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.c2199
-rw-r--r--kernel/sched.c93
-rw-r--r--kernel/sys.c7
-rw-r--r--kernel/sys_ni.c3
7 files changed, 2310 insertions, 7 deletions
diff --git a/kernel/Makefile b/kernel/Makefile
index 170a9213c1b6..5537554ed808 100644
--- a/kernel/Makefile
+++ b/kernel/Makefile
@@ -92,6 +92,7 @@ obj-$(CONFIG_HAVE_GENERIC_DMA_COHERENT) += dma-coherent.o
92obj-$(CONFIG_FUNCTION_TRACER) += trace/ 92obj-$(CONFIG_FUNCTION_TRACER) += trace/
93obj-$(CONFIG_TRACING) += trace/ 93obj-$(CONFIG_TRACING) += trace/
94obj-$(CONFIG_SMP) += sched_cpupri.o 94obj-$(CONFIG_SMP) += sched_cpupri.o
95obj-$(CONFIG_PERF_COUNTERS) += perf_counter.o
95 96
96ifneq ($(CONFIG_SCHED_OMIT_FRAME_POINTER),y) 97ifneq ($(CONFIG_SCHED_OMIT_FRAME_POINTER),y)
97# According to Alan Modra <alan@linuxcare.com.au>, the -fno-omit-frame-pointer is 98# 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 167e1e3ad7c6..f52c24eb8a8f 100644
--- a/kernel/exit.c
+++ b/kernel/exit.c
@@ -162,6 +162,9 @@ static void delayed_put_task_struct(struct rcu_head *rhp)
162{ 162{
163 struct task_struct *tsk = container_of(rhp, struct task_struct, rcu); 163 struct task_struct *tsk = container_of(rhp, struct task_struct, rcu);
164 164
165#ifdef CONFIG_PERF_COUNTERS
166 WARN_ON_ONCE(!list_empty(&tsk->perf_counter_ctx.counter_list));
167#endif
165 trace_sched_process_free(tsk); 168 trace_sched_process_free(tsk);
166 put_task_struct(tsk); 169 put_task_struct(tsk);
167} 170}
@@ -1093,10 +1096,6 @@ NORET_TYPE void do_exit(long code)
1093 tsk->mempolicy = NULL; 1096 tsk->mempolicy = NULL;
1094#endif 1097#endif
1095#ifdef CONFIG_FUTEX 1098#ifdef CONFIG_FUTEX
1096 /*
1097 * This must happen late, after the PID is not
1098 * hashed anymore:
1099 */
1100 if (unlikely(!list_empty(&tsk->pi_state_list))) 1099 if (unlikely(!list_empty(&tsk->pi_state_list)))
1101 exit_pi_state_list(tsk); 1100 exit_pi_state_list(tsk);
1102 if (unlikely(current->pi_state_cache)) 1101 if (unlikely(current->pi_state_cache))
@@ -1363,6 +1362,12 @@ static int wait_task_zombie(struct task_struct *p, int options,
1363 */ 1362 */
1364 read_unlock(&tasklist_lock); 1363 read_unlock(&tasklist_lock);
1365 1364
1365 /*
1366 * Flush inherited counters to the parent - before the parent
1367 * gets woken up by child-exit notifications.
1368 */
1369 perf_counter_exit_task(p);
1370
1366 retval = ru ? getrusage(p, RUSAGE_BOTH, ru) : 0; 1371 retval = ru ? getrusage(p, RUSAGE_BOTH, ru) : 0;
1367 status = (p->signal->flags & SIGNAL_GROUP_EXIT) 1372 status = (p->signal->flags & SIGNAL_GROUP_EXIT)
1368 ? p->signal->group_exit_code : p->exit_code; 1373 ? p->signal->group_exit_code : p->exit_code;
diff --git a/kernel/fork.c b/kernel/fork.c
index 8de303bdd4e5..4640a3e0085e 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -989,6 +989,7 @@ static struct task_struct *copy_process(unsigned long clone_flags,
989 goto fork_out; 989 goto fork_out;
990 990
991 rt_mutex_init_task(p); 991 rt_mutex_init_task(p);
992 perf_counter_init_task(p);
992 993
993#ifdef CONFIG_PROVE_LOCKING 994#ifdef CONFIG_PROVE_LOCKING
994 DEBUG_LOCKS_WARN_ON(!p->hardirqs_enabled); 995 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..fcefb0a726f3
--- /dev/null
+++ b/kernel/perf_counter.c
@@ -0,0 +1,2199 @@
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
448 /*
449 * Don't put the counter on if it is disabled or if
450 * it is in a group and the group isn't on.
451 */
452 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
453 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
454 goto unlock;
455
456 /*
457 * An exclusive counter can't go on if there are already active
458 * hardware counters, and no hardware counter can go on if there
459 * is already an exclusive counter on.
460 */
461 if (!group_can_go_on(counter, cpuctx, 1))
462 err = -EEXIST;
463 else
464 err = counter_sched_in(counter, cpuctx, ctx, cpu);
465
466 if (err) {
467 /*
468 * This counter couldn't go on. If it is in a group
469 * then we have to pull the whole group off.
470 * If the counter group is pinned then put it in error state.
471 */
472 if (leader != counter)
473 group_sched_out(leader, cpuctx, ctx);
474 if (leader->hw_event.pinned)
475 leader->state = PERF_COUNTER_STATE_ERROR;
476 }
477
478 if (!err && !ctx->task && cpuctx->max_pertask)
479 cpuctx->max_pertask--;
480
481 unlock:
482 hw_perf_restore(perf_flags);
483
484 spin_unlock(&ctx->lock);
485 curr_rq_unlock_irq_restore(&flags);
486}
487
488/*
489 * Attach a performance counter to a context
490 *
491 * First we add the counter to the list with the hardware enable bit
492 * in counter->hw_config cleared.
493 *
494 * If the counter is attached to a task which is on a CPU we use a smp
495 * call to enable it in the task context. The task might have been
496 * scheduled away, but we check this in the smp call again.
497 *
498 * Must be called with ctx->mutex held.
499 */
500static void
501perf_install_in_context(struct perf_counter_context *ctx,
502 struct perf_counter *counter,
503 int cpu)
504{
505 struct task_struct *task = ctx->task;
506
507 if (!task) {
508 /*
509 * Per cpu counters are installed via an smp call and
510 * the install is always sucessful.
511 */
512 smp_call_function_single(cpu, __perf_install_in_context,
513 counter, 1);
514 return;
515 }
516
517 counter->task = task;
518retry:
519 task_oncpu_function_call(task, __perf_install_in_context,
520 counter);
521
522 spin_lock_irq(&ctx->lock);
523 /*
524 * we need to retry the smp call.
525 */
526 if (ctx->is_active && list_empty(&counter->list_entry)) {
527 spin_unlock_irq(&ctx->lock);
528 goto retry;
529 }
530
531 /*
532 * The lock prevents that this context is scheduled in so we
533 * can add the counter safely, if it the call above did not
534 * succeed.
535 */
536 if (list_empty(&counter->list_entry)) {
537 list_add_counter(counter, ctx);
538 ctx->nr_counters++;
539 }
540 spin_unlock_irq(&ctx->lock);
541}
542
543/*
544 * Cross CPU call to enable a performance counter
545 */
546static void __perf_counter_enable(void *info)
547{
548 struct perf_counter *counter = info;
549 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
550 struct perf_counter_context *ctx = counter->ctx;
551 struct perf_counter *leader = counter->group_leader;
552 unsigned long flags;
553 int err;
554
555 /*
556 * If this is a per-task counter, need to check whether this
557 * counter's task is the current task on this cpu.
558 */
559 if (ctx->task && cpuctx->task_ctx != ctx)
560 return;
561
562 curr_rq_lock_irq_save(&flags);
563 spin_lock(&ctx->lock);
564
565 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
566 goto unlock;
567 counter->state = PERF_COUNTER_STATE_INACTIVE;
568
569 /*
570 * If the counter is in a group and isn't the group leader,
571 * then don't put it on unless the group is on.
572 */
573 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
574 goto unlock;
575
576 if (!group_can_go_on(counter, cpuctx, 1))
577 err = -EEXIST;
578 else
579 err = counter_sched_in(counter, cpuctx, ctx,
580 smp_processor_id());
581
582 if (err) {
583 /*
584 * If this counter can't go on and it's part of a
585 * group, then the whole group has to come off.
