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 | |
eae3c729 | 55 | #include <lttng/ust-utils.h> |
0466ac28 | 56 | #include <lttng/ringbuffer-context.h> |
3d3a2bb8 MJ |
57 | |
58 | #include "smp.h" | |
0466ac28 | 59 | #include "ringbuffer-config.h" |
2fed87ae | 60 | #include "vatomic.h" |
4931a13e MD |
61 | #include "backend.h" |
62 | #include "frontend.h" | |
a6352fd4 | 63 | #include "shm.h" |
6f97f9c2 | 64 | #include "rb-init.h" |
bdcf8d82 | 65 | #include "../liblttng-ust/compat.h" /* For ENODATA */ |
852c2936 | 66 | |
64493e4f MD |
67 | /* Print DBG() messages about events lost only every 1048576 hits */ |
68 | #define DBG_PRINT_NR_LOST (1UL << 20) | |
69 | ||
34a91bdb MD |
70 | #define LTTNG_UST_RB_SIG_FLUSH SIGRTMIN |
71 | #define LTTNG_UST_RB_SIG_READ SIGRTMIN + 1 | |
72 | #define LTTNG_UST_RB_SIG_TEARDOWN SIGRTMIN + 2 | |
03d2d293 | 73 | #define CLOCKID CLOCK_MONOTONIC |
dddba398 MD |
74 | #define LTTNG_UST_RING_BUFFER_GET_RETRY 10 |
75 | #define LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS 10 | |
6f97f9c2 | 76 | #define RETRY_DELAY_MS 100 /* 100 ms. */ |
03d2d293 | 77 | |
a9ff648c MD |
78 | /* |
79 | * Non-static to ensure the compiler does not optimize away the xor. | |
80 | */ | |
81 | uint8_t lttng_crash_magic_xor[] = RB_CRASH_DUMP_ABI_MAGIC_XOR; | |
82 | ||
2432c3c9 MD |
83 | /* |
84 | * Use POSIX SHM: shm_open(3) and shm_unlink(3). | |
85 | * close(2) to close the fd returned by shm_open. | |
86 | * shm_unlink releases the shared memory object name. | |
87 | * ftruncate(2) sets the size of the memory object. | |
88 | * mmap/munmap maps the shared memory obj to a virtual address in the | |
89 | * calling proceess (should be done both in libust and consumer). | |
90 | * See shm_overview(7) for details. | |
91 | * Pass file descriptor returned by shm_open(3) to ltt-sessiond through | |
92 | * a UNIX socket. | |
93 | * | |
94 | * Since we don't need to access the object using its name, we can | |
95 | * immediately shm_unlink(3) it, and only keep the handle with its file | |
96 | * descriptor. | |
97 | */ | |
98 | ||
852c2936 MD |
99 | /* |
100 | * Internal structure representing offsets to use at a sub-buffer switch. | |
101 | */ | |
102 | struct switch_offsets { | |
103 | unsigned long begin, end, old; | |
104 | size_t pre_header_padding, size; | |
1ad21f70 MD |
105 | unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1, |
106 | switch_old_end:1; | |
852c2936 MD |
107 | }; |
108 | ||
16adecf1 | 109 | DEFINE_URCU_TLS(unsigned int, lib_ring_buffer_nesting); |
852c2936 | 110 | |
cb7378b3 MD |
111 | /* |
112 | * wakeup_fd_mutex protects wakeup fd use by timer from concurrent | |
113 | * close. | |
114 | */ | |
115 | static pthread_mutex_t wakeup_fd_mutex = PTHREAD_MUTEX_INITIALIZER; | |
116 | ||
852c2936 | 117 | static |
5198080d | 118 | void lib_ring_buffer_print_errors(struct lttng_ust_lib_ring_buffer_channel *chan, |
009769ca MD |
119 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
120 | struct lttng_ust_shm_handle *handle); | |
852c2936 | 121 | |
03d2d293 MD |
122 | /* |
123 | * Handle timer teardown race wrt memory free of private data by | |
124 | * ring buffer signals are handled by a single thread, which permits | |
125 | * a synchronization point between handling of each signal. | |
64bf51a6 | 126 | * Protected by the lock within the structure. |
03d2d293 MD |
127 | */ |
128 | struct timer_signal_data { | |
129 | pthread_t tid; /* thread id managing signals */ | |
130 | int setup_done; | |
131 | int qs_done; | |
64bf51a6 | 132 | pthread_mutex_t lock; |
03d2d293 MD |
133 | }; |
134 | ||
64bf51a6 MD |
135 | static struct timer_signal_data timer_signal = { |
136 | .tid = 0, | |
137 | .setup_done = 0, | |
138 | .qs_done = 0, | |
139 | .lock = PTHREAD_MUTEX_INITIALIZER, | |
140 | }; | |
03d2d293 | 141 | |
b2c5f61a | 142 | static bool lttng_ust_allow_blocking; |
6f97f9c2 | 143 | |
b2c5f61a | 144 | void lttng_ust_ringbuffer_set_allow_blocking(void) |
6f97f9c2 | 145 | { |
b2c5f61a MD |
146 | lttng_ust_allow_blocking = true; |
147 | } | |
148 | ||
149 | /* Get blocking timeout, in ms */ | |
5198080d | 150 | static int lttng_ust_ringbuffer_get_timeout(struct lttng_ust_lib_ring_buffer_channel *chan) |
b2c5f61a MD |
151 | { |
152 | if (!lttng_ust_allow_blocking) | |
153 | return 0; | |
154 | return chan->u.s.blocking_timeout_ms; | |
6f97f9c2 MD |
155 | } |
156 | ||
852c2936 MD |
157 | /** |
158 | * lib_ring_buffer_reset - Reset ring buffer to initial values. | |
159 | * @buf: Ring buffer. | |
160 | * | |
161 | * Effectively empty the ring buffer. Should be called when the buffer is not | |
162 | * used for writing. The ring buffer can be opened for reading, but the reader | |
163 | * should not be using the iterator concurrently with reset. The previous | |
164 | * current iterator record is reset. | |
165 | */ | |
4cfec15c | 166 | void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 167 | struct lttng_ust_shm_handle *handle) |
852c2936 | 168 | { |
5198080d | 169 | struct lttng_ust_lib_ring_buffer_channel *chan; |
34daae3e | 170 | const struct lttng_ust_lib_ring_buffer_config *config; |
852c2936 MD |
171 | unsigned int i; |
172 | ||
34daae3e MD |
173 | chan = shmp(handle, buf->backend.chan); |
174 | if (!chan) | |
15500a1b | 175 | return; |
34daae3e | 176 | config = &chan->backend.config; |
852c2936 MD |
177 | /* |
178 | * Reset iterator first. It will put the subbuffer if it currently holds | |
179 | * it. | |
180 | */ | |
852c2936 MD |
181 | v_set(config, &buf->offset, 0); |
182 | for (i = 0; i < chan->backend.num_subbuf; i++) { | |
15500a1b MD |
183 | struct commit_counters_hot *cc_hot; |
184 | struct commit_counters_cold *cc_cold; | |
6c737d05 | 185 | uint64_t *ts_end; |
15500a1b MD |
186 | |
187 | cc_hot = shmp_index(handle, buf->commit_hot, i); | |
188 | if (!cc_hot) | |
189 | return; | |
190 | cc_cold = shmp_index(handle, buf->commit_cold, i); | |
191 | if (!cc_cold) | |
192 | return; | |
6c737d05 MD |
193 | ts_end = shmp_index(handle, buf->ts_end, i); |
194 | if (!ts_end) | |
195 | return; | |
15500a1b MD |
196 | v_set(config, &cc_hot->cc, 0); |
197 | v_set(config, &cc_hot->seq, 0); | |
198 | v_set(config, &cc_cold->cc_sb, 0); | |
6c737d05 | 199 | *ts_end = 0; |
852c2936 | 200 | } |
a6352fd4 MD |
201 | uatomic_set(&buf->consumed, 0); |
202 | uatomic_set(&buf->record_disabled, 0); | |
852c2936 | 203 | v_set(config, &buf->last_tsc, 0); |
1d498196 | 204 | lib_ring_buffer_backend_reset(&buf->backend, handle); |
852c2936 MD |
205 | /* Don't reset number of active readers */ |
206 | v_set(config, &buf->records_lost_full, 0); | |
207 | v_set(config, &buf->records_lost_wrap, 0); | |
208 | v_set(config, &buf->records_lost_big, 0); | |
209 | v_set(config, &buf->records_count, 0); | |
210 | v_set(config, &buf->records_overrun, 0); | |
211 | buf->finalized = 0; | |
212 | } | |
852c2936 MD |
213 | |
214 | /** | |
215 | * channel_reset - Reset channel to initial values. | |
216 | * @chan: Channel. | |
217 | * | |
218 | * Effectively empty the channel. Should be called when the channel is not used | |
219 | * for writing. The channel can be opened for reading, but the reader should not | |
220 | * be using the iterator concurrently with reset. The previous current iterator | |
221 | * record is reset. | |
222 | */ | |
5198080d | 223 | void channel_reset(struct lttng_ust_lib_ring_buffer_channel *chan) |
852c2936 MD |
224 | { |
225 | /* | |
226 | * Reset iterators first. Will put the subbuffer if held for reading. | |
227 | */ | |
a6352fd4 | 228 | uatomic_set(&chan->record_disabled, 0); |
852c2936 MD |
229 | /* Don't reset commit_count_mask, still valid */ |
230 | channel_backend_reset(&chan->backend); | |
231 | /* Don't reset switch/read timer interval */ | |
232 | /* Don't reset notifiers and notifier enable bits */ | |
233 | /* Don't reset reader reference count */ | |
234 | } | |
852c2936 | 235 | |
a9ff648c MD |
236 | static |
237 | void init_crash_abi(const struct lttng_ust_lib_ring_buffer_config *config, | |
238 | struct lttng_crash_abi *crash_abi, | |
239 | struct lttng_ust_lib_ring_buffer *buf, | |
240 | struct channel_backend *chanb, | |
241 | struct shm_object *shmobj, | |
242 | struct lttng_ust_shm_handle *handle) | |
243 | { | |
244 | int i; | |
245 | ||
246 | for (i = 0; i < RB_CRASH_DUMP_ABI_MAGIC_LEN; i++) | |
247 | crash_abi->magic[i] = lttng_crash_magic_xor[i] ^ 0xFF; | |
248 | crash_abi->mmap_length = shmobj->memory_map_size; | |
249 | crash_abi->endian = RB_CRASH_ENDIAN; | |
250 | crash_abi->major = RB_CRASH_DUMP_ABI_MAJOR; | |
251 | crash_abi->minor = RB_CRASH_DUMP_ABI_MINOR; | |
252 | crash_abi->word_size = sizeof(unsigned long); | |
253 | crash_abi->layout_type = LTTNG_CRASH_TYPE_UST; | |
254 | ||
255 | /* Offset of fields */ | |
256 | crash_abi->offset.prod_offset = | |
257 | (uint32_t) ((char *) &buf->offset - (char *) buf); | |
258 | crash_abi->offset.consumed_offset = | |
259 | (uint32_t) ((char *) &buf->consumed - (char *) buf); | |
260 | crash_abi->offset.commit_hot_array = | |
261 | (uint32_t) ((char *) shmp(handle, buf->commit_hot) - (char *) buf); | |
262 | crash_abi->offset.commit_hot_seq = | |
263 | offsetof(struct commit_counters_hot, seq); | |
264 | crash_abi->offset.buf_wsb_array = | |
265 | (uint32_t) ((char *) shmp(handle, buf->backend.buf_wsb) - (char *) buf); | |
266 | crash_abi->offset.buf_wsb_id = | |
267 | offsetof(struct lttng_ust_lib_ring_buffer_backend_subbuffer, id); | |
268 | crash_abi->offset.sb_array = | |
269 | (uint32_t) ((char *) shmp(handle, buf->backend.array) - (char *) buf); | |
270 | crash_abi->offset.sb_array_shmp_offset = | |
271 | offsetof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp, | |
272 | shmp._ref.offset); | |
273 | crash_abi->offset.sb_backend_p_offset = | |
274 | offsetof(struct lttng_ust_lib_ring_buffer_backend_pages, | |
275 | p._ref.offset); | |
276 | ||
277 | /* Field length */ | |
278 | crash_abi->length.prod_offset = sizeof(buf->offset); | |
279 | crash_abi->length.consumed_offset = sizeof(buf->consumed); | |
280 | crash_abi->length.commit_hot_seq = | |
281 | sizeof(((struct commit_counters_hot *) NULL)->seq); | |
282 | crash_abi->length.buf_wsb_id = | |
283 | sizeof(((struct lttng_ust_lib_ring_buffer_backend_subbuffer *) NULL)->id); | |
284 | crash_abi->length.sb_array_shmp_offset = | |
285 | sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages_shmp *) NULL)->shmp._ref.offset); | |
286 | crash_abi->length.sb_backend_p_offset = | |
287 | sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages *) NULL)->p._ref.offset); | |
288 | ||
289 | /* Array stride */ | |
290 | crash_abi->stride.commit_hot_array = | |
291 | sizeof(struct commit_counters_hot); | |
292 | crash_abi->stride.buf_wsb_array = | |
293 | sizeof(struct lttng_ust_lib_ring_buffer_backend_subbuffer); | |
294 | crash_abi->stride.sb_array = | |
295 | sizeof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp); | |
296 | ||
297 | /* Buffer constants */ | |
298 | crash_abi->buf_size = chanb->buf_size; | |
299 | crash_abi->subbuf_size = chanb->subbuf_size; | |
300 | crash_abi->num_subbuf = chanb->num_subbuf; | |
301 | crash_abi->mode = (uint32_t) chanb->config.mode; | |
302 | ||
303 | if (config->cb.content_size_field) { | |
304 | size_t offset, length; | |
305 | ||
306 | config->cb.content_size_field(config, &offset, &length); | |
307 | crash_abi->offset.content_size = offset; | |
308 | crash_abi->length.content_size = length; | |
309 | } else { | |
310 | crash_abi->offset.content_size = 0; | |
311 | crash_abi->length.content_size = 0; | |
312 | } | |
313 | if (config->cb.packet_size_field) { | |
314 | size_t offset, length; | |
315 | ||
316 | config->cb.packet_size_field(config, &offset, &length); | |
317 | crash_abi->offset.packet_size = offset; | |
318 | crash_abi->length.packet_size = length; | |
319 | } else { | |
320 | crash_abi->offset.packet_size = 0; | |
321 | crash_abi->length.packet_size = 0; | |
322 | } | |
323 | } | |
324 | ||
852c2936 MD |
325 | /* |
326 | * Must be called under cpu hotplug protection. | |
327 | */ | |
4cfec15c | 328 | int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf, |
a6352fd4 | 329 | struct channel_backend *chanb, int cpu, |
38fae1d3 | 330 | struct lttng_ust_shm_handle *handle, |
1d498196 | 331 | struct shm_object *shmobj) |
852c2936 | 332 | { |
4cfec15c | 333 | const struct lttng_ust_lib_ring_buffer_config *config = &chanb->config; |
5198080d MJ |
334 | struct lttng_ust_lib_ring_buffer_channel *chan = caa_container_of(chanb, |
335 | struct lttng_ust_lib_ring_buffer_channel, backend); | |
15500a1b | 336 | struct lttng_ust_lib_ring_buffer_backend_subbuffer *wsb; |
5198080d | 337 | struct lttng_ust_lib_ring_buffer_channel *shmp_chan; |
15500a1b | 338 | struct commit_counters_hot *cc_hot; |
f0fde1c3 | 339 | void *priv = channel_get_private_config(chan); |
852c2936 | 340 | size_t subbuf_header_size; |
2fed87ae | 341 | uint64_t tsc; |
852c2936 MD |
342 | int ret; |
343 | ||
344 | /* Test for cpu hotplug */ | |
345 | if (buf->backend.allocated) | |
346 | return 0; | |
347 | ||
1d498196 MD |
348 | align_shm(shmobj, __alignof__(struct commit_counters_hot)); |
349 | set_shmp(buf->commit_hot, | |
350 | zalloc_shm(shmobj, | |
351 | sizeof(struct commit_counters_hot) * chan->backend.num_subbuf)); | |
