Commit | Line | Data |
---|---|---|
d58d1454 MD |
1 | /* |
2 | * lttng-context-perf-counters.c | |
3 | * | |
4 | * LTTng UST performance monitoring counters (perf-counters) integration. | |
5 | * | |
6 | * Copyright (C) 2009-2014 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
7 | * | |
8 | * This library is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU Lesser General Public | |
10 | * License as published by the Free Software Foundation; only | |
11 | * version 2.1 of the License. | |
12 | * | |
13 | * This library is distributed in the hope that it will be useful, | |
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
16 | * Lesser General Public License for more details. | |
17 | * | |
18 | * You should have received a copy of the GNU Lesser General Public | |
19 | * License along with this library; if not, write to the Free Software | |
20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
21 | */ | |
22 | ||
23 | #include <sys/types.h> | |
24 | #include <unistd.h> | |
25 | #include <string.h> | |
26 | #include <stdlib.h> | |
27 | #include <stdio.h> | |
649fb6b3 | 28 | #include <stdbool.h> |
d58d1454 MD |
29 | #include <sys/mman.h> |
30 | #include <sys/syscall.h> | |
31 | #include <linux/perf_event.h> | |
32 | #include <lttng/ust-events.h> | |
33 | #include <lttng/ust-tracer.h> | |
34 | #include <lttng/ringbuffer-config.h> | |
35 | #include <urcu/system.h> | |
36 | #include <urcu/arch.h> | |
37 | #include <urcu/rculist.h> | |
38 | #include <helper.h> | |
39 | #include <urcu/ref.h> | |
40 | #include <usterr-signal-safe.h> | |
41 | #include <signal.h> | |
42 | #include "lttng-tracer-core.h" | |
43 | ||
44 | /* | |
45 | * We use a global perf counter key and iterate on per-thread RCU lists | |
46 | * of fields in the fast path, even though this is not strictly speaking | |
47 | * what would provide the best fast-path complexity, to ensure teardown | |
48 | * of sessions vs thread exit is handled racelessly. | |
49 | * | |
50 | * Updates and traversals of thread_list are protected by UST lock. | |
51 | * Updates to rcu_field_list are protected by UST lock. | |
52 | */ | |
53 | ||
54 | struct lttng_perf_counter_thread_field { | |
55 | struct lttng_perf_counter_field *field; /* Back reference */ | |
56 | struct perf_event_mmap_page *pc; | |
57 | struct cds_list_head thread_field_node; /* Per-field list of thread fields (node) */ | |
58 | struct cds_list_head rcu_field_node; /* RCU per-thread list of fields (node) */ | |
b9389e6e | 59 | int fd; /* Perf FD */ |
d58d1454 MD |
60 | }; |
61 | ||
62 | struct lttng_perf_counter_thread { | |
63 | struct cds_list_head rcu_field_list; /* RCU per-thread list of fields */ | |
64 | }; | |
65 | ||
66 | struct lttng_perf_counter_field { | |
67 | struct perf_event_attr attr; | |
68 | struct cds_list_head thread_field_list; /* Per-field list of thread fields */ | |
69 | }; | |
70 | ||
71 | static pthread_key_t perf_counter_key; | |
72 | ||
73 | static | |
53569322 | 74 | size_t perf_counter_get_size(struct lttng_ctx_field *field, size_t offset) |
d58d1454 MD |
75 | { |
76 | size_t size = 0; | |
77 | ||
78 | size += lib_ring_buffer_align(offset, lttng_alignof(uint64_t)); | |
79 | size += sizeof(uint64_t); | |
80 | return size; | |
81 | } | |
82 | ||
a3a8d943 MD |
83 | static |
84 | uint64_t read_perf_counter_syscall( | |
85 | struct lttng_perf_counter_thread_field *thread_field) | |
86 | { | |
87 | uint64_t count; | |
88 | ||
89 | if (caa_unlikely(thread_field->fd < 0)) | |
90 | return 0; | |
91 | ||
92 | if (caa_unlikely(read(thread_field->fd, &count, sizeof(count)) | |
93 | < sizeof(count))) | |
94 | return 0; | |
95 | ||
