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 | ||
170 | if (counter->active) { | |
621a01ea | 171 | counter->hw_ops->hw_perf_counter_disable(counter); |
0793a61d TG |
172 | counter->active = 0; |
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); |
0793a61d TG |
284 | counter->active = 1; |
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 | /* | |
331 | * If the context is active and the counter has not been added | |
332 | * we need to retry the smp call. | |
333 | */ | |
04289bb9 | 334 | if (ctx->nr_active && list_empty(&counter->list_entry)) { |
0793a61d TG |
335 | spin_unlock_irq(&ctx->lock); |
336 | goto retry; | |
337 | } | |
338 | ||
339 | /* | |
340 | * The lock prevents that this context is scheduled in so we | |
341 | * can add the counter safely, if it the call above did not | |
342 | * succeed. | |
343 | */ | |
04289bb9 IM |
344 | if (list_empty(&counter->list_entry)) { |
345 | list_add_counter(counter, ctx); | |
0793a61d TG |
346 | ctx->nr_counters++; |
347 | } | |
348 | spin_unlock_irq(&ctx->lock); | |
349 | } | |
350 | ||
04289bb9 IM |
351 | static void |
352 | counter_sched_out(struct perf_counter *counter, | |
353 | struct perf_cpu_context *cpuctx, | |
354 | struct perf_counter_context *ctx) | |
355 | { | |
356 | if (!counter->active) | |
357 | return; | |
358 | ||
621a01ea | 359 | counter->hw_ops->hw_perf_counter_disable(counter); |
04289bb9 IM |
360 | counter->active = 0; |
361 | counter->oncpu = -1; | |
362 | ||
363 | cpuctx->active_oncpu--; | |
364 | ctx->nr_active--; | |
365 | } | |
366 | ||
367 | static void | |
368 | group_sched_out(struct perf_counter *group_counter, | |
369 | struct perf_cpu_context *cpuctx, | |
370 | struct perf_counter_context *ctx) | |
371 | { | |
372 | struct perf_counter *counter; | |
373 | ||
374 | counter_sched_out(group_counter, cpuctx, ctx); | |
375 | ||
376 | /* | |
377 | * Schedule out siblings (if any): | |
378 | */ | |
379 | list_for_each_entry(counter, &group_counter->sibling_list, list_entry) | |
380 | counter_sched_out(counter, cpuctx, ctx); | |
381 | } | |
382 | ||
0793a61d TG |
383 | /* |
384 | * Called from scheduler to remove the counters of the current task, | |
385 | * with interrupts disabled. | |
386 | * | |
387 | * We stop each counter and update the counter value in counter->count. | |
388 | * | |
389 | * This does not protect us against NMI, but hw_perf_counter_disable() | |
390 | * sets the disabled bit in the control field of counter _before_ | |
391 | * accessing the counter control register. If a NMI hits, then it will | |
392 | * not restart the counter. | |
393 | */ | |
394 | void perf_counter_task_sched_out(struct task_struct *task, int cpu) | |
395 | { | |
396 | struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); | |
397 | struct perf_counter_context *ctx = &task->perf_counter_ctx; | |
398 | struct perf_counter *counter; | |
399 | ||
400 | if (likely(!cpuctx->task_ctx)) | |
401 | return; | |
402 | ||
403 | spin_lock(&ctx->lock); | |
04289bb9 IM |
404 | if (ctx->nr_active) { |
405 | list_for_each_entry(counter, &ctx->counter_list, list_entry) | |
406 | group_sched_out(counter, cpuctx, ctx); | |
0793a61d TG |
407 | } |
408 | spin_unlock(&ctx->lock); | |
409 | cpuctx->task_ctx = NULL; | |
410 | } | |
411 | ||
04289bb9 IM |
412 | static void |
413 | counter_sched_in(struct perf_counter *counter, | |
414 | struct perf_cpu_context *cpuctx, | |
415 | struct perf_counter_context *ctx, | |
416 | int cpu) | |
417 | { | |
621a01ea | 418 | counter->hw_ops->hw_perf_counter_enable(counter); |
