Lines Matching +full:saw +full:- +full:leader
1 // SPDX-License-Identifier: GPL-2.0
6 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
7 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra
58 #include <linux/percpu-rwsem.h>
76 struct task_struct *p = tfc->p; in remote_function()
79 /* -EAGAIN */ in remote_function()
88 tfc->ret = -ESRCH; /* No such (running) process */ in remote_function()
93 tfc->ret = tfc->func(tfc->info); in remote_function()
97 * task_function_call - call a function on the cpu on which a task runs
107 * returns @func return value or -ESRCH or -ENXIO when the process isn't running
116 .ret = -EAGAIN, in task_function_call()
126 if (ret != -EAGAIN) in task_function_call()
136 * cpu_function_call - call a function on the cpu
143 * returns: @func return value or -ENXIO when the cpu is offline
151 .ret = -ENXIO, /* No such CPU */ in cpu_function_call()
175 raw_spin_lock(&ctx->lock); in __perf_ctx_lock()
176 WARN_ON_ONCE(ctx->is_active & EVENT_FROZEN); in __perf_ctx_lock()
182 __perf_ctx_lock(&cpuctx->ctx); in perf_ctx_lock()
193 if (ctx->is_active & EVENT_FROZEN) { in __perf_ctx_unlock()
194 if (!(ctx->is_active & EVENT_ALL)) in __perf_ctx_unlock()
195 ctx->is_active = 0; in __perf_ctx_unlock()
197 ctx->is_active &= ~EVENT_FROZEN; in __perf_ctx_unlock()
199 raw_spin_unlock(&ctx->lock); in __perf_ctx_unlock()
207 __perf_ctx_unlock(&cpuctx->ctx); in perf_ctx_unlock()
210 #define TASK_TOMBSTONE ((void *)-1L)
214 return READ_ONCE(event->owner) == TASK_TOMBSTONE; in is_kernel_event()
222 return this_cpu_ptr(&perf_cpu_context)->task_ctx; in perf_cpu_task_ctx()
228 * When !ctx->nr_events a task context will not be scheduled. This means
234 * - removing the last event from a task ctx; this is relatively straight
237 * - adding the first event to a task ctx; this is tricky because we cannot
238 * rely on ctx->is_active and therefore cannot use event_function_call().
241 * If ctx->nr_events, then ctx->is_active and cpuctx->task_ctx are set.
256 struct perf_event *event = efs->event; in event_function()
257 struct perf_event_context *ctx = event->ctx; in event_function()
259 struct perf_event_context *task_ctx = cpuctx->task_ctx; in event_function()
266 * Since we do the IPI call without holding ctx->lock things can have in event_function()
269 if (ctx->task) { in event_function()
270 if (ctx->task != current) { in event_function()
271 ret = -ESRCH; in event_function()
279 * above ctx->task != current test), therefore we must have in event_function()
280 * ctx->is_active here. in event_function()
282 WARN_ON_ONCE(!ctx->is_active); in event_function()
284 * And since we have ctx->is_active, cpuctx->task_ctx must in event_function()
289 WARN_ON_ONCE(&cpuctx->ctx != ctx); in event_function()
292 efs->func(event, cpuctx, ctx, efs->data); in event_function()
301 struct perf_event_context *ctx = event->ctx; in event_function_call()
302 struct task_struct *task = READ_ONCE(ctx->task); /* verified in event_function */ in event_function_call()
310 if (!event->parent) { in event_function_call()
313 * stabilize the event->ctx relation. See in event_function_call()
316 lockdep_assert_held(&ctx->mutex); in event_function_call()
320 cpu_function_call(event->cpu, event_function, &efs); in event_function_call()
338 task = ctx->task; in event_function_call()
341 if (ctx->is_active) { in event_function_call()
358 struct perf_event_context *ctx = event->ctx; in event_function_local()
360 struct task_struct *task = READ_ONCE(ctx->task); in event_function_local()
374 task = ctx->task; in event_function_local()
384 if (ctx->is_active) { in event_function_local()
388 if (WARN_ON_ONCE(cpuctx->task_ctx != ctx)) in event_function_local()
392 WARN_ON_ONCE(&cpuctx->ctx != ctx); in event_function_local()
449 * -1 - not paranoid at all
450 * 0 - disallow raw tracepoint access for unpriv
451 * 1 - disallow cpu events for unpriv
452 * 2 - disallow kernel profiling for unpriv
498 return -EINVAL; in perf_event_max_sample_rate_handler()
533 * User-space relies on this file as a feature check for
612 running_len -= running_len/NR_ACCUMULATED_SAMPLES; in perf_sample_event_took()
666 return event->clock(); in perf_event_clock()
672 * The basic idea is to use event->state to determine which (if any) time
678 * rules for a group are that if the group leader is OFF the entire group is
682 * A further ramification is that when a group leader flips between OFF and
694 struct perf_event *leader = event->group_leader; in __perf_effective_state() local
696 if (leader->state <= PERF_EVENT_STATE_OFF) in __perf_effective_state()
697 return leader->state; in __perf_effective_state()
699 return event->state; in __perf_effective_state()
706 u64 delta = now - event->tstamp; in __perf_update_times()
708 *enabled = event->total_time_enabled; in __perf_update_times()
712 *running = event->total_time_running; in __perf_update_times()
721 __perf_update_times(event, now, &event->total_time_enabled, in perf_event_update_time()
722 &event->total_time_running); in perf_event_update_time()
723 event->tstamp = now; in perf_event_update_time()
726 static void perf_event_update_sibling_time(struct perf_event *leader) in perf_event_update_sibling_time() argument
730 for_each_sibling_event(sibling, leader) in perf_event_update_sibling_time()
737 if (event->state == state) in perf_event_set_state()
742 * If a group leader gets enabled/disabled all its siblings in perf_event_set_state()
745 if ((event->state < 0) ^ (state < 0)) in perf_event_set_state()
748 WRITE_ONCE(event->state, state); in perf_event_set_state()
752 * UP store-release, load-acquire
769 list_for_each_entry(_epc, &((_ctx)->pmu_ctx_list), pmu_ctx_entry) \
770 if (_cgroup && !_epc->nr_cgroups) \
772 else if (_pmu && _epc->pmu != _pmu) \
781 perf_pmu_disable(pmu_ctx->pmu); in perf_ctx_disable()
789 perf_pmu_enable(pmu_ctx->pmu); in perf_ctx_enable()
803 if (!event->cgrp) in perf_cgroup_match()
807 if (!cpuctx->cgrp) in perf_cgroup_match()
816 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, in perf_cgroup_match()
817 event->cgrp->css.cgroup); in perf_cgroup_match()
822 css_put(&event->cgrp->css); in perf_detach_cgroup()
823 event->cgrp = NULL; in perf_detach_cgroup()
828 return event->cgrp != NULL; in is_cgroup_event()
835 t = per_cpu_ptr(event->cgrp->info, event->cpu); in perf_cgroup_event_time()
836 return t->time; in perf_cgroup_event_time()
843 t = per_cpu_ptr(event->cgrp->info, event->cpu); in perf_cgroup_event_time_now()
844 if (!__load_acquire(&t->active)) in perf_cgroup_event_time_now()
845 return t->time; in perf_cgroup_event_time_now()
846 now += READ_ONCE(t->timeoffset); in perf_cgroup_event_time_now()
853 info->time += now - info->timestamp; in __update_cgrp_time()
854 info->timestamp = now; in __update_cgrp_time()
858 WRITE_ONCE(info->timeoffset, info->time - info->timestamp); in __update_cgrp_time()
863 struct perf_cgroup *cgrp = cpuctx->cgrp; in update_cgrp_time_from_cpuctx()
870 for (css = &cgrp->css; css; css = css->parent) { in update_cgrp_time_from_cpuctx()
872 info = this_cpu_ptr(cgrp->info); in update_cgrp_time_from_cpuctx()
876 __store_release(&info->active, 0); in update_cgrp_time_from_cpuctx()
892 info = this_cpu_ptr(event->cgrp->info); in update_cgrp_time_from_event()
896 if (info->active) in update_cgrp_time_from_event()
903 struct perf_event_context *ctx = &cpuctx->ctx; in perf_cgroup_set_timestamp()
904 struct perf_cgroup *cgrp = cpuctx->cgrp; in perf_cgroup_set_timestamp()
909 * ctx->lock held by caller in perf_cgroup_set_timestamp()
916 WARN_ON_ONCE(!ctx->nr_cgroups); in perf_cgroup_set_timestamp()
918 for (css = &cgrp->css; css; css = css->parent) { in perf_cgroup_set_timestamp()
920 info = this_cpu_ptr(cgrp->info); in perf_cgroup_set_timestamp()
921 __update_cgrp_time(info, ctx->timestamp, false); in perf_cgroup_set_timestamp()
922 __store_release(&info->active, 1); in perf_cgroup_set_timestamp()
935 * cpuctx->cgrp is set when the first cgroup event enabled, in perf_cgroup_switch()
938 if (READ_ONCE(cpuctx->cgrp) == NULL) in perf_cgroup_switch()
941 WARN_ON_ONCE(cpuctx->ctx.nr_cgroups == 0); in perf_cgroup_switch()
944 if (READ_ONCE(cpuctx->cgrp) == cgrp) in perf_cgroup_switch()
947 perf_ctx_lock(cpuctx, cpuctx->task_ctx); in perf_cgroup_switch()
948 perf_ctx_disable(&cpuctx->ctx, true); in perf_cgroup_switch()
950 ctx_sched_out(&cpuctx->ctx, NULL, EVENT_ALL|EVENT_CGROUP); in perf_cgroup_switch()
956 cpuctx->cgrp = cgrp; in perf_cgroup_switch()
962 ctx_sched_in(&cpuctx->ctx, NULL, EVENT_ALL|EVENT_CGROUP); in perf_cgroup_switch()
964 perf_ctx_enable(&cpuctx->ctx, true); in perf_cgroup_switch()
965 perf_ctx_unlock(cpuctx, cpuctx->task_ctx); in perf_cgroup_switch()
979 for (heap_size = 1; css; css = css->parent) in perf_cgroup_ensure_storage()
984 if (heap_size <= cpuctx->heap_size) in perf_cgroup_ensure_storage()
990 ret = -ENOMEM; in perf_cgroup_ensure_storage()
994 raw_spin_lock_irq(&cpuctx->ctx.lock); in perf_cgroup_ensure_storage()
995 if (cpuctx->heap_size < heap_size) { in perf_cgroup_ensure_storage()
996 swap(cpuctx->heap, storage); in perf_cgroup_ensure_storage()
997 if (storage == cpuctx->heap_default) in perf_cgroup_ensure_storage()
999 cpuctx->heap_size = heap_size; in perf_cgroup_ensure_storage()
1001 raw_spin_unlock_irq(&cpuctx->ctx.lock); in perf_cgroup_ensure_storage()
1019 return -EBADF; in perf_cgroup_connect()
1021 css = css_tryget_online_from_dir(fd_file(f)->f_path.dentry, in perf_cgroup_connect()
1031 event->cgrp = cgrp; in perf_cgroup_connect()
1038 if (group_leader && group_leader->cgrp != cgrp) { in perf_cgroup_connect()
1040 ret = -EINVAL; in perf_cgroup_connect()
1053 event->pmu_ctx->nr_cgroups++; in perf_cgroup_event_enable()
1056 * Because cgroup events are always per-cpu events, in perf_cgroup_event_enable()
1057 * @ctx == &cpuctx->ctx. in perf_cgroup_event_enable()
1061 if (ctx->nr_cgroups++) in perf_cgroup_event_enable()
1064 cpuctx->cgrp = perf_cgroup_from_task(current, ctx); in perf_cgroup_event_enable()
1075 event->pmu_ctx->nr_cgroups--; in perf_cgroup_event_disable()
1078 * Because cgroup events are always per-cpu events, in perf_cgroup_event_disable()
1079 * @ctx == &cpuctx->ctx. in perf_cgroup_event_disable()
1083 if (--ctx->nr_cgroups) in perf_cgroup_event_disable()
1086 cpuctx->cgrp = NULL; in perf_cgroup_event_disable()
1118 return -EINVAL; in perf_cgroup_connect()
1169 raw_spin_lock(&cpc->hrtimer_lock); in perf_mux_hrtimer_handler()
1171 hrtimer_forward_now(hr, cpc->hrtimer_interval); in perf_mux_hrtimer_handler()
1173 cpc->hrtimer_active = 0; in perf_mux_hrtimer_handler()
1174 raw_spin_unlock(&cpc->hrtimer_lock); in perf_mux_hrtimer_handler()
1181 struct hrtimer *timer = &cpc->hrtimer; in __perf_mux_hrtimer_init()
1182 struct pmu *pmu = cpc->epc.pmu; in __perf_mux_hrtimer_init()
1189 interval = pmu->hrtimer_interval_ms; in __perf_mux_hrtimer_init()
1191 interval = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; in __perf_mux_hrtimer_init()
1193 cpc->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * interval); in __perf_mux_hrtimer_init()
1195 raw_spin_lock_init(&cpc->hrtimer_lock); in __perf_mux_hrtimer_init()
1202 struct hrtimer *timer = &cpc->hrtimer; in perf_mux_hrtimer_restart()
1205 raw_spin_lock_irqsave(&cpc->hrtimer_lock, flags); in perf_mux_hrtimer_restart()
1206 if (!cpc->hrtimer_active) { in perf_mux_hrtimer_restart()
1207 cpc->hrtimer_active = 1; in perf_mux_hrtimer_restart()
1208 hrtimer_forward_now(timer, cpc->hrtimer_interval); in perf_mux_hrtimer_restart()
1211 raw_spin_unlock_irqrestore(&cpc->hrtimer_lock, flags); in perf_mux_hrtimer_restart()
1223 return *this_cpu_ptr(pmu->cpu_pmu_context); in this_cpc()
1228 int *count = &this_cpc(pmu)->pmu_disable_count; in perf_pmu_disable()
1230 pmu->pmu_disable(pmu); in perf_pmu_disable()
1235 int *count = &this_cpc(pmu)->pmu_disable_count; in perf_pmu_enable()
1236 if (!--(*count)) in perf_pmu_enable()
1237 pmu->pmu_enable(pmu); in perf_pmu_enable()
1242 int *count = &this_cpc(pmu)->pmu_disable_count; in perf_assert_pmu_disabled()
1248 if (event->state == PERF_EVENT_STATE_ACTIVE) in perf_pmu_read()
1249 event->pmu->read(event); in perf_pmu_read()
1254 refcount_inc(&ctx->refcount); in get_ctx()
1267 if (refcount_dec_and_test(&ctx->refcount)) { in put_ctx()
1268 if (ctx->parent_ctx) in put_ctx()
1269 put_ctx(ctx->parent_ctx); in put_ctx()
1270 if (ctx->task && ctx->task != TASK_TOMBSTONE) in put_ctx()
1271 put_task_struct(ctx->task); in put_ctx()
1272 call_rcu(&ctx->rcu_head, free_ctx); in put_ctx()
1286 * - perf_event_exit_task_context() [ child , 0 ]
1290 * - perf_event_init_context() [ parent, 0 ]
1298 * While it appears there is an obvious deadlock here -- the parent and child
1300 * life-time rules separate them. That is an exiting task cannot fork, and a
1303 * But remember that these are parent<->child context relations, and
1309 * the ctx parent<->child relation, and perf_pmu_migrate_context() is only
1324 * However; because event->ctx can change while we're waiting to acquire
1325 * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
1341 * cpuctx->mutex / perf_event_context::mutex
1350 ctx = READ_ONCE(event->ctx); in perf_event_ctx_lock_nested()
1351 if (!refcount_inc_not_zero(&ctx->refcount)) { in perf_event_ctx_lock_nested()
1357 mutex_lock_nested(&ctx->mutex, nesting); in perf_event_ctx_lock_nested()
1358 if (event->ctx != ctx) { in perf_event_ctx_lock_nested()
1359 mutex_unlock(&ctx->mutex); in perf_event_ctx_lock_nested()
1376 mutex_unlock(&ctx->mutex); in perf_event_ctx_unlock()
1381 * This must be done under the ctx->lock, such as to serialize against
1383 * calling scheduler related locks and ctx->lock nests inside those.
1388 struct perf_event_context *parent_ctx = ctx->parent_ctx; in unclone_ctx()
1390 lockdep_assert_held(&ctx->lock); in unclone_ctx()
1393 ctx->parent_ctx = NULL; in unclone_ctx()
1394 ctx->generation++; in unclone_ctx()
1406 if (event->parent) in perf_event_pid_type()
1407 event = event->parent; in perf_event_pid_type()
1409 nr = __task_pid_nr_ns(p, type, event->ns); in perf_event_pid_type()
1410 /* avoid -1 if it is idle thread or runs in another ns */ in perf_event_pid_type()
1412 nr = -1; in perf_event_pid_type()
1432 u64 id = event->id; in primary_event_id()
1434 if (event->parent) in primary_event_id()
1435 id = event->parent->id; in primary_event_id()
1455 * part of the read side critical section was irqs-enabled -- see in perf_lock_task_context()
1458 * Since ctx->lock nests under rq->lock we must ensure the entire read in perf_lock_task_context()
1463 ctx = rcu_dereference(task->perf_event_ctxp); in perf_lock_task_context()
1475 raw_spin_lock(&ctx->lock); in perf_lock_task_context()
1476 if (ctx != rcu_dereference(task->perf_event_ctxp)) { in perf_lock_task_context()
1477 raw_spin_unlock(&ctx->lock); in perf_lock_task_context()
1483 if (ctx->task == TASK_TOMBSTONE || in perf_lock_task_context()
1484 !refcount_inc_not_zero(&ctx->refcount)) { in perf_lock_task_context()
1485 raw_spin_unlock(&ctx->lock); in perf_lock_task_context()
1488 WARN_ON_ONCE(ctx->task != task); in perf_lock_task_context()
1510 ++ctx->pin_count; in perf_pin_task_context()
1511 raw_spin_unlock_irqrestore(&ctx->lock, flags); in perf_pin_task_context()
1520 raw_spin_lock_irqsave(&ctx->lock, flags); in perf_unpin_context()
1521 --ctx->pin_count; in perf_unpin_context()
1522 raw_spin_unlock_irqrestore(&ctx->lock, flags); in perf_unpin_context()
1532 lockdep_assert_held(&ctx->lock); in __update_context_time()
1535 ctx->time += now - ctx->timestamp; in __update_context_time()
1536 ctx->timestamp = now; in __update_context_time()
1539 * The above: time' = time + (now - timestamp), can be re-arranged in __update_context_time()
1540 * into: time` = now + (time - timestamp), which gives a single value in __update_context_time()
1544 * it's (obviously) not possible to acquire ctx->lock in order to read in __update_context_time()
1547 WRITE_ONCE(ctx->timeoffset, ctx->time - ctx->timestamp); in __update_context_time()
1557 struct perf_event_context *ctx = event->ctx; in perf_event_time()
1565 return ctx->time; in perf_event_time()
1570 struct perf_event_context *ctx = event->ctx; in perf_event_time_now()
1578 if (!(__load_acquire(&ctx->is_active) & EVENT_TIME)) in perf_event_time_now()
1579 return ctx->time; in perf_event_time_now()
1581 now += READ_ONCE(ctx->timeoffset); in perf_event_time_now()
1587 struct perf_event_context *ctx = event->ctx; in get_event_type()
1590 lockdep_assert_held(&ctx->lock); in get_event_type()
1593 * It's 'group type', really, because if our group leader is in get_event_type()
1596 if (event->group_leader != event) in get_event_type()
1597 event = event->group_leader; in get_event_type()
1599 event_type = event->attr.pinned ? EVENT_PINNED : EVENT_FLEXIBLE; in get_event_type()
1600 if (!ctx->task) in get_event_type()
1611 RB_CLEAR_NODE(&event->group_node); in init_event_group()
1612 event->group_index = 0; in init_event_group()
1622 if (event->attr.pinned) in get_event_groups()
1623 return &ctx->pinned_groups; in get_event_groups()
1625 return &ctx->flexible_groups; in get_event_groups()
1633 groups->tree = RB_ROOT; in perf_event_groups_init()
1634 groups->index = 0; in perf_event_groups_init()
1642 if (event->cgrp) in event_cgroup()
1643 cgroup = event->cgrp->css.cgroup; in event_cgroup()
1660 if (left_cpu < right->cpu) in perf_event_groups_cmp()
1661 return -1; in perf_event_groups_cmp()
1662 if (left_cpu > right->cpu) in perf_event_groups_cmp()
1666 if (left_pmu < right->pmu_ctx->pmu) in perf_event_groups_cmp()
1667 return -1; in perf_event_groups_cmp()
1668 if (left_pmu > right->pmu_ctx->pmu) in perf_event_groups_cmp()
1682 return -1; in perf_event_groups_cmp()
1693 return -1; in perf_event_groups_cmp()
1700 if (left_group_index < right->group_index) in perf_event_groups_cmp()
1701 return -1; in perf_event_groups_cmp()
1702 if (left_group_index > right->group_index) in perf_event_groups_cmp()
1714 return perf_event_groups_cmp(e->cpu, e->pmu_ctx->pmu, event_cgroup(e), in __group_less()
1715 e->group_index, __node_2_pe(b)) < 0; in __group_less()
1730 return perf_event_groups_cmp(a->cpu, a->pmu, a->cgroup, b->group_index, b); in __group_cmp()
1740 return perf_event_groups_cmp(a->cpu, a->pmu, event_cgroup(b), in __group_cmp_ignore_cgroup()
1741 b->group_index, b); in __group_cmp_ignore_cgroup()
1746 * {@event->cpu, @event->pmu_ctx->pmu, event_cgroup(@event), ++@groups->index}
1753 event->group_index = ++groups->index; in perf_event_groups_insert()
1755 rb_add(&event->group_node, &groups->tree, __group_less); in perf_event_groups_insert()
1777 WARN_ON_ONCE(RB_EMPTY_NODE(&event->group_node) || in perf_event_groups_delete()
1778 RB_EMPTY_ROOT(&groups->tree)); in perf_event_groups_delete()
1780 rb_erase(&event->group_node, &groups->tree); in perf_event_groups_delete()
1810 node = rb_find_first(&key, &groups->tree, __group_cmp); in perf_event_groups_first()
1821 .cpu = event->cpu, in perf_event_groups_next()
1827 next = rb_next_match(&key, &event->group_node, __group_cmp); in perf_event_groups_next()
1842 for (event = rb_entry_safe(rb_first(&((groups)->tree)), \
1844 event = rb_entry_safe(rb_next(&event->group_node), \
1852 return attr->inherit && (attr->sample_type & PERF_SAMPLE_READ); in has_inherit_and_sample_read()
1857 * Must be called with ctx->mutex and ctx->lock held.
1862 lockdep_assert_held(&ctx->lock); in list_add_event()
1864 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); in list_add_event()
1865 event->attach_state |= PERF_ATTACH_CONTEXT; in list_add_event()
1867 event->tstamp = perf_event_time(event); in list_add_event()
1870 * If we're a stand alone event or group leader, we go to the context in list_add_event()
1874 if (event->group_leader == event) { in list_add_event()
1875 event->group_caps = event->event_caps; in list_add_event()
1879 list_add_rcu(&event->event_entry, &ctx->event_list); in list_add_event()
1880 ctx->nr_events++; in list_add_event()
1881 if (event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT) in list_add_event()
1882 ctx->nr_user++; in list_add_event()
1883 if (event->attr.inherit_stat) in list_add_event()
1884 ctx->nr_stat++; in list_add_event()
1885 if (has_inherit_and_sample_read(&event->attr)) in list_add_event()
1886 local_inc(&ctx->nr_no_switch_fast); in list_add_event()
1888 if (event->state > PERF_EVENT_STATE_OFF) in list_add_event()
1891 ctx->generation++; in list_add_event()
1892 event->pmu_ctx->nr_events++; in list_add_event()
1900 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : in perf_event__state_init()
1940 size += sizeof(data->ip); in __perf_event_header_size()
1943 size += sizeof(data->addr); in __perf_event_header_size()
1946 size += sizeof(data->period); in __perf_event_header_size()
1949 size += sizeof(data->weight.full); in __perf_event_header_size()
1952 size += event->read_size; in __perf_event_header_size()
1955 size += sizeof(data->data_src.val); in __perf_event_header_size()
1958 size += sizeof(data->txn); in __perf_event_header_size()
1961 size += sizeof(data->phys_addr); in __perf_event_header_size()
1964 size += sizeof(data->cgroup); in __perf_event_header_size()
1967 size += sizeof(data->data_page_size); in __perf_event_header_size()
1970 size += sizeof(data->code_page_size); in __perf_event_header_size()
1972 event->header_size = size; in __perf_event_header_size()
1981 event->read_size = in perf_event__header_size()
1982 __perf_event_read_size(event->attr.read_format, in perf_event__header_size()
1983 event->group_leader->nr_siblings); in perf_event__header_size()
1984 __perf_event_header_size(event, event->attr.sample_type); in perf_event__header_size()
1990 u64 sample_type = event->attr.sample_type; in perf_event__id_header_size()
1994 size += sizeof(data->tid_entry); in perf_event__id_header_size()
1997 size += sizeof(data->time); in perf_event__id_header_size()
2000 size += sizeof(data->id); in perf_event__id_header_size()
2003 size += sizeof(data->id); in perf_event__id_header_size()
2006 size += sizeof(data->stream_id); in perf_event__id_header_size()
2009 size += sizeof(data->cpu_entry); in perf_event__id_header_size()
2011 event->id_header_size = size; in perf_event__id_header_size()
2020 * depends on per-event read_format, also (re)check the existing events.
2027 struct perf_event *sibling, *group_leader = event->group_leader; in perf_event_validate_size()
2029 if (__perf_event_read_size(event->attr.read_format, in perf_event_validate_size()
2030 group_leader->nr_siblings + 1) > 16*1024) in perf_event_validate_size()
2033 if (__perf_event_read_size(group_leader->attr.read_format, in perf_event_validate_size()
2034 group_leader->nr_siblings + 1) > 16*1024) in perf_event_validate_size()
2038 * When creating a new group leader, group_leader->ctx is initialized in perf_event_validate_size()
2040 * for_each_sibling_event() until group_leader->ctx is set. A new group in perf_event_validate_size()
2041 * leader cannot have any siblings yet, so we can safely skip checking in perf_event_validate_size()
2042 * the non-existent siblings. in perf_event_validate_size()
2048 if (__perf_event_read_size(sibling->attr.read_format, in perf_event_validate_size()
2049 group_leader->nr_siblings + 1) > 16*1024) in perf_event_validate_size()
2058 struct perf_event *group_leader = event->group_leader, *pos; in perf_group_attach()
2060 lockdep_assert_held(&event->ctx->lock); in perf_group_attach()
2066 if (event->attach_state & PERF_ATTACH_GROUP) in perf_group_attach()
2069 event->attach_state |= PERF_ATTACH_GROUP; in perf_group_attach()
2074 WARN_ON_ONCE(group_leader->ctx != event->ctx); in perf_group_attach()
2076 group_leader->group_caps &= event->event_caps; in perf_group_attach()
2078 list_add_tail(&event->sibling_list, &group_leader->sibling_list); in perf_group_attach()
2079 group_leader->nr_siblings++; in perf_group_attach()
2080 group_leader->group_generation++; in perf_group_attach()
2090 * Must be called with ctx->mutex and ctx->lock held.