586 */
587 if (leader != counter)
588 group_sched_out(leader, cpuctx, ctx);
589 if (leader->hw_event.pinned)
590 leader->state = PERF_COUNTER_STATE_ERROR;
591 }
592
593 unlock:
594 spin_unlock(&ctx->lock);
595 curr_rq_unlock_irq_restore(&flags);
596}
597
598/*
599 * Enable a counter.
600 */
601static void perf_counter_enable(struct perf_counter *counter)
602{
603 struct perf_counter_context *ctx = counter->ctx;
604 struct task_struct *task = ctx->task;
605
606 if (!task) {
607 /*
608 * Enable the counter on the cpu that it's on
609 */
610 smp_call_function_single(counter->cpu, __perf_counter_enable,
611 counter, 1);
612 return;
613 }
614
615 spin_lock_irq(&ctx->lock);
616 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
617 goto out;
618
619 /*
620 * If the counter is in error state, clear that first.
621 * That way, if we see the counter in error state below, we
622 * know that it has gone back into error state, as distinct
623 * from the task having been scheduled away before the
624 * cross-call arrived.
625 */
626 if (counter->state == PERF_COUNTER_STATE_ERROR)
627 counter->state = PERF_COUNTER_STATE_OFF;
628
629 retry:
630 spin_unlock_irq(&ctx->lock);
631 task_oncpu_function_call(task, __perf_counter_enable, counter);
632
633 spin_lock_irq(&ctx->lock);
634
635 /*
636 * If the context is active and the counter is still off,
637 * we need to retry the cross-call.
638 */
639 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
640 goto retry;
641
642 /*
643 * Since we have the lock this context can't be scheduled
644 * in, so we can change the state safely.
645 */
646 if (counter->state == PERF_COUNTER_STATE_OFF)
647 counter->state = PERF_COUNTER_STATE_INACTIVE;
648 out:
649 spin_unlock_irq(&ctx->lock);
650}
651
652/*
653 * Enable a counter and all its children.
654 */
655static void perf_counter_enable_family(struct perf_counter *counter)
656{
657 struct perf_counter *child;
658
659 perf_counter_enable(counter);
660
661 /*
662 * Lock the mutex to protect the list of children
663 */
664 mutex_lock(&counter->mutex);
665 list_for_each_entry(child, &counter->child_list, child_list)
666 perf_counter_enable(child);
667 mutex_unlock(&counter->mutex);
668}
669
670void __perf_counter_sched_out(struct perf_counter_context *ctx,
671 struct perf_cpu_context *cpuctx)
672{
673 struct perf_counter *counter;
674 u64 flags;
675
676 spin_lock(&ctx->lock);
677 ctx->is_active = 0;
678 if (likely(!ctx->nr_counters))
679 goto out;
680
681 flags = hw_perf_save_disable();
682 if (ctx->nr_active) {
683 list_for_each_entry(counter, &ctx->counter_list, list_entry)
684 group_sched_out(counter, cpuctx, ctx);
685 }
686 hw_perf_restore(flags);
687 out:
688 spin_unlock(&ctx->lock);
689}
690
691/*
692 * Called from scheduler to remove the counters of the current task,
693 * with interrupts disabled.
694 *
695 * We stop each counter and update the counter value in counter->count.
696 *
697 * This does not protect us against NMI, but disable()
698 * sets the disabled bit in the control field of counter _before_
699 * accessing the counter control register. If a NMI hits, then it will
700 * not restart the counter.
701 */
702void perf_counter_task_sched_out(struct task_struct *task, int cpu)
703{
704 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
705 struct perf_counter_context *ctx = &task->perf_counter_ctx;
706
707 if (likely(!cpuctx->task_ctx))
708 return;
709
710 __perf_counter_sched_out(ctx, cpuctx);
711
712 cpuctx->task_ctx = NULL;
713}
714
715static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
716{
717 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
718}
719
720static int
721group_sched_in(struct perf_counter *group_counter,
722 struct perf_cpu_context *cpuctx,
723 struct perf_counter_context *ctx,
724 int cpu)
725{
726 struct perf_counter *counter, *partial_group;
727 int ret;
728
729 if (group_counter->state == PERF_COUNTER_STATE_OFF)
730 return 0;
731
732 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
733 if (ret)
734 return ret < 0 ? ret : 0;
735
736 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
737 return -EAGAIN;
738
739 /*
740 * Schedule in siblings as one group (if any):
741 */
742 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
743 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
744 partial_group = counter;
745 goto group_error;
746 }
747 }
748
749 return 0;
750
751group_error:
752 /*
753 * Groups can be scheduled in as one unit only, so undo any
754 * partial group before returning:
755 */
756 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
757 if (counter == partial_group)
758 break;
759 counter_sched_out(counter, cpuctx, ctx);
760 }
761 counter_sched_out(group_counter, cpuctx, ctx);
762
763 return -EAGAIN;
764}
765
766static void
767__perf_counter_sched_in(struct perf_counter_context *ctx,
768 struct perf_cpu_context *cpuctx, int cpu)
769{
770 struct perf_counter *counter;
771 u64 flags;
772 int can_add_hw = 1;
773
774 spin_lock(&ctx->lock);
775 ctx->is_active = 1;
776 if (likely(!ctx->nr_counters))
777 goto out;
778
779 flags = hw_perf_save_disable();
780
781 /*
782 * First go through the list and put on any pinned groups
783 * in order to give them the best chance of going on.
784 */
785 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
786 if (counter->state <= PERF_COUNTER_STATE_OFF ||
787 !counter->hw_event.pinned)
788 continue;
789 if (counter->cpu != -1 && counter->cpu != cpu)
790 continue;
791
792 if (group_can_go_on(counter, cpuctx, 1))
793 group_sched_in(counter, cpuctx, ctx, cpu);
794
795 /*
796 * If this pinned group hasn't been scheduled,
797 * put it in error state.
798 */
799 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
800 counter->state = PERF_COUNTER_STATE_ERROR;
801 }
802
803 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
804 /*
805 * Ignore counters in OFF or ERROR state, and
806 * ignore pinned counters since we did them already.
807 */
808 if (counter->state <= PERF_COUNTER_STATE_OFF ||
809 counter->hw_event.pinned)
810 continue;
811
812 /*
813 * Listen to the 'cpu' scheduling filter constraint
814 * of counters:
815 */
816 if (counter->cpu != -1 && counter->cpu != cpu)
817 continue;
818
819 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
820 if (group_sched_in(counter, cpuctx, ctx, cpu))
821 can_add_hw = 0;
822 }
823 }
824 hw_perf_restore(flags);
825 out:
826 spin_unlock(&ctx->lock);
827}
828
829/*
830 * Called from scheduler to add the counters of the current task
831 * with interrupts disabled.
832 *
833 * We restore the counter value and then enable it.
834 *
835 * This does not protect us against NMI, but enable()
836 * sets the enabled bit in the control field of counter _before_
837 * accessing the counter control register. If a NMI hits, then it will
838 * keep the counter running.
