]>
Commit | Line | Data |
---|---|---|
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 | * | |
8 | * For licensing details see kernel-base/COPYING | |
9 | */ | |
10 | ||
11 | #include <linux/fs.h> | |
12 | #include <linux/mm.h> | |
13 | #include <linux/cpu.h> | |
14 | #include <linux/smp.h> | |
15 | #include <linux/file.h> | |
16 | #include <linux/poll.h> | |
17 | #include <linux/sysfs.h> | |
18 | #include <linux/ptrace.h> | |
19 | #include <linux/percpu.h> | |
20 | #include <linux/vmstat.h> | |
21 | #include <linux/hardirq.h> | |
22 | #include <linux/rculist.h> | |
23 | #include <linux/uaccess.h> | |
24 | #include <linux/syscalls.h> | |
25 | #include <linux/anon_inodes.h> | |
26 | #include <linux/kernel_stat.h> | |
27 | #include <linux/perf_counter.h> | |
28 | ||
29 | #include <asm/irq_regs.h> | |
30 | ||
31 | /* | |
32 | * Each CPU has a list of per CPU counters: | |
33 | */ | |
34 | DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context); | |
35 | ||
36 | int perf_max_counters __read_mostly = 1; | |
37 | static int perf_reserved_percpu __read_mostly; | |
38 | static int perf_overcommit __read_mostly = 1; | |
39 | ||
40 | /* | |
41 | * Mutex for (sysadmin-configurable) counter reservations: | |
42 | */ | |
43 | static DEFINE_MUTEX(perf_resource_mutex); | |
44 | ||
45 | /* | |
46 | * Architecture provided APIs - weak aliases: | |
47 | */ | |
48 | extern __weak const struct hw_perf_counter_ops * | |
49 | hw_perf_counter_init(struct perf_counter *counter) | |
50 | { | |
51 | return NULL; | |
52 | } | |
53 | ||
54 | u64 __weak hw_perf_save_disable(void) { return 0; } | |
55 | void __weak hw_perf_restore(u64 ctrl) { barrier(); } | |
56 | void __weak hw_perf_counter_setup(int cpu) { barrier(); } | |
57 | int __weak hw_perf_group_sched_in(struct perf_counter *group_leader, | |
58 | struct perf_cpu_context *cpuctx, | |
59 | struct perf_counter_context *ctx, int cpu) | |
60 | { | |
61 | return 0; | |
62 | } | |
63 | ||
64 | void __weak perf_counter_print_debug(void) { } | |
65 | ||
66 | static void | |
67 | list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx) | |
68 | { | |
69 | struct perf_counter *group_leader = counter->group_leader; | |
70 | ||
71 | /* | |
72 | * Depending on whether it is a standalone or sibling counter, | |
73 | * add it straight to the context's counter list, or to the group | |
74 | * leader's sibling list: | |
75 | */ | |
76 | if (counter->group_leader == counter) | |
77 | list_add_tail(&counter->list_entry, &ctx->counter_list); | |
78 | else | |
79 | list_add_tail(&counter->list_entry, &group_leader->sibling_list); | |
80 | ||
81 | list_add_rcu(&counter->event_entry, &ctx->event_list); | |
82 | } | |
83 | ||
84 | static void | |
85 | list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx) | |
86 | { | |
87 | struct perf_counter *sibling, *tmp; | |
88 | ||
89 | list_del_init(&counter->list_entry); | |
90 | list_del_rcu(&counter->event_entry); | |
91 | ||
92 | /* | |
93 | * If this was a group counter with sibling counters then | |
94 | * upgrade the siblings to singleton counters by adding them | |
95 | * to the context list directly: | |
96 | */ | |
97 | list_for_each_entry_safe(sibling, tmp, | |
98 | &counter->sibling_list, list_entry) { | |
99 | ||
100 | list_move_tail(&sibling->list_entry, &ctx->counter_list); | |
101 | sibling->group_leader = sibling; | |
102 | } | |
103 | } | |
104 | ||
105 | static void | |
106 | counter_sched_out(struct perf_counter *counter, | |
107 | struct perf_cpu_context *cpuctx, | |
108 | struct perf_counter_context *ctx) | |
109 | { | |
110 | if (counter->state != PERF_COUNTER_STATE_ACTIVE) | |
111 | return; | |
112 | ||
113 | counter->state = PERF_COUNTER_STATE_INACTIVE; | |
114 | counter->hw_ops->disable(counter); | |
115 | counter->oncpu = -1; | |
116 | ||
117 | if (!is_software_counter(counter)) | |
118 | cpuctx->active_oncpu--; | |
119 | ctx->nr_active--; | |
120 | if (counter->hw_event.exclusive || !cpuctx->active_oncpu) | |
121 | cpuctx->exclusive = 0; | |
122 | } | |
123 | ||
124 | static void | |
125 | group_sched_out(struct perf_counter *group_counter, | |
126 | struct perf_cpu_context *cpuctx, | |
127 | struct perf_counter_context *ctx) | |
128 | { | |
129 | struct perf_counter *counter; | |
130 | ||
131 | if (group_counter->state != PERF_COUNTER_STATE_ACTIVE) | |
132 | return; | |
133 | ||
134 | counter_sched_out(group_counter, cpuctx, ctx); | |
135 | ||
136 | /* | |
137 | * Schedule out siblings (if any): | |
138 | */ | |
139 | list_for_each_entry(counter, &group_counter->sibling_list, list_entry) | |
140 | counter_sched_out(counter, cpuctx, ctx); | |
141 | ||
142 | if (group_counter->hw_event.exclusive) | |
143 | cpuctx->exclusive = 0; | |
144 | } | |
145 | ||
146 | /* | |
147 | * Cross CPU call to remove a performance counter | |
148 | * | |
149 | * We disable the counter on the hardware level first. After that we | |
150 | * remove it from the context list. | |
151 | */ | |
152 | static void __perf_counter_remove_from_context(void *info) | |
153 | { | |
154 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
155 | struct perf_counter *counter = info; | |
156 | struct perf_counter_context *ctx = counter->ctx; | |
157 | unsigned long flags; | |
158 | u64 perf_flags; | |
159 | ||
160 | /* | |
161 | * If this is a task context, we need to check whether it is | |
162 | * the current task context of this cpu. If not it has been | |
163 | * scheduled out before the smp call arrived. | |
164 | */ | |
165 | if (ctx->task && cpuctx->task_ctx != ctx) | |
166 | return; | |
167 | ||
168 | curr_rq_lock_irq_save(&flags); | |
169 | spin_lock(&ctx->lock); | |
170 | ||
171 | counter_sched_out(counter, cpuctx, ctx); | |
172 | ||
173 | counter->task = NULL; | |
174 | ctx->nr_counters--; | |
175 | ||
176 | /* | |
177 | * Protect the list operation against NMI by disabling the | |
178 | * counters on a global level. NOP for non NMI based counters. | |
179 | */ | |
180 | perf_flags = hw_perf_save_disable(); | |
181 | list_del_counter(counter, ctx); | |
182 | hw_perf_restore(perf_flags); | |
183 | ||
184 | if (!ctx->task) { | |
185 | /* | |
186 | * Allow more per task counters with respect to the | |
187 | * reservation: | |
188 | */ | |
189 | cpuctx->max_pertask = | |
190 | min(perf_max_counters - ctx->nr_counters, | |
191 | perf_max_counters - perf_reserved_percpu); | |
192 | } | |
193 | ||
194 | spin_unlock(&ctx->lock); | |
195 | curr_rq_unlock_irq_restore(&flags); | |
196 | } | |
197 | ||
198 | ||
199 | /* | |
200 | * Remove the counter from a task's (or a CPU's) list of counters. | |
201 | * | |
202 | * Must be called with counter->mutex and ctx->mutex held. | |
203 | * | |
204 | * CPU counters are removed with a smp call. For task counters we only | |
205 | * call when the task is on a CPU. | |
206 | */ | |
207 | static void perf_counter_remove_from_context(struct perf_counter *counter) | |
208 | { | |
209 | struct perf_counter_context *ctx = counter->ctx; | |
210 | struct task_struct *task = ctx->task; | |
211 | ||
212 | if (!task) { | |
213 | /* | |
214 | * Per cpu counters are removed via an smp call and | |
215 | * the removal is always sucessful. | |
216 | */ | |
217 | smp_call_function_single(counter->cpu, | |
218 | __perf_counter_remove_from_context, | |
219 | counter, 1); | |
220 | return; | |
221 | } | |
222 | ||
223 | retry: | |
224 | task_oncpu_function_call(task, __perf_counter_remove_from_context, | |
225 | counter); | |
226 | ||
227 | spin_lock_irq(&ctx->lock); | |
228 | /* | |
229 | * If the context is active we need to retry the smp call. | |
230 | */ | |
231 | if (ctx->nr_active && !list_empty(&counter->list_entry)) { | |
232 | spin_unlock_irq(&ctx->lock); | |
233 | goto retry; | |
234 | } | |
235 | ||
236 | /* | |
237 | * The lock prevents that this context is scheduled in so we | |
238 | * can remove the counter safely, if the call above did not | |
239 | * succeed. | |
240 | */ | |
241 | if (!list_empty(&counter->list_entry)) { | |
242 | ctx->nr_counters--; | |
243 | list_del_counter(counter, ctx); | |
244 | counter->task = NULL; | |
245 | } | |
246 | spin_unlock_irq(&ctx->lock); | |
247 | } | |
248 | ||
249 | /* | |
250 | * Cross CPU call to disable a performance counter | |
251 | */ | |
252 | static void __perf_counter_disable(void *info) | |
253 | { | |
254 | struct perf_counter *counter = info; | |
255 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
256 | struct perf_counter_context *ctx = counter->ctx; | |
257 | unsigned long flags; | |
258 | ||
259 | /* | |
260 | * If this is a per-task counter, need to check whether this | |
261 | * counter's task is the current task on this cpu. | |
262 | */ | |
263 | if (ctx->task && cpuctx->task_ctx != ctx) | |
264 | return; | |
265 | ||
266 | curr_rq_lock_irq_save(&flags); | |
267 | spin_lock(&ctx->lock); | |
268 | ||
269 | /* | |
270 | * If the counter is on, turn it off. | |
271 | * If it is in error state, leave it in error state. | |
272 | */ | |
273 | if (counter->state >= PERF_COUNTER_STATE_INACTIVE) { | |
274 | if (counter == counter->group_leader) | |
275 | group_sched_out(counter, cpuctx, ctx); | |
276 | else | |
277 | counter_sched_out(counter, cpuctx, ctx); | |
278 | counter->state = PERF_COUNTER_STATE_OFF; | |
279 | } | |
280 | ||
281 | spin_unlock(&ctx->lock); | |
282 | curr_rq_unlock_irq_restore(&flags); | |
283 | } | |
284 | ||
285 | /* | |
286 | * Disable a counter. | |
287 | */ | |
288 | static void perf_counter_disable(struct perf_counter *counter) | |
289 | { | |
290 | struct perf_counter_context *ctx = counter->ctx; | |
291 | struct task_struct *task = ctx->task; | |
292 | ||
293 | if (!task) { | |
294 | /* | |
295 | * Disable the counter on the cpu that it's on | |
296 | */ | |
297 | smp_call_function_single(counter->cpu, __perf_counter_disable, | |
298 | counter, 1); | |
299 | return; | |
300 | } | |
301 | ||
302 | retry: | |
303 | task_oncpu_function_call(task, __perf_counter_disable, counter); | |
304 | ||
305 | spin_lock_irq(&ctx->lock); | |
306 | /* | |
307 | * If the counter is still active, we need to retry the cross-call. | |
308 | */ | |
309 | if (counter->state == PERF_COUNTER_STATE_ACTIVE) { | |
310 | spin_unlock_irq(&ctx->lock); | |
311 | goto retry; | |
312 | } | |
313 | ||
314 | /* | |
315 | * Since we have the lock this context can't be scheduled | |
316 | * in, so we can change the state safely. | |
317 | */ | |
318 | if (counter->state == PERF_COUNTER_STATE_INACTIVE) | |
319 | counter->state = PERF_COUNTER_STATE_OFF; | |
320 | ||
321 | spin_unlock_irq(&ctx->lock); | |
322 | } | |
323 | ||
324 | /* | |
325 | * Disable a counter and all its children. | |
326 | */ | |
327 | static void perf_counter_disable_family(struct perf_counter *counter) | |