2095 WARN_ON_ONCE(event->ctx != ctx); in list_del_event()
2096 lockdep_assert_held(&ctx->lock); in list_del_event()
2099 * We can have double detach due to exit/hot-unplug + close. in list_del_event()
2101 if (!(event->attach_state & PERF_ATTACH_CONTEXT)) in list_del_event()
2104 event->attach_state &= ~PERF_ATTACH_CONTEXT; in list_del_event()
2106 ctx->nr_events--; in list_del_event()
2107 if (event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT) in list_del_event()
2108 ctx->nr_user--; in list_del_event()
2109 if (event->attr.inherit_stat) in list_del_event()
2110 ctx->nr_stat--; in list_del_event()
2111 if (has_inherit_and_sample_read(&event->attr)) in list_del_event()
2112 local_dec(&ctx->nr_no_switch_fast); in list_del_event()
2114 list_del_rcu(&event->event_entry); in list_del_event()
2116 if (event->group_leader == event) in list_del_event()
2123 * of error state is by explicit re-enabling in list_del_event()
2126 if (event->state > PERF_EVENT_STATE_OFF) { in list_del_event()
2131 ctx->generation++; in list_del_event()
2132 event->pmu_ctx->nr_events--; in list_del_event()
2141 if (!event->pmu->aux_output_match) in perf_aux_output_match()
2144 return event->pmu->aux_output_match(aux_event); in perf_aux_output_match()
2153 struct perf_event_context *ctx = event->ctx; in perf_put_aux_event()
2159 if (event->aux_event) { in perf_put_aux_event()
2160 iter = event->aux_event; in perf_put_aux_event()
2161 event->aux_event = NULL; in perf_put_aux_event()
2170 for_each_sibling_event(iter, event->group_leader) { in perf_put_aux_event()
2171 if (iter->aux_event != event) in perf_put_aux_event()
2174 iter->aux_event = NULL; in perf_put_aux_event()
2189 return event->attr.aux_output || has_aux_action(event); in perf_need_aux_event()
2196 * Our group leader must be an aux event if we want to be in perf_get_aux_event()
2207 if (event->attr.aux_output && event->attr.aux_sample_size) in perf_get_aux_event()
2210 if (event->attr.aux_output && in perf_get_aux_event()
2214 if ((event->attr.aux_pause || event->attr.aux_resume) && in perf_get_aux_event()
2215 !(group_leader->pmu->capabilities & PERF_PMU_CAP_AUX_PAUSE)) in perf_get_aux_event()
2218 if (event->attr.aux_sample_size && !group_leader->pmu->snapshot_aux) in perf_get_aux_event()
2221 if (!atomic_long_inc_not_zero(&group_leader->refcount)) in perf_get_aux_event()
2230 event->aux_event = group_leader; in perf_get_aux_event()
2237 return event->attr.pinned ? &event->pmu_ctx->pinned_active : in get_event_list()
2238 &event->pmu_ctx->flexible_active; in get_event_list()
2249 event_sched_out(event, event->ctx); in perf_remove_sibling_event()
2255 struct perf_event *leader = event->group_leader; in perf_group_detach() local
2257 struct perf_event_context *ctx = event->ctx; in perf_group_detach()
2259 lockdep_assert_held(&ctx->lock); in perf_group_detach()
2262 * We can have double detach due to exit/hot-unplug + close. in perf_group_detach()
2264 if (!(event->attach_state & PERF_ATTACH_GROUP)) in perf_group_detach()
2267 event->attach_state &= ~PERF_ATTACH_GROUP; in perf_group_detach()
2274 if (leader != event) { in perf_group_detach()
2275 list_del_init(&event->sibling_list); in perf_group_detach()
2276 event->group_leader->nr_siblings--; in perf_group_detach()
2277 event->group_leader->group_generation++; in perf_group_detach()
2286 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, sibling_list) { in perf_group_detach()
2288 if (sibling->event_caps & PERF_EV_CAP_SIBLING) in perf_group_detach()
2291 sibling->group_leader = sibling; in perf_group_detach()
2292 list_del_init(&sibling->sibling_list); in perf_group_detach()
2294 /* Inherit group flags from the previous leader */ in perf_group_detach()
2295 sibling->group_caps = event->group_caps; in perf_group_detach()
2297 if (sibling->attach_state & PERF_ATTACH_CONTEXT) { in perf_group_detach()
2298 add_event_to_groups(sibling, event->ctx); in perf_group_detach()
2300 if (sibling->state == PERF_EVENT_STATE_ACTIVE) in perf_group_detach()
2301 list_add_tail(&sibling->active_list, get_event_list(sibling)); in perf_group_detach()
2304 WARN_ON_ONCE(sibling->ctx != event->ctx); in perf_group_detach()
2308 for_each_sibling_event(tmp, leader) in perf_group_detach()
2311 perf_event__header_size(leader); in perf_group_detach()
2318 struct perf_event *parent_event = event->parent; in perf_child_detach()
2320 if (!(event->attach_state & PERF_ATTACH_CHILD)) in perf_child_detach()
2323 event->attach_state &= ~PERF_ATTACH_CHILD; in perf_child_detach()
2328 lockdep_assert_held(&parent_event->child_mutex); in perf_child_detach()
2331 list_del_init(&event->child_list); in perf_child_detach()
2336 return event->state == PERF_EVENT_STATE_DEAD; in is_orphaned_event()
2342 return (event->cpu == -1 || event->cpu == smp_processor_id()) && in event_filter_match()
2349 struct perf_event_pmu_context *epc = event->pmu_ctx; in event_sched_out()
2350 struct perf_cpu_pmu_context *cpc = this_cpc(epc->pmu); in event_sched_out()
2353 // XXX cpc serialization, probably per-cpu IRQ disabled in event_sched_out()
2355 WARN_ON_ONCE(event->ctx != ctx); in event_sched_out()
2356 lockdep_assert_held(&ctx->lock); in event_sched_out()
2358 if (event->state != PERF_EVENT_STATE_ACTIVE) in event_sched_out()
2366 list_del_init(&event->active_list); in event_sched_out()
2368 perf_pmu_disable(event->pmu); in event_sched_out()
2370 event->pmu->del(event, 0); in event_sched_out()
2371 event->oncpu = -1; in event_sched_out()
2373 if (event->pending_disable) { in event_sched_out()
2374 event->pending_disable = 0; in event_sched_out()
2382 cpc->active_oncpu--; in event_sched_out()
2383 if (event->attr.freq && event->attr.sample_freq) { in event_sched_out()
2384 ctx->nr_freq--; in event_sched_out()
2385 epc->nr_freq--; in event_sched_out()
2387 if (event->attr.exclusive || !cpc->active_oncpu) in event_sched_out()
2388 cpc->exclusive = 0; in event_sched_out()
2390 perf_pmu_enable(event->pmu); in event_sched_out()
2398 if (group_event->state != PERF_EVENT_STATE_ACTIVE) in group_sched_out()
2401 perf_assert_pmu_disabled(group_event->pmu_ctx->pmu); in group_sched_out()
2415 if (ctx->is_active & EVENT_TIME) { in __ctx_time_update()
2416 if (ctx->is_active & EVENT_FROZEN) in __ctx_time_update()
2436 if (ctx->is_active & EVENT_TIME) in ctx_time_freeze()
2437 ctx->is_active |= EVENT_FROZEN; in ctx_time_freeze()
2443 if (ctx->is_active & EVENT_TIME) { in ctx_time_update_event()
2444 if (ctx->is_active & EVENT_FROZEN) in ctx_time_update_event()
2468 struct perf_event_pmu_context *pmu_ctx = event->pmu_ctx; in __perf_remove_from_context()
2481 event->pending_disable = 1; in __perf_remove_from_context()
2485 perf_event_set_state(event, min(event->state, state)); in __perf_remove_from_context()
2492 if (!pmu_ctx->nr_events) { in __perf_remove_from_context()
2493 pmu_ctx->rotate_necessary = 0; in __perf_remove_from_context()
2495 if (ctx->task && ctx->is_active) { in __perf_remove_from_context()
2496 struct perf_cpu_pmu_context *cpc = this_cpc(pmu_ctx->pmu); in __perf_remove_from_context()
2498 WARN_ON_ONCE(cpc->task_epc && cpc->task_epc != pmu_ctx); in __perf_remove_from_context()
2499 cpc->task_epc = NULL; in __perf_remove_from_context()
2503 if (!ctx->nr_events && ctx->is_active) { in __perf_remove_from_context()
2504 if (ctx == &cpuctx->ctx) in __perf_remove_from_context()
2507 ctx->is_active = 0; in __perf_remove_from_context()
2508 if (ctx->task) { in __perf_remove_from_context()
2509 WARN_ON_ONCE(cpuctx->task_ctx != ctx); in __perf_remove_from_context()
2510 cpuctx->task_ctx = NULL; in __perf_remove_from_context()
2518 * If event->ctx is a cloned context, callers must make sure that
2519 * every task struct that event->ctx->task could possibly point to
2521 * that only calls us on the top-level context, which can't be a clone.
2527 struct perf_event_context *ctx = event->ctx; in perf_remove_from_context()
2529 lockdep_assert_held(&ctx->mutex); in perf_remove_from_context()
2536 raw_spin_lock_irq(&ctx->lock); in perf_remove_from_context()
2537 if (!ctx->is_active) { in perf_remove_from_context()
2540 raw_spin_unlock_irq(&ctx->lock); in perf_remove_from_context()
2543 raw_spin_unlock_irq(&ctx->lock); in perf_remove_from_context()
2556 if (event->state < PERF_EVENT_STATE_INACTIVE) in __perf_event_disable()
2559 perf_pmu_disable(event->pmu_ctx->pmu); in __perf_event_disable()
2562 if (event == event->group_leader) in __perf_event_disable()
2570 perf_pmu_enable(event->pmu_ctx->pmu); in __perf_event_disable()
2576 * If event->ctx is a cloned context, callers must make sure that
2577 * every task struct that event->ctx->task could possibly point to
2580 * hold the top-level event's child_mutex, so any descendant that
2583 * When called from perf_pending_disable it's OK because event->ctx
2589 struct perf_event_context *ctx = event->ctx; in _perf_event_disable()
2591 raw_spin_lock_irq(&ctx->lock); in _perf_event_disable()
2592 if (event->state <= PERF_EVENT_STATE_OFF) { in _perf_event_disable()
2593 raw_spin_unlock_irq(&ctx->lock); in _perf_event_disable()
2596 raw_spin_unlock_irq(&ctx->lock); in _perf_event_disable()
2622 event->pending_disable = 1; in perf_event_disable_inatomic()
2623 irq_work_queue(&event->pending_disable_irq); in perf_event_disable_inatomic()
2634 struct perf_event_pmu_context *epc = event->pmu_ctx; in event_sched_in()
2635 struct perf_cpu_pmu_context *cpc = this_cpc(epc->pmu); in event_sched_in()
2638 WARN_ON_ONCE(event->ctx != ctx); in event_sched_in()
2640 lockdep_assert_held(&ctx->lock); in event_sched_in()
2642 if (event->state <= PERF_EVENT_STATE_OFF) in event_sched_in()
2645 WRITE_ONCE(event->oncpu, smp_processor_id()); in event_sched_in()
2649 * ->oncpu if it sees ACTIVE. in event_sched_in()
2659 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { in event_sched_in()
2661 event->hw.interrupts = 0; in event_sched_in()
2664 perf_pmu_disable(event->pmu); in event_sched_in()
2668 if (event->pmu->add(event, PERF_EF_START)) { in event_sched_in()
2670 event->oncpu = -1; in event_sched_in()
2671 ret = -EAGAIN; in event_sched_in()
2676 cpc->active_oncpu++; in event_sched_in()
2677 if (event->attr.freq && event->attr.sample_freq) { in event_sched_in()
2678 ctx->nr_freq++; in event_sched_in()
2679 epc->nr_freq++; in event_sched_in()
2681 if (event->attr.exclusive) in event_sched_in()
2682 cpc->exclusive = 1; in event_sched_in()
2685 perf_pmu_enable(event->pmu); in event_sched_in()
2694 struct pmu *pmu = group_event->pmu_ctx->pmu; in group_sched_in()
2696 if (group_event->state == PERF_EVENT_STATE_OFF) in group_sched_in()
2699 pmu->start_txn(pmu, PERF_PMU_TXN_ADD); in group_sched_in()
2714 if (!pmu->commit_txn(pmu)) in group_sched_in()
2732 pmu->cancel_txn(pmu); in group_sched_in()
2733 return -EAGAIN; in group_sched_in()
2741 struct perf_event_pmu_context *epc = event->pmu_ctx; in group_can_go_on()
2742 struct perf_cpu_pmu_context *cpc = this_cpc(epc->pmu); in group_can_go_on()
2747 if (event->group_caps & PERF_EV_CAP_SOFTWARE) in group_can_go_on()
2753 if (cpc->exclusive) in group_can_go_on()
2759 if (event->attr.exclusive && !list_empty(get_event_list(event))) in group_can_go_on()
2781 if (!cpuctx->task_ctx) in task_ctx_sched_out()
2784 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) in task_ctx_sched_out()
2794 ctx_sched_in(&cpuctx->ctx, pmu, EVENT_PINNED); in perf_event_sched_in()
2797 ctx_sched_in(&cpuctx->ctx, pmu, EVENT_FLEXIBLE); in perf_event_sched_in()
2804 * - CPU pinned (EVENT_CPU | EVENT_PINNED)
2805 * - task pinned (EVENT_PINNED)
2806 * - CPU flexible (EVENT_CPU | EVENT_FLEXIBLE)
2807 * - task flexible (EVENT_FLEXIBLE).
2833 for_each_epc(epc, &cpuctx->ctx, pmu, false) in ctx_resched()
2834 perf_pmu_disable(epc->pmu); in ctx_resched()
2838 perf_pmu_disable(epc->pmu); in ctx_resched()
2846 * - EVENT_CPU: schedule out corresponding groups; in ctx_resched()
2847 * - EVENT_PINNED task events: schedule out EVENT_FLEXIBLE groups; in ctx_resched()
2848 * - otherwise, do nothing more. in ctx_resched()
2851 ctx_sched_out(&cpuctx->ctx, pmu, event_type); in ctx_resched()
2853 ctx_sched_out(&cpuctx->ctx, pmu, EVENT_FLEXIBLE); in ctx_resched()
2857 for_each_epc(epc, &cpuctx->ctx, pmu, false) in ctx_resched()
2858 perf_pmu_enable(epc->pmu); in ctx_resched()
2862 perf_pmu_enable(epc->pmu); in ctx_resched()
2869 struct perf_event_context *task_ctx = cpuctx->task_ctx; in perf_pmu_resched()
2880 * things like ctx->is_active and cpuctx->task_ctx are set.
2885 struct perf_event_context *ctx = event->ctx; in __perf_install_in_context()
2887 struct perf_event_context *task_ctx = cpuctx->task_ctx; in __perf_install_in_context()
2891 raw_spin_lock(&cpuctx->ctx.lock); in __perf_install_in_context()
2892 if (ctx->task) { in __perf_install_in_context()
2893 raw_spin_lock(&ctx->lock); in __perf_install_in_context()
2896 reprogram = (ctx->task == current); in __perf_install_in_context()
2902 * If its not running, we don't care, ctx->lock will in __perf_install_in_context()
2905 if (task_curr(ctx->task) && !reprogram) { in __perf_install_in_context()
2906 ret = -ESRCH; in __perf_install_in_context()
2910 WARN_ON_ONCE(reprogram && cpuctx->task_ctx && cpuctx->task_ctx != ctx); in __perf_install_in_context()
2912 raw_spin_lock(&task_ctx->lock); in __perf_install_in_context()
2916 if (event->state > PERF_EVENT_STATE_OFF && is_cgroup_event(event)) { in __perf_install_in_context()
2922 reprogram = cgroup_is_descendant(cgrp->css.cgroup, in __perf_install_in_context()
2923 event->cgrp->css.cgroup); in __perf_install_in_context()
2930 ctx_resched(cpuctx, task_ctx, event->pmu_ctx->pmu, in __perf_install_in_context()
2955 struct task_struct *task = READ_ONCE(ctx->task); in perf_install_in_context()
2957 lockdep_assert_held(&ctx->mutex); in perf_install_in_context()
2961 if (event->cpu != -1) in perf_install_in_context()
2962 WARN_ON_ONCE(event->cpu != cpu); in perf_install_in_context()
2965 * Ensures that if we can observe event->ctx, both the event and ctx in perf_install_in_context()
2968 smp_store_release(&event->ctx, ctx); in perf_install_in_context()
2973 * that case we need the magic of the IPI to set ctx->is_active. in perf_install_in_context()
2979 ctx->nr_events && !is_cgroup_event(event)) { in perf_install_in_context()
2980 raw_spin_lock_irq(&ctx->lock); in perf_install_in_context()
2981 if (ctx->task == TASK_TOMBSTONE) { in perf_install_in_context()
2982 raw_spin_unlock_irq(&ctx->lock); in perf_install_in_context()
2986 raw_spin_unlock_irq(&ctx->lock); in perf_install_in_context()
3002 * Installing events is tricky because we cannot rely on ctx->is_active in perf_install_in_context()
3003 * to be set in case this is the nr_events 0 -> 1 transition. in perf_install_in_context()
3017 * our task->perf_event_ctxp[] store, such that it will in fact take in perf_install_in_context()
3026 * This smp_mb() orders the task->perf_event_ctxp[] store with the in perf_install_in_context()
3036 raw_spin_lock_irq(&ctx->lock); in perf_install_in_context()
3037 task = ctx->task; in perf_install_in_context()
3041 * cannot happen), and we hold ctx->mutex, which serializes us in perf_install_in_context()
3044 raw_spin_unlock_irq(&ctx->lock); in perf_install_in_context()
3048 * If the task is not running, ctx->lock will avoid it becoming so, in perf_install_in_context()
3052 raw_spin_unlock_irq(&ctx->lock); in perf_install_in_context()
3056 raw_spin_unlock_irq(&ctx->lock); in perf_install_in_context()
3067 struct perf_event *leader = event->group_leader; in __perf_event_enable() local
3070 if (event->state >= PERF_EVENT_STATE_INACTIVE || in __perf_event_enable()
3071 event->state <= PERF_EVENT_STATE_ERROR) in __perf_event_enable()
3079 if (!ctx->is_active) in __perf_event_enable()
3086 * If the event is in a group and isn't the group leader, in __perf_event_enable()
3089 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) in __perf_event_enable()
3092 task_ctx = cpuctx->task_ctx; in __perf_event_enable()
3093 if (ctx->task) in __perf_event_enable()
3096 ctx_resched(cpuctx, task_ctx, event->pmu_ctx->pmu, get_event_type(event)); in __perf_event_enable()
3102 * If event->ctx is a cloned context, callers must make sure that
3103 * every task struct that event->ctx->task could possibly point to
3110 struct perf_event_context *ctx = event->ctx; in _perf_event_enable()
3112 raw_spin_lock_irq(&ctx->lock); in _perf_event_enable()
3113 if (event->state >= PERF_EVENT_STATE_INACTIVE || in _perf_event_enable()
3114 event->state < PERF_EVENT_STATE_ERROR) { in _perf_event_enable()
3116 raw_spin_unlock_irq(&ctx->lock); in _perf_event_enable()
3125 * been scheduled away before the cross-call arrived. in _perf_event_enable()
3127 if (event->state == PERF_EVENT_STATE_ERROR) { in _perf_event_enable()
3131 if (event->event_caps & PERF_EV_CAP_SIBLING && in _perf_event_enable()
3132 event->group_leader == event) in _perf_event_enable()
3135 event->state = PERF_EVENT_STATE_OFF; in _perf_event_enable()
3137 raw_spin_unlock_irq(&ctx->lock); in _perf_event_enable()
3163 struct perf_event *event = sd->event; in __perf_event_stop()
3166 if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE) in __perf_event_stop()
3176 if (READ_ONCE(event->oncpu) != smp_processor_id()) in __perf_event_stop()
3177 return -EAGAIN; in __perf_event_stop()
3179 event->pmu->stop(event, PERF_EF_UPDATE); in __perf_event_stop()
3187 * Since this is happening on an event-local CPU, no trace is lost in __perf_event_stop()
3190 if (sd->restart) in __perf_event_stop()
3191 event->pmu->start(event, 0); in __perf_event_stop()
3205 if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE) in perf_event_stop()
3213 * inactive here (event->oncpu==-1), there's nothing more to do; in perf_event_stop()
3214 * fall through with ret==-ENXIO. in perf_event_stop()
3216 ret = cpu_function_call(READ_ONCE(event->oncpu), in perf_event_stop()
3218 } while (ret == -EAGAIN); in perf_event_stop()
3237 * pre-existing mappings, called once when new filters arrive via SET_FILTER
3252 raw_spin_lock(&ifh->lock); in perf_event_addr_filters_sync()
3253 if (event->addr_filters_gen != event->hw.addr_filters_gen) { in perf_event_addr_filters_sync()
3254 event->pmu->addr_filters_sync(event); in perf_event_addr_filters_sync()
3255 event->hw.addr_filters_gen = event->addr_filters_gen; in perf_event_addr_filters_sync()
3257 raw_spin_unlock(&ifh->lock); in perf_event_addr_filters_sync()
3266 if (event->attr.inherit || !is_sampling_event(event)) in _perf_event_refresh()
3267 return -EINVAL; in _perf_event_refresh()
3269 atomic_add(refresh, &event->event_limit); in _perf_event_refresh()
3300 if (!bp->attr.disabled) in perf_event_modify_breakpoint()
3307 * Copy event-type-independent attributes that may be modified.
3312 to->sig_data = from->sig_data; in perf_event_modify_copy_attr()
3322 if (event->attr.type != attr->type) in perf_event_modify_attr()
3323 return -EINVAL; in perf_event_modify_attr()
3325 switch (event->attr.type) { in perf_event_modify_attr()
3331 return -EOPNOTSUPP; in perf_event_modify_attr()
3334 WARN_ON_ONCE(event->ctx->parent_ctx); in perf_event_modify_attr()
3336 mutex_lock(&event->child_mutex); in perf_event_modify_attr()
3338 * Event-type-independent attributes must be copied before event-type in perf_event_modify_attr()
3342 perf_event_modify_copy_attr(&event->attr, attr); in perf_event_modify_attr()
3346 list_for_each_entry(child, &event->child_list, child_list) { in perf_event_modify_attr()
3347 perf_event_modify_copy_attr(&child->attr, attr); in perf_event_modify_attr()
3353 mutex_unlock(&event->child_mutex); in perf_event_modify_attr()
3360 struct perf_event_context *ctx = pmu_ctx->ctx; in __pmu_ctx_sched_out()
3362 struct pmu *pmu = pmu_ctx->pmu; in __pmu_ctx_sched_out()
3364 if (ctx->task && !(ctx->is_active & EVENT_ALL)) { in __pmu_ctx_sched_out()
3367 WARN_ON_ONCE(cpc->task_epc && cpc->task_epc != pmu_ctx); in __pmu_ctx_sched_out()
3368 cpc->task_epc = NULL; in __pmu_ctx_sched_out()
3377 &pmu_ctx->pinned_active, in __pmu_ctx_sched_out()
3384 &pmu_ctx->flexible_active, in __pmu_ctx_sched_out()
3392 pmu_ctx->rotate_necessary = 0; in __pmu_ctx_sched_out()
3403 * state, eg. ctx->is_active cleared even though most EPCs would still actually
3411 int is_active = ctx->is_active; in ctx_sched_out()
3416 lockdep_assert_held(&ctx->lock); in ctx_sched_out()
3418 if (likely(!ctx->nr_events)) { in ctx_sched_out()
3422 WARN_ON_ONCE(ctx->is_active); in ctx_sched_out()
3423 if (ctx->task) in ctx_sched_out()
3424 WARN_ON_ONCE(cpuctx->task_ctx); in ctx_sched_out()
3438 __ctx_time_update(cpuctx, ctx, ctx == &cpuctx->ctx); in ctx_sched_out()
3441 * CPU-release for the below ->is_active store, in ctx_sched_out()
3445 ctx->is_active &= ~event_type; in ctx_sched_out()
3447 if (!(ctx->is_active & EVENT_ALL)) { in ctx_sched_out()
3452 if (ctx->is_active & EVENT_FROZEN) in ctx_sched_out()
3453 ctx->is_active &= EVENT_TIME_FROZEN; in ctx_sched_out()
3455 ctx->is_active = 0; in ctx_sched_out()
3458 if (ctx->task) { in ctx_sched_out()
3459 WARN_ON_ONCE(cpuctx->task_ctx != ctx); in ctx_sched_out()
3460 if (!(ctx->is_active & EVENT_ALL)) in ctx_sched_out()
3461 cpuctx->task_ctx = NULL; in ctx_sched_out()
3464 is_active ^= ctx->is_active; /* changed bits */ in ctx_sched_out()
3481 lockdep_assert_held(&ctx1->lock); in context_equiv()
3482 lockdep_assert_held(&ctx2->lock); in context_equiv()
3485 if (ctx1->pin_count || ctx2->pin_count) in context_equiv()
3489 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen) in context_equiv()
3493 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation) in context_equiv()
3500 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx && in context_equiv()
3501 ctx1->parent_gen == ctx2->parent_gen) in context_equiv()
3513 if (!event->attr.inherit_stat) in __perf_event_sync_stat()
3528 * In order to keep per-task stats reliable we need to flip the event in __perf_event_sync_stat()
3531 value = local64_read(&next_event->count); in __perf_event_sync_stat()
3532 value = local64_xchg(&event->count, value); in __perf_event_sync_stat()
3533 local64_set(&next_event->count, value); in __perf_event_sync_stat()
3535 swap(event->total_time_enabled, next_event->total_time_enabled); in __perf_event_sync_stat()
3536 swap(event->total_time_running, next_event->total_time_running); in __perf_event_sync_stat()
3550 if (!ctx->nr_stat) in perf_event_sync_stat()
3555 event = list_first_entry(&ctx->event_list, in perf_event_sync_stat()
3558 next_event = list_first_entry(&next_ctx->event_list, in perf_event_sync_stat()
3561 while (&event->event_entry != &ctx->event_list && in perf_event_sync_stat()
3562 &next_event->event_entry != &next_ctx->event_list) { in perf_event_sync_stat()
3577 list_for_each_entry(pmu_ctx, &ctx->pmu_ctx_list, pmu_ctx_entry) { in perf_ctx_sched_task_cb()
3578 cpc = this_cpc(pmu_ctx->pmu); in perf_ctx_sched_task_cb()
3580 if (cpc->sched_cb_usage && pmu_ctx->pmu->sched_task) in perf_ctx_sched_task_cb()
3581 pmu_ctx->pmu->sched_task(pmu_ctx, task, sched_in); in perf_ctx_sched_task_cb()
3588 struct perf_event_context *ctx = task->perf_event_ctxp; in perf_event_context_sched_out()
3597 next_ctx = rcu_dereference(next->perf_event_ctxp); in perf_event_context_sched_out()
3601 parent = rcu_dereference(ctx->parent_ctx); in perf_event_context_sched_out()
3602 next_parent = rcu_dereference(next_ctx->parent_ctx); in perf_event_context_sched_out()
3613 * lock (including re-checking that neither has been in perf_event_context_sched_out()
3618 raw_spin_lock(&ctx->lock); in perf_event_context_sched_out()
3619 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); in perf_event_context_sched_out()
3624 /* PMIs are disabled; ctx->nr_no_switch_fast is stable. */ in perf_event_context_sched_out()
3625 if (local_read(&ctx->nr_no_switch_fast) || in perf_event_context_sched_out()
3626 local_read(&next_ctx->nr_no_switch_fast)) { in perf_event_context_sched_out()
3629 * events that rely on the ctx->task relation. in perf_event_context_sched_out()
3636 raw_spin_unlock(&next_ctx->lock); in perf_event_context_sched_out()
3641 WRITE_ONCE(ctx->task, next); in perf_event_context_sched_out()
3642 WRITE_ONCE(next_ctx->task, task); in perf_event_context_sched_out()
3651 * ctx->task is immaterial since this value is in perf_event_context_sched_out()
3652 * always verified under ctx->lock which we're now in perf_event_context_sched_out()
3655 RCU_INIT_POINTER(task->perf_event_ctxp, next_ctx); in perf_event_context_sched_out()
3656 RCU_INIT_POINTER(next->perf_event_ctxp, ctx); in perf_event_context_sched_out()
3662 raw_spin_unlock(&next_ctx->lock); in perf_event_context_sched_out()
3663 raw_spin_unlock(&ctx->lock); in perf_event_context_sched_out()
3669 raw_spin_lock(&ctx->lock); in perf_event_context_sched_out()
3677 raw_spin_unlock(&ctx->lock); in perf_event_context_sched_out()
3691 if (!--cpc->sched_cb_usage) in perf_sched_cb_dec()
3692 list_del(&cpc->sched_cb_entry); in perf_sched_cb_dec()
3700 if (!cpc->sched_cb_usage++) in perf_sched_cb_inc()
3701 list_add(&cpc->sched_cb_entry, this_cpu_ptr(&sched_cb_list)); in perf_sched_cb_inc()
3711 * This callback is relevant even to per-cpu events; for example multi event
3721 pmu = cpc->epc.pmu; in __perf_pmu_sched_task()
3724 if (WARN_ON_ONCE(!pmu->sched_task)) in __perf_pmu_sched_task()
3727 perf_ctx_lock(cpuctx, cpuctx->task_ctx); in __perf_pmu_sched_task()
3730 pmu->sched_task(cpc->task_epc, task, sched_in); in __perf_pmu_sched_task()
3733 perf_ctx_unlock(cpuctx, cpuctx->task_ctx); in __perf_pmu_sched_task()
3743 /* cpuctx->task_ctx will be handled in perf_event_context_sched_in/out */ in perf_pmu_sched_task()
3744 if (prev == next || cpuctx->task_ctx) in perf_pmu_sched_task()
3758 * We stop each event and update the event value in event->count.
