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 | ||
3fbec7dc | 23 | #define _LGPL_SOURCE |
d58d1454 MD |
24 | #include <sys/types.h> |
25 | #include <unistd.h> | |
26 | #include <string.h> | |
27 | #include <stdlib.h> | |
28 | #include <stdio.h> | |
649fb6b3 | 29 | #include <stdbool.h> |
d58d1454 MD |
30 | #include <sys/mman.h> |
31 | #include <sys/syscall.h> | |
d58d1454 MD |
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> | |
77d7fa98 | 42 | #include "perf_event.h" |
d58d1454 MD |
43 | #include "lttng-tracer-core.h" |
44 | ||
45 | /* | |
46 | * We use a global perf counter key and iterate on per-thread RCU lists | |
47 | * of fields in the fast path, even though this is not strictly speaking | |
48 | * what would provide the best fast-path complexity, to ensure teardown | |
49 | * of sessions vs thread exit is handled racelessly. | |
50 | * | |
51 | * Updates and traversals of thread_list are protected by UST lock. | |
52 | * Updates to rcu_field_list are protected by UST lock. | |
53 | */ | |
54 | ||
55 | struct lttng_perf_counter_thread_field { | |
56 | struct lttng_perf_counter_field *field; /* Back reference */ | |
57 | struct perf_event_mmap_page *pc; | |
58 | struct cds_list_head thread_field_node; /* Per-field list of thread fields (node) */ | |
59 | struct cds_list_head rcu_field_node; /* RCU per-thread list of fields (node) */ | |
b9389e6e | 60 | int fd; /* Perf FD */ |
d58d1454 MD |
61 | }; |
62 | ||
63 | struct lttng_perf_counter_thread { | |
64 | struct cds_list_head rcu_field_list; /* RCU per-thread list of fields */ | |
65 | }; | |
66 | ||
67 | struct lttng_perf_counter_field { | |
68 | struct perf_event_attr attr; | |
69 | struct cds_list_head thread_field_list; /* Per-field list of thread fields */ | |
70 | }; | |
71 | ||
72 | static pthread_key_t perf_counter_key; | |
73 | ||
74 | static | |
53569322 | 75 | size_t perf_counter_get_size(struct lttng_ctx_field *field, size_t offset) |
d58d1454 MD |
76 | { |
77 | size_t size = 0; | |
78 | ||
79 | size += lib_ring_buffer_align(offset, lttng_alignof(uint64_t)); | |
80 | size += sizeof(uint64_t); | |
81 | return size; | |
82 | } | |
83 | ||
a3a8d943 MD |
84 | static |
85 | uint64_t read_perf_counter_syscall( | |
86 | struct lttng_perf_counter_thread_field *thread_field) | |
87 | { | |
88 | uint64_t count; | |
89 | ||
90 | if (caa_unlikely(thread_field->fd < 0)) | |
91 | return 0; | |
92 | ||
93 | if (caa_unlikely(read(thread_field->fd, &count, sizeof(count)) | |
94 | < sizeof(count))) | |
95 | return 0; | |
96 | ||
97 | return count; | |
98 | } | |
99 | ||
d58d1454 MD |
100 | #if defined(__x86_64__) || defined(__i386__) |
101 | ||
102 | static | |
103 | uint64_t rdpmc(unsigned int counter) | |
104 | { | |
105 | unsigned int low, high; | |
106 | ||
107 | asm volatile("rdpmc" : "=a" (low), "=d" (high) : "c" (counter)); | |
108 | ||
109 | return low | ((uint64_t) high) << 32; | |
110 | } | |
111 | ||
77d7fa98 MD |
112 | static |
113 | bool has_rdpmc(struct perf_event_mmap_page *pc) | |
114 | { | |
115 | if (caa_unlikely(!pc->cap_bit0_is_deprecated)) | |
116 | return false; | |
117 | /* Since Linux kernel 3.12. */ | |
118 | return pc->cap_user_rdpmc; | |
119 | } | |
120 | ||
d58d1454 | 121 | static |
a3a8d943 | 122 | uint64_t arch_read_perf_counter( |
d286ad50 | 123 | struct lttng_perf_counter_thread_field *thread_field) |
d58d1454 MD |
124 | { |
125 | uint32_t seq, idx; | |
126 | uint64_t count; | |
d286ad50 | 127 | struct perf_event_mmap_page *pc = thread_field->pc; |