328 | { | |
329 | struct perf_counter *child; | |
330 | ||
331 | perf_counter_disable(counter); | |
332 | ||
333 | /* | |
334 | * Lock the mutex to protect the list of children | |
335 | */ | |
336 | mutex_lock(&counter->mutex); | |
337 | list_for_each_entry(child, &counter->child_list, child_list) | |
338 | perf_counter_disable(child); | |
339 | mutex_unlock(&counter->mutex); | |
340 | } | |
341 | ||
342 | static int | |
343 | counter_sched_in(struct perf_counter *counter, | |
344 | struct perf_cpu_context *cpuctx, | |
345 | struct perf_counter_context *ctx, | |
346 | int cpu) | |
347 | { | |
348 | if (counter->state <= PERF_COUNTER_STATE_OFF) | |
349 | return 0; | |
350 | ||
351 | counter->state = PERF_COUNTER_STATE_ACTIVE; | |
352 | counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */ | |
353 | /* | |
354 | * The new state must be visible before we turn it on in the hardware: | |
355 | */ | |
356 | smp_wmb(); | |
357 | ||
358 | if (counter->hw_ops->enable(counter)) { | |
359 | counter->state = PERF_COUNTER_STATE_INACTIVE; | |
360 | counter->oncpu = -1; | |
361 | return -EAGAIN; | |
362 | } | |
363 | ||
364 | if (!is_software_counter(counter)) | |
365 | cpuctx->active_oncpu++; | |
366 | ctx->nr_active++; | |
367 | ||
368 | if (counter->hw_event.exclusive) | |
369 | cpuctx->exclusive = 1; | |
370 | ||
371 | return 0; | |
372 | } | |
373 | ||
374 | /* | |
375 | * Return 1 for a group consisting entirely of software counters, | |
376 | * 0 if the group contains any hardware counters. | |
377 | */ | |
378 | static int is_software_only_group(struct perf_counter *leader) | |
379 | { | |
380 | struct perf_counter *counter; | |
381 | ||
382 | if (!is_software_counter(leader)) | |
383 | return 0; | |
384 | list_for_each_entry(counter, &leader->sibling_list, list_entry) | |
385 | if (!is_software_counter(counter)) | |
386 | return 0; | |
387 | return 1; | |
388 | } | |
389 | ||
390 | /* | |
391 | * Work out whether we can put this counter group on the CPU now. | |
392 | */ | |
393 | static int group_can_go_on(struct perf_counter *counter, | |
394 | struct perf_cpu_context *cpuctx, | |
395 | int can_add_hw) | |
396 | { | |
397 | /* | |
398 | * Groups consisting entirely of software counters can always go on. | |
399 | */ | |
400 | if (is_software_only_group(counter)) | |
401 | return 1; | |
402 | /* | |
403 | * If an exclusive group is already on, no other hardware | |
404 | * counters can go on. | |
405 | */ | |
406 | if (cpuctx->exclusive) | |
407 | return 0; | |
408 | /* | |
409 | * If this group is exclusive and there are already | |
410 | * counters on the CPU, it can't go on. | |
411 | */ | |
412 | if (counter->hw_event.exclusive && cpuctx->active_oncpu) | |
413 | return 0; | |
414 | /* | |
415 | * Otherwise, try to add it if all previous groups were able | |
416 | * to go on. | |
417 | */ | |
418 | return can_add_hw; | |
419 | } | |
420 | ||
421 | /* | |
422 | * Cross CPU call to install and enable a performance counter | |
423 | */ | |
424 | static void __perf_install_in_context(void *info) | |
425 | { | |
426 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
427 | struct perf_counter *counter = info; | |
428 | struct perf_counter_context *ctx = counter->ctx; | |
429 | struct perf_counter *leader = counter->group_leader; | |
430 | int cpu = smp_processor_id(); | |
431 | unsigned long flags; | |
432 | u64 perf_flags; | |
433 | int err; | |
434 | ||
435 | /* | |
436 | * If this is a task context, we need to check whether it is | |
437 | * the current task context of this cpu. If not it has been | |
438 | * scheduled out before the smp call arrived. | |
439 | */ | |
440 | if (ctx->task && cpuctx->task_ctx != ctx) | |
441 | return; | |
442 | ||
443 | curr_rq_lock_irq_save(&flags); | |
444 | spin_lock(&ctx->lock); | |
445 | ||
446 | /* | |
447 | * Protect the list operation against NMI by disabling the | |
448 | * counters on a global level. NOP for non NMI based counters. | |
449 | */ | |
450 | perf_flags = hw_perf_save_disable(); | |
451 | ||
452 | list_add_counter(counter, ctx); | |
453 | ctx->nr_counters++; | |
454 | counter->prev_state = PERF_COUNTER_STATE_OFF; | |
455 | ||
456 | /* | |
457 | * Don't put the counter on if it is disabled or if | |
458 | * it is in a group and the group isn't on. | |
459 | */ | |
460 | if (counter->state != PERF_COUNTER_STATE_INACTIVE || | |
461 | (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)) | |
462 | goto unlock; | |
463 | ||
464 | /* | |
465 | * An exclusive counter can't go on if there are already active | |
466 | * hardware counters, and no hardware counter can go on if there | |
467 | * is already an exclusive counter on. | |
468 | */ | |
469 | if (!group_can_go_on(counter, cpuctx, 1)) | |
470 | err = -EEXIST; | |
471 | else | |
472 | err = counter_sched_in(counter, cpuctx, ctx, cpu); | |
473 | ||
474 | if (err) { | |
475 | /* | |
476 | * This counter couldn't go on. If it is in a group | |
477 | * then we have to pull the whole group off. | |
478 | * If the counter group is pinned then put it in error state. | |
479 | */ | |
480 | if (leader != counter) | |
481 | group_sched_out(leader, cpuctx, ctx); | |
482 | if (leader->hw_event.pinned) | |
483 | leader->state = PERF_COUNTER_STATE_ERROR; | |
484 | } | |
485 | ||
486 | if (!err && !ctx->task && cpuctx->max_pertask) | |
487 | cpuctx->max_pertask--; | |
488 | ||
489 | unlock: | |
490 | hw_perf_restore(perf_flags); | |
491 | ||
492 | spin_unlock(&ctx->lock); | |
493 | curr_rq_unlock_irq_restore(&flags); | |
494 | } | |
495 | ||
496 | /* | |
497 | * Attach a performance counter to a context | |
498 | * | |
499 | * First we add the counter to the list with the hardware enable bit | |
500 | * in counter->hw_config cleared. | |
501 | * | |
502 | * If the counter is attached to a task which is on a CPU we use a smp | |
503 | * call to enable it in the task context. The task might have been | |
504 | * scheduled away, but we check this in the smp call again. | |
505 | * | |
506 | * Must be called with ctx->mutex held. | |
507 | */ | |
508 | static void | |
509 | perf_install_in_context(struct perf_counter_context *ctx, | |
510 | struct perf_counter *counter, | |
511 | int cpu) | |
512 | { | |
513 | struct task_struct *task = ctx->task; | |
514 | ||
515 | if (!task) { | |
516 | /* | |
517 | * Per cpu counters are installed via an smp call and | |
518 | * the install is always sucessful. | |
519 | */ | |
520 | smp_call_function_single(cpu, __perf_install_in_context, | |
521 | counter, 1); | |
522 | return; | |
523 | } | |
524 | ||
525 | counter->task = task; | |
526 | retry: | |
527 | task_oncpu_function_call(task, __perf_install_in_context, | |
528 | counter); | |
529 | ||
530 | spin_lock_irq(&ctx->lock); | |
531 | /* | |
532 | * we need to retry the smp call. | |
533 | */ | |
534 | if (ctx->is_active && list_empty(&counter->list_entry)) { | |
535 | spin_unlock_irq(&ctx->lock); | |
536 | goto retry; | |
537 | } | |
538 | ||
539 | /* | |
540 | * The lock prevents that this context is scheduled in so we | |
541 | * can add the counter safely, if it the call above did not | |
542 | * succeed. | |
543 | */ | |
544 | if (list_empty(&counter->list_entry)) { | |
545 | list_add_counter(counter, ctx); | |
546 | ctx->nr_counters++; | |
547 | } | |
548 | spin_unlock_irq(&ctx->lock); | |
549 | } | |
550 | ||
551 | /* | |
552 | * Cross CPU call to enable a performance counter | |
553 | */ | |
554 | static void __perf_counter_enable(void *info) | |
555 | { | |
556 | struct perf_counter *counter = info; | |
557 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
558 | struct perf_counter_context *ctx = counter->ctx; | |
559 | struct perf_counter *leader = counter->group_leader; | |
560 | unsigned long flags; | |
561 | int err; | |
562 | ||
563 | /* | |
564 | * If this is a per-task counter, need to check whether this | |
565 | * counter's task is the current task on this cpu. | |
566 | */ | |
567 | if (ctx->task && cpuctx->task_ctx != ctx) | |
568 | return; | |
569 | ||
570 | curr_rq_lock_irq_save(&flags); | |
571 | spin_lock(&ctx->lock); | |
572 | ||
573 | counter->prev_state = counter->state; | |
574 | if (counter->state >= PERF_COUNTER_STATE_INACTIVE) | |
575 | goto unlock; | |
576 | counter->state = PERF_COUNTER_STATE_INACTIVE; | |
577 | ||
578 | /* | |
579 | * If the counter is in a group and isn't the group leader, | |
580 | * then don't put it on unless the group is on. | |
581 | */ | |
582 | if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE) | |
583 | goto unlock; | |
584 | ||
585 | if (!group_can_go_on(counter, cpuctx, 1)) | |
586 | err = -EEXIST; | |
587 | else | |
588 | err = counter_sched_in(counter, cpuctx, ctx, | |
589 | smp_processor_id()); | |
590 | ||
591 | if (err) { | |
592 | /* | |
593 | * If this counter can't go on and it's part of a | |
594 | * group, then the whole group has to come off. | |
595 | */ | |
596 | if (leader != counter) | |
597 | group_sched_out(leader, cpuctx, ctx); | |
598 | if (leader->hw_event.pinned) | |
599 | leader->state = PERF_COUNTER_STATE_ERROR; | |
600 | } | |
601 | ||
602 | unlock: | |
603 | spin_unlock(&ctx->lock); | |
604 | curr_rq_unlock_irq_restore(&flags); | |
605 | } | |
606 | ||
607 | /* | |
608 | * Enable a counter. | |
609 | */ | |
610 | static void perf_counter_enable(struct perf_counter *counter) | |
611 | { | |
612 | struct perf_counter_context *ctx = counter->ctx; | |
613 | struct task_struct *task = ctx->task; | |
614 | ||
615 | if (!task) { | |
616 | /* | |
617 | * Enable the counter on the cpu that it's on | |
618 | */ | |
619 | smp_call_function_single(counter->cpu, __perf_counter_enable, | |
620 | counter, 1); | |
621 | return; | |
622 | } | |
623 | ||
624 | spin_lock_irq(&ctx->lock); | |
625 | if (counter->state >= PERF_COUNTER_STATE_INACTIVE) | |
626 | goto out; | |
627 | ||
628 | /* | |
629 | * If the counter is in error state, clear that first. | |
630 | * That way, if we see the counter in error state below, we | |
631 | * know that it has gone back into error state, as distinct | |
632 | * from the task having been scheduled away before the | |
633 | * cross-call arrived. | |
634 | */ | |
635 | if (counter->state == PERF_COUNTER_STATE_ERROR) | |
636 | counter->state = PERF_COUNTER_STATE_OFF; | |
637 | ||
638 | retry: | |
639 | spin_unlock_irq(&ctx->lock); | |
640 | task_oncpu_function_call(task, __perf_counter_enable, counter); | |
641 | ||
642 | spin_lock_irq(&ctx->lock); | |
643 | ||
644 | /* | |
645 | * If the context is active and the counter is still off, | |
646 | * we need to retry the cross-call. | |
647 | */ | |
648 | if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF) | |
649 | goto retry; | |
650 | ||