3779 * cgroup event are system-wide mode only in __perf_event_task_sched_out()
3789 return le->group_index < re->group_index; in perf_less_group_idx()
3801 struct perf_event **itrs = heap->data; in __heap_add()
3804 itrs[heap->nr] = event; in __heap_add()
3805 heap->nr++; in __heap_add()
3813 if (!pmu_ctx->ctx->task) in __link_epc()
3816 cpc = this_cpc(pmu_ctx->pmu); in __link_epc()
3817 WARN_ON_ONCE(cpc->task_epc && cpc->task_epc != pmu_ctx); in __link_epc()
3818 cpc->task_epc = pmu_ctx; in __link_epc()
3837 if (pmu->filter && pmu->filter(pmu, cpu)) in visit_groups_merge()
3840 if (!ctx->task) { in visit_groups_merge()
3843 .data = cpuctx->heap, in visit_groups_merge()
3845 .size = cpuctx->heap_size, in visit_groups_merge()
3848 lockdep_assert_held(&cpuctx->ctx.lock); in visit_groups_merge()
3851 if (cpuctx->cgrp) in visit_groups_merge()
3852 css = &cpuctx->cgrp->css; in visit_groups_merge()
3861 __heap_add(&event_heap, perf_event_groups_first(groups, -1, pmu, NULL)); in visit_groups_merge()
3868 for (; css; css = css->parent) in visit_groups_merge()
3869 __heap_add(&event_heap, perf_event_groups_first(groups, cpu, pmu, css->cgroup)); in visit_groups_merge()
3873 __link_epc((*evt)->pmu_ctx); in visit_groups_merge()
3874 perf_assert_pmu_disabled((*evt)->pmu_ctx->pmu); in visit_groups_merge()
3895 * Because the userpage is strictly per-event (there is no concept of context,
3897 * when context time starts :-(
3903 if (likely(!atomic_read(&event->mmap_count))) in event_update_userpage()
3925 struct perf_event_context *ctx = event->ctx; in merge_sched_in()
3928 if (event->state <= PERF_EVENT_STATE_OFF) in merge_sched_in()
3936 list_add_tail(&event->active_list, get_event_list(event)); in merge_sched_in()
3939 if (event->state == PERF_EVENT_STATE_INACTIVE) { in merge_sched_in()
3941 if (event->attr.pinned) { in merge_sched_in()
3946 event->pending_kill = POLL_ERR; in merge_sched_in()
3950 struct perf_cpu_pmu_context *cpc = this_cpc(event->pmu_ctx->pmu); in merge_sched_in()
3952 event->pmu_ctx->rotate_necessary = 1; in merge_sched_in()
3973 struct perf_event_context *ctx = pmu_ctx->ctx; in __pmu_ctx_sched_in()
3976 pmu_groups_sched_in(ctx, &ctx->pinned_groups, pmu_ctx->pmu); in __pmu_ctx_sched_in()
3978 pmu_groups_sched_in(ctx, &ctx->flexible_groups, pmu_ctx->pmu); in __pmu_ctx_sched_in()
3986 int is_active = ctx->is_active; in ctx_sched_in()
3991 lockdep_assert_held(&ctx->lock); in ctx_sched_in()
3993 if (likely(!ctx->nr_events)) in ctx_sched_in()
4001 * CPU-release for the below ->is_active store, in ctx_sched_in()
4007 ctx->is_active |= (event_type | EVENT_TIME); in ctx_sched_in()
4008 if (ctx->task) { in ctx_sched_in()
4010 cpuctx->task_ctx = ctx; in ctx_sched_in()
4012 WARN_ON_ONCE(cpuctx->task_ctx != ctx); in ctx_sched_in()
4015 is_active ^= ctx->is_active; /* changed bits */ in ctx_sched_in()
4039 ctx = rcu_dereference(task->perf_event_ctxp); in perf_event_context_sched_in()
4043 if (cpuctx->task_ctx == ctx) { in perf_event_context_sched_in()
4056 * We must check ctx->nr_events while holding ctx->lock, such in perf_event_context_sched_in()
4059 if (!ctx->nr_events) in perf_event_context_sched_in()
4071 if (!RB_EMPTY_ROOT(&ctx->pinned_groups.tree)) { in perf_event_context_sched_in()
4072 perf_ctx_disable(&cpuctx->ctx, false); in perf_event_context_sched_in()
4073 ctx_sched_out(&cpuctx->ctx, NULL, EVENT_FLEXIBLE); in perf_event_context_sched_in()
4078 perf_ctx_sched_task_cb(cpuctx->task_ctx, task, true); in perf_event_context_sched_in()
4080 if (!RB_EMPTY_ROOT(&ctx->pinned_groups.tree)) in perf_event_context_sched_in()
4081 perf_ctx_enable(&cpuctx->ctx, false); in perf_event_context_sched_in()
4116 u64 frequency = event->attr.sample_freq; in perf_calculate_period()
4132 * period = ------------------- in perf_calculate_period()
4145 a##_fls--; \ in perf_calculate_period()
4148 b##_fls--; \ in perf_calculate_period()
4192 struct hw_perf_event *hwc = &event->hw; in perf_adjust_period()
4198 delta = (s64)(period - hwc->sample_period); in perf_adjust_period()
4202 delta -= 7; in perf_adjust_period()
4205 sample_period = hwc->sample_period + delta; in perf_adjust_period()
4210 hwc->sample_period = sample_period; in perf_adjust_period()
4212 if (local64_read(&hwc->period_left) > 8*sample_period) { in perf_adjust_period()
4214 event->pmu->stop(event, PERF_EF_UPDATE); in perf_adjust_period()
4216 local64_set(&hwc->period_left, 0); in perf_adjust_period()
4219 event->pmu->start(event, PERF_EF_RELOAD); in perf_adjust_period()
4231 if (event->state != PERF_EVENT_STATE_ACTIVE) in perf_adjust_freq_unthr_events()
4234 // XXX use visit thingy to avoid the -1,cpu match in perf_adjust_freq_unthr_events()
4238 hwc = &event->hw; in perf_adjust_freq_unthr_events()
4240 if (hwc->interrupts == MAX_INTERRUPTS) { in perf_adjust_freq_unthr_events()
4241 hwc->interrupts = 0; in perf_adjust_freq_unthr_events()
4243 if (!event->attr.freq || !event->attr.sample_freq) in perf_adjust_freq_unthr_events()
4244 event->pmu->start(event, 0); in perf_adjust_freq_unthr_events()
4247 if (!event->attr.freq || !event->attr.sample_freq) in perf_adjust_freq_unthr_events()
4251 * stop the event and update event->count in perf_adjust_freq_unthr_events()
4253 event->pmu->stop(event, PERF_EF_UPDATE); in perf_adjust_freq_unthr_events()
4255 now = local64_read(&event->count); in perf_adjust_freq_unthr_events()
4256 delta = now - hwc->freq_count_stamp; in perf_adjust_freq_unthr_events()
4257 hwc->freq_count_stamp = now; in perf_adjust_freq_unthr_events()
4269 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); in perf_adjust_freq_unthr_events()
4285 * - context have events in frequency mode (needs freq adjust) in perf_adjust_freq_unthr_context()
4286 * - there are events to unthrottle on this cpu in perf_adjust_freq_unthr_context()
4288 if (!(ctx->nr_freq || unthrottle)) in perf_adjust_freq_unthr_context()
4291 raw_spin_lock(&ctx->lock); in perf_adjust_freq_unthr_context()
4293 list_for_each_entry(pmu_ctx, &ctx->pmu_ctx_list, pmu_ctx_entry) { in perf_adjust_freq_unthr_context()
4294 if (!(pmu_ctx->nr_freq || unthrottle)) in perf_adjust_freq_unthr_context()
4298 if (pmu_ctx->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) in perf_adjust_freq_unthr_context()
4301 perf_pmu_disable(pmu_ctx->pmu); in perf_adjust_freq_unthr_context()
4302 perf_adjust_freq_unthr_events(&pmu_ctx->pinned_active); in perf_adjust_freq_unthr_context()
4303 perf_adjust_freq_unthr_events(&pmu_ctx->flexible_active); in perf_adjust_freq_unthr_context()
4304 perf_pmu_enable(pmu_ctx->pmu); in perf_adjust_freq_unthr_context()
4307 raw_spin_unlock(&ctx->lock); in perf_adjust_freq_unthr_context()
4316 * Rotate the first entry last of non-pinned groups. Rotation might be in rotate_ctx()
4319 if (ctx->rotate_disable) in rotate_ctx()
4322 perf_event_groups_delete(&ctx->flexible_groups, event); in rotate_ctx()
4323 perf_event_groups_insert(&ctx->flexible_groups, event); in rotate_ctx()
4334 .pmu = pmu_ctx->pmu, in ctx_event_to_rotate()
4338 event = list_first_entry_or_null(&pmu_ctx->flexible_active, in ctx_event_to_rotate()
4344 tree = &pmu_ctx->ctx->flexible_groups.tree; in ctx_event_to_rotate()
4346 if (!pmu_ctx->ctx->task) { in ctx_event_to_rotate()
4355 key.cpu = -1; in ctx_event_to_rotate()
4372 pmu_ctx->rotate_necessary = 0; in ctx_event_to_rotate()
4390 cpu_epc = &cpc->epc; in perf_rotate_context()
4391 pmu = cpu_epc->pmu; in perf_rotate_context()
4392 task_epc = cpc->task_epc; in perf_rotate_context()
4394 cpu_rotate = cpu_epc->rotate_necessary; in perf_rotate_context()
4395 task_rotate = task_epc ? task_epc->rotate_necessary : 0; in perf_rotate_context()
4400 perf_ctx_lock(cpuctx, cpuctx->task_ctx); in perf_rotate_context()
4413 update_context_time(task_epc->ctx); in perf_rotate_context()
4418 update_context_time(&cpuctx->ctx); in perf_rotate_context()
4420 rotate_ctx(&cpuctx->ctx, cpu_event); in perf_rotate_context()
4425 rotate_ctx(task_epc->ctx, task_event); in perf_rotate_context()
4431 perf_ctx_unlock(cpuctx, cpuctx->task_ctx); in perf_rotate_context()
4448 perf_adjust_freq_unthr_context(&cpuctx->ctx, !!throttled); in perf_event_task_tick()
4451 ctx = rcu_dereference(current->perf_event_ctxp); in perf_event_task_tick()
4460 if (!event->attr.enable_on_exec) in event_enable_on_exec()
4463 event->attr.enable_on_exec = 0; in event_enable_on_exec()
4464 if (event->state >= PERF_EVENT_STATE_INACTIVE) in event_enable_on_exec()
4473 * Enable all of a task's events that have been marked enable-on-exec.
4486 if (WARN_ON_ONCE(current->perf_event_ctxp != ctx)) in perf_event_enable_on_exec()
4489 if (!ctx->nr_events) in perf_event_enable_on_exec()
4496 list_for_each_entry(event, &ctx->event_list, event_entry) { in perf_event_enable_on_exec()
4523 * remove-on-exec, and feeds their values back to parent events.
4532 mutex_lock(&ctx->mutex); in perf_event_remove_on_exec()
4534 if (WARN_ON_ONCE(ctx->task != current)) in perf_event_remove_on_exec()
4537 list_for_each_entry_safe(event, next, &ctx->event_list, event_entry) { in perf_event_remove_on_exec()
4538 if (!event->attr.remove_on_exec) in perf_event_remove_on_exec()
4549 raw_spin_lock_irqsave(&ctx->lock, flags); in perf_event_remove_on_exec()
4552 raw_spin_unlock_irqrestore(&ctx->lock, flags); in perf_event_remove_on_exec()
4555 mutex_unlock(&ctx->mutex); in perf_event_remove_on_exec()
4577 if (event->group_caps & PERF_EV_CAP_READ_SCOPE) { in __perf_event_read_cpu()
4578 const struct cpumask *cpumask = perf_scope_cpu_topology_cpumask(event->pmu->scope, event_cpu); in __perf_event_read_cpu()
4584 if (event->group_caps & PERF_EV_CAP_READ_ACTIVE_PKG) { in __perf_event_read_cpu()
4601 struct perf_event *sub, *event = data->event; in __perf_event_read()
4602 struct perf_event_context *ctx = event->ctx; in __perf_event_read()
4604 struct pmu *pmu = event->pmu; in __perf_event_read()
4610 * event->count would have been updated to a recent sample in __perf_event_read()
4613 if (ctx->task && cpuctx->task_ctx != ctx) in __perf_event_read()
4616 raw_spin_lock(&ctx->lock); in __perf_event_read()
4620 if (data->group) in __perf_event_read()
4623 if (event->state != PERF_EVENT_STATE_ACTIVE) in __perf_event_read()
4626 if (!data->group) { in __perf_event_read()
4627 pmu->read(event); in __perf_event_read()
4628 data->ret = 0; in __perf_event_read()
4632 pmu->start_txn(pmu, PERF_PMU_TXN_READ); in __perf_event_read()
4634 pmu->read(event); in __perf_event_read()
4639 data->ret = pmu->commit_txn(pmu); in __perf_event_read()
4642 raw_spin_unlock(&ctx->lock); in __perf_event_read()
4648 return local64_read(&event->count); in perf_event_count()
4650 return local64_read(&event->count) + atomic64_read(&event->child_count); in perf_event_count()
4666 * NMI-safe method to read a local event, that is an event that
4668 * - either for the current task, or for this CPU
4669 * - does not have inherit set, for inherited task events
4671 * - must not have a pmu::count method
4691 if (event->attr.inherit) { in perf_event_read_local()
4692 ret = -EOPNOTSUPP; in perf_event_read_local()
4696 /* If this is a per-task event, it must be for current */ in perf_event_read_local()
4697 if ((event->attach_state & PERF_ATTACH_TASK) && in perf_event_read_local()
4698 event->hw.target != current) { in perf_event_read_local()
4699 ret = -EINVAL; in perf_event_read_local()
4705 * a per-package event that can be read locally in perf_event_read_local()
4707 event_oncpu = __perf_event_read_cpu(event, event->oncpu); in perf_event_read_local()
4708 event_cpu = __perf_event_read_cpu(event, event->cpu); in perf_event_read_local()
4710 /* If this is a per-CPU event, it must be for this CPU */ in perf_event_read_local()
4711 if (!(event->attach_state & PERF_ATTACH_TASK) && in perf_event_read_local()
4713 ret = -EINVAL; in perf_event_read_local()
4718 if (event->attr.pinned && event_oncpu != smp_processor_id()) { in perf_event_read_local()
4719 ret = -EBUSY; in perf_event_read_local()
4724 * If the event is currently on this CPU, its either a per-task event, in perf_event_read_local()
4726 * oncpu == -1). in perf_event_read_local()
4729 event->pmu->read(event); in perf_event_read_local()
4731 *value = local64_read(&event->count); in perf_event_read_local()
4749 enum perf_event_state state = READ_ONCE(event->state); in perf_event_read()
4761 * Orders the ->state and ->oncpu loads such that if we see in perf_event_read()
4762 * ACTIVE we must also see the right ->oncpu. in perf_event_read()
4768 event_cpu = READ_ONCE(event->oncpu); in perf_event_read()
4788 * Therefore, either way, we'll have an up-to-date event count in perf_event_read()
4796 struct perf_event_context *ctx = event->ctx; in perf_event_read()
4799 raw_spin_lock_irqsave(&ctx->lock, flags); in perf_event_read()
4800 state = event->state; in perf_event_read()
4802 raw_spin_unlock_irqrestore(&ctx->lock, flags); in perf_event_read()
4815 raw_spin_unlock_irqrestore(&ctx->lock, flags); in perf_event_read()
4826 raw_spin_lock_init(&ctx->lock); in __perf_event_init_context()
4827 mutex_init(&ctx->mutex); in __perf_event_init_context()
4828 INIT_LIST_HEAD(&ctx->pmu_ctx_list); in __perf_event_init_context()
4829 perf_event_groups_init(&ctx->pinned_groups); in __perf_event_init_context()
4830 perf_event_groups_init(&ctx->flexible_groups); in __perf_event_init_context()
4831 INIT_LIST_HEAD(&ctx->event_list); in __perf_event_init_context()
4832 refcount_set(&ctx->refcount, 1); in __perf_event_init_context()
4838 epc->pmu = pmu; in __perf_init_event_pmu_context()
4839 INIT_LIST_HEAD(&epc->pmu_ctx_entry); in __perf_init_event_pmu_context()
4840 INIT_LIST_HEAD(&epc->pinned_active); in __perf_init_event_pmu_context()
4841 INIT_LIST_HEAD(&epc->flexible_active); in __perf_init_event_pmu_context()
4842 atomic_set(&epc->refcount, 1); in __perf_init_event_pmu_context()
4856 ctx->task = get_task_struct(task); in alloc_perf_context()
4876 return ERR_PTR(-ESRCH); in find_lively_task_by_vpid()
4898 cpuctx = per_cpu_ptr(&perf_cpu_context, event->cpu); in find_get_context()
4899 ctx = &cpuctx->ctx; in find_get_context()
4901 raw_spin_lock_irqsave(&ctx->lock, flags); in find_get_context()
4902 ++ctx->pin_count; in find_get_context()
4903 raw_spin_unlock_irqrestore(&ctx->lock, flags); in find_get_context()
4908 err = -EINVAL; in find_get_context()
4913 ++ctx->pin_count; in find_get_context()
4915 raw_spin_unlock_irqrestore(&ctx->lock, flags); in find_get_context()
4921 err = -ENOMEM; in find_get_context()
4926 mutex_lock(&task->perf_event_mutex); in find_get_context()
4931 if (task->flags & PF_EXITING) in find_get_context()
4932 err = -ESRCH; in find_get_context()
4933 else if (task->perf_event_ctxp) in find_get_context()
4934 err = -EAGAIN; in find_get_context()
4937 ++ctx->pin_count; in find_get_context()
4938 rcu_assign_pointer(task->perf_event_ctxp, ctx); in find_get_context()
4940 mutex_unlock(&task->perf_event_mutex); in find_get_context()
4945 if (err == -EAGAIN) in find_get_context()
4963 if (!ctx->task) { in find_get_pmu_context()
4971 cpc = *per_cpu_ptr(pmu->cpu_pmu_context, event->cpu); in find_get_pmu_context()
4972 epc = &cpc->epc; in find_get_pmu_context()
4973 raw_spin_lock_irq(&ctx->lock); in find_get_pmu_context()
4974 if (!epc->ctx) { in find_get_pmu_context()
4978 atomic_set(&epc->refcount, 2); in find_get_pmu_context()
4979 epc->embedded = 1; in find_get_pmu_context()
4980 list_add(&epc->pmu_ctx_entry, &ctx->pmu_ctx_list); in find_get_pmu_context()
4981 epc->ctx = ctx; in find_get_pmu_context()
4983 WARN_ON_ONCE(epc->ctx != ctx); in find_get_pmu_context()
4984 atomic_inc(&epc->refcount); in find_get_pmu_context()
4986 raw_spin_unlock_irq(&ctx->lock); in find_get_pmu_context()
4992 return ERR_PTR(-ENOMEM); in find_get_pmu_context()
4999 * lockdep_assert_held(&ctx->mutex); in find_get_pmu_context()
5002 * child_ctx->mutex. in find_get_pmu_context()
5005 raw_spin_lock_irq(&ctx->lock); in find_get_pmu_context()
5006 list_for_each_entry(epc, &ctx->pmu_ctx_list, pmu_ctx_entry) { in find_get_pmu_context()
5007 if (epc->pmu == pmu) { in find_get_pmu_context()
5008 WARN_ON_ONCE(epc->ctx != ctx); in find_get_pmu_context()
5009 atomic_inc(&epc->refcount); in find_get_pmu_context()
5013 if (!pos && epc->pmu->type > pmu->type) in find_get_pmu_context()
5021 list_add_tail(&epc->pmu_ctx_entry, &ctx->pmu_ctx_list); in find_get_pmu_context()
5023 list_add(&epc->pmu_ctx_entry, pos->pmu_ctx_entry.prev); in find_get_pmu_context()
5025 epc->ctx = ctx; in find_get_pmu_context()
5028 raw_spin_unlock_irq(&ctx->lock); in find_get_pmu_context()
5036 WARN_ON_ONCE(!atomic_inc_not_zero(&epc->refcount)); in get_pmu_ctx()
5056 struct perf_event_context *ctx = epc->ctx; in put_pmu_ctx()
5062 * lockdep_assert_held(&ctx->mutex); in put_pmu_ctx()
5064 * can't because of the call-site in _free_event()/put_event() in put_pmu_ctx()
5065 * which isn't always called under ctx->mutex. in put_pmu_ctx()
5067 if (!atomic_dec_and_raw_lock_irqsave(&epc->refcount, &ctx->lock, flags)) in put_pmu_ctx()
5070 WARN_ON_ONCE(list_empty(&epc->pmu_ctx_entry)); in put_pmu_ctx()
5072 list_del_init(&epc->pmu_ctx_entry); in put_pmu_ctx()
5073 epc->ctx = NULL; in put_pmu_ctx()
5075 WARN_ON_ONCE(!list_empty(&epc->pinned_active)); in put_pmu_ctx()
5076 WARN_ON_ONCE(!list_empty(&epc->flexible_active)); in put_pmu_ctx()
5078 raw_spin_unlock_irqrestore(&ctx->lock, flags); in put_pmu_ctx()
5080 if (epc->embedded) { in put_pmu_ctx()
5081 call_rcu(&epc->rcu_head, free_cpc_rcu); in put_pmu_ctx()
5085 call_rcu(&epc->rcu_head, free_epc_rcu); in put_pmu_ctx()
5094 if (event->ns) in free_event_rcu()
5095 put_pid_ns(event->ns); in free_event_rcu()
5105 struct pmu_event_list *pel = per_cpu_ptr(&pmu_sb_events, event->cpu); in detach_sb_event()
5107 raw_spin_lock(&pel->lock); in detach_sb_event()
5108 list_del_rcu(&event->sb_list); in detach_sb_event()
5109 raw_spin_unlock(&pel->lock); in detach_sb_event()
5114 struct perf_event_attr *attr = &event->attr; in is_sb_event()
5116 if (event->parent) in is_sb_event()
5119 if (event->attach_state & PERF_ATTACH_TASK) in is_sb_event()
5122 if (attr->mmap || attr->mmap_data || attr->mmap2 || in is_sb_event()
5123 attr->comm || attr->comm_exec || in is_sb_event()
5124 attr->task || attr->ksymbol || in is_sb_event()
5125 attr->context_switch || attr->text_poke || in is_sb_event()
5126 attr->bpf_event) in is_sb_event()
5169 cd->data = kmem_cache_zalloc(ctx_cache, GFP_KERNEL); in alloc_perf_ctx_data()
5170 if (!cd->data) { in alloc_perf_ctx_data()
5175 cd->global = global; in alloc_perf_ctx_data()
5176 cd->ctx_cache = ctx_cache; in alloc_perf_ctx_data()
5177 refcount_set(&cd->refcount, 1); in alloc_perf_ctx_data()
5184 kmem_cache_free(cd->ctx_cache, cd->data); in free_perf_ctx_data()
5198 call_rcu(&cd->rcu_head, __free_perf_ctx_data_rcu); in perf_free_ctx_data_rcu()
5209 return -ENOMEM; in attach_task_ctx_data()
5212 if (try_cmpxchg((struct perf_ctx_data **)&task->perf_ctx_data, &old, cd)) { in attach_task_ctx_data()
5226 if (refcount_inc_not_zero(&old->refcount)) { in attach_task_ctx_data()
5260 cd = rcu_dereference(p->perf_ctx_data); in attach_global_ctx_data()
5261 if (cd && !cd->global) { in attach_global_ctx_data()
5262 cd->global = 1; in attach_global_ctx_data()
5263 if (!refcount_inc_not_zero(&cd->refcount)) in attach_global_ctx_data()
5289 struct task_struct *task = event->hw.target; in attach_perf_ctx_data()
5290 struct kmem_cache *ctx_cache = event->pmu->task_ctx_cache; in attach_perf_ctx_data()
5294 return -ENOMEM; in attach_perf_ctx_data()
5303 event->attach_state |= PERF_ATTACH_GLOBAL_DATA; in attach_perf_ctx_data()
5313 cd = rcu_dereference(p->perf_ctx_data); in detach_task_ctx_data()
5314 if (!cd || !refcount_dec_and_test(&cd->refcount)) in detach_task_ctx_data()
5323 if (try_cmpxchg((struct perf_ctx_data **)&p->perf_ctx_data, &cd, NULL)) in detach_task_ctx_data()
5335 cd = rcu_dereference(p->perf_ctx_data); in __detach_global_ctx_data()
5336 if (!cd || !cd->global) in __detach_global_ctx_data()
5338 cd->global = 0; in __detach_global_ctx_data()
5365 struct task_struct *task = event->hw.target; in detach_perf_ctx_data()
5367 event->attach_state &= ~PERF_ATTACH_TASK_DATA; in detach_perf_ctx_data()
5372 if (event->attach_state & PERF_ATTACH_GLOBAL_DATA) { in detach_perf_ctx_data()
5374 event->attach_state &= ~PERF_ATTACH_GLOBAL_DATA; in detach_perf_ctx_data()
5382 if (event->parent) in unaccount_event()
5385 if (event->attach_state & (PERF_ATTACH_TASK | PERF_ATTACH_SCHED_CB)) in unaccount_event()
5387 if (event->attr.mmap || event->attr.mmap_data) in unaccount_event()
5389 if (event->attr.build_id) in unaccount_event()
5391 if (event->attr.comm) in unaccount_event()
5393 if (event->attr.namespaces) in unaccount_event()
5395 if (event->attr.cgroup) in unaccount_event()
5397 if (event->attr.task) in unaccount_event()
5399 if (event->attr.freq) in unaccount_event()
5401 if (event->attr.context_switch) { in unaccount_event()
5409 if (event->attr.ksymbol) in unaccount_event()
5411 if (event->attr.bpf_event) in unaccount_event()
5413 if (event->attr.text_poke) in unaccount_event()
5417 if (!atomic_add_unless(&perf_sched_count, -1, 1)) in unaccount_event()
5437 * 1) cpu-wide events in the presence of per-task events,
5438 * 2) per-task events in the presence of cpu-wide events,
5442 * _free_event()), the latter -- before the first perf_install_in_context().
