Commit | Line | Data |
---|---|---|
852c2936 | 1 | /* |
c0c0989a | 2 | * SPDX-License-Identifier: LGPL-2.1-only |
852c2936 | 3 | * |
e92f3e28 MD |
4 | * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
5 | * | |
852c2936 MD |
6 | * Ring buffer wait-free buffer synchronization. Producer-consumer and flight |
7 | * recorder (overwrite) modes. See thesis: | |
8 | * | |
9 | * Desnoyers, Mathieu (2009), "Low-Impact Operating System Tracing", Ph.D. | |
10 | * dissertation, Ecole Polytechnique de Montreal. | |
11 | * http://www.lttng.org/pub/thesis/desnoyers-dissertation-2009-12.pdf | |
12 | * | |
13 | * - Algorithm presentation in Chapter 5: | |
14 | * "Lockless Multi-Core High-Throughput Buffering". | |
15 | * - Algorithm formal verification in Section 8.6: | |
16 | * "Formal verification of LTTng" | |
17 | * | |
18 | * Author: | |
19 | * Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
20 | * | |
21 | * Inspired from LTT and RelayFS: | |
22 | * Karim Yaghmour <karim@opersys.com> | |
23 | * Tom Zanussi <zanussi@us.ibm.com> | |
24 | * Bob Wisniewski <bob@watson.ibm.com> | |
25 | * And from K42 : | |
26 | * Bob Wisniewski <bob@watson.ibm.com> | |
27 | * | |
28 | * Buffer reader semantic : | |
29 | * | |
30 | * - get_subbuf_size | |
31 | * while buffer is not finalized and empty | |
32 | * - get_subbuf | |
33 | * - if return value != 0, continue | |
34 | * - splice one subbuffer worth of data to a pipe | |
35 | * - splice the data from pipe to disk/network | |
36 | * - put_subbuf | |
852c2936 MD |
37 | */ |
38 | ||
3fbec7dc | 39 | #define _LGPL_SOURCE |
a6352fd4 | 40 | #include <sys/types.h> |
431d5cf0 MD |
41 | #include <sys/mman.h> |
42 | #include <sys/stat.h> | |
03d2d293 | 43 | #include <unistd.h> |
431d5cf0 | 44 | #include <fcntl.h> |
03d2d293 MD |
45 | #include <signal.h> |
46 | #include <time.h> | |
932b85fa | 47 | #include <stdbool.h> |
fb31eb73 | 48 | #include <stdint.h> |
14641deb | 49 | #include <urcu/compiler.h> |
a6352fd4 | 50 | #include <urcu/ref.h> |
8c90a710 | 51 | #include <urcu/tls-compat.h> |
327d9447 | 52 | #include <poll.h> |
864a1eda | 53 | #include <ust-helper.h> |
14641deb | 54 | |
cd61d9bf | 55 | #include <lttng/align.h> |
a6352fd4 | 56 | #include "smp.h" |
4318ae1b | 57 | #include <lttng/ringbuffer-config.h> |
2fed87ae | 58 | #include "vatomic.h" |
4931a13e MD |
59 | #include "backend.h" |
60 | #include "frontend.h" | |
a6352fd4 | 61 | #include "shm.h" |
6f97f9c2 | 62 | #include "rb-init.h" |
bdcf8d82 | 63 | #include "../liblttng-ust/compat.h" /* For ENODATA */ |
852c2936 | 64 | |
64493e4f MD |
65 | /* Print DBG() messages about events lost only every 1048576 hits */ |
66 | #define DBG_PRINT_NR_LOST (1UL << 20) | |
67 | ||
34a91bdb MD |
68 | #define LTTNG_UST_RB_SIG_FLUSH SIGRTMIN |
69 | #define LTTNG_UST_RB_SIG_READ SIGRTMIN + 1 | |
70 | #define LTTNG_UST_RB_SIG_TEARDOWN SIGRTMIN + 2 | |
03d2d293 | 71 | #define CLOCKID CLOCK_MONOTONIC |
dddba398 MD |
72 | #define LTTNG_UST_RING_BUFFER_GET_RETRY 10 |
73 | #define LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS 10 | |
6f97f9c2 | 74 | #define RETRY_DELAY_MS 100 /* 100 ms. */ |
03d2d293 | 75 | |
a9ff648c MD |
76 | /* |
77 | * Non-static to ensure the compiler does not optimize away the xor. | |
78 | */ | |
79 | uint8_t lttng_crash_magic_xor[] = RB_CRASH_DUMP_ABI_MAGIC_XOR; | |
80 | ||
2432c3c9 MD |
81 | /* |
82 | * Use POSIX SHM: shm_open(3) and shm_unlink(3). | |
83 | * close(2) to close the fd returned by shm_open. | |
84 | * shm_unlink releases the shared memory object name. | |
85 | * ftruncate(2) sets the size of the memory object. | |
86 | * mmap/munmap maps the shared memory obj to a virtual address in the | |
87 | * calling proceess (should be done both in libust and consumer). | |
88 | * See shm_overview(7) for details. | |
89 | * Pass file descriptor returned by shm_open(3) to ltt-sessiond through | |
90 | * a UNIX socket. | |
91 | * | |
92 | * Since we don't need to access the object using its name, we can | |
93 | * immediately shm_unlink(3) it, and only keep the handle with its file | |
94 | * descriptor. | |
95 | */ | |
96 | ||
852c2936 MD |
97 | /* |
98 | * Internal structure representing offsets to use at a sub-buffer switch. | |
99 | */ | |
100 | struct switch_offsets { | |
101 | unsigned long begin, end, old; | |
102 | size_t pre_header_padding, size; | |
1ad21f70 MD |
103 | unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1, |
104 | switch_old_end:1; | |
852c2936 MD |
105 | }; |
106 | ||
16adecf1 | 107 | DEFINE_URCU_TLS(unsigned int, lib_ring_buffer_nesting); |
852c2936 | 108 | |
cb7378b3 MD |
109 | /* |
110 | * wakeup_fd_mutex protects wakeup fd use by timer from concurrent | |
111 | * close. | |
112 | */ | |
113 | static pthread_mutex_t wakeup_fd_mutex = PTHREAD_MUTEX_INITIALIZER; | |
114 | ||
852c2936 MD |
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. | |
64bf51a6 | 124 | * Protected by the lock within the structure. |
03d2d293 MD |
125 | */ |
126 | struct timer_signal_data { | |
127 | pthread_t tid; /* thread id managing signals */ | |
128 | int setup_done; | |
129 | int qs_done; | |
64bf51a6 | 130 | pthread_mutex_t lock; |
03d2d293 MD |
131 | }; |
132 | ||
64bf51a6 MD |
133 | static struct timer_signal_data timer_signal = { |
134 | .tid = 0, | |
135 | .setup_done = 0, | |
136 | .qs_done = 0, | |
137 | .lock = PTHREAD_MUTEX_INITIALIZER, | |
138 | }; | |
03d2d293 | 139 | |
b2c5f61a | 140 | static bool lttng_ust_allow_blocking; |
6f97f9c2 | 141 | |
b2c5f61a | 142 | void lttng_ust_ringbuffer_set_allow_blocking(void) |
6f97f9c2 | 143 | { |
b2c5f61a MD |
144 | lttng_ust_allow_blocking = true; |
145 | } | |
146 | ||
147 | /* Get blocking timeout, in ms */ | |
148 | static int lttng_ust_ringbuffer_get_timeout(struct channel *chan) | |
149 | { | |
150 | if (!lttng_ust_allow_blocking) | |
151 | return 0; | |
152 | return chan->u.s.blocking_timeout_ms; | |
6f97f9c2 MD |
153 | } |
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 | { |
34daae3e MD |
167 | struct channel *chan; |
168 | const struct lttng_ust_lib_ring_buffer_config *config; | |
852c2936 MD |
169 | unsigned int i; |
170 | ||
34daae3e MD |
171 | chan = shmp(handle, buf->backend.chan); |
172 | if (!chan) | |
15500a1b | 173 | return; |
34daae3e | 174 | config = &chan->backend.config; |
852c2936 MD |
175 | /* |
176 | * Reset iterator first. It will put the subbuffer if it currently holds | |
177 | * it. | |
178 | */ | |
852c2936 MD |
179 | v_set(config, &buf->offset, 0); |
180 | for (i = 0; i < chan->backend.num_subbuf; i++) { | |
15500a1b MD |
181 | struct commit_counters_hot *cc_hot; |
182 | struct commit_counters_cold *cc_cold; | |
6c737d05 | 183 | uint64_t *ts_end; |
15500a1b MD |
184 | |
185 | cc_hot = shmp_index(handle, buf->commit_hot, i); | |
186 | if (!cc_hot) | |
187 | return; | |
188 | cc_cold = shmp_index(handle, buf->commit_cold, i); | |
189 | if (!cc_cold) | |
190 | return; | |
6c737d05 MD |
191 | ts_end = shmp_index(handle, buf->ts_end, i); |
192 | if (!ts_end) | |
193 | return; | |
15500a1b MD |
194 | v_set(config, &cc_hot->cc, 0); |
195 | v_set(config, &cc_hot->seq, 0); | |
196 | v_set(config, &cc_cold->cc_sb, 0); | |
6c737d05 | 197 | *ts_end = 0; |
852c2936 | 198 | } |
a6352fd4 MD |
199 | uatomic_set(&buf->consumed, 0); |
200 | uatomic_set(&buf->record_disabled, 0); | |
852c2936 | 201 | v_set(config, &buf->last_tsc, 0); |
1d498196 | 202 | lib_ring_buffer_backend_reset(&buf->backend, handle); |
852c2936 MD |
203 | /* Don't reset number of active readers */ |
204 | v_set(config, &buf->records_lost_full, 0); | |
205 | v_set(config, &buf->records_lost_wrap, 0); | |
206 | v_set(config, &buf->records_lost_big, 0); | |
207 | v_set(config, &buf->records_count, 0); | |
208 | v_set(config, &buf->records_overrun, 0); | |
209 | buf->finalized = 0; | |
210 | } | |
852c2936 MD |
211 | |
212 | /** | |
213 | * channel_reset - Reset channel to initial values. | |
214 | * @chan: Channel. | |
215 | * | |
216 | * Effectively empty the channel. Should be called when the channel is not used | |
217 | * for writing. The channel can be opened for reading, but the reader should not | |
218 | * be using the iterator concurrently with reset. The previous current iterator | |
219 | * record is reset. | |
220 | */ | |
221 | void channel_reset(struct channel *chan) | |
222 | { | |
223 | /* | |
224 | * Reset iterators first. Will put the subbuffer if held for reading. | |
225 | */ | |
a6352fd4 | 226 | uatomic_set(&chan->record_disabled, 0); |
852c2936 MD |
227 | /* Don't reset commit_count_mask, still valid */ |
228 | channel_backend_reset(&chan->backend); | |
229 | /* Don't reset switch/read timer interval */ | |
230 | /* Don't reset notifiers and notifier enable bits */ | |
231 | /* Don't reset reader reference count */ | |
232 | } | |
852c2936 | 233 | |
a9ff648c MD |
234 | static |
235 | void init_crash_abi(const struct lttng_ust_lib_ring_buffer_config *config, | |
236 | struct lttng_crash_abi *crash_abi, | |
237 | struct lttng_ust_lib_ring_buffer *buf, | |
238 | struct channel_backend *chanb, | |
239 | struct shm_object *shmobj, | |
240 | struct lttng_ust_shm_handle *handle) | |
241 | { | |
242 | int i; | |
243 | ||
244 | for (i = 0; i < RB_CRASH_DUMP_ABI_MAGIC_LEN; i++) | |
245 | crash_abi->magic[i] = lttng_crash_magic_xor[i] ^ 0xFF; | |
246 | crash_abi->mmap_length = shmobj->memory_map_size; | |
247 | crash_abi->endian = RB_CRASH_ENDIAN; | |
248 | crash_abi->major = RB_CRASH_DUMP_ABI_MAJOR; | |
249 | crash_abi->minor = RB_CRASH_DUMP_ABI_MINOR; | |
250 | crash_abi->word_size = sizeof(unsigned long); | |
251 | crash_abi->layout_type = LTTNG_CRASH_TYPE_UST; | |
252 | ||
253 | /* Offset of fields */ | |
254 | crash_abi->offset.prod_offset = | |
255 | (uint32_t) ((char *) &buf->offset - (char *) buf); | |
256 | crash_abi->offset.consumed_offset = | |
257 | (uint32_t) ((char *) &buf->consumed - (char *) buf); | |
258 | crash_abi->offset.commit_hot_array = | |
259 | (uint32_t) ((char *) shmp(handle, buf->commit_hot) - (char *) buf); | |
260 | crash_abi->offset.commit_hot_seq = | |
261 | offsetof(struct commit_counters_hot, seq); | |
262 | crash_abi->offset.buf_wsb_array = | |
263 | (uint32_t) ((char *) shmp(handle, buf->backend.buf_wsb) - (char *) buf); | |
264 | crash_abi->offset.buf_wsb_id = | |
265 | offsetof(struct lttng_ust_lib_ring_buffer_backend_subbuffer, id); | |
266 | crash_abi->offset.sb_array = | |
267 | (uint32_t) ((char *) shmp(handle, buf->backend.array) - (char *) buf); | |
268 | crash_abi->offset.sb_array_shmp_offset = | |
269 | offsetof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp, | |
270 | shmp._ref.offset); | |
271 | crash_abi->offset.sb_backend_p_offset = | |
272 | offsetof(struct lttng_ust_lib_ring_buffer_backend_pages, | |
273 | p._ref.offset); | |
274 | ||
275 | /* Field length */ | |
276 | crash_abi->length.prod_offset = sizeof(buf->offset); | |
277 | crash_abi->length.consumed_offset = sizeof(buf->consumed); | |
278 | crash_abi->length.commit_hot_seq = | |
279 | sizeof(((struct commit_counters_hot *) NULL)->seq); | |
280 | crash_abi->length.buf_wsb_id = | |
281 | sizeof(((struct lttng_ust_lib_ring_buffer_backend_subbuffer *) NULL)->id); | |
282 | crash_abi->length.sb_array_shmp_offset = | |
283 | sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages_shmp *) NULL)->shmp._ref.offset); | |
284 | crash_abi->length.sb_backend_p_offset = | |
285 | sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages *) NULL)->p._ref.offset); | |
286 | ||
287 | /* Array stride */ | |
288 | crash_abi->stride.commit_hot_array = | |
289 | sizeof(struct commit_counters_hot); | |
290 | crash_abi->stride.buf_wsb_array = | |
291 | sizeof(struct lttng_ust_lib_ring_buffer_backend_subbuffer); | |
292 | crash_abi->stride.sb_array = | |
293 | sizeof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp); | |
294 | ||
295 | /* Buffer constants */ | |
296 | crash_abi->buf_size = chanb->buf_size; | |
297 | crash_abi->subbuf_size = chanb->subbuf_size; | |
298 | crash_abi->num_subbuf = chanb->num_subbuf; | |
299 | crash_abi->mode = (uint32_t) chanb->config.mode; | |
300 | ||
301 | if (config->cb.content_size_field) { | |
302 | size_t offset, length; | |
303 | ||
304 | config->cb.content_size_field(config, &offset, &length); | |
305 | crash_abi->offset.content_size = offset; | |
306 | crash_abi->length.content_size = length; | |
307 | } else { | |
308 | crash_abi->offset.content_size = 0; | |
309 | crash_abi->length.content_size = 0; | |
310 | } | |
311 | if (config->cb.packet_size_field) { | |
312 | size_t offset, length; | |
313 | ||
314 | config->cb.packet_size_field(config, &offset, &length); | |
315 | crash_abi->offset.packet_size = offset; | |
316 | crash_abi->length.packet_size = length; | |
317 | } else { | |
318 | crash_abi->offset.packet_size = 0; | |
319 | crash_abi->length.packet_size = 0; | |
320 | } | |
321 | } | |
322 | ||
852c2936 MD |
323 | /* |
324 | * Must be called under cpu hotplug protection. | |
325 | */ | |
4cfec15c | 326 | int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf, |
a6352fd4 | 327 | struct channel_backend *chanb, int cpu, |
