Commit | Line | Data |
---|---|---|
852c2936 MD |
1 | /* |
2 | * ring_buffer_frontend.c | |
3 | * | |
4 | * (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
5 | * | |
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 | |
37 | * | |
38 | * Dual LGPL v2.1/GPL v2 license. | |
39 | */ | |
40 | ||
a6352fd4 | 41 | #include <sys/types.h> |
431d5cf0 MD |
42 | #include <sys/mman.h> |
43 | #include <sys/stat.h> | |
44 | #include <fcntl.h> | |
14641deb | 45 | #include <urcu/compiler.h> |
a6352fd4 | 46 | #include <urcu/ref.h> |
14641deb | 47 | |
a6352fd4 | 48 | #include "smp.h" |
4318ae1b | 49 | #include <lttng/ringbuffer-config.h> |
2fed87ae | 50 | #include "vatomic.h" |
4931a13e MD |
51 | #include "backend.h" |
52 | #include "frontend.h" | |
a6352fd4 | 53 | #include "shm.h" |
852c2936 | 54 | |
431d5cf0 MD |
55 | #ifndef max |
56 | #define max(a, b) ((a) > (b) ? (a) : (b)) | |
57 | #endif | |
58 | ||
2432c3c9 MD |
59 | /* |
60 | * Use POSIX SHM: shm_open(3) and shm_unlink(3). | |
61 | * close(2) to close the fd returned by shm_open. | |
62 | * shm_unlink releases the shared memory object name. | |
63 | * ftruncate(2) sets the size of the memory object. | |
64 | * mmap/munmap maps the shared memory obj to a virtual address in the | |
65 | * calling proceess (should be done both in libust and consumer). | |
66 | * See shm_overview(7) for details. | |
67 | * Pass file descriptor returned by shm_open(3) to ltt-sessiond through | |
68 | * a UNIX socket. | |
69 | * | |
70 | * Since we don't need to access the object using its name, we can | |
71 | * immediately shm_unlink(3) it, and only keep the handle with its file | |
72 | * descriptor. | |
73 | */ | |
74 | ||
852c2936 MD |
75 | /* |
76 | * Internal structure representing offsets to use at a sub-buffer switch. | |
77 | */ | |
78 | struct switch_offsets { | |
79 | unsigned long begin, end, old; | |
80 | size_t pre_header_padding, size; | |
81 | unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1, | |
82 | switch_old_end:1; | |
83 | }; | |
84 | ||
a6352fd4 | 85 | __thread unsigned int lib_ring_buffer_nesting; |
852c2936 | 86 | |
45e9e699 MD |
87 | /* |
88 | * TODO: this is unused. Errors are saved within the ring buffer. | |
89 | * Eventually, allow consumerd to print these errors. | |
90 | */ | |
852c2936 MD |
91 | static |
92 | void lib_ring_buffer_print_errors(struct channel *chan, | |
4cfec15c | 93 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
38fae1d3 | 94 | struct lttng_ust_shm_handle *handle); |
852c2936 | 95 | |
852c2936 MD |
96 | /** |
97 | * lib_ring_buffer_reset - Reset ring buffer to initial values. | |
98 | * @buf: Ring buffer. | |
99 | * | |
100 | * Effectively empty the ring buffer. Should be called when the buffer is not | |
101 | * used for writing. The ring buffer can be opened for reading, but the reader | |
102 | * should not be using the iterator concurrently with reset. The previous | |
103 | * current iterator record is reset. | |
104 | */ | |
4cfec15c | 105 | void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 106 | struct lttng_ust_shm_handle *handle) |
852c2936 | 107 | { |
1d498196 | 108 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 109 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
110 | unsigned int i; |
111 | ||
112 | /* | |
113 | * Reset iterator first. It will put the subbuffer if it currently holds | |
114 | * it. | |
115 | */ | |
852c2936 MD |
116 | v_set(config, &buf->offset, 0); |
117 | for (i = 0; i < chan->backend.num_subbuf; i++) { | |
4746ae29 MD |
118 | v_set(config, &shmp_index(handle, buf->commit_hot, i)->cc, 0); |
119 | v_set(config, &shmp_index(handle, buf->commit_hot, i)->seq, 0); | |
120 | v_set(config, &shmp_index(handle, buf->commit_cold, i)->cc_sb, 0); | |
852c2936 | 121 | } |
a6352fd4 MD |
122 | uatomic_set(&buf->consumed, 0); |
123 | uatomic_set(&buf->record_disabled, 0); | |
852c2936 | 124 | v_set(config, &buf->last_tsc, 0); |
1d498196 | 125 | lib_ring_buffer_backend_reset(&buf->backend, handle); |
852c2936 MD |
126 | /* Don't reset number of active readers */ |
127 | v_set(config, &buf->records_lost_full, 0); | |
128 | v_set(config, &buf->records_lost_wrap, 0); | |
129 | v_set(config, &buf->records_lost_big, 0); | |
130 | v_set(config, &buf->records_count, 0); | |
131 | v_set(config, &buf->records_overrun, 0); | |
132 | buf->finalized = 0; | |
133 | } | |
852c2936 MD |
134 | |
135 | /** | |
136 | * channel_reset - Reset channel to initial values. | |
137 | * @chan: Channel. | |
138 | * | |
139 | * Effectively empty the channel. Should be called when the channel is not used | |
140 | * for writing. The channel can be opened for reading, but the reader should not | |
141 | * be using the iterator concurrently with reset. The previous current iterator | |
142 | * record is reset. | |
143 | */ | |
144 | void channel_reset(struct channel *chan) | |
145 | { | |
146 | /* | |
147 | * Reset iterators first. Will put the subbuffer if held for reading. | |
148 | */ | |
a6352fd4 | 149 | uatomic_set(&chan->record_disabled, 0); |
852c2936 MD |
150 | /* Don't reset commit_count_mask, still valid */ |
151 | channel_backend_reset(&chan->backend); | |
152 | /* Don't reset switch/read timer interval */ | |
153 | /* Don't reset notifiers and notifier enable bits */ | |
154 | /* Don't reset reader reference count */ | |
155 | } | |
852c2936 MD |
156 | |
157 | /* | |
158 | * Must be called under cpu hotplug protection. | |
159 | */ | |
4cfec15c | 160 | int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf, |
a6352fd4 | 161 | struct channel_backend *chanb, int cpu, |
38fae1d3 | 162 | struct lttng_ust_shm_handle *handle, |
1d498196 | 163 | struct shm_object *shmobj) |
852c2936 | 164 | { |
4cfec15c | 165 | const struct lttng_ust_lib_ring_buffer_config *config = &chanb->config; |
14641deb | 166 | struct channel *chan = caa_container_of(chanb, struct channel, backend); |
a3f61e7f | 167 | void *priv = channel_get_private(chan); |
852c2936 | 168 | size_t subbuf_header_size; |
2fed87ae | 169 | uint64_t tsc; |
852c2936 MD |
170 | int ret; |
171 | ||
172 | /* Test for cpu hotplug */ | |
173 | if (buf->backend.allocated) | |
174 | return 0; | |
175 | ||
a6352fd4 | 176 | ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, |
1d498196 | 177 | cpu, handle, shmobj); |
852c2936 MD |
178 | if (ret) |
179 | return ret; | |
180 | ||
1d498196 MD |
181 | align_shm(shmobj, __alignof__(struct commit_counters_hot)); |
182 | set_shmp(buf->commit_hot, | |
183 | zalloc_shm(shmobj, | |
184 | sizeof(struct commit_counters_hot) * chan->backend.num_subbuf)); | |
185 | if (!shmp(handle, buf->commit_hot)) { | |
852c2936 MD |
186 | ret = -ENOMEM; |
187 | goto free_chanbuf; | |
188 | } | |
189 | ||
1d498196 MD |
190 | align_shm(shmobj, __alignof__(struct commit_counters_cold)); |
191 | set_shmp(buf->commit_cold, | |
192 | zalloc_shm(shmobj, | |
193 | sizeof(struct commit_counters_cold) * chan->backend.num_subbuf)); | |
194 | if (!shmp(handle, buf->commit_cold)) { | |
852c2936 MD |
195 | ret = -ENOMEM; |
196 | goto free_commit; | |
197 | } | |
198 | ||
852c2936 MD |
199 | /* |
200 | * Write the subbuffer header for first subbuffer so we know the total | |
201 | * duration of data gathering. | |
202 | */ | |
203 | subbuf_header_size = config->cb.subbuffer_header_size(); | |
204 | v_set(config, &buf->offset, subbuf_header_size); | |
4746ae29 | 205 | subbuffer_id_clear_noref(config, &shmp_index(handle, buf->backend.buf_wsb, 0)->id); |
1d498196 MD |
206 | tsc = config->cb.ring_buffer_clock_read(shmp(handle, buf->backend.chan)); |
207 | config->cb.buffer_begin(buf, tsc, 0, handle); | |
