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> |
4931a13e MD |
50 | #include "backend.h" |
51 | #include "frontend.h" | |
a6352fd4 | 52 | #include "shm.h" |
852c2936 | 53 | |
431d5cf0 MD |
54 | #ifndef max |
55 | #define max(a, b) ((a) > (b) ? (a) : (b)) | |
56 | #endif | |
57 | ||
2432c3c9 MD |
58 | /* |
59 | * Use POSIX SHM: shm_open(3) and shm_unlink(3). | |
60 | * close(2) to close the fd returned by shm_open. | |
61 | * shm_unlink releases the shared memory object name. | |
62 | * ftruncate(2) sets the size of the memory object. | |
63 | * mmap/munmap maps the shared memory obj to a virtual address in the | |
64 | * calling proceess (should be done both in libust and consumer). | |
65 | * See shm_overview(7) for details. | |
66 | * Pass file descriptor returned by shm_open(3) to ltt-sessiond through | |
67 | * a UNIX socket. | |
68 | * | |
69 | * Since we don't need to access the object using its name, we can | |
70 | * immediately shm_unlink(3) it, and only keep the handle with its file | |
71 | * descriptor. | |
72 | */ | |
73 | ||
852c2936 MD |
74 | /* |
75 | * Internal structure representing offsets to use at a sub-buffer switch. | |
76 | */ | |
77 | struct switch_offsets { | |
78 | unsigned long begin, end, old; | |
79 | size_t pre_header_padding, size; | |
80 | unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1, | |
81 | switch_old_end:1; | |
82 | }; | |
83 | ||
a6352fd4 | 84 | __thread unsigned int lib_ring_buffer_nesting; |
852c2936 | 85 | |
45e9e699 MD |
86 | /* |
87 | * TODO: this is unused. Errors are saved within the ring buffer. | |
88 | * Eventually, allow consumerd to print these errors. | |
89 | */ | |
852c2936 MD |
90 | static |
91 | void lib_ring_buffer_print_errors(struct channel *chan, | |
4cfec15c | 92 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
38fae1d3 | 93 | struct lttng_ust_shm_handle *handle); |
852c2936 | 94 | |
852c2936 MD |
95 | /** |
96 | * lib_ring_buffer_reset - Reset ring buffer to initial values. | |
97 | * @buf: Ring buffer. | |
98 | * | |
99 | * Effectively empty the ring buffer. Should be called when the buffer is not | |
100 | * used for writing. The ring buffer can be opened for reading, but the reader | |
101 | * should not be using the iterator concurrently with reset. The previous | |
102 | * current iterator record is reset. | |
103 | */ | |
4cfec15c | 104 | void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 105 | struct lttng_ust_shm_handle *handle) |
852c2936 | 106 | { |
1d498196 | 107 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 108 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
109 | unsigned int i; |
110 | ||
111 | /* | |
112 | * Reset iterator first. It will put the subbuffer if it currently holds | |
113 | * it. | |
114 | */ | |
852c2936 MD |
115 | v_set(config, &buf->offset, 0); |
116 | for (i = 0; i < chan->backend.num_subbuf; i++) { | |
4746ae29 MD |
117 | v_set(config, &shmp_index(handle, buf->commit_hot, i)->cc, 0); |
118 | v_set(config, &shmp_index(handle, buf->commit_hot, i)->seq, 0); | |
119 | v_set(config, &shmp_index(handle, buf->commit_cold, i)->cc_sb, 0); | |
852c2936 | 120 | } |
a6352fd4 MD |
121 | uatomic_set(&buf->consumed, 0); |
122 | uatomic_set(&buf->record_disabled, 0); | |
852c2936 | 123 | v_set(config, &buf->last_tsc, 0); |
1d498196 | 124 | lib_ring_buffer_backend_reset(&buf->backend, handle); |
852c2936 MD |
125 | /* Don't reset number of active readers */ |
126 | v_set(config, &buf->records_lost_full, 0); | |
127 | v_set(config, &buf->records_lost_wrap, 0); | |
128 | v_set(config, &buf->records_lost_big, 0); | |
129 | v_set(config, &buf->records_count, 0); | |
130 | v_set(config, &buf->records_overrun, 0); | |
131 | buf->finalized = 0; | |
132 | } | |
852c2936 MD |
133 | |
134 | /** | |
135 | * channel_reset - Reset channel to initial values. | |
136 | * @chan: Channel. | |
137 | * | |
138 | * Effectively empty the channel. Should be called when the channel is not used | |
139 | * for writing. The channel can be opened for reading, but the reader should not | |
140 | * be using the iterator concurrently with reset. The previous current iterator | |
141 | * record is reset. | |
142 | */ | |
143 | void channel_reset(struct channel *chan) | |
144 | { | |
145 | /* | |
146 | * Reset iterators first. Will put the subbuffer if held for reading. | |
147 | */ | |
a6352fd4 | 148 | uatomic_set(&chan->record_disabled, 0); |
852c2936 MD |
149 | /* Don't reset commit_count_mask, still valid */ |
150 | channel_backend_reset(&chan->backend); | |
151 | /* Don't reset switch/read timer interval */ | |
152 | /* Don't reset notifiers and notifier enable bits */ | |
153 | /* Don't reset reader reference count */ | |
154 | } | |
852c2936 MD |
155 | |
156 | /* | |
157 | * Must be called under cpu hotplug protection. | |
158 | */ | |
4cfec15c | 159 | int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf, |
a6352fd4 | 160 | struct channel_backend *chanb, int cpu, |
38fae1d3 | 161 | struct lttng_ust_shm_handle *handle, |
1d498196 | 162 | struct shm_object *shmobj) |
852c2936 | 163 | { |
4cfec15c | 164 | const struct lttng_ust_lib_ring_buffer_config *config = &chanb->config; |
14641deb | 165 | struct channel *chan = caa_container_of(chanb, struct channel, backend); |
a3f61e7f | 166 | void *priv = channel_get_private(chan); |
852c2936 MD |