04289bb9 IM |
419 | counter->active = 1; |
420 | counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */ | |
421 | ||
422 | cpuctx->active_oncpu++; | |
423 | ctx->nr_active++; | |
424 | } | |
425 | ||
426 | static void | |
427 | group_sched_in(struct perf_counter *group_counter, | |
428 | struct perf_cpu_context *cpuctx, | |
429 | struct perf_counter_context *ctx, | |
430 | int cpu) | |
431 | { | |
432 | struct perf_counter *counter; | |
433 | ||
434 | counter_sched_in(group_counter, cpuctx, ctx, cpu); | |
435 | ||
436 | /* | |
437 | * Schedule in siblings as one group (if any): | |
438 | */ | |
439 | list_for_each_entry(counter, &group_counter->sibling_list, list_entry) | |
440 | counter_sched_in(counter, cpuctx, ctx, cpu); | |
441 | } | |
442 | ||
0793a61d TG |
443 | /* |
444 | * Called from scheduler to add the counters of the current task | |
445 | * with interrupts disabled. | |
446 | * | |
447 | * We restore the counter value and then enable it. | |
448 | * | |
449 | * This does not protect us against NMI, but hw_perf_counter_enable() | |
450 | * sets the enabled bit in the control field of counter _before_ | |
451 | * accessing the counter control register. If a NMI hits, then it will | |
452 | * keep the counter running. | |
453 | */ | |
454 | void perf_counter_task_sched_in(struct task_struct *task, int cpu) | |
455 | { | |
456 | struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); | |
457 | struct perf_counter_context *ctx = &task->perf_counter_ctx; | |
458 | struct perf_counter *counter; | |
459 | ||
460 | if (likely(!ctx->nr_counters)) | |
461 | return; | |
462 | ||
463 | spin_lock(&ctx->lock); | |
04289bb9 | 464 | list_for_each_entry(counter, &ctx->counter_list, list_entry) { |
0793a61d TG |
465 | if (ctx->nr_active == cpuctx->max_pertask) |
466 | break; | |
04289bb9 IM |
467 | |
468 | /* | |
469 | * Listen to the 'cpu' scheduling filter constraint | |
470 | * of counters: | |
471 | */ | |
0793a61d TG |
472 | if (counter->cpu != -1 && counter->cpu != cpu) |
473 | continue; | |
474 | ||
04289bb9 | 475 | group_sched_in(counter, cpuctx, ctx, cpu); |
0793a61d TG |
476 | } |
477 | spin_unlock(&ctx->lock); | |
04289bb9 | 478 | |
0793a61d TG |
479 | cpuctx->task_ctx = ctx; |
480 | } | |
481 | ||
482 | void perf_counter_task_tick(struct task_struct *curr, int cpu) | |
483 | { | |
484 | struct perf_counter_context *ctx = &curr->perf_counter_ctx; | |
485 | struct perf_counter *counter; | |
5c92d124 | 486 | u64 perf_flags; |
0793a61d TG |
487 | |
488 | if (likely(!ctx->nr_counters)) | |
489 | return; | |
490 | ||
491 | perf_counter_task_sched_out(curr, cpu); | |
492 | ||
493 | spin_lock(&ctx->lock); | |
494 | ||
495 | /* | |
04289bb9 | 496 | * Rotate the first entry last (works just fine for group counters too): |
0793a61d | 497 | */ |
01b2838c | 498 | perf_flags = hw_perf_save_disable(); |
04289bb9 IM |
499 | list_for_each_entry(counter, &ctx->counter_list, list_entry) { |
500 | list_del(&counter->list_entry); | |
501 | list_add_tail(&counter->list_entry, &ctx->counter_list); | |
0793a61d TG |
502 | break; |
503 | } | |
01b2838c | 504 | hw_perf_restore(perf_flags); |
0793a61d TG |
505 | |
506 | spin_unlock(&ctx->lock); | |
507 | ||
508 | perf_counter_task_sched_in(curr, cpu); | |
509 | } | |
510 | ||
04289bb9 IM |
511 | /* |
512 | * Initialize the perf_counter context in a task_struct: | |
513 | */ | |
514 | static void | |
515 | __perf_counter_init_context(struct perf_counter_context *ctx, | |
516 | struct task_struct *task) | |
517 | { | |
518 | spin_lock_init(&ctx->lock); | |
