exception); + + MEMORYSTATUS_DEBUG(3, "memorystatus_update: init: limit on pid %d (%dMB %s) targeting priority(%d) dirty?=0x%x %s\n", + p->p_pid, (p->p_memstat_memlimit > 0 ? p->p_memstat_memlimit : -1), + (p->p_memstat_state & P_MEMSTAT_FATAL_MEMLIMIT ? "F " : "NF"), priority, p->p_memstat_dirty, + (p->p_memstat_dirty ? ((p->p_memstat_dirty & P_DIRTY) ? "isdirty" : "isclean") : "")); } } -#endif +#endif /* CONFIG_JETSAM */ - memorystatus_update_priority_locked(p, priority); + /* + * We can't add to the JETSAM_PRIORITY_IDLE_DEFERRED bucket here. + * But, we could be removing it from the bucket. + * Check and take appropriate steps if so. + */ + + if (p->p_memstat_effectivepriority == JETSAM_PRIORITY_IDLE_DEFERRED) { + + memorystatus_invalidate_idle_demotion_locked(p, TRUE); + } + + memorystatus_update_priority_locked(p, priority, head_insert); proc_list_unlock(); ret = 0; @@ -946,15 +1722,19 @@ memorystatus_remove(proc_t p, boolean_t locked) int ret; memstat_bucket_t *bucket; - MEMORYSTATUS_DEBUG(1, "memorystatus_list_remove: removing process %d\n", pid); + MEMORYSTATUS_DEBUG(1, "memorystatus_list_remove: removing pid %d\n", p->p_pid); if (!locked) { proc_list_lock(); } assert(!(p->p_memstat_state & P_MEMSTAT_INTERNAL)); - + bucket = &memstat_bucket[p->p_memstat_effectivepriority]; + if (p->p_memstat_effectivepriority == JETSAM_PRIORITY_IDLE_DEFERRED) { + assert(bucket->count == memorystatus_scheduled_idle_demotions); + } + TAILQ_REMOVE(&bucket->list, p, p_memstat_list); bucket->count--; @@ -992,32 +1772,46 @@ memorystatus_remove(proc_t p, boolean_t locked) return ret; } -static boolean_t +/* + * Validate dirty tracking flags with process state. + * + * Return: + * 0 on success + * non-0 on failure + */ + +static int memorystatus_validate_track_flags(struct proc *target_p, uint32_t pcontrol) { /* See that the process isn't marked for termination */ if (target_p->p_memstat_dirty & P_DIRTY_TERMINATED) { - return FALSE; + return EBUSY; } /* Idle exit requires that process be tracked */ if ((pcontrol & PROC_DIRTY_ALLOW_IDLE_EXIT) && !(pcontrol & PROC_DIRTY_TRACK)) { - return FALSE; + return EINVAL; + } + + /* 'Launch in progress' tracking requires that process have enabled dirty tracking too. */ + if ((pcontrol & PROC_DIRTY_LAUNCH_IN_PROGRESS) && + !(pcontrol & PROC_DIRTY_TRACK)) { + return EINVAL; } /* Deferral is only relevant if idle exit is specified */ if ((pcontrol & PROC_DIRTY_DEFER) && !(pcontrol & PROC_DIRTY_ALLOWS_IDLE_EXIT)) { - return FALSE; + return EINVAL; } - return TRUE; + return(0); } static void memorystatus_update_idle_priority_locked(proc_t p) { int32_t priority; - + MEMORYSTATUS_DEBUG(1, "memorystatus_update_idle_priority_locked(): pid %d dirty 0x%X\n", p->p_pid, p->p_memstat_dirty); if ((p->p_memstat_dirty & (P_DIRTY_IDLE_EXIT_ENABLED|P_DIRTY_IS_DIRTY)) == P_DIRTY_IDLE_EXIT_ENABLED) { @@ -1026,7 +1820,9 @@ memorystatus_update_idle_priority_locked(proc_t p) { priority = p->p_memstat_requestedpriority; } - memorystatus_update_priority_locked(p, priority); + if (priority != p->p_memstat_effectivepriority) { + memorystatus_update_priority_locked(p, priority, false); + } } /* @@ -1050,19 +1846,32 @@ int memorystatus_dirty_track(proc_t p, uint32_t pcontrol) { unsigned int old_dirty; boolean_t reschedule = FALSE; - int ret; + boolean_t already_deferred = FALSE; + boolean_t defer_now = FALSE; + int ret = 0; + KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_DIRTY_TRACK), + p->p_pid, p->p_memstat_dirty, pcontrol, 0, 0); + proc_list_lock(); + if ((p->p_listflag & P_LIST_EXITED) != 0) { + /* + * Process is on its way out. + */ + ret = EBUSY; + goto exit; + } + if (p->p_memstat_state & P_MEMSTAT_INTERNAL) { ret = EPERM; goto exit; } - if (!memorystatus_validate_track_flags(p, pcontrol)) { - ret = EINVAL; + if ((ret = memorystatus_validate_track_flags(p, pcontrol)) != 0) { + /* error */ goto exit; - } + } old_dirty = p->p_memstat_dirty; @@ -1075,28 +1884,82 @@ memorystatus_dirty_track(proc_t p, uint32_t pcontrol) { p->p_memstat_dirty |= P_DIRTY_ALLOW_IDLE_EXIT; } + if (pcontrol & PROC_DIRTY_LAUNCH_IN_PROGRESS) { + p->p_memstat_dirty |= P_DIRTY_LAUNCH_IN_PROGRESS; + } + + if (old_dirty & P_DIRTY_DEFER_IN_PROGRESS) { + already_deferred = TRUE; + } + /* This can be set and cleared exactly once. */ - if ((pcontrol & PROC_DIRTY_DEFER) && !(old_dirty & P_DIRTY_DEFER)) { - p->p_memstat_dirty |= (P_DIRTY_DEFER|P_DIRTY_DEFER_IN_PROGRESS); - } else { - p->p_memstat_dirty &= ~P_DIRTY_DEFER_IN_PROGRESS; + if (pcontrol & PROC_DIRTY_DEFER) { + + if ( !(old_dirty & P_DIRTY_DEFER)) { + p->p_memstat_dirty |= P_DIRTY_DEFER; + } + + defer_now = TRUE; } - MEMORYSTATUS_DEBUG(1, "memorystatus_on_track_dirty(): set idle-exit %s / deferred %s / dirty %s for process %d\n", + MEMORYSTATUS_DEBUG(1, "memorystatus_on_track_dirty(): set idle-exit %s / defer %s / dirty %s for pid %d\n", ((p->p_memstat_dirty & P_DIRTY_IDLE_EXIT_ENABLED) == P_DIRTY_IDLE_EXIT_ENABLED) ? "Y" : "N", - p->p_memstat_dirty & P_DIRTY_DEFER_IN_PROGRESS ? "Y" : "N", + defer_now ? "Y" : "N", p->p_memstat_dirty & P_DIRTY ? "Y" : "N", p->p_pid); /* Kick off or invalidate the idle exit deferment if there's a state transition. */ if (!(p->p_memstat_dirty & P_DIRTY_IS_DIRTY)) { if (((p->p_memstat_dirty & P_DIRTY_IDLE_EXIT_ENABLED) == P_DIRTY_IDLE_EXIT_ENABLED) && - (p->p_memstat_dirty & P_DIRTY_DEFER_IN_PROGRESS) && !(old_dirty & P_DIRTY_DEFER_IN_PROGRESS)) { + defer_now && !already_deferred) { + + /* + * Request to defer a clean process that's idle-exit enabled + * and not already in the jetsam deferred band. + */ memorystatus_schedule_idle_demotion_locked(p, TRUE); reschedule = TRUE; - } else if (!(p->p_memstat_dirty & P_DIRTY_DEFER_IN_PROGRESS) && (old_dirty & P_DIRTY_DEFER_IN_PROGRESS)) { + + } else if (!defer_now && already_deferred) { + + /* + * Either the process is no longer idle-exit enabled OR + * there's a request to cancel a currently active deferral. + */ + memorystatus_invalidate_idle_demotion_locked(p, TRUE); + reschedule = TRUE; + } + } else { + + /* + * We are trying to operate on a dirty process. Dirty processes have to + * be removed from the deferred band. The question is do we reset the + * deferred state or not? + * + * This could be a legal request like: + * - this process had opted into the JETSAM_DEFERRED band + * - but it's now dirty and requests to opt out. + * In this case, we remove the process from the band and reset its + * state too. It'll opt back in properly when needed. + * + * OR, this request could be a user-space bug. E.g.: + * - this process had opted into the JETSAM_DEFERRED band when clean + * - and, then issues another request to again put it into the band except + * this time the process is dirty. + * The process going dirty, as a transition in memorystatus_dirty_set(), will pull the process out of + * the deferred band with its state intact. So our request below is no-op. + * But we do it here anyways for coverage. + * + * memorystatus_update_idle_priority_locked() + * single-mindedly treats a dirty process as "cannot be in the deferred band". + */ + + if (!defer_now && already_deferred) { memorystatus_invalidate_idle_demotion_locked(p, TRUE); reschedule = TRUE; + } else { + memorystatus_invalidate_idle_demotion_locked(p, FALSE); + reschedule = TRUE; } } @@ -1123,9 +1986,19 @@ memorystatus_dirty_set(proc_t p, boolean_t self, uint32_t pcontrol) { boolean_t now_dirty = FALSE; MEMORYSTATUS_DEBUG(1, "memorystatus_dirty_set(): %d %d 0x%x 0x%x\n", self, p->p_pid, pcontrol, p->p_memstat_dirty); + + KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_DIRTY_SET), p->p_pid, self, pcontrol, 0, 0); proc_list_lock(); + if ((p->p_listflag & P_LIST_EXITED) != 0) { + /* + * Process is on its way out. + */ + ret = EBUSY; + goto exit; + } + if (p->p_memstat_state & P_MEMSTAT_INTERNAL) { ret = EPERM; goto exit; @@ -1151,7 +2024,7 @@ memorystatus_dirty_set(proc_t p, boolean_t self, uint32_t pcontrol) { memorystatus_dirty_count++; ret = 0; } else if ((pcontrol == 0) && (p->p_memstat_dirty & flag)) { - if ((flag == P_DIRTY_SHUTDOWN) && (!p->p_memstat_dirty & P_DIRTY)) { + if ((flag == P_DIRTY_SHUTDOWN) && (!(p->p_memstat_dirty & P_DIRTY))) { /* Clearing the dirty shutdown flag, and the process is otherwise clean - kill */ p->p_memstat_dirty |= P_DIRTY_TERMINATED; kill = true; @@ -1171,7 +2044,7 @@ memorystatus_dirty_set(proc_t p, boolean_t self, uint32_t pcontrol) { if (ret != 0) { goto exit; } - + if (p->p_memstat_dirty & P_DIRTY_IS_DIRTY) now_dirty = TRUE; @@ -1181,31 +2054,130 @@ memorystatus_dirty_set(proc_t p, boolean_t self, uint32_t pcontrol) { /* Manage idle exit deferral, if applied */ if ((p->p_memstat_dirty & (P_DIRTY_IDLE_EXIT_ENABLED|P_DIRTY_DEFER_IN_PROGRESS)) == (P_DIRTY_IDLE_EXIT_ENABLED|P_DIRTY_DEFER_IN_PROGRESS)) { + + /* + * P_DIRTY_DEFER_IN_PROGRESS means the process is in the deferred band OR it might be heading back + * there once it's clean again and has some protection window left. + */ + if (p->p_memstat_dirty & P_DIRTY_IS_DIRTY) { + /* + * New dirty process i.e. "was_dirty == FALSE && now_dirty == TRUE" + * + * The process will move from the deferred band to its higher requested + * jetsam band. But we don't clear its state i.e. we want to remember that + * this process was part of the "deferred" band and will return to it. + * + * This way, we don't let it age beyond the protection + * window when it returns to "clean". All the while giving + * it a chance to perform its work while "dirty". + * + */ memorystatus_invalidate_idle_demotion_locked(p, FALSE); reschedule = TRUE; } else { - /* We evaluate lazily, so reset the idle-deadline if it's expired by the time the process becomes clean. */ + + /* + * Process is back from "dirty" to "clean". + * + * Is its timer up OR does it still have some protection + * window left? + */ + if (mach_absolute_time() >= p->p_memstat_idledeadline) { - p->p_memstat_idledeadline = 0; - p->p_memstat_dirty &= ~P_DIRTY_DEFER_IN_PROGRESS; + /* + * The process' deadline has expired. It currently + * does not reside in the DEFERRED bucket. + * + * It's on its way to the JETSAM_PRIORITY_IDLE + * bucket via memorystatus_update_idle_priority_locked() + * below. + + * So all we need to do is reset all the state on the + * process that's related to the DEFERRED bucket i.e. + * the DIRTY_DEFER_IN_PROGRESS flag and the timer deadline. + * + */ + + memorystatus_invalidate_idle_demotion_locked(p, TRUE); + reschedule = TRUE; } else { + /* + * It still has some protection window left and so + * we just re-arm the timer without modifying any + * state on the process. + */ memorystatus_schedule_idle_demotion_locked(p, FALSE); reschedule = TRUE; } } } - + memorystatus_update_idle_priority_locked(p); + +#if CONFIG_JETSAM + if (memorystatus_highwater_enabled) { + boolean_t trigger_exception; + /* + * We are in this path because this process transitioned between + * dirty <--> clean state. Update the cached memory limits. + */ + + if (proc_jetsam_state_is_active_locked(p) == TRUE) { + /* + * process is dirty + */ + CACHE_ACTIVE_LIMITS_LOCKED(p, trigger_exception); + } else { + /* + * process is clean + */ + CACHE_INACTIVE_LIMITS_LOCKED(p, trigger_exception); + } + + /* + * Enforce the new limits by writing to the ledger. + * + * This is a hot path and holding the proc_list_lock while writing to the ledgers, + * (where the task lock is taken) is bad. So, we temporarily drop the proc_list_lock. + * We aren't traversing the jetsam bucket list here, so we should be safe. + * See rdar://21394491. + */ + + if (proc_ref_locked(p) == p) { + int ledger_limit; + if (p->p_memstat_memlimit > 0) { + ledger_limit = p->p_memstat_memlimit; + } else { + ledger_limit = -1; + } + proc_list_unlock(); + task_set_phys_footprint_limit_internal(p->task, ledger_limit, NULL, trigger_exception); + proc_list_lock(); + proc_rele_locked(p); + + MEMORYSTATUS_DEBUG(3, "memorystatus_dirty_set: new limit on pid %d (%dMB %s) priority(%d) dirty?=0x%x %s\n", + p->p_pid, (p->p_memstat_memlimit > 0 ? p->p_memstat_memlimit : -1), + (p->p_memstat_state & P_MEMSTAT_FATAL_MEMLIMIT ? "F " : "NF"), p->p_memstat_effectivepriority, p->p_memstat_dirty, + (p->p_memstat_dirty ? ((p->p_memstat_dirty & P_DIRTY) ? "isdirty" : "isclean") : "")); + } + + } +#endif /* CONFIG_JETSAM */ /* If the deferral state changed, reschedule the demotion timer */ if (reschedule) { memorystatus_reschedule_idle_demotion_locked(); } } - + if (kill) { - psignal(p, SIGKILL); + if (proc_ref_locked(p) == p) { + proc_list_unlock(); + psignal(p, SIGKILL); + proc_list_lock(); + proc_rele_locked(p); + } } exit: @@ -1214,6 +2186,65 @@ exit: return ret; } +int +memorystatus_dirty_clear(proc_t p, uint32_t pcontrol) { + + int ret = 0; + + MEMORYSTATUS_DEBUG(1, "memorystatus_dirty_clear(): %d 0x%x 0x%x\n", p->p_pid, pcontrol, p->p_memstat_dirty); + + KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_DIRTY_CLEAR), p->p_pid, pcontrol, 0, 0, 0); + + proc_list_lock(); + + if ((p->p_listflag & P_LIST_EXITED) != 0) { + /* + * Process is on its way out. + */ + ret = EBUSY; + goto exit; + } + + if (p->p_memstat_state & P_MEMSTAT_INTERNAL) { + ret = EPERM; + goto exit; + } + + if (!(p->p_memstat_dirty & P_DIRTY_TRACK)) { + /* Dirty tracking not enabled */ + ret = EINVAL; + goto exit; + } + + if (!pcontrol || (pcontrol & (PROC_DIRTY_LAUNCH_IN_PROGRESS | PROC_DIRTY_DEFER)) == 0) { + ret = EINVAL; + goto exit; + } + + if (pcontrol & PROC_DIRTY_LAUNCH_IN_PROGRESS) { + p->p_memstat_dirty &= ~P_DIRTY_LAUNCH_IN_PROGRESS; + } + + /* This can be set and cleared exactly once. */ + if (pcontrol & PROC_DIRTY_DEFER) { + + if (p->p_memstat_dirty & P_DIRTY_DEFER) { + + p->p_memstat_dirty &= ~P_DIRTY_DEFER; + + memorystatus_invalidate_idle_demotion_locked(p, TRUE); + memorystatus_update_idle_priority_locked(p); + memorystatus_reschedule_idle_demotion_locked(); + } + } + + ret = 0; +exit: + proc_list_unlock(); + + return ret; +} + int memorystatus_dirty_get(proc_t p) { int ret = 0; @@ -1228,6 +2259,9 @@ memorystatus_dirty_get(proc_t p) { if (p->p_memstat_dirty & P_DIRTY) { ret |= PROC_DIRTY_IS_DIRTY; } + if (p->p_memstat_dirty & P_DIRTY_LAUNCH_IN_PROGRESS) { + ret |= PROC_DIRTY_LAUNCH_IS_IN_PROGRESS; + } } proc_list_unlock(); @@ -1261,7 +2295,7 @@ memorystatus_on_suspend(proc_t p) { #if CONFIG_FREEZE uint32_t pages; - memorystatus_get_task_page_counts(p->task, &pages, NULL); + memorystatus_get_task_page_counts(p->task, &pages, NULL, NULL, NULL); #endif proc_list_lock(); #if CONFIG_FREEZE @@ -1392,10 +2426,14 @@ kill_idle_exit_proc(void) } #endif +#if CONFIG_JETSAM static void memorystatus_thread_wake(void) { thread_wakeup((event_t)&memorystatus_wakeup); } +#endif /* CONFIG_JETSAM */ + +extern void vm_pressure_response(void); static int memorystatus_thread_block(uint32_t interval_ms, thread_continue_t continuation) @@ -1409,16 +2447,29 @@ memorystatus_thread_block(uint32_t interval_ms, thread_continue_t continuation) return thread_block(continuation); } -extern boolean_t vm_compressor_thrashing_detected; -extern uint64_t vm_compressor_total_compressions(void); - static void memorystatus_thread(void *param __unused, wait_result_t wr __unused) { static boolean_t is_vm_privileged = FALSE; + #if CONFIG_JETSAM boolean_t post_snapshot = FALSE; uint32_t errors = 0; + uint32_t hwm_kill = 0; + boolean_t sort_flag = TRUE; + + /* Jetsam Loop Detection - locals */ + memstat_bucket_t *bucket; + int jld_bucket_count = 0; + struct timeval jld_now_tstamp = {0,0}; + uint64_t jld_now_msecs = 0; + + /* Jetsam Loop Detection - statics */ + static uint64_t jld_timestamp_msecs = 0; + static int jld_idle_kill_candidates = 0; /* Number of available processes in band 0,1 at start */ + static int jld_idle_kills = 0; /* Number of procs killed during eval period */ + static int jld_eval_aggressive_count = 0; /* Bumps the max priority in aggressive loop */ + static