167 | size_t subbuf_header_size; |
168 | u64 tsc; | |
169 | int ret; | |
170 | ||
171 | /* Test for cpu hotplug */ | |
172 | if (buf->backend.allocated) | |
173 | return 0; | |
174 | ||
a6352fd4 | 175 | ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, |
1d498196 | 176 | cpu, handle, shmobj); |
852c2936 MD |
177 | if (ret) |
178 | return ret; | |
179 | ||
1d498196 MD |
180 | align_shm(shmobj, __alignof__(struct commit_counters_hot)); |
181 | set_shmp(buf->commit_hot, | |
182 | zalloc_shm(shmobj, | |
183 | sizeof(struct commit_counters_hot) * chan->backend.num_subbuf)); | |
184 | if (!shmp(handle, buf->commit_hot)) { | |
852c2936 MD |
185 | ret = -ENOMEM; |
186 | goto free_chanbuf; | |
187 | } | |
188 | ||
1d498196 MD |
189 | align_shm(shmobj, __alignof__(struct commit_counters_cold)); |
190 | set_shmp(buf->commit_cold, | |
191 | zalloc_shm(shmobj, | |
192 | sizeof(struct commit_counters_cold) * chan->backend.num_subbuf)); | |
193 | if (!shmp(handle, buf->commit_cold)) { | |
852c2936 MD |
194 | ret = -ENOMEM; |
195 | goto free_commit; | |
196 | } | |
197 | ||
852c2936 MD |
198 | /* |
199 | * Write the subbuffer header for first subbuffer so we know the total | |
200 | * duration of data gathering. | |
201 | */ | |
202 | subbuf_header_size = config->cb.subbuffer_header_size(); | |
203 | v_set(config, &buf->offset, subbuf_header_size); | |
4746ae29 | 204 | subbuffer_id_clear_noref(config, &shmp_index(handle, buf->backend.buf_wsb, 0)->id); |
1d498196 MD |
205 | tsc = config->cb.ring_buffer_clock_read(shmp(handle, buf->backend.chan)); |
206 | config->cb.buffer_begin(buf, tsc, 0, handle); | |
4746ae29 | 207 | v_add(config, subbuf_header_size, &shmp_index(handle, buf->commit_hot, 0)->cc); |
852c2936 MD |
208 | |
209 | if (config->cb.buffer_create) { | |
1d498196 | 210 | ret = config->cb.buffer_create(buf, priv, cpu, chanb->name, handle); |
852c2936 MD |
211 | if (ret) |
212 | goto free_init; | |
213 | } | |
852c2936 | 214 | buf->backend.allocated = 1; |
852c2936 MD |
215 | return 0; |
216 | ||
217 | /* Error handling */ | |
218 | free_init: | |
a6352fd4 | 219 | /* commit_cold will be freed by shm teardown */ |
852c2936 | 220 | free_commit: |
a6352fd4 | 221 | /* commit_hot will be freed by shm teardown */ |
852c2936 | 222 | free_chanbuf: |
852c2936 MD |
223 | return ret; |
224 | } | |
225 | ||
1d498196 | 226 | #if 0 |
852c2936 MD |
227 | static void switch_buffer_timer(unsigned long data) |
228 | { | |
4cfec15c | 229 | struct lttng_ust_lib_ring_buffer *buf = (struct lttng_ust_lib_ring_buffer *)data; |
1d498196 | 230 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 231 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
232 | |
233 | /* | |
234 | * Only flush buffers periodically if readers are active. | |
235 | */ | |
824f40b8 | 236 | if (uatomic_read(&buf->active_readers) || uatomic_read(&buf->active_shadow_readers)) |
1d498196 | 237 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, handle); |
852c2936 | 238 | |
a6352fd4 MD |
239 | //TODO timers |
240 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
241 | // mod_timer_pinned(&buf->switch_timer, | |
242 | // jiffies + chan->switch_timer_interval); | |
243 | //else | |
244 | // mod_timer(&buf->switch_timer, | |
245 | // jiffies + chan->switch_timer_interval); | |
852c2936 | 246 | } |
1d498196 | 247 | #endif //0 |
852c2936 | 248 | |
4cfec15c | 249 | static void lib_ring_buffer_start_switch_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 250 | struct lttng_ust_shm_handle *handle) |
852c2936 | 251 | { |
1d498196 | 252 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 253 | //const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
254 | |
255 | if (!chan->switch_timer_interval || buf->switch_timer_enabled) | |
256 | return; | |
a6352fd4 MD |
257 | //TODO |
258 | //init_timer(&buf->switch_timer); | |
259 | //buf->switch_timer.function = switch_buffer_timer; | |
260 | //buf->switch_timer.expires = jiffies + chan->switch_timer_interval; | |
261 | //buf->switch_timer.data = (unsigned long)buf; | |
262 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
263 | // add_timer_on(&buf->switch_timer, buf->backend.cpu); | |
264 | //else | |
265 | // add_timer(&buf->switch_timer); | |
852c2936 MD |
266 | buf->switch_timer_enabled = 1; |
267 | } | |
268 | ||
4cfec15c | 269 | static void lib_ring_buffer_stop_switch_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 270 | struct lttng_ust_shm_handle *handle) |
852c2936 | 271 | { |
1d498196 | 272 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 MD |
273 | |
274 | if (!chan->switch_timer_interval || !buf->switch_timer_enabled) | |
275 | return; | |
276 | ||
a6352fd4 MD |
277 | //TODO |
278 | //del_timer_sync(&buf->switch_timer); | |
852c2936 MD |
279 | buf->switch_timer_enabled = 0; |
280 | } | |
281 | ||
1d498196 | 282 | #if 0 |
852c2936 MD |
283 | /* |
284 | * Polling timer to check the channels for data. | |
285 | */ | |
286 | static void read_buffer_timer(unsigned long data) | |
287 | { | |
4cfec15c | 288 | struct lttng_ust_lib_ring_buffer *buf = (struct lttng_ust_lib_ring_buffer *)data; |
1d498196 | 289 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 290 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
291 | |
292 | CHAN_WARN_ON(chan, !buf->backend.allocated); | |
293 | ||