38fae1d3 | 328 | struct lttng_ust_shm_handle *handle, |
1d498196 | 329 | struct shm_object *shmobj) |
852c2936 | 330 | { |
4cfec15c | 331 | const struct lttng_ust_lib_ring_buffer_config *config = &chanb->config; |
14641deb | 332 | struct channel *chan = caa_container_of(chanb, struct channel, backend); |
15500a1b MD |
333 | struct lttng_ust_lib_ring_buffer_backend_subbuffer *wsb; |
334 | struct channel *shmp_chan; | |
335 | struct commit_counters_hot *cc_hot; | |
a3f61e7f | 336 | void *priv = channel_get_private(chan); |
852c2936 | 337 | size_t subbuf_header_size; |
2fed87ae | 338 | uint64_t tsc; |
852c2936 MD |
339 | int ret; |
340 | ||
341 | /* Test for cpu hotplug */ | |
342 | if (buf->backend.allocated) | |
343 | return 0; | |
344 | ||
1d498196 MD |
345 | align_shm(shmobj, __alignof__(struct commit_counters_hot)); |
346 | set_shmp(buf->commit_hot, | |
347 | zalloc_shm(shmobj, | |
348 | sizeof(struct commit_counters_hot) * chan->backend.num_subbuf)); | |
349 | if (!shmp(handle, buf->commit_hot)) { | |
a9ff648c | 350 | return -ENOMEM; |
852c2936 MD |
351 | } |
352 | ||
1d498196 MD |
353 | align_shm(shmobj, __alignof__(struct commit_counters_cold)); |
354 | set_shmp(buf->commit_cold, | |
355 | zalloc_shm(shmobj, | |
356 | sizeof(struct commit_counters_cold) * chan->backend.num_subbuf)); | |
357 | if (!shmp(handle, buf->commit_cold)) { | |
852c2936 MD |
358 | ret = -ENOMEM; |
359 | goto free_commit; | |
360 | } | |
361 | ||
6c737d05 MD |
362 | align_shm(shmobj, __alignof__(uint64_t)); |
363 | set_shmp(buf->ts_end, | |
364 | zalloc_shm(shmobj, | |
365 | sizeof(uint64_t) * chan->backend.num_subbuf)); | |
366 | if (!shmp(handle, buf->ts_end)) { | |
367 | ret = -ENOMEM; | |
368 | goto free_commit_cold; | |
369 | } | |
370 | ||
371 | ||
a9ff648c MD |
372 | ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, |
373 | cpu, handle, shmobj); | |
374 | if (ret) { | |
375 | goto free_init; | |
376 | } | |
377 | ||
852c2936 MD |
378 | /* |
379 | * Write the subbuffer header for first subbuffer so we know the total | |
380 | * duration of data gathering. | |
381 | */ | |
382 | subbuf_header_size = config->cb.subbuffer_header_size(); | |
383 | v_set(config, &buf->offset, subbuf_header_size); | |
15500a1b MD |
384 | wsb = shmp_index(handle, buf->backend.buf_wsb, 0); |
385 | if (!wsb) { | |
386 | ret = -EPERM; | |
387 | goto free_chanbuf; | |
388 | } | |
389 | subbuffer_id_clear_noref(config, &wsb->id); | |
390 | shmp_chan = shmp(handle, buf->backend.chan); | |
391 | if (!shmp_chan) { | |
392 | ret = -EPERM; | |
393 | goto free_chanbuf; | |
394 | } | |
395 | tsc = config->cb.ring_buffer_clock_read(shmp_chan); | |
1d498196 | 396 | config->cb.buffer_begin(buf, tsc, 0, handle); |
15500a1b MD |
397 | cc_hot = shmp_index(handle, buf->commit_hot, 0); |
398 | if (!cc_hot) { | |
399 | ret = -EPERM; | |
400 | goto free_chanbuf; | |
401 | } | |
402 | v_add(config, subbuf_header_size, &cc_hot->cc); | |
403 | v_add(config, subbuf_header_size, &cc_hot->seq); | |
852c2936 MD |
404 | |
405 | if (config->cb.buffer_create) { | |
1d498196 | 406 | ret = config->cb.buffer_create(buf, priv, cpu, chanb->name, handle); |
852c2936 | 407 | if (ret) |
a9ff648c | 408 | goto free_chanbuf; |
852c2936 | 409 | } |
a9ff648c MD |
410 | |
411 | init_crash_abi(config, &buf->crash_abi, buf, chanb, shmobj, handle); | |
412 | ||
852c2936 | 413 | buf->backend.allocated = 1; |
852c2936 MD |
414 | return 0; |
415 | ||
416 | /* Error handling */ | |
417 | free_init: | |
6c737d05 MD |
418 | /* ts_end will be freed by shm teardown */ |
419 | free_commit_cold: | |
a6352fd4 | 420 | /* commit_cold will be freed by shm teardown */ |
852c2936 | 421 | free_commit: |
a6352fd4 | 422 | /* commit_hot will be freed by shm teardown */ |
852c2936 | 423 | free_chanbuf: |
852c2936 MD |
424 | return ret; |
425 | } | |
426 | ||
03d2d293 MD |
427 | static |
428 | void lib_ring_buffer_channel_switch_timer(int sig, siginfo_t *si, void *uc) | |
852c2936 | 429 | { |
03d2d293 MD |
430 | const struct lttng_ust_lib_ring_buffer_config *config; |
431 | struct lttng_ust_shm_handle *handle; | |
432 | struct channel *chan; | |
433 | int cpu; | |
434 | ||
435 | assert(CMM_LOAD_SHARED(timer_signal.tid) == pthread_self()); | |
436 | ||
437 | chan = si->si_value.sival_ptr; | |
438 | handle = chan->handle; | |
439 | config = &chan->backend.config; | |
852c2936 | 440 | |
34a91bdb | 441 | DBG("Switch timer for channel %p\n", chan); |
03d2d293 | 442 | |
4dcd7e80 MD |
443 | /* |
444 | * Only flush buffers periodically if readers are active. | |
445 | */ | |
cb7378b3 | 446 | pthread_mutex_lock(&wakeup_fd_mutex); |
03d2d293 MD |
447 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
448 | for_each_possible_cpu(cpu) { | |
449 | struct lttng_ust_lib_ring_buffer *buf = | |
450 | shmp(handle, chan->backend.buf[cpu].shmp); | |
34daae3e MD |
451 | |
452 | if (!buf) | |
15500a1b | 453 | goto end; |
4dcd7e80 MD |
454 | if (uatomic_read(&buf->active_readers)) |
455 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, | |
456 | chan->handle); | |
03d2d293 MD |
457 | } |
458 | } else { | |
459 | struct lttng_ust_lib_ring_buffer *buf = | |
460 | shmp(handle, chan->backend.buf[0].shmp); | |
461 | ||
34daae3e | 462 | if (!buf) |
15500a1b | 463 | goto end; |
34a91bdb MD |
464 | if (uatomic_read(&buf->active_readers)) |
465 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, | |
466 | chan->handle); | |
467 | } | |
15500a1b | 468 | end: |
34a91bdb MD |
469 | pthread_mutex_unlock(&wakeup_fd_mutex); |
470 | return; | |
471 | } | |
472 | ||
b07cd987 MD |
473 | static |
474 | int lib_ring_buffer_poll_deliver(const struct lttng_ust_lib_ring_buffer_config *config, | |
475 | struct lttng_ust_lib_ring_buffer *buf, | |
476 | struct channel *chan, | |
477 | struct lttng_ust_shm_handle *handle) | |
478 | { | |
479 | unsigned long consumed_old, consumed_idx, commit_count, write_offset; | |
15500a1b | 480 | struct commit_counters_cold *cc_cold; |
b07cd987 MD |
481 | |
482 | consumed_old = uatomic_read(&buf->consumed); | |
483 | consumed_idx = subbuf_index(consumed_old, chan); | |
15500a1b MD |
484 | cc_cold = shmp_index(handle, buf->commit_cold, consumed_idx); |
485 | if (!cc_cold) | |
486 | return 0; | |
487 | commit_count = v_read(config, &cc_cold->cc_sb); | |
b07cd987 MD |
488 | /* |
489 | * No memory barrier here, since we are only interested | |
490 | * in a statistically correct polling result. The next poll will | |
491 | * get the data is we are racing. The mb() that ensures correct | |
492 | * memory order is in get_subbuf. | |
493 | */ | |
494 | write_offset = v_read(config, &buf->offset); | |
495 | ||
496 | /* | |
497 | * Check that the subbuffer we are trying to consume has been | |
498 | * already fully committed. | |
499 | */ | |
500 | ||
501 | if (((commit_count - chan->backend.subbuf_size) | |
502 | & chan->commit_count_mask) | |
503 | - (buf_trunc(consumed_old, chan) | |
504 | >> chan->backend.num_subbuf_order) | |
505 | != 0) | |
506 | return 0; | |
507 | ||
508 | /* | |
509 | * Check that we are not about to read the same subbuffer in | |
510 | * which the writer head is. | |
511 | */ | |
512 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan) | |
513 | == 0) | |
514 | return 0; | |
515 | ||
516 | return 1; | |
517 | } | |
518 | ||
519 | static | |
520 | void lib_ring_buffer_wakeup(struct lttng_ust_lib_ring_buffer *buf, | |
521 | struct lttng_ust_shm_handle *handle) | |
522 | { | |
523 | int wakeup_fd = shm_get_wakeup_fd(handle, &buf->self._ref); | |
524 | sigset_t sigpipe_set, pending_set, old_set; | |
525 | int ret, sigpipe_was_pending = 0; | |
526 | ||
527 | if (wakeup_fd < 0) | |
528 | return; | |
529 | ||
530 | /* | |
531 | * Wake-up the other end by writing a null byte in the pipe | |
532 | * (non-blocking). Important note: Because writing into the | |
533 | * pipe is non-blocking (and therefore we allow dropping wakeup | |
534 | * data, as long as there is wakeup data present in the pipe | |
535 | * buffer to wake up the consumer), the consumer should perform | |
536 | * the following sequence for waiting: | |
537 | * 1) empty the pipe (reads). | |
538 | * 2) check if there is data in the buffer. | |
539 | * 3) wait on the pipe (poll). | |
540 | * | |
541 | * Discard the SIGPIPE from write(), not disturbing any SIGPIPE | |
542 | * that might be already pending. If a bogus SIGPIPE is sent to | |
543 | * the entire process concurrently by a malicious user, it may | |
544 | * be simply discarded. | |
545 | */ | |
546 | ret = sigemptyset(&pending_set); | |
547 | assert(!ret); | |
548 | /* | |
549 | * sigpending returns the mask of signals that are _both_ | |
550 | * blocked for the thread _and_ pending for either the thread or | |
551 | * the entire process. | |
552 | */ | |
553 | ret = sigpending(&pending_set); | |
554 | assert(!ret); | |
555 | sigpipe_was_pending = sigismember(&pending_set, SIGPIPE); | |
556 | /* | |
557 | * If sigpipe was pending, it means it was already blocked, so | |
558 | * no need to block it. | |
559 | */ | |
560 | if (!sigpipe_was_pending) { | |
561 | ret = sigemptyset(&sigpipe_set); | |
562 | assert(!ret); | |
563 | ret = sigaddset(&sigpipe_set, SIGPIPE); | |
564 | assert(!ret); | |
565 | ret = pthread_sigmask(SIG_BLOCK, &sigpipe_set, &old_set); | |
566 | assert(!ret); | |
567 | } | |
568 | do { | |
569 | ret = write(wakeup_fd, "", 1); | |
570 | } while (ret == -1L && errno == EINTR); | |
571 | if (ret == -1L && errno == EPIPE && !sigpipe_was_pending) { | |
572 | struct timespec timeout = { 0, 0 }; | |
573 | do { | |
574 | ret = sigtimedwait(&sigpipe_set, NULL, | |
575 | &timeout); | |
576 | } while (ret == -1L && errno == EINTR); | |
577 | } | |
578 | if (!sigpipe_was_pending) { | |
579 | ret = pthread_sigmask(SIG_SETMASK, &old_set, NULL); | |
580 | assert(!ret); | |
581 | } | |
582 | } | |
583 | ||
34a91bdb MD |
584 | static |
585 | void lib_ring_buffer_channel_do_read(struct channel *chan) | |
586 | { | |
587 | const struct lttng_ust_lib_ring_buffer_config *config; | |
588 | struct lttng_ust_shm_handle *handle; | |
589 | int cpu; | |
590 | ||
591 | handle = chan->handle; | |
592 | config = &chan->backend.config; | |
593 | ||
594 | /* | |
595 | * Only flush buffers periodically if readers are active. | |
596 | */ | |
597 | pthread_mutex_lock(&wakeup_fd_mutex); | |
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 | ||
34daae3e | 603 | if (!buf) |
15500a1b | 604 | goto end; |
34a91bdb MD |
605 | if (uatomic_read(&buf->active_readers) |
606 | && lib_ring_buffer_poll_deliver(config, buf, | |
607 | chan, handle)) { | |
608 | lib_ring_buffer_wakeup(buf, handle); | |
609 | } | |
610 | } | |
611 | } else { | |
612 | struct lttng_ust_lib_ring_buffer *buf = | |
613 | shmp(handle, chan->backend.buf[0].shmp); | |
614 | ||
34daae3e | 615 | if (!buf) |
15500a1b | 616 | goto end; |
34a91bdb MD |
617 | if (uatomic_read(&buf->active_readers) |
618 | && lib_ring_buffer_poll_deliver(config, buf, | |
619 | chan, handle)) { | |
620 | lib_ring_buffer_wakeup(buf, handle); | |
621 | } | |
03d2d293 | 622 | } |
15500a1b | 623 | end: |
cb7378b3 | 624 | pthread_mutex_unlock(&wakeup_fd_mutex); |
34a91bdb MD |
625 | } |
626 | ||
627 | static | |
628 | void lib_ring_buffer_channel_read_timer(int sig, siginfo_t *si, void *uc) | |
629 | { | |
630 | struct channel *chan; | |
631 | ||
632 | assert(CMM_LOAD_SHARED(timer_signal.tid) == pthread_self()); | |
633 | chan = si->si_value.sival_ptr; | |
634 | DBG("Read timer for channel %p\n", chan); | |
635 | lib_ring_buffer_channel_do_read(chan); | |
03d2d293 MD |
636 | return; |
637 | } | |
638 | ||
639 | static | |
640 | void rb_setmask(sigset_t *mask) | |
641 | { | |
642 | int ret; | |
643 | ||
644 | ret = sigemptyset(mask); | |
645 | if (ret) { | |
646 | PERROR("sigemptyset"); | |
647 | } | |
34a91bdb MD |
648 | ret = sigaddset(mask, LTTNG_UST_RB_SIG_FLUSH); |
649 | if (ret) { | |
650 | PERROR("sigaddset"); | |
651 | } | |
652 | ret = sigaddset(mask, LTTNG_UST_RB_SIG_READ); | |
03d2d293 MD |
653 | if (ret) { |
654 | PERROR("sigaddset"); | |
655 | } | |
656 | ret = sigaddset(mask, LTTNG_UST_RB_SIG_TEARDOWN); | |
657 | if (ret) { | |
658 | PERROR("sigaddset"); | |
659 | } | |
660 | } | |
661 | ||
662 | static | |
663 | void *sig_thread(void *arg) | |
664 | { | |
665 | sigset_t mask; | |
666 | siginfo_t info; | |
667 | int signr; | |
668 | ||
669 | /* Only self thread will receive signal mask. */ | |
670 | rb_setmask(&mask); | |
671 | CMM_STORE_SHARED(timer_signal.tid, pthread_self()); | |
672 | ||
673 | for (;;) { | |
674 | signr = sigwaitinfo(&mask, &info); | |
675 | if (signr == -1) { | |
34a91bdb MD |
676 | if (errno != EINTR) |
677 | PERROR("sigwaitinfo"); | |
03d2d293 MD |
678 | continue; |
679 | } | |
34a91bdb | 680 | if (signr == LTTNG_UST_RB_SIG_FLUSH) { |
03d2d293 MD |
681 | lib_ring_buffer_channel_switch_timer(info.si_signo, |
682 | &info, NULL); | |
34a91bdb MD |
683 | } else if (signr == LTTNG_UST_RB_SIG_READ) { |
684 | lib_ring_buffer_channel_read_timer(info.si_signo, | |
685 | &info, NULL); | |
03d2d293 MD |
686 | } else if (signr == LTTNG_UST_RB_SIG_TEARDOWN) { |
687 | cmm_smp_mb(); | |
