Lines Matching +full:reserved +full:- +full:ipi +full:- +full:vectors
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
71 struct task_struct *p = tfc->p; in remote_function()
74 /* -EAGAIN */ in remote_function()
83 tfc->ret = -ESRCH; /* No such (running) process */ in remote_function()
88 tfc->ret = tfc->func(tfc->info); in remote_function()
92 * task_function_call - call a function on the cpu on which a task runs
102 * returns @func return value or -ESRCH or -ENXIO when the process isn't running
111 .ret = -EAGAIN, in task_function_call()
121 if (ret != -EAGAIN) in task_function_call()
131 * cpu_function_call - call a function on the cpu
137 * returns: @func return value or -ENXIO when the cpu is offline
145 .ret = -ENXIO, /* No such CPU */ in cpu_function_call()
156 return this_cpu_ptr(ctx->pmu->pmu_cpu_context); in __get_cpu_context()
162 raw_spin_lock(&cpuctx->ctx.lock); in perf_ctx_lock()
164 raw_spin_lock(&ctx->lock); in perf_ctx_lock()
171 raw_spin_unlock(&ctx->lock); in perf_ctx_unlock()
172 raw_spin_unlock(&cpuctx->ctx.lock); in perf_ctx_unlock()
175 #define TASK_TOMBSTONE ((void *)-1L)
179 return READ_ONCE(event->owner) == TASK_TOMBSTONE; in is_kernel_event()
185 * When !ctx->nr_events a task context will not be scheduled. This means
191 * - removing the last event from a task ctx; this is relatively straight
194 * - adding the first event to a task ctx; this is tricky because we cannot
195 * rely on ctx->is_active and therefore cannot use event_function_call().
198 * If ctx->nr_events, then ctx->is_active and cpuctx->task_ctx are set.
213 struct perf_event *event = efs->event; in event_function()
214 struct perf_event_context *ctx = event->ctx; in event_function()
216 struct perf_event_context *task_ctx = cpuctx->task_ctx; in event_function()
223 * Since we do the IPI call without holding ctx->lock things can have in event_function()
226 if (ctx->task) { in event_function()
227 if (ctx->task != current) { in event_function()
228 ret = -ESRCH; in event_function()
236 * above ctx->task != current test), therefore we must have in event_function()
237 * ctx->is_active here. in event_function()
239 WARN_ON_ONCE(!ctx->is_active); in event_function()
241 * And since we have ctx->is_active, cpuctx->task_ctx must in event_function()
246 WARN_ON_ONCE(&cpuctx->ctx != ctx); in event_function()
249 efs->func(event, cpuctx, ctx, efs->data); in event_function()
258 struct perf_event_context *ctx = event->ctx; in event_function_call()
259 struct task_struct *task = READ_ONCE(ctx->task); /* verified in event_function */ in event_function_call()
266 if (!event->parent) { in event_function_call()
269 * stabilize the the event->ctx relation. See in event_function_call()
272 lockdep_assert_held(&ctx->mutex); in event_function_call()
276 cpu_function_call(event->cpu, event_function, &efs); in event_function_call()
287 raw_spin_lock_irq(&ctx->lock); in event_function_call()
292 task = ctx->task; in event_function_call()
294 raw_spin_unlock_irq(&ctx->lock); in event_function_call()
297 if (ctx->is_active) { in event_function_call()
298 raw_spin_unlock_irq(&ctx->lock); in event_function_call()
302 raw_spin_unlock_irq(&ctx->lock); in event_function_call()
311 struct perf_event_context *ctx = event->ctx; in event_function_local()
313 struct task_struct *task = READ_ONCE(ctx->task); in event_function_local()
327 task = ctx->task; in event_function_local()
334 * otherwise we're screwed, since we cannot IPI to somewhere in event_function_local()
337 if (ctx->is_active) { in event_function_local()
341 if (WARN_ON_ONCE(cpuctx->task_ctx != ctx)) in event_function_local()
345 WARN_ON_ONCE(&cpuctx->ctx != ctx); in event_function_local()
376 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
406 * -1 - not paranoid at all
407 * 0 - disallow raw tracepoint access for unpriv
408 * 1 - disallow cpu events for unpriv
409 * 2 - disallow kernel profiling for unpriv
454 return -EINVAL; in perf_proc_update_handler()
524 running_len -= running_len/NR_ACCUMULATED_SAMPLES; in perf_sample_event_took()
590 return event->clock(); in perf_event_clock()
596 * The basic idea is to use event->state to determine which (if any) time
618 struct perf_event *leader = event->group_leader; in __perf_effective_state()
620 if (leader->state <= PERF_EVENT_STATE_OFF) in __perf_effective_state()
621 return leader->state; in __perf_effective_state()
623 return event->state; in __perf_effective_state()
630 u64 delta = now - event->tstamp; in __perf_update_times()
632 *enabled = event->total_time_enabled; in __perf_update_times()
636 *running = event->total_time_running; in __perf_update_times()
645 __perf_update_times(event, now, &event->total_time_enabled, in perf_event_update_time()
646 &event->total_time_running); in perf_event_update_time()
647 event->tstamp = now; in perf_event_update_time()
661 if (event->state == state) in perf_event_set_state()
669 if ((event->state < 0) ^ (state < 0)) in perf_event_set_state()
672 WRITE_ONCE(event->state, state); in perf_event_set_state()
680 struct perf_event_context *ctx = event->ctx; in perf_cgroup_match()
684 if (!event->cgrp) in perf_cgroup_match()
688 if (!cpuctx->cgrp) in perf_cgroup_match()
697 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, in perf_cgroup_match()
698 event->cgrp->css.cgroup); in perf_cgroup_match()
703 css_put(&event->cgrp->css); in perf_detach_cgroup()
704 event->cgrp = NULL; in perf_detach_cgroup()
709 return event->cgrp != NULL; in is_cgroup_event()
716 t = per_cpu_ptr(event->cgrp->info, event->cpu); in perf_cgroup_event_time()
717 return t->time; in perf_cgroup_event_time()
727 info = this_cpu_ptr(cgrp->info); in __update_cgrp_time()
729 info->time += now - info->timestamp; in __update_cgrp_time()
730 info->timestamp = now; in __update_cgrp_time()
735 struct perf_cgroup *cgrp = cpuctx->cgrp; in update_cgrp_time_from_cpuctx()
739 for (css = &cgrp->css; css; css = css->parent) { in update_cgrp_time_from_cpuctx()
757 cgrp = perf_cgroup_from_task(current, event->ctx); in update_cgrp_time_from_event()
761 if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup)) in update_cgrp_time_from_event()
762 __update_cgrp_time(event->cgrp); in update_cgrp_time_from_event()
774 * ctx->lock held by caller in perf_cgroup_set_timestamp()
778 if (!task || !ctx->nr_cgroups) in perf_cgroup_set_timestamp()
783 for (css = &cgrp->css; css; css = css->parent) { in perf_cgroup_set_timestamp()
785 info = this_cpu_ptr(cgrp->info); in perf_cgroup_set_timestamp()
786 info->timestamp = ctx->timestamp; in perf_cgroup_set_timestamp()
815 WARN_ON_ONCE(cpuctx->ctx.nr_cgroups == 0); in perf_cgroup_switch()
817 perf_ctx_lock(cpuctx, cpuctx->task_ctx); in perf_cgroup_switch()
818 perf_pmu_disable(cpuctx->ctx.pmu); in perf_cgroup_switch()
826 cpuctx->cgrp = NULL; in perf_cgroup_switch()
830 WARN_ON_ONCE(cpuctx->cgrp); in perf_cgroup_switch()
835 * we pass the cpuctx->ctx to perf_cgroup_from_task() in perf_cgroup_switch()
836 * because cgorup events are only per-cpu in perf_cgroup_switch()
838 cpuctx->cgrp = perf_cgroup_from_task(task, in perf_cgroup_switch()
839 &cpuctx->ctx); in perf_cgroup_switch()
842 perf_pmu_enable(cpuctx->ctx.pmu); in perf_cgroup_switch()
843 perf_ctx_unlock(cpuctx, cpuctx->task_ctx); in perf_cgroup_switch()
912 for (heap_size = 1; css; css = css->parent) in perf_cgroup_ensure_storage()
916 cpuctx = per_cpu_ptr(event->pmu->pmu_cpu_context, cpu); in perf_cgroup_ensure_storage()
917 if (heap_size <= cpuctx->heap_size) in perf_cgroup_ensure_storage()
923 ret = -ENOMEM; in perf_cgroup_ensure_storage()
927 raw_spin_lock_irq(&cpuctx->ctx.lock); in perf_cgroup_ensure_storage()
928 if (cpuctx->heap_size < heap_size) { in perf_cgroup_ensure_storage()
929 swap(cpuctx->heap, storage); in perf_cgroup_ensure_storage()
930 if (storage == cpuctx->heap_default) in perf_cgroup_ensure_storage()
932 cpuctx->heap_size = heap_size; in perf_cgroup_ensure_storage()
934 raw_spin_unlock_irq(&cpuctx->ctx.lock); in perf_cgroup_ensure_storage()
952 return -EBADF; in perf_cgroup_connect()
954 css = css_tryget_online_from_dir(f.file->f_path.dentry, in perf_cgroup_connect()
966 event->cgrp = cgrp; in perf_cgroup_connect()
973 if (group_leader && group_leader->cgrp != cgrp) { in perf_cgroup_connect()
975 ret = -EINVAL; in perf_cgroup_connect()
986 t = per_cpu_ptr(event->cgrp->info, event->cpu); in perf_cgroup_set_shadow_time()
987 event->shadow_ctx_time = now - t->timestamp; in perf_cgroup_set_shadow_time()
999 * Because cgroup events are always per-cpu events, in perf_cgroup_event_enable()
1000 * @ctx == &cpuctx->ctx. in perf_cgroup_event_enable()
1005 * Since setting cpuctx->cgrp is conditional on the current @cgrp in perf_cgroup_event_enable()
1008 * and we would leave cpuctx->cgrp unset. in perf_cgroup_event_enable()
1010 if (ctx->is_active && !cpuctx->cgrp) { in perf_cgroup_event_enable()
1013 if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup)) in perf_cgroup_event_enable()
1014 cpuctx->cgrp = cgrp; in perf_cgroup_event_enable()
1017 if (ctx->nr_cgroups++) in perf_cgroup_event_enable()
1020 list_add(&cpuctx->cgrp_cpuctx_entry, in perf_cgroup_event_enable()
1021 per_cpu_ptr(&cgrp_cpuctx_list, event->cpu)); in perf_cgroup_event_enable()
1033 * Because cgroup events are always per-cpu events, in perf_cgroup_event_disable()
1034 * @ctx == &cpuctx->ctx. in perf_cgroup_event_disable()
1038 if (--ctx->nr_cgroups) in perf_cgroup_event_disable()
1041 if (ctx->is_active && cpuctx->cgrp) in perf_cgroup_event_disable()
1042 cpuctx->cgrp = NULL; in perf_cgroup_event_disable()
1044 list_del(&cpuctx->cgrp_cpuctx_entry); in perf_cgroup_event_disable()
1085 return -EINVAL; in perf_cgroup_connect()
1138 raw_spin_lock(&cpuctx->hrtimer_lock); in perf_mux_hrtimer_handler()
1140 hrtimer_forward_now(hr, cpuctx->hrtimer_interval); in perf_mux_hrtimer_handler()
1142 cpuctx->hrtimer_active = 0; in perf_mux_hrtimer_handler()
1143 raw_spin_unlock(&cpuctx->hrtimer_lock); in perf_mux_hrtimer_handler()
1150 struct hrtimer *timer = &cpuctx->hrtimer; in __perf_mux_hrtimer_init()
1151 struct pmu *pmu = cpuctx->ctx.pmu; in __perf_mux_hrtimer_init()
1155 if (pmu->task_ctx_nr == perf_sw_context) in __perf_mux_hrtimer_init()
1162 interval = pmu->hrtimer_interval_ms; in __perf_mux_hrtimer_init()
1164 interval = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; in __perf_mux_hrtimer_init()
1166 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * interval); in __perf_mux_hrtimer_init()
1168 raw_spin_lock_init(&cpuctx->hrtimer_lock); in __perf_mux_hrtimer_init()
1170 timer->function = perf_mux_hrtimer_handler; in __perf_mux_hrtimer_init()
1175 struct hrtimer *timer = &cpuctx->hrtimer; in perf_mux_hrtimer_restart()
1176 struct pmu *pmu = cpuctx->ctx.pmu; in perf_mux_hrtimer_restart()
1180 if (pmu->task_ctx_nr == perf_sw_context) in perf_mux_hrtimer_restart()
1183 raw_spin_lock_irqsave(&cpuctx->hrtimer_lock, flags); in perf_mux_hrtimer_restart()
1184 if (!cpuctx->hrtimer_active) { in perf_mux_hrtimer_restart()
1185 cpuctx->hrtimer_active = 1; in perf_mux_hrtimer_restart()
1186 hrtimer_forward_now(timer, cpuctx->hrtimer_interval); in perf_mux_hrtimer_restart()
1189 raw_spin_unlock_irqrestore(&cpuctx->hrtimer_lock, flags); in perf_mux_hrtimer_restart()
1196 int *count = this_cpu_ptr(pmu->pmu_disable_count); in perf_pmu_disable()
1198 pmu->pmu_disable(pmu); in perf_pmu_disable()
1203 int *count = this_cpu_ptr(pmu->pmu_disable_count); in perf_pmu_enable()
1204 if (!--(*count)) in perf_pmu_enable()
1205 pmu->pmu_enable(pmu); in perf_pmu_enable()
1222 WARN_ON(!list_empty(&ctx->active_ctx_list)); in perf_event_ctx_activate()
1224 list_add(&ctx->active_ctx_list, head); in perf_event_ctx_activate()
1231 WARN_ON(list_empty(&ctx->active_ctx_list)); in perf_event_ctx_deactivate()
1233 list_del_init(&ctx->active_ctx_list); in perf_event_ctx_deactivate()
1238 refcount_inc(&ctx->refcount); in get_ctx()
1243 if (pmu->task_ctx_cache) in alloc_task_ctx_data()
1244 return kmem_cache_zalloc(pmu->task_ctx_cache, GFP_KERNEL); in alloc_task_ctx_data()
1251 if (pmu->task_ctx_cache && task_ctx_data) in free_task_ctx_data()
1252 kmem_cache_free(pmu->task_ctx_cache, task_ctx_data); in free_task_ctx_data()
1260 free_task_ctx_data(ctx->pmu, ctx->task_ctx_data); in free_ctx()
1266 if (refcount_dec_and_test(&ctx->refcount)) { in put_ctx()
1267 if (ctx->parent_ctx) in put_ctx()
1268 put_ctx(ctx->parent_ctx); in put_ctx()
1269 if (ctx->task && ctx->task != TASK_TOMBSTONE) in put_ctx()
1270 put_task_struct(ctx->task); in put_ctx()
1271 call_rcu(&ctx->rcu_head, free_ctx); in put_ctx()
1285 * - perf_event_exit_task_context() [ child , 0 ]
1289 * - perf_event_init_context() [ parent, 0 ]
1297 * While it appears there is an obvious deadlock here -- the parent and child
1299 * life-time rules separate them. That is an exiting task cannot fork, and a
1302 * But remember that that these are parent<->child context relations, and
1308 * the ctx parent<->child relation, and perf_pmu_migrate_context() is only
1319 * means that only external vectors (perf_fops, prctl) can perturb the event
1323 * However; because event->ctx can change while we're waiting to acquire
1324 * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
1339 * cpuctx->mutex / perf_event_context::mutex
1348 ctx = READ_ONCE(event->ctx); in perf_event_ctx_lock_nested()
1349 if (!refcount_inc_not_zero(&ctx->refcount)) { in perf_event_ctx_lock_nested()
1355 mutex_lock_nested(&ctx->mutex, nesting); in perf_event_ctx_lock_nested()
1356 if (event->ctx != ctx) { in perf_event_ctx_lock_nested()
1357 mutex_unlock(&ctx->mutex); in perf_event_ctx_lock_nested()
1374 mutex_unlock(&ctx->mutex); in perf_event_ctx_unlock()
1379 * This must be done under the ctx->lock, such as to serialize against
1381 * calling scheduler related locks and ctx->lock nests inside those.
1386 struct perf_event_context *parent_ctx = ctx->parent_ctx; in unclone_ctx()
1388 lockdep_assert_held(&ctx->lock); in unclone_ctx()
1391 ctx->parent_ctx = NULL; in unclone_ctx()
1392 ctx->generation++; in unclone_ctx()
1404 if (event->parent) in perf_event_pid_type()
1405 event = event->parent; in perf_event_pid_type()
1407 nr = __task_pid_nr_ns(p, type, event->ns); in perf_event_pid_type()
1408 /* avoid -1 if it is idle thread or runs in another ns */ in perf_event_pid_type()
1410 nr = -1; in perf_event_pid_type()
1430 u64 id = event->id; in primary_event_id()
1432 if (event->parent) in primary_event_id()
1433 id = event->parent->id; in primary_event_id()
1453 * part of the read side critical section was irqs-enabled -- see in perf_lock_task_context()
1456 * Since ctx->lock nests under rq->lock we must ensure the entire read in perf_lock_task_context()
1461 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); in perf_lock_task_context()
1473 raw_spin_lock(&ctx->lock); in perf_lock_task_context()
1474 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { in perf_lock_task_context()
1475 raw_spin_unlock(&ctx->lock); in perf_lock_task_context()
1481 if (ctx->task == TASK_TOMBSTONE || in perf_lock_task_context()
1482 !refcount_inc_not_zero(&ctx->refcount)) { in perf_lock_task_context()
1483 raw_spin_unlock(&ctx->lock); in perf_lock_task_context()
1486 WARN_ON_ONCE(ctx->task != task); in perf_lock_task_context()
1508 ++ctx->pin_count; in perf_pin_task_context()
1509 raw_spin_unlock_irqrestore(&ctx->lock, flags); in perf_pin_task_context()
1518 raw_spin_lock_irqsave(&ctx->lock, flags); in perf_unpin_context()
1519 --ctx->pin_count; in perf_unpin_context()
1520 raw_spin_unlock_irqrestore(&ctx->lock, flags); in perf_unpin_context()
1530 ctx->time += now - ctx->timestamp; in update_context_time()
1531 ctx->timestamp = now; in update_context_time()
1536 struct perf_event_context *ctx = event->ctx; in perf_event_time()
1541 return ctx ? ctx->time : 0; in perf_event_time()
1546 struct perf_event_context *ctx = event->ctx; in get_event_type()
1549 lockdep_assert_held(&ctx->lock); in get_event_type()
1555 if (event->group_leader != event) in get_event_type()
1556 event = event->group_leader; in get_event_type()
1558 event_type = event->attr.pinned ? EVENT_PINNED : EVENT_FLEXIBLE; in get_event_type()
1559 if (!ctx->task) in get_event_type()
1570 RB_CLEAR_NODE(&event->group_node); in init_event_group()
1571 event->group_index = 0; in init_event_group()
1581 if (event->attr.pinned) in get_event_groups()
1582 return &ctx->pinned_groups; in get_event_groups()
1584 return &ctx->flexible_groups; in get_event_groups()
1592 groups->tree = RB_ROOT; in perf_event_groups_init()
1593 groups->index = 0; in perf_event_groups_init()
1605 if (left->cpu < right->cpu) in perf_event_groups_less()
1607 if (left->cpu > right->cpu) in perf_event_groups_less()
1611 if (left->cgrp != right->cgrp) { in perf_event_groups_less()
1612 if (!left->cgrp || !left->cgrp->css.cgroup) { in perf_event_groups_less()
1619 if (!right->cgrp || !right->cgrp->css.cgroup) { in perf_event_groups_less()
1627 if (left->cgrp->css.cgroup->kn->id < right->cgrp->css.cgroup->kn->id) in perf_event_groups_less()
1634 if (left->group_index < right->group_index) in perf_event_groups_less()
1636 if (left->group_index > right->group_index) in perf_event_groups_less()
1643 * Insert @event into @groups' tree; using {@event->cpu, ++@groups->index} for
1655 event->group_index = ++groups->index; in perf_event_groups_insert()
1657 node = &groups->tree.rb_node; in perf_event_groups_insert()
1665 node = &parent->rb_left; in perf_event_groups_insert()
1667 node = &parent->rb_right; in perf_event_groups_insert()
1670 rb_link_node(&event->group_node, parent, node); in perf_event_groups_insert()
1671 rb_insert_color(&event->group_node, &groups->tree); in perf_event_groups_insert()
1693 WARN_ON_ONCE(RB_EMPTY_NODE(&event->group_node) || in perf_event_groups_delete()
1694 RB_EMPTY_ROOT(&groups->tree)); in perf_event_groups_delete()
1696 rb_erase(&event->group_node, &groups->tree); in perf_event_groups_delete()
1720 struct rb_node *node = groups->tree.rb_node; in perf_event_groups_first()
1725 cgrp_id = cgrp->kn->id; in perf_event_groups_first()
1731 if (cpu < node_event->cpu) { in perf_event_groups_first()
1732 node = node->rb_left; in perf_event_groups_first()
1735 if (cpu > node_event->cpu) { in perf_event_groups_first()
1736 node = node->rb_right; in perf_event_groups_first()
1741 if (node_event->cgrp && node_event->cgrp->css.cgroup) in perf_event_groups_first()
1742 node_cgrp_id = node_event->cgrp->css.cgroup->kn->id; in perf_event_groups_first()
1745 node = node->rb_left; in perf_event_groups_first()
1749 node = node->rb_right; in perf_event_groups_first()
1754 node = node->rb_left; in perf_event_groups_first()
1772 next = rb_entry_safe(rb_next(&event->group_node), typeof(*event), group_node); in perf_event_groups_next()
1773 if (next == NULL || next->cpu != event->cpu) in perf_event_groups_next()
1777 if (event->cgrp && event->cgrp->css.cgroup) in perf_event_groups_next()
1778 curr_cgrp_id = event->cgrp->css.cgroup->kn->id; in perf_event_groups_next()
1780 if (next->cgrp && next->cgrp->css.cgroup) in perf_event_groups_next()
1781 next_cgrp_id = next->cgrp->css.cgroup->kn->id; in perf_event_groups_next()
1793 for (event = rb_entry_safe(rb_first(&((groups)->tree)), \
1795 event = rb_entry_safe(rb_next(&event->group_node), \
1800 * Must be called with ctx->mutex and ctx->lock held.
