Commit | Line | Data |
---|---|---|
852c2936 MD |
1 | /* |
2 | * ring_buffer_frontend.c | |
3 | * | |
e92f3e28 MD |
4 | * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
5 | * | |
6 | * This library is free software; you can redistribute it and/or | |
7 | * modify it under the terms of the GNU Lesser General Public | |
8 | * License as published by the Free Software Foundation; only | |
9 | * version 2.1 of the License. | |
10 | * | |
11 | * This library is distributed in the hope that it will be useful, | |
12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
14 | * Lesser General Public License for more details. | |
15 | * | |
16 | * You should have received a copy of the GNU Lesser General Public | |
17 | * License along with this library; if not, write to the Free Software | |
18 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
19 | * | |
852c2936 MD |
20 | * |
21 | * Ring buffer wait-free buffer synchronization. Producer-consumer and flight | |
22 | * recorder (overwrite) modes. See thesis: | |
23 | * | |
24 | * Desnoyers, Mathieu (2009), "Low-Impact Operating System Tracing", Ph.D. | |
25 | * dissertation, Ecole Polytechnique de Montreal. | |
26 | * http://www.lttng.org/pub/thesis/desnoyers-dissertation-2009-12.pdf | |
27 | * | |
28 | * - Algorithm presentation in Chapter 5: | |
29 | * "Lockless Multi-Core High-Throughput Buffering". | |
30 | * - Algorithm formal verification in Section 8.6: | |
31 | * "Formal verification of LTTng" | |
32 | * | |
33 | * Author: | |
34 | * Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
35 | * | |
36 | * Inspired from LTT and RelayFS: | |
37 | * Karim Yaghmour <karim@opersys.com> | |
38 | * Tom Zanussi <zanussi@us.ibm.com> | |
39 | * Bob Wisniewski <bob@watson.ibm.com> | |
40 | * And from K42 : | |
41 | * Bob Wisniewski <bob@watson.ibm.com> | |
42 | * | |
43 | * Buffer reader semantic : | |
44 | * | |
45 | * - get_subbuf_size | |
46 | * while buffer is not finalized and empty | |
47 | * - get_subbuf | |
48 | * - if return value != 0, continue | |
49 | * - splice one subbuffer worth of data to a pipe | |
50 | * - splice the data from pipe to disk/network | |
51 | * - put_subbuf | |
852c2936 MD |
52 | */ |
53 | ||
5ad63a16 | 54 | #define _GNU_SOURCE |
a6352fd4 | 55 | #include <sys/types.h> |
431d5cf0 MD |
56 | #include <sys/mman.h> |
57 | #include <sys/stat.h> | |
03d2d293 | 58 | #include <unistd.h> |
431d5cf0 | 59 | #include <fcntl.h> |
03d2d293 MD |
60 | #include <signal.h> |
61 | #include <time.h> | |
14641deb | 62 | #include <urcu/compiler.h> |
a6352fd4 | 63 | #include <urcu/ref.h> |
8c90a710 | 64 | #include <urcu/tls-compat.h> |
35897f8b | 65 | #include <helper.h> |
14641deb | 66 | |
a6352fd4 | 67 | #include "smp.h" |
4318ae1b | 68 | #include <lttng/ringbuffer-config.h> |
2fed87ae | 69 | #include "vatomic.h" |
4931a13e MD |
70 | #include "backend.h" |
71 | #include "frontend.h" | |
a6352fd4 | 72 | #include "shm.h" |
f645cfa7 | 73 | #include "tlsfixup.h" |
bdcf8d82 | 74 | #include "../liblttng-ust/compat.h" /* For ENODATA */ |
852c2936 | 75 | |
431d5cf0 MD |
76 | #ifndef max |
77 | #define max(a, b) ((a) > (b) ? (a) : (b)) | |
78 | #endif | |
79 | ||
64493e4f MD |
80 | /* Print DBG() messages about events lost only every 1048576 hits */ |
81 | #define DBG_PRINT_NR_LOST (1UL << 20) | |
82 | ||
03d2d293 MD |
83 | #define LTTNG_UST_RB_SIG SIGRTMIN |
84 | #define LTTNG_UST_RB_SIG_TEARDOWN SIGRTMIN + 1 | |
85 | #define CLOCKID CLOCK_MONOTONIC | |
86 | ||
2432c3c9 MD |
87 | /* |
88 | * Use POSIX SHM: shm_open(3) and shm_unlink(3). | |
89 | * close(2) to close the fd returned by shm_open. | |
90 | * shm_unlink releases the shared memory object name. | |
91 | * ftruncate(2) sets the size of the memory object. | |
92 | * mmap/munmap maps the shared memory obj to a virtual address in the | |
93 | * calling proceess (should be done both in libust and consumer). | |
94 | * See shm_overview(7) for details. | |
95 | * Pass file descriptor returned by shm_open(3) to ltt-sessiond through | |
96 | * a UNIX socket. | |
97 | * | |
98 | * Since we don't need to access the object using its name, we can | |
99 | * immediately shm_unlink(3) it, and only keep the handle with its file | |
100 | * descriptor. | |
101 | */ | |
102 | ||
852c2936 MD |
103 | /* |
104 | * Internal structure representing offsets to use at a sub-buffer switch. | |
105 | */ | |
106 | struct switch_offsets { | |
107 | unsigned long begin, end, old; | |
108 | size_t pre_header_padding, size; | |
109 | unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1, | |
110 | switch_old_end:1; | |
111 | }; | |
112 | ||
8c90a710 | 113 | DEFINE_URCU_TLS(unsigned int, lib_ring_buffer_nesting); |
852c2936 | 114 | |
cb7378b3 MD |
115 | /* |
116 | * wakeup_fd_mutex protects wakeup fd use by timer from concurrent | |
117 | * close. | |
118 | */ | |
119 | static pthread_mutex_t wakeup_fd_mutex = PTHREAD_MUTEX_INITIALIZER; | |
120 | ||
852c2936 MD |
121 | static |
122 | void lib_ring_buffer_print_errors(struct channel *chan, | |
009769ca MD |
123 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
124 | struct lttng_ust_shm_handle *handle); | |
852c2936 | 125 | |
03d2d293 MD |
126 | /* |
127 | * Handle timer teardown race wrt memory free of private data by | |
128 | * ring buffer signals are handled by a single thread, which permits | |
129 | * a synchronization point between handling of each signal. | |
130 | * Protected by the ust mutex. | |
131 | */ | |
132 | struct timer_signal_data { | |
133 | pthread_t tid; /* thread id managing signals */ | |
134 | int setup_done; | |
135 | int qs_done; | |
136 | }; | |
137 | ||
138 | static struct timer_signal_data timer_signal; | |
139 | ||
852c2936 MD |
140 | /** |
141 | * lib_ring_buffer_reset - Reset ring buffer to initial values. | |
142 | * @buf: Ring buffer. | |
143 | * | |
144 | * Effectively empty the ring buffer. Should be called when the buffer is not | |
145 | * used for writing. The ring buffer can be opened for reading, but the reader | |
146 | * should not be using the iterator concurrently with reset. The previous | |
147 | * current iterator record is reset. | |
148 | */ | |
4cfec15c | 149 | void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 150 | struct lttng_ust_shm_handle *handle) |
852c2936 | 151 | { |
1d498196 | 152 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 153 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
154 | unsigned int i; |
155 | ||
156 | /* | |
157 | * Reset iterator first. It will put the subbuffer if it currently holds | |
158 | * it. | |
159 | */ | |
852c2936 MD |
160 | v_set(config, &buf->offset, 0); |
161 | for (i = 0; i < chan->backend.num_subbuf; i++) { | |
4746ae29 MD |
162 | v_set(config, &shmp_index(handle, buf->commit_hot, i)->cc, 0); |
163 | v_set(config, &shmp_index(handle, buf->commit_hot, i)->seq, 0); | |
164 | v_set(config, &shmp_index(handle, buf->commit_cold, i)->cc_sb, 0); | |
852c2936 | 165 | } |
a6352fd4 MD |
166 | uatomic_set(&buf->consumed, 0); |
167 | uatomic_set(&buf->record_disabled, 0); | |
852c2936 | 168 | v_set(config, &buf->last_tsc, 0); |
1d498196 | 169 | lib_ring_buffer_backend_reset(&buf->backend, handle); |
852c2936 MD |
170 | /* Don't reset number of active readers */ |
171 | v_set(config, &buf->records_lost_full, 0); | |
172 | v_set(config, &buf->records_lost_wrap, 0); | |
173 | v_set(config, &buf->records_lost_big, 0); | |
174 | v_set(config, &buf->records_count, 0); | |
175 | v_set(config, &buf->records_overrun, 0); | |
176 | buf->finalized = 0; | |
177 | } | |
852c2936 MD |
178 | |
179 | /** | |
180 | * channel_reset - Reset channel to initial values. | |
181 | * @chan: Channel. | |
182 | * | |
183 | * Effectively empty the channel. Should be called when the channel is not used | |
184 | * for writing. The channel can be opened for reading, but the reader should not | |
185 | * be using the iterator concurrently with reset. The previous current iterator | |
186 | * record is reset. | |
187 | */ | |
188 | void channel_reset(struct channel *chan) | |
189 | { | |
190 | /* | |
191 | * Reset iterators first. Will put the subbuffer if held for reading. | |
192 | */ | |
a6352fd4 | 193 | uatomic_set(&chan->record_disabled, 0); |
852c2936 MD |
194 | /* Don't reset commit_count_mask, still valid */ |
195 | channel_backend_reset(&chan->backend); | |
196 | /* Don't reset switch/read timer interval */ | |
197 | /* Don't reset notifiers and notifier enable bits */ | |
198 | /* Don't reset reader reference count */ | |
199 | } | |
852c2936 MD |
200 | |
201 | /* | |
202 | * Must be called under cpu hotplug protection. | |
203 | */ | |
4cfec15c | 204 | int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf, |
a6352fd4 | 205 | struct channel_backend *chanb, int cpu, |
38fae1d3 | 206 | struct lttng_ust_shm_handle *handle, |
1d498196 | 207 | struct shm_object *shmobj) |
852c2936 | 208 | { |
4cfec15c | 209 | const struct lttng_ust_lib_ring_buffer_config *config = &chanb->config; |
14641deb | 210 | struct channel *chan = caa_container_of(chanb, struct channel, backend); |
a3f61e7f | 211 | void *priv = channel_get_private(chan); |
852c2936 | 212 | size_t subbuf_header_size; |
2fed87ae | 213 | uint64_t tsc; |
852c2936 MD |
214 | int ret; |
215 | ||
216 | /* Test for cpu hotplug */ | |