688 | CMM_STORE_SHARED(timer_signal.qs_done, 1); | |
689 | cmm_smp_mb(); | |
690 | } else { | |
691 | ERR("Unexptected signal %d\n", info.si_signo); | |
692 | } | |
693 | } | |
694 | return NULL; | |
695 | } | |
696 | ||
697 | /* | |
03d2d293 MD |
698 | * Ensure only a single thread listens on the timer signal. |
699 | */ | |
700 | static | |
701 | void lib_ring_buffer_setup_timer_thread(void) | |
702 | { | |
703 | pthread_t thread; | |
704 | int ret; | |
705 | ||
64bf51a6 | 706 | pthread_mutex_lock(&timer_signal.lock); |
03d2d293 | 707 | if (timer_signal.setup_done) |
64bf51a6 | 708 | goto end; |
03d2d293 MD |
709 | |
710 | ret = pthread_create(&thread, NULL, &sig_thread, NULL); | |
711 | if (ret) { | |
712 | errno = ret; | |
713 | PERROR("pthread_create"); | |
714 | } | |
715 | ret = pthread_detach(thread); | |
716 | if (ret) { | |
717 | errno = ret; | |
718 | PERROR("pthread_detach"); | |
719 | } | |
720 | timer_signal.setup_done = 1; | |
64bf51a6 MD |
721 | end: |
722 | pthread_mutex_unlock(&timer_signal.lock); | |
852c2936 MD |
723 | } |
724 | ||
03d2d293 | 725 | /* |
64bf51a6 | 726 | * Wait for signal-handling thread quiescent state. |
03d2d293 | 727 | */ |
64bf51a6 MD |
728 | static |
729 | void lib_ring_buffer_wait_signal_thread_qs(unsigned int signr) | |
730 | { | |
731 | sigset_t pending_set; | |
732 | int ret; | |
733 | ||
734 | /* | |
735 | * We need to be the only thread interacting with the thread | |
736 | * that manages signals for teardown synchronization. | |
737 | */ | |
738 | pthread_mutex_lock(&timer_signal.lock); | |
739 | ||
740 | /* | |
741 | * Ensure we don't have any signal queued for this channel. | |
742 | */ | |
743 | for (;;) { | |
744 | ret = sigemptyset(&pending_set); | |
745 | if (ret == -1) { | |
746 | PERROR("sigemptyset"); | |
747 | } | |
748 | ret = sigpending(&pending_set); | |
749 | if (ret == -1) { | |
750 | PERROR("sigpending"); | |
751 | } | |
752 | if (!sigismember(&pending_set, signr)) | |
753 | break; | |
754 | caa_cpu_relax(); | |
755 | } | |
756 | ||
757 | /* | |
758 | * From this point, no new signal handler will be fired that | |
759 | * would try to access "chan". However, we still need to wait | |
760 | * for any currently executing handler to complete. | |
761 | */ | |
762 | cmm_smp_mb(); | |
763 | CMM_STORE_SHARED(timer_signal.qs_done, 0); | |
764 | cmm_smp_mb(); | |
765 | ||
766 | /* | |
767 | * Kill with LTTNG_UST_RB_SIG_TEARDOWN, so signal management | |
768 | * thread wakes up. | |
769 | */ | |
770 | kill(getpid(), LTTNG_UST_RB_SIG_TEARDOWN); | |
771 | ||
772 | while (!CMM_LOAD_SHARED(timer_signal.qs_done)) | |
773 | caa_cpu_relax(); | |
774 | cmm_smp_mb(); | |
775 | ||
776 | pthread_mutex_unlock(&timer_signal.lock); | |
777 | } | |
778 | ||
03d2d293 MD |
779 | static |
780 | void lib_ring_buffer_channel_switch_timer_start(struct channel *chan) | |
852c2936 | 781 | { |
03d2d293 MD |
782 | struct sigevent sev; |
783 | struct itimerspec its; | |
784 | int ret; | |
852c2936 | 785 | |
03d2d293 | 786 | if (!chan->switch_timer_interval || chan->switch_timer_enabled) |
852c2936 MD |
787 | return; |
788 | ||
03d2d293 MD |
789 | chan->switch_timer_enabled = 1; |
790 | ||
791 | lib_ring_buffer_setup_timer_thread(); | |
792 | ||
ba702529 | 793 | memset(&sev, 0, sizeof(sev)); |
03d2d293 | 794 | sev.sigev_notify = SIGEV_SIGNAL; |
34a91bdb | 795 | sev.sigev_signo = LTTNG_UST_RB_SIG_FLUSH; |
03d2d293 MD |
796 | sev.sigev_value.sival_ptr = chan; |
797 | ret = timer_create(CLOCKID, &sev, &chan->switch_timer); | |
798 | if (ret == -1) { | |
799 | PERROR("timer_create"); | |
800 | } | |
801 | ||
802 | its.it_value.tv_sec = chan->switch_timer_interval / 1000000; | |
f8921d33 | 803 | its.it_value.tv_nsec = (chan->switch_timer_interval % 1000000) * 1000; |
03d2d293 MD |
804 | its.it_interval.tv_sec = its.it_value.tv_sec; |
805 | its.it_interval.tv_nsec = its.it_value.tv_nsec; | |
806 | ||
807 | ret = timer_settime(chan->switch_timer, 0, &its, NULL); | |
808 | if (ret == -1) { | |
809 | PERROR("timer_settime"); | |
810 | } | |
811 | } | |
812 | ||
03d2d293 MD |
813 | static |
814 | void lib_ring_buffer_channel_switch_timer_stop(struct channel *chan) | |
815 | { | |
34a91bdb | 816 | int ret; |
03d2d293 MD |
817 | |
818 | if (!chan->switch_timer_interval || !chan->switch_timer_enabled) | |
819 | return; | |
820 | ||
821 | ret = timer_delete(chan->switch_timer); | |
822 | if (ret == -1) { | |
823 | PERROR("timer_delete"); | |
824 | } | |
825 | ||
64bf51a6 | 826 | lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_FLUSH); |
03d2d293 MD |
827 | |
828 | chan->switch_timer = 0; | |
829 | chan->switch_timer_enabled = 0; | |
852c2936 MD |
830 | } |
831 | ||
34a91bdb MD |
832 | static |
833 | void lib_ring_buffer_channel_read_timer_start(struct channel *chan) | |
852c2936 | 834 | { |
4cfec15c | 835 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
34a91bdb MD |
836 | struct sigevent sev; |
837 | struct itimerspec its; | |
838 | int ret; | |
852c2936 | 839 | |
34a91bdb MD |
840 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER |
841 | || !chan->read_timer_interval || chan->read_timer_enabled) | |
842 | return; | |
852c2936 | 843 | |
34a91bdb | 844 | chan->read_timer_enabled = 1; |
852c2936 | 845 | |
34a91bdb | 846 | lib_ring_buffer_setup_timer_thread(); |
852c2936 | 847 | |
34a91bdb MD |
848 | sev.sigev_notify = SIGEV_SIGNAL; |
849 | sev.sigev_signo = LTTNG_UST_RB_SIG_READ; | |
850 | sev.sigev_value.sival_ptr = chan; | |
851 | ret = timer_create(CLOCKID, &sev, &chan->read_timer); | |
852 | if (ret == -1) { | |
853 | PERROR("timer_create"); | |
854 | } | |
852c2936 | 855 | |
34a91bdb | 856 | its.it_value.tv_sec = chan->read_timer_interval / 1000000; |
f8921d33 | 857 | its.it_value.tv_nsec = (chan->read_timer_interval % 1000000) * 1000; |
34a91bdb MD |
858 | its.it_interval.tv_sec = its.it_value.tv_sec; |
859 | its.it_interval.tv_nsec = its.it_value.tv_nsec; | |
852c2936 | 860 | |
34a91bdb MD |
861 | ret = timer_settime(chan->read_timer, 0, &its, NULL); |
862 | if (ret == -1) { | |
863 | PERROR("timer_settime"); | |
864 | } | |
852c2936 MD |
865 | } |
866 | ||
34a91bdb MD |
867 | static |
868 | void lib_ring_buffer_channel_read_timer_stop(struct channel *chan) | |
852c2936 | 869 | { |
4cfec15c | 870 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
34a91bdb | 871 | int ret; |
852c2936 MD |
872 | |
873 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
34a91bdb | 874 | || !chan->read_timer_interval || !chan->read_timer_enabled) |
852c2936 MD |
875 | return; |
876 | ||
34a91bdb MD |
877 | ret = timer_delete(chan->read_timer); |
878 | if (ret == -1) { | |
879 | PERROR("timer_delete"); | |
880 | } | |
881 | ||
852c2936 MD |
882 | /* |
883 | * do one more check to catch data that has been written in the last | |
884 | * timer period. | |
885 | */ | |
34a91bdb MD |
886 | lib_ring_buffer_channel_do_read(chan); |
887 | ||
64bf51a6 | 888 | lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_READ); |
34a91bdb MD |
889 | |
890 | chan->read_timer = 0; | |
891 | chan->read_timer_enabled = 0; | |
852c2936 MD |
892 | } |
893 | ||
1d498196 | 894 | static void channel_unregister_notifiers(struct channel *chan, |
38fae1d3 | 895 | struct lttng_ust_shm_handle *handle) |
852c2936 | 896 | { |
03d2d293 | 897 | lib_ring_buffer_channel_switch_timer_stop(chan); |
34a91bdb | 898 | lib_ring_buffer_channel_read_timer_stop(chan); |
852c2936 MD |
899 | } |
900 | ||
009769ca MD |
901 | static void channel_print_errors(struct channel *chan, |
902 | struct lttng_ust_shm_handle *handle) | |
903 | { | |
904 | const struct lttng_ust_lib_ring_buffer_config *config = | |
905 | &chan->backend.config; | |
906 | int cpu; | |
907 | ||
908 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
909 | for_each_possible_cpu(cpu) { | |
910 | struct lttng_ust_lib_ring_buffer *buf = | |
911 | shmp(handle, chan->backend.buf[cpu].shmp); | |
15500a1b MD |
912 | if (buf) |
913 | lib_ring_buffer_print_errors(chan, buf, cpu, handle); | |
009769ca MD |
914 | } |
915 | } else { | |
916 | struct lttng_ust_lib_ring_buffer *buf = | |
917 | shmp(handle, chan->backend.buf[0].shmp); | |
918 | ||
15500a1b MD |
919 | if (buf) |
920 | lib_ring_buffer_print_errors(chan, buf, -1, handle); | |
009769ca MD |
921 | } |
922 | } | |
923 | ||
924 | static void channel_free(struct channel *chan, | |
6548fca4 MD |
925 | struct lttng_ust_shm_handle *handle, |
926 | int consumer) | |
852c2936 | 927 | { |
74d81a6c | 928 | channel_backend_free(&chan->backend, handle); |
431d5cf0 | 929 | /* chan is freed by shm teardown */ |
6548fca4 | 930 | shm_object_table_destroy(handle->table, consumer); |
1d498196 | 931 | free(handle); |
852c2936 MD |
932 | } |
933 | ||
934 | /** | |
935 | * channel_create - Create channel. | |
936 | * @config: ring buffer instance configuration | |
937 | * @name: name of the channel | |
a3f61e7f MD |
938 | * @priv_data: ring buffer client private data area pointer (output) |
939 | * @priv_data_size: length, in bytes, of the private data area. | |
d028eddb | 940 | * @priv_data_init: initialization data for private data. |
852c2936 MD |
941 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous |
942 | * address mapping. It is used only by RING_BUFFER_STATIC | |
943 | * configuration. It can be set to NULL for other backends. | |
944 | * @subbuf_size: subbuffer size | |
945 | * @num_subbuf: number of subbuffers | |
946 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
947 | * padding to let readers get those sub-buffers. | |
948 | * Used for live streaming. | |
949 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
5ea386c3 MD |
950 | * @stream_fds: array of stream file descriptors. |
951 | * @nr_stream_fds: number of file descriptors in array. | |
852c2936 MD |
952 | * |
953 | * Holds cpu hotplug. | |
954 | * Returns NULL on failure. | |
955 | */ | |
4cfec15c | 956 | struct lttng_ust_shm_handle *channel_create(const struct lttng_ust_lib_ring_buffer_config *config, |
a3f61e7f MD |
957 | const char *name, |
958 | void **priv_data, | |
959 | size_t priv_data_align, | |
960 | size_t priv_data_size, | |
d028eddb | 961 | void *priv_data_init, |
a3f61e7f | 962 | void *buf_addr, size_t subbuf_size, |
852c2936 | 963 | size_t num_subbuf, unsigned int switch_timer_interval, |
a9ff648c | 964 | unsigned int read_timer_interval, |
b2c5f61a MD |
965 | const int *stream_fds, int nr_stream_fds, |
966 | int64_t blocking_timeout) | |
852c2936 | 967 | { |
24622edc | 968 | int ret; |
a3f61e7f | 969 | size_t shmsize, chansize; |
852c2936 | 970 | struct channel *chan; |
38fae1d3 | 971 | struct lttng_ust_shm_handle *handle; |
1d498196 | 972 | struct shm_object *shmobj; |
74d81a6c | 973 | unsigned int nr_streams; |
b2c5f61a | 974 | int64_t blocking_timeout_ms; |
74d81a6c MD |
975 | |
976 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
977 | nr_streams = num_possible_cpus(); | |
978 | else | |
979 | nr_streams = 1; | |
852c2936 | 980 | |
5ea386c3 MD |
981 | if (nr_stream_fds != nr_streams) |
982 | return NULL; | |
983 | ||
b2c5f61a MD |
984 | if (blocking_timeout < -1) { |
985 | return NULL; | |
986 | } | |
987 | /* usec to msec */ | |
988 | if (blocking_timeout == -1) { | |
989 | blocking_timeout_ms = -1; | |
990 | } else { | |
991 | blocking_timeout_ms = blocking_timeout / 1000; | |
992 | if (blocking_timeout_ms != (int32_t) blocking_timeout_ms) { | |
993 | return NULL; | |
994 | } | |
995 | } | |
996 | ||
852c2936 MD |
997 | if (lib_ring_buffer_check_config(config, switch_timer_interval, |
998 | read_timer_interval)) | |
999 | return NULL; | |
1000 | ||
38fae1d3 | 1001 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
431d5cf0 MD |
1002 | if (!handle) |
1003 | return NULL; | |
1004 | ||
1d498196 MD |
1005 | /* Allocate table for channel + per-cpu buffers */ |
1006 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
1007 | if (!handle->table) | |
1008 | goto error_table_alloc; | |
852c2936 | 1009 | |
1d498196 MD |
1010 | /* Calculate the shm allocation layout */ |
1011 | shmsize = sizeof(struct channel); | |
b72687b8 | 1012 | shmsize += lttng_ust_offset_align(shmsize, __alignof__(struct lttng_ust_lib_ring_buffer_shmp)); |
74d81a6c | 1013 | shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp) * nr_streams; |
a3f61e7f | 1014 | chansize = shmsize; |
74d81a6c | 1015 | if (priv_data_align) |
b72687b8 | 1016 | shmsize += lttng_ust_offset_align(shmsize, priv_data_align); |
a3f61e7f | 1017 | shmsize += priv_data_size; |
a6352fd4 | 1018 | |
74d81a6c | 1019 | /* Allocate normal memory for channel (not shared) */ |
a9ff648c | 1020 | shmobj = shm_object_table_alloc(handle->table, shmsize, SHM_OBJECT_MEM, |
4b68c31f | 1021 | -1, -1); |
b5a14697 MD |
1022 | if (!shmobj) |
1023 | goto error_append; | |
57773204 | 1024 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
a3f61e7f | 1025 | set_shmp(handle->chan, zalloc_shm(shmobj, chansize)); |
57773204 MD |
1026 | assert(handle->chan._ref.index == 0); |
1027 | assert(handle->chan._ref.offset == 0); | |
1d498196 | 1028 | chan = shmp(handle, handle->chan); |
a6352fd4 | 1029 | if (!chan) |
1d498196 | 1030 | goto error_append; |
