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> | |
e336bbe5 | 28 | #include <stdbool.h> |
d58d1454 MD |
29 | #include <sys/mman.h> |
30 | #include <sys/syscall.h> | |
d58d1454 MD |
31 | #include <lttng/ust-events.h> |
32 | #include <lttng/ust-tracer.h> | |
33 | #include <lttng/ringbuffer-config.h> | |
34 | #include <urcu/system.h> | |
35 | #include <urcu/arch.h> | |
36 | #include <urcu/rculist.h> | |
37 | #include <helper.h> | |
38 | #include <urcu/ref.h> | |
39 | #include <usterr-signal-safe.h> | |
40 | #include <signal.h> | |
e336bbe5 | 41 | #include "perf_event.h" |
d58d1454 MD |
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) */ | |
e336bbe5 | 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 | ||
e336bbe5 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 | ||
e336bbe5 MD |
111 | static |
112 | bool has_rdpmc(struct perf_event_mmap_page *pc) | |
113 | { | |
114 | if (caa_unlikely(!pc->cap_bit0_is_deprecated)) | |
115 | return false; | |
116 | /* Since Linux kernel 3.12. */ | |
117 | return pc->cap_user_rdpmc; | |
118 | } | |
d58d1454 MD |
119 | |
120 | static | |
e336bbe5 MD |
121 | uint64_t arch_read_perf_counter( |
122 | struct lttng_perf_counter_thread_field *thread_field) | |
d58d1454 MD |
123 | { |
124 | uint32_t seq, idx; | |
125 | uint64_t count; | |
e336bbe5 | 126 | struct perf_event_mmap_page *pc = thread_field->pc; |
d58d1454 MD |
127 | |
128 | if (caa_unlikely(!pc)) | |
129 | return 0; | |
130 | ||
131 | do { | |
132 | seq = CMM_LOAD_SHARED(pc->lock); | |
133 | cmm_barrier(); | |
134 | ||
135 | idx = pc->index; | |
e336bbe5 | 136 | if (caa_likely(has_rdpmc(pc) && idx)) { |
3046b99e MD |
137 | int64_t pmcval; |
138 | ||
139 | pmcval = rdpmc(idx - 1); | |
140 | /* Sign-extend the pmc register result. */ | |
141 | pmcval <<= 64 - pc->pmc_width; | |
142 | pmcval >>= 64 - pc->pmc_width; | |
143 | count = pc->offset + pmcval; | |
144 | } else { | |
e336bbe5 MD |
145 | /* Fall-back on system call if rdpmc cannot be used. */ |
146 | return read_perf_counter_syscall(thread_field); | |
3046b99e | 147 | } |
d58d1454 MD |
148 | cmm_barrier(); |
149 | } while (CMM_LOAD_SHARED(pc->lock) != seq); | |
150 | ||
151 | return count; | |
152 | } | |
153 | ||
e336bbe5 MD |
154 | static |
155 | int arch_perf_keep_fd(struct lttng_perf_counter_thread_field *thread_field) | |
156 | { | |
157 | struct perf_event_mmap_page *pc = thread_field->pc; | |
158 | ||
159 | if (!pc) | |
160 | return 0; | |
161 | return !has_rdpmc(pc); | |
162 | } | |
163 | ||
164 | #endif | |
165 | ||
d58d1454 MD |
166 | static |
167 | int sys_perf_event_open(struct perf_event_attr *attr, | |
168 | pid_t pid, int cpu, int group_fd, | |
169 | unsigned long flags) | |
170 | { | |
171 | return syscall(SYS_perf_event_open, attr, pid, cpu, | |
172 | group_fd, flags); | |
173 | } | |
174 | ||
175 | static | |
e336bbe5 | 176 | int open_perf_fd(struct perf_event_attr *attr) |
d58d1454 | 177 | { |
e336bbe5 | 178 | int fd; |
d58d1454 MD |
179 | |
180 | fd = sys_perf_event_open(attr, 0, -1, -1, 0); | |
181 | if (fd < 0) | |
e336bbe5 MD |
182 | return -1; |
183 | ||
184 | return fd; | |
185 | } | |
186 | ||
187 | static | |
188 | void close_perf_fd(int fd) | |
189 | { | |
190 | int ret; | |
191 | ||
192 | if (fd < 0) | |
193 | return; | |
d58d1454 | 194 | |
6c2125af MD |
195 | ret = close(fd); |
196 | if (ret) { | |
197 | perror("Error closing LTTng-UST perf memory mapping FD"); | |
198 | } | |
e336bbe5 MD |
199 | } |
200 | ||
201 | static void setup_perf(struct lttng_perf_counter_thread_field *thread_field) | |
