]>
Commit | Line | Data |
---|---|---|
0793a61d TG |
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> | |
04289bb9 | 13 | #include <linux/file.h> |
0793a61d TG |
14 | #include <linux/poll.h> |
15 | #include <linux/sysfs.h> | |
16 | #include <linux/ptrace.h> | |
17 | #include <linux/percpu.h> | |
18 | #include <linux/uaccess.h> | |
19 | #include <linux/syscalls.h> | |
20 | #include <linux/anon_inodes.h> | |
21 | #include <linux/perf_counter.h> | |
22 | ||
23 | /* | |
24 | * Each CPU has a list of per CPU counters: | |
25 | */ | |
26 | DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context); | |
27 | ||
088e2852 | 28 | int perf_max_counters __read_mostly = 1; |
0793a61d TG |
29 | static int perf_reserved_percpu __read_mostly; |
30 | static int perf_overcommit __read_mostly = 1; | |
31 | ||
32 | /* | |
33 | * Mutex for (sysadmin-configurable) counter reservations: | |
34 | */ | |
35 | static DEFINE_MUTEX(perf_resource_mutex); | |
36 | ||
37 | /* | |
38 | * Architecture provided APIs - weak aliases: | |
39 | */ | |
5c92d124 | 40 | extern __weak const struct hw_perf_counter_ops * |
621a01ea | 41 | hw_perf_counter_init(struct perf_counter *counter) |
0793a61d | 42 | { |
621a01ea | 43 | return ERR_PTR(-EINVAL); |
0793a61d TG |
44 | } |
45 | ||
01b2838c | 46 | u64 __weak hw_perf_save_disable(void) { return 0; } |
ee06094f | 47 | void __weak hw_perf_restore(u64 ctrl) { } |
5c92d124 | 48 | void __weak hw_perf_counter_setup(void) { } |
0793a61d | 49 | |
04289bb9 IM |
50 | static void |
51 | list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx) | |
52 | { | |
53 | struct perf_counter *group_leader = counter->group_leader; | |
54 | ||
55 | /* | |
56 | * Depending on whether it is a standalone or sibling counter, | |
57 | * add it straight to the context's counter list, or to the group | |
58 | * leader's sibling list: | |
59 | */ | |
60 | if (counter->group_leader == counter) | |
61 | list_add_tail(&counter->list_entry, &ctx->counter_list); | |
62 | else | |
63 | list_add_tail(&counter->list_entry, &group_leader->sibling_list); | |
64 | } | |
65 | ||
66 | static void | |
67 | list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx) | |
68 | { | |
69 | struct perf_counter *sibling, *tmp; | |
70 | ||
71 | list_del_init(&counter->list_entry); | |
72 | ||
04289bb9 IM |
73 | /* |
74 | * If this was a group counter with sibling counters then | |
75 | * upgrade the siblings to singleton counters by adding them | |
76 | * to the context list directly: | |
77 | */ | |
78 | list_for_each_entry_safe(sibling, tmp, | |
79 | &counter->sibling_list, list_entry) { | |
80 | ||
81 | list_del_init(&sibling->list_entry); | |
82 | list_add_tail(&sibling->list_entry, &ctx->counter_list); | |
04289bb9 IM |
83 | sibling->group_leader = sibling; |
84 | } | |
85 | } | |
86 | ||
0793a61d TG |
87 | /* |
88 | * Cross CPU call to remove a performance counter | |
89 | * | |
90 | * We disable the counter on the hardware level first. After that we | |
91 | * remove it from the context list. | |
92 | */ | |
04289bb9 | 93 | static void __perf_counter_remove_from_context(void *info) |
0793a61d TG |
94 | { |
95 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
96 | struct perf_counter *counter = info; | |
97 | struct perf_counter_context *ctx = counter->ctx; | |
9b51f66d | 98 | unsigned long flags; |
5c92d124 | 99 | u64 perf_flags; |
0793a61d TG |
100 | |
101 | /* | |
102 | * If this is a task context, we need to check whether it is | |
103 | * the current task context of this cpu. If not it has been | |
104 | * scheduled out before the smp call arrived. | |
105 | */ | |
106 | if (ctx->task && cpuctx->task_ctx != ctx) | |
107 | return; | |
108 | ||
9b51f66d | 109 | spin_lock_irqsave(&ctx->lock, flags); |
0793a61d | 110 | |
6a930700 | 111 | if (counter->state == PERF_COUNTER_STATE_ACTIVE) { |
621a01ea | 112 | counter->hw_ops->hw_perf_counter_disable(counter); |
6a930700 | 113 | counter->state = PERF_COUNTER_STATE_INACTIVE; |
0793a61d TG |
114 | ctx->nr_active--; |
115 | cpuctx->active_oncpu--; | |
116 | counter->task = NULL; | |
117 | } | |
118 | ctx->nr_counters--; | |
119 | ||
120 | /* | |
121 | * Protect the list operation against NMI by disabling the | |
122 | * counters on a global level. NOP for non NMI based counters. | |
123 | */ | |
01b2838c | 124 | perf_flags = hw_perf_save_disable(); |
04289bb9 | 125 | list_del_counter(counter, ctx); |
01b2838c | 126 | hw_perf_restore(perf_flags); |
0793a61d TG |
127 | |
128 | if (!ctx->task) { | |
129 | /* | |
130 | * Allow more per task counters with respect to the | |
131 | * reservation: | |
132 | */ | |
133 | cpuctx->max_pertask = | |
134 | min(perf_max_counters - ctx->nr_counters, | |
135 | perf_max_counters - perf_reserved_percpu); | |
136 | } | |
137 | ||
9b51f66d | 138 | spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
139 | } |
140 | ||
141 | ||
142 | /* | |
143 | * Remove the counter from a task's (or a CPU's) list of counters. | |
144 | * | |
145 | * Must be called with counter->mutex held. | |
146 | * | |
147 | * CPU counters are removed with a smp call. For task counters we only | |
148 | * call when the task is on a CPU. | |
149 | */ | |
04289bb9 | 150 | static void perf_counter_remove_from_context(struct perf_counter *counter) |
0793a61d TG |
151 | { |
152 | struct perf_counter_context *ctx = counter->ctx; | |
153 | struct task_struct *task = ctx->task; | |
154 | ||
155 | if (!task) { | |
156 | /* | |
157 | * Per cpu counters are removed via an smp call and | |
158 | * the removal is always sucessful. | |
159 | */ | |
160 | smp_call_function_single(counter->cpu, | |
04289bb9 | 161 | __perf_counter_remove_from_context, |
0793a61d TG |
162 | counter, 1); |
163 | return; | |
164 | } | |
165 | ||
166 | retry: | |
04289bb9 | 167 | task_oncpu_function_call(task, __perf_counter_remove_from_context, |
0793a61d TG |
168 | counter); |
169 | ||
170 | spin_lock_irq(&ctx->lock); | |
171 | /* | |
172 | * If the context is active we need to retry the smp call. | |
173 | */ | |
04289bb9 | 174 | if (ctx->nr_active && !list_empty(&counter->list_entry)) { |
0793a61d TG |
175 | spin_unlock_irq(&ctx->lock); |
176 | goto retry; | |
177 | } | |
178 | ||
179 | /* | |
180 | * The lock prevents that this context is scheduled in so we | |
04289bb9 | 181 | * can remove the counter safely, if the call above did not |
0793a61d TG |
182 | * succeed. |
183 | */ | |
04289bb9 | 184 | if (!list_empty(&counter->list_entry)) { |
0793a61d | 185 | ctx->nr_counters--; |
04289bb9 | 186 | list_del_counter(counter, ctx); |
0793a61d TG |
