aboutsummaryrefslogtreecommitdiffstats
path: root/kernel
diff options
context:
space:
mode:
Diffstat (limited to 'kernel')
-rw-r--r--kernel/Makefile1
-rw-r--r--kernel/exit.c13
-rw-r--r--kernel/fork.c1
-rw-r--r--kernel/perf_counter.c2212
-rw-r--r--kernel/sched.c93
-rw-r--r--kernel/sys.c7
-rw-r--r--kernel/sys_ni.c3
7 files changed, 2323 insertions, 7 deletions
diff --git a/kernel/Makefile b/kernel/Makefile
index bab1dffe37e9..63c697529ca1 100644
--- a/kernel/Makefile
+++ b/kernel/Makefile
@@ -94,6 +94,7 @@ obj-$(CONFIG_FUNCTION_TRACER) += trace/
94obj-$(CONFIG_TRACING) += trace/ 94obj-$(CONFIG_TRACING) += trace/
95obj-$(CONFIG_SMP) += sched_cpupri.o 95obj-$(CONFIG_SMP) += sched_cpupri.o
96obj-$(CONFIG_SLOW_WORK) += slow-work.o 96obj-$(CONFIG_SLOW_WORK) += slow-work.o
97obj-$(CONFIG_PERF_COUNTERS) += perf_counter.o
97 98
98ifneq ($(CONFIG_SCHED_OMIT_FRAME_POINTER),y) 99ifneq ($(CONFIG_SCHED_OMIT_FRAME_POINTER),y)
99# According to Alan Modra <alan@linuxcare.com.au>, the -fno-omit-frame-pointer is 100# According to Alan Modra <alan@linuxcare.com.au>, the -fno-omit-frame-pointer is
diff --git a/kernel/exit.c b/kernel/exit.c
index 6686ed1e4aa3..7a14a2b504f5 100644
--- a/kernel/exit.c
+++ b/kernel/exit.c
@@ -158,6 +158,9 @@ static void delayed_put_task_struct(struct rcu_head *rhp)
158{ 158{
159 struct task_struct *tsk = container_of(rhp, struct task_struct, rcu); 159 struct task_struct *tsk = container_of(rhp, struct task_struct, rcu);
160 160
161#ifdef CONFIG_PERF_COUNTERS
162 WARN_ON_ONCE(!list_empty(&tsk->perf_counter_ctx.counter_list));
163#endif
161 trace_sched_process_free(tsk); 164 trace_sched_process_free(tsk);
162 put_task_struct(tsk); 165 put_task_struct(tsk);
163} 166}
@@ -980,10 +983,6 @@ NORET_TYPE void do_exit(long code)
980 tsk->mempolicy = NULL; 983 tsk->mempolicy = NULL;
981#endif 984#endif
982#ifdef CONFIG_FUTEX 985#ifdef CONFIG_FUTEX
983 /*
984 * This must happen late, after the PID is not
985 * hashed anymore:
986 */
987 if (unlikely(!list_empty(&tsk->pi_state_list))) 986 if (unlikely(!list_empty(&tsk->pi_state_list)))
988 exit_pi_state_list(tsk); 987 exit_pi_state_list(tsk);
989 if (unlikely(current->pi_state_cache)) 988 if (unlikely(current->pi_state_cache))
@@ -1250,6 +1249,12 @@ static int wait_task_zombie(struct task_struct *p, int options,
1250 */ 1249 */
1251 read_unlock(&tasklist_lock); 1250 read_unlock(&tasklist_lock);
1252 1251
1252 /*
1253 * Flush inherited counters to the parent - before the parent
1254 * gets woken up by child-exit notifications.
1255 */
1256 perf_counter_exit_task(p);
1257
1253 retval = ru ? getrusage(p, RUSAGE_BOTH, ru) : 0; 1258 retval = ru ? getrusage(p, RUSAGE_BOTH, ru) : 0;
1254 status = (p->signal->flags & SIGNAL_GROUP_EXIT) 1259 status = (p->signal->flags & SIGNAL_GROUP_EXIT)
1255 ? p->signal->group_exit_code : p->exit_code; 1260 ? p->signal->group_exit_code : p->exit_code;
diff --git a/kernel/fork.c b/kernel/fork.c
index 660c2b8765bc..381d7f9b70fb 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -975,6 +975,7 @@ static struct task_struct *copy_process(unsigned long clone_flags,
975 goto fork_out; 975 goto fork_out;
976 976
977 rt_mutex_init_task(p); 977 rt_mutex_init_task(p);
978 perf_counter_init_task(p);
978 979
979#ifdef CONFIG_PROVE_LOCKING 980#ifdef CONFIG_PROVE_LOCKING
980 DEBUG_LOCKS_WARN_ON(!p->hardirqs_enabled); 981 DEBUG_LOCKS_WARN_ON(!p->hardirqs_enabled);
diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
new file mode 100644
index 000000000000..b2e838959f3e
--- /dev/null
+++ b/kernel/perf_counter.c
@@ -0,0 +1,2212 @@
1/*
2 * Performance counter core code
3 *
4 * Copyright(C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2008 Red Hat, Inc., Ingo Molnar
6 *
7 * For licencing details see kernel-base/COPYING
8 */
9
10#include <linux/fs.h>
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/kernel_stat.h>
22#include <linux/perf_counter.h>
23#include <linux/mm.h>
24#include <linux/vmstat.h>
25
26/*
27 * Each CPU has a list of per CPU counters:
28 */
29DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
30
31int perf_max_counters __read_mostly = 1;
32static int perf_reserved_percpu __read_mostly;
33static int perf_overcommit __read_mostly = 1;
34
35/*
36 * Mutex for (sysadmin-configurable) counter reservations:
37 */
38static DEFINE_MUTEX(perf_resource_mutex);
39
40/*
41 * Architecture provided APIs - weak aliases:
42 */
43extern __weak const struct hw_perf_counter_ops *
44hw_perf_counter_init(struct perf_counter *counter)
45{
46 return NULL;
47}
48
49u64 __weak hw_perf_save_disable(void) { return 0; }
50void __weak hw_perf_restore(u64 ctrl) { barrier(); }
51void __weak hw_perf_counter_setup(int cpu) { barrier(); }
52int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
53 struct perf_cpu_context *cpuctx,
54 struct perf_counter_context *ctx, int cpu)
55{
56 return 0;
57}
58
59void __weak perf_counter_print_debug(void) { }
60
61static void
62list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
63{
64 struct perf_counter *group_leader = counter->group_leader;
65
66 /*
67 * Depending on whether it is a standalone or sibling counter,
68 * add it straight to the context's counter list, or to the group
69 * leader's sibling list:
70 */
71 if (counter->group_leader == counter)
72 list_add_tail(&counter->list_entry, &ctx->counter_list);
73 else
74 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
75}
76
77static void
78list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
79{
80 struct perf_counter *sibling, *tmp;
81
82 list_del_init(&counter->list_entry);
83
84 /*
85 * If this was a group counter with sibling counters then
86 * upgrade the siblings to singleton counters by adding them
87 * to the context list directly:
88 */
89 list_for_each_entry_safe(sibling, tmp,
90 &counter->sibling_list, list_entry) {
91
92 list_del_init(&sibling->list_entry);
93 list_add_tail(&sibling->list_entry, &ctx->counter_list);
94 sibling->group_leader = sibling;
95 }
96}
97
98static void
99counter_sched_out(struct perf_counter *counter,
100 struct perf_cpu_context *cpuctx,
101 struct perf_counter_context *ctx)
102{
103 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
104 return;
105
106 counter->state = PERF_COUNTER_STATE_INACTIVE;
107 counter->hw_ops->disable(counter);
108 counter->oncpu = -1;
109
110 if (!is_software_counter(counter))
111 cpuctx->active_oncpu--;
112 ctx->nr_active--;
113 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
114 cpuctx->exclusive = 0;
115}
116
117static void
118group_sched_out(struct perf_counter *group_counter,
119 struct perf_cpu_context *cpuctx,
120 struct perf_counter_context *ctx)
121{
122 struct perf_counter *counter;
123
124 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
125 return;
126
127 counter_sched_out(group_counter, cpuctx, ctx);
128
129 /*
130 * Schedule out siblings (if any):
131 */
132 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
133 counter_sched_out(counter, cpuctx, ctx);
134
135 if (group_counter->hw_event.exclusive)
136 cpuctx->exclusive = 0;
137}
138
139/*
140 * Cross CPU call to remove a performance counter
141 *
142 * We disable the counter on the hardware level first. After that we
143 * remove it from the context list.
144 */
145static void __perf_counter_remove_from_context(void *info)
146{
147 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
148 struct perf_counter *counter = info;
149 struct perf_counter_context *ctx = counter->ctx;
150 unsigned long flags;
151 u64 perf_flags;
152
153 /*
154 * If this is a task context, we need to check whether it is
155 * the current task context of this cpu. If not it has been
156 * scheduled out before the smp call arrived.
157 */
158 if (ctx->task && cpuctx->task_ctx != ctx)
159 return;
160
161 curr_rq_lock_irq_save(&flags);
162 spin_lock(&ctx->lock);
163
164 counter_sched_out(counter, cpuctx, ctx);
165
166 counter->task = NULL;
167 ctx->nr_counters--;
168
169 /*
170 * Protect the list operation against NMI by disabling the
171 * counters on a global level. NOP for non NMI based counters.
172 */
173 perf_flags = hw_perf_save_disable();
174 list_del_counter(counter, ctx);
175 hw_perf_restore(perf_flags);
176
177 if (!ctx->task) {
178 /*
179 * Allow more per task counters with respect to the
180 * reservation:
181 */
182 cpuctx->max_pertask =
183 min(perf_max_counters - ctx->nr_counters,
184 perf_max_counters - perf_reserved_percpu);
185 }
186
187 spin_unlock(&ctx->lock);
188 curr_rq_unlock_irq_restore(&flags);
189}
190
191
192/*
193 * Remove the counter from a task's (or a CPU's) list of counters.
194 *
195 * Must be called with counter->mutex and ctx->mutex held.
196 *
197 * CPU counters are removed with a smp call. For task counters we only
198 * call when the task is on a CPU.
