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> |
327d9447 | 65 | #include <poll.h> |
35897f8b | 66 | #include <helper.h> |
14641deb | 67 | |
a6352fd4 | 68 | #include "smp.h" |
4318ae1b | 69 | #include <lttng/ringbuffer-config.h> |
2fed87ae | 70 | #include "vatomic.h" |
4931a13e MD |
71 | #include "backend.h" |
72 | #include "frontend.h" | |
a6352fd4 | 73 | #include "shm.h" |
f645cfa7 | 74 | #include "tlsfixup.h" |
bdcf8d82 | 75 | #include "../liblttng-ust/compat.h" /* For ENODATA */ |
852c2936 | 76 | |
431d5cf0 MD |
77 | #ifndef max |
78 | #define max(a, b) ((a) > (b) ? (a) : (b)) | |
79 | #endif | |
80 | ||
64493e4f MD |
81 | /* Print DBG() messages about events lost only every 1048576 hits */ |
82 | #define DBG_PRINT_NR_LOST (1UL << 20) | |
83 | ||
34a91bdb MD |
84 | #define LTTNG_UST_RB_SIG_FLUSH SIGRTMIN |
85 | #define LTTNG_UST_RB_SIG_READ SIGRTMIN + 1 | |
86 | #define LTTNG_UST_RB_SIG_TEARDOWN SIGRTMIN + 2 | |
03d2d293 | 87 | #define CLOCKID CLOCK_MONOTONIC |
dddba398 MD |
88 | #define LTTNG_UST_RING_BUFFER_GET_RETRY 10 |
89 | #define LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS 10 | |
03d2d293 | 90 | |
a9ff648c MD |
91 | /* |
92 | * Non-static to ensure the compiler does not optimize away the xor. | |
93 | */ | |
94 | uint8_t lttng_crash_magic_xor[] = RB_CRASH_DUMP_ABI_MAGIC_XOR; | |
95 | ||
2432c3c9 MD |
96 | /* |
97 | * Use POSIX SHM: shm_open(3) and shm_unlink(3). | |
98 | * close(2) to close the fd returned by shm_open. | |
99 | * shm_unlink releases the shared memory object name. | |
100 | * ftruncate(2) sets the size of the memory object. | |
101 | * mmap/munmap maps the shared memory obj to a virtual address in the | |
102 | * calling proceess (should be done both in libust and consumer). | |
103 | * See shm_overview(7) for details. | |
104 | * Pass file descriptor returned by shm_open(3) to ltt-sessiond through | |
105 | * a UNIX socket. | |
106 | * | |
107 | * Since we don't need to access the object using its name, we can | |
108 | * immediately shm_unlink(3) it, and only keep the handle with its file | |
109 | * descriptor. | |
110 | */ | |
111 | ||
852c2936 MD |
112 | /* |
113 | * Internal structure representing offsets to use at a sub-buffer switch. | |
114 | */ | |
115 | struct switch_offsets { | |
116 | unsigned long begin, end, old; | |
117 | size_t pre_header_padding, size; | |
1ad21f70 MD |
118 | unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1, |
119 | switch_old_end:1; | |
852c2936 MD |
120 | }; |
121 | ||
8c90a710 | 122 | DEFINE_URCU_TLS(unsigned int, lib_ring_buffer_nesting); |
852c2936 | 123 | |
cb7378b3 MD |
124 | /* |
125 | * wakeup_fd_mutex protects wakeup fd use by timer from concurrent | |
126 | * close. | |
127 | */ | |
128 | static pthread_mutex_t wakeup_fd_mutex = PTHREAD_MUTEX_INITIALIZER; | |
129 | ||
852c2936 MD |
130 | static |
131 | void lib_ring_buffer_print_errors(struct channel *chan, | |
009769ca MD |
132 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
133 | struct lttng_ust_shm_handle *handle); | |
852c2936 | 134 | |
03d2d293 MD |
135 | /* |
136 | * Handle timer teardown race wrt memory free of private data by | |
137 | * ring buffer signals are handled by a single thread, which permits | |
138 | * a synchronization point between handling of each signal. | |
64bf51a6 | 139 | * Protected by the lock within the structure. |
03d2d293 MD |
140 | */ |
141 | struct timer_signal_data { | |
142 | pthread_t tid; /* thread id managing signals */ | |
143 | int setup_done; | |
144 | int qs_done; | |
64bf51a6 | 145 | pthread_mutex_t lock; |
03d2d293 MD |
146 | }; |
147 | ||
64bf51a6 MD |
148 | static struct timer_signal_data timer_signal = { |
149 | .tid = 0, | |
150 | .setup_done = 0, | |
151 | .qs_done = 0, | |
152 | .lock = PTHREAD_MUTEX_INITIALIZER, | |
153 | }; | |
03d2d293 | 154 | |
852c2936 MD |
155 | /** |
156 | * lib_ring_buffer_reset - Reset ring buffer to initial values. | |
157 | * @buf: Ring buffer. | |
158 | * | |
159 | * Effectively empty the ring buffer. Should be called when the buffer is not | |
160 | * used for writing. The ring buffer can be opened for reading, but the reader | |
161 | * should not be using the iterator concurrently with reset. The previous | |
162 | * current iterator record is reset. | |
163 | */ | |
4cfec15c | 164 | void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 165 | struct lttng_ust_shm_handle *handle) |
852c2936 | 166 | { |
1d498196 | 167 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 168 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
169 | unsigned int i; |
170 | ||
171 | /* | |
172 | * Reset iterator first. It will put the subbuffer if it currently holds | |
173 | * it. | |
174 | */ | |
852c2936 MD |
175 | v_set(config, &buf->offset, 0); |
176 | for (i = 0; i < chan->backend.num_subbuf; i++) { | |
4746ae29 MD |
177 | v_set(config, &shmp_index(handle, buf->commit_hot, i)->cc, 0); |
178 | v_set(config, &shmp_index(handle, buf->commit_hot, i)->seq, 0); | |
179 | v_set(config, &shmp_index(handle, buf->commit_cold, i)->cc_sb, 0); | |
852c2936 | 180 | } |
a6352fd4 MD |
181 | uatomic_set(&buf->consumed, 0); |
182 | uatomic_set(&buf->record_disabled, 0); | |
852c2936 | 183 | v_set(config, &buf->last_tsc, 0); |
1d498196 | 184 | lib_ring_buffer_backend_reset(&buf->backend, handle); |
852c2936 MD |
185 | /* Don't reset number of active readers */ |
186 | v_set(config, &buf->records_lost_full, 0); | |
187 | v_set(config, &buf->records_lost_wrap, 0); | |
188 | v_set(config, &buf->records_lost_big, 0); | |
189 | v_set(config, &buf->records_count, 0); | |
190 | v_set(config, &buf->records_overrun, 0); | |
191 | buf->finalized = 0; | |
192 | } | |
852c2936 MD |
193 | |
194 | /** | |
195 | * channel_reset - Reset channel to initial values. | |
196 | * @chan: Channel. | |
197 | * | |
198 | * Effectively empty the channel. Should be called when the channel is not used | |
199 | * for writing. The channel can be opened for reading, but the reader should not | |
200 | * be using the iterator concurrently with reset. The previous current iterator | |
201 | * record is reset. | |
202 | */ | |
203 | void channel_reset(struct channel *chan) | |
204 | { | |
205 | /* | |
206 | * Reset iterators first. Will put the subbuffer if held for reading. | |
207 | */ | |
a6352fd4 | 208 | uatomic_set(&chan->record_disabled, 0); |
852c2936 MD |
209 | /* Don't reset commit_count_mask, still valid */ |
210 | channel_backend_reset(&chan->backend); | |
211 | /* Don't reset switch/read timer interval */ | |
212 | /* Don't reset notifiers and notifier enable bits */ | |
213 | /* Don't reset reader reference count */ | |
214 | } | |
852c2936 | 215 | |
a9ff648c MD |
216 | static |
217 | void init_crash_abi(const struct lttng_ust_lib_ring_buffer_config *config, | |
218 | struct lttng_crash_abi *crash_abi, | |
219 | struct lttng_ust_lib_ring_buffer *buf, | |
220 | struct channel_backend *chanb, | |
221 | struct shm_object *shmobj, | |
222 | struct lttng_ust_shm_handle *handle) | |
223 | { | |
224 | int i; | |
225 | ||
226 | for (i = 0; i < RB_CRASH_DUMP_ABI_MAGIC_LEN; i++) | |
227 | crash_abi->magic[i] = lttng_crash_magic_xor[i] ^ 0xFF; | |
228 | crash_abi->mmap_length = shmobj->memory_map_size; | |
229 | crash_abi->endian = RB_CRASH_ENDIAN; | |
230 | crash_abi->major = RB_CRASH_DUMP_ABI_MAJOR; | |
231 | crash_abi->minor = RB_CRASH_DUMP_ABI_MINOR; | |
232 | crash_abi->word_size = sizeof(unsigned long); | |
233 | crash_abi->layout_type = LTTNG_CRASH_TYPE_UST; | |
234 | ||
235 | /* Offset of fields */ | |
236 | crash_abi->offset.prod_offset = | |
237 | (uint32_t) ((char *) &buf->offset - (char *) buf); | |
238 | crash_abi->offset.consumed_offset = | |
239 | (uint32_t) ((char *) &buf->consumed - (char *) buf); | |
240 | crash_abi->offset.commit_hot_array = | |
241 | (uint32_t) ((char *) shmp(handle, buf->commit_hot) - (char *) buf); | |
242 | crash_abi->offset.commit_hot_seq = | |
243 | offsetof(struct commit_counters_hot, seq); | |
244 | crash_abi->offset.buf_wsb_array = | |
245 | (uint32_t) ((char *) shmp(handle, buf->backend.buf_wsb) - (char *) buf); | |
246 | crash_abi->offset.buf_wsb_id = | |
247 | offsetof(struct lttng_ust_lib_ring_buffer_backend_subbuffer, id); | |
248 | crash_abi->offset.sb_array = | |
249 | (uint32_t) ((char *) shmp(handle, buf->backend.array) - (char *) buf); | |
250 | crash_abi->offset.sb_array_shmp_offset = | |
251 | offsetof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp, | |
252 | shmp._ref.offset); | |
253 | crash_abi->offset.sb_backend_p_offset = | |
254 | offsetof(struct lttng_ust_lib_ring_buffer_backend_pages, | |
255 | p._ref.offset); | |
256 | ||
257 | /* Field length */ | |
258 | crash_abi->length.prod_offset = sizeof(buf->offset); | |
259 | crash_abi->length.consumed_offset = sizeof(buf->consumed); | |
260 | crash_abi->length.commit_hot_seq = | |
261 | sizeof(((struct commit_counters_hot *) NULL)->seq); | |
262 | crash_abi->length.buf_wsb_id = | |
263 | sizeof(((struct lttng_ust_lib_ring_buffer_backend_subbuffer *) NULL)->id); | |
264 | crash_abi->length.sb_array_shmp_offset = | |
265 | sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages_shmp *) NULL)->shmp._ref.offset); | |
266 | crash_abi->length.sb_backend_p_offset = | |
267 | sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages *) NULL)->p._ref.offset); | |
268 | ||
269 | /* Array stride */ | |
270 | crash_abi->stride.commit_hot_array = | |
271 | sizeof(struct commit_counters_hot); | |
272 | crash_abi->stride.buf_wsb_array = | |
273 | sizeof(struct lttng_ust_lib_ring_buffer_backend_subbuffer); | |
274 | crash_abi->stride.sb_array = | |
275 | sizeof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp); | |
276 | ||
277 | /* Buffer constants */ | |
278 | crash_abi->buf_size = chanb->buf_size; | |
279 | crash_abi->subbuf_size = chanb->subbuf_size; | |
280 | crash_abi->num_subbuf = chanb->num_subbuf; | |
281 | crash_abi->mode = (uint32_t) chanb->config.mode; | |
282 | ||
283 | if (config->cb.content_size_field) { | |
284 | size_t offset, length; | |
285 | ||
286 | config->cb.content_size_field(config, &offset, &length); | |
287 | crash_abi->offset.content_size = offset; | |
288 | crash_abi->length.content_size = length; | |
289 | } else { | |
290 | crash_abi->offset.content_size = 0; | |