1805 lockdep_assert_held(&ctx->lock); in list_add_event()
1807 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); in list_add_event()
1808 event->attach_state |= PERF_ATTACH_CONTEXT; in list_add_event()
1810 event->tstamp = perf_event_time(event); in list_add_event()
1817 if (event->group_leader == event) { in list_add_event()
1818 event->group_caps = event->event_caps; in list_add_event()
1822 list_add_rcu(&event->event_entry, &ctx->event_list); in list_add_event()
1823 ctx->nr_events++; in list_add_event()
1824 if (event->attr.inherit_stat) in list_add_event()
1825 ctx->nr_stat++; in list_add_event()
1827 if (event->state > PERF_EVENT_STATE_OFF) in list_add_event()
1830 ctx->generation++; in list_add_event()
1838 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : in perf_event__state_init()
1848 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) in __perf_event_read_size()
1851 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) in __perf_event_read_size()
1854 if (event->attr.read_format & PERF_FORMAT_ID) in __perf_event_read_size()
1857 if (event->attr.read_format & PERF_FORMAT_GROUP) { in __perf_event_read_size()
1863 event->read_size = size; in __perf_event_read_size()
1872 size += sizeof(data->ip); in __perf_event_header_size()
1875 size += sizeof(data->addr); in __perf_event_header_size()
1878 size += sizeof(data->period); in __perf_event_header_size()
1881 size += sizeof(data->weight); in __perf_event_header_size()
1884 size += event->read_size; in __perf_event_header_size()
1887 size += sizeof(data->data_src.val); in __perf_event_header_size()
1890 size += sizeof(data->txn); in __perf_event_header_size()
1893 size += sizeof(data->phys_addr); in __perf_event_header_size()
1896 size += sizeof(data->cgroup); in __perf_event_header_size()
1898 event->header_size = size; in __perf_event_header_size()
1908 event->group_leader->nr_siblings); in perf_event__header_size()
1909 __perf_event_header_size(event, event->attr.sample_type); in perf_event__header_size()
1915 u64 sample_type = event->attr.sample_type; in perf_event__id_header_size()
1919 size += sizeof(data->tid_entry); in perf_event__id_header_size()
1922 size += sizeof(data->time); in perf_event__id_header_size()
1925 size += sizeof(data->id); in perf_event__id_header_size()
1928 size += sizeof(data->id); in perf_event__id_header_size()
1931 size += sizeof(data->stream_id); in perf_event__id_header_size()
1934 size += sizeof(data->cpu_entry); in perf_event__id_header_size()
1936 event->id_header_size = size; in perf_event__id_header_size()
1942 * The values computed here will be over-written when we actually in perf_event_validate_size()
1945 __perf_event_read_size(event, event->group_leader->nr_siblings + 1); in perf_event_validate_size()
1946 __perf_event_header_size(event, event->attr.sample_type & ~PERF_SAMPLE_READ); in perf_event_validate_size()
1953 if (event->read_size + event->header_size + in perf_event_validate_size()
1954 event->id_header_size + sizeof(struct perf_event_header) >= 16*1024) in perf_event_validate_size()
1962 struct perf_event *group_leader = event->group_leader, *pos; in perf_group_attach()
1964 lockdep_assert_held(&event->ctx->lock); in perf_group_attach()
1969 if (event->attach_state & PERF_ATTACH_GROUP) in perf_group_attach()
1972 event->attach_state |= PERF_ATTACH_GROUP; in perf_group_attach()
1977 WARN_ON_ONCE(group_leader->ctx != event->ctx); in perf_group_attach()
1979 group_leader->group_caps &= event->event_caps; in perf_group_attach()
1981 list_add_tail(&event->sibling_list, &group_leader->sibling_list); in perf_group_attach()
1982 group_leader->nr_siblings++; in perf_group_attach()
1992 * Must be called with ctx->mutex and ctx->lock held.
1997 WARN_ON_ONCE(event->ctx != ctx); in list_del_event()
1998 lockdep_assert_held(&ctx->lock); in list_del_event()
2001 * We can have double detach due to exit/hot-unplug + close. in list_del_event()
2003 if (!(event->attach_state & PERF_ATTACH_CONTEXT)) in list_del_event()
2006 event->attach_state &= ~PERF_ATTACH_CONTEXT; in list_del_event()
2008 ctx->nr_events--; in list_del_event()
2009 if (event->attr.inherit_stat) in list_del_event()
2010 ctx->nr_stat--; in list_del_event()
2012 list_del_rcu(&event->event_entry); in list_del_event()
2014 if (event->group_leader == event) in list_del_event()
2021 * of error state is by explicit re-enabling in list_del_event()
2024 if (event->state > PERF_EVENT_STATE_OFF) { in list_del_event()
2029 ctx->generation++; in list_del_event()
2038 if (!event->pmu->aux_output_match) in perf_aux_output_match()
2041 return event->pmu->aux_output_match(aux_event); in perf_aux_output_match()
2051 struct perf_event_context *ctx = event->ctx; in perf_put_aux_event()
2058 if (event->aux_event) { in perf_put_aux_event()
2059 iter = event->aux_event; in perf_put_aux_event()
2060 event->aux_event = NULL; in perf_put_aux_event()
2069 for_each_sibling_event(iter, event->group_leader) { in perf_put_aux_event()
2070 if (iter->aux_event != event) in perf_put_aux_event()
2073 iter->aux_event = NULL; in perf_put_aux_event()
2088 return !!event->attr.aux_output || !!event->attr.aux_sample_size; in perf_need_aux_event()
2106 if (event->attr.aux_output && event->attr.aux_sample_size) in perf_get_aux_event()
2109 if (event->attr.aux_output && in perf_get_aux_event()
2113 if (event->attr.aux_sample_size && !group_leader->pmu->snapshot_aux) in perf_get_aux_event()
2116 if (!atomic_long_inc_not_zero(&group_leader->refcount)) in perf_get_aux_event()
2125 event->aux_event = group_leader; in perf_get_aux_event()
2132 struct perf_event_context *ctx = event->ctx; in get_event_list()
2133 return event->attr.pinned ? &ctx->pinned_active : &ctx->flexible_active; in get_event_list()
2144 struct perf_event_context *ctx = event->ctx; in perf_remove_sibling_event()
2153 struct perf_event *leader = event->group_leader; in perf_group_detach()
2155 struct perf_event_context *ctx = event->ctx; in perf_group_detach()
2157 lockdep_assert_held(&ctx->lock); in perf_group_detach()
2160 * We can have double detach due to exit/hot-unplug + close. in perf_group_detach()
2162 if (!(event->attach_state & PERF_ATTACH_GROUP)) in perf_group_detach()
2165 event->attach_state &= ~PERF_ATTACH_GROUP; in perf_group_detach()
2173 list_del_init(&event->sibling_list); in perf_group_detach()
2174 event->group_leader->nr_siblings--; in perf_group_detach()
2183 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, sibling_list) { in perf_group_detach()
2185 if (sibling->event_caps & PERF_EV_CAP_SIBLING) in perf_group_detach()
2188 sibling->group_leader = sibling; in perf_group_detach()
2189 list_del_init(&sibling->sibling_list); in perf_group_detach()
2192 sibling->group_caps = event->group_caps; in perf_group_detach()
2194 if (!RB_EMPTY_NODE(&event->group_node)) { in perf_group_detach()
2195 add_event_to_groups(sibling, event->ctx); in perf_group_detach()
2197 if (sibling->state == PERF_EVENT_STATE_ACTIVE) in perf_group_detach()
2198 list_add_tail(&sibling->active_list, get_event_list(sibling)); in perf_group_detach()
2201 WARN_ON_ONCE(sibling->ctx != event->ctx); in perf_group_detach()
2213 return event->state == PERF_EVENT_STATE_DEAD; in is_orphaned_event()
2218 struct pmu *pmu = event->pmu; in __pmu_filter_match()
2219 return pmu->filter_match ? pmu->filter_match(event) : 1; in __pmu_filter_match()
2224 * PMU-specific filtering. An event group can consist of HW and SW events,
2246 return (event->cpu == -1 || event->cpu == smp_processor_id()) && in event_filter_match()
2257 WARN_ON_ONCE(event->ctx != ctx); in event_sched_out()
2258 lockdep_assert_held(&ctx->lock); in event_sched_out()
2260 if (event->state != PERF_EVENT_STATE_ACTIVE) in event_sched_out()
2268 list_del_init(&event->active_list); in event_sched_out()
2270 perf_pmu_disable(event->pmu); in event_sched_out()
2272 event->pmu->del(event, 0); in event_sched_out()
2273 event->oncpu = -1; in event_sched_out()
2275 if (READ_ONCE(event->pending_disable) >= 0) { in event_sched_out()
2276 WRITE_ONCE(event->pending_disable, -1); in event_sched_out()
2283 cpuctx->active_oncpu--; in event_sched_out()
2284 if (!--ctx->nr_active) in event_sched_out()
2286 if (event->attr.freq && event->attr.sample_freq) in event_sched_out()
2287 ctx->nr_freq--; in event_sched_out()
2288 if (event->attr.exclusive || !cpuctx->active_oncpu) in event_sched_out()
2289 cpuctx->exclusive = 0; in event_sched_out()
2291 perf_pmu_enable(event->pmu); in event_sched_out()
2301 if (group_event->state != PERF_EVENT_STATE_ACTIVE) in group_sched_out()
2304 perf_pmu_disable(ctx->pmu); in group_sched_out()
2314 perf_pmu_enable(ctx->pmu); in group_sched_out()
2333 if (ctx->is_active & EVENT_TIME) { in __perf_remove_from_context()
2343 if (!ctx->nr_events && ctx->is_active) { in __perf_remove_from_context()
2344 ctx->is_active = 0; in __perf_remove_from_context()
2345 ctx->rotate_necessary = 0; in __perf_remove_from_context()
2346 if (ctx->task) { in __perf_remove_from_context()
2347 WARN_ON_ONCE(cpuctx->task_ctx != ctx); in __perf_remove_from_context()
2348 cpuctx->task_ctx = NULL; in __perf_remove_from_context()
2356 * If event->ctx is a cloned context, callers must make sure that
2357 * every task struct that event->ctx->task could possibly point to
2359 * that only calls us on the top-level context, which can't be a clone.
2365 struct perf_event_context *ctx = event->ctx; in perf_remove_from_context()
2367 lockdep_assert_held(&ctx->mutex); in perf_remove_from_context()
2372 * The above event_function_call() can NO-OP when it hits in perf_remove_from_context()
2375 * might still be in-tact. in perf_remove_from_context()
2377 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); in perf_remove_from_context()
2379 (event->attach_state & PERF_ATTACH_GROUP)) { in perf_remove_from_context()
2384 raw_spin_lock_irq(&ctx->lock); in perf_remove_from_context()
2386 raw_spin_unlock_irq(&ctx->lock); in perf_remove_from_context()
2398 if (event->state < PERF_EVENT_STATE_INACTIVE) in __perf_event_disable()
2401 if (ctx->is_active & EVENT_TIME) { in __perf_event_disable()
2406 if (event == event->group_leader) in __perf_event_disable()
2418 * If event->ctx is a cloned context, callers must make sure that
2419 * every task struct that event->ctx->task could possibly point to
2422 * hold the top-level event's child_mutex, so any descendant that
2425 * When called from perf_pending_event it's OK because event->ctx
2431 struct perf_event_context *ctx = event->ctx; in _perf_event_disable()
2433 raw_spin_lock_irq(&ctx->lock); in _perf_event_disable()
2434 if (event->state <= PERF_EVENT_STATE_OFF) { in _perf_event_disable()
2435 raw_spin_unlock_irq(&ctx->lock); in _perf_event_disable()
2438 raw_spin_unlock_irq(&ctx->lock); in _perf_event_disable()
2464 WRITE_ONCE(event->pending_disable, smp_processor_id()); in perf_event_disable_inatomic()
2466 irq_work_queue(&event->pending); in perf_event_disable_inatomic()
2481 * tstamp - ctx->timestamp in perf_set_shadow_time()
2483 * tstamp - cgrp->timestamp. in perf_set_shadow_time()
2487 * - event is guaranteed scheduled in in perf_set_shadow_time()
2488 * - no scheduled out in between in perf_set_shadow_time()
2489 * - thus the timestamp would be the same in perf_set_shadow_time()
2498 perf_cgroup_set_shadow_time(event, event->tstamp); in perf_set_shadow_time()
2500 event->shadow_ctx_time = event->tstamp - ctx->timestamp; in perf_set_shadow_time()
2515 WARN_ON_ONCE(event->ctx != ctx); in event_sched_in()
2517 lockdep_assert_held(&ctx->lock); in event_sched_in()
2519 if (event->state <= PERF_EVENT_STATE_OFF) in event_sched_in()
2522 WRITE_ONCE(event->oncpu, smp_processor_id()); in event_sched_in()
2526 * ->oncpu if it sees ACTIVE. in event_sched_in()
2536 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { in event_sched_in()
2538 event->hw.interrupts = 0; in event_sched_in()
2541 perf_pmu_disable(event->pmu); in event_sched_in()
2547 if (event->pmu->add(event, PERF_EF_START)) { in event_sched_in()
2549 event->oncpu = -1; in event_sched_in()
2550 ret = -EAGAIN; in event_sched_in()
2555 cpuctx->active_oncpu++; in event_sched_in()
2556 if (!ctx->nr_active++) in event_sched_in()
2558 if (event->attr.freq && event->attr.sample_freq) in event_sched_in()
2559 ctx->nr_freq++; in event_sched_in()
2561 if (event->attr.exclusive) in event_sched_in()
2562 cpuctx->exclusive = 1; in event_sched_in()
2565 perf_pmu_enable(event->pmu); in event_sched_in()
2576 struct pmu *pmu = ctx->pmu; in group_sched_in()
2578 if (group_event->state == PERF_EVENT_STATE_OFF) in group_sched_in()
2581 pmu->start_txn(pmu, PERF_PMU_TXN_ADD); in group_sched_in()
2596 if (!pmu->commit_txn(pmu)) in group_sched_in()
2614 pmu->cancel_txn(pmu); in group_sched_in()
2615 return -EAGAIN; in group_sched_in()
2628 if (event->group_caps & PERF_EV_CAP_SOFTWARE) in group_can_go_on()
2634 if (cpuctx->exclusive) in group_can_go_on()
2640 if (event->attr.exclusive && !list_empty(get_event_list(event))) in group_can_go_on()
2669 if (!cpuctx->task_ctx) in task_ctx_sched_out()
2672 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) in task_ctx_sched_out()
2692 * - CPU pinned (EVENT_CPU | EVENT_PINNED)
2693 * - task pinned (EVENT_PINNED)
2694 * - CPU flexible (EVENT_CPU | EVENT_FLEXIBLE)
2695 * - task flexible (EVENT_FLEXIBLE).
2721 perf_pmu_disable(cpuctx->ctx.pmu); in ctx_resched()
2728 * - EVENT_CPU: schedule out corresponding groups; in ctx_resched()
2729 * - EVENT_PINNED task events: schedule out EVENT_FLEXIBLE groups; in ctx_resched()
2730 * - otherwise, do nothing more. in ctx_resched()
2738 perf_pmu_enable(cpuctx->ctx.pmu); in ctx_resched()
2743 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); in perf_pmu_resched()
2744 struct perf_event_context *task_ctx = cpuctx->task_ctx; in perf_pmu_resched()
2755 * things like ctx->is_active and cpuctx->task_ctx are set.
2760 struct perf_event_context *ctx = event->ctx; in __perf_install_in_context()
2762 struct perf_event_context *task_ctx = cpuctx->task_ctx; in __perf_install_in_context()
2766 raw_spin_lock(&cpuctx->ctx.lock); in __perf_install_in_context()
2767 if (ctx->task) { in __perf_install_in_context()
2768 raw_spin_lock(&ctx->lock); in __perf_install_in_context()
2771 reprogram = (ctx->task == current); in __perf_install_in_context()
2777 * If its not running, we don't care, ctx->lock will in __perf_install_in_context()
2780 if (task_curr(ctx->task) && !reprogram) { in __perf_install_in_context()
2781 ret = -ESRCH; in __perf_install_in_context()
2785 WARN_ON_ONCE(reprogram && cpuctx->task_ctx && cpuctx->task_ctx != ctx); in __perf_install_in_context()
2787 raw_spin_lock(&task_ctx->lock); in __perf_install_in_context()
2791 if (event->state > PERF_EVENT_STATE_OFF && is_cgroup_event(event)) { in __perf_install_in_context()
2797 reprogram = cgroup_is_descendant(cgrp->css.cgroup, in __perf_install_in_context()
2798 event->cgrp->css.cgroup); in __perf_install_in_context()
2829 struct task_struct *task = READ_ONCE(ctx->task); in perf_install_in_context()
2831 lockdep_assert_held(&ctx->mutex); in perf_install_in_context()
2835 if (event->cpu != -1) in perf_install_in_context()
2836 event->cpu = cpu; in perf_install_in_context()
2839 * Ensures that if we can observe event->ctx, both the event and ctx in perf_install_in_context()
2842 smp_store_release(&event->ctx, ctx); in perf_install_in_context()
2846 * without IPI. Except when this is the first event for the context, in in perf_install_in_context()
2847 * that case we need the magic of the IPI to set ctx->is_active. in perf_install_in_context()
2850 * event will issue the IPI and reprogram the hardware. in perf_install_in_context()
2852 if (__perf_effective_state(event) == PERF_EVENT_STATE_OFF && ctx->nr_events) { in perf_install_in_context()
2853 raw_spin_lock_irq(&ctx->lock); in perf_install_in_context()
2854 if (ctx->task == TASK_TOMBSTONE) { in perf_install_in_context()
2855 raw_spin_unlock_irq(&ctx->lock); in perf_install_in_context()
2859 raw_spin_unlock_irq(&ctx->lock); in perf_install_in_context()
2875 * Installing events is tricky because we cannot rely on ctx->is_active in perf_install_in_context()
2876 * to be set in case this is the nr_events 0 -> 1 transition. in perf_install_in_context()
2890 * our task->perf_event_ctxp[] store, such that it will in fact take in perf_install_in_context()
2893 * We do this by task_function_call(), if the IPI fails to hit the task in perf_install_in_context()
2899 * This smp_mb() orders the task->perf_event_ctxp[] store with the in perf_install_in_context()
2900 * task_cpu() load, such that if the IPI then does not find the task in perf_install_in_context()
2909 raw_spin_lock_irq(&ctx->lock); in perf_install_in_context()
2910 task = ctx->task; in perf_install_in_context()
2914 * cannot happen), and we hold ctx->mutex, which serializes us in perf_install_in_context()
2917 raw_spin_unlock_irq(&ctx->lock); in perf_install_in_context()
2921 * If the task is not running, ctx->lock will avoid it becoming so, in perf_install_in_context()
2925 raw_spin_unlock_irq(&ctx->lock); in perf_install_in_context()
2929 raw_spin_unlock_irq(&ctx->lock); in perf_install_in_context()
2940 struct perf_event *leader = event->group_leader; in __perf_event_enable()
2943 if (event->state >= PERF_EVENT_STATE_INACTIVE || in __perf_event_enable()
2944 event->state <= PERF_EVENT_STATE_ERROR) in __perf_event_enable()
2947 if (ctx->is_active) in __perf_event_enable()
2953 if (!ctx->is_active) in __perf_event_enable()
2965 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) { in __perf_event_enable()
2970 task_ctx = cpuctx->task_ctx; in __perf_event_enable()
2971 if (ctx->task) in __perf_event_enable()
2980 * If event->ctx is a cloned context, callers must make sure that
2981 * every task struct that event->ctx->task could possibly point to
2988 struct perf_event_context *ctx = event->ctx; in _perf_event_enable()
2990 raw_spin_lock_irq(&ctx->lock); in _perf_event_enable()
2991 if (event->state >= PERF_EVENT_STATE_INACTIVE || in _perf_event_enable()
2992 event->state < PERF_EVENT_STATE_ERROR) { in _perf_event_enable()
2994 raw_spin_unlock_irq(&ctx->lock); in _perf_event_enable()
3003 * been scheduled away before the cross-call arrived. in _perf_event_enable()
3005 if (event->state == PERF_EVENT_STATE_ERROR) { in _perf_event_enable()
3009 if (event->event_caps & PERF_EV_CAP_SIBLING && in _perf_event_enable()
3010 event->group_leader == event) in _perf_event_enable()
3013 event->state = PERF_EVENT_STATE_OFF; in _perf_event_enable()
3015 raw_spin_unlock_irq(&ctx->lock); in _perf_event_enable()
3041 struct perf_event *event = sd->event; in __perf_event_stop()
3044 if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE) in __perf_event_stop()
3054 if (READ_ONCE(event->oncpu) != smp_processor_id()) in __perf_event_stop()
3055 return -EAGAIN; in __perf_event_stop()
3057 event->pmu->stop(event, PERF_EF_UPDATE); in __perf_event_stop()
3065 * Since this is happening on an event-local CPU, no trace is lost in __perf_event_stop()
3068 if (sd->restart) in __perf_event_stop()
3069 event->pmu->start(event, 0); in __perf_event_stop()
3083 if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE) in perf_event_stop()
3091 * inactive here (event->oncpu==-1), there's nothing more to do; in perf_event_stop()
3092 * fall through with ret==-ENXIO. in perf_event_stop()
3094 ret = cpu_function_call(READ_ONCE(event->oncpu), in perf_event_stop()
3096 } while (ret == -EAGAIN); in perf_event_stop()
3115 * pre-existing mappings, called once when new filters arrive via SET_FILTER
3130 raw_spin_lock(&ifh->lock); in perf_event_addr_filters_sync()
3131 if (event->addr_filters_gen != event->hw.addr_filters_gen) { in perf_event_addr_filters_sync()
3132 event->pmu->addr_filters_sync(event); in perf_event_addr_filters_sync()
3133 event->hw.addr_filters_gen = event->addr_filters_gen; in perf_event_addr_filters_sync()
3135 raw_spin_unlock(&ifh->lock); in perf_event_addr_filters_sync()
3144 if (event->attr.inherit || !is_sampling_event(event)) in _perf_event_refresh()
3145 return -EINVAL; in _perf_event_refresh()
3147 atomic_add(refresh, &event->event_limit); in _perf_event_refresh()
3178 if (!bp->attr.disabled) in perf_event_modify_breakpoint()
3187 if (event->attr.type != attr->type) in perf_event_modify_attr()
3188 return -EINVAL; in perf_event_modify_attr()
3190 switch (event->attr.type) { in perf_event_modify_attr()
3195 return -EOPNOTSUPP; in perf_event_modify_attr()
3204 int is_active = ctx->is_active; in ctx_sched_out()
3206 lockdep_assert_held(&ctx->lock); in ctx_sched_out()
3208 if (likely(!ctx->nr_events)) { in ctx_sched_out()
3212 WARN_ON_ONCE(ctx->is_active); in ctx_sched_out()
3213 if (ctx->task) in ctx_sched_out()
3214 WARN_ON_ONCE(cpuctx->task_ctx); in ctx_sched_out()
3218 ctx->is_active &= ~event_type; in ctx_sched_out()
3219 if (!(ctx->is_active & EVENT_ALL)) in ctx_sched_out()
3220 ctx->is_active = 0; in ctx_sched_out()
3222 if (ctx->task) { in ctx_sched_out()
3223 WARN_ON_ONCE(cpuctx->task_ctx != ctx); in ctx_sched_out()
3224 if (!ctx->is_active) in ctx_sched_out()
3225 cpuctx->task_ctx = NULL; in ctx_sched_out()
3244 is_active ^= ctx->is_active; /* changed bits */ in ctx_sched_out()
3246 if (!ctx->nr_active || !(is_active & EVENT_ALL)) in ctx_sched_out()
3249 perf_pmu_disable(ctx->pmu); in ctx_sched_out()
3251 list_for_each_entry_safe(event, tmp, &ctx->pinned_active, active_list) in ctx_sched_out()
3256 list_for_each_entry_safe(event, tmp, &ctx->flexible_active, active_list) in ctx_sched_out()
3264 ctx->rotate_necessary = 0; in ctx_sched_out()
3266 perf_pmu_enable(ctx->pmu); in ctx_sched_out()
3280 lockdep_assert_held(&ctx1->lock); in context_equiv()
3281 lockdep_assert_held(&ctx2->lock); in context_equiv()
3284 if (ctx1->pin_count || ctx2->pin_count) in context_equiv()
3288 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen) in context_equiv()
3292 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation) in context_equiv()
3299 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx && in context_equiv()
3300 ctx1->parent_gen == ctx2->parent_gen) in context_equiv()
3312 if (!event->attr.inherit_stat) in __perf_event_sync_stat()
3322 if (event->state == PERF_EVENT_STATE_ACTIVE) in __perf_event_sync_stat()
3323 event->pmu->read(event); in __perf_event_sync_stat()
3328 * In order to keep per-task stats reliable we need to flip the event in __perf_event_sync_stat()
3331 value = local64_read(&next_event->count); in __perf_event_sync_stat()
3332 value = local64_xchg(&event->count, value); in __perf_event_sync_stat()
3333 local64_set(&next_event->count, value); in __perf_event_sync_stat()
3335 swap(event->total_time_enabled, next_event->total_time_enabled); in __perf_event_sync_stat()
3336 swap(event->total_time_running, next_event->total_time_running); in __perf_event_sync_stat()
3350 if (!ctx->nr_stat) in perf_event_sync_stat()
3355 event = list_first_entry(&ctx->event_list, in perf_event_sync_stat()
3358 next_event = list_first_entry(&next_ctx->event_list, in perf_event_sync_stat()
3361 while (&event->event_entry != &ctx->event_list && in perf_event_sync_stat()
3362 &next_event->event_entry != &next_ctx->event_list) { in perf_event_sync_stat()
3374 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; in perf_event_context_sched_out()
3384 pmu = ctx->pmu; in perf_event_context_sched_out()
3386 if (!cpuctx->task_ctx) in perf_event_context_sched_out()
3390 next_ctx = next->perf_event_ctxp[ctxn]; in perf_event_context_sched_out()
3394 parent = rcu_dereference(ctx->parent_ctx); in perf_event_context_sched_out()
3395 next_parent = rcu_dereference(next_ctx->parent_ctx); in perf_event_context_sched_out()
3406 * lock (including re-checking that neither has been in perf_event_context_sched_out()
3411 raw_spin_lock(&ctx->lock); in perf_event_context_sched_out()
3412 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); in perf_event_context_sched_out()
3415 WRITE_ONCE(ctx->task, next); in perf_event_context_sched_out()
3416 WRITE_ONCE(next_ctx->task, task); in perf_event_context_sched_out()
3420 if (cpuctx->sched_cb_usage && pmu->sched_task) in perf_event_context_sched_out()
3421 pmu->sched_task(ctx, false); in perf_event_context_sched_out()
3429 if (pmu->swap_task_ctx) in perf_event_context_sched_out()
3430 pmu->swap_task_ctx(ctx, next_ctx); in perf_event_context_sched_out()
3432 swap(ctx->task_ctx_data, next_ctx->task_ctx_data); in perf_event_context_sched_out()
3439 * ctx->task and ctx->task_ctx_data are immaterial in perf_event_context_sched_out()
3441 * ctx->lock which we're now holding. in perf_event_context_sched_out()
3443 RCU_INIT_POINTER(task->perf_event_ctxp[ctxn], next_ctx); in perf_event_context_sched_out()
3444 RCU_INIT_POINTER(next->perf_event_ctxp[ctxn], ctx); in perf_event_context_sched_out()
3450 raw_spin_unlock(&next_ctx->lock); in perf_event_context_sched_out()
3451 raw_spin_unlock(&ctx->lock); in perf_event_context_sched_out()
3457 raw_spin_lock(&ctx->lock); in perf_event_context_sched_out()
3460 if (cpuctx->sched_cb_usage && pmu->sched_task) in perf_event_context_sched_out()
3461 pmu->sched_task(ctx, false); in perf_event_context_sched_out()
3465 raw_spin_unlock(&ctx->lock); in perf_event_context_sched_out()
3471 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); in perf_sched_cb_dec()
3473 --cpuctx->sched_cb_usage; in perf_sched_cb_dec()
3479 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); in perf_sched_cb_inc()
3481 cpuctx->sched_cb_usage++; in perf_sched_cb_inc()
3488 * This callback is relevant even to per-cpu events; for example multi event
3496 pmu = cpuctx->ctx.pmu; /* software PMUs will not have sched_task */ in __perf_pmu_sched_task()
3498 if (WARN_ON_ONCE(!pmu->sched_task)) in __perf_pmu_sched_task()
3501 perf_ctx_lock(cpuctx, cpuctx->task_ctx); in __perf_pmu_sched_task()
3504 pmu->sched_task(cpuctx->task_ctx, sched_in); in __perf_pmu_sched_task()
3507 perf_ctx_unlock(cpuctx, cpuctx->task_ctx); in __perf_pmu_sched_task()
3520 * We stop each event and update the event value in event->count.