519 | INIT_LIST_HEAD(&ctx->counter_list); | |
520 | ctx->nr_counters = 0; | |
521 | ctx->task = task; | |
522 | } | |
0793a61d TG |
523 | /* |
524 | * Initialize the perf_counter context in task_struct | |
525 | */ | |
526 | void perf_counter_init_task(struct task_struct *task) | |
527 | { | |
04289bb9 | 528 | __perf_counter_init_context(&task->perf_counter_ctx, task); |
0793a61d TG |
529 | } |
530 | ||
531 | /* | |
532 | * Cross CPU call to read the hardware counter | |
533 | */ | |
534 | static void __hw_perf_counter_read(void *info) | |
535 | { | |
621a01ea IM |
536 | struct perf_counter *counter = info; |
537 | ||
538 | counter->hw_ops->hw_perf_counter_read(counter); | |
0793a61d TG |
539 | } |
540 | ||
04289bb9 | 541 | static u64 perf_counter_read(struct perf_counter *counter) |
0793a61d TG |
542 | { |
543 | /* | |
544 | * If counter is enabled and currently active on a CPU, update the | |
545 | * value in the counter structure: | |
546 | */ | |
547 | if (counter->active) { | |
548 | smp_call_function_single(counter->oncpu, | |
549 | __hw_perf_counter_read, counter, 1); | |
550 | } | |
551 | ||
04289bb9 | 552 | return perf_counter_read_safe(counter); |
0793a61d TG |
553 | } |
554 | ||
555 | /* | |
556 | * Cross CPU call to switch performance data pointers | |
557 | */ | |
558 | static void __perf_switch_irq_data(void *info) | |
559 | { | |
560 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
561 | struct perf_counter *counter = info; | |
562 | struct perf_counter_context *ctx = counter->ctx; | |
563 | struct perf_data *oldirqdata = counter->irqdata; | |
564 | ||
565 | /* | |
566 | * If this is a task context, we need to check whether it is | |
567 | * the current task context of this cpu. If not it has been | |
568 | * scheduled out before the smp call arrived. | |
569 | */ | |
570 | if (ctx->task) { | |
571 | if (cpuctx->task_ctx != ctx) | |
572 | return; | |
573 | spin_lock(&ctx->lock); | |
574 | } | |
575 | ||
576 | /* Change the pointer NMI safe */ | |
577 | atomic_long_set((atomic_long_t *)&counter->irqdata, | |
578 | (unsigned long) counter->usrdata); | |
579 | counter->usrdata = oldirqdata; | |
580 | ||
581 | if (ctx->task) | |
582 | spin_unlock(&ctx->lock); | |
583 | } | |
584 | ||
585 | static struct perf_data *perf_switch_irq_data(struct perf_counter *counter) | |
586 | { | |
587 | struct perf_counter_context *ctx = counter->ctx; | |
588 | struct perf_data *oldirqdata = counter->irqdata; | |
589 | struct task_struct *task = ctx->task; | |
590 | ||
591 | if (!task) { | |
592 | smp_call_function_single(counter->cpu, | |
593 | __perf_switch_irq_data, | |
594 | counter, 1); | |
595 | return counter->usrdata; | |
596 | } | |
597 | ||
598 | retry: | |
599 | spin_lock_irq(&ctx->lock); | |
600 | if (!counter->active) { | |
601 | counter->irqdata = counter->usrdata; | |
602 | counter->usrdata = oldirqdata; | |
603 | spin_unlock_irq(&ctx->lock); | |
604 | return oldirqdata; | |
605 | } | |
606 | spin_unlock_irq(&ctx->lock); | |
607 | task_oncpu_function_call(task, __perf_switch_irq_data, counter); | |
608 | /* Might have failed, because task was scheduled out */ | |
609 | if (counter->irqdata == oldirqdata) | |
610 | goto retry; | |
611 | ||
612 | return counter->usrdata; | |
613 | } | |
614 | ||
615 | static void put_context(struct perf_counter_context *ctx) | |
616 | { | |
617 | if (ctx->task) | |
618 | put_task_struct(ctx->task); | |
619 | } | |
620 | ||
621 | static struct perf_counter_context *find_get_context(pid_t pid, int cpu) | |
622 | { | |