217 | if (buf->backend.allocated) | |
218 | return 0; | |
219 | ||
a6352fd4 | 220 | ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, |
1d498196 | 221 | cpu, handle, shmobj); |
852c2936 MD |
222 | if (ret) |
223 | return ret; | |
224 | ||
1d498196 MD |
225 | align_shm(shmobj, __alignof__(struct commit_counters_hot)); |
226 | set_shmp(buf->commit_hot, | |
227 | zalloc_shm(shmobj, | |
228 | sizeof(struct commit_counters_hot) * chan->backend.num_subbuf)); | |
229 | if (!shmp(handle, buf->commit_hot)) { | |
852c2936 MD |
230 | ret = -ENOMEM; |
231 | goto free_chanbuf; | |
232 | } | |
233 | ||
1d498196 MD |
234 | align_shm(shmobj, __alignof__(struct commit_counters_cold)); |
235 | set_shmp(buf->commit_cold, | |
236 | zalloc_shm(shmobj, | |
237 | sizeof(struct commit_counters_cold) * chan->backend.num_subbuf)); | |
238 | if (!shmp(handle, buf->commit_cold)) { | |
852c2936 MD |
239 | ret = -ENOMEM; |
240 | goto free_commit; | |
241 | } | |
242 | ||
852c2936 MD |
243 | /* |
244 | * Write the subbuffer header for first subbuffer so we know the total | |
245 | * duration of data gathering. | |
246 | */ | |
247 | subbuf_header_size = config->cb.subbuffer_header_size(); | |
248 | v_set(config, &buf->offset, subbuf_header_size); | |
4746ae29 | 249 | subbuffer_id_clear_noref(config, &shmp_index(handle, buf->backend.buf_wsb, 0)->id); |
1d498196 MD |
250 | tsc = config->cb.ring_buffer_clock_read(shmp(handle, buf->backend.chan)); |
251 | config->cb.buffer_begin(buf, tsc, 0, handle); | |
4746ae29 | 252 | v_add(config, subbuf_header_size, &shmp_index(handle, buf->commit_hot, 0)->cc); |
852c2936 MD |
253 | |
254 | if (config->cb.buffer_create) { | |
1d498196 | 255 | ret = config->cb.buffer_create(buf, priv, cpu, chanb->name, handle); |
852c2936 MD |
256 | if (ret) |
257 | goto free_init; | |
258 | } | |
852c2936 | 259 | buf->backend.allocated = 1; |
852c2936 MD |
260 | return 0; |
261 | ||
262 | /* Error handling */ | |
263 | free_init: | |
a6352fd4 | 264 | /* commit_cold will be freed by shm teardown */ |
852c2936 | 265 | free_commit: |
a6352fd4 | 266 | /* commit_hot will be freed by shm teardown */ |
852c2936 | 267 | free_chanbuf: |
852c2936 MD |
268 | return ret; |
269 | } | |
270 | ||
1d498196 | 271 | #if 0 |
852c2936 MD |
272 | static void switch_buffer_timer(unsigned long data) |
273 | { | |
4cfec15c | 274 | struct lttng_ust_lib_ring_buffer *buf = (struct lttng_ust_lib_ring_buffer *)data; |
1d498196 | 275 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 276 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
277 | |
278 | /* | |
279 | * Only flush buffers periodically if readers are active. | |
280 | */ | |
74d81a6c | 281 | if (uatomic_read(&buf->active_readers)) |
1d498196 | 282 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, handle); |
852c2936 | 283 | |
a6352fd4 MD |
284 | //TODO timers |
285 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
286 | // mod_timer_pinned(&buf->switch_timer, | |
287 | // jiffies + chan->switch_timer_interval); | |
288 | //else | |
289 | // mod_timer(&buf->switch_timer, | |
290 | // jiffies + chan->switch_timer_interval); | |
852c2936 | 291 | } |
1d498196 | 292 | #endif //0 |
852c2936 | 293 | |
03d2d293 MD |
294 | static |
295 | void lib_ring_buffer_channel_switch_timer(int sig, siginfo_t *si, void *uc) | |
852c2936 | 296 | { |
03d2d293 MD |
297 | const struct lttng_ust_lib_ring_buffer_config *config; |
298 | struct lttng_ust_shm_handle *handle; | |
299 | struct channel *chan; | |
300 | int cpu; | |
301 | ||
302 | assert(CMM_LOAD_SHARED(timer_signal.tid) == pthread_self()); | |
303 | ||
304 | chan = si->si_value.sival_ptr; | |
305 | handle = chan->handle; | |
306 | config = &chan->backend.config; | |
852c2936 | 307 | |
03d2d293 MD |
308 | DBG("Timer for channel %p\n", chan); |
309 | ||
cb7378b3 | 310 | pthread_mutex_lock(&wakeup_fd_mutex); |
03d2d293 MD |
311 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
312 | for_each_possible_cpu(cpu) { | |
313 | struct lttng_ust_lib_ring_buffer *buf = | |
314 | shmp(handle, chan->backend.buf[cpu].shmp); | |
315 | ||
316 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, | |
317 | chan->handle); | |
318 | } | |
319 | } else { | |
320 | struct lttng_ust_lib_ring_buffer *buf = | |
321 | shmp(handle, chan->backend.buf[0].shmp); | |
322 | ||
323 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, | |
324 | chan->handle); | |
325 | } | |
cb7378b3 | 326 | pthread_mutex_unlock(&wakeup_fd_mutex); |
03d2d293 MD |
327 | return; |
328 | } | |
329 | ||
330 | static | |
331 | void rb_setmask(sigset_t *mask) | |
332 | { | |
333 | int ret; | |
334 | ||
335 | ret = sigemptyset(mask); | |
336 | if (ret) { | |
337 | PERROR("sigemptyset"); | |
338 | } | |
339 | ret = sigaddset(mask, LTTNG_UST_RB_SIG); | |
340 | if (ret) { | |
341 | PERROR("sigaddset"); | |
342 | } | |
343 | ret = sigaddset(mask, LTTNG_UST_RB_SIG_TEARDOWN); | |
344 | if (ret) { | |
345 | PERROR("sigaddset"); | |
346 | } | |
347 | } | |
348 | ||
349 | static | |
350 | void *sig_thread(void *arg) | |
351 | { | |
352 | sigset_t mask; | |
353 | siginfo_t info; | |
354 | int signr; | |
355 | ||
356 | /* Only self thread will receive signal mask. */ | |
357 | rb_setmask(&mask); | |
358 | CMM_STORE_SHARED(timer_signal.tid, pthread_self()); | |
359 | ||
360 | for (;;) { | |
361 | signr = sigwaitinfo(&mask, &info); | |
362 | if (signr == -1) { | |
363 | PERROR("sigwaitinfo"); | |
364 | continue; | |
365 | } | |
366 | if (signr == LTTNG_UST_RB_SIG) { | |
367 | lib_ring_buffer_channel_switch_timer(info.si_signo, | |
368 | &info, NULL); | |
369 | } else if (signr == LTTNG_UST_RB_SIG_TEARDOWN) { | |
370 | cmm_smp_mb(); | |
371 | CMM_STORE_SHARED(timer_signal.qs_done, 1); | |
372 | cmm_smp_mb(); | |
373 | } else { | |
374 | ERR("Unexptected signal %d\n", info.si_signo); | |
375 | } | |
376 | } | |
377 | return NULL; | |
378 | } | |
379 | ||
380 | /* | |
381 | * Called with ust_lock() held. | |
382 | * Ensure only a single thread listens on the timer signal. | |
383 | */ | |
384 | static | |
385 | void lib_ring_buffer_setup_timer_thread(void) | |
386 | { | |
387 | pthread_t thread; | |
388 | int ret; | |
389 | ||
390 | if (timer_signal.setup_done) | |
852c2936 | 391 | return; |
03d2d293 MD |
392 | |
393 | ret = pthread_create(&thread, NULL, &sig_thread, NULL); | |
394 | if (ret) { | |
395 | errno = ret; | |
396 | PERROR("pthread_create"); | |
397 | } | |
398 | ret = pthread_detach(thread); | |
399 | if (ret) { | |
400 | errno = ret; | |
401 | PERROR("pthread_detach"); | |
402 | } | |
403 | timer_signal.setup_done = 1; | |
852c2936 MD |
404 | } |
405 | ||
03d2d293 MD |
406 | /* |
407 | * Called with ust_lock() held. | |
408 | */ | |
409 | static | |
410 | void lib_ring_buffer_channel_switch_timer_start(struct channel *chan) | |
852c2936 | 411 | { |
03d2d293 MD |
412 | struct sigevent sev; |
413 | struct itimerspec its; | |
414 | int ret; | |
852c2936 | 415 | |
03d2d293 | 416 | if (!chan->switch_timer_interval || chan->switch_timer_enabled) |
852c2936 MD |
417 | return; |
418 | ||
03d2d293 MD |
419 | chan->switch_timer_enabled = 1; |
420 | ||
421 | lib_ring_buffer_setup_timer_thread(); | |
422 | ||
423 | sev.sigev_notify = SIGEV_SIGNAL; | |
424 | sev.sigev_signo = LTTNG_UST_RB_SIG; | |
425 | sev.sigev_value.sival_ptr = chan; | |
426 | ret = timer_create(CLOCKID, &sev, &chan->switch_timer); | |
427 | if (ret == -1) { | |
428 | PERROR("timer_create"); | |
429 | } | |
430 | ||
431 | its.it_value.tv_sec = chan->switch_timer_interval / 1000000; | |
432 | its.it_value.tv_nsec = chan->switch_timer_interval % 1000000; | |
433 | its.it_interval.tv_sec = its.it_value.tv_sec; | |
434 | its.it_interval.tv_nsec = its.it_value.tv_nsec; | |
435 | ||
436 | ret = timer_settime(chan->switch_timer, 0, &its, NULL); | |
437 | if (ret == -1) { | |
438 | PERROR("timer_settime"); | |
439 | } | |
440 | } | |
441 | ||
442 | /* | |
443 | * Called with ust_lock() held. | |
444 | */ | |
445 | static | |
446 | void lib_ring_buffer_channel_switch_timer_stop(struct channel *chan) | |
447 | { | |
448 | sigset_t pending_set; | |
449 | int sig_is_pending, ret; | |
450 | ||
451 | if (!chan->switch_timer_interval || !chan->switch_timer_enabled) | |
452 | return; | |
453 | ||
454 | ret = timer_delete(chan->switch_timer); | |
455 | if (ret == -1) { | |
456 | PERROR("timer_delete"); | |
457 | } | |
458 | ||
459 | /* | |
460 | * Ensure we don't have any signal queued for this channel. | |
461 | */ | |
462 | for (;;) { | |
463 | ret = sigemptyset(&pending_set); | |
464 | if (ret == -1) { | |
465 | PERROR("sigemptyset"); | |
466 | } | |
467 | ret = sigpending(&pending_set); | |
468 | if (ret == -1) { | |
469 | PERROR("sigpending"); | |
470 | } | |
471 | sig_is_pending = sigismember(&pending_set, LTTNG_UST_RB_SIG); | |
472 | if (!sig_is_pending) | |
473 | break; | |
474 | caa_cpu_relax(); | |
475 | } | |
476 | ||
477 | /* | |
478 | * From this point, no new signal handler will be fired that | |
479 | * would try to access "chan". However, we still need to wait | |
480 | * for any currently executing handler to complete. | |
481 | */ | |
482 | cmm_smp_mb(); | |
483 | CMM_STORE_SHARED(timer_signal.qs_done, 0); | |
484 | cmm_smp_mb(); | |
485 | ||
486 | /* | |
487 | * Kill with LTTNG_UST_RB_SIG_TEARDOWN, so signal management | |