824f40b8 | 294 | if (uatomic_read(&buf->active_readers) || uatomic_read(&buf->active_shadow_readers)) |
852c2936 | 295 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { |
a6352fd4 MD |
296 | //TODO |
297 | //wake_up_interruptible(&buf->read_wait); | |
298 | //wake_up_interruptible(&chan->read_wait); | |
852c2936 MD |
299 | } |
300 | ||
a6352fd4 MD |
301 | //TODO |
302 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
303 | // mod_timer_pinned(&buf->read_timer, | |
304 | // jiffies + chan->read_timer_interval); | |
305 | //else | |
306 | // mod_timer(&buf->read_timer, | |
307 | // jiffies + chan->read_timer_interval); | |
852c2936 | 308 | } |
1d498196 | 309 | #endif //0 |
852c2936 | 310 | |
4cfec15c | 311 | static void lib_ring_buffer_start_read_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 312 | struct lttng_ust_shm_handle *handle) |
852c2936 | 313 | { |
1d498196 | 314 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 315 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
316 | |
317 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
318 | || !chan->read_timer_interval | |
319 | || buf->read_timer_enabled) | |
320 | return; | |
321 | ||
a6352fd4 MD |
322 | //TODO |
323 | //init_timer(&buf->read_timer); | |
324 | //buf->read_timer.function = read_buffer_timer; | |
325 | //buf->read_timer.expires = jiffies + chan->read_timer_interval; | |
326 | //buf->read_timer.data = (unsigned long)buf; | |
852c2936 | 327 | |
a6352fd4 MD |
328 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
329 | // add_timer_on(&buf->read_timer, buf->backend.cpu); | |
330 | //else | |
331 | // add_timer(&buf->read_timer); | |
852c2936 MD |
332 | buf->read_timer_enabled = 1; |
333 | } | |
334 | ||
4cfec15c | 335 | static void lib_ring_buffer_stop_read_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 336 | struct lttng_ust_shm_handle *handle) |
852c2936 | 337 | { |
1d498196 | 338 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 339 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
340 | |
341 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
342 | || !chan->read_timer_interval | |
343 | || !buf->read_timer_enabled) | |
344 | return; | |
345 | ||
a6352fd4 MD |
346 | //TODO |
347 | //del_timer_sync(&buf->read_timer); | |
852c2936 MD |
348 | /* |
349 | * do one more check to catch data that has been written in the last | |
350 | * timer period. | |
351 | */ | |
1d498196 | 352 | if (lib_ring_buffer_poll_deliver(config, buf, chan, handle)) { |
a6352fd4 MD |
353 | //TODO |
354 | //wake_up_interruptible(&buf->read_wait); | |
355 | //wake_up_interruptible(&chan->read_wait); | |
852c2936 MD |
356 | } |
357 | buf->read_timer_enabled = 0; | |
358 | } | |
359 | ||
1d498196 | 360 | static void channel_unregister_notifiers(struct channel *chan, |
38fae1d3 | 361 | struct lttng_ust_shm_handle *handle) |
852c2936 | 362 | { |
4cfec15c | 363 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
364 | int cpu; |
365 | ||
852c2936 | 366 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
852c2936 | 367 | for_each_possible_cpu(cpu) { |
4cfec15c | 368 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
a6352fd4 | 369 | |
1d498196 MD |
370 | lib_ring_buffer_stop_switch_timer(buf, handle); |
371 | lib_ring_buffer_stop_read_timer(buf, handle); | |
852c2936 | 372 | } |
852c2936 | 373 | } else { |
4cfec15c | 374 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 375 | |
1d498196 MD |
376 | lib_ring_buffer_stop_switch_timer(buf, handle); |
377 | lib_ring_buffer_stop_read_timer(buf, handle); | |
852c2936 | 378 | } |
8d8a24c8 | 379 | //channel_backend_unregister_notifiers(&chan->backend); |
852c2936 MD |
380 | } |
381 | ||
38fae1d3 | 382 | static void channel_free(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 383 | int shadow) |
852c2936 | 384 | { |
824f40b8 MD |
385 | if (!shadow) |
386 | channel_backend_free(&chan->backend, handle); | |
431d5cf0 | 387 | /* chan is freed by shm teardown */ |
1d498196 MD |
388 | shm_object_table_destroy(handle->table); |
389 | free(handle); | |
852c2936 MD |
390 | } |
391 | ||
392 | /** | |
393 | * channel_create - Create channel. | |
394 | * @config: ring buffer instance configuration | |
395 | * @name: name of the channel | |
a3f61e7f MD |
396 | * @priv_data: ring buffer client private data area pointer (output) |
397 | * @priv_data_size: length, in bytes, of the private data area. | |
d028eddb | 398 | * @priv_data_init: initialization data for private data. |
852c2936 MD |
399 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous |
400 | * address mapping. It is used only by RING_BUFFER_STATIC | |
401 | * configuration. It can be set to NULL for other backends. | |
402 | * @subbuf_size: subbuffer size | |
403 | * @num_subbuf: number of subbuffers | |
404 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
405 | * padding to let readers get those sub-buffers. | |
406 | * Used for live streaming. | |
407 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
408 | * | |
409 | * Holds cpu hotplug. | |
410 | * Returns NULL on failure. | |
411 | */ | |
4cfec15c | 412 | struct lttng_ust_shm_handle *channel_create(const struct lttng_ust_lib_ring_buffer_config *config, |
a3f61e7f MD |
413 | const char *name, |
414 | void **priv_data, | |
415 | size_t priv_data_align, | |
416 | size_t priv_data_size, | |
d028eddb | 417 | void *priv_data_init, |