291 | crash_abi->length.content_size = 0; | |
292 | } | |
293 | if (config->cb.packet_size_field) { | |
294 | size_t offset, length; | |
295 | ||
296 | config->cb.packet_size_field(config, &offset, &length); | |
297 | crash_abi->offset.packet_size = offset; | |
298 | crash_abi->length.packet_size = length; | |
299 | } else { | |
300 | crash_abi->offset.packet_size = 0; | |
301 | crash_abi->length.packet_size = 0; | |
302 | } | |
303 | } | |
304 | ||
852c2936 MD |
305 | /* |
306 | * Must be called under cpu hotplug protection. | |
307 | */ | |
4cfec15c | 308 | int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf, |
a6352fd4 | 309 | struct channel_backend *chanb, int cpu, |
38fae1d3 | 310 | struct lttng_ust_shm_handle *handle, |
1d498196 | 311 | struct shm_object *shmobj) |
852c2936 | 312 | { |
4cfec15c | 313 | const struct lttng_ust_lib_ring_buffer_config *config = &chanb->config; |
14641deb | 314 | struct channel *chan = caa_container_of(chanb, struct channel, backend); |
a3f61e7f | 315 | void *priv = channel_get_private(chan); |
852c2936 | 316 | size_t subbuf_header_size; |
2fed87ae | 317 | uint64_t tsc; |
852c2936 MD |
318 | int ret; |
319 | ||
320 | /* Test for cpu hotplug */ | |
321 | if (buf->backend.allocated) | |
322 | return 0; | |
323 | ||
1d498196 MD |
324 | align_shm(shmobj, __alignof__(struct commit_counters_hot)); |
325 | set_shmp(buf->commit_hot, | |
326 | zalloc_shm(shmobj, | |
327 | sizeof(struct commit_counters_hot) * chan->backend.num_subbuf)); | |
328 | if (!shmp(handle, buf->commit_hot)) { | |
a9ff648c | 329 | return -ENOMEM; |
852c2936 MD |
330 | } |
331 | ||
1d498196 MD |
332 | align_shm(shmobj, __alignof__(struct commit_counters_cold)); |
333 | set_shmp(buf->commit_cold, | |
334 | zalloc_shm(shmobj, | |
335 | sizeof(struct commit_counters_cold) * chan->backend.num_subbuf)); | |
336 | if (!shmp(handle, buf->commit_cold)) { | |
852c2936 MD |
337 | ret = -ENOMEM; |
338 | goto free_commit; | |
339 | } | |
340 | ||
a9ff648c MD |
341 | ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, |
342 | cpu, handle, shmobj); | |
343 | if (ret) { | |
344 | goto free_init; | |
345 | } | |
346 | ||
852c2936 MD |
347 | /* |
348 | * Write the subbuffer header for first subbuffer so we know the total | |
349 | * duration of data gathering. | |
350 | */ | |
351 | subbuf_header_size = config->cb.subbuffer_header_size(); | |
352 | v_set(config, &buf->offset, subbuf_header_size); | |
4746ae29 | 353 | subbuffer_id_clear_noref(config, &shmp_index(handle, buf->backend.buf_wsb, 0)->id); |
1d498196 MD |
354 | tsc = config->cb.ring_buffer_clock_read(shmp(handle, buf->backend.chan)); |
355 | config->cb.buffer_begin(buf, tsc, 0, handle); | |
4746ae29 | 356 | v_add(config, subbuf_header_size, &shmp_index(handle, buf->commit_hot, 0)->cc); |
a9ff648c | 357 | v_add(config, subbuf_header_size, &shmp_index(handle, buf->commit_hot, 0)->seq); |
852c2936 MD |
358 | |
359 | if (config->cb.buffer_create) { | |
1d498196 | 360 | ret = config->cb.buffer_create(buf, priv, cpu, chanb->name, handle); |
852c2936 | 361 | if (ret) |
a9ff648c | 362 | goto free_chanbuf; |
852c2936 | 363 | } |
a9ff648c MD |
364 | |
365 | init_crash_abi(config, &buf->crash_abi, buf, chanb, shmobj, handle); | |
366 | ||
852c2936 | 367 | buf->backend.allocated = 1; |
852c2936 MD |
368 | return 0; |
369 | ||
370 | /* Error handling */ | |
371 | free_init: | |
a6352fd4 | 372 | /* commit_cold will be freed by shm teardown */ |
852c2936 | 373 | free_commit: |
a6352fd4 | 374 | /* commit_hot will be freed by shm teardown */ |
852c2936 | 375 | free_chanbuf: |
852c2936 MD |
376 | return ret; |
377 | } | |
378 | ||
03d2d293 MD |
379 | static |
380 | void lib_ring_buffer_channel_switch_timer(int sig, siginfo_t *si, void *uc) | |
852c2936 | 381 | { |
03d2d293 MD |
382 | const struct lttng_ust_lib_ring_buffer_config *config; |
383 | struct lttng_ust_shm_handle *handle; | |
384 | struct channel *chan; | |
385 | int cpu; | |
386 | ||
387 | assert(CMM_LOAD_SHARED(timer_signal.tid) == pthread_self()); | |
388 | ||
389 | chan = si->si_value.sival_ptr; | |
390 | handle = chan->handle; | |
391 | config = &chan->backend.config; | |
852c2936 | 392 | |
34a91bdb | 393 | DBG("Switch timer for channel %p\n", chan); |
03d2d293 | 394 | |
4dcd7e80 MD |
395 | /* |
396 | * Only flush buffers periodically if readers are active. | |
397 | */ | |
cb7378b3 | 398 | pthread_mutex_lock(&wakeup_fd_mutex); |
03d2d293 MD |
399 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
400 | for_each_possible_cpu(cpu) { | |
401 | struct lttng_ust_lib_ring_buffer *buf = | |
402 | shmp(handle, chan->backend.buf[cpu].shmp); | |
4dcd7e80 MD |
403 | if (uatomic_read(&buf->active_readers)) |
404 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, | |
405 | chan->handle); | |
03d2d293 MD |
406 | } |
407 | } else { | |
408 | struct lttng_ust_lib_ring_buffer *buf = | |
409 | shmp(handle, chan->backend.buf[0].shmp); | |
410 | ||
34a91bdb MD |
411 | if (uatomic_read(&buf->active_readers)) |
412 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, | |
413 | chan->handle); | |
414 | } | |
415 | pthread_mutex_unlock(&wakeup_fd_mutex); | |
416 | return; | |
417 | } | |
418 | ||
419 | static | |
420 | void lib_ring_buffer_channel_do_read(struct channel *chan) | |
421 | { | |
422 | const struct lttng_ust_lib_ring_buffer_config *config; | |
423 | struct lttng_ust_shm_handle *handle; | |
424 | int cpu; | |
425 | ||
426 | handle = chan->handle; | |
427 | config = &chan->backend.config; | |
428 | ||
429 | /* | |
430 | * Only flush buffers periodically if readers are active. | |
431 | */ | |
432 | pthread_mutex_lock(&wakeup_fd_mutex); | |
433 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
434 | for_each_possible_cpu(cpu) { | |
435 | struct lttng_ust_lib_ring_buffer *buf = | |
436 | shmp(handle, chan->backend.buf[cpu].shmp); | |
437 | ||
438 | if (uatomic_read(&buf->active_readers) | |
439 | && lib_ring_buffer_poll_deliver(config, buf, | |
440 | chan, handle)) { | |
441 | lib_ring_buffer_wakeup(buf, handle); | |
442 | } | |
443 | } | |
444 | } else { | |
445 | struct lttng_ust_lib_ring_buffer *buf = | |
446 | shmp(handle, chan->backend.buf[0].shmp); | |
447 | ||
448 | if (uatomic_read(&buf->active_readers) | |
449 | && lib_ring_buffer_poll_deliver(config, buf, | |
450 | chan, handle)) { | |
451 | lib_ring_buffer_wakeup(buf, handle); | |
452 | } | |
03d2d293 | 453 | } |
cb7378b3 | 454 | pthread_mutex_unlock(&wakeup_fd_mutex); |
34a91bdb MD |
455 | } |
456 | ||
457 | static | |
458 | void lib_ring_buffer_channel_read_timer(int sig, siginfo_t *si, void *uc) | |
459 | { | |
460 | struct channel *chan; | |
461 | ||
462 | assert(CMM_LOAD_SHARED(timer_signal.tid) == pthread_self()); | |
463 | chan = si->si_value.sival_ptr; | |
464 | DBG("Read timer for channel %p\n", chan); | |
465 | lib_ring_buffer_channel_do_read(chan); | |
03d2d293 MD |
466 | return; |
467 | } | |
468 | ||
469 | static | |
470 | void rb_setmask(sigset_t *mask) | |
471 | { | |
472 | int ret; | |
473 | ||
474 | ret = sigemptyset(mask); | |
475 | if (ret) { | |
476 | PERROR("sigemptyset"); | |
477 | } | |
34a91bdb MD |
478 | ret = sigaddset(mask, LTTNG_UST_RB_SIG_FLUSH); |
479 | if (ret) { | |
480 | PERROR("sigaddset"); | |
481 | } | |
482 | ret = sigaddset(mask, LTTNG_UST_RB_SIG_READ); | |
03d2d293 MD |
483 | if (ret) { |
484 | PERROR("sigaddset"); | |
485 | } | |
486 | ret = sigaddset(mask, LTTNG_UST_RB_SIG_TEARDOWN); | |
487 | if (ret) { | |
488 | PERROR("sigaddset"); | |
489 | } | |
490 | } | |
491 | ||
492 | static | |
493 | void *sig_thread(void *arg) | |
494 | { | |
495 | sigset_t mask; | |
496 | siginfo_t info; | |
497 | int signr; | |
498 | ||
499 | /* Only self thread will receive signal mask. */ | |
500 | rb_setmask(&mask); | |
501 | CMM_STORE_SHARED(timer_signal.tid, pthread_self()); | |
502 | ||
503 | for (;;) { | |
504 | signr = sigwaitinfo(&mask, &info); | |
505 | if (signr == -1) { | |
34a91bdb MD |
506 | if (errno != EINTR) |
507 | PERROR("sigwaitinfo"); | |
03d2d293 MD |
508 | continue; |
509 | } | |
34a91bdb | 510 | if (signr == LTTNG_UST_RB_SIG_FLUSH) { |
03d2d293 MD |
511 | lib_ring_buffer_channel_switch_timer(info.si_signo, |
512 | &info, NULL); | |
34a91bdb MD |
513 | } else if (signr == LTTNG_UST_RB_SIG_READ) { |
514 | lib_ring_buffer_channel_read_timer(info.si_signo, | |
515 | &info, NULL); | |
03d2d293 MD |
516 | } else if (signr == LTTNG_UST_RB_SIG_TEARDOWN) { |
517 | cmm_smp_mb(); | |
518 | CMM_STORE_SHARED(timer_signal.qs_done, 1); | |
519 | cmm_smp_mb(); | |
520 | } else { | |
521 | ERR("Unexptected signal %d\n", info.si_signo); | |
522 | } | |
523 | } | |
524 | return NULL; | |
525 | } | |
526 | ||
527 | /* | |
03d2d293 MD |
528 | * Ensure only a single thread listens on the timer signal. |
529 | */ | |
530 | static | |
531 | void lib_ring_buffer_setup_timer_thread(void) | |
532 | { | |
533 | pthread_t thread; | |
534 | int ret; | |
535 | ||
64bf51a6 | 536 | pthread_mutex_lock(&timer_signal.lock); |
03d2d293 | 537 | if (timer_signal.setup_done) |
64bf51a6 | 538 | goto end; |
03d2d293 MD |
539 | |
540 | ret = pthread_create(&thread, NULL, &sig_thread, NULL); | |
541 | if (ret) { | |
542 | errno = ret; | |
543 | PERROR("pthread_create"); | |
544 | } | |
545 | ret = pthread_detach(thread); | |
546 | if (ret) { | |
547 | errno = ret; | |
548 | PERROR("pthread_detach"); | |
549 | } | |
550 | timer_signal.setup_done = 1; | |
64bf51a6 MD |
551 | end: |
552 | pthread_mutex_unlock(&timer_signal.lock); | |
852c2936 MD |
553 | } |
554 | ||
03d2d293 | 555 | /* |
64bf51a6 | 556 | * Wait for signal-handling thread quiescent state. |
03d2d293 | 557 | */ |
64bf51a6 MD |
558 | static |
559 | void lib_ring_buffer_wait_signal_thread_qs(unsigned int signr) | |
560 | { | |
561 | sigset_t pending_set; | |
562 | int ret; | |
563 | ||
564 | /* | |
565 | * We need to be the only thread interacting with the thread | |
566 | * that manages signals for teardown synchronization. | |
567 | */ | |
568 | pthread_mutex_lock(&timer_signal.lock); | |
569 | ||
570 | /* | |
571 | * Ensure we don't have any signal queued for this channel. | |
572 | */ | |
573 | for (;;) { | |
574 | ret = sigemptyset(&pending_set); | |
575 | if (ret == -1) { | |
576 | PERROR("sigemptyset"); | |
577 | } | |
578 | ret = sigpending(&pending_set); | |