187 | counter->task = NULL; |
188 | } | |
189 | spin_unlock_irq(&ctx->lock); | |
190 | } | |
191 | ||
192 | /* | |
193 | * Cross CPU call to install and enable a preformance counter | |
194 | */ | |
195 | static void __perf_install_in_context(void *info) | |
196 | { | |
197 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
198 | struct perf_counter *counter = info; | |
199 | struct perf_counter_context *ctx = counter->ctx; | |
200 | int cpu = smp_processor_id(); | |
9b51f66d | 201 | unsigned long flags; |
5c92d124 | 202 | u64 perf_flags; |
0793a61d TG |
203 | |
204 | /* | |
205 | * If this is a task context, we need to check whether it is | |
206 | * the current task context of this cpu. If not it has been | |
207 | * scheduled out before the smp call arrived. | |
208 | */ | |
209 | if (ctx->task && cpuctx->task_ctx != ctx) | |
210 | return; | |
211 | ||
9b51f66d | 212 | spin_lock_irqsave(&ctx->lock, flags); |
0793a61d TG |
213 | |
214 | /* | |
215 | * Protect the list operation against NMI by disabling the | |
216 | * counters on a global level. NOP for non NMI based counters. | |
217 | */ | |
01b2838c | 218 | perf_flags = hw_perf_save_disable(); |
04289bb9 | 219 | list_add_counter(counter, ctx); |
01b2838c | 220 | hw_perf_restore(perf_flags); |
0793a61d TG |
221 | |
222 | ctx->nr_counters++; | |
223 | ||
224 | if (cpuctx->active_oncpu < perf_max_counters) { | |
6a930700 | 225 | counter->state = PERF_COUNTER_STATE_ACTIVE; |
0793a61d TG |
226 | counter->oncpu = cpu; |
227 | ctx->nr_active++; | |
228 | cpuctx->active_oncpu++; | |
ee06094f | 229 | counter->hw_ops->hw_perf_counter_enable(counter); |
0793a61d TG |
230 | } |
231 | ||
232 | if (!ctx->task && cpuctx->max_pertask) | |
233 | cpuctx->max_pertask--; | |
234 | ||
9b51f66d | 235 | spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
236 | } |
237 | ||
238 | /* | |
239 | * Attach a performance counter to a context | |
240 | * | |
241 | * First we add the counter to the list with the hardware enable bit | |
242 | * in counter->hw_config cleared. | |
243 | * | |
244 | * If the counter is attached to a task which is on a CPU we use a smp | |
245 | * call to enable it in the task context. The task might have been | |
246 | * scheduled away, but we check this in the smp call again. | |
247 | */ | |
248 | static void | |
249 | perf_install_in_context(struct perf_counter_context *ctx, | |
250 | struct perf_counter *counter, | |
251 | int cpu) | |
252 | { | |
253 | struct task_struct *task = ctx->task; | |
254 | ||
255 | counter->ctx = ctx; | |
256 | if (!task) { | |
257 | /* | |
258 | * Per cpu counters are installed via an smp call and | |
259 | * the install is always sucessful. | |
260 | */ | |
261 | smp_call_function_single(cpu, __perf_install_in_context, | |
262 | counter, 1); | |
263 | return; | |
264 | } | |
265 | ||
266 | counter->task = task; | |
267 | retry: | |
268 | task_oncpu_function_call(task, __perf_install_in_context, | |
269 | counter); | |
270 | ||
271 | spin_lock_irq(&ctx->lock); | |
272 | /* | |
0793a61d TG |
273 | * we need to retry the smp call. |
274 | */ | |
04289bb9 | 275 | if (ctx->nr_active && list_empty(&counter->list_entry)) { |
0793a61d TG |
276 | spin_unlock_irq(&ctx->lock); |
277 | goto retry; | |
278 | } | |
279 | ||
280 | /* | |
281 | * The lock prevents that this context is scheduled in so we | |
282 | * can add the counter safely, if it the call above did not | |
283 | * succeed. | |
284 | */ | |
04289bb9 IM |
285 | if (list_empty(&counter->list_entry)) { |
286 | list_add_counter(counter, ctx); | |
0793a61d TG |
287 | ctx->nr_counters++; |
288 | } | |
289 | spin_unlock_irq(&ctx->lock); | |
290 | } | |
291 | ||
04289bb9 IM |
292 | static void |
293 | counter_sched_out(struct perf_counter *counter, | |
294 | struct perf_cpu_context *cpuctx, | |
295 | struct perf_counter_context *ctx) | |
296 | { | |
6a930700 | 297 | if (counter->state != PERF_COUNTER_STATE_ACTIVE) |
04289bb9 IM |
298 | return; |
299 | ||
621a01ea | 300 | counter->hw_ops->hw_perf_counter_disable(counter); |
6a930700 IM |
301 | counter->state = PERF_COUNTER_STATE_INACTIVE; |
302 | counter->oncpu = -1; | |
04289bb9 IM |
303 | |
304 | cpuctx->active_oncpu--; | |
305 | ctx->nr_active--; | |
306 | } | |
307 | ||
308 | static void | |
309 | group_sched_out(struct perf_counter *group_counter, | |
310 | struct perf_cpu_context *cpuctx, | |
311 | struct perf_counter_context *ctx) | |
312 | { | |
313 | struct perf_counter *counter; | |
314 | ||
315 | counter_sched_out(group_counter, cpuctx, ctx); | |
316 | ||
317 | /* | |
318 | * Schedule out siblings (if any): | |
319 | */ | |
320 | list_for_each_entry(counter, &group_counter->sibling_list, list_entry) | |
321 | counter_sched_out(counter, cpuctx, ctx); | |
322 | } | |
323 | ||
0793a61d TG |
324 | /* |
325 | * Called from scheduler to remove the counters of the current task, | |
326 | * with interrupts disabled. | |
327 | * | |
328 | * We stop each counter and update the counter value in counter->count. | |
329 | * | |
330 | * This does not protect us against NMI, but hw_perf_counter_disable() | |
331 | * sets the disabled bit in the control field of counter _before_ | |
332 | * accessing the counter control register. If a NMI hits, then it will | |
333 | * not restart the counter. | |
334 | */ | |
335 | void perf_counter_task_sched_out(struct task_struct *task, int cpu) | |
336 | { | |
337 | struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); | |
338 | struct perf_counter_context *ctx = &task->perf_counter_ctx; | |
339 | struct perf_counter *counter; | |
340 | ||
341 | if (likely(!cpuctx->task_ctx)) | |
342 | return; | |
343 | ||
344 | spin_lock(&ctx->lock); | |
04289bb9 IM |
345 | if (ctx->nr_active) { |
346 | list_for_each_entry(counter, &ctx->counter_list, list_entry) | |
347 | group_sched_out(counter, cpuctx, ctx); | |
0793a61d TG |
348 | } |
349 | spin_unlock(&ctx->lock); | |
350 | cpuctx->task_ctx = NULL; | |
351 | } | |
352 | ||
04289bb9 IM |
353 | static void |
354 | counter_sched_in(struct perf_counter *counter, | |
355 | struct perf_cpu_context *cpuctx, | |
356 | struct perf_counter_context *ctx, | |
357 | int cpu) | |
358 | { | |
6a930700 | 359 | if (counter->state == PERF_COUNTER_STATE_OFF) |
1d1c7ddb IM |
360 | return; |
361 | ||
621a01ea | 362 | counter->hw_ops->hw_perf_counter_enable(counter); |
6a930700 | 363 | counter->state = PERF_COUNTER_STATE_ACTIVE; |
04289bb9 IM |
364 | counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */ |
365 | ||
366 | cpuctx->active_oncpu++; | |
367 | ctx->nr_active++; | |
368 | } | |
369 | ||
370 | static void | |