74d81a6c | 1031 | chan->nr_streams = nr_streams; |
a6352fd4 | 1032 | |
a3f61e7f MD |
1033 | /* space for private data */ |
1034 | if (priv_data_size) { | |
1035 | DECLARE_SHMP(void, priv_data_alloc); | |
1036 | ||
1037 | align_shm(shmobj, priv_data_align); | |
1038 | chan->priv_data_offset = shmobj->allocated_len; | |
1039 | set_shmp(priv_data_alloc, zalloc_shm(shmobj, priv_data_size)); | |
1040 | if (!shmp(handle, priv_data_alloc)) | |
1041 | goto error_append; | |
1042 | *priv_data = channel_get_private(chan); | |
d028eddb | 1043 | memcpy(*priv_data, priv_data_init, priv_data_size); |
a3f61e7f MD |
1044 | } else { |
1045 | chan->priv_data_offset = -1; | |
74d81a6c MD |
1046 | if (priv_data) |
1047 | *priv_data = NULL; | |
a3f61e7f MD |
1048 | } |
1049 | ||
b2c5f61a MD |
1050 | chan->u.s.blocking_timeout_ms = (int32_t) blocking_timeout_ms; |
1051 | ||
a3f61e7f | 1052 | ret = channel_backend_init(&chan->backend, name, config, |
a9ff648c | 1053 | subbuf_size, num_subbuf, handle, |
5ea386c3 | 1054 | stream_fds); |
852c2936 | 1055 | if (ret) |
1d498196 | 1056 | goto error_backend_init; |
852c2936 | 1057 | |
03d2d293 | 1058 | chan->handle = handle; |
852c2936 | 1059 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); |
852c2936 | 1060 | |
34a91bdb MD |
1061 | chan->switch_timer_interval = switch_timer_interval; |
1062 | chan->read_timer_interval = read_timer_interval; | |
03d2d293 | 1063 | lib_ring_buffer_channel_switch_timer_start(chan); |
34a91bdb | 1064 | lib_ring_buffer_channel_read_timer_start(chan); |
852c2936 | 1065 | |
431d5cf0 | 1066 | return handle; |
852c2936 | 1067 | |
1d498196 MD |
1068 | error_backend_init: |
1069 | error_append: | |
6548fca4 | 1070 | shm_object_table_destroy(handle->table, 1); |
1d498196 | 1071 | error_table_alloc: |
431d5cf0 | 1072 | free(handle); |
852c2936 MD |
1073 | return NULL; |
1074 | } | |
852c2936 | 1075 | |
74d81a6c | 1076 | struct lttng_ust_shm_handle *channel_handle_create(void *data, |
ff0f5728 MD |
1077 | uint64_t memory_map_size, |
1078 | int wakeup_fd) | |
193183fb | 1079 | { |
38fae1d3 | 1080 | struct lttng_ust_shm_handle *handle; |
193183fb MD |
1081 | struct shm_object *object; |
1082 | ||
38fae1d3 | 1083 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
193183fb MD |
1084 | if (!handle) |
1085 | return NULL; | |
1086 | ||
1087 | /* Allocate table for channel + per-cpu buffers */ | |
1088 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
1089 | if (!handle->table) | |
1090 | goto error_table_alloc; | |
1091 | /* Add channel object */ | |
74d81a6c | 1092 | object = shm_object_table_append_mem(handle->table, data, |
ff0f5728 | 1093 | memory_map_size, wakeup_fd); |
193183fb MD |
1094 | if (!object) |
1095 | goto error_table_object; | |
57773204 MD |
1096 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
1097 | handle->chan._ref.index = 0; | |
1098 | handle->chan._ref.offset = 0; | |
193183fb MD |
1099 | return handle; |
1100 | ||
1101 | error_table_object: | |
6548fca4 | 1102 | shm_object_table_destroy(handle->table, 0); |
193183fb MD |
1103 | error_table_alloc: |
1104 | free(handle); | |
1105 | return NULL; | |
1106 | } | |
1107 | ||
38fae1d3 | 1108 | int channel_handle_add_stream(struct lttng_ust_shm_handle *handle, |
74d81a6c MD |
1109 | int shm_fd, int wakeup_fd, uint32_t stream_nr, |
1110 | uint64_t memory_map_size) | |
193183fb MD |
1111 | { |
1112 | struct shm_object *object; | |
1113 | ||
1114 | /* Add stream object */ | |
74d81a6c MD |
1115 | object = shm_object_table_append_shm(handle->table, |
1116 | shm_fd, wakeup_fd, stream_nr, | |
1117 | memory_map_size); | |
193183fb | 1118 | if (!object) |
74d81a6c | 1119 | return -EINVAL; |
193183fb MD |
1120 | return 0; |
1121 | } | |
1122 | ||
74d81a6c MD |
1123 | unsigned int channel_handle_get_nr_streams(struct lttng_ust_shm_handle *handle) |
1124 | { | |
1125 | assert(handle->table); | |
1126 | return handle->table->allocated_len - 1; | |
1127 | } | |
1128 | ||
852c2936 | 1129 | static |
6548fca4 MD |
1130 | void channel_release(struct channel *chan, struct lttng_ust_shm_handle *handle, |
1131 | int consumer) | |
852c2936 | 1132 | { |
6548fca4 | 1133 | channel_free(chan, handle, consumer); |
852c2936 MD |
1134 | } |
1135 | ||
1136 | /** | |
1137 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
1138 | * @chan: channel to destroy | |
1139 | * | |
1140 | * Holds cpu hotplug. | |
431d5cf0 MD |
1141 | * Call "destroy" callback, finalize channels, decrement the channel |
1142 | * reference count. Note that when readers have completed data | |
1143 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
a60af3a5 | 1144 | * They should release their handle at that point. |
852c2936 | 1145 | */ |
a3f61e7f | 1146 | void channel_destroy(struct channel *chan, struct lttng_ust_shm_handle *handle, |
74d81a6c | 1147 | int consumer) |
852c2936 | 1148 | { |
74d81a6c MD |
1149 | if (consumer) { |
1150 | /* | |
1151 | * Note: the consumer takes care of finalizing and | |
1152 | * switching the buffers. | |
1153 | */ | |
1154 | channel_unregister_notifiers(chan, handle); | |
3d0ef9f6 MD |
1155 | /* |
1156 | * The consumer prints errors. | |
1157 | */ | |
1158 | channel_print_errors(chan, handle); | |
824f40b8 MD |
1159 | } |
1160 | ||
431d5cf0 MD |
1161 | /* |
1162 | * sessiond/consumer are keeping a reference on the shm file | |
1163 | * descriptor directly. No need to refcount. | |
1164 | */ | |
6548fca4 | 1165 | channel_release(chan, handle, consumer); |
a3f61e7f | 1166 | return; |
852c2936 | 1167 | } |
852c2936 | 1168 | |
4cfec15c MD |
1169 | struct lttng_ust_lib_ring_buffer *channel_get_ring_buffer( |
1170 | const struct lttng_ust_lib_ring_buffer_config *config, | |
1d498196 | 1171 | struct channel *chan, int cpu, |
38fae1d3 | 1172 | struct lttng_ust_shm_handle *handle, |
74d81a6c MD |
1173 | int *shm_fd, int *wait_fd, |
1174 | int *wakeup_fd, | |
1175 | uint64_t *memory_map_size) | |
852c2936 | 1176 | { |
381c0f1e MD |
1177 | struct shm_ref *ref; |
1178 | ||
1179 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
74d81a6c | 1180 | cpu = 0; |
381c0f1e | 1181 | } else { |
e095d803 MD |
1182 | if (cpu >= num_possible_cpus()) |
1183 | return NULL; | |
381c0f1e | 1184 | } |
74d81a6c MD |
1185 | ref = &chan->backend.buf[cpu].shmp._ref; |
1186 | *shm_fd = shm_get_shm_fd(handle, ref); | |
1187 | *wait_fd = shm_get_wait_fd(handle, ref); | |
1188 | *wakeup_fd = shm_get_wakeup_fd(handle, ref); | |
1189 | if (shm_get_shm_size(handle, ref, memory_map_size)) | |
1190 | return NULL; | |
1191 | return shmp(handle, chan->backend.buf[cpu].shmp); | |
852c2936 | 1192 | } |
852c2936 | 1193 | |
ff0f5728 MD |
1194 | int ring_buffer_channel_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config *config, |
1195 | struct channel *chan, | |
1196 | struct lttng_ust_shm_handle *handle) | |
1197 | { | |
1198 | struct shm_ref *ref; | |
1199 | ||
1200 | ref = &handle->chan._ref; | |
1201 | return shm_close_wait_fd(handle, ref); | |
1202 | } | |
1203 | ||
1204 | int ring_buffer_channel_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config *config, | |
1205 | struct channel *chan, | |
1206 | struct lttng_ust_shm_handle *handle) | |
1207 | { | |
1208 | struct shm_ref *ref; | |
1209 | ||
1210 | ref = &handle->chan._ref; | |
1211 | return shm_close_wakeup_fd(handle, ref); | |
1212 | } | |
1213 | ||
1214 | int ring_buffer_stream_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config *config, | |
74d81a6c MD |
1215 | struct channel *chan, |
1216 | struct lttng_ust_shm_handle *handle, | |
1217 | int cpu) | |
852c2936 | 1218 | { |
74d81a6c MD |
1219 | struct shm_ref *ref; |
1220 | ||
1221 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
1222 | cpu = 0; | |
1223 | } else { | |
1224 | if (cpu >= num_possible_cpus()) | |
1225 | return -EINVAL; | |
824f40b8 | 1226 | } |
74d81a6c MD |
1227 | ref = &chan->backend.buf[cpu].shmp._ref; |
1228 | return shm_close_wait_fd(handle, ref); | |
1229 | } | |
1230 | ||
ff0f5728 | 1231 | int ring_buffer_stream_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config *config, |
74d81a6c MD |
1232 | struct channel *chan, |
1233 | struct lttng_ust_shm_handle *handle, | |
1234 | int cpu) | |
1235 | { | |
1236 | struct shm_ref *ref; | |
cb7378b3 | 1237 | int ret; |
74d81a6c MD |
1238 | |
1239 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
1240 | cpu = 0; | |
1241 | } else { | |
1242 | if (cpu >= num_possible_cpus()) | |
1243 | return -EINVAL; | |
1244 | } | |
1245 | ref = &chan->backend.buf[cpu].shmp._ref; | |
cb7378b3 MD |
1246 | pthread_mutex_lock(&wakeup_fd_mutex); |
1247 | ret = shm_close_wakeup_fd(handle, ref); | |
1248 | pthread_mutex_unlock(&wakeup_fd_mutex); | |
1249 | return ret; | |
74d81a6c MD |
1250 | } |
1251 | ||
1252 | int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer *buf, | |
1253 | struct lttng_ust_shm_handle *handle) | |
1254 | { | |
a6352fd4 | 1255 | if (uatomic_cmpxchg(&buf->active_readers, 0, 1) != 0) |
852c2936 | 1256 | return -EBUSY; |
a6352fd4 | 1257 | cmm_smp_mb(); |
852c2936 MD |
1258 | return 0; |
1259 | } | |
852c2936 | 1260 | |
4cfec15c | 1261 | void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer *buf, |
74d81a6c | 1262 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1263 | { |
1d498196 | 1264 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 | 1265 | |
15500a1b MD |
1266 | if (!chan) |
1267 | return; | |
a6352fd4 MD |
1268 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
1269 | cmm_smp_mb(); | |
1270 | uatomic_dec(&buf->active_readers); | |
852c2936 MD |
1271 | } |
1272 | ||
1273 | /** | |
1274 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
1275 | * @buf: ring buffer | |
1276 | * @consumed: consumed count indicating the position where to read | |
1277 | * @produced: produced count, indicates position when to stop reading | |
1278 | * | |
1279 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
1280 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 MD |
1281 | */ |
1282 | ||
4cfec15c | 1283 | int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 1284 | unsigned long *consumed, unsigned long *produced, |
38fae1d3 | 1285 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1286 | { |
15500a1b MD |
1287 | struct channel *chan; |
1288 | const struct lttng_ust_lib_ring_buffer_config *config; | |
852c2936 MD |
1289 | unsigned long consumed_cur, write_offset; |
1290 | int finalized; | |
1291 | ||
15500a1b MD |
1292 | chan = shmp(handle, buf->backend.chan); |
1293 | if (!chan) | |
1294 | return -EPERM; | |
1295 | config = &chan->backend.config; | |
14641deb | 1296 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
1297 | /* |
1298 | * Read finalized before counters. | |
1299 | */ | |
a6352fd4 MD |
1300 | cmm_smp_rmb(); |
1301 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 MD |
1302 | /* |
1303 | * No need to issue a memory barrier between consumed count read and | |
1304 | * write offset read, because consumed count can only change | |
1305 | * concurrently in overwrite mode, and we keep a sequence counter | |
1306 | * identifier derived from the write offset to check we are getting | |
1307 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
1308 | * "tagged" upon writes, tags are checked upon read). | |
1309 | */ | |
1310 | write_offset = v_read(config, &buf->offset); | |
1311 | ||
1312 | /* | |
1313 | * Check that we are not about to read the same subbuffer in | |
1314 | * which the writer head is. | |
1315 | */ | |
1316 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
1317 | == 0) | |
1318 | goto nodata; | |
1319 | ||
1320 | *consumed = consumed_cur; | |
1321 | *produced = subbuf_trunc(write_offset, chan); | |
1322 | ||
1323 | return 0; | |
1324 | ||
1325 | nodata: | |
1326 | /* | |
1327 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1328 | * of "raw_spin_is_locked" memory ordering. | |
1329 | */ | |
1330 | if (finalized) | |
1331 | return -ENODATA; | |
852c2936 MD |
1332 | else |
1333 | return -EAGAIN; | |
1334 | } | |
852c2936 | 1335 | |
f45930b7 JG |
1336 | /** |
1337 | * Performs the same function as lib_ring_buffer_snapshot(), but the positions | |
1338 | * are saved regardless of whether the consumed and produced positions are | |
1339 | * in the same subbuffer. | |
1340 | * @buf: ring buffer | |
1341 | * @consumed: consumed byte count indicating the last position read | |
1342 | * @produced: produced byte count indicating the last position written | |
1343 | * | |
1344 | * This function is meant to provide information on the exact producer and | |
1345 | * consumer positions without regard for the "snapshot" feature. | |
1346 | */ | |
1347 | int lib_ring_buffer_snapshot_sample_positions( | |
1348 | struct lttng_ust_lib_ring_buffer *buf, | |
1349 | unsigned long *consumed, unsigned long *produced, | |
1350 | struct lttng_ust_shm_handle *handle) | |
1351 | { | |
1352 | struct channel *chan; | |
1353 | const struct lttng_ust_lib_ring_buffer_config *config; | |
f45930b7 JG |
1354 | |
1355 | chan = shmp(handle, buf->backend.chan); | |
1356 | if (!chan) | |
1357 | return -EPERM; | |
1358 | config = &chan->backend.config; | |