a3f61e7f | 418 | void *buf_addr, size_t subbuf_size, |
852c2936 | 419 | size_t num_subbuf, unsigned int switch_timer_interval, |
193183fb MD |
420 | unsigned int read_timer_interval, |
421 | int *shm_fd, int *wait_fd, uint64_t *memory_map_size) | |
852c2936 | 422 | { |
1d498196 | 423 | int ret, cpu; |
a3f61e7f | 424 | size_t shmsize, chansize; |
852c2936 | 425 | struct channel *chan; |
38fae1d3 | 426 | struct lttng_ust_shm_handle *handle; |
1d498196 | 427 | struct shm_object *shmobj; |
193183fb | 428 | struct shm_ref *ref; |
852c2936 MD |
429 | |
430 | if (lib_ring_buffer_check_config(config, switch_timer_interval, | |
431 | read_timer_interval)) | |
432 | return NULL; | |
433 | ||
38fae1d3 | 434 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
431d5cf0 MD |
435 | if (!handle) |
436 | return NULL; | |
437 | ||
1d498196 MD |
438 | /* Allocate table for channel + per-cpu buffers */ |
439 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
440 | if (!handle->table) | |
441 | goto error_table_alloc; | |
852c2936 | 442 | |
1d498196 MD |
443 | /* Calculate the shm allocation layout */ |
444 | shmsize = sizeof(struct channel); | |
c1fca457 | 445 | shmsize += offset_align(shmsize, __alignof__(struct lttng_ust_lib_ring_buffer_shmp)); |
1d498196 | 446 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
4cfec15c | 447 | shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp) * num_possible_cpus(); |
1d498196 | 448 | else |
4cfec15c | 449 | shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp); |
a3f61e7f MD |
450 | chansize = shmsize; |
451 | shmsize += offset_align(shmsize, priv_data_align); | |
452 | shmsize += priv_data_size; | |
a6352fd4 | 453 | |
1d498196 | 454 | shmobj = shm_object_table_append(handle->table, shmsize); |
b5a14697 MD |
455 | if (!shmobj) |
456 | goto error_append; | |
57773204 | 457 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
a3f61e7f | 458 | set_shmp(handle->chan, zalloc_shm(shmobj, chansize)); |
57773204 MD |
459 | assert(handle->chan._ref.index == 0); |
460 | assert(handle->chan._ref.offset == 0); | |
1d498196 | 461 | chan = shmp(handle, handle->chan); |
a6352fd4 | 462 | if (!chan) |
1d498196 | 463 | goto error_append; |
a6352fd4 | 464 | |
a3f61e7f MD |
465 | /* space for private data */ |
466 | if (priv_data_size) { | |
467 | DECLARE_SHMP(void, priv_data_alloc); | |
468 | ||
469 | align_shm(shmobj, priv_data_align); | |
470 | chan->priv_data_offset = shmobj->allocated_len; | |
471 | set_shmp(priv_data_alloc, zalloc_shm(shmobj, priv_data_size)); | |
472 | if (!shmp(handle, priv_data_alloc)) | |
473 | goto error_append; | |
474 | *priv_data = channel_get_private(chan); | |
d028eddb | 475 | memcpy(*priv_data, priv_data_init, priv_data_size); |
a3f61e7f MD |
476 | } else { |
477 | chan->priv_data_offset = -1; | |
478 | *priv_data = NULL; | |
479 | } | |
480 | ||
481 | ret = channel_backend_init(&chan->backend, name, config, | |
1d498196 | 482 | subbuf_size, num_subbuf, handle); |
852c2936 | 483 | if (ret) |
1d498196 | 484 | goto error_backend_init; |
852c2936 MD |
485 | |
486 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); | |
a6352fd4 MD |
487 | //TODO |
488 | //chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval); | |
489 | //chan->read_timer_interval = usecs_to_jiffies(read_timer_interval); | |
a6352fd4 MD |
490 | //TODO |
491 | //init_waitqueue_head(&chan->read_wait); | |
492 | //init_waitqueue_head(&chan->hp_wait); | |
852c2936 MD |
493 | |
494 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
852c2936 MD |
495 | /* |
496 | * In case of non-hotplug cpu, if the ring-buffer is allocated | |
497 | * in early initcall, it will not be notified of secondary cpus. | |
498 | * In that off case, we need to allocate for all possible cpus. | |
499 | */ | |
852c2936 | 500 | for_each_possible_cpu(cpu) { |
4cfec15c | 501 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
1d498196 MD |
502 | lib_ring_buffer_start_switch_timer(buf, handle); |
503 | lib_ring_buffer_start_read_timer(buf, handle); | |
852c2936 | 504 | } |
852c2936 | 505 | } else { |
4cfec15c | 506 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 507 | |
1d498196 MD |
508 | lib_ring_buffer_start_switch_timer(buf, handle); |
509 | lib_ring_buffer_start_read_timer(buf, handle); | |
852c2936 | 510 | } |
193183fb MD |
511 | ref = &handle->chan._ref; |
512 | shm_get_object_data(handle, ref, shm_fd, wait_fd, memory_map_size); | |
431d5cf0 | 513 | return handle; |
852c2936 | 514 | |
1d498196 MD |
515 | error_backend_init: |
516 | error_append: | |
517 | shm_object_table_destroy(handle->table); | |
518 | error_table_alloc: | |
431d5cf0 | 519 | free(handle); |
852c2936 MD |
520 | return NULL; |
521 | } | |
852c2936 | 522 | |
38fae1d3 | 523 | struct lttng_ust_shm_handle *channel_handle_create(int shm_fd, int wait_fd, |
193183fb MD |
524 | uint64_t memory_map_size) |
525 | { | |
38fae1d3 | 526 | struct lttng_ust_shm_handle *handle; |
193183fb MD |
527 | struct shm_object *object; |
528 | ||
38fae1d3 | 529 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
193183fb MD |
530 | if (!handle) |
531 | return NULL; | |
532 | ||
533 | /* Allocate table for channel + per-cpu buffers */ | |
534 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
535 | if (!handle->table) | |
536 | goto error_table_alloc; | |
537 | /* Add channel object */ | |