839 */
840void perf_counter_task_sched_in(struct task_struct *task, int cpu)
841{
842 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
843 struct perf_counter_context *ctx = &task->perf_counter_ctx;
844
845 __perf_counter_sched_in(ctx, cpuctx, cpu);
846 cpuctx->task_ctx = ctx;
847}
848
849static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
850{
851 struct perf_counter_context *ctx = &cpuctx->ctx;
852
853 __perf_counter_sched_in(ctx, cpuctx, cpu);
854}
855
856int perf_counter_task_disable(void)
857{
858 struct task_struct *curr = current;
859 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
860 struct perf_counter *counter;
861 unsigned long flags;
862 u64 perf_flags;
863 int cpu;
864
865 if (likely(!ctx->nr_counters))
866 return 0;
867
868 curr_rq_lock_irq_save(&flags);
869 cpu = smp_processor_id();
870
871 /* force the update of the task clock: */
872 __task_delta_exec(curr, 1);
873
874 perf_counter_task_sched_out(curr, cpu);
875
876 spin_lock(&ctx->lock);
877
878 /*
879 * Disable all the counters:
880 */
881 perf_flags = hw_perf_save_disable();
882
883 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
884 if (counter->state != PERF_COUNTER_STATE_ERROR)
885 counter->state = PERF_COUNTER_STATE_OFF;
886 }
887
888 hw_perf_restore(perf_flags);
889
890 spin_unlock(&ctx->lock);
891
892 curr_rq_unlock_irq_restore(&flags);
893
894 return 0;
895}
896
897int perf_counter_task_enable(void)
898{
899 struct task_struct *curr = current;
900 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
901 struct perf_counter *counter;
902 unsigned long flags;
903 u64 perf_flags;
904 int cpu;
905
906 if (likely(!ctx->nr_counters))
907 return 0;
908
909 curr_rq_lock_irq_save(&flags);
910 cpu = smp_processor_id();
911
912 /* force the update of the task clock: */
913 __task_delta_exec(curr, 1);
914
915 perf_counter_task_sched_out(curr, cpu);
916
917 spin_lock(&ctx->lock);
918
919 /*
920 * Disable all the counters:
921 */
922 perf_flags = hw_perf_save_disable();
923
924 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
925 if (counter->state > PERF_COUNTER_STATE_OFF)
926 continue;
927 counter->state = PERF_COUNTER_STATE_INACTIVE;
928 counter->hw_event.disabled = 0;
929 }
930 hw_perf_restore(perf_flags);
931
932 spin_unlock(&ctx->lock);
933
934 perf_counter_task_sched_in(curr, cpu);
935
936 curr_rq_unlock_irq_restore(&flags);
937
938 return 0;
939}
940
941/*
942 * Round-robin a context's counters:
943 */
944static void rotate_ctx(struct perf_counter_context *ctx)
945{
946 struct perf_counter *counter;
947 u64 perf_flags;
948
949 if (!ctx->nr_counters)
950 return;
951
952 spin_lock(&ctx->lock);
953 /*
954 * Rotate the first entry last (works just fine for group counters too):
955 */
956 perf_flags = hw_perf_save_disable();
957 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
958 list_del(&counter->list_entry);
959 list_add_tail(&counter->list_entry, &ctx->counter_list);
960 break;
961 }
962 hw_perf_restore(perf_flags);
963
964 spin_unlock(&ctx->lock);
965}
966
967void perf_counter_task_tick(struct task_struct *curr, int cpu)
968{
969 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
970 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
971 const int rotate_percpu = 0;
972
973 if (rotate_percpu)
974 perf_counter_cpu_sched_out(cpuctx);
975 perf_counter_task_sched_out(curr, cpu);
976
977 if (rotate_percpu)
978 rotate_ctx(&cpuctx->ctx);
979 rotate_ctx(ctx);
980
981 if (rotate_percpu)
982 perf_counter_cpu_sched_in(cpuctx, cpu);
983 perf_counter_task_sched_in(curr, cpu);
984}
985
986/*
987 * Cross CPU call to read the hardware counter
988 */
989static void __read(void *info)
990{
991 struct perf_counter *counter = info;
992 unsigned long flags;
993
994 curr_rq_lock_irq_save(&flags);
995 counter->hw_ops->read(counter);
996 curr_rq_unlock_irq_restore(&flags);
997}
998
999static u64 perf_counter_read(struct perf_counter *counter)
1000{
1001 /*
1002 * If counter is enabled and currently active on a CPU, update the
1003 * value in the counter structure:
1004 */
1005 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1006 smp_call_function_single(counter->oncpu,
1007 __read, counter, 1);
1008 }
1009
1010 return atomic64_read(&counter->count);
1011}
1012
1013/*
1014 * Cross CPU call to switch performance data pointers
1015 */
1016static void __perf_switch_irq_data(void *info)
1017{
1018 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1019 struct perf_counter *counter = info;
1020 struct perf_counter_context *ctx = counter->ctx;
1021 struct perf_data *oldirqdata = counter->irqdata;
1022
1023 /*
1024 * If this is a task context, we need to check whether it is
1025 * the current task context of this cpu. If not it has been
1026 * scheduled out before the smp call arrived.
1027 */
1028 if (ctx->task) {
1029 if (cpuctx->task_ctx != ctx)
1030 return;
1031 spin_lock(&ctx->lock);
1032 }
1033
1034 /* Change the pointer NMI safe */
1035 atomic_long_set((atomic_long_t *)&counter->irqdata,
1036 (unsigned long) counter->usrdata);
1037 counter->usrdata = oldirqdata;
1038
1039 if (ctx->task)
1040 spin_unlock(&ctx->lock);
1041}
1042
1043static struct perf_data *perf_switch_irq_data(struct perf_counter *counter)
1044{
1045 struct perf_counter_context *ctx = counter->ctx;
1046 struct perf_data *oldirqdata = counter->irqdata;
1047 struct task_struct *task = ctx->task;
1048
1049 if (!task) {
1050 smp_call_function_single(counter->cpu,
1051 __perf_switch_irq_data,
1052 counter, 1);
1053 return counter->usrdata;
1054 }
1055
1056retry:
1057 spin_lock_irq(&ctx->lock);
1058 if (counter->state != PERF_COUNTER_STATE_ACTIVE) {
1059 counter->irqdata = counter->usrdata;
1060 counter->usrdata = oldirqdata;
1061 spin_unlock_irq(&ctx->lock);
1062 return oldirqdata;
1063 }
1064 spin_unlock_irq(&ctx->lock);
1065 task_oncpu_function_call(task, __perf_switch_irq_data, counter);
1066 /* Might have failed, because task was scheduled out */
1067 if (counter->irqdata == oldirqdata)
1068 goto retry;
1069
1070 return counter->usrdata;
1071}
1072
1073static void put_context(struct perf_counter_context *ctx)
1074{
1075 if (ctx->task)
1076 put_task_struct(ctx->task);
1077}
1078
1079static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1080{
1081 struct perf_cpu_context *cpuctx;
1082 struct perf_counter_context *ctx;
1083 struct task_struct *task;
1084
1085 /*
1086 * If cpu is not a wildcard then this is a percpu counter:
1087 */
1088 if (cpu != -1) {
1089 /* Must be root to operate on a CPU counter: */
1090 if (!capable(CAP_SYS_ADMIN))
1091 return ERR_PTR(-EACCES);
1092
1093 if (cpu < 0 || cpu > num_possible_cpus())
1094 return ERR_PTR(-EINVAL);
1095
1096 /*
1097 * We could be clever and allow to attach a counter to an
1098 * offline CPU and activate it when the CPU comes up, but
1099 * that's for later.
1100 */
1101 if (!cpu_isset(cpu, cpu_online_map))
1102 return ERR_PTR(-ENODEV);
1103
1104 cpuctx = &per_cpu(perf_cpu_context, cpu);
1105 ctx = &cpuctx->ctx;
1106
1107 return ctx;
1108 }
1109
1110 rcu_read_lock();
1111 if (!pid)
1112 task = current;
1113 else
1114 task = find_task_by_vpid(pid);
1115 if (task)
1116 get_task_struct(task);
1117 rcu_read_unlock();
1118
1119 if (!task)
1120 return ERR_PTR(-ESRCH);
1121
1122 ctx = &task->perf_counter_ctx;
1123 ctx->task = task;
1124
1125 /* Reuse ptrace permission checks for now. */
1126 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1127 put_context(ctx);
1128 return ERR_PTR(-EACCES);
1129 }
1130
1131 return ctx;
1132}
1133
1134/*
1135 * Called when the last reference to the file is gone.
1136 */
1137static int perf_release(struct inode *inode, struct file *file)
1138{
1139 struct perf_counter *counter = file->private_data;
1140 struct perf_counter_context *ctx = counter->ctx;
1141
1142 file->private_data = NULL;
1143
1144 mutex_lock(&ctx->mutex);
1145 mutex_lock(&counter->mutex);
1146
1147 perf_counter_remove_from_context(counter);
1148
1149 mutex_unlock(&counter->mutex);
1150 mutex_unlock(&ctx->mutex);
1151
1152 kfree(counter);
1153 put_context(ctx);
1154
1155 return 0;
1156}
1157
1158/*
1159 * Read the performance counter - simple non blocking version for now
1160 */
1161static ssize_t
1162perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1163{
1164 u64 cntval;
1165
1166 if (count != sizeof(cntval))
1167 return -EINVAL;
1168
1169 /*
1170 * Return end-of-file for a read on a counter that is in
1171 * error state (i.e. because it was pinned but it couldn't be
1172 * scheduled on to the CPU at some point).