488 | * thread wakes up. | |
489 | */ | |
490 | kill(getpid(), LTTNG_UST_RB_SIG_TEARDOWN); | |
491 | ||
492 | while (!CMM_LOAD_SHARED(timer_signal.qs_done)) | |
493 | caa_cpu_relax(); | |
494 | cmm_smp_mb(); | |
495 | ||
496 | chan->switch_timer = 0; | |
497 | chan->switch_timer_enabled = 0; | |
852c2936 MD |
498 | } |
499 | ||
1d498196 | 500 | #if 0 |
852c2936 MD |
501 | /* |
502 | * Polling timer to check the channels for data. | |
503 | */ | |
504 | static void read_buffer_timer(unsigned long data) | |
505 | { | |
4cfec15c | 506 | struct lttng_ust_lib_ring_buffer *buf = (struct lttng_ust_lib_ring_buffer *)data; |
1d498196 | 507 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 508 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
509 | |
510 | CHAN_WARN_ON(chan, !buf->backend.allocated); | |
511 | ||
74d81a6c | 512 | if (uatomic_read(&buf->active_readers)) |
852c2936 | 513 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { |
a6352fd4 MD |
514 | //TODO |
515 | //wake_up_interruptible(&buf->read_wait); | |
516 | //wake_up_interruptible(&chan->read_wait); | |
852c2936 MD |
517 | } |
518 | ||
a6352fd4 MD |
519 | //TODO |
520 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
521 | // mod_timer_pinned(&buf->read_timer, | |
522 | // jiffies + chan->read_timer_interval); | |
523 | //else | |
524 | // mod_timer(&buf->read_timer, | |
525 | // jiffies + chan->read_timer_interval); | |
852c2936 | 526 | } |
1d498196 | 527 | #endif //0 |
852c2936 | 528 | |
4cfec15c | 529 | static void lib_ring_buffer_start_read_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 530 | struct lttng_ust_shm_handle *handle) |
852c2936 | 531 | { |
1d498196 | 532 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 533 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
534 | |
535 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
536 | || !chan->read_timer_interval | |
537 | || buf->read_timer_enabled) | |
538 | return; | |
539 | ||
a6352fd4 MD |
540 | //TODO |
541 | //init_timer(&buf->read_timer); | |
542 | //buf->read_timer.function = read_buffer_timer; | |
543 | //buf->read_timer.expires = jiffies + chan->read_timer_interval; | |
544 | //buf->read_timer.data = (unsigned long)buf; | |
852c2936 | 545 | |
a6352fd4 MD |
546 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
547 | // add_timer_on(&buf->read_timer, buf->backend.cpu); | |
548 | //else | |
549 | // add_timer(&buf->read_timer); | |
852c2936 MD |
550 | buf->read_timer_enabled = 1; |
551 | } | |
552 | ||
4cfec15c | 553 | static void lib_ring_buffer_stop_read_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 554 | struct lttng_ust_shm_handle *handle) |
852c2936 | 555 | { |
1d498196 | 556 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 557 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
558 | |
559 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
560 | || !chan->read_timer_interval | |
561 | || !buf->read_timer_enabled) | |
562 | return; | |
563 | ||
a6352fd4 MD |
564 | //TODO |
565 | //del_timer_sync(&buf->read_timer); | |
852c2936 MD |
566 | /* |
567 | * do one more check to catch data that has been written in the last | |
568 | * timer period. | |
569 | */ | |
1d498196 | 570 | if (lib_ring_buffer_poll_deliver(config, buf, chan, handle)) { |
a6352fd4 MD |
571 | //TODO |
572 | //wake_up_interruptible(&buf->read_wait); | |
573 | //wake_up_interruptible(&chan->read_wait); | |
852c2936 MD |
574 | } |
575 | buf->read_timer_enabled = 0; | |
576 | } | |
577 | ||
1d498196 | 578 | static void channel_unregister_notifiers(struct channel *chan, |
38fae1d3 | 579 | struct lttng_ust_shm_handle *handle) |
852c2936 | 580 | { |
4cfec15c | 581 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
582 | int cpu; |
583 | ||
03d2d293 | 584 | lib_ring_buffer_channel_switch_timer_stop(chan); |
852c2936 | 585 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
852c2936 | 586 | for_each_possible_cpu(cpu) { |
4cfec15c | 587 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
a6352fd4 | 588 | |
1d498196 | 589 | lib_ring_buffer_stop_read_timer(buf, handle); |
852c2936 | 590 | } |
852c2936 | 591 | } else { |
4cfec15c | 592 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 593 | |
1d498196 | 594 | lib_ring_buffer_stop_read_timer(buf, handle); |
852c2936 | 595 | } |
8d8a24c8 | 596 | //channel_backend_unregister_notifiers(&chan->backend); |
852c2936 MD |
597 | } |
598 | ||
009769ca MD |
599 | static void channel_print_errors(struct channel *chan, |
600 | struct lttng_ust_shm_handle *handle) | |
601 | { | |
602 | const struct lttng_ust_lib_ring_buffer_config *config = | |
603 | &chan->backend.config; | |
604 | int cpu; | |
605 | ||
606 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
607 | for_each_possible_cpu(cpu) { | |
608 | struct lttng_ust_lib_ring_buffer *buf = | |
609 | shmp(handle, chan->backend.buf[cpu].shmp); | |
610 | lib_ring_buffer_print_errors(chan, buf, cpu, handle); | |
611 | } | |
612 | } else { | |
613 | struct lttng_ust_lib_ring_buffer *buf = | |
614 | shmp(handle, chan->backend.buf[0].shmp); | |
615 | ||
616 | lib_ring_buffer_print_errors(chan, buf, -1, handle); | |
617 | } | |
618 | } | |
619 | ||
620 | static void channel_free(struct channel *chan, | |
621 | struct lttng_ust_shm_handle *handle) | |
852c2936 | 622 | { |
74d81a6c | 623 | channel_backend_free(&chan->backend, handle); |
431d5cf0 | 624 | /* chan is freed by shm teardown */ |
1d498196 MD |
625 | shm_object_table_destroy(handle->table); |
626 | free(handle); | |
852c2936 MD |
627 | } |
628 | ||
629 | /** | |
630 | * channel_create - Create channel. | |
631 | * @config: ring buffer instance configuration | |
632 | * @name: name of the channel | |
a3f61e7f MD |
633 | * @priv_data: ring buffer client private data area pointer (output) |
634 | * @priv_data_size: length, in bytes, of the private data area. | |
d028eddb | 635 | * @priv_data_init: initialization data for private data. |
852c2936 MD |
636 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous |
637 | * address mapping. It is used only by RING_BUFFER_STATIC | |
638 | * configuration. It can be set to NULL for other backends. | |
639 | * @subbuf_size: subbuffer size | |
640 | * @num_subbuf: number of subbuffers | |
641 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
642 | * padding to let readers get those sub-buffers. | |
643 | * Used for live streaming. | |
644 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
645 | * | |
646 | * Holds cpu hotplug. | |
647 | * Returns NULL on failure. | |
648 | */ | |
4cfec15c | 649 | struct lttng_ust_shm_handle *channel_create(const struct lttng_ust_lib_ring_buffer_config *config, |
a3f61e7f MD |
650 | const char *name, |
651 | void **priv_data, | |
652 | size_t priv_data_align, | |
653 | size_t priv_data_size, | |
d028eddb | 654 | void *priv_data_init, |
a3f61e7f | 655 | void *buf_addr, size_t subbuf_size, |
852c2936 | 656 | size_t num_subbuf, unsigned int switch_timer_interval, |
74d81a6c | 657 | unsigned int read_timer_interval) |
852c2936 | 658 | { |
1d498196 | 659 | int ret, cpu; |
a3f61e7f | 660 | size_t shmsize, chansize; |
852c2936 | 661 | struct channel *chan; |
38fae1d3 | 662 | struct lttng_ust_shm_handle *handle; |
1d498196 | 663 | struct shm_object *shmobj; |
74d81a6c MD |
664 | unsigned int nr_streams; |
665 | ||
666 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
667 | nr_streams = num_possible_cpus(); | |
668 | else | |
669 | nr_streams = 1; | |
852c2936 MD |
670 | |
671 | if (lib_ring_buffer_check_config(config, switch_timer_interval, | |
672 | read_timer_interval)) | |
673 | return NULL; | |
674 | ||
38fae1d3 | 675 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
431d5cf0 MD |
676 | if (!handle) |
677 | return NULL; | |
678 | ||
1d498196 MD |
679 | /* Allocate table for channel + per-cpu buffers */ |
680 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
681 | if (!handle->table) | |
682 | goto error_table_alloc; | |
852c2936 | 683 | |
1d498196 MD |
684 | /* Calculate the shm allocation layout */ |
685 | shmsize = sizeof(struct channel); | |
c1fca457 | 686 | shmsize += offset_align(shmsize, __alignof__(struct lttng_ust_lib_ring_buffer_shmp)); |
74d81a6c | 687 | shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp) * nr_streams; |
a3f61e7f | 688 | chansize = shmsize; |
74d81a6c MD |
689 | if (priv_data_align) |
690 | shmsize += offset_align(shmsize, priv_data_align); | |
a3f61e7f | 691 | shmsize += priv_data_size; |
a6352fd4 | 692 | |
74d81a6c MD |
693 | /* Allocate normal memory for channel (not shared) */ |
694 | shmobj = shm_object_table_alloc(handle->table, shmsize, SHM_OBJECT_MEM); | |
b5a14697 MD |
695 | if (!shmobj) |
696 | goto error_append; | |
57773204 | 697 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
a3f61e7f | 698 | set_shmp(handle->chan, zalloc_shm(shmobj, chansize)); |
57773204 MD |
699 | assert(handle->chan._ref.index == 0); |
700 | assert(handle->chan._ref.offset == 0); | |
1d498196 | 701 | chan = shmp(handle, handle->chan); |
a6352fd4 | 702 | if (!chan) |
1d498196 | 703 | goto error_append; |
74d81a6c | 704 | chan->nr_streams = nr_streams; |
a6352fd4 | 705 | |
a3f61e7f MD |
706 | /* space for private data */ |
707 | if (priv_data_size) { | |
708 | DECLARE_SHMP(void, priv_data_alloc); | |