d58d1454 MD |
128 | |
129 | if (caa_unlikely(!pc)) | |
130 | return 0; | |
131 | ||
132 | do { | |
133 | seq = CMM_LOAD_SHARED(pc->lock); | |
134 | cmm_barrier(); | |
135 | ||
136 | idx = pc->index; | |
77d7fa98 | 137 | if (caa_likely(has_rdpmc(pc) && idx)) { |
4f58f54f MD |
138 | int64_t pmcval; |
139 | ||
140 | pmcval = rdpmc(idx - 1); | |
141 | /* Sign-extend the pmc register result. */ | |
142 | pmcval <<= 64 - pc->pmc_width; | |
143 | pmcval >>= 64 - pc->pmc_width; | |
144 | count = pc->offset + pmcval; | |
145 | } else { | |
a3a8d943 MD |
146 | /* Fall-back on system call if rdpmc cannot be used. */ |
147 | return read_perf_counter_syscall(thread_field); | |
4f58f54f | 148 | } |
d58d1454 MD |
149 | cmm_barrier(); |
150 | } while (CMM_LOAD_SHARED(pc->lock) != seq); | |
151 | ||
152 | return count; | |
153 | } | |
154 | ||
d286ad50 | 155 | static |
a3a8d943 | 156 | int arch_perf_keep_fd(struct lttng_perf_counter_thread_field *thread_field) |
d286ad50 | 157 | { |
a3a8d943 | 158 | struct perf_event_mmap_page *pc = thread_field->pc; |
d286ad50 | 159 | |
a3a8d943 | 160 | if (!pc) |
d286ad50 | 161 | return 0; |
77d7fa98 | 162 | return !has_rdpmc(pc); |
a3a8d943 | 163 | } |
d286ad50 | 164 | |
a3a8d943 | 165 | #else |
d286ad50 | 166 | |
a3a8d943 MD |
167 | /* Generic (slow) implementation using a read system call. */ |
168 | static | |
169 | uint64_t arch_read_perf_counter( | |
170 | struct lttng_perf_counter_thread_field *thread_field) | |
171 | { | |
172 | return read_perf_counter_syscall(thread_field); | |
d286ad50 JD |
173 | } |
174 | ||
a3a8d943 MD |
175 | static |
176 | int arch_perf_keep_fd(struct lttng_perf_counter_thread_field *thread_field) | |
177 | { | |
178 | return 1; | |
179 | } | |
d286ad50 | 180 | |
a3a8d943 | 181 | #endif |
d286ad50 | 182 | |
d58d1454 MD |
183 | static |
184 | int sys_perf_event_open(struct perf_event_attr *attr, | |
185 | pid_t pid, int cpu, int group_fd, | |
186 | unsigned long flags) | |
187 | { | |
188 | return syscall(SYS_perf_event_open, attr, pid, cpu, | |
189 | group_fd, flags); | |
190 | } | |
191 | ||
192 | static | |
b9389e6e | 193 | int open_perf_fd(struct perf_event_attr *attr) |
d58d1454 | 194 | { |
b9389e6e | 195 | int fd; |
d58d1454 MD |
196 | |
197 | fd = sys_perf_event_open(attr, 0, -1, -1, 0); | |
198 | if (fd < 0) | |
b9389e6e JD |
199 | return -1; |
200 | ||
201 | return fd; | |
202 | } | |
203 | ||
d286ad50 JD |
204 | static |
205 | void close_perf_fd(int fd) | |
206 | { | |
207 | int ret; | |
208 | ||
209 | if (fd < 0) | |
210 | return; | |
211 | ||
212 | ret = close(fd); | |
213 | if (ret) { | |
214 | perror("Error closing LTTng-UST perf memory mapping FD"); | |
215 | } | |
216 | } | |
217 | ||
77d7fa98 | 218 | static void setup_perf(struct lttng_perf_counter_thread_field *thread_field) |
b9389e6e JD |
219 | { |
220 | void *perf_addr; | |
d58d1454 MD |
221 | |
222 | perf_addr = mmap(NULL, sizeof(struct perf_event_mmap_page), | |
b9389e6e | 223 | PROT_READ, MAP_SHARED, thread_field->fd, 0); |
d58d1454 | 224 | if (perf_addr == MAP_FAILED) |
b9389e6e | 225 | perf_addr = NULL; |
77d7fa98 | 226 | thread_field->pc = perf_addr; |
b9389e6e | 227 | |
a3a8d943 | 228 | if (!arch_perf_keep_fd(thread_field)) { |
b9389e6e JD |
229 | close_perf_fd(thread_field->fd); |
230 | thread_field->fd = -1; | |
6c2125af | 231 | } |
d58d1454 MD |
232 | } |
233 | ||
234 | static | |
235 | void unmap_perf_page(struct perf_event_mmap_page *pc) | |
