2 * Performance counter core code
4 * Copyright(C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2008 Red Hat, Inc., Ingo Molnar
7 * For licencing details see kernel-base/COPYING
11 #include <linux/cpu.h>
12 #include <linux/smp.h>
13 #include <linux/file.h>
14 #include <linux/poll.h>
15 #include <linux/sysfs.h>
16 #include <linux/ptrace.h>
17 #include <linux/percpu.h>
18 #include <linux/uaccess.h>
19 #include <linux/syscalls.h>
20 #include <linux/anon_inodes.h>
21 #include <linux/perf_counter.h>
24 * Each CPU has a list of per CPU counters:
26 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
28 int perf_max_counters __read_mostly = 1;
29 static int perf_reserved_percpu __read_mostly;
30 static int perf_overcommit __read_mostly = 1;
33 * Mutex for (sysadmin-configurable) counter reservations:
35 static DEFINE_MUTEX(perf_resource_mutex);
38 * Architecture provided APIs - weak aliases:
40 extern __weak const struct hw_perf_counter_ops *
41 hw_perf_counter_init(struct perf_counter *counter)
43 return ERR_PTR(-EINVAL);
46 u64 __weak hw_perf_save_disable(void) { return 0; }
47 void __weak hw_perf_restore(u64 ctrl) { }
48 void __weak hw_perf_counter_setup(void) { }
51 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
53 struct perf_counter *group_leader = counter->group_leader;
56 * Depending on whether it is a standalone or sibling counter,
57 * add it straight to the context's counter list, or to the group
58 * leader's sibling list:
60 if (counter->group_leader == counter)
61 list_add_tail(&counter->list_entry, &ctx->counter_list);
63 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
67 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
69 struct perf_counter *sibling, *tmp;
71 list_del_init(&counter->list_entry);
74 * If this was a group counter with sibling counters then
75 * upgrade the siblings to singleton counters by adding them
76 * to the context list directly:
78 list_for_each_entry_safe(sibling, tmp,
79 &counter->sibling_list, list_entry) {
81 list_del_init(&sibling->list_entry);
82 list_add_tail(&sibling->list_entry, &ctx->counter_list);
83 sibling->group_leader = sibling;
88 * Cross CPU call to remove a performance counter
90 * We disable the counter on the hardware level first. After that we
91 * remove it from the context list.
93 static void __perf_counter_remove_from_context(void *info)
95 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
96 struct perf_counter *counter = info;
97 struct perf_counter_context *ctx = counter->ctx;
102 * If this is a task context, we need to check whether it is
103 * the current task context of this cpu. If not it has been
104 * scheduled out before the smp call arrived.
106 if (ctx->task && cpuctx->task_ctx != ctx)
109 spin_lock_irqsave(&ctx->lock, flags);
111 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
112 counter->hw_ops->disable(counter);
113 counter->state = PERF_COUNTER_STATE_INACTIVE;
115 cpuctx->active_oncpu--;
116 counter->task = NULL;
121 * Protect the list operation against NMI by disabling the
122 * counters on a global level. NOP for non NMI based counters.
124 perf_flags = hw_perf_save_disable();
125 list_del_counter(counter, ctx);
126 hw_perf_restore(perf_flags);
130 * Allow more per task counters with respect to the
133 cpuctx->max_pertask =
134 min(perf_max_counters - ctx->nr_counters,
135 perf_max_counters - perf_reserved_percpu);
138 spin_unlock_irqrestore(&ctx->lock, flags);
143 * Remove the counter from a task's (or a CPU's) list of counters.
145 * Must be called with counter->mutex held.
147 * CPU counters are removed with a smp call. For task counters we only
148 * call when the task is on a CPU.
150 static void perf_counter_remove_from_context(struct perf_counter *counter)
152 struct perf_counter_context *ctx = counter->ctx;
153 struct task_struct *task = ctx->task;
157 * Per cpu counters are removed via an smp call and
158 * the removal is always sucessful.
160 smp_call_function_single(counter->cpu,
161 __perf_counter_remove_from_context,
167 task_oncpu_function_call(task, __perf_counter_remove_from_context,
170 spin_lock_irq(&ctx->lock);
172 * If the context is active we need to retry the smp call.