199 */
200static void perf_counter_remove_from_context(struct perf_counter *counter)
201{
202 struct perf_counter_context *ctx = counter->ctx;
203 struct task_struct *task = ctx->task;
204
205 if (!task) {
206 /*
207 * Per cpu counters are removed via an smp call and
208 * the removal is always sucessful.
209 */
210 smp_call_function_single(counter->cpu,
211 __perf_counter_remove_from_context,
212 counter, 1);
213 return;
214 }
215
216retry:
217 task_oncpu_function_call(task, __perf_counter_remove_from_context,
218 counter);
219
220 spin_lock_irq(&ctx->lock);
221 /*
222 * If the context is active we need to retry the smp call.
223 */
224 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
225 spin_unlock_irq(&ctx->lock);
226 goto retry;
227 }
228
229 /*
230 * The lock prevents that this context is scheduled in so we
231 * can remove the counter safely, if the call above did not
232 * succeed.
233 */
234 if (!list_empty(&counter->list_entry)) {
235 ctx->nr_counters--;
236 list_del_counter(counter, ctx);
237 counter->task = NULL;
238 }
239 spin_unlock_irq(&ctx->lock);
240}
241
242/*
243 * Cross CPU call to disable a performance counter
244 */
245static void __perf_counter_disable(void *info)
246{
247 struct perf_counter *counter = info;
248 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
249 struct perf_counter_context *ctx = counter->ctx;
250 unsigned long flags;
251
252 /*
253 * If this is a per-task counter, need to check whether this
254 * counter's task is the current task on this cpu.
255 */
256 if (ctx->task && cpuctx->task_ctx != ctx)
257 return;
258
259 curr_rq_lock_irq_save(&flags);
260 spin_lock(&ctx->lock);
261
262 /*
263 * If the counter is on, turn it off.
264 * If it is in error state, leave it in error state.
265 */
266 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
267 if (counter == counter->group_leader)
268 group_sched_out(counter, cpuctx, ctx);
269 else
270 counter_sched_out(counter, cpuctx, ctx);
271 counter->state = PERF_COUNTER_STATE_OFF;
272 }
273
274 spin_unlock(&ctx->lock);
275 curr_rq_unlock_irq_restore(&flags);
276}
277
278/*
279 * Disable a counter.
280 */
281static void perf_counter_disable(struct perf_counter *counter)
282{
283 struct perf_counter_context *ctx = counter->ctx;
284 struct task_struct *task = ctx->task;
285
286 if (!task) {
287 /*
288 * Disable the counter on the cpu that it's on
289 */
290 smp_call_function_single(counter->cpu, __perf_counter_disable,
291 counter, 1);
292 return;
293 }
294
295 retry:
296 task_oncpu_function_call(task, __perf_counter_disable, counter);
297
298 spin_lock_irq(&ctx->lock);
299 /*
300 * If the counter is still active, we need to retry the cross-call.
301 */
302 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
303 spin_unlock_irq(&ctx->lock);
304 goto retry;
305 }
306
307 /*
308 * Since we have the lock this context can't be scheduled
309 * in, so we can change the state safely.
310 */
311 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
312 counter->state = PERF_COUNTER_STATE_OFF;
313
314 spin_unlock_irq(&ctx->lock);
315}
316
317/*
318 * Disable a counter and all its children.
319 */
320static void perf_counter_disable_family(struct perf_counter *counter)
321{
322 struct perf_counter *child;
323
324 perf_counter_disable(counter);
325
326 /*
327 * Lock the mutex to protect the list of children
328 */
329 mutex_lock(&counter->mutex);
330 list_for_each_entry(child, &counter->child_list, child_list)
331 perf_counter_disable(child);
332 mutex_unlock(&counter->mutex);
333}
334
335static int
336counter_sched_in(struct perf_counter *counter,
337 struct perf_cpu_context *cpuctx,
338 struct perf_counter_context *ctx,
339 int cpu)
340{
341 if (counter->state <= PERF_COUNTER_STATE_OFF)
342 return 0;
343
344 counter->state = PERF_COUNTER_STATE_ACTIVE;
345 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
346 /*
347 * The new state must be visible before we turn it on in the hardware:
348 */
349 smp_wmb();
350
351 if (counter->hw_ops->enable(counter)) {
352 counter->state = PERF_COUNTER_STATE_INACTIVE;
353 counter->oncpu = -1;
354 return -EAGAIN;
355 }
356
357 if (!is_software_counter(counter))
358 cpuctx->active_oncpu++;
359 ctx->nr_active++;
360
361 if (counter->hw_event.exclusive)
362 cpuctx->exclusive = 1;
363
364 return 0;
365}
366
367/*
368 * Return 1 for a group consisting entirely of software counters,
369 * 0 if the group contains any hardware counters.
370 */
371static int is_software_only_group(struct perf_counter *leader)
372{
373 struct perf_counter *counter;
374
375 if (!is_software_counter(leader))
376 return 0;
377 list_for_each_entry(counter, &leader->sibling_list, list_entry)
378 if (!is_software_counter(counter))
379 return 0;
380 return 1;
381}
382
383/*
384 * Work out whether we can put this counter group on the CPU now.
385 */
386static int group_can_go_on(struct perf_counter *counter,
387 struct perf_cpu_context *cpuctx,
388 int can_add_hw)
389{
390 /*
391 * Groups consisting entirely of software counters can always go on.
392 */
393 if (is_software_only_group(counter))
394 return 1;
395 /*
396 * If an exclusive group is already on, no other hardware
397 * counters can go on.
398 */
399 if (cpuctx->exclusive)
400 return 0;
401 /*
402 * If this group is exclusive and there are already
403 * counters on the CPU, it can't go on.
404 */
405 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
406 return 0;
407 /*
408 * Otherwise, try to add it if all previous groups were able
409 * to go on.
410 */
411 return can_add_hw;
412}
413
414/*
415 * Cross CPU call to install and enable a performance counter
416 */
417static void __perf_install_in_context(void *info)
418{
419 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
420 struct perf_counter *counter = info;
421 struct perf_counter_context *ctx = counter->ctx;
422 struct perf_counter *leader = counter->group_leader;
423 int cpu = smp_processor_id();
424 unsigned long flags;
425 u64 perf_flags;
426 int err;
427
428 /*
429 * If this is a task context, we need to check whether it is
430 * the current task context of this cpu. If not it has been
431 * scheduled out before the smp call arrived.
432 */
433 if (ctx->task && cpuctx->task_ctx != ctx)
434 return;
435
436 curr_rq_lock_irq_save(&flags);
437 spin_lock(&ctx->lock);
438
439 /*
440 * Protect the list operation against NMI by disabling the
441 * counters on a global level. NOP for non NMI based counters.
442 */
443 perf_flags = hw_perf_save_disable();
444
445 list_add_counter(counter, ctx);
446 ctx->nr_counters++;
447 counter->prev_state = PERF_COUNTER_STATE_OFF;
448
449 /*
450 * Don't put the counter on if it is disabled or if
451 * it is in a group and the group isn't on.
452 */
453 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
454 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
455 goto unlock;
456
457 /*
458 * An exclusive counter can't go on if there are already active
459 * hardware counters, and no hardware counter can go on if there
460 * is already an exclusive counter on.
461 */
462 if (!group_can_go_on(counter, cpuctx, 1))
463 err = -EEXIST;
464 else
465 err = counter_sched_in(counter, cpuctx, ctx, cpu);
466
467 if (err) {
468 /*
469 * This counter couldn't go on. If it is in a group
470 * then we have to pull the whole group off.
471 * If the counter group is pinned then put it in error state.
472 */
473 if (leader != counter)
474 group_sched_out(leader, cpuctx, ctx);
475 if (leader->hw_event.pinned)
476 leader->state = PERF_COUNTER_STATE_ERROR;
477 }
478
479 if (!err && !ctx->task && cpuctx->max_pertask)
480 cpuctx->max_pertask--;
481
482 unlock:
483 hw_perf_restore(perf_flags);
484
485 spin_unlock(&ctx->lock);
486 curr_rq_unlock_irq_restore(&flags);
487}
488
489/*
490 * Attach a performance counter to a context
491 *
492 * First we add the counter to the list with the hardware enable bit
493 * in counter->hw_config cleared.
494 *
495 * If the counter is attached to a task which is on a CPU we use a smp
496 * call to enable it in the task context. The task might have been
497 * scheduled away, but we check this in the smp call again.
498 *
499 * Must be called with ctx->mutex held.
500 */
501static void
502perf_install_in_context(struct perf_counter_context *ctx,
503 struct perf_counter *counter,
504 int cpu)
505{
506 struct task_struct *task = ctx->task;
507
508 if (!task) {
509 /*
510 * Per cpu counters are installed via an smp call and
511 * the install is always sucessful.
512 */
513 smp_call_function_single(cpu, __perf_install_in_context,
514 counter, 1);
515 return;
516 }
517
518 counter->task = task;
519retry:
520 task_oncpu_function_call(task, __perf_install_in_context,
521 counter);
522
523 spin_lock_irq(&ctx->lock);
524 /*
525 * we need to retry the smp call.
526 */
527 if (ctx->is_active && list_empty(&counter->list_entry)) {
528 spin_unlock_irq(&ctx->lock);
529 goto retry;
530 }
531
532 /*
533 * The lock prevents that this context is scheduled in so we
534 * can add the counter safely, if it the call above did not
535 * succeed.
536 */
537 if (list_empty(&counter->list_entry)) {
538 list_add_counter(counter, ctx);
539 ctx->nr_counters++;
540 }
541 spin_unlock_irq(&ctx->lock);
542}
543
544/*
545 * Cross CPU call to enable a performance counter
546 */
547static void __perf_counter_enable(void *info)
548{
549 struct perf_counter *counter = info;
550 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
551 struct perf_counter_context *ctx = counter->ctx;
552 struct perf_counter *leader = counter->group_leader;
553 unsigned long flags;
554 int err;
555
556 /*
557 * If this is a per-task counter, need to check whether this
558 * counter's task is the current task on this cpu.