579 | if (ret == -1) { | |
580 | PERROR("sigpending"); | |
581 | } | |
582 | if (!sigismember(&pending_set, signr)) | |
583 | break; | |
584 | caa_cpu_relax(); | |
585 | } | |
586 | ||
587 | /* | |
588 | * From this point, no new signal handler will be fired that | |
589 | * would try to access "chan". However, we still need to wait | |
590 | * for any currently executing handler to complete. | |
591 | */ | |
592 | cmm_smp_mb(); | |
593 | CMM_STORE_SHARED(timer_signal.qs_done, 0); | |
594 | cmm_smp_mb(); | |
595 | ||
596 | /* | |
597 | * Kill with LTTNG_UST_RB_SIG_TEARDOWN, so signal management | |
598 | * thread wakes up. | |
599 | */ | |
600 | kill(getpid(), LTTNG_UST_RB_SIG_TEARDOWN); | |
601 | ||
602 | while (!CMM_LOAD_SHARED(timer_signal.qs_done)) | |
603 | caa_cpu_relax(); | |
604 | cmm_smp_mb(); | |
605 | ||
606 | pthread_mutex_unlock(&timer_signal.lock); | |
607 | } | |
608 | ||
03d2d293 MD |
609 | static |
610 | void lib_ring_buffer_channel_switch_timer_start(struct channel *chan) | |
852c2936 | 611 | { |
03d2d293 MD |
612 | struct sigevent sev; |
613 | struct itimerspec its; | |
614 | int ret; | |
852c2936 | 615 | |
03d2d293 | 616 | if (!chan->switch_timer_interval || chan->switch_timer_enabled) |
852c2936 MD |
617 | return; |
618 | ||
03d2d293 MD |
619 | chan->switch_timer_enabled = 1; |
620 | ||
621 | lib_ring_buffer_setup_timer_thread(); | |
622 | ||
623 | sev.sigev_notify = SIGEV_SIGNAL; | |
34a91bdb | 624 | sev.sigev_signo = LTTNG_UST_RB_SIG_FLUSH; |
03d2d293 MD |
625 | sev.sigev_value.sival_ptr = chan; |
626 | ret = timer_create(CLOCKID, &sev, &chan->switch_timer); | |
627 | if (ret == -1) { | |
628 | PERROR("timer_create"); | |
629 | } | |
630 | ||
631 | its.it_value.tv_sec = chan->switch_timer_interval / 1000000; | |
632 | its.it_value.tv_nsec = chan->switch_timer_interval % 1000000; | |
633 | its.it_interval.tv_sec = its.it_value.tv_sec; | |
634 | its.it_interval.tv_nsec = its.it_value.tv_nsec; | |
635 | ||
636 | ret = timer_settime(chan->switch_timer, 0, &its, NULL); | |
637 | if (ret == -1) { | |
638 | PERROR("timer_settime"); | |
639 | } | |
640 | } | |
641 | ||
03d2d293 MD |
642 | static |
643 | void lib_ring_buffer_channel_switch_timer_stop(struct channel *chan) | |
644 | { | |
34a91bdb | 645 | int ret; |
03d2d293 MD |
646 | |
647 | if (!chan->switch_timer_interval || !chan->switch_timer_enabled) | |
648 | return; | |
649 | ||
650 | ret = timer_delete(chan->switch_timer); | |
651 | if (ret == -1) { | |
652 | PERROR("timer_delete"); | |
653 | } | |
654 | ||
64bf51a6 | 655 | lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_FLUSH); |
03d2d293 MD |
656 | |
657 | chan->switch_timer = 0; | |
658 | chan->switch_timer_enabled = 0; | |
852c2936 MD |
659 | } |
660 | ||
34a91bdb MD |
661 | static |
662 | void lib_ring_buffer_channel_read_timer_start(struct channel *chan) | |
852c2936 | 663 | { |
4cfec15c | 664 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
34a91bdb MD |
665 | struct sigevent sev; |
666 | struct itimerspec its; | |
667 | int ret; | |
852c2936 | 668 | |
34a91bdb MD |
669 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER |
670 | || !chan->read_timer_interval || chan->read_timer_enabled) | |
671 | return; | |
852c2936 | 672 | |
34a91bdb | 673 | chan->read_timer_enabled = 1; |
852c2936 | 674 | |
34a91bdb | 675 | lib_ring_buffer_setup_timer_thread(); |
852c2936 | 676 | |
34a91bdb MD |
677 | sev.sigev_notify = SIGEV_SIGNAL; |
678 | sev.sigev_signo = LTTNG_UST_RB_SIG_READ; | |
679 | sev.sigev_value.sival_ptr = chan; | |
680 | ret = timer_create(CLOCKID, &sev, &chan->read_timer); | |
681 | if (ret == -1) { | |
682 | PERROR("timer_create"); | |
683 | } | |
852c2936 | 684 | |
34a91bdb MD |
685 | its.it_value.tv_sec = chan->read_timer_interval / 1000000; |
686 | its.it_value.tv_nsec = chan->read_timer_interval % 1000000; | |
687 | its.it_interval.tv_sec = its.it_value.tv_sec; | |
688 | its.it_interval.tv_nsec = its.it_value.tv_nsec; | |
852c2936 | 689 | |
34a91bdb MD |
690 | ret = timer_settime(chan->read_timer, 0, &its, NULL); |
691 | if (ret == -1) { | |
692 | PERROR("timer_settime"); | |
693 | } | |
852c2936 MD |
694 | } |
695 | ||
34a91bdb MD |
696 | static |
697 | void lib_ring_buffer_channel_read_timer_stop(struct channel *chan) | |
852c2936 | 698 | { |
4cfec15c | 699 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
34a91bdb | 700 | int ret; |
852c2936 MD |
701 | |
702 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
34a91bdb | 703 | || !chan->read_timer_interval || !chan->read_timer_enabled) |
852c2936 MD |
704 | return; |
705 | ||
34a91bdb MD |
706 | ret = timer_delete(chan->read_timer); |
707 | if (ret == -1) { | |
708 | PERROR("timer_delete"); | |
709 | } | |
710 | ||
852c2936 MD |
711 | /* |
712 | * do one more check to catch data that has been written in the last | |
713 | * timer period. | |
714 | */ | |
34a91bdb MD |
715 | lib_ring_buffer_channel_do_read(chan); |
716 | ||
64bf51a6 | 717 | lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_READ); |
34a91bdb MD |
718 | |
719 | chan->read_timer = 0; | |
720 | chan->read_timer_enabled = 0; | |
852c2936 MD |
721 | } |
722 | ||
1d498196 | 723 | static void channel_unregister_notifiers(struct channel *chan, |
38fae1d3 | 724 | struct lttng_ust_shm_handle *handle) |
852c2936 | 725 | { |
03d2d293 | 726 | lib_ring_buffer_channel_switch_timer_stop(chan); |
34a91bdb | 727 | lib_ring_buffer_channel_read_timer_stop(chan); |
852c2936 MD |
728 | } |
729 | ||
009769ca MD |
730 | static void channel_print_errors(struct channel *chan, |
731 | struct lttng_ust_shm_handle *handle) | |
732 | { | |
733 | const struct lttng_ust_lib_ring_buffer_config *config = | |
734 | &chan->backend.config; | |
735 | int cpu; | |
736 | ||
737 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
738 | for_each_possible_cpu(cpu) { | |
739 | struct lttng_ust_lib_ring_buffer *buf = | |
740 | shmp(handle, chan->backend.buf[cpu].shmp); | |
741 | lib_ring_buffer_print_errors(chan, buf, cpu, handle); | |
742 | } | |
743 | } else { | |
744 | struct lttng_ust_lib_ring_buffer *buf = | |
745 | shmp(handle, chan->backend.buf[0].shmp); | |
746 | ||
747 | lib_ring_buffer_print_errors(chan, buf, -1, handle); | |
748 | } | |
749 | } | |
750 | ||
751 | static void channel_free(struct channel *chan, | |
752 | struct lttng_ust_shm_handle *handle) | |
852c2936 | 753 | { |
74d81a6c | 754 | channel_backend_free(&chan->backend, handle); |
431d5cf0 | 755 | /* chan is freed by shm teardown */ |
1d498196 MD |
756 | shm_object_table_destroy(handle->table); |
757 | free(handle); | |
852c2936 MD |
758 | } |
759 | ||
760 | /** | |
761 | * channel_create - Create channel. | |
762 | * @config: ring buffer instance configuration | |
763 | * @name: name of the channel | |
a3f61e7f MD |
764 | * @priv_data: ring buffer client private data area pointer (output) |
765 | * @priv_data_size: length, in bytes, of the private data area. | |
d028eddb | 766 | * @priv_data_init: initialization data for private data. |
852c2936 MD |
767 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous |
768 | * address mapping. It is used only by RING_BUFFER_STATIC | |
769 | * configuration. It can be set to NULL for other backends. | |
770 | * @subbuf_size: subbuffer size | |
771 | * @num_subbuf: number of subbuffers | |
772 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
773 | * padding to let readers get those sub-buffers. | |
774 | * Used for live streaming. | |
775 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
a9ff648c | 776 | * @shm_path: Shared memory files path. |
852c2936 MD |
777 | * |
778 | * Holds cpu hotplug. | |
779 | * Returns NULL on failure. | |
780 | */ | |
4cfec15c | 781 | struct lttng_ust_shm_handle *channel_create(const struct lttng_ust_lib_ring_buffer_config *config, |
a3f61e7f MD |
782 | const char *name, |
783 | void **priv_data, | |
784 | size_t priv_data_align, | |
785 | size_t priv_data_size, | |
d028eddb | 786 | void *priv_data_init, |
a3f61e7f | 787 | void *buf_addr, size_t subbuf_size, |
852c2936 | 788 | size_t num_subbuf, unsigned int switch_timer_interval, |
a9ff648c MD |
789 | unsigned int read_timer_interval, |
790 | const char *shm_path) | |
852c2936 | 791 | { |
24622edc | 792 | int ret; |
a3f61e7f | 793 | size_t shmsize, chansize; |
852c2936 | 794 | struct channel *chan; |
38fae1d3 | 795 | struct lttng_ust_shm_handle *handle; |
1d498196 | 796 | struct shm_object *shmobj; |
74d81a6c MD |
797 | unsigned int nr_streams; |
798 | ||
799 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
800 | nr_streams = num_possible_cpus(); | |
801 | else | |
802 | nr_streams = 1; | |
852c2936 MD |
803 | |
804 | if (lib_ring_buffer_check_config(config, switch_timer_interval, | |
805 | read_timer_interval)) | |
806 | return NULL; | |
807 | ||
38fae1d3 | 808 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
431d5cf0 MD |
809 | if (!handle) |
810 | return NULL; | |
811 | ||
1d498196 MD |
812 | /* Allocate table for channel + per-cpu buffers */ |
813 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
814 | if (!handle->table) | |
815 | goto error_table_alloc; | |
852c2936 | 816 | |
1d498196 MD |
817 | /* Calculate the shm allocation layout */ |
818 | shmsize = sizeof(struct channel); | |
c1fca457 | 819 | shmsize += offset_align(shmsize, __alignof__(struct lttng_ust_lib_ring_buffer_shmp)); |
74d81a6c | 820 | shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp) * nr_streams; |
a3f61e7f | 821 | chansize = shmsize; |
74d81a6c MD |
822 | if (priv_data_align) |
823 | shmsize += offset_align(shmsize, priv_data_align); | |
a3f61e7f | 824 | shmsize += priv_data_size; |
a6352fd4 | 825 | |
74d81a6c | 826 | /* Allocate normal memory for channel (not shared) */ |
a9ff648c MD |
827 | shmobj = shm_object_table_alloc(handle->table, shmsize, SHM_OBJECT_MEM, |
828 | NULL); | |
b5a14697 MD |
829 | if (!shmobj) |
830 | goto error_append; | |
57773204 | 831 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
a3f61e7f | 832 | set_shmp(handle->chan, zalloc_shm(shmobj, chansize)); |
57773204 MD |
833 | assert(handle->chan._ref.index == 0); |
834 | assert(handle->chan._ref.offset == 0); | |
1d498196 | 835 | chan = shmp(handle, handle->chan); |
a6352fd4 | 836 | if (!chan) |
1d498196 | 837 | goto error_append; |