174 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
175 spin_unlock_irq(&ctx->lock);
180 * The lock prevents that this context is scheduled in so we
181 * can remove the counter safely, if the call above did not
184 if (!list_empty(&counter->list_entry)) {
186 list_del_counter(counter, ctx);
187 counter->task = NULL;
189 spin_unlock_irq(&ctx->lock);
193 * Cross CPU call to install and enable a preformance counter
195 static void __perf_install_in_context(void *info)
197 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
198 struct perf_counter *counter = info;
199 struct perf_counter_context *ctx = counter->ctx;
200 int cpu = smp_processor_id();
205 * If this is a task context, we need to check whether it is
206 * the current task context of this cpu. If not it has been
207 * scheduled out before the smp call arrived.
209 if (ctx->task && cpuctx->task_ctx != ctx)
212 spin_lock_irqsave(&ctx->lock, flags);
215 * Protect the list operation against NMI by disabling the
216 * counters on a global level. NOP for non NMI based counters.
218 perf_flags = hw_perf_save_disable();
219 list_add_counter(counter, ctx);
220 hw_perf_restore(perf_flags);
224 if (cpuctx->active_oncpu < perf_max_counters) {
225 counter->state = PERF_COUNTER_STATE_ACTIVE;
226 counter->oncpu = cpu;
228 cpuctx->active_oncpu++;
229 counter->hw_ops->enable(counter);
232 if (!ctx->task && cpuctx->max_pertask)
233 cpuctx->max_pertask--;
235 spin_unlock_irqrestore(&ctx->lock, flags);
239 * Attach a performance counter to a context
241 * First we add the counter to the list with the hardware enable bit
242 * in counter->hw_config cleared.
244 * If the counter is attached to a task which is on a CPU we use a smp
245 * call to enable it in the task context. The task might have been
246 * scheduled away, but we check this in the smp call again.
249 perf_install_in_context(struct perf_counter_context *ctx,
250 struct perf_counter *counter,
253 struct task_struct *task = ctx->task;
258 * Per cpu counters are installed via an smp call and
259 * the install is always sucessful.
261 smp_call_function_single(cpu, __perf_install_in_context,
266 counter->task = task;
268 task_oncpu_function_call(task, __perf_install_in_context,
271 spin_lock_irq(&ctx->lock);
273 * we need to retry the smp call.
275 if (ctx->nr_active && list_empty(&counter->list_entry)) {
276 spin_unlock_irq(&ctx->lock);
281 * The lock prevents that this context is scheduled in so we
282 * can add the counter safely, if it the call above did not
285 if (list_empty(&counter->list_entry)) {
286 list_add_counter(counter, ctx);
289 spin_unlock_irq(&ctx->lock);
293 counter_sched_out(struct perf_counter *counter,
294 struct perf_cpu_context *cpuctx,
295 struct perf_counter_context *ctx)
297 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
300 counter->hw_ops->disable(counter);
301 counter->state = PERF_COUNTER_STATE_INACTIVE;
304 cpuctx->active_oncpu--;
309 group_sched_out(struct perf_counter *group_counter,
310 struct perf_cpu_context *cpuctx,
311 struct perf_counter_context *ctx)
313 struct perf_counter *counter;
315 counter_sched_out(group_counter, cpuctx, ctx);
318 * Schedule out siblings (if any):
320 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
321 counter_sched_out(counter, cpuctx, ctx);
325 * Called from scheduler to remove the counters of the current task,
326 * with interrupts disabled.
328 * We stop each counter and update the counter value in counter->count.
330 * This does not protect us against NMI, but disable()
331 * sets the disabled bit in the control field of counter _before_
332 * accessing the counter control register. If a NMI hits, then it will
333 * not restart the counter.
335 void perf_counter_task_sched_out(struct task_struct *task, int cpu)
337 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
338 struct perf_counter_context *ctx = &task->perf_counter_ctx;
339 struct perf_counter *counter;
341 if (likely(!cpuctx->task_ctx))
344 spin_lock(&ctx->lock);
345 if (ctx->nr_active) {
346 list_for_each_entry(counter, &ctx->counter_list, list_entry)
347 group_sched_out(counter, cpuctx, ctx);
349 spin_unlock(&ctx->lock);
350 cpuctx->task_ctx = NULL;
354 counter_sched_in(struct perf_counter *counter,
355 struct perf_cpu_context *cpuctx,
356 struct perf_counter_context *ctx,
359 if (counter->state == PERF_COUNTER_STATE_OFF)
362 counter->hw_ops->enable(counter);
363 counter->state = PERF_COUNTER_STATE_ACTIVE;
364 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
366 cpuctx->active_oncpu++;
371 group_sched_in(struct perf_counter *group_counter,
372 struct perf_cpu_context *cpuctx,
373 struct perf_counter_context *ctx,
376 struct perf_counter *counter;
379 counter_sched_in(group_counter, cpuctx, ctx, cpu);
382 * Schedule in siblings as one group (if any):
384 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
385 counter_sched_in(counter, cpuctx, ctx, cpu);
393 * Called from scheduler to add the counters of the current task
394 * with interrupts disabled.