559 */
560 if (ctx->task && cpuctx->task_ctx != ctx)
561 return;
562
563 curr_rq_lock_irq_save(&flags);
564 spin_lock(&ctx->lock);
565
566 counter->prev_state = counter->state;
567 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
568 goto unlock;
569 counter->state = PERF_COUNTER_STATE_INACTIVE;
570
571 /*
572 * If the counter is in a group and isn't the group leader,
573 * then don't put it on unless the group is on.
574 */
575 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
576 goto unlock;
577
578 if (!group_can_go_on(counter, cpuctx, 1))
579 err = -EEXIST;
580 else
581 err = counter_sched_in(counter, cpuctx, ctx,
582 smp_processor_id());
583
584 if (err) {
585 /*
586 * If this counter can't go on and it's part of a
587 * group, then the whole group has to come off.
588 */
589 if (leader != counter)
590 group_sched_out(leader, cpuctx, ctx);
591 if (leader->hw_event.pinned)
592 leader->state = PERF_COUNTER_STATE_ERROR;
593 }
594
595 unlock:
596 spin_unlock(&ctx->lock);
597 curr_rq_unlock_irq_restore(&flags);
598}
599
600/*
601 * Enable a counter.
602 */
603static void perf_counter_enable(struct perf_counter *counter)
604{
605 struct perf_counter_context *ctx = counter->ctx;
606 struct task_struct *task = ctx->task;
607
608 if (!task) {
609 /*
610 * Enable the counter on the cpu that it's on
611 */
612 smp_call_function_single(counter->cpu, __perf_counter_enable,
613 counter, 1);
614 return;
615 }
616
617 spin_lock_irq(&ctx->lock);
618 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
619 goto out;
620
621 /*
622 * If the counter is in error state, clear that first.
623 * That way, if we see the counter in error state below, we
624 * know that it has gone back into error state, as distinct
625 * from the task having been scheduled away before the
626 * cross-call arrived.
627 */
628 if (counter->state == PERF_COUNTER_STATE_ERROR)
629 counter->state = PERF_COUNTER_STATE_OFF;
630
631 retry:
632 spin_unlock_irq(&ctx->lock);
633 task_oncpu_function_call(task, __perf_counter_enable, counter);
634
635 spin_lock_irq(&ctx->lock);
636
637 /*
638 * If the context is active and the counter is still off,
639 * we need to retry the cross-call.
640 */
641 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
642 goto retry;
643
644 /*
645 * Since we have the lock this context can't be scheduled
646 * in, so we can change the state safely.
647 */
648 if (counter->state == PERF_COUNTER_STATE_OFF)
649 counter->state = PERF_COUNTER_STATE_INACTIVE;
650 out:
651 spin_unlock_irq(&ctx->lock);
652}
653
654/*
655 * Enable a counter and all its children.
656 */
657static void perf_counter_enable_family(struct perf_counter *counter)
658{
659 struct perf_counter *child;
660
661 perf_counter_enable(counter);
662
663 /*
664 * Lock the mutex to protect the list of children
665 */
666 mutex_lock(&counter->mutex);
667 list_for_each_entry(child, &counter->child_list, child_list)
668 perf_counter_enable(child);
669 mutex_unlock(&counter->mutex);
670}
671
672void __perf_counter_sched_out(struct perf_counter_context *ctx,
673 struct perf_cpu_context *cpuctx)
674{
675 struct perf_counter *counter;
676 u64 flags;
677
678 spin_lock(&ctx->lock);
679 ctx->is_active = 0;
680 if (likely(!ctx->nr_counters))
681 goto out;
682
683 flags = hw_perf_save_disable();
684 if (ctx->nr_active) {
685 list_for_each_entry(counter, &ctx->counter_list, list_entry)
686 group_sched_out(counter, cpuctx, ctx);
687 }
688 hw_perf_restore(flags);
689 out:
690 spin_unlock(&ctx->lock);
691}
692
693/*
694 * Called from scheduler to remove the counters of the current task,
695 * with interrupts disabled.
696 *
697 * We stop each counter and update the counter value in counter->count.
698 *
699 * This does not protect us against NMI, but disable()
700 * sets the disabled bit in the control field of counter _before_
701 * accessing the counter control register. If a NMI hits, then it will
702 * not restart the counter.
703 */
704void perf_counter_task_sched_out(struct task_struct *task, int cpu)
705{
706 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
707 struct perf_counter_context *ctx = &task->perf_counter_ctx;
708
709 if (likely(!cpuctx->task_ctx))
710 return;
711
712 __perf_counter_sched_out(ctx, cpuctx);
713
714 cpuctx->task_ctx = NULL;
715}
716
717static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
718{
719 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
720}
721
722static int
723group_sched_in(struct perf_counter *group_counter,
724 struct perf_cpu_context *cpuctx,
725 struct perf_counter_context *ctx,
726 int cpu)
727{
728 struct perf_counter *counter, *partial_group;
729 int ret;
730
731 if (group_counter->state == PERF_COUNTER_STATE_OFF)
732 return 0;
733
734 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
735 if (ret)
736 return ret < 0 ? ret : 0;
737
738 group_counter->prev_state = group_counter->state;
739 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
740 return -EAGAIN;
741
742 /*
743 * Schedule in siblings as one group (if any):
744 */
745 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
746 counter->prev_state = counter->state;
747 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
748 partial_group = counter;
749 goto group_error;
750 }
751 }
752
753 return 0;
754
755group_error:
756 /*
757 * Groups can be scheduled in as one unit only, so undo any
758 * partial group before returning:
759 */
760 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
761 if (counter == partial_group)
762 break;
763 counter_sched_out(counter, cpuctx, ctx);
764 }
765 counter_sched_out(group_counter, cpuctx, ctx);
766
767 return -EAGAIN;
768}
769
770static void
771__perf_counter_sched_in(struct perf_counter_context *ctx,
772 struct perf_cpu_context *cpuctx, int cpu)
773{
774 struct perf_counter *counter;
775 u64 flags;
776 int can_add_hw = 1;
777
778 spin_lock(&ctx->lock);
779 ctx->is_active = 1;
780 if (likely(!ctx->nr_counters))
781 goto out;
782
783 flags = hw_perf_save_disable();
784
785 /*
786 * First go through the list and put on any pinned groups
787 * in order to give them the best chance of going on.
788 */
789 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
790 if (counter->state <= PERF_COUNTER_STATE_OFF ||
791 !counter->hw_event.pinned)
792 continue;
793 if (counter->cpu != -1 && counter->cpu != cpu)
794 continue;
795
796 if (group_can_go_on(counter, cpuctx, 1))
797 group_sched_in(counter, cpuctx, ctx, cpu);
798
799 /*
800 * If this pinned group hasn't been scheduled,
801 * put it in error state.
802 */
803 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
804 counter->state = PERF_COUNTER_STATE_ERROR;
805 }
806
807 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
808 /*
809 * Ignore counters in OFF or ERROR state, and
810 * ignore pinned counters since we did them already.
811 */
812 if (counter->state <= PERF_COUNTER_STATE_OFF ||
813 counter->hw_event.pinned)
814 continue;
815
816 /*
817 * Listen to the 'cpu' scheduling filter constraint
818 * of counters:
819 */
820 if (counter->cpu != -1 && counter->cpu != cpu)
821 continue;
822
823 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
824 if (group_sched_in(counter, cpuctx, ctx, cpu))
825 can_add_hw = 0;
826 }
827 }
828 hw_perf_restore(flags);
829 out:
830 spin_unlock(&ctx->lock);
831}
832
833/*
834 * Called from scheduler to add the counters of the current task
835 * with interrupts disabled.
836 *
837 * We restore the counter value and then enable it.
838 *
839 * This does not protect us against NMI, but enable()
840 * sets the enabled bit in the control field of counter _before_
841 * accessing the counter control register. If a NMI hits, then it will
842 * keep the counter running.
