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> | |
28 | #include <sys/mman.h> | |
29 | #include <sys/syscall.h> | |
30 | #include <linux/perf_event.h> | |
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> | |
41 | #include "lttng-tracer-core.h" | |
42 | ||
43 | /* | |
44 | * We use a global perf counter key and iterate on per-thread RCU lists | |
45 | * of fields in the fast path, even though this is not strictly speaking | |
46 | * what would provide the best fast-path complexity, to ensure teardown | |
47 | * of sessions vs thread exit is handled racelessly. | |
48 | * | |
49 | * Updates and traversals of thread_list are protected by UST lock. | |
50 | * Updates to rcu_field_list are protected by UST lock. | |
51 | */ | |
52 | ||
53 | struct lttng_perf_counter_thread_field { | |
54 | struct lttng_perf_counter_field *field; /* Back reference */ | |
55 | struct perf_event_mmap_page *pc; | |
56 | struct cds_list_head thread_field_node; /* Per-field list of thread fields (node) */ | |
57 | struct cds_list_head rcu_field_node; /* RCU per-thread list of fields (node) */ | |
b9389e6e | 58 | int fd; /* Perf FD */ |
d58d1454 MD |
59 | }; |
60 | ||
61 | struct lttng_perf_counter_thread { | |
62 | struct cds_list_head rcu_field_list; /* RCU per-thread list of fields */ | |
63 | }; | |
64 | ||
65 | struct lttng_perf_counter_field { | |
66 | struct perf_event_attr attr; | |
67 | struct cds_list_head thread_field_list; /* Per-field list of thread fields */ | |
68 | }; | |
69 | ||
70 | static pthread_key_t perf_counter_key; | |
71 | ||
72 | static | |
53569322 | 73 | size_t perf_counter_get_size(struct lttng_ctx_field *field, size_t offset) |
d58d1454 MD |
74 | { |
75 | size_t size = 0; | |
76 | ||
77 | size += lib_ring_buffer_align(offset, lttng_alignof(uint64_t)); | |
78 | size += sizeof(uint64_t); | |
79 | return size; | |
80 | } | |
81 | ||
82 | #if defined(__x86_64__) || defined(__i386__) | |
83 | ||
84 | static | |
85 | uint64_t rdpmc(unsigned int counter) | |
86 | { | |
87 | unsigned int low, high; | |
88 | ||
89 | asm volatile("rdpmc" : "=a" (low), "=d" (high) : "c" (counter)); | |
90 | ||
91 | return low | ((uint64_t) high) << 32; | |
92 | } | |
93 | ||
b9389e6e JD |
94 | static bool arch_perf_use_read(void) |
95 | { | |
96 | return false; | |
97 | } | |
98 | ||
d58d1454 MD |
99 | #else /* defined(__x86_64__) || defined(__i386__) */ |
100 | ||
101 | #error "Perf event counters are only supported on x86 so far." | |
102 | ||
103 | #endif /* #else defined(__x86_64__) || defined(__i386__) */ | |
104 | ||
105 | static | |
106 | uint64_t read_perf_counter(struct perf_event_mmap_page *pc) | |
107 | { | |
108 | uint32_t seq, idx; | |
109 | uint64_t count; | |
110 | ||
111 | if (caa_unlikely(!pc)) | |
112 | return 0; | |
113 | ||
114 | do { | |
115 | seq = CMM_LOAD_SHARED(pc->lock); | |
116 | cmm_barrier(); | |
117 | ||
118 | idx = pc->index; | |
119 | if (idx) | |
120 | count = pc->offset + rdpmc(idx - 1); | |
121 | else | |
122 | count = 0; | |
123 | ||
124 | cmm_barrier(); | |
125 | } while (CMM_LOAD_SHARED(pc->lock) != seq); | |
126 | ||
127 | return count; | |
128 | } | |
129 | ||
130 | static | |
131 | int sys_perf_event_open(struct perf_event_attr *attr, | |
132 | pid_t pid, int cpu, int group_fd, | |
133 | unsigned long flags) | |
134 | { | |
135 | return syscall(SYS_perf_event_open, attr, pid, cpu, | |
136 | group_fd, flags); | |
137 | } | |
138 | ||
139 | static | |
b9389e6e | 140 | int open_perf_fd(struct perf_event_attr *attr) |
d58d1454 | 141 | { |
b9389e6e | 142 | int fd; |
d58d1454 MD |
143 | |
144 | fd = sys_perf_event_open(attr, 0, -1, -1, 0); | |