202 | { | |
203 | void *perf_addr; | |
204 | ||
205 | perf_addr = mmap(NULL, sizeof(struct perf_event_mmap_page), | |
206 | PROT_READ, MAP_SHARED, thread_field->fd, 0); | |
207 | if (perf_addr == MAP_FAILED) | |
208 | perf_addr = NULL; | |
209 | thread_field->pc = perf_addr; | |
210 | ||
211 | if (!arch_perf_keep_fd(thread_field)) { | |
212 | close_perf_fd(thread_field->fd); | |
213 | thread_field->fd = -1; | |
214 | } | |
d58d1454 MD |
215 | } |
216 | ||
217 | static | |
218 | void unmap_perf_page(struct perf_event_mmap_page *pc) | |
219 | { | |
220 | int ret; | |
221 | ||
222 | if (!pc) | |
223 | return; | |
224 | ret = munmap(pc, sizeof(struct perf_event_mmap_page)); | |
225 | if (ret < 0) { | |
226 | PERROR("Error in munmap"); | |
227 | abort(); | |
228 | } | |
229 | } | |
230 | ||
231 | static | |
232 | struct lttng_perf_counter_thread *alloc_perf_counter_thread(void) | |
233 | { | |
234 | struct lttng_perf_counter_thread *perf_thread; | |
235 | sigset_t newmask, oldmask; | |
236 | int ret; | |
237 | ||
238 | ret = sigfillset(&newmask); | |
239 | if (ret) | |
240 | abort(); | |
241 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); | |
242 | if (ret) | |
243 | abort(); | |
244 | /* Check again with signals disabled */ | |
245 | perf_thread = pthread_getspecific(perf_counter_key); | |
246 | if (perf_thread) | |
247 | goto skip; | |
248 | perf_thread = zmalloc(sizeof(*perf_thread)); | |
249 | if (!perf_thread) | |
250 | abort(); | |
251 | CDS_INIT_LIST_HEAD(&perf_thread->rcu_field_list); | |
252 | ret = pthread_setspecific(perf_counter_key, perf_thread); | |
253 | if (ret) | |
254 | abort(); | |
255 | skip: | |
256 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
257 | if (ret) | |
258 | abort(); | |
259 | return perf_thread; | |
260 | } | |
261 | ||
262 | static | |
263 | struct lttng_perf_counter_thread_field * | |
264 | add_thread_field(struct lttng_perf_counter_field *perf_field, | |
265 | struct lttng_perf_counter_thread *perf_thread) | |
266 | { | |
267 | struct lttng_perf_counter_thread_field *thread_field; | |
268 | sigset_t newmask, oldmask; | |
269 | int ret; | |
270 | ||
271 | ret = sigfillset(&newmask); | |
272 | if (ret) | |
273 | abort(); | |
274 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); | |
275 | if (ret) | |
276 | abort(); | |
277 | /* Check again with signals disabled */ | |
278 | cds_list_for_each_entry_rcu(thread_field, &perf_thread->rcu_field_list, | |
279 | rcu_field_node) { | |
280 | if (thread_field->field == perf_field) | |
281 | goto skip; | |
282 | } | |
283 | thread_field = zmalloc(sizeof(*thread_field)); | |
284 | if (!thread_field) | |
285 | abort(); | |
286 | thread_field->field = perf_field; | |
e336bbe5 MD |
287 | thread_field->fd = open_perf_fd(&perf_field->attr); |
288 | if (thread_field->fd >= 0) | |
289 | setup_perf(thread_field); | |
290 | /* | |
291 | * Note: thread_field->pc can be NULL if setup_perf() fails. | |
292 | * Also, thread_field->fd can be -1 if open_perf_fd() fails. | |
293 | */ | |
d58d1454 MD |
294 | ust_lock_nocheck(); |
295 | cds_list_add_rcu(&thread_field->rcu_field_node, | |
296 | &perf_thread->rcu_field_list); | |
297 | cds_list_add(&thread_field->thread_field_node, | |
298 | &perf_field->thread_field_list); | |
299 | ust_unlock(); | |
300 | skip: | |
301 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
302 | if (ret) | |
303 | abort(); | |
304 | return thread_field; | |
305 | } | |
306 | ||
307 | static | |