843 */
844void perf_counter_task_sched_in(struct task_struct *task, int cpu)
845{
846 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
847 struct perf_counter_context *ctx = &task->perf_counter_ctx;
848
849 __perf_counter_sched_in(ctx, cpuctx, cpu);
850 cpuctx->task_ctx = ctx;
851}
852
853static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
854{
855 struct perf_counter_context *ctx = &cpuctx->ctx;
856
857 __perf_counter_sched_in(ctx, cpuctx, cpu);
858}
859
860int perf_counter_task_disable(void)
861{
862 struct task_struct *curr = current;
863 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
864 struct perf_counter *counter;
865 unsigned long flags;
866 u64 perf_flags;
867 int cpu;
868
869 if (likely(!ctx->nr_counters))
870 return 0;
871
872 curr_rq_lock_irq_save(&flags);
873 cpu = smp_processor_id();
874
875 /* force the update of the task clock: */
876 __task_delta_exec(curr, 1);
877
878 perf_counter_task_sched_out(curr, cpu);
879
880 spin_lock(&ctx->lock);
881
882 /*
883 * Disable all the counters:
884 */
885 perf_flags = hw_perf_save_disable();
886
887 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
888 if (counter->state != PERF_COUNTER_STATE_ERROR)
889 counter->state = PERF_COUNTER_STATE_OFF;
890 }
891
892 hw_perf_restore(perf_flags);
893
894 spin_unlock(&ctx->lock);
895
896 curr_rq_unlock_irq_restore(&flags);
897
898 return 0;
899}
900
901int perf_counter_task_enable(void)
902{
903 struct task_struct *curr = current;
904 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
905 struct perf_counter *counter;
906 unsigned long flags;
907 u64 perf_flags;
908 int cpu;
909
910 if (likely(!ctx->nr_counters))
911 return 0;
912
913 curr_rq_lock_irq_save(&flags);
914 cpu = smp_processor_id();
915
916 /* force the update of the task clock: */
917 __task_delta_exec(curr, 1);
918
919 perf_counter_task_sched_out(curr, cpu);
920
921 spin_lock(&ctx->lock);
922
923 /*
924 * Disable all the counters:
925 */
926 perf_flags = hw_perf_save_disable();
927
928 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
929 if (counter->state > PERF_COUNTER_STATE_OFF)
930 continue;
931 counter->state = PERF_COUNTER_STATE_INACTIVE;
932 counter->hw_event.disabled = 0;
933 }
934 hw_perf_restore(perf_flags);
935
936 spin_unlock(&ctx->lock);
937
938 perf_counter_task_sched_in(curr, cpu);
939
940 curr_rq_unlock_irq_restore(&flags);
941
942 return 0;
943}
944
945/*
946 * Round-robin a context's counters:
947 */
948static void rotate_ctx(struct perf_counter_context *ctx)
949{
950 struct perf_counter *counter;
951 u64 perf_flags;
952
953 if (!ctx->nr_counters)
954 return;
955
956 spin_lock(&ctx->lock);
957 /*
958 * Rotate the first entry last (works just fine for group counters too):
959 */
960 perf_flags = hw_perf_save_disable();
961 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
962 list_del(&counter->list_entry);
963 list_add_tail(&counter->list_entry, &ctx->counter_list);
964 break;
965 }
966 hw_perf_restore(perf_flags);
967
968 spin_unlock(&ctx->lock);
969}
970
971void perf_counter_task_tick(struct task_struct *curr, int cpu)
972{
973 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
974 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
975 const int rotate_percpu = 0;
976
977 if (rotate_percpu)
978 perf_counter_cpu_sched_out(cpuctx);
979 perf_counter_task_sched_out(curr, cpu);
980
981 if (rotate_percpu)
982 rotate_ctx(&cpuctx->ctx);
983 rotate_ctx(ctx);
984
985 if (rotate_percpu)
986 perf_counter_cpu_sched_in(cpuctx, cpu);
987 perf_counter_task_sched_in(curr, cpu);
988}
989
990/*
991 * Cross CPU call to read the hardware counter
992 */
993static void __read(void *info)
994{
995 struct perf_counter *counter = info;
996 unsigned long flags;
997
998 curr_rq_lock_irq_save(&flags);
999 counter->hw_ops->read(counter);
1000 curr_rq_unlock_irq_restore(&flags);
1001}
1002
1003static u64 perf_counter_read(struct perf_counter *counter)
1004{
1005 /*
1006 * If counter is enabled and currently active on a CPU, update the
1007 * value in the counter structure:
1008 */
1009 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1010 smp_call_function_single(counter->oncpu,
1011 __read, counter, 1);
1012 }
1013
1014 return atomic64_read(&counter->count);
1015}
1016
1017/*
1018 * Cross CPU call to switch performance data pointers
1019 */
1020static void __perf_switch_irq_data(void *info)
1021{
1022 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1023 struct perf_counter *counter = info;
1024 struct perf_counter_context *ctx = counter->ctx;
1025 struct perf_data *oldirqdata = counter->irqdata;
1026
1027 /*
1028 * If this is a task context, we need to check whether it is
1029 * the current task context of this cpu. If not it has been
1030 * scheduled out before the smp call arrived.
1031 */
1032 if (ctx->task) {
1033 if (cpuctx->task_ctx != ctx)
1034 return;
1035 spin_lock(&ctx->lock);
1036 }
1037
1038 /* Change the pointer NMI safe */
1039 atomic_long_set((atomic_long_t *)&counter->irqdata,
1040 (unsigned long) counter->usrdata);
1041 counter->usrdata = oldirqdata;
1042
1043 if (ctx->task)
1044 spin_unlock(&ctx->lock);
1045}
1046
1047static struct perf_data *perf_switch_irq_data(struct perf_counter *counter)
1048{
1049 struct perf_counter_context *ctx = counter->ctx;
1050 struct perf_data *oldirqdata = counter->irqdata;
1051 struct task_struct *task = ctx->task;
1052
1053 if (!task) {
1054 smp_call_function_single(counter->cpu,
1055 __perf_switch_irq_data,
1056 counter, 1);
1057 return counter->usrdata;
1058 }
1059
1060retry:
1061 spin_lock_irq(&ctx->lock);
1062 if (counter->state != PERF_COUNTER_STATE_ACTIVE) {
1063 counter->irqdata = counter->usrdata;
1064 counter->usrdata = oldirqdata;
1065 spin_unlock_irq(&ctx->lock);
1066 return oldirqdata;
1067 }
1068 spin_unlock_irq(&ctx->lock);
1069 task_oncpu_function_call(task, __perf_switch_irq_data, counter);
1070 /* Might have failed, because task was scheduled out */
1071 if (counter->irqdata == oldirqdata)
1072 goto retry;
1073
1074 return counter->usrdata;
1075}
1076
1077static void put_context(struct perf_counter_context *ctx)
1078{
1079 if (ctx->task)
1080 put_task_struct(ctx->task);
1081}
1082
1083static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1084{
1085 struct perf_cpu_context *cpuctx;
1086 struct perf_counter_context *ctx;
1087 struct task_struct *task;
1088
1089 /*
1090 * If cpu is not a wildcard then this is a percpu counter:
1091 */
1092 if (cpu != -1) {
1093 /* Must be root to operate on a CPU counter: */
1094 if (!capable(CAP_SYS_ADMIN))
1095 return ERR_PTR(-EACCES);
1096
1097 if (cpu < 0 || cpu > num_possible_cpus())
1098 return ERR_PTR(-EINVAL);
1099
1100 /*
1101 * We could be clever and allow to attach a counter to an
1102 * offline CPU and activate it when the CPU comes up, but
1103 * that's for later.
1104 */
1105 if (!cpu_isset(cpu, cpu_online_map))
1106 return ERR_PTR(-ENODEV);
1107
1108 cpuctx = &per_cpu(perf_cpu_context, cpu);
1109 ctx = &cpuctx->ctx;
1110
1111 return ctx;
1112 }
1113
1114 rcu_read_lock();
1115 if (!pid)
1116 task = current;
1117 else
1118 task = find_task_by_vpid(pid);
1119 if (task)
1120 get_task_struct(task);
1121 rcu_read_unlock();
1122
1123 if (!task)
1124 return ERR_PTR(-ESRCH);
1125
1126 ctx = &task->perf_counter_ctx;
1127 ctx->task = task;
1128
1129 /* Reuse ptrace permission checks for now. */
1130 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1131 put_context(ctx);
1132 return ERR_PTR(-EACCES);
1133 }
1134
1135 return ctx;
1136}
1137
1138/*
1139 * Called when the last reference to the file is gone.
1140 */
1141static int perf_release(struct inode *inode, struct file *file)
1142{
1143 struct perf_counter *counter = file->private_data;
1144 struct perf_counter_context *ctx = counter->ctx;
1145
1146 file->private_data = NULL;
1147
1148 mutex_lock(&ctx->mutex);
1149 mutex_lock(&counter->mutex);
1150
1151 perf_counter_remove_from_context(counter);
1152
1153 mutex_unlock(&counter->mutex);
1154 mutex_unlock(&ctx->mutex);
1155
1156 kfree(counter);
1157 put_context(ctx);
1158
1159 return 0;
1160}
1161
1162/*
1163 * Read the performance counter - simple non blocking version for now
1164 */
1165static ssize_t
1166perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1167{
1168 u64 cntval;
1169
1170 if (count != sizeof(cntval))
1171 return -EINVAL;
1172
1173 /*
1174 * Return end-of-file for a read on a counter that is in
1175 * error state (i.e. because it was pinned but it couldn't be
1176 * scheduled on to the CPU at some point).