371 | group_sched_in(struct perf_counter *group_counter, | |
372 | struct perf_cpu_context *cpuctx, | |
373 | struct perf_counter_context *ctx, | |
374 | int cpu) | |
375 | { | |
376 | struct perf_counter *counter; | |
377 | ||
378 | counter_sched_in(group_counter, cpuctx, ctx, cpu); | |
379 | ||
380 | /* | |
381 | * Schedule in siblings as one group (if any): | |
382 | */ | |
383 | list_for_each_entry(counter, &group_counter->sibling_list, list_entry) | |
384 | counter_sched_in(counter, cpuctx, ctx, cpu); | |
385 | } | |
386 | ||
0793a61d TG |
387 | /* |
388 | * Called from scheduler to add the counters of the current task | |
389 | * with interrupts disabled. | |
390 | * | |
391 | * We restore the counter value and then enable it. | |
392 | * | |
393 | * This does not protect us against NMI, but hw_perf_counter_enable() | |
394 | * sets the enabled bit in the control field of counter _before_ | |
395 | * accessing the counter control register. If a NMI hits, then it will | |
396 | * keep the counter running. | |
397 | */ | |
398 | void perf_counter_task_sched_in(struct task_struct *task, int cpu) | |
399 | { | |
400 | struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); | |
401 | struct perf_counter_context *ctx = &task->perf_counter_ctx; | |
402 | struct perf_counter *counter; | |
403 | ||
404 | if (likely(!ctx->nr_counters)) | |
405 | return; | |
406 | ||
407 | spin_lock(&ctx->lock); | |
04289bb9 | 408 | list_for_each_entry(counter, &ctx->counter_list, list_entry) { |
0793a61d TG |
409 | if (ctx->nr_active == cpuctx->max_pertask) |
410 | break; | |
04289bb9 IM |
411 | |
412 | /* | |
413 | * Listen to the 'cpu' scheduling filter constraint | |
414 | * of counters: | |
415 | */ | |
0793a61d TG |
416 | if (counter->cpu != -1 && counter->cpu != cpu) |
417 | continue; | |
418 | ||
04289bb9 | 419 | group_sched_in(counter, cpuctx, ctx, cpu); |
0793a61d TG |
420 | } |
421 | spin_unlock(&ctx->lock); | |
04289bb9 | 422 | |
0793a61d TG |
423 | cpuctx->task_ctx = ctx; |
424 | } | |
425 | ||
1d1c7ddb IM |
426 | int perf_counter_task_disable(void) |
427 | { | |
428 | struct task_struct *curr = current; | |
429 | struct perf_counter_context *ctx = &curr->perf_counter_ctx; | |
430 | struct perf_counter *counter; | |
431 | u64 perf_flags; | |
432 | int cpu; | |
433 | ||
434 | if (likely(!ctx->nr_counters)) | |
435 | return 0; | |
436 | ||
437 | local_irq_disable(); | |
438 | cpu = smp_processor_id(); | |
439 | ||
440 | perf_counter_task_sched_out(curr, cpu); | |
441 | ||
442 | spin_lock(&ctx->lock); | |
443 | ||
444 | /* | |
445 | * Disable all the counters: | |
446 | */ | |
447 | perf_flags = hw_perf_save_disable(); | |
448 | ||
9b51f66d | 449 | list_for_each_entry(counter, &ctx->counter_list, list_entry) |
6a930700 | 450 | counter->state = PERF_COUNTER_STATE_OFF; |
9b51f66d | 451 | |
1d1c7ddb IM |
452 | hw_perf_restore(perf_flags); |
453 | ||
454 | spin_unlock(&ctx->lock); | |
455 | ||
456 | local_irq_enable(); | |
457 | ||
458 | return 0; | |
459 | } | |
460 | ||
461 | int perf_counter_task_enable(void) | |
462 | { | |
463 | struct task_struct *curr = current; | |
464 | struct perf_counter_context *ctx = &curr->perf_counter_ctx; | |
465 | struct perf_counter *counter; | |
466 | u64 perf_flags; | |
467 | int cpu; | |
468 | ||
469 | if (likely(!ctx->nr_counters)) | |
470 | return 0; | |
471 | ||
472 | local_irq_disable(); | |
473 | cpu = smp_processor_id(); | |
474 | ||
475 | spin_lock(&ctx->lock); | |
476 | ||
477 | /* | |
478 | * Disable all the counters: | |
479 | */ | |
480 | perf_flags = hw_perf_save_disable(); | |
481 | ||
482 | list_for_each_entry(counter, &ctx->counter_list, list_entry) { | |
6a930700 | 483 | if (counter->state != PERF_COUNTER_STATE_OFF) |
1d1c7ddb | 484 | continue; |
6a930700 | 485 | counter->state = PERF_COUNTER_STATE_INACTIVE; |
1d1c7ddb IM |
486 | } |
487 | hw_perf_restore(perf_flags); | |
488 | ||
489 | spin_unlock(&ctx->lock); | |
490 | ||
491 | perf_counter_task_sched_in(curr, cpu); | |
492 | ||
493 | local_irq_enable(); | |
494 | ||
495 | return 0; | |
496 | } | |
497 | ||
0793a61d TG |
498 | void perf_counter_task_tick(struct task_struct *curr, int cpu) |
499 | { | |
500 | struct perf_counter_context *ctx = &curr->perf_counter_ctx; | |
501 | struct perf_counter *counter; | |
5c92d124 | 502 | u64 perf_flags; |
0793a61d TG |
503 | |
504 | if (likely(!ctx->nr_counters)) | |
505 | return; | |
506 | ||
507 | perf_counter_task_sched_out(curr, cpu); | |
508 | ||
509 | spin_lock(&ctx->lock); | |
510 | ||
511 | /* | |
04289bb9 | 512 | * Rotate the first entry last (works just fine for group counters too): |
0793a61d | 513 | */ |
01b2838c | 514 | perf_flags = hw_perf_save_disable(); |
04289bb9 IM |
515 | list_for_each_entry(counter, &ctx->counter_list, list_entry) { |
516 | list_del(&counter->list_entry); | |
517 | list_add_tail(&counter->list_entry, &ctx->counter_list); | |
0793a61d TG |
518 | break; |
519 | } | |
01b2838c | 520 | hw_perf_restore(perf_flags); |
0793a61d TG |
521 | |
522 | spin_unlock(&ctx->lock); | |
523 | ||
524 | perf_counter_task_sched_in(curr, cpu); | |
525 | } | |
526 | ||
0793a61d TG |
527 | /* |
528 | * Cross CPU call to read the hardware counter | |
529 | */ | |
530 | static void __hw_perf_counter_read(void *info) | |
531 | { | |
621a01ea IM |
532 | struct perf_counter *counter = info; |
533 | ||
534 | counter->hw_ops->hw_perf_counter_read(counter); | |
0793a61d TG |
535 | } |
536 | ||
04289bb9 | 537 | static u64 perf_counter_read(struct perf_counter *counter) |
0793a61d TG |
538 | { |
539 | /* | |
540 | * If counter is enabled and currently active on a CPU, update the | |
541 | * value in the counter structure: | |
542 | */ | |
6a930700 | 543 | if (counter->state == PERF_COUNTER_STATE_ACTIVE) { |
0793a61d TG |
544 | smp_call_function_single(counter->oncpu, |
545 | __hw_perf_counter_read, counter, 1); | |
546 | } | |
547 | ||
ee06094f | 548 | return atomic64_read(&counter->count); |
0793a61d TG |
549 | } |
550 | ||
551 | /* | |
552 | * Cross CPU call to switch performance data pointers | |
553 | */ | |
554 | static void __perf_switch_irq_data(void *info) | |
555 | { | |
556 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
557 | struct perf_counter *counter = info; | |
558 | struct perf_counter_context *ctx = counter->ctx; | |
559 | struct perf_data *oldirqdata = counter->irqdata; | |
560 | ||
561 | /* | |
562 | * If this is a task context, we need to check whether it is | |
563 | * the current task context of this cpu. If not it has been | |
564 | * scheduled out before the smp call arrived. | |