709 | ||
710 | align_shm(shmobj, priv_data_align); | |
711 | chan->priv_data_offset = shmobj->allocated_len; | |
712 | set_shmp(priv_data_alloc, zalloc_shm(shmobj, priv_data_size)); | |
713 | if (!shmp(handle, priv_data_alloc)) | |
714 | goto error_append; | |
715 | *priv_data = channel_get_private(chan); | |
d028eddb | 716 | memcpy(*priv_data, priv_data_init, priv_data_size); |
a3f61e7f MD |
717 | } else { |
718 | chan->priv_data_offset = -1; | |
74d81a6c MD |
719 | if (priv_data) |
720 | *priv_data = NULL; | |
a3f61e7f MD |
721 | } |
722 | ||
723 | ret = channel_backend_init(&chan->backend, name, config, | |
1d498196 | 724 | subbuf_size, num_subbuf, handle); |
852c2936 | 725 | if (ret) |
1d498196 | 726 | goto error_backend_init; |
852c2936 | 727 | |
03d2d293 | 728 | chan->handle = handle; |
852c2936 | 729 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); |
03d2d293 MD |
730 | chan->switch_timer_interval = switch_timer_interval; |
731 | ||
a6352fd4 | 732 | //TODO |
03d2d293 | 733 | //chan->read_timer_interval = read_timer_interval; |
a6352fd4 MD |
734 | //init_waitqueue_head(&chan->read_wait); |
735 | //init_waitqueue_head(&chan->hp_wait); | |
852c2936 | 736 | |
03d2d293 | 737 | lib_ring_buffer_channel_switch_timer_start(chan); |
852c2936 | 738 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
852c2936 MD |
739 | /* |
740 | * In case of non-hotplug cpu, if the ring-buffer is allocated | |
741 | * in early initcall, it will not be notified of secondary cpus. | |
742 | * In that off case, we need to allocate for all possible cpus. | |
743 | */ | |
852c2936 | 744 | for_each_possible_cpu(cpu) { |
4cfec15c | 745 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
1d498196 | 746 | lib_ring_buffer_start_read_timer(buf, handle); |
852c2936 | 747 | } |
852c2936 | 748 | } else { |
4cfec15c | 749 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 750 | |
1d498196 | 751 | lib_ring_buffer_start_read_timer(buf, handle); |
852c2936 | 752 | } |
431d5cf0 | 753 | return handle; |
852c2936 | 754 | |
1d498196 MD |
755 | error_backend_init: |
756 | error_append: | |
757 | shm_object_table_destroy(handle->table); | |
758 | error_table_alloc: | |
431d5cf0 | 759 | free(handle); |
852c2936 MD |
760 | return NULL; |
761 | } | |
852c2936 | 762 | |
74d81a6c | 763 | struct lttng_ust_shm_handle *channel_handle_create(void *data, |
ff0f5728 MD |
764 | uint64_t memory_map_size, |
765 | int wakeup_fd) | |
193183fb | 766 | { |
38fae1d3 | 767 | struct lttng_ust_shm_handle *handle; |
193183fb MD |
768 | struct shm_object *object; |
769 | ||
38fae1d3 | 770 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
193183fb MD |
771 | if (!handle) |
772 | return NULL; | |
773 | ||
774 | /* Allocate table for channel + per-cpu buffers */ | |
775 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
776 | if (!handle->table) | |
777 | goto error_table_alloc; | |
778 | /* Add channel object */ | |
74d81a6c | 779 | object = shm_object_table_append_mem(handle->table, data, |
ff0f5728 | 780 | memory_map_size, wakeup_fd); |
193183fb MD |
781 | if (!object) |
782 | goto error_table_object; | |
57773204 MD |
783 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
784 | handle->chan._ref.index = 0; | |
785 | handle->chan._ref.offset = 0; | |
193183fb MD |
786 | return handle; |
787 | ||
788 | error_table_object: | |
789 | shm_object_table_destroy(handle->table); | |
790 | error_table_alloc: | |
791 | free(handle); | |
792 | return NULL; | |
793 | } | |
794 | ||
38fae1d3 | 795 | int channel_handle_add_stream(struct lttng_ust_shm_handle *handle, |
74d81a6c MD |
796 | int shm_fd, int wakeup_fd, uint32_t stream_nr, |
797 | uint64_t memory_map_size) | |
193183fb MD |
798 | { |
799 | struct shm_object *object; | |
800 | ||
801 | /* Add stream object */ | |
74d81a6c MD |
802 | object = shm_object_table_append_shm(handle->table, |
803 | shm_fd, wakeup_fd, stream_nr, | |
804 | memory_map_size); | |
193183fb | 805 | if (!object) |
74d81a6c | 806 | return -EINVAL; |
193183fb MD |
807 | return 0; |
808 | } | |
809 | ||
74d81a6c MD |
810 | unsigned int channel_handle_get_nr_streams(struct lttng_ust_shm_handle *handle) |
811 | { | |
812 | assert(handle->table); | |
813 | return handle->table->allocated_len - 1; | |
814 | } | |
815 | ||
852c2936 | 816 | static |
74d81a6c | 817 | void channel_release(struct channel *chan, struct lttng_ust_shm_handle *handle) |
852c2936 | 818 | { |
74d81a6c | 819 | channel_free(chan, handle); |
852c2936 MD |
820 | } |
821 | ||
822 | /** | |
823 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
824 | * @chan: channel to destroy | |
825 | * | |
826 | * Holds cpu hotplug. | |
431d5cf0 MD |
827 | * Call "destroy" callback, finalize channels, decrement the channel |
828 | * reference count. Note that when readers have completed data | |
829 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
a3f61e7f | 830 | * They should release their handle at that point. |
852c2936 | 831 | */ |
a3f61e7f | 832 | void channel_destroy(struct channel *chan, struct lttng_ust_shm_handle *handle, |
74d81a6c | 833 | int consumer) |
852c2936 | 834 | { |
74d81a6c MD |
835 | if (consumer) { |
836 | /* | |
837 | * Note: the consumer takes care of finalizing and | |
838 | * switching the buffers. | |
839 | */ | |
840 | channel_unregister_notifiers(chan, handle); | |
3d0ef9f6 MD |
841 | /* |
842 | * The consumer prints errors. | |
843 | */ | |
844 | channel_print_errors(chan, handle); | |
824f40b8 MD |
845 | } |
846 | ||
431d5cf0 MD |
847 | /* |
848 | * sessiond/consumer are keeping a reference on the shm file | |
849 | * descriptor directly. No need to refcount. | |
850 | */ | |
74d81a6c | 851 | channel_release(chan, handle); |
a3f61e7f | 852 | return; |
852c2936 | 853 | } |
852c2936 | 854 | |
4cfec15c MD |
855 | struct lttng_ust_lib_ring_buffer *channel_get_ring_buffer( |
856 | const struct lttng_ust_lib_ring_buffer_config *config, | |
1d498196 | 857 | struct channel *chan, int cpu, |
38fae1d3 | 858 | struct lttng_ust_shm_handle *handle, |
74d81a6c MD |
859 | int *shm_fd, int *wait_fd, |
860 | int *wakeup_fd, | |
861 | uint64_t *memory_map_size) | |
852c2936 | 862 | { |
381c0f1e MD |
863 | struct shm_ref *ref; |
864 | ||
865 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
74d81a6c | 866 | cpu = 0; |
381c0f1e | 867 | } else { |
e095d803 MD |
868 | if (cpu >= num_possible_cpus()) |
869 | return NULL; | |
381c0f1e | 870 | } |
74d81a6c MD |
871 | ref = &chan->backend.buf[cpu].shmp._ref; |
872 | *shm_fd = shm_get_shm_fd(handle, ref); | |
873 | *wait_fd = shm_get_wait_fd(handle, ref); | |
874 | *wakeup_fd = shm_get_wakeup_fd(handle, ref); | |
875 | if (shm_get_shm_size(handle, ref, memory_map_size)) | |
876 | return NULL; | |
877 | return shmp(handle, chan->backend.buf[cpu].shmp); | |
852c2936 | 878 | } |
852c2936 | 879 | |
ff0f5728 MD |
880 | int ring_buffer_channel_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config *config, |
881 | struct channel *chan, | |
882 | struct lttng_ust_shm_handle *handle) | |
883 | { | |
884 | struct shm_ref *ref; | |
885 | ||
886 | ref = &handle->chan._ref; | |
887 | return shm_close_wait_fd(handle, ref); | |
888 | } | |
889 | ||
890 | int ring_buffer_channel_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config *config, | |
891 | struct channel *chan, | |
892 | struct lttng_ust_shm_handle *handle) | |
893 | { | |
894 | struct shm_ref *ref; | |
895 | ||
896 | ref = &handle->chan._ref; | |
897 | return shm_close_wakeup_fd(handle, ref); | |
898 | } | |
899 | ||
900 | int ring_buffer_stream_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config *config, | |
74d81a6c MD |
901 | struct channel *chan, |
902 | struct lttng_ust_shm_handle *handle, | |
903 | int cpu) | |
852c2936 | 904 | { |
74d81a6c MD |
905 | struct shm_ref *ref; |
906 | ||
907 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
908 | cpu = 0; | |
909 | } else { | |
910 | if (cpu >= num_possible_cpus()) | |
911 | return -EINVAL; | |
824f40b8 | 912 | } |
74d81a6c MD |
913 | ref = &chan->backend.buf[cpu].shmp._ref; |
914 | return shm_close_wait_fd(handle, ref); | |
915 | } | |
916 | ||
ff0f5728 | 917 | int ring_buffer_stream_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config *config, |
74d81a6c MD |
918 | struct channel *chan, |
919 | struct lttng_ust_shm_handle *handle, | |
920 | int cpu) | |
921 | { | |
922 | struct shm_ref *ref; | |
cb7378b3 | 923 | int ret; |
74d81a6c MD |
924 | |
925 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
926 | cpu = 0; | |
927 | } else { | |
928 | if (cpu >= num_possible_cpus()) | |
929 | return -EINVAL; | |
930 | } | |
931 | ref = &chan->backend.buf[cpu].shmp._ref; | |
cb7378b3 MD |
932 | pthread_mutex_lock(&wakeup_fd_mutex); |
933 | ret = shm_close_wakeup_fd(handle, ref); | |
934 | pthread_mutex_unlock(&wakeup_fd_mutex); | |
935 | return ret; | |
74d81a6c MD |
936 | } |
937 | ||
938 | int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer *buf, | |
939 | struct lttng_ust_shm_handle *handle) | |
940 | { | |
a6352fd4 | 941 | if (uatomic_cmpxchg(&buf->active_readers, 0, 1) != 0) |
852c2936 | 942 | return -EBUSY; |
a6352fd4 | 943 | cmm_smp_mb(); |
852c2936 MD |
944 | return 0; |
945 | } | |
852c2936 | 946 | |
4cfec15c | 947 | void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer *buf, |