5446 struct pmu *pmu = event->pmu; in exclusive_event_init()
5452 * Prevent co-existence of per-task and cpu-wide events on the in exclusive_event_init()
5455 * Negative pmu::exclusive_cnt means there are cpu-wide in exclusive_event_init()
5457 * per-task events. in exclusive_event_init()
5461 * to mean "per-task event", because unlike other attach states it in exclusive_event_init()
5464 if (event->attach_state & PERF_ATTACH_TASK) { in exclusive_event_init()
5465 if (!atomic_inc_unless_negative(&pmu->exclusive_cnt)) in exclusive_event_init()
5466 return -EBUSY; in exclusive_event_init()
5468 if (!atomic_dec_unless_positive(&pmu->exclusive_cnt)) in exclusive_event_init()
5469 return -EBUSY; in exclusive_event_init()
5472 event->attach_state |= PERF_ATTACH_EXCLUSIVE; in exclusive_event_init()
5479 struct pmu *pmu = event->pmu; in exclusive_event_destroy()
5482 if (event->attach_state & PERF_ATTACH_TASK) in exclusive_event_destroy()
5483 atomic_dec(&pmu->exclusive_cnt); in exclusive_event_destroy()
5485 atomic_inc(&pmu->exclusive_cnt); in exclusive_event_destroy()
5487 event->attach_state &= ~PERF_ATTACH_EXCLUSIVE; in exclusive_event_destroy()
5492 if ((e1->pmu == e2->pmu) && in exclusive_event_match()
5493 (e1->cpu == e2->cpu || in exclusive_event_match()
5494 e1->cpu == -1 || in exclusive_event_match()
5495 e2->cpu == -1)) in exclusive_event_match()
5504 struct pmu *pmu = event->pmu; in exclusive_event_installable()
5506 lockdep_assert_held(&ctx->mutex); in exclusive_event_installable()
5511 list_for_each_entry(iter_event, &ctx->event_list, event_entry) { in exclusive_event_installable()
5524 if (event->attach_state & PERF_ATTACH_CALLCHAIN) in __free_event()
5527 kfree(event->addr_filter_ranges); in __free_event()
5529 if (event->attach_state & PERF_ATTACH_EXCLUSIVE) in __free_event()
5535 if (event->attach_state & PERF_ATTACH_TASK_DATA) in __free_event()
5538 if (event->destroy) in __free_event()
5539 event->destroy(event); in __free_event()
5542 * Must be after ->destroy(), due to uprobe_perf_close() using in __free_event()
5545 if (event->hw.target) in __free_event()
5546 put_task_struct(event->hw.target); in __free_event()
5548 if (event->pmu_ctx) { in __free_event()
5550 * put_pmu_ctx() needs an event->ctx reference, because of in __free_event()
5551 * epc->ctx. in __free_event()
5553 WARN_ON_ONCE(!event->ctx); in __free_event()
5554 WARN_ON_ONCE(event->pmu_ctx->ctx != event->ctx); in __free_event()
5555 put_pmu_ctx(event->pmu_ctx); in __free_event()
5562 if (event->ctx) in __free_event()
5563 put_ctx(event->ctx); in __free_event()
5565 if (event->pmu) in __free_event()
5566 module_put(event->pmu->module); in __free_event()
5568 call_rcu(&event->rcu_head, free_event_rcu); in __free_event()
5576 irq_work_sync(&event->pending_irq); in DEFINE_FREE()
5577 irq_work_sync(&event->pending_disable_irq); in DEFINE_FREE()
5583 if (event->rb) { in DEFINE_FREE()
5585 * Can happen when we close an event with re-directed output. in DEFINE_FREE()
5590 mutex_lock(&event->mmap_mutex); in DEFINE_FREE()
5592 mutex_unlock(&event->mmap_mutex); in DEFINE_FREE()
5607 if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1, in free_event()
5609 atomic_long_read(&event->refcount), event)) { in free_event()
5610 /* leak to avoid use-after-free */ in free_event()
5629 * owner->perf_event_mutex. in perf_remove_from_owner()
5631 owner = READ_ONCE(event->owner); in perf_remove_from_owner()
5645 * holding ctx->mutex which would be an inversion wrt. the in perf_remove_from_owner()
5649 * ctx->mutex. in perf_remove_from_owner()
5651 mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING); in perf_remove_from_owner()
5654 * We have to re-check the event->owner field, if it is cleared in perf_remove_from_owner()
5659 if (event->owner) { in perf_remove_from_owner()
5660 list_del_init(&event->owner_entry); in perf_remove_from_owner()
5661 smp_store_release(&event->owner, NULL); in perf_remove_from_owner()
5663 mutex_unlock(&owner->perf_event_mutex); in perf_remove_from_owner()
5672 if (!atomic_long_dec_and_test(&event->refcount)) in put_event()
5675 parent = event->parent; in put_event()
5690 struct perf_event_context *ctx = event->ctx; in perf_event_release_kernel()
5699 WARN_ON_ONCE(event->attach_state & in perf_event_release_kernel()
5708 WARN_ON_ONCE(ctx->parent_ctx); in perf_event_release_kernel()
5714 * Anybody acquiring event->child_mutex after the below loop _must_ in perf_event_release_kernel()
5726 mutex_lock(&event->child_mutex); in perf_event_release_kernel()
5727 list_for_each_entry(child, &event->child_list, child_list) { in perf_event_release_kernel()
5734 ctx = READ_ONCE(child->ctx); in perf_event_release_kernel()
5748 * can re-acquire child_mutex. in perf_event_release_kernel()
5750 mutex_unlock(&event->child_mutex); in perf_event_release_kernel()
5751 mutex_lock(&ctx->mutex); in perf_event_release_kernel()
5752 mutex_lock(&event->child_mutex); in perf_event_release_kernel()
5759 tmp = list_first_entry_or_null(&event->child_list, in perf_event_release_kernel()
5763 list_move(&child->child_list, &free_list); in perf_event_release_kernel()
5765 var = &ctx->refcount; in perf_event_release_kernel()
5768 mutex_unlock(&event->child_mutex); in perf_event_release_kernel()
5769 mutex_unlock(&ctx->mutex); in perf_event_release_kernel()
5783 mutex_unlock(&event->child_mutex); in perf_event_release_kernel()
5786 void *var = &child->ctx->refcount; in perf_event_release_kernel()
5788 list_del(&child->child_list); in perf_event_release_kernel()
5789 /* Last reference unless ->pending_task work is pending */ in perf_event_release_kernel()
5802 * Last reference unless ->pending_task work is pending on this event in perf_event_release_kernel()
5815 perf_event_release_kernel(file->private_data); in perf_release()
5827 mutex_lock(&event->child_mutex); in __perf_event_read_value()
5832 *enabled += event->total_time_enabled + in __perf_event_read_value()
5833 atomic64_read(&event->child_total_time_enabled); in __perf_event_read_value()
5834 *running += event->total_time_running + in __perf_event_read_value()
5835 atomic64_read(&event->child_total_time_running); in __perf_event_read_value()
5837 list_for_each_entry(child, &event->child_list, child_list) { in __perf_event_read_value()
5840 *enabled += child->total_time_enabled; in __perf_event_read_value()
5841 *running += child->total_time_running; in __perf_event_read_value()
5843 mutex_unlock(&event->child_mutex); in __perf_event_read_value()
5861 static int __perf_read_group_add(struct perf_event *leader, in __perf_read_group_add() argument
5864 struct perf_event_context *ctx = leader->ctx; in __perf_read_group_add()
5870 ret = perf_event_read(leader, true); in __perf_read_group_add()
5874 raw_spin_lock_irqsave(&ctx->lock, flags); in __perf_read_group_add()
5880 * - leader->ctx->lock pins leader->sibling_list in __perf_read_group_add()
5881 * - parent->child_mutex pins parent->child_list in __perf_read_group_add()
5882 * - parent->ctx->mutex pins parent->sibling_list in __perf_read_group_add()
5884 * Because parent->ctx != leader->ctx (and child_list nests inside in __perf_read_group_add()
5885 * ctx->mutex), group destruction is not atomic between children, also in __perf_read_group_add()
5895 parent = leader->parent; in __perf_read_group_add()
5897 (parent->group_generation != leader->group_generation || in __perf_read_group_add()
5898 parent->nr_siblings != leader->nr_siblings)) { in __perf_read_group_add()
5899 ret = -ECHILD; in __perf_read_group_add()
5904 * Since we co-schedule groups, {enabled,running} times of siblings in __perf_read_group_add()
5905 * will be identical to those of the leader, so we only publish one in __perf_read_group_add()
5909 values[n++] += leader->total_time_enabled + in __perf_read_group_add()
5910 atomic64_read(&leader->child_total_time_enabled); in __perf_read_group_add()
5914 values[n++] += leader->total_time_running + in __perf_read_group_add()
5915 atomic64_read(&leader->child_total_time_running); in __perf_read_group_add()
5921 values[n++] += perf_event_count(leader, false); in __perf_read_group_add()
5923 values[n++] = primary_event_id(leader); in __perf_read_group_add()
5925 values[n++] = atomic64_read(&leader->lost_samples); in __perf_read_group_add()
5927 for_each_sibling_event(sub, leader) { in __perf_read_group_add()
5932 values[n++] = atomic64_read(&sub->lost_samples); in __perf_read_group_add()
5936 raw_spin_unlock_irqrestore(&ctx->lock, flags); in __perf_read_group_add()
5943 struct perf_event *leader = event->group_leader, *child; in perf_read_group() local
5944 struct perf_event_context *ctx = leader->ctx; in perf_read_group()
5948 lockdep_assert_held(&ctx->mutex); in perf_read_group()
5950 values = kzalloc(event->read_size, GFP_KERNEL); in perf_read_group()
5952 return -ENOMEM; in perf_read_group()
5954 values[0] = 1 + leader->nr_siblings; in perf_read_group()
5956 mutex_lock(&leader->child_mutex); in perf_read_group()
5958 ret = __perf_read_group_add(leader, read_format, values); in perf_read_group()
5962 list_for_each_entry(child, &leader->child_list, child_list) { in perf_read_group()
5968 mutex_unlock(&leader->child_mutex); in perf_read_group()
5970 ret = event->read_size; in perf_read_group()
5971 if (copy_to_user(buf, values, event->read_size)) in perf_read_group()
5972 ret = -EFAULT; in perf_read_group()
5976 mutex_unlock(&leader->child_mutex); in perf_read_group()
5997 values[n++] = atomic64_read(&event->lost_samples); in perf_read_one()
6000 return -EFAULT; in perf_read_one()
6009 if (event->state > PERF_EVENT_STATE_EXIT) in is_event_hup()
6012 mutex_lock(&event->child_mutex); in is_event_hup()
6013 no_children = list_empty(&event->child_list); in is_event_hup()
6014 mutex_unlock(&event->child_mutex); in is_event_hup()
6019 * Read the performance event - simple non blocking version for now
6024 u64 read_format = event->attr.read_format; in __perf_read()
6028 * Return end-of-file for a read on an event that is in in __perf_read()
6032 if (event->state == PERF_EVENT_STATE_ERROR) in __perf_read()
6035 if (count < event->read_size) in __perf_read()
6036 return -ENOSPC; in __perf_read()
6038 WARN_ON_ONCE(event->ctx->parent_ctx); in __perf_read()
6050 struct perf_event *event = file->private_data; in perf_read()
6067 struct perf_event *event = file->private_data; in perf_poll()
6071 poll_wait(file, &event->waitq, wait); in perf_poll()
6076 if (unlikely(READ_ONCE(event->state) == PERF_EVENT_STATE_ERROR && in perf_poll()
6077 event->attr.pinned)) in perf_poll()
6081 * Pin the event->rb by taking event->mmap_mutex; otherwise in perf_poll()
6084 mutex_lock(&event->mmap_mutex); in perf_poll()
6085 rb = event->rb; in perf_poll()
6087 events = atomic_xchg(&rb->poll, 0); in perf_poll()
6088 mutex_unlock(&event->mmap_mutex); in perf_poll()
6095 local64_set(&event->count, 0); in _perf_event_reset()
6106 WARN_ON_ONCE(event->attr.inherit); in perf_event_pause()
6108 count = local64_read(&event->count); in perf_event_pause()
6110 local64_set(&event->count, 0); in perf_event_pause()
6118 * Holding the top-level event's child_mutex means that any
6128 WARN_ON_ONCE(event->ctx->parent_ctx); in perf_event_for_each_child()
6130 mutex_lock(&event->child_mutex); in perf_event_for_each_child()
6132 list_for_each_entry(child, &event->child_list, child_list) in perf_event_for_each_child()
6134 mutex_unlock(&event->child_mutex); in perf_event_for_each_child()
6140 struct perf_event_context *ctx = event->ctx; in perf_event_for_each()
6143 lockdep_assert_held(&ctx->mutex); in perf_event_for_each()
6145 event = event->group_leader; in perf_event_for_each()
6160 if (event->attr.freq) { in __perf_event_period()
6161 event->attr.sample_freq = value; in __perf_event_period()
6163 event->attr.sample_period = value; in __perf_event_period()
6164 event->hw.sample_period = value; in __perf_event_period()
6167 active = (event->state == PERF_EVENT_STATE_ACTIVE); in __perf_event_period()
6169 perf_pmu_disable(event->pmu); in __perf_event_period()
6172 * trying to unthrottle while we already re-started the event. in __perf_event_period()
6174 if (event->hw.interrupts == MAX_INTERRUPTS) { in __perf_event_period()
6175 event->hw.interrupts = 0; in __perf_event_period()
6178 event->pmu->stop(event, PERF_EF_UPDATE); in __perf_event_period()
6181 local64_set(&event->hw.period_left, 0); in __perf_event_period()
6184 event->pmu->start(event, PERF_EF_RELOAD); in __perf_event_period()
6185 perf_pmu_enable(event->pmu); in __perf_event_period()
6191 return event->pmu->check_period(event, value); in perf_event_check_period()
6197 return -EINVAL; in _perf_event_period()
6200 return -EINVAL; in _perf_event_period()
6202 if (event->attr.freq) { in _perf_event_period()
6204 return -EINVAL; in _perf_event_period()
6207 return -EINVAL; in _perf_event_period()
6209 return -EINVAL; in _perf_event_period()
6234 return !fd_empty(f) && fd_file(f)->f_op == &perf_fops; in is_perf_file()
6267 return -EFAULT; in _perf_ioctl()
6276 return -EFAULT; in _perf_ioctl()
6282 CLASS(fd, output)(arg); // arg == -1 => empty in _perf_ioctl()
6284 if (arg != -1) { in _perf_ioctl()
6286 return -EBADF; in _perf_ioctl()
6287 output_event = fd_file(output)->private_data; in _perf_ioctl()
6317 rb = rcu_dereference(event->rb); in _perf_ioctl()
6318 if (!rb || !rb->nr_pages) { in _perf_ioctl()
6320 return -EINVAL; in _perf_ioctl()
6341 return -ENOTTY; in _perf_ioctl()
6354 struct perf_event *event = file->private_data; in perf_ioctl()
6379 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */ in perf_compat_ioctl()
6397 mutex_lock(¤t->perf_event_mutex); in perf_event_task_enable()
6398 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { in perf_event_task_enable()
6403 mutex_unlock(¤t->perf_event_mutex); in perf_event_task_enable()
6413 mutex_lock(¤t->perf_event_mutex); in perf_event_task_disable()
6414 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { in perf_event_task_disable()
6419 mutex_unlock(¤t->perf_event_mutex); in perf_event_task_disable()
6426 if (event->hw.state & PERF_HES_STOPPED) in perf_event_index()
6429 if (event->state != PERF_EVENT_STATE_ACTIVE) in perf_event_index()
6432 return event->pmu->event_idx(event); in perf_event_index()
6441 rb = rcu_dereference(event->rb); in perf_event_init_userpage()
6445 userpg = rb->user_page; in perf_event_init_userpage()
6448 userpg->cap_bit0_is_deprecated = 1; in perf_event_init_userpage()
6449 userpg->size = offsetof(struct perf_event_mmap_page, __reserved); in perf_event_init_userpage()
6450 userpg->data_offset = PAGE_SIZE; in perf_event_init_userpage()
6451 userpg->data_size = perf_data_size(rb); in perf_event_init_userpage()
6474 rb = rcu_dereference(event->rb); in perf_event_update_userpage()
6489 userpg = rb->user_page; in perf_event_update_userpage()
6495 ++userpg->lock; in perf_event_update_userpage()
6497 userpg->index = perf_event_index(event); in perf_event_update_userpage()
6498 userpg->offset = perf_event_count(event, false); in perf_event_update_userpage()
6499 if (userpg->index) in perf_event_update_userpage()
6500 userpg->offset -= local64_read(&event->hw.prev_count); in perf_event_update_userpage()
6502 userpg->time_enabled = enabled + in perf_event_update_userpage()
6503 atomic64_read(&event->child_total_time_enabled); in perf_event_update_userpage()
6505 userpg->time_running = running + in perf_event_update_userpage()
6506 atomic64_read(&event->child_total_time_running); in perf_event_update_userpage()
6511 ++userpg->lock; in perf_event_update_userpage()
6524 WARN_ON_ONCE(event->parent); in ring_buffer_attach()
6526 if (event->rb) { in ring_buffer_attach()
6529 * event->rb_entry and wait/clear when adding event->rb_entry. in ring_buffer_attach()
6531 WARN_ON_ONCE(event->rcu_pending); in ring_buffer_attach()
6533 old_rb = event->rb; in ring_buffer_attach()
6534 spin_lock_irqsave(&old_rb->event_lock, flags); in ring_buffer_attach()
6535 list_del_rcu(&event->rb_entry); in ring_buffer_attach()
6536 spin_unlock_irqrestore(&old_rb->event_lock, flags); in ring_buffer_attach()
6538 event->rcu_batches = get_state_synchronize_rcu(); in ring_buffer_attach()
6539 event->rcu_pending = 1; in ring_buffer_attach()
6543 if (event->rcu_pending) { in ring_buffer_attach()
6544 cond_synchronize_rcu(event->rcu_batches); in ring_buffer_attach()
6545 event->rcu_pending = 0; in ring_buffer_attach()
6548 spin_lock_irqsave(&rb->event_lock, flags); in ring_buffer_attach()
6549 list_add_rcu(&event->rb_entry, &rb->event_list); in ring_buffer_attach()
6550 spin_unlock_irqrestore(&rb->event_lock, flags); in ring_buffer_attach()
6560 * mid-air, but then again, whoever does it like this is in ring_buffer_attach()
6566 rcu_assign_pointer(event->rb, rb); in ring_buffer_attach()
6575 wake_up_all(&event->waitq); in ring_buffer_attach()
6583 if (event->parent) in ring_buffer_wakeup()
6584 event = event->parent; in ring_buffer_wakeup()
6587 rb = rcu_dereference(event->rb); in ring_buffer_wakeup()
6589 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) in ring_buffer_wakeup()
6590 wake_up_all(&event->waitq); in ring_buffer_wakeup()
6599 if (event->parent) in ring_buffer_get()
6600 event = event->parent; in ring_buffer_get()
6603 rb = rcu_dereference(event->rb); in ring_buffer_get()
6605 if (!refcount_inc_not_zero(&rb->refcount)) in ring_buffer_get()
6615 if (!refcount_dec_and_test(&rb->refcount)) in ring_buffer_put()
6618 WARN_ON_ONCE(!list_empty(&rb->event_list)); in ring_buffer_put()
6620 call_rcu(&rb->rcu_head, rb_free_rcu); in ring_buffer_put()
6625 struct perf_event *event = vma->vm_file->private_data; in perf_mmap_open()
6627 atomic_inc(&event->mmap_count); in perf_mmap_open()
6628 atomic_inc(&event->rb->mmap_count); in perf_mmap_open()
6630 if (vma->vm_pgoff) in perf_mmap_open()
6631 atomic_inc(&event->rb->aux_mmap_count); in perf_mmap_open()
6633 if (event->pmu->event_mapped) in perf_mmap_open()
6634 event->pmu->event_mapped(event, vma->vm_mm); in perf_mmap_open()
6649 struct perf_event *event = vma->vm_file->private_data; in perf_mmap_close()
6651 struct user_struct *mmap_user = rb->mmap_user; in perf_mmap_close()
6652 int mmap_locked = rb->mmap_locked; in perf_mmap_close()
6656 if (event->pmu->event_unmapped) in perf_mmap_close()
6657 event->pmu->event_unmapped(event, vma->vm_mm); in perf_mmap_close()
6660 * The AUX buffer is strictly a sub-buffer, serialize using aux_mutex in perf_mmap_close()
6663 if (rb_has_aux(rb) && vma->vm_pgoff == rb->aux_pgoff && in perf_mmap_close()
6664 atomic_dec_and_mutex_lock(&rb->aux_mmap_count, &rb->aux_mutex)) { in perf_mmap_close()
6674 atomic_long_sub(rb->aux_nr_pages - rb->aux_mmap_locked, &mmap_user->locked_vm); in perf_mmap_close()
6675 atomic64_sub(rb->aux_mmap_locked, &vma->vm_mm->pinned_vm); in perf_mmap_close()
6679 WARN_ON_ONCE(refcount_read(&rb->aux_refcount)); in perf_mmap_close()
6681 mutex_unlock(&rb->aux_mutex); in perf_mmap_close()
6684 if (atomic_dec_and_test(&rb->mmap_count)) in perf_mmap_close()
6687 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) in perf_mmap_close()
6691 mutex_unlock(&event->mmap_mutex); in perf_mmap_close()
6704 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { in perf_mmap_close()
6705 if (!atomic_long_inc_not_zero(&event->refcount)) { in perf_mmap_close()
6707 * This event is en-route to free_event() which will in perf_mmap_close()
6714 mutex_lock(&event->mmap_mutex); in perf_mmap_close()
6725 if (event->rb == rb) in perf_mmap_close()
6728 mutex_unlock(&event->mmap_mutex); in perf_mmap_close()
6740 * It could be there's still a few 0-ref events on the list; they'll in perf_mmap_close()
6741 * get cleaned up by free_event() -- they'll also still have their in perf_mmap_close()
6748 atomic_long_sub((size >> PAGE_SHIFT) + 1 - mmap_locked, in perf_mmap_close()
6749 &mmap_user->locked_vm); in perf_mmap_close()
6750 atomic64_sub(mmap_locked, &vma->vm_mm->pinned_vm); in perf_mmap_close()
6759 /* The first page is the user control page, others are read-only. */ in perf_mmap_pfn_mkwrite()
6760 return vmf->pgoff == 0 ? 0 : VM_FAULT_SIGBUS; in perf_mmap_pfn_mkwrite()
6779 * referencing memory-mapped I/O ranges or non-system RAM i.e. for which in map_range()
6782 * We are mapping kernel-allocated memory (memory we manage ourselves) in map_range()
6788 * 1. It uses vma->vm_page_prot, but this field has not been completely in map_range()
6789 * setup at the point of the f_op->mmp() hook, so we are unable to in map_range()
6796 * vm_ops->page_mkwrite() will be invoked rather than in map_range()
6797 * vm_ops->pfn_mkwrite(), and this means we have to set page->mapping in map_range()
6814 unsigned long va = vma->vm_start + PAGE_SIZE * pagenum; in map_range()
6815 struct page *page = perf_mmap_to_page(rb, vma->vm_pgoff + pagenum); in map_range()
6818 err = -EINVAL; in map_range()
6824 vm_get_page_prot(vma->vm_flags & ~VM_SHARED)); in map_range()
6832 zap_page_range_single(vma, vma->vm_start, nr_pages * PAGE_SIZE, NULL); in map_range()
6840 struct perf_event *event = file->private_data; in perf_mmap()
6852 * Don't allow mmap() of inherited per-task counters. This would in perf_mmap()
6856 if (event->cpu == -1 && event->attr.inherit) in perf_mmap()
6857 return -EINVAL; in perf_mmap()
6859 if (!(vma->vm_flags & VM_SHARED)) in perf_mmap()
6860 return -EINVAL; in perf_mmap()
6866 vma_size = vma->vm_end - vma->vm_start; in perf_mmap()
6870 return -ENOMEM; in perf_mmap()
6873 return -EINVAL; in perf_mmap()
6877 mutex_lock(&event->mmap_mutex); in perf_mmap()
6878 ret = -EINVAL; in perf_mmap()
6880 if (vma->vm_pgoff == 0) { in perf_mmap()
6881 nr_pages -= 1; in perf_mmap()
6884 * If we have rb pages ensure they're a power-of-two number, so we in perf_mmap()
6890 WARN_ON_ONCE(event->ctx->parent_ctx); in perf_mmap()
6892 if (event->rb) { in perf_mmap()
6893 if (data_page_nr(event->rb) != nr_pages) in perf_mmap()
6896 if (atomic_inc_not_zero(&event->rb->mmap_count)) { in perf_mmap()
6898 * Success -- managed to mmap() the same buffer in perf_mmap()
6903 rb = event->rb; in perf_mmap()
6910 * event and continue as if !event->rb in perf_mmap()
6917 * AUX area mapping: if rb->aux_nr_pages != 0, it's already in perf_mmap()
6923 rb = event->rb; in perf_mmap()
6927 aux_mutex = &rb->aux_mutex; in perf_mmap()
6930 aux_offset = READ_ONCE(rb->user_page->aux_offset); in perf_mmap()
6931 aux_size = READ_ONCE(rb->user_page->aux_size); in perf_mmap()
6936 if (aux_offset != vma->vm_pgoff << PAGE_SHIFT) in perf_mmap()
6940 if (rb_has_aux(rb) && rb->aux_pgoff != vma->vm_pgoff) in perf_mmap()
6947 if (rb_has_aux(rb) && rb->aux_nr_pages != nr_pages) in perf_mmap()
6953 if (!atomic_inc_not_zero(&rb->mmap_count)) in perf_mmap()
6957 atomic_inc(&rb->aux_mmap_count); in perf_mmap()
6962 atomic_set(&rb->aux_mmap_count, 1); in perf_mmap()
6965 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); in perf_mmap()
6972 user_locked = atomic_long_read(&user->locked_vm); in perf_mmap()
6976 * user->locked_vm > user_lock_limit in perf_mmap()
6987 extra = user_locked - user_lock_limit; in perf_mmap()
6988 user_extra -= extra; in perf_mmap()
6993 locked = atomic64_read(&vma->vm_mm->pinned_vm) + extra; in perf_mmap()
6997 ret = -EPERM; in perf_mmap()
7001 WARN_ON(!rb && event->rb); in perf_mmap()
7003 if (vma->vm_flags & VM_WRITE) in perf_mmap()
7008 event->attr.watermark ? event->attr.wakeup_watermark : 0, in perf_mmap()
7009 event->cpu, flags); in perf_mmap()
7012 ret = -ENOMEM; in perf_mmap()
7016 atomic_set(&rb->mmap_count, 1); in perf_mmap()
7017 rb->mmap_user = get_current_user(); in perf_mmap()
7018 rb->mmap_locked = extra; in perf_mmap()
7026 ret = rb_alloc_aux(rb, event, vma->vm_pgoff, nr_pages, in perf_mmap()
7027 event->attr.aux_watermark, flags); in perf_mmap()
7029 rb->aux_mmap_locked = extra; in perf_mmap()
7036 atomic_long_add(user_extra, &user->locked_vm); in perf_mmap()
7037 atomic64_add(extra, &vma->vm_mm->pinned_vm); in perf_mmap()
7039 atomic_inc(&event->mmap_count); in perf_mmap()
7041 atomic_dec(&rb->mmap_count); in perf_mmap()
7046 mutex_unlock(&event->mmap_mutex); in perf_mmap()
7053 vma->vm_ops = &perf_mmap_vmops; in perf_mmap()
7058 if (!ret && event->pmu->event_mapped) in perf_mmap()
7059 event->pmu->event_mapped(event, vma->vm_mm); in perf_mmap()
7067 struct perf_event *event = filp->private_data; in perf_fasync()
7071 retval = fasync_helper(fd, filp, on, &event->fasync); in perf_fasync()
7094 * to user-space before waking everybody up.