396 * We restore the counter value and then enable it.
398 * This does not protect us against NMI, but enable()
399 * sets the enabled bit in the control field of counter _before_
400 * accessing the counter control register. If a NMI hits, then it will
401 * keep the counter running.
403 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
405 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
406 struct perf_counter_context *ctx = &task->perf_counter_ctx;
407 struct perf_counter *counter;
409 if (likely(!ctx->nr_counters))
412 spin_lock(&ctx->lock);
413 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
414 if (ctx->nr_active == cpuctx->max_pertask)
418 * Listen to the 'cpu' scheduling filter constraint
421 if (counter->cpu != -1 && counter->cpu != cpu)
425 * If we scheduled in a group atomically and
426 * exclusively, break out:
428 if (group_sched_in(counter, cpuctx, ctx, cpu))
431 spin_unlock(&ctx->lock);
433 cpuctx->task_ctx = ctx;
436 int perf_counter_task_disable(void)
438 struct task_struct *curr = current;
439 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
440 struct perf_counter *counter;
444 if (likely(!ctx->nr_counters))
448 cpu = smp_processor_id();
450 perf_counter_task_sched_out(curr, cpu);
452 spin_lock(&ctx->lock);
455 * Disable all the counters:
457 perf_flags = hw_perf_save_disable();
459 list_for_each_entry(counter, &ctx->counter_list, list_entry)
460 counter->state = PERF_COUNTER_STATE_OFF;
462 hw_perf_restore(perf_flags);
464 spin_unlock(&ctx->lock);
471 int perf_counter_task_enable(void)
473 struct task_struct *curr = current;
474 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
475 struct perf_counter *counter;
479 if (likely(!ctx->nr_counters))
483 cpu = smp_processor_id();
485 spin_lock(&ctx->lock);
488 * Disable all the counters:
490 perf_flags = hw_perf_save_disable();
492 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
493 if (counter->state != PERF_COUNTER_STATE_OFF)
495 counter->state = PERF_COUNTER_STATE_INACTIVE;
497 hw_perf_restore(perf_flags);
499 spin_unlock(&ctx->lock);
501 perf_counter_task_sched_in(curr, cpu);
508 void perf_counter_task_tick(struct task_struct *curr, int cpu)
510 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
511 struct perf_counter *counter;
514 if (likely(!ctx->nr_counters))
517 perf_counter_task_sched_out(curr, cpu);
519 spin_lock(&ctx->lock);
522 * Rotate the first entry last (works just fine for group counters too):
524 perf_flags = hw_perf_save_disable();
525 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
526 list_del(&counter->list_entry);
527 list_add_tail(&counter->list_entry, &ctx->counter_list);
530 hw_perf_restore(perf_flags);
532 spin_unlock(&ctx->lock);
534 perf_counter_task_sched_in(curr, cpu);
538 * Cross CPU call to read the hardware counter
540 static void __read(void *info)
542 struct perf_counter *counter = info;
544 counter->hw_ops->read(counter);
547 static u64 perf_counter_read(struct perf_counter *counter)
550 * If counter is enabled and currently active on a CPU, update the
551 * value in the counter structure:
553 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
554 smp_call_function_single(counter->oncpu,
558 return atomic64_read(&counter->count);
562 * Cross CPU call to switch performance data pointers
564 static void __perf_switch_irq_data(void *info)
566 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
567 struct perf_counter *counter = info;
568 struct perf_counter_context *ctx = counter->ctx;
569 struct perf_data *oldirqdata = counter->irqdata;
572 * If this is a task context, we need to check whether it is
573 * the current task context of this cpu. If not it has been
574 * scheduled out before the smp call arrived.