565 | */ | |
566 | if (ctx->task) { | |
567 | if (cpuctx->task_ctx != ctx) | |
568 | return; | |
569 | spin_lock(&ctx->lock); | |
570 | } | |
571 | ||
572 | /* Change the pointer NMI safe */ | |
573 | atomic_long_set((atomic_long_t *)&counter->irqdata, | |
574 | (unsigned long) counter->usrdata); | |
575 | counter->usrdata = oldirqdata; | |
576 | ||
577 | if (ctx->task) | |
578 | spin_unlock(&ctx->lock); | |
579 | } | |
580 | ||
581 | static struct perf_data *perf_switch_irq_data(struct perf_counter *counter) | |
582 | { | |
583 | struct perf_counter_context *ctx = counter->ctx; | |
584 | struct perf_data *oldirqdata = counter->irqdata; | |
585 | struct task_struct *task = ctx->task; | |
586 | ||
587 | if (!task) { | |
588 | smp_call_function_single(counter->cpu, | |
589 | __perf_switch_irq_data, | |
590 | counter, 1); | |
591 | return counter->usrdata; | |
592 | } | |
593 | ||
594 | retry: | |
595 | spin_lock_irq(&ctx->lock); | |
6a930700 | 596 | if (counter->state != PERF_COUNTER_STATE_ACTIVE) { |
0793a61d TG |
597 | counter->irqdata = counter->usrdata; |
598 | counter->usrdata = oldirqdata; | |
599 | spin_unlock_irq(&ctx->lock); | |
600 | return oldirqdata; | |
601 | } | |
602 | spin_unlock_irq(&ctx->lock); | |
603 | task_oncpu_function_call(task, __perf_switch_irq_data, counter); | |
604 | /* Might have failed, because task was scheduled out */ | |
605 | if (counter->irqdata == oldirqdata) | |
606 | goto retry; | |
607 | ||
608 | return counter->usrdata; | |
609 | } | |
610 | ||
611 | static void put_context(struct perf_counter_context *ctx) | |
612 | { | |
613 | if (ctx->task) | |
614 | put_task_struct(ctx->task); | |
615 | } | |
616 | ||
617 | static struct perf_counter_context *find_get_context(pid_t pid, int cpu) | |
618 | { | |
619 | struct perf_cpu_context *cpuctx; | |
620 | struct perf_counter_context *ctx; | |
621 | struct task_struct *task; | |
622 | ||
623 | /* | |
624 | * If cpu is not a wildcard then this is a percpu counter: | |
625 | */ | |
626 | if (cpu != -1) { | |
627 | /* Must be root to operate on a CPU counter: */ | |
628 | if (!capable(CAP_SYS_ADMIN)) | |
629 | return ERR_PTR(-EACCES); | |
630 | ||
631 | if (cpu < 0 || cpu > num_possible_cpus()) | |
632 | return ERR_PTR(-EINVAL); | |
633 | ||
634 | /* | |
635 | * We could be clever and allow to attach a counter to an | |
636 | * offline CPU and activate it when the CPU comes up, but | |
637 | * that's for later. | |
638 | */ | |
639 | if (!cpu_isset(cpu, cpu_online_map)) | |
640 | return ERR_PTR(-ENODEV); | |
641 | ||
642 | cpuctx = &per_cpu(perf_cpu_context, cpu); | |
643 | ctx = &cpuctx->ctx; | |
644 | ||
0793a61d TG |
645 | return ctx; |
646 | } | |
647 | ||
648 | rcu_read_lock(); | |
649 | if (!pid) | |
650 | task = current; | |
651 | else | |
652 | task = find_task_by_vpid(pid); | |
653 | if (task) | |
654 | get_task_struct(task); | |
655 | rcu_read_unlock(); | |
656 | ||
657 | if (!task) | |
658 | return ERR_PTR(-ESRCH); | |
659 | ||
660 | ctx = &task->perf_counter_ctx; | |
661 | ctx->task = task; | |
662 | ||
663 | /* Reuse ptrace permission checks for now. */ | |
664 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) { | |
665 | put_context(ctx); | |
666 | return ERR_PTR(-EACCES); | |
667 | } | |
668 | ||
669 | return ctx; | |
670 | } | |
671 | ||
672 | /* | |
673 | * Called when the last reference to the file is gone. | |
674 | */ | |
675 | static int perf_release(struct inode *inode, struct file *file) | |
676 | { | |
677 | struct perf_counter *counter = file->private_data; | |
678 | struct perf_counter_context *ctx = counter->ctx; | |
679 | ||
680 | file->private_data = NULL; | |
681 | ||
682 | mutex_lock(&counter->mutex); | |
683 | ||
04289bb9 | 684 | perf_counter_remove_from_context(counter); |
0793a61d TG |
685 | put_context(ctx); |
686 | ||
687 | mutex_unlock(&counter->mutex); | |
688 | ||
689 | kfree(counter); | |
690 | ||
691 | return 0; | |
692 | } | |
693 | ||
694 | /* | |
695 | * Read the performance counter - simple non blocking version for now | |
696 | */ | |
697 | static ssize_t | |
698 | perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count) | |
699 | { | |
700 | u64 cntval; | |
701 | ||
702 | if (count != sizeof(cntval)) | |
703 | return -EINVAL; | |
704 | ||
705 | mutex_lock(&counter->mutex); | |
04289bb9 | 706 | cntval = perf_counter_read(counter); |
0793a61d TG |
707 | mutex_unlock(&counter->mutex); |
708 | ||
709 | return put_user(cntval, (u64 __user *) buf) ? -EFAULT : sizeof(cntval); | |
710 | } | |
711 | ||
712 | static ssize_t | |
713 | perf_copy_usrdata(struct perf_data *usrdata, char __user *buf, size_t count) | |
714 | { | |
715 | if (!usrdata->len) | |
716 | return 0; | |
717 | ||
718 | count = min(count, (size_t)usrdata->len); | |
719 | if (copy_to_user(buf, usrdata->data + usrdata->rd_idx, count)) | |
720 | return -EFAULT; | |
721 | ||
722 | /* Adjust the counters */ | |
723 | usrdata->len -= count; | |
724 | if (!usrdata->len) | |
725 | usrdata->rd_idx = 0; | |
726 | else | |
727 | usrdata->rd_idx += count; | |
728 | ||
729 | return count; | |
730 | } | |
731 | ||
732 | static ssize_t | |
733 | perf_read_irq_data(struct perf_counter *counter, | |
734 | char __user *buf, | |
735 | size_t count, | |
736 | int nonblocking) | |
737 | { | |
738 | struct perf_data *irqdata, *usrdata; | |
739 | DECLARE_WAITQUEUE(wait, current); | |
740 | ssize_t res; | |
741 | ||
742 | irqdata = counter->irqdata; | |
743 | usrdata = counter->usrdata; | |
744 | ||
745 | if (usrdata->len + irqdata->len >= count) | |
746 | goto read_pending; | |
747 | ||
748 | if (nonblocking) | |
749 | return -EAGAIN; | |
750 | ||
751 | spin_lock_irq(&counter->waitq.lock); | |
752 | __add_wait_queue(&counter->waitq, &wait); | |
753 | for (;;) { | |
754 | set_current_state(TASK_INTERRUPTIBLE); | |
755 | if (usrdata->len + irqdata->len >= count) | |
756 | break; | |
757 | ||
758 | if (signal_pending(current)) | |
759 | break; | |
760 | ||
761 | spin_unlock_irq(&counter->waitq.lock); | |
762 | schedule(); | |
763 | spin_lock_irq(&counter->waitq.lock); | |
764 | } | |
765 | __remove_wait_queue(&counter->waitq, &wait); | |
766 | __set_current_state(TASK_RUNNING); | |
767 | spin_unlock_irq(&counter->waitq.lock); | |
768 | ||
769 | if (usrdata->len + irqdata->len < count) | |
770 | return -ERESTARTSYS; | |
771 | read_pending: | |
772 | mutex_lock(&counter->mutex); | |
773 | ||
774 | /* Drain pending data first: */ | |
775 | res = perf_copy_usrdata(usrdata, buf, count); | |
776 | if (res < 0 || res == count) | |
777 | goto out; | |
778 | ||
779 | /* Switch irq buffer: */ | |