96 | return count; | |
97 | } | |
98 | ||
d58d1454 MD |
99 | #if defined(__x86_64__) || defined(__i386__) |
100 | ||
101 | static | |
102 | uint64_t rdpmc(unsigned int counter) | |
103 | { | |
104 | unsigned int low, high; | |
105 | ||
106 | asm volatile("rdpmc" : "=a" (low), "=d" (high) : "c" (counter)); | |
107 | ||
108 | return low | ((uint64_t) high) << 32; | |
109 | } | |
110 | ||
d58d1454 | 111 | static |
a3a8d943 | 112 | uint64_t arch_read_perf_counter( |
d286ad50 | 113 | struct lttng_perf_counter_thread_field *thread_field) |
d58d1454 MD |
114 | { |
115 | uint32_t seq, idx; | |
116 | uint64_t count; | |
d286ad50 | 117 | struct perf_event_mmap_page *pc = thread_field->pc; |
d58d1454 MD |
118 | |
119 | if (caa_unlikely(!pc)) | |
120 | return 0; | |
121 | ||
122 | do { | |
123 | seq = CMM_LOAD_SHARED(pc->lock); | |
124 | cmm_barrier(); | |
125 | ||
126 | idx = pc->index; | |
a3a8d943 | 127 | if (caa_likely(pc->cap_user_rdpmc && idx)) { |
4f58f54f MD |
128 | int64_t pmcval; |
129 | ||
130 | pmcval = rdpmc(idx - 1); | |
131 | /* Sign-extend the pmc register result. */ | |
132 | pmcval <<= 64 - pc->pmc_width; | |
133 | pmcval >>= 64 - pc->pmc_width; | |
134 | count = pc->offset + pmcval; | |
135 | } else { | |
a3a8d943 MD |
136 | /* Fall-back on system call if rdpmc cannot be used. */ |
137 | return read_perf_counter_syscall(thread_field); | |
4f58f54f | 138 | } |
d58d1454 MD |
139 | cmm_barrier(); |
140 | } while (CMM_LOAD_SHARED(pc->lock) != seq); | |
141 | ||
142 | return count; | |
143 | } | |
144 | ||
d286ad50 | 145 | static |
a3a8d943 | 146 | int arch_perf_keep_fd(struct lttng_perf_counter_thread_field *thread_field) |
d286ad50 | 147 | { |
a3a8d943 | 148 | struct perf_event_mmap_page *pc = thread_field->pc; |
d286ad50 | 149 | |
a3a8d943 | 150 | if (!pc) |
d286ad50 | 151 | return 0; |
a3a8d943 MD |
152 | return !pc->cap_user_rdpmc; |
153 | } | |
d286ad50 | 154 | |
a3a8d943 | 155 | #else |
d286ad50 | 156 | |
a3a8d943 MD |
157 | /* Generic (slow) implementation using a read system call. */ |
158 | static | |
159 | uint64_t arch_read_perf_counter( | |
160 | struct lttng_perf_counter_thread_field *thread_field) | |
161 | { | |
162 | return read_perf_counter_syscall(thread_field); | |
d286ad50 JD |
163 | } |
164 | ||
a3a8d943 MD |
165 | static |
166 | int arch_perf_keep_fd(struct lttng_perf_counter_thread_field *thread_field) | |
167 | { | |
168 | return 1; | |
169 | } | |
d286ad50 | 170 | |
a3a8d943 | 171 | #endif |
d286ad50 | 172 | |
d58d1454 MD |
173 | static |
174 | int sys_perf_event_open(struct perf_event_attr *attr, | |
175 | pid_t pid, int cpu, int group_fd, | |
176 | unsigned long flags) | |
177 | { | |
178 | return syscall(SYS_perf_event_open, attr, pid, cpu, | |
179 | group_fd, flags); | |
180 | } | |
181 | ||
182 | static | |
b9389e6e | 183 | int open_perf_fd(struct perf_event_attr *attr) |
d58d1454 | 184 | { |
b9389e6e | 185 | int fd; |
d58d1454 MD |
186 | |
187 | fd = sys_perf_event_open(attr, 0, -1, -1, 0); | |
188 | if (fd < 0) | |
b9389e6e JD |
189 | return -1; |
190 | ||
191 | return fd; | |
192 | } | |
193 | ||
d286ad50 JD |
194 | static |
195 | void close_perf_fd(int fd) | |
196 | { | |
197 | int ret; | |
198 | ||
199 | if (fd < 0) | |
200 | return; | |
201 | ||
202 | ret = close(fd); | |
203 | if (ret) { | |
204 | perror("Error closing LTTng-UST perf memory mapping FD"); | |
205 | } | |
206 | } | |
207 | ||
b9389e6e JD |
208 | static |
209 | struct perf_event_mmap_page *setup_perf( | |
210 | struct lttng_perf_counter_thread_field *thread_field) | |
211 | { | |
212 | void *perf_addr; | |
d58d1454 MD |
213 | |
214 | perf_addr = mmap(NULL, sizeof(struct perf_event_mmap_page), | |