1177 */
1178 if (counter->state == PERF_COUNTER_STATE_ERROR)
1179 return 0;
1180
1181 mutex_lock(&counter->mutex);
1182 cntval = perf_counter_read(counter);
1183 mutex_unlock(&counter->mutex);
1184
1185 return put_user(cntval, (u64 __user *) buf) ? -EFAULT : sizeof(cntval);
1186}
1187
1188static ssize_t
1189perf_copy_usrdata(struct perf_data *usrdata, char __user *buf, size_t count)
1190{
1191 if (!usrdata->len)
1192 return 0;
1193
1194 count = min(count, (size_t)usrdata->len);
1195 if (copy_to_user(buf, usrdata->data + usrdata->rd_idx, count))
1196 return -EFAULT;
1197
1198 /* Adjust the counters */
1199 usrdata->len -= count;
1200 if (!usrdata->len)
1201 usrdata->rd_idx = 0;
1202 else
1203 usrdata->rd_idx += count;
1204
1205 return count;
1206}
1207
1208static ssize_t
1209perf_read_irq_data(struct perf_counter *counter,
1210 char __user *buf,
1211 size_t count,
1212 int nonblocking)
1213{
1214 struct perf_data *irqdata, *usrdata;
1215 DECLARE_WAITQUEUE(wait, current);
1216 ssize_t res, res2;
1217
1218 irqdata = counter->irqdata;
1219 usrdata = counter->usrdata;
1220
1221 if (usrdata->len + irqdata->len >= count)
1222 goto read_pending;
1223
1224 if (nonblocking)
1225 return -EAGAIN;
1226
1227 spin_lock_irq(&counter->waitq.lock);
1228 __add_wait_queue(&counter->waitq, &wait);
1229 for (;;) {
1230 set_current_state(TASK_INTERRUPTIBLE);
1231 if (usrdata->len + irqdata->len >= count)
1232 break;
1233
1234 if (signal_pending(current))
1235 break;
1236
1237 if (counter->state == PERF_COUNTER_STATE_ERROR)
1238 break;
1239
1240 spin_unlock_irq(&counter->waitq.lock);
1241 schedule();
1242 spin_lock_irq(&counter->waitq.lock);
1243 }
1244 __remove_wait_queue(&counter->waitq, &wait);
1245 __set_current_state(TASK_RUNNING);
1246 spin_unlock_irq(&counter->waitq.lock);
1247
1248 if (usrdata->len + irqdata->len < count &&
1249 counter->state != PERF_COUNTER_STATE_ERROR)
1250 return -ERESTARTSYS;
1251read_pending:
1252 mutex_lock(&counter->mutex);
1253
1254 /* Drain pending data first: */
1255 res = perf_copy_usrdata(usrdata, buf, count);
1256 if (res < 0 || res == count)
1257 goto out;
1258
1259 /* Switch irq buffer: */
1260 usrdata = perf_switch_irq_data(counter);
1261 res2 = perf_copy_usrdata(usrdata, buf + res, count - res);
1262 if (res2 < 0) {
1263 if (!res)
1264 res = -EFAULT;
1265 } else {
1266 res += res2;
1267 }
1268out:
1269 mutex_unlock(&counter->mutex);
1270
1271 return res;
1272}
1273
1274static ssize_t
1275perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1276{
1277 struct perf_counter *counter = file->private_data;
1278
1279 switch (counter->hw_event.record_type) {
1280 case PERF_RECORD_SIMPLE:
1281 return perf_read_hw(counter, buf, count);
1282
1283 case PERF_RECORD_IRQ:
1284 case PERF_RECORD_GROUP:
1285 return perf_read_irq_data(counter, buf, count,
1286 file->f_flags & O_NONBLOCK);
1287 }
1288 return -EINVAL;
1289}
1290
1291static unsigned int perf_poll(struct file *file, poll_table *wait)
1292{
1293 struct perf_counter *counter = file->private_data;
1294 unsigned int events = 0;
1295 unsigned long flags;
1296
1297 poll_wait(file, &counter->waitq, wait);
1298
1299 spin_lock_irqsave(&counter->waitq.lock, flags);
1300 if (counter->usrdata->len || counter->irqdata->len)
1301 events |= POLLIN;
1302 spin_unlock_irqrestore(&counter->waitq.lock, flags);
1303
1304 return events;
1305}
1306
1307static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1308{
1309 struct perf_counter *counter = file->private_data;
1310 int err = 0;
1311
1312 switch (cmd) {
1313 case PERF_COUNTER_IOC_ENABLE:
1314 perf_counter_enable_family(counter);
1315 break;
1316 case PERF_COUNTER_IOC_DISABLE:
1317 perf_counter_disable_family(counter);
1318 break;
1319 default:
1320 err = -ENOTTY;
1321 }
1322 return err;
1323}
1324
1325static const struct file_operations perf_fops = {
1326 .release = perf_release,
1327 .read = perf_read,
1328 .poll = perf_poll,
1329 .unlocked_ioctl = perf_ioctl,
1330 .compat_ioctl = perf_ioctl,
1331};
1332
1333static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
1334{
1335 int cpu = raw_smp_processor_id();
1336
1337 atomic64_set(&counter->hw.prev_count, cpu_clock(cpu));
1338 return 0;
1339}
1340
1341static void cpu_clock_perf_counter_update(struct perf_counter *counter)
1342{
1343 int cpu = raw_smp_processor_id();
1344 s64 prev;
1345 u64 now;
1346
1347 now = cpu_clock(cpu);
1348 prev = atomic64_read(&counter->hw.prev_count);
1349 atomic64_set(&counter->hw.prev_count, now);
1350 atomic64_add(now - prev, &counter->count);
1351}
1352
1353static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
1354{
1355 cpu_clock_perf_counter_update(counter);
1356}
1357
1358static void cpu_clock_perf_counter_read(struct perf_counter *counter)
1359{
1360 cpu_clock_perf_counter_update(counter);
1361}
1362
1363static const struct hw_perf_counter_ops perf_ops_cpu_clock = {
1364 .enable = cpu_clock_perf_counter_enable,
1365 .disable = cpu_clock_perf_counter_disable,
1366 .read = cpu_clock_perf_counter_read,
1367};
1368
1369/*
1370 * Called from within the scheduler:
1371 */
1372static u64 task_clock_perf_counter_val(struct perf_counter *counter, int update)
1373{
1374 struct task_struct *curr = counter->task;
1375 u64 delta;
1376
1377 delta = __task_delta_exec(curr, update);
1378
1379 return curr->se.sum_exec_runtime + delta;
1380}
1381
1382static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
1383{
1384 u64 prev;
1385 s64 delta;
1386
1387 prev = atomic64_read(&counter->hw.prev_count);
1388
1389 atomic64_set(&counter->hw.prev_count, now);
1390
1391 delta = now - prev;
1392
1393 atomic64_add(delta, &counter->count);
1394}
1395
1396static void task_clock_perf_counter_read(struct perf_counter *counter)
1397{
1398 u64 now = task_clock_perf_counter_val(counter, 1);
1399
1400 task_clock_perf_counter_update(counter, now);
1401}
1402
1403static int task_clock_perf_counter_enable(struct perf_counter *counter)
1404{
1405 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
1406 atomic64_set(&counter->hw.prev_count,
1407 task_clock_perf_counter_val(counter, 0));
1408
1409 return 0;
1410}
1411
1412static void task_clock_perf_counter_disable(struct perf_counter *counter)
1413{
1414 u64 now = task_clock_perf_counter_val(counter, 0);
1415
1416 task_clock_perf_counter_update(counter, now);
1417}
1418
1419static const struct hw_perf_counter_ops perf_ops_task_clock = {
1420 .enable = task_clock_perf_counter_enable,
1421 .disable = task_clock_perf_counter_disable,
1422 .read = task_clock_perf_counter_read,
1423};
1424
1425#ifdef CONFIG_VM_EVENT_COUNTERS
1426#define cpu_page_faults() __get_cpu_var(vm_event_states).event[PGFAULT]
1427#else
1428#define cpu_page_faults() 0
1429#endif
1430
1431static u64 get_page_faults(struct perf_counter *counter)
1432{
1433 struct task_struct *curr = counter->ctx->task;
1434
1435 if (curr)
1436 return curr->maj_flt + curr->min_flt;
1437 return cpu_page_faults();
1438}
1439
1440static void page_faults_perf_counter_update(struct perf_counter *counter)
1441{
1442 u64 prev, now;
1443 s64 delta;
1444
1445 prev = atomic64_read(&counter->hw.prev_count);
1446 now = get_page_faults(counter);
1447
1448 atomic64_set(&counter->hw.prev_count, now);
1449
1450 delta = now - prev;
1451
1452 atomic64_add(delta, &counter->count);
1453}
1454
1455static void page_faults_perf_counter_read(struct perf_counter *counter)
1456{
1457 page_faults_perf_counter_update(counter);
1458}
1459
1460static int page_faults_perf_counter_enable(struct perf_counter *counter)
1461{
1462 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
1463 atomic64_set(&counter->hw.prev_count, get_page_faults(counter));
1464 return 0;
1465}
1466
1467static void page_faults_perf_counter_disable(struct perf_counter *counter)
1468{
1469 page_faults_perf_counter_update(counter);
1470}
1471
1472static const struct hw_perf_counter_ops perf_ops_page_faults = {
1473 .enable = page_faults_perf_counter_enable,
1474 .disable = page_faults_perf_counter_disable,
1475 .read = page_faults_perf_counter_read,
1476};
1477
1478static u64 get_context_switches(struct perf_counter *counter)
1479{
1480 struct task_struct *curr = counter->ctx->task;
1481
1482 if (curr)
1483 return curr->nvcsw + curr->nivcsw;
1484 return cpu_nr_switches(smp_processor_id());
1485}
1486
1487static void context_switches_perf_counter_update(struct perf_counter *counter)
1488{
1489 u64 prev, now;
1490 s64 delta;
1491
1492 prev = atomic64_read(&counter->hw.prev_count);
1493 now = get_context_switches(counter);
1494
1495 atomic64_set(&counter->hw.prev_count, now);
1496
1497 delta = now - prev;
1498
1499 atomic64_add(delta, &counter->count);
1500}
1501
1502static void context_switches_perf_counter_read(struct perf_counter *counter)
1503{
1504 context_switches_perf_counter_update(counter);
1505}
1506
1507static int context_switches_perf_counter_enable(struct perf_counter *counter)
1508{
1509 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
1510 atomic64_set(&counter->hw.prev_count,
1511 get_context_switches(counter));
1512 return 0;
1513}
1514
1515static void context_switches_perf_counter_disable(struct perf_counter *counter)
1516{
1517 context_switches_perf_counter_update(counter);
1518}
1519
1520static const struct hw_perf_counter_ops perf_ops_context_switches = {
1521 .enable = context_switches_perf_counter_enable,
1522 .disable = context_switches_perf_counter_disable,
1523 .read = context_switches_perf_counter_read,
1524};
1525
1526static inline u64 get_cpu_migrations(struct perf_counter *counter)
1527{
1528 struct task_struct *curr = counter->ctx->task;
1529
1530 if (curr)
1531 return curr->se.nr_migrations;
1532 return cpu_nr_migrations(smp_processor_id());
1533}
1534
1535static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
1536{
1537 u64 prev, now;
1538 s64 delta;
1539
1540 prev = atomic64_read(&counter->hw.prev_count);
1541 now = get_cpu_migrations(counter);
1542
1543 atomic64_set(&counter->hw.prev_count, now);
1544
1545 delta = now - prev;
1546
1547 atomic64_add(delta, &counter->count);
1548}
1549
1550static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
1551{
1552 cpu_migrations_perf_counter_update(counter);
1553}
1554
1555static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
1556{
1557 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
1558 atomic64_set(&counter->hw.prev_count,
1559 get_cpu_migrations(counter));
1560 return 0;
1561}
1562
1563static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
1564{
1565 cpu_migrations_perf_counter_update(counter);
1566}
1567
1568static const struct hw_perf_counter_ops perf_ops_cpu_migrations = {
1569 .enable = cpu_migrations_perf_counter_enable,
1570 .disable = cpu_migrations_perf_counter_disable,
1571 .read = cpu_migrations_perf_counter_read,
1572};
1573
1574static const struct hw_perf_counter_ops *
1575sw_perf_counter_init(struct perf_counter *counter)
1576{
1577 const struct hw_perf_counter_ops *hw_ops = NULL;
1578
1579 /*
1580 * Software counters (currently) can't in general distinguish
1581 * between user, kernel and hypervisor events.
1582 * However, context switches and cpu migrations are considered
1583 * to be kernel events, and page faults are never hypervisor
1584 * events.