74d81a6c | 948 | struct lttng_ust_shm_handle *handle) |
852c2936 | 949 | { |
1d498196 | 950 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 | 951 | |
a6352fd4 MD |
952 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
953 | cmm_smp_mb(); | |
954 | uatomic_dec(&buf->active_readers); | |
852c2936 MD |
955 | } |
956 | ||
957 | /** | |
958 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
959 | * @buf: ring buffer | |
960 | * @consumed: consumed count indicating the position where to read | |
961 | * @produced: produced count, indicates position when to stop reading | |
962 | * | |
963 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
964 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 MD |
965 | */ |
966 | ||
4cfec15c | 967 | int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 968 | unsigned long *consumed, unsigned long *produced, |
38fae1d3 | 969 | struct lttng_ust_shm_handle *handle) |
852c2936 | 970 | { |
1d498196 | 971 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 972 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
973 | unsigned long consumed_cur, write_offset; |
974 | int finalized; | |
975 | ||
14641deb | 976 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
977 | /* |
978 | * Read finalized before counters. | |
979 | */ | |
a6352fd4 MD |
980 | cmm_smp_rmb(); |
981 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 MD |
982 | /* |
983 | * No need to issue a memory barrier between consumed count read and | |
984 | * write offset read, because consumed count can only change | |
985 | * concurrently in overwrite mode, and we keep a sequence counter | |
986 | * identifier derived from the write offset to check we are getting | |
987 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
988 | * "tagged" upon writes, tags are checked upon read). | |
989 | */ | |
990 | write_offset = v_read(config, &buf->offset); | |
991 | ||
992 | /* | |
993 | * Check that we are not about to read the same subbuffer in | |
994 | * which the writer head is. | |
995 | */ | |
996 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
997 | == 0) | |
998 | goto nodata; | |
999 | ||
1000 | *consumed = consumed_cur; | |
1001 | *produced = subbuf_trunc(write_offset, chan); | |
1002 | ||
1003 | return 0; | |
1004 | ||
1005 | nodata: | |
1006 | /* | |
1007 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1008 | * of "raw_spin_is_locked" memory ordering. | |
1009 | */ | |
1010 | if (finalized) | |
1011 | return -ENODATA; | |
852c2936 MD |
1012 | else |
1013 | return -EAGAIN; | |
1014 | } | |
852c2936 MD |
1015 | |
1016 | /** | |
1017 | * lib_ring_buffer_put_snapshot - move consumed counter forward | |
1018 | * @buf: ring buffer | |
1019 | * @consumed_new: new consumed count value | |
1020 | */ | |
4cfec15c | 1021 | void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 1022 | unsigned long consumed_new, |
38fae1d3 | 1023 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1024 | { |
4cfec15c | 1025 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
1d498196 | 1026 | struct channel *chan = shmp(handle, bufb->chan); |
852c2936 MD |
1027 | unsigned long consumed; |
1028 | ||
74d81a6c | 1029 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
852c2936 MD |
1030 | |
1031 | /* | |
1032 | * Only push the consumed value forward. | |
1033 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
1034 | * the writer in flight recorder mode. | |
1035 | */ | |
a6352fd4 | 1036 | consumed = uatomic_read(&buf->consumed); |
852c2936 | 1037 | while ((long) consumed - (long) consumed_new < 0) |
a6352fd4 MD |
1038 | consumed = uatomic_cmpxchg(&buf->consumed, consumed, |
1039 | consumed_new); | |
852c2936 | 1040 | } |
852c2936 MD |
1041 | |
1042 | /** | |
1043 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
1044 | * @buf: ring buffer | |
1045 | * @consumed: consumed count indicating the position where to read | |
1046 | * | |
1047 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
1048 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 | 1049 | */ |
4cfec15c | 1050 | int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 1051 | unsigned long consumed, |
38fae1d3 | 1052 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1053 | { |
1d498196 | 1054 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 1055 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1056 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
1057 | int ret; | |
1058 | int finalized; | |
1059 | ||
1060 | retry: | |
14641deb | 1061 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
1062 | /* |
1063 | * Read finalized before counters. | |
1064 | */ | |
a6352fd4 MD |
1065 | cmm_smp_rmb(); |
1066 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 | 1067 | consumed_idx = subbuf_index(consumed, chan); |
4746ae29 | 1068 | commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb); |
852c2936 MD |
1069 | /* |
1070 | * Make sure we read the commit count before reading the buffer | |
1071 | * data and the write offset. Correct consumed offset ordering | |
1072 | * wrt commit count is insured by the use of cmpxchg to update | |
1073 | * the consumed offset. | |
852c2936 | 1074 | */ |
a6352fd4 MD |
1075 | /* |
1076 | * Local rmb to match the remote wmb to read the commit count | |
1077 | * before the buffer data and the write offset. | |
1078 | */ | |
1079 | cmm_smp_rmb(); | |
852c2936 MD |
1080 | |
1081 | write_offset = v_read(config, &buf->offset); | |
1082 | ||
1083 | /* | |
1084 | * Check that the buffer we are getting is after or at consumed_cur | |
1085 | * position. | |
1086 | */ | |
1087 | if ((long) subbuf_trunc(consumed, chan) | |
1088 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
1089 | goto nodata; | |
1090 | ||
1091 | /* | |
1092 | * Check that the subbuffer we are trying to consume has been | |
1093 | * already fully committed. | |
1094 | */ | |
1095 | if (((commit_count - chan->backend.subbuf_size) | |
1096 | & chan->commit_count_mask) | |
1097 | - (buf_trunc(consumed_cur, chan) | |
1098 | >> chan->backend.num_subbuf_order) | |
1099 | != 0) | |
1100 | goto nodata; | |
1101 | ||
1102 | /* | |
1103 | * Check that we are not about to read the same subbuffer in | |
1104 | * which the writer head is. | |
1105 | */ | |
1106 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
1107 | == 0) | |
1108 | goto nodata; | |
1109 | ||
1110 | /* | |
1111 | * Failure to get the subbuffer causes a busy-loop retry without going | |
1112 | * to a wait queue. These are caused by short-lived race windows where | |
1113 | * the writer is getting access to a subbuffer we were trying to get | |
1114 | * access to. Also checks that the "consumed" buffer count we are | |
1115 | * looking for matches the one contained in the subbuffer id. | |
1116 | */ | |
1117 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
1118 | consumed_idx, buf_trunc_val(consumed, chan), |
1119 | handle); | |
852c2936 MD |
1120 | if (ret) |
1121 | goto retry; | |
1122 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); | |
1123 | ||
1124 | buf->get_subbuf_consumed = consumed; | |
1125 | buf->get_subbuf = 1; | |
1126 | ||
1127 | return 0; | |
1128 | ||
1129 | nodata: | |
1130 | /* | |
1131 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1132 | * of "raw_spin_is_locked" memory ordering. | |
1133 | */ | |
1134 | if (finalized) | |
1135 | return -ENODATA; | |
852c2936 MD |
1136 | else |
1137 | return -EAGAIN; | |
1138 | } | |
852c2936 MD |
1139 | |
1140 | /** | |
1141 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
1142 | * @buf: ring buffer | |
1143 | */ | |
4cfec15c | 1144 | void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 1145 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1146 | { |
4cfec15c | 1147 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
1d498196 | 1148 | struct channel *chan = shmp(handle, bufb->chan); |
4cfec15c | 1149 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1150 | unsigned long read_sb_bindex, consumed_idx, consumed; |
1151 | ||
74d81a6c | 1152 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
852c2936 MD |
1153 | |
1154 | if (!buf->get_subbuf) { | |
1155 | /* | |
1156 | * Reader puts a subbuffer it did not get. | |
1157 | */ | |
1158 | CHAN_WARN_ON(chan, 1); | |
1159 | return; | |
1160 | } | |
1161 | consumed = buf->get_subbuf_consumed; | |
1162 | buf->get_subbuf = 0; | |
1163 | ||
1164 | /* | |
1165 | * Clear the records_unread counter. (overruns counter) | |
1166 | * Can still be non-zero if a file reader simply grabbed the data | |
1167 | * without using iterators. | |
1168 | * Can be below zero if an iterator is used on a snapshot more than | |
1169 | * once. | |
1170 | */ | |
1171 | read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); | |
1172 | v_add(config, v_read(config, | |
4746ae29 | 1173 | &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread), |
852c2936 | 1174 | &bufb->records_read); |
4746ae29 | 1175 | v_set(config, &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread, 0); |
852c2936 MD |
1176 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE |
1177 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
1178 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
1179 | ||
1180 | /* | |