651 | /* | |
652 | * Since we have the lock this context can't be scheduled | |
653 | * in, so we can change the state safely. | |
654 | */ | |
655 | if (counter->state == PERF_COUNTER_STATE_OFF) | |
656 | counter->state = PERF_COUNTER_STATE_INACTIVE; | |
657 | out: | |
658 | spin_unlock_irq(&ctx->lock); | |
659 | } | |
660 | ||
661 | /* | |
662 | * Enable a counter and all its children. | |
663 | */ | |
664 | static void perf_counter_enable_family(struct perf_counter *counter) | |
665 | { | |
666 | struct perf_counter *child; | |
667 | ||
668 | perf_counter_enable(counter); | |
669 | ||
670 | /* | |
671 | * Lock the mutex to protect the list of children | |
672 | */ | |
673 | mutex_lock(&counter->mutex); | |
674 | list_for_each_entry(child, &counter->child_list, child_list) | |
675 | perf_counter_enable(child); | |
676 | mutex_unlock(&counter->mutex); | |
677 | } | |
678 | ||
679 | void __perf_counter_sched_out(struct perf_counter_context *ctx, | |
680 | struct perf_cpu_context *cpuctx) | |
681 | { | |
682 | struct perf_counter *counter; | |
683 | u64 flags; | |
684 | ||
685 | spin_lock(&ctx->lock); | |
686 | ctx->is_active = 0; | |
687 | if (likely(!ctx->nr_counters)) | |
688 | goto out; | |
689 | ||
690 | flags = hw_perf_save_disable(); | |
691 | if (ctx->nr_active) { | |
692 | list_for_each_entry(counter, &ctx->counter_list, list_entry) | |
693 | group_sched_out(counter, cpuctx, ctx); | |
694 | } | |
695 | hw_perf_restore(flags); | |
696 | out: | |
697 | spin_unlock(&ctx->lock); | |
698 | } | |
699 | ||
700 | /* | |
701 | * Called from scheduler to remove the counters of the current task, | |
702 | * with interrupts disabled. | |
703 | * | |
704 | * We stop each counter and update the counter value in counter->count. | |
705 | * | |
706 | * This does not protect us against NMI, but disable() | |
707 | * sets the disabled bit in the control field of counter _before_ | |
708 | * accessing the counter control register. If a NMI hits, then it will | |
709 | * not restart the counter. | |
710 | */ | |
711 | void perf_counter_task_sched_out(struct task_struct *task, int cpu) | |
712 | { | |
713 | struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); | |
714 | struct perf_counter_context *ctx = &task->perf_counter_ctx; | |
715 | struct pt_regs *regs; | |
716 | ||
717 | if (likely(!cpuctx->task_ctx)) | |
718 | return; | |
719 | ||
720 | regs = task_pt_regs(task); | |
721 | perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs); | |
722 | __perf_counter_sched_out(ctx, cpuctx); | |
723 | ||
724 | cpuctx->task_ctx = NULL; | |
725 | } | |
726 | ||
727 | static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx) | |
728 | { | |
729 | __perf_counter_sched_out(&cpuctx->ctx, cpuctx); | |
730 | } | |
731 | ||
732 | static int | |
733 | group_sched_in(struct perf_counter *group_counter, | |
734 | struct perf_cpu_context *cpuctx, | |
735 | struct perf_counter_context *ctx, | |
736 | int cpu) | |
737 | { | |
738 | struct perf_counter *counter, *partial_group; | |
739 | int ret; | |
740 | ||
741 | if (group_counter->state == PERF_COUNTER_STATE_OFF) | |
742 | return 0; | |
743 | ||
744 | ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu); | |
745 | if (ret) | |
746 | return ret < 0 ? ret : 0; | |
747 | ||
748 | group_counter->prev_state = group_counter->state; | |
749 | if (counter_sched_in(group_counter, cpuctx, ctx, cpu)) | |
750 | return -EAGAIN; | |
751 | ||
752 | /* | |
753 | * Schedule in siblings as one group (if any): | |
754 | */ | |
755 | list_for_each_entry(counter, &group_counter->sibling_list, list_entry) { | |
756 | counter->prev_state = counter->state; | |
757 | if (counter_sched_in(counter, cpuctx, ctx, cpu)) { | |
758 | partial_group = counter; | |
759 | goto group_error; | |
760 | } | |
761 | } | |
762 | ||
763 | return 0; | |
764 | ||
765 | group_error: | |
766 | /* | |
767 | * Groups can be scheduled in as one unit only, so undo any | |
768 | * partial group before returning: | |
769 | */ | |
770 | list_for_each_entry(counter, &group_counter->sibling_list, list_entry) { | |
771 | if (counter == partial_group) | |
772 | break; | |
773 | counter_sched_out(counter, cpuctx, ctx); | |
774 | } | |
775 | counter_sched_out(group_counter, cpuctx, ctx); | |
776 | ||
777 | return -EAGAIN; | |
778 | } | |
779 | ||
780 | static void | |
781 | __perf_counter_sched_in(struct perf_counter_context *ctx, | |
782 | struct perf_cpu_context *cpuctx, int cpu) | |
783 | { | |
784 | struct perf_counter *counter; | |
785 | u64 flags; | |
786 | int can_add_hw = 1; | |
787 | ||
788 | spin_lock(&ctx->lock); | |
789 | ctx->is_active = 1; | |
790 | if (likely(!ctx->nr_counters)) | |
791 | goto out; | |
792 | ||
793 | flags = hw_perf_save_disable(); | |
794 | ||
795 | /* | |
796 | * First go through the list and put on any pinned groups | |
797 | * in order to give them the best chance of going on. | |
798 | */ | |
799 | list_for_each_entry(counter, &ctx->counter_list, list_entry) { | |
800 | if (counter->state <= PERF_COUNTER_STATE_OFF || | |
801 | !counter->hw_event.pinned) | |
802 | continue; | |
803 | if (counter->cpu != -1 && counter->cpu != cpu) | |
804 | continue; | |
805 | ||
806 | if (group_can_go_on(counter, cpuctx, 1)) | |
807 | group_sched_in(counter, cpuctx, ctx, cpu); | |
808 | ||
809 | /* | |
810 | * If this pinned group hasn't been scheduled, | |
811 | * put it in error state. | |
812 | */ | |
813 | if (counter->state == PERF_COUNTER_STATE_INACTIVE) | |
814 | counter->state = PERF_COUNTER_STATE_ERROR; | |
815 | } | |
816 | ||
817 | list_for_each_entry(counter, &ctx->counter_list, list_entry) { | |
818 | /* | |
819 | * Ignore counters in OFF or ERROR state, and | |
820 | * ignore pinned counters since we did them already. | |
821 | */ | |
822 | if (counter->state <= PERF_COUNTER_STATE_OFF || | |
823 | counter->hw_event.pinned) | |
824 | continue; | |
825 | ||
826 | /* | |
827 | * Listen to the 'cpu' scheduling filter constraint | |
828 | * of counters: | |
829 | */ | |
830 | if (counter->cpu != -1 && counter->cpu != cpu) | |
831 | continue; | |
832 | ||
833 | if (group_can_go_on(counter, cpuctx, can_add_hw)) { | |
834 | if (group_sched_in(counter, cpuctx, ctx, cpu)) | |
835 | can_add_hw = 0; | |
836 | } | |
837 | } | |
838 | hw_perf_restore(flags); | |
839 | out: | |
840 | spin_unlock(&ctx->lock); | |
841 | } | |
842 | ||
843 | /* | |
844 | * Called from scheduler to add the counters of the current task | |
845 | * with interrupts disabled. | |
846 | * | |
847 | * We restore the counter value and then enable it. | |
848 | * | |
849 | * This does not protect us against NMI, but enable() | |
850 | * sets the enabled bit in the control field of counter _before_ | |
851 | * accessing the counter control register. If a NMI hits, then it will | |
852 | * keep the counter running. | |
853 | */ | |
854 | void perf_counter_task_sched_in(struct task_struct *task, int cpu) | |
855 | { | |
856 | struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); | |
857 | struct perf_counter_context *ctx = &task->perf_counter_ctx; | |
858 | ||
859 | __perf_counter_sched_in(ctx, cpuctx, cpu); | |
860 | cpuctx->task_ctx = ctx; | |
861 | } | |
862 | ||
863 | static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu) | |
864 | { | |
865 | struct perf_counter_context *ctx = &cpuctx->ctx; | |
866 | ||
867 | __perf_counter_sched_in(ctx, cpuctx, cpu); | |
868 | } | |
869 | ||
870 | int perf_counter_task_disable(void) | |
871 | { | |
872 | struct task_struct *curr = current; | |
873 | struct perf_counter_context *ctx = &curr->perf_counter_ctx; | |
874 | struct perf_counter *counter; | |
875 | unsigned long flags; | |
876 | u64 perf_flags; | |
877 | int cpu; | |
878 | ||
879 | if (likely(!ctx->nr_counters)) | |
880 | return 0; | |
881 | ||
882 | curr_rq_lock_irq_save(&flags); | |
883 | cpu = smp_processor_id(); | |
884 | ||
885 | /* force the update of the task clock: */ | |
886 | __task_delta_exec(curr, 1); | |
887 | ||
888 | perf_counter_task_sched_out(curr, cpu); | |
889 | ||
890 | spin_lock(&ctx->lock); | |
891 | ||
892 | /* | |
893 | * Disable all the counters: | |
894 | */ | |
895 | perf_flags = hw_perf_save_disable(); | |
896 | ||
897 | list_for_each_entry(counter, &ctx->counter_list, list_entry) { | |
898 | if (counter->state != PERF_COUNTER_STATE_ERROR) | |
899 | counter->state = PERF_COUNTER_STATE_OFF; | |
900 | } | |
901 | ||
902 | hw_perf_restore(perf_flags); | |
903 | ||
904 | spin_unlock(&ctx->lock); | |
905 | ||
906 | curr_rq_unlock_irq_restore(&flags); | |
907 | ||
908 | return 0; | |
909 | } | |
910 | ||
911 | int perf_counter_task_enable(void) | |
912 | { | |
913 | struct task_struct *curr = current; | |
914 | struct perf_counter_context *ctx = &curr->perf_counter_ctx; | |
915 | struct perf_counter *counter; | |
916 | unsigned long flags; | |
917 | u64 perf_flags; | |
918 | int cpu; | |
919 | ||
920 | if (likely(!ctx->nr_counters)) | |
921 | return 0; | |
922 | ||
923 | curr_rq_lock_irq_save(&flags); | |
924 | cpu = smp_processor_id(); | |
925 | ||
926 | /* force the update of the task clock: */ | |
927 | __task_delta_exec(curr, 1); | |
928 | ||
929 | perf_counter_task_sched_out(curr, cpu); | |
930 | ||
931 | spin_lock(&ctx->lock); | |
932 | ||
933 | /* | |
934 | * Disable all the counters: | |
935 | */ | |
936 | perf_flags = hw_perf_save_disable(); | |
937 | ||
938 | list_for_each_entry(counter, &ctx->counter_list, list_entry) { | |
939 | if (counter->state > PERF_COUNTER_STATE_OFF) | |
940 | continue; | |
941 | counter->state = PERF_COUNTER_STATE_INACTIVE; | |
942 | counter->hw_event.disabled = 0; | |
943 | } | |
944 | hw_perf_restore(perf_flags); | |
945 | ||
946 | spin_unlock(&ctx->lock); | |
947 | ||
948 | perf_counter_task_sched_in(curr, cpu); | |
949 | ||
950 | curr_rq_unlock_irq_restore(&flags); | |
951 | ||
952 | return 0; | |
953 | } | |
954 | ||
955 | /* | |
956 | * Round-robin a context's counters: | |
957 | */ | |
958 | static void rotate_ctx(struct perf_counter_context *ctx) | |
959 | { | |
960 | struct perf_counter *counter; | |
961 | u64 perf_flags; | |
962 | ||
963 | if (!ctx->nr_counters) | |
964 | return; | |
965 | ||
966 | spin_lock(&ctx->lock); | |
967 | /* | |
968 | * Rotate the first entry last (works just fine for group counters too): | |
969 | */ | |
970 | perf_flags = hw_perf_save_disable(); | |
971 | list_for_each_entry(counter, &ctx->counter_list, list_entry) { | |
972 | list_move_tail(&counter->list_entry, &ctx->counter_list); | |
973 | break; | |
974 | } | |
975 | hw_perf_restore(perf_flags); | |
976 | ||
977 | spin_unlock(&ctx->lock); | |
978 | } | |
979 | ||
980 | void perf_counter_task_tick(struct task_struct *curr, int cpu) | |
981 | { | |
982 | struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); | |