352 | if (!shmp(handle, buf->commit_hot)) { | |
a9ff648c | 353 | return -ENOMEM; |
852c2936 MD |
354 | } |
355 | ||
1d498196 MD |
356 | align_shm(shmobj, __alignof__(struct commit_counters_cold)); |
357 | set_shmp(buf->commit_cold, | |
358 | zalloc_shm(shmobj, | |
359 | sizeof(struct commit_counters_cold) * chan->backend.num_subbuf)); | |
360 | if (!shmp(handle, buf->commit_cold)) { | |
852c2936 MD |
361 | ret = -ENOMEM; |
362 | goto free_commit; | |
363 | } | |
364 | ||
6c737d05 MD |
365 | align_shm(shmobj, __alignof__(uint64_t)); |
366 | set_shmp(buf->ts_end, | |
367 | zalloc_shm(shmobj, | |
368 | sizeof(uint64_t) * chan->backend.num_subbuf)); | |
369 | if (!shmp(handle, buf->ts_end)) { | |
370 | ret = -ENOMEM; | |
371 | goto free_commit_cold; | |
372 | } | |
373 | ||
374 | ||
a9ff648c MD |
375 | ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, |
376 | cpu, handle, shmobj); | |
377 | if (ret) { | |
378 | goto free_init; | |
379 | } | |
380 | ||
852c2936 MD |
381 | /* |
382 | * Write the subbuffer header for first subbuffer so we know the total | |
383 | * duration of data gathering. | |
384 | */ | |
385 | subbuf_header_size = config->cb.subbuffer_header_size(); | |
386 | v_set(config, &buf->offset, subbuf_header_size); | |
15500a1b MD |
387 | wsb = shmp_index(handle, buf->backend.buf_wsb, 0); |
388 | if (!wsb) { | |
389 | ret = -EPERM; | |
390 | goto free_chanbuf; | |
391 | } | |
392 | subbuffer_id_clear_noref(config, &wsb->id); | |
393 | shmp_chan = shmp(handle, buf->backend.chan); | |
394 | if (!shmp_chan) { | |
395 | ret = -EPERM; | |
396 | goto free_chanbuf; | |
397 | } | |
398 | tsc = config->cb.ring_buffer_clock_read(shmp_chan); | |
1d498196 | 399 | config->cb.buffer_begin(buf, tsc, 0, handle); |
15500a1b MD |
400 | cc_hot = shmp_index(handle, buf->commit_hot, 0); |
401 | if (!cc_hot) { | |
402 | ret = -EPERM; | |
403 | goto free_chanbuf; | |
404 | } | |
405 | v_add(config, subbuf_header_size, &cc_hot->cc); | |
406 | v_add(config, subbuf_header_size, &cc_hot->seq); | |
852c2936 MD |
407 | |
408 | if (config->cb.buffer_create) { | |
1d498196 | 409 | ret = config->cb.buffer_create(buf, priv, cpu, chanb->name, handle); |
852c2936 | 410 | if (ret) |
a9ff648c | 411 | goto free_chanbuf; |
852c2936 | 412 | } |
a9ff648c MD |
413 | |
414 | init_crash_abi(config, &buf->crash_abi, buf, chanb, shmobj, handle); | |
415 | ||
852c2936 | 416 | buf->backend.allocated = 1; |
852c2936 MD |
417 | return 0; |
418 | ||
419 | /* Error handling */ | |
420 | free_init: | |
6c737d05 MD |
421 | /* ts_end will be freed by shm teardown */ |
422 | free_commit_cold: | |
a6352fd4 | 423 | /* commit_cold will be freed by shm teardown */ |
852c2936 | 424 | free_commit: |
a6352fd4 | 425 | /* commit_hot will be freed by shm teardown */ |
852c2936 | 426 | free_chanbuf: |
852c2936 MD |
427 | return ret; |
428 | } | |
429 | ||
03d2d293 MD |
430 | static |
431 | void lib_ring_buffer_channel_switch_timer(int sig, siginfo_t *si, void *uc) | |
852c2936 | 432 | { |
03d2d293 MD |
433 | const struct lttng_ust_lib_ring_buffer_config *config; |
434 | struct lttng_ust_shm_handle *handle; | |
5198080d | 435 | struct lttng_ust_lib_ring_buffer_channel *chan; |
03d2d293 MD |
436 | int cpu; |
437 | ||
438 | assert(CMM_LOAD_SHARED(timer_signal.tid) == pthread_self()); | |
439 | ||
440 | chan = si->si_value.sival_ptr; | |
441 | handle = chan->handle; | |
442 | config = &chan->backend.config; | |
852c2936 | 443 | |
34a91bdb | 444 | DBG("Switch timer for channel %p\n", chan); |
03d2d293 | 445 | |
4dcd7e80 MD |
446 | /* |
447 | * Only flush buffers periodically if readers are active. | |
448 | */ | |
cb7378b3 | 449 | pthread_mutex_lock(&wakeup_fd_mutex); |
03d2d293 MD |
450 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
451 | for_each_possible_cpu(cpu) { | |
452 | struct lttng_ust_lib_ring_buffer *buf = | |
453 | shmp(handle, chan->backend.buf[cpu].shmp); | |
34daae3e MD |
454 | |
455 | if (!buf) | |
15500a1b | 456 | goto end; |
4dcd7e80 MD |
457 | if (uatomic_read(&buf->active_readers)) |
458 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, | |
459 | chan->handle); | |
03d2d293 MD |
460 | } |
461 | } else { | |
462 | struct lttng_ust_lib_ring_buffer *buf = | |
463 | shmp(handle, chan->backend.buf[0].shmp); | |
464 | ||
34daae3e | 465 | if (!buf) |
15500a1b | 466 | goto end; |
34a91bdb MD |
467 | if (uatomic_read(&buf->active_readers)) |
468 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, | |
469 | chan->handle); | |
470 | } | |
15500a1b | 471 | end: |
34a91bdb MD |
472 | pthread_mutex_unlock(&wakeup_fd_mutex); |
473 | return; | |
474 | } | |
475 | ||
b07cd987 MD |
476 | static |
477 | int lib_ring_buffer_poll_deliver(const struct lttng_ust_lib_ring_buffer_config *config, | |
478 | struct lttng_ust_lib_ring_buffer *buf, | |
5198080d | 479 | struct lttng_ust_lib_ring_buffer_channel *chan, |
b07cd987 MD |
480 | struct lttng_ust_shm_handle *handle) |
481 | { | |
482 | unsigned long consumed_old, consumed_idx, commit_count, write_offset; | |
15500a1b | 483 | struct commit_counters_cold *cc_cold; |
b07cd987 MD |
484 | |
485 | consumed_old = uatomic_read(&buf->consumed); | |
486 | consumed_idx = subbuf_index(consumed_old, chan); | |
15500a1b MD |
487 | cc_cold = shmp_index(handle, buf->commit_cold, consumed_idx); |
488 | if (!cc_cold) | |
489 | return 0; | |
490 | commit_count = v_read(config, &cc_cold->cc_sb); | |
b07cd987 MD |
491 | /* |
492 | * No memory barrier here, since we are only interested | |
493 | * in a statistically correct polling result. The next poll will | |
494 | * get the data is we are racing. The mb() that ensures correct | |
495 | * memory order is in get_subbuf. | |
496 | */ | |
497 | write_offset = v_read(config, &buf->offset); | |
498 | ||
499 | /* | |
500 | * Check that the subbuffer we are trying to consume has been | |
501 | * already fully committed. | |
502 | */ | |
503 | ||
504 | if (((commit_count - chan->backend.subbuf_size) | |
505 | & chan->commit_count_mask) | |
506 | - (buf_trunc(consumed_old, chan) | |
507 | >> chan->backend.num_subbuf_order) | |
508 | != 0) | |
509 | return 0; | |
510 | ||
511 | /* | |
512 | * Check that we are not about to read the same subbuffer in | |
513 | * which the writer head is. | |
514 | */ | |
515 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan) | |
516 | == 0) | |
517 | return 0; | |
518 | ||
519 | return 1; | |
520 | } | |
521 | ||
522 | static | |
523 | void lib_ring_buffer_wakeup(struct lttng_ust_lib_ring_buffer *buf, | |
524 | struct lttng_ust_shm_handle *handle) | |
525 | { | |
526 | int wakeup_fd = shm_get_wakeup_fd(handle, &buf->self._ref); | |
527 | sigset_t sigpipe_set, pending_set, old_set; | |
528 | int ret, sigpipe_was_pending = 0; | |
529 | ||
530 | if (wakeup_fd < 0) | |
531 | return; | |
532 | ||
533 | /* | |
534 | * Wake-up the other end by writing a null byte in the pipe | |
535 | * (non-blocking). Important note: Because writing into the | |
536 | * pipe is non-blocking (and therefore we allow dropping wakeup | |
537 | * data, as long as there is wakeup data present in the pipe | |
538 | * buffer to wake up the consumer), the consumer should perform | |
539 | * the following sequence for waiting: | |
540 | * 1) empty the pipe (reads). | |
541 | * 2) check if there is data in the buffer. | |
542 | * 3) wait on the pipe (poll). | |
543 | * | |
544 | * Discard the SIGPIPE from write(), not disturbing any SIGPIPE | |
545 | * that might be already pending. If a bogus SIGPIPE is sent to | |
546 | * the entire process concurrently by a malicious user, it may | |
547 | * be simply discarded. | |
548 | */ | |
549 | ret = sigemptyset(&pending_set); | |
550 | assert(!ret); | |
551 | /* | |
552 | * sigpending returns the mask of signals that are _both_ | |
553 | * blocked for the thread _and_ pending for either the thread or | |
554 | * the entire process. | |
555 | */ | |
556 | ret = sigpending(&pending_set); | |
557 | assert(!ret); | |
558 | sigpipe_was_pending = sigismember(&pending_set, SIGPIPE); | |
559 | /* | |
560 | * If sigpipe was pending, it means it was already blocked, so | |
561 | * no need to block it. | |
562 | */ | |
563 | if (!sigpipe_was_pending) { | |
564 | ret = sigemptyset(&sigpipe_set); | |
565 | assert(!ret); | |
566 | ret = sigaddset(&sigpipe_set, SIGPIPE); | |
567 | assert(!ret); | |
568 | ret = pthread_sigmask(SIG_BLOCK, &sigpipe_set, &old_set); | |
569 | assert(!ret); | |
570 | } | |
571 | do { | |
572 | ret = write(wakeup_fd, "", 1); | |
573 | } while (ret == -1L && errno == EINTR); | |
574 | if (ret == -1L && errno == EPIPE && !sigpipe_was_pending) { | |
575 | struct timespec timeout = { 0, 0 }; | |
576 | do { | |
577 | ret = sigtimedwait(&sigpipe_set, NULL, | |
578 | &timeout); | |
579 | } while (ret == -1L && errno == EINTR); | |
580 | } | |
581 | if (!sigpipe_was_pending) { | |
582 | ret = pthread_sigmask(SIG_SETMASK, &old_set, NULL); | |
583 | assert(!ret); | |
584 | } | |
585 | } | |
586 | ||
34a91bdb | 587 | static |
5198080d | 588 | void lib_ring_buffer_channel_do_read(struct lttng_ust_lib_ring_buffer_channel *chan) |
34a91bdb MD |
589 | { |
590 | const struct lttng_ust_lib_ring_buffer_config *config; | |
591 | struct lttng_ust_shm_handle *handle; | |
592 | int cpu; | |
593 | ||
594 | handle = chan->handle; | |
595 | config = &chan->backend.config; | |
596 | ||
597 | /* | |
598 | * Only flush buffers periodically if readers are active. | |
599 | */ | |
600 | pthread_mutex_lock(&wakeup_fd_mutex); | |
601 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
602 | for_each_possible_cpu(cpu) { | |
603 | struct lttng_ust_lib_ring_buffer *buf = | |
604 | shmp(handle, chan->backend.buf[cpu].shmp); | |
605 | ||
34daae3e | 606 | if (!buf) |
15500a1b | 607 | goto end; |
34a91bdb MD |
608 | if (uatomic_read(&buf->active_readers) |
609 | && lib_ring_buffer_poll_deliver(config, buf, | |
610 | chan, handle)) { | |
611 | lib_ring_buffer_wakeup(buf, handle); | |
612 | } | |
613 | } | |
614 | } else { | |
615 | struct lttng_ust_lib_ring_buffer *buf = | |
616 | shmp(handle, chan->backend.buf[0].shmp); | |
617 | ||
34daae3e | 618 | if (!buf) |
15500a1b | 619 | goto end; |
34a91bdb MD |
620 | if (uatomic_read(&buf->active_readers) |
621 | && lib_ring_buffer_poll_deliver(config, buf, | |
622 | chan, handle)) { | |
623 | lib_ring_buffer_wakeup(buf, handle); | |
624 | } | |
03d2d293 | 625 | } |
15500a1b | 626 | end: |
cb7378b3 | 627 | pthread_mutex_unlock(&wakeup_fd_mutex); |
34a91bdb MD |
628 | } |
629 | ||
630 | static | |
631 | void lib_ring_buffer_channel_read_timer(int sig, siginfo_t *si, void *uc) | |
632 | { | |
5198080d | 633 | struct lttng_ust_lib_ring_buffer_channel *chan; |
34a91bdb MD |
634 | |
635 | assert(CMM_LOAD_SHARED(timer_signal.tid) == pthread_self()); | |
636 | chan = si->si_value.sival_ptr; | |
637 | DBG("Read timer for channel %p\n", chan); | |
638 | lib_ring_buffer_channel_do_read(chan); | |
03d2d293 MD |
639 | return; |
640 | } | |
641 | ||
642 | static | |
643 | void rb_setmask(sigset_t *mask) | |
644 | { | |
645 | int ret; | |
646 | ||
647 | ret = sigemptyset(mask); | |
648 | if (ret) { | |
649 | PERROR("sigemptyset"); | |
650 | } | |
34a91bdb MD |
651 | ret = sigaddset(mask, LTTNG_UST_RB_SIG_FLUSH); |
652 | if (ret) { | |
653 | PERROR("sigaddset"); | |
654 | } | |
655 | ret = sigaddset(mask, LTTNG_UST_RB_SIG_READ); | |
03d2d293 MD |
656 | if (ret) { |
657 | PERROR("sigaddset"); | |
658 | } | |
659 | ret = sigaddset(mask, LTTNG_UST_RB_SIG_TEARDOWN); | |
660 | if (ret) { | |
661 | PERROR("sigaddset"); | |
662 | } | |
663 | } | |
664 | ||
665 | static | |
666 | void *sig_thread(void *arg) | |
667 | { | |
668 | sigset_t mask; | |
669 | siginfo_t info; | |
670 | int signr; | |
671 | ||
672 | /* Only self thread will receive signal mask. */ | |
673 | rb_setmask(&mask); | |
674 | CMM_STORE_SHARED(timer_signal.tid, pthread_self()); | |
675 | ||
676 | for (;;) { | |
677 | signr = sigwaitinfo(&mask, &info); | |
678 | if (signr == -1) { | |
34a91bdb MD |
679 | if (errno != EINTR) |
680 | PERROR("sigwaitinfo"); | |
03d2d293 MD |
681 | continue; |
682 | } | |
34a91bdb | 683 | if (signr == LTTNG_UST_RB_SIG_FLUSH) { |
03d2d293 MD |
684 | lib_ring_buffer_channel_switch_timer(info.si_signo, |
685 | &info, NULL); | |
34a91bdb MD |
686 | } else if (signr == LTTNG_UST_RB_SIG_READ) { |
687 | lib_ring_buffer_channel_read_timer(info.si_signo, | |
688 | &info, NULL); | |
03d2d293 MD |
689 | } else if (signr == LTTNG_UST_RB_SIG_TEARDOWN) { |
690 | cmm_smp_mb(); | |
691 | CMM_STORE_SHARED(timer_signal.qs_done, 1); | |
692 | cmm_smp_mb(); | |
693 | } else { | |
694 | ERR("Unexptected signal %d\n", info.si_signo); | |
695 | } | |
696 | } | |
697 | return NULL; | |
698 | } | |
699 | ||
700 | /* | |
03d2d293 MD |
701 | * Ensure only a single thread listens on the timer signal. |
702 | */ | |
703 | static | |
704 | void lib_ring_buffer_setup_timer_thread(void) | |
705 | { | |
706 | pthread_t thread; | |
707 | int ret; | |
708 | ||
64bf51a6 | 709 | pthread_mutex_lock(&timer_signal.lock); |
03d2d293 | 710 | if (timer_signal.setup_done) |
64bf51a6 | 711 | goto end; |
03d2d293 MD |
712 | |
713 | ret = pthread_create(&thread, NULL, &sig_thread, NULL); | |
714 | if (ret) { | |
715 | errno = ret; | |
716 | PERROR("pthread_create"); | |
717 | } | |
718 | ret = pthread_detach(thread); | |
