]>
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 | ||
28 | int perf_max_counters __read_mostly; | |
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 IM |
46 | u64 __weak hw_perf_save_disable(void) { return 0; } |
47 | void __weak hw_perf_restore(u64 ctrl) { } | |
5c92d124 | 48 | void __weak hw_perf_counter_setup(void) { } |
0793a61d TG |
49 | |
50 | #if BITS_PER_LONG == 64 | |
51 | ||
52 | /* | |
53 | * Read the cached counter in counter safe against cross CPU / NMI | |
54 | * modifications. 64 bit version - no complications. | |
55 | */ | |
04289bb9 | 56 | static inline u64 perf_counter_read_safe(struct perf_counter *counter) |
0793a61d TG |
57 | { |
58 | return (u64) atomic64_read(&counter->count); | |
59 | } | |
60 | ||
5c92d124 IM |
61 | void atomic64_counter_set(struct perf_counter *counter, u64 val) |
62 | { | |
63 | atomic64_set(&counter->count, val); | |
64 | } | |
65 | ||
66 | u64 atomic64_counter_read(struct perf_counter *counter) | |
67 | { | |
68 | return atomic64_read(&counter->count); | |
69 | } | |
70 | ||
0793a61d TG |
71 | #else |
72 | ||
73 | /* | |
74 | * Read the cached counter in counter safe against cross CPU / NMI | |
75 | * modifications. 32 bit version. | |
76 | */ | |
04289bb9 | 77 | static u64 perf_counter_read_safe(struct perf_counter *counter) |
0793a61d TG |
78 | { |
79 | u32 cntl, cnth; | |
80 | ||
81 | local_irq_disable(); | |
82 | do { | |
83 | cnth = atomic_read(&counter->count32[1]); | |
84 | cntl = atomic_read(&counter->count32[0]); | |
85 | } while (cnth != atomic_read(&counter->count32[1])); | |
86 | ||
87 | local_irq_enable(); | |
88 | ||
89 | return cntl | ((u64) cnth) << 32; | |
90 | } | |
91 | ||
5c92d124 IM |
92 | void atomic64_counter_set(struct perf_counter *counter, u64 val64) |
93 | { | |
94 | u32 *val32 = (void *)&val64; | |
95 | ||
96 | atomic_set(counter->count32 + 0, *(val32 + 0)); | |
97 | atomic_set(counter->count32 + 1, *(val32 + 1)); | |
98 | } | |
99 | ||
100 | u64 atomic64_counter_read(struct perf_counter *counter) | |
101 | { | |
102 | return atomic_read(counter->count32 + 0) | | |
103 | (u64) atomic_read(counter->count32 + 1) << 32; | |
104 | } | |
105 | ||
0793a61d TG |
106 | #endif |
107 | ||
04289bb9 IM |
108 | static void |
109 | list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx) | |
110 | { | |
111 | struct perf_counter *group_leader = counter->group_leader; | |
112 | ||
113 | /* | |
114 | * Depending on whether it is a standalone or sibling counter, | |
115 | * add it straight to the context's counter list, or to the group | |
116 | * leader's sibling list: | |
117 | */ | |
118 | if (counter->group_leader == counter) | |
119 | list_add_tail(&counter->list_entry, &ctx->counter_list); | |
120 | else | |
121 | list_add_tail(&counter->list_entry, &group_leader->sibling_list); | |
122 | } | |
123 | ||
124 | static void | |
125 | list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx) | |
126 | { | |
127 | struct perf_counter *sibling, *tmp; | |
128 | ||
129 | list_del_init(&counter->list_entry); | |
130 | ||
04289bb9 IM |
131 | /* |
132 | * If this was a group counter with sibling counters then | |
133 | * upgrade the siblings to singleton counters by adding them | |
134 | * to the context list directly: | |
135 | */ | |
136 | list_for_each_entry_safe(sibling, tmp, | |
137 | &counter->sibling_list, list_entry) { | |
138 | ||
139 | list_del_init(&sibling->list_entry); | |
140 | list_add_tail(&sibling->list_entry, &ctx->counter_list); | |
141 | WARN_ON_ONCE(!sibling->group_leader); | |
142 | WARN_ON_ONCE(sibling->group_leader == sibling); | |
143 | sibling->group_leader = sibling; | |
144 | } | |
145 | } | |
146 | ||
0793a61d TG |
147 | /* |
148 | * Cross CPU call to remove a performance counter | |
149 | * | |
150 | * We disable the counter on the hardware level first. After that we | |
151 | * remove it from the context list. | |
152 | */ | |
04289bb9 | 153 | static void __perf_counter_remove_from_context(void *info) |
0793a61d TG |
154 | { |
155 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
156 | struct perf_counter *counter = info; | |
157 | struct perf_counter_context *ctx = counter->ctx; | |
5c92d124 | 158 | u64 perf_flags; |
0793a61d TG |
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 | spin_lock(&ctx->lock); | |
169 | ||
6a930700 | 170 | if (counter->state == PERF_COUNTER_STATE_ACTIVE) { |
621a01ea | 171 | counter->hw_ops->hw_perf_counter_disable(counter); |
6a930700 | 172 | counter->state = PERF_COUNTER_STATE_INACTIVE; |
0793a61d TG |
173 | ctx->nr_active--; |
174 | cpuctx->active_oncpu--; | |
175 | counter->task = NULL; | |
176 | } | |
177 | ctx->nr_counters--; | |
178 | ||
179 | /* | |
180 | * Protect the list operation against NMI by disabling the | |
181 | * counters on a global level. NOP for non NMI based counters. | |
182 | */ | |
01b2838c | 183 | perf_flags = hw_perf_save_disable(); |
04289bb9 | 184 | list_del_counter(counter, ctx); |
01b2838c | 185 | hw_perf_restore(perf_flags); |
0793a61d TG |
186 | |
187 | if (!ctx->task) { | |
188 | /* | |
189 | * Allow more per task counters with respect to the | |
190 | * reservation: | |
191 | */ | |