577 if (cpuctx->task_ctx != ctx)
579 spin_lock(&ctx->lock);
582 /* Change the pointer NMI safe */
583 atomic_long_set((atomic_long_t *)&counter->irqdata,
584 (unsigned long) counter->usrdata);
585 counter->usrdata = oldirqdata;
588 spin_unlock(&ctx->lock);
591 static struct perf_data *perf_switch_irq_data(struct perf_counter *counter)
593 struct perf_counter_context *ctx = counter->ctx;
594 struct perf_data *oldirqdata = counter->irqdata;
595 struct task_struct *task = ctx->task;
598 smp_call_function_single(counter->cpu,
599 __perf_switch_irq_data,
601 return counter->usrdata;
605 spin_lock_irq(&ctx->lock);
606 if (counter->state != PERF_COUNTER_STATE_ACTIVE) {
607 counter->irqdata = counter->usrdata;
608 counter->usrdata = oldirqdata;
609 spin_unlock_irq(&ctx->lock);
612 spin_unlock_irq(&ctx->lock);
613 task_oncpu_function_call(task, __perf_switch_irq_data, counter);
614 /* Might have failed, because task was scheduled out */
615 if (counter->irqdata == oldirqdata)
618 return counter->usrdata;
621 static void put_context(struct perf_counter_context *ctx)
624 put_task_struct(ctx->task);
627 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
629 struct perf_cpu_context *cpuctx;
630 struct perf_counter_context *ctx;
631 struct task_struct *task;
634 * If cpu is not a wildcard then this is a percpu counter:
637 /* Must be root to operate on a CPU counter: */
638 if (!capable(CAP_SYS_ADMIN))
639 return ERR_PTR(-EACCES);
641 if (cpu < 0 || cpu > num_possible_cpus())
642 return ERR_PTR(-EINVAL);
645 * We could be clever and allow to attach a counter to an
646 * offline CPU and activate it when the CPU comes up, but
649 if (!cpu_isset(cpu, cpu_online_map))
650 return ERR_PTR(-ENODEV);
652 cpuctx = &per_cpu(perf_cpu_context, cpu);
662 task = find_task_by_vpid(pid);
664 get_task_struct(task);
668 return ERR_PTR(-ESRCH);
670 ctx = &task->perf_counter_ctx;
673 /* Reuse ptrace permission checks for now. */
674 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
676 return ERR_PTR(-EACCES);
683 * Called when the last reference to the file is gone.
685 static int perf_release(struct inode *inode, struct file *file)
687 struct perf_counter *counter = file->private_data;
688 struct perf_counter_context *ctx = counter->ctx;
690 file->private_data = NULL;
692 mutex_lock(&counter->mutex);
694 perf_counter_remove_from_context(counter);
697 mutex_unlock(&counter->mutex);
705 * Read the performance counter - simple non blocking version for now
708 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
712 if (count != sizeof(cntval))
715 mutex_lock(&counter->mutex);
716 cntval = perf_counter_read(counter);
717 mutex_unlock(&counter->mutex);
719 return put_user(cntval, (u64 __user *) buf) ? -EFAULT : sizeof(cntval);
723 perf_copy_usrdata(struct perf_data *usrdata, char __user *buf, size_t count)
728 count = min(count, (size_t)usrdata->len);
729 if (copy_to_user(buf, usrdata->data + usrdata->rd_idx, count))
732 /* Adjust the counters */
733 usrdata->len -= count;
737 usrdata->rd_idx += count;
743 perf_read_irq_data(struct perf_counter *counter,
748 struct perf_data *irqdata, *usrdata;
749 DECLARE_WAITQUEUE(wait, current);
752 irqdata = counter->irqdata;
753 usrdata = counter->usrdata;
755 if (usrdata->len + irqdata->len >= count)
761 spin_lock_irq(&counter->waitq.lock);
762 __add_wait_queue(&counter->waitq, &wait);
764 set_current_state(TASK_INTERRUPTIBLE);
765 if (usrdata->len + irqdata->len >= count)
768 if (signal_pending(current))
771 spin_unlock_irq(&counter->waitq.lock);
773 spin_lock_irq(&counter->waitq.lock);
775 __remove_wait_queue(&counter->waitq, &wait);
776 __set_current_state(TASK_RUNNING);
777 spin_unlock_irq(&counter->waitq.lock);
779 if (usrdata->len + irqdata->len < count)
782 mutex_lock(&counter->mutex);
784 /* Drain pending data first: */
785 res = perf_copy_usrdata(usrdata, buf, count);
786 if (res < 0 || res == count)
789 /* Switch irq buffer: */
790 usrdata = perf_switch_irq_data(counter);