719 | if (ret) { | |
720 | errno = ret; | |
721 | PERROR("pthread_detach"); | |
722 | } | |
723 | timer_signal.setup_done = 1; | |
64bf51a6 MD |
724 | end: |
725 | pthread_mutex_unlock(&timer_signal.lock); | |
852c2936 MD |
726 | } |
727 | ||
03d2d293 | 728 | /* |
64bf51a6 | 729 | * Wait for signal-handling thread quiescent state. |
03d2d293 | 730 | */ |
64bf51a6 MD |
731 | static |
732 | void lib_ring_buffer_wait_signal_thread_qs(unsigned int signr) | |
733 | { | |
734 | sigset_t pending_set; | |
735 | int ret; | |
736 | ||
737 | /* | |
738 | * We need to be the only thread interacting with the thread | |
739 | * that manages signals for teardown synchronization. | |
740 | */ | |
741 | pthread_mutex_lock(&timer_signal.lock); | |
742 | ||
743 | /* | |
744 | * Ensure we don't have any signal queued for this channel. | |
745 | */ | |
746 | for (;;) { | |
747 | ret = sigemptyset(&pending_set); | |
748 | if (ret == -1) { | |
749 | PERROR("sigemptyset"); | |
750 | } | |
751 | ret = sigpending(&pending_set); | |
752 | if (ret == -1) { | |
753 | PERROR("sigpending"); | |
754 | } | |
755 | if (!sigismember(&pending_set, signr)) | |
756 | break; | |
757 | caa_cpu_relax(); | |
758 | } | |
759 | ||
760 | /* | |
761 | * From this point, no new signal handler will be fired that | |
762 | * would try to access "chan". However, we still need to wait | |
763 | * for any currently executing handler to complete. | |
764 | */ | |
765 | cmm_smp_mb(); | |
766 | CMM_STORE_SHARED(timer_signal.qs_done, 0); | |
767 | cmm_smp_mb(); | |
768 | ||
769 | /* | |
770 | * Kill with LTTNG_UST_RB_SIG_TEARDOWN, so signal management | |
771 | * thread wakes up. | |
772 | */ | |
773 | kill(getpid(), LTTNG_UST_RB_SIG_TEARDOWN); | |
774 | ||
775 | while (!CMM_LOAD_SHARED(timer_signal.qs_done)) | |
776 | caa_cpu_relax(); | |
777 | cmm_smp_mb(); | |
778 | ||
779 | pthread_mutex_unlock(&timer_signal.lock); | |
780 | } | |
781 | ||
03d2d293 | 782 | static |
5198080d | 783 | void lib_ring_buffer_channel_switch_timer_start(struct lttng_ust_lib_ring_buffer_channel *chan) |
852c2936 | 784 | { |
03d2d293 MD |
785 | struct sigevent sev; |
786 | struct itimerspec its; | |
787 | int ret; | |
852c2936 | 788 | |
03d2d293 | 789 | if (!chan->switch_timer_interval || chan->switch_timer_enabled) |
852c2936 MD |
790 | return; |
791 | ||
03d2d293 MD |
792 | chan->switch_timer_enabled = 1; |
793 | ||
794 | lib_ring_buffer_setup_timer_thread(); | |
795 | ||
ba702529 | 796 | memset(&sev, 0, sizeof(sev)); |
03d2d293 | 797 | sev.sigev_notify = SIGEV_SIGNAL; |
34a91bdb | 798 | sev.sigev_signo = LTTNG_UST_RB_SIG_FLUSH; |
03d2d293 MD |
799 | sev.sigev_value.sival_ptr = chan; |
800 | ret = timer_create(CLOCKID, &sev, &chan->switch_timer); | |
801 | if (ret == -1) { | |
802 | PERROR("timer_create"); | |
803 | } | |
804 | ||
805 | its.it_value.tv_sec = chan->switch_timer_interval / 1000000; | |
f8921d33 | 806 | its.it_value.tv_nsec = (chan->switch_timer_interval % 1000000) * 1000; |
03d2d293 MD |
807 | its.it_interval.tv_sec = its.it_value.tv_sec; |
808 | its.it_interval.tv_nsec = its.it_value.tv_nsec; | |
809 | ||
810 | ret = timer_settime(chan->switch_timer, 0, &its, NULL); | |
811 | if (ret == -1) { | |
812 | PERROR("timer_settime"); | |
813 | } | |
814 | } | |
815 | ||
03d2d293 | 816 | static |
5198080d | 817 | void lib_ring_buffer_channel_switch_timer_stop(struct lttng_ust_lib_ring_buffer_channel *chan) |
03d2d293 | 818 | { |
34a91bdb | 819 | int ret; |
03d2d293 MD |
820 | |
821 | if (!chan->switch_timer_interval || !chan->switch_timer_enabled) | |
822 | return; | |
823 | ||
824 | ret = timer_delete(chan->switch_timer); | |
825 | if (ret == -1) { | |
826 | PERROR("timer_delete"); | |
827 | } | |
828 | ||
64bf51a6 | 829 | lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_FLUSH); |
03d2d293 MD |
830 | |
831 | chan->switch_timer = 0; | |
832 | chan->switch_timer_enabled = 0; | |
852c2936 MD |
833 | } |
834 | ||
34a91bdb | 835 | static |
5198080d | 836 | void lib_ring_buffer_channel_read_timer_start(struct lttng_ust_lib_ring_buffer_channel *chan) |
852c2936 | 837 | { |
4cfec15c | 838 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
34a91bdb MD |
839 | struct sigevent sev; |
840 | struct itimerspec its; | |
841 | int ret; | |
852c2936 | 842 | |
34a91bdb MD |
843 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER |
844 | || !chan->read_timer_interval || chan->read_timer_enabled) | |
845 | return; | |
852c2936 | 846 | |
34a91bdb | 847 | chan->read_timer_enabled = 1; |
852c2936 | 848 | |
34a91bdb | 849 | lib_ring_buffer_setup_timer_thread(); |
852c2936 | 850 | |
34a91bdb MD |
851 | sev.sigev_notify = SIGEV_SIGNAL; |
852 | sev.sigev_signo = LTTNG_UST_RB_SIG_READ; | |
853 | sev.sigev_value.sival_ptr = chan; | |
854 | ret = timer_create(CLOCKID, &sev, &chan->read_timer); | |
855 | if (ret == -1) { | |
856 | PERROR("timer_create"); | |
857 | } | |
852c2936 | 858 | |
34a91bdb | 859 | its.it_value.tv_sec = chan->read_timer_interval / 1000000; |
f8921d33 | 860 | its.it_value.tv_nsec = (chan->read_timer_interval % 1000000) * 1000; |
34a91bdb MD |
861 | its.it_interval.tv_sec = its.it_value.tv_sec; |
862 | its.it_interval.tv_nsec = its.it_value.tv_nsec; | |
852c2936 | 863 | |
34a91bdb MD |
864 | ret = timer_settime(chan->read_timer, 0, &its, NULL); |
865 | if (ret == -1) { | |
866 | PERROR("timer_settime"); | |
867 | } | |
852c2936 MD |
868 | } |
869 | ||
34a91bdb | 870 | static |
5198080d | 871 | void lib_ring_buffer_channel_read_timer_stop(struct lttng_ust_lib_ring_buffer_channel *chan) |
852c2936 | 872 | { |
4cfec15c | 873 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
34a91bdb | 874 | int ret; |
852c2936 MD |
875 | |
876 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
34a91bdb | 877 | || !chan->read_timer_interval || !chan->read_timer_enabled) |
852c2936 MD |
878 | return; |
879 | ||
34a91bdb MD |
880 | ret = timer_delete(chan->read_timer); |
881 | if (ret == -1) { | |
882 | PERROR("timer_delete"); | |
883 | } | |
884 | ||
852c2936 MD |
885 | /* |
886 | * do one more check to catch data that has been written in the last | |
887 | * timer period. | |
888 | */ | |
34a91bdb MD |
889 | lib_ring_buffer_channel_do_read(chan); |
890 | ||
64bf51a6 | 891 | lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_READ); |
34a91bdb MD |
892 | |
893 | chan->read_timer = 0; | |
894 | chan->read_timer_enabled = 0; | |
852c2936 MD |
895 | } |
896 | ||
5198080d | 897 | static void channel_unregister_notifiers(struct lttng_ust_lib_ring_buffer_channel *chan, |
38fae1d3 | 898 | struct lttng_ust_shm_handle *handle) |
852c2936 | 899 | { |
03d2d293 | 900 | lib_ring_buffer_channel_switch_timer_stop(chan); |
34a91bdb | 901 | lib_ring_buffer_channel_read_timer_stop(chan); |
852c2936 MD |
902 | } |
903 | ||
5198080d | 904 | static void channel_print_errors(struct lttng_ust_lib_ring_buffer_channel *chan, |
009769ca MD |
905 | struct lttng_ust_shm_handle *handle) |
906 | { | |
907 | const struct lttng_ust_lib_ring_buffer_config *config = | |
908 | &chan->backend.config; | |
909 | int cpu; | |
910 | ||
911 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
912 | for_each_possible_cpu(cpu) { | |
913 | struct lttng_ust_lib_ring_buffer *buf = | |
914 | shmp(handle, chan->backend.buf[cpu].shmp); | |
15500a1b MD |
915 | if (buf) |
916 | lib_ring_buffer_print_errors(chan, buf, cpu, handle); | |
009769ca MD |
917 | } |
918 | } else { | |
919 | struct lttng_ust_lib_ring_buffer *buf = | |
920 | shmp(handle, chan->backend.buf[0].shmp); | |
921 | ||
15500a1b MD |
922 | if (buf) |
923 | lib_ring_buffer_print_errors(chan, buf, -1, handle); | |
009769ca MD |
924 | } |
925 | } | |
926 | ||
5198080d | 927 | static void channel_free(struct lttng_ust_lib_ring_buffer_channel *chan, |
6548fca4 MD |
928 | struct lttng_ust_shm_handle *handle, |
929 | int consumer) | |
852c2936 | 930 | { |
74d81a6c | 931 | channel_backend_free(&chan->backend, handle); |
431d5cf0 | 932 | /* chan is freed by shm teardown */ |
6548fca4 | 933 | shm_object_table_destroy(handle->table, consumer); |
1d498196 | 934 | free(handle); |
852c2936 MD |
935 | } |
936 | ||
937 | /** | |
938 | * channel_create - Create channel. | |
939 | * @config: ring buffer instance configuration | |
940 | * @name: name of the channel | |
f0fde1c3 MD |
941 | * @priv_data_align: alignment, in bytes, of the private data area. (config) |
942 | * @priv_data_size: length, in bytes, of the private data area. (config) | |
943 | * @priv_data_init: initialization data for private data. (config) | |
944 | * @priv: local private data (memory owner by caller) | |
852c2936 MD |
945 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous |
946 | * address mapping. It is used only by RING_BUFFER_STATIC | |
947 | * configuration. It can be set to NULL for other backends. | |
948 | * @subbuf_size: subbuffer size | |
949 | * @num_subbuf: number of subbuffers | |
950 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
951 | * padding to let readers get those sub-buffers. | |
952 | * Used for live streaming. | |
953 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
5ea386c3 MD |
954 | * @stream_fds: array of stream file descriptors. |
955 | * @nr_stream_fds: number of file descriptors in array. | |
852c2936 MD |
956 | * |
957 | * Holds cpu hotplug. | |
958 | * Returns NULL on failure. | |
959 | */ | |
4cfec15c | 960 | struct lttng_ust_shm_handle *channel_create(const struct lttng_ust_lib_ring_buffer_config *config, |
a3f61e7f | 961 | const char *name, |
a3f61e7f MD |
962 | size_t priv_data_align, |
963 | size_t priv_data_size, | |
d028eddb | 964 | void *priv_data_init, |
f0fde1c3 | 965 | void *priv, |
a3f61e7f | 966 | void *buf_addr, size_t subbuf_size, |
852c2936 | 967 | size_t num_subbuf, unsigned int switch_timer_interval, |
a9ff648c | 968 | unsigned int read_timer_interval, |
b2c5f61a MD |
969 | const int *stream_fds, int nr_stream_fds, |
970 | int64_t blocking_timeout) | |
852c2936 | 971 | { |
24622edc | 972 | int ret; |
a3f61e7f | 973 | size_t shmsize, chansize; |
5198080d | 974 | struct lttng_ust_lib_ring_buffer_channel *chan; |
38fae1d3 | 975 | struct lttng_ust_shm_handle *handle; |
1d498196 | 976 | struct shm_object *shmobj; |
74d81a6c | 977 | unsigned int nr_streams; |
b2c5f61a | 978 | int64_t blocking_timeout_ms; |
74d81a6c MD |
979 | |
980 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
981 | nr_streams = num_possible_cpus(); | |
982 | else | |
983 | nr_streams = 1; | |
852c2936 | 984 | |
5ea386c3 MD |
985 | if (nr_stream_fds != nr_streams) |
986 | return NULL; | |
987 | ||
b2c5f61a MD |
988 | if (blocking_timeout < -1) { |
989 | return NULL; | |
990 | } | |
991 | /* usec to msec */ | |
992 | if (blocking_timeout == -1) { | |
993 | blocking_timeout_ms = -1; | |
994 | } else { | |
995 | blocking_timeout_ms = blocking_timeout / 1000; | |
996 | if (blocking_timeout_ms != (int32_t) blocking_timeout_ms) { | |
997 | return NULL; | |
998 | } | |
999 | } | |
1000 | ||
852c2936 MD |
1001 | if (lib_ring_buffer_check_config(config, switch_timer_interval, |
1002 | read_timer_interval)) | |
1003 | return NULL; | |
1004 | ||
38fae1d3 | 1005 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
431d5cf0 MD |
1006 | if (!handle) |
1007 | return NULL; | |
1008 | ||
1d498196 MD |
1009 | /* Allocate table for channel + per-cpu buffers */ |
1010 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
1011 | if (!handle->table) | |
1012 | goto error_table_alloc; | |
852c2936 | 1013 | |
1d498196 | 1014 | /* Calculate the shm allocation layout */ |
5198080d | 1015 | shmsize = sizeof(struct lttng_ust_lib_ring_buffer_channel); |
b72687b8 | 1016 | shmsize += lttng_ust_offset_align(shmsize, __alignof__(struct lttng_ust_lib_ring_buffer_shmp)); |
74d81a6c | 1017 | shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp) * nr_streams; |
a3f61e7f | 1018 | chansize = shmsize; |
74d81a6c | 1019 | if (priv_data_align) |
b72687b8 | 1020 | shmsize += lttng_ust_offset_align(shmsize, priv_data_align); |
a3f61e7f | 1021 | shmsize += priv_data_size; |
a6352fd4 | 1022 | |
74d81a6c | 1023 | /* Allocate normal memory for channel (not shared) */ |
a9ff648c | 1024 | shmobj = shm_object_table_alloc(handle->table, shmsize, SHM_OBJECT_MEM, |
4b68c31f | 1025 | -1, -1); |
b5a14697 MD |
1026 | if (!shmobj) |
1027 | goto error_append; | |
5198080d | 1028 | /* struct lttng_ust_lib_ring_buffer_channel is at object 0, offset 0 (hardcoded) */ |
a3f61e7f | 1029 | set_shmp(handle->chan, zalloc_shm(shmobj, chansize)); |
57773204 MD |
1030 | assert(handle->chan._ref.index == 0); |
1031 | assert(handle->chan._ref.offset == 0); | |
1d498196 | 1032 | chan = shmp(handle, handle->chan); |
a6352fd4 | 1033 | if (!chan) |
1d498196 | 1034 | goto error_append; |
74d81a6c | 1035 | chan->nr_streams = nr_streams; |
a6352fd4 | 1036 | |
a3f61e7f MD |
1037 | /* space for private data */ |
1038 | if (priv_data_size) { | |
f0fde1c3 MD |
1039 | void *priv_config; |
1040 | ||
a3f61e7f MD |
1041 | DECLARE_SHMP(void, priv_data_alloc); |
1042 | ||
1043 | align_shm(shmobj, priv_data_align); | |
1044 | chan->priv_data_offset = shmobj->allocated_len; | |
1045 | set_shmp(priv_data_alloc, zalloc_shm(shmobj, priv_data_size)); | |
1046 | if (!shmp(handle, priv_data_alloc)) | |
1047 | goto error_append; | |
f0fde1c3 MD |
1048 | priv_config = channel_get_private_config(chan); |