983 | struct perf_counter_context *ctx = &curr->perf_counter_ctx; | |
984 | const int rotate_percpu = 0; | |
985 | ||
986 | if (rotate_percpu) | |
987 | perf_counter_cpu_sched_out(cpuctx); | |
988 | perf_counter_task_sched_out(curr, cpu); | |
989 | ||
990 | if (rotate_percpu) | |
991 | rotate_ctx(&cpuctx->ctx); | |
992 | rotate_ctx(ctx); | |
993 | ||
994 | if (rotate_percpu) | |
995 | perf_counter_cpu_sched_in(cpuctx, cpu); | |
996 | perf_counter_task_sched_in(curr, cpu); | |
997 | } | |
998 | ||
999 | /* | |
1000 | * Cross CPU call to read the hardware counter | |
1001 | */ | |
1002 | static void __read(void *info) | |
1003 | { | |
1004 | struct perf_counter *counter = info; | |
1005 | unsigned long flags; | |
1006 | ||
1007 | curr_rq_lock_irq_save(&flags); | |
1008 | counter->hw_ops->read(counter); | |
1009 | curr_rq_unlock_irq_restore(&flags); | |
1010 | } | |
1011 | ||
1012 | static u64 perf_counter_read(struct perf_counter *counter) | |
1013 | { | |
1014 | /* | |
1015 | * If counter is enabled and currently active on a CPU, update the | |
1016 | * value in the counter structure: | |
1017 | */ | |
1018 | if (counter->state == PERF_COUNTER_STATE_ACTIVE) { | |
1019 | smp_call_function_single(counter->oncpu, | |
1020 | __read, counter, 1); | |
1021 | } | |
1022 | ||
1023 | return atomic64_read(&counter->count); | |
1024 | } | |
1025 | ||
1026 | static void put_context(struct perf_counter_context *ctx) | |
1027 | { | |
1028 | if (ctx->task) | |
1029 | put_task_struct(ctx->task); | |
1030 | } | |
1031 | ||
1032 | static struct perf_counter_context *find_get_context(pid_t pid, int cpu) | |
1033 | { | |
1034 | struct perf_cpu_context *cpuctx; | |
1035 | struct perf_counter_context *ctx; | |
1036 | struct task_struct *task; | |
1037 | ||
1038 | /* | |
1039 | * If cpu is not a wildcard then this is a percpu counter: | |
1040 | */ | |
1041 | if (cpu != -1) { | |
1042 | /* Must be root to operate on a CPU counter: */ | |
1043 | if (!capable(CAP_SYS_ADMIN)) | |
1044 | return ERR_PTR(-EACCES); | |
1045 | ||
1046 | if (cpu < 0 || cpu > num_possible_cpus()) | |
1047 | return ERR_PTR(-EINVAL); | |
1048 | ||
1049 | /* | |
1050 | * We could be clever and allow to attach a counter to an | |
1051 | * offline CPU and activate it when the CPU comes up, but | |
1052 | * that's for later. | |
1053 | */ | |
1054 | if (!cpu_isset(cpu, cpu_online_map)) | |
1055 | return ERR_PTR(-ENODEV); | |
1056 | ||
1057 | cpuctx = &per_cpu(perf_cpu_context, cpu); | |
1058 | ctx = &cpuctx->ctx; | |
1059 | ||
1060 | return ctx; | |
1061 | } | |
1062 | ||
1063 | rcu_read_lock(); | |
1064 | if (!pid) | |
1065 | task = current; | |
1066 | else | |
1067 | task = find_task_by_vpid(pid); | |
1068 | if (task) | |
1069 | get_task_struct(task); | |
1070 | rcu_read_unlock(); | |
1071 | ||
1072 | if (!task) | |
1073 | return ERR_PTR(-ESRCH); | |
1074 | ||
1075 | ctx = &task->perf_counter_ctx; | |
1076 | ctx->task = task; | |
1077 | ||
1078 | /* Reuse ptrace permission checks for now. */ | |
1079 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) { | |
1080 | put_context(ctx); | |
1081 | return ERR_PTR(-EACCES); | |
1082 | } | |
1083 | ||
1084 | return ctx; | |
1085 | } | |
1086 | ||
1087 | static void free_counter_rcu(struct rcu_head *head) | |
1088 | { | |
1089 | struct perf_counter *counter; | |
1090 | ||
1091 | counter = container_of(head, struct perf_counter, rcu_head); | |
1092 | kfree(counter); | |
1093 | } | |
1094 | ||
1095 | static void free_counter(struct perf_counter *counter) | |
1096 | { | |
1097 | if (counter->destroy) | |
1098 | counter->destroy(counter); | |
1099 | ||
1100 | call_rcu(&counter->rcu_head, free_counter_rcu); | |
1101 | } | |
1102 | ||
1103 | /* | |
1104 | * Called when the last reference to the file is gone. | |
1105 | */ | |
1106 | static int perf_release(struct inode *inode, struct file *file) | |
1107 | { | |
1108 | struct perf_counter *counter = file->private_data; | |
1109 | struct perf_counter_context *ctx = counter->ctx; | |
1110 | ||
1111 | file->private_data = NULL; | |
1112 | ||
1113 | mutex_lock(&ctx->mutex); | |
1114 | mutex_lock(&counter->mutex); | |
1115 | ||
1116 | perf_counter_remove_from_context(counter); | |
1117 | ||
1118 | mutex_unlock(&counter->mutex); | |
1119 | mutex_unlock(&ctx->mutex); | |
1120 | ||
1121 | free_counter(counter); | |
1122 | put_context(ctx); | |
1123 | ||
1124 | return 0; | |
1125 | } | |
1126 | ||
1127 | /* | |
1128 | * Read the performance counter - simple non blocking version for now | |
1129 | */ | |
1130 | static ssize_t | |
1131 | perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count) | |
1132 | { | |
1133 | u64 cntval; | |
1134 | ||
1135 | if (count < sizeof(cntval)) | |
1136 | return -EINVAL; | |
1137 | ||
1138 | /* | |
1139 | * Return end-of-file for a read on a counter that is in | |
1140 | * error state (i.e. because it was pinned but it couldn't be | |
1141 | * scheduled on to the CPU at some point). | |
1142 | */ | |
1143 | if (counter->state == PERF_COUNTER_STATE_ERROR) | |
1144 | return 0; | |
1145 | ||
1146 | mutex_lock(&counter->mutex); | |
1147 | cntval = perf_counter_read(counter); | |
1148 | mutex_unlock(&counter->mutex); | |
1149 | ||
1150 | return put_user(cntval, (u64 __user *) buf) ? -EFAULT : sizeof(cntval); | |
1151 | } | |
1152 | ||
1153 | static ssize_t | |
1154 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
1155 | { | |
1156 | struct perf_counter *counter = file->private_data; | |
1157 | ||
1158 | return perf_read_hw(counter, buf, count); | |
1159 | } | |
1160 | ||
1161 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
1162 | { | |
1163 | struct perf_counter *counter = file->private_data; | |
1164 | struct perf_mmap_data *data; | |
1165 | unsigned int events; | |
1166 | ||
1167 | rcu_read_lock(); | |
1168 | data = rcu_dereference(counter->data); | |
1169 | if (data) | |
1170 | events = atomic_xchg(&data->wakeup, 0); | |
1171 | else | |
1172 | events = POLL_HUP; | |
1173 | rcu_read_unlock(); | |
1174 | ||
1175 | poll_wait(file, &counter->waitq, wait); | |
1176 | ||
1177 | return events; | |
1178 | } | |
1179 | ||
1180 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) | |
1181 | { | |
1182 | struct perf_counter *counter = file->private_data; | |
1183 | int err = 0; | |
1184 | ||
1185 | switch (cmd) { | |
1186 | case PERF_COUNTER_IOC_ENABLE: | |
1187 | perf_counter_enable_family(counter); | |
1188 | break; | |
1189 | case PERF_COUNTER_IOC_DISABLE: | |
1190 | perf_counter_disable_family(counter); | |
1191 | break; | |
1192 | default: | |
1193 | err = -ENOTTY; | |
1194 | } | |
1195 | return err; | |
1196 | } | |
1197 | ||
1198 | static void __perf_counter_update_userpage(struct perf_counter *counter, | |
1199 | struct perf_mmap_data *data) | |
1200 | { | |
1201 | struct perf_counter_mmap_page *userpg = data->user_page; | |
1202 | ||
1203 | /* | |
1204 | * Disable preemption so as to not let the corresponding user-space | |
1205 | * spin too long if we get preempted. | |
1206 | */ | |
1207 | preempt_disable(); | |
1208 | ++userpg->lock; | |
1209 | smp_wmb(); | |
1210 | userpg->index = counter->hw.idx; | |
1211 | userpg->offset = atomic64_read(&counter->count); | |
1212 | if (counter->state == PERF_COUNTER_STATE_ACTIVE) | |
1213 | userpg->offset -= atomic64_read(&counter->hw.prev_count); | |
1214 | ||
1215 | userpg->data_head = atomic_read(&data->head); | |
1216 | smp_wmb(); | |
1217 | ++userpg->lock; | |
1218 | preempt_enable(); | |
1219 | } | |
1220 | ||
1221 | void perf_counter_update_userpage(struct perf_counter *counter) | |
1222 | { | |
1223 | struct perf_mmap_data *data; | |
1224 | ||
1225 | rcu_read_lock(); | |
1226 | data = rcu_dereference(counter->data); | |
1227 | if (data) | |
1228 | __perf_counter_update_userpage(counter, data); | |
1229 | rcu_read_unlock(); | |
1230 | } | |
1231 | ||
1232 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) | |
1233 | { | |
1234 | struct perf_counter *counter = vma->vm_file->private_data; | |
1235 | struct perf_mmap_data *data; | |
1236 | int ret = VM_FAULT_SIGBUS; | |
1237 | ||
1238 | rcu_read_lock(); | |
1239 | data = rcu_dereference(counter->data); | |
1240 | if (!data) | |
1241 | goto unlock; | |
1242 | ||
1243 | if (vmf->pgoff == 0) { | |
1244 | vmf->page = virt_to_page(data->user_page); | |
1245 | } else { | |
1246 | int nr = vmf->pgoff - 1; | |
1247 | ||
1248 | if ((unsigned)nr > data->nr_pages) | |
1249 | goto unlock; | |
1250 | ||
1251 | vmf->page = virt_to_page(data->data_pages[nr]); | |
1252 | } | |
1253 | get_page(vmf->page); | |
1254 | ret = 0; | |
1255 | unlock: | |
1256 | rcu_read_unlock(); | |
1257 | ||
1258 | return ret; | |
1259 | } | |
1260 | ||
1261 | static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages) | |
1262 | { | |
1263 | struct perf_mmap_data *data; | |
1264 | unsigned long size; | |
1265 | int i; | |
1266 | ||
1267 | WARN_ON(atomic_read(&counter->mmap_count)); | |
1268 | ||
1269 | size = sizeof(struct perf_mmap_data); | |
1270 | size += nr_pages * sizeof(void *); | |
1271 | ||
1272 | data = kzalloc(size, GFP_KERNEL); | |
1273 | if (!data) | |
1274 | goto fail; | |
1275 | ||
1276 | data->user_page = (void *)get_zeroed_page(GFP_KERNEL); | |
1277 | if (!data->user_page) | |
1278 | goto fail_user_page; | |
1279 | ||
1280 | for (i = 0; i < nr_pages; i++) { | |
1281 | data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL); | |
1282 | if (!data->data_pages[i]) | |
1283 | goto fail_data_pages; | |
1284 | } | |
1285 | ||
1286 | data->nr_pages = nr_pages; | |
1287 | ||
1288 | rcu_assign_pointer(counter->data, data); | |
1289 | ||
1290 | return 0; | |
1291 | ||
1292 | fail_data_pages: | |
1293 | for (i--; i >= 0; i--) | |
1294 | free_page((unsigned long)data->data_pages[i]); | |
1295 | ||
1296 | free_page((unsigned long)data->user_page); | |
1297 | ||
1298 | fail_user_page: | |
1299 | kfree(data); | |
1300 | ||
1301 | fail: | |
1302 | return -ENOMEM; | |
1303 | } | |
1304 | ||
1305 | static void __perf_mmap_data_free(struct rcu_head *rcu_head) | |
1306 | { | |
1307 | struct perf_mmap_data *data = container_of(rcu_head, | |
1308 | struct perf_mmap_data, rcu_head); | |
1309 | int i; | |
1310 | ||
1311 | free_page((unsigned long)data->user_page); | |
1312 | for (i = 0; i < data->nr_pages; i++) | |
1313 | free_page((unsigned long)data->data_pages[i]); | |
1314 | kfree(data); | |
1315 | } | |
1316 | ||
1317 | static void perf_mmap_data_free(struct perf_counter *counter) | |
1318 | { | |
1319 | struct perf_mmap_data *data = counter->data; | |
1320 | ||
1321 | WARN_ON(atomic_read(&counter->mmap_count)); | |
1322 | ||
1323 | rcu_assign_pointer(counter->data, NULL); | |
1324 | call_rcu(&data->rcu_head, __perf_mmap_data_free); | |