192 | cpuctx->max_pertask = | |
193 | min(perf_max_counters - ctx->nr_counters, | |
194 | perf_max_counters - perf_reserved_percpu); | |
195 | } | |
196 | ||
197 | spin_unlock(&ctx->lock); | |
198 | } | |
199 | ||
200 | ||
201 | /* | |
202 | * Remove the counter from a task's (or a CPU's) list of counters. | |
203 | * | |
204 | * Must be called with counter->mutex held. | |
205 | * | |
206 | * CPU counters are removed with a smp call. For task counters we only | |
207 | * call when the task is on a CPU. | |
208 | */ | |
04289bb9 | 209 | static void perf_counter_remove_from_context(struct perf_counter *counter) |
0793a61d TG |
210 | { |
211 | struct perf_counter_context *ctx = counter->ctx; | |
212 | struct task_struct *task = ctx->task; | |
213 | ||
214 | if (!task) { | |
215 | /* | |
216 | * Per cpu counters are removed via an smp call and | |
217 | * the removal is always sucessful. | |
218 | */ | |
219 | smp_call_function_single(counter->cpu, | |
04289bb9 | 220 | __perf_counter_remove_from_context, |
0793a61d TG |
221 | counter, 1); |
222 | return; | |
223 | } | |
224 | ||
225 | retry: | |
04289bb9 | 226 | task_oncpu_function_call(task, __perf_counter_remove_from_context, |
0793a61d TG |
227 | counter); |
228 | ||
229 | spin_lock_irq(&ctx->lock); | |
230 | /* | |
231 | * If the context is active we need to retry the smp call. | |
232 | */ | |
04289bb9 | 233 | if (ctx->nr_active && !list_empty(&counter->list_entry)) { |
0793a61d TG |
234 | spin_unlock_irq(&ctx->lock); |
235 | goto retry; | |
236 | } | |
237 | ||
238 | /* | |
239 | * The lock prevents that this context is scheduled in so we | |
04289bb9 | 240 | * can remove the counter safely, if the call above did not |
0793a61d TG |
241 | * succeed. |
242 | */ | |
04289bb9 | 243 | if (!list_empty(&counter->list_entry)) { |
0793a61d | 244 | ctx->nr_counters--; |
04289bb9 | 245 | list_del_counter(counter, ctx); |
0793a61d TG |
246 | counter->task = NULL; |
247 | } | |
248 | spin_unlock_irq(&ctx->lock); | |
249 | } | |
250 | ||
251 | /* | |
252 | * Cross CPU call to install and enable a preformance counter | |
253 | */ | |
254 | static void __perf_install_in_context(void *info) | |
255 | { | |
256 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
257 | struct perf_counter *counter = info; | |
258 | struct perf_counter_context *ctx = counter->ctx; | |
259 | int cpu = smp_processor_id(); | |
5c92d124 | 260 | u64 perf_flags; |
0793a61d TG |
261 | |
262 | /* | |
263 | * If this is a task context, we need to check whether it is | |
264 | * the current task context of this cpu. If not it has been | |
265 | * scheduled out before the smp call arrived. | |
266 | */ | |
267 | if (ctx->task && cpuctx->task_ctx != ctx) | |
268 | return; | |
269 | ||
270 | spin_lock(&ctx->lock); | |
271 | ||
272 | /* | |
273 | * Protect the list operation against NMI by disabling the | |
274 | * counters on a global level. NOP for non NMI based counters. | |
275 | */ | |
01b2838c | 276 | perf_flags = hw_perf_save_disable(); |
04289bb9 | 277 | list_add_counter(counter, ctx); |
01b2838c | 278 | hw_perf_restore(perf_flags); |
0793a61d TG |
279 | |
280 | ctx->nr_counters++; | |
281 | ||
282 | if (cpuctx->active_oncpu < perf_max_counters) { | |
621a01ea | 283 | counter->hw_ops->hw_perf_counter_enable(counter); |
6a930700 | 284 | counter->state = PERF_COUNTER_STATE_ACTIVE; |
0793a61d TG |
285 | counter->oncpu = cpu; |
286 | ctx->nr_active++; | |
287 | cpuctx->active_oncpu++; | |
288 | } | |
289 | ||
290 | if (!ctx->task && cpuctx->max_pertask) | |
291 | cpuctx->max_pertask--; | |
292 | ||
293 | spin_unlock(&ctx->lock); | |
294 | } | |
295 | ||
296 | /* | |
297 | * Attach a performance counter to a context | |
298 | * | |
299 | * First we add the counter to the list with the hardware enable bit | |
300 | * in counter->hw_config cleared. | |
301 | * | |
302 | * If the counter is attached to a task which is on a CPU we use a smp | |
303 | * call to enable it in the task context. The task might have been | |
304 | * scheduled away, but we check this in the smp call again. | |
305 | */ | |
306 | static void | |
307 | perf_install_in_context(struct perf_counter_context *ctx, | |
308 | struct perf_counter *counter, | |
309 | int cpu) | |
310 | { | |
311 | struct task_struct *task = ctx->task; | |
312 | ||
313 | counter->ctx = ctx; | |
314 | if (!task) { | |
315 | /* | |
316 | * Per cpu counters are installed via an smp call and | |
317 | * the install is always sucessful. | |
318 | */ | |
319 | smp_call_function_single(cpu, __perf_install_in_context, | |
320 | counter, 1); | |
321 | return; | |
322 | } | |
323 | ||
324 | counter->task = task; | |
325 | retry: | |
326 | task_oncpu_function_call(task, __perf_install_in_context, | |
327 | counter); | |
328 | ||
329 | spin_lock_irq(&ctx->lock); | |
330 | /* | |
0793a61d TG |
331 | * we need to retry the smp call. |
332 | */ | |
04289bb9 | 333 | if (ctx->nr_active && list_empty(&counter->list_entry)) { |
0793a61d TG |
334 | spin_unlock_irq(&ctx->lock); |
335 | goto retry; | |
336 | } | |
337 | ||
338 | /* | |
339 | * The lock prevents that this context is scheduled in so we | |