3541 * cgroup event are system-wide mode only in __perf_event_task_sched_out()
3553 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); in cpu_ctx_sched_out()
3561 return le->group_index < re->group_index; in perf_less_group_idx()
3579 struct perf_event **itrs = heap->data; in __heap_add()
3582 itrs[heap->nr] = event; in __heap_add()
3583 heap->nr++; in __heap_add()
3603 .data = cpuctx->heap, in visit_groups_merge()
3605 .size = cpuctx->heap_size, in visit_groups_merge()
3608 lockdep_assert_held(&cpuctx->ctx.lock); in visit_groups_merge()
3611 if (cpuctx->cgrp) in visit_groups_merge()
3612 css = &cpuctx->cgrp->css; in visit_groups_merge()
3621 __heap_add(&event_heap, perf_event_groups_first(groups, -1, NULL)); in visit_groups_merge()
3628 for (; css; css = css->parent) in visit_groups_merge()
3629 __heap_add(&event_heap, perf_event_groups_first(groups, cpu, css->cgroup)); in visit_groups_merge()
3651 struct perf_event_context *ctx = event->ctx; in merge_sched_in()
3655 if (event->state <= PERF_EVENT_STATE_OFF) in merge_sched_in()
3663 list_add_tail(&event->active_list, get_event_list(event)); in merge_sched_in()
3666 if (event->state == PERF_EVENT_STATE_INACTIVE) { in merge_sched_in()
3667 if (event->attr.pinned) { in merge_sched_in()
3673 ctx->rotate_necessary = 1; in merge_sched_in()
3686 if (ctx != &cpuctx->ctx) in ctx_pinned_sched_in()
3689 visit_groups_merge(cpuctx, &ctx->pinned_groups, in ctx_pinned_sched_in()
3700 if (ctx != &cpuctx->ctx) in ctx_flexible_sched_in()
3703 visit_groups_merge(cpuctx, &ctx->flexible_groups, in ctx_flexible_sched_in()
3714 int is_active = ctx->is_active; in ctx_sched_in()
3717 lockdep_assert_held(&ctx->lock); in ctx_sched_in()
3719 if (likely(!ctx->nr_events)) in ctx_sched_in()
3722 ctx->is_active |= (event_type | EVENT_TIME); in ctx_sched_in()
3723 if (ctx->task) { in ctx_sched_in()
3725 cpuctx->task_ctx = ctx; in ctx_sched_in()
3727 WARN_ON_ONCE(cpuctx->task_ctx != ctx); in ctx_sched_in()
3730 is_active ^= ctx->is_active; /* changed bits */ in ctx_sched_in()
3735 ctx->timestamp = now; in ctx_sched_in()
3755 struct perf_event_context *ctx = &cpuctx->ctx; in cpu_ctx_sched_in()
3764 struct pmu *pmu = ctx->pmu; in perf_event_context_sched_in()
3767 if (cpuctx->task_ctx == ctx) { in perf_event_context_sched_in()
3768 if (cpuctx->sched_cb_usage) in perf_event_context_sched_in()
3775 * We must check ctx->nr_events while holding ctx->lock, such in perf_event_context_sched_in()
3778 if (!ctx->nr_events) in perf_event_context_sched_in()
3790 if (!RB_EMPTY_ROOT(&ctx->pinned_groups.tree)) in perf_event_context_sched_in()
3794 if (cpuctx->sched_cb_usage && pmu->sched_task) in perf_event_context_sched_in()
3795 pmu->sched_task(cpuctx->task_ctx, true); in perf_event_context_sched_in()
3822 * to switch in PMU state; cgroup event are system-wide mode only. in __perf_event_task_sched_in()
3831 ctx = task->perf_event_ctxp[ctxn]; in __perf_event_task_sched_in()
3844 u64 frequency = event->attr.sample_freq; in perf_calculate_period()
3860 * period = ------------------- in perf_calculate_period()
3873 a##_fls--; \ in perf_calculate_period()
3876 b##_fls--; \ in perf_calculate_period()
3920 struct hw_perf_event *hwc = &event->hw; in perf_adjust_period()
3926 delta = (s64)(period - hwc->sample_period); in perf_adjust_period()
3929 sample_period = hwc->sample_period + delta; in perf_adjust_period()
3934 hwc->sample_period = sample_period; in perf_adjust_period()
3936 if (local64_read(&hwc->period_left) > 8*sample_period) { in perf_adjust_period()
3938 event->pmu->stop(event, PERF_EF_UPDATE); in perf_adjust_period()
3940 local64_set(&hwc->period_left, 0); in perf_adjust_period()
3943 event->pmu->start(event, PERF_EF_RELOAD); in perf_adjust_period()
3962 * - context have events in frequency mode (needs freq adjust) in perf_adjust_freq_unthr_context()
3963 * - there are events to unthrottle on this cpu in perf_adjust_freq_unthr_context()
3965 if (!(ctx->nr_freq || needs_unthr)) in perf_adjust_freq_unthr_context()
3968 raw_spin_lock(&ctx->lock); in perf_adjust_freq_unthr_context()
3969 perf_pmu_disable(ctx->pmu); in perf_adjust_freq_unthr_context()
3971 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { in perf_adjust_freq_unthr_context()
3972 if (event->state != PERF_EVENT_STATE_ACTIVE) in perf_adjust_freq_unthr_context()
3978 perf_pmu_disable(event->pmu); in perf_adjust_freq_unthr_context()
3980 hwc = &event->hw; in perf_adjust_freq_unthr_context()
3982 if (hwc->interrupts == MAX_INTERRUPTS) { in perf_adjust_freq_unthr_context()
3983 hwc->interrupts = 0; in perf_adjust_freq_unthr_context()
3985 event->pmu->start(event, 0); in perf_adjust_freq_unthr_context()
3988 if (!event->attr.freq || !event->attr.sample_freq) in perf_adjust_freq_unthr_context()
3992 * stop the event and update event->count in perf_adjust_freq_unthr_context()
3994 event->pmu->stop(event, PERF_EF_UPDATE); in perf_adjust_freq_unthr_context()
3996 now = local64_read(&event->count); in perf_adjust_freq_unthr_context()
3997 delta = now - hwc->freq_count_stamp; in perf_adjust_freq_unthr_context()
3998 hwc->freq_count_stamp = now; in perf_adjust_freq_unthr_context()
4010 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); in perf_adjust_freq_unthr_context()
4012 perf_pmu_enable(event->pmu); in perf_adjust_freq_unthr_context()
4015 perf_pmu_enable(ctx->pmu); in perf_adjust_freq_unthr_context()
4016 raw_spin_unlock(&ctx->lock); in perf_adjust_freq_unthr_context()
4025 * Rotate the first entry last of non-pinned groups. Rotation might be in rotate_ctx()
4028 if (ctx->rotate_disable) in rotate_ctx()
4031 perf_event_groups_delete(&ctx->flexible_groups, event); in rotate_ctx()
4032 perf_event_groups_insert(&ctx->flexible_groups, event); in rotate_ctx()
4042 event = list_first_entry_or_null(&ctx->flexible_active, in ctx_event_to_rotate()
4047 event = rb_entry_safe(rb_first(&ctx->flexible_groups.tree), in ctx_event_to_rotate()
4055 ctx->rotate_necessary = 0; in ctx_event_to_rotate()
4071 cpu_rotate = cpuctx->ctx.rotate_necessary; in perf_rotate_context()
4072 task_ctx = cpuctx->task_ctx; in perf_rotate_context()
4073 task_rotate = task_ctx ? task_ctx->rotate_necessary : 0; in perf_rotate_context()
4078 perf_ctx_lock(cpuctx, cpuctx->task_ctx); in perf_rotate_context()
4079 perf_pmu_disable(cpuctx->ctx.pmu); in perf_rotate_context()
4084 cpu_event = ctx_event_to_rotate(&cpuctx->ctx); in perf_rotate_context()
4098 rotate_ctx(&cpuctx->ctx, cpu_event); in perf_rotate_context()
4102 perf_pmu_enable(cpuctx->ctx.pmu); in perf_rotate_context()
4103 perf_ctx_unlock(cpuctx, cpuctx->task_ctx); in perf_rotate_context()
4127 if (!event->attr.enable_on_exec) in event_enable_on_exec()
4130 event->attr.enable_on_exec = 0; in event_enable_on_exec()
4131 if (event->state >= PERF_EVENT_STATE_INACTIVE) in event_enable_on_exec()
4140 * Enable all of a task's events that have been marked enable-on-exec.
4153 ctx = current->perf_event_ctxp[ctxn]; in perf_event_enable_on_exec()
4154 if (!ctx || !ctx->nr_events) in perf_event_enable_on_exec()
4160 list_for_each_entry(event, &ctx->event_list, event_entry) { in perf_event_enable_on_exec()
4193 if (event->group_caps & PERF_EV_CAP_READ_ACTIVE_PKG) { in __perf_event_read_cpu()
4212 struct perf_event *sub, *event = data->event; in __perf_event_read()
4213 struct perf_event_context *ctx = event->ctx; in __perf_event_read()
4215 struct pmu *pmu = event->pmu; in __perf_event_read()
4221 * event->count would have been updated to a recent sample in __perf_event_read()
4224 if (ctx->task && cpuctx->task_ctx != ctx) in __perf_event_read()
4227 raw_spin_lock(&ctx->lock); in __perf_event_read()
4228 if (ctx->is_active & EVENT_TIME) { in __perf_event_read()
4234 if (data->group) in __perf_event_read()
4237 if (event->state != PERF_EVENT_STATE_ACTIVE) in __perf_event_read()
4240 if (!data->group) { in __perf_event_read()
4241 pmu->read(event); in __perf_event_read()
4242 data->ret = 0; in __perf_event_read()
4246 pmu->start_txn(pmu, PERF_PMU_TXN_READ); in __perf_event_read()
4248 pmu->read(event); in __perf_event_read()
4251 if (sub->state == PERF_EVENT_STATE_ACTIVE) { in __perf_event_read()
4256 sub->pmu->read(sub); in __perf_event_read()
4260 data->ret = pmu->commit_txn(pmu); in __perf_event_read()
4263 raw_spin_unlock(&ctx->lock); in __perf_event_read()
4268 return local64_read(&event->count) + atomic64_read(&event->child_count); in perf_event_count()
4272 * NMI-safe method to read a local event, that is an event that
4274 * - either for the current task, or for this CPU
4275 * - does not have inherit set, for inherited task events
4277 * - must not have a pmu::count method
4295 if (event->attr.inherit) { in perf_event_read_local()
4296 ret = -EOPNOTSUPP; in perf_event_read_local()
4300 /* If this is a per-task event, it must be for current */ in perf_event_read_local()
4301 if ((event->attach_state & PERF_ATTACH_TASK) && in perf_event_read_local()
4302 event->hw.target != current) { in perf_event_read_local()
4303 ret = -EINVAL; in perf_event_read_local()
4307 /* If this is a per-CPU event, it must be for this CPU */ in perf_event_read_local()
4308 if (!(event->attach_state & PERF_ATTACH_TASK) && in perf_event_read_local()
4309 event->cpu != smp_processor_id()) { in perf_event_read_local()
4310 ret = -EINVAL; in perf_event_read_local()
4315 if (event->attr.pinned && event->oncpu != smp_processor_id()) { in perf_event_read_local()
4316 ret = -EBUSY; in perf_event_read_local()
4321 * If the event is currently on this CPU, its either a per-task event, in perf_event_read_local()
4323 * oncpu == -1). in perf_event_read_local()
4325 if (event->oncpu == smp_processor_id()) in perf_event_read_local()
4326 event->pmu->read(event); in perf_event_read_local()
4328 *value = local64_read(&event->count); in perf_event_read_local()
4330 u64 now = event->shadow_ctx_time + perf_clock(); in perf_event_read_local()
4347 enum perf_event_state state = READ_ONCE(event->state); in perf_event_read()
4359 * Orders the ->state and ->oncpu loads such that if we see in perf_event_read()
4360 * ACTIVE we must also see the right ->oncpu. in perf_event_read()
4366 event_cpu = READ_ONCE(event->oncpu); in perf_event_read()
4386 * Therefore, either way, we'll have an up-to-date event count in perf_event_read()
4394 struct perf_event_context *ctx = event->ctx; in perf_event_read()
4397 raw_spin_lock_irqsave(&ctx->lock, flags); in perf_event_read()
4398 state = event->state; in perf_event_read()
4400 raw_spin_unlock_irqrestore(&ctx->lock, flags); in perf_event_read()
4408 if (ctx->is_active & EVENT_TIME) { in perf_event_read()
4416 raw_spin_unlock_irqrestore(&ctx->lock, flags); in perf_event_read()
4427 raw_spin_lock_init(&ctx->lock); in __perf_event_init_context()
4428 mutex_init(&ctx->mutex); in __perf_event_init_context()
4429 INIT_LIST_HEAD(&ctx->active_ctx_list); in __perf_event_init_context()
4430 perf_event_groups_init(&ctx->pinned_groups); in __perf_event_init_context()
4431 perf_event_groups_init(&ctx->flexible_groups); in __perf_event_init_context()
4432 INIT_LIST_HEAD(&ctx->event_list); in __perf_event_init_context()
4433 INIT_LIST_HEAD(&ctx->pinned_active); in __perf_event_init_context()
4434 INIT_LIST_HEAD(&ctx->flexible_active); in __perf_event_init_context()
4435 refcount_set(&ctx->refcount, 1); in __perf_event_init_context()
4449 ctx->task = get_task_struct(task); in alloc_perf_context()
4450 ctx->pmu = pmu; in alloc_perf_context()
4470 return ERR_PTR(-ESRCH); in find_lively_task_by_vpid()
4487 int cpu = event->cpu; in find_get_context()
4491 err = perf_allow_cpu(&event->attr); in find_get_context()
4495 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); in find_get_context()
4496 ctx = &cpuctx->ctx; in find_get_context()
4498 ++ctx->pin_count; in find_get_context()
4503 err = -EINVAL; in find_get_context()
4504 ctxn = pmu->task_ctx_nr; in find_get_context()
4508 if (event->attach_state & PERF_ATTACH_TASK_DATA) { in find_get_context()
4511 err = -ENOMEM; in find_get_context()
4520 ++ctx->pin_count; in find_get_context()
4522 if (task_ctx_data && !ctx->task_ctx_data) { in find_get_context()
4523 ctx->task_ctx_data = task_ctx_data; in find_get_context()
4526 raw_spin_unlock_irqrestore(&ctx->lock, flags); in find_get_context()
4532 err = -ENOMEM; in find_get_context()
4537 ctx->task_ctx_data = task_ctx_data; in find_get_context()
4542 mutex_lock(&task->perf_event_mutex); in find_get_context()
4547 if (task->flags & PF_EXITING) in find_get_context()
4548 err = -ESRCH; in find_get_context()
4549 else if (task->perf_event_ctxp[ctxn]) in find_get_context()
4550 err = -EAGAIN; in find_get_context()
4553 ++ctx->pin_count; in find_get_context()
4554 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); in find_get_context()
4556 mutex_unlock(&task->perf_event_mutex); in find_get_context()
4561 if (err == -EAGAIN) in find_get_context()
4583 if (event->ns) in free_event_rcu()
4584 put_pid_ns(event->ns); in free_event_rcu()
4594 struct pmu_event_list *pel = per_cpu_ptr(&pmu_sb_events, event->cpu); in detach_sb_event()
4596 raw_spin_lock(&pel->lock); in detach_sb_event()
4597 list_del_rcu(&event->sb_list); in detach_sb_event()
4598 raw_spin_unlock(&pel->lock); in detach_sb_event()
4603 struct perf_event_attr *attr = &event->attr; in is_sb_event()
4605 if (event->parent) in is_sb_event()
4608 if (event->attach_state & PERF_ATTACH_TASK) in is_sb_event()
4611 if (attr->mmap || attr->mmap_data || attr->mmap2 || in is_sb_event()
4612 attr->comm || attr->comm_exec || in is_sb_event()
4613 attr->task || attr->ksymbol || in is_sb_event()
4614 attr->context_switch || attr->text_poke || in is_sb_event()
4615 attr->bpf_event) in is_sb_event()
4628 if (event->parent) in unaccount_event_cpu()
4661 if (event->parent) in unaccount_event()
4664 if (event->attach_state & PERF_ATTACH_TASK) in unaccount_event()
4666 if (event->attr.mmap || event->attr.mmap_data) in unaccount_event()
4668 if (event->attr.comm) in unaccount_event()
4670 if (event->attr.namespaces) in unaccount_event()
4672 if (event->attr.cgroup) in unaccount_event()
4674 if (event->attr.task) in unaccount_event()
4676 if (event->attr.freq) in unaccount_event()
4678 if (event->attr.context_switch) { in unaccount_event()
4686 if (event->attr.ksymbol) in unaccount_event()
4688 if (event->attr.bpf_event) in unaccount_event()
4690 if (event->attr.text_poke) in unaccount_event()
4694 if (!atomic_add_unless(&perf_sched_count, -1, 1)) in unaccount_event()
4698 unaccount_event_cpu(event, event->cpu); in unaccount_event()
4716 * 1) cpu-wide events in the presence of per-task events,
4717 * 2) per-task events in the presence of cpu-wide events,
4721 * _free_event()), the latter -- before the first perf_install_in_context().