1325 | } | |
1326 | ||
1327 | static void perf_mmap_open(struct vm_area_struct *vma) | |
1328 | { | |
1329 | struct perf_counter *counter = vma->vm_file->private_data; | |
1330 | ||
1331 | atomic_inc(&counter->mmap_count); | |
1332 | } | |
1333 | ||
1334 | static void perf_mmap_close(struct vm_area_struct *vma) | |
1335 | { | |
1336 | struct perf_counter *counter = vma->vm_file->private_data; | |
1337 | ||
1338 | if (atomic_dec_and_mutex_lock(&counter->mmap_count, | |
1339 | &counter->mmap_mutex)) { | |
1340 | perf_mmap_data_free(counter); | |
1341 | mutex_unlock(&counter->mmap_mutex); | |
1342 | } | |
1343 | } | |
1344 | ||
1345 | static struct vm_operations_struct perf_mmap_vmops = { | |
1346 | .open = perf_mmap_open, | |
1347 | .close = perf_mmap_close, | |
1348 | .fault = perf_mmap_fault, | |
1349 | }; | |
1350 | ||
1351 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
1352 | { | |
1353 | struct perf_counter *counter = file->private_data; | |
1354 | unsigned long vma_size; | |
1355 | unsigned long nr_pages; | |
1356 | unsigned long locked, lock_limit; | |
1357 | int ret = 0; | |
1358 | ||
1359 | if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE)) | |
1360 | return -EINVAL; | |
1361 | ||
1362 | vma_size = vma->vm_end - vma->vm_start; | |
1363 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
1364 | ||
1365 | if (nr_pages == 0 || !is_power_of_2(nr_pages)) | |
1366 | return -EINVAL; | |
1367 | ||
1368 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) | |
1369 | return -EINVAL; | |
1370 | ||
1371 | if (vma->vm_pgoff != 0) | |
1372 | return -EINVAL; | |
1373 | ||
1374 | locked = vma_size >> PAGE_SHIFT; | |
1375 | locked += vma->vm_mm->locked_vm; | |
1376 | ||
1377 | lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur; | |
1378 | lock_limit >>= PAGE_SHIFT; | |
1379 | ||
1380 | if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) | |
1381 | return -EPERM; | |
1382 | ||
1383 | mutex_lock(&counter->mmap_mutex); | |
1384 | if (atomic_inc_not_zero(&counter->mmap_count)) | |
1385 | goto out; | |
1386 | ||
1387 | WARN_ON(counter->data); | |
1388 | ret = perf_mmap_data_alloc(counter, nr_pages); | |
1389 | if (!ret) | |
1390 | atomic_set(&counter->mmap_count, 1); | |
1391 | out: | |
1392 | mutex_unlock(&counter->mmap_mutex); | |
1393 | ||
1394 | vma->vm_flags &= ~VM_MAYWRITE; | |
1395 | vma->vm_flags |= VM_RESERVED; | |
1396 | vma->vm_ops = &perf_mmap_vmops; | |
1397 | ||
1398 | return ret; | |
1399 | } | |
1400 | ||
1401 | static const struct file_operations perf_fops = { | |
1402 | .release = perf_release, | |
1403 | .read = perf_read, | |
1404 | .poll = perf_poll, | |
1405 | .unlocked_ioctl = perf_ioctl, | |
1406 | .compat_ioctl = perf_ioctl, | |
1407 | .mmap = perf_mmap, | |
1408 | }; | |
1409 | ||
1410 | /* | |
1411 | * Output | |
1412 | */ | |
1413 | ||
1414 | struct perf_output_handle { | |
1415 | struct perf_counter *counter; | |
1416 | struct perf_mmap_data *data; | |
1417 | unsigned int offset; | |
1418 | int wakeup; | |
1419 | }; | |
1420 | ||
1421 | static int perf_output_begin(struct perf_output_handle *handle, | |
1422 | struct perf_counter *counter, unsigned int size) | |
1423 | { | |
1424 | struct perf_mmap_data *data; | |
1425 | unsigned int offset, head; | |
1426 | ||
1427 | rcu_read_lock(); | |
1428 | data = rcu_dereference(counter->data); | |
1429 | if (!data) | |
1430 | goto out; | |
1431 | ||
1432 | if (!data->nr_pages) | |
1433 | goto out; | |
1434 | ||
1435 | do { | |
1436 | offset = head = atomic_read(&data->head); | |
1437 | head += size; | |
1438 | } while (atomic_cmpxchg(&data->head, offset, head) != offset); | |
1439 | ||
1440 | handle->counter = counter; | |
1441 | handle->data = data; | |
1442 | handle->offset = offset; | |
1443 | handle->wakeup = (offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT); | |
1444 | ||
1445 | return 0; | |
1446 | ||
1447 | out: | |
1448 | rcu_read_unlock(); | |
1449 | ||
1450 | return -ENOSPC; | |
1451 | } | |
1452 | ||
1453 | static void perf_output_copy(struct perf_output_handle *handle, | |
1454 | void *buf, unsigned int len) | |
1455 | { | |
1456 | unsigned int pages_mask; | |
1457 | unsigned int offset; | |
1458 | unsigned int size; | |
1459 | void **pages; | |
1460 | ||
1461 | offset = handle->offset; | |
1462 | pages_mask = handle->data->nr_pages - 1; | |
1463 | pages = handle->data->data_pages; | |
1464 | ||
1465 | do { | |
1466 | unsigned int page_offset; | |
1467 | int nr; | |
1468 | ||
1469 | nr = (offset >> PAGE_SHIFT) & pages_mask; | |
1470 | page_offset = offset & (PAGE_SIZE - 1); | |
1471 | size = min_t(unsigned int, PAGE_SIZE - page_offset, len); | |
1472 | ||
1473 | memcpy(pages[nr] + page_offset, buf, size); | |
1474 | ||
1475 | len -= size; | |
1476 | buf += size; | |
1477 | offset += size; | |
1478 | } while (len); | |
1479 | ||
1480 | handle->offset = offset; | |
1481 | } | |
1482 | ||
1483 | static void perf_output_end(struct perf_output_handle *handle, int nmi) | |
1484 | { | |
1485 | if (handle->wakeup) { | |
1486 | (void)atomic_xchg(&handle->data->wakeup, POLL_IN); | |
1487 | __perf_counter_update_userpage(handle->counter, handle->data); | |
1488 | if (nmi) { | |
1489 | handle->counter->wakeup_pending = 1; | |
1490 | set_perf_counter_pending(); | |
1491 | } else | |
1492 | wake_up(&handle->counter->waitq); | |
1493 | } | |
1494 | rcu_read_unlock(); | |
1495 | } | |
1496 | ||
1497 | static int perf_output_write(struct perf_counter *counter, int nmi, | |
1498 | void *buf, ssize_t size) | |
1499 | { | |
1500 | struct perf_output_handle handle; | |
1501 | int ret; | |
1502 | ||
1503 | ret = perf_output_begin(&handle, counter, size); | |
1504 | if (ret) | |
1505 | goto out; | |
1506 | ||
1507 | perf_output_copy(&handle, buf, size); | |
1508 | perf_output_end(&handle, nmi); | |
1509 | ||
1510 | out: | |
1511 | return ret; | |
1512 | } | |
1513 | ||
1514 | static void perf_output_simple(struct perf_counter *counter, | |
1515 | int nmi, struct pt_regs *regs) | |
1516 | { | |
1517 | u64 entry; | |
1518 | ||
1519 | entry = instruction_pointer(regs); | |
1520 | ||
1521 | perf_output_write(counter, nmi, &entry, sizeof(entry)); | |
1522 | } | |
1523 | ||
1524 | struct group_entry { | |
1525 | u64 event; | |
1526 | u64 counter; | |
1527 | }; | |
1528 | ||
1529 | static void perf_output_group(struct perf_counter *counter, int nmi) | |
1530 | { | |
1531 | struct perf_counter *leader, *sub; | |
1532 | ||
1533 | leader = counter->group_leader; | |
1534 | list_for_each_entry(sub, &leader->sibling_list, list_entry) { | |
1535 | struct group_entry entry; | |
1536 | ||
1537 | if (sub != counter) | |
1538 | sub->hw_ops->read(sub); | |
1539 | ||
1540 | entry.event = sub->hw_event.config; | |
1541 | entry.counter = atomic64_read(&sub->count); | |
1542 | ||
1543 | perf_output_write(counter, nmi, &entry, sizeof(entry)); | |
1544 | } | |
1545 | } | |
1546 | ||
1547 | void perf_counter_output(struct perf_counter *counter, | |
1548 | int nmi, struct pt_regs *regs) | |
1549 | { | |
1550 | switch (counter->hw_event.record_type) { | |
1551 | case PERF_RECORD_SIMPLE: | |
1552 | return; | |
1553 | ||
1554 | case PERF_RECORD_IRQ: | |
1555 | perf_output_simple(counter, nmi, regs); | |
1556 | break; | |
1557 | ||
1558 | case PERF_RECORD_GROUP: | |
1559 | perf_output_group(counter, nmi); | |
1560 | break; | |
1561 | } | |
1562 | } | |
1563 | ||
1564 | /* | |
1565 | * Generic software counter infrastructure | |
1566 | */ | |
1567 | ||
1568 | static void perf_swcounter_update(struct perf_counter *counter) | |
1569 | { | |
1570 | struct hw_perf_counter *hwc = &counter->hw; | |
1571 | u64 prev, now; | |
1572 | s64 delta; | |
1573 | ||
1574 | again: | |
1575 | prev = atomic64_read(&hwc->prev_count); | |
1576 | now = atomic64_read(&hwc->count); | |
1577 | if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev) | |
1578 | goto again; | |
1579 | ||
1580 | delta = now - prev; | |
1581 | ||
1582 | atomic64_add(delta, &counter->count); | |
1583 | atomic64_sub(delta, &hwc->period_left); | |
1584 | } | |
1585 | ||
1586 | static void perf_swcounter_set_period(struct perf_counter *counter) | |
1587 | { | |
1588 | struct hw_perf_counter *hwc = &counter->hw; | |
1589 | s64 left = atomic64_read(&hwc->period_left); | |
1590 | s64 period = hwc->irq_period; | |
1591 | ||
1592 | if (unlikely(left <= -period)) { | |
1593 | left = period; | |
1594 | atomic64_set(&hwc->period_left, left); | |
1595 | } | |
1596 | ||
1597 | if (unlikely(left <= 0)) { | |
1598 | left += period; | |
1599 | atomic64_add(period, &hwc->period_left); | |
1600 | } | |
1601 | ||
1602 | atomic64_set(&hwc->prev_count, -left); | |
1603 | atomic64_set(&hwc->count, -left); | |
1604 | } | |
1605 | ||
1606 | static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer) | |
1607 | { | |
1608 | struct perf_counter *counter; | |
1609 | struct pt_regs *regs; | |
1610 | ||
1611 | counter = container_of(hrtimer, struct perf_counter, hw.hrtimer); | |
1612 | counter->hw_ops->read(counter); | |
1613 | ||
1614 | regs = get_irq_regs(); | |
1615 | /* | |
1616 | * In case we exclude kernel IPs or are somehow not in interrupt | |
1617 | * context, provide the next best thing, the user IP. | |
1618 | */ | |
1619 | if ((counter->hw_event.exclude_kernel || !regs) && | |
1620 | !counter->hw_event.exclude_user) | |
1621 | regs = task_pt_regs(current); | |
1622 | ||
1623 | if (regs) | |
1624 | perf_counter_output(counter, 0, regs); | |
1625 | ||
1626 | hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period)); | |
1627 | ||
1628 | return HRTIMER_RESTART; | |
1629 | } | |
1630 | ||
1631 | static void perf_swcounter_overflow(struct perf_counter *counter, | |
1632 | int nmi, struct pt_regs *regs) | |
1633 | { | |
1634 | perf_swcounter_update(counter); | |
1635 | perf_swcounter_set_period(counter); | |
1636 | perf_counter_output(counter, nmi, regs); | |
1637 | } | |
1638 | ||
1639 | static int perf_swcounter_match(struct perf_counter *counter, | |
1640 | enum perf_event_types type, | |
1641 | u32 event, struct pt_regs *regs) | |
1642 | { | |
1643 | if (counter->state != PERF_COUNTER_STATE_ACTIVE) | |
1644 | return 0; | |
1645 | ||
1646 | if (perf_event_raw(&counter->hw_event)) | |
1647 | return 0; | |
1648 | ||
1649 | if (perf_event_type(&counter->hw_event) != type) | |
1650 | return 0; | |
1651 | ||
1652 | if (perf_event_id(&counter->hw_event) != event) | |
1653 | return 0; | |
1654 | ||
1655 | if (counter->hw_event.exclude_user && user_mode(regs)) | |
1656 | return 0; | |
1657 | ||
1658 | if (counter->hw_event.exclude_kernel && !user_mode(regs)) | |
1659 | return 0; | |
1660 | ||
1661 | return 1; | |
1662 | } | |
1663 | ||
1664 | static void perf_swcounter_add(struct perf_counter *counter, u64 nr, | |