145 | if (fd < 0) | |
b9389e6e JD |
146 | return -1; |
147 | ||
148 | return fd; | |
149 | } | |
150 | ||
151 | static | |
152 | struct perf_event_mmap_page *setup_perf( | |
153 | struct lttng_perf_counter_thread_field *thread_field) | |
154 | { | |
155 | void *perf_addr; | |
d58d1454 MD |
156 | |
157 | perf_addr = mmap(NULL, sizeof(struct perf_event_mmap_page), | |
b9389e6e | 158 | PROT_READ, MAP_SHARED, thread_field->fd, 0); |
d58d1454 | 159 | if (perf_addr == MAP_FAILED) |
b9389e6e JD |
160 | perf_addr = NULL; |
161 | ||
162 | if (!arch_perf_use_read()) { | |
163 | close_perf_fd(thread_field->fd); | |
164 | thread_field->fd = -1; | |
6c2125af | 165 | } |
b9389e6e JD |
166 | |
167 | end: | |
d58d1454 MD |
168 | return perf_addr; |
169 | } | |
170 | ||
b9389e6e JD |
171 | static |
172 | void close_perf_fd(int fd) | |
173 | { | |
174 | int ret; | |
175 | ||
176 | if (fd < 0) | |
177 | return; | |
178 | ||
179 | ret = close(fd); | |
180 | if (ret) | |
181 | perror("Error closing LTTng-UST perf memory mapping FD"); | |
182 | } | |
183 | ||
d58d1454 MD |
184 | static |
185 | void unmap_perf_page(struct perf_event_mmap_page *pc) | |
186 | { | |
187 | int ret; | |
188 | ||
189 | if (!pc) | |
190 | return; | |
191 | ret = munmap(pc, sizeof(struct perf_event_mmap_page)); | |
192 | if (ret < 0) { | |
193 | PERROR("Error in munmap"); | |
194 | abort(); | |
195 | } | |
196 | } | |
197 | ||
198 | static | |
199 | struct lttng_perf_counter_thread *alloc_perf_counter_thread(void) | |
200 | { | |
201 | struct lttng_perf_counter_thread *perf_thread; | |
202 | sigset_t newmask, oldmask; | |
203 | int ret; | |
204 | ||
205 | ret = sigfillset(&newmask); | |
206 | if (ret) | |
207 | abort(); | |
208 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); | |
209 | if (ret) | |
210 | abort(); | |
211 | /* Check again with signals disabled */ | |
212 | perf_thread = pthread_getspecific(perf_counter_key); | |
213 | if (perf_thread) | |
214 | goto skip; | |
215 | perf_thread = zmalloc(sizeof(*perf_thread)); | |
216 | if (!perf_thread) | |
217 | abort(); | |
218 | CDS_INIT_LIST_HEAD(&perf_thread->rcu_field_list); | |
219 | ret = pthread_setspecific(perf_counter_key, perf_thread); | |
220 | if (ret) | |
221 | abort(); | |
222 | skip: | |
223 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
224 | if (ret) | |
225 | abort(); | |
226 | return perf_thread; | |
227 | } | |
228 | ||
229 | static | |
230 | struct lttng_perf_counter_thread_field * | |
231 | add_thread_field(struct lttng_perf_counter_field *perf_field, | |
232 | struct lttng_perf_counter_thread *perf_thread) | |
233 | { | |
234 | struct lttng_perf_counter_thread_field *thread_field; | |
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 | cds_list_for_each_entry_rcu(thread_field, &perf_thread->rcu_field_list, | |
246 | rcu_field_node) { | |
247 | if (thread_field->field == perf_field) | |
248 | goto skip; | |
249 | } | |
250 | thread_field = zmalloc(sizeof(*thread_field)); | |
251 | if (!thread_field) | |
252 | abort(); | |
253 | thread_field->field = perf_field; | |
b9389e6e JD |
254 | thread_field->fd = open_perf_fd(&perf_field->attr); |
255 | if (thread_field->fd >= 0) | |
256 | thread_field->pc = setup_perf(thread_field); | |
257 | /* | |
258 | * Note: thread_field->pc can be NULL if setup_perf() fails. | |
259 | * Also, thread_field->fd can be -1 if open_perf_fd() fails. | |
260 | */ | |
d58d1454 MD |
261 | ust_lock_nocheck(); |
262 | cds_list_add_rcu(&thread_field->rcu_field_node, | |
263 | &perf_thread->rcu_field_list); | |
264 | cds_list_add(&thread_field->thread_field_node, | |