623 | struct perf_cpu_context *cpuctx; | |
624 | struct perf_counter_context *ctx; | |
625 | struct task_struct *task; | |
626 | ||
627 | /* | |
628 | * If cpu is not a wildcard then this is a percpu counter: | |
629 | */ | |
630 | if (cpu != -1) { | |
631 | /* Must be root to operate on a CPU counter: */ | |
632 | if (!capable(CAP_SYS_ADMIN)) | |
633 | return ERR_PTR(-EACCES); | |
634 | ||
635 | if (cpu < 0 || cpu > num_possible_cpus()) | |
636 | return ERR_PTR(-EINVAL); | |
637 | ||
638 | /* | |
639 | * We could be clever and allow to attach a counter to an | |
640 | * offline CPU and activate it when the CPU comes up, but | |
641 | * that's for later. | |
642 | */ | |
643 | if (!cpu_isset(cpu, cpu_online_map)) | |
644 | return ERR_PTR(-ENODEV); | |
645 | ||
646 | cpuctx = &per_cpu(perf_cpu_context, cpu); | |
647 | ctx = &cpuctx->ctx; | |
648 | ||
649 | WARN_ON_ONCE(ctx->task); | |
650 | return ctx; | |
651 | } | |
652 | ||
653 | rcu_read_lock(); | |
654 | if (!pid) | |
655 | task = current; | |
656 | else | |
657 | task = find_task_by_vpid(pid); | |
658 | if (task) | |
659 | get_task_struct(task); | |
660 | rcu_read_unlock(); | |
661 | ||
662 | if (!task) | |
663 | return ERR_PTR(-ESRCH); | |
664 | ||
665 | ctx = &task->perf_counter_ctx; | |
666 | ctx->task = task; | |
667 | ||
668 | /* Reuse ptrace permission checks for now. */ | |
669 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) { | |
670 | put_context(ctx); | |
671 | return ERR_PTR(-EACCES); | |
672 | } | |
673 | ||
674 | return ctx; | |
675 | } | |
676 | ||
677 | /* | |
678 | * Called when the last reference to the file is gone. | |
679 | */ | |
680 | static int perf_release(struct inode *inode, struct file *file) | |
681 | { | |
682 | struct perf_counter *counter = file->private_data; | |
683 | struct perf_counter_context *ctx = counter->ctx; | |
684 | ||
685 | file->private_data = NULL; | |
686 | ||
687 | mutex_lock(&counter->mutex); | |
688 | ||
04289bb9 | 689 | perf_counter_remove_from_context(counter); |
0793a61d TG |
690 | put_context(ctx); |
691 | ||
692 | mutex_unlock(&counter->mutex); | |
693 | ||
694 | kfree(counter); | |
695 | ||
696 | return 0; | |
697 | } | |
698 | ||
699 | /* | |
700 | * Read the performance counter - simple non blocking version for now | |
701 | */ | |
702 | static ssize_t | |
703 | perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count) | |
704 | { | |
705 | u64 cntval; | |
706 | ||
707 | if (count != sizeof(cntval)) | |
708 | return -EINVAL; | |
709 | ||
710 | mutex_lock(&counter->mutex); | |
04289bb9 | 711 | cntval = perf_counter_read(counter); |
0793a61d TG |
712 | mutex_unlock(&counter->mutex); |
713 | ||
714 | return put_user(cntval, (u64 __user *) buf) ? -EFAULT : sizeof(cntval); | |
715 | } | |
716 | ||
717 | static ssize_t | |
718 | perf_copy_usrdata(struct perf_data *usrdata, char __user *buf, size_t count) | |
719 | { | |
720 | if (!usrdata->len) | |
721 | return 0; | |
722 | ||
723 | count = min(count, (size_t)usrdata->len); | |
724 | if (copy_to_user(buf, usrdata->data + usrdata->rd_idx, count)) | |
725 | return -EFAULT; | |
726 | ||
727 | /* Adjust the counters */ | |
728 | usrdata->len -= count; | |
729 | if (!usrdata->len) | |
730 | usrdata->rd_idx = 0; | |
731 | else | |
732 | usrdata->rd_idx += count; | |
733 | ||
734 | return count; | |
735 | } | |
736 | ||
737 | static ssize_t | |
738 | perf_read_irq_data(struct perf_counter *counter, | |
739 | char __user *buf, | |
740 | size_t count, | |
741 | int nonblocking) | |
742 | { | |
743 | struct perf_data *irqdata, *usrdata; | |