4725 struct pmu *pmu = event->pmu; in exclusive_event_init()
4731 * Prevent co-existence of per-task and cpu-wide events on the in exclusive_event_init()
4734 * Negative pmu::exclusive_cnt means there are cpu-wide in exclusive_event_init()
4736 * per-task events. in exclusive_event_init()
4740 * to mean "per-task event", because unlike other attach states it in exclusive_event_init()
4743 if (event->attach_state & PERF_ATTACH_TASK) { in exclusive_event_init()
4744 if (!atomic_inc_unless_negative(&pmu->exclusive_cnt)) in exclusive_event_init()
4745 return -EBUSY; in exclusive_event_init()
4747 if (!atomic_dec_unless_positive(&pmu->exclusive_cnt)) in exclusive_event_init()
4748 return -EBUSY; in exclusive_event_init()
4756 struct pmu *pmu = event->pmu; in exclusive_event_destroy()
4762 if (event->attach_state & PERF_ATTACH_TASK) in exclusive_event_destroy()
4763 atomic_dec(&pmu->exclusive_cnt); in exclusive_event_destroy()
4765 atomic_inc(&pmu->exclusive_cnt); in exclusive_event_destroy()
4770 if ((e1->pmu == e2->pmu) && in exclusive_event_match()
4771 (e1->cpu == e2->cpu || in exclusive_event_match()
4772 e1->cpu == -1 || in exclusive_event_match()
4773 e2->cpu == -1)) in exclusive_event_match()
4782 struct pmu *pmu = event->pmu; in exclusive_event_installable()
4784 lockdep_assert_held(&ctx->mutex); in exclusive_event_installable()
4789 list_for_each_entry(iter_event, &ctx->event_list, event_entry) { in exclusive_event_installable()
4802 irq_work_sync(&event->pending); in _free_event()
4808 if (event->rb) { in _free_event()
4810 * Can happen when we close an event with re-directed output. in _free_event()
4815 mutex_lock(&event->mmap_mutex); in _free_event()
4817 mutex_unlock(&event->mmap_mutex); in _free_event()
4823 if (!event->parent) { in _free_event()
4824 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) in _free_event()
4830 kfree(event->addr_filter_ranges); in _free_event()
4832 if (event->destroy) in _free_event()
4833 event->destroy(event); in _free_event()
4836 * Must be after ->destroy(), due to uprobe_perf_close() using in _free_event()
4839 if (event->hw.target) in _free_event()
4840 put_task_struct(event->hw.target); in _free_event()
4846 if (event->ctx) in _free_event()
4847 put_ctx(event->ctx); in _free_event()
4850 module_put(event->pmu->module); in _free_event()
4852 call_rcu(&event->rcu_head, free_event_rcu); in _free_event()
4861 if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1, in free_event()
4863 atomic_long_read(&event->refcount), event)) { in free_event()
4864 /* leak to avoid use-after-free */ in free_event()
4883 * owner->perf_event_mutex. in perf_remove_from_owner()
4885 owner = READ_ONCE(event->owner); in perf_remove_from_owner()
4899 * holding ctx->mutex which would be an inversion wrt. the in perf_remove_from_owner()
4903 * ctx->mutex. in perf_remove_from_owner()
4905 mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING); in perf_remove_from_owner()
4908 * We have to re-check the event->owner field, if it is cleared in perf_remove_from_owner()
4913 if (event->owner) { in perf_remove_from_owner()
4914 list_del_init(&event->owner_entry); in perf_remove_from_owner()
4915 smp_store_release(&event->owner, NULL); in perf_remove_from_owner()
4917 mutex_unlock(&owner->perf_event_mutex); in perf_remove_from_owner()
4924 if (!atomic_long_dec_and_test(&event->refcount)) in put_event()
4937 struct perf_event_context *ctx = event->ctx; in perf_event_release_kernel()
4946 WARN_ON_ONCE(event->attach_state & in perf_event_release_kernel()
4955 WARN_ON_ONCE(ctx->parent_ctx); in perf_event_release_kernel()
4958 raw_spin_lock_irq(&ctx->lock); in perf_event_release_kernel()
4963 * Anybody acquiring event->child_mutex after the below loop _must_ in perf_event_release_kernel()
4970 event->state = PERF_EVENT_STATE_DEAD; in perf_event_release_kernel()
4971 raw_spin_unlock_irq(&ctx->lock); in perf_event_release_kernel()
4976 mutex_lock(&event->child_mutex); in perf_event_release_kernel()
4977 list_for_each_entry(child, &event->child_list, child_list) { in perf_event_release_kernel()
4983 ctx = READ_ONCE(child->ctx); in perf_event_release_kernel()
4997 * can re-acquire child_mutex. in perf_event_release_kernel()
4999 mutex_unlock(&event->child_mutex); in perf_event_release_kernel()
5000 mutex_lock(&ctx->mutex); in perf_event_release_kernel()
5001 mutex_lock(&event->child_mutex); in perf_event_release_kernel()
5008 tmp = list_first_entry_or_null(&event->child_list, in perf_event_release_kernel()
5012 list_move(&child->child_list, &free_list); in perf_event_release_kernel()
5020 mutex_unlock(&event->child_mutex); in perf_event_release_kernel()
5021 mutex_unlock(&ctx->mutex); in perf_event_release_kernel()
5025 mutex_unlock(&event->child_mutex); in perf_event_release_kernel()
5028 void *var = &child->ctx->refcount; in perf_event_release_kernel()
5030 list_del(&child->child_list); in perf_event_release_kernel()
5052 perf_event_release_kernel(file->private_data); in perf_release()
5064 mutex_lock(&event->child_mutex); in __perf_event_read_value()
5069 *enabled += event->total_time_enabled + in __perf_event_read_value()
5070 atomic64_read(&event->child_total_time_enabled); in __perf_event_read_value()
5071 *running += event->total_time_running + in __perf_event_read_value()
5072 atomic64_read(&event->child_total_time_running); in __perf_event_read_value()
5074 list_for_each_entry(child, &event->child_list, child_list) { in __perf_event_read_value()
5077 *enabled += child->total_time_enabled; in __perf_event_read_value()
5078 *running += child->total_time_running; in __perf_event_read_value()
5080 mutex_unlock(&event->child_mutex); in __perf_event_read_value()
5101 struct perf_event_context *ctx = leader->ctx; in __perf_read_group_add()
5111 raw_spin_lock_irqsave(&ctx->lock, flags); in __perf_read_group_add()
5114 * Since we co-schedule groups, {enabled,running} times of siblings in __perf_read_group_add()
5119 values[n++] += leader->total_time_enabled + in __perf_read_group_add()
5120 atomic64_read(&leader->child_total_time_enabled); in __perf_read_group_add()
5124 values[n++] += leader->total_time_running + in __perf_read_group_add()
5125 atomic64_read(&leader->child_total_time_running); in __perf_read_group_add()
5141 raw_spin_unlock_irqrestore(&ctx->lock, flags); in __perf_read_group_add()
5148 struct perf_event *leader = event->group_leader, *child; in perf_read_group()
5149 struct perf_event_context *ctx = leader->ctx; in perf_read_group()
5153 lockdep_assert_held(&ctx->mutex); in perf_read_group()
5155 values = kzalloc(event->read_size, GFP_KERNEL); in perf_read_group()
5157 return -ENOMEM; in perf_read_group()
5159 values[0] = 1 + leader->nr_siblings; in perf_read_group()
5165 mutex_lock(&leader->child_mutex); in perf_read_group()
5171 list_for_each_entry(child, &leader->child_list, child_list) { in perf_read_group()
5177 mutex_unlock(&leader->child_mutex); in perf_read_group()
5179 ret = event->read_size; in perf_read_group()
5180 if (copy_to_user(buf, values, event->read_size)) in perf_read_group()
5181 ret = -EFAULT; in perf_read_group()
5185 mutex_unlock(&leader->child_mutex); in perf_read_group()
5207 return -EFAULT; in perf_read_one()
5216 if (event->state > PERF_EVENT_STATE_EXIT) in is_event_hup()
5219 mutex_lock(&event->child_mutex); in is_event_hup()
5220 no_children = list_empty(&event->child_list); in is_event_hup()
5221 mutex_unlock(&event->child_mutex); in is_event_hup()
5226 * Read the performance event - simple non blocking version for now
5231 u64 read_format = event->attr.read_format; in __perf_read()
5235 * Return end-of-file for a read on an event that is in in __perf_read()
5239 if (event->state == PERF_EVENT_STATE_ERROR) in __perf_read()
5242 if (count < event->read_size) in __perf_read()
5243 return -ENOSPC; in __perf_read()
5245 WARN_ON_ONCE(event->ctx->parent_ctx); in __perf_read()
5257 struct perf_event *event = file->private_data; in perf_read()
5274 struct perf_event *event = file->private_data; in perf_poll()
5278 poll_wait(file, &event->waitq, wait); in perf_poll()
5284 * Pin the event->rb by taking event->mmap_mutex; otherwise in perf_poll()
5287 mutex_lock(&event->mmap_mutex); in perf_poll()
5288 rb = event->rb; in perf_poll()
5290 events = atomic_xchg(&rb->poll, 0); in perf_poll()
5291 mutex_unlock(&event->mmap_mutex); in perf_poll()
5298 local64_set(&event->count, 0); in _perf_event_reset()
5309 WARN_ON_ONCE(event->attr.inherit); in perf_event_pause()
5311 count = local64_read(&event->count); in perf_event_pause()
5313 local64_set(&event->count, 0); in perf_event_pause()
5321 * Holding the top-level event's child_mutex means that any
5331 WARN_ON_ONCE(event->ctx->parent_ctx); in perf_event_for_each_child()
5333 mutex_lock(&event->child_mutex); in perf_event_for_each_child()
5335 list_for_each_entry(child, &event->child_list, child_list) in perf_event_for_each_child()
5337 mutex_unlock(&event->child_mutex); in perf_event_for_each_child()
5343 struct perf_event_context *ctx = event->ctx; in perf_event_for_each()
5346 lockdep_assert_held(&ctx->mutex); in perf_event_for_each()
5348 event = event->group_leader; in perf_event_for_each()
5363 if (event->attr.freq) { in __perf_event_period()
5364 event->attr.sample_freq = value; in __perf_event_period()
5366 event->attr.sample_period = value; in __perf_event_period()
5367 event->hw.sample_period = value; in __perf_event_period()
5370 active = (event->state == PERF_EVENT_STATE_ACTIVE); in __perf_event_period()
5372 perf_pmu_disable(ctx->pmu); in __perf_event_period()
5375 * trying to unthrottle while we already re-started the event. in __perf_event_period()
5377 if (event->hw.interrupts == MAX_INTERRUPTS) { in __perf_event_period()
5378 event->hw.interrupts = 0; in __perf_event_period()
5381 event->pmu->stop(event, PERF_EF_UPDATE); in __perf_event_period()
5384 local64_set(&event->hw.period_left, 0); in __perf_event_period()
5387 event->pmu->start(event, PERF_EF_RELOAD); in __perf_event_period()
5388 perf_pmu_enable(ctx->pmu); in __perf_event_period()
5394 return event->pmu->check_period(event, value); in perf_event_check_period()
5400 return -EINVAL; in _perf_event_period()
5403 return -EINVAL; in _perf_event_period()
5405 if (event->attr.freq && value > sysctl_perf_event_sample_rate) in _perf_event_period()
5406 return -EINVAL; in _perf_event_period()
5409 return -EINVAL; in _perf_event_period()
5411 if (!event->attr.freq && (value & (1ULL << 63))) in _perf_event_period()
5412 return -EINVAL; in _perf_event_period()
5438 return -EBADF; in perf_fget_light()
5440 if (f.file->f_op != &perf_fops) { in perf_fget_light()
5442 return -EBADF; in perf_fget_light()
5479 return -EFAULT; in _perf_ioctl()
5488 return -EFAULT; in _perf_ioctl()
5495 if (arg != -1) { in _perf_ioctl()
5501 output_event = output.file->private_data; in _perf_ioctl()
5520 rb = rcu_dereference(event->rb); in _perf_ioctl()
5521 if (!rb || !rb->nr_pages) { in _perf_ioctl()
5523 return -EINVAL; in _perf_ioctl()
5544 return -ENOTTY; in _perf_ioctl()
5557 struct perf_event *event = file->private_data; in perf_ioctl()
5582 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */ in perf_compat_ioctl()
5600 mutex_lock(¤t->perf_event_mutex); in perf_event_task_enable()
5601 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { in perf_event_task_enable()
5606 mutex_unlock(¤t->perf_event_mutex); in perf_event_task_enable()
5616 mutex_lock(¤t->perf_event_mutex); in perf_event_task_disable()
5617 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { in perf_event_task_disable()
5622 mutex_unlock(¤t->perf_event_mutex); in perf_event_task_disable()
5629 if (event->hw.state & PERF_HES_STOPPED) in perf_event_index()
5632 if (event->state != PERF_EVENT_STATE_ACTIVE) in perf_event_index()
5635 return event->pmu->event_idx(event); in perf_event_index()
5646 ctx_time = event->shadow_ctx_time + *now; in calc_timer_values()
5656 rb = rcu_dereference(event->rb); in perf_event_init_userpage()
5660 userpg = rb->user_page; in perf_event_init_userpage()
5663 userpg->cap_bit0_is_deprecated = 1; in perf_event_init_userpage()
5664 userpg->size = offsetof(struct perf_event_mmap_page, __reserved); in perf_event_init_userpage()
5665 userpg->data_offset = PAGE_SIZE; in perf_event_init_userpage()
5666 userpg->data_size = perf_data_size(rb); in perf_event_init_userpage()
5689 rb = rcu_dereference(event->rb); in perf_event_update_userpage()
5704 userpg = rb->user_page; in perf_event_update_userpage()
5710 ++userpg->lock; in perf_event_update_userpage()
5712 userpg->index = perf_event_index(event); in perf_event_update_userpage()
5713 userpg->offset = perf_event_count(event); in perf_event_update_userpage()
5714 if (userpg->index) in perf_event_update_userpage()
5715 userpg->offset -= local64_read(&event->hw.prev_count); in perf_event_update_userpage()
5717 userpg->time_enabled = enabled + in perf_event_update_userpage()
5718 atomic64_read(&event->child_total_time_enabled); in perf_event_update_userpage()
5720 userpg->time_running = running + in perf_event_update_userpage()
5721 atomic64_read(&event->child_total_time_running); in perf_event_update_userpage()
5726 ++userpg->lock; in perf_event_update_userpage()
5735 struct perf_event *event = vmf->vma->vm_file->private_data; in perf_mmap_fault()
5739 if (vmf->flags & FAULT_FLAG_MKWRITE) { in perf_mmap_fault()
5740 if (vmf->pgoff == 0) in perf_mmap_fault()
5746 rb = rcu_dereference(event->rb); in perf_mmap_fault()
5750 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) in perf_mmap_fault()
5753 vmf->page = perf_mmap_to_page(rb, vmf->pgoff); in perf_mmap_fault()
5754 if (!vmf->page) in perf_mmap_fault()
5757 get_page(vmf->page); in perf_mmap_fault()
5758 vmf->page->mapping = vmf->vma->vm_file->f_mapping; in perf_mmap_fault()
5759 vmf->page->index = vmf->pgoff; in perf_mmap_fault()
5774 if (event->rb) { in ring_buffer_attach()
5777 * event->rb_entry and wait/clear when adding event->rb_entry. in ring_buffer_attach()
5779 WARN_ON_ONCE(event->rcu_pending); in ring_buffer_attach()
5781 old_rb = event->rb; in ring_buffer_attach()
5782 spin_lock_irqsave(&old_rb->event_lock, flags); in ring_buffer_attach()
5783 list_del_rcu(&event->rb_entry); in ring_buffer_attach()
5784 spin_unlock_irqrestore(&old_rb->event_lock, flags); in ring_buffer_attach()
5786 event->rcu_batches = get_state_synchronize_rcu(); in ring_buffer_attach()
5787 event->rcu_pending = 1; in ring_buffer_attach()
5791 if (event->rcu_pending) { in ring_buffer_attach()
5792 cond_synchronize_rcu(event->rcu_batches); in ring_buffer_attach()
5793 event->rcu_pending = 0; in ring_buffer_attach()
5796 spin_lock_irqsave(&rb->event_lock, flags); in ring_buffer_attach()
5797 list_add_rcu(&event->rb_entry, &rb->event_list); in ring_buffer_attach()
5798 spin_unlock_irqrestore(&rb->event_lock, flags); in ring_buffer_attach()
5808 * mid-air, but then again, whoever does it like this is in ring_buffer_attach()
5814 rcu_assign_pointer(event->rb, rb); in ring_buffer_attach()
5823 wake_up_all(&event->waitq); in ring_buffer_attach()
5832 rb = rcu_dereference(event->rb); in ring_buffer_wakeup()
5834 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) in ring_buffer_wakeup()
5835 wake_up_all(&event->waitq); in ring_buffer_wakeup()
5845 rb = rcu_dereference(event->rb); in ring_buffer_get()
5847 if (!refcount_inc_not_zero(&rb->refcount)) in ring_buffer_get()
5857 if (!refcount_dec_and_test(&rb->refcount)) in ring_buffer_put()
5860 WARN_ON_ONCE(!list_empty(&rb->event_list)); in ring_buffer_put()
5862 call_rcu(&rb->rcu_head, rb_free_rcu); in ring_buffer_put()
5867 struct perf_event *event = vma->vm_file->private_data; in perf_mmap_open()
5869 atomic_inc(&event->mmap_count); in perf_mmap_open()
5870 atomic_inc(&event->rb->mmap_count); in perf_mmap_open()
5872 if (vma->vm_pgoff) in perf_mmap_open()
5873 atomic_inc(&event->rb->aux_mmap_count); in perf_mmap_open()
5875 if (event->pmu->event_mapped) in perf_mmap_open()
5876 event->pmu->event_mapped(event, vma->vm_mm); in perf_mmap_open()
5891 struct perf_event *event = vma->vm_file->private_data; in perf_mmap_close()
5893 struct user_struct *mmap_user = rb->mmap_user; in perf_mmap_close()
5894 int mmap_locked = rb->mmap_locked; in perf_mmap_close()
5898 if (event->pmu->event_unmapped) in perf_mmap_close()
5899 event->pmu->event_unmapped(event, vma->vm_mm); in perf_mmap_close()
5902 * rb->aux_mmap_count will always drop before rb->mmap_count and in perf_mmap_close()
5903 * event->mmap_count, so it is ok to use event->mmap_mutex to in perf_mmap_close()
5906 if (rb_has_aux(rb) && vma->vm_pgoff == rb->aux_pgoff && in perf_mmap_close()
5907 atomic_dec_and_mutex_lock(&rb->aux_mmap_count, &event->mmap_mutex)) { in perf_mmap_close()
5917 atomic_long_sub(rb->aux_nr_pages - rb->aux_mmap_locked, &mmap_user->locked_vm); in perf_mmap_close()
5918 atomic64_sub(rb->aux_mmap_locked, &vma->vm_mm->pinned_vm); in perf_mmap_close()
5922 WARN_ON_ONCE(refcount_read(&rb->aux_refcount)); in perf_mmap_close()
5924 mutex_unlock(&event->mmap_mutex); in perf_mmap_close()
5927 if (atomic_dec_and_test(&rb->mmap_count)) in perf_mmap_close()
5930 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) in perf_mmap_close()
5934 mutex_unlock(&event->mmap_mutex); in perf_mmap_close()
5947 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { in perf_mmap_close()
5948 if (!atomic_long_inc_not_zero(&event->refcount)) { in perf_mmap_close()
5950 * This event is en-route to free_event() which will in perf_mmap_close()
5957 mutex_lock(&event->mmap_mutex); in perf_mmap_close()
5968 if (event->rb == rb) in perf_mmap_close()
5971 mutex_unlock(&event->mmap_mutex); in perf_mmap_close()
5983 * It could be there's still a few 0-ref events on the list; they'll in perf_mmap_close()
5984 * get cleaned up by free_event() -- they'll also still have their in perf_mmap_close()
5991 atomic_long_sub((size >> PAGE_SHIFT) + 1 - mmap_locked, in perf_mmap_close()
5992 &mmap_user->locked_vm); in perf_mmap_close()
5993 atomic64_sub(mmap_locked, &vma->vm_mm->pinned_vm); in perf_mmap_close()
6009 struct perf_event *event = file->private_data; in perf_mmap()
6020 * Don't allow mmap() of inherited per-task counters. This would in perf_mmap()
6024 if (event->cpu == -1 && event->attr.inherit) in perf_mmap()
6025 return -EINVAL; in perf_mmap()
6027 if (!(vma->vm_flags & VM_SHARED)) in perf_mmap()
6028 return -EINVAL; in perf_mmap()
6034 vma_size = vma->vm_end - vma->vm_start; in perf_mmap()
6036 if (vma->vm_pgoff == 0) { in perf_mmap()
6037 nr_pages = (vma_size / PAGE_SIZE) - 1; in perf_mmap()
6040 * AUX area mapping: if rb->aux_nr_pages != 0, it's already in perf_mmap()
6046 if (!event->rb) in perf_mmap()
6047 return -EINVAL; in perf_mmap()
6051 mutex_lock(&event->mmap_mutex); in perf_mmap()
6052 ret = -EINVAL; in perf_mmap()
6054 rb = event->rb; in perf_mmap()
6058 aux_offset = READ_ONCE(rb->user_page->aux_offset); in perf_mmap()
6059 aux_size = READ_ONCE(rb->user_page->aux_size); in perf_mmap()
6064 if (aux_offset != vma->vm_pgoff << PAGE_SHIFT) in perf_mmap()
6068 if (rb_has_aux(rb) && rb->aux_pgoff != vma->vm_pgoff) in perf_mmap()
6075 if (rb_has_aux(rb) && rb->aux_nr_pages != nr_pages) in perf_mmap()
6081 if (!atomic_inc_not_zero(&rb->mmap_count)) in perf_mmap()
6085 atomic_inc(&rb->aux_mmap_count); in perf_mmap()
6090 atomic_set(&rb->aux_mmap_count, 1); in perf_mmap()
6097 * If we have rb pages ensure they're a power-of-two number, so we in perf_mmap()
6101 return -EINVAL; in perf_mmap()
6104 return -EINVAL; in perf_mmap()
6106 WARN_ON_ONCE(event->ctx->parent_ctx); in perf_mmap()
6108 mutex_lock(&event->mmap_mutex); in perf_mmap()
6109 if (event->rb) { in perf_mmap()
6110 if (event->rb->nr_pages != nr_pages) { in perf_mmap()
6111 ret = -EINVAL; in perf_mmap()
6115 if (!atomic_inc_not_zero(&event->rb->mmap_count)) { in perf_mmap()
6121 mutex_unlock(&event->mmap_mutex); in perf_mmap()
6131 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); in perf_mmap()
6138 user_locked = atomic_long_read(&user->locked_vm); in perf_mmap()
6142 * user->locked_vm > user_lock_limit in perf_mmap()
6153 extra = user_locked - user_lock_limit; in perf_mmap()
6154 user_extra -= extra; in perf_mmap()
6159 locked = atomic64_read(&vma->vm_mm->pinned_vm) + extra; in perf_mmap()
6163 ret = -EPERM; in perf_mmap()
6167 WARN_ON(!rb && event->rb); in perf_mmap()
6169 if (vma->vm_flags & VM_WRITE) in perf_mmap()
6174 event->attr.watermark ? event->attr.wakeup_watermark : 0, in perf_mmap()
6175 event->cpu, flags); in perf_mmap()
6178 ret = -ENOMEM; in perf_mmap()
6182 atomic_set(&rb->mmap_count, 1); in perf_mmap()
6183 rb->mmap_user = get_current_user(); in perf_mmap()
6184 rb->mmap_locked = extra; in perf_mmap()
6191 ret = rb_alloc_aux(rb, event, vma->vm_pgoff, nr_pages, in perf_mmap()
6192 event->attr.aux_watermark, flags); in perf_mmap()
6194 rb->aux_mmap_locked = extra; in perf_mmap()
6199 atomic_long_add(user_extra, &user->locked_vm); in perf_mmap()
6200 atomic64_add(extra, &vma->vm_mm->pinned_vm); in perf_mmap()
6202 atomic_inc(&event->mmap_count); in perf_mmap()
6204 atomic_dec(&rb->mmap_count); in perf_mmap()
6207 mutex_unlock(&event->mmap_mutex); in perf_mmap()
6213 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; in perf_mmap()
6214 vma->vm_ops = &perf_mmap_vmops; in perf_mmap()
6216 if (event->pmu->event_mapped) in perf_mmap()
6217 event->pmu->event_mapped(event, vma->vm_mm); in perf_mmap()
6225 struct perf_event *event = filp->private_data; in perf_fasync()
6229 retval = fasync_helper(fd, filp, on, &event->fasync); in perf_fasync()
6253 * to user-space before waking everybody up.