340 | * can add the counter safely, if it the call above did not | |
341 | * succeed. | |
342 | */ | |
04289bb9 IM |
343 | if (list_empty(&counter->list_entry)) { |
344 | list_add_counter(counter, ctx); | |
0793a61d TG |
345 | ctx->nr_counters++; |
346 | } | |
347 | spin_unlock_irq(&ctx->lock); | |
348 | } | |
349 | ||
04289bb9 IM |
350 | static void |
351 | counter_sched_out(struct perf_counter *counter, | |
352 | struct perf_cpu_context *cpuctx, | |
353 | struct perf_counter_context *ctx) | |
354 | { | |
6a930700 | 355 | if (counter->state != PERF_COUNTER_STATE_ACTIVE) |
04289bb9 IM |
356 | return; |
357 | ||
621a01ea | 358 | counter->hw_ops->hw_perf_counter_disable(counter); |
6a930700 IM |
359 | counter->state = PERF_COUNTER_STATE_INACTIVE; |
360 | counter->oncpu = -1; | |
04289bb9 IM |
361 | |
362 | cpuctx->active_oncpu--; | |
363 | ctx->nr_active--; | |
364 | } | |
365 | ||
366 | static void | |
367 | group_sched_out(struct perf_counter *group_counter, | |
368 | struct perf_cpu_context *cpuctx, | |
369 | struct perf_counter_context *ctx) | |
370 | { | |
371 | struct perf_counter *counter; | |
372 | ||
373 | counter_sched_out(group_counter, cpuctx, ctx); | |
374 | ||
375 | /* | |
376 | * Schedule out siblings (if any): | |
377 | */ | |
378 | list_for_each_entry(counter, &group_counter->sibling_list, list_entry) | |
379 | counter_sched_out(counter, cpuctx, ctx); | |
380 | } | |
381 | ||
0793a61d TG |
382 | /* |
383 | * Called from scheduler to remove the counters of the current task, | |
384 | * with interrupts disabled. | |
385 | * | |
386 | * We stop each counter and update the counter value in counter->count. | |
387 | * | |
388 | * This does not protect us against NMI, but hw_perf_counter_disable() | |
389 | * sets the disabled bit in the control field of counter _before_ | |
390 | * accessing the counter control register. If a NMI hits, then it will | |
391 | * not restart the counter. | |
392 | */ | |
393 | void perf_counter_task_sched_out(struct task_struct *task, int cpu) | |
394 | { | |
395 | struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); | |
396 | struct perf_counter_context *ctx = &task->perf_counter_ctx; | |
397 | struct perf_counter *counter; | |
398 | ||
399 | if (likely(!cpuctx->task_ctx)) | |
400 | return; | |
401 | ||
402 | spin_lock(&ctx->lock); | |
04289bb9 IM |
403 | if (ctx->nr_active) { |
404 | list_for_each_entry(counter, &ctx->counter_list, list_entry) | |
405 | group_sched_out(counter, cpuctx, ctx); | |
0793a61d TG |
406 | } |
407 | spin_unlock(&ctx->lock); | |
408 | cpuctx->task_ctx = NULL; | |
409 | } | |
410 | ||
04289bb9 IM |
411 | static void |
412 | counter_sched_in(struct perf_counter *counter, | |
413 | struct perf_cpu_context *cpuctx, | |
414 | struct perf_counter_context *ctx, | |
415 | int cpu) | |
416 | { | |
6a930700 | 417 | if (counter->state == PERF_COUNTER_STATE_OFF) |
1d1c7ddb IM |
418 | return; |
419 | ||
621a01ea | 420 | counter->hw_ops->hw_perf_counter_enable(counter); |
6a930700 | 421 | counter->state = PERF_COUNTER_STATE_ACTIVE; |
04289bb9 IM |
422 | counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */ |
423 | ||
424 | cpuctx->active_oncpu++; | |
425 | ctx->nr_active++; | |
426 | } | |
427 | ||
428 | static void | |
429 | group_sched_in(struct perf_counter *group_counter, | |
430 | struct perf_cpu_context *cpuctx, | |
431 | struct perf_counter_context *ctx, | |
432 | int cpu) | |
433 | { | |
434 | struct perf_counter *counter; | |
435 | ||
436 | counter_sched_in(group_counter, cpuctx, ctx, cpu); | |
437 | ||
438 | /* | |
439 | * Schedule in siblings as one group (if any): | |
440 | */ | |
441 | list_for_each_entry(counter, &group_counter->sibling_list, list_entry) | |
442 | counter_sched_in(counter, cpuctx, ctx, cpu); | |
443 | } | |
444 | ||
0793a61d TG |
445 | /* |
446 | * Called from scheduler to add the counters of the current task | |
447 | * with interrupts disabled. | |
448 | * | |
449 | * We restore the counter value and then enable it. | |
450 | * | |
451 | * This does not protect us against NMI, but hw_perf_counter_enable() | |
452 | * sets the enabled bit in the control field of counter _before_ | |
453 | * accessing the counter control register. If a NMI hits, then it will | |
454 | * keep the counter running. | |
455 | */ | |
456 | void perf_counter_task_sched_in(struct task_struct *task, int cpu) | |
457 | { | |
458 | struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); | |
459 | struct perf_counter_context *ctx = &task->perf_counter_ctx; | |
460 | struct perf_counter *counter; | |
461 | ||
462 | if (likely(!ctx->nr_counters)) | |
463 | return; | |
464 | ||
465 | spin_lock(&ctx->lock); | |
04289bb9 | 466 | list_for_each_entry(counter, &ctx->counter_list, list_entry) { |
0793a61d TG |
467 | if (ctx->nr_active == cpuctx->max_pertask) |
468 | break; | |
04289bb9 IM |
469 | |
470 | /* | |
471 | * Listen to the 'cpu' scheduling filter constraint | |
472 | * of counters: | |
473 | */ | |
0793a61d TG |
474 | if (counter->cpu != -1 && counter->cpu != cpu) |
475 | continue; | |
476 | ||
04289bb9 | 477 | group_sched_in(counter, cpuctx, ctx, cpu); |
0793a61d TG |
478 | } |
479 | spin_unlock(&ctx->lock); | |
