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