780 | usrdata = perf_switch_irq_data(counter); | |
781 | if (perf_copy_usrdata(usrdata, buf + res, count - res) < 0) { | |
782 | if (!res) | |
783 | res = -EFAULT; | |
784 | } else { | |
785 | res = count; | |
786 | } | |
787 | out: | |
788 | mutex_unlock(&counter->mutex); | |
789 | ||
790 | return res; | |
791 | } | |
792 | ||
793 | static ssize_t | |
794 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
795 | { | |
796 | struct perf_counter *counter = file->private_data; | |
797 | ||
9f66a381 | 798 | switch (counter->hw_event.record_type) { |
0793a61d TG |
799 | case PERF_RECORD_SIMPLE: |
800 | return perf_read_hw(counter, buf, count); | |
801 | ||
802 | case PERF_RECORD_IRQ: | |
803 | case PERF_RECORD_GROUP: | |
804 | return perf_read_irq_data(counter, buf, count, | |
805 | file->f_flags & O_NONBLOCK); | |
806 | } | |
807 | return -EINVAL; | |
808 | } | |
809 | ||
810 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
811 | { | |
812 | struct perf_counter *counter = file->private_data; | |
813 | unsigned int events = 0; | |
814 | unsigned long flags; | |
815 | ||
816 | poll_wait(file, &counter->waitq, wait); | |
817 | ||
818 | spin_lock_irqsave(&counter->waitq.lock, flags); | |
819 | if (counter->usrdata->len || counter->irqdata->len) | |
820 | events |= POLLIN; | |
821 | spin_unlock_irqrestore(&counter->waitq.lock, flags); | |
822 | ||
823 | return events; | |
824 | } | |
825 | ||
826 | static const struct file_operations perf_fops = { | |
827 | .release = perf_release, | |
828 | .read = perf_read, | |
829 | .poll = perf_poll, | |
830 | }; | |
831 | ||
5c92d124 IM |
832 | static void cpu_clock_perf_counter_enable(struct perf_counter *counter) |
833 | { | |
834 | } | |
835 | ||
836 | static void cpu_clock_perf_counter_disable(struct perf_counter *counter) | |
837 | { | |
838 | } | |
839 | ||
840 | static void cpu_clock_perf_counter_read(struct perf_counter *counter) | |
841 | { | |
842 | int cpu = raw_smp_processor_id(); | |
843 | ||
ee06094f | 844 | atomic64_set(&counter->count, cpu_clock(cpu)); |
5c92d124 IM |
845 | } |
846 | ||
847 | static const struct hw_perf_counter_ops perf_ops_cpu_clock = { | |
848 | .hw_perf_counter_enable = cpu_clock_perf_counter_enable, | |
849 | .hw_perf_counter_disable = cpu_clock_perf_counter_disable, | |
850 | .hw_perf_counter_read = cpu_clock_perf_counter_read, | |
851 | }; | |
852 | ||
8cb391e8 | 853 | static void task_clock_perf_counter_update(struct perf_counter *counter) |
bae43c99 | 854 | { |
8cb391e8 IM |
855 | u64 prev, now; |
856 | s64 delta; | |
857 | ||
858 | prev = atomic64_read(&counter->hw.prev_count); | |
859 | now = current->se.sum_exec_runtime; | |
860 | ||
861 | atomic64_set(&counter->hw.prev_count, now); | |
862 | ||
863 | delta = now - prev; | |
864 | if (WARN_ON_ONCE(delta < 0)) | |
865 | delta = 0; | |
866 | ||
867 | atomic64_add(delta, &counter->count); | |
bae43c99 IM |
868 | } |
869 | ||
8cb391e8 | 870 | static void task_clock_perf_counter_read(struct perf_counter *counter) |
bae43c99 | 871 | { |
8cb391e8 | 872 | task_clock_perf_counter_update(counter); |
bae43c99 IM |
873 | } |
874 | ||
8cb391e8 IM |
875 | static void task_clock_perf_counter_enable(struct perf_counter *counter) |
876 | { | |
877 | atomic64_set(&counter->hw.prev_count, current->se.sum_exec_runtime); | |
878 | } | |
879 | ||
880 | static void task_clock_perf_counter_disable(struct perf_counter *counter) | |
bae43c99 | 881 | { |
8cb391e8 | 882 | task_clock_perf_counter_update(counter); |
bae43c99 IM |
883 | } |
884 | ||
885 | static const struct hw_perf_counter_ops perf_ops_task_clock = { | |
886 | .hw_perf_counter_enable = task_clock_perf_counter_enable, | |
887 | .hw_perf_counter_disable = task_clock_perf_counter_disable, | |
888 | .hw_perf_counter_read = task_clock_perf_counter_read, | |
889 | }; | |
890 | ||
e06c61a8 IM |
891 | static u64 get_page_faults(void) |
892 | { | |
893 | struct task_struct *curr = current; | |
894 | ||
895 | return curr->maj_flt + curr->min_flt; | |
896 | } | |
897 | ||
898 | static void page_faults_perf_counter_update(struct perf_counter *counter) | |
899 | { | |
900 | u64 prev, now; | |
901 | s64 delta; | |
902 | ||
903 | prev = atomic64_read(&counter->hw.prev_count); | |
904 | now = get_page_faults(); | |
905 | ||
906 | atomic64_set(&counter->hw.prev_count, now); | |
907 | ||
908 | delta = now - prev; | |
909 | if (WARN_ON_ONCE(delta < 0)) | |
910 | delta = 0; | |
911 | ||
912 | atomic64_add(delta, &counter->count); | |
913 | } | |
914 | ||
915 | static void page_faults_perf_counter_read(struct perf_counter *counter) | |
916 | { | |
917 | page_faults_perf_counter_update(counter); | |
918 | } | |
919 | ||
920 | static void page_faults_perf_counter_enable(struct perf_counter *counter) | |
921 | { | |
922 | /* | |
923 | * page-faults is a per-task value already, | |
924 | * so we dont have to clear it on switch-in. | |
925 | */ | |
926 | } | |
927 | ||
928 | static void page_faults_perf_counter_disable(struct perf_counter *counter) | |
929 | { | |
930 | page_faults_perf_counter_update(counter); | |
931 | } | |
932 | ||
933 | static const struct hw_perf_counter_ops perf_ops_page_faults = { | |
934 | .hw_perf_counter_enable = page_faults_perf_counter_enable, | |
935 | .hw_perf_counter_disable = page_faults_perf_counter_disable, | |
936 | .hw_perf_counter_read = page_faults_perf_counter_read, | |
937 | }; | |
938 | ||
5d6a27d8 IM |
939 | static u64 get_context_switches(void) |
940 | { | |
941 | struct task_struct *curr = current; | |
942 | ||
943 | return curr->nvcsw + curr->nivcsw; | |
944 | } | |
945 | ||
946 | static void context_switches_perf_counter_update(struct perf_counter *counter) | |
947 | { | |
948 | u64 prev, now; | |
949 | s64 delta; | |
950 | ||
951 | prev = atomic64_read(&counter->hw.prev_count); | |
952 | now = get_context_switches(); | |
953 | ||
954 | atomic64_set(&counter->hw.prev_count, now); | |
955 | ||
956 | delta = now - prev; | |
957 | if (WARN_ON_ONCE(delta < 0)) | |
958 | delta = 0; | |
959 | ||
960 | atomic64_add(delta, &counter->count); | |
961 | } | |
962 | ||
963 | static void context_switches_perf_counter_read(struct perf_counter *counter) | |
964 | { | |
965 | context_switches_perf_counter_update(counter); | |
966 | } | |
967 | ||
968 | static void context_switches_perf_counter_enable(struct perf_counter *counter) | |
969 | { | |
970 | /* | |
971 | * ->nvcsw + curr->nivcsw is a per-task value already, | |
972 | * so we dont have to clear it on switch-in. | |
973 | */ | |
974 | } | |
975 | ||
976 | static void context_switches_perf_counter_disable(struct perf_counter *counter) | |