7101 if (event->pending_kill) { in perf_event_wakeup()
7102 kill_fasync(perf_event_fasync(event), SIGIO, event->pending_kill); in perf_event_wakeup()
7103 event->pending_kill = 0; in perf_event_wakeup()
7111 * ctx->task or current has changed in the meantime. This can be the in perf_sigtrap()
7114 if (WARN_ON_ONCE(event->ctx->task != current)) in perf_sigtrap()
7121 if (current->flags & PF_EXITING) in perf_sigtrap()
7124 send_sig_perf((void __user *)event->pending_addr, in perf_sigtrap()
7125 event->orig_type, event->attr.sig_data); in perf_sigtrap()
7129 * Deliver the pending work in-event-context or follow the context.
7133 int cpu = READ_ONCE(event->oncpu); in __perf_pending_disable()
7146 if (event->pending_disable) { in __perf_pending_disable()
7147 event->pending_disable = 0; in __perf_pending_disable()
7154 * CPU-A CPU-B in __perf_pending_disable()
7157 * @pending_disable = CPU-A; in __perf_pending_disable()
7160 * sched-out in __perf_pending_disable()
7161 * @pending_disable = -1; in __perf_pending_disable()
7163 * sched-in in __perf_pending_disable()
7165 * @pending_disable = CPU-B; in __perf_pending_disable()
7171 * But the event runs on CPU-B and wants disabling there. in __perf_pending_disable()
7173 irq_work_queue_on(&event->pending_disable_irq, cpu); in __perf_pending_disable()
7203 * The wakeup isn't bound to the context of the event -- it can happen in perf_pending_irq()
7206 if (event->pending_wakeup) { in perf_pending_irq()
7207 event->pending_wakeup = 0; in perf_pending_irq()
7226 if (event->pending_work) { in perf_pending_task()
7227 event->pending_work = 0; in perf_pending_task()
7229 local_dec(&event->ctx->nr_no_switch_fast); in perf_pending_task()
7240 DEFINE_STATIC_CALL_RET0(__perf_guest_state, *perf_guest_cbs->state);
7241 DEFINE_STATIC_CALL_RET0(__perf_guest_get_ip, *perf_guest_cbs->get_ip);
7242 DEFINE_STATIC_CALL_RET0(__perf_guest_handle_intel_pt_intr, *perf_guest_cbs->handle_intel_pt_intr);
7250 static_call_update(__perf_guest_state, cbs->state); in perf_register_guest_info_callbacks()
7251 static_call_update(__perf_guest_get_ip, cbs->get_ip); in perf_register_guest_info_callbacks()
7253 /* Implementing ->handle_intel_pt_intr is optional. */ in perf_register_guest_info_callbacks()
7254 if (cbs->handle_intel_pt_intr) in perf_register_guest_info_callbacks()
7256 cbs->handle_intel_pt_intr); in perf_register_guest_info_callbacks()
7277 return !event->attr.exclude_guest && perf_guest_state(); in should_sample_guest()
7318 regs_user->abi = perf_reg_abi(current); in perf_sample_regs_user()
7319 regs_user->regs = regs; in perf_sample_regs_user()
7320 } else if (!(current->flags & PF_KTHREAD)) { in perf_sample_regs_user()
7323 regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE; in perf_sample_regs_user()
7324 regs_user->regs = NULL; in perf_sample_regs_user()
7331 regs_intr->regs = regs; in perf_sample_regs_intr()
7332 regs_intr->abi = perf_reg_abi(current); in perf_sample_regs_intr()
7350 return TASK_SIZE - addr; in perf_ustack_task_size()
7365 * - TASK_SIZE in perf_sample_ustack_size()
7368 * - remaining sample size in perf_sample_ustack_size()
7385 stack_size = USHRT_MAX - header_size - sizeof(u64); in perf_sample_ustack_size()
7408 * - the size requested by user or the best one we can fit in perf_output_sample_ustack()
7411 * - user stack dump data in perf_output_sample_ustack()
7413 * - the actual dumped size in perf_output_sample_ustack()
7422 dyn_size = dump_size - rem; in perf_output_sample_ustack()
7435 struct perf_event *sampler = event->aux_event; in perf_prepare_sample_aux()
7438 data->aux_size = 0; in perf_prepare_sample_aux()
7443 if (WARN_ON_ONCE(READ_ONCE(sampler->state) != PERF_EVENT_STATE_ACTIVE)) in perf_prepare_sample_aux()
7446 if (WARN_ON_ONCE(READ_ONCE(sampler->oncpu) != smp_processor_id())) in perf_prepare_sample_aux()
7457 if (READ_ONCE(rb->aux_in_sampling)) { in perf_prepare_sample_aux()
7458 data->aux_size = 0; in perf_prepare_sample_aux()
7461 data->aux_size = ALIGN(size, sizeof(u64)); in perf_prepare_sample_aux()
7466 return data->aux_size; in perf_prepare_sample_aux()
7478 * Normal ->start()/->stop() callbacks run in IRQ mode in scheduler in perf_pmu_snapshot_aux()
7480 * the IRQ ones, that is, for example, re-starting an event that's just in perf_pmu_snapshot_aux()
7491 WRITE_ONCE(rb->aux_in_sampling, 1); in perf_pmu_snapshot_aux()
7494 ret = event->pmu->snapshot_aux(event, handle, size); in perf_pmu_snapshot_aux()
7497 WRITE_ONCE(rb->aux_in_sampling, 0); in perf_pmu_snapshot_aux()
7507 struct perf_event *sampler = event->aux_event; in perf_aux_sample_output()
7512 if (WARN_ON_ONCE(!sampler || !data->aux_size)) in perf_aux_sample_output()
7519 size = perf_pmu_snapshot_aux(rb, sampler, handle, data->aux_size); in perf_aux_sample_output()
7523 * non-zero surplus that it didn't copy), which in its current in perf_aux_sample_output()
7531 * The pad comes from ALIGN()ing data->aux_size up to u64 in in perf_aux_sample_output()
7534 pad = data->aux_size - size; in perf_aux_sample_output()
7548 * A set of common sample data types saved even for non-sample records
7549 * when event->attr.sample_id_all is set.
7559 data->type = event->attr.sample_type; in __perf_event_header__init_id()
7560 data->sample_flags |= data->type & PERF_SAMPLE_ID_ALL; in __perf_event_header__init_id()
7564 data->tid_entry.pid = perf_event_pid(event, current); in __perf_event_header__init_id()
7565 data->tid_entry.tid = perf_event_tid(event, current); in __perf_event_header__init_id()
7569 data->time = perf_event_clock(event); in __perf_event_header__init_id()
7572 data->id = primary_event_id(event); in __perf_event_header__init_id()
7575 data->stream_id = event->id; in __perf_event_header__init_id()
7578 data->cpu_entry.cpu = raw_smp_processor_id(); in __perf_event_header__init_id()
7579 data->cpu_entry.reserved = 0; in __perf_event_header__init_id()
7587 if (event->attr.sample_id_all) { in perf_event_header__init_id()
7588 header->size += event->id_header_size; in perf_event_header__init_id()
7589 __perf_event_header__init_id(data, event, event->attr.sample_type); in perf_event_header__init_id()
7596 u64 sample_type = data->type; in __perf_event__output_id_sample()
7599 perf_output_put(handle, data->tid_entry); in __perf_event__output_id_sample()
7602 perf_output_put(handle, data->time); in __perf_event__output_id_sample()
7605 perf_output_put(handle, data->id); in __perf_event__output_id_sample()
7608 perf_output_put(handle, data->stream_id); in __perf_event__output_id_sample()
7611 perf_output_put(handle, data->cpu_entry); in __perf_event__output_id_sample()
7614 perf_output_put(handle, data->id); in __perf_event__output_id_sample()
7621 if (event->attr.sample_id_all) in perf_event__output_id_sample()
7629 u64 read_format = event->attr.read_format; in perf_output_read_one()
7633 values[n++] = perf_event_count(event, has_inherit_and_sample_read(&event->attr)); in perf_output_read_one()
7636 atomic64_read(&event->child_total_time_enabled); in perf_output_read_one()
7640 atomic64_read(&event->child_total_time_running); in perf_output_read_one()
7645 values[n++] = atomic64_read(&event->lost_samples); in perf_output_read_one()
7654 struct perf_event *leader = event->group_leader, *sub; in perf_output_read_group() local
7655 u64 read_format = event->attr.read_format; in perf_output_read_group()
7659 bool self = has_inherit_and_sample_read(&event->attr); in perf_output_read_group()
7667 values[n++] = 1 + leader->nr_siblings; in perf_output_read_group()
7675 if ((leader != event) && !handle->skip_read) in perf_output_read_group()
7676 perf_pmu_read(leader); in perf_output_read_group()
7678 values[n++] = perf_event_count(leader, self); in perf_output_read_group()
7680 values[n++] = primary_event_id(leader); in perf_output_read_group()
7682 values[n++] = atomic64_read(&leader->lost_samples); in perf_output_read_group()
7686 for_each_sibling_event(sub, leader) { in perf_output_read_group()
7689 if ((sub != event) && !handle->skip_read) in perf_output_read_group()
7696 values[n++] = atomic64_read(&sub->lost_samples); in perf_output_read_group()
7714 * Instead the combination of PERF_SAMPLE_READ and inherit will track per-thread
7722 u64 read_format = event->attr.read_format; in perf_output_read()
7736 if (event->attr.read_format & PERF_FORMAT_GROUP) in perf_output_read()
7747 u64 sample_type = data->type; in perf_output_sample()
7749 if (data->sample_flags & PERF_SAMPLE_READ) in perf_output_sample()
7750 handle->skip_read = 1; in perf_output_sample()
7755 perf_output_put(handle, data->id); in perf_output_sample()
7758 perf_output_put(handle, data->ip); in perf_output_sample()
7761 perf_output_put(handle, data->tid_entry); in perf_output_sample()
7764 perf_output_put(handle, data->time); in perf_output_sample()
7767 perf_output_put(handle, data->addr); in perf_output_sample()
7770 perf_output_put(handle, data->id); in perf_output_sample()
7773 perf_output_put(handle, data->stream_id); in perf_output_sample()
7776 perf_output_put(handle, data->cpu_entry); in perf_output_sample()
7779 perf_output_put(handle, data->period); in perf_output_sample()
7787 size += data->callchain->nr; in perf_output_sample()
7789 __output_copy(handle, data->callchain, size); in perf_output_sample()
7793 struct perf_raw_record *raw = data->raw; in perf_output_sample()
7796 struct perf_raw_frag *frag = &raw->frag; in perf_output_sample()
7798 perf_output_put(handle, raw->size); in perf_output_sample()
7800 if (frag->copy) { in perf_output_sample()
7801 __output_custom(handle, frag->copy, in perf_output_sample()
7802 frag->data, frag->size); in perf_output_sample()
7804 __output_copy(handle, frag->data, in perf_output_sample()
7805 frag->size); in perf_output_sample()
7809 frag = frag->next; in perf_output_sample()
7811 if (frag->pad) in perf_output_sample()
7812 __output_skip(handle, NULL, frag->pad); in perf_output_sample()
7826 if (data->br_stack) { in perf_output_sample()
7829 size = data->br_stack->nr in perf_output_sample()
7832 perf_output_put(handle, data->br_stack->nr); in perf_output_sample()
7834 perf_output_put(handle, data->br_stack->hw_idx); in perf_output_sample()
7835 perf_output_copy(handle, data->br_stack->entries, size); in perf_output_sample()
7840 if (data->br_stack_cntr) { in perf_output_sample()
7841 size = data->br_stack->nr * sizeof(u64); in perf_output_sample()
7842 perf_output_copy(handle, data->br_stack_cntr, size); in perf_output_sample()
7854 u64 abi = data->regs_user.abi; in perf_output_sample()
7863 u64 mask = event->attr.sample_regs_user; in perf_output_sample()
7865 data->regs_user.regs, in perf_output_sample()
7872 data->stack_user_size, in perf_output_sample()
7873 data->regs_user.regs); in perf_output_sample()
7877 perf_output_put(handle, data->weight.full); in perf_output_sample()
7880 perf_output_put(handle, data->data_src.val); in perf_output_sample()
7883 perf_output_put(handle, data->txn); in perf_output_sample()
7886 u64 abi = data->regs_intr.abi; in perf_output_sample()
7894 u64 mask = event->attr.sample_regs_intr; in perf_output_sample()
7897 data->regs_intr.regs, in perf_output_sample()
7903 perf_output_put(handle, data->phys_addr); in perf_output_sample()
7906 perf_output_put(handle, data->cgroup); in perf_output_sample()
7909 perf_output_put(handle, data->data_page_size); in perf_output_sample()
7912 perf_output_put(handle, data->code_page_size); in perf_output_sample()
7915 perf_output_put(handle, data->aux_size); in perf_output_sample()
7917 if (data->aux_size) in perf_output_sample()
7921 if (!event->attr.watermark) { in perf_output_sample()
7922 int wakeup_events = event->attr.wakeup_events; in perf_output_sample()
7925 struct perf_buffer *rb = handle->rb; in perf_output_sample()
7926 int events = local_inc_return(&rb->events); in perf_output_sample()
7929 local_sub(wakeup_events, &rb->events); in perf_output_sample()
7930 local_inc(&rb->wakeup); in perf_output_sample()
7953 * Try IRQ-safe get_user_page_fast_only first. in perf_virt_to_phys()
7956 if (current->mm != NULL) { in perf_virt_to_phys()
8041 * Software page-table walkers must disable IRQs, in perf_get_page_size()
8046 mm = current->mm; in perf_get_page_size()
8067 bool kernel = !event->attr.exclude_callchain_kernel; in perf_callchain()
8068 bool user = !event->attr.exclude_callchain_user; in perf_callchain()
8069 /* Disallow cross-task user callchains. */ in perf_callchain()
8070 bool crosstask = event->ctx->task && event->ctx->task != current; in perf_callchain()
8071 const u32 max_stack = event->attr.sample_max_stack; in perf_callchain()
8091 u64 sample_type = event->attr.sample_type; in perf_prepare_sample()
8105 filtered_sample_type &= ~data->sample_flags; in perf_prepare_sample()
8108 /* Make sure it has the correct data->type for output */ in perf_prepare_sample()
8109 data->type = event->attr.sample_type; in perf_prepare_sample()
8116 data->ip = perf_instruction_pointer(event, regs); in perf_prepare_sample()
8117 data->sample_flags |= PERF_SAMPLE_IP; in perf_prepare_sample()
8124 data->raw = NULL; in perf_prepare_sample()
8125 data->dyn_size += sizeof(u64); in perf_prepare_sample()
8126 data->sample_flags |= PERF_SAMPLE_RAW; in perf_prepare_sample()
8130 data->br_stack = NULL; in perf_prepare_sample()
8131 data->dyn_size += sizeof(u64); in perf_prepare_sample()
8132 data->sample_flags |= PERF_SAMPLE_BRANCH_STACK; in perf_prepare_sample()
8136 perf_sample_regs_user(&data->regs_user, regs); in perf_prepare_sample()
8143 if ((sample_type & ~data->sample_flags) & PERF_SAMPLE_REGS_USER) { in perf_prepare_sample()
8147 if (data->regs_user.regs) { in perf_prepare_sample()
8148 u64 mask = event->attr.sample_regs_user; in perf_prepare_sample()
8152 data->dyn_size += size; in perf_prepare_sample()
8153 data->sample_flags |= PERF_SAMPLE_REGS_USER; in perf_prepare_sample()
8163 u16 stack_size = event->attr.sample_stack_user; in perf_prepare_sample()
8168 data->regs_user.regs); in perf_prepare_sample()
8178 data->stack_user_size = stack_size; in perf_prepare_sample()
8179 data->dyn_size += size; in perf_prepare_sample()
8180 data->sample_flags |= PERF_SAMPLE_STACK_USER; in perf_prepare_sample()
8184 data->weight.full = 0; in perf_prepare_sample()
8185 data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE; in perf_prepare_sample()
8189 data->data_src.val = PERF_MEM_NA; in perf_prepare_sample()
8190 data->sample_flags |= PERF_SAMPLE_DATA_SRC; in perf_prepare_sample()
8194 data->txn = 0; in perf_prepare_sample()
8195 data->sample_flags |= PERF_SAMPLE_TRANSACTION; in perf_prepare_sample()
8199 data->addr = 0; in perf_prepare_sample()
8200 data->sample_flags |= PERF_SAMPLE_ADDR; in perf_prepare_sample()
8207 perf_sample_regs_intr(&data->regs_intr, regs); in perf_prepare_sample()
8209 if (data->regs_intr.regs) { in perf_prepare_sample()
8210 u64 mask = event->attr.sample_regs_intr; in perf_prepare_sample()
8215 data->dyn_size += size; in perf_prepare_sample()
8216 data->sample_flags |= PERF_SAMPLE_REGS_INTR; in perf_prepare_sample()
8220 data->phys_addr = perf_virt_to_phys(data->addr); in perf_prepare_sample()
8221 data->sample_flags |= PERF_SAMPLE_PHYS_ADDR; in perf_prepare_sample()
8229 cgrp = task_css_check(current, perf_event_cgrp_id, 1)->cgroup; in perf_prepare_sample()
8230 data->cgroup = cgroup_id(cgrp); in perf_prepare_sample()
8231 data->sample_flags |= PERF_SAMPLE_CGROUP; in perf_prepare_sample()
8237 * require PERF_SAMPLE_ADDR, kernel implicitly retrieve the data->addr, in perf_prepare_sample()
8241 data->data_page_size = perf_get_page_size(data->addr); in perf_prepare_sample()
8242 data->sample_flags |= PERF_SAMPLE_DATA_PAGE_SIZE; in perf_prepare_sample()
8246 data->code_page_size = perf_get_page_size(data->ip); in perf_prepare_sample()
8247 data->sample_flags |= PERF_SAMPLE_CODE_PAGE_SIZE; in perf_prepare_sample()
8262 size = min_t(size_t, U16_MAX - header_size, in perf_prepare_sample()
8263 event->attr.aux_sample_size); in perf_prepare_sample()
8268 data->dyn_size += size + sizeof(u64); /* size above */ in perf_prepare_sample()
8269 data->sample_flags |= PERF_SAMPLE_AUX; in perf_prepare_sample()
8278 header->type = PERF_RECORD_SAMPLE; in perf_prepare_header()
8279 header->size = perf_sample_data_size(data, event); in perf_prepare_header()
8280 header->misc = perf_misc_flags(event, regs); in perf_prepare_header()
8290 WARN_ON_ONCE(header->size & 7); in perf_prepare_header()
8296 if (!event->hw.aux_paused) { in __perf_event_aux_pause()
8297 event->hw.aux_paused = 1; in __perf_event_aux_pause()
8298 event->pmu->stop(event, PERF_EF_PAUSE); in __perf_event_aux_pause()
8301 if (event->hw.aux_paused) { in __perf_event_aux_pause()
8302 event->hw.aux_paused = 0; in __perf_event_aux_pause()
8303 event->pmu->start(event, PERF_EF_RESUME); in __perf_event_aux_pause()
8321 * Guard against self-recursion here. Another event could trip in perf_event_aux_pause()
8324 if (READ_ONCE(rb->aux_in_pause_resume)) in perf_event_aux_pause()
8327 WRITE_ONCE(rb->aux_in_pause_resume, 1); in perf_event_aux_pause()
8331 WRITE_ONCE(rb->aux_in_pause_resume, 0); in perf_event_aux_pause()
8413 .size = sizeof(read_event) + event->read_size, in perf_event_read_event()
8441 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { in perf_iterate_ctx()
8443 if (event->state < PERF_EVENT_STATE_INACTIVE) in perf_iterate_ctx()
8458 list_for_each_entry_rcu(event, &pel->list, sb_list) { in perf_iterate_sb_cpu()
8461 * if we observe event->ctx, both event and ctx will be in perf_iterate_sb_cpu()
8464 if (!smp_load_acquire(&event->ctx)) in perf_iterate_sb_cpu()
8467 if (event->state < PERF_EVENT_STATE_INACTIVE) in perf_iterate_sb_cpu()
8476 * Iterate all events that need to receive side-band events.
8502 ctx = rcu_dereference(current->perf_event_ctxp); in perf_iterate_sb()
8511 * Clear all file-based filters at exec, they'll have to be
8512 * re-instated when/if these objects are mmapped again.