1173 */
1174 if (counter->state == PERF_COUNTER_STATE_ERROR)
1175 return 0;
1176
1177 mutex_lock(&counter->mutex);
1178 cntval = perf_counter_read(counter);
1179 mutex_unlock(&counter->mutex);
1180
1181 return put_user(cntval, (u64 __user *) buf) ? -EFAULT : sizeof(cntval);
1182}
1183
1184static ssize_t
1185perf_copy_usrdata(struct perf_data *usrdata, char __user *buf, size_t count)
1186{
1187 if (!usrdata->len)
1188 return 0;
1189
1190 count = min(count, (size_t)usrdata->len);
1191 if (copy_to_user(buf, usrdata->data + usrdata->rd_idx, count))
1192 return -EFAULT;
1193
1194 /* Adjust the counters */
1195 usrdata->len -= count;
1196 if (!usrdata->len)
1197 usrdata->rd_idx = 0;
1198 else
1199 usrdata->rd_idx += count;
1200
1201 return count;
1202}
1203
1204static ssize_t
1205perf_read_irq_data(struct perf_counter *counter,
1206 char __user *buf,
1207 size_t count,
1208 int nonblocking)
1209{
1210 struct perf_data *irqdata, *usrdata;
1211 DECLARE_WAITQUEUE(wait, current);
1212 ssize_t res, res2;
1213
1214 irqdata = counter->irqdata;
1215 usrdata = counter->usrdata;
1216
1217 if (usrdata->len + irqdata->len >= count)
1218 goto read_pending;
1219
1220 if (nonblocking)
1221 return -EAGAIN;
1222
1223 spin_lock_irq(&counter->waitq.lock);
1224 __add_wait_queue(&counter->waitq, &wait);
1225 for (;;) {
1226 set_current_state(TASK_INTERRUPTIBLE);
1227 if (usrdata->len + irqdata->len >= count)
1228 break;
1229
1230 if (signal_pending(current))
1231 break;
1232
1233 if (counter->state == PERF_COUNTER_STATE_ERROR)
1234 break;
1235
1236 spin_unlock_irq(&counter->waitq.lock);
1237 schedule();
1238 spin_lock_irq(&counter->waitq.lock);
1239 }
1240 __remove_wait_queue(&counter->waitq, &wait);
1241 __set_current_state(TASK_RUNNING);
1242 spin_unlock_irq(&counter->waitq.lock);
1243
1244 if (usrdata->len + irqdata->len < count &&
1245 counter->state != PERF_COUNTER_STATE_ERROR)
1246 return -ERESTARTSYS;
1247read_pending:
1248 mutex_lock(&counter->mutex);
1249
1250 /* Drain pending data first: */
1251 res = perf_copy_usrdata(usrdata, buf, count);
1252 if (res < 0 || res == count)
1253 goto out;
1254
1255 /* Switch irq buffer: */
1256 usrdata = perf_switch_irq_data(counter);
1257 res2 = perf_copy_usrdata(usrdata, buf + res, count - res);
1258 if (res2 < 0) {
1259 if (!res)
1260 res = -EFAULT;
1261 } else {
1262 res += res2;
1263 }
1264out:
1265 mutex_unlock(&counter->mutex);
1266
1267 return res;
1268}
1269
1270static ssize_t
1271perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1272{
1273 struct perf_counter *counter = file->private_data;
1274
1275 switch (counter->hw_event.record_type) {
1276 case PERF_RECORD_SIMPLE:
1277 return perf_read_hw(counter, buf, count);
1278
1279 case PERF_RECORD_IRQ:
1280 case PERF_RECORD_GROUP:
1281 return perf_read_irq_data(counter, buf, count,
1282 file->f_flags & O_NONBLOCK);
1283 }
1284 return -EINVAL;
1285}
1286
1287static unsigned int perf_poll(struct file *file, poll_table *wait)
1288{
1289 struct perf_counter *counter = file->private_data;
1290 unsigned int events = 0;
1291 unsigned long flags;
1292
1293 poll_wait(file, &counter->waitq, wait);
1294
1295 spin_lock_irqsave(&counter->waitq.lock, flags);
1296 if (counter->usrdata->len || counter->irqdata->len)
1297 events |= POLLIN;
1298 spin_unlock_irqrestore(&counter->waitq.lock, flags);
1299
1300 return events;
1301}
1302
1303static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1304{
1305 struct perf_counter *counter = file->private_data;
1306 int err = 0;
1307
1308 switch (cmd) {
1309 case PERF_COUNTER_IOC_ENABLE:
1310 perf_counter_enable_family(counter);
1311 break;
1312 case PERF_COUNTER_IOC_DISABLE:
1313 perf_counter_disable_family(counter);
1314 break;
1315 default:
1316 err = -ENOTTY;
1317 }
1318 return err;
1319}
1320
1321static const struct file_operations perf_fops = {
1322 .release = perf_release,
1323 .read = perf_read,
1324 .poll = perf_poll,
1325 .unlocked_ioctl = perf_ioctl,
1326 .compat_ioctl = perf_ioctl,
1327};
1328
1329static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
1330{
1331 int cpu = raw_smp_processor_id();
1332
1333 atomic64_set(&counter->hw.prev_count, cpu_clock(cpu));
1334 return 0;
1335}
1336
1337static void cpu_clock_perf_counter_update(struct perf_counter *counter)
1338{
1339 int cpu = raw_smp_processor_id();
1340 s64 prev;
1341 u64 now;
1342
1343 now = cpu_clock(cpu);
1344 prev = atomic64_read(&counter->hw.prev_count);
1345 atomic64_set(&counter->hw.prev_count, now);
1346 atomic64_add(now - prev, &counter->count);
1347}
1348
1349static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
1350{
1351 cpu_clock_perf_counter_update(counter);
1352}
1353
1354static void cpu_clock_perf_counter_read(struct perf_counter *counter)
1355{
1356 cpu_clock_perf_counter_update(counter);
1357}
1358
1359static const struct hw_perf_counter_ops perf_ops_cpu_clock = {
1360 .enable = cpu_clock_perf_counter_enable,
1361 .disable = cpu_clock_perf_counter_disable,
1362 .read = cpu_clock_perf_counter_read,
1363};
1364
1365/*
1366 * Called from within the scheduler:
1367 */
1368static u64 task_clock_perf_counter_val(struct perf_counter *counter, int update)
1369{
1370 struct task_struct *curr = counter->task;
1371 u64 delta;
1372
1373 delta = __task_delta_exec(curr, update);
1374
1375 return curr->se.sum_exec_runtime + delta;
1376}
1377
1378static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
1379{
1380 u64 prev;
1381 s64 delta;
1382
1383 prev = atomic64_read(&counter->hw.prev_count);
1384
1385 atomic64_set(&counter->hw.prev_count, now);
1386
1387 delta = now - prev;
1388
1389 atomic64_add(delta, &counter->count);
1390}
1391
1392static void task_clock_perf_counter_read(struct perf_counter *counter)
1393{
1394 u64 now = task_clock_perf_counter_val(counter, 1);
1395
1396 task_clock_perf_counter_update(counter, now);
1397}
1398
1399static int task_clock_perf_counter_enable(struct perf_counter *counter)
1400{
1401 u64 now = task_clock_perf_counter_val(counter, 0);
1402
1403 atomic64_set(&counter->hw.prev_count, now);
1404
1405 return 0;
1406}
1407
1408static void task_clock_perf_counter_disable(struct perf_counter *counter)
1409{
1410 u64 now = task_clock_perf_counter_val(counter, 0);
1411
1412 task_clock_perf_counter_update(counter, now);
1413}
1414
1415static const struct hw_perf_counter_ops perf_ops_task_clock = {
1416 .enable = task_clock_perf_counter_enable,
1417 .disable = task_clock_perf_counter_disable,
1418 .read = task_clock_perf_counter_read,
1419};
1420
1421#ifdef CONFIG_VM_EVENT_COUNTERS
1422#define cpu_page_faults() __get_cpu_var(vm_event_states).event[PGFAULT]
1423#else
1424#define cpu_page_faults() 0
1425#endif
1426
1427static u64 get_page_faults(struct perf_counter *counter)
1428{
1429 struct task_struct *curr = counter->ctx->task;
1430
1431 if (curr)
1432 return curr->maj_flt + curr->min_flt;
1433 return cpu_page_faults();
1434}
1435
1436static void page_faults_perf_counter_update(struct perf_counter *counter)
1437{
1438 u64 prev, now;
1439 s64 delta;
1440
1441 prev = atomic64_read(&counter->hw.prev_count);
1442 now = get_page_faults(counter);
1443
1444 atomic64_set(&counter->hw.prev_count, now);
1445
1446 delta = now - prev;
1447
1448 atomic64_add(delta, &counter->count);
1449}
1450
1451static void page_faults_perf_counter_read(struct perf_counter *counter)