1665 | int nmi, struct pt_regs *regs) | |
1666 | { | |
1667 | int neg = atomic64_add_negative(nr, &counter->hw.count); | |
1668 | if (counter->hw.irq_period && !neg) | |
1669 | perf_swcounter_overflow(counter, nmi, regs); | |
1670 | } | |
1671 | ||
1672 | static void perf_swcounter_ctx_event(struct perf_counter_context *ctx, | |
1673 | enum perf_event_types type, u32 event, | |
1674 | u64 nr, int nmi, struct pt_regs *regs) | |
1675 | { | |
1676 | struct perf_counter *counter; | |
1677 | ||
1678 | if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list)) | |
1679 | return; | |
1680 | ||
1681 | rcu_read_lock(); | |
1682 | list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) { | |
1683 | if (perf_swcounter_match(counter, type, event, regs)) | |
1684 | perf_swcounter_add(counter, nr, nmi, regs); | |
1685 | } | |
1686 | rcu_read_unlock(); | |
1687 | } | |
1688 | ||
1689 | static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx) | |
1690 | { | |
1691 | if (in_nmi()) | |
1692 | return &cpuctx->recursion[3]; | |
1693 | ||
1694 | if (in_irq()) | |
1695 | return &cpuctx->recursion[2]; | |
1696 | ||
1697 | if (in_softirq()) | |
1698 | return &cpuctx->recursion[1]; | |
1699 | ||
1700 | return &cpuctx->recursion[0]; | |
1701 | } | |
1702 | ||
1703 | static void __perf_swcounter_event(enum perf_event_types type, u32 event, | |
1704 | u64 nr, int nmi, struct pt_regs *regs) | |
1705 | { | |
1706 | struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context); | |
1707 | int *recursion = perf_swcounter_recursion_context(cpuctx); | |
1708 | ||
1709 | if (*recursion) | |
1710 | goto out; | |
1711 | ||
1712 | (*recursion)++; | |
1713 | barrier(); | |
1714 | ||
1715 | perf_swcounter_ctx_event(&cpuctx->ctx, type, event, nr, nmi, regs); | |
1716 | if (cpuctx->task_ctx) { | |
1717 | perf_swcounter_ctx_event(cpuctx->task_ctx, type, event, | |
1718 | nr, nmi, regs); | |
1719 | } | |
1720 | ||
1721 | barrier(); | |
1722 | (*recursion)--; | |
1723 | ||
1724 | out: | |
1725 | put_cpu_var(perf_cpu_context); | |
1726 | } | |
1727 | ||
1728 | void perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs) | |
1729 | { | |
1730 | __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs); | |
1731 | } | |
1732 | ||
1733 | static void perf_swcounter_read(struct perf_counter *counter) | |
1734 | { | |
1735 | perf_swcounter_update(counter); | |
1736 | } | |
1737 | ||
1738 | static int perf_swcounter_enable(struct perf_counter *counter) | |
1739 | { | |
1740 | perf_swcounter_set_period(counter); | |
1741 | return 0; | |
1742 | } | |
1743 | ||
1744 | static void perf_swcounter_disable(struct perf_counter *counter) | |
1745 | { | |
1746 | perf_swcounter_update(counter); | |
1747 | } | |
1748 | ||
1749 | static const struct hw_perf_counter_ops perf_ops_generic = { | |
1750 | .enable = perf_swcounter_enable, | |
1751 | .disable = perf_swcounter_disable, | |
1752 | .read = perf_swcounter_read, | |
1753 | }; | |
1754 | ||
1755 | /* | |
1756 | * Software counter: cpu wall time clock | |
1757 | */ | |
1758 | ||
1759 | static void cpu_clock_perf_counter_update(struct perf_counter *counter) | |
1760 | { | |
1761 | int cpu = raw_smp_processor_id(); | |
1762 | s64 prev; | |
1763 | u64 now; | |
1764 | ||
1765 | now = cpu_clock(cpu); | |
1766 | prev = atomic64_read(&counter->hw.prev_count); | |
1767 | atomic64_set(&counter->hw.prev_count, now); | |
1768 | atomic64_add(now - prev, &counter->count); | |
1769 | } | |
1770 | ||
1771 | static int cpu_clock_perf_counter_enable(struct perf_counter *counter) | |
1772 | { | |
1773 | struct hw_perf_counter *hwc = &counter->hw; | |
1774 | int cpu = raw_smp_processor_id(); | |
1775 | ||
1776 | atomic64_set(&hwc->prev_count, cpu_clock(cpu)); | |
1777 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
1778 | hwc->hrtimer.function = perf_swcounter_hrtimer; | |
1779 | if (hwc->irq_period) { | |
1780 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
1781 | ns_to_ktime(hwc->irq_period), 0, | |
1782 | HRTIMER_MODE_REL, 0); | |
1783 | } | |
1784 | ||
1785 | return 0; | |
1786 | } | |
1787 | ||
1788 | static void cpu_clock_perf_counter_disable(struct perf_counter *counter) | |
1789 | { | |
1790 | hrtimer_cancel(&counter->hw.hrtimer); | |
1791 | cpu_clock_perf_counter_update(counter); | |
1792 | } | |
1793 | ||
1794 | static void cpu_clock_perf_counter_read(struct perf_counter *counter) | |
1795 | { | |
1796 | cpu_clock_perf_counter_update(counter); | |
1797 | } | |
1798 | ||
1799 | static const struct hw_perf_counter_ops perf_ops_cpu_clock = { | |
1800 | .enable = cpu_clock_perf_counter_enable, | |
1801 | .disable = cpu_clock_perf_counter_disable, | |
1802 | .read = cpu_clock_perf_counter_read, | |
1803 | }; | |
1804 | ||
1805 | /* | |
1806 | * Software counter: task time clock | |
1807 | */ | |
1808 | ||
1809 | /* | |
1810 | * Called from within the scheduler: | |
1811 | */ | |
1812 | static u64 task_clock_perf_counter_val(struct perf_counter *counter, int update) | |
1813 | { | |
1814 | struct task_struct *curr = counter->task; | |
1815 | u64 delta; | |
1816 | ||
1817 | delta = __task_delta_exec(curr, update); | |
1818 | ||
1819 | return curr->se.sum_exec_runtime + delta; | |
1820 | } | |
1821 | ||
1822 | static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now) | |
1823 | { | |
1824 | u64 prev; | |
1825 | s64 delta; | |
1826 | ||
1827 | prev = atomic64_read(&counter->hw.prev_count); | |
1828 | ||
1829 | atomic64_set(&counter->hw.prev_count, now); | |
1830 | ||
1831 | delta = now - prev; | |
1832 | ||
1833 | atomic64_add(delta, &counter->count); | |
1834 | } | |
1835 | ||
1836 | static int task_clock_perf_counter_enable(struct perf_counter *counter) | |
1837 | { | |
1838 | struct hw_perf_counter *hwc = &counter->hw; | |
1839 | ||
1840 | atomic64_set(&hwc->prev_count, task_clock_perf_counter_val(counter, 0)); | |
1841 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
1842 | hwc->hrtimer.function = perf_swcounter_hrtimer; | |
1843 | if (hwc->irq_period) { | |
1844 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
1845 | ns_to_ktime(hwc->irq_period), 0, | |
1846 | HRTIMER_MODE_REL, 0); | |
1847 | } | |
1848 | ||
1849 | return 0; | |
1850 | } | |
1851 | ||
1852 | static void task_clock_perf_counter_disable(struct perf_counter *counter) | |
1853 | { | |
1854 | hrtimer_cancel(&counter->hw.hrtimer); | |
1855 | task_clock_perf_counter_update(counter, | |
1856 | task_clock_perf_counter_val(counter, 0)); | |
1857 | } | |
1858 | ||
1859 | static void task_clock_perf_counter_read(struct perf_counter *counter) | |
1860 | { | |
1861 | task_clock_perf_counter_update(counter, | |
1862 | task_clock_perf_counter_val(counter, 1)); | |
1863 | } | |
1864 | ||
1865 | static const struct hw_perf_counter_ops perf_ops_task_clock = { | |
1866 | .enable = task_clock_perf_counter_enable, | |
1867 | .disable = task_clock_perf_counter_disable, | |
1868 | .read = task_clock_perf_counter_read, | |
1869 | }; | |
1870 | ||
1871 | /* | |
1872 | * Software counter: cpu migrations | |
1873 | */ | |
1874 | ||
1875 | static inline u64 get_cpu_migrations(struct perf_counter *counter) | |
1876 | { | |
1877 | struct task_struct *curr = counter->ctx->task; | |
1878 | ||
1879 | if (curr) | |
1880 | return curr->se.nr_migrations; | |
1881 | return cpu_nr_migrations(smp_processor_id()); | |
1882 | } | |
1883 | ||
1884 | static void cpu_migrations_perf_counter_update(struct perf_counter *counter) | |
1885 | { | |
1886 | u64 prev, now; | |
1887 | s64 delta; | |
1888 | ||
1889 | prev = atomic64_read(&counter->hw.prev_count); | |
1890 | now = get_cpu_migrations(counter); | |
1891 | ||
1892 | atomic64_set(&counter->hw.prev_count, now); | |
1893 | ||
1894 | delta = now - prev; | |
1895 | ||
1896 | atomic64_add(delta, &counter->count); | |
1897 | } | |
1898 | ||
1899 | static void cpu_migrations_perf_counter_read(struct perf_counter *counter) | |
1900 | { | |
1901 | cpu_migrations_perf_counter_update(counter); | |
1902 | } | |
1903 | ||
1904 | static int cpu_migrations_perf_counter_enable(struct perf_counter *counter) | |
1905 | { | |
1906 | if (counter->prev_state <= PERF_COUNTER_STATE_OFF) | |
1907 | atomic64_set(&counter->hw.prev_count, | |
1908 | get_cpu_migrations(counter)); | |
1909 | return 0; | |
1910 | } | |
1911 | ||
1912 | static void cpu_migrations_perf_counter_disable(struct perf_counter *counter) | |
1913 | { | |
1914 | cpu_migrations_perf_counter_update(counter); | |
1915 | } | |
1916 | ||
1917 | static const struct hw_perf_counter_ops perf_ops_cpu_migrations = { | |
1918 | .enable = cpu_migrations_perf_counter_enable, | |
1919 | .disable = cpu_migrations_perf_counter_disable, | |
1920 | .read = cpu_migrations_perf_counter_read, | |
1921 | }; | |
1922 | ||
1923 | #ifdef CONFIG_EVENT_PROFILE | |
1924 | void perf_tpcounter_event(int event_id) | |
1925 | { | |
1926 | struct pt_regs *regs = get_irq_regs(); | |
1927 | ||
1928 | if (!regs) | |
1929 | regs = task_pt_regs(current); | |
1930 | ||
1931 | __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs); | |
1932 | } | |
1933 | ||
1934 | extern int ftrace_profile_enable(int); | |
1935 | extern void ftrace_profile_disable(int); | |
1936 | ||
1937 | static void tp_perf_counter_destroy(struct perf_counter *counter) | |
1938 | { | |
1939 | ftrace_profile_disable(perf_event_id(&counter->hw_event)); | |
1940 | } | |
1941 | ||
1942 | static const struct hw_perf_counter_ops * | |
1943 | tp_perf_counter_init(struct perf_counter *counter) | |
1944 | { | |
1945 | int event_id = perf_event_id(&counter->hw_event); | |
1946 | int ret; | |
1947 | ||
1948 | ret = ftrace_profile_enable(event_id); | |
1949 | if (ret) | |
1950 | return NULL; | |
1951 | ||
1952 | counter->destroy = tp_perf_counter_destroy; | |
1953 | counter->hw.irq_period = counter->hw_event.irq_period; | |
1954 | ||
1955 | return &perf_ops_generic; | |
1956 | } | |
1957 | #else | |
1958 | static const struct hw_perf_counter_ops * | |
1959 | tp_perf_counter_init(struct perf_counter *counter) | |
1960 | { | |
1961 | return NULL; | |
1962 | } | |
1963 | #endif | |
1964 | ||
1965 | static const struct hw_perf_counter_ops * | |
1966 | sw_perf_counter_init(struct perf_counter *counter) | |
1967 | { | |
1968 | struct perf_counter_hw_event *hw_event = &counter->hw_event; | |
1969 | const struct hw_perf_counter_ops *hw_ops = NULL; | |
1970 | struct hw_perf_counter *hwc = &counter->hw; | |
1971 | ||
1972 | /* | |
1973 | * Software counters (currently) can't in general distinguish | |
1974 | * between user, kernel and hypervisor events. | |
1975 | * However, context switches and cpu migrations are considered | |
1976 | * to be kernel events, and page faults are never hypervisor | |
1977 | * events. | |
1978 | */ | |
