2 * Performance counter core code
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/ptrace.h>
20 #include <linux/percpu.h>
21 #include <linux/vmstat.h>
22 #include <linux/hardirq.h>
23 #include <linux/rculist.h>
24 #include <linux/uaccess.h>
25 #include <linux/syscalls.h>
26 #include <linux/anon_inodes.h>
27 #include <linux/kernel_stat.h>
28 #include <linux/perf_counter.h>
29 #include <linux/dcache.h>
31 #include <asm/irq_regs.h>
34 * Each CPU has a list of per CPU counters:
36 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
38 int perf_max_counters __read_mostly = 1;
39 static int perf_reserved_percpu __read_mostly;
40 static int perf_overcommit __read_mostly = 1;
42 static atomic_t nr_counters __read_mostly;
43 static atomic_t nr_mmap_tracking __read_mostly;
44 static atomic_t nr_munmap_tracking __read_mostly;
45 static atomic_t nr_comm_tracking __read_mostly;
47 int sysctl_perf_counter_priv __read_mostly; /* do we need to be privileged */
48 int sysctl_perf_counter_mlock __read_mostly = 512; /* 'free' kb per user */
51 * Lock for (sysadmin-configurable) counter reservations:
53 static DEFINE_SPINLOCK(perf_resource_lock);
56 * Architecture provided APIs - weak aliases:
58 extern __weak const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
63 void __weak hw_perf_disable(void) { barrier(); }
64 void __weak hw_perf_enable(void) { barrier(); }
66 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
67 int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
68 struct perf_cpu_context *cpuctx,
69 struct perf_counter_context *ctx, int cpu)
74 void __weak perf_counter_print_debug(void) { }
76 static DEFINE_PER_CPU(int, disable_count);
78 void __perf_disable(void)
80 __get_cpu_var(disable_count)++;
83 bool __perf_enable(void)
85 return !--__get_cpu_var(disable_count);
88 void perf_disable(void)
94 void perf_enable(void)
100 static void get_ctx(struct perf_counter_context *ctx)
102 atomic_inc(&ctx->refcount);
105 static void put_ctx(struct perf_counter_context *ctx)
107 if (atomic_dec_and_test(&ctx->refcount)) {
109 put_ctx(ctx->parent_ctx);
115 * Add a counter from the lists for its context.
116 * Must be called with ctx->mutex and ctx->lock held.
119 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
121 struct perf_counter *group_leader = counter->group_leader;
124 * Depending on whether it is a standalone or sibling counter,
125 * add it straight to the context's counter list, or to the group
126 * leader's sibling list:
128 if (group_leader == counter)
129 list_add_tail(&counter->list_entry, &ctx->counter_list);
131 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
132 group_leader->nr_siblings++;
135 list_add_rcu(&counter->event_entry, &ctx->event_list);
137 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
142 * Remove a counter from the lists for its context.
143 * Must be called with ctx->mutex and ctx->lock held.
146 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
148 struct perf_counter *sibling, *tmp;
150 if (list_empty(&counter->list_entry))
153 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
156 list_del_init(&counter->list_entry);
157 list_del_rcu(&counter->event_entry);
159 if (counter->group_leader != counter)
160 counter->group_leader->nr_siblings--;
163 * If this was a group counter with sibling counters then
164 * upgrade the siblings to singleton counters by adding them
165 * to the context list directly:
167 list_for_each_entry_safe(sibling, tmp,
168 &counter->sibling_list, list_entry) {
170 list_move_tail(&sibling->list_entry, &ctx->counter_list);
171 sibling->group_leader = sibling;
176 counter_sched_out(struct perf_counter *counter,
177 struct perf_cpu_context *cpuctx,
178 struct perf_counter_context *ctx)
180 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
183 counter->state = PERF_COUNTER_STATE_INACTIVE;
184 counter->tstamp_stopped = ctx->time;
185 counter->pmu->disable(counter);
188 if (!is_software_counter(counter))
189 cpuctx->active_oncpu--;
191 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
192 cpuctx->exclusive = 0;
196 group_sched_out(struct perf_counter *group_counter,
197 struct perf_cpu_context *cpuctx,
198 struct perf_counter_context *ctx)
200 struct perf_counter *counter;
202 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
205 counter_sched_out(group_counter, cpuctx, ctx);
208 * Schedule out siblings (if any):
210 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
211 counter_sched_out(counter, cpuctx, ctx);
213 if (group_counter->hw_event.exclusive)
214 cpuctx->exclusive = 0;
218 * Mark this context as not being a clone of another.
219 * Called when counters are added to or removed from this context.
220 * We also increment our generation number so that anything that
221 * was cloned from this context before this will not match anything
222 * cloned from this context after this.
224 static void unclone_ctx(struct perf_counter_context *ctx)
227 if (!ctx->parent_ctx)
229 put_ctx(ctx->parent_ctx);
230 ctx->parent_ctx = NULL;
234 * Cross CPU call to remove a performance counter
236 * We disable the counter on the hardware level first. After that we
237 * remove it from the context list.
239 static void __perf_counter_remove_from_context(void *info)
241 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
242 struct perf_counter *counter = info;
243 struct perf_counter_context *ctx = counter->ctx;
247 * If this is a task context, we need to check whether it is
248 * the current task context of this cpu. If not it has been
249 * scheduled out before the smp call arrived.
251 if (ctx->task && cpuctx->task_ctx != ctx)
254 spin_lock_irqsave(&ctx->lock, flags);
256 * Protect the list operation against NMI by disabling the
257 * counters on a global level.
261 counter_sched_out(counter, cpuctx, ctx);
263 list_del_counter(counter, ctx);
267 * Allow more per task counters with respect to the
270 cpuctx->max_pertask =
271 min(perf_max_counters - ctx->nr_counters,
272 perf_max_counters - perf_reserved_percpu);
276 spin_unlock_irqrestore(&ctx->lock, flags);
281 * Remove the counter from a task's (or a CPU's) list of counters.
283 * Must be called with ctx->mutex held.
285 * CPU counters are removed with a smp call. For task counters we only
286 * call when the task is on a CPU.
288 static void perf_counter_remove_from_context(struct perf_counter *counter)
290 struct perf_counter_context *ctx = counter->ctx;
291 struct task_struct *task = ctx->task;
296 * Per cpu counters are removed via an smp call and
297 * the removal is always sucessful.
299 smp_call_function_single(counter->cpu,
300 __perf_counter_remove_from_context,
306 task_oncpu_function_call(task, __perf_counter_remove_from_context,
309 spin_lock_irq(&ctx->lock);
311 * If the context is active we need to retry the smp call.
313 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
314 spin_unlock_irq(&ctx->lock);
319 * The lock prevents that this context is scheduled in so we
320 * can remove the counter safely, if the call above did not
323 if (!list_empty(&counter->list_entry)) {
324 list_del_counter(counter, ctx);
326 spin_unlock_irq(&ctx->lock);
329 static inline u64 perf_clock(void)
331 return cpu_clock(smp_processor_id());
335 * Update the record of the current time in a context.
337 static void update_context_time(struct perf_counter_context *ctx)
339 u64 now = perf_clock();
341 ctx->time += now - ctx->timestamp;
342 ctx->timestamp = now;
346 * Update the total_time_enabled and total_time_running fields for a counter.
348 static void update_counter_times(struct perf_counter *counter)
350 struct perf_counter_context *ctx = counter->ctx;
353 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
356 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
358 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
359 run_end = counter->tstamp_stopped;
363 counter->total_time_running = run_end - counter->tstamp_running;
367 * Update total_time_enabled and total_time_running for all counters in a group.
369 static void update_group_times(struct perf_counter *leader)
371 struct perf_counter *counter;
373 update_counter_times(leader);
374 list_for_each_entry(counter, &leader->sibling_list, list_entry)
375 update_counter_times(counter);
379 * Cross CPU call to disable a performance counter
381 static void __perf_counter_disable(void *info)
383 struct perf_counter *counter = info;
384 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
385 struct perf_counter_context *ctx = counter->ctx;
389 * If this is a per-task counter, need to check whether this
390 * counter's task is the current task on this cpu.
392 if (ctx->task && cpuctx->task_ctx != ctx)
395 spin_lock_irqsave(&ctx->lock, flags);
398 * If the counter is on, turn it off.
399 * If it is in error state, leave it in error state.
401 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
402 update_context_time(ctx);
403 update_counter_times(counter);
404 if (counter == counter->group_leader)
405 group_sched_out(counter, cpuctx, ctx);
407 counter_sched_out(counter, cpuctx, ctx);
408 counter->state = PERF_COUNTER_STATE_OFF;
412 spin_unlock_irqrestore(&ctx->lock, flags);
418 static void perf_counter_disable(struct perf_counter *counter)
420 struct perf_counter_context *ctx = counter->ctx;
421 struct task_struct *task = ctx->task;
425 * Disable the counter on the cpu that it's on
427 smp_call_function_single(counter->cpu, __perf_counter_disable,
433 task_oncpu_function_call(task, __perf_counter_disable, counter);
435 spin_lock_irq(&ctx->lock);
437 * If the counter is still active, we need to retry the cross-call.
439 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
440 spin_unlock_irq(&ctx->lock);
445 * Since we have the lock this context can't be scheduled
446 * in, so we can change the state safely.
448 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
449 update_counter_times(counter);
450 counter->state = PERF_COUNTER_STATE_OFF;
454 spin_unlock_irq(&ctx->lock);
458 counter_sched_in(struct perf_counter *counter,
459 struct perf_cpu_context *cpuctx,
460 struct perf_counter_context *ctx,
463 if (counter->state <= PERF_COUNTER_STATE_OFF)
466 counter->state = PERF_COUNTER_STATE_ACTIVE;
467 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
469 * The new state must be visible before we turn it on in the hardware:
473 if (counter->pmu->enable(counter)) {
474 counter->state = PERF_COUNTER_STATE_INACTIVE;
479 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
481 if (!is_software_counter(counter))
482 cpuctx->active_oncpu++;
485 if (counter->hw_event.exclusive)
486 cpuctx->exclusive = 1;
492 group_sched_in(struct perf_counter *group_counter,
493 struct perf_cpu_context *cpuctx,
494 struct perf_counter_context *ctx,
497 struct perf_counter *counter, *partial_group;
500 if (group_counter->state == PERF_COUNTER_STATE_OFF)
503 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
505 return ret < 0 ? ret : 0;
507 group_counter->prev_state = group_counter->state;
508 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
512 * Schedule in siblings as one group (if any):
514 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
515 counter->prev_state = counter->state;
516 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
517 partial_group = counter;
526 * Groups can be scheduled in as one unit only, so undo any
527 * partial group before returning:
529 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
530 if (counter == partial_group)
532 counter_sched_out(counter, cpuctx, ctx);
534 counter_sched_out(group_counter, cpuctx, ctx);
540 * Return 1 for a group consisting entirely of software counters,
541 * 0 if the group contains any hardware counters.