538 | object = shm_object_table_append_shadow(handle->table, | |
539 | shm_fd, wait_fd, memory_map_size); | |
540 | if (!object) | |
541 | goto error_table_object; | |
57773204 MD |
542 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
543 | handle->chan._ref.index = 0; | |
544 | handle->chan._ref.offset = 0; | |
193183fb MD |
545 | return handle; |
546 | ||
547 | error_table_object: | |
548 | shm_object_table_destroy(handle->table); | |
549 | error_table_alloc: | |
550 | free(handle); | |
551 | return NULL; | |
552 | } | |
553 | ||
38fae1d3 | 554 | int channel_handle_add_stream(struct lttng_ust_shm_handle *handle, |
193183fb MD |
555 | int shm_fd, int wait_fd, uint64_t memory_map_size) |
556 | { | |
557 | struct shm_object *object; | |
558 | ||
559 | /* Add stream object */ | |
560 | object = shm_object_table_append_shadow(handle->table, | |
561 | shm_fd, wait_fd, memory_map_size); | |
562 | if (!object) | |
563 | return -1; | |
564 | return 0; | |
565 | } | |
566 | ||
852c2936 | 567 | static |
38fae1d3 | 568 | void channel_release(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 569 | int shadow) |
852c2936 | 570 | { |
824f40b8 | 571 | channel_free(chan, handle, shadow); |
852c2936 MD |
572 | } |
573 | ||
574 | /** | |
575 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
576 | * @chan: channel to destroy | |
577 | * | |
578 | * Holds cpu hotplug. | |
431d5cf0 MD |
579 | * Call "destroy" callback, finalize channels, decrement the channel |
580 | * reference count. Note that when readers have completed data | |
581 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
a3f61e7f | 582 | * They should release their handle at that point. |
852c2936 | 583 | */ |
a3f61e7f | 584 | void channel_destroy(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 585 | int shadow) |
852c2936 | 586 | { |
824f40b8 MD |
587 | if (shadow) { |
588 | channel_release(chan, handle, shadow); | |
a3f61e7f | 589 | return; |
824f40b8 MD |
590 | } |
591 | ||
1d498196 | 592 | channel_unregister_notifiers(chan, handle); |
852c2936 | 593 | |
45e9e699 MD |
594 | /* |
595 | * Note: the consumer takes care of finalizing and switching the | |
596 | * buffers. | |
597 | */ | |
852c2936 | 598 | |
431d5cf0 MD |
599 | /* |
600 | * sessiond/consumer are keeping a reference on the shm file | |
601 | * descriptor directly. No need to refcount. | |
602 | */ | |
824f40b8 | 603 | channel_release(chan, handle, shadow); |
a3f61e7f | 604 | return; |
852c2936 | 605 | } |
852c2936 | 606 | |
4cfec15c MD |
607 | struct lttng_ust_lib_ring_buffer *channel_get_ring_buffer( |
608 | const struct lttng_ust_lib_ring_buffer_config *config, | |
1d498196 | 609 | struct channel *chan, int cpu, |
38fae1d3 | 610 | struct lttng_ust_shm_handle *handle, |
381c0f1e MD |
611 | int *shm_fd, int *wait_fd, |
612 | uint64_t *memory_map_size) | |
852c2936 | 613 | { |
381c0f1e MD |
614 | struct shm_ref *ref; |
615 | ||
616 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
617 | ref = &chan->backend.buf[0].shmp._ref; | |
618 | shm_get_object_data(handle, ref, shm_fd, wait_fd, | |
619 | memory_map_size); | |
1d498196 | 620 | return shmp(handle, chan->backend.buf[0].shmp); |
381c0f1e | 621 | } else { |
e095d803 MD |
622 | if (cpu >= num_possible_cpus()) |
623 | return NULL; | |
381c0f1e MD |
624 | ref = &chan->backend.buf[cpu].shmp._ref; |
625 | shm_get_object_data(handle, ref, shm_fd, wait_fd, | |
626 | memory_map_size); | |
1d498196 | 627 | return shmp(handle, chan->backend.buf[cpu].shmp); |
381c0f1e | 628 | } |
852c2936 | 629 | } |
852c2936 | 630 | |
4cfec15c | 631 | int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 632 | struct lttng_ust_shm_handle *handle, |
824f40b8 | 633 | int shadow) |
852c2936 | 634 | { |
824f40b8 MD |
635 | if (shadow) { |
636 | if (uatomic_cmpxchg(&buf->active_shadow_readers, 0, 1) != 0) | |
637 | return -EBUSY; | |
638 | cmm_smp_mb(); | |
639 | return 0; | |
640 | } | |
a6352fd4 | 641 | if (uatomic_cmpxchg(&buf->active_readers, 0, 1) != 0) |
852c2936 | 642 | return -EBUSY; |
a6352fd4 | 643 | cmm_smp_mb(); |
852c2936 MD |
644 | return 0; |
645 | } | |
852c2936 | 646 | |
4cfec15c | 647 | void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 648 | struct lttng_ust_shm_handle *handle, |
824f40b8 | 649 | int shadow) |
852c2936 | 650 | { |
1d498196 | 651 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 | 652 | |
824f40b8 MD |
653 | if (shadow) { |
654 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_shadow_readers) != 1); | |
655 | cmm_smp_mb(); | |
656 | uatomic_dec(&buf->active_shadow_readers); | |
657 | return; | |
658 | } | |
a6352fd4 MD |
659 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
660 | cmm_smp_mb(); | |
661 | uatomic_dec(&buf->active_readers); | |
852c2936 MD |
662 | } |
663 | ||
664 | /** | |
665 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
666 | * @buf: ring buffer | |
667 | * @consumed: consumed count indicating the position where to read | |
668 | * @produced: produced count, indicates position when to stop reading | |
669 | * | |
670 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
671 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 MD |
672 | */ |
673 | ||
4cfec15c | 674 | int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 675 | unsigned long *consumed, unsigned long *produced, |