04289bb9 | 480 | |
0793a61d TG |
481 | cpuctx->task_ctx = ctx; |
482 | } | |
483 | ||
1d1c7ddb IM |
484 | int perf_counter_task_disable(void) |
485 | { | |
486 | struct task_struct *curr = current; | |
487 | struct perf_counter_context *ctx = &curr->perf_counter_ctx; | |
488 | struct perf_counter *counter; | |
489 | u64 perf_flags; | |
490 | int cpu; | |
491 | ||
492 | if (likely(!ctx->nr_counters)) | |
493 | return 0; | |
494 | ||
495 | local_irq_disable(); | |
496 | cpu = smp_processor_id(); | |
497 | ||
498 | perf_counter_task_sched_out(curr, cpu); | |
499 | ||
500 | spin_lock(&ctx->lock); | |
501 | ||
502 | /* | |
503 | * Disable all the counters: | |
504 | */ | |
505 | perf_flags = hw_perf_save_disable(); | |
506 | ||
507 | list_for_each_entry(counter, &ctx->counter_list, list_entry) { | |
6a930700 IM |
508 | WARN_ON_ONCE(counter->state == PERF_COUNTER_STATE_ACTIVE); |
509 | counter->state = PERF_COUNTER_STATE_OFF; | |
1d1c7ddb IM |
510 | } |
511 | hw_perf_restore(perf_flags); | |
512 | ||
513 | spin_unlock(&ctx->lock); | |
514 | ||
515 | local_irq_enable(); | |
516 | ||
517 | return 0; | |
518 | } | |
519 | ||
520 | int perf_counter_task_enable(void) | |
521 | { | |
522 | struct task_struct *curr = current; | |
523 | struct perf_counter_context *ctx = &curr->perf_counter_ctx; | |
524 | struct perf_counter *counter; | |
525 | u64 perf_flags; | |
526 | int cpu; | |
527 | ||
528 | if (likely(!ctx->nr_counters)) | |
529 | return 0; | |
530 | ||
531 | local_irq_disable(); | |
532 | cpu = smp_processor_id(); | |
533 | ||
534 | spin_lock(&ctx->lock); | |
535 | ||
536 | /* | |
537 | * Disable all the counters: | |
538 | */ | |
539 | perf_flags = hw_perf_save_disable(); | |
540 | ||
541 | list_for_each_entry(counter, &ctx->counter_list, list_entry) { | |
6a930700 | 542 | if (counter->state != PERF_COUNTER_STATE_OFF) |
1d1c7ddb | 543 | continue; |
6a930700 | 544 | counter->state = PERF_COUNTER_STATE_INACTIVE; |
1d1c7ddb IM |
545 | } |
546 | hw_perf_restore(perf_flags); | |
547 | ||
548 | spin_unlock(&ctx->lock); | |
549 | ||
550 | perf_counter_task_sched_in(curr, cpu); | |
551 | ||
552 | local_irq_enable(); | |
553 | ||
554 | return 0; | |
555 | } | |
556 | ||
0793a61d TG |
557 | void perf_counter_task_tick(struct task_struct *curr, int cpu) |
558 | { | |
559 | struct perf_counter_context *ctx = &curr->perf_counter_ctx; | |
560 | struct perf_counter *counter; | |
5c92d124 | 561 | u64 perf_flags; |
0793a61d TG |
562 | |
563 | if (likely(!ctx->nr_counters)) | |
564 | return; | |
565 | ||
566 | perf_counter_task_sched_out(curr, cpu); | |
567 | ||
568 | spin_lock(&ctx->lock); | |
569 | ||
570 | /* | |
04289bb9 | 571 | * Rotate the first entry last (works just fine for group counters too): |
0793a61d | 572 | */ |
01b2838c | 573 | perf_flags = hw_perf_save_disable(); |
04289bb9 IM |
574 | list_for_each_entry(counter, &ctx->counter_list, list_entry) { |
575 | list_del(&counter->list_entry); | |
576 | list_add_tail(&counter->list_entry, &ctx->counter_list); | |
0793a61d TG |
577 | break; |
578 | } | |
01b2838c | 579 | hw_perf_restore(perf_flags); |
0793a61d TG |
580 | |
581 | spin_unlock(&ctx->lock); | |
582 | ||
583 | perf_counter_task_sched_in(curr, cpu); | |
584 | } | |
585 | ||
04289bb9 IM |
586 | /* |
587 | * Initialize the perf_counter context in a task_struct: | |
588 | */ | |
589 | static void | |
590 | __perf_counter_init_context(struct perf_counter_context *ctx, | |
591 | struct task_struct *task) | |
592 | { | |
593 | spin_lock_init(&ctx->lock); | |
594 | INIT_LIST_HEAD(&ctx->counter_list); | |
595 | ctx->nr_counters = 0; | |
596 | ctx->task = task; | |
597 | } | |
0793a61d TG |
598 | /* |
599 | * Initialize the perf_counter context in task_struct | |
600 | */ | |
601 | void perf_counter_init_task(struct task_struct *task) | |
602 | { | |
04289bb9 | 603 | __perf_counter_init_context(&task->perf_counter_ctx, task); |
0793a61d TG |
604 | } |
605 | ||
606 | /* | |
607 | * Cross CPU call to read the hardware counter | |
608 | */ | |
609 | static void __hw_perf_counter_read(void *info) | |
610 | { | |
621a01ea IM |
611 | struct perf_counter *counter = info; |
612 | ||
613 | counter->hw_ops->hw_perf_counter_read(counter); | |
0793a61d TG |
614 | } |
615 | ||
04289bb9 | 616 | static u64 perf_counter_read(struct perf_counter *counter) |
0793a61d TG |
617 | { |
618 | /* | |
619 | * If counter is enabled and currently active on a CPU, update the | |
620 | * value in the counter structure: | |
621 | */ | |
6a930700 | 622 | if (counter->state == PERF_COUNTER_STATE_ACTIVE) { |
0793a61d TG |
623 | smp_call_function_single(counter->oncpu, |
624 | __hw_perf_counter_read, counter, 1); | |
625 | } | |
626 | ||
04289bb9 | 627 | return perf_counter_read_safe(counter); |
0793a61d TG |
628 | } |
629 | ||
630 | /* | |
631 | * Cross CPU call to switch performance data pointers | |
632 | */ | |
633 | static void __perf_switch_irq_data(void *info) | |
634 | { | |
635 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
636 | struct perf_counter *counter = info; | |
637 | struct perf_counter_context *ctx = counter->ctx; | |
638 | struct perf_data *oldirqdata = counter->irqdata; | |