8524 raw_spin_lock_irqsave(&ifh->lock, flags); in perf_event_addr_filters_exec()
8525 list_for_each_entry(filter, &ifh->list, entry) { in perf_event_addr_filters_exec()
8526 if (filter->path.dentry) { in perf_event_addr_filters_exec()
8527 event->addr_filter_ranges[count].start = 0; in perf_event_addr_filters_exec()
8528 event->addr_filter_ranges[count].size = 0; in perf_event_addr_filters_exec()
8536 event->addr_filters_gen++; in perf_event_addr_filters_exec()
8537 raw_spin_unlock_irqrestore(&ifh->lock, flags); in perf_event_addr_filters_exec()
8567 struct perf_event *parent = event->parent; in __perf_event_output_stop()
8569 struct perf_buffer *rb = ro->rb; in __perf_event_output_stop()
8582 * ring-buffer, but it will be the child that's actually using it. in __perf_event_output_stop()
8587 * So ring_buffer_attach() has to stop an aux event before re-assigning in __perf_event_output_stop()
8590 if (rcu_dereference(parent->rb) == rb) in __perf_event_output_stop()
8591 ro->err = __perf_event_stop(&sd); in __perf_event_output_stop()
8599 .rb = event->rb, in __perf_pmu_output_stop()
8603 perf_iterate_ctx(&cpuctx->ctx, __perf_event_output_stop, &ro, false); in __perf_pmu_output_stop()
8604 if (cpuctx->task_ctx) in __perf_pmu_output_stop()
8605 perf_iterate_ctx(cpuctx->task_ctx, __perf_event_output_stop, in __perf_pmu_output_stop()
8619 list_for_each_entry_rcu(iter, &event->rb->event_list, rb_entry) { in perf_pmu_output_stop()
8621 * For per-CPU events, we need to make sure that neither they in perf_pmu_output_stop()
8622 * nor their children are running; for cpu==-1 events it's in perf_pmu_output_stop()
8626 cpu = iter->cpu; in perf_pmu_output_stop()
8627 if (cpu == -1) in perf_pmu_output_stop()
8628 cpu = READ_ONCE(iter->oncpu); in perf_pmu_output_stop()
8630 if (cpu == -1) in perf_pmu_output_stop()
8634 if (err == -EAGAIN) { in perf_pmu_output_stop()
8643 * task tracking -- fork/exit
8665 return event->attr.comm || event->attr.mmap || in perf_event_task_match()
8666 event->attr.mmap2 || event->attr.mmap_data || in perf_event_task_match()
8667 event->attr.task; in perf_event_task_match()
8676 struct task_struct *task = task_event->task; in perf_event_task_output()
8677 int ret, size = task_event->event_id.header.size; in perf_event_task_output()
8682 perf_event_header__init_id(&task_event->event_id.header, &sample, event); in perf_event_task_output()
8685 task_event->event_id.header.size); in perf_event_task_output()
8689 task_event->event_id.pid = perf_event_pid(event, task); in perf_event_task_output()
8690 task_event->event_id.tid = perf_event_tid(event, task); in perf_event_task_output()
8692 if (task_event->event_id.header.type == PERF_RECORD_EXIT) { in perf_event_task_output()
8693 task_event->event_id.ppid = perf_event_pid(event, in perf_event_task_output()
8694 task->real_parent); in perf_event_task_output()
8695 task_event->event_id.ptid = perf_event_pid(event, in perf_event_task_output()
8696 task->real_parent); in perf_event_task_output()
8698 task_event->event_id.ppid = perf_event_pid(event, current); in perf_event_task_output()
8699 task_event->event_id.ptid = perf_event_tid(event, current); in perf_event_task_output()
8702 task_event->event_id.time = perf_event_clock(event); in perf_event_task_output()
8704 perf_output_put(&handle, task_event->event_id); in perf_event_task_output()
8710 task_event->event_id.header.size = size; in perf_event_task_output()
8747 * Allocate data for a new task when profiling system-wide
8761 cd = rcu_dereference(parent->perf_ctx_data); in perf_event_alloc_task_data()
8763 ctx_cache = cd->ctx_cache; in perf_event_alloc_task_data()
8771 cd = rcu_dereference(child->perf_ctx_data); in perf_event_alloc_task_data()
8774 * A system-wide event may be unaccount, in perf_event_alloc_task_data()
8782 if (!cd->global) { in perf_event_alloc_task_data()
8783 cd->global = 1; in perf_event_alloc_task_data()
8784 refcount_inc(&cd->refcount); in perf_event_alloc_task_data()
8819 return event->attr.comm; in perf_event_comm_match()
8828 int size = comm_event->event_id.header.size; in perf_event_comm_output()
8834 perf_event_header__init_id(&comm_event->event_id.header, &sample, event); in perf_event_comm_output()
8836 comm_event->event_id.header.size); in perf_event_comm_output()
8841 comm_event->event_id.pid = perf_event_pid(event, comm_event->task); in perf_event_comm_output()
8842 comm_event->event_id.tid = perf_event_tid(event, comm_event->task); in perf_event_comm_output()
8844 perf_output_put(&handle, comm_event->event_id); in perf_event_comm_output()
8845 __output_copy(&handle, comm_event->comm, in perf_event_comm_output()
8846 comm_event->comm_size); in perf_event_comm_output()
8852 comm_event->event_id.header.size = size; in perf_event_comm_output()
8861 strscpy(comm, comm_event->task->comm); in perf_event_comm_event()
8864 comm_event->comm = comm; in perf_event_comm_event()
8865 comm_event->comm_size = size; in perf_event_comm_event()
8867 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; in perf_event_comm_event()
8918 return event->attr.namespaces; in perf_event_namespaces_match()
8927 u16 header_size = namespaces_event->event_id.header.size; in perf_event_namespaces_output()
8933 perf_event_header__init_id(&namespaces_event->event_id.header, in perf_event_namespaces_output()
8936 namespaces_event->event_id.header.size); in perf_event_namespaces_output()
8940 namespaces_event->event_id.pid = perf_event_pid(event, in perf_event_namespaces_output()
8941 namespaces_event->task); in perf_event_namespaces_output()
8942 namespaces_event->event_id.tid = perf_event_tid(event, in perf_event_namespaces_output()
8943 namespaces_event->task); in perf_event_namespaces_output()
8945 perf_output_put(&handle, namespaces_event->event_id); in perf_event_namespaces_output()
8951 namespaces_event->event_id.header.size = header_size; in perf_event_namespaces_output()
8964 ns_inode = ns_path.dentry->d_inode; in perf_fill_ns_link_info()
8965 ns_link_info->dev = new_encode_dev(ns_inode->i_sb->s_dev); in perf_fill_ns_link_info()
8966 ns_link_info->ino = ns_inode->i_ino; in perf_fill_ns_link_info()
9046 return event->attr.cgroup; in perf_event_cgroup_match()
9054 u16 header_size = cgroup_event->event_id.header.size; in perf_event_cgroup_output()
9060 perf_event_header__init_id(&cgroup_event->event_id.header, in perf_event_cgroup_output()
9063 cgroup_event->event_id.header.size); in perf_event_cgroup_output()
9067 perf_output_put(&handle, cgroup_event->event_id); in perf_event_cgroup_output()
9068 __output_copy(&handle, cgroup_event->path, cgroup_event->path_size); in perf_event_cgroup_output()
9074 cgroup_event->event_id.header.size = header_size; in perf_event_cgroup_output()
9103 cgroup_path(cgrp, pathname, PATH_MAX - sizeof(u64)); in perf_event_cgroup()
9159 struct vm_area_struct *vma = mmap_event->vma; in perf_event_mmap_match()
9160 int executable = vma->vm_flags & VM_EXEC; in perf_event_mmap_match()
9162 return (!executable && event->attr.mmap_data) || in perf_event_mmap_match()
9163 (executable && (event->attr.mmap || event->attr.mmap2)); in perf_event_mmap_match()
9172 int size = mmap_event->event_id.header.size; in perf_event_mmap_output()
9173 u32 type = mmap_event->event_id.header.type; in perf_event_mmap_output()
9180 if (event->attr.mmap2) { in perf_event_mmap_output()
9181 mmap_event->event_id.header.type = PERF_RECORD_MMAP2; in perf_event_mmap_output()
9182 mmap_event->event_id.header.size += sizeof(mmap_event->maj); in perf_event_mmap_output()
9183 mmap_event->event_id.header.size += sizeof(mmap_event->min); in perf_event_mmap_output()
9184 mmap_event->event_id.header.size += sizeof(mmap_event->ino); in perf_event_mmap_output()
9185 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation); in perf_event_mmap_output()
9186 mmap_event->event_id.header.size += sizeof(mmap_event->prot); in perf_event_mmap_output()
9187 mmap_event->event_id.header.size += sizeof(mmap_event->flags); in perf_event_mmap_output()
9190 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); in perf_event_mmap_output()
9192 mmap_event->event_id.header.size); in perf_event_mmap_output()
9196 mmap_event->event_id.pid = perf_event_pid(event, current); in perf_event_mmap_output()
9197 mmap_event->event_id.tid = perf_event_tid(event, current); in perf_event_mmap_output()
9199 use_build_id = event->attr.build_id && mmap_event->build_id_size; in perf_event_mmap_output()
9201 if (event->attr.mmap2 && use_build_id) in perf_event_mmap_output()
9202 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_BUILD_ID; in perf_event_mmap_output()
9204 perf_output_put(&handle, mmap_event->event_id); in perf_event_mmap_output()
9206 if (event->attr.mmap2) { in perf_event_mmap_output()
9208 u8 size[4] = { (u8) mmap_event->build_id_size, 0, 0, 0 }; in perf_event_mmap_output()
9211 __output_copy(&handle, mmap_event->build_id, BUILD_ID_SIZE_MAX); in perf_event_mmap_output()
9213 perf_output_put(&handle, mmap_event->maj); in perf_event_mmap_output()
9214 perf_output_put(&handle, mmap_event->min); in perf_event_mmap_output()
9215 perf_output_put(&handle, mmap_event->ino); in perf_event_mmap_output()
9216 perf_output_put(&handle, mmap_event->ino_generation); in perf_event_mmap_output()
9218 perf_output_put(&handle, mmap_event->prot); in perf_event_mmap_output()
9219 perf_output_put(&handle, mmap_event->flags); in perf_event_mmap_output()
9222 __output_copy(&handle, mmap_event->file_name, in perf_event_mmap_output()
9223 mmap_event->file_size); in perf_event_mmap_output()
9229 mmap_event->event_id.header.size = size; in perf_event_mmap_output()
9230 mmap_event->event_id.header.type = type; in perf_event_mmap_output()
9235 struct vm_area_struct *vma = mmap_event->vma; in perf_event_mmap_event()
9236 struct file *file = vma->vm_file; in perf_event_mmap_event()
9245 if (vma->vm_flags & VM_READ) in perf_event_mmap_event()
9247 if (vma->vm_flags & VM_WRITE) in perf_event_mmap_event()
9249 if (vma->vm_flags & VM_EXEC) in perf_event_mmap_event()
9252 if (vma->vm_flags & VM_MAYSHARE) in perf_event_mmap_event()
9257 if (vma->vm_flags & VM_LOCKED) in perf_event_mmap_event()
9276 name = file_path(file, buf, PATH_MAX - sizeof(u64)); in perf_event_mmap_event()
9281 inode = file_inode(vma->vm_file); in perf_event_mmap_event()
9282 dev = inode->i_sb->s_dev; in perf_event_mmap_event()
9283 ino = inode->i_ino; in perf_event_mmap_event()
9284 gen = inode->i_generation; in perf_event_mmap_event()
9290 if (vma->vm_ops && vma->vm_ops->name) in perf_event_mmap_event()
9291 name = (char *) vma->vm_ops->name(vma); in perf_event_mmap_event()
9317 mmap_event->file_name = name; in perf_event_mmap_event()
9318 mmap_event->file_size = size; in perf_event_mmap_event()
9319 mmap_event->maj = maj; in perf_event_mmap_event()
9320 mmap_event->min = min; in perf_event_mmap_event()
9321 mmap_event->ino = ino; in perf_event_mmap_event()
9322 mmap_event->ino_generation = gen; in perf_event_mmap_event()
9323 mmap_event->prot = prot; in perf_event_mmap_event()
9324 mmap_event->flags = flags; in perf_event_mmap_event()
9326 if (!(vma->vm_flags & VM_EXEC)) in perf_event_mmap_event()
9327 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; in perf_event_mmap_event()
9329 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; in perf_event_mmap_event()
9332 build_id_parse_nofault(vma, mmap_event->build_id, &mmap_event->build_id_size); in perf_event_mmap_event()
9348 /* d_inode(NULL) won't be equal to any mapped user-space file */ in perf_addr_filter_match()
9349 if (!filter->path.dentry) in perf_addr_filter_match()
9352 if (d_inode(filter->path.dentry) != file_inode(file)) in perf_addr_filter_match()
9355 if (filter->offset > offset + size) in perf_addr_filter_match()
9358 if (filter->offset + filter->size < offset) in perf_addr_filter_match()
9368 unsigned long vma_size = vma->vm_end - vma->vm_start; in perf_addr_filter_vma_adjust()
9369 unsigned long off = vma->vm_pgoff << PAGE_SHIFT; in perf_addr_filter_vma_adjust()
9370 struct file *file = vma->vm_file; in perf_addr_filter_vma_adjust()
9375 if (filter->offset < off) { in perf_addr_filter_vma_adjust()
9376 fr->start = vma->vm_start; in perf_addr_filter_vma_adjust()
9377 fr->size = min(vma_size, filter->size - (off - filter->offset)); in perf_addr_filter_vma_adjust()
9379 fr->start = vma->vm_start + filter->offset - off; in perf_addr_filter_vma_adjust()
9380 fr->size = min(vma->vm_end - fr->start, filter->size); in perf_addr_filter_vma_adjust()
9397 if (!vma->vm_file) in __perf_addr_filters_adjust()
9400 raw_spin_lock_irqsave(&ifh->lock, flags); in __perf_addr_filters_adjust()
9401 list_for_each_entry(filter, &ifh->list, entry) { in __perf_addr_filters_adjust()
9403 &event->addr_filter_ranges[count])) in __perf_addr_filters_adjust()
9410 event->addr_filters_gen++; in __perf_addr_filters_adjust()
9411 raw_spin_unlock_irqrestore(&ifh->lock, flags); in __perf_addr_filters_adjust()
9428 if (!(vma->vm_flags & VM_EXEC)) in perf_addr_filters_adjust()
9432 ctx = rcu_dereference(current->perf_event_ctxp); in perf_addr_filters_adjust()
9457 .start = vma->vm_start, in perf_event_mmap()
9458 .len = vma->vm_end - vma->vm_start, in perf_event_mmap()
9459 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, in perf_event_mmap()
9557 return event->attr.context_switch; in perf_event_switch_match()
9570 /* Only CPU-wide events are allowed to see next/prev pid/tid */ in perf_event_switch_output()
9571 if (event->ctx->task) { in perf_event_switch_output()
9572 se->event_id.header.type = PERF_RECORD_SWITCH; in perf_event_switch_output()
9573 se->event_id.header.size = sizeof(se->event_id.header); in perf_event_switch_output()
9575 se->event_id.header.type = PERF_RECORD_SWITCH_CPU_WIDE; in perf_event_switch_output()
9576 se->event_id.header.size = sizeof(se->event_id); in perf_event_switch_output()
9577 se->event_id.next_prev_pid = in perf_event_switch_output()
9578 perf_event_pid(event, se->next_prev); in perf_event_switch_output()
9579 se->event_id.next_prev_tid = in perf_event_switch_output()
9580 perf_event_tid(event, se->next_prev); in perf_event_switch_output()
9583 perf_event_header__init_id(&se->event_id.header, &sample, event); in perf_event_switch_output()
9585 ret = perf_output_begin(&handle, &sample, event, se->event_id.header.size); in perf_event_switch_output()
9589 if (event->ctx->task) in perf_event_switch_output()
9590 perf_output_put(&handle, se->event_id.header); in perf_event_switch_output()
9592 perf_output_put(&handle, se->event_id); in perf_event_switch_output()
9651 .stream_id = event->id, in perf_log_throttle()
9687 return event->attr.ksymbol; in perf_event_ksymbol_match()
9700 perf_event_header__init_id(&ksymbol_event->event_id.header, in perf_event_ksymbol_output()
9703 ksymbol_event->event_id.header.size); in perf_event_ksymbol_output()
9707 perf_output_put(&handle, ksymbol_event->event_id); in perf_event_ksymbol_output()
9708 __output_copy(&handle, ksymbol_event->name, ksymbol_event->name_len); in perf_event_ksymbol_output()
9777 return event->attr.bpf_event; in perf_event_bpf_match()
9790 perf_event_header__init_id(&bpf_event->event_id.header, in perf_event_bpf_output()
9793 bpf_event->event_id.header.size); in perf_event_bpf_output()
9797 perf_output_put(&handle, bpf_event->event_id); in perf_event_bpf_output()
9810 (u64)(unsigned long)prog->bpf_func, in perf_event_bpf_emit_ksymbols()
9811 prog->jited_len, unregister, in perf_event_bpf_emit_ksymbols()
9812 prog->aux->ksym.name); in perf_event_bpf_emit_ksymbols()
9814 for (i = 1; i < prog->aux->func_cnt; i++) { in perf_event_bpf_emit_ksymbols()
9815 struct bpf_prog *subprog = prog->aux->func[i]; in perf_event_bpf_emit_ksymbols()
9819 (u64)(unsigned long)subprog->bpf_func, in perf_event_bpf_emit_ksymbols()
9820 subprog->jited_len, unregister, in perf_event_bpf_emit_ksymbols()
9821 subprog->aux->ksym.name); in perf_event_bpf_emit_ksymbols()
9853 .id = prog->aux->id, in perf_event_bpf_event()
9859 memcpy(bpf_event.event_id.tag, prog->tag, BPF_TAG_SIZE); in perf_event_bpf_event()
9879 return event->attr.text_poke; in perf_event_text_poke_match()
9893 perf_event_header__init_id(&text_poke_event->event_id.header, &sample, event); in perf_event_text_poke_output()
9896 text_poke_event->event_id.header.size); in perf_event_text_poke_output()
9900 perf_output_put(&handle, text_poke_event->event_id); in perf_event_text_poke_output()
9901 perf_output_put(&handle, text_poke_event->old_len); in perf_event_text_poke_output()
9902 perf_output_put(&handle, text_poke_event->new_len); in perf_event_text_poke_output()
9904 __output_copy(&handle, text_poke_event->old_bytes, text_poke_event->old_len); in perf_event_text_poke_output()
9905 __output_copy(&handle, text_poke_event->new_bytes, text_poke_event->new_len); in perf_event_text_poke_output()
9907 if (text_poke_event->pad) in perf_event_text_poke_output()
9908 __output_copy(&handle, &padding, text_poke_event->pad); in perf_event_text_poke_output()
9926 pad = ALIGN(tot, sizeof(u64)) - tot; in perf_event_text_poke()
9949 event->attach_state |= PERF_ATTACH_ITRACE; in perf_event_itrace_started()
9963 if (event->parent) in perf_log_itrace_start()
9964 event = event->parent; in perf_log_itrace_start()
9966 if (!(event->pmu->capabilities & PERF_PMU_CAP_ITRACE) || in perf_log_itrace_start()
9967 event->attach_state & PERF_ATTACH_ITRACE) in perf_log_itrace_start()
9998 if (event->parent) in perf_report_aux_output_id()
9999 event = event->parent; in perf_report_aux_output_id()
10022 struct hw_perf_event *hwc = &event->hw; in __perf_event_account_interrupt()
10027 if (seq != hwc->interrupts_seq) { in __perf_event_account_interrupt()
10028 hwc->interrupts_seq = seq; in __perf_event_account_interrupt()
10029 hwc->interrupts = 1; in __perf_event_account_interrupt()
10031 hwc->interrupts++; in __perf_event_account_interrupt()
10033 hwc->interrupts > max_samples_per_tick)) { in __perf_event_account_interrupt()
10036 hwc->interrupts = MAX_INTERRUPTS; in __perf_event_account_interrupt()
10042 if (event->attr.freq) { in __perf_event_account_interrupt()
10044 s64 delta = now - hwc->freq_time_stamp; in __perf_event_account_interrupt()
10046 hwc->freq_time_stamp = now; in __perf_event_account_interrupt()
10049 perf_adjust_period(event, delta, hwc->last_period, true); in __perf_event_account_interrupt()
10067 if (event->attr.exclude_kernel && !user_mode(regs)) in sample_is_allowed()
10089 prog = READ_ONCE(event->prog); in bpf_overflow_handler()
10105 if (event->overflow_handler_context) in perf_event_set_bpf_handler()
10107 return -EINVAL; in perf_event_set_bpf_handler()
10109 if (event->prog) in perf_event_set_bpf_handler()
10110 return -EEXIST; in perf_event_set_bpf_handler()
10112 if (prog->type != BPF_PROG_TYPE_PERF_EVENT) in perf_event_set_bpf_handler()
10113 return -EINVAL; in perf_event_set_bpf_handler()
10115 if (event->attr.precise_ip && in perf_event_set_bpf_handler()
10116 prog->call_get_stack && in perf_event_set_bpf_handler()
10117 (!(event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) || in perf_event_set_bpf_handler()
10118 event->attr.exclude_callchain_kernel || in perf_event_set_bpf_handler()
10119 event->attr.exclude_callchain_user)) { in perf_event_set_bpf_handler()
10129 return -EPROTO; in perf_event_set_bpf_handler()
10132 event->prog = prog; in perf_event_set_bpf_handler()
10133 event->bpf_cookie = bpf_cookie; in perf_event_set_bpf_handler()
10139 struct bpf_prog *prog = event->prog; in perf_event_free_bpf_handler()
10144 event->prog = NULL; in perf_event_free_bpf_handler()
10159 return -EOPNOTSUPP; in perf_event_set_bpf_handler()
10175 int events = atomic_read(&event->event_limit); in __perf_event_overflow()
10179 * Non-sampling counters might still use the PMI to fold short in __perf_event_overflow()
10187 if (event->attr.aux_pause) in __perf_event_overflow()
10188 perf_event_aux_pause(event->aux_event, true); in __perf_event_overflow()
10190 if (event->prog && event->prog->type == BPF_PROG_TYPE_PERF_EVENT && in __perf_event_overflow()
10199 event->pending_kill = POLL_IN; in __perf_event_overflow()
10200 if (events && atomic_dec_and_test(&event->event_limit)) { in __perf_event_overflow()
10202 event->pending_kill = POLL_HUP; in __perf_event_overflow()
10206 if (event->attr.sigtrap) { in __perf_event_overflow()
10222 if (!event->pending_work && in __perf_event_overflow()
10223 !task_work_add(current, &event->pending_task, notify_mode)) { in __perf_event_overflow()
10224 event->pending_work = pending_id; in __perf_event_overflow()
10225 local_inc(&event->ctx->nr_no_switch_fast); in __perf_event_overflow()
10226 WARN_ON_ONCE(!atomic_long_inc_not_zero(&event->refcount)); in __perf_event_overflow()
10228 event->pending_addr = 0; in __perf_event_overflow()
10229 if (valid_sample && (data->sample_flags & PERF_SAMPLE_ADDR)) in __perf_event_overflow()
10230 event->pending_addr = data->addr; in __perf_event_overflow()
10232 } else if (event->attr.exclude_kernel && valid_sample) { in __perf_event_overflow()
10240 * 2. Events that can overflow again before the IRQ- in __perf_event_overflow()
10243 * check 32-bit hash of the current IP. in __perf_event_overflow()
10245 WARN_ON_ONCE(event->pending_work != pending_id); in __perf_event_overflow()
10249 READ_ONCE(event->overflow_handler)(event, data, regs); in __perf_event_overflow()
10251 if (*perf_event_fasync(event) && event->pending_kill) { in __perf_event_overflow()
10252 event->pending_wakeup = 1; in __perf_event_overflow()
10253 irq_work_queue(&event->pending_irq); in __perf_event_overflow()
10256 if (event->attr.aux_resume) in __perf_event_overflow()
10257 perf_event_aux_pause(event->aux_event, false); in __perf_event_overflow()
10281 * We directly increment event->count and keep a second value in
10282 * event->hw.period_left to count intervals. This period event
10283 * is kept in the range [-sample_period, 0] so that we can use the
10289 struct hw_perf_event *hwc = &event->hw; in perf_swevent_set_period()
10290 u64 period = hwc->last_period; in perf_swevent_set_period()
10294 hwc->last_period = hwc->sample_period; in perf_swevent_set_period()
10296 old = local64_read(&hwc->period_left); in perf_swevent_set_period()
10304 val -= offset; in perf_swevent_set_period()
10305 } while (!local64_try_cmpxchg(&hwc->period_left, &old, val)); in perf_swevent_set_period()
10314 struct hw_perf_event *hwc = &event->hw; in perf_swevent_overflow()
10320 if (hwc->interrupts == MAX_INTERRUPTS) in perf_swevent_overflow()
10323 for (; overflow; overflow--) { in perf_swevent_overflow()
10328 * hwc->interrupts == MAX_INTERRUPTS. in perf_swevent_overflow()
10340 struct hw_perf_event *hwc = &event->hw; in perf_swevent_event()
10342 local64_add(nr, &event->count); in perf_swevent_event()
10350 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { in perf_swevent_event()
10351 data->period = nr; in perf_swevent_event()
10354 data->period = event->hw.last_period; in perf_swevent_event()
10356 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) in perf_swevent_event()
10359 if (local64_add_negative(nr, &hwc->period_left)) in perf_swevent_event()
10367 if (event->hw.state & PERF_HES_STOPPED) in perf_exclude_event()
10371 if (event->attr.exclude_user && user_mode(regs)) in perf_exclude_event()
10374 if (event->attr.exclude_kernel && !user_mode(regs)) in perf_exclude_event()
10387 if (event->attr.type != type) in perf_swevent_match()
10390 if (event->attr.config != event_id) in perf_swevent_match()
10411 return &hlist->heads[hash]; in __find_swevent_head()
10420 hlist = rcu_dereference(swhash->swevent_hlist); in find_swevent_head_rcu()
10432 u32 event_id = event->attr.config; in find_swevent_head()
10433 u64 type = event->attr.type; in find_swevent_head()
10440 hlist = rcu_dereference_protected(swhash->swevent_hlist, in find_swevent_head()
10441 lockdep_is_held(&event->ctx->lock)); in find_swevent_head()
10474 return get_recursion_context(current->perf_recursion); in perf_swevent_get_recursion_context()
10480 put_recursion_context(current->perf_recursion, rctx); in perf_swevent_put_recursion_context()
10517 struct hw_perf_event *hwc = &event->hw; in perf_swevent_add()
10521 hwc->last_period = hwc->sample_period; in perf_swevent_add()
10525 hwc->state = !(flags & PERF_EF_START); in perf_swevent_add()
10529 return -EINVAL; in perf_swevent_add()
10531 hlist_add_head_rcu(&event->hlist_entry, head); in perf_swevent_add()
10539 hlist_del_rcu(&event->hlist_entry); in perf_swevent_del()
10544 event->hw.state = 0; in perf_swevent_start()
10549 event->hw.state = PERF_HES_STOPPED; in perf_swevent_stop()
10556 return rcu_dereference_protected(swhash->swevent_hlist, in swevent_hlist_deref()
10557 lockdep_is_held(&swhash->hlist_mutex)); in swevent_hlist_deref()
10567 RCU_INIT_POINTER(swhash->swevent_hlist, NULL); in swevent_hlist_release()
10575 mutex_lock(&swhash->hlist_mutex); in swevent_hlist_put_cpu()
10577 if (!--swhash->hlist_refcount) in swevent_hlist_put_cpu()
10580 mutex_unlock(&swhash->hlist_mutex); in swevent_hlist_put_cpu()
10596 mutex_lock(&swhash->hlist_mutex); in swevent_hlist_get_cpu()
10603 err = -ENOMEM; in swevent_hlist_get_cpu()
10606 rcu_assign_pointer(swhash->swevent_hlist, hlist); in swevent_hlist_get_cpu()
10608 swhash->hlist_refcount++; in swevent_hlist_get_cpu()
10610 mutex_unlock(&swhash->hlist_mutex); in swevent_hlist_get_cpu()
10643 u64 event_id = event->attr.config; in sw_perf_event_destroy()
10645 WARN_ON(event->parent); in sw_perf_event_destroy()
10656 u64 event_id = event->attr.config; in perf_swevent_init()
10658 if (event->attr.type != PERF_TYPE_SOFTWARE) in perf_swevent_init()
10659 return -ENOENT; in perf_swevent_init()
10665 return -EOPNOTSUPP; in perf_swevent_init()
10669 event->attr.type = perf_cpu_clock.type; in perf_swevent_init()
10670 return -ENOENT; in perf_swevent_init()
10672 event->attr.type = perf_task_clock.type; in perf_swevent_init()
10673 return -ENOENT; in perf_swevent_init()
10680 return -ENOENT; in perf_swevent_init()