4746ae29 | 208 | v_add(config, subbuf_header_size, &shmp_index(handle, buf->commit_hot, 0)->cc); |
852c2936 MD |
209 | |
210 | if (config->cb.buffer_create) { | |
1d498196 | 211 | ret = config->cb.buffer_create(buf, priv, cpu, chanb->name, handle); |
852c2936 MD |
212 | if (ret) |
213 | goto free_init; | |
214 | } | |
852c2936 | 215 | buf->backend.allocated = 1; |
852c2936 MD |
216 | return 0; |
217 | ||
218 | /* Error handling */ | |
219 | free_init: | |
a6352fd4 | 220 | /* commit_cold will be freed by shm teardown */ |
852c2936 | 221 | free_commit: |
a6352fd4 | 222 | /* commit_hot will be freed by shm teardown */ |
852c2936 | 223 | free_chanbuf: |
852c2936 MD |
224 | return ret; |
225 | } | |
226 | ||
1d498196 | 227 | #if 0 |
852c2936 MD |
228 | static void switch_buffer_timer(unsigned long data) |
229 | { | |
4cfec15c | 230 | struct lttng_ust_lib_ring_buffer *buf = (struct lttng_ust_lib_ring_buffer *)data; |
1d498196 | 231 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 232 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
233 | |
234 | /* | |
235 | * Only flush buffers periodically if readers are active. | |
236 | */ | |
824f40b8 | 237 | if (uatomic_read(&buf->active_readers) || uatomic_read(&buf->active_shadow_readers)) |
1d498196 | 238 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, handle); |
852c2936 | 239 | |
a6352fd4 MD |
240 | //TODO timers |
241 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
242 | // mod_timer_pinned(&buf->switch_timer, | |
243 | // jiffies + chan->switch_timer_interval); | |
244 | //else | |
245 | // mod_timer(&buf->switch_timer, | |
246 | // jiffies + chan->switch_timer_interval); | |
852c2936 | 247 | } |
1d498196 | 248 | #endif //0 |
852c2936 | 249 | |
4cfec15c | 250 | static void lib_ring_buffer_start_switch_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 251 | struct lttng_ust_shm_handle *handle) |
852c2936 | 252 | { |
1d498196 | 253 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 254 | //const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
255 | |
256 | if (!chan->switch_timer_interval || buf->switch_timer_enabled) | |
257 | return; | |
a6352fd4 MD |
258 | //TODO |
259 | //init_timer(&buf->switch_timer); | |
260 | //buf->switch_timer.function = switch_buffer_timer; | |
261 | //buf->switch_timer.expires = jiffies + chan->switch_timer_interval; | |
262 | //buf->switch_timer.data = (unsigned long)buf; | |
263 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
264 | // add_timer_on(&buf->switch_timer, buf->backend.cpu); | |
265 | //else | |
266 | // add_timer(&buf->switch_timer); | |
852c2936 MD |
267 | buf->switch_timer_enabled = 1; |
268 | } | |
269 | ||
4cfec15c | 270 | static void lib_ring_buffer_stop_switch_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 271 | struct lttng_ust_shm_handle *handle) |
852c2936 | 272 | { |
1d498196 | 273 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 MD |
274 | |
275 | if (!chan->switch_timer_interval || !buf->switch_timer_enabled) | |
276 | return; | |
277 | ||
a6352fd4 MD |
278 | //TODO |
279 | //del_timer_sync(&buf->switch_timer); | |
852c2936 MD |
280 | buf->switch_timer_enabled = 0; |
281 | } | |
282 | ||
1d498196 | 283 | #if 0 |
852c2936 MD |
284 | /* |
285 | * Polling timer to check the channels for data. | |
286 | */ | |
287 | static void read_buffer_timer(unsigned long data) | |
288 | { | |
4cfec15c | 289 | struct lttng_ust_lib_ring_buffer *buf = (struct lttng_ust_lib_ring_buffer *)data; |
1d498196 | 290 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 291 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
292 | |
293 | CHAN_WARN_ON(chan, !buf->backend.allocated); | |
294 | ||
824f40b8 | 295 | if (uatomic_read(&buf->active_readers) || uatomic_read(&buf->active_shadow_readers)) |
852c2936 | 296 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { |
a6352fd4 MD |
297 | //TODO |
298 | //wake_up_interruptible(&buf->read_wait); | |
299 | //wake_up_interruptible(&chan->read_wait); | |
852c2936 MD |
300 | } |
301 | ||
a6352fd4 MD |
302 | //TODO |
303 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
304 | // mod_timer_pinned(&buf->read_timer, | |
305 | // jiffies + chan->read_timer_interval); | |
306 | //else | |
307 | // mod_timer(&buf->read_timer, | |
308 | // jiffies + chan->read_timer_interval); | |
852c2936 | 309 | } |
1d498196 | 310 | #endif //0 |
852c2936 | 311 | |
4cfec15c | 312 | static void lib_ring_buffer_start_read_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 313 | struct lttng_ust_shm_handle *handle) |
852c2936 | 314 | { |
1d498196 | 315 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 316 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
317 | |
318 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
319 | || !chan->read_timer_interval | |
320 | || buf->read_timer_enabled) | |
321 | return; | |
322 | ||
a6352fd4 MD |
323 | //TODO |
324 | //init_timer(&buf->read_timer); | |
325 | //buf->read_timer.function = read_buffer_timer; | |
326 | //buf->read_timer.expires = jiffies + chan->read_timer_interval; | |
327 | //buf->read_timer.data = (unsigned long)buf; | |
852c2936 | 328 | |
a6352fd4 MD |
329 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
330 | // add_timer_on(&buf->read_timer, buf->backend.cpu); | |
331 | //else | |
332 | // add_timer(&buf->read_timer); | |
852c2936 MD |
333 | buf->read_timer_enabled = 1; |
334 | } | |
335 | ||
4cfec15c | 336 | static void lib_ring_buffer_stop_read_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 337 | struct lttng_ust_shm_handle *handle) |
852c2936 | 338 | { |
1d498196 | 339 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 340 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
341 | |
342 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
343 | || !chan->read_timer_interval | |
344 | || !buf->read_timer_enabled) | |
345 | return; | |
346 | ||
a6352fd4 MD |
347 | //TODO |
348 | //del_timer_sync(&buf->read_timer); | |
852c2936 MD |
349 | /* |
350 | * do one more check to catch data that has been written in the last | |
351 | * timer period. | |
352 | */ | |
1d498196 | 353 | if (lib_ring_buffer_poll_deliver(config, buf, chan, handle)) { |
a6352fd4 MD |
354 | //TODO |
355 | //wake_up_interruptible(&buf->read_wait); | |
356 | //wake_up_interruptible(&chan->read_wait); | |
852c2936 MD |
357 | } |
358 | buf->read_timer_enabled = 0; | |
359 | } | |
360 | ||
1d498196 | 361 | static void channel_unregister_notifiers(struct channel *chan, |
38fae1d3 | 362 | struct lttng_ust_shm_handle *handle) |
852c2936 | 363 | { |
4cfec15c | 364 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
365 | int cpu; |
366 | ||
852c2936 | 367 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
852c2936 | 368 | for_each_possible_cpu(cpu) { |
4cfec15c | 369 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
a6352fd4 | 370 | |
1d498196 MD |
371 | lib_ring_buffer_stop_switch_timer(buf, handle); |
372 | lib_ring_buffer_stop_read_timer(buf, handle); | |
852c2936 | 373 | } |
852c2936 | 374 | } else { |
4cfec15c | 375 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 376 | |
1d498196 MD |
377 | lib_ring_buffer_stop_switch_timer(buf, handle); |
378 | lib_ring_buffer_stop_read_timer(buf, handle); | |
852c2936 | 379 | } |
8d8a24c8 | 380 | //channel_backend_unregister_notifiers(&chan->backend); |
852c2936 MD |
381 | } |
382 | ||
38fae1d3 | 383 | static void channel_free(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 384 | int shadow) |
852c2936 | 385 | { |
824f40b8 MD |
386 | if (!shadow) |
387 | channel_backend_free(&chan->backend, handle); | |