791 if (perf_copy_usrdata(usrdata, buf + res, count - res) < 0) {
798 mutex_unlock(&counter->mutex);
804 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
806 struct perf_counter *counter = file->private_data;
808 switch (counter->hw_event.record_type) {
809 case PERF_RECORD_SIMPLE:
810 return perf_read_hw(counter, buf, count);
812 case PERF_RECORD_IRQ:
813 case PERF_RECORD_GROUP:
814 return perf_read_irq_data(counter, buf, count,
815 file->f_flags & O_NONBLOCK);
820 static unsigned int perf_poll(struct file *file, poll_table *wait)
822 struct perf_counter *counter = file->private_data;
823 unsigned int events = 0;
826 poll_wait(file, &counter->waitq, wait);
828 spin_lock_irqsave(&counter->waitq.lock, flags);
829 if (counter->usrdata->len || counter->irqdata->len)
831 spin_unlock_irqrestore(&counter->waitq.lock, flags);
836 static const struct file_operations perf_fops = {
837 .release = perf_release,
842 static void cpu_clock_perf_counter_enable(struct perf_counter *counter)
846 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
850 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
852 int cpu = raw_smp_processor_id();
854 atomic64_set(&counter->count, cpu_clock(cpu));
857 static const struct hw_perf_counter_ops perf_ops_cpu_clock = {
858 .enable = cpu_clock_perf_counter_enable,
859 .disable = cpu_clock_perf_counter_disable,
860 .read = cpu_clock_perf_counter_read,
863 static void task_clock_perf_counter_update(struct perf_counter *counter)
868 prev = atomic64_read(&counter->hw.prev_count);
869 now = current->se.sum_exec_runtime;
871 atomic64_set(&counter->hw.prev_count, now);
875 atomic64_add(delta, &counter->count);
878 static void task_clock_perf_counter_read(struct perf_counter *counter)
880 task_clock_perf_counter_update(counter);
883 static void task_clock_perf_counter_enable(struct perf_counter *counter)
885 atomic64_set(&counter->hw.prev_count, current->se.sum_exec_runtime);
888 static void task_clock_perf_counter_disable(struct perf_counter *counter)
890 task_clock_perf_counter_update(counter);
893 static const struct hw_perf_counter_ops perf_ops_task_clock = {
894 .enable = task_clock_perf_counter_enable,
895 .disable = task_clock_perf_counter_disable,
896 .read = task_clock_perf_counter_read,
899 static u64 get_page_faults(void)
901 struct task_struct *curr = current;
903 return curr->maj_flt + curr->min_flt;
906 static void page_faults_perf_counter_update(struct perf_counter *counter)
911 prev = atomic64_read(&counter->hw.prev_count);
912 now = get_page_faults();
914 atomic64_set(&counter->hw.prev_count, now);
918 atomic64_add(delta, &counter->count);
921 static void page_faults_perf_counter_read(struct perf_counter *counter)
923 page_faults_perf_counter_update(counter);
926 static void page_faults_perf_counter_enable(struct perf_counter *counter)
929 * page-faults is a per-task value already,
930 * so we dont have to clear it on switch-in.
934 static void page_faults_perf_counter_disable(struct perf_counter *counter)
936 page_faults_perf_counter_update(counter);
939 static const struct hw_perf_counter_ops perf_ops_page_faults = {
940 .enable = page_faults_perf_counter_enable,
941 .disable = page_faults_perf_counter_disable,
942 .read = page_faults_perf_counter_read,
945 static u64 get_context_switches(void)
947 struct task_struct *curr = current;
949 return curr->nvcsw + curr->nivcsw;
952 static void context_switches_perf_counter_update(struct perf_counter *counter)
957 prev = atomic64_read(&counter->hw.prev_count);
958 now = get_context_switches();
960 atomic64_set(&counter->hw.prev_count, now);
964 atomic64_add(delta, &counter->count);
967 static void context_switches_perf_counter_read(struct perf_counter *counter)
969 context_switches_perf_counter_update(counter);
972 static void context_switches_perf_counter_enable(struct perf_counter *counter)
975 * ->nvcsw + curr->nivcsw is a per-task value already,
976 * so we dont have to clear it on switch-in.