1359 | cmm_smp_rmb(); | |
1360 | *consumed = uatomic_read(&buf->consumed); | |
1361 | /* | |
1362 | * No need to issue a memory barrier between consumed count read and | |
1363 | * write offset read, because consumed count can only change | |
1364 | * concurrently in overwrite mode, and we keep a sequence counter | |
1365 | * identifier derived from the write offset to check we are getting | |
1366 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
1367 | * "tagged" upon writes, tags are checked upon read). | |
1368 | */ | |
1369 | *produced = v_read(config, &buf->offset); | |
1370 | return 0; | |
1371 | } | |
1372 | ||
852c2936 | 1373 | /** |
d1a996f5 | 1374 | * lib_ring_buffer_move_consumer - move consumed counter forward |
852c2936 MD |
1375 | * @buf: ring buffer |
1376 | * @consumed_new: new consumed count value | |
1377 | */ | |
4cfec15c | 1378 | void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 1379 | unsigned long consumed_new, |
38fae1d3 | 1380 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1381 | { |
4cfec15c | 1382 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
15500a1b | 1383 | struct channel *chan; |
852c2936 MD |
1384 | unsigned long consumed; |
1385 | ||
15500a1b MD |
1386 | chan = shmp(handle, bufb->chan); |
1387 | if (!chan) | |
1388 | return; | |
74d81a6c | 1389 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
852c2936 MD |
1390 | |
1391 | /* | |
1392 | * Only push the consumed value forward. | |
1393 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
1394 | * the writer in flight recorder mode. | |
1395 | */ | |
a6352fd4 | 1396 | consumed = uatomic_read(&buf->consumed); |
852c2936 | 1397 | while ((long) consumed - (long) consumed_new < 0) |
a6352fd4 MD |
1398 | consumed = uatomic_cmpxchg(&buf->consumed, consumed, |
1399 | consumed_new); | |
852c2936 | 1400 | } |
852c2936 MD |
1401 | |
1402 | /** | |
1403 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
1404 | * @buf: ring buffer | |
1405 | * @consumed: consumed count indicating the position where to read | |
1406 | * | |
1407 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
1408 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 | 1409 | */ |
4cfec15c | 1410 | int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 1411 | unsigned long consumed, |
38fae1d3 | 1412 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1413 | { |
15500a1b MD |
1414 | struct channel *chan; |
1415 | const struct lttng_ust_lib_ring_buffer_config *config; | |
852c2936 | 1416 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
dddba398 | 1417 | int ret, finalized, nr_retry = LTTNG_UST_RING_BUFFER_GET_RETRY; |
15500a1b | 1418 | struct commit_counters_cold *cc_cold; |
852c2936 | 1419 | |
15500a1b MD |
1420 | chan = shmp(handle, buf->backend.chan); |
1421 | if (!chan) | |
1422 | return -EPERM; | |
1423 | config = &chan->backend.config; | |
852c2936 | 1424 | retry: |
14641deb | 1425 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
1426 | /* |
1427 | * Read finalized before counters. | |
1428 | */ | |
a6352fd4 MD |
1429 | cmm_smp_rmb(); |
1430 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 | 1431 | consumed_idx = subbuf_index(consumed, chan); |
15500a1b MD |
1432 | cc_cold = shmp_index(handle, buf->commit_cold, consumed_idx); |
1433 | if (!cc_cold) | |
1434 | return -EPERM; | |
1435 | commit_count = v_read(config, &cc_cold->cc_sb); | |
852c2936 MD |
1436 | /* |
1437 | * Make sure we read the commit count before reading the buffer | |
1438 | * data and the write offset. Correct consumed offset ordering | |
1439 | * wrt commit count is insured by the use of cmpxchg to update | |
1440 | * the consumed offset. | |
852c2936 | 1441 | */ |
a6352fd4 MD |
1442 | /* |
1443 | * Local rmb to match the remote wmb to read the commit count | |
1444 | * before the buffer data and the write offset. | |
1445 | */ | |
1446 | cmm_smp_rmb(); | |
852c2936 MD |
1447 | |
1448 | write_offset = v_read(config, &buf->offset); | |
1449 | ||
1450 | /* | |
1451 | * Check that the buffer we are getting is after or at consumed_cur | |
1452 | * position. | |
1453 | */ | |
1454 | if ((long) subbuf_trunc(consumed, chan) | |
1455 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
1456 | goto nodata; | |
1457 | ||
1458 | /* | |
1459 | * Check that the subbuffer we are trying to consume has been | |
dddba398 MD |
1460 | * already fully committed. There are a few causes that can make |
1461 | * this unavailability situation occur: | |
1462 | * | |
1463 | * Temporary (short-term) situation: | |
1464 | * - Application is running on a different CPU, between reserve | |
1465 | * and commit ring buffer operations, | |
1466 | * - Application is preempted between reserve and commit ring | |
1467 | * buffer operations, | |
1468 | * | |
1469 | * Long-term situation: | |
1470 | * - Application is stopped (SIGSTOP) between reserve and commit | |
1471 | * ring buffer operations. Could eventually be resumed by | |
1472 | * SIGCONT. | |
1473 | * - Application is killed (SIGTERM, SIGINT, SIGKILL) between | |
1474 | * reserve and commit ring buffer operation. | |
1475 | * | |
1476 | * From a consumer perspective, handling short-term | |
1477 | * unavailability situations is performed by retrying a few | |
1478 | * times after a delay. Handling long-term unavailability | |
1479 | * situations is handled by failing to get the sub-buffer. | |
1480 | * | |
1481 | * In all of those situations, if the application is taking a | |
1482 | * long time to perform its commit after ring buffer space | |
1483 | * reservation, we can end up in a situation where the producer | |
1484 | * will fill the ring buffer and try to write into the same | |
1485 | * sub-buffer again (which has a missing commit). This is | |
1486 | * handled by the producer in the sub-buffer switch handling | |
1487 | * code of the reserve routine by detecting unbalanced | |
1488 | * reserve/commit counters and discarding all further events | |
1489 | * until the situation is resolved in those situations. Two | |
1490 | * scenarios can occur: | |
1491 | * | |
1492 | * 1) The application causing the reserve/commit counters to be | |
1493 | * unbalanced has been terminated. In this situation, all | |
1494 | * further events will be discarded in the buffers, and no | |
1495 | * further buffer data will be readable by the consumer | |
1496 | * daemon. Tearing down the UST tracing session and starting | |
1497 | * anew is a work-around for those situations. Note that this | |
1498 | * only affects per-UID tracing. In per-PID tracing, the | |
1499 | * application vanishes with the termination, and therefore | |
1500 | * no more data needs to be written to the buffers. | |
1501 | * 2) The application causing the unbalance has been delayed for | |
1502 | * a long time, but will eventually try to increment the | |
1503 | * commit counter after eventually writing to the sub-buffer. | |
1504 | * This situation can cause events to be discarded until the | |
1505 | * application resumes its operations. | |
852c2936 MD |
1506 | */ |
1507 | if (((commit_count - chan->backend.subbuf_size) | |
1508 | & chan->commit_count_mask) | |
0bade047 | 1509 | - (buf_trunc(consumed, chan) |
852c2936 | 1510 | >> chan->backend.num_subbuf_order) |
dddba398 MD |
1511 | != 0) { |
1512 | if (nr_retry-- > 0) { | |
1513 | if (nr_retry <= (LTTNG_UST_RING_BUFFER_GET_RETRY >> 1)) | |
1514 | (void) poll(NULL, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS); | |
1515 | goto retry; | |
1516 | } else { | |
1517 | goto nodata; | |
1518 | } | |
1519 | } | |
852c2936 MD |
1520 | |
1521 | /* | |
1522 | * Check that we are not about to read the same subbuffer in | |
1523 | * which the writer head is. | |
1524 | */ | |
0bade047 | 1525 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed, chan) |
852c2936 MD |
1526 | == 0) |
1527 | goto nodata; | |
1528 | ||
1529 | /* | |
1530 | * Failure to get the subbuffer causes a busy-loop retry without going | |
1531 | * to a wait queue. These are caused by short-lived race windows where | |
1532 | * the writer is getting access to a subbuffer we were trying to get | |
1533 | * access to. Also checks that the "consumed" buffer count we are | |
1534 | * looking for matches the one contained in the subbuffer id. | |
dddba398 MD |
1535 | * |
1536 | * The short-lived race window described here can be affected by | |
1537 | * application signals and preemption, thus requiring to bound | |
1538 | * the loop to a maximum number of retry. | |
852c2936 MD |
1539 | */ |
1540 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
1541 | consumed_idx, buf_trunc_val(consumed, chan), |
1542 | handle); | |
dddba398 MD |
1543 | if (ret) { |
1544 | if (nr_retry-- > 0) { | |
1545 | if (nr_retry <= (LTTNG_UST_RING_BUFFER_GET_RETRY >> 1)) | |
1546 | (void) poll(NULL, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS); | |
1547 | goto retry; | |
1548 | } else { | |
1549 | goto nodata; | |
1550 | } | |
1551 | } | |
852c2936 MD |
1552 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); |
1553 | ||
1554 | buf->get_subbuf_consumed = consumed; | |
1555 | buf->get_subbuf = 1; | |
1556 | ||
1557 | return 0; | |
1558 | ||
1559 | nodata: | |
1560 | /* | |
1561 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1562 | * of "raw_spin_is_locked" memory ordering. | |
1563 | */ | |
1564 | if (finalized) | |
1565 | return -ENODATA; | |
852c2936 MD |
1566 | else |
1567 | return -EAGAIN; | |
1568 | } | |
852c2936 MD |
1569 | |
1570 | /** | |
1571 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
1572 | * @buf: ring buffer | |
1573 | */ | |
4cfec15c | 1574 | void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 1575 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1576 | { |
4cfec15c | 1577 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
15500a1b MD |
1578 | struct channel *chan; |
1579 | const struct lttng_ust_lib_ring_buffer_config *config; | |
1580 | unsigned long sb_bindex, consumed_idx, consumed; | |
1581 | struct lttng_ust_lib_ring_buffer_backend_pages_shmp *rpages; | |
1582 | struct lttng_ust_lib_ring_buffer_backend_pages *backend_pages; | |
852c2936 | 1583 | |
15500a1b MD |
1584 | chan = shmp(handle, bufb->chan); |
1585 | if (!chan) | |
1586 | return; | |
1587 | config = &chan->backend.config; | |
74d81a6c | 1588 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
852c2936 MD |
1589 | |
1590 | if (!buf->get_subbuf) { | |
1591 | /* | |
1592 | * Reader puts a subbuffer it did not get. | |
1593 | */ | |
1594 | CHAN_WARN_ON(chan, 1); | |
1595 | return; | |
1596 | } | |
1597 | consumed = buf->get_subbuf_consumed; | |
1598 | buf->get_subbuf = 0; | |
1599 | ||
1600 | /* | |
1601 | * Clear the records_unread counter. (overruns counter) | |
1602 | * Can still be non-zero if a file reader simply grabbed the data | |
1603 | * without using iterators. | |
1604 | * Can be below zero if an iterator is used on a snapshot more than | |
1605 | * once. | |
1606 | */ | |
15500a1b MD |
1607 | sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); |
1608 | rpages = shmp_index(handle, bufb->array, sb_bindex); | |
1609 | if (!rpages) | |
1610 | return; | |
1611 | backend_pages = shmp(handle, rpages->shmp); | |
1612 | if (!backend_pages) | |
1613 | return; | |
1614 | v_add(config, v_read(config, &backend_pages->records_unread), | |
1615 | &bufb->records_read); | |
1616 | v_set(config, &backend_pages->records_unread, 0); | |
852c2936 MD |
1617 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE |
1618 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
1619 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
1620 | ||
1621 | /* | |
1622 | * Exchange the reader subbuffer with the one we put in its place in the | |
1623 | * writer subbuffer table. Expect the original consumed count. If | |
1624 | * update_read_sb_index fails, this is because the writer updated the | |
1625 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
1626 | * currently have: it has become invalid to try reading this sub-buffer | |
1627 | * consumed count value anyway. | |
1628 | */ | |
1629 | consumed_idx = subbuf_index(consumed, chan); | |
1630 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
1631 | consumed_idx, buf_trunc_val(consumed, chan), |
1632 | handle); | |
852c2936 MD |
1633 | /* |
1634 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
1635 | * if the writer concurrently updated it. | |
1636 | */ | |
1637 | } | |
852c2936 MD |
1638 | |
1639 | /* | |
1640 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
1641 | * position and the writer position. (inclusive) | |
1642 | */ | |
1643 | static | |
4cfec15c | 1644 | void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1645 | struct channel *chan, |
1646 | unsigned long cons_offset, | |
1d498196 | 1647 | int cpu, |
38fae1d3 | 1648 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1649 | { |
4cfec15c | 1650 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 | 1651 | unsigned long cons_idx, commit_count, commit_count_sb; |
15500a1b MD |
1652 | struct commit_counters_hot *cc_hot; |
1653 | struct commit_counters_cold *cc_cold; | |