38fae1d3 | 676 | struct lttng_ust_shm_handle *handle) |
852c2936 | 677 | { |
1d498196 | 678 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 679 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
680 | unsigned long consumed_cur, write_offset; |
681 | int finalized; | |
682 | ||
14641deb | 683 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
684 | /* |
685 | * Read finalized before counters. | |
686 | */ | |
a6352fd4 MD |
687 | cmm_smp_rmb(); |
688 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 MD |
689 | /* |
690 | * No need to issue a memory barrier between consumed count read and | |
691 | * write offset read, because consumed count can only change | |
692 | * concurrently in overwrite mode, and we keep a sequence counter | |
693 | * identifier derived from the write offset to check we are getting | |
694 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
695 | * "tagged" upon writes, tags are checked upon read). | |
696 | */ | |
697 | write_offset = v_read(config, &buf->offset); | |
698 | ||
699 | /* | |
700 | * Check that we are not about to read the same subbuffer in | |
701 | * which the writer head is. | |
702 | */ | |
703 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
704 | == 0) | |
705 | goto nodata; | |
706 | ||
707 | *consumed = consumed_cur; | |
708 | *produced = subbuf_trunc(write_offset, chan); | |
709 | ||
710 | return 0; | |
711 | ||
712 | nodata: | |
713 | /* | |
714 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
715 | * of "raw_spin_is_locked" memory ordering. | |
716 | */ | |
717 | if (finalized) | |
718 | return -ENODATA; | |
852c2936 MD |
719 | else |
720 | return -EAGAIN; | |
721 | } | |
852c2936 MD |
722 | |
723 | /** | |
724 | * lib_ring_buffer_put_snapshot - move consumed counter forward | |
725 | * @buf: ring buffer | |
726 | * @consumed_new: new consumed count value | |
727 | */ | |
4cfec15c | 728 | void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 729 | unsigned long consumed_new, |
38fae1d3 | 730 | struct lttng_ust_shm_handle *handle) |
852c2936 | 731 | { |
4cfec15c | 732 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
1d498196 | 733 | struct channel *chan = shmp(handle, bufb->chan); |
852c2936 MD |
734 | unsigned long consumed; |
735 | ||
824f40b8 MD |
736 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1 |
737 | && uatomic_read(&buf->active_shadow_readers) != 1); | |
852c2936 MD |
738 | |
739 | /* | |
740 | * Only push the consumed value forward. | |
741 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
742 | * the writer in flight recorder mode. | |
743 | */ | |
a6352fd4 | 744 | consumed = uatomic_read(&buf->consumed); |
852c2936 | 745 | while ((long) consumed - (long) consumed_new < 0) |
a6352fd4 MD |
746 | consumed = uatomic_cmpxchg(&buf->consumed, consumed, |
747 | consumed_new); | |
852c2936 | 748 | } |
852c2936 MD |
749 | |
750 | /** | |
751 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
752 | * @buf: ring buffer | |
753 | * @consumed: consumed count indicating the position where to read | |
754 | * | |
755 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
756 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 | 757 | */ |
4cfec15c | 758 | int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 759 | unsigned long consumed, |
38fae1d3 | 760 | struct lttng_ust_shm_handle *handle) |
852c2936 | 761 | { |
1d498196 | 762 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 763 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
764 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
765 | int ret; | |
766 | int finalized; | |
767 | ||
768 | retry: | |
14641deb | 769 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
770 | /* |
771 | * Read finalized before counters. | |
772 | */ | |
a6352fd4 MD |
773 | cmm_smp_rmb(); |
774 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 | 775 | consumed_idx = subbuf_index(consumed, chan); |
4746ae29 | 776 | commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb); |
852c2936 MD |
777 | /* |
778 | * Make sure we read the commit count before reading the buffer | |
779 | * data and the write offset. Correct consumed offset ordering | |
780 | * wrt commit count is insured by the use of cmpxchg to update | |
781 | * the consumed offset. | |
852c2936 | 782 | */ |
a6352fd4 MD |
783 | /* |
784 | * Local rmb to match the remote wmb to read the commit count | |
785 | * before the buffer data and the write offset. | |
786 | */ | |
787 | cmm_smp_rmb(); | |
852c2936 MD |
788 | |
789 | write_offset = v_read(config, &buf->offset); | |
790 | ||
791 | /* | |
792 | * Check that the buffer we are getting is after or at consumed_cur | |
793 | * position. | |
794 | */ | |
795 | if ((long) subbuf_trunc(consumed, chan) | |
796 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
797 | goto nodata; | |
798 | ||
799 | /* | |
800 | * Check that the subbuffer we are trying to consume has been | |
801 | * already fully committed. | |
802 | */ | |
803 | if (((commit_count - chan->backend.subbuf_size) | |
804 | & chan->commit_count_mask) | |
805 | - (buf_trunc(consumed_cur, chan) | |
806 | >> chan->backend.num_subbuf_order) | |
807 | != 0) | |
808 | goto nodata; | |
809 | ||
810 | /* | |
811 | * Check that we are not about to read the same subbuffer in | |
812 | * which the writer head is. | |
813 | */ | |