b9389e6e | 215 | PROT_READ, MAP_SHARED, thread_field->fd, 0); |
d58d1454 | 216 | if (perf_addr == MAP_FAILED) |
b9389e6e JD |
217 | perf_addr = NULL; |
218 | ||
a3a8d943 | 219 | if (!arch_perf_keep_fd(thread_field)) { |
b9389e6e JD |
220 | close_perf_fd(thread_field->fd); |
221 | thread_field->fd = -1; | |
6c2125af | 222 | } |
b9389e6e | 223 | |
d58d1454 MD |
224 | return perf_addr; |
225 | } | |
226 | ||
227 | static | |
228 | void unmap_perf_page(struct perf_event_mmap_page *pc) | |
229 | { | |
230 | int ret; | |
231 | ||
232 | if (!pc) | |
233 | return; | |
234 | ret = munmap(pc, sizeof(struct perf_event_mmap_page)); | |
235 | if (ret < 0) { | |
236 | PERROR("Error in munmap"); | |
237 | abort(); | |
238 | } | |
239 | } | |
240 | ||
241 | static | |
242 | struct lttng_perf_counter_thread *alloc_perf_counter_thread(void) | |
243 | { | |
244 | struct lttng_perf_counter_thread *perf_thread; | |
245 | sigset_t newmask, oldmask; | |
246 | int ret; | |
247 | ||
248 | ret = sigfillset(&newmask); | |
249 | if (ret) | |
250 | abort(); | |
251 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); | |
252 | if (ret) | |
253 | abort(); | |
254 | /* Check again with signals disabled */ | |
255 | perf_thread = pthread_getspecific(perf_counter_key); | |
256 | if (perf_thread) | |
257 | goto skip; | |
258 | perf_thread = zmalloc(sizeof(*perf_thread)); | |
259 | if (!perf_thread) | |
260 | abort(); | |
261 | CDS_INIT_LIST_HEAD(&perf_thread->rcu_field_list); | |
262 | ret = pthread_setspecific(perf_counter_key, perf_thread); | |
263 | if (ret) | |
264 | abort(); | |
265 | skip: | |
266 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
267 | if (ret) | |
268 | abort(); | |
269 | return perf_thread; | |
270 | } | |
271 | ||
272 | static | |
273 | struct lttng_perf_counter_thread_field * | |
274 | add_thread_field(struct lttng_perf_counter_field *perf_field, | |
275 | struct lttng_perf_counter_thread *perf_thread) | |
276 | { | |
277 | struct lttng_perf_counter_thread_field *thread_field; | |
278 | sigset_t newmask, oldmask; | |
279 | int ret; | |
280 | ||
281 | ret = sigfillset(&newmask); | |
282 | if (ret) | |
283 | abort(); | |
284 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); | |
285 | if (ret) | |
286 | abort(); | |
287 | /* Check again with signals disabled */ | |
288 | cds_list_for_each_entry_rcu(thread_field, &perf_thread->rcu_field_list, | |
289 | rcu_field_node) { | |
290 | if (thread_field->field == perf_field) | |
291 | goto skip; | |
292 | } | |
293 | thread_field = zmalloc(sizeof(*thread_field)); | |
294 | if (!thread_field) | |
295 | abort(); | |
296 | thread_field->field = perf_field; | |
b9389e6e JD |
297 | thread_field->fd = open_perf_fd(&perf_field->attr); |
298 | if (thread_field->fd >= 0) | |
299 | thread_field->pc = setup_perf(thread_field); | |
300 | /* | |
301 | * Note: thread_field->pc can be NULL if setup_perf() fails. | |
302 | * Also, thread_field->fd can be -1 if open_perf_fd() fails. | |
303 | */ | |
d58d1454 MD |
304 | ust_lock_nocheck(); |
305 | cds_list_add_rcu(&thread_field->rcu_field_node, | |
306 | &perf_thread->rcu_field_list); | |
307 | cds_list_add(&thread_field->thread_field_node, | |
308 | &perf_field->thread_field_list); | |
309 | ust_unlock(); | |
310 | skip: | |
311 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
312 | if (ret) | |
313 | abort(); | |
314 | return thread_field; | |
315 | } | |
316 | ||
317 | static | |
318 | struct lttng_perf_counter_thread_field * | |
319 | get_thread_field(struct lttng_perf_counter_field *field) | |
320 | { | |
321 | struct lttng_perf_counter_thread *perf_thread; | |