543 static int is_software_only_group(struct perf_counter *leader)
545 struct perf_counter *counter;
547 if (!is_software_counter(leader))
550 list_for_each_entry(counter, &leader->sibling_list, list_entry)
551 if (!is_software_counter(counter))
558 * Work out whether we can put this counter group on the CPU now.
560 static int group_can_go_on(struct perf_counter *counter,
561 struct perf_cpu_context *cpuctx,
565 * Groups consisting entirely of software counters can always go on.
567 if (is_software_only_group(counter))
570 * If an exclusive group is already on, no other hardware
571 * counters can go on.
573 if (cpuctx->exclusive)
576 * If this group is exclusive and there are already
577 * counters on the CPU, it can't go on.
579 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
582 * Otherwise, try to add it if all previous groups were able
588 static void add_counter_to_ctx(struct perf_counter *counter,
589 struct perf_counter_context *ctx)
591 list_add_counter(counter, ctx);
592 counter->prev_state = PERF_COUNTER_STATE_OFF;
593 counter->tstamp_enabled = ctx->time;
594 counter->tstamp_running = ctx->time;
595 counter->tstamp_stopped = ctx->time;
599 * Cross CPU call to install and enable a performance counter
601 * Must be called with ctx->mutex held
603 static void __perf_install_in_context(void *info)
605 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
606 struct perf_counter *counter = info;
607 struct perf_counter_context *ctx = counter->ctx;
608 struct perf_counter *leader = counter->group_leader;
609 int cpu = smp_processor_id();
614 * If this is a task context, we need to check whether it is
615 * the current task context of this cpu. If not it has been
616 * scheduled out before the smp call arrived.
617 * Or possibly this is the right context but it isn't
618 * on this cpu because it had no counters.
620 if (ctx->task && cpuctx->task_ctx != ctx) {
621 if (cpuctx->task_ctx || ctx->task != current)
623 cpuctx->task_ctx = ctx;
626 spin_lock_irqsave(&ctx->lock, flags);
628 update_context_time(ctx);
631 * Protect the list operation against NMI by disabling the
632 * counters on a global level. NOP for non NMI based counters.
636 add_counter_to_ctx(counter, ctx);
639 * Don't put the counter on if it is disabled or if
640 * it is in a group and the group isn't on.
642 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
643 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
647 * An exclusive counter can't go on if there are already active
648 * hardware counters, and no hardware counter can go on if there
649 * is already an exclusive counter on.
651 if (!group_can_go_on(counter, cpuctx, 1))
654 err = counter_sched_in(counter, cpuctx, ctx, cpu);
658 * This counter couldn't go on. If it is in a group
659 * then we have to pull the whole group off.
660 * If the counter group is pinned then put it in error state.
662 if (leader != counter)
663 group_sched_out(leader, cpuctx, ctx);
664 if (leader->hw_event.pinned) {
665 update_group_times(leader);
666 leader->state = PERF_COUNTER_STATE_ERROR;
670 if (!err && !ctx->task && cpuctx->max_pertask)
671 cpuctx->max_pertask--;
676 spin_unlock_irqrestore(&ctx->lock, flags);
680 * Attach a performance counter to a context
682 * First we add the counter to the list with the hardware enable bit
683 * in counter->hw_config cleared.
685 * If the counter is attached to a task which is on a CPU we use a smp
686 * call to enable it in the task context. The task might have been
687 * scheduled away, but we check this in the smp call again.
689 * Must be called with ctx->mutex held.
692 perf_install_in_context(struct perf_counter_context *ctx,
693 struct perf_counter *counter,
696 struct task_struct *task = ctx->task;
700 * Per cpu counters are installed via an smp call and
701 * the install is always sucessful.
703 smp_call_function_single(cpu, __perf_install_in_context,
709 task_oncpu_function_call(task, __perf_install_in_context,
712 spin_lock_irq(&ctx->lock);
714 * we need to retry the smp call.
716 if (ctx->is_active && list_empty(&counter->list_entry)) {
717 spin_unlock_irq(&ctx->lock);
722 * The lock prevents that this context is scheduled in so we
723 * can add the counter safely, if it the call above did not
726 if (list_empty(&counter->list_entry))
727 add_counter_to_ctx(counter, ctx);
728 spin_unlock_irq(&ctx->lock);
732 * Cross CPU call to enable a performance counter
734 static void __perf_counter_enable(void *info)
736 struct perf_counter *counter = info;
737 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
738 struct perf_counter_context *ctx = counter->ctx;
739 struct perf_counter *leader = counter->group_leader;
744 * If this is a per-task counter, need to check whether this
745 * counter's task is the current task on this cpu.
747 if (ctx->task && cpuctx->task_ctx != ctx) {
748 if (cpuctx->task_ctx || ctx->task != current)
750 cpuctx->task_ctx = ctx;
753 spin_lock_irqsave(&ctx->lock, flags);
755 update_context_time(ctx);
757 counter->prev_state = counter->state;
758 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
760 counter->state = PERF_COUNTER_STATE_INACTIVE;
761 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
765 * If the counter is in a group and isn't the group leader,
766 * then don't put it on unless the group is on.
768 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
771 if (!group_can_go_on(counter, cpuctx, 1)) {
775 if (counter == leader)
776 err = group_sched_in(counter, cpuctx, ctx,
779 err = counter_sched_in(counter, cpuctx, ctx,
786 * If this counter can't go on and it's part of a
787 * group, then the whole group has to come off.
789 if (leader != counter)
790 group_sched_out(leader, cpuctx, ctx);
791 if (leader->hw_event.pinned) {
792 update_group_times(leader);
793 leader->state = PERF_COUNTER_STATE_ERROR;
798 spin_unlock_irqrestore(&ctx->lock, flags);
804 static void perf_counter_enable(struct perf_counter *counter)
806 struct perf_counter_context *ctx = counter->ctx;
807 struct task_struct *task = ctx->task;
811 * Enable the counter on the cpu that it's on
813 smp_call_function_single(counter->cpu, __perf_counter_enable,
818 spin_lock_irq(&ctx->lock);
819 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
823 * If the counter is in error state, clear that first.
824 * That way, if we see the counter in error state below, we
825 * know that it has gone back into error state, as distinct
826 * from the task having been scheduled away before the
827 * cross-call arrived.
829 if (counter->state == PERF_COUNTER_STATE_ERROR)
830 counter->state = PERF_COUNTER_STATE_OFF;
833 spin_unlock_irq(&ctx->lock);
834 task_oncpu_function_call(task, __perf_counter_enable, counter);
836 spin_lock_irq(&ctx->lock);
839 * If the context is active and the counter is still off,
840 * we need to retry the cross-call.
842 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
846 * Since we have the lock this context can't be scheduled
847 * in, so we can change the state safely.
849 if (counter->state == PERF_COUNTER_STATE_OFF) {
850 counter->state = PERF_COUNTER_STATE_INACTIVE;
851 counter->tstamp_enabled =
852 ctx->time - counter->total_time_enabled;
856 spin_unlock_irq(&ctx->lock);
859 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
862 * not supported on inherited counters
864 if (counter->hw_event.inherit)
867 atomic_add(refresh, &counter->event_limit);
868 perf_counter_enable(counter);
873 void __perf_counter_sched_out(struct perf_counter_context *ctx,
874 struct perf_cpu_context *cpuctx)
876 struct perf_counter *counter;
878 spin_lock(&ctx->lock);
880 if (likely(!ctx->nr_counters))
882 update_context_time(ctx);
885 if (ctx->nr_active) {
886 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
887 if (counter != counter->group_leader)
888 counter_sched_out(counter, cpuctx, ctx);
890 group_sched_out(counter, cpuctx, ctx);
895 spin_unlock(&ctx->lock);
899 * Test whether two contexts are equivalent, i.e. whether they
900 * have both been cloned from the same version of the same context
901 * and they both have the same number of enabled counters.
902 * If the number of enabled counters is the same, then the set
903 * of enabled counters should be the same, because these are both
904 * inherited contexts, therefore we can't access individual counters
905 * in them directly with an fd; we can only enable/disable all
906 * counters via prctl, or enable/disable all counters in a family
907 * via ioctl, which will have the same effect on both contexts.
909 static int context_equiv(struct perf_counter_context *ctx1,
910 struct perf_counter_context *ctx2)
912 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
913 && ctx1->parent_gen == ctx2->parent_gen
914 && ctx1->nr_enabled == ctx2->nr_enabled;
918 * Called from scheduler to remove the counters of the current task,
919 * with interrupts disabled.
921 * We stop each counter and update the counter value in counter->count.
923 * This does not protect us against NMI, but disable()
924 * sets the disabled bit in the control field of counter _before_
925 * accessing the counter control register. If a NMI hits, then it will
926 * not restart the counter.