980 static void context_switches_perf_counter_disable(struct perf_counter *counter)
982 context_switches_perf_counter_update(counter);
985 static const struct hw_perf_counter_ops perf_ops_context_switches = {
986 .enable = context_switches_perf_counter_enable,
987 .disable = context_switches_perf_counter_disable,
988 .read = context_switches_perf_counter_read,
991 static inline u64 get_cpu_migrations(void)
993 return current->se.nr_migrations;
996 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
1001 prev = atomic64_read(&counter->hw.prev_count);
1002 now = get_cpu_migrations();
1004 atomic64_set(&counter->hw.prev_count, now);
1008 atomic64_add(delta, &counter->count);
1011 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
1013 cpu_migrations_perf_counter_update(counter);
1016 static void cpu_migrations_perf_counter_enable(struct perf_counter *counter)
1019 * se.nr_migrations is a per-task value already,
1020 * so we dont have to clear it on switch-in.
1024 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
1026 cpu_migrations_perf_counter_update(counter);
1029 static const struct hw_perf_counter_ops perf_ops_cpu_migrations = {
1030 .enable = cpu_migrations_perf_counter_enable,
1031 .disable = cpu_migrations_perf_counter_disable,
1032 .read = cpu_migrations_perf_counter_read,
1035 static const struct hw_perf_counter_ops *
1036 sw_perf_counter_init(struct perf_counter *counter)
1038 const struct hw_perf_counter_ops *hw_ops = NULL;
1040 switch (counter->hw_event.type) {
1041 case PERF_COUNT_CPU_CLOCK:
1042 hw_ops = &perf_ops_cpu_clock;
1044 case PERF_COUNT_TASK_CLOCK:
1045 hw_ops = &perf_ops_task_clock;
1047 case PERF_COUNT_PAGE_FAULTS:
1048 hw_ops = &perf_ops_page_faults;
1050 case PERF_COUNT_CONTEXT_SWITCHES:
1051 hw_ops = &perf_ops_context_switches;
1053 case PERF_COUNT_CPU_MIGRATIONS:
1054 hw_ops = &perf_ops_cpu_migrations;
1063 * Allocate and initialize a counter structure
1065 static struct perf_counter *
1066 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
1068 struct perf_counter *group_leader,
1071 const struct hw_perf_counter_ops *hw_ops;
1072 struct perf_counter *counter;
1074 counter = kzalloc(sizeof(*counter), gfpflags);
1079 * Single counters are their own group leaders, with an
1080 * empty sibling list:
1083 group_leader = counter;
1085 mutex_init(&counter->mutex);
1086 INIT_LIST_HEAD(&counter->list_entry);
1087 INIT_LIST_HEAD(&counter->sibling_list);
1088 init_waitqueue_head(&counter->waitq);
1090 counter->irqdata = &counter->data[0];
1091 counter->usrdata = &counter->data[1];
1093 counter->hw_event = *hw_event;
1094 counter->wakeup_pending = 0;
1095 counter->group_leader = group_leader;
1096 counter->hw_ops = NULL;
1098 if (hw_event->disabled)
1099 counter->state = PERF_COUNTER_STATE_OFF;
1102 if (!hw_event->raw && hw_event->type < 0)
1103 hw_ops = sw_perf_counter_init(counter);
1105 hw_ops = hw_perf_counter_init(counter);
1111 counter->hw_ops = hw_ops;
1117 * sys_perf_task_open - open a performance counter, associate it to a task/cpu
1119 * @hw_event_uptr: event type attributes for monitoring/sampling
1122 * @group_fd: group leader counter fd
1125 sys_perf_counter_open(struct perf_counter_hw_event *hw_event_uptr __user,
1126 pid_t pid, int cpu, int group_fd)
1128 struct perf_counter *counter, *group_leader;
1129 struct perf_counter_hw_event hw_event;
1130 struct perf_counter_context *ctx;
1131 struct file *counter_file = NULL;
1132 struct file *group_file = NULL;
1133 int fput_needed = 0;
1134 int fput_needed2 = 0;
1137 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
1141 * Get the target context (task or percpu):
1143 ctx = find_get_context(pid, cpu);
1145 return PTR_ERR(ctx);
1148 * Look up the group leader (we will attach this counter to it):
1150 group_leader = NULL;
1151 if (group_fd != -1) {
1153 group_file = fget_light(group_fd, &fput_needed);
1155 goto err_put_context;
1156 if (group_file->f_op != &perf_fops)
1157 goto err_put_context;
1159 group_leader = group_file->private_data;
1161 * Do not allow a recursive hierarchy (this new sibling
1162 * becoming part of another group-sibling):
1164 if (group_leader->group_leader != group_leader)
1165 goto err_put_context;
1167 * Do not allow to attach to a group in a different
1168 * task or CPU context:
1170 if (group_leader->ctx != ctx)
1171 goto err_put_context;
1175 counter = perf_counter_alloc(&hw_event, cpu, group_leader, GFP_KERNEL);