639 | ||
640 | /* | |
641 | * If this is a task context, we need to check whether it is | |
642 | * the current task context of this cpu. If not it has been | |
643 | * scheduled out before the smp call arrived. | |
644 | */ | |
645 | if (ctx->task) { | |
646 | if (cpuctx->task_ctx != ctx) | |
647 | return; | |
648 | spin_lock(&ctx->lock); | |
649 | } | |
650 | ||
651 | /* Change the pointer NMI safe */ | |
652 | atomic_long_set((atomic_long_t *)&counter->irqdata, | |
653 | (unsigned long) counter->usrdata); | |
654 | counter->usrdata = oldirqdata; | |
655 | ||
656 | if (ctx->task) | |
657 | spin_unlock(&ctx->lock); | |
658 | } | |
659 | ||
660 | static struct perf_data *perf_switch_irq_data(struct perf_counter *counter) | |
661 | { | |
662 | struct perf_counter_context *ctx = counter->ctx; | |
663 | struct perf_data *oldirqdata = counter->irqdata; | |
664 | struct task_struct *task = ctx->task; | |
665 | ||
666 | if (!task) { | |
667 | smp_call_function_single(counter->cpu, | |
668 | __perf_switch_irq_data, | |
669 | counter, 1); | |
670 | return counter->usrdata; | |
671 | } | |
672 | ||
673 | retry: | |
674 | spin_lock_irq(&ctx->lock); | |
6a930700 | 675 | if (counter->state != PERF_COUNTER_STATE_ACTIVE) { |
0793a61d TG |
676 | counter->irqdata = counter->usrdata; |
677 | counter->usrdata = oldirqdata; | |
678 | spin_unlock_irq(&ctx->lock); | |
679 | return oldirqdata; | |
680 | } | |
681 | spin_unlock_irq(&ctx->lock); | |
682 | task_oncpu_function_call(task, __perf_switch_irq_data, counter); | |
683 | /* Might have failed, because task was scheduled out */ | |
684 | if (counter->irqdata == oldirqdata) | |
685 | goto retry; | |
686 | ||
687 | return counter->usrdata; | |
688 | } | |
689 | ||
690 | static void put_context(struct perf_counter_context *ctx) | |
691 | { | |
692 | if (ctx->task) | |
693 | put_task_struct(ctx->task); | |
694 | } | |
695 | ||
696 | static struct perf_counter_context *find_get_context(pid_t pid, int cpu) | |
697 | { | |
698 | struct perf_cpu_context *cpuctx; | |
699 | struct perf_counter_context *ctx; | |
700 | struct task_struct *task; | |
701 | ||
702 | /* | |
703 | * If cpu is not a wildcard then this is a percpu counter: | |
704 | */ | |
705 | if (cpu != -1) { | |
706 | /* Must be root to operate on a CPU counter: */ | |
707 | if (!capable(CAP_SYS_ADMIN)) | |
708 | return ERR_PTR(-EACCES); | |
709 | ||
710 | if (cpu < 0 || cpu > num_possible_cpus()) | |
711 | return ERR_PTR(-EINVAL); | |
712 | ||
713 | /* | |
714 | * We could be clever and allow to attach a counter to an | |
715 | * offline CPU and activate it when the CPU comes up, but | |
716 | * that's for later. | |
717 | */ | |
718 | if (!cpu_isset(cpu, cpu_online_map)) | |
719 | return ERR_PTR(-ENODEV); | |
720 | ||
721 | cpuctx = &per_cpu(perf_cpu_context, cpu); | |
722 | ctx = &cpuctx->ctx; | |
723 | ||
724 | WARN_ON_ONCE(ctx->task); | |
725 | return ctx; | |
726 | } | |
727 | ||
728 | rcu_read_lock(); | |
729 | if (!pid) | |
730 | task = current; | |
731 | else | |
732 | task = find_task_by_vpid(pid); | |
733 | if (task) | |
734 | get_task_struct(task); | |
735 | rcu_read_unlock(); | |
736 | ||
737 | if (!task) | |
738 | return ERR_PTR(-ESRCH); | |
739 | ||
740 | ctx = &task->perf_counter_ctx; | |
741 | ctx->task = task; | |
742 | ||
743 | /* Reuse ptrace permission checks for now. */ | |
744 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) { | |
745 | put_context(ctx); | |
746 | return ERR_PTR(-EACCES); | |
747 | } | |
748 | ||
749 | return ctx; | |
750 | } | |
751 | ||
752 | /* | |
753 | * Called when the last reference to the file is gone. | |
754 | */ | |
755 | static int perf_release(struct inode *inode, struct file *file) | |
756 | { | |
757 | struct perf_counter *counter = file->private_data; | |
758 | struct perf_counter_context *ctx = counter->ctx; | |
759 | ||
760 | file->private_data = NULL; | |
761 | ||
762 | mutex_lock(&counter->mutex); | |
763 | ||
04289bb9 | 764 | perf_counter_remove_from_context(counter); |
0793a61d TG |
765 | put_context(ctx); |
766 | ||
767 | mutex_unlock(&counter->mutex); | |
768 | ||
769 | kfree(counter); | |
770 | ||
771 | return 0; | |
772 | } | |
773 | ||
774 | /* | |
775 | * Read the performance counter - simple non blocking version for now | |
776 | */ | |
777 | static ssize_t | |
778 | perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count) | |
779 | { | |
780 | u64 cntval; | |
781 | ||
782 | if (count != sizeof(cntval)) | |
783 | return -EINVAL; | |
784 | ||
785 | mutex_lock(&counter->mutex); | |
04289bb9 | 786 | cntval = perf_counter_read(counter); |
0793a61d TG |
787 | mutex_unlock(&counter->mutex); |
788 | ||
789 | return put_user(cntval, (u64 __user *) buf) ? -EFAULT : sizeof(cntval); | |
790 | } | |
791 | ||
792 | static ssize_t | |
793 | perf_copy_usrdata(struct perf_data *usrdata, char __user *buf, size_t count) | |
794 | { | |
795 | if (!usrdata->len) | |
796 | return 0; | |
797 | ||
798 | count = min(count, (size_t)usrdata->len); | |
799 | if (copy_to_user(buf, usrdata->data + usrdata->rd_idx, count)) | |
800 | return -EFAULT; | |
801 | ||
802 | /* Adjust the counters */ | |
803 | usrdata->len -= count; | |