1181 | * Exchange the reader subbuffer with the one we put in its place in the | |
1182 | * writer subbuffer table. Expect the original consumed count. If | |
1183 | * update_read_sb_index fails, this is because the writer updated the | |
1184 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
1185 | * currently have: it has become invalid to try reading this sub-buffer | |
1186 | * consumed count value anyway. | |
1187 | */ | |
1188 | consumed_idx = subbuf_index(consumed, chan); | |
1189 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
1190 | consumed_idx, buf_trunc_val(consumed, chan), |
1191 | handle); | |
852c2936 MD |
1192 | /* |
1193 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
1194 | * if the writer concurrently updated it. | |
1195 | */ | |
1196 | } | |
852c2936 MD |
1197 | |
1198 | /* | |
1199 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
1200 | * position and the writer position. (inclusive) | |
1201 | */ | |
1202 | static | |
4cfec15c | 1203 | void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1204 | struct channel *chan, |
1205 | unsigned long cons_offset, | |
1d498196 | 1206 | int cpu, |
38fae1d3 | 1207 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1208 | { |
4cfec15c | 1209 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1210 | unsigned long cons_idx, commit_count, commit_count_sb; |
1211 | ||
1212 | cons_idx = subbuf_index(cons_offset, chan); | |
4746ae29 MD |
1213 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, cons_idx)->cc); |
1214 | commit_count_sb = v_read(config, &shmp_index(handle, buf->commit_cold, cons_idx)->cc_sb); | |
852c2936 MD |
1215 | |
1216 | if (subbuf_offset(commit_count, chan) != 0) | |
4d3c9523 | 1217 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
1218 | "commit count in subbuffer %lu,\n" |
1219 | "expecting multiples of %lu bytes\n" | |
1220 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
1221 | chan->backend.name, cpu, cons_idx, | |
1222 | chan->backend.subbuf_size, | |
1223 | commit_count, commit_count_sb); | |
1224 | ||
4d3c9523 | 1225 | DBG("ring buffer: %s, cpu %d: %lu bytes committed\n", |
852c2936 MD |
1226 | chan->backend.name, cpu, commit_count); |
1227 | } | |
1228 | ||
1229 | static | |
4cfec15c | 1230 | void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 | 1231 | struct channel *chan, |
1d498196 | 1232 | void *priv, int cpu, |
38fae1d3 | 1233 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1234 | { |
4cfec15c | 1235 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1236 | unsigned long write_offset, cons_offset; |
1237 | ||
852c2936 MD |
1238 | /* |
1239 | * No need to order commit_count, write_offset and cons_offset reads | |
1240 | * because we execute at teardown when no more writer nor reader | |
1241 | * references are left. | |
1242 | */ | |
1243 | write_offset = v_read(config, &buf->offset); | |
a6352fd4 | 1244 | cons_offset = uatomic_read(&buf->consumed); |
852c2936 | 1245 | if (write_offset != cons_offset) |
4d3c9523 | 1246 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
1247 | "non-consumed data\n" |
1248 | " [ %lu bytes written, %lu bytes read ]\n", | |
1249 | chan->backend.name, cpu, write_offset, cons_offset); | |
1250 | ||
a6352fd4 | 1251 | for (cons_offset = uatomic_read(&buf->consumed); |
852c2936 MD |
1252 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), |
1253 | chan) | |
1254 | - cons_offset) > 0; | |
1255 | cons_offset = subbuf_align(cons_offset, chan)) | |
1256 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1d498196 | 1257 | cpu, handle); |
852c2936 MD |
1258 | } |
1259 | ||
1260 | static | |
1261 | void lib_ring_buffer_print_errors(struct channel *chan, | |
009769ca MD |
1262 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
1263 | struct lttng_ust_shm_handle *handle) | |
852c2936 | 1264 | { |
4cfec15c | 1265 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
a3f61e7f | 1266 | void *priv = channel_get_private(chan); |
852c2936 | 1267 | |
a1360615 MD |
1268 | if (!strcmp(chan->backend.name, "relay-metadata-mmap")) { |
1269 | DBG("ring buffer %s: %lu records written, " | |
1270 | "%lu records overrun\n", | |
1271 | chan->backend.name, | |
1272 | v_read(config, &buf->records_count), | |
1273 | v_read(config, &buf->records_overrun)); | |
1274 | } else { | |
1275 | DBG("ring buffer %s, cpu %d: %lu records written, " | |
1276 | "%lu records overrun\n", | |
1277 | chan->backend.name, cpu, | |
1278 | v_read(config, &buf->records_count), | |
1279 | v_read(config, &buf->records_overrun)); | |
1280 | ||
1281 | if (v_read(config, &buf->records_lost_full) | |
1282 | || v_read(config, &buf->records_lost_wrap) | |
1283 | || v_read(config, &buf->records_lost_big)) | |
1284 | DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n" | |
1285 | " [ %lu buffer full, %lu nest buffer wrap-around, " | |
1286 | "%lu event too big ]\n", | |
1287 | chan->backend.name, cpu, | |
1288 | v_read(config, &buf->records_lost_full), | |
1289 | v_read(config, &buf->records_lost_wrap), | |
1290 | v_read(config, &buf->records_lost_big)); | |
1291 | } | |
1d498196 | 1292 | lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu, handle); |
852c2936 MD |
1293 | } |
1294 | ||
1295 | /* | |
1296 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
1297 | * | |
1298 | * Only executed when the buffer is finalized, in SWITCH_FLUSH. | |
1299 | */ | |
1300 | static | |
4cfec15c | 1301 | void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1302 | struct channel *chan, |
1303 | struct switch_offsets *offsets, | |
2fed87ae | 1304 | uint64_t tsc, |
38fae1d3 | 1305 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1306 | { |
4cfec15c | 1307 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1308 | unsigned long oldidx = subbuf_index(offsets->old, chan); |
1309 | unsigned long commit_count; | |
1310 | ||
1d498196 | 1311 | config->cb.buffer_begin(buf, tsc, oldidx, handle); |
852c2936 MD |
1312 | |
1313 | /* | |
1314 | * Order all writes to buffer before the commit count update that will | |
1315 | * determine that the subbuffer is full. | |
1316 | */ | |
a6352fd4 | 1317 | cmm_smp_wmb(); |
852c2936 | 1318 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1319 | &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1320 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 MD |
1321 | /* Check if the written buffer has to be delivered */ |
1322 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
1d498196 | 1323 | commit_count, oldidx, handle); |
852c2936 MD |
1324 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
1325 | offsets->old, commit_count, | |
1d498196 MD |
1326 | config->cb.subbuffer_header_size(), |
1327 | handle); | |
852c2936 MD |
1328 | } |
1329 | ||
1330 | /* | |
1331 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
1332 | * | |
1333 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
1334 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
1335 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
1336 | * subbuffer. | |
1337 | */ | |
1338 | static | |
4cfec15c | 1339 | void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1340 | struct channel *chan, |
1341 | struct switch_offsets *offsets, | |
2fed87ae | 1342 | uint64_t tsc, |
38fae1d3 | 1343 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1344 | { |
4cfec15c | 1345 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1346 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); |
1347 | unsigned long commit_count, padding_size, data_size; | |
1348 | ||
1349 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1350 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1351 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size, |
1352 | handle); | |
852c2936 MD |
1353 | |
1354 | /* | |
1355 | * Order all writes to buffer before the commit count update that will | |
1356 | * determine that the subbuffer is full. | |
1357 | */ | |
a6352fd4 | 1358 | cmm_smp_wmb(); |
4746ae29 MD |
1359 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1360 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 | 1361 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, |
1d498196 | 1362 | commit_count, oldidx, handle); |
852c2936 MD |
1363 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
1364 | offsets->old, commit_count, | |
1d498196 | 1365 | padding_size, handle); |
852c2936 MD |
1366 | } |
1367 | ||
1368 | /* | |
1369 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1370 | * | |
1371 | * This code can be executed unordered : writers may already have written to the | |
1372 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1373 | * that this code is executed before the deliver of this sub-buffer. | |
1374 | */ | |
1375 | static | |
4cfec15c | 1376 | void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1377 | struct channel *chan, |
1378 | struct switch_offsets *offsets, | |
2fed87ae | 1379 | uint64_t tsc, |
38fae1d3 | 1380 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1381 | { |
4cfec15c | 1382 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1383 | unsigned long beginidx = subbuf_index(offsets->begin, chan); |
1384 | unsigned long commit_count; | |
1385 | ||
1d498196 | 1386 | config->cb.buffer_begin(buf, tsc, beginidx, handle); |
852c2936 MD |
1387 | |
1388 | /* | |