928 void perf_counter_task_sched_out(struct task_struct *task,
929 struct task_struct *next, int cpu)
931 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
932 struct perf_counter_context *ctx = task->perf_counter_ctxp;
933 struct perf_counter_context *next_ctx;
934 struct pt_regs *regs;
936 if (likely(!ctx || !cpuctx->task_ctx))
939 update_context_time(ctx);
941 regs = task_pt_regs(task);
942 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs, 0);
944 next_ctx = next->perf_counter_ctxp;
945 if (next_ctx && context_equiv(ctx, next_ctx)) {
946 task->perf_counter_ctxp = next_ctx;
947 next->perf_counter_ctxp = ctx;
949 next_ctx->task = task;
953 __perf_counter_sched_out(ctx, cpuctx);
955 cpuctx->task_ctx = NULL;
958 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
960 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
962 if (!cpuctx->task_ctx)
964 __perf_counter_sched_out(ctx, cpuctx);
965 cpuctx->task_ctx = NULL;
968 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
970 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
974 __perf_counter_sched_in(struct perf_counter_context *ctx,
975 struct perf_cpu_context *cpuctx, int cpu)
977 struct perf_counter *counter;
980 spin_lock(&ctx->lock);
982 if (likely(!ctx->nr_counters))
985 ctx->timestamp = perf_clock();
990 * First go through the list and put on any pinned groups
991 * in order to give them the best chance of going on.
993 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
994 if (counter->state <= PERF_COUNTER_STATE_OFF ||
995 !counter->hw_event.pinned)
997 if (counter->cpu != -1 && counter->cpu != cpu)
1000 if (counter != counter->group_leader)
1001 counter_sched_in(counter, cpuctx, ctx, cpu);
1003 if (group_can_go_on(counter, cpuctx, 1))
1004 group_sched_in(counter, cpuctx, ctx, cpu);
1008 * If this pinned group hasn't been scheduled,
1009 * put it in error state.
1011 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1012 update_group_times(counter);
1013 counter->state = PERF_COUNTER_STATE_ERROR;
1017 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1019 * Ignore counters in OFF or ERROR state, and
1020 * ignore pinned counters since we did them already.
1022 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1023 counter->hw_event.pinned)
1027 * Listen to the 'cpu' scheduling filter constraint
1030 if (counter->cpu != -1 && counter->cpu != cpu)
1033 if (counter != counter->group_leader) {
1034 if (counter_sched_in(counter, cpuctx, ctx, cpu))
1037 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
1038 if (group_sched_in(counter, cpuctx, ctx, cpu))
1045 spin_unlock(&ctx->lock);
1049 * Called from scheduler to add the counters of the current task
1050 * with interrupts disabled.
1052 * We restore the counter value and then enable it.
1054 * This does not protect us against NMI, but enable()
1055 * sets the enabled bit in the control field of counter _before_
1056 * accessing the counter control register. If a NMI hits, then it will
1057 * keep the counter running.
1059 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
1061 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1062 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1066 if (cpuctx->task_ctx == ctx)
1068 __perf_counter_sched_in(ctx, cpuctx, cpu);
1069 cpuctx->task_ctx = ctx;
1072 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1074 struct perf_counter_context *ctx = &cpuctx->ctx;
1076 __perf_counter_sched_in(ctx, cpuctx, cpu);
1079 int perf_counter_task_enable(void)
1081 struct perf_counter *counter;
1083 mutex_lock(¤t->perf_counter_mutex);
1084 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1085 perf_counter_enable(counter);
1086 mutex_unlock(¤t->perf_counter_mutex);
1091 int perf_counter_task_disable(void)
1093 struct perf_counter *counter;
1095 mutex_lock(¤t->perf_counter_mutex);
1096 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1097 perf_counter_disable(counter);
1098 mutex_unlock(¤t->perf_counter_mutex);
1103 static void perf_log_period(struct perf_counter *counter, u64 period);
1105 static void perf_adjust_freq(struct perf_counter_context *ctx)
1107 struct perf_counter *counter;
1112 spin_lock(&ctx->lock);
1113 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1114 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1117 if (!counter->hw_event.freq || !counter->hw_event.irq_freq)
1120 events = HZ * counter->hw.interrupts * counter->hw.irq_period;
1121 period = div64_u64(events, counter->hw_event.irq_freq);
1123 delta = (s64)(1 + period - counter->hw.irq_period);
1126 irq_period = counter->hw.irq_period + delta;
1131 perf_log_period(counter, irq_period);
1133 counter->hw.irq_period = irq_period;
1134 counter->hw.interrupts = 0;
1136 spin_unlock(&ctx->lock);
1140 * Round-robin a context's counters:
1142 static void rotate_ctx(struct perf_counter_context *ctx)
1144 struct perf_counter *counter;
1146 if (!ctx->nr_counters)
1149 spin_lock(&ctx->lock);
1151 * Rotate the first entry last (works just fine for group counters too):
1154 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1155 list_move_tail(&counter->list_entry, &ctx->counter_list);
1160 spin_unlock(&ctx->lock);
1163 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1165 struct perf_cpu_context *cpuctx;
1166 struct perf_counter_context *ctx;
1168 if (!atomic_read(&nr_counters))
1171 cpuctx = &per_cpu(perf_cpu_context, cpu);
1172 ctx = curr->perf_counter_ctxp;
1174 perf_adjust_freq(&cpuctx->ctx);
1176 perf_adjust_freq(ctx);
1178 perf_counter_cpu_sched_out(cpuctx);
1180 __perf_counter_task_sched_out(ctx);
1182 rotate_ctx(&cpuctx->ctx);
1186 perf_counter_cpu_sched_in(cpuctx, cpu);
1188 perf_counter_task_sched_in(curr, cpu);
1192 * Cross CPU call to read the hardware counter
1194 static void __read(void *info)
1196 struct perf_counter *counter = info;
1197 struct perf_counter_context *ctx = counter->ctx;
1198 unsigned long flags;
1200 local_irq_save(flags);
1202 update_context_time(ctx);
1203 counter->pmu->read(counter);
1204 update_counter_times(counter);
1205 local_irq_restore(flags);
1208 static u64 perf_counter_read(struct perf_counter *counter)
1211 * If counter is enabled and currently active on a CPU, update the
1212 * value in the counter structure:
1214 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1215 smp_call_function_single(counter->oncpu,
1216 __read, counter, 1);
1217 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1218 update_counter_times(counter);
1221 return atomic64_read(&counter->count);
1225 * Initialize the perf_counter context in a task_struct:
1228 __perf_counter_init_context(struct perf_counter_context *ctx,
1229 struct task_struct *task)
1231 memset(ctx, 0, sizeof(*ctx));
1232 spin_lock_init(&ctx->lock);
1233 mutex_init(&ctx->mutex);
1234 INIT_LIST_HEAD(&ctx->counter_list);
1235 INIT_LIST_HEAD(&ctx->event_list);
1236 atomic_set(&ctx->refcount, 1);
1240 static void put_context(struct perf_counter_context *ctx)
1243 put_task_struct(ctx->task);
1246 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1248 struct perf_cpu_context *cpuctx;
1249 struct perf_counter_context *ctx;
1250 struct perf_counter_context *tctx;
1251 struct task_struct *task;
1254 * If cpu is not a wildcard then this is a percpu counter:
1257 /* Must be root to operate on a CPU counter: */
1258 if (sysctl_perf_counter_priv && !capable(CAP_SYS_ADMIN))
1259 return ERR_PTR(-EACCES);
1261 if (cpu < 0 || cpu > num_possible_cpus())
1262 return ERR_PTR(-EINVAL);
1265 * We could be clever and allow to attach a counter to an
1266 * offline CPU and activate it when the CPU comes up, but
1269 if (!cpu_isset(cpu, cpu_online_map))
1270 return ERR_PTR(-ENODEV);
1272 cpuctx = &per_cpu(perf_cpu_context, cpu);
1282 task = find_task_by_vpid(pid);
1284 get_task_struct(task);
1288 return ERR_PTR(-ESRCH);
1290 /* Reuse ptrace permission checks for now. */
1291 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1292 put_task_struct(task);
1293 return ERR_PTR(-EACCES);
1296 ctx = task->perf_counter_ctxp;
1298 ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
1300 put_task_struct(task);
1301 return ERR_PTR(-ENOMEM);
1303 __perf_counter_init_context(ctx, task);
1305 * Make sure other cpus see correct values for *ctx
1306 * once task->perf_counter_ctxp is visible to them.
1309 tctx = cmpxchg(&task->perf_counter_ctxp, NULL, ctx);
1312 * We raced with some other task; use
1313 * the context they set.
1323 static void free_counter_rcu(struct rcu_head *head)
1325 struct perf_counter *counter;
1327 counter = container_of(head, struct perf_counter, rcu_head);
1328 put_ctx(counter->ctx);
1332 static void perf_pending_sync(struct perf_counter *counter);
1334 static void free_counter(struct perf_counter *counter)
1336 perf_pending_sync(counter);
1338 atomic_dec(&nr_counters);
1339 if (counter->hw_event.mmap)
1340 atomic_dec(&nr_mmap_tracking);
1341 if (counter->hw_event.munmap)
1342 atomic_dec(&nr_munmap_tracking);
1343 if (counter->hw_event.comm)
1344 atomic_dec(&nr_comm_tracking);
1346 if (counter->destroy)
1347 counter->destroy(counter);
1349 call_rcu(&counter->rcu_head, free_counter_rcu);
1353 * Called when the last reference to the file is gone.
1355 static int perf_release(struct inode *inode, struct file *file)
1357 struct perf_counter *counter = file->private_data;
1358 struct perf_counter_context *ctx = counter->ctx;
1360 file->private_data = NULL;
1362 mutex_lock(&ctx->mutex);
1363 perf_counter_remove_from_context(counter);
1364 mutex_unlock(&ctx->mutex);
1366 mutex_lock(&counter->owner->perf_counter_mutex);
1367 list_del_init(&counter->owner_entry);
1368 mutex_unlock(&counter->owner->perf_counter_mutex);
1369 put_task_struct(counter->owner);
1371 free_counter(counter);
1378 * Read the performance counter - simple non blocking version for now
1381 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1387 * Return end-of-file for a read on a counter that is in
1388 * error state (i.e. because it was pinned but it couldn't be
1389 * scheduled on to the CPU at some point).