1452{
1453 page_faults_perf_counter_update(counter);
1454}
1455
1456static int page_faults_perf_counter_enable(struct perf_counter *counter)
1457{
1458 atomic64_set(&counter->hw.prev_count, get_page_faults(counter));
1459 return 0;
1460}
1461
1462static void page_faults_perf_counter_disable(struct perf_counter *counter)
1463{
1464 page_faults_perf_counter_update(counter);
1465}
1466
1467static const struct hw_perf_counter_ops perf_ops_page_faults = {
1468 .enable = page_faults_perf_counter_enable,
1469 .disable = page_faults_perf_counter_disable,
1470 .read = page_faults_perf_counter_read,
1471};
1472
1473static u64 get_context_switches(struct perf_counter *counter)
1474{
1475 struct task_struct *curr = counter->ctx->task;
1476
1477 if (curr)
1478 return curr->nvcsw + curr->nivcsw;
1479 return cpu_nr_switches(smp_processor_id());
1480}
1481
1482static void context_switches_perf_counter_update(struct perf_counter *counter)
1483{
1484 u64 prev, now;
1485 s64 delta;
1486
1487 prev = atomic64_read(&counter->hw.prev_count);
1488 now = get_context_switches(counter);
1489
1490 atomic64_set(&counter->hw.prev_count, now);
1491
1492 delta = now - prev;
1493
1494 atomic64_add(delta, &counter->count);
1495}
1496
1497static void context_switches_perf_counter_read(struct perf_counter *counter)
1498{
1499 context_switches_perf_counter_update(counter);
1500}
1501
1502static int context_switches_perf_counter_enable(struct perf_counter *counter)
1503{
1504 atomic64_set(&counter->hw.prev_count, get_context_switches(counter));
1505 return 0;
1506}
1507
1508static void context_switches_perf_counter_disable(struct perf_counter *counter)
1509{
1510 context_switches_perf_counter_update(counter);
1511}
1512
1513static const struct hw_perf_counter_ops perf_ops_context_switches = {
1514 .enable = context_switches_perf_counter_enable,
1515 .disable = context_switches_perf_counter_disable,
1516 .read = context_switches_perf_counter_read,
1517};
1518
1519static inline u64 get_cpu_migrations(struct perf_counter *counter)
1520{
1521 struct task_struct *curr = counter->ctx->task;
1522
1523 if (curr)
1524 return curr->se.nr_migrations;
1525 return cpu_nr_migrations(smp_processor_id());
1526}
1527
1528static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
1529{
1530 u64 prev, now;
1531 s64 delta;
1532
1533 prev = atomic64_read(&counter->hw.prev_count);
1534 now = get_cpu_migrations(counter);
1535
1536 atomic64_set(&counter->hw.prev_count, now);
1537
1538 delta = now - prev;
1539
1540 atomic64_add(delta, &counter->count);
1541}
1542
1543static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
1544{
1545 cpu_migrations_perf_counter_update(counter);
1546}
1547
1548static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
1549{
1550 atomic64_set(&counter->hw.prev_count, get_cpu_migrations(counter));
1551 return 0;
1552}
1553
1554static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
1555{
1556 cpu_migrations_perf_counter_update(counter);
1557}
1558
1559static const struct hw_perf_counter_ops perf_ops_cpu_migrations = {
1560 .enable = cpu_migrations_perf_counter_enable,
1561 .disable = cpu_migrations_perf_counter_disable,
1562 .read = cpu_migrations_perf_counter_read,
1563};
1564
1565static const struct hw_perf_counter_ops *
1566sw_perf_counter_init(struct perf_counter *counter)
1567{
1568 const struct hw_perf_counter_ops *hw_ops = NULL;
1569
1570 /*
1571 * Software counters (currently) can't in general distinguish
1572 * between user, kernel and hypervisor events.
1573 * However, context switches and cpu migrations are considered
1574 * to be kernel events, and page faults are never hypervisor
1575 * events.
1576 */
1577 switch (counter->hw_event.type) {
1578 case PERF_COUNT_CPU_CLOCK:
1579 if (!(counter->hw_event.exclude_user ||
1580 counter->hw_event.exclude_kernel ||
1581 counter->hw_event.exclude_hv))
1582 hw_ops = &perf_ops_cpu_clock;
1583 break;
1584 case PERF_COUNT_TASK_CLOCK:
1585 if (counter->hw_event.exclude_user ||
1586 counter->hw_event.exclude_kernel ||
1587 counter->hw_event.exclude_hv)
1588 break;
1589 /*
1590 * If the user instantiates this as a per-cpu counter,
1591 * use the cpu_clock counter instead.
1592 */
1593 if (counter->ctx->task)
1594 hw_ops = &perf_ops_task_clock;
1595 else
1596 hw_ops = &perf_ops_cpu_clock;
1597 break;
1598 case PERF_COUNT_PAGE_FAULTS:
1599 if (!(counter->hw_event.exclude_user ||
1600 counter->hw_event.exclude_kernel))
1601 hw_ops = &perf_ops_page_faults;
1602 break;
1603 case PERF_COUNT_CONTEXT_SWITCHES:
1604 if (!counter->hw_event.exclude_kernel)
1605 hw_ops = &perf_ops_context_switches;
1606 break;
1607 case PERF_COUNT_CPU_MIGRATIONS:
1608 if (!counter->hw_event.exclude_kernel)
1609 hw_ops = &perf_ops_cpu_migrations;
1610 break;
1611 default:
1612 break;
1613 }
1614 return hw_ops;
1615}
1616
1617/*
1618 * Allocate and initialize a counter structure
1619 */
1620static struct perf_counter *
1621perf_counter_alloc(struct perf_counter_hw_event *hw_event,
1622 int cpu,
1623 struct perf_counter_context *ctx,
1624 struct perf_counter *group_leader,
1625 gfp_t gfpflags)
1626{
1627 const struct hw_perf_counter_ops *hw_ops;
1628 struct perf_counter *counter;
1629
1630 counter = kzalloc(sizeof(*counter), gfpflags);
1631 if (!counter)
1632 return NULL;
1633
1634 /*
1635 * Single counters are their own group leaders, with an
1636 * empty sibling list:
1637 */
1638 if (!group_leader)
1639 group_leader = counter;
1640
1641 mutex_init(&counter->mutex);
1642 INIT_LIST_HEAD(&counter->list_entry);
1643 INIT_LIST_HEAD(&counter->sibling_list);
1644 init_waitqueue_head(&counter->waitq);
1645
1646 INIT_LIST_HEAD(&counter->child_list);
1647
1648 counter->irqdata = &counter->data[0];
1649 counter->usrdata = &counter->data[1];
1650 counter->cpu = cpu;
1651 counter->hw_event = *hw_event;
1652 counter->wakeup_pending = 0;
1653 counter->group_leader = group_leader;
1654 counter->hw_ops = NULL;
1655 counter->ctx = ctx;
1656
1657 counter->state = PERF_COUNTER_STATE_INACTIVE;
1658 if (hw_event->disabled)
1659 counter->state = PERF_COUNTER_STATE_OFF;
1660
1661 hw_ops = NULL;
1662 if (!hw_event->raw && hw_event->type < 0)
1663 hw_ops = sw_perf_counter_init(counter);
1664 else
1665 hw_ops = hw_perf_counter_init(counter);
1666
1667 if (!hw_ops) {
1668 kfree(counter);
1669 return NULL;
1670 }
1671 counter->hw_ops = hw_ops;
1672
1673 return counter;
1674}
1675
1676/**
1677 * sys_perf_task_open - open a performance counter, associate it to a task/cpu
1678 *
1679 * @hw_event_uptr: event type attributes for monitoring/sampling
1680 * @pid: target pid
1681 * @cpu: target cpu
1682 * @group_fd: group leader counter fd
1683 */
1684asmlinkage int
1685sys_perf_counter_open(struct perf_counter_hw_event *hw_event_uptr __user,
1686 pid_t pid, int cpu, int group_fd)
1687{
1688 struct perf_counter *counter, *group_leader;
1689 struct perf_counter_hw_event hw_event;
1690 struct perf_counter_context *ctx;
1691 struct file *counter_file = NULL;
1692 struct file *group_file = NULL;
1693 int fput_needed = 0;
1694 int fput_needed2 = 0;
1695 int ret;
1696
1697 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
1698 return -EFAULT;
1699
1700 /*
1701 * Get the target context (task or percpu):
1702 */
1703 ctx = find_get_context(pid, cpu);