322 | struct lttng_perf_counter_thread_field *thread_field; | |
323 | ||
324 | perf_thread = pthread_getspecific(perf_counter_key); | |
325 | if (!perf_thread) | |
326 | perf_thread = alloc_perf_counter_thread(); | |
327 | cds_list_for_each_entry_rcu(thread_field, &perf_thread->rcu_field_list, | |
328 | rcu_field_node) { | |
329 | if (thread_field->field == field) | |
330 | return thread_field; | |
331 | } | |
332 | /* perf_counter_thread_field not found, need to add one */ | |
333 | return add_thread_field(field, perf_thread); | |
334 | } | |
335 | ||
336 | static | |
337 | uint64_t wrapper_perf_counter_read(struct lttng_ctx_field *field) | |
338 | { | |
339 | struct lttng_perf_counter_field *perf_field; | |
340 | struct lttng_perf_counter_thread_field *perf_thread_field; | |
341 | ||
342 | perf_field = field->u.perf_counter; | |
343 | perf_thread_field = get_thread_field(perf_field); | |
a3a8d943 | 344 | return arch_read_perf_counter(perf_thread_field); |
d58d1454 MD |
345 | } |
346 | ||
347 | static | |
348 | void perf_counter_record(struct lttng_ctx_field *field, | |
349 | struct lttng_ust_lib_ring_buffer_ctx *ctx, | |
350 | struct lttng_channel *chan) | |
351 | { | |
352 | uint64_t value; | |
353 | ||
354 | value = wrapper_perf_counter_read(field); | |
355 | lib_ring_buffer_align_ctx(ctx, lttng_alignof(value)); | |
356 | chan->ops->event_write(ctx, &value, sizeof(value)); | |
357 | } | |
358 | ||
359 | static | |
360 | void perf_counter_get_value(struct lttng_ctx_field *field, | |
53569322 | 361 | struct lttng_ctx_value *value) |
d58d1454 MD |
362 | { |
363 | uint64_t v; | |
364 | ||
365 | v = wrapper_perf_counter_read(field); | |
53569322 | 366 | value->u.s64 = v; |
d58d1454 MD |
367 | } |
368 | ||
369 | /* Called with UST lock held */ | |
370 | static | |
371 | void lttng_destroy_perf_thread_field( | |
372 | struct lttng_perf_counter_thread_field *thread_field) | |
373 | { | |
b9389e6e | 374 | close_perf_fd(thread_field->fd); |
d58d1454 MD |
375 | unmap_perf_page(thread_field->pc); |
376 | cds_list_del_rcu(&thread_field->rcu_field_node); | |
377 | cds_list_del(&thread_field->thread_field_node); | |
378 | free(thread_field); | |
379 | } | |
380 | ||
381 | static | |
382 | void lttng_destroy_perf_thread_key(void *_key) | |
383 | { | |
384 | struct lttng_perf_counter_thread *perf_thread = _key; | |
385 | struct lttng_perf_counter_thread_field *pos, *p; | |
386 | ||
387 | ust_lock_nocheck(); | |
388 | cds_list_for_each_entry_safe(pos, p, &perf_thread->rcu_field_list, | |
389 | rcu_field_node) | |
390 | lttng_destroy_perf_thread_field(pos); | |
391 | ust_unlock(); | |
392 | free(perf_thread); | |
393 | } | |
394 | ||
395 | /* Called with UST lock held */ | |
396 | static | |
397 | void lttng_destroy_perf_counter_field(struct lttng_ctx_field *field) | |
398 | { | |
399 | struct lttng_perf_counter_field *perf_field; | |
400 | struct lttng_perf_counter_thread_field *pos, *p; | |
401 | ||
402 | free((char *) field->event_field.name); | |
403 | perf_field = field->u.perf_counter; | |
404 | /* | |
405 | * This put is performed when no threads can concurrently | |
406 | * perform a "get" concurrently, thanks to urcu-bp grace | |
407 | * period. | |
408 | */ | |
409 | cds_list_for_each_entry_safe(pos, p, &perf_field->thread_field_list, | |
410 | thread_field_node) | |
411 | lttng_destroy_perf_thread_field(pos); | |
412 | free(perf_field); | |
413 | } | |
414 | ||
d286ad50 JD |
415 | #ifdef __ARM_ARCH_7A__ |
416 | ||
417 | static | |
418 | int perf_get_exclude_kernel(void) | |
419 | { | |
420 | return 0; | |
421 | } | |
422 | ||
423 | #else /* __ARM_ARCH_7A__ */ | |
424 | ||
425 | static | |