1391 if (counter->state == PERF_COUNTER_STATE_ERROR)
1394 mutex_lock(&counter->child_mutex);
1395 values[0] = perf_counter_read(counter);
1397 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1398 values[n++] = counter->total_time_enabled +
1399 atomic64_read(&counter->child_total_time_enabled);
1400 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1401 values[n++] = counter->total_time_running +
1402 atomic64_read(&counter->child_total_time_running);
1403 mutex_unlock(&counter->child_mutex);
1405 if (count < n * sizeof(u64))
1407 count = n * sizeof(u64);
1409 if (copy_to_user(buf, values, count))
1416 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1418 struct perf_counter *counter = file->private_data;
1420 return perf_read_hw(counter, buf, count);
1423 static unsigned int perf_poll(struct file *file, poll_table *wait)
1425 struct perf_counter *counter = file->private_data;
1426 struct perf_mmap_data *data;
1427 unsigned int events = POLL_HUP;
1430 data = rcu_dereference(counter->data);
1432 events = atomic_xchg(&data->poll, 0);
1435 poll_wait(file, &counter->waitq, wait);
1440 static void perf_counter_reset(struct perf_counter *counter)
1442 (void)perf_counter_read(counter);
1443 atomic64_set(&counter->count, 0);
1444 perf_counter_update_userpage(counter);
1447 static void perf_counter_for_each_sibling(struct perf_counter *counter,
1448 void (*func)(struct perf_counter *))
1450 struct perf_counter_context *ctx = counter->ctx;
1451 struct perf_counter *sibling;
1453 mutex_lock(&ctx->mutex);
1454 counter = counter->group_leader;
1457 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1459 mutex_unlock(&ctx->mutex);
1462 static void perf_counter_for_each_child(struct perf_counter *counter,
1463 void (*func)(struct perf_counter *))
1465 struct perf_counter *child;
1467 mutex_lock(&counter->child_mutex);
1469 list_for_each_entry(child, &counter->child_list, child_list)
1471 mutex_unlock(&counter->child_mutex);
1474 static void perf_counter_for_each(struct perf_counter *counter,
1475 void (*func)(struct perf_counter *))
1477 struct perf_counter *child;
1479 mutex_lock(&counter->child_mutex);
1480 perf_counter_for_each_sibling(counter, func);
1481 list_for_each_entry(child, &counter->child_list, child_list)
1482 perf_counter_for_each_sibling(child, func);
1483 mutex_unlock(&counter->child_mutex);
1486 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1488 struct perf_counter *counter = file->private_data;
1489 void (*func)(struct perf_counter *);
1493 case PERF_COUNTER_IOC_ENABLE:
1494 func = perf_counter_enable;
1496 case PERF_COUNTER_IOC_DISABLE:
1497 func = perf_counter_disable;
1499 case PERF_COUNTER_IOC_RESET:
1500 func = perf_counter_reset;
1503 case PERF_COUNTER_IOC_REFRESH:
1504 return perf_counter_refresh(counter, arg);
1509 if (flags & PERF_IOC_FLAG_GROUP)
1510 perf_counter_for_each(counter, func);
1512 perf_counter_for_each_child(counter, func);
1518 * Callers need to ensure there can be no nesting of this function, otherwise
1519 * the seqlock logic goes bad. We can not serialize this because the arch
1520 * code calls this from NMI context.
1522 void perf_counter_update_userpage(struct perf_counter *counter)
1524 struct perf_mmap_data *data;
1525 struct perf_counter_mmap_page *userpg;
1528 data = rcu_dereference(counter->data);
1532 userpg = data->user_page;
1535 * Disable preemption so as to not let the corresponding user-space
1536 * spin too long if we get preempted.
1541 userpg->index = counter->hw.idx;
1542 userpg->offset = atomic64_read(&counter->count);
1543 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1544 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1553 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1555 struct perf_counter *counter = vma->vm_file->private_data;
1556 struct perf_mmap_data *data;
1557 int ret = VM_FAULT_SIGBUS;
1560 data = rcu_dereference(counter->data);
1564 if (vmf->pgoff == 0) {
1565 vmf->page = virt_to_page(data->user_page);
1567 int nr = vmf->pgoff - 1;
1569 if ((unsigned)nr > data->nr_pages)
1572 vmf->page = virt_to_page(data->data_pages[nr]);
1574 get_page(vmf->page);
1582 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1584 struct perf_mmap_data *data;
1588 WARN_ON(atomic_read(&counter->mmap_count));
1590 size = sizeof(struct perf_mmap_data);
1591 size += nr_pages * sizeof(void *);
1593 data = kzalloc(size, GFP_KERNEL);
1597 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1598 if (!data->user_page)
1599 goto fail_user_page;
1601 for (i = 0; i < nr_pages; i++) {
1602 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1603 if (!data->data_pages[i])
1604 goto fail_data_pages;
1607 data->nr_pages = nr_pages;
1608 atomic_set(&data->lock, -1);
1610 rcu_assign_pointer(counter->data, data);
1615 for (i--; i >= 0; i--)
1616 free_page((unsigned long)data->data_pages[i]);
1618 free_page((unsigned long)data->user_page);
1627 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1629 struct perf_mmap_data *data = container_of(rcu_head,
1630 struct perf_mmap_data, rcu_head);
1633 free_page((unsigned long)data->user_page);
1634 for (i = 0; i < data->nr_pages; i++)
1635 free_page((unsigned long)data->data_pages[i]);
1639 static void perf_mmap_data_free(struct perf_counter *counter)
1641 struct perf_mmap_data *data = counter->data;
1643 WARN_ON(atomic_read(&counter->mmap_count));
1645 rcu_assign_pointer(counter->data, NULL);
1646 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1649 static void perf_mmap_open(struct vm_area_struct *vma)
1651 struct perf_counter *counter = vma->vm_file->private_data;
1653 atomic_inc(&counter->mmap_count);
1656 static void perf_mmap_close(struct vm_area_struct *vma)
1658 struct perf_counter *counter = vma->vm_file->private_data;
1660 if (atomic_dec_and_mutex_lock(&counter->mmap_count,
1661 &counter->mmap_mutex)) {
1662 struct user_struct *user = current_user();
1664 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
1665 vma->vm_mm->locked_vm -= counter->data->nr_locked;
1666 perf_mmap_data_free(counter);
1667 mutex_unlock(&counter->mmap_mutex);
1671 static struct vm_operations_struct perf_mmap_vmops = {
1672 .open = perf_mmap_open,
1673 .close = perf_mmap_close,
1674 .fault = perf_mmap_fault,
1677 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
1679 struct perf_counter *counter = file->private_data;
1680 struct user_struct *user = current_user();
1681 unsigned long vma_size;
1682 unsigned long nr_pages;
1683 unsigned long user_locked, user_lock_limit;
1684 unsigned long locked, lock_limit;
1685 long user_extra, extra;
1688 if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
1691 vma_size = vma->vm_end - vma->vm_start;
1692 nr_pages = (vma_size / PAGE_SIZE) - 1;
1695 * If we have data pages ensure they're a power-of-two number, so we
1696 * can do bitmasks instead of modulo.
1698 if (nr_pages != 0 && !is_power_of_2(nr_pages))
1701 if (vma_size != PAGE_SIZE * (1 + nr_pages))
1704 if (vma->vm_pgoff != 0)
1707 mutex_lock(&counter->mmap_mutex);
1708 if (atomic_inc_not_zero(&counter->mmap_count)) {
1709 if (nr_pages != counter->data->nr_pages)
1714 user_extra = nr_pages + 1;
1715 user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
1716 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
1719 if (user_locked > user_lock_limit)
1720 extra = user_locked - user_lock_limit;
1722 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1723 lock_limit >>= PAGE_SHIFT;
1724 locked = vma->vm_mm->locked_vm + extra;
1726 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
1731 WARN_ON(counter->data);
1732 ret = perf_mmap_data_alloc(counter, nr_pages);
1736 atomic_set(&counter->mmap_count, 1);
1737 atomic_long_add(user_extra, &user->locked_vm);
1738 vma->vm_mm->locked_vm += extra;
1739 counter->data->nr_locked = extra;
1741 mutex_unlock(&counter->mmap_mutex);
1743 vma->vm_flags &= ~VM_MAYWRITE;
1744 vma->vm_flags |= VM_RESERVED;
1745 vma->vm_ops = &perf_mmap_vmops;
1750 static int perf_fasync(int fd, struct file *filp, int on)
1752 struct perf_counter *counter = filp->private_data;
1753 struct inode *inode = filp->f_path.dentry->d_inode;
1756 mutex_lock(&inode->i_mutex);
1757 retval = fasync_helper(fd, filp, on, &counter->fasync);
1758 mutex_unlock(&inode->i_mutex);
1766 static const struct file_operations perf_fops = {
1767 .release = perf_release,
1770 .unlocked_ioctl = perf_ioctl,
1771 .compat_ioctl = perf_ioctl,
1773 .fasync = perf_fasync,
1777 * Perf counter wakeup
1779 * If there's data, ensure we set the poll() state and publish everything
1780 * to user-space before waking everybody up.
1783 void perf_counter_wakeup(struct perf_counter *counter)
1785 wake_up_all(&counter->waitq);
1787 if (counter->pending_kill) {
1788 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
1789 counter->pending_kill = 0;
1796 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1798 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1799 * single linked list and use cmpxchg() to add entries lockless.
1802 static void perf_pending_counter(struct perf_pending_entry *entry)
1804 struct perf_counter *counter = container_of(entry,
1805 struct perf_counter, pending);
1807 if (counter->pending_disable) {
1808 counter->pending_disable = 0;
1809 perf_counter_disable(counter);
1812 if (counter->pending_wakeup) {
1813 counter->pending_wakeup = 0;
1814 perf_counter_wakeup(counter);
1818 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1820 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
1824 static void perf_pending_queue(struct perf_pending_entry *entry,
1825 void (*func)(struct perf_pending_entry *))
1827 struct perf_pending_entry **head;
1829 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
1834 head = &get_cpu_var(perf_pending_head);
1837 entry->next = *head;
1838 } while (cmpxchg(head, entry->next, entry) != entry->next);
1840 set_perf_counter_pending();
1842 put_cpu_var(perf_pending_head);
1845 static int __perf_pending_run(void)
1847 struct perf_pending_entry *list;
1850 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
1851 while (list != PENDING_TAIL) {
1852 void (*func)(struct perf_pending_entry *);
1853 struct perf_pending_entry *entry = list;
1860 * Ensure we observe the unqueue before we issue the wakeup,
1861 * so that we won't be waiting forever.