1704 if (IS_ERR(ctx))
1705 return PTR_ERR(ctx);
1706
1707 /*
1708 * Look up the group leader (we will attach this counter to it):
1709 */
1710 group_leader = NULL;
1711 if (group_fd != -1) {
1712 ret = -EINVAL;
1713 group_file = fget_light(group_fd, &fput_needed);
1714 if (!group_file)
1715 goto err_put_context;
1716 if (group_file->f_op != &perf_fops)
1717 goto err_put_context;
1718
1719 group_leader = group_file->private_data;
1720 /*
1721 * Do not allow a recursive hierarchy (this new sibling
1722 * becoming part of another group-sibling):
1723 */
1724 if (group_leader->group_leader != group_leader)
1725 goto err_put_context;
1726 /*
1727 * Do not allow to attach to a group in a different
1728 * task or CPU context:
1729 */
1730 if (group_leader->ctx != ctx)
1731 goto err_put_context;
1732 /*
1733 * Only a group leader can be exclusive or pinned
1734 */
1735 if (hw_event.exclusive || hw_event.pinned)
1736 goto err_put_context;
1737 }
1738
1739 ret = -EINVAL;
1740 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
1741 GFP_KERNEL);
1742 if (!counter)
1743 goto err_put_context;
1744
1745 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
1746 if (ret < 0)
1747 goto err_free_put_context;
1748
1749 counter_file = fget_light(ret, &fput_needed2);
1750 if (!counter_file)
1751 goto err_free_put_context;
1752
1753 counter->filp = counter_file;
1754 mutex_lock(&ctx->mutex);
1755 perf_install_in_context(ctx, counter, cpu);
1756 mutex_unlock(&ctx->mutex);
1757
1758 fput_light(counter_file, fput_needed2);
1759
1760out_fput:
1761 fput_light(group_file, fput_needed);
1762
1763 return ret;
1764
1765err_free_put_context:
1766 kfree(counter);
1767
1768err_put_context:
1769 put_context(ctx);
1770
1771 goto out_fput;
1772}
1773
1774/*
1775 * Initialize the perf_counter context in a task_struct:
1776 */
1777static void
1778__perf_counter_init_context(struct perf_counter_context *ctx,
1779 struct task_struct *task)
1780{
1781 memset(ctx, 0, sizeof(*ctx));
1782 spin_lock_init(&ctx->lock);
1783 mutex_init(&ctx->mutex);
1784 INIT_LIST_HEAD(&ctx->counter_list);
1785 ctx->task = task;
1786}
1787
1788/*
1789 * inherit a counter from parent task to child task:
1790 */
1791static struct perf_counter *
1792inherit_counter(struct perf_counter *parent_counter,
1793 struct task_struct *parent,
1794 struct perf_counter_context *parent_ctx,
1795 struct task_struct *child,
1796 struct perf_counter *group_leader,
1797 struct perf_counter_context *child_ctx)
1798{
1799 struct perf_counter *child_counter;
1800
1801 /*
1802 * Instead of creating recursive hierarchies of counters,
1803 * we link inherited counters back to the original parent,
1804 * which has a filp for sure, which we use as the reference
1805 * count:
1806 */
1807 if (parent_counter->parent)
1808 parent_counter = parent_counter->parent;
1809
1810 child_counter = perf_counter_alloc(&parent_counter->hw_event,
1811 parent_counter->cpu, child_ctx,
1812 group_leader, GFP_KERNEL);
1813 if (!child_counter)
1814 return NULL;
1815
1816 /*
1817 * Link it up in the child's context:
1818 */
1819 child_counter->task = child;
1820 list_add_counter(child_counter, child_ctx);
1821 child_ctx->nr_counters++;
1822
1823 child_counter->parent = parent_counter;
1824 /*
1825 * inherit into child's child as well:
1826 */
1827 child_counter->hw_event.inherit = 1;
1828
1829 /*
1830 * Get a reference to the parent filp - we will fput it
1831 * when the child counter exits. This is safe to do because
1832 * we are in the parent and we know that the filp still
1833 * exists and has a nonzero count:
1834 */
1835 atomic_long_inc(&parent_counter->filp->f_count);
1836
1837 /*
1838 * Link this into the parent counter's child list
1839 */
1840 mutex_lock(&parent_counter->mutex);
1841 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
1842
1843 /*
1844 * Make the child state follow the state of the parent counter,
1845 * not its hw_event.disabled bit. We hold the parent's mutex,
1846 * so we won't race with perf_counter_{en,dis}able_family.
1847 */
1848 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
1849 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
1850 else
1851 child_counter->state = PERF_COUNTER_STATE_OFF;
1852
1853 mutex_unlock(&parent_counter->mutex);
1854
1855 return child_counter;
1856}
1857
1858static int inherit_group(struct perf_counter *parent_counter,
1859 struct task_struct *parent,
1860 struct perf_counter_context *parent_ctx,
1861 struct task_struct *child,
1862 struct perf_counter_context *child_ctx)
1863{
1864 struct perf_counter *leader;
1865 struct perf_counter *sub;
1866
1867 leader = inherit_counter(parent_counter, parent, parent_ctx,
1868 child, NULL, child_ctx);
1869 if (!leader)
1870 return -ENOMEM;
1871 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
1872 if (!inherit_counter(sub, parent, parent_ctx,
1873 child, leader, child_ctx))
1874 return -ENOMEM;
1875 }
1876 return 0;
1877}
1878
1879static void sync_child_counter(struct perf_counter *child_counter,
1880 struct perf_counter *parent_counter)
1881{
1882 u64 parent_val, child_val;
1883
1884 parent_val = atomic64_read(&parent_counter->count);
1885 child_val = atomic64_read(&child_counter->count);
1886
1887 /*
1888 * Add back the child's count to the parent's count:
1889 */
1890 atomic64_add(child_val, &parent_counter->count);
1891
1892 /*
1893 * Remove this counter from the parent's list
1894 */
1895 mutex_lock(&parent_counter->mutex);
1896 list_del_init(&child_counter->child_list);
1897 mutex_unlock(&parent_counter->mutex);
1898
1899 /*
1900 * Release the parent counter, if this was the last
1901 * reference to it.
1902 */
1903 fput(parent_counter->filp);
1904}
1905
1906static void
1907__perf_counter_exit_task(struct task_struct *child,
1908 struct perf_counter *child_counter,
1909 struct perf_counter_context *child_ctx)
1910{
1911 struct perf_counter *parent_counter;
1912 struct perf_counter *sub, *tmp;
1913
1914 /*
1915 * If we do not self-reap then we have to wait for the
1916 * child task to unschedule (it will happen for sure),
1917 * so that its counter is at its final count. (This
1918 * condition triggers rarely - child tasks usually get
1919 * off their CPU before the parent has a chance to
1920 * get this far into the reaping action)
1921 */
1922 if (child != current) {
1923 wait_task_inactive(child, 0);
1924 list_del_init(&child_counter->list_entry);
1925 } else {
1926 struct perf_cpu_context *cpuctx;
1927 unsigned long flags;
1928 u64 perf_flags;
1929
1930 /*
1931 * Disable and unlink this counter.
1932 *
1933 * Be careful about zapping the list - IRQ/NMI context
1934 * could still be processing it:
1935 */
1936 curr_rq_lock_irq_save(&flags);
1937 perf_flags = hw_perf_save_disable();
1938
1939 cpuctx = &__get_cpu_var(perf_cpu_context);
1940
1941 group_sched_out(child_counter, cpuctx, child_ctx);
1942
1943 list_del_init(&child_counter->list_entry);
1944
1945 child_ctx->nr_counters--;
1946
1947 hw_perf_restore(perf_flags);
1948 curr_rq_unlock_irq_restore(&flags);
1949 }
1950
1951 parent_counter = child_counter->parent;
1952 /*
1953 * It can happen that parent exits first, and has counters
1954 * that are still around due to the child reference. These
1955 * counters need to be zapped - but otherwise linger.