1177 goto err_put_context;
1179 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
1181 goto err_free_put_context;
1183 counter_file = fget_light(ret, &fput_needed2);
1185 goto err_free_put_context;
1187 counter->filp = counter_file;
1188 perf_install_in_context(ctx, counter, cpu);
1190 fput_light(counter_file, fput_needed2);
1193 fput_light(group_file, fput_needed);
1197 err_free_put_context:
1207 * Initialize the perf_counter context in a task_struct:
1210 __perf_counter_init_context(struct perf_counter_context *ctx,
1211 struct task_struct *task)
1213 memset(ctx, 0, sizeof(*ctx));
1214 spin_lock_init(&ctx->lock);
1215 INIT_LIST_HEAD(&ctx->counter_list);
1220 * inherit a counter from parent task to child task:
1223 inherit_counter(struct perf_counter *parent_counter,
1224 struct task_struct *parent,
1225 struct perf_counter_context *parent_ctx,
1226 struct task_struct *child,
1227 struct perf_counter_context *child_ctx)
1229 struct perf_counter *child_counter;
1231 child_counter = perf_counter_alloc(&parent_counter->hw_event,
1232 parent_counter->cpu, NULL,
1238 * Link it up in the child's context:
1240 child_counter->ctx = child_ctx;
1241 child_counter->task = child;
1242 list_add_counter(child_counter, child_ctx);
1243 child_ctx->nr_counters++;
1245 child_counter->parent = parent_counter;
1246 parent_counter->nr_inherited++;
1248 * inherit into child's child as well:
1250 child_counter->hw_event.inherit = 1;
1253 * Get a reference to the parent filp - we will fput it
1254 * when the child counter exits. This is safe to do because
1255 * we are in the parent and we know that the filp still
1256 * exists and has a nonzero count:
1258 atomic_long_inc(&parent_counter->filp->f_count);
1264 __perf_counter_exit_task(struct task_struct *child,
1265 struct perf_counter *child_counter,
1266 struct perf_counter_context *child_ctx)
1268 struct perf_counter *parent_counter;
1269 u64 parent_val, child_val;
1273 * Disable and unlink this counter.
1275 * Be careful about zapping the list - IRQ/NMI context
1276 * could still be processing it:
1278 local_irq_disable();
1279 perf_flags = hw_perf_save_disable();
1281 if (child_counter->state == PERF_COUNTER_STATE_ACTIVE) {
1282 struct perf_cpu_context *cpuctx;
1284 cpuctx = &__get_cpu_var(perf_cpu_context);
1286 child_counter->hw_ops->disable(child_counter);
1287 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
1288 child_counter->oncpu = -1;
1290 cpuctx->active_oncpu--;
1291 child_ctx->nr_active--;
1294 list_del_init(&child_counter->list_entry);
1296 hw_perf_restore(perf_flags);
1299 parent_counter = child_counter->parent;
1301 * It can happen that parent exits first, and has counters
1302 * that are still around due to the child reference. These
1303 * counters need to be zapped - but otherwise linger.
1305 if (!parent_counter)
1308 parent_val = atomic64_read(&parent_counter->count);
1309 child_val = atomic64_read(&child_counter->count);
1312 * Add back the child's count to the parent's count:
1314 atomic64_add(child_val, &parent_counter->count);
1316 fput(parent_counter->filp);
1318 kfree(child_counter);
1322 * When a child task exist, feed back counter values to parent counters.
1324 * Note: we are running in child context, but the PID is not hashed
1325 * anymore so new counters will not be added.
1327 void perf_counter_exit_task(struct task_struct *child)
1329 struct perf_counter *child_counter, *tmp;
1330 struct perf_counter_context *child_ctx;
1332 child_ctx = &child->perf_counter_ctx;
1334 if (likely(!child_ctx->nr_counters))
1337 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
1339 __perf_counter_exit_task(child, child_counter, child_ctx);
1343 * Initialize the perf_counter context in task_struct
1345 void perf_counter_init_task(struct task_struct *child)
1347 struct perf_counter_context *child_ctx, *parent_ctx;
1348 struct perf_counter *counter, *parent_counter;
1349 struct task_struct *parent = current;
1350 unsigned long flags;
1352 child_ctx = &child->perf_counter_ctx;
1353 parent_ctx = &parent->perf_counter_ctx;
1355 __perf_counter_init_context(child_ctx, child);
1358 * This is executed from the parent task context, so inherit
1359 * counters that have been marked for cloning:
1362 if (likely(!parent_ctx->nr_counters))
1366 * Lock the parent list. No need to lock the child - not PID
1367 * hashed yet and not running, so nobody can access it.