744 | DECLARE_WAITQUEUE(wait, current); | |
745 | ssize_t res; | |
746 | ||
747 | irqdata = counter->irqdata; | |
748 | usrdata = counter->usrdata; | |
749 | ||
750 | if (usrdata->len + irqdata->len >= count) | |
751 | goto read_pending; | |
752 | ||
753 | if (nonblocking) | |
754 | return -EAGAIN; | |
755 | ||
756 | spin_lock_irq(&counter->waitq.lock); | |
757 | __add_wait_queue(&counter->waitq, &wait); | |
758 | for (;;) { | |
759 | set_current_state(TASK_INTERRUPTIBLE); | |
760 | if (usrdata->len + irqdata->len >= count) | |
761 | break; | |
762 | ||
763 | if (signal_pending(current)) | |
764 | break; | |
765 | ||
766 | spin_unlock_irq(&counter->waitq.lock); | |
767 | schedule(); | |
768 | spin_lock_irq(&counter->waitq.lock); | |
769 | } | |
770 | __remove_wait_queue(&counter->waitq, &wait); | |
771 | __set_current_state(TASK_RUNNING); | |
772 | spin_unlock_irq(&counter->waitq.lock); | |
773 | ||
774 | if (usrdata->len + irqdata->len < count) | |
775 | return -ERESTARTSYS; | |
776 | read_pending: | |
777 | mutex_lock(&counter->mutex); | |
778 | ||
779 | /* Drain pending data first: */ | |
780 | res = perf_copy_usrdata(usrdata, buf, count); | |
781 | if (res < 0 || res == count) | |
782 | goto out; | |
783 | ||
784 | /* Switch irq buffer: */ | |
785 | usrdata = perf_switch_irq_data(counter); | |
786 | if (perf_copy_usrdata(usrdata, buf + res, count - res) < 0) { | |
787 | if (!res) | |
788 | res = -EFAULT; | |
789 | } else { | |
790 | res = count; | |
791 | } | |
792 | out: | |
793 | mutex_unlock(&counter->mutex); | |
794 | ||
795 | return res; | |
796 | } | |
797 | ||
798 | static ssize_t | |
799 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
800 | { | |
801 | struct perf_counter *counter = file->private_data; | |
802 | ||
9f66a381 | 803 | switch (counter->hw_event.record_type) { |
0793a61d TG |
804 | case PERF_RECORD_SIMPLE: |
805 | return perf_read_hw(counter, buf, count); | |
806 | ||
807 | case PERF_RECORD_IRQ: | |
808 | case PERF_RECORD_GROUP: | |
809 | return perf_read_irq_data(counter, buf, count, | |
810 | file->f_flags & O_NONBLOCK); | |
811 | } | |
812 | return -EINVAL; | |
813 | } | |
814 | ||
815 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
816 | { | |
817 | struct perf_counter *counter = file->private_data; | |
818 | unsigned int events = 0; | |
819 | unsigned long flags; | |
820 | ||
821 | poll_wait(file, &counter->waitq, wait); | |
822 | ||
823 | spin_lock_irqsave(&counter->waitq.lock, flags); | |
824 | if (counter->usrdata->len || counter->irqdata->len) | |
825 | events |= POLLIN; | |
826 | spin_unlock_irqrestore(&counter->waitq.lock, flags); | |
827 | ||
828 | return events; | |
829 | } | |
830 | ||
831 | static const struct file_operations perf_fops = { | |
832 | .release = perf_release, | |
833 | .read = perf_read, | |
834 | .poll = perf_poll, | |
835 | }; | |
836 | ||
5c92d124 IM |
837 | static void cpu_clock_perf_counter_enable(struct perf_counter *counter) |
838 | { | |
839 | } | |
840 | ||
841 | static void cpu_clock_perf_counter_disable(struct perf_counter *counter) | |
842 | { | |
843 | } | |
844 | ||
845 | static void cpu_clock_perf_counter_read(struct perf_counter *counter) | |
846 | { | |
847 | int cpu = raw_smp_processor_id(); | |
848 | ||
849 | atomic64_counter_set(counter, cpu_clock(cpu)); | |
850 | } | |
851 | ||
852 | static const struct hw_perf_counter_ops perf_ops_cpu_clock = { | |
853 | .hw_perf_counter_enable = cpu_clock_perf_counter_enable, | |
854 | .hw_perf_counter_disable = cpu_clock_perf_counter_disable, | |