6259 if (event->parent) in perf_event_fasync()
6260 event = event->parent; in perf_event_fasync()
6261 return &event->fasync; in perf_event_fasync()
6268 if (event->pending_kill) { in perf_event_wakeup()
6269 kill_fasync(perf_event_fasync(event), SIGIO, event->pending_kill); in perf_event_wakeup()
6270 event->pending_kill = 0; in perf_event_wakeup()
6276 int cpu = READ_ONCE(event->pending_disable); in perf_pending_event_disable()
6282 WRITE_ONCE(event->pending_disable, -1); in perf_pending_event_disable()
6288 * CPU-A CPU-B in perf_pending_event_disable()
6291 * @pending_disable = CPU-A; in perf_pending_event_disable()
6294 * sched-out in perf_pending_event_disable()
6295 * @pending_disable = -1; in perf_pending_event_disable()
6297 * sched-in in perf_pending_event_disable()
6299 * @pending_disable = CPU-B; in perf_pending_event_disable()
6305 * But the event runs on CPU-B and wants disabling there. in perf_pending_event_disable()
6307 irq_work_queue_on(&event->pending, cpu); in perf_pending_event_disable()
6323 if (event->pending_wakeup) { in perf_pending_event()
6324 event->pending_wakeup = 0; in perf_pending_event()
6373 regs_user->abi = perf_reg_abi(current); in perf_sample_regs_user()
6374 regs_user->regs = regs; in perf_sample_regs_user()
6375 } else if (!(current->flags & PF_KTHREAD)) { in perf_sample_regs_user()
6378 regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE; in perf_sample_regs_user()
6379 regs_user->regs = NULL; in perf_sample_regs_user()
6386 regs_intr->regs = regs; in perf_sample_regs_intr()
6387 regs_intr->abi = perf_reg_abi(current); in perf_sample_regs_intr()
6405 return TASK_SIZE - addr; in perf_ustack_task_size()
6420 * - TASK_SIZE in perf_sample_ustack_size()
6423 * - remaining sample size in perf_sample_ustack_size()
6440 stack_size = USHRT_MAX - header_size - sizeof(u64); in perf_sample_ustack_size()
6464 * - the size requested by user or the best one we can fit in perf_output_sample_ustack()
6467 * - user stack dump data in perf_output_sample_ustack()
6469 * - the actual dumped size in perf_output_sample_ustack()
6480 dyn_size = dump_size - rem; in perf_output_sample_ustack()
6493 struct perf_event *sampler = event->aux_event; in perf_prepare_sample_aux()
6496 data->aux_size = 0; in perf_prepare_sample_aux()
6501 if (WARN_ON_ONCE(READ_ONCE(sampler->state) != PERF_EVENT_STATE_ACTIVE)) in perf_prepare_sample_aux()
6504 if (WARN_ON_ONCE(READ_ONCE(sampler->oncpu) != smp_processor_id())) in perf_prepare_sample_aux()
6507 rb = ring_buffer_get(sampler->parent ? sampler->parent : sampler); in perf_prepare_sample_aux()
6515 if (READ_ONCE(rb->aux_in_sampling)) { in perf_prepare_sample_aux()
6516 data->aux_size = 0; in perf_prepare_sample_aux()
6519 data->aux_size = ALIGN(size, sizeof(u64)); in perf_prepare_sample_aux()
6524 return data->aux_size; in perf_prepare_sample_aux()
6536 * Normal ->start()/->stop() callbacks run in IRQ mode in scheduler in perf_pmu_snapshot_aux()
6538 * the IRQ ones, that is, for example, re-starting an event that's just in perf_pmu_snapshot_aux()
6549 WRITE_ONCE(rb->aux_in_sampling, 1); in perf_pmu_snapshot_aux()
6552 ret = event->pmu->snapshot_aux(event, handle, size); in perf_pmu_snapshot_aux()
6555 WRITE_ONCE(rb->aux_in_sampling, 0); in perf_pmu_snapshot_aux()
6565 struct perf_event *sampler = event->aux_event; in perf_aux_sample_output()
6570 if (WARN_ON_ONCE(!sampler || !data->aux_size)) in perf_aux_sample_output()
6573 rb = ring_buffer_get(sampler->parent ? sampler->parent : sampler); in perf_aux_sample_output()
6577 size = perf_pmu_snapshot_aux(rb, sampler, handle, data->aux_size); in perf_aux_sample_output()
6581 * non-zero surplus that it didn't copy), which in its current in perf_aux_sample_output()
6589 * The pad comes from ALIGN()ing data->aux_size up to u64 in in perf_aux_sample_output()
6592 pad = data->aux_size - size; in perf_aux_sample_output()
6609 u64 sample_type = event->attr.sample_type; in __perf_event_header__init_id()
6611 data->type = sample_type; in __perf_event_header__init_id()
6612 header->size += event->id_header_size; in __perf_event_header__init_id()
6616 data->tid_entry.pid = perf_event_pid(event, current); in __perf_event_header__init_id()
6617 data->tid_entry.tid = perf_event_tid(event, current); in __perf_event_header__init_id()
6621 data->time = perf_event_clock(event); in __perf_event_header__init_id()
6624 data->id = primary_event_id(event); in __perf_event_header__init_id()
6627 data->stream_id = event->id; in __perf_event_header__init_id()
6630 data->cpu_entry.cpu = raw_smp_processor_id(); in __perf_event_header__init_id()
6631 data->cpu_entry.reserved = 0; in __perf_event_header__init_id()
6639 if (event->attr.sample_id_all) in perf_event_header__init_id()
6646 u64 sample_type = data->type; in __perf_event__output_id_sample()
6649 perf_output_put(handle, data->tid_entry); in __perf_event__output_id_sample()
6652 perf_output_put(handle, data->time); in __perf_event__output_id_sample()
6655 perf_output_put(handle, data->id); in __perf_event__output_id_sample()
6658 perf_output_put(handle, data->stream_id); in __perf_event__output_id_sample()
6661 perf_output_put(handle, data->cpu_entry); in __perf_event__output_id_sample()
6664 perf_output_put(handle, data->id); in __perf_event__output_id_sample()
6671 if (event->attr.sample_id_all) in perf_event__output_id_sample()
6679 u64 read_format = event->attr.read_format; in perf_output_read_one()
6686 atomic64_read(&event->child_total_time_enabled); in perf_output_read_one()
6690 atomic64_read(&event->child_total_time_running); in perf_output_read_one()
6702 struct perf_event *leader = event->group_leader, *sub; in perf_output_read_group()
6703 u64 read_format = event->attr.read_format; in perf_output_read_group()
6707 values[n++] = 1 + leader->nr_siblings; in perf_output_read_group()
6716 (leader->state == PERF_EVENT_STATE_ACTIVE)) in perf_output_read_group()
6717 leader->pmu->read(leader); in perf_output_read_group()
6729 (sub->state == PERF_EVENT_STATE_ACTIVE)) in perf_output_read_group()
6730 sub->pmu->read(sub); in perf_output_read_group()
6747 * child list, not to mention that its impossible to IPI the children running
6754 u64 read_format = event->attr.read_format; in perf_output_read()
6768 if (event->attr.read_format & PERF_FORMAT_GROUP) in perf_output_read()
6776 return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX; in perf_sample_save_hw_index()
6784 u64 sample_type = data->type; in perf_output_sample()
6789 perf_output_put(handle, data->id); in perf_output_sample()
6792 perf_output_put(handle, data->ip); in perf_output_sample()
6795 perf_output_put(handle, data->tid_entry); in perf_output_sample()
6798 perf_output_put(handle, data->time); in perf_output_sample()
6801 perf_output_put(handle, data->addr); in perf_output_sample()
6804 perf_output_put(handle, data->id); in perf_output_sample()
6807 perf_output_put(handle, data->stream_id); in perf_output_sample()
6810 perf_output_put(handle, data->cpu_entry); in perf_output_sample()
6813 perf_output_put(handle, data->period); in perf_output_sample()
6821 size += data->callchain->nr; in perf_output_sample()
6823 __output_copy(handle, data->callchain, size); in perf_output_sample()
6827 struct perf_raw_record *raw = data->raw; in perf_output_sample()
6830 struct perf_raw_frag *frag = &raw->frag; in perf_output_sample()
6832 perf_output_put(handle, raw->size); in perf_output_sample()
6834 if (frag->copy) { in perf_output_sample()
6835 __output_custom(handle, frag->copy, in perf_output_sample()
6836 frag->data, frag->size); in perf_output_sample()
6838 __output_copy(handle, frag->data, in perf_output_sample()
6839 frag->size); in perf_output_sample()
6843 frag = frag->next; in perf_output_sample()
6845 if (frag->pad) in perf_output_sample()
6846 __output_skip(handle, NULL, frag->pad); in perf_output_sample()
6860 if (data->br_stack) { in perf_output_sample()
6863 size = data->br_stack->nr in perf_output_sample()
6866 perf_output_put(handle, data->br_stack->nr); in perf_output_sample()
6868 perf_output_put(handle, data->br_stack->hw_idx); in perf_output_sample()
6869 perf_output_copy(handle, data->br_stack->entries, size); in perf_output_sample()
6880 u64 abi = data->regs_user.abi; in perf_output_sample()
6889 u64 mask = event->attr.sample_regs_user; in perf_output_sample()
6891 data->regs_user.regs, in perf_output_sample()
6898 data->stack_user_size, in perf_output_sample()
6899 data->regs_user.regs); in perf_output_sample()
6903 perf_output_put(handle, data->weight); in perf_output_sample()
6906 perf_output_put(handle, data->data_src.val); in perf_output_sample()
6909 perf_output_put(handle, data->txn); in perf_output_sample()
6912 u64 abi = data->regs_intr.abi; in perf_output_sample()
6920 u64 mask = event->attr.sample_regs_intr; in perf_output_sample()
6923 data->regs_intr.regs, in perf_output_sample()
6929 perf_output_put(handle, data->phys_addr); in perf_output_sample()
6932 perf_output_put(handle, data->cgroup); in perf_output_sample()
6935 perf_output_put(handle, data->aux_size); in perf_output_sample()
6937 if (data->aux_size) in perf_output_sample()
6941 if (!event->attr.watermark) { in perf_output_sample()
6942 int wakeup_events = event->attr.wakeup_events; in perf_output_sample()
6945 struct perf_buffer *rb = handle->rb; in perf_output_sample()
6946 int events = local_inc_return(&rb->events); in perf_output_sample()
6949 local_sub(wakeup_events, &rb->events); in perf_output_sample()
6950 local_inc(&rb->wakeup); in perf_output_sample()
6974 * Try IRQ-safe get_user_page_fast_only first. in perf_virt_to_phys()
6977 if (current->mm != NULL) { in perf_virt_to_phys()
6996 bool kernel = !event->attr.exclude_callchain_kernel; in perf_callchain()
6997 bool user = !event->attr.exclude_callchain_user; in perf_callchain()
6998 /* Disallow cross-task user callchains. */ in perf_callchain()
6999 bool crosstask = event->ctx->task && event->ctx->task != current; in perf_callchain()
7000 const u32 max_stack = event->attr.sample_max_stack; in perf_callchain()
7016 u64 sample_type = event->attr.sample_type; in perf_prepare_sample()
7018 header->type = PERF_RECORD_SAMPLE; in perf_prepare_sample()
7019 header->size = sizeof(*header) + event->header_size; in perf_prepare_sample()
7021 header->misc = 0; in perf_prepare_sample()
7022 header->misc |= perf_misc_flags(regs); in perf_prepare_sample()
7027 data->ip = perf_instruction_pointer(regs); in perf_prepare_sample()
7033 data->callchain = perf_callchain(event, regs); in perf_prepare_sample()
7035 size += data->callchain->nr; in perf_prepare_sample()
7037 header->size += size * sizeof(u64); in perf_prepare_sample()
7041 struct perf_raw_record *raw = data->raw; in perf_prepare_sample()
7045 struct perf_raw_frag *frag = &raw->frag; in perf_prepare_sample()
7049 sum += frag->size; in perf_prepare_sample()
7052 frag = frag->next; in perf_prepare_sample()
7056 raw->size = size - sizeof(u32); in perf_prepare_sample()
7057 frag->pad = raw->size - sum; in perf_prepare_sample()
7062 header->size += size; in perf_prepare_sample()
7067 if (data->br_stack) { in perf_prepare_sample()
7071 size += data->br_stack->nr in perf_prepare_sample()
7074 header->size += size; in perf_prepare_sample()
7078 perf_sample_regs_user(&data->regs_user, regs); in perf_prepare_sample()
7084 if (data->regs_user.regs) { in perf_prepare_sample()
7085 u64 mask = event->attr.sample_regs_user; in perf_prepare_sample()
7089 header->size += size; in perf_prepare_sample()
7099 u16 stack_size = event->attr.sample_stack_user; in perf_prepare_sample()
7102 stack_size = perf_sample_ustack_size(stack_size, header->size, in perf_prepare_sample()
7103 data->regs_user.regs); in perf_prepare_sample()
7113 data->stack_user_size = stack_size; in perf_prepare_sample()
7114 header->size += size; in perf_prepare_sample()
7121 perf_sample_regs_intr(&data->regs_intr, regs); in perf_prepare_sample()
7123 if (data->regs_intr.regs) { in perf_prepare_sample()
7124 u64 mask = event->attr.sample_regs_intr; in perf_prepare_sample()
7129 header->size += size; in perf_prepare_sample()
7133 data->phys_addr = perf_virt_to_phys(data->addr); in perf_prepare_sample()
7140 cgrp = task_css_check(current, perf_event_cgrp_id, 1)->cgroup; in perf_prepare_sample()
7141 data->cgroup = cgroup_id(cgrp); in perf_prepare_sample()
7148 header->size += sizeof(u64); /* size */ in perf_prepare_sample()
7156 size = min_t(size_t, U16_MAX - header->size, in perf_prepare_sample()
7157 event->attr.aux_sample_size); in perf_prepare_sample()
7161 WARN_ON_ONCE(size + header->size > U16_MAX); in perf_prepare_sample()
7162 header->size += size; in perf_prepare_sample()
7172 WARN_ON_ONCE(header->size & 7); in perf_prepare_sample()
7251 .size = sizeof(read_event) + event->read_size, in perf_event_read_event()
7279 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { in perf_iterate_ctx()
7281 if (event->state < PERF_EVENT_STATE_INACTIVE) in perf_iterate_ctx()
7296 list_for_each_entry_rcu(event, &pel->list, sb_list) { in perf_iterate_sb_cpu()
7299 * if we observe event->ctx, both event and ctx will be in perf_iterate_sb_cpu()
7302 if (!smp_load_acquire(&event->ctx)) in perf_iterate_sb_cpu()
7305 if (event->state < PERF_EVENT_STATE_INACTIVE) in perf_iterate_sb_cpu()
7314 * Iterate all events that need to receive side-band events.
7342 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); in perf_iterate_sb()
7352 * Clear all file-based filters at exec, they'll have to be
7353 * re-instated when/if these objects are mmapped again.
7365 raw_spin_lock_irqsave(&ifh->lock, flags); in perf_event_addr_filters_exec()
7366 list_for_each_entry(filter, &ifh->list, entry) { in perf_event_addr_filters_exec()
7367 if (filter->path.dentry) { in perf_event_addr_filters_exec()
7368 event->addr_filter_ranges[count].start = 0; in perf_event_addr_filters_exec()
7369 event->addr_filter_ranges[count].size = 0; in perf_event_addr_filters_exec()
7377 event->addr_filters_gen++; in perf_event_addr_filters_exec()
7378 raw_spin_unlock_irqrestore(&ifh->lock, flags); in perf_event_addr_filters_exec()
7391 ctx = current->perf_event_ctxp[ctxn]; in perf_event_exec()
7410 struct perf_event *parent = event->parent; in __perf_event_output_stop()
7412 struct perf_buffer *rb = ro->rb; in __perf_event_output_stop()
7425 * ring-buffer, but it will be the child that's actually using it. in __perf_event_output_stop()
7430 * So ring_buffer_attach() has to stop an aux event before re-assigning in __perf_event_output_stop()
7433 if (rcu_dereference(parent->rb) == rb) in __perf_event_output_stop()
7434 ro->err = __perf_event_stop(&sd); in __perf_event_output_stop()
7440 struct pmu *pmu = event->ctx->pmu; in __perf_pmu_output_stop()
7441 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); in __perf_pmu_output_stop()
7443 .rb = event->rb, in __perf_pmu_output_stop()
7447 perf_iterate_ctx(&cpuctx->ctx, __perf_event_output_stop, &ro, false); in __perf_pmu_output_stop()
7448 if (cpuctx->task_ctx) in __perf_pmu_output_stop()
7449 perf_iterate_ctx(cpuctx->task_ctx, __perf_event_output_stop, in __perf_pmu_output_stop()
7463 list_for_each_entry_rcu(iter, &event->rb->event_list, rb_entry) { in perf_pmu_output_stop()
7465 * For per-CPU events, we need to make sure that neither they in perf_pmu_output_stop()
7466 * nor their children are running; for cpu==-1 events it's in perf_pmu_output_stop()
7470 cpu = iter->cpu; in perf_pmu_output_stop()
7471 if (cpu == -1) in perf_pmu_output_stop()
7472 cpu = READ_ONCE(iter->oncpu); in perf_pmu_output_stop()
7474 if (cpu == -1) in perf_pmu_output_stop()
7478 if (err == -EAGAIN) { in perf_pmu_output_stop()
7487 * task tracking -- fork/exit
7509 return event->attr.comm || event->attr.mmap || in perf_event_task_match()
7510 event->attr.mmap2 || event->attr.mmap_data || in perf_event_task_match()
7511 event->attr.task; in perf_event_task_match()
7520 struct task_struct *task = task_event->task; in perf_event_task_output()
7521 int ret, size = task_event->event_id.header.size; in perf_event_task_output()
7526 perf_event_header__init_id(&task_event->event_id.header, &sample, event); in perf_event_task_output()
7529 task_event->event_id.header.size); in perf_event_task_output()
7533 task_event->event_id.pid = perf_event_pid(event, task); in perf_event_task_output()
7534 task_event->event_id.tid = perf_event_tid(event, task); in perf_event_task_output()
7536 if (task_event->event_id.header.type == PERF_RECORD_EXIT) { in perf_event_task_output()
7537 task_event->event_id.ppid = perf_event_pid(event, in perf_event_task_output()
7538 task->real_parent); in perf_event_task_output()
7539 task_event->event_id.ptid = perf_event_pid(event, in perf_event_task_output()
7540 task->real_parent); in perf_event_task_output()
7542 task_event->event_id.ppid = perf_event_pid(event, current); in perf_event_task_output()
7543 task_event->event_id.ptid = perf_event_tid(event, current); in perf_event_task_output()
7546 task_event->event_id.time = perf_event_clock(event); in perf_event_task_output()
7548 perf_output_put(&handle, task_event->event_id); in perf_event_task_output()
7554 task_event->event_id.header.size = size; in perf_event_task_output()
7615 return event->attr.comm; in perf_event_comm_match()
7624 int size = comm_event->event_id.header.size; in perf_event_comm_output()
7630 perf_event_header__init_id(&comm_event->event_id.header, &sample, event); in perf_event_comm_output()
7632 comm_event->event_id.header.size); in perf_event_comm_output()
7637 comm_event->event_id.pid = perf_event_pid(event, comm_event->task); in perf_event_comm_output()
7638 comm_event->event_id.tid = perf_event_tid(event, comm_event->task); in perf_event_comm_output()
7640 perf_output_put(&handle, comm_event->event_id); in perf_event_comm_output()
7641 __output_copy(&handle, comm_event->comm, in perf_event_comm_output()
7642 comm_event->comm_size); in perf_event_comm_output()
7648 comm_event->event_id.header.size = size; in perf_event_comm_output()
7657 strlcpy(comm, comm_event->task->comm, sizeof(comm)); in perf_event_comm_event()
7660 comm_event->comm = comm; in perf_event_comm_event()
7661 comm_event->comm_size = size; in perf_event_comm_event()
7663 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; in perf_event_comm_event()
7714 return event->attr.namespaces; in perf_event_namespaces_match()
7723 u16 header_size = namespaces_event->event_id.header.size; in perf_event_namespaces_output()
7729 perf_event_header__init_id(&namespaces_event->event_id.header, in perf_event_namespaces_output()
7732 namespaces_event->event_id.header.size); in perf_event_namespaces_output()
7736 namespaces_event->event_id.pid = perf_event_pid(event, in perf_event_namespaces_output()
7737 namespaces_event->task); in perf_event_namespaces_output()
7738 namespaces_event->event_id.tid = perf_event_tid(event, in perf_event_namespaces_output()
7739 namespaces_event->task); in perf_event_namespaces_output()
7741 perf_output_put(&handle, namespaces_event->event_id); in perf_event_namespaces_output()
7747 namespaces_event->event_id.header.size = header_size; in perf_event_namespaces_output()
7760 ns_inode = ns_path.dentry->d_inode; in perf_fill_ns_link_info()
7761 ns_link_info->dev = new_encode_dev(ns_inode->i_sb->s_dev); in perf_fill_ns_link_info()
7762 ns_link_info->ino = ns_inode->i_ino; in perf_fill_ns_link_info()
7842 return event->attr.cgroup; in perf_event_cgroup_match()
7850 u16 header_size = cgroup_event->event_id.header.size; in perf_event_cgroup_output()
7856 perf_event_header__init_id(&cgroup_event->event_id.header, in perf_event_cgroup_output()
7859 cgroup_event->event_id.header.size); in perf_event_cgroup_output()
7863 perf_output_put(&handle, cgroup_event->event_id); in perf_event_cgroup_output()
7864 __output_copy(&handle, cgroup_event->path, cgroup_event->path_size); in perf_event_cgroup_output()
7870 cgroup_event->event_id.header.size = header_size; in perf_event_cgroup_output()
7899 cgroup_path(cgrp, pathname, PATH_MAX - sizeof(u64)); in perf_event_cgroup()
7953 struct vm_area_struct *vma = mmap_event->vma; in perf_event_mmap_match()
7954 int executable = vma->vm_flags & VM_EXEC; in perf_event_mmap_match()
7956 return (!executable && event->attr.mmap_data) || in perf_event_mmap_match()
7957 (executable && (event->attr.mmap || event->attr.mmap2)); in perf_event_mmap_match()
7966 int size = mmap_event->event_id.header.size; in perf_event_mmap_output()
7967 u32 type = mmap_event->event_id.header.type; in perf_event_mmap_output()
7973 if (event->attr.mmap2) { in perf_event_mmap_output()
7974 mmap_event->event_id.header.type = PERF_RECORD_MMAP2; in perf_event_mmap_output()
7975 mmap_event->event_id.header.size += sizeof(mmap_event->maj); in perf_event_mmap_output()
7976 mmap_event->event_id.header.size += sizeof(mmap_event->min); in perf_event_mmap_output()
7977 mmap_event->event_id.header.size += sizeof(mmap_event->ino); in perf_event_mmap_output()
7978 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation); in perf_event_mmap_output()
7979 mmap_event->event_id.header.size += sizeof(mmap_event->prot); in perf_event_mmap_output()