852c2936 MD |
1654 | |
1655 | cons_idx = subbuf_index(cons_offset, chan); | |
15500a1b MD |
1656 | cc_hot = shmp_index(handle, buf->commit_hot, cons_idx); |
1657 | if (!cc_hot) | |
1658 | return; | |
1659 | cc_cold = shmp_index(handle, buf->commit_cold, cons_idx); | |
1660 | if (!cc_cold) | |
1661 | return; | |
1662 | commit_count = v_read(config, &cc_hot->cc); | |
1663 | commit_count_sb = v_read(config, &cc_cold->cc_sb); | |
852c2936 MD |
1664 | |
1665 | if (subbuf_offset(commit_count, chan) != 0) | |
4d3c9523 | 1666 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
1667 | "commit count in subbuffer %lu,\n" |
1668 | "expecting multiples of %lu bytes\n" | |
1669 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
1670 | chan->backend.name, cpu, cons_idx, | |
1671 | chan->backend.subbuf_size, | |
1672 | commit_count, commit_count_sb); | |
1673 | ||
4d3c9523 | 1674 | DBG("ring buffer: %s, cpu %d: %lu bytes committed\n", |
852c2936 MD |
1675 | chan->backend.name, cpu, commit_count); |
1676 | } | |
1677 | ||
1678 | static | |
4cfec15c | 1679 | void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 | 1680 | struct channel *chan, |
1d498196 | 1681 | void *priv, int cpu, |
38fae1d3 | 1682 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1683 | { |
4cfec15c | 1684 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1685 | unsigned long write_offset, cons_offset; |
1686 | ||
852c2936 MD |
1687 | /* |
1688 | * No need to order commit_count, write_offset and cons_offset reads | |
1689 | * because we execute at teardown when no more writer nor reader | |
1690 | * references are left. | |
1691 | */ | |
1692 | write_offset = v_read(config, &buf->offset); | |
a6352fd4 | 1693 | cons_offset = uatomic_read(&buf->consumed); |
852c2936 | 1694 | if (write_offset != cons_offset) |
4d3c9523 | 1695 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
1696 | "non-consumed data\n" |
1697 | " [ %lu bytes written, %lu bytes read ]\n", | |
1698 | chan->backend.name, cpu, write_offset, cons_offset); | |
1699 | ||
a6352fd4 | 1700 | for (cons_offset = uatomic_read(&buf->consumed); |
852c2936 MD |
1701 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), |
1702 | chan) | |
1703 | - cons_offset) > 0; | |
1704 | cons_offset = subbuf_align(cons_offset, chan)) | |
1705 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1d498196 | 1706 | cpu, handle); |
852c2936 MD |
1707 | } |
1708 | ||
1709 | static | |
1710 | void lib_ring_buffer_print_errors(struct channel *chan, | |
009769ca MD |
1711 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
1712 | struct lttng_ust_shm_handle *handle) | |
852c2936 | 1713 | { |
4cfec15c | 1714 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
a3f61e7f | 1715 | void *priv = channel_get_private(chan); |
852c2936 | 1716 | |
a1360615 MD |
1717 | if (!strcmp(chan->backend.name, "relay-metadata-mmap")) { |
1718 | DBG("ring buffer %s: %lu records written, " | |
1719 | "%lu records overrun\n", | |
1720 | chan->backend.name, | |
1721 | v_read(config, &buf->records_count), | |
1722 | v_read(config, &buf->records_overrun)); | |
1723 | } else { | |
1724 | DBG("ring buffer %s, cpu %d: %lu records written, " | |
1725 | "%lu records overrun\n", | |
1726 | chan->backend.name, cpu, | |
1727 | v_read(config, &buf->records_count), | |
1728 | v_read(config, &buf->records_overrun)); | |
1729 | ||
1730 | if (v_read(config, &buf->records_lost_full) | |
1731 | || v_read(config, &buf->records_lost_wrap) | |
1732 | || v_read(config, &buf->records_lost_big)) | |
1733 | DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n" | |
1734 | " [ %lu buffer full, %lu nest buffer wrap-around, " | |
1735 | "%lu event too big ]\n", | |
1736 | chan->backend.name, cpu, | |
1737 | v_read(config, &buf->records_lost_full), | |
1738 | v_read(config, &buf->records_lost_wrap), | |
1739 | v_read(config, &buf->records_lost_big)); | |
1740 | } | |
1d498196 | 1741 | lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu, handle); |
852c2936 MD |
1742 | } |
1743 | ||
1744 | /* | |
1745 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
1746 | * | |
48f22436 MD |
1747 | * Only executed by SWITCH_FLUSH, which can be issued while tracing is |
1748 | * active or at buffer finalization (destroy). | |
852c2936 MD |
1749 | */ |
1750 | static | |
4cfec15c | 1751 | void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1752 | struct channel *chan, |
1753 | struct switch_offsets *offsets, | |
2fed87ae | 1754 | uint64_t tsc, |
38fae1d3 | 1755 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1756 | { |
4cfec15c | 1757 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1758 | unsigned long oldidx = subbuf_index(offsets->old, chan); |
1759 | unsigned long commit_count; | |
3d3c3833 | 1760 | struct commit_counters_hot *cc_hot; |
852c2936 | 1761 | |
1d498196 | 1762 | config->cb.buffer_begin(buf, tsc, oldidx, handle); |
852c2936 MD |
1763 | |
1764 | /* | |
1765 | * Order all writes to buffer before the commit count update that will | |
1766 | * determine that the subbuffer is full. | |
1767 | */ | |
a6352fd4 | 1768 | cmm_smp_wmb(); |
3d3c3833 | 1769 | cc_hot = shmp_index(handle, buf->commit_hot, oldidx); |
15500a1b MD |
1770 | if (!cc_hot) |
1771 | return; | |
852c2936 | 1772 | v_add(config, config->cb.subbuffer_header_size(), |
3d3c3833 MD |
1773 | &cc_hot->cc); |
1774 | commit_count = v_read(config, &cc_hot->cc); | |
852c2936 MD |
1775 | /* Check if the written buffer has to be delivered */ |
1776 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
1b7b0501 | 1777 | commit_count, oldidx, handle, tsc); |
d2fe4771 | 1778 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
80249235 | 1779 | offsets->old + config->cb.subbuffer_header_size(), |
3d3c3833 | 1780 | commit_count, handle, cc_hot); |
852c2936 MD |
1781 | } |
1782 | ||
1783 | /* | |
1784 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
1785 | * | |
1786 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
1787 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
1788 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
1789 | * subbuffer. | |
1790 | */ | |
1791 | static | |
4cfec15c | 1792 | void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1793 | struct channel *chan, |
1794 | struct switch_offsets *offsets, | |
2fed87ae | 1795 | uint64_t tsc, |
38fae1d3 | 1796 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1797 | { |
4cfec15c | 1798 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1799 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); |
1800 | unsigned long commit_count, padding_size, data_size; | |
3d3c3833 | 1801 | struct commit_counters_hot *cc_hot; |
6c737d05 | 1802 | uint64_t *ts_end; |
852c2936 MD |
1803 | |
1804 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1805 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1806 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size, |
1807 | handle); | |
852c2936 | 1808 | |
6c737d05 MD |
1809 | ts_end = shmp_index(handle, buf->ts_end, oldidx); |
1810 | if (!ts_end) | |
1811 | return; | |
852c2936 | 1812 | /* |
6c737d05 MD |
1813 | * This is the last space reservation in that sub-buffer before |
1814 | * it gets delivered. This provides exclusive access to write to | |
1815 | * this sub-buffer's ts_end. There are also no concurrent | |
1816 | * readers of that ts_end because delivery of that sub-buffer is | |
1817 | * postponed until the commit counter is incremented for the | |
1818 | * current space reservation. | |
1819 | */ | |
1820 | *ts_end = tsc; | |
1821 | ||
1822 | /* | |
1823 | * Order all writes to buffer and store to ts_end before the commit | |
1824 | * count update that will determine that the subbuffer is full. | |
852c2936 | 1825 | */ |
a6352fd4 | 1826 | cmm_smp_wmb(); |
3d3c3833 | 1827 | cc_hot = shmp_index(handle, buf->commit_hot, oldidx); |
15500a1b MD |
1828 | if (!cc_hot) |
1829 | return; | |
3d3c3833 MD |
1830 | v_add(config, padding_size, &cc_hot->cc); |
1831 | commit_count = v_read(config, &cc_hot->cc); | |
852c2936 | 1832 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, |
1b7b0501 | 1833 | commit_count, oldidx, handle, tsc); |
d2fe4771 MD |
1834 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
1835 | offsets->old + padding_size, commit_count, handle, | |
3d3c3833 | 1836 | cc_hot); |
852c2936 MD |
1837 | } |
1838 | ||
1839 | /* | |
1840 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1841 | * | |
1842 | * This code can be executed unordered : writers may already have written to the | |
1843 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1844 | * that this code is executed before the deliver of this sub-buffer. | |
1845 | */ | |
1846 | static | |
4cfec15c | 1847 | void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1848 | struct channel *chan, |
1849 | struct switch_offsets *offsets, | |
2fed87ae | 1850 | uint64_t tsc, |
38fae1d3 | 1851 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1852 | { |
4cfec15c | 1853 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1854 | unsigned long beginidx = subbuf_index(offsets->begin, chan); |
1855 | unsigned long commit_count; | |
3d3c3833 | 1856 | struct commit_counters_hot *cc_hot; |
852c2936 | 1857 | |
1d498196 | 1858 | config->cb.buffer_begin(buf, tsc, beginidx, handle); |
852c2936 MD |
1859 | |
1860 | /* | |
1861 | * Order all writes to buffer before the commit count update that will | |
1862 | * determine that the subbuffer is full. | |
1863 | */ | |
a6352fd4 | 1864 | cmm_smp_wmb(); |
3d3c3833 | 1865 | cc_hot = shmp_index(handle, buf->commit_hot, beginidx); |
15500a1b MD |
1866 | if (!cc_hot) |
1867 | return; | |
1868 | v_add(config, config->cb.subbuffer_header_size(), &cc_hot->cc); | |
3d3c3833 | 1869 | commit_count = v_read(config, &cc_hot->cc); |
852c2936 MD |
1870 | /* Check if the written buffer has to be delivered */ |
1871 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
1b7b0501 | 1872 | commit_count, beginidx, handle, tsc); |
d2fe4771 | 1873 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
80249235 | 1874 | offsets->begin + config->cb.subbuffer_header_size(), |
3d3c3833 | 1875 | commit_count, handle, cc_hot); |
852c2936 MD |
1876 | } |
1877 | ||
1ad21f70 MD |
1878 | /* |
1879 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1880 | * | |
3962118d MD |
1881 | * Calls subbuffer_set_data_size() to set the data size of the current |
1882 | * sub-buffer. We do not need to perform check_deliver nor commit here, | |
1883 | * since this task will be done by the "commit" of the event for which | |
1884 | * we are currently doing the space reservation. | |
1ad21f70 MD |
1885 | */ |
1886 | static | |
1887 | void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer *buf, | |
1888 | struct channel *chan, | |
1889 | struct switch_offsets *offsets, | |
1890 | uint64_t tsc, | |
1891 | struct lttng_ust_shm_handle *handle) | |
1892 | { | |
1893 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; | |
3962118d | 1894 | unsigned long endidx, data_size; |
6c737d05 | 1895 | uint64_t *ts_end; |
1ad21f70 | 1896 | |
3962118d | 1897 | endidx = subbuf_index(offsets->end - 1, chan); |
1ad21f70 | 1898 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; |
1ad21f70 MD |
1899 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size, |
1900 | handle); | |
6c737d05 MD |
1901 | ts_end = shmp_index(handle, buf->ts_end, endidx); |
1902 | if (!ts_end) | |
1903 | return; | |
1904 | /* | |
1905 | * This is the last space reservation in that sub-buffer before | |
1906 | * it gets delivered. This provides exclusive access to write to | |
1907 | * this sub-buffer's ts_end. There are also no concurrent | |
1908 | * readers of that ts_end because delivery of that sub-buffer is | |
1909 | * postponed until the commit counter is incremented for the | |
1910 | * current space reservation. | |
1911 | */ | |
1912 | *ts_end = tsc; | |
1ad21f70 MD |
1913 | } |
1914 | ||
852c2936 MD |
1915 | /* |
1916 | * Returns : | |
1917 | * 0 if ok | |
1918 | * !0 if execution must be aborted. | |
1919 | */ | |
1920 | static | |
1921 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
4cfec15c | 1922 | struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1923 | struct channel *chan, |
1924 | struct switch_offsets *offsets, | |
2dac206d MD |
1925 | uint64_t *tsc, |
1926 | struct lttng_ust_shm_handle *handle) | |
852c2936 | 1927 | { |
4cfec15c | 1928 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
2dac206d | 1929 | unsigned long off, reserve_commit_diff; |
852c2936 MD |
1930 | |
1931 | offsets->begin = v_read(config, &buf->offset); | |
1932 | offsets->old = offsets->begin; | |
1933 | offsets->switch_old_start = 0; | |
1934 | off = subbuf_offset(offsets->begin, chan); | |
1935 | ||
1936 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1937 | ||
1938 | /* | |
1939 | * Ensure we flush the header of an empty subbuffer when doing the | |
1940 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1941 | * total data gathering duration even if there were no records saved | |
1942 | * after the last buffer switch. | |