1979 | switch (perf_event_id(&counter->hw_event)) { | |
1980 | case PERF_COUNT_CPU_CLOCK: | |
1981 | hw_ops = &perf_ops_cpu_clock; | |
1982 | ||
1983 | if (hw_event->irq_period && hw_event->irq_period < 10000) | |
1984 | hw_event->irq_period = 10000; | |
1985 | break; | |
1986 | case PERF_COUNT_TASK_CLOCK: | |
1987 | /* | |
1988 | * If the user instantiates this as a per-cpu counter, | |
1989 | * use the cpu_clock counter instead. | |
1990 | */ | |
1991 | if (counter->ctx->task) | |
1992 | hw_ops = &perf_ops_task_clock; | |
1993 | else | |
1994 | hw_ops = &perf_ops_cpu_clock; | |
1995 | ||
1996 | if (hw_event->irq_period && hw_event->irq_period < 10000) | |
1997 | hw_event->irq_period = 10000; | |
1998 | break; | |
1999 | case PERF_COUNT_PAGE_FAULTS: | |
2000 | case PERF_COUNT_PAGE_FAULTS_MIN: | |
2001 | case PERF_COUNT_PAGE_FAULTS_MAJ: | |
2002 | case PERF_COUNT_CONTEXT_SWITCHES: | |
2003 | hw_ops = &perf_ops_generic; | |
2004 | break; | |
2005 | case PERF_COUNT_CPU_MIGRATIONS: | |
2006 | if (!counter->hw_event.exclude_kernel) | |
2007 | hw_ops = &perf_ops_cpu_migrations; | |
2008 | break; | |
2009 | } | |
2010 | ||
2011 | if (hw_ops) | |
2012 | hwc->irq_period = hw_event->irq_period; | |
2013 | ||
2014 | return hw_ops; | |
2015 | } | |
2016 | ||
2017 | /* | |
2018 | * Allocate and initialize a counter structure | |
2019 | */ | |
2020 | static struct perf_counter * | |
2021 | perf_counter_alloc(struct perf_counter_hw_event *hw_event, | |
2022 | int cpu, | |
2023 | struct perf_counter_context *ctx, | |
2024 | struct perf_counter *group_leader, | |
2025 | gfp_t gfpflags) | |
2026 | { | |
2027 | const struct hw_perf_counter_ops *hw_ops; | |
2028 | struct perf_counter *counter; | |
2029 | ||
2030 | counter = kzalloc(sizeof(*counter), gfpflags); | |
2031 | if (!counter) | |
2032 | return NULL; | |
2033 | ||
2034 | /* | |
2035 | * Single counters are their own group leaders, with an | |
2036 | * empty sibling list: | |
2037 | */ | |
2038 | if (!group_leader) | |
2039 | group_leader = counter; | |
2040 | ||
2041 | mutex_init(&counter->mutex); | |
2042 | INIT_LIST_HEAD(&counter->list_entry); | |
2043 | INIT_LIST_HEAD(&counter->event_entry); | |
2044 | INIT_LIST_HEAD(&counter->sibling_list); | |
2045 | init_waitqueue_head(&counter->waitq); | |
2046 | ||
2047 | mutex_init(&counter->mmap_mutex); | |
2048 | ||
2049 | INIT_LIST_HEAD(&counter->child_list); | |
2050 | ||
2051 | counter->cpu = cpu; | |
2052 | counter->hw_event = *hw_event; | |
2053 | counter->wakeup_pending = 0; | |
2054 | counter->group_leader = group_leader; | |
2055 | counter->hw_ops = NULL; | |
2056 | counter->ctx = ctx; | |
2057 | ||
2058 | counter->state = PERF_COUNTER_STATE_INACTIVE; | |
2059 | if (hw_event->disabled) | |
2060 | counter->state = PERF_COUNTER_STATE_OFF; | |
2061 | ||
2062 | hw_ops = NULL; | |
2063 | ||
2064 | if (perf_event_raw(hw_event)) { | |
2065 | hw_ops = hw_perf_counter_init(counter); | |
2066 | goto done; | |
2067 | } | |
2068 | ||
2069 | switch (perf_event_type(hw_event)) { | |
2070 | case PERF_TYPE_HARDWARE: | |
2071 | hw_ops = hw_perf_counter_init(counter); | |
2072 | break; | |
2073 | ||
2074 | case PERF_TYPE_SOFTWARE: | |
2075 | hw_ops = sw_perf_counter_init(counter); | |
2076 | break; | |
2077 | ||
2078 | case PERF_TYPE_TRACEPOINT: | |
2079 | hw_ops = tp_perf_counter_init(counter); | |
2080 | break; | |
2081 | } | |
2082 | ||
2083 | if (!hw_ops) { | |
2084 | kfree(counter); | |
2085 | return NULL; | |
2086 | } | |
2087 | done: | |
2088 | counter->hw_ops = hw_ops; | |
2089 | ||
2090 | return counter; | |
2091 | } | |
2092 | ||
2093 | /** | |
2094 | * sys_perf_counter_open - open a performance counter, associate it to a task/cpu | |
2095 | * | |
2096 | * @hw_event_uptr: event type attributes for monitoring/sampling | |
2097 | * @pid: target pid | |
2098 | * @cpu: target cpu | |
2099 | * @group_fd: group leader counter fd | |
2100 | */ | |
2101 | SYSCALL_DEFINE5(perf_counter_open, | |
2102 | const struct perf_counter_hw_event __user *, hw_event_uptr, | |
2103 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) | |
2104 | { | |
2105 | struct perf_counter *counter, *group_leader; | |
2106 | struct perf_counter_hw_event hw_event; | |
2107 | struct perf_counter_context *ctx; | |
2108 | struct file *counter_file = NULL; | |
2109 | struct file *group_file = NULL; | |
2110 | int fput_needed = 0; | |
2111 | int fput_needed2 = 0; | |
2112 | int ret; | |
2113 | ||
2114 | /* for future expandability... */ | |
2115 | if (flags) | |
2116 | return -EINVAL; | |
2117 | ||
2118 | if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0) | |
2119 | return -EFAULT; | |
2120 | ||
2121 | /* | |
2122 | * Get the target context (task or percpu): | |
2123 | */ | |
2124 | ctx = find_get_context(pid, cpu); | |
2125 | if (IS_ERR(ctx)) | |
2126 | return PTR_ERR(ctx); | |
2127 | ||
2128 | /* | |
2129 | * Look up the group leader (we will attach this counter to it): | |
2130 | */ | |
2131 | group_leader = NULL; | |
2132 | if (group_fd != -1) { | |
2133 | ret = -EINVAL; | |
2134 | group_file = fget_light(group_fd, &fput_needed); | |
2135 | if (!group_file) | |
2136 | goto err_put_context; | |
2137 | if (group_file->f_op != &perf_fops) | |
2138 | goto err_put_context; | |
2139 | ||
2140 | group_leader = group_file->private_data; | |
2141 | /* | |
2142 | * Do not allow a recursive hierarchy (this new sibling | |
2143 | * becoming part of another group-sibling): | |
2144 | */ | |
2145 | if (group_leader->group_leader != group_leader) | |
2146 | goto err_put_context; | |
2147 | /* | |
2148 | * Do not allow to attach to a group in a different | |
2149 | * task or CPU context: | |
2150 | */ | |
2151 | if (group_leader->ctx != ctx) | |
2152 | goto err_put_context; | |
2153 | /* | |
2154 | * Only a group leader can be exclusive or pinned | |
2155 | */ | |
2156 | if (hw_event.exclusive || hw_event.pinned) | |
2157 | goto err_put_context; | |
2158 | } | |
2159 | ||
2160 | ret = -EINVAL; | |
2161 | counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader, | |
2162 | GFP_KERNEL); | |
2163 | if (!counter) | |
2164 | goto err_put_context; | |
2165 | ||
2166 | ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0); | |
2167 | if (ret < 0) | |
2168 | goto err_free_put_context; | |
2169 | ||
2170 | counter_file = fget_light(ret, &fput_needed2); | |
2171 | if (!counter_file) | |
2172 | goto err_free_put_context; | |
2173 | ||
2174 | counter->filp = counter_file; | |
2175 | mutex_lock(&ctx->mutex); | |
2176 | perf_install_in_context(ctx, counter, cpu); | |
2177 | mutex_unlock(&ctx->mutex); | |
2178 | ||
2179 | fput_light(counter_file, fput_needed2); | |
2180 | ||
2181 | out_fput: | |
2182 | fput_light(group_file, fput_needed); | |
2183 | ||
2184 | return ret; | |
2185 | ||
2186 | err_free_put_context: | |
2187 | kfree(counter); | |
2188 | ||
2189 | err_put_context: | |
2190 | put_context(ctx); | |
2191 | ||
2192 | goto out_fput; | |
2193 | } | |
2194 | ||
2195 | /* | |
2196 | * Initialize the perf_counter context in a task_struct: | |
2197 | */ | |
2198 | static void | |
2199 | __perf_counter_init_context(struct perf_counter_context *ctx, | |
2200 | struct task_struct *task) | |
2201 | { | |
2202 | memset(ctx, 0, sizeof(*ctx)); | |
2203 | spin_lock_init(&ctx->lock); | |
2204 | mutex_init(&ctx->mutex); | |
2205 | INIT_LIST_HEAD(&ctx->counter_list); | |
2206 | INIT_LIST_HEAD(&ctx->event_list); | |
2207 | ctx->task = task; | |
2208 | } | |
2209 | ||
2210 | /* | |
2211 | * inherit a counter from parent task to child task: | |
2212 | */ | |
2213 | static struct perf_counter * | |
2214 | inherit_counter(struct perf_counter *parent_counter, | |
2215 | struct task_struct *parent, | |
2216 | struct perf_counter_context *parent_ctx, | |
2217 | struct task_struct *child, | |
2218 | struct perf_counter *group_leader, | |
2219 | struct perf_counter_context *child_ctx) | |
2220 | { | |
2221 | struct perf_counter *child_counter; | |
2222 | ||
2223 | /* | |
2224 | * Instead of creating recursive hierarchies of counters, | |
2225 | * we link inherited counters back to the original parent, | |
2226 | * which has a filp for sure, which we use as the reference | |
2227 | * count: | |
2228 | */ | |
2229 | if (parent_counter->parent) | |
2230 | parent_counter = parent_counter->parent; | |
2231 | ||
2232 | child_counter = perf_counter_alloc(&parent_counter->hw_event, | |
2233 | parent_counter->cpu, child_ctx, | |
2234 | group_leader, GFP_KERNEL); | |
2235 | if (!child_counter) | |
2236 | return NULL; | |
2237 | ||
2238 | /* | |
2239 | * Link it up in the child's context: | |
2240 | */ | |
2241 | child_counter->task = child; | |
2242 | list_add_counter(child_counter, child_ctx); | |
2243 | child_ctx->nr_counters++; | |
2244 | ||
2245 | child_counter->parent = parent_counter; | |
2246 | /* | |
2247 | * inherit into child's child as well: | |
2248 | */ | |
2249 | child_counter->hw_event.inherit = 1; | |
2250 | ||
2251 | /* | |
2252 | * Get a reference to the parent filp - we will fput it | |
2253 | * when the child counter exits. This is safe to do because | |
2254 | * we are in the parent and we know that the filp still | |
2255 | * exists and has a nonzero count: | |
2256 | */ | |
2257 | atomic_long_inc(&parent_counter->filp->f_count); | |
2258 | ||
2259 | /* | |
2260 | * Link this into the parent counter's child list | |
2261 | */ | |
2262 | mutex_lock(&parent_counter->mutex); | |
2263 | list_add_tail(&child_counter->child_list, &parent_counter->child_list); | |
2264 | ||
2265 | /* | |
2266 | * Make the child state follow the state of the parent counter, | |
2267 | * not its hw_event.disabled bit. We hold the parent's mutex, | |
2268 | * so we won't race with perf_counter_{en,dis}able_family. | |
2269 | */ | |
2270 | if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE) | |
2271 | child_counter->state = PERF_COUNTER_STATE_INACTIVE; | |
2272 | else | |
2273 | child_counter->state = PERF_COUNTER_STATE_OFF; | |
2274 | ||
2275 | mutex_unlock(&parent_counter->mutex); | |
2276 | ||
2277 | return child_counter; | |
2278 | } | |
2279 | ||
2280 | static int inherit_group(struct perf_counter *parent_counter, | |
2281 | struct task_struct *parent, | |
2282 | struct perf_counter_context *parent_ctx, | |
2283 | struct task_struct *child, | |
2284 | struct perf_counter_context *child_ctx) | |
2285 | { | |
2286 | struct perf_counter *leader; | |
2287 | struct perf_counter *sub; | |
2288 | ||
2289 | leader = inherit_counter(parent_counter, parent, parent_ctx, | |
2290 | child, NULL, child_ctx); | |
2291 | if (!leader) | |
2292 | return -ENOMEM; | |
2293 | list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) { | |
2294 | if (!inherit_counter(sub, parent, parent_ctx, | |
2295 | child, leader, child_ctx)) | |