7980 mmap_event->event_id.header.size += sizeof(mmap_event->flags); in perf_event_mmap_output()
7983 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); in perf_event_mmap_output()
7985 mmap_event->event_id.header.size); in perf_event_mmap_output()
7989 mmap_event->event_id.pid = perf_event_pid(event, current); in perf_event_mmap_output()
7990 mmap_event->event_id.tid = perf_event_tid(event, current); in perf_event_mmap_output()
7992 perf_output_put(&handle, mmap_event->event_id); in perf_event_mmap_output()
7994 if (event->attr.mmap2) { in perf_event_mmap_output()
7995 perf_output_put(&handle, mmap_event->maj); in perf_event_mmap_output()
7996 perf_output_put(&handle, mmap_event->min); in perf_event_mmap_output()
7997 perf_output_put(&handle, mmap_event->ino); in perf_event_mmap_output()
7998 perf_output_put(&handle, mmap_event->ino_generation); in perf_event_mmap_output()
7999 perf_output_put(&handle, mmap_event->prot); in perf_event_mmap_output()
8000 perf_output_put(&handle, mmap_event->flags); in perf_event_mmap_output()
8003 __output_copy(&handle, mmap_event->file_name, in perf_event_mmap_output()
8004 mmap_event->file_size); in perf_event_mmap_output()
8010 mmap_event->event_id.header.size = size; in perf_event_mmap_output()
8011 mmap_event->event_id.header.type = type; in perf_event_mmap_output()
8016 struct vm_area_struct *vma = mmap_event->vma; in perf_event_mmap_event()
8017 struct file *file = vma->vm_file; in perf_event_mmap_event()
8026 if (vma->vm_flags & VM_READ) in perf_event_mmap_event()
8028 if (vma->vm_flags & VM_WRITE) in perf_event_mmap_event()
8030 if (vma->vm_flags & VM_EXEC) in perf_event_mmap_event()
8033 if (vma->vm_flags & VM_MAYSHARE) in perf_event_mmap_event()
8038 if (vma->vm_flags & VM_DENYWRITE) in perf_event_mmap_event()
8040 if (vma->vm_flags & VM_MAYEXEC) in perf_event_mmap_event()
8042 if (vma->vm_flags & VM_LOCKED) in perf_event_mmap_event()
8061 name = file_path(file, buf, PATH_MAX - sizeof(u64)); in perf_event_mmap_event()
8066 inode = file_inode(vma->vm_file); in perf_event_mmap_event()
8067 dev = inode->i_sb->s_dev; in perf_event_mmap_event()
8068 ino = inode->i_ino; in perf_event_mmap_event()
8069 gen = inode->i_generation; in perf_event_mmap_event()
8075 if (vma->vm_ops && vma->vm_ops->name) { in perf_event_mmap_event()
8076 name = (char *) vma->vm_ops->name(vma); in perf_event_mmap_event()
8085 if (vma->vm_start <= vma->vm_mm->start_brk && in perf_event_mmap_event()
8086 vma->vm_end >= vma->vm_mm->brk) { in perf_event_mmap_event()
8090 if (vma->vm_start <= vma->vm_mm->start_stack && in perf_event_mmap_event()
8091 vma->vm_end >= vma->vm_mm->start_stack) { in perf_event_mmap_event()
8113 mmap_event->file_name = name; in perf_event_mmap_event()
8114 mmap_event->file_size = size; in perf_event_mmap_event()
8115 mmap_event->maj = maj; in perf_event_mmap_event()
8116 mmap_event->min = min; in perf_event_mmap_event()
8117 mmap_event->ino = ino; in perf_event_mmap_event()
8118 mmap_event->ino_generation = gen; in perf_event_mmap_event()
8119 mmap_event->prot = prot; in perf_event_mmap_event()
8120 mmap_event->flags = flags; in perf_event_mmap_event()
8122 if (!(vma->vm_flags & VM_EXEC)) in perf_event_mmap_event()
8123 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; in perf_event_mmap_event()
8125 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; in perf_event_mmap_event()
8141 /* d_inode(NULL) won't be equal to any mapped user-space file */ in perf_addr_filter_match()
8142 if (!filter->path.dentry) in perf_addr_filter_match()
8145 if (d_inode(filter->path.dentry) != file_inode(file)) in perf_addr_filter_match()
8148 if (filter->offset > offset + size) in perf_addr_filter_match()
8151 if (filter->offset + filter->size < offset) in perf_addr_filter_match()
8161 unsigned long vma_size = vma->vm_end - vma->vm_start; in perf_addr_filter_vma_adjust()
8162 unsigned long off = vma->vm_pgoff << PAGE_SHIFT; in perf_addr_filter_vma_adjust()
8163 struct file *file = vma->vm_file; in perf_addr_filter_vma_adjust()
8168 if (filter->offset < off) { in perf_addr_filter_vma_adjust()
8169 fr->start = vma->vm_start; in perf_addr_filter_vma_adjust()
8170 fr->size = min(vma_size, filter->size - (off - filter->offset)); in perf_addr_filter_vma_adjust()
8172 fr->start = vma->vm_start + filter->offset - off; in perf_addr_filter_vma_adjust()
8173 fr->size = min(vma->vm_end - fr->start, filter->size); in perf_addr_filter_vma_adjust()
8190 if (!vma->vm_file) in __perf_addr_filters_adjust()
8193 raw_spin_lock_irqsave(&ifh->lock, flags); in __perf_addr_filters_adjust()
8194 list_for_each_entry(filter, &ifh->list, entry) { in __perf_addr_filters_adjust()
8196 &event->addr_filter_ranges[count])) in __perf_addr_filters_adjust()
8203 event->addr_filters_gen++; in __perf_addr_filters_adjust()
8204 raw_spin_unlock_irqrestore(&ifh->lock, flags); in __perf_addr_filters_adjust()
8222 if (!(vma->vm_flags & VM_EXEC)) in perf_addr_filters_adjust()
8227 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); in perf_addr_filters_adjust()
8255 .start = vma->vm_start, in perf_event_mmap()
8256 .len = vma->vm_end - vma->vm_start, in perf_event_mmap()
8257 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, in perf_event_mmap()
8355 return event->attr.context_switch; in perf_event_switch_match()
8368 /* Only CPU-wide events are allowed to see next/prev pid/tid */ in perf_event_switch_output()
8369 if (event->ctx->task) { in perf_event_switch_output()
8370 se->event_id.header.type = PERF_RECORD_SWITCH; in perf_event_switch_output()
8371 se->event_id.header.size = sizeof(se->event_id.header); in perf_event_switch_output()
8373 se->event_id.header.type = PERF_RECORD_SWITCH_CPU_WIDE; in perf_event_switch_output()
8374 se->event_id.header.size = sizeof(se->event_id); in perf_event_switch_output()
8375 se->event_id.next_prev_pid = in perf_event_switch_output()
8376 perf_event_pid(event, se->next_prev); in perf_event_switch_output()
8377 se->event_id.next_prev_tid = in perf_event_switch_output()
8378 perf_event_tid(event, se->next_prev); in perf_event_switch_output()
8381 perf_event_header__init_id(&se->event_id.header, &sample, event); in perf_event_switch_output()
8383 ret = perf_output_begin(&handle, &sample, event, se->event_id.header.size); in perf_event_switch_output()
8387 if (event->ctx->task) in perf_event_switch_output()
8388 perf_output_put(&handle, se->event_id.header); in perf_event_switch_output()
8390 perf_output_put(&handle, se->event_id); in perf_event_switch_output()
8418 if (!sched_in && task->state == TASK_RUNNING) in perf_event_switch()
8450 .stream_id = event->id, in perf_log_throttle()
8486 return event->attr.ksymbol; in perf_event_ksymbol_match()
8499 perf_event_header__init_id(&ksymbol_event->event_id.header, in perf_event_ksymbol_output()
8502 ksymbol_event->event_id.header.size); in perf_event_ksymbol_output()
8506 perf_output_put(&handle, ksymbol_event->event_id); in perf_event_ksymbol_output()
8507 __output_copy(&handle, ksymbol_event->name, ksymbol_event->name_len); in perf_event_ksymbol_output()
8576 return event->attr.bpf_event; in perf_event_bpf_match()
8589 perf_event_header__init_id(&bpf_event->event_id.header, in perf_event_bpf_output()
8592 bpf_event->event_id.header.size); in perf_event_bpf_output()
8596 perf_output_put(&handle, bpf_event->event_id); in perf_event_bpf_output()
8608 if (prog->aux->func_cnt == 0) { in perf_event_bpf_emit_ksymbols()
8610 (u64)(unsigned long)prog->bpf_func, in perf_event_bpf_emit_ksymbols()
8611 prog->jited_len, unregister, in perf_event_bpf_emit_ksymbols()
8612 prog->aux->ksym.name); in perf_event_bpf_emit_ksymbols()
8614 for (i = 0; i < prog->aux->func_cnt; i++) { in perf_event_bpf_emit_ksymbols()
8615 struct bpf_prog *subprog = prog->aux->func[i]; in perf_event_bpf_emit_ksymbols()
8619 (u64)(unsigned long)subprog->bpf_func, in perf_event_bpf_emit_ksymbols()
8620 subprog->jited_len, unregister, in perf_event_bpf_emit_ksymbols()
8621 prog->aux->ksym.name); in perf_event_bpf_emit_ksymbols()
8658 .id = prog->aux->id, in perf_event_bpf_event()
8664 memcpy(bpf_event.event_id.tag, prog->tag, BPF_TAG_SIZE); in perf_event_bpf_event()
8684 return event->attr.text_poke; in perf_event_text_poke_match()
8698 perf_event_header__init_id(&text_poke_event->event_id.header, &sample, event); in perf_event_text_poke_output()
8701 text_poke_event->event_id.header.size); in perf_event_text_poke_output()
8705 perf_output_put(&handle, text_poke_event->event_id); in perf_event_text_poke_output()
8706 perf_output_put(&handle, text_poke_event->old_len); in perf_event_text_poke_output()
8707 perf_output_put(&handle, text_poke_event->new_len); in perf_event_text_poke_output()
8709 __output_copy(&handle, text_poke_event->old_bytes, text_poke_event->old_len); in perf_event_text_poke_output()
8710 __output_copy(&handle, text_poke_event->new_bytes, text_poke_event->new_len); in perf_event_text_poke_output()
8712 if (text_poke_event->pad) in perf_event_text_poke_output()
8713 __output_copy(&handle, &padding, text_poke_event->pad); in perf_event_text_poke_output()
8731 pad = ALIGN(tot, sizeof(u64)) - tot; in perf_event_text_poke()
8754 event->attach_state |= PERF_ATTACH_ITRACE; in perf_event_itrace_started()
8768 if (event->parent) in perf_log_itrace_start()
8769 event = event->parent; in perf_log_itrace_start()
8771 if (!(event->pmu->capabilities & PERF_PMU_CAP_ITRACE) || in perf_log_itrace_start()
8772 event->attach_state & PERF_ATTACH_ITRACE) in perf_log_itrace_start()
8796 struct hw_perf_event *hwc = &event->hw; in __perf_event_account_interrupt()
8801 if (seq != hwc->interrupts_seq) { in __perf_event_account_interrupt()
8802 hwc->interrupts_seq = seq; in __perf_event_account_interrupt()
8803 hwc->interrupts = 1; in __perf_event_account_interrupt()
8805 hwc->interrupts++; in __perf_event_account_interrupt()
8807 && hwc->interrupts >= max_samples_per_tick)) { in __perf_event_account_interrupt()
8810 hwc->interrupts = MAX_INTERRUPTS; in __perf_event_account_interrupt()
8816 if (event->attr.freq) { in __perf_event_account_interrupt()
8818 s64 delta = now - hwc->freq_time_stamp; in __perf_event_account_interrupt()
8820 hwc->freq_time_stamp = now; in __perf_event_account_interrupt()
8823 perf_adjust_period(event, delta, hwc->last_period, true); in __perf_event_account_interrupt()
8842 int events = atomic_read(&event->event_limit); in __perf_event_overflow()
8846 * Non-sampling counters might still use the PMI to fold short in __perf_event_overflow()
8859 event->pending_kill = POLL_IN; in __perf_event_overflow()
8860 if (events && atomic_dec_and_test(&event->event_limit)) { in __perf_event_overflow()
8862 event->pending_kill = POLL_HUP; in __perf_event_overflow()
8867 READ_ONCE(event->overflow_handler)(event, data, regs); in __perf_event_overflow()
8869 if (*perf_event_fasync(event) && event->pending_kill) { in __perf_event_overflow()
8870 event->pending_wakeup = 1; in __perf_event_overflow()
8871 irq_work_queue(&event->pending); in __perf_event_overflow()
8900 * We directly increment event->count and keep a second value in
8901 * event->hw.period_left to count intervals. This period event
8902 * is kept in the range [-sample_period, 0] so that we can use the
8908 struct hw_perf_event *hwc = &event->hw; in perf_swevent_set_period()
8909 u64 period = hwc->last_period; in perf_swevent_set_period()
8913 hwc->last_period = hwc->sample_period; in perf_swevent_set_period()
8916 old = val = local64_read(&hwc->period_left); in perf_swevent_set_period()
8922 val -= offset; in perf_swevent_set_period()
8923 if (local64_cmpxchg(&hwc->period_left, old, val) != old) in perf_swevent_set_period()
8933 struct hw_perf_event *hwc = &event->hw; in perf_swevent_overflow()
8939 if (hwc->interrupts == MAX_INTERRUPTS) in perf_swevent_overflow()
8942 for (; overflow; overflow--) { in perf_swevent_overflow()
8947 * hwc->interrupts == MAX_INTERRUPTS. in perf_swevent_overflow()
8959 struct hw_perf_event *hwc = &event->hw; in perf_swevent_event()
8961 local64_add(nr, &event->count); in perf_swevent_event()
8969 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { in perf_swevent_event()
8970 data->period = nr; in perf_swevent_event()
8973 data->period = event->hw.last_period; in perf_swevent_event()
8975 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) in perf_swevent_event()
8978 if (local64_add_negative(nr, &hwc->period_left)) in perf_swevent_event()
8987 if (event->hw.state & PERF_HES_STOPPED) in perf_exclude_event()
8991 if (event->attr.exclude_user && user_mode(regs)) in perf_exclude_event()
8994 if (event->attr.exclude_kernel && !user_mode(regs)) in perf_exclude_event()
9007 if (event->attr.type != type) in perf_swevent_match()
9010 if (event->attr.config != event_id) in perf_swevent_match()
9031 return &hlist->heads[hash]; in __find_swevent_head()
9040 hlist = rcu_dereference(swhash->swevent_hlist); in find_swevent_head_rcu()
9052 u32 event_id = event->attr.config; in find_swevent_head()
9053 u64 type = event->attr.type; in find_swevent_head()
9060 hlist = rcu_dereference_protected(swhash->swevent_hlist, in find_swevent_head()
9061 lockdep_is_held(&event->ctx->lock)); in find_swevent_head()
9096 return get_recursion_context(swhash->recursion); in perf_swevent_get_recursion_context()
9104 put_recursion_context(swhash->recursion, rctx); in perf_swevent_put_recursion_context()
9141 struct hw_perf_event *hwc = &event->hw; in perf_swevent_add()
9145 hwc->last_period = hwc->sample_period; in perf_swevent_add()
9149 hwc->state = !(flags & PERF_EF_START); in perf_swevent_add()
9153 return -EINVAL; in perf_swevent_add()
9155 hlist_add_head_rcu(&event->hlist_entry, head); in perf_swevent_add()
9163 hlist_del_rcu(&event->hlist_entry); in perf_swevent_del()
9168 event->hw.state = 0; in perf_swevent_start()
9173 event->hw.state = PERF_HES_STOPPED; in perf_swevent_stop()
9180 return rcu_dereference_protected(swhash->swevent_hlist, in swevent_hlist_deref()
9181 lockdep_is_held(&swhash->hlist_mutex)); in swevent_hlist_deref()
9191 RCU_INIT_POINTER(swhash->swevent_hlist, NULL); in swevent_hlist_release()
9199 mutex_lock(&swhash->hlist_mutex); in swevent_hlist_put_cpu()
9201 if (!--swhash->hlist_refcount) in swevent_hlist_put_cpu()
9204 mutex_unlock(&swhash->hlist_mutex); in swevent_hlist_put_cpu()
9220 mutex_lock(&swhash->hlist_mutex); in swevent_hlist_get_cpu()
9227 err = -ENOMEM; in swevent_hlist_get_cpu()
9230 rcu_assign_pointer(swhash->swevent_hlist, hlist); in swevent_hlist_get_cpu()
9232 swhash->hlist_refcount++; in swevent_hlist_get_cpu()
9234 mutex_unlock(&swhash->hlist_mutex); in swevent_hlist_get_cpu()
9267 u64 event_id = event->attr.config; in sw_perf_event_destroy()
9269 WARN_ON(event->parent); in sw_perf_event_destroy()
9277 u64 event_id = event->attr.config; in perf_swevent_init()
9279 if (event->attr.type != PERF_TYPE_SOFTWARE) in perf_swevent_init()
9280 return -ENOENT; in perf_swevent_init()
9286 return -EOPNOTSUPP; in perf_swevent_init()
9291 return -ENOENT; in perf_swevent_init()
9298 return -ENOENT; in perf_swevent_init()
9300 if (!event->parent) { in perf_swevent_init()
9308 event->destroy = sw_perf_event_destroy; in perf_swevent_init()
9332 void *record = data->raw->frag.data; in perf_tp_filter_match()
9335 if (event->parent) in perf_tp_filter_match()
9336 event = event->parent; in perf_tp_filter_match()
9338 if (likely(!event->filter) || filter_match_preds(event->filter, record)) in perf_tp_filter_match()
9347 if (event->hw.state & PERF_HES_STOPPED) in perf_tp_event_match()
9350 * If exclude_kernel, only trace user-space tracepoints (uprobes) in perf_tp_event_match()
9352 if (event->attr.exclude_kernel && !user_mode(regs)) in perf_tp_event_match()
9373 perf_tp_event(call->event.type, count, raw_data, size, regs, head, in perf_trace_run_bpf_submit()
9411 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); in perf_tp_event()
9415 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { in perf_tp_event()
9416 if (event->cpu != smp_processor_id()) in perf_tp_event()
9418 if (event->attr.type != PERF_TYPE_TRACEPOINT) in perf_tp_event()
9420 if (event->attr.config != entry->type) in perf_tp_event()
9442 if (event->attr.type != PERF_TYPE_TRACEPOINT) in perf_tp_event_init()
9443 return -ENOENT; in perf_tp_event_init()
9449 return -EOPNOTSUPP; in perf_tp_event_init()
9455 event->destroy = tp_perf_event_destroy; in perf_tp_event_init()
9489 PERF_UPROBE_REF_CTR_OFFSET_SHIFT = 64 - PERF_UPROBE_REF_CTR_OFFSET_BITS,
9528 if (event->attr.type != perf_kprobe.type) in perf_kprobe_event_init()
9529 return -ENOENT; in perf_kprobe_event_init()
9532 return -EACCES; in perf_kprobe_event_init()
9538 return -EOPNOTSUPP; in perf_kprobe_event_init()
9540 is_retprobe = event->attr.config & PERF_PROBE_CONFIG_IS_RETPROBE; in perf_kprobe_event_init()
9545 event->destroy = perf_kprobe_destroy; in perf_kprobe_event_init()
9552 PMU_FORMAT_ATTR(ref_ctr_offset, "config:32-63");
9588 if (event->attr.type != perf_uprobe.type) in perf_uprobe_event_init()
9589 return -ENOENT; in perf_uprobe_event_init()
9592 return -EACCES; in perf_uprobe_event_init()
9598 return -EOPNOTSUPP; in perf_uprobe_event_init()
9600 is_retprobe = event->attr.config & PERF_PROBE_CONFIG_IS_RETPROBE; in perf_uprobe_event_init()
9601 ref_ctr_offset = event->attr.config >> PERF_UPROBE_REF_CTR_OFFSET_SHIFT; in perf_uprobe_event_init()
9606 event->destroy = perf_uprobe_destroy; in perf_uprobe_event_init()
9616 perf_pmu_register(&perf_kprobe, "kprobe", -1); in perf_tp_register()
9619 perf_pmu_register(&perf_uprobe, "uprobe", -1); in perf_tp_register()
9643 ret = BPF_PROG_RUN(event->prog, &ctx); in bpf_overflow_handler()
9650 event->orig_overflow_handler(event, data, regs); in bpf_overflow_handler()
9657 if (event->overflow_handler_context) in perf_event_set_bpf_handler()
9659 return -EINVAL; in perf_event_set_bpf_handler()
9661 if (event->prog) in perf_event_set_bpf_handler()
9662 return -EEXIST; in perf_event_set_bpf_handler()
9668 if (event->attr.precise_ip && in perf_event_set_bpf_handler()
9669 prog->call_get_stack && in perf_event_set_bpf_handler()
9670 (!(event->attr.sample_type & __PERF_SAMPLE_CALLCHAIN_EARLY) || in perf_event_set_bpf_handler()
9671 event->attr.exclude_callchain_kernel || in perf_event_set_bpf_handler()
9672 event->attr.exclude_callchain_user)) { in perf_event_set_bpf_handler()
9683 return -EPROTO; in perf_event_set_bpf_handler()
9686 event->prog = prog; in perf_event_set_bpf_handler()
9687 event->orig_overflow_handler = READ_ONCE(event->overflow_handler); in perf_event_set_bpf_handler()
9688 WRITE_ONCE(event->overflow_handler, bpf_overflow_handler); in perf_event_set_bpf_handler()
9694 struct bpf_prog *prog = event->prog; in perf_event_free_bpf_handler()
9699 WRITE_ONCE(event->overflow_handler, event->orig_overflow_handler); in perf_event_free_bpf_handler()
9700 event->prog = NULL; in perf_event_free_bpf_handler()
9706 return -EOPNOTSUPP; in perf_event_set_bpf_handler()
9719 if (event->pmu == &perf_tracepoint) in perf_event_is_tracing()
9722 if (event->pmu == &perf_kprobe) in perf_event_is_tracing()
9726 if (event->pmu == &perf_uprobe) in perf_event_is_tracing()
9741 is_kprobe = event->tp_event->flags & TRACE_EVENT_FL_UKPROBE; in perf_event_set_bpf_prog()
9742 is_tracepoint = event->tp_event->flags & TRACE_EVENT_FL_TRACEPOINT; in perf_event_set_bpf_prog()
9743 is_syscall_tp = is_syscall_trace_event(event->tp_event); in perf_event_set_bpf_prog()
9746 return -EINVAL; in perf_event_set_bpf_prog()
9752 if ((is_kprobe && prog->type != BPF_PROG_TYPE_KPROBE) || in perf_event_set_bpf_prog()
9753 (is_tracepoint && prog->type != BPF_PROG_TYPE_TRACEPOINT) || in perf_event_set_bpf_prog()
9754 (is_syscall_tp && prog->type != BPF_PROG_TYPE_TRACEPOINT)) { in perf_event_set_bpf_prog()
9757 return -EINVAL; in perf_event_set_bpf_prog()
9761 if (prog->kprobe_override && in perf_event_set_bpf_prog()
9762 !(event->tp_event->flags & TRACE_EVENT_FL_KPROBE)) { in perf_event_set_bpf_prog()
9764 return -EINVAL; in perf_event_set_bpf_prog()
9768 int off = trace_event_get_offsets(event->tp_event); in perf_event_set_bpf_prog()
9770 if (prog->aux->max_ctx_offset > off) { in perf_event_set_bpf_prog()
9772 return -EACCES; in perf_event_set_bpf_prog()
9803 return -ENOENT; in perf_event_set_bpf_prog()
9817 perf_sample_data_init(&sample, bp->attr.bp_addr, 0); in perf_bp_event()
9819 if (!bp->hw.state && !perf_exclude_event(bp, regs)) in perf_bp_event()
9830 int node = cpu_to_node(event->cpu == -1 ? 0 : event->cpu); in perf_addr_filter_new()
9837 INIT_LIST_HEAD(&filter->entry); in perf_addr_filter_new()
9838 list_add_tail(&filter->entry, filters); in perf_addr_filter_new()
9848 path_put(&filter->path); in free_filters_list()
9849 list_del(&filter->entry); in free_filters_list()
9867 if (event->parent) in perf_addr_filters_splice()
9870 raw_spin_lock_irqsave(&event->addr_filters.lock, flags); in perf_addr_filters_splice()
9872 list_splice_init(&event->addr_filters.list, &list); in perf_addr_filters_splice()
9874 list_splice(head, &event->addr_filters.list); in perf_addr_filters_splice()