10682 if (!event->parent) { in perf_swevent_init()
10690 event->destroy = sw_perf_event_destroy; in perf_swevent_init()
10720 if (event->attr.type != PERF_TYPE_TRACEPOINT) in perf_tp_event_init()
10721 return -ENOENT; in perf_tp_event_init()
10727 return -EOPNOTSUPP; in perf_tp_event_init()
10733 event->destroy = tp_perf_event_destroy; in perf_tp_event_init()
10752 void *record = raw->frag.data; in perf_tp_filter_match()
10755 if (event->parent) in perf_tp_filter_match()
10756 event = event->parent; in perf_tp_filter_match()
10758 if (likely(!event->filter) || filter_match_preds(event->filter, record)) in perf_tp_filter_match()
10767 if (event->hw.state & PERF_HES_STOPPED) in perf_tp_event_match()
10770 * If exclude_kernel, only trace user-space tracepoints (uprobes) in perf_tp_event_match()
10772 if (event->attr.exclude_kernel && !user_mode(regs)) in perf_tp_event_match()
10793 perf_tp_event(call->event.type, count, raw_data, size, regs, head, in perf_trace_run_bpf_submit()
10806 if (event->attr.config != entry->type) in __perf_tp_event_target_task()
10809 if (event->attr.sigtrap) in __perf_tp_event_target_task()
10828 perf_event_groups_for_cpu_pmu(event, &ctx->pinned_groups, cpu, pmu) { in perf_tp_event_target_task()
10834 perf_event_groups_for_cpu_pmu(event, &ctx->flexible_groups, cpu, pmu) { in perf_tp_event_target_task()
10860 * Here use the same on-stack perf_sample_data, in perf_tp_event()
10861 * some members in data are event-specific and in perf_tp_event()
10862 * need to be re-computed for different sweveents. in perf_tp_event()
10863 * Re-initialize data->sample_flags safely to avoid in perf_tp_event()
10865 * because data->sample_flags is set. in perf_tp_event()
10881 ctx = rcu_dereference(task->perf_event_ctxp); in perf_tp_event()
10885 raw_spin_lock(&ctx->lock); in perf_tp_event()
10887 raw_spin_unlock(&ctx->lock); in perf_tp_event()
10914 PERF_UPROBE_REF_CTR_OFFSET_SHIFT = 64 - PERF_UPROBE_REF_CTR_OFFSET_BITS,
10953 if (event->attr.type != perf_kprobe.type) in perf_kprobe_event_init()
10954 return -ENOENT; in perf_kprobe_event_init()
10957 return -EACCES; in perf_kprobe_event_init()
10963 return -EOPNOTSUPP; in perf_kprobe_event_init()
10965 is_retprobe = event->attr.config & PERF_PROBE_CONFIG_IS_RETPROBE; in perf_kprobe_event_init()
10970 event->destroy = perf_kprobe_destroy; in perf_kprobe_event_init()
10977 PMU_FORMAT_ATTR(ref_ctr_offset, "config:32-63");
11013 if (event->attr.type != perf_uprobe.type) in perf_uprobe_event_init()
11014 return -ENOENT; in perf_uprobe_event_init()
11017 return -EACCES; in perf_uprobe_event_init()
11023 return -EOPNOTSUPP; in perf_uprobe_event_init()
11025 is_retprobe = event->attr.config & PERF_PROBE_CONFIG_IS_RETPROBE; in perf_uprobe_event_init()
11026 ref_ctr_offset = event->attr.config >> PERF_UPROBE_REF_CTR_OFFSET_SHIFT; in perf_uprobe_event_init()
11031 event->destroy = perf_uprobe_destroy; in perf_uprobe_event_init()
11041 perf_pmu_register(&perf_kprobe, "kprobe", -1); in perf_tp_register()
11044 perf_pmu_register(&perf_uprobe, "uprobe", -1); in perf_tp_register()
11059 if (event->pmu == &perf_tracepoint) in perf_event_is_tracing()
11062 if (event->pmu == &perf_kprobe) in perf_event_is_tracing()
11066 if (event->pmu == &perf_uprobe) in perf_event_is_tracing()
11080 is_kprobe = event->tp_event->flags & TRACE_EVENT_FL_KPROBE; in perf_event_set_bpf_prog()
11081 is_uprobe = event->tp_event->flags & TRACE_EVENT_FL_UPROBE; in perf_event_set_bpf_prog()
11082 is_tracepoint = event->tp_event->flags & TRACE_EVENT_FL_TRACEPOINT; in perf_event_set_bpf_prog()
11083 is_syscall_tp = is_syscall_trace_event(event->tp_event); in perf_event_set_bpf_prog()
11086 return -EINVAL; in perf_event_set_bpf_prog()
11088 if (((is_kprobe || is_uprobe) && prog->type != BPF_PROG_TYPE_KPROBE) || in perf_event_set_bpf_prog()
11089 (is_tracepoint && prog->type != BPF_PROG_TYPE_TRACEPOINT) || in perf_event_set_bpf_prog()
11090 (is_syscall_tp && prog->type != BPF_PROG_TYPE_TRACEPOINT)) in perf_event_set_bpf_prog()
11091 return -EINVAL; in perf_event_set_bpf_prog()
11093 if (prog->type == BPF_PROG_TYPE_KPROBE && prog->sleepable && !is_uprobe) in perf_event_set_bpf_prog()
11095 return -EINVAL; in perf_event_set_bpf_prog()
11098 if (prog->kprobe_override && !is_kprobe) in perf_event_set_bpf_prog()
11099 return -EINVAL; in perf_event_set_bpf_prog()
11102 int off = trace_event_get_offsets(event->tp_event); in perf_event_set_bpf_prog()
11104 if (prog->aux->max_ctx_offset > off) in perf_event_set_bpf_prog()
11105 return -EACCES; in perf_event_set_bpf_prog()
11113 if (!event->prog) in perf_event_free_bpf_prog()
11136 return -ENOENT; in perf_event_set_bpf_prog()
11150 perf_sample_data_init(&sample, bp->attr.bp_addr, 0); in perf_bp_event()
11152 if (!bp->hw.state && !perf_exclude_event(bp, regs)) in perf_bp_event()
11163 int node = cpu_to_node(event->cpu == -1 ? 0 : event->cpu); in perf_addr_filter_new()
11170 INIT_LIST_HEAD(&filter->entry); in perf_addr_filter_new()
11171 list_add_tail(&filter->entry, filters); in perf_addr_filter_new()
11181 path_put(&filter->path); in free_filters_list()
11182 list_del(&filter->entry); in free_filters_list()
11200 if (event->parent) in perf_addr_filters_splice()
11203 raw_spin_lock_irqsave(&event->addr_filters.lock, flags); in perf_addr_filters_splice()
11205 list_splice_init(&event->addr_filters.list, &list); in perf_addr_filters_splice()
11207 list_splice(head, &event->addr_filters.list); in perf_addr_filters_splice()
11209 raw_spin_unlock_irqrestore(&event->addr_filters.lock, flags); in perf_addr_filters_splice()
11219 if (list_empty(&event->addr_filters.list)) in perf_free_addr_filters()
11238 if (!vma->vm_file) in perf_addr_filter_apply()
11253 struct task_struct *task = READ_ONCE(event->ctx->task); in perf_event_addr_filters_apply()
11260 * We may observe TASK_TOMBSTONE, which means that the event tear-down in perf_event_addr_filters_apply()
11266 if (ifh->nr_file_filters) { in perf_event_addr_filters_apply()
11274 raw_spin_lock_irqsave(&ifh->lock, flags); in perf_event_addr_filters_apply()
11275 list_for_each_entry(filter, &ifh->list, entry) { in perf_event_addr_filters_apply()
11276 if (filter->path.dentry) { in perf_event_addr_filters_apply()
11281 event->addr_filter_ranges[count].start = 0; in perf_event_addr_filters_apply()
11282 event->addr_filter_ranges[count].size = 0; in perf_event_addr_filters_apply()
11284 perf_addr_filter_apply(filter, mm, &event->addr_filter_ranges[count]); in perf_event_addr_filters_apply()
11286 event->addr_filter_ranges[count].start = filter->offset; in perf_event_addr_filters_apply()
11287 event->addr_filter_ranges[count].size = filter->size; in perf_event_addr_filters_apply()
11293 event->addr_filters_gen++; in perf_event_addr_filters_apply()
11294 raw_spin_unlock_irqrestore(&ifh->lock, flags); in perf_event_addr_filters_apply()
11296 if (ifh->nr_file_filters) { in perf_event_addr_filters_apply()
11326 IF_ACT_NONE = -1,
11365 int ret = -EINVAL; in perf_event_parse_addr_filter()
11369 return -ENOMEM; in perf_event_parse_addr_filter()
11377 ret = -EINVAL; in perf_event_parse_addr_filter()
11397 filter->action = actions[token]; in perf_event_parse_addr_filter()
11412 ret = kstrtoul(args[0].from, 0, &filter->offset); in perf_event_parse_addr_filter()
11418 ret = kstrtoul(args[1].from, 0, &filter->size); in perf_event_parse_addr_filter()
11429 ret = -ENOMEM; in perf_event_parse_addr_filter()
11447 ret = -EINVAL; in perf_event_parse_addr_filter()
11450 * ACTION "filter" must have a non-zero length region in perf_event_parse_addr_filter()
11453 if (filter->action == PERF_ADDR_FILTER_ACTION_FILTER && in perf_event_parse_addr_filter()
11454 !filter->size) in perf_event_parse_addr_filter()
11462 * For now, we only support file-based filters in perf_event_parse_addr_filter()
11463 * in per-task events; doing so for CPU-wide in perf_event_parse_addr_filter()
11469 ret = -EOPNOTSUPP; in perf_event_parse_addr_filter()
11470 if (!event->ctx->task) in perf_event_parse_addr_filter()
11475 &filter->path); in perf_event_parse_addr_filter()
11479 ret = -EINVAL; in perf_event_parse_addr_filter()
11480 if (!filter->path.dentry || in perf_event_parse_addr_filter()
11481 !S_ISREG(d_inode(filter->path.dentry) in perf_event_parse_addr_filter()
11482 ->i_mode)) in perf_event_parse_addr_filter()
11485 event->addr_filters.nr_file_filters++; in perf_event_parse_addr_filter()
11523 lockdep_assert_held(&event->ctx->mutex); in perf_event_set_addr_filter()
11525 if (WARN_ON_ONCE(event->parent)) in perf_event_set_addr_filter()
11526 return -EINVAL; in perf_event_set_addr_filter()
11532 ret = event->pmu->addr_filters_validate(&filters); in perf_event_set_addr_filter()
11548 event->addr_filters.nr_file_filters = 0; in perf_event_set_addr_filter()
11555 int ret = -EINVAL; in perf_event_set_filter()
11564 struct perf_event_context *ctx = event->ctx; in perf_event_set_filter()
11569 * the tracepoint muck will deadlock against ctx->mutex, but in perf_event_set_filter()
11571 * temporarily drop ctx->mutex. As per perf_event_ctx_lock() we in perf_event_set_filter()
11577 mutex_unlock(&ctx->mutex); in perf_event_set_filter()
11578 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); in perf_event_set_filter()
11579 mutex_lock(&ctx->mutex); in perf_event_set_filter()
11603 if (event->state != PERF_EVENT_STATE_ACTIVE) in perf_swevent_hrtimer()
11606 event->pmu->read(event); in perf_swevent_hrtimer()
11608 perf_sample_data_init(&data, 0, event->hw.last_period); in perf_swevent_hrtimer()
11612 if (!(event->attr.exclude_idle && is_idle_task(current))) in perf_swevent_hrtimer()
11617 period = max_t(u64, 10000, event->hw.sample_period); in perf_swevent_hrtimer()
11625 struct hw_perf_event *hwc = &event->hw; in perf_swevent_start_hrtimer()
11631 period = local64_read(&hwc->period_left); in perf_swevent_start_hrtimer()
11636 local64_set(&hwc->period_left, 0); in perf_swevent_start_hrtimer()
11638 period = max_t(u64, 10000, hwc->sample_period); in perf_swevent_start_hrtimer()
11640 hrtimer_start(&hwc->hrtimer, ns_to_ktime(period), in perf_swevent_start_hrtimer()
11646 struct hw_perf_event *hwc = &event->hw; in perf_swevent_cancel_hrtimer()
11649 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); in perf_swevent_cancel_hrtimer()
11650 local64_set(&hwc->period_left, ktime_to_ns(remaining)); in perf_swevent_cancel_hrtimer()
11652 hrtimer_cancel(&hwc->hrtimer); in perf_swevent_cancel_hrtimer()
11658 struct hw_perf_event *hwc = &event->hw; in perf_swevent_init_hrtimer()
11663 hrtimer_setup(&hwc->hrtimer, perf_swevent_hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD); in perf_swevent_init_hrtimer()
11666 * Since hrtimers have a fixed rate, we can do a static freq->period in perf_swevent_init_hrtimer()
11669 if (event->attr.freq) { in perf_swevent_init_hrtimer()
11670 long freq = event->attr.sample_freq; in perf_swevent_init_hrtimer()
11672 event->attr.sample_period = NSEC_PER_SEC / freq; in perf_swevent_init_hrtimer()
11673 hwc->sample_period = event->attr.sample_period; in perf_swevent_init_hrtimer()
11674 local64_set(&hwc->period_left, hwc->sample_period); in perf_swevent_init_hrtimer()
11675 hwc->last_period = hwc->sample_period; in perf_swevent_init_hrtimer()
11676 event->attr.freq = 0; in perf_swevent_init_hrtimer()
11690 prev = local64_xchg(&event->hw.prev_count, now); in cpu_clock_event_update()
11691 local64_add(now - prev, &event->count); in cpu_clock_event_update()
11696 local64_set(&event->hw.prev_count, local_clock()); in cpu_clock_event_start()
11727 if (event->attr.type != perf_cpu_clock.type) in cpu_clock_event_init()
11728 return -ENOENT; in cpu_clock_event_init()
11730 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) in cpu_clock_event_init()
11731 return -ENOENT; in cpu_clock_event_init()
11737 return -EOPNOTSUPP; in cpu_clock_event_init()
11767 prev = local64_xchg(&event->hw.prev_count, now); in task_clock_event_update()
11768 delta = now - prev; in task_clock_event_update()
11769 local64_add(delta, &event->count); in task_clock_event_update()
11774 local64_set(&event->hw.prev_count, event->ctx->time); in task_clock_event_start()
11781 task_clock_event_update(event, event->ctx->time); in task_clock_event_stop()
11801 u64 delta = now - event->ctx->timestamp; in task_clock_event_read()
11802 u64 time = event->ctx->time + delta; in task_clock_event_read()
11809 if (event->attr.type != perf_task_clock.type) in task_clock_event_init()
11810 return -ENOENT; in task_clock_event_init()
11812 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) in task_clock_event_init()
11813 return -ENOENT; in task_clock_event_init()
11819 return -EOPNOTSUPP; in task_clock_event_init()
11909 return sysfs_emit(page, "%d\n", pmu->nr_addr_filters); in nr_addr_filters_show()
11920 return sysfs_emit(page, "%d\n", pmu->type); in type_show()
11931 return sysfs_emit(page, "%d\n", pmu->hrtimer_interval_ms); in perf_event_mux_interval_ms_show()
11949 return -EINVAL; in perf_event_mux_interval_ms_store()
11952 if (timer == pmu->hrtimer_interval_ms) in perf_event_mux_interval_ms_store()
11956 pmu->hrtimer_interval_ms = timer; in perf_event_mux_interval_ms_store()
11962 cpc = *per_cpu_ptr(pmu->cpu_pmu_context, cpu); in perf_event_mux_interval_ms_store()
11963 cpc->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); in perf_event_mux_interval_ms_store()
12014 struct cpumask *mask = perf_scope_cpumask(pmu->scope); in cpumask_show()
12036 if (n == 2 && !pmu->nr_addr_filters) in pmu_dev_is_visible()
12040 if (n == 3 && pmu->scope == PERF_PMU_SCOPE_NONE) in pmu_dev_is_visible()
12043 return a->mode; in pmu_dev_is_visible()
12069 int ret = -ENOMEM; in pmu_dev_alloc()
12071 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); in pmu_dev_alloc()
12072 if (!pmu->dev) in pmu_dev_alloc()
12075 pmu->dev->groups = pmu->attr_groups; in pmu_dev_alloc()
12076 device_initialize(pmu->dev); in pmu_dev_alloc()
12078 dev_set_drvdata(pmu->dev, pmu); in pmu_dev_alloc()
12079 pmu->dev->bus = &pmu_bus; in pmu_dev_alloc()
12080 pmu->dev->parent = pmu->parent; in pmu_dev_alloc()
12081 pmu->dev->release = pmu_dev_release; in pmu_dev_alloc()
12083 ret = dev_set_name(pmu->dev, "%s", pmu->name); in pmu_dev_alloc()
12087 ret = device_add(pmu->dev); in pmu_dev_alloc()
12091 if (pmu->attr_update) { in pmu_dev_alloc()
12092 ret = sysfs_update_groups(&pmu->dev->kobj, pmu->attr_update); in pmu_dev_alloc()
12101 device_del(pmu->dev); in pmu_dev_alloc()
12104 put_device(pmu->dev); in pmu_dev_alloc()
12105 pmu->dev = NULL; in pmu_dev_alloc()
12129 if (pmu_bus_running && pmu->dev && pmu->dev != PMU_NULL_DEV) { in perf_pmu_free()
12130 if (pmu->nr_addr_filters) in perf_pmu_free()
12131 device_remove_file(pmu->dev, &dev_attr_nr_addr_filters); in perf_pmu_free()
12132 device_del(pmu->dev); in perf_pmu_free()
12133 put_device(pmu->dev); in perf_pmu_free()
12136 if (pmu->cpu_pmu_context) { in perf_pmu_free()
12142 cpc = *per_cpu_ptr(pmu->cpu_pmu_context, cpu); in perf_pmu_free()
12145 if (cpc->epc.embedded) { in perf_pmu_free()
12147 put_pmu_ctx(&cpc->epc); in perf_pmu_free()
12152 free_percpu(pmu->cpu_pmu_context); in perf_pmu_free()
12166 return -EINVAL; in DEFINE_FREE()
12168 if (WARN_ONCE(pmu->scope >= PERF_PMU_MAX_SCOPE, in DEFINE_FREE()
12170 return -EINVAL; in DEFINE_FREE()
12172 pmu->name = name; in DEFINE_FREE()
12183 pmu->type = pmu_type.id; in DEFINE_FREE()
12184 atomic_set(&pmu->exclusive_cnt, 0); in DEFINE_FREE()
12186 if (pmu_bus_running && !pmu->dev) { in DEFINE_FREE()
12192 pmu->cpu_pmu_context = alloc_percpu(struct perf_cpu_pmu_context *); in DEFINE_FREE()
12193 if (!pmu->cpu_pmu_context) in DEFINE_FREE()
12194 return -ENOMEM; in DEFINE_FREE()
12203 return -ENOMEM; in DEFINE_FREE()
12205 *per_cpu_ptr(pmu->cpu_pmu_context, cpu) = cpc; in DEFINE_FREE()
12206 __perf_init_event_pmu_context(&cpc->epc, pmu); in DEFINE_FREE()
12210 if (!pmu->start_txn) { in DEFINE_FREE()
12211 if (pmu->pmu_enable) { in DEFINE_FREE()
12217 pmu->start_txn = perf_pmu_start_txn; in DEFINE_FREE()
12218 pmu->commit_txn = perf_pmu_commit_txn; in DEFINE_FREE()
12219 pmu->cancel_txn = perf_pmu_cancel_txn; in DEFINE_FREE()
12221 pmu->start_txn = perf_pmu_nop_txn; in DEFINE_FREE()
12222 pmu->commit_txn = perf_pmu_nop_int; in DEFINE_FREE()
12223 pmu->cancel_txn = perf_pmu_nop_void; in DEFINE_FREE()
12227 if (!pmu->pmu_enable) { in DEFINE_FREE()
12228 pmu->pmu_enable = perf_pmu_nop_void; in DEFINE_FREE()
12229 pmu->pmu_disable = perf_pmu_nop_void; in DEFINE_FREE()
12232 if (!pmu->check_period) in DEFINE_FREE()
12233 pmu->check_period = perf_event_nop_int; in DEFINE_FREE()
12235 if (!pmu->event_idx) in DEFINE_FREE()
12236 pmu->event_idx = perf_event_idx_default; in DEFINE_FREE()
12241 if (!idr_cmpxchg(&pmu_idr, pmu->type, NULL, pmu)) in DEFINE_FREE()
12242 return -EINVAL; in DEFINE_FREE()
12243 list_add_rcu(&pmu->entry, &pmus); in DEFINE_FREE()
12254 list_del_rcu(&pmu->entry); in perf_pmu_unregister()
12255 idr_remove(&pmu_idr, pmu->type); in perf_pmu_unregister()
12271 return (event->attr.sample_regs_user & PERF_REG_EXTENDED_MASK) || in has_extended_regs()
12272 (event->attr.sample_regs_intr & PERF_REG_EXTENDED_MASK); in has_extended_regs()
12280 if (!try_module_get(pmu->module)) in perf_try_init_event()
12281 return -ENODEV; in perf_try_init_event()
12284 * A number of pmu->event_init() methods iterate the sibling_list to, in perf_try_init_event()
12286 * if this is a sibling event, acquire the ctx->mutex to protect in perf_try_init_event()
12289 if (event->group_leader != event && pmu->task_ctx_nr != perf_sw_context) { in perf_try_init_event()
12291 * This ctx->mutex can nest when we're called through in perf_try_init_event()
12294 ctx = perf_event_ctx_lock_nested(event->group_leader, in perf_try_init_event()
12299 event->pmu = pmu; in perf_try_init_event()
12300 ret = pmu->event_init(event); in perf_try_init_event()
12303 perf_event_ctx_unlock(event->group_leader, ctx); in perf_try_init_event()
12308 if (!(pmu->capabilities & PERF_PMU_CAP_EXTENDED_REGS) && in perf_try_init_event()
12310 ret = -EOPNOTSUPP; in perf_try_init_event()
12314 if (pmu->capabilities & PERF_PMU_CAP_NO_EXCLUDE && in perf_try_init_event()
12316 ret = -EINVAL; in perf_try_init_event()
12320 if (pmu->scope != PERF_PMU_SCOPE_NONE && event->cpu >= 0) { in perf_try_init_event()
12325 cpumask = perf_scope_cpu_topology_cpumask(pmu->scope, event->cpu); in perf_try_init_event()
12326 pmu_cpumask = perf_scope_cpumask(pmu->scope); in perf_try_init_event()
12328 ret = -ENODEV; in perf_try_init_event()
12336 event->event_caps |= PERF_EV_CAP_READ_SCOPE; in perf_try_init_event()
12342 if (event->destroy) { in perf_try_init_event()
12343 event->destroy(event); in perf_try_init_event()
12344 event->destroy = NULL; in perf_try_init_event()
12348 event->pmu = NULL; in perf_try_init_event()
12349 module_put(pmu->module); in perf_try_init_event()
12362 * Save original type before calling pmu->event_init() since certain in perf_init_event()
12363 * pmus overwrites event->attr.type to forward event to another pmu. in perf_init_event()
12365 event->orig_type = event->attr.type; in perf_init_event()
12368 if (event->parent && event->parent->pmu) { in perf_init_event()
12369 pmu = event->parent->pmu; in perf_init_event()
12379 type = event->attr.type; in perf_init_event()
12381 type = event->attr.config >> PERF_PMU_TYPE_SHIFT; in perf_init_event()
12386 event->attr.config &= PERF_HW_EVENT_MASK; in perf_init_event()
12394 if (event->attr.type != type && type != PERF_TYPE_RAW && in perf_init_event()
12395 !(pmu->capabilities & PERF_PMU_CAP_EXTENDED_HW_TYPE)) in perf_init_event()
12396 return ERR_PTR(-ENOENT); in perf_init_event()
12399 if (ret == -ENOENT && event->attr.type != type && !extended_type) { in perf_init_event()
12400 type = event->attr.type; in perf_init_event()
12415 if (ret != -ENOENT) in perf_init_event()
12419 return ERR_PTR(-ENOENT); in perf_init_event()
12424 struct pmu_event_list *pel = per_cpu_ptr(&pmu_sb_events, event->cpu); in attach_sb_event()
12426 raw_spin_lock(&pel->lock); in attach_sb_event()
12427 list_add_rcu(&event->sb_list, &pel->list); in attach_sb_event()
12428 raw_spin_unlock(&pel->lock); in attach_sb_event()
12432 * We keep a list of all !task (and therefore per-cpu) events
12433 * that need to receive side-band records.
12435 * This avoids having to scan all the various PMU per-cpu contexts
12469 if (event->parent) in account_event()
12472 if (event->attach_state & (PERF_ATTACH_TASK | PERF_ATTACH_SCHED_CB)) in account_event()
12474 if (event->attr.mmap || event->attr.mmap_data) in account_event()
12476 if (event->attr.build_id) in account_event()
12478 if (event->attr.comm) in account_event()
12480 if (event->attr.namespaces) in account_event()
12482 if (event->attr.cgroup) in account_event()
12484 if (event->attr.task) in account_event()
12486 if (event->attr.freq) in account_event()
12488 if (event->attr.context_switch) { in account_event()
12496 if (event->attr.ksymbol) in account_event()
12498 if (event->attr.bpf_event) in account_event()
12500 if (event->attr.text_poke) in account_event()
12524 * increments to by-pass the mutex. in account_event()
12547 long err = -EINVAL; in perf_event_alloc()
12551 if (!task || cpu != -1) in perf_event_alloc()
12552 return ERR_PTR(-EINVAL); in perf_event_alloc()
12554 if (attr->sigtrap && !task) { in perf_event_alloc()
12556 return ERR_PTR(-EINVAL); in perf_event_alloc()
12559 node = (cpu >= 0) ? cpu_to_node(cpu) : -1; in perf_event_alloc()
12563 return ERR_PTR(-ENOMEM); in perf_event_alloc()
12572 mutex_init(&event->child_mutex); in perf_event_alloc()
12573 INIT_LIST_HEAD(&event->child_list); in perf_event_alloc()
12575 INIT_LIST_HEAD(&event->event_entry); in perf_event_alloc()
12576 INIT_LIST_HEAD(&event->sibling_list); in perf_event_alloc()
12577 INIT_LIST_HEAD(&event->active_list); in perf_event_alloc()
12579 INIT_LIST_HEAD(&event->rb_entry); in perf_event_alloc()
12580 INIT_LIST_HEAD(&event->active_entry); in perf_event_alloc()
12581 INIT_LIST_HEAD(&event->addr_filters.list); in perf_event_alloc()
12582 INIT_HLIST_NODE(&event->hlist_entry); in perf_event_alloc()
12585 init_waitqueue_head(&event->waitq); in perf_event_alloc()
12586 init_irq_work(&event->pending_irq, perf_pending_irq); in perf_event_alloc()
12587 event->pending_disable_irq = IRQ_WORK_INIT_HARD(perf_pending_disable); in perf_event_alloc()
12588 init_task_work(&event->pending_task, perf_pending_task); in perf_event_alloc()
12590 mutex_init(&event->mmap_mutex); in perf_event_alloc()
12591 raw_spin_lock_init(&event->addr_filters.lock); in perf_event_alloc()
12593 atomic_long_set(&event->refcount, 1); in perf_event_alloc()
12594 event->cpu = cpu; in perf_event_alloc()
12595 event->attr = *attr; in perf_event_alloc()
12596 event->group_leader = group_leader; in perf_event_alloc()
12597 event->pmu = NULL; in perf_event_alloc()
12598 event->oncpu = -1; in perf_event_alloc()
12600 event->parent = parent_event; in perf_event_alloc()
12602 event->ns = get_pid_ns(task_active_pid_ns(current)); in perf_event_alloc()
12603 event->id = atomic64_inc_return(&perf_event_id); in perf_event_alloc()
12605 event->state = PERF_EVENT_STATE_INACTIVE; in perf_event_alloc()
12608 event->event_caps = parent_event->event_caps; in perf_event_alloc()
12611 event->attach_state = PERF_ATTACH_TASK; in perf_event_alloc()
12617 event->hw.target = get_task_struct(task); in perf_event_alloc()
12620 event->clock = &local_clock; in perf_event_alloc()
12622 event->clock = parent_event->clock; in perf_event_alloc()
12625 overflow_handler = parent_event->overflow_handler; in perf_event_alloc()
12626 context = parent_event->overflow_handler_context; in perf_event_alloc()
12628 if (parent_event->prog) { in perf_event_alloc()
12629 struct bpf_prog *prog = parent_event->prog; in perf_event_alloc()
12632 event->prog = prog; in perf_event_alloc()
12638 event->overflow_handler = overflow_handler; in perf_event_alloc()
12639 event->overflow_handler_context = context; in perf_event_alloc()
12641 event->overflow_handler = perf_event_output_backward; in perf_event_alloc()
12642 event->overflow_handler_context = NULL; in perf_event_alloc()
12644 event->overflow_handler = perf_event_output_forward; in perf_event_alloc()
12645 event->overflow_handler_context = NULL; in perf_event_alloc()
12652 hwc = &event->hw; in perf_event_alloc()
12653 hwc->sample_period = attr->sample_period; in perf_event_alloc()
12654 if (attr->freq && attr->sample_freq) in perf_event_alloc()
12655 hwc->sample_period = 1; in perf_event_alloc()
12656 hwc->last_period = hwc->sample_period; in perf_event_alloc()
12658 local64_set(&hwc->period_left, hwc->sample_period); in perf_event_alloc()
12663 * collect per-thread samples. in perf_event_alloc()
12666 if (has_inherit_and_sample_read(attr) && !(attr->sample_type & PERF_SAMPLE_TID)) in perf_event_alloc()
12667 return ERR_PTR(-EINVAL); in perf_event_alloc()
12670 event->attr.branch_sample_type = 0; in perf_event_alloc()
12677 * The PERF_ATTACH_TASK_DATA is set in the event_init()->hw_config(). in perf_event_alloc()
12679 * Otherwise, the __free_event() would mistakenly detach the non-exist in perf_event_alloc()
12682 if (event->attach_state & PERF_ATTACH_TASK_DATA) { in perf_event_alloc()
12689 * Disallow uncore-task events. Similarly, disallow uncore-cgroup in perf_event_alloc()
12693 if (pmu->task_ctx_nr == perf_invalid_context && (task || cgroup_fd != -1)) in perf_event_alloc()
12694 return ERR_PTR(-EINVAL); in perf_event_alloc()
12696 if (event->attr.aux_output && in perf_event_alloc()
12697 (!(pmu->capabilities & PERF_PMU_CAP_AUX_OUTPUT) || in perf_event_alloc()
12698 event->attr.aux_pause || event->attr.aux_resume)) in perf_event_alloc()
12699 return ERR_PTR(-EOPNOTSUPP); in perf_event_alloc()
12701 if (event->attr.aux_pause && event->attr.aux_resume) in perf_event_alloc()
12702 return ERR_PTR(-EINVAL); in perf_event_alloc()
12704 if (event->attr.aux_start_paused) { in perf_event_alloc()