236 | { | |
237 | int ret; | |
238 | ||
239 | if (!pc) | |
240 | return; | |
241 | ret = munmap(pc, sizeof(struct perf_event_mmap_page)); | |
242 | if (ret < 0) { | |
243 | PERROR("Error in munmap"); | |
244 | abort(); | |
245 | } | |
246 | } | |
247 | ||
248 | static | |
249 | struct lttng_perf_counter_thread *alloc_perf_counter_thread(void) | |
250 | { | |
251 | struct lttng_perf_counter_thread *perf_thread; | |
252 | sigset_t newmask, oldmask; | |
253 | int ret; | |
254 | ||
255 | ret = sigfillset(&newmask); | |
256 | if (ret) | |
257 | abort(); | |
258 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); | |
259 | if (ret) | |
260 | abort(); | |
261 | /* Check again with signals disabled */ | |
262 | perf_thread = pthread_getspecific(perf_counter_key); | |
263 | if (perf_thread) | |
264 | goto skip; | |
265 | perf_thread = zmalloc(sizeof(*perf_thread)); | |
266 | if (!perf_thread) | |
267 | abort(); | |
268 | CDS_INIT_LIST_HEAD(&perf_thread->rcu_field_list); | |
269 | ret = pthread_setspecific(perf_counter_key, perf_thread); | |
270 | if (ret) | |
271 | abort(); | |
272 | skip: | |
273 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
274 | if (ret) | |
275 | abort(); | |
276 | return perf_thread; | |
277 | } | |
278 | ||
279 | static | |
280 | struct lttng_perf_counter_thread_field * | |
281 | add_thread_field(struct lttng_perf_counter_field *perf_field, | |
282 | struct lttng_perf_counter_thread *perf_thread) | |
283 | { | |
284 | struct lttng_perf_counter_thread_field *thread_field; | |
285 | sigset_t newmask, oldmask; | |
286 | int ret; | |
287 | ||
288 | ret = sigfillset(&newmask); | |
289 | if (ret) | |
290 | abort(); | |
291 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); | |
292 | if (ret) | |
293 | abort(); | |
294 | /* Check again with signals disabled */ | |
295 | cds_list_for_each_entry_rcu(thread_field, &perf_thread->rcu_field_list, | |
296 | rcu_field_node) { | |
297 | if (thread_field->field == perf_field) | |
298 | goto skip; | |
299 | } | |
300 | thread_field = zmalloc(sizeof(*thread_field)); | |
301 | if (!thread_field) | |
302 | abort(); | |
303 | thread_field->field = perf_field; | |
b9389e6e JD |
304 | thread_field->fd = open_perf_fd(&perf_field->attr); |
305 | if (thread_field->fd >= 0) | |
77d7fa98 | 306 | setup_perf(thread_field); |
b9389e6e JD |
307 | /* |
308 | * Note: thread_field->pc can be NULL if setup_perf() fails. | |
309 | * Also, thread_field->fd can be -1 if open_perf_fd() fails. | |
310 | */ | |
d58d1454 MD |
311 | ust_lock_nocheck(); |
312 | cds_list_add_rcu(&thread_field->rcu_field_node, | |
313 | &perf_thread->rcu_field_list); | |
314 | cds_list_add(&thread_field->thread_field_node, | |
315 | &perf_field->thread_field_list); | |
316 | ust_unlock(); | |
317 | skip: | |
318 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
319 | if (ret) | |
320 | abort(); | |
321 | return thread_field; | |
322 | } | |
323 | ||
324 | static | |
325 | struct lttng_perf_counter_thread_field * | |
326 | get_thread_field(struct lttng_perf_counter_field *field) | |
327 | { | |
328 | struct lttng_perf_counter_thread *perf_thread; | |
329 | struct lttng_perf_counter_thread_field *thread_field; | |
330 | ||
331 | perf_thread = pthread_getspecific(perf_counter_key); | |
332 | if (!perf_thread) | |
333 | perf_thread = alloc_perf_counter_thread(); | |
334 | cds_list_for_each_entry_rcu(thread_field, &perf_thread->rcu_field_list, | |
335 | rcu_field_node) { | |