1862 * -- see perf_not_pending().
1873 static inline int perf_not_pending(struct perf_counter *counter)
1876 * If we flush on whatever cpu we run, there is a chance we don't
1880 __perf_pending_run();
1884 * Ensure we see the proper queue state before going to sleep
1885 * so that we do not miss the wakeup. -- see perf_pending_handle()
1888 return counter->pending.next == NULL;
1891 static void perf_pending_sync(struct perf_counter *counter)
1893 wait_event(counter->waitq, perf_not_pending(counter));
1896 void perf_counter_do_pending(void)
1898 __perf_pending_run();
1902 * Callchain support -- arch specific
1905 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1914 struct perf_output_handle {
1915 struct perf_counter *counter;
1916 struct perf_mmap_data *data;
1917 unsigned int offset;
1922 unsigned long flags;
1925 static void perf_output_wakeup(struct perf_output_handle *handle)
1927 atomic_set(&handle->data->poll, POLL_IN);
1930 handle->counter->pending_wakeup = 1;
1931 perf_pending_queue(&handle->counter->pending,
1932 perf_pending_counter);
1934 perf_counter_wakeup(handle->counter);
1938 * Curious locking construct.
1940 * We need to ensure a later event doesn't publish a head when a former
1941 * event isn't done writing. However since we need to deal with NMIs we
1942 * cannot fully serialize things.
1944 * What we do is serialize between CPUs so we only have to deal with NMI
1945 * nesting on a single CPU.
1947 * We only publish the head (and generate a wakeup) when the outer-most
1950 static void perf_output_lock(struct perf_output_handle *handle)
1952 struct perf_mmap_data *data = handle->data;
1957 local_irq_save(handle->flags);
1958 cpu = smp_processor_id();
1960 if (in_nmi() && atomic_read(&data->lock) == cpu)
1963 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1969 static void perf_output_unlock(struct perf_output_handle *handle)
1971 struct perf_mmap_data *data = handle->data;
1974 data->done_head = data->head;
1976 if (!handle->locked)
1981 * The xchg implies a full barrier that ensures all writes are done
1982 * before we publish the new head, matched by a rmb() in userspace when
1983 * reading this position.
1985 while ((head = atomic_xchg(&data->done_head, 0)))
1986 data->user_page->data_head = head;
1989 * NMI can happen here, which means we can miss a done_head update.
1992 cpu = atomic_xchg(&data->lock, -1);
1993 WARN_ON_ONCE(cpu != smp_processor_id());
1996 * Therefore we have to validate we did not indeed do so.
1998 if (unlikely(atomic_read(&data->done_head))) {
2000 * Since we had it locked, we can lock it again.
2002 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2008 if (atomic_xchg(&data->wakeup, 0))
2009 perf_output_wakeup(handle);
2011 local_irq_restore(handle->flags);
2014 static int perf_output_begin(struct perf_output_handle *handle,
2015 struct perf_counter *counter, unsigned int size,
2016 int nmi, int overflow)
2018 struct perf_mmap_data *data;
2019 unsigned int offset, head;
2022 * For inherited counters we send all the output towards the parent.
2024 if (counter->parent)
2025 counter = counter->parent;
2028 data = rcu_dereference(counter->data);
2032 handle->data = data;
2033 handle->counter = counter;
2035 handle->overflow = overflow;
2037 if (!data->nr_pages)
2040 perf_output_lock(handle);
2043 offset = head = atomic_read(&data->head);
2045 } while (atomic_cmpxchg(&data->head, offset, head) != offset);
2047 handle->offset = offset;
2048 handle->head = head;
2050 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
2051 atomic_set(&data->wakeup, 1);
2056 perf_output_wakeup(handle);
2063 static void perf_output_copy(struct perf_output_handle *handle,
2064 void *buf, unsigned int len)
2066 unsigned int pages_mask;
2067 unsigned int offset;
2071 offset = handle->offset;
2072 pages_mask = handle->data->nr_pages - 1;
2073 pages = handle->data->data_pages;
2076 unsigned int page_offset;
2079 nr = (offset >> PAGE_SHIFT) & pages_mask;
2080 page_offset = offset & (PAGE_SIZE - 1);
2081 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
2083 memcpy(pages[nr] + page_offset, buf, size);
2090 handle->offset = offset;
2093 * Check we didn't copy past our reservation window, taking the
2094 * possible unsigned int wrap into account.
2096 WARN_ON_ONCE(((int)(handle->head - handle->offset)) < 0);
2099 #define perf_output_put(handle, x) \
2100 perf_output_copy((handle), &(x), sizeof(x))
2102 static void perf_output_end(struct perf_output_handle *handle)
2104 struct perf_counter *counter = handle->counter;
2105 struct perf_mmap_data *data = handle->data;
2107 int wakeup_events = counter->hw_event.wakeup_events;
2109 if (handle->overflow && wakeup_events) {
2110 int events = atomic_inc_return(&data->events);
2111 if (events >= wakeup_events) {
2112 atomic_sub(wakeup_events, &data->events);
2113 atomic_set(&data->wakeup, 1);
2117 perf_output_unlock(handle);
2121 static void perf_counter_output(struct perf_counter *counter,
2122 int nmi, struct pt_regs *regs, u64 addr)
2125 u64 record_type = counter->hw_event.record_type;
2126 struct perf_output_handle handle;
2127 struct perf_event_header header;
2136 struct perf_callchain_entry *callchain = NULL;
2137 int callchain_size = 0;
2144 header.size = sizeof(header);
2146 header.misc = PERF_EVENT_MISC_OVERFLOW;
2147 header.misc |= perf_misc_flags(regs);
2149 if (record_type & PERF_RECORD_IP) {
2150 ip = perf_instruction_pointer(regs);
2151 header.type |= PERF_RECORD_IP;
2152 header.size += sizeof(ip);
2155 if (record_type & PERF_RECORD_TID) {
2156 /* namespace issues */
2157 tid_entry.pid = current->group_leader->pid;
2158 tid_entry.tid = current->pid;
2160 header.type |= PERF_RECORD_TID;
2161 header.size += sizeof(tid_entry);
2164 if (record_type & PERF_RECORD_TIME) {
2166 * Maybe do better on x86 and provide cpu_clock_nmi()
2168 time = sched_clock();
2170 header.type |= PERF_RECORD_TIME;
2171 header.size += sizeof(u64);
2174 if (record_type & PERF_RECORD_ADDR) {
2175 header.type |= PERF_RECORD_ADDR;
2176 header.size += sizeof(u64);
2179 if (record_type & PERF_RECORD_CONFIG) {
2180 header.type |= PERF_RECORD_CONFIG;
2181 header.size += sizeof(u64);
2184 if (record_type & PERF_RECORD_CPU) {
2185 header.type |= PERF_RECORD_CPU;
2186 header.size += sizeof(cpu_entry);
2188 cpu_entry.cpu = raw_smp_processor_id();
2191 if (record_type & PERF_RECORD_GROUP) {
2192 header.type |= PERF_RECORD_GROUP;
2193 header.size += sizeof(u64) +
2194 counter->nr_siblings * sizeof(group_entry);
2197 if (record_type & PERF_RECORD_CALLCHAIN) {
2198 callchain = perf_callchain(regs);
2201 callchain_size = (1 + callchain->nr) * sizeof(u64);
2203 header.type |= PERF_RECORD_CALLCHAIN;
2204 header.size += callchain_size;
2208 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2212 perf_output_put(&handle, header);
2214 if (record_type & PERF_RECORD_IP)
2215 perf_output_put(&handle, ip);
2217 if (record_type & PERF_RECORD_TID)
2218 perf_output_put(&handle, tid_entry);
2220 if (record_type & PERF_RECORD_TIME)
2221 perf_output_put(&handle, time);
2223 if (record_type & PERF_RECORD_ADDR)
2224 perf_output_put(&handle, addr);
2226 if (record_type & PERF_RECORD_CONFIG)
2227 perf_output_put(&handle, counter->hw_event.config);
2229 if (record_type & PERF_RECORD_CPU)
2230 perf_output_put(&handle, cpu_entry);
2233 * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
2235 if (record_type & PERF_RECORD_GROUP) {
2236 struct perf_counter *leader, *sub;
2237 u64 nr = counter->nr_siblings;
2239 perf_output_put(&handle, nr);
2241 leader = counter->group_leader;
2242 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2244 sub->pmu->read(sub);
2246 group_entry.event = sub->hw_event.config;
2247 group_entry.counter = atomic64_read(&sub->count);
2249 perf_output_put(&handle, group_entry);
2254 perf_output_copy(&handle, callchain, callchain_size);
2256 perf_output_end(&handle);
2263 struct perf_comm_event {
2264 struct task_struct *task;
2269 struct perf_event_header header;
2276 static void perf_counter_comm_output(struct perf_counter *counter,
2277 struct perf_comm_event *comm_event)
2279 struct perf_output_handle handle;
2280 int size = comm_event->event.header.size;
2281 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2286 perf_output_put(&handle, comm_event->event);
2287 perf_output_copy(&handle, comm_event->comm,
2288 comm_event->comm_size);
2289 perf_output_end(&handle);
2292 static int perf_counter_comm_match(struct perf_counter *counter,
2293 struct perf_comm_event *comm_event)
2295 if (counter->hw_event.comm &&
2296 comm_event->event.header.type == PERF_EVENT_COMM)
2302 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2303 struct perf_comm_event *comm_event)
2305 struct perf_counter *counter;
2307 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2311 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2312 if (perf_counter_comm_match(counter, comm_event))
2313 perf_counter_comm_output(counter, comm_event);
2318 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2320 struct perf_cpu_context *cpuctx;
2322 char *comm = comm_event->task->comm;
2324 size = ALIGN(strlen(comm)+1, sizeof(u64));
2326 comm_event->comm = comm;
2327 comm_event->comm_size = size;
2329 comm_event->event.header.size = sizeof(comm_event->event) + size;
2331 cpuctx = &get_cpu_var(perf_cpu_context);
2332 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2333 put_cpu_var(perf_cpu_context);
2335 perf_counter_comm_ctx(current->perf_counter_ctxp, comm_event);
2338 void perf_counter_comm(struct task_struct *task)
2340 struct perf_comm_event comm_event;
2342 if (!atomic_read(&nr_comm_tracking))
2344 if (!current->perf_counter_ctxp)
2347 comm_event = (struct perf_comm_event){