1049 | memcpy(priv_config, priv_data_init, priv_data_size); | |
a3f61e7f MD |
1050 | } else { |
1051 | chan->priv_data_offset = -1; | |
a3f61e7f MD |
1052 | } |
1053 | ||
b2c5f61a MD |
1054 | chan->u.s.blocking_timeout_ms = (int32_t) blocking_timeout_ms; |
1055 | ||
f0fde1c3 MD |
1056 | channel_set_private(chan, priv); |
1057 | ||
a3f61e7f | 1058 | ret = channel_backend_init(&chan->backend, name, config, |
a9ff648c | 1059 | subbuf_size, num_subbuf, handle, |
5ea386c3 | 1060 | stream_fds); |
852c2936 | 1061 | if (ret) |
1d498196 | 1062 | goto error_backend_init; |
852c2936 | 1063 | |
03d2d293 | 1064 | chan->handle = handle; |
852c2936 | 1065 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); |
852c2936 | 1066 | |
34a91bdb MD |
1067 | chan->switch_timer_interval = switch_timer_interval; |
1068 | chan->read_timer_interval = read_timer_interval; | |
03d2d293 | 1069 | lib_ring_buffer_channel_switch_timer_start(chan); |
34a91bdb | 1070 | lib_ring_buffer_channel_read_timer_start(chan); |
852c2936 | 1071 | |
431d5cf0 | 1072 | return handle; |
852c2936 | 1073 | |
1d498196 MD |
1074 | error_backend_init: |
1075 | error_append: | |
6548fca4 | 1076 | shm_object_table_destroy(handle->table, 1); |
1d498196 | 1077 | error_table_alloc: |
431d5cf0 | 1078 | free(handle); |
852c2936 MD |
1079 | return NULL; |
1080 | } | |
852c2936 | 1081 | |
74d81a6c | 1082 | struct lttng_ust_shm_handle *channel_handle_create(void *data, |
ff0f5728 MD |
1083 | uint64_t memory_map_size, |
1084 | int wakeup_fd) | |
193183fb | 1085 | { |
38fae1d3 | 1086 | struct lttng_ust_shm_handle *handle; |
193183fb MD |
1087 | struct shm_object *object; |
1088 | ||
38fae1d3 | 1089 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
193183fb MD |
1090 | if (!handle) |
1091 | return NULL; | |
1092 | ||
1093 | /* Allocate table for channel + per-cpu buffers */ | |
1094 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
1095 | if (!handle->table) | |
1096 | goto error_table_alloc; | |
1097 | /* Add channel object */ | |
74d81a6c | 1098 | object = shm_object_table_append_mem(handle->table, data, |
ff0f5728 | 1099 | memory_map_size, wakeup_fd); |
193183fb MD |
1100 | if (!object) |
1101 | goto error_table_object; | |
5198080d | 1102 | /* struct lttng_ust_lib_ring_buffer_channel is at object 0, offset 0 (hardcoded) */ |
57773204 MD |
1103 | handle->chan._ref.index = 0; |
1104 | handle->chan._ref.offset = 0; | |
193183fb MD |
1105 | return handle; |
1106 | ||
1107 | error_table_object: | |
6548fca4 | 1108 | shm_object_table_destroy(handle->table, 0); |
193183fb MD |
1109 | error_table_alloc: |
1110 | free(handle); | |
1111 | return NULL; | |
1112 | } | |
1113 | ||
38fae1d3 | 1114 | int channel_handle_add_stream(struct lttng_ust_shm_handle *handle, |
74d81a6c MD |
1115 | int shm_fd, int wakeup_fd, uint32_t stream_nr, |
1116 | uint64_t memory_map_size) | |
193183fb MD |
1117 | { |
1118 | struct shm_object *object; | |
1119 | ||
1120 | /* Add stream object */ | |
74d81a6c MD |
1121 | object = shm_object_table_append_shm(handle->table, |
1122 | shm_fd, wakeup_fd, stream_nr, | |
1123 | memory_map_size); | |
193183fb | 1124 | if (!object) |
74d81a6c | 1125 | return -EINVAL; |
193183fb MD |
1126 | return 0; |
1127 | } | |
1128 | ||
74d81a6c MD |
1129 | unsigned int channel_handle_get_nr_streams(struct lttng_ust_shm_handle *handle) |
1130 | { | |
1131 | assert(handle->table); | |
1132 | return handle->table->allocated_len - 1; | |
1133 | } | |
1134 | ||
852c2936 | 1135 | static |
5198080d | 1136 | void channel_release(struct lttng_ust_lib_ring_buffer_channel *chan, struct lttng_ust_shm_handle *handle, |
6548fca4 | 1137 | int consumer) |
852c2936 | 1138 | { |
6548fca4 | 1139 | channel_free(chan, handle, consumer); |
852c2936 MD |
1140 | } |
1141 | ||
1142 | /** | |
1143 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
1144 | * @chan: channel to destroy | |
1145 | * | |
1146 | * Holds cpu hotplug. | |
431d5cf0 MD |
1147 | * Call "destroy" callback, finalize channels, decrement the channel |
1148 | * reference count. Note that when readers have completed data | |
1149 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
a60af3a5 | 1150 | * They should release their handle at that point. |
852c2936 | 1151 | */ |
5198080d | 1152 | void channel_destroy(struct lttng_ust_lib_ring_buffer_channel *chan, struct lttng_ust_shm_handle *handle, |
74d81a6c | 1153 | int consumer) |
852c2936 | 1154 | { |
74d81a6c MD |
1155 | if (consumer) { |
1156 | /* | |
1157 | * Note: the consumer takes care of finalizing and | |
1158 | * switching the buffers. | |
1159 | */ | |
1160 | channel_unregister_notifiers(chan, handle); | |
3d0ef9f6 MD |
1161 | /* |
1162 | * The consumer prints errors. | |
1163 | */ | |
1164 | channel_print_errors(chan, handle); | |
824f40b8 MD |
1165 | } |
1166 | ||
431d5cf0 MD |
1167 | /* |
1168 | * sessiond/consumer are keeping a reference on the shm file | |
1169 | * descriptor directly. No need to refcount. | |
1170 | */ | |
6548fca4 | 1171 | channel_release(chan, handle, consumer); |
a3f61e7f | 1172 | return; |
852c2936 | 1173 | } |
852c2936 | 1174 | |
4cfec15c MD |
1175 | struct lttng_ust_lib_ring_buffer *channel_get_ring_buffer( |
1176 | const struct lttng_ust_lib_ring_buffer_config *config, | |
5198080d | 1177 | struct lttng_ust_lib_ring_buffer_channel *chan, int cpu, |
38fae1d3 | 1178 | struct lttng_ust_shm_handle *handle, |
74d81a6c MD |
1179 | int *shm_fd, int *wait_fd, |
1180 | int *wakeup_fd, | |
1181 | uint64_t *memory_map_size) | |
852c2936 | 1182 | { |
381c0f1e MD |
1183 | struct shm_ref *ref; |
1184 | ||
1185 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
74d81a6c | 1186 | cpu = 0; |
381c0f1e | 1187 | } else { |
e095d803 MD |
1188 | if (cpu >= num_possible_cpus()) |
1189 | return NULL; | |
381c0f1e | 1190 | } |
74d81a6c MD |
1191 | ref = &chan->backend.buf[cpu].shmp._ref; |
1192 | *shm_fd = shm_get_shm_fd(handle, ref); | |
1193 | *wait_fd = shm_get_wait_fd(handle, ref); | |
1194 | *wakeup_fd = shm_get_wakeup_fd(handle, ref); | |
1195 | if (shm_get_shm_size(handle, ref, memory_map_size)) | |
1196 | return NULL; | |
1197 | return shmp(handle, chan->backend.buf[cpu].shmp); | |
852c2936 | 1198 | } |
852c2936 | 1199 | |
ff0f5728 | 1200 | int ring_buffer_channel_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config *config, |
5198080d | 1201 | struct lttng_ust_lib_ring_buffer_channel *chan, |
ff0f5728 MD |
1202 | struct lttng_ust_shm_handle *handle) |
1203 | { | |
1204 | struct shm_ref *ref; | |
1205 | ||
1206 | ref = &handle->chan._ref; | |
1207 | return shm_close_wait_fd(handle, ref); | |
1208 | } | |
1209 | ||
1210 | int ring_buffer_channel_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config *config, | |
5198080d | 1211 | struct lttng_ust_lib_ring_buffer_channel *chan, |
ff0f5728 MD |
1212 | struct lttng_ust_shm_handle *handle) |
1213 | { | |
1214 | struct shm_ref *ref; | |
1215 | ||
1216 | ref = &handle->chan._ref; | |
1217 | return shm_close_wakeup_fd(handle, ref); | |
1218 | } | |
1219 | ||
1220 | int ring_buffer_stream_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config *config, | |
5198080d | 1221 | struct lttng_ust_lib_ring_buffer_channel *chan, |
74d81a6c MD |
1222 | struct lttng_ust_shm_handle *handle, |
1223 | int cpu) | |
852c2936 | 1224 | { |
74d81a6c MD |
1225 | struct shm_ref *ref; |
1226 | ||
1227 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
1228 | cpu = 0; | |
1229 | } else { | |
1230 | if (cpu >= num_possible_cpus()) | |
1231 | return -EINVAL; | |
824f40b8 | 1232 | } |
74d81a6c MD |
1233 | ref = &chan->backend.buf[cpu].shmp._ref; |
1234 | return shm_close_wait_fd(handle, ref); | |
1235 | } | |
1236 | ||
ff0f5728 | 1237 | int ring_buffer_stream_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config *config, |
5198080d | 1238 | struct lttng_ust_lib_ring_buffer_channel *chan, |
74d81a6c MD |
1239 | struct lttng_ust_shm_handle *handle, |
1240 | int cpu) | |
1241 | { | |
1242 | struct shm_ref *ref; | |
cb7378b3 | 1243 | int ret; |
74d81a6c MD |
1244 | |
1245 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
1246 | cpu = 0; | |
1247 | } else { | |
1248 | if (cpu >= num_possible_cpus()) | |
1249 | return -EINVAL; | |
1250 | } | |
1251 | ref = &chan->backend.buf[cpu].shmp._ref; | |
cb7378b3 MD |
1252 | pthread_mutex_lock(&wakeup_fd_mutex); |
1253 | ret = shm_close_wakeup_fd(handle, ref); | |
1254 | pthread_mutex_unlock(&wakeup_fd_mutex); | |
1255 | return ret; | |
74d81a6c MD |
1256 | } |
1257 | ||
1258 | int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer *buf, | |
1259 | struct lttng_ust_shm_handle *handle) | |
1260 | { | |
a6352fd4 | 1261 | if (uatomic_cmpxchg(&buf->active_readers, 0, 1) != 0) |
852c2936 | 1262 | return -EBUSY; |
a6352fd4 | 1263 | cmm_smp_mb(); |
852c2936 MD |
1264 | return 0; |
1265 | } | |
852c2936 | 1266 | |
4cfec15c | 1267 | void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer *buf, |
74d81a6c | 1268 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1269 | { |
5198080d | 1270 | struct lttng_ust_lib_ring_buffer_channel *chan = shmp(handle, buf->backend.chan); |
852c2936 | 1271 | |
15500a1b MD |
1272 | if (!chan) |
1273 | return; | |
a6352fd4 MD |
1274 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
1275 | cmm_smp_mb(); | |
1276 | uatomic_dec(&buf->active_readers); | |
852c2936 MD |
1277 | } |
1278 | ||
1279 | /** | |
1280 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
1281 | * @buf: ring buffer | |
1282 | * @consumed: consumed count indicating the position where to read | |
1283 | * @produced: produced count, indicates position when to stop reading | |
1284 | * | |
1285 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
1286 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 MD |
1287 | */ |
1288 | ||
4cfec15c | 1289 | int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 1290 | unsigned long *consumed, unsigned long *produced, |
38fae1d3 | 1291 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1292 | { |
5198080d | 1293 | struct lttng_ust_lib_ring_buffer_channel *chan; |
15500a1b | 1294 | const struct lttng_ust_lib_ring_buffer_config *config; |
852c2936 MD |
1295 | unsigned long consumed_cur, write_offset; |
1296 | int finalized; | |
1297 | ||
15500a1b MD |
1298 | chan = shmp(handle, buf->backend.chan); |
1299 | if (!chan) | |
1300 | return -EPERM; | |
1301 | config = &chan->backend.config; | |
14641deb | 1302 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
1303 | /* |
1304 | * Read finalized before counters. | |
1305 | */ | |
a6352fd4 MD |
1306 | cmm_smp_rmb(); |
1307 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 MD |
1308 | /* |
1309 | * No need to issue a memory barrier between consumed count read and | |
1310 | * write offset read, because consumed count can only change | |
1311 | * concurrently in overwrite mode, and we keep a sequence counter | |
1312 | * identifier derived from the write offset to check we are getting | |
1313 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
1314 | * "tagged" upon writes, tags are checked upon read). | |
1315 | */ | |
1316 | write_offset = v_read(config, &buf->offset); | |
1317 | ||
1318 | /* | |
1319 | * Check that we are not about to read the same subbuffer in | |
1320 | * which the writer head is. | |
1321 | */ | |
1322 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
1323 | == 0) | |
1324 | goto nodata; | |
1325 | ||
1326 | *consumed = consumed_cur; | |
1327 | *produced = subbuf_trunc(write_offset, chan); | |
1328 | ||
1329 | return 0; | |
1330 | ||
1331 | nodata: | |
1332 | /* | |
1333 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1334 | * of "raw_spin_is_locked" memory ordering. | |
1335 | */ | |
1336 | if (finalized) | |
1337 | return -ENODATA; | |
852c2936 MD |
1338 | else |
1339 | return -EAGAIN; | |
1340 | } | |
852c2936 | 1341 | |
f45930b7 JG |
1342 | /** |
1343 | * Performs the same function as lib_ring_buffer_snapshot(), but the positions | |
1344 | * are saved regardless of whether the consumed and produced positions are | |
1345 | * in the same subbuffer. | |
1346 | * @buf: ring buffer | |
1347 | * @consumed: consumed byte count indicating the last position read | |
1348 | * @produced: produced byte count indicating the last position written | |
1349 | * | |
1350 | * This function is meant to provide information on the exact producer and | |
1351 | * consumer positions without regard for the "snapshot" feature. | |
1352 | */ | |
1353 | int lib_ring_buffer_snapshot_sample_positions( | |
1354 | struct lttng_ust_lib_ring_buffer *buf, | |
1355 | unsigned long *consumed, unsigned long *produced, | |
1356 | struct lttng_ust_shm_handle *handle) | |
1357 | { | |
5198080d | 1358 | struct lttng_ust_lib_ring_buffer_channel *chan; |
f45930b7 | 1359 | const struct lttng_ust_lib_ring_buffer_config *config; |
f45930b7 JG |
1360 | |
1361 | chan = shmp(handle, buf->backend.chan); | |
1362 | if (!chan) | |
1363 | return -EPERM; | |
1364 | config = &chan->backend.config; | |
1365 | cmm_smp_rmb(); | |
1366 | *consumed = uatomic_read(&buf->consumed); | |
1367 | /* | |
1368 | * No need to issue a memory barrier between consumed count read and | |
1369 | * write offset read, because consumed count can only change | |
1370 | * concurrently in overwrite mode, and we keep a sequence counter | |