2296 | return -ENOMEM; | |
2297 | } | |
2298 | return 0; | |
2299 | } | |
2300 | ||
2301 | static void sync_child_counter(struct perf_counter *child_counter, | |
2302 | struct perf_counter *parent_counter) | |
2303 | { | |
2304 | u64 parent_val, child_val; | |
2305 | ||
2306 | parent_val = atomic64_read(&parent_counter->count); | |
2307 | child_val = atomic64_read(&child_counter->count); | |
2308 | ||
2309 | /* | |
2310 | * Add back the child's count to the parent's count: | |
2311 | */ | |
2312 | atomic64_add(child_val, &parent_counter->count); | |
2313 | ||
2314 | /* | |
2315 | * Remove this counter from the parent's list | |
2316 | */ | |
2317 | mutex_lock(&parent_counter->mutex); | |
2318 | list_del_init(&child_counter->child_list); | |
2319 | mutex_unlock(&parent_counter->mutex); | |
2320 | ||
2321 | /* | |
2322 | * Release the parent counter, if this was the last | |
2323 | * reference to it. | |
2324 | */ | |
2325 | fput(parent_counter->filp); | |
2326 | } | |
2327 | ||
2328 | static void | |
2329 | __perf_counter_exit_task(struct task_struct *child, | |
2330 | struct perf_counter *child_counter, | |
2331 | struct perf_counter_context *child_ctx) | |
2332 | { | |
2333 | struct perf_counter *parent_counter; | |
2334 | struct perf_counter *sub, *tmp; | |
2335 | ||
2336 | /* | |
2337 | * If we do not self-reap then we have to wait for the | |
2338 | * child task to unschedule (it will happen for sure), | |
2339 | * so that its counter is at its final count. (This | |
2340 | * condition triggers rarely - child tasks usually get | |
2341 | * off their CPU before the parent has a chance to | |
2342 | * get this far into the reaping action) | |
2343 | */ | |
2344 | if (child != current) { | |
2345 | wait_task_inactive(child, 0); | |
2346 | list_del_init(&child_counter->list_entry); | |
2347 | } else { | |
2348 | struct perf_cpu_context *cpuctx; | |
2349 | unsigned long flags; | |
2350 | u64 perf_flags; | |
2351 | ||
2352 | /* | |
2353 | * Disable and unlink this counter. | |
2354 | * | |
2355 | * Be careful about zapping the list - IRQ/NMI context | |
2356 | * could still be processing it: | |
2357 | */ | |
2358 | curr_rq_lock_irq_save(&flags); | |
2359 | perf_flags = hw_perf_save_disable(); | |
2360 | ||
2361 | cpuctx = &__get_cpu_var(perf_cpu_context); | |
2362 | ||
2363 | group_sched_out(child_counter, cpuctx, child_ctx); | |
2364 | ||
2365 | list_del_init(&child_counter->list_entry); | |
2366 | ||
2367 | child_ctx->nr_counters--; | |
2368 | ||
2369 | hw_perf_restore(perf_flags); | |
2370 | curr_rq_unlock_irq_restore(&flags); | |
2371 | } | |
2372 | ||
2373 | parent_counter = child_counter->parent; | |
2374 | /* | |
2375 | * It can happen that parent exits first, and has counters | |
2376 | * that are still around due to the child reference. These | |
2377 | * counters need to be zapped - but otherwise linger. | |
2378 | */ | |
2379 | if (parent_counter) { | |
2380 | sync_child_counter(child_counter, parent_counter); | |
2381 | list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list, | |
2382 | list_entry) { | |
2383 | if (sub->parent) { | |
2384 | sync_child_counter(sub, sub->parent); | |
2385 | free_counter(sub); | |
2386 | } | |
2387 | } | |
2388 | free_counter(child_counter); | |
2389 | } | |
2390 | } | |
2391 | ||
2392 | /* | |
2393 | * When a child task exits, feed back counter values to parent counters. | |
2394 | * | |
2395 | * Note: we may be running in child context, but the PID is not hashed | |
2396 | * anymore so new counters will not be added. | |
2397 | */ | |
2398 | void perf_counter_exit_task(struct task_struct *child) | |
2399 | { | |
2400 | struct perf_counter *child_counter, *tmp; | |
2401 | struct perf_counter_context *child_ctx; | |
2402 | ||
2403 | child_ctx = &child->perf_counter_ctx; | |
2404 | ||
2405 | if (likely(!child_ctx->nr_counters)) | |
2406 | return; | |
2407 | ||
2408 | list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list, | |
2409 | list_entry) | |
2410 | __perf_counter_exit_task(child, child_counter, child_ctx); | |
2411 | } | |
2412 | ||
2413 | /* | |
2414 | * Initialize the perf_counter context in task_struct | |
2415 | */ | |
2416 | void perf_counter_init_task(struct task_struct *child) | |
2417 | { | |
2418 | struct perf_counter_context *child_ctx, *parent_ctx; | |
2419 | struct perf_counter *counter; | |
2420 | struct task_struct *parent = current; | |
2421 | ||
2422 | child_ctx = &child->perf_counter_ctx; | |
2423 | parent_ctx = &parent->perf_counter_ctx; | |
2424 | ||
2425 | __perf_counter_init_context(child_ctx, child); | |
2426 | ||
2427 | /* | |
2428 | * This is executed from the parent task context, so inherit | |
2429 | * counters that have been marked for cloning: | |
2430 | */ | |
2431 | ||
2432 | if (likely(!parent_ctx->nr_counters)) | |
2433 | return; | |
2434 | ||
2435 | /* | |
2436 | * Lock the parent list. No need to lock the child - not PID | |
2437 | * hashed yet and not running, so nobody can access it. | |
2438 | */ | |
2439 | mutex_lock(&parent_ctx->mutex); | |
2440 | ||
2441 | /* | |
2442 | * We dont have to disable NMIs - we are only looking at | |
2443 | * the list, not manipulating it: | |
2444 | */ | |
2445 | list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) { | |
2446 | if (!counter->hw_event.inherit) | |
2447 | continue; | |
2448 | ||
2449 | if (inherit_group(counter, parent, | |
2450 | parent_ctx, child, child_ctx)) | |
2451 | break; | |
2452 | } | |
2453 | ||
2454 | mutex_unlock(&parent_ctx->mutex); | |
2455 | } | |
2456 | ||
2457 | static void __cpuinit perf_counter_init_cpu(int cpu) | |
2458 | { | |
2459 | struct perf_cpu_context *cpuctx; | |
2460 | ||
2461 | cpuctx = &per_cpu(perf_cpu_context, cpu); | |
2462 | __perf_counter_init_context(&cpuctx->ctx, NULL); | |
2463 | ||
2464 | mutex_lock(&perf_resource_mutex); | |
2465 | cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu; | |
2466 | mutex_unlock(&perf_resource_mutex); | |
2467 | ||
2468 | hw_perf_counter_setup(cpu); | |
2469 | } | |
2470 | ||
2471 | #ifdef CONFIG_HOTPLUG_CPU | |
2472 | static void __perf_counter_exit_cpu(void *info) | |
2473 | { | |
2474 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
2475 | struct perf_counter_context *ctx = &cpuctx->ctx; | |
2476 | struct perf_counter *counter, *tmp; | |
2477 | ||
2478 | list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry) | |
2479 | __perf_counter_remove_from_context(counter); | |
2480 | } | |
2481 | static void perf_counter_exit_cpu(int cpu) | |
2482 | { | |
2483 | struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); | |
2484 | struct perf_counter_context *ctx = &cpuctx->ctx; | |
2485 | ||
2486 | mutex_lock(&ctx->mutex); | |
2487 | smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1); | |
2488 | mutex_unlock(&ctx->mutex); | |
2489 | } | |
2490 | #else | |
2491 | static inline void perf_counter_exit_cpu(int cpu) { } | |
2492 | #endif | |
2493 | ||
2494 | static int __cpuinit | |
2495 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) | |
2496 | { | |
2497 | unsigned int cpu = (long)hcpu; | |
2498 | ||
2499 | switch (action) { | |
2500 | ||
2501 | case CPU_UP_PREPARE: | |
2502 | case CPU_UP_PREPARE_FROZEN: | |
2503 | perf_counter_init_cpu(cpu); | |
2504 | break; | |
2505 | ||
2506 | case CPU_DOWN_PREPARE: | |
2507 | case CPU_DOWN_PREPARE_FROZEN: | |
2508 | perf_counter_exit_cpu(cpu); | |
2509 | break; | |
2510 | ||
2511 | default: | |
2512 | break; | |
2513 | } | |
2514 | ||
2515 | return NOTIFY_OK; | |
2516 | } | |
2517 | ||
2518 | static struct notifier_block __cpuinitdata perf_cpu_nb = { | |
2519 | .notifier_call = perf_cpu_notify, | |
2520 | }; | |
2521 | ||
2522 | static int __init perf_counter_init(void) | |
2523 | { | |
2524 | perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE, | |
2525 | (void *)(long)smp_processor_id()); | |
2526 | register_cpu_notifier(&perf_cpu_nb); | |
2527 | ||
2528 | return 0; | |
2529 | } | |
2530 | early_initcall(perf_counter_init); | |
2531 | ||
2532 | static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf) | |
2533 | { | |
2534 | return sprintf(buf, "%d\n", perf_reserved_percpu); | |
2535 | } | |
2536 | ||
2537 | static ssize_t | |
2538 | perf_set_reserve_percpu(struct sysdev_class *class, | |
2539 | const char *buf, | |
2540 | size_t count) | |
2541 | { | |
2542 | struct perf_cpu_context *cpuctx; | |
2543 | unsigned long val; | |
2544 | int err, cpu, mpt; | |
2545 | ||
2546 | err = strict_strtoul(buf, 10, &val); | |
2547 | if (err) | |
2548 | return err; | |
2549 | if (val > perf_max_counters) | |
2550 | return -EINVAL; | |
2551 | ||
2552 | mutex_lock(&perf_resource_mutex); | |
2553 | perf_reserved_percpu = val; | |
2554 | for_each_online_cpu(cpu) { | |
2555 | cpuctx = &per_cpu(perf_cpu_context, cpu); | |
2556 | spin_lock_irq(&cpuctx->ctx.lock); | |
2557 | mpt = min(perf_max_counters - cpuctx->ctx.nr_counters, | |
2558 | perf_max_counters - perf_reserved_percpu); | |
2559 | cpuctx->max_pertask = mpt; | |
2560 | spin_unlock_irq(&cpuctx->ctx.lock); | |
2561 | } | |
2562 | mutex_unlock(&perf_resource_mutex); | |
2563 | ||
2564 | return count; | |
2565 | } | |
2566 | ||
2567 | static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf) | |
2568 | { | |
2569 | return sprintf(buf, "%d\n", perf_overcommit); | |
2570 | } | |
2571 | ||
2572 | static ssize_t | |
2573 | perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count) | |
2574 | { | |
2575 | unsigned long val; | |
2576 | int err; | |
2577 | ||
2578 | err = strict_strtoul(buf, 10, &val); | |
2579 | if (err) | |
2580 | return err; | |
2581 | if (val > 1) | |
2582 | return -EINVAL; | |
2583 | ||
2584 | mutex_lock(&perf_resource_mutex); | |
2585 | perf_overcommit = val; | |
2586 | mutex_unlock(&perf_resource_mutex); | |
2587 | ||
2588 | return count; | |
2589 | } | |
2590 | ||
2591 | static SYSDEV_CLASS_ATTR( | |
2592 | reserve_percpu, | |
2593 | 0644, | |
2594 | perf_show_reserve_percpu, | |
2595 | perf_set_reserve_percpu | |
2596 | ); | |
2597 | ||
2598 | static SYSDEV_CLASS_ATTR( | |
2599 | overcommit, | |
2600 | 0644, | |
2601 | perf_show_overcommit, | |
2602 | perf_set_overcommit | |
2603 | ); | |
2604 | ||
2605 | static struct attribute *perfclass_attrs[] = { | |
2606 | &attr_reserve_percpu.attr, | |
2607 | &attr_overcommit.attr, | |
2608 | NULL | |
2609 | }; | |
2610 | ||
2611 | static struct attribute_group perfclass_attr_group = { | |
2612 | .attrs = perfclass_attrs, | |
2613 | .name = "perf_counters", | |
2614 | }; | |
2615 | ||
2616 | static int __init perf_counter_sysfs_init(void) | |
2617 | { | |
2618 | return sysfs_create_group(&cpu_sysdev_class.kset.kobj, | |
2619 | &perfclass_attr_group); | |
2620 | } | |
2621 | device_initcall(perf_counter_sysfs_init); |