431d5cf0 | 388 | /* chan is freed by shm teardown */ |
1d498196 MD |
389 | shm_object_table_destroy(handle->table); |
390 | free(handle); | |
852c2936 MD |
391 | } |
392 | ||
393 | /** | |
394 | * channel_create - Create channel. | |
395 | * @config: ring buffer instance configuration | |
396 | * @name: name of the channel | |
a3f61e7f MD |
397 | * @priv_data: ring buffer client private data area pointer (output) |
398 | * @priv_data_size: length, in bytes, of the private data area. | |
d028eddb | 399 | * @priv_data_init: initialization data for private data. |
852c2936 MD |
400 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous |
401 | * address mapping. It is used only by RING_BUFFER_STATIC | |
402 | * configuration. It can be set to NULL for other backends. | |
403 | * @subbuf_size: subbuffer size | |
404 | * @num_subbuf: number of subbuffers | |
405 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
406 | * padding to let readers get those sub-buffers. | |
407 | * Used for live streaming. | |
408 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
409 | * | |
410 | * Holds cpu hotplug. | |
411 | * Returns NULL on failure. | |
412 | */ | |
4cfec15c | 413 | struct lttng_ust_shm_handle *channel_create(const struct lttng_ust_lib_ring_buffer_config *config, |
a3f61e7f MD |
414 | const char *name, |
415 | void **priv_data, | |
416 | size_t priv_data_align, | |
417 | size_t priv_data_size, | |
d028eddb | 418 | void *priv_data_init, |
a3f61e7f | 419 | void *buf_addr, size_t subbuf_size, |
852c2936 | 420 | size_t num_subbuf, unsigned int switch_timer_interval, |
193183fb MD |
421 | unsigned int read_timer_interval, |
422 | int *shm_fd, int *wait_fd, uint64_t *memory_map_size) | |
852c2936 | 423 | { |
1d498196 | 424 | int ret, cpu; |
a3f61e7f | 425 | size_t shmsize, chansize; |
852c2936 | 426 | struct channel *chan; |
38fae1d3 | 427 | struct lttng_ust_shm_handle *handle; |
1d498196 | 428 | struct shm_object *shmobj; |
193183fb | 429 | struct shm_ref *ref; |
852c2936 MD |
430 | |
431 | if (lib_ring_buffer_check_config(config, switch_timer_interval, | |
432 | read_timer_interval)) | |
433 | return NULL; | |
434 | ||
38fae1d3 | 435 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
431d5cf0 MD |
436 | if (!handle) |
437 | return NULL; | |
438 | ||
1d498196 MD |
439 | /* Allocate table for channel + per-cpu buffers */ |
440 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
441 | if (!handle->table) | |
442 | goto error_table_alloc; | |
852c2936 | 443 | |
1d498196 MD |
444 | /* Calculate the shm allocation layout */ |
445 | shmsize = sizeof(struct channel); | |
c1fca457 | 446 | shmsize += offset_align(shmsize, __alignof__(struct lttng_ust_lib_ring_buffer_shmp)); |
1d498196 | 447 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
4cfec15c | 448 | shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp) * num_possible_cpus(); |
1d498196 | 449 | else |
4cfec15c | 450 | shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp); |
a3f61e7f MD |
451 | chansize = shmsize; |
452 | shmsize += offset_align(shmsize, priv_data_align); | |
453 | shmsize += priv_data_size; | |
a6352fd4 | 454 | |
1d498196 | 455 | shmobj = shm_object_table_append(handle->table, shmsize); |
b5a14697 MD |
456 | if (!shmobj) |
457 | goto error_append; | |
57773204 | 458 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
a3f61e7f | 459 | set_shmp(handle->chan, zalloc_shm(shmobj, chansize)); |
57773204 MD |
460 | assert(handle->chan._ref.index == 0); |
461 | assert(handle->chan._ref.offset == 0); | |
1d498196 | 462 | chan = shmp(handle, handle->chan); |
a6352fd4 | 463 | if (!chan) |
1d498196 | 464 | goto error_append; |
a6352fd4 | 465 | |
a3f61e7f MD |
466 | /* space for private data */ |
467 | if (priv_data_size) { | |
468 | DECLARE_SHMP(void, priv_data_alloc); | |
469 | ||
470 | align_shm(shmobj, priv_data_align); | |
471 | chan->priv_data_offset = shmobj->allocated_len; | |
472 | set_shmp(priv_data_alloc, zalloc_shm(shmobj, priv_data_size)); | |
473 | if (!shmp(handle, priv_data_alloc)) | |
474 | goto error_append; | |
475 | *priv_data = channel_get_private(chan); | |
d028eddb | 476 | memcpy(*priv_data, priv_data_init, priv_data_size); |
a3f61e7f MD |
477 | } else { |
478 | chan->priv_data_offset = -1; | |
479 | *priv_data = NULL; | |
480 | } | |
481 | ||
482 | ret = channel_backend_init(&chan->backend, name, config, | |
1d498196 | 483 | subbuf_size, num_subbuf, handle); |
852c2936 | 484 | if (ret) |
1d498196 | 485 | goto error_backend_init; |
852c2936 MD |
486 | |
487 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); | |
a6352fd4 MD |
488 | //TODO |
489 | //chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval); | |
490 | //chan->read_timer_interval = usecs_to_jiffies(read_timer_interval); | |
a6352fd4 MD |
491 | //TODO |
492 | //init_waitqueue_head(&chan->read_wait); | |
493 | //init_waitqueue_head(&chan->hp_wait); | |
852c2936 MD |
494 | |
495 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
852c2936 MD |
496 | /* |
497 | * In case of non-hotplug cpu, if the ring-buffer is allocated | |
498 | * in early initcall, it will not be notified of secondary cpus. | |
499 | * In that off case, we need to allocate for all possible cpus. | |
500 | */ | |
852c2936 | 501 | for_each_possible_cpu(cpu) { |
4cfec15c | 502 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
1d498196 MD |
503 | lib_ring_buffer_start_switch_timer(buf, handle); |
504 | lib_ring_buffer_start_read_timer(buf, handle); | |
852c2936 | 505 | } |
852c2936 | 506 | } else { |
4cfec15c | 507 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 508 | |
1d498196 MD |
509 | lib_ring_buffer_start_switch_timer(buf, handle); |
510 | lib_ring_buffer_start_read_timer(buf, handle); | |
852c2936 | 511 | } |
193183fb MD |
512 | ref = &handle->chan._ref; |
513 | shm_get_object_data(handle, ref, shm_fd, wait_fd, memory_map_size); | |
431d5cf0 | 514 | return handle; |
852c2936 | 515 | |
1d498196 MD |
516 | error_backend_init: |
517 | error_append: | |
518 | shm_object_table_destroy(handle->table); | |
519 | error_table_alloc: | |
431d5cf0 | 520 | free(handle); |
852c2936 MD |
521 | return NULL; |
522 | } | |
852c2936 | 523 | |
38fae1d3 | 524 | struct lttng_ust_shm_handle *channel_handle_create(int shm_fd, int wait_fd, |
193183fb MD |
525 | uint64_t memory_map_size) |
526 | { | |
38fae1d3 | 527 | struct lttng_ust_shm_handle *handle; |
193183fb MD |
528 | struct shm_object *object; |
529 | ||
38fae1d3 | 530 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
193183fb MD |
531 | if (!handle) |
532 | return NULL; | |
533 | ||
534 | /* Allocate table for channel + per-cpu buffers */ | |
535 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
536 | if (!handle->table) | |
537 | goto error_table_alloc; | |
538 | /* Add channel object */ | |
539 | object = shm_object_table_append_shadow(handle->table, | |
540 | shm_fd, wait_fd, memory_map_size); | |
541 | if (!object) | |
542 | goto error_table_object; | |
57773204 MD |
543 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
544 | handle->chan._ref.index = 0; | |
545 | handle->chan._ref.offset = 0; | |
193183fb MD |
546 | return handle; |
547 | ||
548 | error_table_object: | |
549 | shm_object_table_destroy(handle->table); | |
550 | error_table_alloc: | |
551 | free(handle); | |
552 | return NULL; | |
553 | } | |
554 | ||
38fae1d3 | 555 | int channel_handle_add_stream(struct lttng_ust_shm_handle *handle, |
193183fb MD |
556 | int shm_fd, int wait_fd, uint64_t memory_map_size) |
557 | { | |
558 | struct shm_object *object; | |
559 | ||