1943 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1944 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1945 | * subbuffer header as appropriate. | |
1946 | * The next record that reserves space will be responsible for | |
1947 | * populating the following subbuffer header. We choose not to populate | |
1948 | * the next subbuffer header here because we want to be able to use | |
a6352fd4 MD |
1949 | * SWITCH_ACTIVE for periodical buffer flush, which must |
1950 | * guarantee that all the buffer content (records and header | |
1951 | * timestamps) are visible to the reader. This is required for | |
1952 | * quiescence guarantees for the fusion merge. | |
852c2936 | 1953 | */ |
2dac206d MD |
1954 | if (mode != SWITCH_FLUSH && !off) |
1955 | return -1; /* we do not have to switch : buffer is empty */ | |
1956 | ||
1957 | if (caa_unlikely(off == 0)) { | |
1958 | unsigned long sb_index, commit_count; | |
15500a1b | 1959 | struct commit_counters_cold *cc_cold; |
2dac206d MD |
1960 | |
1961 | /* | |
48f22436 MD |
1962 | * We are performing a SWITCH_FLUSH. There may be concurrent |
1963 | * writes into the buffer if e.g. invoked while performing a | |
1964 | * snapshot on an active trace. | |
2dac206d | 1965 | * |
48f22436 MD |
1966 | * If the client does not save any header information |
1967 | * (sub-buffer header size == 0), don't switch empty subbuffer | |
1968 | * on finalize, because it is invalid to deliver a completely | |
1969 | * empty subbuffer. | |
2dac206d MD |
1970 | */ |
1971 | if (!config->cb.subbuffer_header_size()) | |
1972 | return -1; | |
1973 | ||
1974 | /* Test new buffer integrity */ | |
1975 | sb_index = subbuf_index(offsets->begin, chan); | |
15500a1b MD |
1976 | cc_cold = shmp_index(handle, buf->commit_cold, sb_index); |
1977 | if (!cc_cold) | |
1978 | return -1; | |
1979 | commit_count = v_read(config, &cc_cold->cc_sb); | |
2dac206d MD |
1980 | reserve_commit_diff = |
1981 | (buf_trunc(offsets->begin, chan) | |
1982 | >> chan->backend.num_subbuf_order) | |
1983 | - (commit_count & chan->commit_count_mask); | |
1984 | if (caa_likely(reserve_commit_diff == 0)) { | |
1985 | /* Next subbuffer not being written to. */ | |
1986 | if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
1987 | subbuf_trunc(offsets->begin, chan) | |
1988 | - subbuf_trunc((unsigned long) | |
1989 | uatomic_read(&buf->consumed), chan) | |
1990 | >= chan->backend.buf_size)) { | |
1991 | /* | |
1992 | * We do not overwrite non consumed buffers | |
1993 | * and we are full : don't switch. | |
1994 | */ | |
852c2936 | 1995 | return -1; |
2dac206d MD |
1996 | } else { |
1997 | /* | |
1998 | * Next subbuffer not being written to, and we | |
1999 | * are either in overwrite mode or the buffer is | |
2000 | * not full. It's safe to write in this new | |
2001 | * subbuffer. | |
2002 | */ | |
2003 | } | |
2004 | } else { | |
852c2936 | 2005 | /* |
2dac206d MD |
2006 | * Next subbuffer reserve offset does not match the |
2007 | * commit offset. Don't perform switch in | |
2008 | * producer-consumer and overwrite mode. Caused by | |
2009 | * either a writer OOPS or too many nested writes over a | |
2010 | * reserve/commit pair. | |
852c2936 | 2011 | */ |
2dac206d | 2012 | return -1; |
852c2936 | 2013 | } |
2dac206d MD |
2014 | |
2015 | /* | |
2016 | * Need to write the subbuffer start header on finalize. | |
2017 | */ | |
2018 | offsets->switch_old_start = 1; | |
2019 | } | |
2020 | offsets->begin = subbuf_align(offsets->begin, chan); | |
852c2936 MD |
2021 | /* Note: old points to the next subbuf at offset 0 */ |
2022 | offsets->end = offsets->begin; | |
2023 | return 0; | |
2024 | } | |
2025 | ||
2026 | /* | |
2027 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
2028 | * called while tracing is active with absolutely no lock held. | |
2029 | * | |
ef1a13e0 MD |
2030 | * For RING_BUFFER_SYNC_PER_CPU ring buffers, as a v_cmpxchg is used for |
2031 | * some atomic operations, this function must be called from the CPU | |
2032 | * which owns the buffer for a ACTIVE flush. However, for | |
2033 | * RING_BUFFER_SYNC_GLOBAL ring buffers, this function can be called | |
2034 | * from any CPU. | |
852c2936 | 2035 | */ |
4cfec15c | 2036 | void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode, |
38fae1d3 | 2037 | struct lttng_ust_shm_handle *handle) |
852c2936 | 2038 | { |
15500a1b MD |
2039 | struct channel *chan; |
2040 | const struct lttng_ust_lib_ring_buffer_config *config; | |
852c2936 MD |
2041 | struct switch_offsets offsets; |
2042 | unsigned long oldidx; | |
2fed87ae | 2043 | uint64_t tsc; |
852c2936 | 2044 | |
15500a1b MD |
2045 | chan = shmp(handle, buf->backend.chan); |
2046 | if (!chan) | |
2047 | return; | |
2048 | config = &chan->backend.config; | |
2049 | ||
852c2936 MD |
2050 | offsets.size = 0; |
2051 | ||
2052 | /* | |
2053 | * Perform retryable operations. | |
2054 | */ | |
2055 | do { | |
2056 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
2dac206d | 2057 | &tsc, handle)) |
852c2936 MD |
2058 | return; /* Switch not needed */ |
2059 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
2060 | != offsets.old); | |
2061 | ||
2062 | /* | |
2063 | * Atomically update last_tsc. This update races against concurrent | |
2064 | * atomic updates, but the race will always cause supplementary full TSC | |
2065 | * records, never the opposite (missing a full TSC record when it would | |
2066 | * be needed). | |
2067 | */ | |
2068 | save_last_tsc(config, buf, tsc); | |
2069 | ||
2070 | /* | |
2071 | * Push the reader if necessary | |
2072 | */ | |
2073 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
2074 | ||
2075 | oldidx = subbuf_index(offsets.old, chan); | |
1d498196 | 2076 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx, handle); |
852c2936 MD |
2077 | |
2078 | /* | |
2079 | * May need to populate header start on SWITCH_FLUSH. | |
2080 | */ | |
2081 | if (offsets.switch_old_start) { | |
1d498196 | 2082 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc, handle); |
852c2936 MD |
2083 | offsets.old += config->cb.subbuffer_header_size(); |
2084 | } | |
2085 | ||
2086 | /* | |
2087 | * Switch old subbuffer. | |
2088 | */ | |
1d498196 | 2089 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc, handle); |
852c2936 | 2090 | } |
852c2936 | 2091 | |
6f97f9c2 MD |
2092 | static |
2093 | bool handle_blocking_retry(int *timeout_left_ms) | |
2094 | { | |
2095 | int timeout = *timeout_left_ms, delay; | |
2096 | ||
2097 | if (caa_likely(!timeout)) | |
2098 | return false; /* Do not retry, discard event. */ | |
2099 | if (timeout < 0) /* Wait forever. */ | |
2100 | delay = RETRY_DELAY_MS; | |
2101 | else | |
2102 | delay = min_t(int, timeout, RETRY_DELAY_MS); | |
2103 | (void) poll(NULL, 0, delay); | |
2104 | if (timeout > 0) | |
2105 | *timeout_left_ms -= delay; | |
2106 | return true; /* Retry. */ | |
2107 | } | |
2108 | ||
852c2936 MD |
2109 | /* |
2110 | * Returns : | |
2111 | * 0 if ok | |
2112 | * -ENOSPC if event size is too large for packet. | |
2113 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
2114 | * -EIO if data cannot be written into the buffer for any other reason. | |
2115 | */ | |
2116 | static | |
4cfec15c | 2117 | int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
2118 | struct channel *chan, |
2119 | struct switch_offsets *offsets, | |
e56bb47c MD |
2120 | struct lttng_ust_lib_ring_buffer_ctx *ctx, |
2121 | void *client_ctx) | |
852c2936 | 2122 | { |
4cfec15c | 2123 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
38fae1d3 | 2124 | struct lttng_ust_shm_handle *handle = ctx->handle; |
e185cf4b | 2125 | unsigned long reserve_commit_diff, offset_cmp; |
b2c5f61a | 2126 | int timeout_left_ms = lttng_ust_ringbuffer_get_timeout(chan); |
852c2936 | 2127 | |
e185cf4b MD |
2128 | retry: |
2129 | offsets->begin = offset_cmp = v_read(config, &buf->offset); | |
852c2936 MD |
2130 | offsets->old = offsets->begin; |
2131 | offsets->switch_new_start = 0; | |
1ad21f70 | 2132 | offsets->switch_new_end = 0; |
852c2936 MD |
2133 | offsets->switch_old_end = 0; |
2134 | offsets->pre_header_padding = 0; | |
2135 | ||
2136 | ctx->tsc = config->cb.ring_buffer_clock_read(chan); | |
2137 | if ((int64_t) ctx->tsc == -EIO) | |
2138 | return -EIO; | |
2139 | ||
2140 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
2141 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
2142 | ||
b5a3dfa5 | 2143 | if (caa_unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) { |
852c2936 MD |
2144 | offsets->switch_new_start = 1; /* For offsets->begin */ |
2145 | } else { | |
2146 | offsets->size = config->cb.record_header_size(config, chan, | |
2147 | offsets->begin, | |
2148 | &offsets->pre_header_padding, | |
e56bb47c | 2149 | ctx, client_ctx); |
852c2936 MD |
2150 | offsets->size += |
2151 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
2152 | ctx->largest_align) | |
2153 | + ctx->data_size; | |
b5a3dfa5 | 2154 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) + |
852c2936 MD |
2155 | offsets->size > chan->backend.subbuf_size)) { |
2156 | offsets->switch_old_end = 1; /* For offsets->old */ | |
2157 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
2158 | } | |
2159 | } | |
b5a3dfa5 | 2160 | if (caa_unlikely(offsets->switch_new_start)) { |
e185cf4b | 2161 | unsigned long sb_index, commit_count; |
15500a1b | 2162 | struct commit_counters_cold *cc_cold; |
852c2936 MD |
2163 | |
2164 | /* | |
2165 | * We are typically not filling the previous buffer completely. | |
2166 | */ | |
b5a3dfa5 | 2167 | if (caa_likely(offsets->switch_old_end)) |
852c2936 MD |
2168 | offsets->begin = subbuf_align(offsets->begin, chan); |
2169 | offsets->begin = offsets->begin | |
2170 | + config->cb.subbuffer_header_size(); | |
2171 | /* Test new buffer integrity */ | |
2172 | sb_index = subbuf_index(offsets->begin, chan); | |
e185cf4b MD |
2173 | /* |
2174 | * Read buf->offset before buf->commit_cold[sb_index].cc_sb. | |
2175 | * lib_ring_buffer_check_deliver() has the matching | |
2176 | * memory barriers required around commit_cold cc_sb | |
2177 | * updates to ensure reserve and commit counter updates | |
2178 | * are not seen reordered when updated by another CPU. | |
2179 | */ | |
2180 | cmm_smp_rmb(); | |
15500a1b MD |
2181 | cc_cold = shmp_index(handle, buf->commit_cold, sb_index); |
2182 | if (!cc_cold) | |
2183 | return -1; | |
2184 | commit_count = v_read(config, &cc_cold->cc_sb); | |
e185cf4b MD |
2185 | /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */ |
2186 | cmm_smp_rmb(); | |
2187 | if (caa_unlikely(offset_cmp != v_read(config, &buf->offset))) { | |
2188 | /* | |
2189 | * The reserve counter have been concurrently updated | |
2190 | * while we read the commit counter. This means the | |
2191 | * commit counter we read might not match buf->offset | |
2192 | * due to concurrent update. We therefore need to retry. | |
2193 | */ | |
2194 | goto retry; | |
2195 | } | |
852c2936 MD |
2196 | reserve_commit_diff = |
2197 | (buf_trunc(offsets->begin, chan) | |
2198 | >> chan->backend.num_subbuf_order) | |
e185cf4b | 2199 | - (commit_count & chan->commit_count_mask); |
b5a3dfa5 | 2200 | if (caa_likely(reserve_commit_diff == 0)) { |
852c2936 | 2201 | /* Next subbuffer not being written to. */ |
b5a3dfa5 | 2202 | if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE && |
852c2936 MD |
2203 | subbuf_trunc(offsets->begin, chan) |
2204 | - subbuf_trunc((unsigned long) | |
a6352fd4 | 2205 | uatomic_read(&buf->consumed), chan) |
852c2936 | 2206 | >= chan->backend.buf_size)) { |
64493e4f MD |
2207 | unsigned long nr_lost; |
2208 | ||
6f97f9c2 MD |
2209 | if (handle_blocking_retry(&timeout_left_ms)) |
2210 | goto retry; | |
2211 | ||
852c2936 MD |
2212 | /* |
2213 | * We do not overwrite non consumed buffers | |
2214 | * and we are full : record is lost. | |
2215 | */ | |
64493e4f | 2216 | nr_lost = v_read(config, &buf->records_lost_full); |
852c2936 | 2217 | v_inc(config, &buf->records_lost_full); |
64493e4f MD |
2218 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
2219 | DBG("%lu or more records lost in (%s:%d) (buffer full)\n", | |
2220 | nr_lost + 1, chan->backend.name, | |
2221 | buf->backend.cpu); | |
2222 | } | |
852c2936 MD |
2223 | return -ENOBUFS; |
2224 | } else { | |
2225 | /* | |
2226 | * Next subbuffer not being written to, and we | |
2227 | * are either in overwrite mode or the buffer is | |
2228 | * not full. It's safe to write in this new | |
2229 | * subbuffer. | |
2230 | */ | |
2231 | } | |
2232 | } else { | |
64493e4f MD |
2233 | unsigned long nr_lost; |
2234 | ||
852c2936 MD |
2235 | /* |
2236 | * Next subbuffer reserve offset does not match the | |
e185cf4b MD |
2237 | * commit offset, and this did not involve update to the |
2238 | * reserve counter. Drop record in producer-consumer and | |
852c2936 MD |
2239 | * overwrite mode. Caused by either a writer OOPS or too |
2240 | * many nested writes over a reserve/commit pair. | |
2241 | */ | |
64493e4f | 2242 | nr_lost = v_read(config, &buf->records_lost_wrap); |
852c2936 | 2243 | v_inc(config, &buf->records_lost_wrap); |
64493e4f MD |
2244 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