804 | if (!usrdata->len) | |
805 | usrdata->rd_idx = 0; | |
806 | else | |
807 | usrdata->rd_idx += count; | |
808 | ||
809 | return count; | |
810 | } | |
811 | ||
812 | static ssize_t | |
813 | perf_read_irq_data(struct perf_counter *counter, | |
814 | char __user *buf, | |
815 | size_t count, | |
816 | int nonblocking) | |
817 | { | |
818 | struct perf_data *irqdata, *usrdata; | |
819 | DECLARE_WAITQUEUE(wait, current); | |
820 | ssize_t res; | |
821 | ||
822 | irqdata = counter->irqdata; | |
823 | usrdata = counter->usrdata; | |
824 | ||
825 | if (usrdata->len + irqdata->len >= count) | |
826 | goto read_pending; | |
827 | ||
828 | if (nonblocking) | |
829 | return -EAGAIN; | |
830 | ||
831 | spin_lock_irq(&counter->waitq.lock); | |
832 | __add_wait_queue(&counter->waitq, &wait); | |
833 | for (;;) { | |
834 | set_current_state(TASK_INTERRUPTIBLE); | |
835 | if (usrdata->len + irqdata->len >= count) | |
836 | break; | |
837 | ||
838 | if (signal_pending(current)) | |
839 | break; | |
840 | ||
841 | spin_unlock_irq(&counter->waitq.lock); | |
842 | schedule(); | |
843 | spin_lock_irq(&counter->waitq.lock); | |
844 | } | |
845 | __remove_wait_queue(&counter->waitq, &wait); | |
846 | __set_current_state(TASK_RUNNING); | |
847 | spin_unlock_irq(&counter->waitq.lock); | |
848 | ||
849 | if (usrdata->len + irqdata->len < count) | |
850 | return -ERESTARTSYS; | |
851 | read_pending: | |
852 | mutex_lock(&counter->mutex); | |
853 | ||
854 | /* Drain pending data first: */ | |
855 | res = perf_copy_usrdata(usrdata, buf, count); | |
856 | if (res < 0 || res == count) | |
857 | goto out; | |
858 | ||
859 | /* Switch irq buffer: */ | |
860 | usrdata = perf_switch_irq_data(counter); | |
861 | if (perf_copy_usrdata(usrdata, buf + res, count - res) < 0) { | |
862 | if (!res) | |
863 | res = -EFAULT; | |
864 | } else { | |
865 | res = count; | |
866 | } | |
867 | out: | |
868 | mutex_unlock(&counter->mutex); | |
869 | ||
870 | return res; | |
871 | } | |
872 | ||
873 | static ssize_t | |
874 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
875 | { | |
876 | struct perf_counter *counter = file->private_data; | |
877 | ||
9f66a381 | 878 | switch (counter->hw_event.record_type) { |
0793a61d TG |
879 | case PERF_RECORD_SIMPLE: |
880 | return perf_read_hw(counter, buf, count); | |
881 | ||
882 | case PERF_RECORD_IRQ: | |
883 | case PERF_RECORD_GROUP: | |
884 | return perf_read_irq_data(counter, buf, count, | |
885 | file->f_flags & O_NONBLOCK); | |
886 | } | |
887 | return -EINVAL; | |
888 | } | |
889 | ||
890 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
891 | { | |
892 | struct perf_counter *counter = file->private_data; | |
893 | unsigned int events = 0; | |
894 | unsigned long flags; | |
895 | ||
896 | poll_wait(file, &counter->waitq, wait); | |
897 | ||
898 | spin_lock_irqsave(&counter->waitq.lock, flags); | |
899 | if (counter->usrdata->len || counter->irqdata->len) | |
900 | events |= POLLIN; | |
901 | spin_unlock_irqrestore(&counter->waitq.lock, flags); | |
902 | ||
903 | return events; | |
904 | } | |
905 | ||
906 | static const struct file_operations perf_fops = { | |
907 | .release = perf_release, | |
908 | .read = perf_read, | |
909 | .poll = perf_poll, | |
910 | }; | |
911 | ||
5c92d124 IM |
912 | static void cpu_clock_perf_counter_enable(struct perf_counter *counter) |
913 | { | |
914 | } | |
915 | ||
916 | static void cpu_clock_perf_counter_disable(struct perf_counter *counter) | |
917 | { | |
918 | } | |
919 | ||
920 | static void cpu_clock_perf_counter_read(struct perf_counter *counter) | |
921 | { | |
922 | int cpu = raw_smp_processor_id(); | |
923 | ||
924 | atomic64_counter_set(counter, cpu_clock(cpu)); | |
925 | } | |
926 | ||
927 | static const struct hw_perf_counter_ops perf_ops_cpu_clock = { | |
928 | .hw_perf_counter_enable = cpu_clock_perf_counter_enable, | |
929 | .hw_perf_counter_disable = cpu_clock_perf_counter_disable, | |
930 | .hw_perf_counter_read = cpu_clock_perf_counter_read, | |
931 | }; | |
932 | ||
bae43c99 IM |
933 | static void task_clock_perf_counter_enable(struct perf_counter *counter) |
934 | { | |
935 | } | |
936 | ||
937 | static void task_clock_perf_counter_disable(struct perf_counter *counter) | |
938 | { | |
939 | } | |
940 | ||
941 | static void task_clock_perf_counter_read(struct perf_counter *counter) | |
942 | { | |
943 | atomic64_counter_set(counter, current->se.sum_exec_runtime); | |
944 | } | |
945 | ||
946 | static const struct hw_perf_counter_ops perf_ops_task_clock = { | |
947 | .hw_perf_counter_enable = task_clock_perf_counter_enable, | |
948 | .hw_perf_counter_disable = task_clock_perf_counter_disable, | |
949 | .hw_perf_counter_read = task_clock_perf_counter_read, | |
950 | }; | |
951 | ||
5c92d124 IM |
952 | static const struct hw_perf_counter_ops * |
953 | sw_perf_counter_init(struct perf_counter *counter) | |
954 | { | |
955 | const struct hw_perf_counter_ops *hw_ops = NULL; | |
956 | ||
957 | switch (counter->hw_event.type) { | |
958 | case PERF_COUNT_CPU_CLOCK: | |
959 | hw_ops = &perf_ops_cpu_clock; | |
960 | break; | |
bae43c99 IM |
961 | case PERF_COUNT_TASK_CLOCK: |