1369 spin_lock_irqsave(&parent_ctx->lock, flags);
1372 * We dont have to disable NMIs - we are only looking at
1373 * the list, not manipulating it:
1375 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
1376 if (!counter->hw_event.inherit || counter->group_leader != counter)
1380 * Instead of creating recursive hierarchies of counters,
1381 * we link inheritd counters back to the original parent,
1382 * which has a filp for sure, which we use as the reference
1385 parent_counter = counter;
1386 if (counter->parent)
1387 parent_counter = counter->parent;
1389 if (inherit_counter(parent_counter, parent,
1390 parent_ctx, child, child_ctx))
1394 spin_unlock_irqrestore(&parent_ctx->lock, flags);
1397 static void __cpuinit perf_counter_init_cpu(int cpu)
1399 struct perf_cpu_context *cpuctx;
1401 cpuctx = &per_cpu(perf_cpu_context, cpu);
1402 __perf_counter_init_context(&cpuctx->ctx, NULL);
1404 mutex_lock(&perf_resource_mutex);
1405 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
1406 mutex_unlock(&perf_resource_mutex);
1408 hw_perf_counter_setup();
1411 #ifdef CONFIG_HOTPLUG_CPU
1412 static void __perf_counter_exit_cpu(void *info)
1414 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1415 struct perf_counter_context *ctx = &cpuctx->ctx;
1416 struct perf_counter *counter, *tmp;
1418 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
1419 __perf_counter_remove_from_context(counter);
1422 static void perf_counter_exit_cpu(int cpu)
1424 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
1427 static inline void perf_counter_exit_cpu(int cpu) { }
1430 static int __cpuinit
1431 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
1433 unsigned int cpu = (long)hcpu;
1437 case CPU_UP_PREPARE:
1438 case CPU_UP_PREPARE_FROZEN:
1439 perf_counter_init_cpu(cpu);
1442 case CPU_DOWN_PREPARE:
1443 case CPU_DOWN_PREPARE_FROZEN:
1444 perf_counter_exit_cpu(cpu);
1454 static struct notifier_block __cpuinitdata perf_cpu_nb = {
1455 .notifier_call = perf_cpu_notify,
1458 static int __init perf_counter_init(void)
1460 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
1461 (void *)(long)smp_processor_id());
1462 register_cpu_notifier(&perf_cpu_nb);
1466 early_initcall(perf_counter_init);
1468 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
1470 return sprintf(buf, "%d\n", perf_reserved_percpu);
1474 perf_set_reserve_percpu(struct sysdev_class *class,
1478 struct perf_cpu_context *cpuctx;
1482 err = strict_strtoul(buf, 10, &val);
1485 if (val > perf_max_counters)
1488 mutex_lock(&perf_resource_mutex);
1489 perf_reserved_percpu = val;
1490 for_each_online_cpu(cpu) {
1491 cpuctx = &per_cpu(perf_cpu_context, cpu);
1492 spin_lock_irq(&cpuctx->ctx.lock);
1493 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
1494 perf_max_counters - perf_reserved_percpu);
1495 cpuctx->max_pertask = mpt;
1496 spin_unlock_irq(&cpuctx->ctx.lock);
1498 mutex_unlock(&perf_resource_mutex);
1503 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
1505 return sprintf(buf, "%d\n", perf_overcommit);
1509 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
1514 err = strict_strtoul(buf, 10, &val);
1520 mutex_lock(&perf_resource_mutex);
1521 perf_overcommit = val;
1522 mutex_unlock(&perf_resource_mutex);
1527 static SYSDEV_CLASS_ATTR(
1530 perf_show_reserve_percpu,
1531 perf_set_reserve_percpu
1534 static SYSDEV_CLASS_ATTR(
1537 perf_show_overcommit,
1541 static struct attribute *perfclass_attrs[] = {
1542 &attr_reserve_percpu.attr,
1543 &attr_overcommit.attr,
1547 static struct attribute_group perfclass_attr_group = {
1548 .attrs = perfclass_attrs,
1549 .name = "perf_counters",
1552 static int __init perf_counter_sysfs_init(void)
1554 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
1555 &perfclass_attr_group);
1557 device_initcall(perf_counter_sysfs_init);