1956 */
1957 if (parent_counter) {
1958 sync_child_counter(child_counter, parent_counter);
1959 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
1960 list_entry) {
1961 if (sub->parent) {
1962 sync_child_counter(sub, sub->parent);
1963 kfree(sub);
1964 }
1965 }
1966 kfree(child_counter);
1967 }
1968}
1969
1970/*
1971 * When a child task exits, feed back counter values to parent counters.
1972 *
1973 * Note: we may be running in child context, but the PID is not hashed
1974 * anymore so new counters will not be added.
1975 */
1976void perf_counter_exit_task(struct task_struct *child)
1977{
1978 struct perf_counter *child_counter, *tmp;
1979 struct perf_counter_context *child_ctx;
1980
1981 child_ctx = &child->perf_counter_ctx;
1982
1983 if (likely(!child_ctx->nr_counters))
1984 return;
1985
1986 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
1987 list_entry)
1988 __perf_counter_exit_task(child, child_counter, child_ctx);
1989}
1990
1991/*
1992 * Initialize the perf_counter context in task_struct
1993 */
1994void perf_counter_init_task(struct task_struct *child)
1995{
1996 struct perf_counter_context *child_ctx, *parent_ctx;
1997 struct perf_counter *counter;
1998 struct task_struct *parent = current;
1999
2000 child_ctx = &child->perf_counter_ctx;
2001 parent_ctx = &parent->perf_counter_ctx;
2002
2003 __perf_counter_init_context(child_ctx, child);
2004
2005 /*
2006 * This is executed from the parent task context, so inherit
2007 * counters that have been marked for cloning:
2008 */
2009
2010 if (likely(!parent_ctx->nr_counters))
2011 return;
2012
2013 /*
2014 * Lock the parent list. No need to lock the child - not PID
2015 * hashed yet and not running, so nobody can access it.
2016 */
2017 mutex_lock(&parent_ctx->mutex);
2018
2019 /*
2020 * We dont have to disable NMIs - we are only looking at
2021 * the list, not manipulating it:
2022 */
2023 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
2024 if (!counter->hw_event.inherit)
2025 continue;
2026
2027 if (inherit_group(counter, parent,
2028 parent_ctx, child, child_ctx))
2029 break;
2030 }
2031
2032 mutex_unlock(&parent_ctx->mutex);
2033}
2034
2035static void __cpuinit perf_counter_init_cpu(int cpu)
2036{
2037 struct perf_cpu_context *cpuctx;
2038
2039 cpuctx = &per_cpu(perf_cpu_context, cpu);
2040 __perf_counter_init_context(&cpuctx->ctx, NULL);
2041
2042 mutex_lock(&perf_resource_mutex);
2043 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
2044 mutex_unlock(&perf_resource_mutex);
2045
2046 hw_perf_counter_setup(cpu);
2047}
2048
2049#ifdef CONFIG_HOTPLUG_CPU
2050static void __perf_counter_exit_cpu(void *info)
2051{
2052 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
2053 struct perf_counter_context *ctx = &cpuctx->ctx;
2054 struct perf_counter *counter, *tmp;
2055
2056 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
2057 __perf_counter_remove_from_context(counter);
2058}
2059static void perf_counter_exit_cpu(int cpu)
2060{
2061 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
2062 struct perf_counter_context *ctx = &cpuctx->ctx;
2063
2064 mutex_lock(&ctx->mutex);
2065 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
2066 mutex_unlock(&ctx->mutex);
2067}
2068#else
2069static inline void perf_counter_exit_cpu(int cpu) { }
2070#endif
2071
2072static int __cpuinit
2073perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
2074{
2075 unsigned int cpu = (long)hcpu;
2076
2077 switch (action) {
2078
2079 case CPU_UP_PREPARE:
2080 case CPU_UP_PREPARE_FROZEN:
2081 perf_counter_init_cpu(cpu);
2082 break;
2083
2084 case CPU_DOWN_PREPARE:
2085 case CPU_DOWN_PREPARE_FROZEN:
2086 perf_counter_exit_cpu(cpu);
2087 break;
2088
2089 default:
2090 break;
2091 }
2092
2093 return NOTIFY_OK;
2094}
2095
2096static struct notifier_block __cpuinitdata perf_cpu_nb = {
2097 .notifier_call = perf_cpu_notify,
2098};
2099
2100static int __init perf_counter_init(void)
2101{
2102 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
2103 (void *)(long)smp_processor_id());
2104 register_cpu_notifier(&perf_cpu_nb);
2105
2106 return 0;
2107}
2108early_initcall(perf_counter_init);
2109
2110static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
2111{
2112 return sprintf(buf, "%d\n", perf_reserved_percpu);
2113}
2114
2115static ssize_t
2116perf_set_reserve_percpu(struct sysdev_class *class,
2117 const char *buf,
2118 size_t count)
2119{
2120 struct perf_cpu_context *cpuctx;
2121 unsigned long val;
2122 int err, cpu, mpt;
2123
2124 err = strict_strtoul(buf, 10, &val);
2125 if (err)
2126 return err;
2127 if (val > perf_max_counters)
2128 return -EINVAL;
2129
2130 mutex_lock(&perf_resource_mutex);
2131 perf_reserved_percpu = val;
2132 for_each_online_cpu(cpu) {
2133 cpuctx = &per_cpu(perf_cpu_context, cpu);
2134 spin_lock_irq(&cpuctx->ctx.lock);
2135 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
2136 perf_max_counters - perf_reserved_percpu);
2137 cpuctx->max_pertask = mpt;
2138 spin_unlock_irq(&cpuctx->ctx.lock);
2139 }
2140 mutex_unlock(&perf_resource_mutex);
2141
2142 return count;
2143}
2144
2145static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
2146{
2147 return sprintf(buf, "%d\n", perf_overcommit);
2148}
2149
2150static ssize_t
2151perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
2152{
2153 unsigned long val;
2154 int err;
2155
2156 err = strict_strtoul(buf, 10, &val);
2157 if (err)
2158 return err;
2159 if (val > 1)
2160 return -EINVAL;
2161
2162 mutex_lock(&perf_resource_mutex);
2163 perf_overcommit = val;
2164 mutex_unlock(&perf_resource_mutex);
2165
2166 return count;
2167}
2168
2169static SYSDEV_CLASS_ATTR(
2170 reserve_percpu,
2171 0644,
2172 perf_show_reserve_percpu,
2173 perf_set_reserve_percpu
2174 );
2175
2176static SYSDEV_CLASS_ATTR(
2177 overcommit,
2178 0644,
2179 perf_show_overcommit,
2180 perf_set_overcommit
2181 );
2182
2183static struct attribute *perfclass_attrs[] = {
2184 &attr_reserve_percpu.attr,
2185 &attr_overcommit.attr,
2186 NULL
2187};
2188
2189static struct attribute_group perfclass_attr_group = {
2190 .attrs = perfclass_attrs,
2191 .name = "perf_counters",
2192};
2193
2194static int __init perf_counter_sysfs_init(void)
2195{
2196 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
2197 &perfclass_attr_group);
2198}
2199device_initcall(perf_counter_sysfs_init);
diff --git a/kernel/sched.c b/kernel/sched.c
index 61245b8d0f16..83b68ff6df80 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -558,6 +558,7 @@ struct rq {
558 struct load_weight load; 558 struct load_weight load;
559 unsigned long nr_load_updates; 559 unsigned long nr_load_updates;
560 u64 nr_switches; 560 u64 nr_switches;
561 u64 nr_migrations_in;
561 562
562 struct cfs_rq cfs; 563 struct cfs_rq cfs;
563 struct rt_rq rt; 564 struct rt_rq rt;
@@ -668,7 +669,7 @@ static inline int cpu_of(struct rq *rq)
668#define task_rq(p) cpu_rq(task_cpu(p)) 669#define task_rq(p) cpu_rq(task_cpu(p))
669#define cpu_curr(cpu) (cpu_rq(cpu)->curr) 670#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