265 | &perf_field->thread_field_list); | |
266 | ust_unlock(); | |
267 | skip: | |
268 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
269 | if (ret) | |
270 | abort(); | |
271 | return thread_field; | |
272 | } | |
273 | ||
274 | static | |
275 | struct lttng_perf_counter_thread_field * | |
276 | get_thread_field(struct lttng_perf_counter_field *field) | |
277 | { | |
278 | struct lttng_perf_counter_thread *perf_thread; | |
279 | struct lttng_perf_counter_thread_field *thread_field; | |
280 | ||
281 | perf_thread = pthread_getspecific(perf_counter_key); | |
282 | if (!perf_thread) | |
283 | perf_thread = alloc_perf_counter_thread(); | |
284 | cds_list_for_each_entry_rcu(thread_field, &perf_thread->rcu_field_list, | |
285 | rcu_field_node) { | |
286 | if (thread_field->field == field) | |
287 | return thread_field; | |
288 | } | |
289 | /* perf_counter_thread_field not found, need to add one */ | |
290 | return add_thread_field(field, perf_thread); | |
291 | } | |
292 | ||
293 | static | |
294 | uint64_t wrapper_perf_counter_read(struct lttng_ctx_field *field) | |
295 | { | |
296 | struct lttng_perf_counter_field *perf_field; | |
297 | struct lttng_perf_counter_thread_field *perf_thread_field; | |
298 | ||
299 | perf_field = field->u.perf_counter; | |
300 | perf_thread_field = get_thread_field(perf_field); | |
301 | return read_perf_counter(perf_thread_field->pc); | |
302 | } | |
303 | ||
304 | static | |
305 | void perf_counter_record(struct lttng_ctx_field *field, | |
306 | struct lttng_ust_lib_ring_buffer_ctx *ctx, | |
307 | struct lttng_channel *chan) | |
308 | { | |
309 | uint64_t value; | |
310 | ||
311 | value = wrapper_perf_counter_read(field); | |
312 | lib_ring_buffer_align_ctx(ctx, lttng_alignof(value)); | |
313 | chan->ops->event_write(ctx, &value, sizeof(value)); | |
314 | } | |
315 | ||
316 | static | |
317 | void perf_counter_get_value(struct lttng_ctx_field *field, | |
53569322 | 318 | struct lttng_ctx_value *value) |
d58d1454 MD |
319 | { |
320 | uint64_t v; | |
321 | ||
322 | v = wrapper_perf_counter_read(field); | |
53569322 | 323 | value->u.s64 = v; |
d58d1454 MD |
324 | } |
325 | ||
326 | /* Called with UST lock held */ | |
327 | static | |
328 | void lttng_destroy_perf_thread_field( | |
329 | struct lttng_perf_counter_thread_field *thread_field) | |
330 | { | |
b9389e6e | 331 | close_perf_fd(thread_field->fd); |
d58d1454 MD |
332 | unmap_perf_page(thread_field->pc); |
333 | cds_list_del_rcu(&thread_field->rcu_field_node); | |
334 | cds_list_del(&thread_field->thread_field_node); | |
335 | free(thread_field); | |
336 | } | |
337 | ||
338 | static | |
339 | void lttng_destroy_perf_thread_key(void *_key) | |
340 | { | |
341 | struct lttng_perf_counter_thread *perf_thread = _key; | |
342 | struct lttng_perf_counter_thread_field *pos, *p; | |
343 | ||
344 | ust_lock_nocheck(); | |
345 | cds_list_for_each_entry_safe(pos, p, &perf_thread->rcu_field_list, | |
346 | rcu_field_node) | |
347 | lttng_destroy_perf_thread_field(pos); | |
348 | ust_unlock(); | |
349 | free(perf_thread); | |
350 | } | |
351 | ||
352 | /* Called with UST lock held */ | |
353 | static | |
354 | void lttng_destroy_perf_counter_field(struct lttng_ctx_field *field) | |
355 | { | |
356 | struct lttng_perf_counter_field *perf_field; | |
357 | struct lttng_perf_counter_thread_field *pos, *p; | |
358 | ||
359 | free((char *) field->event_field.name); | |
360 | perf_field = field->u.perf_counter; | |
361 | /* | |
362 | * This put is performed when no threads can concurrently | |
363 | * perform a "get" concurrently, thanks to urcu-bp grace | |
364 | * period. | |
365 | */ | |