74d81a6c | 838 | chan->nr_streams = nr_streams; |
a6352fd4 | 839 | |
a3f61e7f MD |
840 | /* space for private data */ |
841 | if (priv_data_size) { | |
842 | DECLARE_SHMP(void, priv_data_alloc); | |
843 | ||
844 | align_shm(shmobj, priv_data_align); | |
845 | chan->priv_data_offset = shmobj->allocated_len; | |
846 | set_shmp(priv_data_alloc, zalloc_shm(shmobj, priv_data_size)); | |
847 | if (!shmp(handle, priv_data_alloc)) | |
848 | goto error_append; | |
849 | *priv_data = channel_get_private(chan); | |
d028eddb | 850 | memcpy(*priv_data, priv_data_init, priv_data_size); |
a3f61e7f MD |
851 | } else { |
852 | chan->priv_data_offset = -1; | |
74d81a6c MD |
853 | if (priv_data) |
854 | *priv_data = NULL; | |
a3f61e7f MD |
855 | } |
856 | ||
857 | ret = channel_backend_init(&chan->backend, name, config, | |
a9ff648c MD |
858 | subbuf_size, num_subbuf, handle, |
859 | shm_path); | |
852c2936 | 860 | if (ret) |
1d498196 | 861 | goto error_backend_init; |
852c2936 | 862 | |
03d2d293 | 863 | chan->handle = handle; |
852c2936 | 864 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); |
852c2936 | 865 | |
34a91bdb MD |
866 | chan->switch_timer_interval = switch_timer_interval; |
867 | chan->read_timer_interval = read_timer_interval; | |
03d2d293 | 868 | lib_ring_buffer_channel_switch_timer_start(chan); |
34a91bdb | 869 | lib_ring_buffer_channel_read_timer_start(chan); |
852c2936 | 870 | |
431d5cf0 | 871 | return handle; |
852c2936 | 872 | |
1d498196 MD |
873 | error_backend_init: |
874 | error_append: | |
875 | shm_object_table_destroy(handle->table); | |
876 | error_table_alloc: | |
431d5cf0 | 877 | free(handle); |
852c2936 MD |
878 | return NULL; |
879 | } | |
852c2936 | 880 | |
74d81a6c | 881 | struct lttng_ust_shm_handle *channel_handle_create(void *data, |
ff0f5728 MD |
882 | uint64_t memory_map_size, |
883 | int wakeup_fd) | |
193183fb | 884 | { |
38fae1d3 | 885 | struct lttng_ust_shm_handle *handle; |
193183fb MD |
886 | struct shm_object *object; |
887 | ||
38fae1d3 | 888 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
193183fb MD |
889 | if (!handle) |
890 | return NULL; | |
891 | ||
892 | /* Allocate table for channel + per-cpu buffers */ | |
893 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
894 | if (!handle->table) | |
895 | goto error_table_alloc; | |
896 | /* Add channel object */ | |
74d81a6c | 897 | object = shm_object_table_append_mem(handle->table, data, |
ff0f5728 | 898 | memory_map_size, wakeup_fd); |
193183fb MD |
899 | if (!object) |
900 | goto error_table_object; | |
57773204 MD |
901 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
902 | handle->chan._ref.index = 0; | |
903 | handle->chan._ref.offset = 0; | |
193183fb MD |
904 | return handle; |
905 | ||
906 | error_table_object: | |
907 | shm_object_table_destroy(handle->table); | |
908 | error_table_alloc: | |
909 | free(handle); | |
910 | return NULL; | |
911 | } | |
912 | ||
38fae1d3 | 913 | int channel_handle_add_stream(struct lttng_ust_shm_handle *handle, |
74d81a6c MD |
914 | int shm_fd, int wakeup_fd, uint32_t stream_nr, |
915 | uint64_t memory_map_size) | |
193183fb MD |
916 | { |
917 | struct shm_object *object; | |
918 | ||
919 | /* Add stream object */ | |
74d81a6c MD |
920 | object = shm_object_table_append_shm(handle->table, |
921 | shm_fd, wakeup_fd, stream_nr, | |
922 | memory_map_size); | |
193183fb | 923 | if (!object) |
74d81a6c | 924 | return -EINVAL; |
193183fb MD |
925 | return 0; |
926 | } | |
927 | ||
74d81a6c MD |
928 | unsigned int channel_handle_get_nr_streams(struct lttng_ust_shm_handle *handle) |
929 | { | |
930 | assert(handle->table); | |
931 | return handle->table->allocated_len - 1; | |
932 | } | |
933 | ||
852c2936 | 934 | static |
74d81a6c | 935 | void channel_release(struct channel *chan, struct lttng_ust_shm_handle *handle) |
852c2936 | 936 | { |
74d81a6c | 937 | channel_free(chan, handle); |
852c2936 MD |
938 | } |
939 | ||
940 | /** | |
941 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
942 | * @chan: channel to destroy | |
943 | * | |
944 | * Holds cpu hotplug. | |
431d5cf0 MD |
945 | * Call "destroy" callback, finalize channels, decrement the channel |
946 | * reference count. Note that when readers have completed data | |
947 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
a3f61e7f | 948 | * They should release their handle at that point. |
852c2936 | 949 | */ |
a3f61e7f | 950 | void channel_destroy(struct channel *chan, struct lttng_ust_shm_handle *handle, |
74d81a6c | 951 | int consumer) |
852c2936 | 952 | { |
74d81a6c MD |
953 | if (consumer) { |
954 | /* | |
955 | * Note: the consumer takes care of finalizing and | |
956 | * switching the buffers. | |
957 | */ | |
958 | channel_unregister_notifiers(chan, handle); | |
3d0ef9f6 MD |
959 | /* |
960 | * The consumer prints errors. | |
961 | */ | |
962 | channel_print_errors(chan, handle); | |
824f40b8 MD |
963 | } |
964 | ||
431d5cf0 MD |
965 | /* |
966 | * sessiond/consumer are keeping a reference on the shm file | |
967 | * descriptor directly. No need to refcount. | |
968 | */ | |
74d81a6c | 969 | channel_release(chan, handle); |
a3f61e7f | 970 | return; |
852c2936 | 971 | } |
852c2936 | 972 | |
4cfec15c MD |
973 | struct lttng_ust_lib_ring_buffer *channel_get_ring_buffer( |
974 | const struct lttng_ust_lib_ring_buffer_config *config, | |
1d498196 | 975 | struct channel *chan, int cpu, |
38fae1d3 | 976 | struct lttng_ust_shm_handle *handle, |
74d81a6c MD |
977 | int *shm_fd, int *wait_fd, |
978 | int *wakeup_fd, | |
979 | uint64_t *memory_map_size) | |
852c2936 | 980 | { |
381c0f1e MD |
981 | struct shm_ref *ref; |
982 | ||
983 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
74d81a6c | 984 | cpu = 0; |
381c0f1e | 985 | } else { |
e095d803 MD |
986 | if (cpu >= num_possible_cpus()) |
987 | return NULL; | |
381c0f1e | 988 | } |
74d81a6c MD |
989 | ref = &chan->backend.buf[cpu].shmp._ref; |
990 | *shm_fd = shm_get_shm_fd(handle, ref); | |
991 | *wait_fd = shm_get_wait_fd(handle, ref); | |
992 | *wakeup_fd = shm_get_wakeup_fd(handle, ref); | |
993 | if (shm_get_shm_size(handle, ref, memory_map_size)) | |
994 | return NULL; | |
995 | return shmp(handle, chan->backend.buf[cpu].shmp); | |
852c2936 | 996 | } |
852c2936 | 997 | |
ff0f5728 MD |
998 | int ring_buffer_channel_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config *config, |
999 | struct channel *chan, | |
1000 | struct lttng_ust_shm_handle *handle) | |
1001 | { | |
1002 | struct shm_ref *ref; | |
1003 | ||
1004 | ref = &handle->chan._ref; | |
1005 | return shm_close_wait_fd(handle, ref); | |
1006 | } | |
1007 | ||
1008 | int ring_buffer_channel_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config *config, | |
1009 | struct channel *chan, | |
1010 | struct lttng_ust_shm_handle *handle) | |
1011 | { | |
1012 | struct shm_ref *ref; | |
1013 | ||
1014 | ref = &handle->chan._ref; | |
1015 | return shm_close_wakeup_fd(handle, ref); | |
1016 | } | |
1017 | ||
1018 | int ring_buffer_stream_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config *config, | |
74d81a6c MD |
1019 | struct channel *chan, |
1020 | struct lttng_ust_shm_handle *handle, | |
1021 | int cpu) | |
852c2936 | 1022 | { |
74d81a6c MD |
1023 | struct shm_ref *ref; |
1024 | ||
1025 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
1026 | cpu = 0; | |
1027 | } else { | |
1028 | if (cpu >= num_possible_cpus()) | |
1029 | return -EINVAL; | |
824f40b8 | 1030 | } |
74d81a6c MD |
1031 | ref = &chan->backend.buf[cpu].shmp._ref; |
1032 | return shm_close_wait_fd(handle, ref); | |
1033 | } | |
1034 | ||
ff0f5728 | 1035 | int ring_buffer_stream_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config *config, |
74d81a6c MD |
1036 | struct channel *chan, |
1037 | struct lttng_ust_shm_handle *handle, | |
1038 | int cpu) | |
1039 | { | |
1040 | struct shm_ref *ref; | |
cb7378b3 | 1041 | int ret; |
74d81a6c MD |
1042 | |
1043 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
1044 | cpu = 0; | |
1045 | } else { | |
1046 | if (cpu >= num_possible_cpus()) | |
1047 | return -EINVAL; | |
1048 | } | |
1049 | ref = &chan->backend.buf[cpu].shmp._ref; | |
cb7378b3 MD |
1050 | pthread_mutex_lock(&wakeup_fd_mutex); |
1051 | ret = shm_close_wakeup_fd(handle, ref); | |
1052 | pthread_mutex_unlock(&wakeup_fd_mutex); | |
1053 | return ret; | |
74d81a6c MD |
1054 | } |
1055 | ||
1056 | int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer *buf, | |
1057 | struct lttng_ust_shm_handle *handle) | |
1058 | { | |
a6352fd4 | 1059 | if (uatomic_cmpxchg(&buf->active_readers, 0, 1) != 0) |
852c2936 | 1060 | return -EBUSY; |
a6352fd4 | 1061 | cmm_smp_mb(); |
852c2936 MD |
1062 | return 0; |
1063 | } | |
852c2936 | 1064 | |
4cfec15c | 1065 | void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer *buf, |
74d81a6c | 1066 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1067 | { |
1d498196 | 1068 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 | 1069 | |
a6352fd4 MD |
1070 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
1071 | cmm_smp_mb(); | |
1072 | uatomic_dec(&buf->active_readers); | |
852c2936 MD |
1073 | } |
1074 | ||
1075 | /** | |
1076 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
1077 | * @buf: ring buffer | |
1078 | * @consumed: consumed count indicating the position where to read | |
1079 | * @produced: produced count, indicates position when to stop reading | |
1080 | * | |
1081 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
1082 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 MD |
1083 | */ |
1084 | ||
4cfec15c | 1085 | int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 1086 | unsigned long *consumed, unsigned long *produced, |
38fae1d3 | 1087 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1088 | { |
1d498196 | 1089 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 1090 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1091 | unsigned long consumed_cur, write_offset; |
1092 | int finalized; | |
1093 | ||
14641deb | 1094 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
1095 | /* |
1096 | * Read finalized before counters. | |
1097 | */ | |
a6352fd4 MD |
1098 | cmm_smp_rmb(); |
1099 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 MD |
1100 | /* |
1101 | * No need to issue a memory barrier between consumed count read and | |