1371 | * identifier derived from the write offset to check we are getting | |
1372 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
1373 | * "tagged" upon writes, tags are checked upon read). | |
1374 | */ | |
1375 | *produced = v_read(config, &buf->offset); | |
1376 | return 0; | |
1377 | } | |
1378 | ||
852c2936 | 1379 | /** |
d1a996f5 | 1380 | * lib_ring_buffer_move_consumer - move consumed counter forward |
852c2936 MD |
1381 | * @buf: ring buffer |
1382 | * @consumed_new: new consumed count value | |
1383 | */ | |
4cfec15c | 1384 | void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 1385 | unsigned long consumed_new, |
38fae1d3 | 1386 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1387 | { |
4cfec15c | 1388 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
5198080d | 1389 | struct lttng_ust_lib_ring_buffer_channel *chan; |
852c2936 MD |
1390 | unsigned long consumed; |
1391 | ||
15500a1b MD |
1392 | chan = shmp(handle, bufb->chan); |
1393 | if (!chan) | |
1394 | return; | |
74d81a6c | 1395 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
852c2936 MD |
1396 | |
1397 | /* | |
1398 | * Only push the consumed value forward. | |
1399 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
1400 | * the writer in flight recorder mode. | |
1401 | */ | |
a6352fd4 | 1402 | consumed = uatomic_read(&buf->consumed); |
852c2936 | 1403 | while ((long) consumed - (long) consumed_new < 0) |
a6352fd4 MD |
1404 | consumed = uatomic_cmpxchg(&buf->consumed, consumed, |
1405 | consumed_new); | |
852c2936 | 1406 | } |
852c2936 MD |
1407 | |
1408 | /** | |
1409 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
1410 | * @buf: ring buffer | |
1411 | * @consumed: consumed count indicating the position where to read | |
1412 | * | |
1413 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
1414 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 | 1415 | */ |
4cfec15c | 1416 | int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 1417 | unsigned long consumed, |
38fae1d3 | 1418 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1419 | { |
5198080d | 1420 | struct lttng_ust_lib_ring_buffer_channel *chan; |
15500a1b | 1421 | const struct lttng_ust_lib_ring_buffer_config *config; |
852c2936 | 1422 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
dddba398 | 1423 | int ret, finalized, nr_retry = LTTNG_UST_RING_BUFFER_GET_RETRY; |
15500a1b | 1424 | struct commit_counters_cold *cc_cold; |
852c2936 | 1425 | |
15500a1b MD |
1426 | chan = shmp(handle, buf->backend.chan); |
1427 | if (!chan) | |
1428 | return -EPERM; | |
1429 | config = &chan->backend.config; | |
852c2936 | 1430 | retry: |
14641deb | 1431 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
1432 | /* |
1433 | * Read finalized before counters. | |
1434 | */ | |
a6352fd4 MD |
1435 | cmm_smp_rmb(); |
1436 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 | 1437 | consumed_idx = subbuf_index(consumed, chan); |
15500a1b MD |
1438 | cc_cold = shmp_index(handle, buf->commit_cold, consumed_idx); |
1439 | if (!cc_cold) | |
1440 | return -EPERM; | |
1441 | commit_count = v_read(config, &cc_cold->cc_sb); | |
852c2936 MD |
1442 | /* |
1443 | * Make sure we read the commit count before reading the buffer | |
1444 | * data and the write offset. Correct consumed offset ordering | |
1445 | * wrt commit count is insured by the use of cmpxchg to update | |
1446 | * the consumed offset. | |
852c2936 | 1447 | */ |
a6352fd4 MD |
1448 | /* |
1449 | * Local rmb to match the remote wmb to read the commit count | |
1450 | * before the buffer data and the write offset. | |
1451 | */ | |
1452 | cmm_smp_rmb(); | |
852c2936 MD |
1453 | |
1454 | write_offset = v_read(config, &buf->offset); | |
1455 | ||
1456 | /* | |
1457 | * Check that the buffer we are getting is after or at consumed_cur | |
1458 | * position. | |
1459 | */ | |
1460 | if ((long) subbuf_trunc(consumed, chan) | |
1461 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
1462 | goto nodata; | |
1463 | ||
1464 | /* | |
1465 | * Check that the subbuffer we are trying to consume has been | |
dddba398 MD |
1466 | * already fully committed. There are a few causes that can make |
1467 | * this unavailability situation occur: | |
1468 | * | |
1469 | * Temporary (short-term) situation: | |
1470 | * - Application is running on a different CPU, between reserve | |
1471 | * and commit ring buffer operations, | |
1472 | * - Application is preempted between reserve and commit ring | |
1473 | * buffer operations, | |
1474 | * | |
1475 | * Long-term situation: | |
1476 | * - Application is stopped (SIGSTOP) between reserve and commit | |
1477 | * ring buffer operations. Could eventually be resumed by | |
1478 | * SIGCONT. | |
1479 | * - Application is killed (SIGTERM, SIGINT, SIGKILL) between | |
1480 | * reserve and commit ring buffer operation. | |
1481 | * | |
1482 | * From a consumer perspective, handling short-term | |
1483 | * unavailability situations is performed by retrying a few | |
1484 | * times after a delay. Handling long-term unavailability | |
1485 | * situations is handled by failing to get the sub-buffer. | |
1486 | * | |
1487 | * In all of those situations, if the application is taking a | |
1488 | * long time to perform its commit after ring buffer space | |
1489 | * reservation, we can end up in a situation where the producer | |
1490 | * will fill the ring buffer and try to write into the same | |
1491 | * sub-buffer again (which has a missing commit). This is | |
1492 | * handled by the producer in the sub-buffer switch handling | |
1493 | * code of the reserve routine by detecting unbalanced | |
1494 | * reserve/commit counters and discarding all further events | |
1495 | * until the situation is resolved in those situations. Two | |
1496 | * scenarios can occur: | |
1497 | * | |
1498 | * 1) The application causing the reserve/commit counters to be | |
1499 | * unbalanced has been terminated. In this situation, all | |
1500 | * further events will be discarded in the buffers, and no | |
1501 | * further buffer data will be readable by the consumer | |
1502 | * daemon. Tearing down the UST tracing session and starting | |
1503 | * anew is a work-around for those situations. Note that this | |
1504 | * only affects per-UID tracing. In per-PID tracing, the | |
1505 | * application vanishes with the termination, and therefore | |
1506 | * no more data needs to be written to the buffers. | |
1507 | * 2) The application causing the unbalance has been delayed for | |
1508 | * a long time, but will eventually try to increment the | |
1509 | * commit counter after eventually writing to the sub-buffer. | |
1510 | * This situation can cause events to be discarded until the | |
1511 | * application resumes its operations. | |
852c2936 MD |
1512 | */ |
1513 | if (((commit_count - chan->backend.subbuf_size) | |
1514 | & chan->commit_count_mask) | |
0bade047 | 1515 | - (buf_trunc(consumed, chan) |
852c2936 | 1516 | >> chan->backend.num_subbuf_order) |
dddba398 MD |
1517 | != 0) { |
1518 | if (nr_retry-- > 0) { | |
1519 | if (nr_retry <= (LTTNG_UST_RING_BUFFER_GET_RETRY >> 1)) | |
1520 | (void) poll(NULL, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS); | |
1521 | goto retry; | |
1522 | } else { | |
1523 | goto nodata; | |
1524 | } | |
1525 | } | |
852c2936 MD |
1526 | |
1527 | /* | |
1528 | * Check that we are not about to read the same subbuffer in | |
1529 | * which the writer head is. | |
1530 | */ | |
0bade047 | 1531 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed, chan) |
852c2936 MD |
1532 | == 0) |
1533 | goto nodata; | |
1534 | ||
1535 | /* | |
1536 | * Failure to get the subbuffer causes a busy-loop retry without going | |
1537 | * to a wait queue. These are caused by short-lived race windows where | |
1538 | * the writer is getting access to a subbuffer we were trying to get | |
1539 | * access to. Also checks that the "consumed" buffer count we are | |
1540 | * looking for matches the one contained in the subbuffer id. | |
dddba398 MD |
1541 | * |
1542 | * The short-lived race window described here can be affected by | |
1543 | * application signals and preemption, thus requiring to bound | |
1544 | * the loop to a maximum number of retry. | |
852c2936 MD |
1545 | */ |
1546 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
1547 | consumed_idx, buf_trunc_val(consumed, chan), |
1548 | handle); | |
dddba398 MD |
1549 | if (ret) { |
1550 | if (nr_retry-- > 0) { | |
1551 | if (nr_retry <= (LTTNG_UST_RING_BUFFER_GET_RETRY >> 1)) | |
1552 | (void) poll(NULL, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS); | |
1553 | goto retry; | |
1554 | } else { | |
1555 | goto nodata; | |
1556 | } | |
1557 | } | |
852c2936 MD |
1558 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); |
1559 | ||
1560 | buf->get_subbuf_consumed = consumed; | |
1561 | buf->get_subbuf = 1; | |
1562 | ||
1563 | return 0; | |
1564 | ||
1565 | nodata: | |
1566 | /* | |
1567 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1568 | * of "raw_spin_is_locked" memory ordering. | |
1569 | */ | |
1570 | if (finalized) | |
1571 | return -ENODATA; | |
852c2936 MD |
1572 | else |
1573 | return -EAGAIN; | |
1574 | } | |
852c2936 MD |
1575 | |
1576 | /** | |
1577 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
1578 | * @buf: ring buffer | |
1579 | */ | |
4cfec15c | 1580 | void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 1581 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1582 | { |
4cfec15c | 1583 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
5198080d | 1584 | struct lttng_ust_lib_ring_buffer_channel *chan; |
15500a1b MD |
1585 | const struct lttng_ust_lib_ring_buffer_config *config; |
1586 | unsigned long sb_bindex, consumed_idx, consumed; | |
1587 | struct lttng_ust_lib_ring_buffer_backend_pages_shmp *rpages; | |
1588 | struct lttng_ust_lib_ring_buffer_backend_pages *backend_pages; | |
852c2936 | 1589 | |
15500a1b MD |
1590 | chan = shmp(handle, bufb->chan); |
1591 | if (!chan) | |
1592 | return; | |
1593 | config = &chan->backend.config; | |
74d81a6c | 1594 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
852c2936 MD |
1595 | |
1596 | if (!buf->get_subbuf) { | |
1597 | /* | |
1598 | * Reader puts a subbuffer it did not get. | |
1599 | */ | |
1600 | CHAN_WARN_ON(chan, 1); | |
1601 | return; | |
1602 | } | |
1603 | consumed = buf->get_subbuf_consumed; | |
1604 | buf->get_subbuf = 0; | |
1605 | ||
1606 | /* | |
1607 | * Clear the records_unread counter. (overruns counter) | |
1608 | * Can still be non-zero if a file reader simply grabbed the data | |
1609 | * without using iterators. | |
1610 | * Can be below zero if an iterator is used on a snapshot more than | |
1611 | * once. | |
1612 | */ | |
15500a1b MD |
1613 | sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); |
1614 | rpages = shmp_index(handle, bufb->array, sb_bindex); | |
1615 | if (!rpages) | |
1616 | return; | |
1617 | backend_pages = shmp(handle, rpages->shmp); | |
1618 | if (!backend_pages) | |
1619 | return; | |
1620 | v_add(config, v_read(config, &backend_pages->records_unread), | |
1621 | &bufb->records_read); | |
1622 | v_set(config, &backend_pages->records_unread, 0); | |
852c2936 MD |
1623 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE |
1624 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
1625 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
1626 | ||
1627 | /* | |
1628 | * Exchange the reader subbuffer with the one we put in its place in the | |
1629 | * writer subbuffer table. Expect the original consumed count. If | |
1630 | * update_read_sb_index fails, this is because the writer updated the | |
1631 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
1632 | * currently have: it has become invalid to try reading this sub-buffer | |
1633 | * consumed count value anyway. | |
1634 | */ | |
1635 | consumed_idx = subbuf_index(consumed, chan); | |
1636 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
1637 | consumed_idx, buf_trunc_val(consumed, chan), |
1638 | handle); | |
852c2936 MD |
1639 | /* |
1640 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
1641 | * if the writer concurrently updated it. | |
1642 | */ | |
1643 | } | |
852c2936 MD |
1644 | |
1645 | /* | |
1646 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
1647 | * position and the writer position. (inclusive) | |
1648 | */ | |
1649 | static | |
4cfec15c | 1650 | void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
5198080d | 1651 | struct lttng_ust_lib_ring_buffer_channel *chan, |
852c2936 | 1652 | unsigned long cons_offset, |
1d498196 | 1653 | int cpu, |
38fae1d3 | 1654 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1655 | { |
4cfec15c | 1656 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 | 1657 | unsigned long cons_idx, commit_count, commit_count_sb; |
15500a1b MD |
1658 | struct commit_counters_hot *cc_hot; |
1659 | struct commit_counters_cold *cc_cold; | |
852c2936 MD |
1660 | |
1661 | cons_idx = subbuf_index(cons_offset, chan); | |
15500a1b MD |
1662 | cc_hot = shmp_index(handle, buf->commit_hot, cons_idx); |
1663 | if (!cc_hot) | |
1664 | return; | |
1665 | cc_cold = shmp_index(handle, buf->commit_cold, cons_idx); | |
1666 | if (!cc_cold) | |
1667 | return; | |