336 | if (thread_field->field == field) | |
337 | return thread_field; | |
338 | } | |
339 | /* perf_counter_thread_field not found, need to add one */ | |
340 | return add_thread_field(field, perf_thread); | |
341 | } | |
342 | ||
343 | static | |
344 | uint64_t wrapper_perf_counter_read(struct lttng_ctx_field *field) | |
345 | { | |
346 | struct lttng_perf_counter_field *perf_field; | |
347 | struct lttng_perf_counter_thread_field *perf_thread_field; | |
348 | ||
349 | perf_field = field->u.perf_counter; | |
350 | perf_thread_field = get_thread_field(perf_field); | |
a3a8d943 | 351 | return arch_read_perf_counter(perf_thread_field); |
d58d1454 MD |
352 | } |
353 | ||
354 | static | |
355 | void perf_counter_record(struct lttng_ctx_field *field, | |
356 | struct lttng_ust_lib_ring_buffer_ctx *ctx, | |
357 | struct lttng_channel *chan) | |
358 | { | |
359 | uint64_t value; | |
360 | ||
361 | value = wrapper_perf_counter_read(field); | |
362 | lib_ring_buffer_align_ctx(ctx, lttng_alignof(value)); | |
363 | chan->ops->event_write(ctx, &value, sizeof(value)); | |
364 | } | |
365 | ||
366 | static | |
367 | void perf_counter_get_value(struct lttng_ctx_field *field, | |
53569322 | 368 | struct lttng_ctx_value *value) |
d58d1454 MD |
369 | { |
370 | uint64_t v; | |
371 | ||
372 | v = wrapper_perf_counter_read(field); | |
53569322 | 373 | value->u.s64 = v; |
d58d1454 MD |
374 | } |
375 | ||
376 | /* Called with UST lock held */ | |
377 | static | |
378 | void lttng_destroy_perf_thread_field( | |
379 | struct lttng_perf_counter_thread_field *thread_field) | |
380 | { | |
b9389e6e | 381 | close_perf_fd(thread_field->fd); |
d58d1454 MD |
382 | unmap_perf_page(thread_field->pc); |
383 | cds_list_del_rcu(&thread_field->rcu_field_node); | |
384 | cds_list_del(&thread_field->thread_field_node); | |
385 | free(thread_field); | |
386 | } | |
387 | ||
388 | static | |
389 | void lttng_destroy_perf_thread_key(void *_key) | |
390 | { | |
391 | struct lttng_perf_counter_thread *perf_thread = _key; | |
392 | struct lttng_perf_counter_thread_field *pos, *p; | |
393 | ||
394 | ust_lock_nocheck(); | |
395 | cds_list_for_each_entry_safe(pos, p, &perf_thread->rcu_field_list, | |
396 | rcu_field_node) | |
397 | lttng_destroy_perf_thread_field(pos); | |
398 | ust_unlock(); | |
399 | free(perf_thread); | |
400 | } | |
401 | ||
402 | /* Called with UST lock held */ | |
403 | static | |
404 | void lttng_destroy_perf_counter_field(struct lttng_ctx_field *field) | |
405 | { | |
406 | struct lttng_perf_counter_field *perf_field; | |
407 | struct lttng_perf_counter_thread_field *pos, *p; | |
408 | ||
409 | free((char *) field->event_field.name); | |
410 | perf_field = field->u.perf_counter; | |
411 | /* | |
412 | * This put is performed when no threads can concurrently | |
413 | * perform a "get" concurrently, thanks to urcu-bp grace | |
414 | * period. | |
415 | */ | |
416 | cds_list_for_each_entry_safe(pos, p, &perf_field->thread_field_list, | |
417 | thread_field_node) | |
418 | lttng_destroy_perf_thread_field(pos); | |
419 | free(perf_field); | |
420 | } | |
421 | ||
d286ad50 JD |
422 | #ifdef __ARM_ARCH_7A__ |
423 | ||
424 | static | |
425 | int perf_get_exclude_kernel(void) | |
426 | { | |
427 | return 0; | |
428 | } | |
429 | ||
430 | #else /* __ARM_ARCH_7A__ */ | |
431 | ||
432 | static | |
433 | int perf_get_exclude_kernel(void) | |
434 | { | |
435 | return 1; | |
436 | } | |
437 | ||
438 | #endif /* __ARM_ARCH_7A__ */ | |
439 | ||
d58d1454 MD |