426 | int perf_get_exclude_kernel(void) | |
427 | { | |
428 | return 1; | |
429 | } | |
430 | ||
431 | #endif /* __ARM_ARCH_7A__ */ | |
432 | ||
d58d1454 MD |
433 | /* Called with UST lock held */ |
434 | int lttng_add_perf_counter_to_ctx(uint32_t type, | |
435 | uint64_t config, | |
436 | const char *name, | |
437 | struct lttng_ctx **ctx) | |
438 | { | |
439 | struct lttng_ctx_field *field; | |
440 | struct lttng_perf_counter_field *perf_field; | |
d58d1454 MD |
441 | char *name_alloc; |
442 | int ret; | |
443 | ||
444 | name_alloc = strdup(name); | |
445 | if (!name_alloc) { | |
446 | ret = -ENOMEM; | |
447 | goto name_alloc_error; | |
448 | } | |
449 | perf_field = zmalloc(sizeof(*perf_field)); | |
450 | if (!perf_field) { | |
451 | ret = -ENOMEM; | |
452 | goto perf_field_alloc_error; | |
453 | } | |
454 | field = lttng_append_context(ctx); | |
455 | if (!field) { | |
456 | ret = -ENOMEM; | |
457 | goto append_context_error; | |
458 | } | |
459 | if (lttng_find_context(*ctx, name_alloc)) { | |
460 | ret = -EEXIST; | |
461 | goto find_error; | |
462 | } | |
463 | ||
464 | field->destroy = lttng_destroy_perf_counter_field; | |
465 | ||
466 | field->event_field.name = name_alloc; | |
467 | field->event_field.type.atype = atype_integer; | |
468 | field->event_field.type.u.basic.integer.size = | |
469 | sizeof(uint64_t) * CHAR_BIT; | |
470 | field->event_field.type.u.basic.integer.alignment = | |
471 | lttng_alignof(uint64_t) * CHAR_BIT; | |
472 | field->event_field.type.u.basic.integer.signedness = | |
473 | lttng_is_signed_type(uint64_t); | |
474 | field->event_field.type.u.basic.integer.reverse_byte_order = 0; | |
475 | field->event_field.type.u.basic.integer.base = 10; | |
476 | field->event_field.type.u.basic.integer.encoding = lttng_encode_none; | |
477 | field->get_size = perf_counter_get_size; | |
478 | field->record = perf_counter_record; | |
479 | field->get_value = perf_counter_get_value; | |
480 | ||
481 | perf_field->attr.type = type; | |
482 | perf_field->attr.config = config; | |
d286ad50 | 483 | perf_field->attr.exclude_kernel = perf_get_exclude_kernel(); |
d58d1454 MD |
484 | CDS_INIT_LIST_HEAD(&perf_field->thread_field_list); |
485 | field->u.perf_counter = perf_field; | |
486 | ||
487 | /* Ensure that this perf counter can be used in this process. */ | |
b9389e6e JD |
488 | ret = open_perf_fd(&perf_field->attr); |
489 | if (ret < 0) { | |
d58d1454 MD |
490 | ret = -ENODEV; |
491 | goto setup_error; | |
492 | } | |
b9389e6e | 493 | close_perf_fd(ret); |
d58d1454 MD |
494 | |
495 | /* | |
496 | * Contexts can only be added before tracing is started, so we | |
497 | * don't have to synchronize against concurrent threads using | |
498 | * the field here. | |
499 | */ | |
500 | ||
b2cc986a | 501 | lttng_context_update(*ctx); |
d58d1454 MD |
502 | return 0; |
503 | ||
504 | setup_error: | |
505 | find_error: | |
506 | lttng_remove_context_field(ctx, field); | |
507 | append_context_error: | |
508 | free(perf_field); | |
509 | perf_field_alloc_error: | |
510 | free(name_alloc); | |
511 | name_alloc_error: | |
512 | return ret; | |
513 | } | |
514 | ||
515 | int lttng_perf_counter_init(void) | |
516 | { | |
517 | int ret; | |
518 | ||
519 | ret = pthread_key_create(&perf_counter_key, | |
520 | lttng_destroy_perf_thread_key); | |
521 | if (ret) | |
522 | ret = -ret; | |
523 | return ret; | |
524 | } | |
525 | ||
526 | void lttng_perf_counter_exit(void) | |
527 | { | |
528 | int ret; | |
529 | ||
530 | ret = pthread_key_delete(perf_counter_key); | |
531 | if (ret) { | |
532 | errno = ret; | |
533 | PERROR("Error in pthread_key_delete"); | |
534 | } | |
535 | } |