977 | { | |
978 | context_switches_perf_counter_update(counter); | |
979 | } | |
980 | ||
981 | static const struct hw_perf_counter_ops perf_ops_context_switches = { | |
982 | .hw_perf_counter_enable = context_switches_perf_counter_enable, | |
983 | .hw_perf_counter_disable = context_switches_perf_counter_disable, | |
984 | .hw_perf_counter_read = context_switches_perf_counter_read, | |
985 | }; | |
986 | ||
6c594c21 IM |
987 | static inline u64 get_cpu_migrations(void) |
988 | { | |
989 | return current->se.nr_migrations; | |
990 | } | |
991 | ||
992 | static void cpu_migrations_perf_counter_update(struct perf_counter *counter) | |
993 | { | |
994 | u64 prev, now; | |
995 | s64 delta; | |
996 | ||
997 | prev = atomic64_read(&counter->hw.prev_count); | |
998 | now = get_cpu_migrations(); | |
999 | ||
1000 | atomic64_set(&counter->hw.prev_count, now); | |
1001 | ||
1002 | delta = now - prev; | |
1003 | if (WARN_ON_ONCE(delta < 0)) | |
1004 | delta = 0; | |
1005 | ||
1006 | atomic64_add(delta, &counter->count); | |
1007 | } | |
1008 | ||
1009 | static void cpu_migrations_perf_counter_read(struct perf_counter *counter) | |
1010 | { | |
1011 | cpu_migrations_perf_counter_update(counter); | |
1012 | } | |
1013 | ||
1014 | static void cpu_migrations_perf_counter_enable(struct perf_counter *counter) | |
1015 | { | |
1016 | /* | |
1017 | * se.nr_migrations is a per-task value already, | |
1018 | * so we dont have to clear it on switch-in. | |
1019 | */ | |
1020 | } | |
1021 | ||
1022 | static void cpu_migrations_perf_counter_disable(struct perf_counter *counter) | |
1023 | { | |
1024 | cpu_migrations_perf_counter_update(counter); | |
1025 | } | |
1026 | ||
1027 | static const struct hw_perf_counter_ops perf_ops_cpu_migrations = { | |
1028 | .hw_perf_counter_enable = cpu_migrations_perf_counter_enable, | |
1029 | .hw_perf_counter_disable = cpu_migrations_perf_counter_disable, | |
1030 | .hw_perf_counter_read = cpu_migrations_perf_counter_read, | |
1031 | }; | |
1032 | ||
5c92d124 IM |
1033 | static const struct hw_perf_counter_ops * |
1034 | sw_perf_counter_init(struct perf_counter *counter) | |
1035 | { | |
1036 | const struct hw_perf_counter_ops *hw_ops = NULL; | |
1037 | ||
1038 | switch (counter->hw_event.type) { | |
1039 | case PERF_COUNT_CPU_CLOCK: | |
1040 | hw_ops = &perf_ops_cpu_clock; | |
1041 | break; | |
bae43c99 IM |
1042 | case PERF_COUNT_TASK_CLOCK: |
1043 | hw_ops = &perf_ops_task_clock; | |
1044 | break; | |
e06c61a8 IM |
1045 | case PERF_COUNT_PAGE_FAULTS: |
1046 | hw_ops = &perf_ops_page_faults; | |
1047 | break; | |
5d6a27d8 IM |
1048 | case PERF_COUNT_CONTEXT_SWITCHES: |
1049 | hw_ops = &perf_ops_context_switches; | |
1050 | break; | |
6c594c21 IM |
1051 | case PERF_COUNT_CPU_MIGRATIONS: |
1052 | hw_ops = &perf_ops_cpu_migrations; | |
1053 | break; | |
5c92d124 IM |
1054 | default: |
1055 | break; | |
1056 | } | |
1057 | return hw_ops; | |
1058 | } | |
1059 | ||
0793a61d TG |
1060 | /* |
1061 | * Allocate and initialize a counter structure | |
1062 | */ | |
1063 | static struct perf_counter * | |
04289bb9 IM |
1064 | perf_counter_alloc(struct perf_counter_hw_event *hw_event, |
1065 | int cpu, | |
9b51f66d IM |
1066 | struct perf_counter *group_leader, |
1067 | gfp_t gfpflags) | |
0793a61d | 1068 | { |
5c92d124 | 1069 | const struct hw_perf_counter_ops *hw_ops; |
621a01ea | 1070 | struct perf_counter *counter; |
0793a61d | 1071 | |
9b51f66d | 1072 | counter = kzalloc(sizeof(*counter), gfpflags); |
0793a61d TG |
1073 | if (!counter) |
1074 | return NULL; | |
1075 | ||
04289bb9 IM |
1076 | /* |
1077 | * Single counters are their own group leaders, with an | |
1078 | * empty sibling list: | |
1079 | */ | |
1080 | if (!group_leader) | |
1081 | group_leader = counter; | |
1082 | ||
0793a61d | 1083 | mutex_init(&counter->mutex); |
04289bb9 IM |
1084 | INIT_LIST_HEAD(&counter->list_entry); |
1085 | INIT_LIST_HEAD(&counter->sibling_list); | |
0793a61d TG |
1086 | init_waitqueue_head(&counter->waitq); |
1087 | ||
9f66a381 IM |
1088 | counter->irqdata = &counter->data[0]; |
1089 | counter->usrdata = &counter->data[1]; | |
1090 | counter->cpu = cpu; | |
1091 | counter->hw_event = *hw_event; | |
1092 | counter->wakeup_pending = 0; | |
04289bb9 | 1093 | counter->group_leader = group_leader; |
621a01ea IM |
1094 | counter->hw_ops = NULL; |
1095 | ||
5c92d124 IM |
1096 | hw_ops = NULL; |
1097 | if (!hw_event->raw && hw_event->type < 0) | |
1098 | hw_ops = sw_perf_counter_init(counter); | |
9b51f66d | 1099 | if (!hw_ops) |
5c92d124 | 1100 | hw_ops = hw_perf_counter_init(counter); |
5c92d124 | 1101 | |
621a01ea IM |
1102 | if (!hw_ops) { |
1103 | kfree(counter); | |
1104 | return NULL; | |
1105 | } | |
1106 | counter->hw_ops = hw_ops; | |
0793a61d TG |
1107 | |
1108 | return counter; | |
1109 | } | |
1110 | ||
1111 | /** | |
9f66a381 IM |
1112 | * sys_perf_task_open - open a performance counter, associate it to a task/cpu |
1113 | * | |
1114 | * @hw_event_uptr: event type attributes for monitoring/sampling | |
0793a61d | 1115 | * @pid: target pid |
9f66a381 IM |
1116 | * @cpu: target cpu |
1117 | * @group_fd: group leader counter fd | |
0793a61d | 1118 | */ |
1d1c7ddb IM |
1119 | asmlinkage int |
1120 | sys_perf_counter_open(struct perf_counter_hw_event *hw_event_uptr __user, | |
1121 | pid_t pid, int cpu, int group_fd) | |
0793a61d | 1122 | { |
04289bb9 | 1123 | struct perf_counter *counter, *group_leader; |
9f66a381 | 1124 | struct perf_counter_hw_event hw_event; |
04289bb9 | 1125 | struct perf_counter_context *ctx; |
9b51f66d | 1126 | struct file *counter_file = NULL; |
04289bb9 IM |
1127 | struct file *group_file = NULL; |
1128 | int fput_needed = 0; | |
9b51f66d | 1129 | int fput_needed2 = 0; |
0793a61d TG |
1130 | int ret; |
1131 | ||
9f66a381 | 1132 | if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0) |
eab656ae TG |
1133 | return -EFAULT; |
1134 | ||
04289bb9 | 1135 | /* |
ccff286d IM |
1136 | * Get the target context (task or percpu): |
1137 | */ | |
1138 | ctx = find_get_context(pid, cpu); | |
1139 | if (IS_ERR(ctx)) | |
1140 | return PTR_ERR(ctx); | |
1141 | ||
1142 | /* | |
1143 | * Look up the group leader (we will attach this counter to it): | |
04289bb9 IM |
1144 | */ |
1145 | group_leader = NULL; | |
1146 | if (group_fd != -1) { | |
1147 | ret = -EINVAL; | |
1148 | group_file = fget_light(group_fd, &fput_needed); | |
1149 | if (!group_file) | |
ccff286d | 1150 | goto err_put_context; |
04289bb9 | 1151 | if (group_file->f_op != &perf_fops) |
ccff286d | 1152 | goto err_put_context; |
04289bb9 IM |
1153 | |