1668 | commit_count = v_read(config, &cc_hot->cc); | |
1669 | commit_count_sb = v_read(config, &cc_cold->cc_sb); | |
852c2936 MD |
1670 | |
1671 | if (subbuf_offset(commit_count, chan) != 0) | |
4d3c9523 | 1672 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
1673 | "commit count in subbuffer %lu,\n" |
1674 | "expecting multiples of %lu bytes\n" | |
1675 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
1676 | chan->backend.name, cpu, cons_idx, | |
1677 | chan->backend.subbuf_size, | |
1678 | commit_count, commit_count_sb); | |
1679 | ||
4d3c9523 | 1680 | DBG("ring buffer: %s, cpu %d: %lu bytes committed\n", |
852c2936 MD |
1681 | chan->backend.name, cpu, commit_count); |
1682 | } | |
1683 | ||
1684 | static | |
4cfec15c | 1685 | void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
5198080d | 1686 | struct lttng_ust_lib_ring_buffer_channel *chan, |
f0fde1c3 | 1687 | int cpu, struct lttng_ust_shm_handle *handle) |
852c2936 | 1688 | { |
4cfec15c | 1689 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1690 | unsigned long write_offset, cons_offset; |
1691 | ||
852c2936 MD |
1692 | /* |
1693 | * No need to order commit_count, write_offset and cons_offset reads | |
1694 | * because we execute at teardown when no more writer nor reader | |
1695 | * references are left. | |
1696 | */ | |
1697 | write_offset = v_read(config, &buf->offset); | |
a6352fd4 | 1698 | cons_offset = uatomic_read(&buf->consumed); |
852c2936 | 1699 | if (write_offset != cons_offset) |
4d3c9523 | 1700 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
1701 | "non-consumed data\n" |
1702 | " [ %lu bytes written, %lu bytes read ]\n", | |
1703 | chan->backend.name, cpu, write_offset, cons_offset); | |
1704 | ||
a6352fd4 | 1705 | for (cons_offset = uatomic_read(&buf->consumed); |
852c2936 MD |
1706 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), |
1707 | chan) | |
1708 | - cons_offset) > 0; | |
1709 | cons_offset = subbuf_align(cons_offset, chan)) | |
1710 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1d498196 | 1711 | cpu, handle); |
852c2936 MD |
1712 | } |
1713 | ||
1714 | static | |
5198080d | 1715 | void lib_ring_buffer_print_errors(struct lttng_ust_lib_ring_buffer_channel *chan, |
009769ca MD |
1716 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
1717 | struct lttng_ust_shm_handle *handle) | |
852c2936 | 1718 | { |
4cfec15c | 1719 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 | 1720 | |
a1360615 MD |
1721 | if (!strcmp(chan->backend.name, "relay-metadata-mmap")) { |
1722 | DBG("ring buffer %s: %lu records written, " | |
1723 | "%lu records overrun\n", | |
1724 | chan->backend.name, | |
1725 | v_read(config, &buf->records_count), | |
1726 | v_read(config, &buf->records_overrun)); | |
1727 | } else { | |
1728 | DBG("ring buffer %s, cpu %d: %lu records written, " | |
1729 | "%lu records overrun\n", | |
1730 | chan->backend.name, cpu, | |
1731 | v_read(config, &buf->records_count), | |
1732 | v_read(config, &buf->records_overrun)); | |
1733 | ||
1734 | if (v_read(config, &buf->records_lost_full) | |
1735 | || v_read(config, &buf->records_lost_wrap) | |
1736 | || v_read(config, &buf->records_lost_big)) | |
1737 | DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n" | |
1738 | " [ %lu buffer full, %lu nest buffer wrap-around, " | |
1739 | "%lu event too big ]\n", | |
1740 | chan->backend.name, cpu, | |
1741 | v_read(config, &buf->records_lost_full), | |
1742 | v_read(config, &buf->records_lost_wrap), | |
1743 | v_read(config, &buf->records_lost_big)); | |
1744 | } | |
f0fde1c3 | 1745 | lib_ring_buffer_print_buffer_errors(buf, chan, cpu, handle); |
852c2936 MD |
1746 | } |
1747 | ||
1748 | /* | |
1749 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
1750 | * | |
48f22436 MD |
1751 | * Only executed by SWITCH_FLUSH, which can be issued while tracing is |
1752 | * active or at buffer finalization (destroy). | |
852c2936 MD |
1753 | */ |
1754 | static | |
4cfec15c | 1755 | void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer *buf, |
5198080d | 1756 | struct lttng_ust_lib_ring_buffer_channel *chan, |
852c2936 | 1757 | struct switch_offsets *offsets, |
2fed87ae | 1758 | uint64_t tsc, |
38fae1d3 | 1759 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1760 | { |
4cfec15c | 1761 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1762 | unsigned long oldidx = subbuf_index(offsets->old, chan); |
1763 | unsigned long commit_count; | |
3d3c3833 | 1764 | struct commit_counters_hot *cc_hot; |
852c2936 | 1765 | |
1d498196 | 1766 | config->cb.buffer_begin(buf, tsc, oldidx, handle); |
852c2936 MD |
1767 | |
1768 | /* | |
1769 | * Order all writes to buffer before the commit count update that will | |
1770 | * determine that the subbuffer is full. | |
1771 | */ | |
a6352fd4 | 1772 | cmm_smp_wmb(); |
3d3c3833 | 1773 | cc_hot = shmp_index(handle, buf->commit_hot, oldidx); |
15500a1b MD |
1774 | if (!cc_hot) |
1775 | return; | |
852c2936 | 1776 | v_add(config, config->cb.subbuffer_header_size(), |
3d3c3833 MD |
1777 | &cc_hot->cc); |
1778 | commit_count = v_read(config, &cc_hot->cc); | |
852c2936 MD |
1779 | /* Check if the written buffer has to be delivered */ |
1780 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
1b7b0501 | 1781 | commit_count, oldidx, handle, tsc); |
d2fe4771 | 1782 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
80249235 | 1783 | offsets->old + config->cb.subbuffer_header_size(), |
3d3c3833 | 1784 | commit_count, handle, cc_hot); |
852c2936 MD |
1785 | } |
1786 | ||
1787 | /* | |
1788 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
1789 | * | |
1790 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
1791 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
1792 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
1793 | * subbuffer. | |
1794 | */ | |
1795 | static | |
4cfec15c | 1796 | void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer *buf, |
5198080d | 1797 | struct lttng_ust_lib_ring_buffer_channel *chan, |
852c2936 | 1798 | struct switch_offsets *offsets, |
2fed87ae | 1799 | uint64_t tsc, |
38fae1d3 | 1800 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1801 | { |
4cfec15c | 1802 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1803 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); |
1804 | unsigned long commit_count, padding_size, data_size; | |
3d3c3833 | 1805 | struct commit_counters_hot *cc_hot; |
6c737d05 | 1806 | uint64_t *ts_end; |
852c2936 MD |
1807 | |
1808 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1809 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1810 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size, |
1811 | handle); | |
852c2936 | 1812 | |
6c737d05 MD |
1813 | ts_end = shmp_index(handle, buf->ts_end, oldidx); |
1814 | if (!ts_end) | |
1815 | return; | |
852c2936 | 1816 | /* |
6c737d05 MD |
1817 | * This is the last space reservation in that sub-buffer before |
1818 | * it gets delivered. This provides exclusive access to write to | |
1819 | * this sub-buffer's ts_end. There are also no concurrent | |
1820 | * readers of that ts_end because delivery of that sub-buffer is | |
1821 | * postponed until the commit counter is incremented for the | |
1822 | * current space reservation. | |
1823 | */ | |
1824 | *ts_end = tsc; | |
1825 | ||
1826 | /* | |
1827 | * Order all writes to buffer and store to ts_end before the commit | |
1828 | * count update that will determine that the subbuffer is full. | |
852c2936 | 1829 | */ |
a6352fd4 | 1830 | cmm_smp_wmb(); |
3d3c3833 | 1831 | cc_hot = shmp_index(handle, buf->commit_hot, oldidx); |
15500a1b MD |
1832 | if (!cc_hot) |
1833 | return; | |
3d3c3833 MD |
1834 | v_add(config, padding_size, &cc_hot->cc); |
1835 | commit_count = v_read(config, &cc_hot->cc); | |
852c2936 | 1836 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, |
1b7b0501 | 1837 | commit_count, oldidx, handle, tsc); |
d2fe4771 MD |
1838 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
1839 | offsets->old + padding_size, commit_count, handle, | |
3d3c3833 | 1840 | cc_hot); |
852c2936 MD |
1841 | } |
1842 | ||
1843 | /* | |
1844 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1845 | * | |
1846 | * This code can be executed unordered : writers may already have written to the | |
1847 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1848 | * that this code is executed before the deliver of this sub-buffer. | |
1849 | */ | |
1850 | static | |
4cfec15c | 1851 | void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer *buf, |
5198080d | 1852 | struct lttng_ust_lib_ring_buffer_channel *chan, |
852c2936 | 1853 | struct switch_offsets *offsets, |
2fed87ae | 1854 | uint64_t tsc, |
38fae1d3 | 1855 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1856 | { |
4cfec15c | 1857 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1858 | unsigned long beginidx = subbuf_index(offsets->begin, chan); |
1859 | unsigned long commit_count; | |
3d3c3833 | 1860 | struct commit_counters_hot *cc_hot; |
852c2936 | 1861 | |
1d498196 | 1862 | config->cb.buffer_begin(buf, tsc, beginidx, handle); |
852c2936 MD |
1863 | |
1864 | /* | |
1865 | * Order all writes to buffer before the commit count update that will | |
1866 | * determine that the subbuffer is full. | |
1867 | */ | |
a6352fd4 | 1868 | cmm_smp_wmb(); |
3d3c3833 | 1869 | cc_hot = shmp_index(handle, buf->commit_hot, beginidx); |
15500a1b MD |
1870 | if (!cc_hot) |
1871 | return; | |
1872 | v_add(config, config->cb.subbuffer_header_size(), &cc_hot->cc); | |
3d3c3833 | 1873 | commit_count = v_read(config, &cc_hot->cc); |
852c2936 MD |
1874 | /* Check if the written buffer has to be delivered */ |
1875 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
1b7b0501 | 1876 | commit_count, beginidx, handle, tsc); |
d2fe4771 | 1877 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
80249235 | 1878 | offsets->begin + config->cb.subbuffer_header_size(), |
3d3c3833 | 1879 | commit_count, handle, cc_hot); |
852c2936 MD |
1880 | } |
1881 | ||
1ad21f70 MD |
1882 | /* |
1883 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1884 | * | |
3962118d MD |
1885 | * Calls subbuffer_set_data_size() to set the data size of the current |
1886 | * sub-buffer. We do not need to perform check_deliver nor commit here, | |
1887 | * since this task will be done by the "commit" of the event for which | |
1888 | * we are currently doing the space reservation. | |
1ad21f70 MD |
1889 | */ |
1890 | static | |
1891 | void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer *buf, | |
5198080d | 1892 | struct lttng_ust_lib_ring_buffer_channel *chan, |
1ad21f70 MD |
1893 | struct switch_offsets *offsets, |
1894 | uint64_t tsc, | |
1895 | struct lttng_ust_shm_handle *handle) | |
1896 | { | |
1897 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; | |
3962118d | 1898 | unsigned long endidx, data_size; |
6c737d05 | 1899 | uint64_t *ts_end; |
1ad21f70 | 1900 | |
3962118d | 1901 | endidx = subbuf_index(offsets->end - 1, chan); |
1ad21f70 | 1902 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; |
1ad21f70 MD |
1903 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size, |
1904 | handle); | |
6c737d05 MD |
1905 | ts_end = shmp_index(handle, buf->ts_end, endidx); |
1906 | if (!ts_end) | |
1907 | return; | |
1908 | /* | |
1909 | * This is the last space reservation in that sub-buffer before | |
1910 | * it gets delivered. This provides exclusive access to write to | |
1911 | * this sub-buffer's ts_end. There are also no concurrent | |
1912 | * readers of that ts_end because delivery of that sub-buffer is | |
1913 | * postponed until the commit counter is incremented for the | |
1914 | * current space reservation. | |
1915 | */ | |
1916 | *ts_end = tsc; | |
1ad21f70 MD |
1917 | } |
1918 | ||
852c2936 MD |
1919 | /* |
1920 | * Returns : | |
1921 | * 0 if ok | |
1922 | * !0 if execution must be aborted. | |
1923 | */ | |
1924 | static | |
1925 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
4cfec15c | 1926 | struct lttng_ust_lib_ring_buffer *buf, |
5198080d | 1927 | struct lttng_ust_lib_ring_buffer_channel *chan, |
852c2936 | 1928 | struct switch_offsets *offsets, |
2dac206d MD |
1929 | uint64_t *tsc, |
1930 | struct lttng_ust_shm_handle *handle) | |
852c2936 | 1931 | { |
4cfec15c | 1932 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
2dac206d | 1933 | unsigned long off, reserve_commit_diff; |
852c2936 MD |
1934 | |
1935 | offsets->begin = v_read(config, &buf->offset); | |
1936 | offsets->old = offsets->begin; | |
1937 | offsets->switch_old_start = 0; | |
1938 | off = subbuf_offset(offsets->begin, chan); | |
1939 | ||
1940 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1941 | ||
1942 | /* | |
1943 | * Ensure we flush the header of an empty subbuffer when doing the | |
1944 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1945 | * total data gathering duration even if there were no records saved | |
1946 | * after the last buffer switch. | |
1947 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1948 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1949 | * subbuffer header as appropriate. | |
1950 | * The next record that reserves space will be responsible for | |
1951 | * populating the following subbuffer header. We choose not to populate | |