962 | hw_ops = &perf_ops_task_clock; | |
963 | break; | |
5c92d124 IM |
964 | default: |
965 | break; | |
966 | } | |
967 | return hw_ops; | |
968 | } | |
969 | ||
0793a61d TG |
970 | /* |
971 | * Allocate and initialize a counter structure | |
972 | */ | |
973 | static struct perf_counter * | |
04289bb9 IM |
974 | perf_counter_alloc(struct perf_counter_hw_event *hw_event, |
975 | int cpu, | |
976 | struct perf_counter *group_leader) | |
0793a61d | 977 | { |
5c92d124 | 978 | const struct hw_perf_counter_ops *hw_ops; |
621a01ea | 979 | struct perf_counter *counter; |
0793a61d | 980 | |
621a01ea | 981 | counter = kzalloc(sizeof(*counter), GFP_KERNEL); |
0793a61d TG |
982 | if (!counter) |
983 | return NULL; | |
984 | ||
04289bb9 IM |
985 | /* |
986 | * Single counters are their own group leaders, with an | |
987 | * empty sibling list: | |
988 | */ | |
989 | if (!group_leader) | |
990 | group_leader = counter; | |
991 | ||
0793a61d | 992 | mutex_init(&counter->mutex); |
04289bb9 IM |
993 | INIT_LIST_HEAD(&counter->list_entry); |
994 | INIT_LIST_HEAD(&counter->sibling_list); | |
0793a61d TG |
995 | init_waitqueue_head(&counter->waitq); |
996 | ||
9f66a381 IM |
997 | counter->irqdata = &counter->data[0]; |
998 | counter->usrdata = &counter->data[1]; | |
999 | counter->cpu = cpu; | |
1000 | counter->hw_event = *hw_event; | |
1001 | counter->wakeup_pending = 0; | |
04289bb9 | 1002 | counter->group_leader = group_leader; |
621a01ea IM |
1003 | counter->hw_ops = NULL; |
1004 | ||
5c92d124 IM |
1005 | hw_ops = NULL; |
1006 | if (!hw_event->raw && hw_event->type < 0) | |
1007 | hw_ops = sw_perf_counter_init(counter); | |
1008 | if (!hw_ops) { | |
1009 | hw_ops = hw_perf_counter_init(counter); | |
1010 | } | |
1011 | ||
621a01ea IM |
1012 | if (!hw_ops) { |
1013 | kfree(counter); | |
1014 | return NULL; | |
1015 | } | |
1016 | counter->hw_ops = hw_ops; | |
0793a61d TG |
1017 | |
1018 | return counter; | |
1019 | } | |
1020 | ||
1021 | /** | |
9f66a381 IM |
1022 | * sys_perf_task_open - open a performance counter, associate it to a task/cpu |
1023 | * | |
1024 | * @hw_event_uptr: event type attributes for monitoring/sampling | |
0793a61d | 1025 | * @pid: target pid |
9f66a381 IM |
1026 | * @cpu: target cpu |
1027 | * @group_fd: group leader counter fd | |
0793a61d | 1028 | */ |
1d1c7ddb IM |
1029 | asmlinkage int |
1030 | sys_perf_counter_open(struct perf_counter_hw_event *hw_event_uptr __user, | |
1031 | pid_t pid, int cpu, int group_fd) | |
0793a61d | 1032 | { |
04289bb9 | 1033 | struct perf_counter *counter, *group_leader; |
9f66a381 | 1034 | struct perf_counter_hw_event hw_event; |
04289bb9 IM |
1035 | struct perf_counter_context *ctx; |
1036 | struct file *group_file = NULL; | |
1037 | int fput_needed = 0; | |
0793a61d TG |
1038 | int ret; |
1039 | ||
9f66a381 | 1040 | if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0) |
eab656ae TG |
1041 | return -EFAULT; |
1042 | ||
04289bb9 | 1043 | /* |
ccff286d IM |
1044 | * Get the target context (task or percpu): |
1045 | */ | |
1046 | ctx = find_get_context(pid, cpu); | |
1047 | if (IS_ERR(ctx)) | |
1048 | return PTR_ERR(ctx); | |
1049 | ||
1050 | /* | |
1051 | * Look up the group leader (we will attach this counter to it): | |
04289bb9 IM |
1052 | */ |
1053 | group_leader = NULL; | |
1054 | if (group_fd != -1) { | |
1055 | ret = -EINVAL; | |
1056 | group_file = fget_light(group_fd, &fput_needed); | |
1057 | if (!group_file) | |
ccff286d | 1058 | goto err_put_context; |
04289bb9 | 1059 | if (group_file->f_op != &perf_fops) |
ccff286d | 1060 | goto err_put_context; |
04289bb9 IM |
1061 | |
1062 | group_leader = group_file->private_data; | |
1063 | /* | |
ccff286d IM |
1064 | * Do not allow a recursive hierarchy (this new sibling |
1065 | * becoming part of another group-sibling): | |
1066 | */ | |
1067 | if (group_leader->group_leader != group_leader) | |
1068 | goto err_put_context; | |
1069 | /* | |
1070 | * Do not allow to attach to a group in a different | |
1071 | * task or CPU context: | |
04289bb9 | 1072 | */ |
ccff286d IM |
1073 | if (group_leader->ctx != ctx) |
1074 | goto err_put_context; | |
04289bb9 IM |
1075 | } |
1076 | ||
5c92d124 | 1077 | ret = -EINVAL; |
04289bb9 | 1078 | counter = perf_counter_alloc(&hw_event, cpu, group_leader); |
0793a61d TG |
1079 | if (!counter) |
1080 | goto err_put_context; | |
1081 | ||
0793a61d TG |
1082 | perf_install_in_context(ctx, counter, cpu); |
1083 | ||
1084 | ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0); | |
1085 | if (ret < 0) | |
1086 | goto err_remove_free_put_context; | |
1087 | ||
04289bb9 IM |
1088 | out_fput: |
1089 | fput_light(group_file, fput_needed); | |
1090 | ||
0793a61d TG |
1091 | return ret; |
1092 | ||
1093 | err_remove_free_put_context: | |
1094 | mutex_lock(&counter->mutex); | |
04289bb9 | 1095 | perf_counter_remove_from_context(counter); |
0793a61d | 1096 | mutex_unlock(&counter->mutex); |
0793a61d TG |
1097 | kfree(counter); |
1098 | ||
1099 | err_put_context: | |
1100 | put_context(ctx); | |
1101 | ||
04289bb9 | 1102 | goto out_fput; |
0793a61d TG |
1103 | } |
1104 | ||
04289bb9 | 1105 | static void __cpuinit perf_counter_init_cpu(int cpu) |