2350 .header = { .type = PERF_EVENT_COMM, },
2351 .pid = task->group_leader->pid,
2356 perf_counter_comm_event(&comm_event);
2363 struct perf_mmap_event {
2369 struct perf_event_header header;
2379 static void perf_counter_mmap_output(struct perf_counter *counter,
2380 struct perf_mmap_event *mmap_event)
2382 struct perf_output_handle handle;
2383 int size = mmap_event->event.header.size;
2384 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2389 perf_output_put(&handle, mmap_event->event);
2390 perf_output_copy(&handle, mmap_event->file_name,
2391 mmap_event->file_size);
2392 perf_output_end(&handle);
2395 static int perf_counter_mmap_match(struct perf_counter *counter,
2396 struct perf_mmap_event *mmap_event)
2398 if (counter->hw_event.mmap &&
2399 mmap_event->event.header.type == PERF_EVENT_MMAP)
2402 if (counter->hw_event.munmap &&
2403 mmap_event->event.header.type == PERF_EVENT_MUNMAP)
2409 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
2410 struct perf_mmap_event *mmap_event)
2412 struct perf_counter *counter;
2414 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2418 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2419 if (perf_counter_mmap_match(counter, mmap_event))
2420 perf_counter_mmap_output(counter, mmap_event);
2425 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
2427 struct perf_cpu_context *cpuctx;
2428 struct file *file = mmap_event->file;
2435 buf = kzalloc(PATH_MAX, GFP_KERNEL);
2437 name = strncpy(tmp, "//enomem", sizeof(tmp));
2440 name = d_path(&file->f_path, buf, PATH_MAX);
2442 name = strncpy(tmp, "//toolong", sizeof(tmp));
2446 name = strncpy(tmp, "//anon", sizeof(tmp));
2451 size = ALIGN(strlen(name)+1, sizeof(u64));
2453 mmap_event->file_name = name;
2454 mmap_event->file_size = size;
2456 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
2458 cpuctx = &get_cpu_var(perf_cpu_context);
2459 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
2460 put_cpu_var(perf_cpu_context);
2462 perf_counter_mmap_ctx(current->perf_counter_ctxp, mmap_event);
2467 void perf_counter_mmap(unsigned long addr, unsigned long len,
2468 unsigned long pgoff, struct file *file)
2470 struct perf_mmap_event mmap_event;
2472 if (!atomic_read(&nr_mmap_tracking))
2474 if (!current->perf_counter_ctxp)
2477 mmap_event = (struct perf_mmap_event){
2480 .header = { .type = PERF_EVENT_MMAP, },
2481 .pid = current->group_leader->pid,
2482 .tid = current->pid,
2489 perf_counter_mmap_event(&mmap_event);
2492 void perf_counter_munmap(unsigned long addr, unsigned long len,
2493 unsigned long pgoff, struct file *file)
2495 struct perf_mmap_event mmap_event;
2497 if (!atomic_read(&nr_munmap_tracking))
2500 mmap_event = (struct perf_mmap_event){
2503 .header = { .type = PERF_EVENT_MUNMAP, },
2504 .pid = current->group_leader->pid,
2505 .tid = current->pid,
2512 perf_counter_mmap_event(&mmap_event);
2516 * Log irq_period changes so that analyzing tools can re-normalize the
2520 static void perf_log_period(struct perf_counter *counter, u64 period)
2522 struct perf_output_handle handle;
2526 struct perf_event_header header;
2531 .type = PERF_EVENT_PERIOD,
2533 .size = sizeof(freq_event),
2535 .time = sched_clock(),
2539 if (counter->hw.irq_period == period)
2542 ret = perf_output_begin(&handle, counter, sizeof(freq_event), 0, 0);
2546 perf_output_put(&handle, freq_event);
2547 perf_output_end(&handle);
2551 * Generic counter overflow handling.
2554 int perf_counter_overflow(struct perf_counter *counter,
2555 int nmi, struct pt_regs *regs, u64 addr)
2557 int events = atomic_read(&counter->event_limit);
2560 counter->hw.interrupts++;
2563 * XXX event_limit might not quite work as expected on inherited
2567 counter->pending_kill = POLL_IN;
2568 if (events && atomic_dec_and_test(&counter->event_limit)) {
2570 counter->pending_kill = POLL_HUP;
2572 counter->pending_disable = 1;
2573 perf_pending_queue(&counter->pending,
2574 perf_pending_counter);
2576 perf_counter_disable(counter);
2579 perf_counter_output(counter, nmi, regs, addr);
2584 * Generic software counter infrastructure
2587 static void perf_swcounter_update(struct perf_counter *counter)
2589 struct hw_perf_counter *hwc = &counter->hw;
2594 prev = atomic64_read(&hwc->prev_count);
2595 now = atomic64_read(&hwc->count);
2596 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
2601 atomic64_add(delta, &counter->count);
2602 atomic64_sub(delta, &hwc->period_left);
2605 static void perf_swcounter_set_period(struct perf_counter *counter)
2607 struct hw_perf_counter *hwc = &counter->hw;
2608 s64 left = atomic64_read(&hwc->period_left);
2609 s64 period = hwc->irq_period;
2611 if (unlikely(left <= -period)) {
2613 atomic64_set(&hwc->period_left, left);
2616 if (unlikely(left <= 0)) {
2618 atomic64_add(period, &hwc->period_left);
2621 atomic64_set(&hwc->prev_count, -left);
2622 atomic64_set(&hwc->count, -left);
2625 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
2627 enum hrtimer_restart ret = HRTIMER_RESTART;
2628 struct perf_counter *counter;
2629 struct pt_regs *regs;
2632 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
2633 counter->pmu->read(counter);
2635 regs = get_irq_regs();
2637 * In case we exclude kernel IPs or are somehow not in interrupt
2638 * context, provide the next best thing, the user IP.
2640 if ((counter->hw_event.exclude_kernel || !regs) &&
2641 !counter->hw_event.exclude_user)
2642 regs = task_pt_regs(current);
2645 if (perf_counter_overflow(counter, 0, regs, 0))
2646 ret = HRTIMER_NORESTART;
2649 period = max_t(u64, 10000, counter->hw.irq_period);
2650 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
2655 static void perf_swcounter_overflow(struct perf_counter *counter,
2656 int nmi, struct pt_regs *regs, u64 addr)
2658 perf_swcounter_update(counter);
2659 perf_swcounter_set_period(counter);
2660 if (perf_counter_overflow(counter, nmi, regs, addr))
2661 /* soft-disable the counter */
2666 static int perf_swcounter_match(struct perf_counter *counter,
2667 enum perf_event_types type,
2668 u32 event, struct pt_regs *regs)
2670 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2673 if (perf_event_raw(&counter->hw_event))
2676 if (perf_event_type(&counter->hw_event) != type)
2679 if (perf_event_id(&counter->hw_event) != event)
2682 if (counter->hw_event.exclude_user && user_mode(regs))
2685 if (counter->hw_event.exclude_kernel && !user_mode(regs))
2691 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
2692 int nmi, struct pt_regs *regs, u64 addr)
2694 int neg = atomic64_add_negative(nr, &counter->hw.count);
2695 if (counter->hw.irq_period && !neg)
2696 perf_swcounter_overflow(counter, nmi, regs, addr);
2699 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
2700 enum perf_event_types type, u32 event,
2701 u64 nr, int nmi, struct pt_regs *regs,
2704 struct perf_counter *counter;
2706 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2710 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2711 if (perf_swcounter_match(counter, type, event, regs))
2712 perf_swcounter_add(counter, nr, nmi, regs, addr);
2717 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
2720 return &cpuctx->recursion[3];
2723 return &cpuctx->recursion[2];
2726 return &cpuctx->recursion[1];
2728 return &cpuctx->recursion[0];
2731 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
2732 u64 nr, int nmi, struct pt_regs *regs,
2735 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
2736 int *recursion = perf_swcounter_recursion_context(cpuctx);
2744 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
2745 nr, nmi, regs, addr);
2746 if (cpuctx->task_ctx) {
2747 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
2748 nr, nmi, regs, addr);
2755 put_cpu_var(perf_cpu_context);
2759 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
2761 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs, addr);
2764 static void perf_swcounter_read(struct perf_counter *counter)
2766 perf_swcounter_update(counter);
2769 static int perf_swcounter_enable(struct perf_counter *counter)
2771 perf_swcounter_set_period(counter);
2775 static void perf_swcounter_disable(struct perf_counter *counter)
2777 perf_swcounter_update(counter);
2780 static const struct pmu perf_ops_generic = {
2781 .enable = perf_swcounter_enable,
2782 .disable = perf_swcounter_disable,
2783 .read = perf_swcounter_read,
2787 * Software counter: cpu wall time clock
2790 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
2792 int cpu = raw_smp_processor_id();
2796 now = cpu_clock(cpu);
2797 prev = atomic64_read(&counter->hw.prev_count);
2798 atomic64_set(&counter->hw.prev_count, now);
2799 atomic64_add(now - prev, &counter->count);
2802 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
2804 struct hw_perf_counter *hwc = &counter->hw;
2805 int cpu = raw_smp_processor_id();
2807 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
2808 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2809 hwc->hrtimer.function = perf_swcounter_hrtimer;
2810 if (hwc->irq_period) {
2811 u64 period = max_t(u64, 10000, hwc->irq_period);
2812 __hrtimer_start_range_ns(&hwc->hrtimer,
2813 ns_to_ktime(period), 0,
2814 HRTIMER_MODE_REL, 0);
2820 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
2822 if (counter->hw.irq_period)
2823 hrtimer_cancel(&counter->hw.hrtimer);
2824 cpu_clock_perf_counter_update(counter);
2827 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
2829 cpu_clock_perf_counter_update(counter);
2832 static const struct pmu perf_ops_cpu_clock = {
2833 .enable = cpu_clock_perf_counter_enable,
2834 .disable = cpu_clock_perf_counter_disable,
2835 .read = cpu_clock_perf_counter_read,
2839 * Software counter: task time clock
2842 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