1952 | * the next subbuffer header here because we want to be able to use | |
a6352fd4 MD |
1953 | * SWITCH_ACTIVE for periodical buffer flush, which must |
1954 | * guarantee that all the buffer content (records and header | |
1955 | * timestamps) are visible to the reader. This is required for | |
1956 | * quiescence guarantees for the fusion merge. | |
852c2936 | 1957 | */ |
2dac206d MD |
1958 | if (mode != SWITCH_FLUSH && !off) |
1959 | return -1; /* we do not have to switch : buffer is empty */ | |
1960 | ||
1961 | if (caa_unlikely(off == 0)) { | |
1962 | unsigned long sb_index, commit_count; | |
15500a1b | 1963 | struct commit_counters_cold *cc_cold; |
2dac206d MD |
1964 | |
1965 | /* | |
48f22436 MD |
1966 | * We are performing a SWITCH_FLUSH. There may be concurrent |
1967 | * writes into the buffer if e.g. invoked while performing a | |
1968 | * snapshot on an active trace. | |
2dac206d | 1969 | * |
48f22436 MD |
1970 | * If the client does not save any header information |
1971 | * (sub-buffer header size == 0), don't switch empty subbuffer | |
1972 | * on finalize, because it is invalid to deliver a completely | |
1973 | * empty subbuffer. | |
2dac206d MD |
1974 | */ |
1975 | if (!config->cb.subbuffer_header_size()) | |
1976 | return -1; | |
1977 | ||
1978 | /* Test new buffer integrity */ | |
1979 | sb_index = subbuf_index(offsets->begin, chan); | |
15500a1b MD |
1980 | cc_cold = shmp_index(handle, buf->commit_cold, sb_index); |
1981 | if (!cc_cold) | |
1982 | return -1; | |
1983 | commit_count = v_read(config, &cc_cold->cc_sb); | |
2dac206d MD |
1984 | reserve_commit_diff = |
1985 | (buf_trunc(offsets->begin, chan) | |
1986 | >> chan->backend.num_subbuf_order) | |
1987 | - (commit_count & chan->commit_count_mask); | |
1988 | if (caa_likely(reserve_commit_diff == 0)) { | |
1989 | /* Next subbuffer not being written to. */ | |
1990 | if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
1991 | subbuf_trunc(offsets->begin, chan) | |
1992 | - subbuf_trunc((unsigned long) | |
1993 | uatomic_read(&buf->consumed), chan) | |
1994 | >= chan->backend.buf_size)) { | |
1995 | /* | |
1996 | * We do not overwrite non consumed buffers | |
1997 | * and we are full : don't switch. | |
1998 | */ | |
852c2936 | 1999 | return -1; |
2dac206d MD |
2000 | } else { |
2001 | /* | |
2002 | * Next subbuffer not being written to, and we | |
2003 | * are either in overwrite mode or the buffer is | |
2004 | * not full. It's safe to write in this new | |
2005 | * subbuffer. | |
2006 | */ | |
2007 | } | |
2008 | } else { | |
852c2936 | 2009 | /* |
2dac206d MD |
2010 | * Next subbuffer reserve offset does not match the |
2011 | * commit offset. Don't perform switch in | |
2012 | * producer-consumer and overwrite mode. Caused by | |
2013 | * either a writer OOPS or too many nested writes over a | |
2014 | * reserve/commit pair. | |
852c2936 | 2015 | */ |
2dac206d | 2016 | return -1; |
852c2936 | 2017 | } |
2dac206d MD |
2018 | |
2019 | /* | |
2020 | * Need to write the subbuffer start header on finalize. | |
2021 | */ | |
2022 | offsets->switch_old_start = 1; | |
2023 | } | |
2024 | offsets->begin = subbuf_align(offsets->begin, chan); | |
852c2936 MD |
2025 | /* Note: old points to the next subbuf at offset 0 */ |
2026 | offsets->end = offsets->begin; | |
2027 | return 0; | |
2028 | } | |
2029 | ||
2030 | /* | |
2031 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
2032 | * called while tracing is active with absolutely no lock held. | |
2033 | * | |
ef1a13e0 MD |
2034 | * For RING_BUFFER_SYNC_PER_CPU ring buffers, as a v_cmpxchg is used for |
2035 | * some atomic operations, this function must be called from the CPU | |
2036 | * which owns the buffer for a ACTIVE flush. However, for | |
2037 | * RING_BUFFER_SYNC_GLOBAL ring buffers, this function can be called | |
2038 | * from any CPU. | |
852c2936 | 2039 | */ |
4cfec15c | 2040 | void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode, |
38fae1d3 | 2041 | struct lttng_ust_shm_handle *handle) |
852c2936 | 2042 | { |
5198080d | 2043 | struct lttng_ust_lib_ring_buffer_channel *chan; |
15500a1b | 2044 | const struct lttng_ust_lib_ring_buffer_config *config; |
852c2936 MD |
2045 | struct switch_offsets offsets; |
2046 | unsigned long oldidx; | |
2fed87ae | 2047 | uint64_t tsc; |
852c2936 | 2048 | |
15500a1b MD |
2049 | chan = shmp(handle, buf->backend.chan); |
2050 | if (!chan) | |
2051 | return; | |
2052 | config = &chan->backend.config; | |
2053 | ||
852c2936 MD |
2054 | offsets.size = 0; |
2055 | ||
2056 | /* | |
2057 | * Perform retryable operations. | |
2058 | */ | |
2059 | do { | |
2060 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
2dac206d | 2061 | &tsc, handle)) |
852c2936 MD |
2062 | return; /* Switch not needed */ |
2063 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
2064 | != offsets.old); | |
2065 | ||
2066 | /* | |
2067 | * Atomically update last_tsc. This update races against concurrent | |
2068 | * atomic updates, but the race will always cause supplementary full TSC | |
2069 | * records, never the opposite (missing a full TSC record when it would | |
2070 | * be needed). | |
2071 | */ | |
2072 | save_last_tsc(config, buf, tsc); | |
2073 | ||
2074 | /* | |
2075 | * Push the reader if necessary | |
2076 | */ | |
2077 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
2078 | ||
2079 | oldidx = subbuf_index(offsets.old, chan); | |
1d498196 | 2080 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx, handle); |
852c2936 MD |
2081 | |
2082 | /* | |
2083 | * May need to populate header start on SWITCH_FLUSH. | |
2084 | */ | |
2085 | if (offsets.switch_old_start) { | |
1d498196 | 2086 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc, handle); |
852c2936 MD |
2087 | offsets.old += config->cb.subbuffer_header_size(); |
2088 | } | |
2089 | ||
2090 | /* | |
2091 | * Switch old subbuffer. | |
2092 | */ | |
1d498196 | 2093 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc, handle); |
852c2936 | 2094 | } |
852c2936 | 2095 | |
6f97f9c2 MD |
2096 | static |
2097 | bool handle_blocking_retry(int *timeout_left_ms) | |
2098 | { | |
2099 | int timeout = *timeout_left_ms, delay; | |
2100 | ||
2101 | if (caa_likely(!timeout)) | |
2102 | return false; /* Do not retry, discard event. */ | |
2103 | if (timeout < 0) /* Wait forever. */ | |
2104 | delay = RETRY_DELAY_MS; | |
2105 | else | |
2106 | delay = min_t(int, timeout, RETRY_DELAY_MS); | |
2107 | (void) poll(NULL, 0, delay); | |
2108 | if (timeout > 0) | |
2109 | *timeout_left_ms -= delay; | |
2110 | return true; /* Retry. */ | |
2111 | } | |
2112 | ||
852c2936 MD |
2113 | /* |
2114 | * Returns : | |
2115 | * 0 if ok | |
2116 | * -ENOSPC if event size is too large for packet. | |
2117 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
2118 | * -EIO if data cannot be written into the buffer for any other reason. | |
2119 | */ | |
2120 | static | |
4cfec15c | 2121 | int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf, |
5198080d | 2122 | struct lttng_ust_lib_ring_buffer_channel *chan, |
852c2936 | 2123 | struct switch_offsets *offsets, |
e56bb47c MD |
2124 | struct lttng_ust_lib_ring_buffer_ctx *ctx, |
2125 | void *client_ctx) | |
852c2936 | 2126 | { |
8936b6c0 | 2127 | struct lttng_ust_lib_ring_buffer_ctx_private *ctx_private = ctx->priv; |
4cfec15c | 2128 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
8936b6c0 | 2129 | struct lttng_ust_shm_handle *handle = chan->handle; |
e185cf4b | 2130 | unsigned long reserve_commit_diff, offset_cmp; |
b2c5f61a | 2131 | int timeout_left_ms = lttng_ust_ringbuffer_get_timeout(chan); |
852c2936 | 2132 | |
e185cf4b MD |
2133 | retry: |
2134 | offsets->begin = offset_cmp = v_read(config, &buf->offset); | |
852c2936 MD |
2135 | offsets->old = offsets->begin; |
2136 | offsets->switch_new_start = 0; | |
1ad21f70 | 2137 | offsets->switch_new_end = 0; |
852c2936 MD |
2138 | offsets->switch_old_end = 0; |
2139 | offsets->pre_header_padding = 0; | |
2140 | ||
8936b6c0 MD |
2141 | ctx_private->tsc = config->cb.ring_buffer_clock_read(chan); |
2142 | if ((int64_t) ctx_private->tsc == -EIO) | |
852c2936 MD |
2143 | return -EIO; |
2144 | ||
8936b6c0 MD |
2145 | if (last_tsc_overflow(config, buf, ctx_private->tsc)) |
2146 | ctx_private->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
852c2936 | 2147 | |
8936b6c0 | 2148 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) == 0)) { |
852c2936 MD |
2149 | offsets->switch_new_start = 1; /* For offsets->begin */ |
2150 | } else { | |
2151 | offsets->size = config->cb.record_header_size(config, chan, | |
2152 | offsets->begin, | |
2153 | &offsets->pre_header_padding, | |
e56bb47c | 2154 | ctx, client_ctx); |
852c2936 | 2155 | offsets->size += |
3b8bedd8 | 2156 | lttng_ust_lib_ring_buffer_align(offsets->begin + offsets->size, |
8936b6c0 | 2157 | ctx->largest_align) |
852c2936 | 2158 | + ctx->data_size; |
b5a3dfa5 | 2159 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) + |
852c2936 MD |
2160 | offsets->size > chan->backend.subbuf_size)) { |
2161 | offsets->switch_old_end = 1; /* For offsets->old */ | |
2162 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
2163 | } | |
2164 | } | |
b5a3dfa5 | 2165 | if (caa_unlikely(offsets->switch_new_start)) { |
e185cf4b | 2166 | unsigned long sb_index, commit_count; |
15500a1b | 2167 | struct commit_counters_cold *cc_cold; |
852c2936 MD |
2168 | |
2169 | /* | |
2170 | * We are typically not filling the previous buffer completely. | |
2171 | */ | |
b5a3dfa5 | 2172 | if (caa_likely(offsets->switch_old_end)) |
852c2936 MD |
2173 | offsets->begin = subbuf_align(offsets->begin, chan); |
2174 | offsets->begin = offsets->begin | |
2175 | + config->cb.subbuffer_header_size(); | |
2176 | /* Test new buffer integrity */ | |
2177 | sb_index = subbuf_index(offsets->begin, chan); | |
e185cf4b MD |
2178 | /* |
2179 | * Read buf->offset before buf->commit_cold[sb_index].cc_sb. | |
2180 | * lib_ring_buffer_check_deliver() has the matching | |
2181 | * memory barriers required around commit_cold cc_sb | |
2182 | * updates to ensure reserve and commit counter updates | |
2183 | * are not seen reordered when updated by another CPU. | |
2184 | */ | |
2185 | cmm_smp_rmb(); | |
15500a1b MD |
2186 | cc_cold = shmp_index(handle, buf->commit_cold, sb_index); |
2187 | if (!cc_cold) | |
2188 | return -1; | |
2189 | commit_count = v_read(config, &cc_cold->cc_sb); | |
e185cf4b MD |
2190 | /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */ |
2191 | cmm_smp_rmb(); | |
2192 | if (caa_unlikely(offset_cmp != v_read(config, &buf->offset))) { | |
2193 | /* | |
2194 | * The reserve counter have been concurrently updated | |
2195 | * while we read the commit counter. This means the | |
2196 | * commit counter we read might not match buf->offset | |
2197 | * due to concurrent update. We therefore need to retry. | |
2198 | */ | |
2199 | goto retry; | |
2200 | } | |
852c2936 MD |
2201 | reserve_commit_diff = |
2202 | (buf_trunc(offsets->begin, chan) | |
2203 | >> chan->backend.num_subbuf_order) | |
e185cf4b | 2204 | - (commit_count & chan->commit_count_mask); |
b5a3dfa5 | 2205 | if (caa_likely(reserve_commit_diff == 0)) { |
852c2936 | 2206 | /* Next subbuffer not being written to. */ |
b5a3dfa5 | 2207 | if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE && |
852c2936 MD |
2208 | subbuf_trunc(offsets->begin, chan) |
2209 | - subbuf_trunc((unsigned long) | |
a6352fd4 | 2210 | uatomic_read(&buf->consumed), chan) |
852c2936 | 2211 | >= chan->backend.buf_size)) { |
64493e4f MD |
2212 | unsigned long nr_lost; |
2213 | ||
6f97f9c2 MD |
2214 | if (handle_blocking_retry(&timeout_left_ms)) |
2215 | goto retry; | |
2216 | ||
852c2936 MD |
2217 | /* |
2218 | * We do not overwrite non consumed buffers | |
2219 | * and we are full : record is lost. | |
2220 | */ | |
64493e4f | 2221 | nr_lost = v_read(config, &buf->records_lost_full); |
852c2936 | 2222 | v_inc(config, &buf->records_lost_full); |
64493e4f MD |
2223 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
2224 | DBG("%lu or more records lost in (%s:%d) (buffer full)\n", | |
2225 | nr_lost + 1, chan->backend.name, | |
2226 | buf->backend.cpu); | |
2227 | } | |
852c2936 MD |
2228 | return -ENOBUFS; |
2229 | } else { | |
2230 | /* | |
2231 | * Next subbuffer not being written to, and we | |
2232 | * are either in overwrite mode or the buffer is | |
2233 | * not full. It's safe to write in this new | |
2234 | * subbuffer. | |
2235 | */ | |
2236 | } | |
2237 | } else { | |
64493e4f MD |
2238 | unsigned long nr_lost; |
2239 | ||
852c2936 MD |
2240 | /* |
2241 | * Next subbuffer reserve offset does not match the | |
e185cf4b MD |
2242 | * commit offset, and this did not involve update to the |
2243 | * reserve counter. Drop record in producer-consumer and | |
852c2936 MD |
2244 | * overwrite mode. Caused by either a writer OOPS or too |
2245 | * many nested writes over a reserve/commit pair. | |
2246 | */ | |
64493e4f | 2247 | nr_lost = v_read(config, &buf->records_lost_wrap); |
852c2936 | 2248 | v_inc(config, &buf->records_lost_wrap); |
64493e4f MD |
2249 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
2250 | DBG("%lu or more records lost in (%s:%d) (wrap-around)\n", | |
2251 | nr_lost + 1, chan->backend.name, | |
2252 | buf->backend.cpu); | |
2253 | } | |
852c2936 MD |
2254 | return -EIO; |