560 | /* Add stream object */ | |
561 | object = shm_object_table_append_shadow(handle->table, | |
562 | shm_fd, wait_fd, memory_map_size); | |
563 | if (!object) | |
564 | return -1; | |
565 | return 0; | |
566 | } | |
567 | ||
852c2936 | 568 | static |
38fae1d3 | 569 | void channel_release(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 570 | int shadow) |
852c2936 | 571 | { |
824f40b8 | 572 | channel_free(chan, handle, shadow); |
852c2936 MD |
573 | } |
574 | ||
575 | /** | |
576 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
577 | * @chan: channel to destroy | |
578 | * | |
579 | * Holds cpu hotplug. | |
431d5cf0 MD |
580 | * Call "destroy" callback, finalize channels, decrement the channel |
581 | * reference count. Note that when readers have completed data | |
582 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
a3f61e7f | 583 | * They should release their handle at that point. |
852c2936 | 584 | */ |
a3f61e7f | 585 | void channel_destroy(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 586 | int shadow) |
852c2936 | 587 | { |
824f40b8 MD |
588 | if (shadow) { |
589 | channel_release(chan, handle, shadow); | |
a3f61e7f | 590 | return; |
824f40b8 MD |
591 | } |
592 | ||
1d498196 | 593 | channel_unregister_notifiers(chan, handle); |
852c2936 | 594 | |
45e9e699 MD |
595 | /* |
596 | * Note: the consumer takes care of finalizing and switching the | |
597 | * buffers. | |
598 | */ | |
852c2936 | 599 | |
431d5cf0 MD |
600 | /* |
601 | * sessiond/consumer are keeping a reference on the shm file | |
602 | * descriptor directly. No need to refcount. | |
603 | */ | |
824f40b8 | 604 | channel_release(chan, handle, shadow); |
a3f61e7f | 605 | return; |
852c2936 | 606 | } |
852c2936 | 607 | |
4cfec15c MD |
608 | struct lttng_ust_lib_ring_buffer *channel_get_ring_buffer( |
609 | const struct lttng_ust_lib_ring_buffer_config *config, | |
1d498196 | 610 | struct channel *chan, int cpu, |
38fae1d3 | 611 | struct lttng_ust_shm_handle *handle, |
381c0f1e MD |
612 | int *shm_fd, int *wait_fd, |
613 | uint64_t *memory_map_size) | |
852c2936 | 614 | { |
381c0f1e MD |
615 | struct shm_ref *ref; |
616 | ||
617 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
618 | ref = &chan->backend.buf[0].shmp._ref; | |
619 | shm_get_object_data(handle, ref, shm_fd, wait_fd, | |
620 | memory_map_size); | |
1d498196 | 621 | return shmp(handle, chan->backend.buf[0].shmp); |
381c0f1e | 622 | } else { |
e095d803 MD |
623 | if (cpu >= num_possible_cpus()) |
624 | return NULL; | |
381c0f1e MD |
625 | ref = &chan->backend.buf[cpu].shmp._ref; |
626 | shm_get_object_data(handle, ref, shm_fd, wait_fd, | |
627 | memory_map_size); | |
1d498196 | 628 | return shmp(handle, chan->backend.buf[cpu].shmp); |
381c0f1e | 629 | } |
852c2936 | 630 | } |
852c2936 | 631 | |
4cfec15c | 632 | int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 633 | struct lttng_ust_shm_handle *handle, |
824f40b8 | 634 | int shadow) |
852c2936 | 635 | { |
824f40b8 MD |
636 | if (shadow) { |
637 | if (uatomic_cmpxchg(&buf->active_shadow_readers, 0, 1) != 0) | |
638 | return -EBUSY; | |
639 | cmm_smp_mb(); | |
640 | return 0; | |
641 | } | |
a6352fd4 | 642 | if (uatomic_cmpxchg(&buf->active_readers, 0, 1) != 0) |
852c2936 | 643 | return -EBUSY; |
a6352fd4 | 644 | cmm_smp_mb(); |
852c2936 MD |
645 | return 0; |
646 | } | |
852c2936 | 647 | |
4cfec15c | 648 | void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 649 | struct lttng_ust_shm_handle *handle, |
824f40b8 | 650 | int shadow) |
852c2936 | 651 | { |
1d498196 | 652 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 | 653 | |
824f40b8 MD |
654 | if (shadow) { |
655 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_shadow_readers) != 1); | |
656 | cmm_smp_mb(); | |
657 | uatomic_dec(&buf->active_shadow_readers); | |
658 | return; | |
659 | } | |
a6352fd4 MD |
660 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
661 | cmm_smp_mb(); | |
662 | uatomic_dec(&buf->active_readers); | |
852c2936 MD |
663 | } |
664 | ||
665 | /** | |
666 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
667 | * @buf: ring buffer | |
668 | * @consumed: consumed count indicating the position where to read | |
669 | * @produced: produced count, indicates position when to stop reading | |
670 | * | |
671 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
672 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 MD |
673 | */ |
674 | ||
4cfec15c | 675 | int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 676 | unsigned long *consumed, unsigned long *produced, |
38fae1d3 | 677 | struct lttng_ust_shm_handle *handle) |
852c2936 | 678 | { |
1d498196 | 679 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 680 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
681 | unsigned long consumed_cur, write_offset; |
682 | int finalized; | |
683 | ||
14641deb | 684 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
685 | /* |
686 | * Read finalized before counters. | |
687 | */ | |
a6352fd4 MD |
688 | cmm_smp_rmb(); |
689 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 MD |
690 | /* |
691 | * No need to issue a memory barrier between consumed count read and | |
692 | * write offset read, because consumed count can only change | |
693 | * concurrently in overwrite mode, and we keep a sequence counter | |
694 | * identifier derived from the write offset to check we are getting | |
695 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
696 | * "tagged" upon writes, tags are checked upon read). | |
697 | */ | |
698 | write_offset = v_read(config, &buf->offset); | |
699 | ||
700 | /* | |
701 | * Check that we are not about to read the same subbuffer in | |
702 | * which the writer head is. | |
703 | */ | |
704 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
705 | == 0) | |
706 | goto nodata; | |
707 | ||
708 | *consumed = consumed_cur; | |
709 | *produced = subbuf_trunc(write_offset, chan); | |
710 | ||
711 | return 0; | |
712 | ||
713 | nodata: | |
714 | /* | |
715 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
716 | * of "raw_spin_is_locked" memory ordering. | |
717 | */ | |
718 | if (finalized) | |
719 | return -ENODATA; | |
852c2936 MD |
720 | else |
721 | return -EAGAIN; | |
722 | } | |
852c2936 MD |
723 | |
724 | /** | |
725 | * lib_ring_buffer_put_snapshot - move consumed counter forward | |
726 | * @buf: ring buffer | |
727 | * @consumed_new: new consumed count value | |
728 | */ | |
4cfec15c | 729 | void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 730 | unsigned long consumed_new, |
38fae1d3 | 731 | struct lttng_ust_shm_handle *handle) |
852c2936 | 732 | { |
4cfec15c | 733 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
1d498196 | 734 | struct channel *chan = shmp(handle, bufb->chan); |
852c2936 MD |
735 | unsigned long consumed; |
736 | ||
824f40b8 MD |
737 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1 |
738 | && uatomic_read(&buf->active_shadow_readers) != 1); | |
852c2936 MD |
739 | |
740 | /* | |
741 | * Only push the consumed value forward. | |
742 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
743 | * the writer in flight recorder mode. | |
744 | */ | |
a6352fd4 | 745 | consumed = uatomic_read(&buf->consumed); |
852c2936 | 746 | while ((long) consumed - (long) consumed_new < 0) |
a6352fd4 MD |
747 | consumed = uatomic_cmpxchg(&buf->consumed, consumed, |
748 | consumed_new); | |
852c2936 | 749 | } |
852c2936 MD |
750 | |
751 | /** | |
752 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
753 | * @buf: ring buffer | |
754 | * @consumed: consumed count indicating the position where to read | |