1102 | * write offset read, because consumed count can only change | |
1103 | * concurrently in overwrite mode, and we keep a sequence counter | |
1104 | * identifier derived from the write offset to check we are getting | |
1105 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
1106 | * "tagged" upon writes, tags are checked upon read). | |
1107 | */ | |
1108 | write_offset = v_read(config, &buf->offset); | |
1109 | ||
1110 | /* | |
1111 | * Check that we are not about to read the same subbuffer in | |
1112 | * which the writer head is. | |
1113 | */ | |
1114 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
1115 | == 0) | |
1116 | goto nodata; | |
1117 | ||
1118 | *consumed = consumed_cur; | |
1119 | *produced = subbuf_trunc(write_offset, chan); | |
1120 | ||
1121 | return 0; | |
1122 | ||
1123 | nodata: | |
1124 | /* | |
1125 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1126 | * of "raw_spin_is_locked" memory ordering. | |
1127 | */ | |
1128 | if (finalized) | |
1129 | return -ENODATA; | |
852c2936 MD |
1130 | else |
1131 | return -EAGAIN; | |
1132 | } | |
852c2936 MD |
1133 | |
1134 | /** | |
d1a996f5 | 1135 | * lib_ring_buffer_move_consumer - move consumed counter forward |
852c2936 MD |
1136 | * @buf: ring buffer |
1137 | * @consumed_new: new consumed count value | |
1138 | */ | |
4cfec15c | 1139 | void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 1140 | unsigned long consumed_new, |
38fae1d3 | 1141 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1142 | { |
4cfec15c | 1143 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
1d498196 | 1144 | struct channel *chan = shmp(handle, bufb->chan); |
852c2936 MD |
1145 | unsigned long consumed; |
1146 | ||
74d81a6c | 1147 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
852c2936 MD |
1148 | |
1149 | /* | |
1150 | * Only push the consumed value forward. | |
1151 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
1152 | * the writer in flight recorder mode. | |
1153 | */ | |
a6352fd4 | 1154 | consumed = uatomic_read(&buf->consumed); |
852c2936 | 1155 | while ((long) consumed - (long) consumed_new < 0) |
a6352fd4 MD |
1156 | consumed = uatomic_cmpxchg(&buf->consumed, consumed, |
1157 | consumed_new); | |
852c2936 | 1158 | } |
852c2936 MD |
1159 | |
1160 | /** | |
1161 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
1162 | * @buf: ring buffer | |
1163 | * @consumed: consumed count indicating the position where to read | |
1164 | * | |
1165 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
1166 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 | 1167 | */ |
4cfec15c | 1168 | int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 1169 | unsigned long consumed, |
38fae1d3 | 1170 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1171 | { |
1d498196 | 1172 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 1173 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 | 1174 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
dddba398 | 1175 | int ret, finalized, nr_retry = LTTNG_UST_RING_BUFFER_GET_RETRY; |
852c2936 MD |
1176 | |
1177 | retry: | |
14641deb | 1178 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
1179 | /* |
1180 | * Read finalized before counters. | |
1181 | */ | |
a6352fd4 MD |
1182 | cmm_smp_rmb(); |
1183 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 | 1184 | consumed_idx = subbuf_index(consumed, chan); |
4746ae29 | 1185 | commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb); |
852c2936 MD |
1186 | /* |
1187 | * Make sure we read the commit count before reading the buffer | |
1188 | * data and the write offset. Correct consumed offset ordering | |
1189 | * wrt commit count is insured by the use of cmpxchg to update | |
1190 | * the consumed offset. | |
852c2936 | 1191 | */ |
a6352fd4 MD |
1192 | /* |
1193 | * Local rmb to match the remote wmb to read the commit count | |
1194 | * before the buffer data and the write offset. | |
1195 | */ | |
1196 | cmm_smp_rmb(); | |
852c2936 MD |
1197 | |
1198 | write_offset = v_read(config, &buf->offset); | |
1199 | ||
1200 | /* | |
1201 | * Check that the buffer we are getting is after or at consumed_cur | |
1202 | * position. | |
1203 | */ | |
1204 | if ((long) subbuf_trunc(consumed, chan) | |
1205 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
1206 | goto nodata; | |
1207 | ||
1208 | /* | |
1209 | * Check that the subbuffer we are trying to consume has been | |
dddba398 MD |
1210 | * already fully committed. There are a few causes that can make |
1211 | * this unavailability situation occur: | |
1212 | * | |
1213 | * Temporary (short-term) situation: | |
1214 | * - Application is running on a different CPU, between reserve | |
1215 | * and commit ring buffer operations, | |
1216 | * - Application is preempted between reserve and commit ring | |
1217 | * buffer operations, | |
1218 | * | |
1219 | * Long-term situation: | |
1220 | * - Application is stopped (SIGSTOP) between reserve and commit | |
1221 | * ring buffer operations. Could eventually be resumed by | |
1222 | * SIGCONT. | |
1223 | * - Application is killed (SIGTERM, SIGINT, SIGKILL) between | |
1224 | * reserve and commit ring buffer operation. | |
1225 | * | |
1226 | * From a consumer perspective, handling short-term | |
1227 | * unavailability situations is performed by retrying a few | |
1228 | * times after a delay. Handling long-term unavailability | |
1229 | * situations is handled by failing to get the sub-buffer. | |
1230 | * | |
1231 | * In all of those situations, if the application is taking a | |
1232 | * long time to perform its commit after ring buffer space | |
1233 | * reservation, we can end up in a situation where the producer | |
1234 | * will fill the ring buffer and try to write into the same | |
1235 | * sub-buffer again (which has a missing commit). This is | |
1236 | * handled by the producer in the sub-buffer switch handling | |
1237 | * code of the reserve routine by detecting unbalanced | |
1238 | * reserve/commit counters and discarding all further events | |
1239 | * until the situation is resolved in those situations. Two | |
1240 | * scenarios can occur: | |
1241 | * | |
1242 | * 1) The application causing the reserve/commit counters to be | |
1243 | * unbalanced has been terminated. In this situation, all | |
1244 | * further events will be discarded in the buffers, and no | |
1245 | * further buffer data will be readable by the consumer | |
1246 | * daemon. Tearing down the UST tracing session and starting | |
1247 | * anew is a work-around for those situations. Note that this | |
1248 | * only affects per-UID tracing. In per-PID tracing, the | |
1249 | * application vanishes with the termination, and therefore | |
1250 | * no more data needs to be written to the buffers. | |
1251 | * 2) The application causing the unbalance has been delayed for | |
1252 | * a long time, but will eventually try to increment the | |
1253 | * commit counter after eventually writing to the sub-buffer. | |
1254 | * This situation can cause events to be discarded until the | |
1255 | * application resumes its operations. | |
852c2936 MD |
1256 | */ |
1257 | if (((commit_count - chan->backend.subbuf_size) | |
1258 | & chan->commit_count_mask) | |
0bade047 | 1259 | - (buf_trunc(consumed, chan) |
852c2936 | 1260 | >> chan->backend.num_subbuf_order) |
dddba398 MD |
1261 | != 0) { |
1262 | if (nr_retry-- > 0) { | |
1263 | if (nr_retry <= (LTTNG_UST_RING_BUFFER_GET_RETRY >> 1)) | |
1264 | (void) poll(NULL, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS); | |
1265 | goto retry; | |
1266 | } else { | |
1267 | goto nodata; | |
1268 | } | |
1269 | } | |
852c2936 MD |
1270 | |
1271 | /* | |
1272 | * Check that we are not about to read the same subbuffer in | |
1273 | * which the writer head is. | |
1274 | */ | |
0bade047 | 1275 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed, chan) |
852c2936 MD |
1276 | == 0) |
1277 | goto nodata; | |
1278 | ||
1279 | /* | |
1280 | * Failure to get the subbuffer causes a busy-loop retry without going | |
1281 | * to a wait queue. These are caused by short-lived race windows where | |
1282 | * the writer is getting access to a subbuffer we were trying to get | |
1283 | * access to. Also checks that the "consumed" buffer count we are | |
1284 | * looking for matches the one contained in the subbuffer id. | |
dddba398 MD |
1285 | * |
1286 | * The short-lived race window described here can be affected by | |
1287 | * application signals and preemption, thus requiring to bound | |
1288 | * the loop to a maximum number of retry. | |
852c2936 MD |
1289 | */ |
1290 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
1291 | consumed_idx, buf_trunc_val(consumed, chan), |
1292 | handle); | |
dddba398 MD |
1293 | if (ret) { |
1294 | if (nr_retry-- > 0) { | |
1295 | if (nr_retry <= (LTTNG_UST_RING_BUFFER_GET_RETRY >> 1)) | |
1296 | (void) poll(NULL, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS); | |
1297 | goto retry; | |
1298 | } else { | |
1299 | goto nodata; | |
1300 | } | |
1301 | } | |
852c2936 MD |
1302 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); |
1303 | ||
1304 | buf->get_subbuf_consumed = consumed; | |
1305 | buf->get_subbuf = 1; | |
1306 | ||
1307 | return 0; | |
1308 | ||
1309 | nodata: | |
1310 | /* | |
1311 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1312 | * of "raw_spin_is_locked" memory ordering. | |
1313 | */ | |
1314 | if (finalized) | |
1315 | return -ENODATA; | |
852c2936 MD |
1316 | else |
1317 | return -EAGAIN; | |
1318 | } | |
852c2936 MD |
1319 | |
1320 | /** | |
1321 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
1322 | * @buf: ring buffer | |
1323 | */ | |
4cfec15c | 1324 | void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 1325 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1326 | { |
4cfec15c | 1327 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
1d498196 | 1328 | struct channel *chan = shmp(handle, bufb->chan); |
4cfec15c | 1329 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1330 | unsigned long read_sb_bindex, consumed_idx, consumed; |
1331 | ||
74d81a6c | 1332 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
852c2936 MD |
1333 | |
1334 | if (!buf->get_subbuf) { | |
1335 | /* | |