855 | .hw_perf_counter_read = cpu_clock_perf_counter_read, | |
856 | }; | |
857 | ||
858 | static const struct hw_perf_counter_ops * | |
859 | sw_perf_counter_init(struct perf_counter *counter) | |
860 | { | |
861 | const struct hw_perf_counter_ops *hw_ops = NULL; | |
862 | ||
863 | switch (counter->hw_event.type) { | |
864 | case PERF_COUNT_CPU_CLOCK: | |
865 | hw_ops = &perf_ops_cpu_clock; | |
866 | break; | |
867 | default: | |
868 | break; | |
869 | } | |
870 | return hw_ops; | |
871 | } | |
872 | ||
0793a61d TG |
873 | /* |
874 | * Allocate and initialize a counter structure | |
875 | */ | |
876 | static struct perf_counter * | |
04289bb9 IM |
877 | perf_counter_alloc(struct perf_counter_hw_event *hw_event, |
878 | int cpu, | |
879 | struct perf_counter *group_leader) | |
0793a61d | 880 | { |
5c92d124 | 881 | const struct hw_perf_counter_ops *hw_ops; |
621a01ea | 882 | struct perf_counter *counter; |
0793a61d | 883 | |
621a01ea | 884 | counter = kzalloc(sizeof(*counter), GFP_KERNEL); |
0793a61d TG |
885 | if (!counter) |
886 | return NULL; | |
887 | ||
04289bb9 IM |
888 | /* |
889 | * Single counters are their own group leaders, with an | |
890 | * empty sibling list: | |
891 | */ | |
892 | if (!group_leader) | |
893 | group_leader = counter; | |
894 | ||
0793a61d | 895 | mutex_init(&counter->mutex); |
04289bb9 IM |
896 | INIT_LIST_HEAD(&counter->list_entry); |
897 | INIT_LIST_HEAD(&counter->sibling_list); | |
0793a61d TG |
898 | init_waitqueue_head(&counter->waitq); |
899 | ||
9f66a381 IM |
900 | counter->irqdata = &counter->data[0]; |
901 | counter->usrdata = &counter->data[1]; | |
902 | counter->cpu = cpu; | |
903 | counter->hw_event = *hw_event; | |
904 | counter->wakeup_pending = 0; | |
04289bb9 | 905 | counter->group_leader = group_leader; |
621a01ea IM |
906 | counter->hw_ops = NULL; |
907 | ||
5c92d124 IM |
908 | hw_ops = NULL; |
909 | if (!hw_event->raw && hw_event->type < 0) | |
910 | hw_ops = sw_perf_counter_init(counter); | |
911 | if (!hw_ops) { | |
912 | hw_ops = hw_perf_counter_init(counter); | |
913 | } | |
914 | ||
621a01ea IM |
915 | if (!hw_ops) { |
916 | kfree(counter); | |
917 | return NULL; | |
918 | } | |
919 | counter->hw_ops = hw_ops; | |
0793a61d TG |
920 | |
921 | return counter; | |
922 | } | |
923 | ||
924 | /** | |
9f66a381 IM |
925 | * sys_perf_task_open - open a performance counter, associate it to a task/cpu |
926 | * | |
927 | * @hw_event_uptr: event type attributes for monitoring/sampling | |
0793a61d | 928 | * @pid: target pid |
9f66a381 IM |
929 | * @cpu: target cpu |
930 | * @group_fd: group leader counter fd | |
0793a61d | 931 | */ |
9f66a381 IM |
932 | asmlinkage int sys_perf_counter_open( |
933 | ||
934 | struct perf_counter_hw_event *hw_event_uptr __user, | |
935 | pid_t pid, | |
936 | int cpu, | |
937 | int group_fd) | |
938 | ||
0793a61d | 939 | { |
04289bb9 | 940 | struct perf_counter *counter, *group_leader; |
9f66a381 | 941 | struct perf_counter_hw_event hw_event; |
04289bb9 IM |
942 | struct perf_counter_context *ctx; |
943 | struct file *group_file = NULL; | |
944 | int fput_needed = 0; | |
0793a61d TG |
945 | int ret; |
946 | ||
9f66a381 | 947 | if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0) |
eab656ae TG |
948 | return -EFAULT; |
949 | ||
04289bb9 | 950 | /* |
ccff286d IM |
951 | * Get the target context (task or percpu): |
952 | */ | |
953 | ctx = find_get_context(pid, cpu); | |
954 | if (IS_ERR(ctx)) | |
955 | return PTR_ERR(ctx); | |
956 | ||
957 | /* | |
958 | * Look up the group leader (we will attach this counter to it): | |