366 | cds_list_for_each_entry_safe(pos, p, &perf_field->thread_field_list, | |
367 | thread_field_node) | |
368 | lttng_destroy_perf_thread_field(pos); | |
369 | free(perf_field); | |
370 | } | |
371 | ||
372 | /* Called with UST lock held */ | |
373 | int lttng_add_perf_counter_to_ctx(uint32_t type, | |
374 | uint64_t config, | |
375 | const char *name, | |
376 | struct lttng_ctx **ctx) | |
377 | { | |
378 | struct lttng_ctx_field *field; | |
379 | struct lttng_perf_counter_field *perf_field; | |
d58d1454 MD |
380 | char *name_alloc; |
381 | int ret; | |
382 | ||
383 | name_alloc = strdup(name); | |
384 | if (!name_alloc) { | |
385 | ret = -ENOMEM; | |
386 | goto name_alloc_error; | |
387 | } | |
388 | perf_field = zmalloc(sizeof(*perf_field)); | |
389 | if (!perf_field) { | |
390 | ret = -ENOMEM; | |
391 | goto perf_field_alloc_error; | |
392 | } | |
393 | field = lttng_append_context(ctx); | |
394 | if (!field) { | |
395 | ret = -ENOMEM; | |
396 | goto append_context_error; | |
397 | } | |
398 | if (lttng_find_context(*ctx, name_alloc)) { | |
399 | ret = -EEXIST; | |
400 | goto find_error; | |
401 | } | |
402 | ||
403 | field->destroy = lttng_destroy_perf_counter_field; | |
404 | ||
405 | field->event_field.name = name_alloc; | |
406 | field->event_field.type.atype = atype_integer; | |
407 | field->event_field.type.u.basic.integer.size = | |
408 | sizeof(uint64_t) * CHAR_BIT; | |
409 | field->event_field.type.u.basic.integer.alignment = | |
410 | lttng_alignof(uint64_t) * CHAR_BIT; | |
411 | field->event_field.type.u.basic.integer.signedness = | |
412 | lttng_is_signed_type(uint64_t); | |
413 | field->event_field.type.u.basic.integer.reverse_byte_order = 0; | |
414 | field->event_field.type.u.basic.integer.base = 10; | |
415 | field->event_field.type.u.basic.integer.encoding = lttng_encode_none; | |
416 | field->get_size = perf_counter_get_size; | |
417 | field->record = perf_counter_record; | |
418 | field->get_value = perf_counter_get_value; | |
419 | ||
420 | perf_field->attr.type = type; | |
421 | perf_field->attr.config = config; | |
422 | perf_field->attr.exclude_kernel = 1; | |
423 | CDS_INIT_LIST_HEAD(&perf_field->thread_field_list); | |
424 | field->u.perf_counter = perf_field; | |
425 | ||
426 | /* Ensure that this perf counter can be used in this process. */ | |
b9389e6e JD |
427 | ret = open_perf_fd(&perf_field->attr); |
428 | if (ret < 0) { | |
d58d1454 MD |
429 | ret = -ENODEV; |
430 | goto setup_error; | |
431 | } | |
b9389e6e | 432 | close_perf_fd(ret); |
d58d1454 MD |
433 | |
434 | /* | |
435 | * Contexts can only be added before tracing is started, so we | |
436 | * don't have to synchronize against concurrent threads using | |
437 | * the field here. | |
438 | */ | |
439 | ||
b2cc986a | 440 | lttng_context_update(*ctx); |
d58d1454 MD |
441 | return 0; |
442 | ||
443 | setup_error: | |
444 | find_error: | |
445 | lttng_remove_context_field(ctx, field); | |
446 | append_context_error: | |
447 | free(perf_field); | |
448 | perf_field_alloc_error: | |
449 | free(name_alloc); | |
450 | name_alloc_error: | |
451 | return ret; | |
452 | } | |
453 | ||
454 | int lttng_perf_counter_init(void) | |
455 | { | |
456 | int ret; | |
457 | ||
458 | ret = pthread_key_create(&perf_counter_key, | |
459 | lttng_destroy_perf_thread_key); | |
460 | if (ret) | |
461 | ret = -ret; | |
462 | return ret; | |
463 | } | |
464 | ||
465 | void lttng_perf_counter_exit(void) | |
466 | { | |
467 | int ret; | |
468 | ||
469 | ret = pthread_key_delete(perf_counter_key); | |
470 | if (ret) { | |
471 | errno = ret; | |
472 | PERROR("Error in pthread_key_delete"); | |
473 | } | |
474 | } |