int32_t jld_priority_band_max = JETSAM_PRIORITY_UI_SUPPORT; #endif if (is_vm_privileged == FALSE) { @@ -1429,120 +2480,247 @@ memorystatus_thread(void *param __unused, wait_result_t wr __unused) thread_wire(host_priv_self(), current_thread(), TRUE); is_vm_privileged = TRUE; + if (vm_restricted_to_single_processor == TRUE) + thread_vm_bind_group_add(); + memorystatus_thread_block(0, memorystatus_thread); } #if CONFIG_JETSAM KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_SCAN) | DBG_FUNC_START, - memorystatus_available_pages, 0, 0, 0, 0); + memorystatus_available_pages, memorystatus_jld_enabled, memorystatus_jld_eval_period_msecs, memorystatus_jld_eval_aggressive_count,0); - uint32_t cause = vm_compressor_thrashing_detected ? kMemorystatusKilledVMThrashing : kMemorystatusKilledVMPageShortage; - - /* Jetsam aware version. + /* + * Jetsam aware version. * - * If woken under pressure, go down the path of killing: + * The VM pressure notification thread is working it's way through clients in parallel. * - * - processes exceeding their highwater mark if no clean victims available - * - the least recently used process if no highwater mark victims available + * So, while the pressure notification thread is targeting processes in order of + * increasing jetsam priority, we can hopefully reduce / stop it's work by killing + * any processes that have exceeded their highwater mark. + * + * If we run out of HWM processes and our available pages drops below the critical threshold, then, + * we target the least recently used process in order of increasing jetsam priority (exception: the FG band). */ -#if !LATENCY_JETSAM - while (vm_compressor_thrashing_detected || memorystatus_available_pages <= memorystatus_available_pages_critical) { -#else - while (kill_under_pressure) { - const uint32_t SNAPSHOT_WAIT_TIMEOUT_MS = 100; - wait_result_t wait_result; -#endif + while (is_thrashing(kill_under_pressure_cause) || + memorystatus_available_pages <= memorystatus_available_pages_pressure) { boolean_t killed; int32_t priority; + uint32_t cause; + + if (kill_under_pressure_cause) { + cause = kill_under_pressure_cause; + } else { + cause = kMemorystatusKilledVMPageShortage; + } #if LEGACY_HIWATER /* Highwater */ killed = memorystatus_kill_hiwat_proc(&errors); if (killed) { + hwm_kill++; post_snapshot = TRUE; goto done; + } else { + memorystatus_hwm_candidates = FALSE; + } + + /* No highwater processes to kill. Continue or stop for now? */ + if (!is_thrashing(kill_under_pressure_cause) && + (memorystatus_available_pages > memorystatus_available_pages_critical)) { + /* + * We are _not_ out of pressure but we are above the critical threshold and there's: + * - no compressor thrashing + * - no more HWM processes left. + * For now, don't kill any other processes. + */ + + if (hwm_kill == 0) { + memorystatus_thread_wasted_wakeup++; + } + + break; } #endif + if (memorystatus_jld_enabled == TRUE) { + + /* + * Jetsam Loop Detection: attempt to detect + * rapid daemon relaunches in the lower bands. + */ + + microuptime(&jld_now_tstamp); + + /* + * Ignore usecs in this calculation. + * msecs granularity is close enough. + */ + jld_now_msecs = (jld_now_tstamp.tv_sec * 1000); + + proc_list_lock(); + bucket = &memstat_bucket[JETSAM_PRIORITY_IDLE]; + jld_bucket_count = bucket->count; + bucket = &memstat_bucket[JETSAM_PRIORITY_IDLE_DEFERRED]; + jld_bucket_count += bucket->count; + proc_list_unlock(); + + /* + * memorystatus_jld_eval_period_msecs is a tunable + * memorystatus_jld_eval_aggressive_count is a tunable + * memorystatus_jld_eval_aggressive_priority_band_max is a tunable + */ + if ( (jld_bucket_count == 0) || + (jld_now_msecs > (jld_timestamp_msecs + memorystatus_jld_eval_period_msecs))) { + + /* + * Refresh evaluation parameters + */ + jld_timestamp_msecs = jld_now_msecs; + jld_idle_kill_candidates = jld_bucket_count; + jld_idle_kills = 0; + jld_eval_aggressive_count = 0; + jld_priority_band_max = JETSAM_PRIORITY_UI_SUPPORT; + } + + if (jld_idle_kills > jld_idle_kill_candidates) { + jld_eval_aggressive_count++; + if (jld_eval_aggressive_count > memorystatus_jld_eval_aggressive_count) { + /* + * Bump up the jetsam priority limit (eg: the bucket index) + * Enforce bucket index sanity. + */ + if ((memorystatus_jld_eval_aggressive_priority_band_max < 0) || + (memorystatus_jld_eval_aggressive_priority_band_max >= MEMSTAT_BUCKET_COUNT)) { + /* + * Do nothing. Stick with the default level. + */ + } else { + jld_priority_band_max = memorystatus_jld_eval_aggressive_priority_band_max; + } + } + + killed = memorystatus_kill_top_process_aggressive( + TRUE, + kMemorystatusKilledVMThrashing, + jld_eval_aggressive_count, + jld_priority_band_max, + &errors); + + + if (killed) { + /* Always generate logs after aggressive kill */ + post_snapshot = TRUE; + goto done; + } + } + } /* LRU */ - killed = memorystatus_kill_top_process(TRUE, cause, &priority, &errors); + killed = memorystatus_kill_top_process(TRUE, sort_flag, cause, &priority, &errors); + sort_flag = FALSE; + if (killed) { - if (!kill_under_pressure && (priority != JETSAM_PRIORITY_IDLE)) { - /* Don't generate logs for steady-state idle-exit kills */ + /* + * Don't generate logs for steady-state idle-exit kills, + * unless it is overridden for debug or by the device + * tree. + */ + if ((priority != JETSAM_PRIORITY_IDLE) || memorystatus_idle_snapshot) { post_snapshot = TRUE; } + + /* Jetsam Loop Detection */ + if (memorystatus_jld_enabled == TRUE) { + if ((priority == JETSAM_PRIORITY_IDLE) || (priority == JETSAM_PRIORITY_IDLE_DEFERRED)) { + jld_idle_kills++; + } else { + /* + * We've reached into bands beyond idle deferred. + * We make no attempt to monitor them + */ + } + } goto done; } - - /* Under pressure and unable to kill a process - panic */ - panic("memorystatus_jetsam_thread: no victim! available pages:%d\n", memorystatus_available_pages); + + if (memorystatus_available_pages <= memorystatus_available_pages_critical) { + /* Under pressure and unable to kill a process - panic */ + panic("memorystatus_jetsam_thread: no victim! available pages:%d\n", memorystatus_available_pages); + } done: - kill_under_pressure = FALSE; - vm_compressor_thrashing_detected = FALSE; - -#if LATENCY_JETSAM - KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_LATENCY_COALESCE) | DBG_FUNC_START, - memorystatus_available_pages, 0, 0, 0, 0); - thread_wakeup((event_t)&latency_jetsam_wakeup); - /* - * Coalesce snapshot reports in the face of repeated jetsams by blocking here with a timeout. - * If the wait expires, issue the note. + + /* + * We do not want to over-kill when thrashing has been detected. + * To avoid that, we reset the flag here and notify the + * compressor. */ - wait_result = memorystatus_thread_block(SNAPSHOT_WAIT_TIMEOUT_MS, THREAD_CONTINUE_NULL); - KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_LATENCY_COALESCE) | DBG_FUNC_END, - memorystatus_available_pages, 0, 0, 0, 0); - if (wait_result != THREAD_AWAKENED) { - /* Catch-all */ - break; + if (is_thrashing(kill_under_pressure_cause)) { + kill_under_pressure_cause = 0; + vm_thrashing_jetsam_done(); } -#endif } - + + kill_under_pressure_cause = 0; + if (errors) { memorystatus_clear_errors(); } #if VM_PRESSURE_EVENTS - memorystatus_update_vm_pressure(TRUE); + /* + * LD: We used to target the foreground process first and foremost here. + * Now, we target all processes, starting from the non-suspended, background + * processes first. We will target foreground too. + * + * memorystatus_update_vm_pressure(TRUE); + */ + //vm_pressure_response(); #endif if (post_snapshot) { size_t snapshot_size = sizeof(memorystatus_jetsam_snapshot_t) + sizeof(memorystatus_jetsam_snapshot_entry_t) * (memorystatus_jetsam_snapshot_count); - memorystatus_jetsam_snapshot->notification_time = mach_absolute_time(); - memorystatus_send_note(kMemorystatusSnapshotNote, &snapshot_size, sizeof(snapshot_size)); + uint64_t timestamp_now = mach_absolute_time(); + memorystatus_jetsam_snapshot->notification_time = timestamp_now; + if (memorystatus_jetsam_snapshot_last_timestamp == 0 || + timestamp_now > memorystatus_jetsam_snapshot_last_timestamp + memorystatus_jetsam_snapshot_timeout) { + int ret = memorystatus_send_note(kMemorystatusSnapshotNote, &snapshot_size, sizeof(snapshot_size)); + if (!ret) { + proc_list_lock(); + memorystatus_jetsam_snapshot_last_timestamp = timestamp_now; + proc_list_unlock(); + } + } } - + KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_SCAN) | DBG_FUNC_END, memorystatus_available_pages, 0, 0, 0, 0); #else /* CONFIG_JETSAM */ - /* Simple version. - * - * Jetsam not enabled, so just kill the first suitable clean process - * and sleep. + /* + * Jetsam not enabled */ - if (kill_idle_exit) { - kill_idle_exit_proc(); - kill_idle_exit = FALSE; - } - #endif /* CONFIG_JETSAM */ memorystatus_thread_block(0, memorystatus_thread); } #if !CONFIG_JETSAM +/* + * Returns TRUE: + * when an idle-exitable proc was killed + * Returns FALSE: + * when there are no more idle-exitable procs found + * when the attempt to kill an idle-exitable proc failed + */ boolean_t memorystatus_idle_exit_from_VM(void) { - kill_idle_exit = TRUE; - memorystatus_thread_wake(); - return TRUE; + return(kill_idle_exit_proc()); } -#endif +#endif /* !CONFIG_JETSAM */ #if CONFIG_JETSAM @@ -1551,41 +2729,130 @@ boolean_t memorystatus_idle_exit_from_VM(void) { * (dirty pages + IOKit mappings) * * This is invoked for both advisory, non-fatal per-task high watermarks, - * as well as the fatal system-wide task memory limit. + * as well as the fatal task memory limits. */ void memorystatus_on_ledger_footprint_exceeded(boolean_t warning, const int max_footprint_mb) { + boolean_t is_active; + boolean_t is_fatal; + proc_t p = current_proc(); - - printf("process %d (%s) %s physical memory footprint limit of %d MB\n", - p->p_pid, p->p_comm, - warning ? "approaching" : "exceeded", - max_footprint_mb); + + proc_list_lock(); + + is_active = proc_jetsam_state_is_active_locked(p); + is_fatal = (p->p_memstat_state & P_MEMSTAT_FATAL_MEMLIMIT); + + if (warning == FALSE) { + /* + * We only want the EXC_RESOURCE to trigger once per lifetime + * of the active/inactive limit state. So, here, we detect the + * active/inactive state of the process and mark the + * state as exception has been triggered. + */ + if (is_active == TRUE) { + /* + * turn off exceptions for active state + */ + p->p_memstat_state |= P_MEMSTAT_MEMLIMIT_ACTIVE_EXC_TRIGGERED; + } else { + /* + * turn off exceptions for inactive state + */ + p->p_memstat_state |= P_MEMSTAT_MEMLIMIT_INACTIVE_EXC_TRIGGERED; + } + + /* + * Soft memory limit is a non-fatal high-water-mark + * Hard memory limit is a fatal custom-task-limit or system-wide per-task memory limit. + */ + printf("process %d (%s) exceeded physical memory footprint, the %s%sMemoryLimit of %d MB\n", + p->p_pid, p->p_comm, (is_active ? "Active" : "Inactive"), + (is_fatal ? "Hard" : "Soft"), max_footprint_mb); + + } + + proc_list_unlock(); #if VM_PRESSURE_EVENTS if (warning == TRUE) { - if (memorystatus_warn_process(p->p_pid) != TRUE) { + if (memorystatus_warn_process(p->p_pid, TRUE /* critical? */) != TRUE) { /* Print warning, since it's possible that task has not registered for pressure notifications */ - printf("task_exceeded_footprint: failed to warn the current task (exiting?).\n"); + printf("task_exceeded_footprint: failed to warn the current task (exiting, or no handler registered?).\n"); } return; } #endif /* VM_PRESSURE_EVENTS */ - if (p->p_memstat_memlimit <= 0) { + if (is_fatal) { /* - * If this process has no high watermark, then we have been invoked because the task - * has violated the system-wide per-task memory limit. + * If this process has no high watermark or has a fatal task limit, then we have been invoked because the task + * has violated either the system-wide per-task memory limit OR its own task limit. */ if (memorystatus_kill_process_sync(p->p_pid, kMemorystatusKilledPerProcessLimit) != TRUE) { printf("task_exceeded_footprint: failed to kill the current task (exiting?).\n"); } + } else { + /* + * HWM offender exists. Done without locks or synchronization. + * See comment near its declaration for more details. + */ + memorystatus_hwm_candidates = TRUE; + } +} + +/* + * Toggle the P_MEMSTAT_TERMINATED state. + * Takes the proc_list_lock. + */ +void +proc_memstat_terminated(proc_t p, boolean_t set) +{ +#if DEVELOPMENT || DEBUG + if (p) { + proc_list_lock(); + if (set == TRUE) { + p->p_memstat_state |= P_MEMSTAT_TERMINATED; + } else { + p->p_memstat_state &= ~P_MEMSTAT_TERMINATED; + } + proc_list_unlock(); + } +#else +#pragma unused(p, set) + /* + * do nothing + */ +#endif /* DEVELOPMENT || DEBUG */ + return; +} + +/* + * This is invoked when cpulimits have been exceeded while in fatal mode. + * The jetsam_flags do not apply as those are for memory related kills. + * We call this routine so that the offending process is killed with + * a non-zero exit status. + */ +void +jetsam_on_ledger_cpulimit_exceeded(void) +{ + int retval = 0; + int jetsam_flags = 0; /* make it obvious */ + proc_t p = current_proc(); + + printf("task_exceeded_cpulimit: killing pid %d [%s]\n", + p->p_pid, (p->p_comm ? p->p_comm : "(unknown)")); + + retval = jetsam_do_kill(p, jetsam_flags); + + if (retval) { + printf("task_exceeded_cpulimit: failed to kill current task (exiting?).\n"); } } static void -memorystatus_get_task_page_counts(task_t task, uint32_t *footprint, uint32_t *max_footprint) +memorystatus_get_task_page_counts(task_t task, uint32_t *footprint, uint32_t *max_footprint, uint32_t *max_footprint_lifetime, uint32_t *purgeable_pages) { assert(task); assert(footprint); @@ -1594,33 +2861,16 @@ memorystatus_get_task_page_counts(task_t task, uint32_t *footprint, uint32_t *ma if (max_footprint) { *max_footprint = (uint32_t)(get_task_phys_footprint_max(task) / PAGE_SIZE_64); } -} - -static int -memorystatus_send_note(int event_code, void *data, size_t data_length) { - int ret; - struct kev_msg ev_msg; - - ev_msg.vendor_code = KEV_VENDOR_APPLE; - ev_msg.kev_class = KEV_SYSTEM_CLASS; - ev_msg.kev_subclass = KEV_MEMORYSTATUS_SUBCLASS; - - ev_msg.event_code = event_code; - - ev_msg.dv[0].data_length = data_length; - ev_msg.dv[0].data_ptr = data; - ev_msg.dv[1].data_length = 0; - - ret = kev_post_msg(&ev_msg); - if (ret) { - printf("%s: kev_post_msg() failed, err %d\n", __func__, ret); + if (max_footprint_lifetime) { + *max_footprint_lifetime = (uint32_t)(get_task_resident_max(task) / PAGE_SIZE_64); + } + if (purgeable_pages) { + *purgeable_pages = (uint32_t)(get_task_purgeable_size(task) / PAGE_SIZE_64); } - - return ret; } static void -memorystatus_update_snapshot_locked(proc_t p, uint32_t kill_cause) +memorystatus_update_jetsam_snapshot_entry_locked(proc_t p, uint32_t kill_cause) { unsigned int i; @@ -1642,6 +2892,25 @@ memorystatus_update_snapshot_locked(proc_t p, uint32_t kill_cause) void memorystatus_pages_update(unsigned int pages_avail) { + memorystatus_available_pages = pages_avail; + +#if VM_PRESSURE_EVENTS + /* + * Since memorystatus_available_pages changes, we should + * re-evaluate the pressure levels on the system and + * check if we need to wake the pressure thread. + * We also update memorystatus_level in that routine. + */ + vm_pressure_response(); + + if (memorystatus_available_pages <= memorystatus_available_pages_pressure) { + + if (memorystatus_hwm_candidates || (memorystatus_available_pages <= memorystatus_available_pages_critical)) { + memorystatus_thread_wake(); + } + } +#else /* VM_PRESSURE_EVENTS */ + boolean_t critical, delta; if (!memorystatus_delta) { @@ -1653,87 +2922,131 @@ void memorystatus_pages_update(unsigned int pages_avail) || (memorystatus_available_pages >= (pages_avail + memorystatus_delta))) ? TRUE : FALSE; if (critical || delta) { - memorystatus_available_pages = pages_avail; memorystatus_level = memorystatus_available_pages * 100 / atop_64(max_mem); - -#if LATENCY_JETSAM - /* Bail early to avoid excessive wake-ups */ - if (critical) { - return; - } -#endif - memorystatus_thread_wake(); } +#endif /* VM_PRESSURE_EVENTS */ } static boolean_t -memorystatus_get_snapshot_properties_for_proc_locked(proc_t p, memorystatus_jetsam_snapshot_entry_t *entry) +memorystatus_init_jetsam_snapshot_entry_locked(proc_t p, memorystatus_jetsam_snapshot_entry_t *entry) { + clock_sec_t tv_sec; + clock_usec_t tv_usec; + memset(entry, 0, sizeof(memorystatus_jetsam_snapshot_entry_t)); entry->pid = p->p_pid; strlcpy(&entry->name[0], p->p_comm, MAXCOMLEN+1); entry->priority = p->p_memstat_effectivepriority; - memorystatus_get_task_page_counts(p->task, &entry->pages, &entry->max_pages); + memorystatus_get_task_page_counts(p->task, &entry->pages, &entry->max_pages, &entry->max_pages_lifetime, &entry->purgeable_pages); entry->state = memorystatus_build_state(p); entry->user_data = p->p_memstat_userdata; memcpy(&entry->uuid[0], &p->p_uuid[0], sizeof(p->p_uuid)); + entry->fds = p->p_fd->fd_nfiles; + + absolutetime_to_microtime(get_task_cpu_time(p->task), &tv_sec, &tv_usec); + entry->cpu_time.tv_sec = tv_sec; + entry->cpu_time.tv_usec = tv_usec; return TRUE; } static void -memorystatus_jetsam_snapshot_procs_locked(void) +memorystatus_init_snapshot_vmstats(memorystatus_jetsam_snapshot_t *snapshot) { - proc_t p, next_p; - unsigned int b = 0, i = 0; kern_return_t kr = KERN_SUCCESS; - mach_msg_type_number_t count = HOST_VM_INFO64_COUNT; vm_statistics64_data_t vm_stat; if ((kr = host_statistics64(host_self(), HOST_VM_INFO64, (host_info64_t)&vm_stat, &count) != KERN_SUCCESS)) { - printf("memorystatus_jetsam_snapshot_procs_locked: host_statistics64 failed with %d\n", kr); - memset(&memorystatus_jetsam_snapshot->stats, 0, sizeof(memorystatus_jetsam_snapshot->stats)); + printf("memorystatus_init_jetsam_snapshot_stats: host_statistics64 failed with %d\n", kr); + memset(&snapshot->stats, 0, sizeof(snapshot->stats)); } else { - memorystatus_jetsam_snapshot->stats.free_pages = vm_stat.free_count; - memorystatus_jetsam_snapshot->stats.active_pages = vm_stat.active_count; - memorystatus_jetsam_snapshot->stats.inactive_pages = vm_stat.inactive_count; - memorystatus_jetsam_snapshot->stats.throttled_pages = vm_stat.throttled_count; - memorystatus_jetsam_snapshot->stats.purgeable_pages = vm_stat.purgeable_count; - memorystatus_jetsam_snapshot->stats.wired_pages = vm_stat.wire_count; - - memorystatus_jetsam_snapshot->stats.speculative_pages = vm_stat.speculative_count; - memorystatus_jetsam_snapshot->stats.filebacked_pages = vm_stat.external_page_count; - memorystatus_jetsam_snapshot->stats.anonymous_pages = vm_stat.internal_page_count; - memorystatus_jetsam_snapshot->stats.compressions = vm_stat.compressions; - memorystatus_jetsam_snapshot->stats.decompressions = vm_stat.decompressions; - memorystatus_jetsam_snapshot->stats.compressor_pages = vm_stat.compressor_page_count; - memorystatus_jetsam_snapshot->stats.total_uncompressed_pages_in_compressor = vm_stat.total_uncompressed_pages_in_compressor; + snapshot->stats.free_pages = vm_stat.free_count; + snapshot->stats.active_pages = vm_stat.active_count; + snapshot->stats.inactive_pages = vm_stat.inactive_count; + snapshot->stats.throttled_pages = vm_stat.throttled_count; + snapshot->stats.purgeable_pages = vm_stat.purgeable_count; + snapshot->stats.wired_pages = vm_stat.wire_count; + + snapshot->stats.speculative_pages = vm_stat.speculative_count; + snapshot->stats.filebacked_pages = vm_stat.external_page_count; + snapshot->stats.anonymous_pages = vm_stat.internal_page_count; + snapshot->stats.compressions = vm_stat.compressions; + snapshot->stats.decompressions = vm_stat.decompressions; + snapshot->stats.compressor_pages = vm_stat.compressor_page_count; + snapshot->stats.total_uncompressed_pages_in_compressor = vm_stat.total_uncompressed_pages_in_compressor; + } +} + +/* + * Collect vm statistics at boot. + * Called only once (see kern_exec.c) + * Data can be consumed at any time. + */ +void +memorystatus_init_at_boot_snapshot() { + memorystatus_init_snapshot_vmstats(&memorystatus_at_boot_snapshot); + memorystatus_at_boot_snapshot.entry_count = 0; + memorystatus_at_boot_snapshot.notification_time = 0; /* updated when consumed */ + memorystatus_at_boot_snapshot.snapshot_time = mach_absolute_time(); +} + +static void +memorystatus_init_jetsam_snapshot_locked(memorystatus_jetsam_snapshot_t *od_snapshot, uint32_t ods_list_count ) +{ + proc_t p, next_p; + unsigned int b = 0, i = 0; + + memorystatus_jetsam_snapshot_t *snapshot = NULL; + memorystatus_jetsam_snapshot_entry_t *snapshot_list = NULL; + unsigned int snapshot_max = 0; + + if (od_snapshot) { + /* + * This is an on_demand snapshot + */ + snapshot = od_snapshot; + snapshot_list = od_snapshot->entries; + snapshot_max = ods_list_count; + } else { + /* + * This is a jetsam event snapshot + */ + snapshot = memorystatus_jetsam_snapshot; + snapshot_list = memorystatus_jetsam_snapshot->entries; + snapshot_max = memorystatus_jetsam_snapshot_max; } + memorystatus_init_snapshot_vmstats(snapshot); + next_p = memorystatus_get_first_proc_locked(&b, TRUE); while (next_p) { p = next_p; next_p = memorystatus_get_next_proc_locked(&b, p, TRUE); - if (FALSE == memorystatus_get_snapshot_properties_for_proc_locked(p, &memorystatus_jetsam_snapshot_list[i])) { + if (FALSE == memorystatus_init_jetsam_snapshot_entry_locked(p, &snapshot_list[i])) { continue; } - MEMORYSTATUS_DEBUG(0, "jetsam snapshot pid = %d, uuid = %02x%02x%02x%02x%02x%02x%02x%02x%02x%02x%02x%02x%02x%02x%02x%02x\n", + MEMORYSTATUS_DEBUG(0, "jetsam snapshot pid %d, uuid = %02x%02x%02x%02x%02x%02x%02x%02x%02x%02x%02x%02x%02x%02x%02x%02x\n", p->p_pid, p->p_uuid[0], p->p_uuid[1], p->p_uuid[2], p->p_uuid[3], p->p_uuid[4], p->p_uuid[5], p->p_uuid[6], p->p_uuid[7], p->p_uuid[8], p->p_uuid[9], p->p_uuid[10], p->p_uuid[11], p->p_uuid[12], p->p_uuid[13], p->p_uuid[14], p->p_uuid[15]); - if (++i == memorystatus_jetsam_snapshot_max) { + if (++i == snapshot_max) { break; } } - memorystatus_jetsam_snapshot->snapshot_time = mach_absolute_time(); - memorystatus_jetsam_snapshot->entry_count = memorystatus_jetsam_snapshot_count = i; + snapshot->snapshot_time = mach_absolute_time(); + snapshot->entry_count = i; + + if (!od_snapshot) { + /* update the system buffer count */ + memorystatus_jetsam_snapshot_count = i; + } } #if DEVELOPMENT || DEBUG @@ -1762,6 +3075,30 @@ memorystatus_cmd_set_panic_bits(user_addr_t buffer, uint32_t buffer_size) { return ret; } +/* + * Triggers a sort_order on a specified jetsam priority band. + * This is for testing only, used to force a path through the sort + * function. + */ +static int +memorystatus_cmd_test_jetsam_sort(int priority, int sort_order) { + + int error = 0; + + unsigned int bucket_index = 0; + + if (priority == -1) { + /* Use as shorthand for default priority */ + bucket_index = JETSAM_PRIORITY_DEFAULT; + } else { + bucket_index = (unsigned int)priority; + } + + error = memorystatus_sort_bucket(bucket_index, sort_order); + + return (error); +} + #endif /* @@ -1773,23 +3110,24 @@ memorystatus_kill_specific_process(pid_t victim_pid, uint32_t cause) { proc_t p; /* TODO - add a victim queue and push this into the main jetsam thread */ - p = proc_find(victim_pid); if (!p) { return FALSE; } - printf("memorystatus: specifically killing pid %d [%s] - memorystatus_available_pages: %d\n", - victim_pid, (p->p_comm ? p->p_comm : "(unknown)"), memorystatus_available_pages); - proc_list_lock(); if (memorystatus_jetsam_snapshot_count == 0) { - memorystatus_jetsam_snapshot_procs_locked(); + memorystatus_init_jetsam_snapshot_locked(NULL,0); } - memorystatus_update_snapshot_locked(p, cause); + memorystatus_update_jetsam_snapshot_entry_locked(p, cause); proc_list_unlock(); + + printf("memorystatus: specifically killing pid %d [%s] (%s %d) - memorystatus_available_pages: %d\n", + victim_pid, (p->p_comm ? p->p_comm : "(unknown)"), + jetsam_kill_cause_name[cause], p->p_memstat_effectivepriority, memorystatus_available_pages); + killed = memorystatus_do_kill(p, cause); proc_rele(p); @@ -1801,12 +3139,14 @@ memorystatus_kill_specific_process(pid_t victim_pid, uint32_t cause) { * Jetsam the first process in the queue. */ static boolean_t -memorystatus_kill_top_process(boolean_t any, uint32_t cause, int32_t *priority, uint32_t *errors) +memorystatus_kill_top_process(boolean_t any, boolean_t sort_flag, uint32_t cause, int32_t *priority, uint32_t *errors) { pid_t aPid; proc_t p = PROC_NULL, next_p = PROC_NULL; boolean_t new_snapshot = FALSE, killed = FALSE; + int kill_count = 0; unsigned int i = 0; + uint32_t aPid_ep; #ifndef CONFIG_FREEZE #pragma unused(any) @@ -1815,6 +3155,11 @@ memorystatus_kill_top_process(boolean_t any, uint32_t cause, int32_t *priority, KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_JETSAM) | DBG_FUNC_START, memorystatus_available_pages, 0, 0, 0, 0); + + if (sort_flag == TRUE) { + (void)memorystatus_sort_bucket(JETSAM_PRIORITY_FOREGROUND, JETSAM_SORT_DEFAULT); + } + proc_list_lock(); next_p = memorystatus_get_first_proc_locked(&i, TRUE); @@ -1833,6 +3178,7 @@ memorystatus_kill_top_process(boolean_t any, uint32_t cause, int32_t *priority, #endif /* DEVELOPMENT || DEBUG */ aPid = p->p_pid; + aPid_ep = p->p_memstat_effectivepriority; if (p->p_memstat_state & (P_MEMSTAT_ERROR | P_MEMSTAT_TERMINATED)) { continue; @@ -1845,6 +3191,22 @@ memorystatus_kill_top_process(boolean_t any, uint32_t cause, int32_t *priority, } #endif /* DEVELOPMENT || DEBUG */ + if (cause == kMemorystatusKilledVnodes) + { + /* + * If the system runs out of vnodes, we systematically jetsam + * processes in hopes of stumbling onto a vnode gain that helps + * the system recover. The process that happens to trigger + * this path has no known relationship to the vnode consumption. + * We attempt to safeguard that process e.g: do not jetsam it. + */ + + if (p == current_proc()) { + /* do not jetsam the current process */ + continue; + } + } + #if CONFIG_FREEZE boolean_t skip; boolean_t reclaim_proc = !(p->p_memstat_state & (P_MEMSTAT_LOCKED | P_MEMSTAT_NORECLAIM)); @@ -1859,18 +3221,14 @@ memorystatus_kill_top_process(boolean_t any, uint32_t cause, int32_t *priority, } else #endif { - if (priority) { - *priority = p->p_memstat_effectivepriority; - } - /* * Capture a snapshot if none exists and: * - priority was not requested (this is something other than an ambient kill) * - the priority was requested *and* the targeted process is not at idle priority */ if ((memorystatus_jetsam_snapshot_count == 0) && - ((!priority) || (priority && (*priority != JETSAM_PRIORITY_IDLE)))) { - memorystatus_jetsam_snapshot_procs_locked(); + (memorystatus_idle_snapshot || ((!priority) || (priority && (*priority != JETSAM_PRIORITY_IDLE))))) { + memorystatus_init_jetsam_snapshot_locked(NULL,0); new_snapshot = TRUE; } @@ -1886,7 +3244,7 @@ memorystatus_kill_top_process(boolean_t any, uint32_t cause, int32_t *priority, if ((memorystatus_jetsam_policy & kPolicyDiagnoseActive) && activeProcess) { MEMORYSTATUS_DEBUG(1, "jetsam: suspending pid %d [%s] (active) for diagnosis - memory_status_level: %d\n", aPid, (p->p_comm ? p->p_comm: "(unknown)"), memorystatus_level); - memorystatus_update_snapshot_locked(p, kMemorystatusKilledDiagnostic); + memorystatus_update_jetsam_snapshot_entry_locked(p, kMemorystatusKilledDiagnostic); p->p_memstat_state |= P_MEMSTAT_DIAG_SUSPENDED; if (memorystatus_jetsam_policy & kPolicyDiagnoseFirst) { jetsam_diagnostic_suspended_one_active_proc = 1; @@ -1897,6 +3255,9 @@ memorystatus_kill_top_process(boolean_t any, uint32_t cause, int32_t *priority, proc_list_unlock(); if (p) { task_suspend(p->task); + if (priority) { + *priority = aPid_ep; + } proc_rele(p); killed = TRUE; } @@ -1906,28 +3267,51 @@ memorystatus_kill_top_process(boolean_t any, uint32_t cause, int32_t *priority, #endif /* DEVELOPMENT || DEBUG */ { /* Shift queue, update stats */ - memorystatus_update_snapshot_locked(p, cause); - - p = proc_ref_locked(p); - proc_list_unlock(); - if (p) { - printf("memorystatus: jetsam killing pid %d [%s] - memorystatus_available_pages: %d\n", - aPid, (p->p_comm ? p->p_comm : "(unknown)"), memorystatus_available_pages); + memorystatus_update_jetsam_snapshot_entry_locked(p, cause); + + if (proc_ref_locked(p) == p) { + proc_list_unlock(); + printf("memorystatus: %s %d [%s] (%s %d) - memorystatus_available_pages: %d\n", + ((aPid_ep == JETSAM_PRIORITY_IDLE) ? + "idle exiting pid" : "jetsam killing pid"), + aPid, (p->p_comm ? p->p_comm : "(unknown)"), + jetsam_kill_cause_name[cause], aPid_ep, memorystatus_available_pages); + killed = memorystatus_do_kill(p, cause); - } + + /* Success? */ + if (killed) { + if (priority) { + *priority = aPid_ep; + } + proc_rele(p); + kill_count++; + goto exit; + } - /* Success? */ - if (killed) { - proc_rele(p); - goto exit; + /* + * Failure - first unwind the state, + * then fall through to restart the search. + */ + proc_list_lock(); + proc_rele_locked(p); + p->p_memstat_state &= ~P_MEMSTAT_TERMINATED; + p->p_memstat_state |= P_MEMSTAT_ERROR; + *errors += 1; } - /* Failure - unwind and restart. */ - proc_list_lock(); - proc_rele_locked(p); - p->p_memstat_state &= ~P_MEMSTAT_TERMINATED; - p->p_memstat_state |= P_MEMSTAT_ERROR; - *errors += 1; + /* + * Failure - restart the search. + * + * We might have raced with "p" exiting on another core, resulting in no + * ref on "p". Or, we may have failed to kill "p". + * + * Either way, we fall thru to here, leaving the proc in the + * P_MEMSTAT_TERMINATED state. + * + * And, we hold the the proc_list_lock at this point. + */ + i = 0; next_p = memorystatus_get_first_proc_locked(&i, TRUE); } @@ -1943,11 +3327,227 @@ exit: } KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_JETSAM) | DBG_FUNC_END, - memorystatus_available_pages, killed ? aPid : 0, 0, 0, 0); + memorystatus_available_pages, killed ? aPid : 0, kill_count, 0, 0); return killed; } +/* + * Jetsam aggressively + */ +static boolean_t +memorystatus_kill_top_process_aggressive(boolean_t any, uint32_t cause, int aggr_count, int32_t priority_max, + uint32_t *errors) +{ + pid_t aPid; + proc_t p = PROC_NULL, next_p = PROC_NULL; + boolean_t new_snapshot = FALSE, killed = FALSE; + int kill_count = 0; + unsigned int i = 0; + int32_t aPid_ep = 0; + unsigned int memorystatus_level_snapshot = 0; + +#pragma unused(any) + + KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_JETSAM) | DBG_FUNC_START, + memorystatus_available_pages, priority_max, 0, 0, 0); + + memorystatus_sort_bucket(JETSAM_PRIORITY_FOREGROUND, JETSAM_SORT_DEFAULT); + + proc_list_lock(); + + next_p = memorystatus_get_first_proc_locked(&i, TRUE); + while (next_p) { +#if DEVELOPMENT || DEBUG + int activeProcess; + int procSuspendedForDiagnosis; +#endif /* DEVELOPMENT || DEBUG */ + + if ((unsigned int)(next_p->p_memstat_effectivepriority) != i) { + + /* + * We have raced with next_p running on another core, as it has + * moved to a different jetsam priority band. This means we have + * lost our place in line while traversing the jetsam list. We + * attempt to recover by rewinding to the beginning of the band + * we were already traversing. By doing this, we do not guarantee + * that no process escapes this aggressive march, but we can make + * skipping an entire range of processes less likely. (PR-21069019) + */ + + MEMORYSTATUS_DEBUG(1, "memorystatus: aggressive%d: rewinding %s moved from band %d --> %d\n", + aggr_count, next_p->p_comm, i, next_p->p_memstat_effectivepriority); + + next_p = memorystatus_get_first_proc_locked(&i, TRUE); + continue; + } + + p = next_p; + next_p = memorystatus_get_next_proc_locked(&i, p, TRUE); + + if (p->p_memstat_effectivepriority > priority_max) { + /* + * Bail out of this killing spree if we have + * reached beyond the priority_max jetsam band. + * That is, we kill up to and through the + * priority_max jetsam band. + */ + proc_list_unlock(); + goto exit; + } + +#if DEVELOPMENT || DEBUG + activeProcess = p->p_memstat_state & P_MEMSTAT_FOREGROUND; + procSuspendedForDiagnosis = p->p_memstat_state & P_MEMSTAT_DIAG_SUSPENDED; +#endif /* DEVELOPMENT || DEBUG */ + + aPid = p->p_pid; + aPid_ep = p->p_memstat_effectivepriority; + + if (p->p_memstat_state & (P_MEMSTAT_ERROR | P_MEMSTAT_TERMINATED)) { + continue; + } + +#if DEVELOPMENT || DEBUG + if ((memorystatus_jetsam_policy & kPolicyDiagnoseActive) && procSuspendedForDiagnosis) { + printf("jetsam: continuing after ignoring proc suspended already for diagnosis - %d\n", aPid); + continue; + } +#endif /* DEVELOPMENT || DEBUG */ + + /* + * Capture a snapshot if none exists. + */ + if (memorystatus_jetsam_snapshot_count == 0) { + memorystatus_init_jetsam_snapshot_locked(NULL,0); + new_snapshot = TRUE; + } + + /* + * Mark as terminated so that if exit1() indicates success, but the process (for example) + * is blocked in task_exception_notify(), it'll be skipped if encountered again - see + * . This is cheaper than examining P_LEXIT, which requires the + * acquisition of the proc lock. + */ + p->p_memstat_state |= P_MEMSTAT_TERMINATED; + + /* Shift queue, update stats */ + memorystatus_update_jetsam_snapshot_entry_locked(p, cause); + + /* + * In order to kill the target process, we will drop the proc_list_lock. + * To guaranteee that p and next_p don't disappear out from under the lock, + * we must take a ref on both. + * If we cannot get a reference, then it's likely we've raced with + * that process exiting on another core. + */ + if (proc_ref_locked(p) == p) { + if (next_p) { + while (next_p && (proc_ref_locked(next_p) != next_p)) { + proc_t temp_p; + + /* + * We must have raced with next_p exiting on another core. + * Recover by getting the next eligible process in the band. + */ + + MEMORYSTATUS_DEBUG(1, "memorystatus: aggressive%d: skipping %d [%s] (exiting?)