755 | * | |
756 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
757 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 | 758 | */ |
4cfec15c | 759 | int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 760 | unsigned long consumed, |
38fae1d3 | 761 | struct lttng_ust_shm_handle *handle) |
852c2936 | 762 | { |
1d498196 | 763 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 764 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
765 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
766 | int ret; | |
767 | int finalized; | |
768 | ||
769 | retry: | |
14641deb | 770 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
771 | /* |
772 | * Read finalized before counters. | |
773 | */ | |
a6352fd4 MD |
774 | cmm_smp_rmb(); |
775 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 | 776 | consumed_idx = subbuf_index(consumed, chan); |
4746ae29 | 777 | commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb); |
852c2936 MD |
778 | /* |
779 | * Make sure we read the commit count before reading the buffer | |
780 | * data and the write offset. Correct consumed offset ordering | |
781 | * wrt commit count is insured by the use of cmpxchg to update | |
782 | * the consumed offset. | |
852c2936 | 783 | */ |
a6352fd4 MD |
784 | /* |
785 | * Local rmb to match the remote wmb to read the commit count | |
786 | * before the buffer data and the write offset. | |
787 | */ | |
788 | cmm_smp_rmb(); | |
852c2936 MD |
789 | |
790 | write_offset = v_read(config, &buf->offset); | |
791 | ||
792 | /* | |
793 | * Check that the buffer we are getting is after or at consumed_cur | |
794 | * position. | |
795 | */ | |
796 | if ((long) subbuf_trunc(consumed, chan) | |
797 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
798 | goto nodata; | |
799 | ||
800 | /* | |
801 | * Check that the subbuffer we are trying to consume has been | |
802 | * already fully committed. | |
803 | */ | |
804 | if (((commit_count - chan->backend.subbuf_size) | |
805 | & chan->commit_count_mask) | |
806 | - (buf_trunc(consumed_cur, chan) | |
807 | >> chan->backend.num_subbuf_order) | |
808 | != 0) | |
809 | goto nodata; | |
810 | ||
811 | /* | |
812 | * Check that we are not about to read the same subbuffer in | |
813 | * which the writer head is. | |
814 | */ | |
815 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
816 | == 0) | |
817 | goto nodata; | |
818 | ||
819 | /* | |
820 | * Failure to get the subbuffer causes a busy-loop retry without going | |
821 | * to a wait queue. These are caused by short-lived race windows where | |
822 | * the writer is getting access to a subbuffer we were trying to get | |
823 | * access to. Also checks that the "consumed" buffer count we are | |
824 | * looking for matches the one contained in the subbuffer id. | |
825 | */ | |
826 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
827 | consumed_idx, buf_trunc_val(consumed, chan), |
828 | handle); | |
852c2936 MD |
829 | if (ret) |
830 | goto retry; | |
831 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); | |
832 | ||
833 | buf->get_subbuf_consumed = consumed; | |
834 | buf->get_subbuf = 1; | |
835 | ||
836 | return 0; | |
837 | ||
838 | nodata: | |
839 | /* | |
840 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
841 | * of "raw_spin_is_locked" memory ordering. | |
842 | */ | |
843 | if (finalized) | |
844 | return -ENODATA; | |
852c2936 MD |
845 | else |
846 | return -EAGAIN; | |
847 | } | |
852c2936 MD |
848 | |
849 | /** | |
850 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
851 | * @buf: ring buffer | |
852 | */ | |
4cfec15c | 853 | void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 854 | struct lttng_ust_shm_handle *handle) |
852c2936 | 855 | { |
4cfec15c | 856 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
1d498196 | 857 | struct channel *chan = shmp(handle, bufb->chan); |
4cfec15c | 858 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
859 | unsigned long read_sb_bindex, consumed_idx, consumed; |
860 | ||
824f40b8 MD |
861 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1 |
862 | && uatomic_read(&buf->active_shadow_readers) != 1); | |
852c2936 MD |
863 | |
864 | if (!buf->get_subbuf) { | |
865 | /* | |
866 | * Reader puts a subbuffer it did not get. | |
867 | */ | |
868 | CHAN_WARN_ON(chan, 1); | |
869 | return; | |
870 | } | |
871 | consumed = buf->get_subbuf_consumed; | |
872 | buf->get_subbuf = 0; | |
873 | ||
874 | /* | |
875 | * Clear the records_unread counter. (overruns counter) | |
876 | * Can still be non-zero if a file reader simply grabbed the data | |
877 | * without using iterators. | |
878 | * Can be below zero if an iterator is used on a snapshot more than | |
879 | * once. | |
880 | */ | |
881 | read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); | |
882 | v_add(config, v_read(config, | |
4746ae29 | 883 | &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread), |
852c2936 | 884 | &bufb->records_read); |
4746ae29 | 885 | v_set(config, &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread, 0); |
852c2936 MD |
886 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE |
887 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
888 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
889 | ||
890 | /* | |
891 | * Exchange the reader subbuffer with the one we put in its place in the | |
892 | * writer subbuffer table. Expect the original consumed count. If | |
893 | * update_read_sb_index fails, this is because the writer updated the | |
894 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
895 | * currently have: it has become invalid to try reading this sub-buffer | |
896 | * consumed count value anyway. | |
897 | */ | |
898 | consumed_idx = subbuf_index(consumed, chan); | |
899 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
900 | consumed_idx, buf_trunc_val(consumed, chan), |
901 | handle); | |
852c2936 MD |
902 | /* |
903 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
904 | * if the writer concurrently updated it. | |
905 | */ | |
906 | } | |
852c2936 MD |
907 | |
908 | /* | |
909 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
910 | * position and the writer position. (inclusive) | |
911 | */ | |
912 | static | |
4cfec15c | 913 | void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
914 | struct channel *chan, |
915 | unsigned long cons_offset, | |
1d498196 | 916 | int cpu, |
38fae1d3 | 917 | struct lttng_ust_shm_handle *handle) |
852c2936 | 918 | { |
4cfec15c | 919 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
920 | unsigned long cons_idx, commit_count, commit_count_sb; |
921 | ||
922 | cons_idx = subbuf_index(cons_offset, chan); | |
4746ae29 MD |
923 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, cons_idx)->cc); |
924 | commit_count_sb = v_read(config, &shmp_index(handle, buf->commit_cold, cons_idx)->cc_sb); | |
852c2936 MD |
925 | |
926 | if (subbuf_offset(commit_count, chan) != 0) | |
4d3c9523 | 927 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
928 | "commit count in subbuffer %lu,\n" |
929 | "expecting multiples of %lu bytes\n" | |
930 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
931 | chan->backend.name, cpu, cons_idx, | |
932 | chan->backend.subbuf_size, | |
933 | commit_count, commit_count_sb); | |
934 | ||
4d3c9523 | 935 | DBG("ring buffer: %s, cpu %d: %lu bytes committed\n", |
852c2936 MD |
936 | chan->backend.name, cpu, commit_count); |
937 | } | |
938 | ||
939 | static | |
4cfec15c | 940 | void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 | 941 | struct channel *chan, |