1154 | group_leader = group_file->private_data; | |
1155 | /* | |
ccff286d IM |
1156 | * Do not allow a recursive hierarchy (this new sibling |
1157 | * becoming part of another group-sibling): | |
1158 | */ | |
1159 | if (group_leader->group_leader != group_leader) | |
1160 | goto err_put_context; | |
1161 | /* | |
1162 | * Do not allow to attach to a group in a different | |
1163 | * task or CPU context: | |
04289bb9 | 1164 | */ |
ccff286d IM |
1165 | if (group_leader->ctx != ctx) |
1166 | goto err_put_context; | |
04289bb9 IM |
1167 | } |
1168 | ||
5c92d124 | 1169 | ret = -EINVAL; |
9b51f66d | 1170 | counter = perf_counter_alloc(&hw_event, cpu, group_leader, GFP_KERNEL); |
0793a61d TG |
1171 | if (!counter) |
1172 | goto err_put_context; | |
1173 | ||
0793a61d TG |
1174 | ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0); |
1175 | if (ret < 0) | |
9b51f66d IM |
1176 | goto err_free_put_context; |
1177 | ||
1178 | counter_file = fget_light(ret, &fput_needed2); | |
1179 | if (!counter_file) | |
1180 | goto err_free_put_context; | |
1181 | ||
1182 | counter->filp = counter_file; | |
1183 | perf_install_in_context(ctx, counter, cpu); | |
1184 | ||
1185 | fput_light(counter_file, fput_needed2); | |
0793a61d | 1186 | |
04289bb9 IM |
1187 | out_fput: |
1188 | fput_light(group_file, fput_needed); | |
1189 | ||
0793a61d TG |
1190 | return ret; |
1191 | ||
9b51f66d | 1192 | err_free_put_context: |
0793a61d TG |
1193 | kfree(counter); |
1194 | ||
1195 | err_put_context: | |
1196 | put_context(ctx); | |
1197 | ||
04289bb9 | 1198 | goto out_fput; |
0793a61d TG |
1199 | } |
1200 | ||
9b51f66d IM |
1201 | /* |
1202 | * Initialize the perf_counter context in a task_struct: | |
1203 | */ | |
1204 | static void | |
1205 | __perf_counter_init_context(struct perf_counter_context *ctx, | |
1206 | struct task_struct *task) | |
1207 | { | |
1208 | memset(ctx, 0, sizeof(*ctx)); | |
1209 | spin_lock_init(&ctx->lock); | |
1210 | INIT_LIST_HEAD(&ctx->counter_list); | |
1211 | ctx->task = task; | |
1212 | } | |
1213 | ||
1214 | /* | |
1215 | * inherit a counter from parent task to child task: | |
1216 | */ | |
1217 | static int | |
1218 | inherit_counter(struct perf_counter *parent_counter, | |
1219 | struct task_struct *parent, | |
1220 | struct perf_counter_context *parent_ctx, | |
1221 | struct task_struct *child, | |
1222 | struct perf_counter_context *child_ctx) | |
1223 | { | |
1224 | struct perf_counter *child_counter; | |
1225 | ||
1226 | child_counter = perf_counter_alloc(&parent_counter->hw_event, | |
1227 | parent_counter->cpu, NULL, | |
1228 | GFP_ATOMIC); | |
1229 | if (!child_counter) | |
1230 | return -ENOMEM; | |
1231 | ||
1232 | /* | |
1233 | * Link it up in the child's context: | |
1234 | */ | |
1235 | child_counter->ctx = child_ctx; | |
1236 | child_counter->task = child; | |
1237 | list_add_counter(child_counter, child_ctx); | |
1238 | child_ctx->nr_counters++; | |
1239 | ||
1240 | child_counter->parent = parent_counter; | |
1241 | parent_counter->nr_inherited++; | |
1242 | /* | |
1243 | * inherit into child's child as well: | |
1244 | */ | |
1245 | child_counter->hw_event.inherit = 1; | |
1246 | ||
1247 | /* | |
1248 | * Get a reference to the parent filp - we will fput it | |
1249 | * when the child counter exits. This is safe to do because | |
1250 | * we are in the parent and we know that the filp still | |
1251 | * exists and has a nonzero count: | |
1252 | */ | |
1253 | atomic_long_inc(&parent_counter->filp->f_count); | |
1254 | ||
1255 | return 0; | |
1256 | } | |
1257 | ||
1258 | static void | |
1259 | __perf_counter_exit_task(struct task_struct *child, | |
1260 | struct perf_counter *child_counter, | |
1261 | struct perf_counter_context *child_ctx) | |
1262 | { | |
1263 | struct perf_counter *parent_counter; | |
1264 | u64 parent_val, child_val; | |
1265 | u64 perf_flags; | |
1266 | ||
1267 | /* | |
1268 | * Disable and unlink this counter. | |
1269 | * | |
1270 | * Be careful about zapping the list - IRQ/NMI context | |
1271 | * could still be processing it: | |
1272 | */ | |
1273 | local_irq_disable(); | |
1274 | perf_flags = hw_perf_save_disable(); | |
1275 | ||
1276 | if (child_counter->state == PERF_COUNTER_STATE_ACTIVE) | |
1277 | child_counter->hw_ops->hw_perf_counter_disable(child_counter); | |
1278 | list_del_init(&child_counter->list_entry); | |
1279 | ||
1280 | hw_perf_restore(perf_flags); | |
1281 | local_irq_enable(); | |
1282 | ||
1283 | parent_counter = child_counter->parent; | |
1284 | /* | |
1285 | * It can happen that parent exits first, and has counters | |
1286 | * that are still around due to the child reference. These | |
1287 | * counters need to be zapped - but otherwise linger. | |
1288 | */ | |
1289 | if (!parent_counter) | |
1290 | return; | |
1291 | ||
1292 | parent_val = atomic64_read(&parent_counter->count); | |
1293 | child_val = atomic64_read(&child_counter->count); | |
1294 | ||
1295 | /* | |
1296 | * Add back the child's count to the parent's count: | |
1297 | */ | |
1298 | atomic64_add(child_val, &parent_counter->count); | |
1299 | ||
1300 | fput(parent_counter->filp); | |
1301 | ||
1302 | kfree(child_counter); | |
1303 | } | |
1304 | ||
1305 | /* | |
1306 | * When a child task exist, feed back counter values to parent counters. | |
1307 | * | |
1308 | * Note: we are running in child context, but the PID is not hashed | |
1309 | * anymore so new counters will not be added. | |
1310 | */ | |
1311 | void perf_counter_exit_task(struct task_struct *child) | |
1312 | { | |
1313 | struct perf_counter *child_counter, *tmp; | |
1314 | struct perf_counter_context *child_ctx; | |
1315 | ||
1316 | child_ctx = &child->perf_counter_ctx; | |
1317 | ||
1318 | if (likely(!child_ctx->nr_counters)) | |
1319 | return; | |
1320 | ||
1321 | list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list, | |
1322 | list_entry) | |
1323 | __perf_counter_exit_task(child, child_counter, child_ctx); | |
1324 | } | |
1325 | ||
1326 | /* | |
1327 | * Initialize the perf_counter context in task_struct | |
1328 | */ | |
1329 | void perf_counter_init_task(struct task_struct *child) | |
1330 | { | |
1331 | struct perf_counter_context *child_ctx, *parent_ctx; | |
1332 | struct perf_counter *counter, *parent_counter; | |
1333 | struct task_struct *parent = current; | |
1334 | unsigned long flags; | |
1335 | ||
1336 | child_ctx = &child->perf_counter_ctx; | |
1337 | parent_ctx = &parent->perf_counter_ctx; | |
1338 | ||
1339 | __perf_counter_init_context(child_ctx, child); | |
1340 | ||
1341 | /* | |
1342 | * This is executed from the parent task context, so inherit | |