308 | struct lttng_perf_counter_thread_field * | |
309 | get_thread_field(struct lttng_perf_counter_field *field) | |
310 | { | |
311 | struct lttng_perf_counter_thread *perf_thread; | |
312 | struct lttng_perf_counter_thread_field *thread_field; | |
313 | ||
314 | perf_thread = pthread_getspecific(perf_counter_key); | |
315 | if (!perf_thread) | |
316 | perf_thread = alloc_perf_counter_thread(); | |
317 | cds_list_for_each_entry_rcu(thread_field, &perf_thread->rcu_field_list, | |
318 | rcu_field_node) { | |
319 | if (thread_field->field == field) | |
320 | return thread_field; | |
321 | } | |
322 | /* perf_counter_thread_field not found, need to add one */ | |
323 | return add_thread_field(field, perf_thread); | |
324 | } | |
325 | ||
326 | static | |
327 | uint64_t wrapper_perf_counter_read(struct lttng_ctx_field *field) | |
328 | { | |
329 | struct lttng_perf_counter_field *perf_field; | |
330 | struct lttng_perf_counter_thread_field *perf_thread_field; | |
331 | ||
332 | perf_field = field->u.perf_counter; | |
333 | perf_thread_field = get_thread_field(perf_field); | |
e336bbe5 | 334 | return arch_read_perf_counter(perf_thread_field); |
d58d1454 MD |
335 | } |
336 | ||
337 | static | |
338 | void perf_counter_record(struct lttng_ctx_field *field, | |
339 | struct lttng_ust_lib_ring_buffer_ctx *ctx, | |
340 | struct lttng_channel *chan) | |
341 | { | |
342 | uint64_t value; | |
343 | ||
344 | value = wrapper_perf_counter_read(field); | |
345 | lib_ring_buffer_align_ctx(ctx, lttng_alignof(value)); | |
346 | chan->ops->event_write(ctx, &value, sizeof(value)); | |
347 | } | |
348 | ||
349 | static | |
350 | void perf_counter_get_value(struct lttng_ctx_field *field, | |
53569322 | 351 | struct lttng_ctx_value *value) |
d58d1454 MD |
352 | { |
353 | uint64_t v; | |
354 | ||
355 | v = wrapper_perf_counter_read(field); | |
53569322 | 356 | value->u.s64 = v; |
d58d1454 MD |
357 | } |
358 | ||
359 | /* Called with UST lock held */ | |
360 | static | |
361 | void lttng_destroy_perf_thread_field( | |
362 | struct lttng_perf_counter_thread_field *thread_field) | |
363 | { | |
e336bbe5 | 364 | close_perf_fd(thread_field->fd); |
d58d1454 MD |
365 | unmap_perf_page(thread_field->pc); |
366 | cds_list_del_rcu(&thread_field->rcu_field_node); | |
367 | cds_list_del(&thread_field->thread_field_node); | |
368 | free(thread_field); | |
369 | } | |
370 | ||
371 | static | |
372 | void lttng_destroy_perf_thread_key(void *_key) | |
373 | { | |
374 | struct lttng_perf_counter_thread *perf_thread = _key; | |
375 | struct lttng_perf_counter_thread_field *pos, *p; | |
376 | ||
377 | ust_lock_nocheck(); | |
378 | cds_list_for_each_entry_safe(pos, p, &perf_thread->rcu_field_list, | |
379 | rcu_field_node) | |
380 | lttng_destroy_perf_thread_field(pos); | |
381 | ust_unlock(); | |
382 | free(perf_thread); | |
383 | } | |
384 | ||
385 | /* Called with UST lock held */ | |
386 | static | |
387 | void lttng_destroy_perf_counter_field(struct lttng_ctx_field *field) | |
388 | { | |
389 | struct lttng_perf_counter_field *perf_field; | |
390 | struct lttng_perf_counter_thread_field *pos, *p; | |
391 | ||
392 | free((char *) field->event_field.name); | |
393 | perf_field = field->u.perf_counter; | |
394 | /* | |
395 | * This put is performed when no threads can concurrently | |
396 | * perform a "get" concurrently, thanks to urcu-bp grace | |
397 | * period. | |
398 | */ | |
399 | cds_list_for_each_entry_safe(pos, p, &perf_field->thread_field_list, | |
400 | thread_field_node) | |