9876 raw_spin_unlock_irqrestore(&event->addr_filters.lock, flags); in perf_addr_filters_splice()
9892 for (vma = mm->mmap; vma; vma = vma->vm_next) { in perf_addr_filter_apply()
9893 if (!vma->vm_file) in perf_addr_filter_apply()
9908 struct task_struct *task = READ_ONCE(event->ctx->task); in perf_event_addr_filters_apply()
9915 * We may observe TASK_TOMBSTONE, which means that the event tear-down in perf_event_addr_filters_apply()
9921 if (ifh->nr_file_filters) { in perf_event_addr_filters_apply()
9922 mm = get_task_mm(event->ctx->task); in perf_event_addr_filters_apply()
9929 raw_spin_lock_irqsave(&ifh->lock, flags); in perf_event_addr_filters_apply()
9930 list_for_each_entry(filter, &ifh->list, entry) { in perf_event_addr_filters_apply()
9931 if (filter->path.dentry) { in perf_event_addr_filters_apply()
9936 event->addr_filter_ranges[count].start = 0; in perf_event_addr_filters_apply()
9937 event->addr_filter_ranges[count].size = 0; in perf_event_addr_filters_apply()
9939 perf_addr_filter_apply(filter, mm, &event->addr_filter_ranges[count]); in perf_event_addr_filters_apply()
9941 event->addr_filter_ranges[count].start = filter->offset; in perf_event_addr_filters_apply()
9942 event->addr_filter_ranges[count].size = filter->size; in perf_event_addr_filters_apply()
9948 event->addr_filters_gen++; in perf_event_addr_filters_apply()
9949 raw_spin_unlock_irqrestore(&ifh->lock, flags); in perf_event_addr_filters_apply()
9951 if (ifh->nr_file_filters) { in perf_event_addr_filters_apply()
9981 IF_ACT_NONE = -1,
10020 int ret = -EINVAL; in perf_event_parse_addr_filter()
10024 return -ENOMEM; in perf_event_parse_addr_filter()
10032 ret = -EINVAL; in perf_event_parse_addr_filter()
10052 filter->action = actions[token]; in perf_event_parse_addr_filter()
10067 ret = kstrtoul(args[0].from, 0, &filter->offset); in perf_event_parse_addr_filter()
10073 ret = kstrtoul(args[1].from, 0, &filter->size); in perf_event_parse_addr_filter()
10084 ret = -ENOMEM; in perf_event_parse_addr_filter()
10102 ret = -EINVAL; in perf_event_parse_addr_filter()
10103 if (kernel && event->attr.exclude_kernel) in perf_event_parse_addr_filter()
10107 * ACTION "filter" must have a non-zero length region in perf_event_parse_addr_filter()
10110 if (filter->action == PERF_ADDR_FILTER_ACTION_FILTER && in perf_event_parse_addr_filter()
10111 !filter->size) in perf_event_parse_addr_filter()
10119 * For now, we only support file-based filters in perf_event_parse_addr_filter()
10120 * in per-task events; doing so for CPU-wide in perf_event_parse_addr_filter()
10126 ret = -EOPNOTSUPP; in perf_event_parse_addr_filter()
10127 if (!event->ctx->task) in perf_event_parse_addr_filter()
10132 &filter->path); in perf_event_parse_addr_filter()
10136 ret = -EINVAL; in perf_event_parse_addr_filter()
10137 if (!filter->path.dentry || in perf_event_parse_addr_filter()
10138 !S_ISREG(d_inode(filter->path.dentry) in perf_event_parse_addr_filter()
10139 ->i_mode)) in perf_event_parse_addr_filter()
10142 event->addr_filters.nr_file_filters++; in perf_event_parse_addr_filter()
10177 lockdep_assert_held(&event->ctx->mutex); in perf_event_set_addr_filter()
10179 if (WARN_ON_ONCE(event->parent)) in perf_event_set_addr_filter()
10180 return -EINVAL; in perf_event_set_addr_filter()
10186 ret = event->pmu->addr_filters_validate(&filters); in perf_event_set_addr_filter()
10202 event->addr_filters.nr_file_filters = 0; in perf_event_set_addr_filter()
10209 int ret = -EINVAL; in perf_event_set_filter()
10218 struct perf_event_context *ctx = event->ctx; in perf_event_set_filter()
10223 * the tracepoint muck will deadlock against ctx->mutex, but in perf_event_set_filter()
10225 * temporarily drop ctx->mutex. As per perf_event_ctx_lock() we in perf_event_set_filter()
10231 mutex_unlock(&ctx->mutex); in perf_event_set_filter()
10232 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); in perf_event_set_filter()
10233 mutex_lock(&ctx->mutex); in perf_event_set_filter()
10257 if (event->state != PERF_EVENT_STATE_ACTIVE) in perf_swevent_hrtimer()
10260 event->pmu->read(event); in perf_swevent_hrtimer()
10262 perf_sample_data_init(&data, 0, event->hw.last_period); in perf_swevent_hrtimer()
10266 if (!(event->attr.exclude_idle && is_idle_task(current))) in perf_swevent_hrtimer()
10271 period = max_t(u64, 10000, event->hw.sample_period); in perf_swevent_hrtimer()
10279 struct hw_perf_event *hwc = &event->hw; in perf_swevent_start_hrtimer()
10285 period = local64_read(&hwc->period_left); in perf_swevent_start_hrtimer()
10290 local64_set(&hwc->period_left, 0); in perf_swevent_start_hrtimer()
10292 period = max_t(u64, 10000, hwc->sample_period); in perf_swevent_start_hrtimer()
10294 hrtimer_start(&hwc->hrtimer, ns_to_ktime(period), in perf_swevent_start_hrtimer()
10300 struct hw_perf_event *hwc = &event->hw; in perf_swevent_cancel_hrtimer()
10303 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); in perf_swevent_cancel_hrtimer()
10304 local64_set(&hwc->period_left, ktime_to_ns(remaining)); in perf_swevent_cancel_hrtimer()
10306 hrtimer_cancel(&hwc->hrtimer); in perf_swevent_cancel_hrtimer()
10312 struct hw_perf_event *hwc = &event->hw; in perf_swevent_init_hrtimer()
10317 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD); in perf_swevent_init_hrtimer()
10318 hwc->hrtimer.function = perf_swevent_hrtimer; in perf_swevent_init_hrtimer()
10321 * Since hrtimers have a fixed rate, we can do a static freq->period in perf_swevent_init_hrtimer()
10324 if (event->attr.freq) { in perf_swevent_init_hrtimer()
10325 long freq = event->attr.sample_freq; in perf_swevent_init_hrtimer()
10327 event->attr.sample_period = NSEC_PER_SEC / freq; in perf_swevent_init_hrtimer()
10328 hwc->sample_period = event->attr.sample_period; in perf_swevent_init_hrtimer()
10329 local64_set(&hwc->period_left, hwc->sample_period); in perf_swevent_init_hrtimer()
10330 hwc->last_period = hwc->sample_period; in perf_swevent_init_hrtimer()
10331 event->attr.freq = 0; in perf_swevent_init_hrtimer()
10345 prev = local64_xchg(&event->hw.prev_count, now); in cpu_clock_event_update()
10346 local64_add(now - prev, &event->count); in cpu_clock_event_update()
10351 local64_set(&event->hw.prev_count, local_clock()); in cpu_clock_event_start()
10382 if (event->attr.type != PERF_TYPE_SOFTWARE) in cpu_clock_event_init()
10383 return -ENOENT; in cpu_clock_event_init()
10385 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) in cpu_clock_event_init()
10386 return -ENOENT; in cpu_clock_event_init()
10392 return -EOPNOTSUPP; in cpu_clock_event_init()
10421 prev = local64_xchg(&event->hw.prev_count, now); in task_clock_event_update()
10422 delta = now - prev; in task_clock_event_update()
10423 local64_add(delta, &event->count); in task_clock_event_update()
10428 local64_set(&event->hw.prev_count, event->ctx->time); in task_clock_event_start()
10435 task_clock_event_update(event, event->ctx->time); in task_clock_event_stop()
10455 u64 delta = now - event->ctx->timestamp; in task_clock_event_read()
10456 u64 time = event->ctx->time + delta; in task_clock_event_read()
10463 if (event->attr.type != PERF_TYPE_SOFTWARE) in task_clock_event_init()
10464 return -ENOENT; in task_clock_event_init()
10466 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) in task_clock_event_init()
10467 return -ENOENT; in task_clock_event_init()
10473 return -EOPNOTSUPP; in task_clock_event_init()
10565 if (pmu->task_ctx_nr == ctxn) in find_pmu_context()
10566 return pmu->pmu_cpu_context; in find_pmu_context()
10579 if (pmu->task_ctx_nr > perf_invalid_context) in free_pmu_context()
10582 free_percpu(pmu->pmu_cpu_context); in free_pmu_context()
10594 return snprintf(page, PAGE_SIZE - 1, "%d\n", pmu->nr_addr_filters); in nr_addr_filters_show()
10605 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); in type_show()
10616 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); in perf_event_mux_interval_ms_show()
10634 return -EINVAL; in perf_event_mux_interval_ms_store()
10637 if (timer == pmu->hrtimer_interval_ms) in perf_event_mux_interval_ms_store()
10641 pmu->hrtimer_interval_ms = timer; in perf_event_mux_interval_ms_store()
10647 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); in perf_event_mux_interval_ms_store()
10648 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); in perf_event_mux_interval_ms_store()
10680 int ret = -ENOMEM; in pmu_dev_alloc()
10682 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); in pmu_dev_alloc()
10683 if (!pmu->dev) in pmu_dev_alloc()
10686 pmu->dev->groups = pmu->attr_groups; in pmu_dev_alloc()
10687 device_initialize(pmu->dev); in pmu_dev_alloc()
10688 ret = dev_set_name(pmu->dev, "%s", pmu->name); in pmu_dev_alloc()
10692 dev_set_drvdata(pmu->dev, pmu); in pmu_dev_alloc()
10693 pmu->dev->bus = &pmu_bus; in pmu_dev_alloc()
10694 pmu->dev->release = pmu_dev_release; in pmu_dev_alloc()
10695 ret = device_add(pmu->dev); in pmu_dev_alloc()
10700 if (pmu->nr_addr_filters) in pmu_dev_alloc()
10701 ret = device_create_file(pmu->dev, &dev_attr_nr_addr_filters); in pmu_dev_alloc()
10706 if (pmu->attr_update) in pmu_dev_alloc()
10707 ret = sysfs_update_groups(&pmu->dev->kobj, pmu->attr_update); in pmu_dev_alloc()
10716 device_del(pmu->dev); in pmu_dev_alloc()
10719 put_device(pmu->dev); in pmu_dev_alloc()
10731 ret = -ENOMEM; in perf_pmu_register()
10732 pmu->pmu_disable_count = alloc_percpu(int); in perf_pmu_register()
10733 if (!pmu->pmu_disable_count) in perf_pmu_register()
10736 pmu->type = -1; in perf_pmu_register()
10739 pmu->name = name; in perf_pmu_register()
10753 pmu->type = type; in perf_pmu_register()
10762 if (pmu->task_ctx_nr == perf_hw_context) { in perf_pmu_register()
10771 !(pmu->capabilities & PERF_PMU_CAP_HETEROGENEOUS_CPUS))) in perf_pmu_register()
10772 pmu->task_ctx_nr = perf_invalid_context; in perf_pmu_register()
10777 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); in perf_pmu_register()
10778 if (pmu->pmu_cpu_context) in perf_pmu_register()
10781 ret = -ENOMEM; in perf_pmu_register()
10782 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); in perf_pmu_register()
10783 if (!pmu->pmu_cpu_context) in perf_pmu_register()
10789 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); in perf_pmu_register()
10790 __perf_event_init_context(&cpuctx->ctx); in perf_pmu_register()
10791 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); in perf_pmu_register()
10792 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); in perf_pmu_register()
10793 cpuctx->ctx.pmu = pmu; in perf_pmu_register()
10794 cpuctx->online = cpumask_test_cpu(cpu, perf_online_mask); in perf_pmu_register()
10798 cpuctx->heap_size = ARRAY_SIZE(cpuctx->heap_default); in perf_pmu_register()
10799 cpuctx->heap = cpuctx->heap_default; in perf_pmu_register()
10803 if (!pmu->start_txn) { in perf_pmu_register()
10804 if (pmu->pmu_enable) { in perf_pmu_register()
10810 pmu->start_txn = perf_pmu_start_txn; in perf_pmu_register()
10811 pmu->commit_txn = perf_pmu_commit_txn; in perf_pmu_register()
10812 pmu->cancel_txn = perf_pmu_cancel_txn; in perf_pmu_register()
10814 pmu->start_txn = perf_pmu_nop_txn; in perf_pmu_register()
10815 pmu->commit_txn = perf_pmu_nop_int; in perf_pmu_register()
10816 pmu->cancel_txn = perf_pmu_nop_void; in perf_pmu_register()
10820 if (!pmu->pmu_enable) { in perf_pmu_register()
10821 pmu->pmu_enable = perf_pmu_nop_void; in perf_pmu_register()
10822 pmu->pmu_disable = perf_pmu_nop_void; in perf_pmu_register()
10825 if (!pmu->check_period) in perf_pmu_register()
10826 pmu->check_period = perf_event_nop_int; in perf_pmu_register()
10828 if (!pmu->event_idx) in perf_pmu_register()
10829 pmu->event_idx = perf_event_idx_default; in perf_pmu_register()
10837 list_add_rcu(&pmu->entry, &pmus); in perf_pmu_register()
10839 list_add_tail_rcu(&pmu->entry, &pmus); in perf_pmu_register()
10841 atomic_set(&pmu->exclusive_cnt, 0); in perf_pmu_register()
10849 device_del(pmu->dev); in perf_pmu_register()
10850 put_device(pmu->dev); in perf_pmu_register()
10853 if (pmu->type != PERF_TYPE_SOFTWARE) in perf_pmu_register()
10854 idr_remove(&pmu_idr, pmu->type); in perf_pmu_register()
10857 free_percpu(pmu->pmu_disable_count); in perf_pmu_register()
10865 list_del_rcu(&pmu->entry); in perf_pmu_unregister()
10874 free_percpu(pmu->pmu_disable_count); in perf_pmu_unregister()
10875 if (pmu->type != PERF_TYPE_SOFTWARE) in perf_pmu_unregister()
10876 idr_remove(&pmu_idr, pmu->type); in perf_pmu_unregister()
10878 if (pmu->nr_addr_filters) in perf_pmu_unregister()
10879 device_remove_file(pmu->dev, &dev_attr_nr_addr_filters); in perf_pmu_unregister()
10880 device_del(pmu->dev); in perf_pmu_unregister()
10881 put_device(pmu->dev); in perf_pmu_unregister()
10890 return (event->attr.sample_regs_user & PERF_REG_EXTENDED_MASK) || in has_extended_regs()
10891 (event->attr.sample_regs_intr & PERF_REG_EXTENDED_MASK); in has_extended_regs()
10899 if (!try_module_get(pmu->module)) in perf_try_init_event()
10900 return -ENODEV; in perf_try_init_event()
10903 * A number of pmu->event_init() methods iterate the sibling_list to, in perf_try_init_event()
10905 * if this is a sibling event, acquire the ctx->mutex to protect in perf_try_init_event()
10908 if (event->group_leader != event && pmu->task_ctx_nr != perf_sw_context) { in perf_try_init_event()
10910 * This ctx->mutex can nest when we're called through in perf_try_init_event()
10913 ctx = perf_event_ctx_lock_nested(event->group_leader, in perf_try_init_event()
10918 event->pmu = pmu; in perf_try_init_event()
10919 ret = pmu->event_init(event); in perf_try_init_event()
10922 perf_event_ctx_unlock(event->group_leader, ctx); in perf_try_init_event()
10925 if (!(pmu->capabilities & PERF_PMU_CAP_EXTENDED_REGS) && in perf_try_init_event()
10927 ret = -EOPNOTSUPP; in perf_try_init_event()
10929 if (pmu->capabilities & PERF_PMU_CAP_NO_EXCLUDE && in perf_try_init_event()
10931 ret = -EINVAL; in perf_try_init_event()
10933 if (ret && event->destroy) in perf_try_init_event()
10934 event->destroy(event); in perf_try_init_event()
10938 module_put(pmu->module); in perf_try_init_event()
10951 if (event->parent && event->parent->pmu) { in perf_init_event()
10952 pmu = event->parent->pmu; in perf_init_event()
10962 type = event->attr.type; in perf_init_event()
10972 if (ret == -ENOENT && event->attr.type != type) { in perf_init_event()
10973 type = event->attr.type; in perf_init_event()
10988 if (ret != -ENOENT) { in perf_init_event()
10993 pmu = ERR_PTR(-ENOENT); in perf_init_event()
11002 struct pmu_event_list *pel = per_cpu_ptr(&pmu_sb_events, event->cpu); in attach_sb_event()
11004 raw_spin_lock(&pel->lock); in attach_sb_event()
11005 list_add_rcu(&event->sb_list, &pel->list); in attach_sb_event()
11006 raw_spin_unlock(&pel->lock); in attach_sb_event()
11010 * We keep a list of all !task (and therefore per-cpu) events
11011 * that need to receive side-band records.
11013 * This avoids having to scan all the various PMU per-cpu contexts
11024 if (event->parent) in account_event_cpu()
11056 if (event->parent) in account_event()
11059 if (event->attach_state & PERF_ATTACH_TASK) in account_event()
11061 if (event->attr.mmap || event->attr.mmap_data) in account_event()
11063 if (event->attr.comm) in account_event()
11065 if (event->attr.namespaces) in account_event()
11067 if (event->attr.cgroup) in account_event()
11069 if (event->attr.task) in account_event()
11071 if (event->attr.freq) in account_event()
11073 if (event->attr.context_switch) { in account_event()
11081 if (event->attr.ksymbol) in account_event()
11083 if (event->attr.bpf_event) in account_event()
11085 if (event->attr.text_poke) in account_event()
11109 * increments to by-pass the mutex. in account_event()
11116 account_event_cpu(event, event->cpu); in account_event()
11135 long err = -EINVAL; in perf_event_alloc()
11138 if (!task || cpu != -1) in perf_event_alloc()
11139 return ERR_PTR(-EINVAL); in perf_event_alloc()
11144 return ERR_PTR(-ENOMEM); in perf_event_alloc()
11153 mutex_init(&event->child_mutex); in perf_event_alloc()
11154 INIT_LIST_HEAD(&event->child_list); in perf_event_alloc()
11156 INIT_LIST_HEAD(&event->event_entry); in perf_event_alloc()
11157 INIT_LIST_HEAD(&event->sibling_list); in perf_event_alloc()
11158 INIT_LIST_HEAD(&event->active_list); in perf_event_alloc()
11160 INIT_LIST_HEAD(&event->rb_entry); in perf_event_alloc()
11161 INIT_LIST_HEAD(&event->active_entry); in perf_event_alloc()
11162 INIT_LIST_HEAD(&event->addr_filters.list); in perf_event_alloc()
11163 INIT_HLIST_NODE(&event->hlist_entry); in perf_event_alloc()
11166 init_waitqueue_head(&event->waitq); in perf_event_alloc()
11167 event->pending_disable = -1; in perf_event_alloc()
11168 init_irq_work(&event->pending, perf_pending_event); in perf_event_alloc()
11170 mutex_init(&event->mmap_mutex); in perf_event_alloc()
11171 raw_spin_lock_init(&event->addr_filters.lock); in perf_event_alloc()
11173 atomic_long_set(&event->refcount, 1); in perf_event_alloc()
11174 event->cpu = cpu; in perf_event_alloc()
11175 event->attr = *attr; in perf_event_alloc()
11176 event->group_leader = group_leader; in perf_event_alloc()
11177 event->pmu = NULL; in perf_event_alloc()
11178 event->oncpu = -1; in perf_event_alloc()
11180 event->parent = parent_event; in perf_event_alloc()
11182 event->ns = get_pid_ns(task_active_pid_ns(current)); in perf_event_alloc()
11183 event->id = atomic64_inc_return(&perf_event_id); in perf_event_alloc()
11185 event->state = PERF_EVENT_STATE_INACTIVE; in perf_event_alloc()
11188 event->attach_state = PERF_ATTACH_TASK; in perf_event_alloc()
11194 event->hw.target = get_task_struct(task); in perf_event_alloc()
11197 event->clock = &local_clock; in perf_event_alloc()
11199 event->clock = parent_event->clock; in perf_event_alloc()
11202 overflow_handler = parent_event->overflow_handler; in perf_event_alloc()
11203 context = parent_event->overflow_handler_context; in perf_event_alloc()
11206 struct bpf_prog *prog = parent_event->prog; in perf_event_alloc()
11209 event->prog = prog; in perf_event_alloc()
11210 event->orig_overflow_handler = in perf_event_alloc()
11211 parent_event->orig_overflow_handler; in perf_event_alloc()
11217 event->overflow_handler = overflow_handler; in perf_event_alloc()
11218 event->overflow_handler_context = context; in perf_event_alloc()
11220 event->overflow_handler = perf_event_output_backward; in perf_event_alloc()
11221 event->overflow_handler_context = NULL; in perf_event_alloc()
11223 event->overflow_handler = perf_event_output_forward; in perf_event_alloc()
11224 event->overflow_handler_context = NULL; in perf_event_alloc()
11231 hwc = &event->hw; in perf_event_alloc()
11232 hwc->sample_period = attr->sample_period; in perf_event_alloc()
11233 if (attr->freq && attr->sample_freq) in perf_event_alloc()
11234 hwc->sample_period = 1; in perf_event_alloc()
11235 hwc->last_period = hwc->sample_period; in perf_event_alloc()
11237 local64_set(&hwc->period_left, hwc->sample_period); in perf_event_alloc()
11243 if (attr->inherit && (attr->sample_type & PERF_SAMPLE_READ)) in perf_event_alloc()
11247 event->attr.branch_sample_type = 0; in perf_event_alloc()
11256 * Disallow uncore-cgroup events, they don't make sense as the cgroup will in perf_event_alloc()
11259 if (pmu->task_ctx_nr == perf_invalid_context && cgroup_fd != -1) { in perf_event_alloc()
11260 err = -EINVAL; in perf_event_alloc()
11264 if (event->attr.aux_output && in perf_event_alloc()
11265 !(pmu->capabilities & PERF_PMU_CAP_AUX_OUTPUT)) { in perf_event_alloc()
11266 err = -EOPNOTSUPP; in perf_event_alloc()
11270 if (cgroup_fd != -1) { in perf_event_alloc()
11281 event->addr_filter_ranges = kcalloc(pmu->nr_addr_filters, in perf_event_alloc()
11284 if (!event->addr_filter_ranges) { in perf_event_alloc()
11285 err = -ENOMEM; in perf_event_alloc()
11293 if (event->parent) { in perf_event_alloc()
11296 raw_spin_lock_irq(&ifh->lock); in perf_event_alloc()
11297 memcpy(event->addr_filter_ranges, in perf_event_alloc()
11298 event->parent->addr_filter_ranges, in perf_event_alloc()
11299 pmu->nr_addr_filters * sizeof(struct perf_addr_filter_range)); in perf_event_alloc()
11300 raw_spin_unlock_irq(&ifh->lock); in perf_event_alloc()
11304 event->addr_filters_gen = 1; in perf_event_alloc()
11307 if (!event->parent) { in perf_event_alloc()
11308 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { in perf_event_alloc()
11309 err = get_callchain_buffers(attr->sample_max_stack); in perf_event_alloc()
11325 if (!event->parent) { in perf_event_alloc()
11326 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) in perf_event_alloc()
11330 kfree(event->addr_filter_ranges); in perf_event_alloc()
11338 if (event->destroy) in perf_event_alloc()
11339 event->destroy(event); in perf_event_alloc()
11340 module_put(pmu->module); in perf_event_alloc()
11342 if (event->ns) in perf_event_alloc()
11343 put_pid_ns(event->ns); in perf_event_alloc()
11344 if (event->hw.target) in perf_event_alloc()
11345 put_task_struct(event->hw.target); in perf_event_alloc()
11360 ret = get_user(size, &uattr->size); in perf_copy_attr()
11372 if (ret == -E2BIG) in perf_copy_attr()
11377 attr->size = size; in perf_copy_attr()