\n", + aggr_count, next_p->p_pid, (next_p->p_comm ? next_p->p_comm : "(unknown)")); + + temp_p = next_p; + next_p = memorystatus_get_next_proc_locked(&i, temp_p, TRUE); + } + } + proc_list_unlock(); + + printf("memorystatus: aggressive%d: %s %d [%s] (%s %d) - memorystatus_available_pages: %d\n", + aggr_count, + ((aPid_ep == JETSAM_PRIORITY_IDLE) ? "idle exiting pid" : "jetsam killing pid"), + aPid, (p->p_comm ? p->p_comm : "(unknown)"), + jetsam_kill_cause_name[cause], aPid_ep, memorystatus_available_pages); + + memorystatus_level_snapshot = memorystatus_level; + + killed = memorystatus_do_kill(p, cause); + + /* Success? */ + if (killed) { + proc_rele(p); + kill_count++; + p = NULL; + killed = FALSE; + + /* + * Continue the killing spree. + */ + proc_list_lock(); + if (next_p) { + proc_rele_locked(next_p); + } + + if (aPid_ep == JETSAM_PRIORITY_FOREGROUND && memorystatus_aggressive_jetsam_lenient == TRUE) { + if (memorystatus_level > memorystatus_level_snapshot && ((memorystatus_level - memorystatus_level_snapshot) >= AGGRESSIVE_JETSAM_LENIENT_MODE_THRESHOLD)) { +#if DEVELOPMENT || DEBUG + printf("Disabling Lenient mode after one-time deployment.\n"); +#endif /* DEVELOPMENT || DEBUG */ + memorystatus_aggressive_jetsam_lenient = FALSE; + break; + } + } + + continue; + } + + /* + * Failure - first unwind the state, + * then fall through to restart the search. + */ + proc_list_lock(); + proc_rele_locked(p); + if (next_p) { + proc_rele_locked(next_p); + } + p->p_memstat_state &= ~P_MEMSTAT_TERMINATED; + p->p_memstat_state |= P_MEMSTAT_ERROR; + *errors += 1; + p = NULL; + } + + /* + * Failure - restart the search at the beginning of + * the band we were already traversing. + * + * We might have raced with "p" exiting on another core, resulting in no + * ref on "p". Or, we may have failed to kill "p". + * + * Either way, we fall thru to here, leaving the proc in the + * P_MEMSTAT_TERMINATED or P_MEMSTAT_ERROR state. + * + * And, we hold the the proc_list_lock at this point. + */ + + next_p = memorystatus_get_first_proc_locked(&i, TRUE); + } + + proc_list_unlock(); + +exit: + /* Clear snapshot if freshly captured and no target was found */ + if (new_snapshot && (kill_count == 0)) { + memorystatus_jetsam_snapshot->entry_count = memorystatus_jetsam_snapshot_count = 0; + } + + KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_JETSAM) | DBG_FUNC_END, + memorystatus_available_pages, killed ? aPid : 0, kill_count, 0, 0); + + if (kill_count > 0) { + return(TRUE); + } + else { + return(FALSE); + } +} + #if LEGACY_HIWATER static boolean_t @@ -1956,7 +3556,9 @@ memorystatus_kill_hiwat_proc(uint32_t *errors) pid_t aPid = 0; proc_t p = PROC_NULL, next_p = PROC_NULL; boolean_t new_snapshot = FALSE, killed = FALSE; + int kill_count = 0; unsigned int i = 0; + uint32_t aPid_ep; KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_JETSAM_HIWAT) | DBG_FUNC_START, memorystatus_available_pages, 0, 0, 0, 0); @@ -1972,6 +3574,7 @@ memorystatus_kill_hiwat_proc(uint32_t *errors) next_p = memorystatus_get_next_proc_locked(&i, p, TRUE); aPid = p->p_pid; + aPid_ep = p->p_memstat_effectivepriority; if (p->p_memstat_state & (P_MEMSTAT_ERROR | P_MEMSTAT_TERMINATED)) { continue; @@ -1981,14 +3584,24 @@ memorystatus_kill_hiwat_proc(uint32_t *errors) if (p->p_memstat_memlimit <= 0) { continue; } - + +#if 0 + /* + * No need to consider P_MEMSTAT_MEMLIMIT_BACKGROUND anymore. + * Background limits are described via the inactive limit slots. + * Their fatal/non-fatal setting will drive whether or not to be + * considered in this kill path. + */ + /* skip if a currently inapplicable limit is encountered */ if ((p->p_memstat_state & P_MEMSTAT_MEMLIMIT_BACKGROUND) && (p->p_memstat_effectivepriority >= JETSAM_PRIORITY_FOREGROUND)) { continue; } +#endif footprint = (uint32_t)(get_task_phys_footprint(p->task) / (1024 * 1024)); skip = (((int32_t)footprint) <= p->p_memstat_memlimit); + #if DEVELOPMENT || DEBUG if (!skip && (memorystatus_jetsam_policy & kPolicyDiagnoseActive)) { if (p->p_memstat_state & P_MEMSTAT_DIAG_SUSPENDED) { @@ -2010,11 +3623,11 @@ memorystatus_kill_hiwat_proc(uint32_t *errors) if (skip) { continue; } else { - MEMORYSTATUS_DEBUG(1, "jetsam: %s pid %d [%s] - %d pages > 1 (%d)\n", - (memorystatus_jetsam_policy & kPolicyDiagnoseActive) ? "suspending": "killing", aPid, p->p_comm, pages, hiwat); + MEMORYSTATUS_DEBUG(1, "jetsam: %s pid %d [%s] - %d Mb > 1 (%d Mb)\n", + (memorystatus_jetsam_policy & kPolicyDiagnoseActive) ? "suspending": "killing", aPid, p->p_comm, footprint, p->p_memstat_memlimit); if (memorystatus_jetsam_snapshot_count == 0) { - memorystatus_jetsam_snapshot_procs_locked(); + memorystatus_init_jetsam_snapshot_locked(NULL,0); new_snapshot = TRUE; } @@ -2023,7 +3636,7 @@ memorystatus_kill_hiwat_proc(uint32_t *errors) #if DEVELOPMENT || DEBUG if (memorystatus_jetsam_policy & kPolicyDiagnoseActive) { MEMORYSTATUS_DEBUG(1, "jetsam: pid %d suspended for diagnosis - memorystatus_available_pages: %d\n", aPid, memorystatus_available_pages); - memorystatus_update_snapshot_locked(p, kMemorystatusKilledDiagnostic); + memorystatus_update_jetsam_snapshot_entry_locked(p, kMemorystatusKilledDiagnostic); p->p_memstat_state |= P_MEMSTAT_DIAG_SUSPENDED; p = proc_ref_locked(p); @@ -2038,28 +3651,46 @@ memorystatus_kill_hiwat_proc(uint32_t *errors) } else #endif /* DEVELOPMENT || DEBUG */ { - memorystatus_update_snapshot_locked(p, kMemorystatusKilledHiwat); + memorystatus_update_jetsam_snapshot_entry_locked(p, kMemorystatusKilledHiwat); - p = proc_ref_locked(p); - proc_list_unlock(); - if (p) { - printf("memorystatus: jetsam killing pid %d [%s] (highwater) - memorystatus_available_pages: %d\n", - aPid, (p->p_comm ? p->p_comm : "(unknown)"), memorystatus_available_pages); - killed = memorystatus_do_kill(p, kMemorystatusKilledHiwat); - } + if (proc_ref_locked(p) == p) { + proc_list_unlock(); + + printf("memorystatus: jetsam killing pid %d [%s] (highwater %d) - memorystatus_available_pages: %d\n", + aPid, (p->p_comm ? p->p_comm : "(unknown)"), aPid_ep, memorystatus_available_pages); + + killed = memorystatus_do_kill(p, kMemorystatusKilledHiwat); - /* Success? */ - if (killed) { - proc_rele(p); - goto exit; + /* Success? */ + if (killed) { + proc_rele(p); + kill_count++; + goto exit; + } + + /* + * Failure - first unwind the state, + * then fall through to restart the search. + */ + proc_list_lock(); + proc_rele_locked(p); + p->p_memstat_state &= ~P_MEMSTAT_TERMINATED; + p->p_memstat_state |= P_MEMSTAT_ERROR; + *errors += 1; } - /* Failure - unwind and restart. */ - proc_list_lock(); - proc_rele_locked(p); - p->p_memstat_state &= ~P_MEMSTAT_TERMINATED; - p->p_memstat_state |= P_MEMSTAT_ERROR; - *errors += 1; + /* + * Failure - restart the search. + * + * We might have raced with "p" exiting on another core, resulting in no + * ref on "p". Or, we may have failed to kill "p". + * + * Either way, we fall thru to here, leaving the proc in the + * P_MEMSTAT_TERMINATED state. + * + * And, we hold the the proc_list_lock at this point. + */ + i = 0; next_p = memorystatus_get_first_proc_locked(&i, TRUE); } @@ -2075,7 +3706,7 @@ exit: } KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_JETSAM_HIWAT) | DBG_FUNC_END, - memorystatus_available_pages, killed ? aPid : 0, 0, 0, 0); + memorystatus_available_pages, killed ? aPid : 0, kill_count, 0, 0); return killed; } @@ -2085,11 +3716,12 @@ exit: static boolean_t memorystatus_kill_process_async(pid_t victim_pid, uint32_t cause) { /* TODO: allow a general async path */ - if ((victim_pid != -1) || (cause != kMemorystatusKilledVMPageShortage || cause != kMemorystatusKilledVMThrashing)) { + if ((victim_pid != -1) || (cause != kMemorystatusKilledVMPageShortage && cause != kMemorystatusKilledVMThrashing && + cause != kMemorystatusKilledFCThrashing)) { return FALSE; } - kill_under_pressure = TRUE; + kill_under_pressure_cause = cause; memorystatus_thread_wake(); return TRUE; } @@ -2101,7 +3733,7 @@ memorystatus_kill_process_sync(pid_t victim_pid, uint32_t cause) { if (victim_pid == -1) { /* No pid, so kill first process */ - res = memorystatus_kill_top_process(TRUE, cause, NULL, &errors); + res = memorystatus_kill_top_process(TRUE, TRUE, cause, NULL, &errors); } else { res = memorystatus_kill_specific_process(victim_pid, cause); } @@ -2114,8 +3746,17 @@ memorystatus_kill_process_sync(pid_t victim_pid, uint32_t cause) { /* Fire off snapshot notification */ size_t snapshot_size = sizeof(memorystatus_jetsam_snapshot_t) + sizeof(memorystatus_jetsam_snapshot_entry_t) * memorystatus_jetsam_snapshot_count; - memorystatus_jetsam_snapshot->notification_time = mach_absolute_time(); - memorystatus_send_note(kMemorystatusSnapshotNote, &snapshot_size, sizeof(snapshot_size)); + uint64_t timestamp_now = mach_absolute_time(); + memorystatus_jetsam_snapshot->notification_time = timestamp_now; + if (memorystatus_jetsam_snapshot_last_timestamp == 0 || + timestamp_now > memorystatus_jetsam_snapshot_last_timestamp + memorystatus_jetsam_snapshot_timeout) { + int ret = memorystatus_send_note(kMemorystatusSnapshotNote, &snapshot_size, sizeof(snapshot_size)); + if (!ret) { + proc_list_lock(); + memorystatus_jetsam_snapshot_last_timestamp = timestamp_now; + proc_list_unlock(); + } + } } return res; @@ -2139,6 +3780,15 @@ memorystatus_kill_on_VM_thrashing(boolean_t async) { } } +boolean_t +memorystatus_kill_on_FC_thrashing(boolean_t async) { + if (async) { + return memorystatus_kill_process_async(-1, kMemorystatusKilledFCThrashing); + } else { + return memorystatus_kill_process_sync(-1, kMemorystatusKilledFCThrashing); + } +} + boolean_t memorystatus_kill_on_vnode_limit(void) { return memorystatus_kill_process_sync(-1, kMemorystatusKilledVnodes); @@ -2153,6 +3803,11 @@ memorystatus_freeze_init(void) { kern_return_t result; thread_t thread; + + freezer_lck_grp_attr = lck_grp_attr_alloc_init(); + freezer_lck_grp = lck_grp_alloc_init("freezer", freezer_lck_grp_attr); + + lck_mtx_init(&freezer_mutex, freezer_lck_grp, NULL); result = kernel_thread_start(memorystatus_freeze_thread, NULL, &thread); if (result == KERN_SUCCESS) { @@ -2162,63 +3817,219 @@ memorystatus_freeze_init(void) } } -static int -memorystatus_freeze_top_process(boolean_t *memorystatus_freeze_swap_low) +/* + * Synchronously freeze the passed proc. Called with a reference to the proc held. + * + * Returns EINVAL or the value returned by task_freeze(). + */ +int +memorystatus_freeze_process_sync(proc_t p) { + int ret = EINVAL; pid_t aPid = 0; - int ret = -1; - proc_t p = PROC_NULL, next_p = PROC_NULL; - unsigned int i = 0; + boolean_t memorystatus_freeze_swap_low = FALSE; KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_FREEZE) | DBG_FUNC_START, memorystatus_available_pages, 0, 0, 0, 0); + lck_mtx_lock(&freezer_mutex); + + if (p == NULL) { + goto exit; + } + + if (memorystatus_freeze_enabled == FALSE) { + goto exit; + } + + if (!memorystatus_can_freeze(&memorystatus_freeze_swap_low)) { + goto exit; + } + + if (memorystatus_freeze_update_throttle()) { + printf("memorystatus_freeze_process_sync: in throttle, ignorning freeze\n"); + memorystatus_freeze_throttle_count++; + goto exit; + } + proc_list_lock(); - - next_p = memorystatus_get_first_proc_locked(&i, TRUE); - while (next_p) { - kern_return_t kr; - uint32_t purgeable, wired, clean, dirty; + + if (p != NULL) { + uint32_t purgeable, wired, clean, dirty, state; + uint32_t max_pages, pages, i; boolean_t shared; - uint32_t pages; - uint32_t max_pages = 0; - uint32_t state; - - p = next_p; - next_p = memorystatus_get_next_proc_locked(&i, p, TRUE); aPid = p->p_pid; state = p->p_memstat_state; /* Ensure the process is eligible for freezing */ if ((state & (P_MEMSTAT_TERMINATED | P_MEMSTAT_LOCKED | P_MEMSTAT_FROZEN)) || !(state & P_MEMSTAT_SUSPENDED)) { - continue; // with lock held + proc_list_unlock(); + goto exit; } - + /* Only freeze processes meeting our minimum resident page criteria */ - memorystatus_get_task_page_counts(p->task, &pages, NULL); + memorystatus_get_task_page_counts(p->task, &pages, NULL, NULL, NULL); if (pages < memorystatus_freeze_pages_min) { - continue; // with lock held - } + proc_list_unlock(); + goto exit; + } + + if (DEFAULT_FREEZER_IS_ACTIVE || DEFAULT_FREEZER_COMPRESSED_PAGER_IS_SWAPBACKED) { + + unsigned int avail_swap_space = 0; /* in pages. */ + + if (DEFAULT_FREEZER_IS_ACTIVE) { + /* + * Freezer backed by default pager and swap file(s). + */ + avail_swap_space = default_pager_swap_pages_free(); + } else { + /* + * Freezer backed by the compressor and swap file(s) + * while will hold compressed data. + */ + avail_swap_space = vm_swap_get_free_space() / PAGE_SIZE_64; + } + + max_pages = MIN(avail_swap_space, memorystatus_freeze_pages_max); - if (DEFAULT_FREEZER_IS_ACTIVE || DEFAULT_FREEZER_COMPRESSED_PAGER_IS_ACTIVE) { - /* Ensure there's enough free space to freeze this process. */ - max_pages = MIN(default_pager_swap_pages_free(), memorystatus_freeze_pages_max); if (max_pages < memorystatus_freeze_pages_min) { - *memorystatus_freeze_swap_low = TRUE; proc_list_unlock(); goto exit; } } else { + /* + * We only have the compressor without any swap. + */ max_pages = UINT32_MAX - 1; } - + /* Mark as locked temporarily to avoid kill */ p->p_memstat_state |= P_MEMSTAT_LOCKED; + proc_list_unlock(); - p = proc_ref_locked(p); - proc_list_unlock(); - if (!p) { + ret = task_freeze(p->task, &purgeable, &wired, &clean, &dirty, max_pages, &shared, FALSE); + + MEMORYSTATUS_DEBUG(1, "memorystatus_freeze_process_sync: task_freeze %s for pid %d [%s] - " + "memorystatus_pages: %d, purgeable: %d, wired: %d, clean: %d, dirty: %d, shared %d, free swap: %d\n", + (ret == KERN_SUCCESS) ? "SUCCEEDED" : "FAILED", aPid, (p->p_comm ? p->p_comm : "(unknown)"), + memorystatus_available_pages, purgeable, wired, clean, dirty, shared, default_pager_swap_pages_free()); + + proc_list_lock(); + p->p_memstat_state &= ~P_MEMSTAT_LOCKED; + + if (ret == KERN_SUCCESS) { + memorystatus_freeze_entry_t data = { aPid, TRUE, dirty }; + + memorystatus_frozen_count++; + + p->p_memstat_state |= (P_MEMSTAT_FROZEN | (shared ? 