401 | lttng_destroy_perf_thread_field(pos); | |
402 | free(perf_field); | |
403 | } | |
404 | ||
405 | /* Called with UST lock held */ | |
406 | int lttng_add_perf_counter_to_ctx(uint32_t type, | |
407 | uint64_t config, | |
408 | const char *name, | |
409 | struct lttng_ctx **ctx) | |
410 | { | |
411 | struct lttng_ctx_field *field; | |
412 | struct lttng_perf_counter_field *perf_field; | |
d58d1454 MD |
413 | char *name_alloc; |
414 | int ret; | |
415 | ||
416 | name_alloc = strdup(name); | |
417 | if (!name_alloc) { | |
418 | ret = -ENOMEM; | |
419 | goto name_alloc_error; | |
420 | } | |
421 | perf_field = zmalloc(sizeof(*perf_field)); | |
422 | if (!perf_field) { | |
423 | ret = -ENOMEM; | |
424 | goto perf_field_alloc_error; | |
425 | } | |
426 | field = lttng_append_context(ctx); | |
427 | if (!field) { | |
428 | ret = -ENOMEM; | |
429 | goto append_context_error; | |
430 | } | |
431 | if (lttng_find_context(*ctx, name_alloc)) { | |
432 | ret = -EEXIST; | |
433 | goto find_error; | |
434 | } | |
435 | ||
436 | field->destroy = lttng_destroy_perf_counter_field; | |
437 | ||
438 | field->event_field.name = name_alloc; | |
439 | field->event_field.type.atype = atype_integer; | |
440 | field->event_field.type.u.basic.integer.size = | |
441 | sizeof(uint64_t) * CHAR_BIT; | |
442 | field->event_field.type.u.basic.integer.alignment = | |
443 | lttng_alignof(uint64_t) * CHAR_BIT; | |
444 | field->event_field.type.u.basic.integer.signedness = | |
445 | lttng_is_signed_type(uint64_t); | |
446 | field->event_field.type.u.basic.integer.reverse_byte_order = 0; | |
447 | field->event_field.type.u.basic.integer.base = 10; | |
448 | field->event_field.type.u.basic.integer.encoding = lttng_encode_none; | |
449 | field->get_size = perf_counter_get_size; | |
450 | field->record = perf_counter_record; | |
451 | field->get_value = perf_counter_get_value; | |
452 | ||
453 | perf_field->attr.type = type; | |
454 | perf_field->attr.config = config; | |
455 | perf_field->attr.exclude_kernel = 1; | |
456 | CDS_INIT_LIST_HEAD(&perf_field->thread_field_list); | |
457 | field->u.perf_counter = perf_field; | |
458 | ||
459 | /* Ensure that this perf counter can be used in this process. */ | |
e336bbe5 MD |
460 | ret = open_perf_fd(&perf_field->attr); |
461 | if (ret < 0) { | |
d58d1454 MD |
462 | ret = -ENODEV; |
463 | goto setup_error; | |
464 | } | |
e336bbe5 | 465 | close_perf_fd(ret); |
d58d1454 MD |
466 | |
467 | /* | |
468 | * Contexts can only be added before tracing is started, so we | |
469 | * don't have to synchronize against concurrent threads using | |
470 | * the field here. | |
471 | */ | |
472 | ||
b2cc986a | 473 | lttng_context_update(*ctx); |
d58d1454 MD |
474 | return 0; |
475 | ||
476 | setup_error: | |
477 | find_error: | |
478 | lttng_remove_context_field(ctx, field); | |
479 | append_context_error: | |
480 | free(perf_field); | |
481 | perf_field_alloc_error: | |
482 | free(name_alloc); | |
483 | name_alloc_error: | |
484 | return ret; | |
485 | } | |
486 | ||
487 | int lttng_perf_counter_init(void) | |
488 | { | |
489 | int ret; | |
490 | ||
491 | ret = pthread_key_create(&perf_counter_key, | |
492 | lttng_destroy_perf_thread_key); | |
493 | if (ret) | |
494 | ret = -ret; | |
495 | return ret; | |
496 | } | |
497 | ||
498 | void lttng_perf_counter_exit(void) | |
499 | { | |
500 | int ret; | |
501 | ||
502 | ret = pthread_key_delete(perf_counter_key); | |
503 | if (ret) { | |
504 | errno = ret; | |
505 | PERROR("Error in pthread_key_delete"); | |
506 | } | |
507 | } |