1d498196 | 942 | void *priv, int cpu, |
38fae1d3 | 943 | struct lttng_ust_shm_handle *handle) |
852c2936 | 944 | { |
4cfec15c | 945 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
946 | unsigned long write_offset, cons_offset; |
947 | ||
948 | /* | |
949 | * Can be called in the error path of allocation when | |
950 | * trans_channel_data is not yet set. | |
951 | */ | |
952 | if (!chan) | |
953 | return; | |
954 | /* | |
955 | * No need to order commit_count, write_offset and cons_offset reads | |
956 | * because we execute at teardown when no more writer nor reader | |
957 | * references are left. | |
958 | */ | |
959 | write_offset = v_read(config, &buf->offset); | |
a6352fd4 | 960 | cons_offset = uatomic_read(&buf->consumed); |
852c2936 | 961 | if (write_offset != cons_offset) |
4d3c9523 | 962 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
963 | "non-consumed data\n" |
964 | " [ %lu bytes written, %lu bytes read ]\n", | |
965 | chan->backend.name, cpu, write_offset, cons_offset); | |
966 | ||
a6352fd4 | 967 | for (cons_offset = uatomic_read(&buf->consumed); |
852c2936 MD |
968 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), |
969 | chan) | |
970 | - cons_offset) > 0; | |
971 | cons_offset = subbuf_align(cons_offset, chan)) | |
972 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1d498196 | 973 | cpu, handle); |
852c2936 MD |
974 | } |
975 | ||
976 | static | |
977 | void lib_ring_buffer_print_errors(struct channel *chan, | |
4cfec15c | 978 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
38fae1d3 | 979 | struct lttng_ust_shm_handle *handle) |
852c2936 | 980 | { |
4cfec15c | 981 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
a3f61e7f | 982 | void *priv = channel_get_private(chan); |
852c2936 | 983 | |
4d3c9523 | 984 | DBG("ring buffer %s, cpu %d: %lu records written, " |
852c2936 MD |
985 | "%lu records overrun\n", |
986 | chan->backend.name, cpu, | |
987 | v_read(config, &buf->records_count), | |
988 | v_read(config, &buf->records_overrun)); | |
989 | ||
990 | if (v_read(config, &buf->records_lost_full) | |
991 | || v_read(config, &buf->records_lost_wrap) | |
992 | || v_read(config, &buf->records_lost_big)) | |
4d3c9523 | 993 | DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n" |
852c2936 MD |
994 | " [ %lu buffer full, %lu nest buffer wrap-around, " |
995 | "%lu event too big ]\n", | |
996 | chan->backend.name, cpu, | |
997 | v_read(config, &buf->records_lost_full), | |
998 | v_read(config, &buf->records_lost_wrap), | |
999 | v_read(config, &buf->records_lost_big)); | |
1000 | ||
1d498196 | 1001 | lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu, handle); |
852c2936 MD |
1002 | } |
1003 | ||
1004 | /* | |
1005 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
1006 | * | |
1007 | * Only executed when the buffer is finalized, in SWITCH_FLUSH. | |
1008 | */ | |
1009 | static | |
4cfec15c | 1010 | void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1011 | struct channel *chan, |
1012 | struct switch_offsets *offsets, | |
2fed87ae | 1013 | uint64_t tsc, |
38fae1d3 | 1014 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1015 | { |
4cfec15c | 1016 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1017 | unsigned long oldidx = subbuf_index(offsets->old, chan); |
1018 | unsigned long commit_count; | |
1019 | ||
1d498196 | 1020 | config->cb.buffer_begin(buf, tsc, oldidx, handle); |
852c2936 MD |
1021 | |
1022 | /* | |
1023 | * Order all writes to buffer before the commit count update that will | |
1024 | * determine that the subbuffer is full. | |
1025 | */ | |
a6352fd4 | 1026 | cmm_smp_wmb(); |
852c2936 | 1027 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1028 | &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1029 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 MD |
1030 | /* Check if the written buffer has to be delivered */ |
1031 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
1d498196 | 1032 | commit_count, oldidx, handle); |
852c2936 MD |
1033 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
1034 | offsets->old, commit_count, | |
1d498196 MD |
1035 | config->cb.subbuffer_header_size(), |
1036 | handle); | |
852c2936 MD |
1037 | } |
1038 | ||
1039 | /* | |
1040 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
1041 | * | |
1042 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
1043 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
1044 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
1045 | * subbuffer. | |
1046 | */ | |
1047 | static | |
4cfec15c | 1048 | void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1049 | struct channel *chan, |
1050 | struct switch_offsets *offsets, | |
2fed87ae | 1051 | uint64_t tsc, |
38fae1d3 | 1052 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1053 | { |
4cfec15c | 1054 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1055 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); |
1056 | unsigned long commit_count, padding_size, data_size; | |
1057 | ||
1058 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1059 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1060 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size, |
1061 | handle); | |
852c2936 MD |
1062 | |
1063 | /* | |
1064 | * Order all writes to buffer before the commit count update that will | |
1065 | * determine that the subbuffer is full. | |
1066 | */ | |
a6352fd4 | 1067 | cmm_smp_wmb(); |
4746ae29 MD |
1068 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1069 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 | 1070 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, |
1d498196 | 1071 | commit_count, oldidx, handle); |
852c2936 MD |
1072 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
1073 | offsets->old, commit_count, | |
1d498196 | 1074 | padding_size, handle); |
852c2936 MD |
1075 | } |
1076 | ||
1077 | /* | |
1078 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1079 | * | |
1080 | * This code can be executed unordered : writers may already have written to the | |
1081 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1082 | * that this code is executed before the deliver of this sub-buffer. | |
1083 | */ | |
1084 | static | |
4cfec15c | 1085 | void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1086 | struct channel *chan, |
1087 | struct switch_offsets *offsets, | |
2fed87ae | 1088 | uint64_t tsc, |
38fae1d3 | 1089 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1090 | { |
4cfec15c | 1091 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1092 | unsigned long beginidx = subbuf_index(offsets->begin, chan); |
1093 | unsigned long commit_count; | |
1094 | ||
1d498196 | 1095 | config->cb.buffer_begin(buf, tsc, beginidx, handle); |
852c2936 MD |
1096 | |
1097 | /* | |
1098 | * Order all writes to buffer before the commit count update that will | |
1099 | * determine that the subbuffer is full. | |
1100 | */ | |
a6352fd4 | 1101 | cmm_smp_wmb(); |
852c2936 | 1102 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1103 | &shmp_index(handle, buf->commit_hot, beginidx)->cc); |
1104 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, beginidx)->cc); | |
852c2936 MD |
1105 | /* Check if the written buffer has to be delivered */ |
1106 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
1d498196 | 1107 | commit_count, beginidx, handle); |
852c2936 MD |
1108 | lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx, |
1109 | offsets->begin, commit_count, | |
1d498196 MD |
1110 | config->cb.subbuffer_header_size(), |
1111 | handle); | |
852c2936 MD |
1112 | } |
1113 | ||
1114 | /* | |