12705 if (!(pmu->capabilities & PERF_PMU_CAP_AUX_PAUSE)) in perf_event_alloc()
12706 return ERR_PTR(-EOPNOTSUPP); in perf_event_alloc()
12707 event->hw.aux_paused = 1; in perf_event_alloc()
12710 if (cgroup_fd != -1) { in perf_event_alloc()
12721 event->addr_filter_ranges = kcalloc(pmu->nr_addr_filters, in perf_event_alloc()
12724 if (!event->addr_filter_ranges) in perf_event_alloc()
12725 return ERR_PTR(-ENOMEM); in perf_event_alloc()
12731 if (event->parent) { in perf_event_alloc()
12734 raw_spin_lock_irq(&ifh->lock); in perf_event_alloc()
12735 memcpy(event->addr_filter_ranges, in perf_event_alloc()
12736 event->parent->addr_filter_ranges, in perf_event_alloc()
12737 pmu->nr_addr_filters * sizeof(struct perf_addr_filter_range)); in perf_event_alloc()
12738 raw_spin_unlock_irq(&ifh->lock); in perf_event_alloc()
12742 event->addr_filters_gen = 1; in perf_event_alloc()
12745 if (!event->parent) { in perf_event_alloc()
12746 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { in perf_event_alloc()
12747 err = get_callchain_buffers(attr->sample_max_stack); in perf_event_alloc()
12750 event->attach_state |= PERF_ATTACH_CALLCHAIN; in perf_event_alloc()
12773 ret = get_user(size, &uattr->size); in perf_copy_attr()
12785 if (ret == -E2BIG) in perf_copy_attr()
12790 attr->size = size; in perf_copy_attr()
12792 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3) in perf_copy_attr()
12793 return -EINVAL; in perf_copy_attr()
12795 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) in perf_copy_attr()
12796 return -EINVAL; in perf_copy_attr()
12798 if (attr->read_format & ~(PERF_FORMAT_MAX-1)) in perf_copy_attr()
12799 return -EINVAL; in perf_copy_attr()
12801 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { in perf_copy_attr()
12802 u64 mask = attr->branch_sample_type; in perf_copy_attr()
12805 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) in perf_copy_attr()
12806 return -EINVAL; in perf_copy_attr()
12810 return -EINVAL; in perf_copy_attr()
12816 if (!attr->exclude_kernel) in perf_copy_attr()
12819 if (!attr->exclude_user) in perf_copy_attr()
12822 if (!attr->exclude_hv) in perf_copy_attr()
12827 attr->branch_sample_type = mask; in perf_copy_attr()
12837 if (attr->sample_type & PERF_SAMPLE_REGS_USER) { in perf_copy_attr()
12838 ret = perf_reg_validate(attr->sample_regs_user); in perf_copy_attr()
12843 if (attr->sample_type & PERF_SAMPLE_STACK_USER) { in perf_copy_attr()
12845 return -ENOSYS; in perf_copy_attr()
12852 if (attr->sample_stack_user >= USHRT_MAX) in perf_copy_attr()
12853 return -EINVAL; in perf_copy_attr()
12854 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) in perf_copy_attr()
12855 return -EINVAL; in perf_copy_attr()
12858 if (!attr->sample_max_stack) in perf_copy_attr()
12859 attr->sample_max_stack = sysctl_perf_event_max_stack; in perf_copy_attr()
12861 if (attr->sample_type & PERF_SAMPLE_REGS_INTR) in perf_copy_attr()
12862 ret = perf_reg_validate(attr->sample_regs_intr); in perf_copy_attr()
12865 if (attr->sample_type & PERF_SAMPLE_CGROUP) in perf_copy_attr()
12866 return -EINVAL; in perf_copy_attr()
12868 if ((attr->sample_type & PERF_SAMPLE_WEIGHT) && in perf_copy_attr()
12869 (attr->sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) in perf_copy_attr()
12870 return -EINVAL; in perf_copy_attr()
12872 if (!attr->inherit && attr->inherit_thread) in perf_copy_attr()
12873 return -EINVAL; in perf_copy_attr()
12875 if (attr->remove_on_exec && attr->enable_on_exec) in perf_copy_attr()
12876 return -EINVAL; in perf_copy_attr()
12878 if (attr->sigtrap && !attr->remove_on_exec) in perf_copy_attr()
12879 return -EINVAL; in perf_copy_attr()
12885 put_user(sizeof(*attr), &uattr->size); in perf_copy_attr()
12886 ret = -E2BIG; in perf_copy_attr()
12903 int ret = -EINVAL; in perf_event_set_output()
12906 mutex_lock(&event->mmap_mutex); in perf_event_set_output()
12915 * Don't allow cross-cpu buffers in perf_event_set_output()
12917 if (output_event->cpu != event->cpu) in perf_event_set_output()
12921 * If its not a per-cpu rb, it must be the same task. in perf_event_set_output()
12923 if (output_event->cpu == -1 && output_event->hw.target != event->hw.target) in perf_event_set_output()
12929 if (output_event->clock != event->clock) in perf_event_set_output()
12943 event->pmu != output_event->pmu) in perf_event_set_output()
12948 * output_event is already on rb->event_list, and the list iteration in perf_event_set_output()
12951 * observe !rb->mmap_count. in perf_event_set_output()
12953 mutex_lock_double(&event->mmap_mutex, &output_event->mmap_mutex); in perf_event_set_output()
12956 if (atomic_read(&event->mmap_count)) in perf_event_set_output()
12966 if (!atomic_read(&rb->mmap_count)) { in perf_event_set_output()
12976 mutex_unlock(&event->mmap_mutex); in perf_event_set_output()
12978 mutex_unlock(&output_event->mmap_mutex); in perf_event_set_output()
12990 event->clock = &ktime_get_mono_fast_ns; in perf_event_set_clock()
12995 event->clock = &ktime_get_raw_fast_ns; in perf_event_set_clock()
13000 event->clock = &ktime_get_real_ns; in perf_event_set_clock()
13004 event->clock = &ktime_get_boottime_ns; in perf_event_set_clock()
13008 event->clock = &ktime_get_clocktai_ns; in perf_event_set_clock()
13012 return -EINVAL; in perf_event_set_clock()
13015 if (!nmi_safe && !(event->pmu->capabilities & PERF_PMU_CAP_NO_NMI)) in perf_event_set_clock()
13016 return -EINVAL; in perf_event_set_clock()
13027 if (attr->sigtrap) { in perf_check_permission()
13033 is_capable &= ns_capable(__task_cred(task)->user_ns, CAP_KILL); in perf_check_permission()
13053 * sys_perf_event_open - open a performance event, associate it to a task/cpu
13058 * @group_fd: group leader event fd
13077 int cgroup_fd = -1; in SYSCALL_DEFINE5()
13081 return -EINVAL; in SYSCALL_DEFINE5()
13100 return -EACCES; in SYSCALL_DEFINE5()
13105 return -EINVAL; in SYSCALL_DEFINE5()
13108 return -EINVAL; in SYSCALL_DEFINE5()
13131 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) in SYSCALL_DEFINE5()
13132 return -EINVAL; in SYSCALL_DEFINE5()
13141 CLASS(fd, group)(group_fd); // group_fd == -1 => empty in SYSCALL_DEFINE5()
13142 if (group_fd != -1) { in SYSCALL_DEFINE5()
13144 err = -EBADF; in SYSCALL_DEFINE5()
13147 group_leader = fd_file(group)->private_data; in SYSCALL_DEFINE5()
13154 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { in SYSCALL_DEFINE5()
13163 group_leader->attr.inherit != attr.inherit) { in SYSCALL_DEFINE5()
13164 err = -EINVAL; in SYSCALL_DEFINE5()
13179 if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) { in SYSCALL_DEFINE5()
13180 err = -EOPNOTSUPP; in SYSCALL_DEFINE5()
13189 pmu = event->pmu; in SYSCALL_DEFINE5()
13197 if (pmu->task_ctx_nr == perf_sw_context) in SYSCALL_DEFINE5()
13198 event->event_caps |= PERF_EV_CAP_SOFTWARE; in SYSCALL_DEFINE5()
13201 err = down_read_interruptible(&task->signal->exec_update_lock); in SYSCALL_DEFINE5()
13211 err = -EACCES; in SYSCALL_DEFINE5()
13225 mutex_lock(&ctx->mutex); in SYSCALL_DEFINE5()
13227 if (ctx->task == TASK_TOMBSTONE) { in SYSCALL_DEFINE5()
13228 err = -ESRCH; in SYSCALL_DEFINE5()
13239 struct perf_cpu_context *cpuctx = per_cpu_ptr(&perf_cpu_context, event->cpu); in SYSCALL_DEFINE5()
13241 if (!cpuctx->online) { in SYSCALL_DEFINE5()
13242 err = -ENODEV; in SYSCALL_DEFINE5()
13248 err = -EINVAL; in SYSCALL_DEFINE5()
13252 * becoming part of another group-sibling): in SYSCALL_DEFINE5()
13254 if (group_leader->group_leader != group_leader) in SYSCALL_DEFINE5()
13258 if (group_leader->clock != event->clock) in SYSCALL_DEFINE5()
13266 if (group_leader->cpu != event->cpu) in SYSCALL_DEFINE5()
13272 if (group_leader->ctx != ctx) in SYSCALL_DEFINE5()
13276 * Only a group leader can be exclusive or pinned in SYSCALL_DEFINE5()
13294 pmu = group_leader->pmu_ctx->pmu; in SYSCALL_DEFINE5()
13297 (group_leader->group_caps & PERF_EV_CAP_SOFTWARE)) { in SYSCALL_DEFINE5()
13308 group_leader->pmu_ctx->pmu != pmu) in SYSCALL_DEFINE5()
13321 event->pmu_ctx = pmu_ctx; in SYSCALL_DEFINE5()
13330 err = -E2BIG; in SYSCALL_DEFINE5()
13335 err = -EINVAL; in SYSCALL_DEFINE5()
13344 err = -EBUSY; in SYSCALL_DEFINE5()
13348 WARN_ON_ONCE(ctx->parent_ctx); in SYSCALL_DEFINE5()
13364 put_pmu_ctx(group_leader->pmu_ctx); in SYSCALL_DEFINE5()
13368 put_pmu_ctx(sibling->pmu_ctx); in SYSCALL_DEFINE5()
13372 * Install the group siblings before the group leader. in SYSCALL_DEFINE5()
13374 * Because a group leader will try and install the entire group in SYSCALL_DEFINE5()
13375 * (through the sibling list, which is still in-tact), we can in SYSCALL_DEFINE5()
13378 * By installing siblings first we NO-OP because they're not in SYSCALL_DEFINE5()
13382 sibling->pmu_ctx = pmu_ctx; in SYSCALL_DEFINE5()
13385 perf_install_in_context(ctx, sibling, sibling->cpu); in SYSCALL_DEFINE5()
13393 group_leader->pmu_ctx = pmu_ctx; in SYSCALL_DEFINE5()
13396 perf_install_in_context(ctx, group_leader, group_leader->cpu); in SYSCALL_DEFINE5()
13408 event->owner = current; in SYSCALL_DEFINE5()
13410 perf_install_in_context(ctx, event, event->cpu); in SYSCALL_DEFINE5()
13413 mutex_unlock(&ctx->mutex); in SYSCALL_DEFINE5()
13416 up_read(&task->signal->exec_update_lock); in SYSCALL_DEFINE5()
13420 mutex_lock(¤t->perf_event_mutex); in SYSCALL_DEFINE5()
13421 list_add_tail(&event->owner_entry, ¤t->perf_event_list); in SYSCALL_DEFINE5()
13422 mutex_unlock(¤t->perf_event_mutex); in SYSCALL_DEFINE5()
13427 * This ensures destruction of the group leader will find in SYSCALL_DEFINE5()
13434 put_pmu_ctx(event->pmu_ctx); in SYSCALL_DEFINE5()
13435 event->pmu_ctx = NULL; /* _free_event() */ in SYSCALL_DEFINE5()
13437 mutex_unlock(&ctx->mutex); in SYSCALL_DEFINE5()
13442 up_read(&task->signal->exec_update_lock); in SYSCALL_DEFINE5()
13478 if (attr->aux_output || attr->aux_action) in perf_event_create_kernel_counter()
13479 return ERR_PTR(-EINVAL); in perf_event_create_kernel_counter()
13482 overflow_handler, context, -1); in perf_event_create_kernel_counter()
13489 event->owner = TASK_TOMBSTONE; in perf_event_create_kernel_counter()
13490 pmu = event->pmu; in perf_event_create_kernel_counter()
13492 if (pmu->task_ctx_nr == perf_sw_context) in perf_event_create_kernel_counter()
13493 event->event_caps |= PERF_EV_CAP_SOFTWARE; in perf_event_create_kernel_counter()
13504 WARN_ON_ONCE(ctx->parent_ctx); in perf_event_create_kernel_counter()
13505 mutex_lock(&ctx->mutex); in perf_event_create_kernel_counter()
13506 if (ctx->task == TASK_TOMBSTONE) { in perf_event_create_kernel_counter()
13507 err = -ESRCH; in perf_event_create_kernel_counter()
13516 event->pmu_ctx = pmu_ctx; in perf_event_create_kernel_counter()
13527 if (!cpuctx->online) { in perf_event_create_kernel_counter()
13528 err = -ENODEV; in perf_event_create_kernel_counter()
13534 err = -EBUSY; in perf_event_create_kernel_counter()
13538 perf_install_in_context(ctx, event, event->cpu); in perf_event_create_kernel_counter()
13540 mutex_unlock(&ctx->mutex); in perf_event_create_kernel_counter()
13546 event->pmu_ctx = NULL; /* _free_event() */ in perf_event_create_kernel_counter()
13548 mutex_unlock(&ctx->mutex); in perf_event_create_kernel_counter()
13567 put_pmu_ctx(event->pmu_ctx); in __perf_pmu_remove()
13568 list_add(&event->migrate_entry, events); in __perf_pmu_remove()
13572 put_pmu_ctx(sibling->pmu_ctx); in __perf_pmu_remove()
13573 list_add(&sibling->migrate_entry, events); in __perf_pmu_remove()
13583 struct perf_event_context *old_ctx = event->ctx; in __perf_pmu_install_event()
13587 event->cpu = cpu; in __perf_pmu_install_event()
13589 event->pmu_ctx = epc; in __perf_pmu_install_event()
13591 if (event->state >= PERF_EVENT_STATE_OFF) in __perf_pmu_install_event()
13592 event->state = PERF_EVENT_STATE_INACTIVE; in __perf_pmu_install_event()
13596 * Now that event->ctx is updated and visible, put the old ctx. in __perf_pmu_install_event()
13607 * Re-instate events in 2 passes. in __perf_pmu_install()
13610 * pass, siblings will not get enabled without a leader, however a in __perf_pmu_install()
13611 * leader will enable its siblings, even if those are still on the old in __perf_pmu_install()
13615 if (event->group_leader == event) in __perf_pmu_install()
13618 list_del(&event->migrate_entry); in __perf_pmu_install()
13627 list_del(&event->migrate_entry); in __perf_pmu_install()
13638 * Since per-cpu context is persistent, no need to grab an extra in perf_pmu_migrate_context()
13641 src_ctx = &per_cpu_ptr(&perf_cpu_context, src_cpu)->ctx; in perf_pmu_migrate_context()
13642 dst_ctx = &per_cpu_ptr(&perf_cpu_context, dst_cpu)->ctx; in perf_pmu_migrate_context()
13648 mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex); in perf_pmu_migrate_context()
13650 __perf_pmu_remove(src_ctx, src_cpu, pmu, &src_ctx->pinned_groups, &events); in perf_pmu_migrate_context()
13651 __perf_pmu_remove(src_ctx, src_cpu, pmu, &src_ctx->flexible_groups, &events); in perf_pmu_migrate_context()
13655 * Wait for the events to quiesce before re-instating them. in perf_pmu_migrate_context()
13662 mutex_unlock(&dst_ctx->mutex); in perf_pmu_migrate_context()
13663 mutex_unlock(&src_ctx->mutex); in perf_pmu_migrate_context()
13669 struct perf_event *parent_event = child_event->parent; in sync_child_event()
13672 if (child_event->attr.inherit_stat) { in sync_child_event()
13673 struct task_struct *task = child_event->ctx->task; in sync_child_event()
13684 atomic64_add(child_val, &parent_event->child_count); in sync_child_event()
13685 atomic64_add(child_event->total_time_enabled, in sync_child_event()
13686 &parent_event->child_total_time_enabled); in sync_child_event()
13687 atomic64_add(child_event->total_time_running, in sync_child_event()
13688 &parent_event->child_total_time_running); in sync_child_event()
13694 struct perf_event *parent_event = event->parent; in perf_event_exit_event()
13711 mutex_lock(&parent_event->child_mutex); in perf_event_exit_event()
13720 mutex_unlock(&parent_event->child_mutex); in perf_event_exit_event()
13747 * In order to reduce the amount of tricky in ctx tear-down, we hold in perf_event_exit_task_context()
13756 mutex_lock(&child_ctx->mutex); in perf_event_exit_task_context()
13759 * In a single ctx::lock section, de-schedule the events and detach the in perf_event_exit_task_context()
13763 raw_spin_lock_irq(&child_ctx->lock); in perf_event_exit_task_context()
13767 * Now that the context is inactive, destroy the task <-> ctx relation in perf_event_exit_task_context()
13770 RCU_INIT_POINTER(child->perf_event_ctxp, NULL); in perf_event_exit_task_context()
13772 WRITE_ONCE(child_ctx->task, TASK_TOMBSTONE); in perf_event_exit_task_context()
13776 raw_spin_unlock_irq(&child_ctx->lock); in perf_event_exit_task_context()
13788 list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry) in perf_event_exit_task_context()
13791 mutex_unlock(&child_ctx->mutex); in perf_event_exit_task_context()
13806 mutex_lock(&child->perf_event_mutex); in perf_event_exit_task()
13807 list_for_each_entry_safe(event, tmp, &child->perf_event_list, in perf_event_exit_task()
13809 list_del_init(&event->owner_entry); in perf_event_exit_task()
13814 * we need to serialize on the owner->perf_event_mutex. in perf_event_exit_task()
13816 smp_store_release(&event->owner, NULL); in perf_event_exit_task()
13818 mutex_unlock(&child->perf_event_mutex); in perf_event_exit_task()
13825 * child contexts and sets child->perf_event_ctxp[] to NULL. in perf_event_exit_task()
13831 * Detach the perf_ctx_data for the system-wide event. in perf_event_exit_task()
13840 struct perf_event *parent = event->parent; in perf_free_event()
13845 mutex_lock(&parent->child_mutex); in perf_free_event()
13846 list_del_init(&event->child_list); in perf_free_event()
13847 mutex_unlock(&parent->child_mutex); in perf_free_event()
13849 raw_spin_lock_irq(&ctx->lock); in perf_free_event()
13852 raw_spin_unlock_irq(&ctx->lock); in perf_free_event()
13868 ctx = rcu_access_pointer(task->perf_event_ctxp); in perf_event_free_task()
13872 mutex_lock(&ctx->mutex); in perf_event_free_task()
13873 raw_spin_lock_irq(&ctx->lock); in perf_event_free_task()
13875 * Destroy the task <-> ctx relation and mark the context dead. in perf_event_free_task()
13880 RCU_INIT_POINTER(task->perf_event_ctxp, NULL); in perf_event_free_task()
13881 WRITE_ONCE(ctx->task, TASK_TOMBSTONE); in perf_event_free_task()
13883 raw_spin_unlock_irq(&ctx->lock); in perf_event_free_task()
13886 list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) in perf_event_free_task()
13889 mutex_unlock(&ctx->mutex); in perf_event_free_task()
13900 * _free_event()'s put_task_struct(event->hw.target) will be a in perf_event_free_task()
13901 * use-after-free. in perf_event_free_task()
13905 wait_var_event(&ctx->refcount, refcount_read(&ctx->refcount) == 1); in perf_event_free_task()
13911 WARN_ON_ONCE(task->perf_event_ctxp); in perf_event_delayed_put()
13918 return ERR_PTR(-EBADF); in perf_event_get()
13920 if (file->f_op != &perf_fops) { in perf_event_get()
13922 return ERR_PTR(-EBADF); in perf_event_get()
13930 if (file->f_op != &perf_fops) in perf_get_event()
13931 return ERR_PTR(-EINVAL); in perf_get_event()
13933 return file->private_data; in perf_get_event()
13939 return ERR_PTR(-EINVAL); in perf_event_attrs()
13941 return &event->attr; in perf_event_attrs()
13947 return -EACCES; in perf_allow_kernel()
13957 * - valid pointer on success
13958 * - NULL for orphaned events
13959 * - IS_ERR() on error
13969 enum perf_event_state parent_state = parent_event->state; in inherit_event()
13980 if (parent_event->parent) in inherit_event()
13981 parent_event = parent_event->parent; in inherit_event()
13983 child_event = perf_event_alloc(&parent_event->attr, in inherit_event()
13984 parent_event->cpu, in inherit_event()
13987 NULL, NULL, -1); in inherit_event()
13992 child_event->ctx = child_ctx; in inherit_event()
13994 pmu_ctx = find_get_pmu_context(child_event->pmu, child_ctx, child_event); in inherit_event()
13999 child_event->pmu_ctx = pmu_ctx; in inherit_event()
14002 * is_orphaned_event() and list_add_tail(&parent_event->child_list) in inherit_event()
14007 mutex_lock(&parent_event->child_mutex); in inherit_event()
14009 !atomic_long_inc_not_zero(&parent_event->refcount)) { in inherit_event()
14010 mutex_unlock(&parent_event->child_mutex); in inherit_event()
14021 child_event->state = PERF_EVENT_STATE_INACTIVE; in inherit_event()
14023 child_event->state = PERF_EVENT_STATE_OFF; in inherit_event()
14025 if (parent_event->attr.freq) { in inherit_event()
14026 u64 sample_period = parent_event->hw.sample_period; in inherit_event()
14027 struct hw_perf_event *hwc = &child_event->hw; in inherit_event()
14029 hwc->sample_period = sample_period; in inherit_event()
14030 hwc->last_period = sample_period; in inherit_event()
14032 local64_set(&hwc->period_left, sample_period); in inherit_event()
14035 child_event->overflow_handler = parent_event->overflow_handler; in inherit_event()
14036 child_event->overflow_handler_context in inherit_event()
14037 = parent_event->overflow_handler_context; in inherit_event()
14048 raw_spin_lock_irqsave(&child_ctx->lock, flags); in inherit_event()
14050 child_event->attach_state |= PERF_ATTACH_CHILD; in inherit_event()
14051 raw_spin_unlock_irqrestore(&child_ctx->lock, flags); in inherit_event()
14056 list_add_tail(&child_event->child_list, &parent_event->child_list); in inherit_event()
14057 mutex_unlock(&parent_event->child_mutex); in inherit_event()
14069 * - 0 on success
14070 * - <0 on error
14078 struct perf_event *leader; in inherit_group() local
14082 leader = inherit_event(parent_event, parent, parent_ctx, in inherit_group()
14084 if (IS_ERR(leader)) in inherit_group()
14085 return PTR_ERR(leader); in inherit_group()
14087 * @leader can be NULL here because of is_orphaned_event(). In this in inherit_group()
14093 child, leader, child_ctx); in inherit_group()
14097 if (sub->aux_event == parent_event && child_ctr && in inherit_group()
14098 !perf_get_aux_event(child_ctr, leader)) in inherit_group()
14099 return -EINVAL; in inherit_group()
14101 if (leader) in inherit_group()
14102 leader->group_generation = parent_event->group_generation; in inherit_group()
14107 * Creates the child task context and tries to inherit the event-group.
14114 * - 0 on success
14115 * - <0 on error
14126 if (!event->attr.inherit || in inherit_task_group()
14127 (event->attr.inherit_thread && !(clone_flags & CLONE_THREAD)) || in inherit_task_group()
14129 (event->attr.sigtrap && (clone_flags & CLONE_CLEAR_SIGHAND))) { in inherit_task_group()
14134 child_ctx = child->perf_event_ctxp; in inherit_task_group()
14144 return -ENOMEM; in inherit_task_group()
14146 child->perf_event_ctxp = child_ctx; in inherit_task_group()
14169 if (likely(!parent->perf_event_ctxp)) in perf_event_init_context()
14181 * No need to check if parent_ctx != NULL here; since we saw in perf_event_init_context()
14182 * it non-NULL earlier, the only reason for it to become NULL in perf_event_init_context()
14188 * Lock the parent list. No need to lock the child - not PID in perf_event_init_context()
14191 mutex_lock(&parent_ctx->mutex); in perf_event_init_context()
14194 * We dont have to disable NMIs - we are only looking at in perf_event_init_context()
14197 perf_event_groups_for_each(event, &parent_ctx->pinned_groups) { in perf_event_init_context()
14205 * We can't hold ctx->lock when iterating the ->flexible_group list due in perf_event_init_context()
14209 raw_spin_lock_irqsave(&parent_ctx->lock, flags); in perf_event_init_context()
14210 parent_ctx->rotate_disable = 1; in perf_event_init_context()
14211 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); in perf_event_init_context()
14213 perf_event_groups_for_each(event, &parent_ctx->flexible_groups) { in perf_event_init_context()
14220 raw_spin_lock_irqsave(&parent_ctx->lock, flags); in perf_event_init_context()
14221 parent_ctx->rotate_disable = 0; in perf_event_init_context()
14223 child_ctx = child->perf_event_ctxp; in perf_event_init_context()
14231 * parent_ctx->lock avoids it from being uncloned. in perf_event_init_context()
14233 cloned_ctx = parent_ctx->parent_ctx; in perf_event_init_context()
14235 child_ctx->parent_ctx = cloned_ctx; in perf_event_init_context()
14236 child_ctx->parent_gen = parent_ctx->parent_gen; in perf_event_init_context()
14238 child_ctx->parent_ctx = parent_ctx; in perf_event_init_context()
14239 child_ctx->parent_gen = parent_ctx->generation; in perf_event_init_context()
14241 get_ctx(child_ctx->parent_ctx); in perf_event_init_context()
14244 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); in perf_event_init_context()
14246 mutex_unlock(&parent_ctx->mutex); in perf_event_init_context()
14261 memset(child->perf_recursion, 0, sizeof(child->perf_recursion)); in perf_event_init_task()
14262 child->perf_event_ctxp = NULL; in perf_event_init_task()
14263 mutex_init(&child->perf_event_mutex); in perf_event_init_task()
14264 INIT_LIST_HEAD(&child->perf_event_list); in perf_event_init_task()
14265 child->perf_ctx_data = NULL; in perf_event_init_task()
14292 mutex_init(&swhash->hlist_mutex); in perf_event_init_all_cpus()
14300 __perf_event_init_context(&cpuctx->ctx); in perf_event_init_all_cpus()
14301 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); in perf_event_init_all_cpus()
14302 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); in perf_event_init_all_cpus()
14303 cpuctx->online = cpumask_test_cpu(cpu, perf_online_mask); in perf_event_init_all_cpus()
14304 cpuctx->heap_size = ARRAY_SIZE(cpuctx->heap_default); in perf_event_init_all_cpus()
14305 cpuctx->heap = cpuctx->heap_default; in perf_event_init_all_cpus()
14313 mutex_lock(&swhash->hlist_mutex); in perf_swevent_init_cpu()
14314 if (swhash->hlist_refcount > 0 && !swevent_hlist_deref(swhash)) { in perf_swevent_init_cpu()
14319 rcu_assign_pointer(swhash->swevent_hlist, hlist); in perf_swevent_init_cpu()
14321 mutex_unlock(&swhash->hlist_mutex); in perf_swevent_init_cpu()
14331 raw_spin_lock(&ctx->lock); in __perf_event_exit_context()
14333 list_for_each_entry(event, &ctx->event_list, event_entry) in __perf_event_exit_context()
14335 raw_spin_unlock(&ctx->lock); in __perf_event_exit_context()
14350 target[scope] = -1; in perf_event_clear_cpumask()
14363 if (pmu->scope == PERF_PMU_SCOPE_NONE || in perf_event_clear_cpumask()
14364 WARN_ON_ONCE(pmu->scope >= PERF_PMU_MAX_SCOPE)) in perf_event_clear_cpumask()
14367 if (target[pmu->scope] >= 0 && target[pmu->scope] < nr_cpu_ids) in perf_event_clear_cpumask()
14368 perf_pmu_migrate_context(pmu, cpu, target[pmu->scope]); in perf_event_clear_cpumask()
14377 // XXX simplify cpuctx->online in perf_event_exit_cpu_context()
14385 ctx = &cpuctx->ctx; in perf_event_exit_cpu_context()
14387 mutex_lock(&ctx->mutex); in perf_event_exit_cpu_context()
14389 cpuctx->online = 0; in perf_event_exit_cpu_context()
14390 mutex_unlock(&ctx->mutex); in perf_event_exit_cpu_context()
14445 ctx = &cpuctx->ctx; in perf_event_init_cpu()
14447 mutex_lock(&ctx->mutex); in perf_event_init_cpu()
14448 cpuctx->online = 1; in perf_event_init_cpu()
14449 mutex_unlock(&ctx->mutex); in perf_event_init_cpu()
14490 perf_pmu_register(&perf_cpu_clock, "cpu_clock", -1); in perf_event_init()
14491 perf_pmu_register(&perf_task_clock, "task_clock", -1); in perf_event_init()
14515 if (pmu_attr->event_str) in perf_event_sysfs_show()
14516 return sprintf(page, "%s\n", pmu_attr->event_str); in perf_event_sysfs_show()
14534 if (pmu->dev) in perf_event_sysfs_init()
14538 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); in perf_event_sysfs_init()
14558 return ERR_PTR(-ENOMEM); in perf_cgroup_css_alloc()
14560 jc->info = alloc_percpu(struct perf_cgroup_info); in perf_cgroup_css_alloc()
14561 if (!jc->info) { in perf_cgroup_css_alloc()
14563 return ERR_PTR(-ENOMEM); in perf_cgroup_css_alloc()
14566 return &jc->css; in perf_cgroup_css_alloc()
14573 free_percpu(jc->info); in perf_cgroup_css_free()
14579 perf_event_cgroup(css->cgroup); in perf_cgroup_css_online()