1585 */
1586 switch (counter->hw_event.type) {
1587 case PERF_COUNT_CPU_CLOCK:
1588 if (!(counter->hw_event.exclude_user ||
1589 counter->hw_event.exclude_kernel ||
1590 counter->hw_event.exclude_hv))
1591 hw_ops = &perf_ops_cpu_clock;
1592 break;
1593 case PERF_COUNT_TASK_CLOCK:
1594 if (counter->hw_event.exclude_user ||
1595 counter->hw_event.exclude_kernel ||
1596 counter->hw_event.exclude_hv)
1597 break;
1598 /*
1599 * If the user instantiates this as a per-cpu counter,
1600 * use the cpu_clock counter instead.
1601 */
1602 if (counter->ctx->task)
1603 hw_ops = &perf_ops_task_clock;
1604 else
1605 hw_ops = &perf_ops_cpu_clock;
1606 break;
1607 case PERF_COUNT_PAGE_FAULTS:
1608 if (!(counter->hw_event.exclude_user ||
1609 counter->hw_event.exclude_kernel))
1610 hw_ops = &perf_ops_page_faults;
1611 break;
1612 case PERF_COUNT_CONTEXT_SWITCHES:
1613 if (!counter->hw_event.exclude_kernel)
1614 hw_ops = &perf_ops_context_switches;
1615 break;
1616 case PERF_COUNT_CPU_MIGRATIONS:
1617 if (!counter->hw_event.exclude_kernel)
1618 hw_ops = &perf_ops_cpu_migrations;
1619 break;
1620 default:
1621 break;
1622 }
1623 return hw_ops;
1624}
1625
1626/*
1627 * Allocate and initialize a counter structure
1628 */
1629static struct perf_counter *
1630perf_counter_alloc(struct perf_counter_hw_event *hw_event,
1631 int cpu,
1632 struct perf_counter_context *ctx,
1633 struct perf_counter *group_leader,
1634 gfp_t gfpflags)
1635{
1636 const struct hw_perf_counter_ops *hw_ops;
1637 struct perf_counter *counter;
1638
1639 counter = kzalloc(sizeof(*counter), gfpflags);
1640 if (!counter)
1641 return NULL;
1642
1643 /*
1644 * Single counters are their own group leaders, with an
1645 * empty sibling list:
1646 */
1647 if (!group_leader)
1648 group_leader = counter;
1649
1650 mutex_init(&counter->mutex);
1651 INIT_LIST_HEAD(&counter->list_entry);
1652 INIT_LIST_HEAD(&counter->sibling_list);
1653 init_waitqueue_head(&counter->waitq);
1654
1655 INIT_LIST_HEAD(&counter->child_list);
1656
1657 counter->irqdata = &counter->data[0];
1658 counter->usrdata = &counter->data[1];
1659 counter->cpu = cpu;
1660 counter->hw_event = *hw_event;
1661 counter->wakeup_pending = 0;
1662 counter->group_leader = group_leader;
1663 counter->hw_ops = NULL;
1664 counter->ctx = ctx;
1665
1666 counter->state = PERF_COUNTER_STATE_INACTIVE;
1667 if (hw_event->disabled)
1668 counter->state = PERF_COUNTER_STATE_OFF;
1669
1670 hw_ops = NULL;
1671 if (!hw_event->raw && hw_event->type < 0)
1672 hw_ops = sw_perf_counter_init(counter);
1673 else
1674 hw_ops = hw_perf_counter_init(counter);
1675
1676 if (!hw_ops) {
1677 kfree(counter);
1678 return NULL;
1679 }
1680 counter->hw_ops = hw_ops;
1681
1682 return counter;
1683}
1684
1685/**
1686 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
1687 *
1688 * @hw_event_uptr: event type attributes for monitoring/sampling
1689 * @pid: target pid
1690 * @cpu: target cpu
1691 * @group_fd: group leader counter fd
1692 */
1693SYSCALL_DEFINE5(perf_counter_open,
1694 const struct perf_counter_hw_event __user *, hw_event_uptr,
1695 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
1696{
1697 struct perf_counter *counter, *group_leader;
1698 struct perf_counter_hw_event hw_event;
1699 struct perf_counter_context *ctx;
1700 struct file *counter_file = NULL;
1701 struct file *group_file = NULL;
1702 int fput_needed = 0;
1703 int fput_needed2 = 0;
1704 int ret;
1705
1706 /* for future expandability... */
1707 if (flags)
1708 return -EINVAL;
1709
1710 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
1711 return -EFAULT;
1712
1713 /*
1714 * Get the target context (task or percpu):
1715 */
1716 ctx = find_get_context(pid, cpu);
1717 if (IS_ERR(ctx))
1718 return PTR_ERR(ctx);
1719
1720 /*
1721 * Look up the group leader (we will attach this counter to it):
1722 */
1723 group_leader = NULL;
1724 if (group_fd != -1) {
1725 ret = -EINVAL;
1726 group_file = fget_light(group_fd, &fput_needed);
1727 if (!group_file)
1728 goto err_put_context;
1729 if (group_file->f_op != &perf_fops)
1730 goto err_put_context;
1731
1732 group_leader = group_file->private_data;
1733 /*
1734 * Do not allow a recursive hierarchy (this new sibling
1735 * becoming part of another group-sibling):
1736 */
1737 if (group_leader->group_leader != group_leader)
1738 goto err_put_context;
1739 /*
1740 * Do not allow to attach to a group in a different
1741 * task or CPU context:
1742 */
1743 if (group_leader->ctx != ctx)
1744 goto err_put_context;
1745 /*
1746 * Only a group leader can be exclusive or pinned
1747 */
1748 if (hw_event.exclusive || hw_event.pinned)
1749 goto err_put_context;
1750 }
1751
1752 ret = -EINVAL;
1753 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
1754 GFP_KERNEL);
1755 if (!counter)
1756 goto err_put_context;
1757
1758 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
1759 if (ret < 0)
1760 goto err_free_put_context;
1761
1762 counter_file = fget_light(ret, &fput_needed2);
1763 if (!counter_file)
1764 goto err_free_put_context;
1765
1766 counter->filp = counter_file;
1767 mutex_lock(&ctx->mutex);
1768 perf_install_in_context(ctx, counter, cpu);
1769 mutex_unlock(&ctx->mutex);
1770
1771 fput_light(counter_file, fput_needed2);
1772
1773out_fput:
1774 fput_light(group_file, fput_needed);
1775
1776 return ret;
1777
1778err_free_put_context:
1779 kfree(counter);
1780
1781err_put_context:
1782 put_context(ctx);
1783
1784 goto out_fput;
1785}
1786
1787/*
1788 * Initialize the perf_counter context in a task_struct:
1789 */
1790static void
1791__perf_counter_init_context(struct perf_counter_context *ctx,
1792 struct task_struct *task)
1793{
1794 memset(ctx, 0, sizeof(*ctx));
1795 spin_lock_init(&ctx->lock);
1796 mutex_init(&ctx->mutex);
1797 INIT_LIST_HEAD(&ctx->counter_list);
1798 ctx->task = task;
1799}
1800
1801/*
1802 * inherit a counter from parent task to child task:
1803 */
1804static struct perf_counter *
1805inherit_counter(struct perf_counter *parent_counter,
1806 struct task_struct *parent,
1807 struct perf_counter_context *parent_ctx,
1808 struct task_struct *child,
1809 struct perf_counter *group_leader,
1810 struct perf_counter_context *child_ctx)
1811{
1812 struct perf_counter *child_counter;
1813
1814 /*
1815 * Instead of creating recursive hierarchies of counters,
1816 * we link inherited counters back to the original parent,
1817 * which has a filp for sure, which we use as the reference
1818 * count:
1819 */
1820 if (parent_counter->parent)
1821 parent_counter = parent_counter->parent;
1822
1823 child_counter = perf_counter_alloc(&parent_counter->hw_event,
1824 parent_counter->cpu, child_ctx,
1825 group_leader, GFP_KERNEL);
1826 if (!child_counter)
1827 return NULL;
1828
1829 /*
1830 * Link it up in the child's context:
1831 */
1832 child_counter->task = child;
1833 list_add_counter(child_counter, child_ctx);
1834 child_ctx->nr_counters++;
1835
1836 child_counter->parent = parent_counter;
1837 /*
1838 * inherit into child's child as well:
1839 */
1840 child_counter->hw_event.inherit = 1;
1841
1842 /*
1843 * Get a reference to the parent filp - we will fput it
1844 * when the child counter exits. This is safe to do because
1845 * we are in the parent and we know that the filp still
1846 * exists and has a nonzero count:
1847 */
1848 atomic_long_inc(&parent_counter->filp->f_count);
1849
1850 /*
1851 * Link this into the parent counter's child list
1852 */
1853 mutex_lock(&parent_counter->mutex);
1854 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
1855
1856 /*
1857 * Make the child state follow the state of the parent counter,
1858 * not its hw_event.disabled bit. We hold the parent's mutex,
1859 * so we won't race with perf_counter_{en,dis}able_family.
1860 */
1861 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
1862 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
1863 else
1864 child_counter->state = PERF_COUNTER_STATE_OFF;
1865
1866 mutex_unlock(&parent_counter->mutex);
1867
1868 return child_counter;
1869}
1870
1871static int inherit_group(struct perf_counter *parent_counter,
1872 struct task_struct *parent,
1873 struct perf_counter_context *parent_ctx,
1874 struct task_struct *child,
1875 struct perf_counter_context *child_ctx)
1876{
1877 struct perf_counter *leader;
1878 struct perf_counter *sub;
1879
1880 leader = inherit_counter(parent_counter, parent, parent_ctx,
1881 child, NULL, child_ctx);
1882 if (!leader)
1883 return -ENOMEM;
1884 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
1885 if (!inherit_counter(sub, parent, parent_ctx,
1886 child, leader, child_ctx))
1887 return -ENOMEM;
1888 }
1889 return 0;
1890}
1891
1892static void sync_child_counter(struct perf_counter *child_counter,
1893 struct perf_counter *parent_counter)
1894{
1895 u64 parent_val, child_val;
1896
1897 parent_val = atomic64_read(&parent_counter->count);
1898 child_val = atomic64_read(&child_counter->count);
1899
1900 /*
1901 * Add back the child's count to the parent's count:
1902 */
1903 atomic64_add(child_val, &parent_counter->count);
1904
1905 /*
1906 * Remove this counter from the parent's list
1907 */
1908 mutex_lock(&parent_counter->mutex);
1909 list_del_init(&child_counter->child_list);
1910 mutex_unlock(&parent_counter->mutex);
1911
1912 /*
1913 * Release the parent counter, if this was the last
1914 * reference to it.