670 671
671static inline void update_rq_clock(struct rq *rq) 672inline void update_rq_clock(struct rq *rq)
672{ 673{
673 rq->clock = sched_clock_cpu(cpu_of(rq)); 674 rq->clock = sched_clock_cpu(cpu_of(rq));
674} 675}
@@ -979,6 +980,26 @@ static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
979 } 980 }
980} 981}
981 982
983void curr_rq_lock_irq_save(unsigned long *flags)
984 __acquires(rq->lock)
985{
986 struct rq *rq;
987
988 local_irq_save(*flags);
989 rq = cpu_rq(smp_processor_id());
990 spin_lock(&rq->lock);
991}
992
993void curr_rq_unlock_irq_restore(unsigned long *flags)
994 __releases(rq->lock)
995{
996 struct rq *rq;
997
998 rq = cpu_rq(smp_processor_id());
999 spin_unlock(&rq->lock);
1000 local_irq_restore(*flags);
1001}
1002
982void task_rq_unlock_wait(struct task_struct *p) 1003void task_rq_unlock_wait(struct task_struct *p)
983{ 1004{
984 struct rq *rq = task_rq(p); 1005 struct rq *rq = task_rq(p);
@@ -1885,12 +1906,15 @@ void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1885 p->se.sleep_start -= clock_offset; 1906 p->se.sleep_start -= clock_offset;
1886 if (p->se.block_start) 1907 if (p->se.block_start)
1887 p->se.block_start -= clock_offset; 1908 p->se.block_start -= clock_offset;
1909#endif
1888 if (old_cpu != new_cpu) { 1910 if (old_cpu != new_cpu) {
1889 schedstat_inc(p, se.nr_migrations); 1911 p->se.nr_migrations++;
1912 new_rq->nr_migrations_in++;
1913#ifdef CONFIG_SCHEDSTATS
1890 if (task_hot(p, old_rq->clock, NULL)) 1914 if (task_hot(p, old_rq->clock, NULL))
1891 schedstat_inc(p, se.nr_forced2_migrations); 1915 schedstat_inc(p, se.nr_forced2_migrations);
1892 }
1893#endif 1916#endif
1917 }
1894 p->se.vruntime -= old_cfsrq->min_vruntime - 1918 p->se.vruntime -= old_cfsrq->min_vruntime -
1895 new_cfsrq->min_vruntime; 1919 new_cfsrq->min_vruntime;
1896 1920
@@ -2242,6 +2266,27 @@ static int sched_balance_self(int cpu, int flag)
2242 2266
2243#endif /* CONFIG_SMP */ 2267#endif /* CONFIG_SMP */
2244 2268
2269/**
2270 * task_oncpu_function_call - call a function on the cpu on which a task runs
2271 * @p: the task to evaluate
2272 * @func: the function to be called
2273 * @info: the function call argument
2274 *
2275 * Calls the function @func when the task is currently running. This might
2276 * be on the current CPU, which just calls the function directly
2277 */
2278void task_oncpu_function_call(struct task_struct *p,
2279 void (*func) (void *info), void *info)
2280{
2281 int cpu;
2282
2283 preempt_disable();
2284 cpu = task_cpu(p);
2285 if (task_curr(p))
2286 smp_call_function_single(cpu, func, info, 1);
2287 preempt_enable();
2288}
2289
2245/*** 2290/***
2246 * try_to_wake_up - wake up a thread 2291 * try_to_wake_up - wake up a thread
2247 * @p: the to-be-woken-up thread 2292 * @p: the to-be-woken-up thread
@@ -2384,6 +2429,7 @@ static void __sched_fork(struct task_struct *p)
2384 p->se.exec_start = 0; 2429 p->se.exec_start = 0;
2385 p->se.sum_exec_runtime = 0; 2430 p->se.sum_exec_runtime = 0;
2386 p->se.prev_sum_exec_runtime = 0; 2431 p->se.prev_sum_exec_runtime = 0;
2432 p->se.nr_migrations = 0;
2387 p->se.last_wakeup = 0; 2433 p->se.last_wakeup = 0;
2388 p->se.avg_overlap = 0; 2434 p->se.avg_overlap = 0;
2389 2435
@@ -2604,6 +2650,7 @@ static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2604 */ 2650 */
2605 prev_state = prev->state; 2651 prev_state = prev->state;
2606 finish_arch_switch(prev); 2652 finish_arch_switch(prev);
2653 perf_counter_task_sched_in(current, cpu_of(rq));
2607 finish_lock_switch(rq, prev); 2654 finish_lock_switch(rq, prev);
2608#ifdef CONFIG_SMP 2655#ifdef CONFIG_SMP
2609 if (current->sched_class->post_schedule) 2656 if (current->sched_class->post_schedule)
@@ -2766,6 +2813,21 @@ unsigned long nr_active(void)
2766} 2813}
2767 2814
2768/* 2815/*
2816 * Externally visible per-cpu scheduler statistics:
2817 * cpu_nr_switches(cpu) - number of context switches on that cpu
2818 * cpu_nr_migrations(cpu) - number of migrations into that cpu
2819 */
2820u64 cpu_nr_switches(int cpu)
2821{
2822 return cpu_rq(cpu)->nr_switches;
2823}
2824
2825u64 cpu_nr_migrations(int cpu)
2826{
2827 return cpu_rq(cpu)->nr_migrations_in;
2828}
2829
2830/*
2769 * Update rq->cpu_load[] statistics. This function is usually called every 2831 * Update rq->cpu_load[] statistics. This function is usually called every
2770 * scheduler tick (TICK_NSEC). 2832 * scheduler tick (TICK_NSEC).
2771 */ 2833 */
@@ -4137,6 +4199,29 @@ EXPORT_PER_CPU_SYMBOL(kstat);
4137 * Return any ns on the sched_clock that have not yet been banked in 4199 * Return any ns on the sched_clock that have not yet been banked in
4138 * @p in case that task is currently running. 4200 * @p in case that task is currently running.
4139 */ 4201 */
4202unsigned long long __task_delta_exec(struct task_struct *p, int update)
4203{
4204 s64 delta_exec;
4205 struct rq *rq;
4206
4207 rq = task_rq(p);
4208 WARN_ON_ONCE(!runqueue_is_locked());
4209 WARN_ON_ONCE(!task_current(rq, p));
4210
4211 if (update)
4212 update_rq_clock(rq);
4213
4214 delta_exec = rq->clock - p->se.exec_start;
4215
4216 WARN_ON_ONCE(delta_exec < 0);
4217
4218 return delta_exec;
4219}
4220
4221/*
4222 * Return any ns on the sched_clock that have not yet been banked in
4223 * @p in case that task is currently running.
4224 */
4140unsigned long long task_delta_exec(struct task_struct *p) 4225unsigned long long task_delta_exec(struct task_struct *p)
4141{ 4226{
4142 unsigned long flags; 4227 unsigned long flags;
@@ -4396,6 +4481,7 @@ void scheduler_tick(void)
4396 update_rq_clock(rq); 4481 update_rq_clock(rq);
4397 update_cpu_load(rq); 4482 update_cpu_load(rq);
4398 curr->sched_class->task_tick(rq, curr, 0); 4483 curr->sched_class->task_tick(rq, curr, 0);
4484 perf_counter_task_tick(curr, cpu);
4399 spin_unlock(&rq->lock); 4485 spin_unlock(&rq->lock);
4400 4486
4401#ifdef CONFIG_SMP 4487#ifdef CONFIG_SMP
@@ -4591,6 +4677,7 @@ need_resched_nonpreemptible:
4591 4677
4592 if (likely(prev != next)) { 4678 if (likely(prev != next)) {
4593 sched_info_switch(prev, next); 4679 sched_info_switch(prev, next);
4680 perf_counter_task_sched_out(prev, cpu);
4594 4681
4595 rq->nr_switches++; 4682 rq->nr_switches++;
4596 rq->curr = next; 4683 rq->curr = next;
diff --git a/kernel/sys.c b/kernel/sys.c
index f145c415bc16..c5e7dec4966e 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>
@@ -1791,6 +1792,12 @@ SYSCALL_DEFINE5(prctl, int, option, unsigned long, arg2, unsigned long, arg3,
1791 case PR_SET_TSC: 1792 case PR_SET_TSC:
1792 error = SET_TSC_CTL(arg2); 1793 error = SET_TSC_CTL(arg2);
1793 break; 1794 break;
1795 case PR_TASK_PERF_COUNTERS_DISABLE:
1796 error = perf_counter_task_disable();
1797 break;
1798 case PR_TASK_PERF_COUNTERS_ENABLE:
1799 error = perf_counter_task_enable();
1800 break;
1794 case PR_GET_TIMERSLACK: 1801 case PR_GET_TIMERSLACK:
1795 error = current->timer_slack_ns; 1802 error = current->timer_slack_ns;
1796 break; 1803 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 2025-03-01 23:14:42 -0500