440 | /* Called with UST lock held */ |
441 | int lttng_add_perf_counter_to_ctx(uint32_t type, | |
442 | uint64_t config, | |
443 | const char *name, | |
444 | struct lttng_ctx **ctx) | |
445 | { | |
446 | struct lttng_ctx_field *field; | |
447 | struct lttng_perf_counter_field *perf_field; | |
d58d1454 MD |
448 | char *name_alloc; |
449 | int ret; | |
450 | ||
451 | name_alloc = strdup(name); | |
452 | if (!name_alloc) { | |
453 | ret = -ENOMEM; | |
454 | goto name_alloc_error; | |
455 | } | |
456 | perf_field = zmalloc(sizeof(*perf_field)); | |
457 | if (!perf_field) { | |
458 | ret = -ENOMEM; | |
459 | goto perf_field_alloc_error; | |
460 | } | |
461 | field = lttng_append_context(ctx); | |
462 | if (!field) { | |
463 | ret = -ENOMEM; | |
464 | goto append_context_error; | |
465 | } | |
466 | if (lttng_find_context(*ctx, name_alloc)) { | |
467 | ret = -EEXIST; | |
468 | goto find_error; | |
469 | } | |
470 | ||
471 | field->destroy = lttng_destroy_perf_counter_field; | |
472 | ||
473 | field->event_field.name = name_alloc; | |
474 | field->event_field.type.atype = atype_integer; | |
475 | field->event_field.type.u.basic.integer.size = | |
476 | sizeof(uint64_t) * CHAR_BIT; | |
477 | field->event_field.type.u.basic.integer.alignment = | |
478 | lttng_alignof(uint64_t) * CHAR_BIT; | |
479 | field->event_field.type.u.basic.integer.signedness = | |
480 | lttng_is_signed_type(uint64_t); | |
481 | field->event_field.type.u.basic.integer.reverse_byte_order = 0; | |
482 | field->event_field.type.u.basic.integer.base = 10; | |
483 | field->event_field.type.u.basic.integer.encoding = lttng_encode_none; | |
484 | field->get_size = perf_counter_get_size; | |
485 | field->record = perf_counter_record; | |
486 | field->get_value = perf_counter_get_value; | |
487 | ||
488 | perf_field->attr.type = type; | |
489 | perf_field->attr.config = config; | |
d286ad50 | 490 | perf_field->attr.exclude_kernel = perf_get_exclude_kernel(); |
d58d1454 MD |
491 | CDS_INIT_LIST_HEAD(&perf_field->thread_field_list); |
492 | field->u.perf_counter = perf_field; | |
493 | ||
494 | /* Ensure that this perf counter can be used in this process. */ | |
b9389e6e JD |
495 | ret = open_perf_fd(&perf_field->attr); |
496 | if (ret < 0) { | |
d58d1454 MD |
497 | ret = -ENODEV; |
498 | goto setup_error; | |
499 | } | |
b9389e6e | 500 | close_perf_fd(ret); |
d58d1454 MD |
501 | |
502 | /* | |
503 | * Contexts can only be added before tracing is started, so we | |
504 | * don't have to synchronize against concurrent threads using | |
505 | * the field here. | |
506 | */ | |
507 | ||
b2cc986a | 508 | lttng_context_update(*ctx); |
d58d1454 MD |
509 | return 0; |
510 | ||
511 | setup_error: | |
512 | find_error: | |
513 | lttng_remove_context_field(ctx, field); | |
514 | append_context_error: | |
515 | free(perf_field); | |
516 | perf_field_alloc_error: | |
517 | free(name_alloc); | |
518 | name_alloc_error: | |
519 | return ret; | |
520 | } | |
521 | ||
522 | int lttng_perf_counter_init(void) | |
523 | { | |
524 | int ret; | |
525 | ||
526 | ret = pthread_key_create(&perf_counter_key, | |
527 | lttng_destroy_perf_thread_key); | |
528 | if (ret) | |
529 | ret = -ret; | |
530 | return ret; | |
531 | } | |
532 | ||
533 | void lttng_perf_counter_exit(void) | |
534 | { | |
535 | int ret; | |
536 | ||
537 | ret = pthread_key_delete(perf_counter_key); | |
538 | if (ret) { | |
539 | errno = ret; | |
540 | PERROR("Error in pthread_key_delete"); | |
541 | } | |
542 | } |