2847 prev = atomic64_xchg(&counter->hw.prev_count, now);
2849 atomic64_add(delta, &counter->count);
2852 static int task_clock_perf_counter_enable(struct perf_counter *counter)
2854 struct hw_perf_counter *hwc = &counter->hw;
2857 now = counter->ctx->time;
2859 atomic64_set(&hwc->prev_count, now);
2860 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2861 hwc->hrtimer.function = perf_swcounter_hrtimer;
2862 if (hwc->irq_period) {
2863 u64 period = max_t(u64, 10000, hwc->irq_period);
2864 __hrtimer_start_range_ns(&hwc->hrtimer,
2865 ns_to_ktime(period), 0,
2866 HRTIMER_MODE_REL, 0);
2872 static void task_clock_perf_counter_disable(struct perf_counter *counter)
2874 if (counter->hw.irq_period)
2875 hrtimer_cancel(&counter->hw.hrtimer);
2876 task_clock_perf_counter_update(counter, counter->ctx->time);
2880 static void task_clock_perf_counter_read(struct perf_counter *counter)
2885 update_context_time(counter->ctx);
2886 time = counter->ctx->time;
2888 u64 now = perf_clock();
2889 u64 delta = now - counter->ctx->timestamp;
2890 time = counter->ctx->time + delta;
2893 task_clock_perf_counter_update(counter, time);
2896 static const struct pmu perf_ops_task_clock = {
2897 .enable = task_clock_perf_counter_enable,
2898 .disable = task_clock_perf_counter_disable,
2899 .read = task_clock_perf_counter_read,
2903 * Software counter: cpu migrations
2906 static inline u64 get_cpu_migrations(struct perf_counter *counter)
2908 struct task_struct *curr = counter->ctx->task;
2911 return curr->se.nr_migrations;
2912 return cpu_nr_migrations(smp_processor_id());
2915 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
2920 prev = atomic64_read(&counter->hw.prev_count);
2921 now = get_cpu_migrations(counter);
2923 atomic64_set(&counter->hw.prev_count, now);
2927 atomic64_add(delta, &counter->count);
2930 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
2932 cpu_migrations_perf_counter_update(counter);
2935 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
2937 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
2938 atomic64_set(&counter->hw.prev_count,
2939 get_cpu_migrations(counter));
2943 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
2945 cpu_migrations_perf_counter_update(counter);
2948 static const struct pmu perf_ops_cpu_migrations = {
2949 .enable = cpu_migrations_perf_counter_enable,
2950 .disable = cpu_migrations_perf_counter_disable,
2951 .read = cpu_migrations_perf_counter_read,
2954 #ifdef CONFIG_EVENT_PROFILE
2955 void perf_tpcounter_event(int event_id)
2957 struct pt_regs *regs = get_irq_regs();
2960 regs = task_pt_regs(current);
2962 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs, 0);
2964 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
2966 extern int ftrace_profile_enable(int);
2967 extern void ftrace_profile_disable(int);
2969 static void tp_perf_counter_destroy(struct perf_counter *counter)
2971 ftrace_profile_disable(perf_event_id(&counter->hw_event));
2974 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2976 int event_id = perf_event_id(&counter->hw_event);
2979 ret = ftrace_profile_enable(event_id);
2983 counter->destroy = tp_perf_counter_destroy;
2984 counter->hw.irq_period = counter->hw_event.irq_period;
2986 return &perf_ops_generic;
2989 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2995 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
2997 const struct pmu *pmu = NULL;
3000 * Software counters (currently) can't in general distinguish
3001 * between user, kernel and hypervisor events.
3002 * However, context switches and cpu migrations are considered
3003 * to be kernel events, and page faults are never hypervisor
3006 switch (perf_event_id(&counter->hw_event)) {
3007 case PERF_COUNT_CPU_CLOCK:
3008 pmu = &perf_ops_cpu_clock;
3011 case PERF_COUNT_TASK_CLOCK:
3013 * If the user instantiates this as a per-cpu counter,
3014 * use the cpu_clock counter instead.
3016 if (counter->ctx->task)
3017 pmu = &perf_ops_task_clock;
3019 pmu = &perf_ops_cpu_clock;
3022 case PERF_COUNT_PAGE_FAULTS:
3023 case PERF_COUNT_PAGE_FAULTS_MIN:
3024 case PERF_COUNT_PAGE_FAULTS_MAJ:
3025 case PERF_COUNT_CONTEXT_SWITCHES:
3026 pmu = &perf_ops_generic;
3028 case PERF_COUNT_CPU_MIGRATIONS:
3029 if (!counter->hw_event.exclude_kernel)
3030 pmu = &perf_ops_cpu_migrations;
3038 * Allocate and initialize a counter structure
3040 static struct perf_counter *
3041 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
3043 struct perf_counter_context *ctx,
3044 struct perf_counter *group_leader,
3047 const struct pmu *pmu;
3048 struct perf_counter *counter;
3049 struct hw_perf_counter *hwc;
3052 counter = kzalloc(sizeof(*counter), gfpflags);
3054 return ERR_PTR(-ENOMEM);
3057 * Single counters are their own group leaders, with an
3058 * empty sibling list:
3061 group_leader = counter;
3063 mutex_init(&counter->child_mutex);
3064 INIT_LIST_HEAD(&counter->child_list);
3066 INIT_LIST_HEAD(&counter->list_entry);
3067 INIT_LIST_HEAD(&counter->event_entry);
3068 INIT_LIST_HEAD(&counter->sibling_list);
3069 init_waitqueue_head(&counter->waitq);
3071 mutex_init(&counter->mmap_mutex);
3074 counter->hw_event = *hw_event;
3075 counter->group_leader = group_leader;
3076 counter->pmu = NULL;
3080 counter->state = PERF_COUNTER_STATE_INACTIVE;
3081 if (hw_event->disabled)
3082 counter->state = PERF_COUNTER_STATE_OFF;
3087 if (hw_event->freq && hw_event->irq_freq)
3088 hwc->irq_period = div64_u64(TICK_NSEC, hw_event->irq_freq);
3090 hwc->irq_period = hw_event->irq_period;
3093 * we currently do not support PERF_RECORD_GROUP on inherited counters
3095 if (hw_event->inherit && (hw_event->record_type & PERF_RECORD_GROUP))
3098 if (perf_event_raw(hw_event)) {
3099 pmu = hw_perf_counter_init(counter);
3103 switch (perf_event_type(hw_event)) {
3104 case PERF_TYPE_HARDWARE:
3105 pmu = hw_perf_counter_init(counter);
3108 case PERF_TYPE_SOFTWARE:
3109 pmu = sw_perf_counter_init(counter);
3112 case PERF_TYPE_TRACEPOINT:
3113 pmu = tp_perf_counter_init(counter);
3120 else if (IS_ERR(pmu))
3125 return ERR_PTR(err);
3130 atomic_inc(&nr_counters);
3131 if (counter->hw_event.mmap)
3132 atomic_inc(&nr_mmap_tracking);
3133 if (counter->hw_event.munmap)
3134 atomic_inc(&nr_munmap_tracking);
3135 if (counter->hw_event.comm)
3136 atomic_inc(&nr_comm_tracking);
3142 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3144 * @hw_event_uptr: event type attributes for monitoring/sampling
3147 * @group_fd: group leader counter fd
3149 SYSCALL_DEFINE5(perf_counter_open,
3150 const struct perf_counter_hw_event __user *, hw_event_uptr,
3151 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
3153 struct perf_counter *counter, *group_leader;
3154 struct perf_counter_hw_event hw_event;
3155 struct perf_counter_context *ctx;
3156 struct file *counter_file = NULL;
3157 struct file *group_file = NULL;
3158 int fput_needed = 0;
3159 int fput_needed2 = 0;
3162 /* for future expandability... */
3166 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
3170 * Get the target context (task or percpu):
3172 ctx = find_get_context(pid, cpu);
3174 return PTR_ERR(ctx);
3177 * Look up the group leader (we will attach this counter to it):
3179 group_leader = NULL;
3180 if (group_fd != -1) {
3182 group_file = fget_light(group_fd, &fput_needed);
3184 goto err_put_context;
3185 if (group_file->f_op != &perf_fops)
3186 goto err_put_context;
3188 group_leader = group_file->private_data;
3190 * Do not allow a recursive hierarchy (this new sibling
3191 * becoming part of another group-sibling):
3193 if (group_leader->group_leader != group_leader)
3194 goto err_put_context;
3196 * Do not allow to attach to a group in a different
3197 * task or CPU context:
3199 if (group_leader->ctx != ctx)
3200 goto err_put_context;
3202 * Only a group leader can be exclusive or pinned
3204 if (hw_event.exclusive || hw_event.pinned)
3205 goto err_put_context;
3208 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
3210 ret = PTR_ERR(counter);
3211 if (IS_ERR(counter))
3212 goto err_put_context;
3214 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
3216 goto err_free_put_context;
3218 counter_file = fget_light(ret, &fput_needed2);
3220 goto err_free_put_context;
3222 counter->filp = counter_file;
3223 mutex_lock(&ctx->mutex);
3224 perf_install_in_context(ctx, counter, cpu);
3225 mutex_unlock(&ctx->mutex);
3227 counter->owner = current;
3228 get_task_struct(current);
3229 mutex_lock(¤t->perf_counter_mutex);
3230 list_add_tail(&counter->owner_entry, ¤t->perf_counter_list);
3231 mutex_unlock(¤t->perf_counter_mutex);
3233 fput_light(counter_file, fput_needed2);
3236 fput_light(group_file, fput_needed);
3240 err_free_put_context:
3250 * inherit a counter from parent task to child task:
3252 static struct perf_counter *
3253 inherit_counter(struct perf_counter *parent_counter,
3254 struct task_struct *parent,
3255 struct perf_counter_context *parent_ctx,
3256 struct task_struct *child,
3257 struct perf_counter *group_leader,
3258 struct perf_counter_context *child_ctx)
3260 struct perf_counter *child_counter;
3263 * Instead of creating recursive hierarchies of counters,
3264 * we link inherited counters back to the original parent,
3265 * which has a filp for sure, which we use as the reference
3268 if (parent_counter->parent)
3269 parent_counter = parent_counter->parent;
3271 child_counter = perf_counter_alloc(&parent_counter->hw_event,
3272 parent_counter->cpu, child_ctx,
3273 group_leader, GFP_KERNEL);
3274 if (IS_ERR(child_counter))
3275 return child_counter;
3278 * Make the child state follow the state of the parent counter,
3279 * not its hw_event.disabled bit. We hold the parent's mutex,
3280 * so we won't race with perf_counter_{en,dis}able_family.