1336 | * Reader puts a subbuffer it did not get. | |
1337 | */ | |
1338 | CHAN_WARN_ON(chan, 1); | |
1339 | return; | |
1340 | } | |
1341 | consumed = buf->get_subbuf_consumed; | |
1342 | buf->get_subbuf = 0; | |
1343 | ||
1344 | /* | |
1345 | * Clear the records_unread counter. (overruns counter) | |
1346 | * Can still be non-zero if a file reader simply grabbed the data | |
1347 | * without using iterators. | |
1348 | * Can be below zero if an iterator is used on a snapshot more than | |
1349 | * once. | |
1350 | */ | |
1351 | read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); | |
1352 | v_add(config, v_read(config, | |
4746ae29 | 1353 | &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread), |
852c2936 | 1354 | &bufb->records_read); |
4746ae29 | 1355 | v_set(config, &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread, 0); |
852c2936 MD |
1356 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE |
1357 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
1358 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
1359 | ||
1360 | /* | |
1361 | * Exchange the reader subbuffer with the one we put in its place in the | |
1362 | * writer subbuffer table. Expect the original consumed count. If | |
1363 | * update_read_sb_index fails, this is because the writer updated the | |
1364 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
1365 | * currently have: it has become invalid to try reading this sub-buffer | |
1366 | * consumed count value anyway. | |
1367 | */ | |
1368 | consumed_idx = subbuf_index(consumed, chan); | |
1369 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
1370 | consumed_idx, buf_trunc_val(consumed, chan), |
1371 | handle); | |
852c2936 MD |
1372 | /* |
1373 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
1374 | * if the writer concurrently updated it. | |
1375 | */ | |
1376 | } | |
852c2936 MD |
1377 | |
1378 | /* | |
1379 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
1380 | * position and the writer position. (inclusive) | |
1381 | */ | |
1382 | static | |
4cfec15c | 1383 | void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1384 | struct channel *chan, |
1385 | unsigned long cons_offset, | |
1d498196 | 1386 | int cpu, |
38fae1d3 | 1387 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1388 | { |
4cfec15c | 1389 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1390 | unsigned long cons_idx, commit_count, commit_count_sb; |
1391 | ||
1392 | cons_idx = subbuf_index(cons_offset, chan); | |
4746ae29 MD |
1393 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, cons_idx)->cc); |
1394 | commit_count_sb = v_read(config, &shmp_index(handle, buf->commit_cold, cons_idx)->cc_sb); | |
852c2936 MD |
1395 | |
1396 | if (subbuf_offset(commit_count, chan) != 0) | |
4d3c9523 | 1397 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
1398 | "commit count in subbuffer %lu,\n" |
1399 | "expecting multiples of %lu bytes\n" | |
1400 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
1401 | chan->backend.name, cpu, cons_idx, | |
1402 | chan->backend.subbuf_size, | |
1403 | commit_count, commit_count_sb); | |
1404 | ||
4d3c9523 | 1405 | DBG("ring buffer: %s, cpu %d: %lu bytes committed\n", |
852c2936 MD |
1406 | chan->backend.name, cpu, commit_count); |
1407 | } | |
1408 | ||
1409 | static | |
4cfec15c | 1410 | void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 | 1411 | struct channel *chan, |
1d498196 | 1412 | void *priv, int cpu, |
38fae1d3 | 1413 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1414 | { |
4cfec15c | 1415 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1416 | unsigned long write_offset, cons_offset; |
1417 | ||
852c2936 MD |
1418 | /* |
1419 | * No need to order commit_count, write_offset and cons_offset reads | |
1420 | * because we execute at teardown when no more writer nor reader | |
1421 | * references are left. | |
1422 | */ | |
1423 | write_offset = v_read(config, &buf->offset); | |
a6352fd4 | 1424 | cons_offset = uatomic_read(&buf->consumed); |
852c2936 | 1425 | if (write_offset != cons_offset) |
4d3c9523 | 1426 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
1427 | "non-consumed data\n" |
1428 | " [ %lu bytes written, %lu bytes read ]\n", | |
1429 | chan->backend.name, cpu, write_offset, cons_offset); | |
1430 | ||
a6352fd4 | 1431 | for (cons_offset = uatomic_read(&buf->consumed); |
852c2936 MD |
1432 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), |
1433 | chan) | |
1434 | - cons_offset) > 0; | |
1435 | cons_offset = subbuf_align(cons_offset, chan)) | |
1436 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1d498196 | 1437 | cpu, handle); |
852c2936 MD |
1438 | } |
1439 | ||
1440 | static | |
1441 | void lib_ring_buffer_print_errors(struct channel *chan, | |
009769ca MD |
1442 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
1443 | struct lttng_ust_shm_handle *handle) | |
852c2936 | 1444 | { |
4cfec15c | 1445 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
a3f61e7f | 1446 | void *priv = channel_get_private(chan); |
852c2936 | 1447 | |
a1360615 MD |
1448 | if (!strcmp(chan->backend.name, "relay-metadata-mmap")) { |
1449 | DBG("ring buffer %s: %lu records written, " | |
1450 | "%lu records overrun\n", | |
1451 | chan->backend.name, | |
1452 | v_read(config, &buf->records_count), | |
1453 | v_read(config, &buf->records_overrun)); | |
1454 | } else { | |
1455 | DBG("ring buffer %s, cpu %d: %lu records written, " | |
1456 | "%lu records overrun\n", | |
1457 | chan->backend.name, cpu, | |
1458 | v_read(config, &buf->records_count), | |
1459 | v_read(config, &buf->records_overrun)); | |
1460 | ||
1461 | if (v_read(config, &buf->records_lost_full) | |
1462 | || v_read(config, &buf->records_lost_wrap) | |
1463 | || v_read(config, &buf->records_lost_big)) | |
1464 | DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n" | |
1465 | " [ %lu buffer full, %lu nest buffer wrap-around, " | |
1466 | "%lu event too big ]\n", | |
1467 | chan->backend.name, cpu, | |
1468 | v_read(config, &buf->records_lost_full), | |
1469 | v_read(config, &buf->records_lost_wrap), | |
1470 | v_read(config, &buf->records_lost_big)); | |
1471 | } | |
1d498196 | 1472 | lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu, handle); |
852c2936 MD |
1473 | } |
1474 | ||
1475 | /* | |
1476 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
1477 | * | |
1478 | * Only executed when the buffer is finalized, in SWITCH_FLUSH. | |
1479 | */ | |
1480 | static | |
4cfec15c | 1481 | void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1482 | struct channel *chan, |
1483 | struct switch_offsets *offsets, | |
2fed87ae | 1484 | uint64_t tsc, |
38fae1d3 | 1485 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1486 | { |
4cfec15c | 1487 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1488 | unsigned long oldidx = subbuf_index(offsets->old, chan); |
1489 | unsigned long commit_count; | |
1490 | ||
1d498196 | 1491 | config->cb.buffer_begin(buf, tsc, oldidx, handle); |
852c2936 MD |
1492 | |
1493 | /* | |
1494 | * Order all writes to buffer before the commit count update that will | |
1495 | * determine that the subbuffer is full. | |
1496 | */ | |
a6352fd4 | 1497 | cmm_smp_wmb(); |
852c2936 | 1498 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1499 | &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1500 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 MD |
1501 | /* Check if the written buffer has to be delivered */ |
1502 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
1b7b0501 | 1503 | commit_count, oldidx, handle, tsc); |
852c2936 | 1504 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
80249235 MD |
1505 | offsets->old + config->cb.subbuffer_header_size(), |
1506 | commit_count, handle); | |
852c2936 MD |
1507 | } |
1508 | ||
1509 | /* | |
1510 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
1511 | * | |
1512 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
1513 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
1514 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
1515 | * subbuffer. | |
1516 | */ | |
1517 | static | |
4cfec15c | 1518 | void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1519 | struct channel *chan, |
1520 | struct switch_offsets *offsets, | |
2fed87ae | 1521 | uint64_t tsc, |
38fae1d3 | 1522 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1523 | { |
4cfec15c | 1524 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1525 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); |
1526 | unsigned long commit_count, padding_size, data_size; | |
1527 | ||
1528 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1529 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1530 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size, |
1531 | handle); | |
852c2936 MD |
1532 | |
1533 | /* | |
1534 | * Order all writes to buffer before the commit count update that will | |
1535 | * determine that the subbuffer is full. | |
1536 | */ | |
a6352fd4 | 1537 | cmm_smp_wmb(); |
4746ae29 MD |
1538 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1539 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 | 1540 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, |
1b7b0501 | 1541 | commit_count, oldidx, handle, tsc); |
852c2936 | 1542 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
80249235 | 1543 | offsets->old + padding_size, commit_count, handle); |
852c2936 MD |
1544 | } |
1545 | ||
1546 | /* | |
1547 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1548 | * | |
1549 | * This code can be executed unordered : writers may already have written to the | |
1550 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1551 | * that this code is executed before the deliver of this sub-buffer. | |
1552 | */ | |
1553 | static | |
4cfec15c | 1554 | void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1555 | struct channel *chan, |
1556 | struct switch_offsets *offsets, | |
2fed87ae | 1557 | uint64_t tsc, |
38fae1d3 | 1558 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1559 | { |
4cfec15c | 1560 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1561 | unsigned long beginidx = subbuf_index(offsets->begin, chan); |
1562 | unsigned long commit_count; | |
1563 | ||
1d498196 | 1564 | config->cb.buffer_begin(buf, tsc, beginidx, handle); |
