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