2255 | } | |
2256 | offsets->size = | |
2257 | config->cb.record_header_size(config, chan, | |
2258 | offsets->begin, | |
2259 | &offsets->pre_header_padding, | |
e56bb47c | 2260 | ctx, client_ctx); |
852c2936 | 2261 | offsets->size += |
3b8bedd8 | 2262 | lttng_ust_lib_ring_buffer_align(offsets->begin + offsets->size, |
852c2936 MD |
2263 | ctx->largest_align) |
2264 | + ctx->data_size; | |
b5a3dfa5 | 2265 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) |
852c2936 | 2266 | + offsets->size > chan->backend.subbuf_size)) { |
64493e4f MD |
2267 | unsigned long nr_lost; |
2268 | ||
852c2936 MD |
2269 | /* |
2270 | * Record too big for subbuffers, report error, don't | |
2271 | * complete the sub-buffer switch. | |
2272 | */ | |
64493e4f | 2273 | nr_lost = v_read(config, &buf->records_lost_big); |
852c2936 | 2274 | v_inc(config, &buf->records_lost_big); |
64493e4f MD |
2275 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
2276 | DBG("%lu or more records lost in (%s:%d) record size " | |
2277 | " of %zu bytes is too large for buffer\n", | |
2278 | nr_lost + 1, chan->backend.name, | |
2279 | buf->backend.cpu, offsets->size); | |
2280 | } | |
852c2936 MD |
2281 | return -ENOSPC; |
2282 | } else { | |
2283 | /* | |
2284 | * We just made a successful buffer switch and the | |
2285 | * record fits in the new subbuffer. Let's write. | |
2286 | */ | |
2287 | } | |
2288 | } else { | |
2289 | /* | |
2290 | * Record fits in the current buffer and we are not on a switch | |
2291 | * boundary. It's safe to write. | |
2292 | */ | |
2293 | } | |
2294 | offsets->end = offsets->begin + offsets->size; | |
1ad21f70 MD |
2295 | |
2296 | if (caa_unlikely(subbuf_offset(offsets->end, chan) == 0)) { | |
2297 | /* | |
2298 | * The offset_end will fall at the very beginning of the next | |
2299 | * subbuffer. | |
2300 | */ | |
2301 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
2302 | } | |
852c2936 MD |
2303 | return 0; |
2304 | } | |
2305 | ||
2306 | /** | |
2307 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
2308 | * @ctx: ring buffer context. | |
2309 | * | |
2310 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, | |
2311 | * -EIO for other errors, else returns 0. | |
2312 | * It will take care of sub-buffer switching. | |
2313 | */ | |
e56bb47c MD |
2314 | int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx, |
2315 | void *client_ctx) | |
852c2936 | 2316 | { |
8936b6c0 MD |
2317 | struct lttng_ust_lib_ring_buffer_ctx_private *ctx_private = ctx->priv; |
2318 | struct lttng_ust_lib_ring_buffer_channel *chan = ctx_private->chan; | |
2319 | struct lttng_ust_shm_handle *handle = chan->handle; | |
4cfec15c MD |
2320 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
2321 | struct lttng_ust_lib_ring_buffer *buf; | |
852c2936 MD |
2322 | struct switch_offsets offsets; |
2323 | int ret; | |
2324 | ||
2325 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
8936b6c0 | 2326 | buf = shmp(handle, chan->backend.buf[ctx_private->reserve_cpu].shmp); |
852c2936 | 2327 | else |
1d498196 | 2328 | buf = shmp(handle, chan->backend.buf[0].shmp); |
15500a1b MD |
2329 | if (!buf) |
2330 | return -EIO; | |
8936b6c0 | 2331 | ctx_private->buf = buf; |
852c2936 MD |
2332 | |
2333 | offsets.size = 0; | |
2334 | ||
2335 | do { | |
2336 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, | |
e56bb47c | 2337 | ctx, client_ctx); |
b5a3dfa5 | 2338 | if (caa_unlikely(ret)) |
852c2936 | 2339 | return ret; |
b5a3dfa5 | 2340 | } while (caa_unlikely(v_cmpxchg(config, &buf->offset, offsets.old, |
852c2936 MD |
2341 | offsets.end) |
2342 | != offsets.old)); | |
2343 | ||
2344 | /* | |
2345 | * Atomically update last_tsc. This update races against concurrent | |
2346 | * atomic updates, but the race will always cause supplementary full TSC | |
2347 | * records, never the opposite (missing a full TSC record when it would | |
2348 | * be needed). | |
2349 | */ | |
8936b6c0 | 2350 | save_last_tsc(config, buf, ctx_private->tsc); |
852c2936 MD |
2351 | |
2352 | /* | |
2353 | * Push the reader if necessary | |
2354 | */ | |
2355 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
2356 | ||
2357 | /* | |
2358 | * Clear noref flag for this subbuffer. | |
2359 | */ | |
2360 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1d498196 MD |
2361 | subbuf_index(offsets.end - 1, chan), |
2362 | handle); | |
852c2936 MD |
2363 | |
2364 | /* | |
2365 | * Switch old subbuffer if needed. | |
2366 | */ | |
b5a3dfa5 | 2367 | if (caa_unlikely(offsets.switch_old_end)) { |
852c2936 | 2368 | lib_ring_buffer_clear_noref(config, &buf->backend, |
1d498196 MD |
2369 | subbuf_index(offsets.old - 1, chan), |
2370 | handle); | |
8936b6c0 | 2371 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx_private->tsc, handle); |
852c2936 MD |
2372 | } |
2373 | ||
2374 | /* | |
2375 | * Populate new subbuffer. | |
2376 | */ | |
b5a3dfa5 | 2377 | if (caa_unlikely(offsets.switch_new_start)) |
8936b6c0 | 2378 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx_private->tsc, handle); |
852c2936 | 2379 | |
1ad21f70 | 2380 | if (caa_unlikely(offsets.switch_new_end)) |
8936b6c0 | 2381 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx_private->tsc, handle); |
1ad21f70 | 2382 | |
8936b6c0 MD |
2383 | ctx_private->slot_size = offsets.size; |
2384 | ctx_private->pre_offset = offsets.begin; | |
2385 | ctx_private->buf_offset = offsets.begin + offsets.pre_header_padding; | |
852c2936 MD |
2386 | return 0; |
2387 | } | |
f645cfa7 | 2388 | |
b07cd987 MD |
2389 | static |
2390 | void lib_ring_buffer_vmcore_check_deliver(const struct lttng_ust_lib_ring_buffer_config *config, | |
2391 | struct lttng_ust_lib_ring_buffer *buf, | |
2392 | unsigned long commit_count, | |
2393 | unsigned long idx, | |
2394 | struct lttng_ust_shm_handle *handle) | |
2395 | { | |
15500a1b MD |
2396 | struct commit_counters_hot *cc_hot; |
2397 | ||
2398 | if (config->oops != RING_BUFFER_OOPS_CONSISTENCY) | |
2399 | return; | |
2400 | cc_hot = shmp_index(handle, buf->commit_hot, idx); | |
2401 | if (!cc_hot) | |
2402 | return; | |
2403 | v_set(config, &cc_hot->seq, commit_count); | |
b07cd987 MD |
2404 | } |
2405 | ||
9c995331 MD |
2406 | /* |
2407 | * The ring buffer can count events recorded and overwritten per buffer, | |
2408 | * but it is disabled by default due to its performance overhead. | |
2409 | */ | |
2410 | #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS | |
2411 | static | |
2412 | void deliver_count_events(const struct lttng_ust_lib_ring_buffer_config *config, | |
2413 | struct lttng_ust_lib_ring_buffer *buf, | |
2414 | unsigned long idx, | |
2415 | struct lttng_ust_shm_handle *handle) | |
2416 | { | |
2417 | v_add(config, subbuffer_get_records_count(config, | |
2418 | &buf->backend, idx, handle), | |
2419 | &buf->records_count); | |
2420 | v_add(config, subbuffer_count_records_overrun(config, | |
2421 | &buf->backend, idx, handle), | |
2422 | &buf->records_overrun); | |
2423 | } | |
2424 | #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */ | |
2425 | static | |
2426 | void deliver_count_events(const struct lttng_ust_lib_ring_buffer_config *config, | |
2427 | struct lttng_ust_lib_ring_buffer *buf, | |
2428 | unsigned long idx, | |
2429 | struct lttng_ust_shm_handle *handle) | |
2430 | { | |
2431 | } | |
2432 | #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */ | |
2433 | ||
b07cd987 MD |
2434 | void lib_ring_buffer_check_deliver_slow(const struct lttng_ust_lib_ring_buffer_config *config, |
2435 | struct lttng_ust_lib_ring_buffer *buf, | |
5198080d | 2436 | struct lttng_ust_lib_ring_buffer_channel *chan, |
b07cd987 MD |
2437 | unsigned long offset, |
2438 | unsigned long commit_count, | |
2439 | unsigned long idx, | |
2440 | struct lttng_ust_shm_handle *handle, | |
2441 | uint64_t tsc) | |
2442 | { | |
2443 | unsigned long old_commit_count = commit_count | |
2444 | - chan->backend.subbuf_size; | |
15500a1b | 2445 | struct commit_counters_cold *cc_cold; |
b07cd987 MD |
2446 | |
2447 | /* | |
2448 | * If we succeeded at updating cc_sb below, we are the subbuffer | |
2449 | * writer delivering the subbuffer. Deals with concurrent | |
2450 | * updates of the "cc" value without adding a add_return atomic | |
2451 | * operation to the fast path. | |
2452 | * | |
2453 | * We are doing the delivery in two steps: | |
2454 | * - First, we cmpxchg() cc_sb to the new value | |
2455 | * old_commit_count + 1. This ensures that we are the only | |
2456 | * subbuffer user successfully filling the subbuffer, but we | |
2457 | * do _not_ set the cc_sb value to "commit_count" yet. | |
2458 | * Therefore, other writers that would wrap around the ring | |
2459 | * buffer and try to start writing to our subbuffer would | |
2460 | * have to drop records, because it would appear as | |
2461 | * non-filled. | |
2462 | * We therefore have exclusive access to the subbuffer control | |
2463 | * structures. This mutual exclusion with other writers is | |
2464 | * crucially important to perform record overruns count in | |
2465 | * flight recorder mode locklessly. | |
2466 | * - When we are ready to release the subbuffer (either for | |
2467 | * reading or for overrun by other writers), we simply set the | |
2468 | * cc_sb value to "commit_count" and perform delivery. | |
2469 | * | |
2470 | * The subbuffer size is least 2 bytes (minimum size: 1 page). | |
2471 | * This guarantees that old_commit_count + 1 != commit_count. | |
2472 | */ | |
2473 | ||
2474 | /* | |
2475 | * Order prior updates to reserve count prior to the | |
2476 | * commit_cold cc_sb update. | |
2477 | */ | |
2478 | cmm_smp_wmb(); | |
15500a1b MD |
2479 | cc_cold = shmp_index(handle, buf->commit_cold, idx); |
2480 | if (!cc_cold) | |
2481 | return; | |
2482 | if (caa_likely(v_cmpxchg(config, &cc_cold->cc_sb, | |
b07cd987 MD |
2483 | old_commit_count, old_commit_count + 1) |
2484 | == old_commit_count)) { | |
6c737d05 MD |
2485 | uint64_t *ts_end; |
2486 | ||
b07cd987 MD |
2487 | /* |
2488 | * Start of exclusive subbuffer access. We are | |
2489 | * guaranteed to be the last writer in this subbuffer | |
2490 | * and any other writer trying to access this subbuffer | |
2491 | * in this state is required to drop records. | |
6c737d05 MD |
2492 | * |
2493 | * We can read the ts_end for the current sub-buffer | |
2494 | * which has been saved by the very last space | |
2495 | * reservation for the current sub-buffer. | |
2496 | * | |
2497 | * Order increment of commit counter before reading ts_end. | |
b07cd987 | 2498 | */ |
6c737d05 MD |
2499 | cmm_smp_mb(); |
2500 | ts_end = shmp_index(handle, buf->ts_end, idx); | |
2501 | if (!ts_end) | |
2502 | return; | |
9c995331 | 2503 | deliver_count_events(config, buf, idx, handle); |
6c737d05 | 2504 | config->cb.buffer_end(buf, *ts_end, idx, |
b07cd987 MD |
2505 | lib_ring_buffer_get_data_size(config, |
2506 | buf, | |
2507 | idx, | |
2508 | handle), | |
2509 | handle); | |
2510 | ||
2511 | /* | |
2512 | * Increment the packet counter while we have exclusive | |
2513 | * access. | |
2514 | */ | |
2515 | subbuffer_inc_packet_count(config, &buf->backend, idx, handle); | |
2516 | ||
2517 | /* | |
2518 | * Set noref flag and offset for this subbuffer id. | |
2519 | * Contains a memory barrier that ensures counter stores | |
2520 | * are ordered before set noref and offset. | |
2521 | */ | |
2522 | lib_ring_buffer_set_noref_offset(config, &buf->backend, idx, | |
2523 | buf_trunc_val(offset, chan), handle); | |
2524 | ||
2525 | /* | |
2526 | * Order set_noref and record counter updates before the | |
2527 | * end of subbuffer exclusive access. Orders with | |
2528 | * respect to writers coming into the subbuffer after | |
2529 | * wrap around, and also order wrt concurrent readers. | |
2530 | */ | |
2531 | cmm_smp_mb(); | |
2532 | /* End of exclusive subbuffer access */ | |
15500a1b | 2533 | v_set(config, &cc_cold->cc_sb, commit_count); |
b07cd987 MD |
2534 | /* |
2535 | * Order later updates to reserve count after | |
2536 | * the commit cold cc_sb update. | |
2537 | */ | |
2538 | cmm_smp_wmb(); | |
2539 | lib_ring_buffer_vmcore_check_deliver(config, buf, | |
2540 | commit_count, idx, handle); | |
2541 | ||
2542 | /* | |
2543 | * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free. | |
2544 | */ | |
2545 | if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER | |
2546 | && uatomic_read(&buf->active_readers) | |
2547 | && lib_ring_buffer_poll_deliver(config, buf, chan, handle)) { | |
2548 | lib_ring_buffer_wakeup(buf, handle); | |
2549 | } | |
2550 | } | |
2551 | } | |
2552 | ||
f645cfa7 MD |
2553 | /* |
2554 | * Force a read (imply TLS fixup for dlopen) of TLS variables. | |
2555 | */ | |
2556 | void lttng_fixup_ringbuffer_tls(void) | |
2557 | { | |
8c90a710 | 2558 | asm volatile ("" : : "m" (URCU_TLS(lib_ring_buffer_nesting))); |
f645cfa7 | 2559 | } |
03d2d293 MD |
2560 | |
2561 | void lib_ringbuffer_signal_init(void) | |
2562 | { | |
2563 | sigset_t mask; | |
2564 | int ret; | |
2565 | ||
2566 | /* | |
2567 | * Block signal for entire process, so only our thread processes | |
2568 | * it. | |
2569 | */ | |
2570 | rb_setmask(&mask); | |
2571 | ret = pthread_sigmask(SIG_BLOCK, &mask, NULL); | |
2572 | if (ret) { | |
2573 | errno = ret; | |
2574 | PERROR("pthread_sigmask"); | |
2575 | } | |
2576 | } |