0: P_MEMSTAT_NORECLAIM)); + + if (DEFAULT_FREEZER_IS_ACTIVE || DEFAULT_FREEZER_COMPRESSED_PAGER_IS_SWAPBACKED) { + /* Update stats */ + for (i = 0; i < sizeof(throttle_intervals) / sizeof(struct throttle_interval_t); i++) { + throttle_intervals[i].pageouts += dirty; + } + } + + memorystatus_freeze_pageouts += dirty; + memorystatus_freeze_count++; + + proc_list_unlock(); + + memorystatus_send_note(kMemorystatusFreezeNote, &data, sizeof(data)); + } else { + proc_list_unlock(); + } + } + +exit: + lck_mtx_unlock(&freezer_mutex); + KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_FREEZE) | DBG_FUNC_END, + memorystatus_available_pages, aPid, 0, 0, 0); + + return ret; +} + +static int +memorystatus_freeze_top_process(boolean_t *memorystatus_freeze_swap_low) +{ + pid_t aPid = 0; + int ret = -1; + proc_t p = PROC_NULL, next_p = PROC_NULL; + unsigned int i = 0; + + KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_FREEZE) | DBG_FUNC_START, + memorystatus_available_pages, 0, 0, 0, 0); + + proc_list_lock(); + + next_p = memorystatus_get_first_proc_locked(&i, TRUE); + while (next_p) { + kern_return_t kr; + uint32_t purgeable, wired, clean, dirty; + boolean_t shared; + uint32_t pages; + uint32_t max_pages = 0; + uint32_t state; + + p = next_p; + next_p = memorystatus_get_next_proc_locked(&i, p, TRUE); + + aPid = p->p_pid; + state = p->p_memstat_state; + + /* Ensure the process is eligible for freezing */ + if ((state & (P_MEMSTAT_TERMINATED | P_MEMSTAT_LOCKED | P_MEMSTAT_FROZEN)) || !(state & P_MEMSTAT_SUSPENDED)) { + continue; // with lock held + } + + /* Only freeze processes meeting our minimum resident page criteria */ + memorystatus_get_task_page_counts(p->task, &pages, NULL, NULL, NULL); + if (pages < memorystatus_freeze_pages_min) { + continue; // with lock held + } + + if (DEFAULT_FREEZER_IS_ACTIVE || DEFAULT_FREEZER_COMPRESSED_PAGER_IS_SWAPBACKED) { + + /* Ensure there's enough free space to freeze this process. */ + + unsigned int avail_swap_space = 0; /* in pages. */ + + if (DEFAULT_FREEZER_IS_ACTIVE) { + /* + * Freezer backed by default pager and swap file(s). + */ + avail_swap_space = default_pager_swap_pages_free(); + } else { + /* + * Freezer backed by the compressor and swap file(s) + * while will hold compressed data. + */ + avail_swap_space = vm_swap_get_free_space() / PAGE_SIZE_64; + } + + max_pages = MIN(avail_swap_space, memorystatus_freeze_pages_max); + + if (max_pages < memorystatus_freeze_pages_min) { + *memorystatus_freeze_swap_low = TRUE; + proc_list_unlock(); + goto exit; + } + } else { + /* + * We only have the compressor pool. + */ + max_pages = UINT32_MAX - 1; + } + + /* Mark as locked temporarily to avoid kill */ + p->p_memstat_state |= P_MEMSTAT_LOCKED; + + p = proc_ref_locked(p); + proc_list_unlock(); + if (!p) { goto exit; } @@ -2240,11 +4051,13 @@ memorystatus_freeze_top_process(boolean_t *memorystatus_freeze_swap_low) p->p_memstat_state |= (P_MEMSTAT_FROZEN | (shared ? 0: P_MEMSTAT_NORECLAIM)); - /* Update stats */ - for (i = 0; i < sizeof(throttle_intervals) / sizeof(struct throttle_interval_t); i++) { - throttle_intervals[i].pageouts += dirty; + if (DEFAULT_FREEZER_IS_ACTIVE || DEFAULT_FREEZER_COMPRESSED_PAGER_IS_SWAPBACKED) { + /* Update stats */ + for (i = 0; i < sizeof(throttle_intervals) / sizeof(struct throttle_interval_t); i++) { + throttle_intervals[i].pageouts += dirty; + } } - + memorystatus_freeze_pageouts += dirty; memorystatus_freeze_count++; @@ -2252,8 +4065,8 @@ memorystatus_freeze_top_process(boolean_t *memorystatus_freeze_swap_low) memorystatus_send_note(kMemorystatusFreezeNote, &data, sizeof(data)); - /* Return the number of reclaimed pages */ - ret = dirty; + /* Return KERN_SUCESS */ + ret = kr; } else { proc_list_unlock(); @@ -2314,6 +4127,8 @@ memorystatus_can_freeze_processes(void) static boolean_t memorystatus_can_freeze(boolean_t *memorystatus_freeze_swap_low) { + boolean_t can_freeze = TRUE; + /* Only freeze if we're sufficiently low on memory; this holds off freeze right after boot, and is generally is a no-op once we've reached steady state. */ if (memorystatus_available_pages > memorystatus_freeze_threshold) { @@ -2325,27 +4140,68 @@ memorystatus_can_freeze(boolean_t *memorystatus_freeze_swap_low) return FALSE; } - /* Is swap running low? */ - if (*memorystatus_freeze_swap_low) { - /* If there's been no movement in free swap pages since we last attempted freeze, return. */ - if (default_pager_swap_pages_free() < memorystatus_freeze_pages_min) { - return FALSE; + if (COMPRESSED_PAGER_IS_SWAPLESS || DEFAULT_FREEZER_COMPRESSED_PAGER_IS_SWAPLESS) { + /* + * In-core compressor used for freezing WITHOUT on-disk swap support. + */ + + if (vm_compressor_low_on_space()) { + if (*memorystatus_freeze_swap_low) { + *memorystatus_freeze_swap_low = TRUE; + } + + can_freeze = FALSE; + + } else { + if (*memorystatus_freeze_swap_low) { + *memorystatus_freeze_swap_low = FALSE; + } + + can_freeze = TRUE; + } + } else { + /* + * Freezing WITH on-disk swap support. + */ + + if (DEFAULT_FREEZER_COMPRESSED_PAGER_IS_SWAPBACKED) { + /* + * In-core compressor fronts the swap. + */ + if (vm_swap_low_on_space()) { + if (*memorystatus_freeze_swap_low) { + *memorystatus_freeze_swap_low = TRUE; + } + + can_freeze = FALSE; + } + + } else if (DEFAULT_FREEZER_IS_ACTIVE) { + /* + * Legacy freeze mode with no compressor support. + */ + if (default_pager_swap_pages_free() < memorystatus_freeze_pages_min) { + if (*memorystatus_freeze_swap_low) { + *memorystatus_freeze_swap_low = TRUE; + } + + can_freeze = FALSE; + } + } else { + panic("Not a valid freeze configuration.\n"); } - - /* Pages have been freed - we can retry. */ - *memorystatus_freeze_swap_low = FALSE; } - /* OK */ - return TRUE; + return can_freeze; } static void memorystatus_freeze_update_throttle_interval(mach_timespec_t *ts, struct throttle_interval_t *interval) { + unsigned int freeze_daily_pageouts_max = memorystatus_freeze_daily_mb_max * (1024 * 1024 / PAGE_SIZE); if (CMP_MACH_TIMESPEC(ts, &interval->ts) >= 0) { if (!interval->max_pageouts) { - interval->max_pageouts = (interval->burst_multiple * (((uint64_t)interval->mins * FREEZE_DAILY_PAGEOUTS_MAX) / (24 * 60))); + interval->max_pageouts = (interval->burst_multiple * (((uint64_t)interval->mins * freeze_daily_pageouts_max) / (24 * 60))); } else { printf("memorystatus_freeze_update_throttle_interval: %d minute throttle timeout, resetting\n", interval->mins); } @@ -2406,10 +4262,11 @@ static void memorystatus_freeze_thread(void *param __unused, wait_result_t wr __unused) { static boolean_t memorystatus_freeze_swap_low = FALSE; - + + lck_mtx_lock(&freezer_mutex); if (memorystatus_freeze_enabled) { if (memorystatus_can_freeze(&memorystatus_freeze_swap_low)) { - /* Only freeze if we've not exceeded our pageout budgets */ + /* Only freeze if we've not exceeded our pageout budgets.*/ if (!memorystatus_freeze_update_throttle()) { memorystatus_freeze_top_process(&memorystatus_freeze_swap_low); } else { @@ -2418,6 +4275,7 @@ memorystatus_freeze_thread(void *param __unused, wait_result_t wr __unused) } } } + lck_mtx_unlock(&freezer_mutex); assert_wait((event_t) &memorystatus_freeze_wakeup, THREAD_UNINT); thread_block((thread_continue_t) memorystatus_freeze_thread); @@ -2425,113 +4283,109 @@ memorystatus_freeze_thread(void *param __unused, wait_result_t wr __unused) #endif /* CONFIG_FREEZE */ -#if CONFIG_JETSAM && VM_PRESSURE_EVENTS +#if VM_PRESSURE_EVENTS -boolean_t -memorystatus_warn_process(pid_t pid) { - return (vm_dispatch_pressure_note_to_pid(pid, FALSE) == 0); -} +#if CONFIG_MEMORYSTATUS -static inline boolean_t -memorystatus_update_pressure_locked(boolean_t *pressured) { - vm_pressure_level_t old_level, new_level; - - old_level = memorystatus_vm_pressure_level; - - if (memorystatus_available_pages > memorystatus_available_pages_pressure) { - /* Too many free pages */ - new_level = kVMPressureNormal; - } -#if CONFIG_FREEZE - else if (memorystatus_frozen_count > 0) { - /* Frozen processes exist */ - new_level = kVMPressureNormal; - } -#endif - else if (memorystatus_suspended_count > MEMORYSTATUS_SUSPENDED_THRESHOLD) { - /* Too many supended processes */ - new_level = kVMPressureNormal; - } - else if (memorystatus_suspended_count > 0) { - /* Some suspended processes - warn */ - new_level = kVMPressureWarning; - } - else { - /* Otherwise, pressure level is urgent */ - new_level = kVMPressureUrgent; - } - - *pressured = (new_level != kVMPressureNormal); +static int +memorystatus_send_note(int event_code, void *data, size_t data_length) { + int ret; + struct kev_msg ev_msg; - /* Did the pressure level change? */ - if (old_level != new_level) { - MEMORYSTATUS_DEBUG(1, "memorystatus_update_pressure_locked(): memory pressure changed %d -> %d; memorystatus_available_pages: %d\n ", - old_level, new_level, memorystatus_available_pages); - memorystatus_vm_pressure_level = new_level; - return TRUE; + ev_msg.vendor_code = KEV_VENDOR_APPLE; + ev_msg.kev_class = KEV_SYSTEM_CLASS; + ev_msg.kev_subclass = KEV_MEMORYSTATUS_SUBCLASS; + + ev_msg.event_code = event_code; + + ev_msg.dv[0].data_length = data_length; + ev_msg.dv[0].data_ptr = data; + ev_msg.dv[1].data_length = 0; + + ret = kev_post_msg(&ev_msg); + if (ret) { + printf("%s: kev_post_msg() failed, err %d\n", __func__, ret); } - return FALSE; + return ret; } -kern_return_t -memorystatus_update_vm_pressure(boolean_t target_foreground) { - boolean_t pressure_changed, pressured; - boolean_t warn = FALSE; +boolean_t +memorystatus_warn_process(pid_t pid, boolean_t critical) { - /* - * Centralised pressure handling routine. Called from: - * - The main jetsam thread. In this case, we update the pressure level and dispatch warnings to the foreground - * process *only*, each time the available page % drops. - * - The pageout scan path. In this scenario, every other registered process is targeted in footprint order. - * - * This scheme guarantees delivery to the foreground app, while providing for warnings to the remaining processes - * driven by the pageout scan. + boolean_t ret = FALSE; + boolean_t found_knote = FALSE; + struct knote *kn = NULL; + + /* + * See comment in sysctl_memorystatus_vm_pressure_send. */ - MEMORYSTATUS_DEBUG(1, "memorystatus_update_vm_pressure(): foreground %d; available %d, critical %d, pressure %d\n", - target_foreground, memorystatus_available_pages, memorystatus_available_pages_critical, memorystatus_available_pages_pressure); + memorystatus_klist_lock(); - proc_list_lock(); + SLIST_FOREACH(kn, &memorystatus_klist, kn_selnext) { + proc_t knote_proc = kn->kn_kq->kq_p; + pid_t knote_pid = knote_proc->p_pid; - pressure_changed = memorystatus_update_pressure_locked(&pressured); + if (knote_pid == pid) { + /* + * By setting the "fflags" here, we are forcing + * a process to deal with the case where it's + * bumping up into its memory limits. If we don't + * do this here, we will end up depending on the + * system pressure snapshot evaluation in + * filt_memorystatus(). + */ - if (pressured) { - if (target_foreground) { - if (memorystatus_available_pages != memorystatus_last_foreground_pressure_pages) { - if (memorystatus_available_pages < memorystatus_last_foreground_pressure_pages) { - warn = TRUE; + if (critical) { + if (kn->kn_sfflags & NOTE_MEMORYSTATUS_PRESSURE_CRITICAL) { + kn->kn_fflags = NOTE_MEMORYSTATUS_PRESSURE_CRITICAL; + } else if (kn->kn_sfflags & NOTE_MEMORYSTATUS_PRESSURE_WARN) { + kn->kn_fflags = NOTE_MEMORYSTATUS_PRESSURE_WARN; + } + } else { + if (kn->kn_sfflags & NOTE_MEMORYSTATUS_PRESSURE_WARN) { + kn->kn_fflags = NOTE_MEMORYSTATUS_PRESSURE_WARN; } - memorystatus_last_foreground_pressure_pages = memorystatus_available_pages; } - } else { - warn = TRUE; + + found_knote = TRUE; } - } else if (pressure_changed) { - memorystatus_last_foreground_pressure_pages = (unsigned int)-1; } - - proc_list_unlock(); - /* Target foreground processes if specified */ - if (warn) { - if (target_foreground) { - MEMORYSTATUS_DEBUG(1, "memorystatus_update_vm_pressure(): invoking vm_find_pressure_foreground_candidates()\n"); - vm_find_pressure_foreground_candidates(); - } else { - MEMORYSTATUS_DEBUG(1, "memorystatus_update_vm_pressure(): invoking vm_find_pressure_candidate()\n"); - /* Defer to VM code. This can race with the foreground priority, but - * it's preferable to holding onto locks for an extended period. */ - vm_find_pressure_candidate(); + if (found_knote) { + KNOTE(&memorystatus_klist, 0); + ret = TRUE; + } else { + if (vm_dispatch_pressure_note_to_pid(pid, FALSE) == 0) { + ret = TRUE; } } - - /* Dispatch the global kevent to privileged listeners */ - if (pressure_changed) { - memorystatus_issue_pressure_kevent(pressured); + + memorystatus_klist_unlock(); + + return ret; +} + +/* + * Can only be set by the current task on itself. + */ +int +memorystatus_low_mem_privileged_listener(uint32_t op_flags) +{ + boolean_t set_privilege = FALSE; + /* + * Need an entitlement check here? + */ + if (op_flags == MEMORYSTATUS_CMD_PRIVILEGED_LISTENER_ENABLE) { + set_privilege = TRUE; + } else if (op_flags == MEMORYSTATUS_CMD_PRIVILEGED_LISTENER_DISABLE) { + set_privilege = FALSE; + } else { + return EINVAL; } - return KERN_SUCCESS; + return (task_low_mem_privileged_listener(current_task(), set_privilege, NULL)); } int @@ -2540,6 +4394,26 @@ memorystatus_send_pressure_note(pid_t pid) { return memorystatus_send_note(kMemorystatusPressureNote, &pid, sizeof(pid)); } +void +memorystatus_send_low_swap_note(void) { + + struct knote *kn = NULL; + + memorystatus_klist_lock(); + SLIST_FOREACH(kn, &memorystatus_klist, kn_selnext) { + /* We call is_knote_registered_modify_task_pressure_bits to check if the sfflags for the + * current note contain NOTE_MEMORYSTATUS_LOW_SWAP. Once we find one note in the memorystatus_klist + * that has the NOTE_MEMORYSTATUS_LOW_SWAP flags in its sfflags set, we call KNOTE with + * kMemoryStatusLowSwap as the hint to process and update all knotes on the memorystatus_klist accordingly. */ + if (is_knote_registered_modify_task_pressure_bits(kn, NOTE_MEMORYSTATUS_LOW_SWAP, NULL, 0, 0) == TRUE) { + KNOTE(&memorystatus_klist, kMemorystatusLowSwap); + break; + } + } + + memorystatus_klist_unlock(); +} + boolean_t memorystatus_bg_pressure_eligible(proc_t p) { boolean_t eligible = FALSE; @@ -2563,8 +4437,7 @@ memorystatus_is_foreground_locked(proc_t p) { return ((p->p_memstat_effectivepriority == JETSAM_PRIORITY_FOREGROUND) || (p->p_memstat_effectivepriority == JETSAM_PRIORITY_FOREGROUND_SUPPORT)); } - -#else /* CONFIG_JETSAM && VM_PRESSURE_EVENTS */ +#endif /* CONFIG_MEMORYSTATUS */ /* * Trigger levels to test the mechanism. @@ -2581,10 +4454,10 @@ boolean_t memorystatus_manual_testing_on = FALSE; vm_pressure_level_t memorystatus_manual_testing_level = kVMPressureNormal; extern struct knote * -vm_pressure_select_optimal_candidate_to_notify(struct klist *, int); +vm_pressure_select_optimal_candidate_to_notify(struct klist *, int, boolean_t); extern -kern_return_t vm_pressure_notification_without_levels(void); +kern_return_t vm_pressure_notification_without_levels(boolean_t); extern void vm_pressure_klist_lock(void); extern void vm_pressure_klist_unlock(void); @@ -2611,8 +4484,6 @@ void memorystatus_on_pageout_scan_end(void) { * pressure_level_to_set - the task is about to be notified of this new level. Update the task's bit notification information appropriately. * */ -boolean_t -is_knote_registered_modify_task_pressure_bits(struct knote*, int, task_t, vm_pressure_level_t, vm_pressure_level_t); boolean_t is_knote_registered_modify_task_pressure_bits(struct knote *kn_max, int knote_pressure_level, task_t task, vm_pressure_level_t