0793a61d | 1106 | { |
04289bb9 | 1107 | struct perf_cpu_context *cpuctx; |
0793a61d | 1108 | |
04289bb9 IM |
1109 | cpuctx = &per_cpu(perf_cpu_context, cpu); |
1110 | __perf_counter_init_context(&cpuctx->ctx, NULL); | |
0793a61d TG |
1111 | |
1112 | mutex_lock(&perf_resource_mutex); | |
04289bb9 | 1113 | cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu; |
0793a61d | 1114 | mutex_unlock(&perf_resource_mutex); |
04289bb9 | 1115 | |
0793a61d TG |
1116 | hw_perf_counter_setup(); |
1117 | } | |
1118 | ||
1119 | #ifdef CONFIG_HOTPLUG_CPU | |
04289bb9 | 1120 | static void __perf_counter_exit_cpu(void *info) |
0793a61d TG |
1121 | { |
1122 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
1123 | struct perf_counter_context *ctx = &cpuctx->ctx; | |
1124 | struct perf_counter *counter, *tmp; | |
1125 | ||
04289bb9 IM |
1126 | list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry) |
1127 | __perf_counter_remove_from_context(counter); | |
0793a61d TG |
1128 | |
1129 | } | |
04289bb9 | 1130 | static void perf_counter_exit_cpu(int cpu) |
0793a61d | 1131 | { |
04289bb9 | 1132 | smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1); |
0793a61d TG |
1133 | } |
1134 | #else | |
04289bb9 | 1135 | static inline void perf_counter_exit_cpu(int cpu) { } |
0793a61d TG |
1136 | #endif |
1137 | ||
1138 | static int __cpuinit | |
1139 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) | |
1140 | { | |
1141 | unsigned int cpu = (long)hcpu; | |
1142 | ||
1143 | switch (action) { | |
1144 | ||
1145 | case CPU_UP_PREPARE: | |
1146 | case CPU_UP_PREPARE_FROZEN: | |
04289bb9 | 1147 | perf_counter_init_cpu(cpu); |
0793a61d TG |
1148 | break; |
1149 | ||
1150 | case CPU_DOWN_PREPARE: | |
1151 | case CPU_DOWN_PREPARE_FROZEN: | |
04289bb9 | 1152 | perf_counter_exit_cpu(cpu); |
0793a61d TG |
1153 | break; |
1154 | ||
1155 | default: | |
1156 | break; | |
1157 | } | |
1158 | ||
1159 | return NOTIFY_OK; | |
1160 | } | |
1161 | ||
1162 | static struct notifier_block __cpuinitdata perf_cpu_nb = { | |
1163 | .notifier_call = perf_cpu_notify, | |
1164 | }; | |
1165 | ||
1166 | static int __init perf_counter_init(void) | |
1167 | { | |
1168 | perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE, | |
1169 | (void *)(long)smp_processor_id()); | |
1170 | register_cpu_notifier(&perf_cpu_nb); | |
1171 | ||
1172 | return 0; | |
1173 | } | |
1174 | early_initcall(perf_counter_init); | |
1175 | ||
1176 | static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf) | |
1177 | { | |
1178 | return sprintf(buf, "%d\n", perf_reserved_percpu); | |
1179 | } | |
1180 | ||
1181 | static ssize_t | |
1182 | perf_set_reserve_percpu(struct sysdev_class *class, | |
1183 | const char *buf, | |
1184 | size_t count) | |
1185 | { | |
1186 | struct perf_cpu_context *cpuctx; | |
1187 | unsigned long val; | |
1188 | int err, cpu, mpt; | |
1189 | ||
1190 | err = strict_strtoul(buf, 10, &val); | |
1191 | if (err) | |
1192 | return err; | |
1193 | if (val > perf_max_counters) | |
1194 | return -EINVAL; | |
1195 | ||
1196 | mutex_lock(&perf_resource_mutex); | |
1197 | perf_reserved_percpu = val; | |
1198 | for_each_online_cpu(cpu) { | |
1199 | cpuctx = &per_cpu(perf_cpu_context, cpu); | |
1200 | spin_lock_irq(&cpuctx->ctx.lock); | |
1201 | mpt = min(perf_max_counters - cpuctx->ctx.nr_counters, | |
1202 | perf_max_counters - perf_reserved_percpu); | |
1203 | cpuctx->max_pertask = mpt; | |
1204 | spin_unlock_irq(&cpuctx->ctx.lock); | |
1205 | } | |
1206 | mutex_unlock(&perf_resource_mutex); | |
1207 | ||
1208 | return count; | |
1209 | } | |
1210 | ||
1211 | static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf) | |
1212 | { | |
1213 | return sprintf(buf, "%d\n", perf_overcommit); | |
1214 | } | |
1215 | ||
1216 | static ssize_t | |
1217 | perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count) | |
1218 | { | |
1219 | unsigned long val; | |
1220 | int err; | |
1221 | ||
1222 | err = strict_strtoul(buf, 10, &val); | |
1223 | if (err) | |
1224 | return err; | |
1225 | if (val > 1) | |
1226 | return -EINVAL; | |
1227 | ||
1228 | mutex_lock(&perf_resource_mutex); | |
1229 | perf_overcommit = val; | |
1230 | mutex_unlock(&perf_resource_mutex); | |
1231 | ||
1232 | return count; | |
1233 | } | |
1234 | ||
1235 | static SYSDEV_CLASS_ATTR( | |
1236 | reserve_percpu, | |
1237 | 0644, | |
1238 | perf_show_reserve_percpu, | |
1239 | perf_set_reserve_percpu | |
1240 | ); | |
1241 | ||
1242 | static SYSDEV_CLASS_ATTR( | |
1243 | overcommit, | |
1244 | 0644, | |
1245 | perf_show_overcommit, | |
1246 | perf_set_overcommit | |
1247 | ); | |
1248 | ||
1249 | static struct attribute *perfclass_attrs[] = { | |
1250 | &attr_reserve_percpu.attr, | |
1251 | &attr_overcommit.attr, | |
1252 | NULL | |
1253 | }; | |
1254 | ||
1255 | static struct attribute_group perfclass_attr_group = { | |
1256 | .attrs = perfclass_attrs, | |
1257 | .name = "perf_counters", | |
1258 | }; | |
1259 | ||
1260 | static int __init perf_counter_sysfs_init(void) | |
1261 | { | |
1262 | return sysfs_create_group(&cpu_sysdev_class.kset.kobj, | |
1263 | &perfclass_attr_group); | |
1264 | } | |
1265 | device_initcall(perf_counter_sysfs_init); | |
1266 |