3282 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
3283 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
3285 child_counter->state = PERF_COUNTER_STATE_OFF;
3288 * Link it up in the child's context:
3290 add_counter_to_ctx(child_counter, child_ctx);
3292 child_counter->parent = parent_counter;
3294 * inherit into child's child as well:
3296 child_counter->hw_event.inherit = 1;
3299 * Get a reference to the parent filp - we will fput it
3300 * when the child counter exits. This is safe to do because
3301 * we are in the parent and we know that the filp still
3302 * exists and has a nonzero count:
3304 atomic_long_inc(&parent_counter->filp->f_count);
3307 * Link this into the parent counter's child list
3309 mutex_lock(&parent_counter->child_mutex);
3310 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
3311 mutex_unlock(&parent_counter->child_mutex);
3313 return child_counter;
3316 static int inherit_group(struct perf_counter *parent_counter,
3317 struct task_struct *parent,
3318 struct perf_counter_context *parent_ctx,
3319 struct task_struct *child,
3320 struct perf_counter_context *child_ctx)
3322 struct perf_counter *leader;
3323 struct perf_counter *sub;
3324 struct perf_counter *child_ctr;
3326 leader = inherit_counter(parent_counter, parent, parent_ctx,
3327 child, NULL, child_ctx);
3329 return PTR_ERR(leader);
3330 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
3331 child_ctr = inherit_counter(sub, parent, parent_ctx,
3332 child, leader, child_ctx);
3333 if (IS_ERR(child_ctr))
3334 return PTR_ERR(child_ctr);
3339 static void sync_child_counter(struct perf_counter *child_counter,
3340 struct perf_counter *parent_counter)
3344 child_val = atomic64_read(&child_counter->count);
3347 * Add back the child's count to the parent's count:
3349 atomic64_add(child_val, &parent_counter->count);
3350 atomic64_add(child_counter->total_time_enabled,
3351 &parent_counter->child_total_time_enabled);
3352 atomic64_add(child_counter->total_time_running,
3353 &parent_counter->child_total_time_running);
3356 * Remove this counter from the parent's list
3358 mutex_lock(&parent_counter->child_mutex);
3359 list_del_init(&child_counter->child_list);
3360 mutex_unlock(&parent_counter->child_mutex);
3363 * Release the parent counter, if this was the last
3366 fput(parent_counter->filp);
3370 __perf_counter_exit_task(struct task_struct *child,
3371 struct perf_counter *child_counter,
3372 struct perf_counter_context *child_ctx)
3374 struct perf_counter *parent_counter;
3376 update_counter_times(child_counter);
3377 perf_counter_remove_from_context(child_counter);
3379 parent_counter = child_counter->parent;
3381 * It can happen that parent exits first, and has counters
3382 * that are still around due to the child reference. These
3383 * counters need to be zapped - but otherwise linger.
3385 if (parent_counter) {
3386 sync_child_counter(child_counter, parent_counter);
3387 free_counter(child_counter);
3392 * When a child task exits, feed back counter values to parent counters.
3394 * Note: we may be running in child context, but the PID is not hashed
3395 * anymore so new counters will not be added.
3396 * (XXX not sure that is true when we get called from flush_old_exec.
3399 void perf_counter_exit_task(struct task_struct *child)
3401 struct perf_counter *child_counter, *tmp;
3402 struct perf_counter_context *child_ctx;
3403 unsigned long flags;
3405 WARN_ON_ONCE(child != current);
3407 child_ctx = child->perf_counter_ctxp;
3409 if (likely(!child_ctx))
3412 local_irq_save(flags);
3413 __perf_counter_task_sched_out(child_ctx);
3414 child->perf_counter_ctxp = NULL;
3415 local_irq_restore(flags);
3417 mutex_lock(&child_ctx->mutex);
3420 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
3422 __perf_counter_exit_task(child, child_counter, child_ctx);
3425 * If the last counter was a group counter, it will have appended all
3426 * its siblings to the list, but we obtained 'tmp' before that which
3427 * will still point to the list head terminating the iteration.
3429 if (!list_empty(&child_ctx->counter_list))
3432 mutex_unlock(&child_ctx->mutex);
3438 * Initialize the perf_counter context in task_struct
3440 void perf_counter_init_task(struct task_struct *child)
3442 struct perf_counter_context *child_ctx, *parent_ctx;
3443 struct perf_counter *counter;
3444 struct task_struct *parent = current;
3445 int inherited_all = 1;
3447 child->perf_counter_ctxp = NULL;
3449 mutex_init(&child->perf_counter_mutex);
3450 INIT_LIST_HEAD(&child->perf_counter_list);
3453 * This is executed from the parent task context, so inherit
3454 * counters that have been marked for cloning.
3455 * First allocate and initialize a context for the child.
3458 child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
3462 parent_ctx = parent->perf_counter_ctxp;
3463 if (likely(!parent_ctx || !parent_ctx->nr_counters))
3466 __perf_counter_init_context(child_ctx, child);
3467 child->perf_counter_ctxp = child_ctx;
3470 * Lock the parent list. No need to lock the child - not PID
3471 * hashed yet and not running, so nobody can access it.
3473 mutex_lock(&parent_ctx->mutex);
3476 * We dont have to disable NMIs - we are only looking at
3477 * the list, not manipulating it:
3479 list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
3480 if (counter != counter->group_leader)
3483 if (!counter->hw_event.inherit) {
3488 if (inherit_group(counter, parent,
3489 parent_ctx, child, child_ctx)) {
3495 if (inherited_all) {
3497 * Mark the child context as a clone of the parent
3498 * context, or of whatever the parent is a clone of.
3500 if (parent_ctx->parent_ctx) {
3501 child_ctx->parent_ctx = parent_ctx->parent_ctx;
3502 child_ctx->parent_gen = parent_ctx->parent_gen;
3504 child_ctx->parent_ctx = parent_ctx;
3505 child_ctx->parent_gen = parent_ctx->generation;
3507 get_ctx(child_ctx->parent_ctx);
3510 mutex_unlock(&parent_ctx->mutex);
3513 static void __cpuinit perf_counter_init_cpu(int cpu)
3515 struct perf_cpu_context *cpuctx;
3517 cpuctx = &per_cpu(perf_cpu_context, cpu);
3518 __perf_counter_init_context(&cpuctx->ctx, NULL);
3520 spin_lock(&perf_resource_lock);
3521 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
3522 spin_unlock(&perf_resource_lock);
3524 hw_perf_counter_setup(cpu);
3527 #ifdef CONFIG_HOTPLUG_CPU
3528 static void __perf_counter_exit_cpu(void *info)
3530 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3531 struct perf_counter_context *ctx = &cpuctx->ctx;
3532 struct perf_counter *counter, *tmp;
3534 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
3535 __perf_counter_remove_from_context(counter);
3537 static void perf_counter_exit_cpu(int cpu)
3539 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3540 struct perf_counter_context *ctx = &cpuctx->ctx;
3542 mutex_lock(&ctx->mutex);
3543 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
3544 mutex_unlock(&ctx->mutex);
3547 static inline void perf_counter_exit_cpu(int cpu) { }
3550 static int __cpuinit
3551 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
3553 unsigned int cpu = (long)hcpu;
3557 case CPU_UP_PREPARE:
3558 case CPU_UP_PREPARE_FROZEN:
3559 perf_counter_init_cpu(cpu);
3562 case CPU_DOWN_PREPARE:
3563 case CPU_DOWN_PREPARE_FROZEN:
3564 perf_counter_exit_cpu(cpu);
3574 static struct notifier_block __cpuinitdata perf_cpu_nb = {
3575 .notifier_call = perf_cpu_notify,
3578 void __init perf_counter_init(void)
3580 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
3581 (void *)(long)smp_processor_id());
3582 register_cpu_notifier(&perf_cpu_nb);
3585 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
3587 return sprintf(buf, "%d\n", perf_reserved_percpu);
3591 perf_set_reserve_percpu(struct sysdev_class *class,
3595 struct perf_cpu_context *cpuctx;
3599 err = strict_strtoul(buf, 10, &val);
3602 if (val > perf_max_counters)
3605 spin_lock(&perf_resource_lock);
3606 perf_reserved_percpu = val;
3607 for_each_online_cpu(cpu) {
3608 cpuctx = &per_cpu(perf_cpu_context, cpu);
3609 spin_lock_irq(&cpuctx->ctx.lock);
3610 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
3611 perf_max_counters - perf_reserved_percpu);
3612 cpuctx->max_pertask = mpt;
3613 spin_unlock_irq(&cpuctx->ctx.lock);
3615 spin_unlock(&perf_resource_lock);
3620 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
3622 return sprintf(buf, "%d\n", perf_overcommit);
3626 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
3631 err = strict_strtoul(buf, 10, &val);
3637 spin_lock(&perf_resource_lock);
3638 perf_overcommit = val;
3639 spin_unlock(&perf_resource_lock);
3644 static SYSDEV_CLASS_ATTR(
3647 perf_show_reserve_percpu,
3648 perf_set_reserve_percpu
3651 static SYSDEV_CLASS_ATTR(
3654 perf_show_overcommit,
3658 static struct attribute *perfclass_attrs[] = {
3659 &attr_reserve_percpu.attr,
3660 &attr_overcommit.attr,
3664 static struct attribute_group perfclass_attr_group = {
3665 .attrs = perfclass_attrs,
3666 .name = "perf_counters",
3669 static int __init perf_counter_sysfs_init(void)
3671 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
3672 &perfclass_attr_group);
3674 device_initcall(perf_counter_sysfs_init);