852c2936 MD |
1565 | |
1566 | /* | |
1567 | * Order all writes to buffer before the commit count update that will | |
1568 | * determine that the subbuffer is full. | |
1569 | */ | |
a6352fd4 | 1570 | cmm_smp_wmb(); |
852c2936 | 1571 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1572 | &shmp_index(handle, buf->commit_hot, beginidx)->cc); |
1573 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, beginidx)->cc); | |
852c2936 MD |
1574 | /* Check if the written buffer has to be delivered */ |
1575 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
1b7b0501 | 1576 | commit_count, beginidx, handle, tsc); |
852c2936 | 1577 | lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx, |
80249235 MD |
1578 | offsets->begin + config->cb.subbuffer_header_size(), |
1579 | commit_count, handle); | |
852c2936 MD |
1580 | } |
1581 | ||
1ad21f70 MD |
1582 | /* |
1583 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1584 | * | |
3962118d MD |
1585 | * Calls subbuffer_set_data_size() to set the data size of the current |
1586 | * sub-buffer. We do not need to perform check_deliver nor commit here, | |
1587 | * since this task will be done by the "commit" of the event for which | |
1588 | * we are currently doing the space reservation. | |
1ad21f70 MD |
1589 | */ |
1590 | static | |
1591 | void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer *buf, | |
1592 | struct channel *chan, | |
1593 | struct switch_offsets *offsets, | |
1594 | uint64_t tsc, | |
1595 | struct lttng_ust_shm_handle *handle) | |
1596 | { | |
1597 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; | |
3962118d | 1598 | unsigned long endidx, data_size; |
1ad21f70 | 1599 | |
3962118d | 1600 | endidx = subbuf_index(offsets->end - 1, chan); |
1ad21f70 | 1601 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; |
1ad21f70 MD |
1602 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size, |
1603 | handle); | |
1ad21f70 MD |
1604 | } |
1605 | ||
852c2936 MD |
1606 | /* |
1607 | * Returns : | |
1608 | * 0 if ok | |
1609 | * !0 if execution must be aborted. | |
1610 | */ | |
1611 | static | |
1612 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
4cfec15c | 1613 | struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1614 | struct channel *chan, |
1615 | struct switch_offsets *offsets, | |
2dac206d MD |
1616 | uint64_t *tsc, |
1617 | struct lttng_ust_shm_handle *handle) | |
852c2936 | 1618 | { |
4cfec15c | 1619 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
2dac206d | 1620 | unsigned long off, reserve_commit_diff; |
852c2936 MD |
1621 | |
1622 | offsets->begin = v_read(config, &buf->offset); | |
1623 | offsets->old = offsets->begin; | |
1624 | offsets->switch_old_start = 0; | |
1625 | off = subbuf_offset(offsets->begin, chan); | |
1626 | ||
1627 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1628 | ||
1629 | /* | |
1630 | * Ensure we flush the header of an empty subbuffer when doing the | |
1631 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1632 | * total data gathering duration even if there were no records saved | |
1633 | * after the last buffer switch. | |
1634 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1635 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1636 | * subbuffer header as appropriate. | |
1637 | * The next record that reserves space will be responsible for | |
1638 | * populating the following subbuffer header. We choose not to populate | |
1639 | * the next subbuffer header here because we want to be able to use | |
a6352fd4 MD |
1640 | * SWITCH_ACTIVE for periodical buffer flush, which must |
1641 | * guarantee that all the buffer content (records and header | |
1642 | * timestamps) are visible to the reader. This is required for | |
1643 | * quiescence guarantees for the fusion merge. | |
852c2936 | 1644 | */ |
2dac206d MD |
1645 | if (mode != SWITCH_FLUSH && !off) |
1646 | return -1; /* we do not have to switch : buffer is empty */ | |
1647 | ||
1648 | if (caa_unlikely(off == 0)) { | |
1649 | unsigned long sb_index, commit_count; | |
1650 | ||
1651 | /* | |
1652 | * We are performing a SWITCH_FLUSH. At this stage, there are no | |
1653 | * concurrent writes into the buffer. | |
1654 | * | |
1655 | * The client does not save any header information. Don't | |
1656 | * switch empty subbuffer on finalize, because it is invalid to | |
1657 | * deliver a completely empty subbuffer. | |
1658 | */ | |
1659 | if (!config->cb.subbuffer_header_size()) | |
1660 | return -1; | |
1661 | ||
1662 | /* Test new buffer integrity */ | |
1663 | sb_index = subbuf_index(offsets->begin, chan); | |
1664 | commit_count = v_read(config, | |
1665 | &shmp_index(handle, buf->commit_cold, | |
1666 | sb_index)->cc_sb); | |
1667 | reserve_commit_diff = | |
1668 | (buf_trunc(offsets->begin, chan) | |
1669 | >> chan->backend.num_subbuf_order) | |
1670 | - (commit_count & chan->commit_count_mask); | |
1671 | if (caa_likely(reserve_commit_diff == 0)) { | |
1672 | /* Next subbuffer not being written to. */ | |
1673 | if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
1674 | subbuf_trunc(offsets->begin, chan) | |
1675 | - subbuf_trunc((unsigned long) | |
1676 | uatomic_read(&buf->consumed), chan) | |
1677 | >= chan->backend.buf_size)) { | |
1678 | /* | |
1679 | * We do not overwrite non consumed buffers | |
1680 | * and we are full : don't switch. | |
1681 | */ | |
852c2936 | 1682 | return -1; |
2dac206d MD |
1683 | } else { |
1684 | /* | |
1685 | * Next subbuffer not being written to, and we | |
1686 | * are either in overwrite mode or the buffer is | |
1687 | * not full. It's safe to write in this new | |
1688 | * subbuffer. | |
1689 | */ | |
1690 | } | |
1691 | } else { | |
852c2936 | 1692 | /* |
2dac206d MD |
1693 | * Next subbuffer reserve offset does not match the |
1694 | * commit offset. Don't perform switch in | |
1695 | * producer-consumer and overwrite mode. Caused by | |
1696 | * either a writer OOPS or too many nested writes over a | |
1697 | * reserve/commit pair. | |
852c2936 | 1698 | */ |
2dac206d | 1699 | return -1; |
852c2936 | 1700 | } |
2dac206d MD |
1701 | |
1702 | /* | |
1703 | * Need to write the subbuffer start header on finalize. | |
1704 | */ | |
1705 | offsets->switch_old_start = 1; | |
1706 | } | |
1707 | offsets->begin = subbuf_align(offsets->begin, chan); | |
852c2936 MD |
1708 | /* Note: old points to the next subbuf at offset 0 */ |
1709 | offsets->end = offsets->begin; | |
1710 | return 0; | |
1711 | } | |
1712 | ||
1713 | /* | |
1714 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
1715 | * called while tracing is active with absolutely no lock held. | |
1716 | * | |
1717 | * Note, however, that as a v_cmpxchg is used for some atomic | |
1718 | * operations, this function must be called from the CPU which owns the buffer | |
1719 | * for a ACTIVE flush. | |
1720 | */ | |
4cfec15c | 1721 | void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode, |
38fae1d3 | 1722 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1723 | { |
1d498196 | 1724 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 1725 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1726 | struct switch_offsets offsets; |
1727 | unsigned long oldidx; | |
2fed87ae | 1728 | uint64_t tsc; |
852c2936 MD |
1729 | |
1730 | offsets.size = 0; | |
1731 | ||
1732 | /* | |
1733 | * Perform retryable operations. | |
1734 | */ | |
1735 | do { | |
1736 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
2dac206d | 1737 | &tsc, handle)) |
852c2936 MD |
1738 | return; /* Switch not needed */ |
1739 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
1740 | != offsets.old); | |
1741 | ||
1742 | /* | |
1743 | * Atomically update last_tsc. This update races against concurrent | |
1744 | * atomic updates, but the race will always cause supplementary full TSC | |
1745 | * records, never the opposite (missing a full TSC record when it would | |
1746 | * be needed). | |
1747 | */ | |
1748 | save_last_tsc(config, buf, tsc); | |
1749 | ||
1750 | /* | |
1751 | * Push the reader if necessary | |
1752 | */ | |
1753 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
1754 | ||
1755 | oldidx = subbuf_index(offsets.old, chan); | |
1d498196 | 1756 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx, handle); |
852c2936 MD |
1757 | |
1758 | /* | |
1759 | * May need to populate header start on SWITCH_FLUSH. | |
1760 | */ | |
1761 | if (offsets.switch_old_start) { | |
1d498196 | 1762 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc, handle); |
852c2936 MD |
1763 | offsets.old += config->cb.subbuffer_header_size(); |
1764 | } | |
1765 | ||
1766 | /* | |
1767 | * Switch old subbuffer. | |
1768 | */ | |
1d498196 | 1769 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc, handle); |
852c2936 | 1770 | } |
852c2936 MD |
1771 | |
1772 | /* | |
1773 | * Returns : | |
1774 | * 0 if ok | |
1775 | * -ENOSPC if event size is too large for packet. | |
1776 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
1777 | * -EIO if data cannot be written into the buffer for any other reason. | |
1778 | */ | |
1779 | static | |
4cfec15c | 1780 | int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1781 | struct channel *chan, |
1782 | struct switch_offsets *offsets, | |
4cfec15c | 1783 | struct lttng_ust_lib_ring_buffer_ctx *ctx) |
852c2936 | 1784 | { |
4cfec15c | 1785 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
38fae1d3 | 1786 | struct lttng_ust_shm_handle *handle = ctx->handle; |
e185cf4b | 1787 | unsigned long reserve_commit_diff, offset_cmp; |
852c2936 | 1788 | |
e185cf4b MD |
1789 | retry: |
1790 | offsets->begin = offset_cmp = v_read(config, &buf->offset); | |
852c2936 MD |
1791 | offsets->old = offsets->begin; |
1792 | offsets->switch_new_start = 0; | |
1ad21f70 | 1793 | offsets->switch_new_end = 0; |
852c2936 MD |
1794 | offsets->switch_old_end = 0; |
1795 | offsets->pre_header_padding = 0; | |
1796 | ||
1797 | ctx->tsc = config->cb.ring_buffer_clock_read(chan); | |
1798 | if ((int64_t) ctx->tsc == -EIO) | |
1799 | return -EIO; | |
1800 | ||
1801 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
1802 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
1803 | ||
b5a3dfa5 | 1804 | if (caa_unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) { |
852c2936 MD |