1115 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1116 | * | |
1117 | * The only remaining threads could be the ones with pending commits. They will | |
1118 | * have to do the deliver themselves. | |
1119 | */ | |
1120 | static | |
4cfec15c | 1121 | void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 MD |
1122 | struct channel *chan, |
1123 | struct switch_offsets *offsets, | |
2fed87ae | 1124 | uint64_t tsc, |
38fae1d3 | 1125 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1126 | { |
4cfec15c | 1127 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1128 | unsigned long endidx = subbuf_index(offsets->end - 1, chan); |
1129 | unsigned long commit_count, padding_size, data_size; | |
1130 | ||
1131 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; | |
1132 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1133 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size, |
1134 | handle); | |
852c2936 MD |
1135 | |
1136 | /* | |
1137 | * Order all writes to buffer before the commit count update that will | |
1138 | * determine that the subbuffer is full. | |
1139 | */ | |
a6352fd4 | 1140 | cmm_smp_wmb(); |
4746ae29 MD |
1141 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, endidx)->cc); |
1142 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc); | |
852c2936 | 1143 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->end - 1, |
1d498196 | 1144 | commit_count, endidx, handle); |
852c2936 MD |
1145 | lib_ring_buffer_write_commit_counter(config, buf, chan, endidx, |
1146 | offsets->end, commit_count, | |
1d498196 | 1147 | padding_size, handle); |
852c2936 MD |
1148 | } |
1149 | ||
1150 | /* | |
1151 | * Returns : | |
1152 | * 0 if ok | |
1153 | * !0 if execution must be aborted. | |
1154 | */ | |
1155 | static | |
1156 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
4cfec15c | 1157 | struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1158 | struct channel *chan, |
1159 | struct switch_offsets *offsets, | |
2fed87ae | 1160 | uint64_t *tsc) |
852c2936 | 1161 | { |
4cfec15c | 1162 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1163 | unsigned long off; |
1164 | ||
1165 | offsets->begin = v_read(config, &buf->offset); | |
1166 | offsets->old = offsets->begin; | |
1167 | offsets->switch_old_start = 0; | |
1168 | off = subbuf_offset(offsets->begin, chan); | |
1169 | ||
1170 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1171 | ||
1172 | /* | |
1173 | * Ensure we flush the header of an empty subbuffer when doing the | |
1174 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1175 | * total data gathering duration even if there were no records saved | |
1176 | * after the last buffer switch. | |
1177 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1178 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1179 | * subbuffer header as appropriate. | |
1180 | * The next record that reserves space will be responsible for | |
1181 | * populating the following subbuffer header. We choose not to populate | |
1182 | * the next subbuffer header here because we want to be able to use | |
a6352fd4 MD |
1183 | * SWITCH_ACTIVE for periodical buffer flush, which must |
1184 | * guarantee that all the buffer content (records and header | |
1185 | * timestamps) are visible to the reader. This is required for | |
1186 | * quiescence guarantees for the fusion merge. | |
852c2936 MD |
1187 | */ |
1188 | if (mode == SWITCH_FLUSH || off > 0) { | |
b5a3dfa5 | 1189 | if (caa_unlikely(off == 0)) { |
852c2936 MD |
1190 | /* |
1191 | * The client does not save any header information. | |
1192 | * Don't switch empty subbuffer on finalize, because it | |
1193 | * is invalid to deliver a completely empty subbuffer. | |
1194 | */ | |
1195 | if (!config->cb.subbuffer_header_size()) | |
1196 | return -1; | |
1197 | /* | |
1198 | * Need to write the subbuffer start header on finalize. | |
1199 | */ | |
1200 | offsets->switch_old_start = 1; | |
1201 | } | |
1202 | offsets->begin = subbuf_align(offsets->begin, chan); | |
1203 | } else | |
1204 | return -1; /* we do not have to switch : buffer is empty */ | |
1205 | /* Note: old points to the next subbuf at offset 0 */ | |
1206 | offsets->end = offsets->begin; | |
1207 | return 0; | |
1208 | } | |
1209 | ||
1210 | /* | |
1211 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
1212 | * called while tracing is active with absolutely no lock held. | |
1213 | * | |
1214 | * Note, however, that as a v_cmpxchg is used for some atomic | |
1215 | * operations, this function must be called from the CPU which owns the buffer | |
1216 | * for a ACTIVE flush. | |
1217 | */ | |
4cfec15c | 1218 | void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode, |
38fae1d3 | 1219 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1220 | { |
1d498196 | 1221 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 1222 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1223 | struct switch_offsets offsets; |
1224 | unsigned long oldidx; | |
2fed87ae | 1225 | uint64_t tsc; |
852c2936 MD |
1226 | |
1227 | offsets.size = 0; | |
1228 | ||
1229 | /* | |
1230 | * Perform retryable operations. | |
1231 | */ | |
1232 | do { | |
1233 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
1234 | &tsc)) | |
1235 | return; /* Switch not needed */ | |
1236 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
1237 | != offsets.old); | |
1238 | ||
1239 | /* | |
1240 | * Atomically update last_tsc. This update races against concurrent | |
1241 | * atomic updates, but the race will always cause supplementary full TSC | |
1242 | * records, never the opposite (missing a full TSC record when it would | |
1243 | * be needed). | |
1244 | */ | |
1245 | save_last_tsc(config, buf, tsc); | |
1246 | ||
1247 | /* | |
1248 | * Push the reader if necessary | |
1249 | */ | |
1250 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
1251 | ||
1252 | oldidx = subbuf_index(offsets.old, chan); | |
1d498196 | 1253 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx, handle); |
852c2936 MD |
1254 | |
1255 | /* | |
1256 | * May need to populate header start on SWITCH_FLUSH. | |
1257 | */ | |
1258 | if (offsets.switch_old_start) { | |
1d498196 | 1259 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc, handle); |
852c2936 MD |
1260 | offsets.old += config->cb.subbuffer_header_size(); |
1261 | } | |
1262 | ||
1263 | /* | |
1264 | * Switch old subbuffer. | |
1265 | */ | |
1d498196 | 1266 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc, handle); |
852c2936 | 1267 | } |
852c2936 MD |
1268 | |
1269 | /* | |
1270 | * Returns : | |
1271 | * 0 if ok | |
1272 | * -ENOSPC if event size is too large for packet. | |
1273 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
1274 | * -EIO if data cannot be written into the buffer for any other reason. | |
1275 | */ | |
1276 | static | |
4cfec15c | 1277 | int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1278 | struct channel *chan, |
1279 | struct switch_offsets *offsets, | |
4cfec15c | 1280 | struct lttng_ust_lib_ring_buffer_ctx *ctx) |
852c2936 | 1281 | { |
4cfec15c | 1282 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
38fae1d3 | 1283 | struct lttng_ust_shm_handle *handle = ctx->handle; |
852c2936 MD |
1284 | unsigned long reserve_commit_diff; |
1285 | ||
1286 | offsets->begin = v_read(config, &buf->offset); | |
1287 | offsets->old = offsets->begin; | |
1288 | offsets->switch_new_start = 0; | |
1289 | offsets->switch_new_end = 0; | |
1290 | offsets->switch_old_end = 0; | |
1291 | offsets->pre_header_padding = 0; | |
1292 | ||
1293 | ctx->tsc = config->cb.ring_buffer_clock_read(chan); | |