1389 | * Order all writes to buffer before the commit count update that will | |
1390 | * determine that the subbuffer is full. | |
1391 | */ | |
a6352fd4 | 1392 | cmm_smp_wmb(); |
852c2936 | 1393 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1394 | &shmp_index(handle, buf->commit_hot, beginidx)->cc); |
1395 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, beginidx)->cc); | |
852c2936 MD |
1396 | /* Check if the written buffer has to be delivered */ |
1397 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
1d498196 | 1398 | commit_count, beginidx, handle); |
852c2936 MD |
1399 | lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx, |
1400 | offsets->begin, commit_count, | |
1d498196 MD |
1401 | config->cb.subbuffer_header_size(), |
1402 | handle); | |
852c2936 MD |
1403 | } |
1404 | ||
1405 | /* | |
1406 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1407 | * | |
1408 | * The only remaining threads could be the ones with pending commits. They will | |
1409 | * have to do the deliver themselves. | |
1410 | */ | |
1411 | static | |
4cfec15c | 1412 | void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 MD |
1413 | struct channel *chan, |
1414 | struct switch_offsets *offsets, | |
2fed87ae | 1415 | uint64_t tsc, |
38fae1d3 | 1416 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1417 | { |
4cfec15c | 1418 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1419 | unsigned long endidx = subbuf_index(offsets->end - 1, chan); |
1420 | unsigned long commit_count, padding_size, data_size; | |
1421 | ||
1422 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; | |
1423 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1424 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size, |
1425 | handle); | |
852c2936 MD |
1426 | |
1427 | /* | |
1428 | * Order all writes to buffer before the commit count update that will | |
1429 | * determine that the subbuffer is full. | |
1430 | */ | |
a6352fd4 | 1431 | cmm_smp_wmb(); |
4746ae29 MD |
1432 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, endidx)->cc); |
1433 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc); | |
852c2936 | 1434 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->end - 1, |
1d498196 | 1435 | commit_count, endidx, handle); |
852c2936 MD |
1436 | lib_ring_buffer_write_commit_counter(config, buf, chan, endidx, |
1437 | offsets->end, commit_count, | |
1d498196 | 1438 | padding_size, handle); |
852c2936 MD |
1439 | } |
1440 | ||
1441 | /* | |
1442 | * Returns : | |
1443 | * 0 if ok | |
1444 | * !0 if execution must be aborted. | |
1445 | */ | |
1446 | static | |
1447 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
4cfec15c | 1448 | struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1449 | struct channel *chan, |
1450 | struct switch_offsets *offsets, | |
2fed87ae | 1451 | uint64_t *tsc) |
852c2936 | 1452 | { |
4cfec15c | 1453 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1454 | unsigned long off; |
1455 | ||
1456 | offsets->begin = v_read(config, &buf->offset); | |
1457 | offsets->old = offsets->begin; | |
1458 | offsets->switch_old_start = 0; | |
1459 | off = subbuf_offset(offsets->begin, chan); | |
1460 | ||
1461 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1462 | ||
1463 | /* | |
1464 | * Ensure we flush the header of an empty subbuffer when doing the | |
1465 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1466 | * total data gathering duration even if there were no records saved | |
1467 | * after the last buffer switch. | |
1468 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1469 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1470 | * subbuffer header as appropriate. | |
1471 | * The next record that reserves space will be responsible for | |
1472 | * populating the following subbuffer header. We choose not to populate | |
1473 | * the next subbuffer header here because we want to be able to use | |
a6352fd4 MD |
1474 | * SWITCH_ACTIVE for periodical buffer flush, which must |
1475 | * guarantee that all the buffer content (records and header | |
1476 | * timestamps) are visible to the reader. This is required for | |
1477 | * quiescence guarantees for the fusion merge. | |
852c2936 MD |
1478 | */ |
1479 | if (mode == SWITCH_FLUSH || off > 0) { | |
b5a3dfa5 | 1480 | if (caa_unlikely(off == 0)) { |
c352d3a4 MD |
1481 | /* |
1482 | * A final flush that encounters an empty | |
1483 | * sub-buffer cannot switch buffer if a | |
1484 | * reader is located within this sub-buffer. | |
1485 | * Anyway, the purpose of final flushing of a | |
1486 | * sub-buffer at offset 0 is to handle the case | |
1487 | * of entirely empty stream. | |
1488 | */ | |
1489 | if (caa_unlikely(subbuf_trunc(offsets->begin, chan) | |
1490 | - subbuf_trunc((unsigned long) | |
1491 | uatomic_read(&buf->consumed), chan) | |
1492 | >= chan->backend.buf_size)) | |
1493 | return -1; | |
852c2936 MD |
1494 | /* |
1495 | * The client does not save any header information. | |
1496 | * Don't switch empty subbuffer on finalize, because it | |
1497 | * is invalid to deliver a completely empty subbuffer. | |
1498 | */ | |
1499 | if (!config->cb.subbuffer_header_size()) | |
1500 | return -1; | |
1501 | /* | |
1502 | * Need to write the subbuffer start header on finalize. | |
1503 | */ | |
1504 | offsets->switch_old_start = 1; | |
1505 | } | |
1506 | offsets->begin = subbuf_align(offsets->begin, chan); | |
1507 | } else | |
1508 | return -1; /* we do not have to switch : buffer is empty */ | |
1509 | /* Note: old points to the next subbuf at offset 0 */ | |
1510 | offsets->end = offsets->begin; | |
1511 | return 0; | |
1512 | } | |
1513 | ||
1514 | /* | |
1515 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
1516 | * called while tracing is active with absolutely no lock held. | |
1517 | * | |
1518 | * Note, however, that as a v_cmpxchg is used for some atomic | |
1519 | * operations, this function must be called from the CPU which owns the buffer | |
1520 | * for a ACTIVE flush. | |
1521 | */ | |
4cfec15c | 1522 | void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode, |
38fae1d3 | 1523 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1524 | { |
1d498196 | 1525 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 1526 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1527 | struct switch_offsets offsets; |
1528 | unsigned long oldidx; | |
2fed87ae | 1529 | uint64_t tsc; |
852c2936 MD |
1530 | |
1531 | offsets.size = 0; | |
1532 | ||
1533 | /* | |
1534 | * Perform retryable operations. | |
1535 | */ | |
1536 | do { | |
1537 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
1538 | &tsc)) | |
1539 | return; /* Switch not needed */ | |
1540 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
1541 | != offsets.old); | |
1542 | ||
1543 | /* | |
1544 | * Atomically update last_tsc. This update races against concurrent | |
1545 | * atomic updates, but the race will always cause supplementary full TSC | |
1546 | * records, never the opposite (missing a full TSC record when it would | |
1547 | * be needed). | |
1548 | */ | |
1549 | save_last_tsc(config, buf, tsc); | |
1550 | ||
1551 | /* | |
1552 | * Push the reader if necessary | |
1553 | */ | |
1554 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
1555 | ||
1556 | oldidx = subbuf_index(offsets.old, chan); | |
1d498196 | 1557 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx, handle); |
852c2936 MD |
1558 | |
1559 | /* | |
1560 | * May need to populate header start on SWITCH_FLUSH. | |
1561 | */ | |
1562 | if (offsets.switch_old_start) { | |
1d498196 | 1563 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc, handle); |
852c2936 MD |
1564 | offsets.old += config->cb.subbuffer_header_size(); |
1565 | } | |
1566 | ||
1567 | /* | |
1568 | * Switch old subbuffer. | |
1569 | */ | |
1d498196 | 1570 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc, handle); |
852c2936 | 1571 | } |
852c2936 MD |
1572 | |
1573 | /* | |
1574 | * Returns : | |
1575 | * 0 if ok | |
1576 | * -ENOSPC if event size is too large for packet. | |
1577 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
1578 | * -EIO if data cannot be written into the buffer for any other reason. | |
1579 | */ | |
1580 | static | |
4cfec15c | 1581 | int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1582 | struct channel *chan, |
1583 | struct switch_offsets *offsets, | |
4cfec15c | 1584 | struct lttng_ust_lib_ring_buffer_ctx *ctx) |
852c2936 | 1585 | { |
4cfec15c | 1586 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
38fae1d3 | 1587 | struct lttng_ust_shm_handle *handle = ctx->handle; |
852c2936 MD |
1588 | unsigned long reserve_commit_diff; |
1589 | ||
1590 | offsets->begin = v_read(config, &buf->offset); | |
1591 | offsets->old = offsets->begin; | |
1592 | offsets->switch_new_start = 0; | |
1593 | offsets->switch_new_end = 0; | |
1594 | offsets->switch_old_end = 0; | |
1595 | offsets->pre_header_padding = 0; | |
1596 | ||
1597 | ctx->tsc = config->cb.ring_buffer_clock_read(chan); | |
1598 | if ((int64_t) ctx->tsc == -EIO) | |
1599 | return -EIO; | |
1600 | ||
1601 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
1602 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
1603 | ||
b5a3dfa5 | 1604 | if (caa_unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) { |