11379 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3) in perf_copy_attr()
11380 return -EINVAL; in perf_copy_attr()
11382 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) in perf_copy_attr()
11383 return -EINVAL; in perf_copy_attr()
11385 if (attr->read_format & ~(PERF_FORMAT_MAX-1)) in perf_copy_attr()
11386 return -EINVAL; in perf_copy_attr()
11388 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { in perf_copy_attr()
11389 u64 mask = attr->branch_sample_type; in perf_copy_attr()
11392 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) in perf_copy_attr()
11393 return -EINVAL; in perf_copy_attr()
11397 return -EINVAL; in perf_copy_attr()
11403 if (!attr->exclude_kernel) in perf_copy_attr()
11406 if (!attr->exclude_user) in perf_copy_attr()
11409 if (!attr->exclude_hv) in perf_copy_attr()
11414 attr->branch_sample_type = mask; in perf_copy_attr()
11424 if (attr->sample_type & PERF_SAMPLE_REGS_USER) { in perf_copy_attr()
11425 ret = perf_reg_validate(attr->sample_regs_user); in perf_copy_attr()
11430 if (attr->sample_type & PERF_SAMPLE_STACK_USER) { in perf_copy_attr()
11432 return -ENOSYS; in perf_copy_attr()
11439 if (attr->sample_stack_user >= USHRT_MAX) in perf_copy_attr()
11440 return -EINVAL; in perf_copy_attr()
11441 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) in perf_copy_attr()
11442 return -EINVAL; in perf_copy_attr()
11445 if (!attr->sample_max_stack) in perf_copy_attr()
11446 attr->sample_max_stack = sysctl_perf_event_max_stack; in perf_copy_attr()
11448 if (attr->sample_type & PERF_SAMPLE_REGS_INTR) in perf_copy_attr()
11449 ret = perf_reg_validate(attr->sample_regs_intr); in perf_copy_attr()
11452 if (attr->sample_type & PERF_SAMPLE_CGROUP) in perf_copy_attr()
11453 return -EINVAL; in perf_copy_attr()
11460 put_user(sizeof(*attr), &uattr->size); in perf_copy_attr()
11461 ret = -E2BIG; in perf_copy_attr()
11469 int ret = -EINVAL; in perf_event_set_output()
11479 * Don't allow cross-cpu buffers in perf_event_set_output()
11481 if (output_event->cpu != event->cpu) in perf_event_set_output()
11485 * If its not a per-cpu rb, it must be the same task. in perf_event_set_output()
11487 if (output_event->cpu == -1 && output_event->ctx != event->ctx) in perf_event_set_output()
11493 if (output_event->clock != event->clock) in perf_event_set_output()
11507 event->pmu != output_event->pmu) in perf_event_set_output()
11511 mutex_lock(&event->mmap_mutex); in perf_event_set_output()
11513 if (atomic_read(&event->mmap_count)) in perf_event_set_output()
11527 mutex_unlock(&event->mmap_mutex); in perf_event_set_output()
11548 event->clock = &ktime_get_mono_fast_ns; in perf_event_set_clock()
11553 event->clock = &ktime_get_raw_fast_ns; in perf_event_set_clock()
11558 event->clock = &ktime_get_real_ns; in perf_event_set_clock()
11562 event->clock = &ktime_get_boottime_ns; in perf_event_set_clock()
11566 event->clock = &ktime_get_clocktai_ns; in perf_event_set_clock()
11570 return -EINVAL; in perf_event_set_clock()
11573 if (!nmi_safe && !(event->pmu->capabilities & PERF_PMU_CAP_NO_NMI)) in perf_event_set_clock()
11574 return -EINVAL; in perf_event_set_clock()
11591 gctx = READ_ONCE(group_leader->ctx); in __perf_event_ctx_lock_double()
11592 if (!refcount_inc_not_zero(&gctx->refcount)) { in __perf_event_ctx_lock_double()
11598 mutex_lock_double(&gctx->mutex, &ctx->mutex); in __perf_event_ctx_lock_double()
11600 if (group_leader->ctx != gctx) { in __perf_event_ctx_lock_double()
11601 mutex_unlock(&ctx->mutex); in __perf_event_ctx_lock_double()
11602 mutex_unlock(&gctx->mutex); in __perf_event_ctx_lock_double()
11611 * sys_perf_event_open - open a performance event, associate it to a task/cpu
11634 int cgroup_fd = -1; in SYSCALL_DEFINE5()
11638 return -EINVAL; in SYSCALL_DEFINE5()
11657 return -EACCES; in SYSCALL_DEFINE5()
11662 return -EINVAL; in SYSCALL_DEFINE5()
11665 return -EINVAL; in SYSCALL_DEFINE5()
11688 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) in SYSCALL_DEFINE5()
11689 return -EINVAL; in SYSCALL_DEFINE5()
11698 if (group_fd != -1) { in SYSCALL_DEFINE5()
11702 group_leader = group.file->private_data; in SYSCALL_DEFINE5()
11709 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { in SYSCALL_DEFINE5()
11718 group_leader->attr.inherit != attr.inherit) { in SYSCALL_DEFINE5()
11719 err = -EINVAL; in SYSCALL_DEFINE5()
11724 err = mutex_lock_interruptible(&task->signal->exec_update_mutex); in SYSCALL_DEFINE5()
11736 err = -EACCES; in SYSCALL_DEFINE5()
11752 if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) { in SYSCALL_DEFINE5()
11753 err = -EOPNOTSUPP; in SYSCALL_DEFINE5()
11762 pmu = event->pmu; in SYSCALL_DEFINE5()
11770 if (pmu->task_ctx_nr == perf_sw_context) in SYSCALL_DEFINE5()
11771 event->event_caps |= PERF_EV_CAP_SOFTWARE; in SYSCALL_DEFINE5()
11784 pmu = group_leader->ctx->pmu; in SYSCALL_DEFINE5()
11787 (group_leader->group_caps & PERF_EV_CAP_SOFTWARE)) { in SYSCALL_DEFINE5()
11810 err = -EINVAL; in SYSCALL_DEFINE5()
11814 * becoming part of another group-sibling): in SYSCALL_DEFINE5()
11816 if (group_leader->group_leader != group_leader) in SYSCALL_DEFINE5()
11820 if (group_leader->clock != event->clock) in SYSCALL_DEFINE5()
11828 if (group_leader->cpu != event->cpu) in SYSCALL_DEFINE5()
11833 * per-CPU events. in SYSCALL_DEFINE5()
11835 if (group_leader->ctx->task != ctx->task) in SYSCALL_DEFINE5()
11843 if (!move_group && group_leader->ctx != ctx) in SYSCALL_DEFINE5()
11870 if (gctx->task == TASK_TOMBSTONE) { in SYSCALL_DEFINE5()
11871 err = -ESRCH; in SYSCALL_DEFINE5()
11879 if (!(group_leader->group_caps & PERF_EV_CAP_SOFTWARE)) { in SYSCALL_DEFINE5()
11886 err = -EINVAL; in SYSCALL_DEFINE5()
11895 * Failure to create exclusive events returns -EBUSY. in SYSCALL_DEFINE5()
11897 err = -EBUSY; in SYSCALL_DEFINE5()
11906 mutex_lock(&ctx->mutex); in SYSCALL_DEFINE5()
11909 if (ctx->task == TASK_TOMBSTONE) { in SYSCALL_DEFINE5()
11910 err = -ESRCH; in SYSCALL_DEFINE5()
11915 err = -E2BIG; in SYSCALL_DEFINE5()
11929 if (!cpuctx->online) { in SYSCALL_DEFINE5()
11930 err = -ENODEV; in SYSCALL_DEFINE5()
11936 err = -EINVAL; in SYSCALL_DEFINE5()
11945 err = -EBUSY; in SYSCALL_DEFINE5()
11949 WARN_ON_ONCE(ctx->parent_ctx); in SYSCALL_DEFINE5()
11979 * (through the sibling list, which is still in-tact), we can in SYSCALL_DEFINE5()
11982 * By installing siblings first we NO-OP because they're not in SYSCALL_DEFINE5()
11987 perf_install_in_context(ctx, sibling, sibling->cpu); in SYSCALL_DEFINE5()
11997 perf_install_in_context(ctx, group_leader, group_leader->cpu); in SYSCALL_DEFINE5()
12010 event->owner = current; in SYSCALL_DEFINE5()
12012 perf_install_in_context(ctx, event, event->cpu); in SYSCALL_DEFINE5()
12017 mutex_unlock(&ctx->mutex); in SYSCALL_DEFINE5()
12020 mutex_unlock(&task->signal->exec_update_mutex); in SYSCALL_DEFINE5()
12024 mutex_lock(¤t->perf_event_mutex); in SYSCALL_DEFINE5()
12025 list_add_tail(&event->owner_entry, ¤t->perf_event_list); in SYSCALL_DEFINE5()
12026 mutex_unlock(¤t->perf_event_mutex); in SYSCALL_DEFINE5()
12041 mutex_unlock(&ctx->mutex); in SYSCALL_DEFINE5()
12049 * If event_file is set, the fput() above will have called ->release() in SYSCALL_DEFINE5()
12056 mutex_unlock(&task->signal->exec_update_mutex); in SYSCALL_DEFINE5()
12088 if (attr->aux_output) in perf_event_create_kernel_counter()
12089 return ERR_PTR(-EINVAL); in perf_event_create_kernel_counter()
12092 overflow_handler, context, -1); in perf_event_create_kernel_counter()
12099 event->owner = TASK_TOMBSTONE; in perf_event_create_kernel_counter()
12104 ctx = find_get_context(event->pmu, task, event); in perf_event_create_kernel_counter()
12110 WARN_ON_ONCE(ctx->parent_ctx); in perf_event_create_kernel_counter()
12111 mutex_lock(&ctx->mutex); in perf_event_create_kernel_counter()
12112 if (ctx->task == TASK_TOMBSTONE) { in perf_event_create_kernel_counter()
12113 err = -ESRCH; in perf_event_create_kernel_counter()
12126 if (!cpuctx->online) { in perf_event_create_kernel_counter()
12127 err = -ENODEV; in perf_event_create_kernel_counter()
12133 err = -EBUSY; in perf_event_create_kernel_counter()
12137 perf_install_in_context(ctx, event, event->cpu); in perf_event_create_kernel_counter()
12139 mutex_unlock(&ctx->mutex); in perf_event_create_kernel_counter()
12144 mutex_unlock(&ctx->mutex); in perf_event_create_kernel_counter()
12161 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; in perf_pmu_migrate_context()
12162 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; in perf_pmu_migrate_context()
12168 mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex); in perf_pmu_migrate_context()
12169 list_for_each_entry_safe(event, tmp, &src_ctx->event_list, in perf_pmu_migrate_context()
12174 list_add(&event->migrate_entry, &events); in perf_pmu_migrate_context()
12178 * Wait for the events to quiesce before re-instating them. in perf_pmu_migrate_context()
12183 * Re-instate events in 2 passes. in perf_pmu_migrate_context()
12191 if (event->group_leader == event) in perf_pmu_migrate_context()
12194 list_del(&event->migrate_entry); in perf_pmu_migrate_context()
12195 if (event->state >= PERF_EVENT_STATE_OFF) in perf_pmu_migrate_context()
12196 event->state = PERF_EVENT_STATE_INACTIVE; in perf_pmu_migrate_context()
12207 list_del(&event->migrate_entry); in perf_pmu_migrate_context()
12208 if (event->state >= PERF_EVENT_STATE_OFF) in perf_pmu_migrate_context()
12209 event->state = PERF_EVENT_STATE_INACTIVE; in perf_pmu_migrate_context()
12214 mutex_unlock(&dst_ctx->mutex); in perf_pmu_migrate_context()
12215 mutex_unlock(&src_ctx->mutex); in perf_pmu_migrate_context()
12222 struct perf_event *parent_event = child_event->parent; in sync_child_event()
12225 if (child_event->attr.inherit_stat) in sync_child_event()
12233 atomic64_add(child_val, &parent_event->child_count); in sync_child_event()
12234 atomic64_add(child_event->total_time_enabled, in sync_child_event()
12235 &parent_event->child_total_time_enabled); in sync_child_event()
12236 atomic64_add(child_event->total_time_running, in sync_child_event()
12237 &parent_event->child_total_time_running); in sync_child_event()
12245 struct perf_event *parent_event = child_event->parent; in perf_event_exit_event()
12259 raw_spin_lock_irq(&child_ctx->lock); in perf_event_exit_event()
12260 WARN_ON_ONCE(child_ctx->is_active); in perf_event_exit_event()
12266 raw_spin_unlock_irq(&child_ctx->lock); in perf_event_exit_event()
12284 WARN_ON_ONCE(parent_event->ctx->parent_ctx); in perf_event_exit_event()
12285 mutex_lock(&parent_event->child_mutex); in perf_event_exit_event()
12286 list_del_init(&child_event->child_list); in perf_event_exit_event()
12287 mutex_unlock(&parent_event->child_mutex); in perf_event_exit_event()
12309 * In order to reduce the amount of tricky in ctx tear-down, we hold in perf_event_exit_task_context()
12318 mutex_lock(&child_ctx->mutex); in perf_event_exit_task_context()
12321 * In a single ctx::lock section, de-schedule the events and detach the in perf_event_exit_task_context()
12325 raw_spin_lock_irq(&child_ctx->lock); in perf_event_exit_task_context()
12329 * Now that the context is inactive, destroy the task <-> ctx relation in perf_event_exit_task_context()
12332 RCU_INIT_POINTER(child->perf_event_ctxp[ctxn], NULL); in perf_event_exit_task_context()
12334 WRITE_ONCE(child_ctx->task, TASK_TOMBSTONE); in perf_event_exit_task_context()
12338 raw_spin_unlock_irq(&child_ctx->lock); in perf_event_exit_task_context()
12350 list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry) in perf_event_exit_task_context()
12353 mutex_unlock(&child_ctx->mutex); in perf_event_exit_task_context()
12369 mutex_lock(&child->perf_event_mutex); in perf_event_exit_task()
12370 list_for_each_entry_safe(event, tmp, &child->perf_event_list, in perf_event_exit_task()
12372 list_del_init(&event->owner_entry); in perf_event_exit_task()
12377 * we need to serialize on the owner->perf_event_mutex. in perf_event_exit_task()
12379 smp_store_release(&event->owner, NULL); in perf_event_exit_task()
12381 mutex_unlock(&child->perf_event_mutex); in perf_event_exit_task()
12389 * child contexts and sets child->perf_event_ctxp[] to NULL. in perf_event_exit_task()
12398 struct perf_event *parent = event->parent; in perf_free_event()
12403 mutex_lock(&parent->child_mutex); in perf_free_event()
12404 list_del_init(&event->child_list); in perf_free_event()
12405 mutex_unlock(&parent->child_mutex); in perf_free_event()
12409 raw_spin_lock_irq(&ctx->lock); in perf_free_event()
12412 raw_spin_unlock_irq(&ctx->lock); in perf_free_event()
12430 ctx = task->perf_event_ctxp[ctxn]; in perf_event_free_task()
12434 mutex_lock(&ctx->mutex); in perf_event_free_task()
12435 raw_spin_lock_irq(&ctx->lock); in perf_event_free_task()
12437 * Destroy the task <-> ctx relation and mark the context dead. in perf_event_free_task()
12442 RCU_INIT_POINTER(task->perf_event_ctxp[ctxn], NULL); in perf_event_free_task()
12443 WRITE_ONCE(ctx->task, TASK_TOMBSTONE); in perf_event_free_task()
12445 raw_spin_unlock_irq(&ctx->lock); in perf_event_free_task()
12447 list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) in perf_event_free_task()
12450 mutex_unlock(&ctx->mutex); in perf_event_free_task()
12461 * _free_event()'s put_task_struct(event->hw.target) will be a in perf_event_free_task()
12462 * use-after-free. in perf_event_free_task()
12466 wait_var_event(&ctx->refcount, refcount_read(&ctx->refcount) == 1); in perf_event_free_task()
12476 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); in perf_event_delayed_put()
12483 return ERR_PTR(-EBADF); in perf_event_get()
12485 if (file->f_op != &perf_fops) { in perf_event_get()
12487 return ERR_PTR(-EBADF); in perf_event_get()
12495 if (file->f_op != &perf_fops) in perf_get_event()
12496 return ERR_PTR(-EINVAL); in perf_get_event()
12498 return file->private_data; in perf_get_event()
12504 return ERR_PTR(-EINVAL); in perf_event_attrs()
12506 return &event->attr; in perf_event_attrs()
12513 * - valid pointer on success
12514 * - NULL for orphaned events
12515 * - IS_ERR() on error
12525 enum perf_event_state parent_state = parent_event->state; in inherit_event()
12535 if (parent_event->parent) in inherit_event()
12536 parent_event = parent_event->parent; in inherit_event()
12538 child_event = perf_event_alloc(&parent_event->attr, in inherit_event()
12539 parent_event->cpu, in inherit_event()
12542 NULL, NULL, -1); in inherit_event()
12547 if ((child_event->attach_state & PERF_ATTACH_TASK_DATA) && in inherit_event()
12548 !child_ctx->task_ctx_data) { in inherit_event()
12549 struct pmu *pmu = child_event->pmu; in inherit_event()
12551 child_ctx->task_ctx_data = alloc_task_ctx_data(pmu); in inherit_event()
12552 if (!child_ctx->task_ctx_data) { in inherit_event()
12554 return ERR_PTR(-ENOMEM); in inherit_event()
12559 * is_orphaned_event() and list_add_tail(&parent_event->child_list) in inherit_event()
12564 mutex_lock(&parent_event->child_mutex); in inherit_event()
12566 !atomic_long_inc_not_zero(&parent_event->refcount)) { in inherit_event()
12567 mutex_unlock(&parent_event->child_mutex); in inherit_event()
12581 child_event->state = PERF_EVENT_STATE_INACTIVE; in inherit_event()
12583 child_event->state = PERF_EVENT_STATE_OFF; in inherit_event()
12585 if (parent_event->attr.freq) { in inherit_event()
12586 u64 sample_period = parent_event->hw.sample_period; in inherit_event()
12587 struct hw_perf_event *hwc = &child_event->hw; in inherit_event()
12589 hwc->sample_period = sample_period; in inherit_event()
12590 hwc->last_period = sample_period; in inherit_event()
12592 local64_set(&hwc->period_left, sample_period); in inherit_event()
12595 child_event->ctx = child_ctx; in inherit_event()
12596 child_event->overflow_handler = parent_event->overflow_handler; in inherit_event()
12597 child_event->overflow_handler_context in inherit_event()
12598 = parent_event->overflow_handler_context; in inherit_event()
12609 raw_spin_lock_irqsave(&child_ctx->lock, flags); in inherit_event()
12611 raw_spin_unlock_irqrestore(&child_ctx->lock, flags); in inherit_event()
12616 list_add_tail(&child_event->child_list, &parent_event->child_list); in inherit_event()
12617 mutex_unlock(&parent_event->child_mutex); in inherit_event()
12629 * - 0 on success
12630 * - <0 on error
12657 if (sub->aux_event == parent_event && child_ctr && in inherit_group()
12659 return -EINVAL; in inherit_group()
12665 * Creates the child task context and tries to inherit the event-group.
12672 * - 0 on success
12673 * - <0 on error
12684 if (!event->attr.inherit) { in inherit_task_group()
12689 child_ctx = child->perf_event_ctxp[ctxn]; in inherit_task_group()
12697 child_ctx = alloc_perf_context(parent_ctx->pmu, child); in inherit_task_group()
12699 return -ENOMEM; in inherit_task_group()
12701 child->perf_event_ctxp[ctxn] = child_ctx; in inherit_task_group()
12726 if (likely(!parent->perf_event_ctxp[ctxn])) in perf_event_init_context()
12739 * it non-NULL earlier, the only reason for it to become NULL in perf_event_init_context()
12745 * Lock the parent list. No need to lock the child - not PID in perf_event_init_context()
12748 mutex_lock(&parent_ctx->mutex); in perf_event_init_context()
12751 * We dont have to disable NMIs - we are only looking at in perf_event_init_context()
12754 perf_event_groups_for_each(event, &parent_ctx->pinned_groups) { in perf_event_init_context()
12762 * We can't hold ctx->lock when iterating the ->flexible_group list due in perf_event_init_context()
12766 raw_spin_lock_irqsave(&parent_ctx->lock, flags); in perf_event_init_context()
12767 parent_ctx->rotate_disable = 1; in perf_event_init_context()
12768 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); in perf_event_init_context()
12770 perf_event_groups_for_each(event, &parent_ctx->flexible_groups) { in perf_event_init_context()
12777 raw_spin_lock_irqsave(&parent_ctx->lock, flags); in perf_event_init_context()
12778 parent_ctx->rotate_disable = 0; in perf_event_init_context()
12780 child_ctx = child->perf_event_ctxp[ctxn]; in perf_event_init_context()
12788 * parent_ctx->lock avoids it from being uncloned. in perf_event_init_context()
12790 cloned_ctx = parent_ctx->parent_ctx; in perf_event_init_context()
12792 child_ctx->parent_ctx = cloned_ctx; in perf_event_init_context()
12793 child_ctx->parent_gen = parent_ctx->parent_gen; in perf_event_init_context()
12795 child_ctx->parent_ctx = parent_ctx; in perf_event_init_context()
12796 child_ctx->parent_gen = parent_ctx->generation; in perf_event_init_context()
12798 get_ctx(child_ctx->parent_ctx); in perf_event_init_context()
12801 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); in perf_event_init_context()
12803 mutex_unlock(&parent_ctx->mutex); in perf_event_init_context()
12818 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); in perf_event_init_task()
12819 mutex_init(&child->perf_event_mutex); in perf_event_init_task()
12820 INIT_LIST_HEAD(&child->perf_event_list); in perf_event_init_task()
12842 mutex_init(&swhash->hlist_mutex); in perf_event_init_all_cpus()
12858 mutex_lock(&swhash->hlist_mutex); in perf_swevent_init_cpu()
12859 if (swhash->hlist_refcount > 0 && !swevent_hlist_deref(swhash)) { in perf_swevent_init_cpu()
12864 rcu_assign_pointer(swhash->swevent_hlist, hlist); in perf_swevent_init_cpu()
12866 mutex_unlock(&swhash->hlist_mutex); in perf_swevent_init_cpu()
12876 raw_spin_lock(&ctx->lock); in __perf_event_exit_context()
12878 list_for_each_entry(event, &ctx->event_list, event_entry) in __perf_event_exit_context()
12880 raw_spin_unlock(&ctx->lock); in __perf_event_exit_context()
12891 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); in perf_event_exit_cpu_context()
12892 ctx = &cpuctx->ctx; in perf_event_exit_cpu_context()
12894 mutex_lock(&ctx->mutex); in perf_event_exit_cpu_context()
12896 cpuctx->online = 0; in perf_event_exit_cpu_context()
12897 mutex_unlock(&ctx->mutex); in perf_event_exit_cpu_context()
12919 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); in perf_event_init_cpu()
12920 ctx = &cpuctx->ctx; in perf_event_init_cpu()
12922 mutex_lock(&ctx->mutex); in perf_event_init_cpu()
12923 cpuctx->online = 1; in perf_event_init_cpu()
12924 mutex_unlock(&ctx->mutex); in perf_event_init_cpu()
12966 perf_pmu_register(&perf_cpu_clock, NULL, -1); in perf_event_init()
12967 perf_pmu_register(&perf_task_clock, NULL, -1); in perf_event_init()
12989 if (pmu_attr->event_str) in perf_event_sysfs_show()
12990 return sprintf(page, "%s\n", pmu_attr->event_str); in perf_event_sysfs_show()
13008 if (!pmu->name || pmu->type < 0) in perf_event_sysfs_init()
13012 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); in perf_event_sysfs_init()
13032 return ERR_PTR(-ENOMEM); in perf_cgroup_css_alloc()
13034 jc->info = alloc_percpu(struct perf_cgroup_info); in perf_cgroup_css_alloc()
13035 if (!jc->info) { in perf_cgroup_css_alloc()
13037 return ERR_PTR(-ENOMEM); in perf_cgroup_css_alloc()
13040 return &jc->css; in perf_cgroup_css_alloc()
13047 free_percpu(jc->info); in perf_cgroup_css_free()
13053 perf_event_cgroup(css->cgroup); in perf_cgroup_css_online()