2245 | DBG("%lu or more records lost in (%s:%d) (wrap-around)\n", | |
2246 | nr_lost + 1, chan->backend.name, | |
2247 | buf->backend.cpu); | |
2248 | } | |
852c2936 MD |
2249 | return -EIO; |
2250 | } | |
2251 | offsets->size = | |
2252 | config->cb.record_header_size(config, chan, | |
2253 | offsets->begin, | |
2254 | &offsets->pre_header_padding, | |
e56bb47c | 2255 | ctx, client_ctx); |
852c2936 MD |
2256 | offsets->size += |
2257 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
2258 | ctx->largest_align) | |
2259 | + ctx->data_size; | |
b5a3dfa5 | 2260 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) |
852c2936 | 2261 | + offsets->size > chan->backend.subbuf_size)) { |
64493e4f MD |
2262 | unsigned long nr_lost; |
2263 | ||
852c2936 MD |
2264 | /* |
2265 | * Record too big for subbuffers, report error, don't | |
2266 | * complete the sub-buffer switch. | |
2267 | */ | |
64493e4f | 2268 | nr_lost = v_read(config, &buf->records_lost_big); |
852c2936 | 2269 | v_inc(config, &buf->records_lost_big); |
64493e4f MD |
2270 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
2271 | DBG("%lu or more records lost in (%s:%d) record size " | |
2272 | " of %zu bytes is too large for buffer\n", | |
2273 | nr_lost + 1, chan->backend.name, | |
2274 | buf->backend.cpu, offsets->size); | |
2275 | } | |
852c2936 MD |
2276 | return -ENOSPC; |
2277 | } else { | |
2278 | /* | |
2279 | * We just made a successful buffer switch and the | |
2280 | * record fits in the new subbuffer. Let's write. | |
2281 | */ | |
2282 | } | |
2283 | } else { | |
2284 | /* | |
2285 | * Record fits in the current buffer and we are not on a switch | |
2286 | * boundary. It's safe to write. | |
2287 | */ | |
2288 | } | |
2289 | offsets->end = offsets->begin + offsets->size; | |
1ad21f70 MD |
2290 | |
2291 | if (caa_unlikely(subbuf_offset(offsets->end, chan) == 0)) { | |
2292 | /* | |
2293 | * The offset_end will fall at the very beginning of the next | |
2294 | * subbuffer. | |
2295 | */ | |
2296 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
2297 | } | |
852c2936 MD |
2298 | return 0; |
2299 | } | |
2300 | ||
2301 | /** | |
2302 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
2303 | * @ctx: ring buffer context. | |
2304 | * | |
2305 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, | |
2306 | * -EIO for other errors, else returns 0. | |
2307 | * It will take care of sub-buffer switching. | |
2308 | */ | |
e56bb47c MD |
2309 | int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx, |
2310 | void *client_ctx) | |
852c2936 MD |
2311 | { |
2312 | struct channel *chan = ctx->chan; | |
38fae1d3 | 2313 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c MD |
2314 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
2315 | struct lttng_ust_lib_ring_buffer *buf; | |
852c2936 MD |
2316 | struct switch_offsets offsets; |
2317 | int ret; | |
2318 | ||
2319 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1d498196 | 2320 | buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp); |
852c2936 | 2321 | else |
1d498196 | 2322 | buf = shmp(handle, chan->backend.buf[0].shmp); |
15500a1b MD |
2323 | if (!buf) |
2324 | return -EIO; | |
852c2936 MD |
2325 | ctx->buf = buf; |
2326 | ||
2327 | offsets.size = 0; | |
2328 | ||
2329 | do { | |
2330 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, | |
e56bb47c | 2331 | ctx, client_ctx); |
b5a3dfa5 | 2332 | if (caa_unlikely(ret)) |
852c2936 | 2333 | return ret; |
b5a3dfa5 | 2334 | } while (caa_unlikely(v_cmpxchg(config, &buf->offset, offsets.old, |
852c2936 MD |
2335 | offsets.end) |
2336 | != offsets.old)); | |
2337 | ||
2338 | /* | |
2339 | * Atomically update last_tsc. This update races against concurrent | |
2340 | * atomic updates, but the race will always cause supplementary full TSC | |
2341 | * records, never the opposite (missing a full TSC record when it would | |
2342 | * be needed). | |
2343 | */ | |
2344 | save_last_tsc(config, buf, ctx->tsc); | |
2345 | ||
2346 | /* | |
2347 | * Push the reader if necessary | |
2348 | */ | |
2349 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
2350 | ||
2351 | /* | |
2352 | * Clear noref flag for this subbuffer. | |
2353 | */ | |
2354 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1d498196 MD |
2355 | subbuf_index(offsets.end - 1, chan), |
2356 | handle); | |
852c2936 MD |
2357 | |
2358 | /* | |
2359 | * Switch old subbuffer if needed. | |
2360 | */ | |
b5a3dfa5 | 2361 | if (caa_unlikely(offsets.switch_old_end)) { |
852c2936 | 2362 | lib_ring_buffer_clear_noref(config, &buf->backend, |
1d498196 MD |
2363 | subbuf_index(offsets.old - 1, chan), |
2364 | handle); | |
2365 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc, handle); | |
852c2936 MD |
2366 | } |
2367 | ||
2368 | /* | |
2369 | * Populate new subbuffer. | |
2370 | */ | |
b5a3dfa5 | 2371 | if (caa_unlikely(offsets.switch_new_start)) |
1d498196 | 2372 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 | 2373 | |
1ad21f70 MD |
2374 | if (caa_unlikely(offsets.switch_new_end)) |
2375 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc, handle); | |
2376 | ||
852c2936 MD |
2377 | ctx->slot_size = offsets.size; |
2378 | ctx->pre_offset = offsets.begin; | |
2379 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
2380 | return 0; | |
2381 | } | |
f645cfa7 | 2382 | |
b07cd987 MD |
2383 | static |
2384 | void lib_ring_buffer_vmcore_check_deliver(const struct lttng_ust_lib_ring_buffer_config *config, | |
2385 | struct lttng_ust_lib_ring_buffer *buf, | |
2386 | unsigned long commit_count, | |
2387 | unsigned long idx, | |
2388 | struct lttng_ust_shm_handle *handle) | |
2389 | { | |
15500a1b MD |
2390 | struct commit_counters_hot *cc_hot; |
2391 | ||
2392 | if (config->oops != RING_BUFFER_OOPS_CONSISTENCY) | |
2393 | return; | |
2394 | cc_hot = shmp_index(handle, buf->commit_hot, idx); | |
2395 | if (!cc_hot) | |
2396 | return; | |
2397 | v_set(config, &cc_hot->seq, commit_count); | |
b07cd987 MD |
2398 | } |
2399 | ||
9c995331 MD |
2400 | /* |
2401 | * The ring buffer can count events recorded and overwritten per buffer, | |
2402 | * but it is disabled by default due to its performance overhead. | |
2403 | */ | |
2404 | #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS | |
2405 | static | |
2406 | void deliver_count_events(const struct lttng_ust_lib_ring_buffer_config *config, | |
2407 | struct lttng_ust_lib_ring_buffer *buf, | |
2408 | unsigned long idx, | |
2409 | struct lttng_ust_shm_handle *handle) | |
2410 | { | |
2411 | v_add(config, subbuffer_get_records_count(config, | |
2412 | &buf->backend, idx, handle), | |
2413 | &buf->records_count); | |
2414 | v_add(config, subbuffer_count_records_overrun(config, | |
2415 | &buf->backend, idx, handle), | |
2416 | &buf->records_overrun); | |
2417 | } | |
2418 | #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */ | |
2419 | static | |
2420 | void deliver_count_events(const struct lttng_ust_lib_ring_buffer_config *config, | |
2421 | struct lttng_ust_lib_ring_buffer *buf, | |
2422 | unsigned long idx, | |
2423 | struct lttng_ust_shm_handle *handle) | |
2424 | { | |
2425 | } | |
2426 | #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */ | |
2427 | ||
b07cd987 MD |
2428 | void lib_ring_buffer_check_deliver_slow(const struct lttng_ust_lib_ring_buffer_config *config, |
2429 | struct lttng_ust_lib_ring_buffer *buf, | |
2430 | struct channel *chan, | |
2431 | unsigned long offset, | |
2432 | unsigned long commit_count, | |
2433 | unsigned long idx, | |
2434 | struct lttng_ust_shm_handle *handle, | |
2435 | uint64_t tsc) | |
2436 | { | |
2437 | unsigned long old_commit_count = commit_count | |
2438 | - chan->backend.subbuf_size; | |
15500a1b | 2439 | struct commit_counters_cold *cc_cold; |
b07cd987 MD |
2440 | |
2441 | /* | |
2442 | * If we succeeded at updating cc_sb below, we are the subbuffer | |
2443 | * writer delivering the subbuffer. Deals with concurrent | |
2444 | * updates of the "cc" value without adding a add_return atomic | |
2445 | * operation to the fast path. | |
2446 | * | |
2447 | * We are doing the delivery in two steps: | |
2448 | * - First, we cmpxchg() cc_sb to the new value | |
2449 | * old_commit_count + 1. This ensures that we are the only | |
2450 | * subbuffer user successfully filling the subbuffer, but we | |
2451 | * do _not_ set the cc_sb value to "commit_count" yet. | |
2452 | * Therefore, other writers that would wrap around the ring | |
2453 | * buffer and try to start writing to our subbuffer would | |
2454 | * have to drop records, because it would appear as | |
2455 | * non-filled. | |
2456 | * We therefore have exclusive access to the subbuffer control | |
2457 | * structures. This mutual exclusion with other writers is | |
2458 | * crucially important to perform record overruns count in | |
2459 | * flight recorder mode locklessly. | |
2460 | * - When we are ready to release the subbuffer (either for | |
2461 | * reading or for overrun by other writers), we simply set the | |
2462 | * cc_sb value to "commit_count" and perform delivery. | |
2463 | * | |
2464 | * The subbuffer size is least 2 bytes (minimum size: 1 page). | |
2465 | * This guarantees that old_commit_count + 1 != commit_count. | |
2466 | */ | |
2467 | ||
2468 | /* | |
2469 | * Order prior updates to reserve count prior to the | |
2470 | * commit_cold cc_sb update. | |
2471 | */ | |
2472 | cmm_smp_wmb(); | |
15500a1b MD |
2473 | cc_cold = shmp_index(handle, buf->commit_cold, idx); |
2474 | if (!cc_cold) | |
2475 | return; | |
2476 | if (caa_likely(v_cmpxchg(config, &cc_cold->cc_sb, | |
b07cd987 MD |
2477 | old_commit_count, old_commit_count + 1) |
2478 | == old_commit_count)) { | |
6c737d05 MD |
2479 | uint64_t *ts_end; |
2480 | ||
b07cd987 MD |
2481 | /* |
2482 | * Start of exclusive subbuffer access. We are | |
2483 | * guaranteed to be the last writer in this subbuffer | |
2484 | * and any other writer trying to access this subbuffer | |
2485 | * in this state is required to drop records. | |
6c737d05 MD |
2486 | * |
2487 | * We can read the ts_end for the current sub-buffer | |
2488 | * which has been saved by the very last space | |
2489 | * reservation for the current sub-buffer. | |
2490 | * | |
2491 | * Order increment of commit counter before reading ts_end. | |
b07cd987 | 2492 | */ |
6c737d05 MD |
2493 | cmm_smp_mb(); |
2494 | ts_end = shmp_index(handle, buf->ts_end, idx); | |
2495 | if (!ts_end) | |
2496 | return; | |
9c995331 | 2497 | deliver_count_events(config, buf, idx, handle); |
6c737d05 | 2498 | config->cb.buffer_end(buf, *ts_end, idx, |
b07cd987 MD |
2499 | lib_ring_buffer_get_data_size(config, |
2500 | buf, | |
2501 | idx, | |
2502 | handle), | |
2503 | handle); | |
2504 | ||
2505 | /* | |
2506 | * Increment the packet counter while we have exclusive | |
2507 | * access. | |
2508 | */ | |
2509 | subbuffer_inc_packet_count(config, &buf->backend, idx, handle); | |
2510 | ||
2511 | /* | |
2512 | * Set noref flag and offset for this subbuffer id. | |
2513 | * Contains a memory barrier that ensures counter stores | |
2514 | * are ordered before set noref and offset. | |
2515 | */ | |
2516 | lib_ring_buffer_set_noref_offset(config, &buf->backend, idx, | |
2517 | buf_trunc_val(offset, chan), handle); | |
2518 | ||
2519 | /* | |
2520 | * Order set_noref and record counter updates before the | |
2521 | * end of subbuffer exclusive access. Orders with | |
2522 | * respect to writers coming into the subbuffer after | |
2523 | * wrap around, and also order wrt concurrent readers. | |
2524 | */ | |
2525 | cmm_smp_mb(); | |
2526 | /* End of exclusive subbuffer access */ | |
15500a1b | 2527 | v_set(config, &cc_cold->cc_sb, commit_count); |
b07cd987 MD |
2528 | /* |
2529 | * Order later updates to reserve count after | |
2530 | * the commit cold cc_sb update. | |
2531 | */ | |
2532 | cmm_smp_wmb(); | |
2533 | lib_ring_buffer_vmcore_check_deliver(config, buf, | |
2534 | commit_count, idx, handle); | |
2535 | ||
2536 | /* | |
2537 | * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free. | |
2538 | */ | |
2539 | if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER | |
2540 | && uatomic_read(&buf->active_readers) | |
2541 | && lib_ring_buffer_poll_deliver(config, buf, chan, handle)) { | |
2542 | lib_ring_buffer_wakeup(buf, handle); | |
2543 | } | |
2544 | } | |
2545 | } | |
2546 | ||
f645cfa7 MD |
2547 | /* |
2548 | * Force a read (imply TLS fixup for dlopen) of TLS variables. | |
2549 | */ | |
2550 | void lttng_fixup_ringbuffer_tls(void) | |
2551 | { | |
8c90a710 | 2552 | asm volatile ("" : : "m" (URCU_TLS(lib_ring_buffer_nesting))); |
f645cfa7 | 2553 | } |
03d2d293 MD |
2554 | |
2555 | void lib_ringbuffer_signal_init(void) | |
2556 | { | |
2557 | sigset_t mask; | |
2558 | int ret; | |
2559 | ||
2560 | /* | |
2561 | * Block signal for entire process, so only our thread processes | |
2562 | * it. | |
2563 | */ | |
2564 | rb_setmask(&mask); | |
2565 | ret = pthread_sigmask(SIG_BLOCK, &mask, NULL); | |
2566 | if (ret) { | |
2567 | errno = ret; | |
2568 | PERROR("pthread_sigmask"); | |
2569 | } | |
2570 | } |