1915 */
1916 fput(parent_counter->filp);
1917}
1918
1919static void
1920__perf_counter_exit_task(struct task_struct *child,
1921 struct perf_counter *child_counter,
1922 struct perf_counter_context *child_ctx)
1923{
1924 struct perf_counter *parent_counter;
1925 struct perf_counter *sub, *tmp;
1926
1927 /*
1928 * If we do not self-reap then we have to wait for the
1929 * child task to unschedule (it will happen for sure),
1930 * so that its counter is at its final count. (This
1931 * condition triggers rarely - child tasks usually get
1932 * off their CPU before the parent has a chance to
1933 * get this far into the reaping action)
1934 */
1935 if (child != current) {
1936 wait_task_inactive(child, 0);
1937 list_del_init(&child_counter->list_entry);
1938 } else {
1939 struct perf_cpu_context *cpuctx;
1940 unsigned long flags;
1941 u64 perf_flags;
1942
1943 /*
1944 * Disable and unlink this counter.
1945 *
1946 * Be careful about zapping the list - IRQ/NMI context
1947 * could still be processing it:
1948 */
1949 curr_rq_lock_irq_save(&flags);
1950 perf_flags = hw_perf_save_disable();
1951
1952 cpuctx = &__get_cpu_var(perf_cpu_context);
1953
1954 group_sched_out(child_counter, cpuctx, child_ctx);
1955
1956 list_del_init(&child_counter->list_entry);
1957
1958 child_ctx->nr_counters--;
1959
1960 hw_perf_restore(perf_flags);
1961 curr_rq_unlock_irq_restore(&flags);
1962 }
1963
1964 parent_counter = child_counter->parent;
1965 /*
1966 * It can happen that parent exits first, and has counters
1967 * that are still around due to the child reference. These
1968 * counters need to be zapped - but otherwise linger.
1969 */
1970 if (parent_counter) {
1971 sync_child_counter(child_counter, parent_counter);
1972 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
1973 list_entry) {
1974 if (sub->parent) {
1975 sync_child_counter(sub, sub->parent);
1976 kfree(sub);
1977 }
1978 }
1979 kfree(child_counter);
1980 }
1981}
1982
1983/*
1984 * When a child task exits, feed back counter values to parent counters.
1985 *
1986 * Note: we may be running in child context, but the PID is not hashed
1987 * anymore so new counters will not be added.
1988 */
1989void perf_counter_exit_task(struct task_struct *child)
1990{
1991 struct perf_counter *child_counter, *tmp;
1992 struct perf_counter_context *child_ctx;
1993
1994 child_ctx = &child->perf_counter_ctx;
1995
1996 if (likely(!child_ctx->nr_counters))
1997 return;
1998
1999 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
2000 list_entry)
2001 __perf_counter_exit_task(child, child_counter, child_ctx);
2002}
2003
2004/*
2005 * Initialize the perf_counter context in task_struct
2006 */
2007void perf_counter_init_task(struct task_struct *child)
2008{
2009 struct perf_counter_context *child_ctx, *parent_ctx;
2010 struct perf_counter *counter;
2011 struct task_struct *parent = current;
2012
2013 child_ctx = &child->perf_counter_ctx;
2014 parent_ctx = &parent->perf_counter_ctx;
2015
2016 __perf_counter_init_context(child_ctx, child);
2017
2018 /*
2019 * This is executed from the parent task context, so inherit
2020 * counters that have been marked for cloning:
2021 */
2022
2023 if (likely(!parent_ctx->nr_counters))
2024 return;
2025
2026 /*
2027 * Lock the parent list. No need to lock the child - not PID
2028 * hashed yet and not running, so nobody can access it.
2029 */
2030 mutex_lock(&parent_ctx->mutex);
2031
2032 /*
2033 * We dont have to disable NMIs - we are only looking at
2034 * the list, not manipulating it:
2035 */
2036 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
2037 if (!counter->hw_event.inherit)
2038 continue;
2039
2040 if (inherit_group(counter, parent,
2041 parent_ctx, child, child_ctx))
2042 break;
2043 }
2044
2045 mutex_unlock(&parent_ctx->mutex);
2046}
2047
2048static void __cpuinit perf_counter_init_cpu(int cpu)
2049{
2050 struct perf_cpu_context *cpuctx;
2051
2052 cpuctx = &per_cpu(perf_cpu_context, cpu);
2053 __perf_counter_init_context(&cpuctx->ctx, NULL);
2054
2055 mutex_lock(&perf_resource_mutex);
2056 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
2057 mutex_unlock(&perf_resource_mutex);
2058
2059 hw_perf_counter_setup(cpu);
2060}
2061
2062#ifdef CONFIG_HOTPLUG_CPU
2063static void __perf_counter_exit_cpu(void *info)
2064{
2065 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
2066 struct perf_counter_context *ctx = &cpuctx->ctx;
2067 struct perf_counter *counter, *tmp;
2068
2069 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
2070 __perf_counter_remove_from_context(counter);
2071}
2072static void perf_counter_exit_cpu(int cpu)
2073{
2074 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
2075 struct perf_counter_context *ctx = &cpuctx->ctx;
2076
2077 mutex_lock(&ctx->mutex);
2078 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
2079 mutex_unlock(&ctx->mutex);
2080}
2081#else
2082static inline void perf_counter_exit_cpu(int cpu) { }
2083#endif
2084
2085static int __cpuinit
2086perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
2087{
2088 unsigned int cpu = (long)hcpu;
2089
2090 switch (action) {
2091
2092 case CPU_UP_PREPARE:
2093 case CPU_UP_PREPARE_FROZEN:
2094 perf_counter_init_cpu(cpu);
2095 break;
2096
2097 case CPU_DOWN_PREPARE:
2098 case CPU_DOWN_PREPARE_FROZEN:
2099 perf_counter_exit_cpu(cpu);
2100 break;
2101
2102 default:
2103 break;
2104 }
2105
2106 return NOTIFY_OK;
2107}
2108
2109static struct notifier_block __cpuinitdata perf_cpu_nb = {
2110 .notifier_call = perf_cpu_notify,
2111};
2112
2113static int __init perf_counter_init(void)
2114{
2115 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
2116 (void *)(long)smp_processor_id());
2117 register_cpu_notifier(&perf_cpu_nb);
2118
2119 return 0;
2120}
2121early_initcall(perf_counter_init);
2122
2123static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
2124{
2125 return sprintf(buf, "%d\n", perf_reserved_percpu);
2126}
2127
2128static ssize_t
2129perf_set_reserve_percpu(struct sysdev_class *class,
2130 const char *buf,
2131 size_t count)
2132{
2133 struct perf_cpu_context *cpuctx;
2134 unsigned long val;
2135 int err, cpu, mpt;
2136
2137 err = strict_strtoul(buf, 10, &val);
2138 if (err)
2139 return err;
2140 if (val > perf_max_counters)
2141 return -EINVAL;
2142
2143 mutex_lock(&perf_resource_mutex);
2144 perf_reserved_percpu = val;
2145 for_each_online_cpu(cpu) {
2146 cpuctx = &per_cpu(perf_cpu_context, cpu);
2147 spin_lock_irq(&cpuctx->ctx.lock);
2148 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
2149 perf_max_counters - perf_reserved_percpu);
2150 cpuctx->max_pertask = mpt;
2151 spin_unlock_irq(&cpuctx->ctx.lock);
2152 }
2153 mutex_unlock(&perf_resource_mutex);
2154
2155 return count;
2156}
2157
2158static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
2159{
2160 return sprintf(buf, "%d\n", perf_overcommit);
2161}
2162
2163static ssize_t
2164perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
2165{
2166 unsigned long val;
2167 int err;
2168
2169 err = strict_strtoul(buf, 10, &val);
2170 if (err)
2171 return err;
2172 if (val > 1)
2173 return -EINVAL;
2174
2175 mutex_lock(&perf_resource_mutex);
2176 perf_overcommit = val;
2177 mutex_unlock(&perf_resource_mutex);
2178
2179 return count;
2180}
2181
2182static SYSDEV_CLASS_ATTR(
2183 reserve_percpu,
2184 0644,
2185 perf_show_reserve_percpu,
2186 perf_set_reserve_percpu
2187 );
2188
2189static SYSDEV_CLASS_ATTR(
2190 overcommit,
2191 0644,
2192 perf_show_overcommit,
2193 perf_set_overcommit
2194 );
2195
2196static struct attribute *perfclass_attrs[] = {
2197 &attr_reserve_percpu.attr,
2198 &attr_overcommit.attr,
2199 NULL
2200};
2201
2202static struct attribute_group perfclass_attr_group = {
2203 .attrs = perfclass_attrs,
2204 .name = "perf_counters",
2205};
2206
2207static int __init perf_counter_sysfs_init(void)
2208{
2209 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
2210 &perfclass_attr_group);
2211}
2212device_initcall(perf_counter_sysfs_init);
diff --git a/kernel/sched.c b/kernel/sched.c
index bec249885e17..39e708602169 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -577,6 +577,7 @@ struct rq {
577 struct load_weight load; 577 struct load_weight load;
578 unsigned long nr_load_updates; 578 unsigned long nr_load_updates;
579 u64 nr_switches; 579 u64 nr_switches;