814 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
815 | == 0) | |
816 | goto nodata; | |
817 | ||
818 | /* | |
819 | * Failure to get the subbuffer causes a busy-loop retry without going | |
820 | * to a wait queue. These are caused by short-lived race windows where | |
821 | * the writer is getting access to a subbuffer we were trying to get | |
822 | * access to. Also checks that the "consumed" buffer count we are | |
823 | * looking for matches the one contained in the subbuffer id. | |
824 | */ | |
825 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
826 | consumed_idx, buf_trunc_val(consumed, chan), |
827 | handle); | |
852c2936 MD |
828 | if (ret) |
829 | goto retry; | |
830 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); | |
831 | ||
832 | buf->get_subbuf_consumed = consumed; | |
833 | buf->get_subbuf = 1; | |
834 | ||
835 | return 0; | |
836 | ||
837 | nodata: | |
838 | /* | |
839 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
840 | * of "raw_spin_is_locked" memory ordering. | |
841 | */ | |
842 | if (finalized) | |
843 | return -ENODATA; | |
852c2936 MD |
844 | else |
845 | return -EAGAIN; | |
846 | } | |
852c2936 MD |
847 | |
848 | /** | |
849 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
850 | * @buf: ring buffer | |
851 | */ | |
4cfec15c | 852 | void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 853 | struct lttng_ust_shm_handle *handle) |
852c2936 | 854 | { |
4cfec15c | 855 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
1d498196 | 856 | struct channel *chan = shmp(handle, bufb->chan); |
4cfec15c | 857 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
858 | unsigned long read_sb_bindex, consumed_idx, consumed; |
859 | ||
824f40b8 MD |
860 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1 |
861 | && uatomic_read(&buf->active_shadow_readers) != 1); | |
852c2936 MD |
862 | |
863 | if (!buf->get_subbuf) { | |
864 | /* | |
865 | * Reader puts a subbuffer it did not get. | |
866 | */ | |
867 | CHAN_WARN_ON(chan, 1); | |
868 | return; | |
869 | } | |
870 | consumed = buf->get_subbuf_consumed; | |
871 | buf->get_subbuf = 0; | |
872 | ||
873 | /* | |
874 | * Clear the records_unread counter. (overruns counter) | |
875 | * Can still be non-zero if a file reader simply grabbed the data | |
876 | * without using iterators. | |
877 | * Can be below zero if an iterator is used on a snapshot more than | |
878 | * once. | |
879 | */ | |
880 | read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); | |
881 | v_add(config, v_read(config, | |
4746ae29 | 882 | &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread), |
852c2936 | 883 | &bufb->records_read); |
4746ae29 | 884 | v_set(config, &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread, 0); |
852c2936 MD |
885 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE |
886 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
887 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
888 | ||
889 | /* | |
890 | * Exchange the reader subbuffer with the one we put in its place in the | |
891 | * writer subbuffer table. Expect the original consumed count. If | |
892 | * update_read_sb_index fails, this is because the writer updated the | |
893 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
894 | * currently have: it has become invalid to try reading this sub-buffer | |
895 | * consumed count value anyway. | |
896 | */ | |
897 | consumed_idx = subbuf_index(consumed, chan); | |
898 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
899 | consumed_idx, buf_trunc_val(consumed, chan), |
900 | handle); | |
852c2936 MD |
901 | /* |
902 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
903 | * if the writer concurrently updated it. | |
904 | */ | |
905 | } | |
852c2936 MD |
906 | |
907 | /* | |
908 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
909 | * position and the writer position. (inclusive) | |
910 | */ | |
911 | static | |
4cfec15c | 912 | void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
913 | struct channel *chan, |
914 | unsigned long cons_offset, | |
1d498196 | 915 | int cpu, |
38fae1d3 | 916 | struct lttng_ust_shm_handle *handle) |
852c2936 | 917 | { |
4cfec15c | 918 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
919 | unsigned long cons_idx, commit_count, commit_count_sb; |
920 | ||
921 | cons_idx = subbuf_index(cons_offset, chan); | |
4746ae29 MD |
922 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, cons_idx)->cc); |
923 | commit_count_sb = v_read(config, &shmp_index(handle, buf->commit_cold, cons_idx)->cc_sb); | |
852c2936 MD |
924 | |
925 | if (subbuf_offset(commit_count, chan) != 0) | |
4d3c9523 | 926 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
927 | "commit count in subbuffer %lu,\n" |
928 | "expecting multiples of %lu bytes\n" | |
929 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
930 | chan->backend.name, cpu, cons_idx, | |
931 | chan->backend.subbuf_size, | |
932 | commit_count, commit_count_sb); | |
933 | ||
4d3c9523 | 934 | DBG("ring buffer: %s, cpu %d: %lu bytes committed\n", |
852c2936 MD |
935 | chan->backend.name, cpu, commit_count); |
936 | } | |
937 | ||
938 | static | |
4cfec15c | 939 | void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 | 940 | struct channel *chan, |
1d498196 | 941 | void *priv, int cpu, |
38fae1d3 | 942 | struct lttng_ust_shm_handle *handle) |