852c2936 MD |
1605 | offsets->switch_new_start = 1; /* For offsets->begin */ |
1606 | } else { | |
1607 | offsets->size = config->cb.record_header_size(config, chan, | |
1608 | offsets->begin, | |
1609 | &offsets->pre_header_padding, | |
1610 | ctx); | |
1611 | offsets->size += | |
1612 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1613 | ctx->largest_align) | |
1614 | + ctx->data_size; | |
b5a3dfa5 | 1615 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) + |
852c2936 MD |
1616 | offsets->size > chan->backend.subbuf_size)) { |
1617 | offsets->switch_old_end = 1; /* For offsets->old */ | |
1618 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1619 | } | |
1620 | } | |
b5a3dfa5 | 1621 | if (caa_unlikely(offsets->switch_new_start)) { |
852c2936 MD |
1622 | unsigned long sb_index; |
1623 | ||
1624 | /* | |
1625 | * We are typically not filling the previous buffer completely. | |
1626 | */ | |
b5a3dfa5 | 1627 | if (caa_likely(offsets->switch_old_end)) |
852c2936 MD |
1628 | offsets->begin = subbuf_align(offsets->begin, chan); |
1629 | offsets->begin = offsets->begin | |
1630 | + config->cb.subbuffer_header_size(); | |
1631 | /* Test new buffer integrity */ | |
1632 | sb_index = subbuf_index(offsets->begin, chan); | |
1633 | reserve_commit_diff = | |
1634 | (buf_trunc(offsets->begin, chan) | |
1635 | >> chan->backend.num_subbuf_order) | |
1636 | - ((unsigned long) v_read(config, | |
4746ae29 | 1637 | &shmp_index(handle, buf->commit_cold, sb_index)->cc_sb) |
852c2936 | 1638 | & chan->commit_count_mask); |
b5a3dfa5 | 1639 | if (caa_likely(reserve_commit_diff == 0)) { |
852c2936 | 1640 | /* Next subbuffer not being written to. */ |
b5a3dfa5 | 1641 | if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE && |
852c2936 MD |
1642 | subbuf_trunc(offsets->begin, chan) |
1643 | - subbuf_trunc((unsigned long) | |
a6352fd4 | 1644 | uatomic_read(&buf->consumed), chan) |
852c2936 | 1645 | >= chan->backend.buf_size)) { |
64493e4f MD |
1646 | unsigned long nr_lost; |
1647 | ||
852c2936 MD |
1648 | /* |
1649 | * We do not overwrite non consumed buffers | |
1650 | * and we are full : record is lost. | |
1651 | */ | |
64493e4f | 1652 | nr_lost = v_read(config, &buf->records_lost_full); |
852c2936 | 1653 | v_inc(config, &buf->records_lost_full); |
64493e4f MD |
1654 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
1655 | DBG("%lu or more records lost in (%s:%d) (buffer full)\n", | |
1656 | nr_lost + 1, chan->backend.name, | |
1657 | buf->backend.cpu); | |
1658 | } | |
852c2936 MD |
1659 | return -ENOBUFS; |
1660 | } else { | |
1661 | /* | |
1662 | * Next subbuffer not being written to, and we | |
1663 | * are either in overwrite mode or the buffer is | |
1664 | * not full. It's safe to write in this new | |
1665 | * subbuffer. | |
1666 | */ | |
1667 | } | |
1668 | } else { | |
64493e4f MD |
1669 | unsigned long nr_lost; |
1670 | ||
852c2936 MD |
1671 | /* |
1672 | * Next subbuffer reserve offset does not match the | |
1673 | * commit offset. Drop record in producer-consumer and | |
1674 | * overwrite mode. Caused by either a writer OOPS or too | |
1675 | * many nested writes over a reserve/commit pair. | |
1676 | */ | |
64493e4f | 1677 | nr_lost = v_read(config, &buf->records_lost_wrap); |
852c2936 | 1678 | v_inc(config, &buf->records_lost_wrap); |
64493e4f MD |
1679 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
1680 | DBG("%lu or more records lost in (%s:%d) (wrap-around)\n", | |
1681 | nr_lost + 1, chan->backend.name, | |
1682 | buf->backend.cpu); | |
1683 | } | |
852c2936 MD |
1684 | return -EIO; |
1685 | } | |
1686 | offsets->size = | |
1687 | config->cb.record_header_size(config, chan, | |
1688 | offsets->begin, | |
1689 | &offsets->pre_header_padding, | |
1690 | ctx); | |
1691 | offsets->size += | |
1692 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1693 | ctx->largest_align) | |
1694 | + ctx->data_size; | |
b5a3dfa5 | 1695 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) |
852c2936 | 1696 | + offsets->size > chan->backend.subbuf_size)) { |
64493e4f MD |
1697 | unsigned long nr_lost; |
1698 | ||
852c2936 MD |
1699 | /* |
1700 | * Record too big for subbuffers, report error, don't | |
1701 | * complete the sub-buffer switch. | |
1702 | */ | |
64493e4f | 1703 | nr_lost = v_read(config, &buf->records_lost_big); |
852c2936 | 1704 | v_inc(config, &buf->records_lost_big); |
64493e4f MD |
1705 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
1706 | DBG("%lu or more records lost in (%s:%d) record size " | |
1707 | " of %zu bytes is too large for buffer\n", | |
1708 | nr_lost + 1, chan->backend.name, | |
1709 | buf->backend.cpu, offsets->size); | |
1710 | } | |
852c2936 MD |
1711 | return -ENOSPC; |
1712 | } else { | |
1713 | /* | |
1714 | * We just made a successful buffer switch and the | |
1715 | * record fits in the new subbuffer. Let's write. | |
1716 | */ | |
1717 | } | |
1718 | } else { | |
1719 | /* | |
1720 | * Record fits in the current buffer and we are not on a switch | |
1721 | * boundary. It's safe to write. | |
1722 | */ | |
1723 | } | |
1724 | offsets->end = offsets->begin + offsets->size; | |
1725 | ||
b5a3dfa5 | 1726 | if (caa_unlikely(subbuf_offset(offsets->end, chan) == 0)) { |
852c2936 MD |
1727 | /* |
1728 | * The offset_end will fall at the very beginning of the next | |
1729 | * subbuffer. | |
1730 | */ | |
1731 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
1732 | } | |
1733 | return 0; | |
1734 | } | |
1735 | ||
1736 | /** | |
1737 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
1738 | * @ctx: ring buffer context. | |
1739 | * | |
1740 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, | |
1741 | * -EIO for other errors, else returns 0. | |
1742 | * It will take care of sub-buffer switching. | |
1743 | */ | |
4cfec15c | 1744 | int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx) |
852c2936 MD |
1745 | { |
1746 | struct channel *chan = ctx->chan; | |
38fae1d3 | 1747 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c MD |
1748 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
1749 | struct lttng_ust_lib_ring_buffer *buf; | |
852c2936 MD |
1750 | struct switch_offsets offsets; |
1751 | int ret; | |
1752 | ||
1753 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1d498196 | 1754 | buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp); |
852c2936 | 1755 | else |
1d498196 | 1756 | buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 MD |
1757 | ctx->buf = buf; |
1758 | ||
1759 | offsets.size = 0; | |
1760 | ||
1761 | do { | |
1762 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, | |
1763 | ctx); | |
b5a3dfa5 | 1764 | if (caa_unlikely(ret)) |
852c2936 | 1765 | return ret; |
b5a3dfa5 | 1766 | } while (caa_unlikely(v_cmpxchg(config, &buf->offset, offsets.old, |
852c2936 MD |
1767 | offsets.end) |
1768 | != offsets.old)); | |
1769 | ||
1770 | /* | |
1771 | * Atomically update last_tsc. This update races against concurrent | |
1772 | * atomic updates, but the race will always cause supplementary full TSC | |
1773 | * records, never the opposite (missing a full TSC record when it would | |
1774 | * be needed). | |
1775 | */ | |
1776 | save_last_tsc(config, buf, ctx->tsc); | |
1777 | ||
1778 | /* | |
1779 | * Push the reader if necessary | |
1780 | */ | |
1781 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
1782 | ||
1783 | /* | |
1784 | * Clear noref flag for this subbuffer. | |
1785 | */ | |
1786 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1d498196 MD |
1787 | subbuf_index(offsets.end - 1, chan), |
1788 | handle); | |
852c2936 MD |
1789 | |
1790 | /* | |
1791 | * Switch old subbuffer if needed. | |
1792 | */ | |
b5a3dfa5 | 1793 | if (caa_unlikely(offsets.switch_old_end)) { |
852c2936 | 1794 | lib_ring_buffer_clear_noref(config, &buf->backend, |
1d498196 MD |
1795 | subbuf_index(offsets.old - 1, chan), |
1796 | handle); | |
1797 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc, handle); | |
852c2936 MD |
1798 | } |
1799 | ||
1800 | /* | |
1801 | * Populate new subbuffer. | |
1802 | */ | |
b5a3dfa5 | 1803 | if (caa_unlikely(offsets.switch_new_start)) |
1d498196 | 1804 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 | 1805 | |
b5a3dfa5 | 1806 | if (caa_unlikely(offsets.switch_new_end)) |
1d498196 | 1807 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 MD |
1808 | |
1809 | ctx->slot_size = offsets.size; | |
1810 | ctx->pre_offset = offsets.begin; | |
1811 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
1812 | return 0; | |
1813 | } | |
f645cfa7 MD |
1814 | |
1815 | /* | |
1816 | * Force a read (imply TLS fixup for dlopen) of TLS variables. | |
1817 | */ | |
1818 | void lttng_fixup_ringbuffer_tls(void) | |
1819 | { | |
8c90a710 | 1820 | asm volatile ("" : : "m" (URCU_TLS(lib_ring_buffer_nesting))); |
f645cfa7 | 1821 | } |
03d2d293 MD |
1822 | |
1823 | void lib_ringbuffer_signal_init(void) | |
1824 | { | |
1825 | sigset_t mask; | |
1826 | int ret; | |
1827 | ||
1828 | /* | |
1829 | * Block signal for entire process, so only our thread processes | |
1830 | * it. | |
1831 | */ | |
1832 | rb_setmask(&mask); | |
1833 | ret = pthread_sigmask(SIG_BLOCK, &mask, NULL); | |
1834 | if (ret) { | |
1835 | errno = ret; | |
1836 | PERROR("pthread_sigmask"); | |
1837 | } | |
1838 | } |