1294 | if ((int64_t) ctx->tsc == -EIO) | |
1295 | return -EIO; | |
1296 | ||
1297 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
1298 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
1299 | ||
b5a3dfa5 | 1300 | if (caa_unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) { |
852c2936 MD |
1301 | offsets->switch_new_start = 1; /* For offsets->begin */ |
1302 | } else { | |
1303 | offsets->size = config->cb.record_header_size(config, chan, | |
1304 | offsets->begin, | |
1305 | &offsets->pre_header_padding, | |
1306 | ctx); | |
1307 | offsets->size += | |
1308 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1309 | ctx->largest_align) | |
1310 | + ctx->data_size; | |
b5a3dfa5 | 1311 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) + |
852c2936 MD |
1312 | offsets->size > chan->backend.subbuf_size)) { |
1313 | offsets->switch_old_end = 1; /* For offsets->old */ | |
1314 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1315 | } | |
1316 | } | |
b5a3dfa5 | 1317 | if (caa_unlikely(offsets->switch_new_start)) { |
852c2936 MD |
1318 | unsigned long sb_index; |
1319 | ||
1320 | /* | |
1321 | * We are typically not filling the previous buffer completely. | |
1322 | */ | |
b5a3dfa5 | 1323 | if (caa_likely(offsets->switch_old_end)) |
852c2936 MD |
1324 | offsets->begin = subbuf_align(offsets->begin, chan); |
1325 | offsets->begin = offsets->begin | |
1326 | + config->cb.subbuffer_header_size(); | |
1327 | /* Test new buffer integrity */ | |
1328 | sb_index = subbuf_index(offsets->begin, chan); | |
1329 | reserve_commit_diff = | |
1330 | (buf_trunc(offsets->begin, chan) | |
1331 | >> chan->backend.num_subbuf_order) | |
1332 | - ((unsigned long) v_read(config, | |
4746ae29 | 1333 | &shmp_index(handle, buf->commit_cold, sb_index)->cc_sb) |
852c2936 | 1334 | & chan->commit_count_mask); |
b5a3dfa5 | 1335 | if (caa_likely(reserve_commit_diff == 0)) { |
852c2936 | 1336 | /* Next subbuffer not being written to. */ |
b5a3dfa5 | 1337 | if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE && |
852c2936 MD |
1338 | subbuf_trunc(offsets->begin, chan) |
1339 | - subbuf_trunc((unsigned long) | |
a6352fd4 | 1340 | uatomic_read(&buf->consumed), chan) |
852c2936 MD |
1341 | >= chan->backend.buf_size)) { |
1342 | /* | |
1343 | * We do not overwrite non consumed buffers | |
1344 | * and we are full : record is lost. | |
1345 | */ | |
1346 | v_inc(config, &buf->records_lost_full); | |
1347 | return -ENOBUFS; | |
1348 | } else { | |
1349 | /* | |
1350 | * Next subbuffer not being written to, and we | |
1351 | * are either in overwrite mode or the buffer is | |
1352 | * not full. It's safe to write in this new | |
1353 | * subbuffer. | |
1354 | */ | |
1355 | } | |
1356 | } else { | |
1357 | /* | |
1358 | * Next subbuffer reserve offset does not match the | |
1359 | * commit offset. Drop record in producer-consumer and | |
1360 | * overwrite mode. Caused by either a writer OOPS or too | |
1361 | * many nested writes over a reserve/commit pair. | |
1362 | */ | |
1363 | v_inc(config, &buf->records_lost_wrap); | |
1364 | return -EIO; | |
1365 | } | |
1366 | offsets->size = | |
1367 | config->cb.record_header_size(config, chan, | |
1368 | offsets->begin, | |
1369 | &offsets->pre_header_padding, | |
1370 | ctx); | |
1371 | offsets->size += | |
1372 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1373 | ctx->largest_align) | |
1374 | + ctx->data_size; | |
b5a3dfa5 | 1375 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) |
852c2936 MD |
1376 | + offsets->size > chan->backend.subbuf_size)) { |
1377 | /* | |
1378 | * Record too big for subbuffers, report error, don't | |
1379 | * complete the sub-buffer switch. | |
1380 | */ | |
1381 | v_inc(config, &buf->records_lost_big); | |
1382 | return -ENOSPC; | |
1383 | } else { | |
1384 | /* | |
1385 | * We just made a successful buffer switch and the | |
1386 | * record fits in the new subbuffer. Let's write. | |
1387 | */ | |
1388 | } | |
1389 | } else { | |
1390 | /* | |
1391 | * Record fits in the current buffer and we are not on a switch | |
1392 | * boundary. It's safe to write. | |
1393 | */ | |
1394 | } | |
1395 | offsets->end = offsets->begin + offsets->size; | |
1396 | ||
b5a3dfa5 | 1397 | if (caa_unlikely(subbuf_offset(offsets->end, chan) == 0)) { |
852c2936 MD |
1398 | /* |
1399 | * The offset_end will fall at the very beginning of the next | |
1400 | * subbuffer. | |
1401 | */ | |
1402 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
1403 | } | |
1404 | return 0; | |
1405 | } | |
1406 | ||
1407 | /** | |
1408 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
1409 | * @ctx: ring buffer context. | |
1410 | * | |
1411 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, | |
1412 | * -EIO for other errors, else returns 0. | |
1413 | * It will take care of sub-buffer switching. | |
1414 | */ | |
4cfec15c | 1415 | int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx) |
852c2936 MD |
1416 | { |
1417 | struct channel *chan = ctx->chan; | |
38fae1d3 | 1418 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c MD |
1419 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
1420 | struct lttng_ust_lib_ring_buffer *buf; | |
852c2936 MD |
1421 | struct switch_offsets offsets; |
1422 | int ret; | |
1423 | ||
1424 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1d498196 | 1425 | buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp); |
852c2936 | 1426 | else |
1d498196 | 1427 | buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 MD |
1428 | ctx->buf = buf; |
1429 | ||
1430 | offsets.size = 0; | |
1431 | ||
1432 | do { | |
1433 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, | |
1434 | ctx); | |
b5a3dfa5 | 1435 | if (caa_unlikely(ret)) |
852c2936 | 1436 | return ret; |
b5a3dfa5 | 1437 | } while (caa_unlikely(v_cmpxchg(config, &buf->offset, offsets.old, |
852c2936 MD |
1438 | offsets.end) |
1439 | != offsets.old)); | |
1440 | ||
1441 | /* | |
1442 | * Atomically update last_tsc. This update races against concurrent | |
1443 | * atomic updates, but the race will always cause supplementary full TSC | |
1444 | * records, never the opposite (missing a full TSC record when it would | |
1445 | * be needed). | |
1446 | */ | |
1447 | save_last_tsc(config, buf, ctx->tsc); | |
1448 | ||
1449 | /* | |
1450 | * Push the reader if necessary | |
1451 | */ | |
1452 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
1453 | ||
1454 | /* | |
1455 | * Clear noref flag for this subbuffer. | |
1456 | */ | |
1457 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1d498196 MD |
1458 | subbuf_index(offsets.end - 1, chan), |
1459 | handle); | |
852c2936 MD |
1460 | |
1461 | /* | |
1462 | * Switch old subbuffer if needed. | |
1463 | */ | |
b5a3dfa5 | 1464 | if (caa_unlikely(offsets.switch_old_end)) { |
852c2936 | 1465 | lib_ring_buffer_clear_noref(config, &buf->backend, |
1d498196 MD |
1466 | subbuf_index(offsets.old - 1, chan), |
1467 | handle); | |
1468 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc, handle); | |
852c2936 MD |
1469 | } |
1470 | ||
1471 | /* | |
1472 | * Populate new subbuffer. | |
1473 | */ | |
b5a3dfa5 | 1474 | if (caa_unlikely(offsets.switch_new_start)) |
1d498196 | 1475 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 | 1476 | |
b5a3dfa5 | 1477 | if (caa_unlikely(offsets.switch_new_end)) |
1d498196 | 1478 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 MD |
1479 | |
1480 | ctx->slot_size = offsets.size; | |
1481 | ctx->pre_offset = offsets.begin; | |
1482 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
1483 | return 0; | |
1484 | } |