pressure_level_to_clear, vm_pressure_level_t pressure_level_to_set) @@ -2631,19 +4502,57 @@ is_knote_registered_modify_task_pressure_bits(struct knote *kn_max, int knote_pr return FALSE; } -extern kern_return_t vm_pressure_notify_dispatch_vm_clients(void); +extern kern_return_t vm_pressure_notify_dispatch_vm_clients(boolean_t target_foreground_process); + +#define VM_PRESSURE_DECREASED_SMOOTHING_PERIOD 5000 /* milliseconds */ kern_return_t -memorystatus_update_vm_pressure(boolean_t target_best_process) +memorystatus_update_vm_pressure(boolean_t target_foreground_process) { struct knote *kn_max = NULL; + struct knote *kn_cur = NULL, *kn_temp = NULL; /* for safe list traversal */ pid_t target_pid = -1; struct klist dispatch_klist = { NULL }; proc_t target_proc = PROC_NULL; - static vm_pressure_level_t level_snapshot = kVMPressureNormal; struct task *task = NULL; boolean_t found_candidate = FALSE; + static vm_pressure_level_t level_snapshot = kVMPressureNormal; + static vm_pressure_level_t prev_level_snapshot = kVMPressureNormal; + boolean_t smoothing_window_started = FALSE; + struct timeval smoothing_window_start_tstamp = {0, 0}; + struct timeval curr_tstamp = {0, 0}; + int elapsed_msecs = 0; + +#if !CONFIG_JETSAM +#define MAX_IDLE_KILLS 100 /* limit the number of idle kills allowed */ + + int idle_kill_counter = 0; + + /* + * On desktop we take this opportunity to free up memory pressure + * by immediately killing idle exitable processes. We use a delay + * to avoid overkill. And we impose a max counter as a fail safe + * in case daemons re-launch too fast. + */ + while ((memorystatus_vm_pressure_level != kVMPressureNormal) && (idle_kill_counter < MAX_IDLE_KILLS)) { + if (memorystatus_idle_exit_from_VM() == FALSE) { + /* No idle exitable processes left to kill */ + break; + } + idle_kill_counter++; + + if (memorystatus_manual_testing_on == TRUE) { + /* + * Skip the delay when testing + * the pressure notification scheme. + */ + } else { + delay(1000000); /* 1 second */ + } + } +#endif /* !CONFIG_JETSAM */ + while (1) { /* @@ -2652,8 +4561,33 @@ memorystatus_update_vm_pressure(boolean_t target_best_process) */ level_snapshot = memorystatus_vm_pressure_level; + if (prev_level_snapshot > level_snapshot) { + /* + * Pressure decreased? Let's take a little breather + * and see if this condition stays. + */ + if (smoothing_window_started == FALSE) { + + smoothing_window_started = TRUE; + microuptime(&smoothing_window_start_tstamp); + } + + microuptime(&curr_tstamp); + timevalsub(&curr_tstamp, &smoothing_window_start_tstamp); + elapsed_msecs = curr_tstamp.tv_sec * 1000 + curr_tstamp.tv_usec / 1000; + + if (elapsed_msecs < VM_PRESSURE_DECREASED_SMOOTHING_PERIOD) { + + delay(INTER_NOTIFICATION_DELAY); + continue; + } + } + + prev_level_snapshot = level_snapshot; + smoothing_window_started = FALSE; + memorystatus_klist_lock(); - kn_max = vm_pressure_select_optimal_candidate_to_notify(&memorystatus_klist, level_snapshot); + kn_max = vm_pressure_select_optimal_candidate_to_notify(&memorystatus_klist, level_snapshot, target_foreground_process); if (kn_max == NULL) { memorystatus_klist_unlock(); @@ -2687,7 +4621,6 @@ memorystatus_update_vm_pressure(boolean_t target_best_process) continue; } proc_list_unlock(); - memorystatus_klist_unlock(); target_pid = target_proc->p_pid; @@ -2719,43 +4652,43 @@ memorystatus_update_vm_pressure(boolean_t target_best_process) } if (found_candidate == FALSE) { + proc_rele(target_proc); + memorystatus_klist_unlock(); continue; } - memorystatus_klist_lock(); - KNOTE_DETACH(&memorystatus_klist, kn_max); - KNOTE_ATTACH(&dispatch_klist, kn_max); - memorystatus_klist_unlock(); + SLIST_FOREACH_SAFE(kn_cur, &memorystatus_klist, kn_selnext, kn_temp) { + proc_t knote_proc = kn_cur->kn_kq->kq_p; + pid_t knote_pid = knote_proc->p_pid; + if (knote_pid == target_pid) { + KNOTE_DETACH(&memorystatus_klist, kn_cur); + KNOTE_ATTACH(&dispatch_klist, kn_cur); + } + } KNOTE(&dispatch_klist, (level_snapshot != kVMPressureNormal) ? kMemorystatusPressure : kMemorystatusNoPressure); - memorystatus_klist_lock(); - KNOTE_DETACH(&dispatch_klist, kn_max); - KNOTE_ATTACH(&memorystatus_klist, kn_max); + SLIST_FOREACH_SAFE(kn_cur, &dispatch_klist, kn_selnext, kn_temp) { + KNOTE_DETACH(&dispatch_klist, kn_cur); + KNOTE_ATTACH(&memorystatus_klist, kn_cur); + } + memorystatus_klist_unlock(); microuptime(&target_proc->vm_pressure_last_notify_tstamp); proc_rele(target_proc); - if (target_best_process == TRUE) { + if (memorystatus_manual_testing_on == TRUE && target_foreground_process == TRUE) { break; } try_dispatch_vm_clients: - if (level_snapshot != kVMPressureNormal) { - /* - * Wake up idle-exit thread. - * Targets one process per invocation. - * - * TODO: memorystatus_idle_exit_from_VM should return FALSE once it's - * done with all idle-exitable processes. Currently, we will exit this - * loop when we are done with notification clients (level and non-level based) - * but we may still have some idle-exitable processes around. - * + if (kn_max == NULL && level_snapshot != kVMPressureNormal) { + /* + * We will exit this loop when we are done with + * notification clients (level and non-level based). */ - memorystatus_idle_exit_from_VM(); - - if ((vm_pressure_notify_dispatch_vm_clients() == KERN_FAILURE) && (kn_max == NULL)) { + if ((vm_pressure_notify_dispatch_vm_clients(target_foreground_process) == KERN_FAILURE) && (kn_max == NULL)) { /* * kn_max == NULL i.e. we didn't find any eligible clients for the level-based notifications * AND @@ -2767,8 +4700,45 @@ try_dispatch_vm_clients: } } - if (memorystatus_manual_testing_on == FALSE) { - delay(INTER_NOTIFICATION_DELAY); + /* + * LD: This block of code below used to be invoked in the older memory notification scheme on embedded everytime + * a process was sent a memory pressure notification. The "memorystatus_klist" list was used to hold these + * privileged listeners. But now we have moved to the newer scheme and are trying to move away from the extra + * notifications. So the code is here in case we break compat. and need to send out notifications to the privileged + * apps. + */ +#if 0 +#endif /* 0 */ + + if (memorystatus_manual_testing_on == TRUE) { + /* + * Testing out the pressure notification scheme. + * No need for delays etc. + */ + } else { + + uint32_t sleep_interval = INTER_NOTIFICATION_DELAY; +#if CONFIG_JETSAM + unsigned int page_delta = 0; + unsigned int skip_delay_page_threshold = 0; + + assert(memorystatus_available_pages_pressure >= memorystatus_available_pages_critical_base); + + page_delta = (memorystatus_available_pages_pressure - memorystatus_available_pages_critical_base) / 2; + skip_delay_page_threshold = memorystatus_available_pages_pressure - page_delta; + + if (memorystatus_available_pages <= skip_delay_page_threshold) { + /* + * We are nearing the critcal mark fast and can't afford to wait between + * notifications. + */ + sleep_interval = 0; + } +#endif /* CONFIG_JETSAM */ + + if (sleep_interval) { + delay(sleep_interval); + } } } @@ -2815,15 +4785,22 @@ static int sysctl_memorystatus_vm_pressure_level SYSCTL_HANDLER_ARGS { #pragma unused(arg1, arg2, oidp) - vm_pressure_level_t dispatch_level = convert_internal_pressure_level_to_dispatch_level(memorystatus_vm_pressure_level); return SYSCTL_OUT(req, &dispatch_level, sizeof(dispatch_level)); } +#if DEBUG || DEVELOPMENT + SYSCTL_PROC(_kern, OID_AUTO, memorystatus_vm_pressure_level, CTLTYPE_INT|CTLFLAG_RD|CTLFLAG_LOCKED, 0, 0, &sysctl_memorystatus_vm_pressure_level, "I", ""); +#else /* DEBUG || DEVELOPMENT */ + +SYSCTL_PROC(_kern, OID_AUTO, memorystatus_vm_pressure_level, CTLTYPE_INT|CTLFLAG_RD|CTLFLAG_LOCKED|CTLFLAG_MASKED, + 0, 0, &sysctl_memorystatus_vm_pressure_level, "I", ""); + +#endif /* DEBUG || DEVELOPMENT */ extern int memorystatus_purge_on_warning; extern int memorystatus_purge_on_critical; @@ -2928,7 +4905,7 @@ sysctl_memorypressure_manual_trigger SYSCTL_HANDLER_ARGS } else { vm_pressure_klist_lock(); - vm_pressure_notification_without_levels(); + vm_pressure_notification_without_levels(FALSE); vm_pressure_klist_unlock(); } @@ -2948,7 +4925,7 @@ SYSCTL_INT(_kern, OID_AUTO, memorystatus_purge_on_urgent, CTLTYPE_INT|CTLFLAG_RW SYSCTL_INT(_kern, OID_AUTO, memorystatus_purge_on_critical, CTLTYPE_INT|CTLFLAG_RW|CTLFLAG_LOCKED, &memorystatus_purge_on_critical, 0, ""); -#endif /* CONFIG_JETSAM && VM_PRESSURE_EVENTS */ +#endif /* VM_PRESSURE_EVENTS */ /* Return both allocated and actual size, since there's a race between allocation and list compilation */ static int @@ -2991,12 +4968,18 @@ memorystatus_get_priority_list(memorystatus_priority_entry_t **list_ptr, size_t list_entry->priority = p->p_memstat_effectivepriority; list_entry->user_data = p->p_memstat_userdata; #if LEGACY_HIWATER - if (((p->p_memstat_state & P_MEMSTAT_MEMLIMIT_BACKGROUND) && (p->p_memstat_effectivepriority >= JETSAM_PRIORITY_FOREGROUND)) || - (p->p_memstat_memlimit <= 0)) { - task_get_phys_footprint_limit(p->task, &list_entry->limit); - } else { - list_entry->limit = p->p_memstat_memlimit; - } + + /* + * No need to consider P_MEMSTAT_MEMLIMIT_BACKGROUND anymore. + * Background limits are described via the inactive limit slots. + * So, here, the cached limit should always be valid. + */ + + if (p->p_memstat_memlimit <= 0) { + task_get_phys_footprint_limit(p->task, &list_entry->limit); + } else { + list_entry->limit = p->p_memstat_memlimit; + } #else task_get_phys_footprint_limit(p->task, &list_entry->limit); #endif @@ -3072,19 +5055,34 @@ memorystatus_clear_errors(void) static void memorystatus_update_levels_locked(boolean_t critical_only) { + memorystatus_available_pages_critical = memorystatus_available_pages_critical_base; -#if !LATENCY_JETSAM - { - // If there's an entry in the first bucket, we have idle processes - memstat_bucket_t *first_bucket = &memstat_bucket[JETSAM_PRIORITY_IDLE]; - if (first_bucket->count) { - memorystatus_available_pages_critical += memorystatus_available_pages_critical_idle_offset; + + /* + * If there's an entry in the first bucket, we have idle processes. + */ + memstat_bucket_t *first_bucket = &memstat_bucket[JETSAM_PRIORITY_IDLE]; + if (first_bucket->count) { + memorystatus_available_pages_critical += memorystatus_available_pages_critical_idle_offset; + + if (memorystatus_available_pages_critical > memorystatus_available_pages_pressure ) { + /* + * The critical threshold must never exceed the pressure threshold + */ + memorystatus_available_pages_critical = memorystatus_available_pages_pressure; } } -#endif + #if DEBUG || DEVELOPMENT if (memorystatus_jetsam_policy & kPolicyDiagnoseActive) { memorystatus_available_pages_critical += memorystatus_jetsam_policy_offset_pages_diagnostic; + + if (memorystatus_available_pages_critical > memorystatus_available_pages_pressure ) { + /* + * The critical threshold must never exceed the pressure threshold + */ + memorystatus_available_pages_critical = memorystatus_available_pages_pressure; + } } #endif @@ -3102,10 +5100,94 @@ memorystatus_update_levels_locked(boolean_t critical_only) { #endif } +/* + * Get the at_boot snapshot + */ +static int +memorystatus_get_at_boot_snapshot(memorystatus_jetsam_snapshot_t **snapshot, size_t *snapshot_size, boolean_t size_only) { + size_t input_size = *snapshot_size; + + /* + * The at_boot snapshot has no entry list. + */ + *snapshot_size = sizeof(memorystatus_jetsam_snapshot_t); + + if (size_only) { + return 0; + } + + /* + * Validate the size of the snapshot buffer + */ + if (input_size < *snapshot_size) { + return EINVAL; + } + + /* + * Update the notification_time only + */ + memorystatus_at_boot_snapshot.notification_time = mach_absolute_time(); + *snapshot = &memorystatus_at_boot_snapshot; + + MEMORYSTATUS_DEBUG(7, "memorystatus_get_at_boot_snapshot: returned inputsize (%ld), snapshot_size(%ld), listcount(%d)\n", + (long)input_size, (long)*snapshot_size, 0); + return 0; +} + static int -memorystatus_get_snapshot(memorystatus_jetsam_snapshot_t **snapshot, size_t *snapshot_size, boolean_t size_only) { +memorystatus_get_on_demand_snapshot(memorystatus_jetsam_snapshot_t **snapshot, size_t *snapshot_size, boolean_t size_only) { size_t input_size = *snapshot_size; + uint32_t ods_list_count = memorystatus_list_count; + memorystatus_jetsam_snapshot_t *ods = NULL; /* The on_demand snapshot buffer */ + + *snapshot_size = sizeof(memorystatus_jetsam_snapshot_t) + (sizeof(memorystatus_jetsam_snapshot_entry_t) * (ods_list_count)); + + if (size_only) { + return 0; + } + + /* + * Validate the size of the snapshot buffer. + * This is inherently racey. May want to revisit + * this error condition and trim the output when + * it doesn't fit. + */ + if (input_size < *snapshot_size) { + return EINVAL; + } + + /* + * Allocate and initialize a snapshot buffer. + */ + ods = (memorystatus_jetsam_snapshot_t *)kalloc(*snapshot_size); + if (!ods) { + return (ENOMEM); + } + + memset(ods, 0, *snapshot_size); + + proc_list_lock(); + memorystatus_init_jetsam_snapshot_locked(ods, ods_list_count); + proc_list_unlock(); + + /* + * Return the kernel allocated, on_demand buffer. + * The caller of this routine will copy the data out + * to user space and then free the kernel allocated + * buffer. + */ + *snapshot = ods; + + MEMORYSTATUS_DEBUG(7, "memorystatus_get_on_demand_snapshot: returned inputsize (%ld), snapshot_size(%ld), listcount(%ld)\n", + (long)input_size, (long)*snapshot_size, (long)ods_list_count); + return 0; +} + +static int +memorystatus_get_jetsam_snapshot(memorystatus_jetsam_snapshot_t **snapshot, size_t *snapshot_size, boolean_t size_only) { + size_t input_size = *snapshot_size; + if (memorystatus_jetsam_snapshot_count > 0) { *snapshot_size = sizeof(memorystatus_jetsam_snapshot_t) + (sizeof(memorystatus_jetsam_snapshot_entry_t) * (memorystatus_jetsam_snapshot_count)); } else { @@ -3121,29 +5203,99 @@ memorystatus_get_snapshot(memorystatus_jetsam_snapshot_t **snapshot, size_t *sna } *snapshot = memorystatus_jetsam_snapshot; - - MEMORYSTATUS_DEBUG(1, "memorystatus_snapshot: returning %ld for size\n", (long)*snapshot_size); - + + MEMORYSTATUS_DEBUG(7, "memorystatus_get_jetsam_snapshot: returned inputsize (%ld), snapshot_size(%ld), listcount(%ld)\n", + (long)input_size, (long)*snapshot_size, (long)memorystatus_jetsam_snapshot_count); + return 0; } + static int -memorystatus_cmd_get_jetsam_snapshot(user_addr_t buffer, size_t buffer_size, int32_t *retval) { +memorystatus_cmd_get_jetsam_snapshot(int32_t flags, user_addr_t buffer, size_t buffer_size, int32_t *retval) { int error = EINVAL; boolean_t size_only; + boolean_t is_default_snapshot = FALSE; + boolean_t is_on_demand_snapshot = FALSE; + boolean_t is_at_boot_snapshot = FALSE; memorystatus_jetsam_snapshot_t *snapshot; - + size_only = ((buffer == USER_ADDR_NULL) ? TRUE : FALSE); - - error = memorystatus_get_snapshot(&snapshot, &buffer_size, size_only); + + if (flags == 0) { + /* Default */ + is_default_snapshot = TRUE; + error = memorystatus_get_jetsam_snapshot(&snapshot, &buffer_size, size_only); + } else { + if (flags & ~(MEMORYSTATUS_SNAPSHOT_ON_DEMAND | MEMORYSTATUS_SNAPSHOT_AT_BOOT)) { + /* + * Unsupported bit set in flag. + */ + return EINVAL; + } + + if ((flags & (MEMORYSTATUS_SNAPSHOT_ON_DEMAND | MEMORYSTATUS_SNAPSHOT_AT_BOOT)) == + (MEMORYSTATUS_SNAPSHOT_ON_DEMAND | MEMORYSTATUS_SNAPSHOT_AT_BOOT)) { + /* + * Can't have both set at the same time. + */ + return EINVAL; + } + + if (flags & MEMORYSTATUS_SNAPSHOT_ON_DEMAND) { + is_on_demand_snapshot = TRUE; + /* + * When not requesting the size only, the following call will allocate + * an on_demand snapshot buffer, which is freed below. + */ + error = memorystatus_get_on_demand_snapshot(&snapshot, &buffer_size, size_only); + + } else if (flags & MEMORYSTATUS_SNAPSHOT_AT_BOOT) { + is_at_boot_snapshot = TRUE; + error = memorystatus_get_at_boot_snapshot(&snapshot, &buffer_size, size_only); + } else { + /* + * Invalid flag setting. + */ + return EINVAL; + } + } + if (error) { goto out; } - /* Copy out and reset */ + /* + * Copy the data out to user space and clear the snapshot buffer. + * If working with the jetsam snapshot, + * clearing the buffer means, reset the count. + * If working with an on_demand snapshot + * clearing the buffer means, free it. + * If working with the at_boot snapshot + * there is nothing to clear or update. + */ if (!size_only) { if ((error = copyout(snapshot, buffer, buffer_size)) == 0) { - snapshot->entry_count = memorystatus_jetsam_snapshot_count = 0; + if (is_default_snapshot) { + /* + * The jetsam snapshot is never freed, its count is simply reset. + */ + snapshot->entry_count = memorystatus_jetsam_snapshot_count = 0; + + proc_list_lock(); + memorystatus_jetsam_snapshot_last_timestamp = 0; + proc_list_unlock(); + } + } + + if (is_on_demand_snapshot) { + /* + * The on_demand snapshot is always freed, + * even if the copyout failed. + */ + if(snapshot) { + kfree(snapshot, buffer_size); + } } } @@ -3154,58 +5306,313 @@ out: return error; } +/* + * Routine: memorystatus_cmd_grp_set_properties + * Purpose: Update properties for a group of processes. + * + * Supported Properties: + * [priority] + * Move each process out of its effective priority + * band and into a new priority band. + * Maintains relative order from lowest to highest priority. + * In single band, maintains relative order from head to tail. + * + * eg: before [effectivepriority | pid] + * [18 | p101 ] + * [17 | p55, p67, p19 ] + * [12 | p103 p10 ] + * [ 7 | p25 ] + * [ 0 | p71, p82, ] + * + * after [ new band | pid] + * [ xxx | p71, p82, p25, p103, p10, p55, p67, p19, p101] + * + * Returns: 0 on success, else non-zero. + * + * Caveat: We know there is a race window regarding recycled pids. + * A process could be killed before the kernel can act on it here. + * If a pid cannot be found in any of the jetsam priority bands, + * then we simply ignore it. No harm. + * But, if the pid has been recycled then it could be an issue. + * In that scenario, we might move an unsuspecting process to the new + * priority band. It's not clear how the kernel can safeguard + * against this, but it would be an extremely rare case anyway. + * The caller of this api might avoid such race conditions by + * ensuring that the processes passed in the pid list are suspended. + */ + + +/* This internal structure can expand when we add support for more properties */ +typedef struct memorystatus_internal_properties +{ + proc_t proc; + int32_t priority; /* see memorytstatus_priority_entry_t : priority */ +} memorystatus_internal_properties_t; + + static int -memorystatus_cmd_set_priority_properties(pid_t pid, user_addr_t buffer, size_t buffer_size, __unused int32_t *retval) { - const uint32_t MAX_ENTRY_COUNT = 2; /* Cap the entry count */ +memorystatus_cmd_grp_set_properties(int32_t flags, user_addr_t buffer, size_t buffer_size, __unused int32_t *retval) { + +#pragma unused (flags) + + /* + * We only handle setting priority + * per process + */ + + int error = 0; + memorystatus_priority_entry_t *entries = NULL; + uint32_t entry_count = 0; + + /* This will be the ordered proc list */ + memorystatus_internal_properties_t *table = NULL; + size_t table_size = 0; + uint32_t table_count = 0; + + uint32_t i = 0; + uint32_t bucket_index = 0; + boolean_t head_insert; + int32_t new_priority; - int error; - uint32_t i; - uint32_t entry_count; - memorystatus_priority_properties_t *entries; + proc_t p; + + /* Verify inputs */ + if ((buffer == USER_ADDR_NULL) || (buffer_size == 0) || ((buffer_size % sizeof(memorystatus_priority_entry_t)) != 0)) { + error = EINVAL; + goto out; + } + + entry_count = (buffer_size / sizeof(memorystatus_priority_entry_t)); + if ((entries = (memorystatus_priority_entry_t *)kalloc(buffer_size)) == NULL) { + error = ENOMEM; + goto out; + } + + KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_GRP_SET_PROP) | DBG_FUNC_START, entry_count, 0, 0, 0, 0); + + if ((error = copyin(buffer, entries, buffer_size)) != 0) { + goto out; + } + + /* Verify sanity of input priorities */ + for (i=0; i < entry_count; i++) { + if (entries[i].priority == -1) { + /* Use as shorthand for default priority */ + entries[i].priority = JETSAM_PRIORITY_DEFAULT; + } else if (entries[i].priority == JETSAM_PRIORITY_IDLE_DEFERRED) { + /* JETSAM_PRIORITY_IDLE_DEFERRED is reserved for internal use; + * if requested, adjust to JETSAM_PRIORITY_IDLE. */ + entries[i].priority = JETSAM_PRIORITY_IDLE; + } else if (entries[i].priority == JETSAM_PRIORITY_IDLE_HEAD) { + /* JETSAM_PRIORITY_IDLE_HEAD inserts at the head of the idle + * queue */ + /* Deal with this later */ + } else if ((entries[i].priority < 0) || (entries[i].priority >= MEMSTAT_BUCKET_COUNT)) { + /* Sanity check */ + error = EINVAL; + goto out; + } + } + + table_size = sizeof(memorystatus_internal_properties_t) * entry_count; + if ( (table = (memorystatus_internal_properties_t *)kalloc(table_size)) == NULL) { + error = ENOMEM; + goto out; + } + memset(table, 0, table_size); + + + /* + * For each jetsam bucket entry, spin through the input property list. + * When a matching pid is found, populate an adjacent table with the + * appropriate proc pointer and new property values. + * This traversal automatically preserves order from lowest + * to highest priority. + */ + + bucket_index=0; - /* Validate inputs */ - if ((pid == 0) || (buffer == USER_ADDR_NULL) || (buffer_size == 0)) { - return EINVAL; + proc_list_lock(); + + /* Create the ordered table */ + p = memorystatus_get_first_proc_locked(&bucket_index, TRUE); + while (p && (table_count < entry_count)) { + for (i=0; i < entry_count; i++ ) { + if (p->p_pid == entries[i].pid) { + /* Build the table data */ + table[table_count].proc = p; + table[table_count].priority = entries[i].priority; + table_count++; + break; + } + } + p = memorystatus_get_next_proc_locked(&bucket_index, p, TRUE); + } + + /* We now have ordered list of procs ready to move */ + for (i=0; i < table_count; i++) { + p = table[i].proc; + assert(p != NULL); + + /* Allow head inserts -- but relative order is now */ + if (table[i].priority == JETSAM_PRIORITY_IDLE_HEAD) { + new_priority = JETSAM_PRIORITY_IDLE; + head_insert = true; + } else { + new_priority = table[i].priority; + head_insert = false; + } + + /* Not allowed */ + if (p->p_memstat_state & P_MEMSTAT_INTERNAL) { + continue; + } + + /* + * Take appropriate steps if moving proc out of the + * JETSAM_PRIORITY_IDLE_DEFERRED band. + */ + if (p->p_memstat_effectivepriority == JETSAM_PRIORITY_IDLE_DEFERRED) { + memorystatus_invalidate_idle_demotion_locked(p, TRUE); + } + + memorystatus_update_priority_locked(p, new_priority, head_insert); } + + proc_list_unlock(); + + /* + * if (table_count != entry_count) + * then some pids were not found in a jetsam band. + * harmless but interesting... + */ + KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_MEMSTAT, BSD_MEMSTAT_GRP_SET_PROP) | DBG_FUNC_END, entry_count, table_count, 0, 0, 0); + +out: + if (entries) + kfree(entries, buffer_size); + if (table) + kfree(table, table_size); + + return (error); +} + + +/* + * This routine is used to update a process's jetsam priority position and stored user_data. + * It is not used for the setting of memory limits, which is why the last 6 args to the + * memorystatus_update() call are 0 or FALSE. + */ - /* Make sure the buffer is a multiple of the entry size, and that an excessive size isn't specified */ - entry_count = (buffer_size / sizeof(memorystatus_priority_properties_t)); - if (((buffer_size % sizeof(memorystatus_priority_properties_t)) != 0) || (entry_count > MAX_ENTRY_COUNT)) { +static int +memorystatus_cmd_set_priority_properties(pid_t pid, user_addr_t buffer, size_t buffer_size, __unused int32_t *retval) { + int error = 0; + memorystatus_priority_properties_t mpp_entry; + + /* Validate inputs */ + if ((pid == 0) || (buffer == USER_ADDR_NULL) || (buffer_size != sizeof(memorystatus_priority_properties_t))) { return EINVAL; } - - entries = (memorystatus_priority_properties_t *)kalloc(buffer_size); - - error = copyin(buffer, entries, buffer_size); - for (i = 0; i < entry_count; i++) { + error = copyin(buffer, &mpp_entry, buffer_size); + + if (error == 0) { proc_t p; - if (error) { - break; - } - p = proc_find(pid); if (!p) { - error = ESRCH; - break; + return ESRCH; } if (p->p_memstat_state & P_MEMSTAT_INTERNAL) { - error = EPERM; proc_rele(p); - break; + return EPERM; } - - error = memorystatus_update(p, entries[i].priority, entries[i].user_data, FALSE, FALSE, 0, 0); + + error = memorystatus_update(p, mpp_entry.priority, mpp_entry.user_data, FALSE, FALSE, 0, 0, FALSE, FALSE, FALSE); proc_rele(p); } - kfree(entries, buffer_size); - - return error; + return(error); } +static int +memorystatus_cmd_set_memlimit_properties(pid_t pid, user_addr_t buffer, size_t buffer_size, __unused int32_t *retval) { + int error = 0; + memorystatus_memlimit_properties_t mmp_entry; + + /* Validate inputs */ + if ((pid == 0) || (buffer == USER_ADDR_NULL) || (buffer_size != sizeof(memorystatus_memlimit_properties_t))) { + return EINVAL; + } + + error = copyin(buffer, &mmp_entry, buffer_size); + + if (error == 0) { + error = memorystatus_set_memlimit_properties(pid, &mmp_entry); + } + + return(error); +} + +/* + * When getting the memlimit settings, we can't simply call task_get_phys_footprint_limit(). + * That gets the proc's cached memlimit and there is no guarantee that the active/inactive + * limits will be the same in the no-limit case. Instead we convert limits <= 0 using + * task_convert_phys_footprint_limit(). It computes the same limit value that would be written + * to the task's ledgers via task_set_phys_footprint_limit(). + */ +static int +memorystatus_cmd_get_memlimit_properties(pid_t pid, user_addr_t buffer, size_t buffer_size, __unused int32_t *retval) { + int error = 0; + memorystatus_memlimit_properties_t mmp_entry; + + /* Validate inputs */ + if ((pid == 0) || (buffer == USER_ADDR_NULL) || (buffer_size != sizeof(memorystatus_memlimit_properties_t))) { + return EINVAL; + } + + memset (&mmp_entry, 0, sizeof(memorystatus_memlimit_properties_t)); + + proc_t p = proc_find(pid); + if (!p) { + return ESRCH; + } + + /* + * Get the active limit and attributes. + * No locks taken since we hold a reference to the proc. + */ + + if (p->p_memstat_memlimit_active > 0 ) { + mmp_entry.memlimit_active = p->p_memstat_memlimit_active; + } else { + task_convert_phys_footprint_limit(-1, &mmp_entry.memlimit_active); + } + + if (p->p_memstat_state & P_MEMSTAT_MEMLIMIT_ACTIVE_FATAL) { + mmp_entry.memlimit_active_attr |= MEMORYSTATUS_MEMLIMIT_ATTR_FATAL; + } + + /* + * Get the inactive limit and attributes + */ + if (p->p_memstat_memlimit_inactive <= 0) { + task_convert_phys_footprint_limit(-1, &mmp_entry.memlimit_inactive); + } else { + mmp_entry.memlimit_inactive = p->p_memstat_memlimit_inactive; + } + if (p->p_memstat_state & P_MEMSTAT_MEMLIMIT_INACTIVE_FATAL) { + mmp_entry.memlimit_inactive_attr |= MEMORYSTATUS_MEMLIMIT_ATTR_FATAL; + } + proc_rele(p); + + error = copyout(&mmp_entry, buffer, buffer_size); + + return(error); +} + + static int memorystatus_cmd_get_pressure_status(int32_t *retval) { int error; @@ -3222,42 +5629,235 @@ memorystatus_cmd_get_pressure_status(int32_t *retval) { return error; } +int +memorystatus_get_pressure_status_kdp() { + return (kVMPressureNormal != memorystatus_vm_pressure_level) ? 1 : 0; +} + +/* + * Every process, including a P_MEMSTAT_INTERNAL process (currently only pid 1), is allowed to set a HWM. + * + * This call is inflexible -- it does not distinguish between active/inactive, fatal/non-fatal + * So, with 2-level HWM preserving previous behavior will map as follows. + * - treat the limit passed in as both an active and inactive limit. + * - treat the is_fatal_limit flag as though it applies to both active and inactive limits. + * + * When invoked via MEMORYSTATUS_CMD_SET_JETSAM_HIGH_WATER_MARK + * - the is_fatal_limit is FALSE, meaning the active and inactive limits are non-fatal/soft + * - so mapping is (active/non-fatal, inactive/non-fatal) + * + * When invoked via MEMORYSTATUS_CMD_SET_JETSAM_TASK_LIMIT + * - the is_fatal_limit is TRUE, meaning the process's active and inactive limits are fatal/hard + * - so mapping is (active/fatal, inactive/fatal) + */ + static int -memorystatus_cmd_set_jetsam_high_water_mark(pid_t pid, int32_t high_water_mark, __unused int32_t *retval) { +memorystatus_cmd_set_jetsam_memory_limit(pid_t pid, int32_t high_water_mark, __unused int32_t *retval, boolean_t is_fatal_limit) { int error = 0; + memorystatus_memlimit_properties_t entry; + + entry.memlimit_active = high_water_mark; + entry.memlimit_active_attr = 0; + entry.memlimit_inactive = high_water_mark; + entry.memlimit_inactive_attr = 0; + + if (is_fatal_limit == TRUE) { + entry.memlimit_active_attr |= MEMORYSTATUS_MEMLIMIT_ATTR_FATAL; + entry.memlimit_inactive_attr |= MEMORYSTATUS_MEMLIMIT_ATTR_FATAL; + } + + error = memorystatus_set_memlimit_properties(pid, &entry); + return (error); +} + +static int +memorystatus_set_memlimit_properties(pid_t pid, memorystatus_memlimit_properties_t *entry) { + + int32_t memlimit_active; + boolean_t memlimit_active_is_fatal; + int32_t memlimit_inactive; + boolean_t memlimit_inactive_is_fatal; + uint32_t valid_attrs = 0; + int error = 0; proc_t p = proc_find(pid); if (!p) { return ESRCH; } - - if (high_water_mark <= 0) { - high_water_mark = -1; /* Disable */ + + /* + * Check for valid attribute flags. + */ + valid_attrs |= (MEMORYSTATUS_MEMLIMIT_ATTR_FATAL); + if ((entry->memlimit_active_attr & (~valid_attrs)) != 0) { + proc_rele(p); + return EINVAL; } - - proc_list_lock(); - - if (p->p_memstat_state & P_MEMSTAT_INTERNAL) { - error = EPERM; - goto exit; + if ((entry->memlimit_inactive_attr & (~valid_attrs)) != 0) { + proc_rele(p); + return EINVAL; } - - p->p_memstat_memlimit = high_water_mark; + + /* + * Setup the active memlimit properties + */ + memlimit_active = entry->memlimit_active; + if (entry->memlimit_active_attr & MEMORYSTATUS_MEMLIMIT_ATTR_FATAL) { + memlimit_active_is_fatal = TRUE; + } else { + memlimit_active_is_fatal = FALSE; + } + + /* + * Setup the inactive memlimit properties + */ + memlimit_inactive = entry->memlimit_inactive; + if (entry->memlimit_inactive_attr & MEMORYSTATUS_MEMLIMIT_ATTR_FATAL) { + memlimit_inactive_is_fatal = TRUE; + } else { + memlimit_inactive_is_fatal = FALSE; + } + + /* + * Setting a limit of <= 0 implies that the process has no + * high-water-mark and has no per-task-limit. That means + * the system_wide task limit is in place, which by the way, + * is always fatal. + */ + + if (memlimit_active <= 0) { + /* + * Enforce the fatal system_wide task limit while process is active. + */ + memlimit_active = -1; + memlimit_active_is_fatal = TRUE; + } + + if (memlimit_inactive <= 0) { + /* + * Enforce the fatal system_wide task limit while process is inactive. + */ + memlimit_inactive = -1; + memlimit_inactive_is_fatal = TRUE; + } + + proc_list_lock(); + + /* + * Store the active limit variants in the proc. + */ + SET_ACTIVE_LIMITS_LOCKED(p, memlimit_active, memlimit_active_is_fatal); + + /* + * Store the inactive limit variants in the proc. + */ + SET_INACTIVE_LIMITS_LOCKED(p, memlimit_inactive, memlimit_inactive_is_fatal); + + /* + * Enforce appropriate limit variant by updating the cached values + * and writing the ledger. + * Limit choice is based on process active/inactive state. + */ + if (memorystatus_highwater_enabled) { - if (p->p_memstat_state & P_MEMSTAT_MEMLIMIT_BACKGROUND) { - memorystatus_update_priority_locked(p, p->p_memstat_effectivepriority); - } else { - error = (task_set_phys_footprint_limit_internal(p->task, high_water_mark, NULL, TRUE) == 0) ? 