04289bb9 IM |
959 | */ |
960 | group_leader = NULL; | |
961 | if (group_fd != -1) { | |
962 | ret = -EINVAL; | |
963 | group_file = fget_light(group_fd, &fput_needed); | |
964 | if (!group_file) | |
ccff286d | 965 | goto err_put_context; |
04289bb9 | 966 | if (group_file->f_op != &perf_fops) |
ccff286d | 967 | goto err_put_context; |
04289bb9 IM |
968 | |
969 | group_leader = group_file->private_data; | |
970 | /* | |
ccff286d IM |
971 | * Do not allow a recursive hierarchy (this new sibling |
972 | * becoming part of another group-sibling): | |
973 | */ | |
974 | if (group_leader->group_leader != group_leader) | |
975 | goto err_put_context; | |
976 | /* | |
977 | * Do not allow to attach to a group in a different | |
978 | * task or CPU context: | |
04289bb9 | 979 | */ |
ccff286d IM |
980 | if (group_leader->ctx != ctx) |
981 | goto err_put_context; | |
04289bb9 IM |
982 | } |
983 | ||
5c92d124 | 984 | ret = -EINVAL; |
04289bb9 | 985 | counter = perf_counter_alloc(&hw_event, cpu, group_leader); |
0793a61d TG |
986 | if (!counter) |
987 | goto err_put_context; | |
988 | ||
0793a61d TG |
989 | perf_install_in_context(ctx, counter, cpu); |
990 | ||
991 | ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0); | |
992 | if (ret < 0) | |
993 | goto err_remove_free_put_context; | |
994 | ||
04289bb9 IM |
995 | out_fput: |
996 | fput_light(group_file, fput_needed); | |
997 | ||
0793a61d TG |
998 | return ret; |
999 | ||
1000 | err_remove_free_put_context: | |
1001 | mutex_lock(&counter->mutex); | |
04289bb9 | 1002 | perf_counter_remove_from_context(counter); |
0793a61d | 1003 | mutex_unlock(&counter->mutex); |
0793a61d TG |
1004 | kfree(counter); |
1005 | ||
1006 | err_put_context: | |
1007 | put_context(ctx); | |
1008 | ||
04289bb9 | 1009 | goto out_fput; |
0793a61d TG |
1010 | } |
1011 | ||
04289bb9 | 1012 | static void __cpuinit perf_counter_init_cpu(int cpu) |
0793a61d | 1013 | { |
04289bb9 | 1014 | struct perf_cpu_context *cpuctx; |
0793a61d | 1015 | |
04289bb9 IM |
1016 | cpuctx = &per_cpu(perf_cpu_context, cpu); |
1017 | __perf_counter_init_context(&cpuctx->ctx, NULL); | |
0793a61d TG |
1018 | |
1019 | mutex_lock(&perf_resource_mutex); | |
04289bb9 | 1020 | cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu; |
0793a61d | 1021 | mutex_unlock(&perf_resource_mutex); |
04289bb9 | 1022 | |
0793a61d TG |
1023 | hw_perf_counter_setup(); |
1024 | } | |
1025 | ||
1026 | #ifdef CONFIG_HOTPLUG_CPU | |
04289bb9 | 1027 | static void __perf_counter_exit_cpu(void *info) |
0793a61d TG |
1028 | { |
1029 | struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); | |
1030 | struct perf_counter_context *ctx = &cpuctx->ctx; | |
1031 | struct perf_counter *counter, *tmp; | |
1032 | ||
04289bb9 IM |
1033 | list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry) |
1034 | __perf_counter_remove_from_context(counter); | |
0793a61d TG |
1035 | |
1036 | } | |
04289bb9 | 1037 | static void perf_counter_exit_cpu(int cpu) |
0793a61d | 1038 | { |
04289bb9 | 1039 | smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1); |
0793a61d TG |
1040 | } |
1041 | #else | |
04289bb9 | 1042 | static inline void perf_counter_exit_cpu(int cpu) { } |
0793a61d TG |
1043 | #endif |
1044 | ||
1045 | static int __cpuinit | |
1046 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) | |
1047 | { | |
1048 | unsigned int cpu = (long)hcpu; | |
1049 | ||
1050 | switch (action) { | |
1051 | ||
1052 | case CPU_UP_PREPARE: | |
1053 | case CPU_UP_PREPARE_FROZEN: | |
04289bb9 | 1054 | perf_counter_init_cpu(cpu); |