580 u64 nr_migrations_in;
580 581
581 struct cfs_rq cfs; 582 struct cfs_rq cfs;
582 struct rt_rq rt; 583 struct rt_rq rt;
@@ -685,7 +686,7 @@ static inline int cpu_of(struct rq *rq)
685#define task_rq(p) cpu_rq(task_cpu(p)) 686#define task_rq(p) cpu_rq(task_cpu(p))
686#define cpu_curr(cpu) (cpu_rq(cpu)->curr) 687#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
687 688
688static inline void update_rq_clock(struct rq *rq) 689inline void update_rq_clock(struct rq *rq)
689{ 690{
690 rq->clock = sched_clock_cpu(cpu_of(rq)); 691 rq->clock = sched_clock_cpu(cpu_of(rq));
691} 692}
@@ -996,6 +997,26 @@ static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
996 } 997 }
997} 998}
998 999
1000void curr_rq_lock_irq_save(unsigned long *flags)
1001 __acquires(rq->lock)
1002{
1003 struct rq *rq;
1004
1005 local_irq_save(*flags);
1006 rq = cpu_rq(smp_processor_id());
1007 spin_lock(&rq->lock);
1008}
1009
1010void curr_rq_unlock_irq_restore(unsigned long *flags)
1011 __releases(rq->lock)
1012{
1013 struct rq *rq;
1014
1015 rq = cpu_rq(smp_processor_id());
1016 spin_unlock(&rq->lock);
1017 local_irq_restore(*flags);
1018}
1019
999void task_rq_unlock_wait(struct task_struct *p) 1020void task_rq_unlock_wait(struct task_struct *p)
1000{ 1021{
1001 struct rq *rq = task_rq(p); 1022 struct rq *rq = task_rq(p);
@@ -1947,12 +1968,15 @@ void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1947 p->se.sleep_start -= clock_offset; 1968 p->se.sleep_start -= clock_offset;
1948 if (p->se.block_start) 1969 if (p->se.block_start)
1949 p->se.block_start -= clock_offset; 1970 p->se.block_start -= clock_offset;
1971#endif
1950 if (old_cpu != new_cpu) { 1972 if (old_cpu != new_cpu) {
1951 schedstat_inc(p, se.nr_migrations); 1973 p->se.nr_migrations++;
1974 new_rq->nr_migrations_in++;
1975#ifdef CONFIG_SCHEDSTATS
1952 if (task_hot(p, old_rq->clock, NULL)) 1976 if (task_hot(p, old_rq->clock, NULL))
1953 schedstat_inc(p, se.nr_forced2_migrations); 1977 schedstat_inc(p, se.nr_forced2_migrations);
1954 }
1955#endif 1978#endif
1979 }
1956 p->se.vruntime -= old_cfsrq->min_vruntime - 1980 p->se.vruntime -= old_cfsrq->min_vruntime -
1957 new_cfsrq->min_vruntime; 1981 new_cfsrq->min_vruntime;
1958 1982
@@ -2304,6 +2328,27 @@ static int sched_balance_self(int cpu, int flag)
2304 2328
2305#endif /* CONFIG_SMP */ 2329#endif /* CONFIG_SMP */
2306 2330
2331/**
2332 * task_oncpu_function_call - call a function on the cpu on which a task runs
2333 * @p: the task to evaluate
2334 * @func: the function to be called
2335 * @info: the function call argument
2336 *
2337 * Calls the function @func when the task is currently running. This might
2338 * be on the current CPU, which just calls the function directly
2339 */
2340void task_oncpu_function_call(struct task_struct *p,
2341 void (*func) (void *info), void *info)
2342{
2343 int cpu;
2344
2345 preempt_disable();
2346 cpu = task_cpu(p);
2347 if (task_curr(p))
2348 smp_call_function_single(cpu, func, info, 1);
2349 preempt_enable();
2350}
2351
2307/*** 2352/***
2308 * try_to_wake_up - wake up a thread 2353 * try_to_wake_up - wake up a thread
2309 * @p: the to-be-woken-up thread 2354 * @p: the to-be-woken-up thread
@@ -2460,6 +2505,7 @@ static void __sched_fork(struct task_struct *p)
2460 p->se.exec_start = 0; 2505 p->se.exec_start = 0;
2461 p->se.sum_exec_runtime = 0; 2506 p->se.sum_exec_runtime = 0;
2462 p->se.prev_sum_exec_runtime = 0; 2507 p->se.prev_sum_exec_runtime = 0;
2508 p->se.nr_migrations = 0;
2463 p->se.last_wakeup = 0; 2509 p->se.last_wakeup = 0;
2464 p->se.avg_overlap = 0; 2510 p->se.avg_overlap = 0;
2465 p->se.start_runtime = 0; 2511 p->se.start_runtime = 0;
@@ -2690,6 +2736,7 @@ static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2690 */ 2736 */
2691 prev_state = prev->state; 2737 prev_state = prev->state;
2692 finish_arch_switch(prev); 2738 finish_arch_switch(prev);
2739 perf_counter_task_sched_in(current, cpu_of(rq));
2693 finish_lock_switch(rq, prev); 2740 finish_lock_switch(rq, prev);
2694#ifdef CONFIG_SMP 2741#ifdef CONFIG_SMP
2695 if (post_schedule) 2742 if (post_schedule)
@@ -2852,6 +2899,21 @@ unsigned long nr_active(void)
2852} 2899}
2853 2900
2854/* 2901/*
2902 * Externally visible per-cpu scheduler statistics:
2903 * cpu_nr_switches(cpu) - number of context switches on that cpu
2904 * cpu_nr_migrations(cpu) - number of migrations into that cpu
2905 */
2906u64 cpu_nr_switches(int cpu)
2907{
2908 return cpu_rq(cpu)->nr_switches;
2909}
2910
2911u64 cpu_nr_migrations(int cpu)
2912{
2913 return cpu_rq(cpu)->nr_migrations_in;
2914}
2915
2916/*
2855 * Update rq->cpu_load[] statistics. This function is usually called every 2917 * Update rq->cpu_load[] statistics. This function is usually called every
2856 * scheduler tick (TICK_NSEC). 2918 * scheduler tick (TICK_NSEC).
2857 */ 2919 */
@@ -4506,6 +4568,29 @@ EXPORT_PER_CPU_SYMBOL(kstat);
4506 * Return any ns on the sched_clock that have not yet been banked in 4568 * Return any ns on the sched_clock that have not yet been banked in
4507 * @p in case that task is currently running. 4569 * @p in case that task is currently running.
4508 */ 4570 */
4571unsigned long long __task_delta_exec(struct task_struct *p, int update)
4572{
4573 s64 delta_exec;
4574 struct rq *rq;
4575
4576 rq = task_rq(p);
4577 WARN_ON_ONCE(!runqueue_is_locked());
4578 WARN_ON_ONCE(!task_current(rq, p));
4579
4580 if (update)
4581 update_rq_clock(rq);
4582
4583 delta_exec = rq->clock - p->se.exec_start;
4584
4585 WARN_ON_ONCE(delta_exec < 0);
4586
4587 return delta_exec;
4588}
4589
4590/*
4591 * Return any ns on the sched_clock that have not yet been banked in
4592 * @p in case that task is currently running.
4593 */
4509unsigned long long task_delta_exec(struct task_struct *p) 4594unsigned long long task_delta_exec(struct task_struct *p)
4510{ 4595{
4511 unsigned long flags; 4596 unsigned long flags;
@@ -4765,6 +4850,7 @@ void scheduler_tick(void)
4765 update_rq_clock(rq); 4850 update_rq_clock(rq);
4766 update_cpu_load(rq); 4851 update_cpu_load(rq);
4767 curr->sched_class->task_tick(rq, curr, 0); 4852 curr->sched_class->task_tick(rq, curr, 0);
4853 perf_counter_task_tick(curr, cpu);
4768 spin_unlock(&rq->lock); 4854 spin_unlock(&rq->lock);
4769 4855
4770#ifdef CONFIG_SMP 4856#ifdef CONFIG_SMP
@@ -4980,6 +5066,7 @@ need_resched_nonpreemptible:
4980 5066
4981 if (likely(prev != next)) { 5067 if (likely(prev != next)) {
4982 sched_info_switch(prev, next); 5068 sched_info_switch(prev, next);
5069 perf_counter_task_sched_out(prev, cpu);
4983 5070
4984 rq->nr_switches++; 5071 rq->nr_switches++;
4985 rq->curr = next; 5072 rq->curr = next;
diff --git a/kernel/sys.c b/kernel/sys.c
index 51dbb55604e8..14c4c5613118 100644
--- a/kernel/sys.c
+++ b/kernel/sys.c
@@ -14,6 +14,7 @@
14#include <linux/prctl.h> 14#include <linux/prctl.h>
15#include <linux/highuid.h> 15#include <linux/highuid.h>
16#include <linux/fs.h> 16#include <linux/fs.h>
17#include <linux/perf_counter.h>
17#include <linux/resource.h> 18#include <linux/resource.h>
18#include <linux/kernel.h> 19#include <linux/kernel.h>
19#include <linux/kexec.h> 20#include <linux/kexec.h>
@@ -1799,6 +1800,12 @@ SYSCALL_DEFINE5(prctl, int, option, unsigned long, arg2, unsigned long, arg3,
1799 case PR_SET_TSC: 1800 case PR_SET_TSC:
1800 error = SET_TSC_CTL(arg2); 1801 error = SET_TSC_CTL(arg2);
1801 break; 1802 break;
1803 case PR_TASK_PERF_COUNTERS_DISABLE:
1804 error = perf_counter_task_disable();
1805 break;
1806 case PR_TASK_PERF_COUNTERS_ENABLE:
1807 error = perf_counter_task_enable();
1808 break;
1802 case PR_GET_TIMERSLACK: 1809 case PR_GET_TIMERSLACK:
1803 error = current->timer_slack_ns; 1810 error = current->timer_slack_ns;
1804 break; 1811 break;
diff --git a/kernel/sys_ni.c b/kernel/sys_ni.c
index 27dad2967387..68320f6b07b5 100644
--- a/kernel/sys_ni.c
+++ b/kernel/sys_ni.c
@@ -175,3 +175,6 @@ cond_syscall(compat_sys_timerfd_settime);
175cond_syscall(compat_sys_timerfd_gettime); 175cond_syscall(compat_sys_timerfd_gettime);
176cond_syscall(sys_eventfd); 176cond_syscall(sys_eventfd);
177cond_syscall(sys_eventfd2); 177cond_syscall(sys_eventfd2);
178
179/* performance counters: */
180cond_syscall(sys_perf_counter_open);
generated by cgit v1.2.2 (git 2.25.0) at 2025-01-03 19:30:07 -0500