1343 | * counters that have been marked for cloning: | |
1344 | */ | |
1345 | ||
1346 | if (likely(!parent_ctx->nr_counters)) | |
1347 | return; | |
1348 | ||
1349 | /* | |
1350 | * Lock the parent list. No need to lock the child - not PID | |
1351 | * hashed yet and not running, so nobody can access it. | |
1352 | */ | |
1353 | spin_lock_irqsave(&parent_ctx->lock, flags); | |
1354 | ||
1355 | /* | |
1356 | * We dont have to disable NMIs - we are only looking at | |
1357 | * the list, not manipulating it: | |
1358 | */ | |
1359 | list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) { | |
1360 | if (!counter->hw_event.inherit || counter->group_leader != counter) | |
1361 | continue; | |
1362 | ||
1363 | /* | |
1364 | * Instead of creating recursive hierarchies of counters, | |
1365 | * we link inheritd counters back to the original parent, | |
1366 | * which has a filp for sure, which we use as the reference | |
1367 | * count: | |
1368 | */ | |
1369 | parent_counter = counter; | |
1370 | if (counter->parent) | |
1371 | parent_counter = counter->parent; | |
1372 | ||
1373 | if (inherit_counter(parent_counter, parent, | |
1374 | parent_ctx, child, child_ctx)) | |
1375 | break; | |
1376 | } | |
1377 | ||
1378 | spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
1379 | } | |
1380 | ||
04289bb9 | 1381 | static void __cpuinit perf_counter_init_cpu(int cpu) |
0793a61d | 1382 | { |
04289bb9 | 1383 | struct perf_cpu_context *cpuctx; |
0793a61d | 1384 | |
04289bb9 IM |
1385 | cpuctx = &per_cpu(perf_cpu_context, cpu); |
1386 | __perf_counter_init_context(&cpuctx->ctx, NULL); | |
0793a61d TG |
1387 | |
1388 | mutex_lock(&perf_resource_mutex); | |
04289bb9 | 1389 | cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu; |
0793a61d | 1390 | mutex_unlock(&perf_resource_mutex); |
04289bb9 | 1391 | |
0793a61d TG |
1392 | hw_perf_counter_setup(); |
1393 | } | |
1394 | ||
1395 | #ifdef CONFIG_HOTPLUG_CPU | |
04289bb9 | 1396 | static void __perf_counter_exit_cpu(void *info) |
0793a61d TG |
1397 | { |
1398 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
1399 | struct perf_counter_context *ctx = &cpuctx->ctx; | |
1400 | struct perf_counter *counter, *tmp; | |
1401 | ||
04289bb9 IM |
1402 | list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry) |
1403 | __perf_counter_remove_from_context(counter); | |
0793a61d TG |
1404 | |
1405 | } | |
04289bb9 | 1406 | static void perf_counter_exit_cpu(int cpu) |
0793a61d | 1407 | { |
04289bb9 | 1408 | smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1); |
0793a61d TG |
1409 | } |
1410 | #else | |
04289bb9 | 1411 | static inline void perf_counter_exit_cpu(int cpu) { } |
0793a61d TG |
1412 | #endif |
1413 | ||
1414 | static int __cpuinit | |
1415 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) | |
1416 | { | |
1417 | unsigned int cpu = (long)hcpu; | |
1418 | ||
1419 | switch (action) { | |
1420 | ||
1421 | case CPU_UP_PREPARE: | |
1422 | case CPU_UP_PREPARE_FROZEN: | |
04289bb9 | 1423 | perf_counter_init_cpu(cpu); |
0793a61d TG |
1424 | break; |
1425 | ||
1426 | case CPU_DOWN_PREPARE: | |
1427 | case CPU_DOWN_PREPARE_FROZEN: | |
04289bb9 | 1428 | perf_counter_exit_cpu(cpu); |
0793a61d TG |
1429 | break; |
1430 | ||
1431 | default: | |
1432 | break; | |
1433 | } | |
1434 | ||
1435 | return NOTIFY_OK; | |
1436 | } | |
1437 | ||
1438 | static struct notifier_block __cpuinitdata perf_cpu_nb = { | |
1439 | .notifier_call = perf_cpu_notify, | |
1440 | }; | |
1441 | ||
1442 | static int __init perf_counter_init(void) | |
1443 | { | |
1444 | perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE, | |
1445 | (void *)(long)smp_processor_id()); | |
1446 | register_cpu_notifier(&perf_cpu_nb); | |
1447 | ||
1448 | return 0; | |
1449 | } | |
1450 | early_initcall(perf_counter_init); | |
1451 | ||
1452 | static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf) | |
1453 | { | |
1454 | return sprintf(buf, "%d\n", perf_reserved_percpu); | |
1455 | } | |
1456 | ||
1457 | static ssize_t | |
1458 | perf_set_reserve_percpu(struct sysdev_class *class, | |
1459 | const char *buf, | |
1460 | size_t count) | |
1461 | { | |
1462 | struct perf_cpu_context *cpuctx; | |
1463 | unsigned long val; | |
1464 | int err, cpu, mpt; | |
1465 | ||
1466 | err = strict_strtoul(buf, 10, &val); | |
1467 | if (err) | |
1468 | return err; | |
1469 | if (val > perf_max_counters) | |
1470 | return -EINVAL; | |
1471 | ||
1472 | mutex_lock(&perf_resource_mutex); | |
1473 | perf_reserved_percpu = val; | |
1474 | for_each_online_cpu(cpu) { | |
1475 | cpuctx = &per_cpu(perf_cpu_context, cpu); | |
1476 | spin_lock_irq(&cpuctx->ctx.lock); | |
1477 | mpt = min(perf_max_counters - cpuctx->ctx.nr_counters, | |
1478 | perf_max_counters - perf_reserved_percpu); | |
1479 | cpuctx->max_pertask = mpt; | |
1480 | spin_unlock_irq(&cpuctx->ctx.lock); | |
1481 | } | |
1482 | mutex_unlock(&perf_resource_mutex); | |
1483 | ||
1484 | return count; | |
1485 | } | |
1486 | ||
1487 | static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf) | |
1488 | { | |
1489 | return sprintf(buf, "%d\n", perf_overcommit); | |
1490 | } | |
1491 | ||
1492 | static ssize_t | |
1493 | perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count) | |
1494 | { | |
1495 | unsigned long val; | |
1496 | int err; | |
1497 | ||
1498 | err = strict_strtoul(buf, 10, &val); | |
1499 | if (err) | |
1500 | return err; | |
1501 | if (val > 1) | |
1502 | return -EINVAL; | |
1503 | ||
1504 | mutex_lock(&perf_resource_mutex); | |
1505 | perf_overcommit = val; | |
1506 | mutex_unlock(&perf_resource_mutex); | |
1507 | ||
1508 | return count; | |
1509 | } | |
1510 | ||
1511 | static SYSDEV_CLASS_ATTR( | |
1512 | reserve_percpu, | |
1513 | 0644, | |
1514 | perf_show_reserve_percpu, | |
1515 | perf_set_reserve_percpu | |
1516 | ); | |
1517 | ||
1518 | static SYSDEV_CLASS_ATTR( | |
1519 | overcommit, | |
1520 | 0644, | |
1521 | perf_show_overcommit, | |
1522 | perf_set_overcommit | |
1523 | ); | |
1524 | ||
1525 | static struct attribute *perfclass_attrs[] = { | |
1526 | &attr_reserve_percpu.attr, | |
1527 | &attr_overcommit.attr, | |
1528 | NULL | |
1529 | }; | |
1530 | ||
1531 | static struct attribute_group perfclass_attr_group = { | |
1532 | .attrs = perfclass_attrs, | |
1533 | .name = "perf_counters", | |
1534 | }; | |
1535 | ||
1536 | static int __init perf_counter_sysfs_init(void) | |
1537 | { | |
1538 | return sysfs_create_group(&cpu_sysdev_class.kset.kobj, | |
1539 | &perfclass_attr_group); | |
1540 | } | |
1541 | device_initcall(perf_counter_sysfs_init); | |
1542 |