1805 | offsets->switch_new_start = 1; /* For offsets->begin */ |
1806 | } else { | |
1807 | offsets->size = config->cb.record_header_size(config, chan, | |
1808 | offsets->begin, | |
1809 | &offsets->pre_header_padding, | |
1810 | ctx); | |
1811 | offsets->size += | |
1812 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1813 | ctx->largest_align) | |
1814 | + ctx->data_size; | |
b5a3dfa5 | 1815 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) + |
852c2936 MD |
1816 | offsets->size > chan->backend.subbuf_size)) { |
1817 | offsets->switch_old_end = 1; /* For offsets->old */ | |
1818 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1819 | } | |
1820 | } | |
b5a3dfa5 | 1821 | if (caa_unlikely(offsets->switch_new_start)) { |
e185cf4b | 1822 | unsigned long sb_index, commit_count; |
852c2936 MD |
1823 | |
1824 | /* | |
1825 | * We are typically not filling the previous buffer completely. | |
1826 | */ | |
b5a3dfa5 | 1827 | if (caa_likely(offsets->switch_old_end)) |
852c2936 MD |
1828 | offsets->begin = subbuf_align(offsets->begin, chan); |
1829 | offsets->begin = offsets->begin | |
1830 | + config->cb.subbuffer_header_size(); | |
1831 | /* Test new buffer integrity */ | |
1832 | sb_index = subbuf_index(offsets->begin, chan); | |
e185cf4b MD |
1833 | /* |
1834 | * Read buf->offset before buf->commit_cold[sb_index].cc_sb. | |
1835 | * lib_ring_buffer_check_deliver() has the matching | |
1836 | * memory barriers required around commit_cold cc_sb | |
1837 | * updates to ensure reserve and commit counter updates | |
1838 | * are not seen reordered when updated by another CPU. | |
1839 | */ | |
1840 | cmm_smp_rmb(); | |
1841 | commit_count = v_read(config, | |
1842 | &shmp_index(handle, buf->commit_cold, | |
1843 | sb_index)->cc_sb); | |
1844 | /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */ | |
1845 | cmm_smp_rmb(); | |
1846 | if (caa_unlikely(offset_cmp != v_read(config, &buf->offset))) { | |
1847 | /* | |
1848 | * The reserve counter have been concurrently updated | |
1849 | * while we read the commit counter. This means the | |
1850 | * commit counter we read might not match buf->offset | |
1851 | * due to concurrent update. We therefore need to retry. | |
1852 | */ | |
1853 | goto retry; | |
1854 | } | |
852c2936 MD |
1855 | reserve_commit_diff = |
1856 | (buf_trunc(offsets->begin, chan) | |
1857 | >> chan->backend.num_subbuf_order) | |
e185cf4b | 1858 | - (commit_count & chan->commit_count_mask); |
b5a3dfa5 | 1859 | if (caa_likely(reserve_commit_diff == 0)) { |
852c2936 | 1860 | /* Next subbuffer not being written to. */ |
b5a3dfa5 | 1861 | if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE && |
852c2936 MD |
1862 | subbuf_trunc(offsets->begin, chan) |
1863 | - subbuf_trunc((unsigned long) | |
a6352fd4 | 1864 | uatomic_read(&buf->consumed), chan) |
852c2936 | 1865 | >= chan->backend.buf_size)) { |
64493e4f MD |
1866 | unsigned long nr_lost; |
1867 | ||
852c2936 MD |
1868 | /* |
1869 | * We do not overwrite non consumed buffers | |
1870 | * and we are full : record is lost. | |
1871 | */ | |
64493e4f | 1872 | nr_lost = v_read(config, &buf->records_lost_full); |
852c2936 | 1873 | v_inc(config, &buf->records_lost_full); |
64493e4f MD |
1874 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
1875 | DBG("%lu or more records lost in (%s:%d) (buffer full)\n", | |
1876 | nr_lost + 1, chan->backend.name, | |
1877 | buf->backend.cpu); | |
1878 | } | |
852c2936 MD |
1879 | return -ENOBUFS; |
1880 | } else { | |
1881 | /* | |
1882 | * Next subbuffer not being written to, and we | |
1883 | * are either in overwrite mode or the buffer is | |
1884 | * not full. It's safe to write in this new | |
1885 | * subbuffer. | |
1886 | */ | |
1887 | } | |
1888 | } else { | |
64493e4f MD |
1889 | unsigned long nr_lost; |
1890 | ||
852c2936 MD |
1891 | /* |
1892 | * Next subbuffer reserve offset does not match the | |
e185cf4b MD |
1893 | * commit offset, and this did not involve update to the |
1894 | * reserve counter. Drop record in producer-consumer and | |
852c2936 MD |
1895 | * overwrite mode. Caused by either a writer OOPS or too |
1896 | * many nested writes over a reserve/commit pair. | |
1897 | */ | |
64493e4f | 1898 | nr_lost = v_read(config, &buf->records_lost_wrap); |
852c2936 | 1899 | v_inc(config, &buf->records_lost_wrap); |
64493e4f MD |
1900 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
1901 | DBG("%lu or more records lost in (%s:%d) (wrap-around)\n", | |
1902 | nr_lost + 1, chan->backend.name, | |
1903 | buf->backend.cpu); | |
1904 | } | |
852c2936 MD |
1905 | return -EIO; |
1906 | } | |
1907 | offsets->size = | |
1908 | config->cb.record_header_size(config, chan, | |
1909 | offsets->begin, | |
1910 | &offsets->pre_header_padding, | |
1911 | ctx); | |
1912 | offsets->size += | |
1913 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1914 | ctx->largest_align) | |
1915 | + ctx->data_size; | |
b5a3dfa5 | 1916 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) |
852c2936 | 1917 | + offsets->size > chan->backend.subbuf_size)) { |
64493e4f MD |
1918 | unsigned long nr_lost; |
1919 | ||
852c2936 MD |
1920 | /* |
1921 | * Record too big for subbuffers, report error, don't | |
1922 | * complete the sub-buffer switch. | |
1923 | */ | |
64493e4f | 1924 | nr_lost = v_read(config, &buf->records_lost_big); |
852c2936 | 1925 | v_inc(config, &buf->records_lost_big); |
64493e4f MD |
1926 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
1927 | DBG("%lu or more records lost in (%s:%d) record size " | |
1928 | " of %zu bytes is too large for buffer\n", | |
1929 | nr_lost + 1, chan->backend.name, | |
1930 | buf->backend.cpu, offsets->size); | |
1931 | } | |
852c2936 MD |
1932 | return -ENOSPC; |
1933 | } else { | |
1934 | /* | |
1935 | * We just made a successful buffer switch and the | |
1936 | * record fits in the new subbuffer. Let's write. | |
1937 | */ | |
1938 | } | |
1939 | } else { | |
1940 | /* | |
1941 | * Record fits in the current buffer and we are not on a switch | |
1942 | * boundary. It's safe to write. | |
1943 | */ | |
1944 | } | |
1945 | offsets->end = offsets->begin + offsets->size; | |
1ad21f70 MD |
1946 | |
1947 | if (caa_unlikely(subbuf_offset(offsets->end, chan) == 0)) { | |
1948 | /* | |
1949 | * The offset_end will fall at the very beginning of the next | |
1950 | * subbuffer. | |
1951 | */ | |
1952 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
1953 | } | |
852c2936 MD |
1954 | return 0; |
1955 | } | |
1956 | ||
1957 | /** | |
1958 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
1959 | * @ctx: ring buffer context. | |
1960 | * | |
1961 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, | |
1962 | * -EIO for other errors, else returns 0. | |
1963 | * It will take care of sub-buffer switching. | |
1964 | */ | |
4cfec15c | 1965 | int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx) |
852c2936 MD |
1966 | { |
1967 | struct channel *chan = ctx->chan; | |
38fae1d3 | 1968 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c MD |
1969 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
1970 | struct lttng_ust_lib_ring_buffer *buf; | |
852c2936 MD |
1971 | struct switch_offsets offsets; |
1972 | int ret; | |
1973 | ||
1974 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1d498196 | 1975 | buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp); |
852c2936 | 1976 | else |
1d498196 | 1977 | buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 MD |
1978 | ctx->buf = buf; |
1979 | ||
1980 | offsets.size = 0; | |
1981 | ||
1982 | do { | |
1983 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, | |
1984 | ctx); | |
b5a3dfa5 | 1985 | if (caa_unlikely(ret)) |
852c2936 | 1986 | return ret; |
b5a3dfa5 | 1987 | } while (caa_unlikely(v_cmpxchg(config, &buf->offset, offsets.old, |
852c2936 MD |
1988 | offsets.end) |
1989 | != offsets.old)); | |
1990 | ||
1991 | /* | |
1992 | * Atomically update last_tsc. This update races against concurrent | |
1993 | * atomic updates, but the race will always cause supplementary full TSC | |
1994 | * records, never the opposite (missing a full TSC record when it would | |
1995 | * be needed). | |
1996 | */ | |
1997 | save_last_tsc(config, buf, ctx->tsc); | |
1998 | ||
1999 | /* | |
2000 | * Push the reader if necessary | |
2001 | */ | |
2002 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
2003 | ||
2004 | /* | |
2005 | * Clear noref flag for this subbuffer. | |
2006 | */ | |
2007 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1d498196 MD |
2008 | subbuf_index(offsets.end - 1, chan), |
2009 | handle); | |
852c2936 MD |
2010 | |
2011 | /* | |
2012 | * Switch old subbuffer if needed. | |
2013 | */ | |
b5a3dfa5 | 2014 | if (caa_unlikely(offsets.switch_old_end)) { |
852c2936 | 2015 | lib_ring_buffer_clear_noref(config, &buf->backend, |
1d498196 MD |
2016 | subbuf_index(offsets.old - 1, chan), |
2017 | handle); | |
2018 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc, handle); | |
852c2936 MD |
2019 | } |
2020 | ||
2021 | /* | |
2022 | * Populate new subbuffer. | |
2023 | */ | |
b5a3dfa5 | 2024 | if (caa_unlikely(offsets.switch_new_start)) |
1d498196 | 2025 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 | 2026 | |
1ad21f70 MD |
2027 | if (caa_unlikely(offsets.switch_new_end)) |
2028 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc, handle); | |
2029 | ||
852c2936 MD |
2030 | ctx->slot_size = offsets.size; |
2031 | ctx->pre_offset = offsets.begin; | |
2032 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
2033 | return 0; | |
2034 | } | |
f645cfa7 MD |
2035 | |
2036 | /* | |
2037 | * Force a read (imply TLS fixup for dlopen) of TLS variables. | |
2038 | */ | |
2039 | void lttng_fixup_ringbuffer_tls(void) | |
2040 | { | |
8c90a710 | 2041 | asm volatile ("" : : "m" (URCU_TLS(lib_ring_buffer_nesting))); |
f645cfa7 | 2042 | } |
03d2d293 MD |
2043 | |
2044 | void lib_ringbuffer_signal_init(void) | |
2045 | { | |
2046 | sigset_t mask; | |
2047 | int ret; | |
2048 | ||
2049 | /* | |
2050 | * Block signal for entire process, so only our thread processes | |
2051 | * it. | |
2052 | */ | |
2053 | rb_setmask(&mask); | |
2054 | ret = pthread_sigmask(SIG_BLOCK, &mask, NULL); | |
2055 | if (ret) { | |
2056 | errno = ret; | |
2057 | PERROR("pthread_sigmask"); | |
2058 | } | |
2059 | } |