0 : EINVAL; - } + boolean_t trigger_exception; + /* + * No need to consider P_MEMSTAT_MEMLIMIT_BACKGROUND anymore. + * Background limits are described via the inactive limit slots. + */ + + if (proc_jetsam_state_is_active_locked(p) == TRUE) { + CACHE_ACTIVE_LIMITS_LOCKED(p, trigger_exception); + } else { + CACHE_INACTIVE_LIMITS_LOCKED(p, trigger_exception); + } + + /* Enforce the limit by writing to the ledgers */ + assert(trigger_exception == TRUE); + error = (task_set_phys_footprint_limit_internal(p->task, ((p->p_memstat_memlimit > 0) ? p->p_memstat_memlimit : -1), NULL, trigger_exception) == 0) ? 0 : EINVAL; + + MEMORYSTATUS_DEBUG(3, "memorystatus_set_memlimit_properties: new limit on pid %d (%dMB %s) current priority (%d) dirty_state?=0x%x %s\n", + p->p_pid, (p->p_memstat_memlimit > 0 ? p->p_memstat_memlimit : -1), + (p->p_memstat_state & P_MEMSTAT_FATAL_MEMLIMIT ? "F " : "NF"), p->p_memstat_effectivepriority, p->p_memstat_dirty, + (p->p_memstat_dirty ? ((p->p_memstat_dirty & P_DIRTY) ? "isdirty" : "isclean") : "")); } -exit: proc_list_unlock(); proc_rele(p); return error; } +/* + * Returns the jetsam priority (effective or requested) of the process + * associated with this task. + */ +int +proc_get_memstat_priority(proc_t p, boolean_t effective_priority) +{ + if (p) { + if (effective_priority) { + return p->p_memstat_effectivepriority; + } else { + return p->p_memstat_requestedpriority; + } + } + return 0; +} + +/* + * Description: + * Evaluates active vs. inactive process state. + * Processes that opt into dirty tracking are evaluated + * based on clean vs dirty state. + * dirty ==> active + * clean ==> inactive + * + * Process that do not opt into dirty tracking are + * evalulated based on priority level. + * Foreground or above ==> active + * Below Foreground ==> inactive + * + * Return: TRUE if active + * False if inactive + */ + +static boolean_t +proc_jetsam_state_is_active_locked(proc_t p) { + + if (p->p_memstat_dirty & P_DIRTY_TRACK) { + /* + * process has opted into dirty tracking + * active state is based on dirty vs. clean + */ + if (p->p_memstat_dirty & P_DIRTY_IS_DIRTY) { + /* + * process is dirty + * implies active state + */ + return TRUE; + } else { + /* + * process is clean + * implies inactive state + */ + return FALSE; + } + } else if (p->p_memstat_effectivepriority >= JETSAM_PRIORITY_FOREGROUND) { + /* + * process is Foreground or higher + * implies active state + */ + return TRUE; + } else { + /* + * process found below Foreground + * implies inactive state + */ + return FALSE; + } +} + #endif /* CONFIG_JETSAM */ int @@ -3288,26 +5888,71 @@ memorystatus_control(struct proc *p __unused, struct memorystatus_control_args * case MEMORYSTATUS_CMD_SET_PRIORITY_PROPERTIES: error = memorystatus_cmd_set_priority_properties(args->pid, args->buffer, args->buffersize, ret); break; + case MEMORYSTATUS_CMD_SET_MEMLIMIT_PROPERTIES: + error = memorystatus_cmd_set_memlimit_properties(args->pid, args->buffer, args->buffersize, ret); + break; + case MEMORYSTATUS_CMD_GET_MEMLIMIT_PROPERTIES: + error = memorystatus_cmd_get_memlimit_properties(args->pid, args->buffer, args->buffersize, ret); + break; + case MEMORYSTATUS_CMD_GRP_SET_PROPERTIES: + error = memorystatus_cmd_grp_set_properties((int32_t)args->flags, args->buffer, args->buffersize, ret); + break; case MEMORYSTATUS_CMD_GET_JETSAM_SNAPSHOT: - error = memorystatus_cmd_get_jetsam_snapshot(args->buffer, args->buffersize, ret); + error = memorystatus_cmd_get_jetsam_snapshot((int32_t)args->flags, args->buffer, args->buffersize, ret); break; case MEMORYSTATUS_CMD_GET_PRESSURE_STATUS: error = memorystatus_cmd_get_pressure_status(ret); break; case MEMORYSTATUS_CMD_SET_JETSAM_HIGH_WATER_MARK: - /* TODO: deprecate. Keeping it in as there's no pid based way to set the ledger limit right now. */ - error = memorystatus_cmd_set_jetsam_high_water_mark(args->pid, (int32_t)args->flags, ret); + /* + * This call does not distinguish between active and inactive limits. + * Default behavior in 2-level HWM world is to set both. + * Non-fatal limit is also assumed for both. + */ + error = memorystatus_cmd_set_jetsam_memory_limit(args->pid, (int32_t)args->flags, ret, FALSE); + break; + case MEMORYSTATUS_CMD_SET_JETSAM_TASK_LIMIT: + /* + * This call does not distinguish between active and inactive limits. + * Default behavior in 2-level HWM world is to set both. + * Fatal limit is also assumed for both. + */ + error = memorystatus_cmd_set_jetsam_memory_limit(args->pid, (int32_t)args->flags, ret, TRUE); break; /* Test commands */ #if DEVELOPMENT || DEBUG case MEMORYSTATUS_CMD_TEST_JETSAM: error = memorystatus_kill_process_sync(args->pid, kMemorystatusKilled) ? 0 : EINVAL; break; + case MEMORYSTATUS_CMD_TEST_JETSAM_SORT: + error = memorystatus_cmd_test_jetsam_sort(args->pid, (int32_t)args->flags); + break; case MEMORYSTATUS_CMD_SET_JETSAM_PANIC_BITS: error = memorystatus_cmd_set_panic_bits(args->buffer, args->buffersize); break; #endif /* DEVELOPMENT || DEBUG */ + case MEMORYSTATUS_CMD_AGGRESSIVE_JETSAM_LENIENT_MODE_ENABLE: + if (memorystatus_aggressive_jetsam_lenient_allowed == FALSE) { +#if DEVELOPMENT || DEBUG + printf("Enabling Lenient Mode\n"); +#endif /* DEVELOPMENT || DEBUG */ + + memorystatus_aggressive_jetsam_lenient_allowed = TRUE; + memorystatus_aggressive_jetsam_lenient = TRUE; + } + break; + case MEMORYSTATUS_CMD_AGGRESSIVE_JETSAM_LENIENT_MODE_DISABLE: +#if DEVELOPMENT || DEBUG + printf("Disabling Lenient mode\n"); +#endif /* DEVELOPMENT || DEBUG */ + memorystatus_aggressive_jetsam_lenient_allowed = FALSE; + memorystatus_aggressive_jetsam_lenient = FALSE; + break; #endif /* CONFIG_JETSAM */ + case MEMORYSTATUS_CMD_PRIVILEGED_LISTENER_ENABLE: + case MEMORYSTATUS_CMD_PRIVILEGED_LISTENER_DISABLE: + error = memorystatus_low_mem_privileged_listener(args->command); + break; default: break; } @@ -3337,21 +5982,26 @@ filt_memorystatus(struct knote *kn __unused, long hint) switch (hint) { case kMemorystatusNoPressure: if (kn->kn_sfflags & NOTE_MEMORYSTATUS_PRESSURE_NORMAL) { - kn->kn_fflags |= NOTE_MEMORYSTATUS_PRESSURE_NORMAL; + kn->kn_fflags = NOTE_MEMORYSTATUS_PRESSURE_NORMAL; } break; case kMemorystatusPressure: if (memorystatus_vm_pressure_level == kVMPressureWarning || memorystatus_vm_pressure_level == kVMPressureUrgent) { if (kn->kn_sfflags & NOTE_MEMORYSTATUS_PRESSURE_WARN) { - kn->kn_fflags |= NOTE_MEMORYSTATUS_PRESSURE_WARN; + kn->kn_fflags = NOTE_MEMORYSTATUS_PRESSURE_WARN; } } else if (memorystatus_vm_pressure_level == kVMPressureCritical) { if (kn->kn_sfflags & NOTE_MEMORYSTATUS_PRESSURE_CRITICAL) { - kn->kn_fflags |= NOTE_MEMORYSTATUS_PRESSURE_CRITICAL; + kn->kn_fflags = NOTE_MEMORYSTATUS_PRESSURE_CRITICAL; } } break; + case kMemorystatusLowSwap: + if (kn->kn_sfflags & NOTE_MEMORYSTATUS_LOW_SWAP) { + kn->kn_fflags = NOTE_MEMORYSTATUS_LOW_SWAP; + } + break; default: break; } @@ -3382,16 +6032,10 @@ memorystatus_knote_register(struct knote *kn) { memorystatus_klist_lock(); - if (kn->kn_sfflags & (NOTE_MEMORYSTATUS_PRESSURE_NORMAL | NOTE_MEMORYSTATUS_PRESSURE_WARN | NOTE_MEMORYSTATUS_PRESSURE_CRITICAL)) { + if (kn->kn_sfflags & (NOTE_MEMORYSTATUS_PRESSURE_NORMAL | NOTE_MEMORYSTATUS_PRESSURE_WARN | NOTE_MEMORYSTATUS_PRESSURE_CRITICAL | NOTE_MEMORYSTATUS_LOW_SWAP)) { -#if CONFIG_JETSAM && VM_PRESSURE_EVENTS - /* Need a privilege to register */ - error = priv_check_cred(kauth_cred_get(), PRIV_VM_PRESSURE, 0); -#endif /* CONFIG_JETSAM && VM_PRESSURE_EVENTS */ + KNOTE_ATTACH(&memorystatus_klist, kn); - if (!error) { - KNOTE_ATTACH(&memorystatus_klist, kn); - } } else { error = ENOTSUP; } @@ -3408,6 +6052,8 @@ memorystatus_knote_unregister(struct knote *kn __unused) { memorystatus_klist_unlock(); } + +#if 0 #if CONFIG_JETSAM && VM_PRESSURE_EVENTS static boolean_t memorystatus_issue_pressure_kevent(boolean_t pressured) { @@ -3416,5 +6062,269 @@ memorystatus_issue_pressure_kevent(boolean_t pressured) { memorystatus_klist_unlock(); return TRUE; } - #endif /* CONFIG_JETSAM && VM_PRESSURE_EVENTS */ +#endif /* 0 */ + +#if CONFIG_JETSAM +/* Coalition support */ + +/* sorting info for a particular priority bucket */ +typedef struct memstat_sort_info { + coalition_t msi_coal; + uint64_t msi_page_count; + pid_t msi_pid; + int msi_ntasks; +} memstat_sort_info_t; + +/* + * qsort from smallest page count to largest page count + * + * return < 0 for a < b + * 0 for a == b + * > 0 for a > b + */ +static int memstat_asc_cmp(const void *a, const void *b) +{ + const memstat_sort_info_t *msA = (const memstat_sort_info_t *)a; + const memstat_sort_info_t *msB = (const memstat_sort_info_t *)b; + + return (int)((uint64_t)msA->msi_page_count - (uint64_t)msB->msi_page_count); +} + +/* + * Return the number of pids rearranged during this sort. + */ +static int +memorystatus_sort_by_largest_coalition_locked(unsigned int bucket_index, int coal_sort_order) +{ +#define MAX_SORT_PIDS 80 +#define MAX_COAL_LEADERS 10 + + unsigned int b = bucket_index; + int nleaders = 0; + int ntasks = 0; + proc_t p = NULL; + coalition_t coal = COALITION_NULL; + int pids_moved = 0; + int total_pids_moved = 0; + int i; + + /* + * The system is typically under memory pressure when in this + * path, hence, we want to avoid dynamic memory allocation. + */ + memstat_sort_info_t leaders[MAX_COAL_LEADERS]; + pid_t pid_list[MAX_SORT_PIDS]; + + if (bucket_index >= MEMSTAT_BUCKET_COUNT) { + return(0); + } + + /* + * Clear the array that holds coalition leader information + */ + for (i=0; i < MAX_COAL_LEADERS; i++) { + leaders[i].msi_coal = COALITION_NULL; + leaders[i].msi_page_count = 0; /* will hold total coalition page count */ + leaders[i].msi_pid = 0; /* will hold coalition leader pid */ + leaders[i].msi_ntasks = 0; /* will hold the number of tasks in a coalition */ + } + + p = memorystatus_get_first_proc_locked(&b, FALSE); + while (p) { + if (coalition_is_leader(p->task, COALITION_TYPE_JETSAM, &coal)) { + if (nleaders < MAX_COAL_LEADERS) { + int coal_ntasks = 0; + uint64_t coal_page_count = coalition_get_page_count(coal, &coal_ntasks); + leaders[nleaders].msi_coal = coal; + leaders[nleaders].msi_page_count = coal_page_count; + leaders[nleaders].msi_pid = p->p_pid; /* the coalition leader */ + leaders[nleaders].msi_ntasks = coal_ntasks; + nleaders++; + } else { + /* + * We've hit MAX_COAL_LEADERS meaning we can handle no more coalitions. + * Abandoned coalitions will linger at the tail of the priority band + * when this sort session ends. + * TODO: should this be an assert? + */ + printf("%s: WARNING: more than %d leaders in priority band [%d]\n", + __FUNCTION__, MAX_COAL_LEADERS, bucket_index); + break; + } + } + p=memorystatus_get_next_proc_locked(&b, p, FALSE); + } + + if (nleaders == 0) { + /* Nothing to sort */ + return(0); + } + + /* + * Sort the coalition leader array, from smallest coalition page count + * to largest coalition page count. When inserted in the priority bucket, + * smallest coalition is handled first, resulting in the last to be jetsammed. + */ + if (nleaders > 1) { + qsort(leaders, nleaders, sizeof(memstat_sort_info_t), memstat_asc_cmp); + } + +#if 0 + for (i = 0; i < nleaders; i++) { + printf("%s: coal_leader[%d of %d] pid[%d] pages[%llu] ntasks[%d]\n", + __FUNCTION__, i, nleaders, leaders[i].msi_pid, leaders[i].msi_page_count, + leaders[i].msi_ntasks); + } +#endif + + /* + * During coalition sorting, processes in a priority band are rearranged + * by being re-inserted at the head of the queue. So, when handling a + * list, the first process that gets moved to the head of the queue, + * ultimately gets pushed toward the queue tail, and hence, jetsams last. + * + * So, for example, the coalition leader is expected to jetsam last, + * after its coalition members. Therefore, the coalition leader is + * inserted at the head of the queue first. + * + * After processing a coalition, the jetsam order is as follows: + * undefs(jetsam first), extensions, xpc services, leader(jetsam last) + */ + + /* + * Coalition members are rearranged in the priority bucket here, + * based on their coalition role. + */ + total_pids_moved = 0; + for (i=0; i < nleaders; i++) { + + /* a bit of bookkeeping */ + pids_moved = 0; + + /* Coalition leaders are jetsammed last, so move into place first */ + pid_list[0] = leaders[i].msi_pid; + pids_moved += memorystatus_move_list_locked(bucket_index, pid_list, 1); + + /* xpc services should jetsam after extensions */ + ntasks = coalition_get_pid_list (leaders[i].msi_coal, COALITION_ROLEMASK_XPC, + coal_sort_order, pid_list, MAX_SORT_PIDS); + + if (ntasks > 0) { + pids_moved += memorystatus_move_list_locked(bucket_index, pid_list, + (ntasks <= MAX_SORT_PIDS ? ntasks : MAX_SORT_PIDS)); + } + + /* extensions should jetsam after unmarked processes */ + ntasks = coalition_get_pid_list (leaders[i].msi_coal, COALITION_ROLEMASK_EXT, + coal_sort_order, pid_list, MAX_SORT_PIDS); + + if (ntasks > 0) { + pids_moved += memorystatus_move_list_locked(bucket_index, pid_list, + (ntasks <= MAX_SORT_PIDS ? ntasks : MAX_SORT_PIDS)); + } + + /* undefined coalition members should be the first to jetsam */ + ntasks = coalition_get_pid_list (leaders[i].msi_coal, COALITION_ROLEMASK_UNDEF, + coal_sort_order, pid_list, MAX_SORT_PIDS); + + if (ntasks > 0) { + pids_moved += memorystatus_move_list_locked(bucket_index, pid_list, + (ntasks <= MAX_SORT_PIDS ? ntasks : MAX_SORT_PIDS)); + } + +#if 0 + if (pids_moved == leaders[i].msi_ntasks) { + /* + * All the pids in the coalition were found in this band. + */ + printf("%s: pids_moved[%d] equal total coalition ntasks[%d] \n", __FUNCTION__, + pids_moved, leaders[i].msi_ntasks); + } else if (pids_moved > leaders[i].msi_ntasks) { + /* + * Apparently new coalition members showed up during the sort? + */ + printf("%s: pids_moved[%d] were greater than expected coalition ntasks[%d] \n", __FUNCTION__, + pids_moved, leaders[i].msi_ntasks); + } else { + /* + * Apparently not all the pids in the coalition were found in this band? + */ + printf("%s: pids_moved[%d] were less than expected coalition ntasks[%d] \n", __FUNCTION__, + pids_moved, leaders[i].msi_ntasks); + } +#endif + + total_pids_moved += pids_moved; + + } /* end for */ + + return(total_pids_moved); +} + + +/* + * Traverse a list of pids, searching for each within the priority band provided. + * If pid is found, move it to the front of the priority band. + * Never searches outside the priority band provided. + * + * Input: + * bucket_index - jetsam priority band. + * pid_list - pointer to a list of pids. + * list_sz - number of pids in the list. + * + * Pid list ordering is important in that, + * pid_list[n] is expected to jetsam ahead of pid_list[n+1]. + * The sort_order is set by the coalition default. + * + * Return: + * the number of pids found and hence moved within the priority band. + */ +static int +memorystatus_move_list_locked(unsigned int bucket_index, pid_t *pid_list, int list_sz) +{ + memstat_bucket_t *current_bucket; + int i; + int found_pids = 0; + + if ((pid_list == NULL) || (list_sz <= 0)) { + return(0); + } + + if (bucket_index >= MEMSTAT_BUCKET_COUNT) { + return(0); + } + + current_bucket = &memstat_bucket[bucket_index]; + for (i=0; i < list_sz; i++) { + unsigned int b = bucket_index; + proc_t p = NULL; + proc_t aProc = NULL; + pid_t aPid; + int list_index; + + list_index = ((list_sz - 1) - i); + aPid = pid_list[list_index]; + + /* never search beyond bucket_index provided */ + p = memorystatus_get_first_proc_locked(&b, FALSE); + while (p) { + if (p->p_pid == aPid) { + aProc = p; + break; + } + p = memorystatus_get_next_proc_locked(&b, p, FALSE); + } + + if (aProc == NULL) { + /* pid not found in this band, just skip it */ + continue; + } else { + TAILQ_REMOVE(¤t_bucket->list, aProc, p_memstat_list); + TAILQ_INSERT_HEAD(¤t_bucket->list, aProc, p_memstat_list); + found_pids++; + } + } + return(found_pids); +} +#endif /* CONFIG_JETSAM */