0793a61d TG |
1055 | break; |
1056 | ||
1057 | case CPU_DOWN_PREPARE: | |
1058 | case CPU_DOWN_PREPARE_FROZEN: | |
04289bb9 | 1059 | perf_counter_exit_cpu(cpu); |
0793a61d TG |
1060 | break; |
1061 | ||
1062 | default: | |
1063 | break; | |
1064 | } | |
1065 | ||
1066 | return NOTIFY_OK; | |
1067 | } | |
1068 | ||
1069 | static struct notifier_block __cpuinitdata perf_cpu_nb = { | |
1070 | .notifier_call = perf_cpu_notify, | |
1071 | }; | |
1072 | ||
1073 | static int __init perf_counter_init(void) | |
1074 | { | |
1075 | perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE, | |
1076 | (void *)(long)smp_processor_id()); | |
1077 | register_cpu_notifier(&perf_cpu_nb); | |
1078 | ||
1079 | return 0; | |
1080 | } | |
1081 | early_initcall(perf_counter_init); | |
1082 | ||
1083 | static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf) | |
1084 | { | |
1085 | return sprintf(buf, "%d\n", perf_reserved_percpu); | |
1086 | } | |
1087 | ||
1088 | static ssize_t | |
1089 | perf_set_reserve_percpu(struct sysdev_class *class, | |
1090 | const char *buf, | |
1091 | size_t count) | |
1092 | { | |
1093 | struct perf_cpu_context *cpuctx; | |
1094 | unsigned long val; | |
1095 | int err, cpu, mpt; | |
1096 | ||
1097 | err = strict_strtoul(buf, 10, &val); | |
1098 | if (err) | |
1099 | return err; | |
1100 | if (val > perf_max_counters) | |
1101 | return -EINVAL; | |
1102 | ||
1103 | mutex_lock(&perf_resource_mutex); | |
1104 | perf_reserved_percpu = val; | |
1105 | for_each_online_cpu(cpu) { | |
1106 | cpuctx = &per_cpu(perf_cpu_context, cpu); | |
1107 | spin_lock_irq(&cpuctx->ctx.lock); | |
1108 | mpt = min(perf_max_counters - cpuctx->ctx.nr_counters, | |
1109 | perf_max_counters - perf_reserved_percpu); | |
1110 | cpuctx->max_pertask = mpt; | |
1111 | spin_unlock_irq(&cpuctx->ctx.lock); | |
1112 | } | |
1113 | mutex_unlock(&perf_resource_mutex); | |
1114 | ||
1115 | return count; | |
1116 | } | |
1117 | ||
1118 | static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf) | |
1119 | { | |
1120 | return sprintf(buf, "%d\n", perf_overcommit); | |
1121 | } | |
1122 | ||
1123 | static ssize_t | |
1124 | perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count) | |
1125 | { | |
1126 | unsigned long val; | |
1127 | int err; | |
1128 | ||
1129 | err = strict_strtoul(buf, 10, &val); | |
1130 | if (err) | |
1131 | return err; | |
1132 | if (val > 1) | |
1133 | return -EINVAL; | |
1134 | ||
1135 | mutex_lock(&perf_resource_mutex); | |
1136 | perf_overcommit = val; | |
1137 | mutex_unlock(&perf_resource_mutex); | |
1138 | ||
1139 | return count; | |
1140 | } | |
1141 | ||
1142 | static SYSDEV_CLASS_ATTR( | |
1143 | reserve_percpu, | |
1144 | 0644, | |
1145 | perf_show_reserve_percpu, | |
1146 | perf_set_reserve_percpu | |
1147 | ); | |
1148 | ||
1149 | static SYSDEV_CLASS_ATTR( | |
1150 | overcommit, | |
1151 | 0644, | |
1152 | perf_show_overcommit, | |
1153 | perf_set_overcommit | |
1154 | ); | |
1155 | ||
1156 | static struct attribute *perfclass_attrs[] = { | |
1157 | &attr_reserve_percpu.attr, | |
1158 | &attr_overcommit.attr, | |
1159 | NULL | |
1160 | }; | |
1161 | ||
1162 | static struct attribute_group perfclass_attr_group = { | |
1163 | .attrs = perfclass_attrs, | |
1164 | .name = "perf_counters", | |
1165 | }; | |
1166 | ||
1167 | static int __init perf_counter_sysfs_init(void) | |
1168 | { | |
1169 | return sysfs_create_group(&cpu_sysdev_class.kset.kobj, | |
1170 | &perfclass_attr_group); | |
1171 | } | |
1172 | device_initcall(perf_counter_sysfs_init); | |
1173 |