852c2936 | 943 | { |
4cfec15c | 944 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
945 | unsigned long write_offset, cons_offset; |
946 | ||
947 | /* | |
948 | * Can be called in the error path of allocation when | |
949 | * trans_channel_data is not yet set. | |
950 | */ | |
951 | if (!chan) | |
952 | return; | |
953 | /* | |
954 | * No need to order commit_count, write_offset and cons_offset reads | |
955 | * because we execute at teardown when no more writer nor reader | |
956 | * references are left. | |
957 | */ | |
958 | write_offset = v_read(config, &buf->offset); | |
a6352fd4 | 959 | cons_offset = uatomic_read(&buf->consumed); |
852c2936 | 960 | if (write_offset != cons_offset) |
4d3c9523 | 961 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
962 | "non-consumed data\n" |
963 | " [ %lu bytes written, %lu bytes read ]\n", | |
964 | chan->backend.name, cpu, write_offset, cons_offset); | |
965 | ||
a6352fd4 | 966 | for (cons_offset = uatomic_read(&buf->consumed); |
852c2936 MD |
967 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), |
968 | chan) | |
969 | - cons_offset) > 0; | |
970 | cons_offset = subbuf_align(cons_offset, chan)) | |
971 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1d498196 | 972 | cpu, handle); |
852c2936 MD |
973 | } |
974 | ||
975 | static | |
976 | void lib_ring_buffer_print_errors(struct channel *chan, | |
4cfec15c | 977 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
38fae1d3 | 978 | struct lttng_ust_shm_handle *handle) |
852c2936 | 979 | { |
4cfec15c | 980 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
a3f61e7f | 981 | void *priv = channel_get_private(chan); |
852c2936 | 982 | |
4d3c9523 | 983 | DBG("ring buffer %s, cpu %d: %lu records written, " |
852c2936 MD |
984 | "%lu records overrun\n", |
985 | chan->backend.name, cpu, | |
986 | v_read(config, &buf->records_count), | |
987 | v_read(config, &buf->records_overrun)); | |
988 | ||
989 | if (v_read(config, &buf->records_lost_full) | |
990 | || v_read(config, &buf->records_lost_wrap) | |
991 | || v_read(config, &buf->records_lost_big)) | |
4d3c9523 | 992 | DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n" |
852c2936 MD |
993 | " [ %lu buffer full, %lu nest buffer wrap-around, " |
994 | "%lu event too big ]\n", | |
995 | chan->backend.name, cpu, | |
996 | v_read(config, &buf->records_lost_full), | |
997 | v_read(config, &buf->records_lost_wrap), | |
998 | v_read(config, &buf->records_lost_big)); | |
999 | ||
1d498196 | 1000 | lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu, handle); |
852c2936 MD |
1001 | } |
1002 | ||
1003 | /* | |
1004 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
1005 | * | |
1006 | * Only executed when the buffer is finalized, in SWITCH_FLUSH. | |
1007 | */ | |
1008 | static | |
4cfec15c | 1009 | void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1010 | struct channel *chan, |
1011 | struct switch_offsets *offsets, | |
1d498196 | 1012 | u64 tsc, |
38fae1d3 | 1013 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1014 | { |
4cfec15c | 1015 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1016 | unsigned long oldidx = subbuf_index(offsets->old, chan); |
1017 | unsigned long commit_count; | |
1018 | ||
1d498196 | 1019 | config->cb.buffer_begin(buf, tsc, oldidx, handle); |
852c2936 MD |
1020 | |
1021 | /* | |
1022 | * Order all writes to buffer before the commit count update that will | |
1023 | * determine that the subbuffer is full. | |
1024 | */ | |
a6352fd4 | 1025 | cmm_smp_wmb(); |
852c2936 | 1026 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1027 | &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1028 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 MD |
1029 | /* Check if the written buffer has to be delivered */ |
1030 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
1d498196 | 1031 | commit_count, oldidx, handle); |
852c2936 MD |
1032 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
1033 | offsets->old, commit_count, | |
1d498196 MD |
1034 | config->cb.subbuffer_header_size(), |
1035 | handle); | |
852c2936 MD |
1036 | } |
1037 | ||
1038 | /* | |
1039 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
1040 | * | |
1041 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
1042 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
1043 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
1044 | * subbuffer. | |
1045 | */ | |
1046 | static | |
4cfec15c | 1047 | void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1048 | struct channel *chan, |
1049 | struct switch_offsets *offsets, | |
1d498196 | 1050 | u64 tsc, |
38fae1d3 | 1051 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1052 | { |
4cfec15c | 1053 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1054 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); |
1055 | unsigned long commit_count, padding_size, data_size; | |
1056 | ||
1057 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1058 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1059 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size, |
1060 | handle); | |
852c2936 MD |
1061 | |
1062 | /* | |
1063 | * Order all writes to buffer before the commit count update that will | |
1064 | * determine that the subbuffer is full. | |
1065 | */ | |
a6352fd4 | 1066 | cmm_smp_wmb(); |
4746ae29 MD |
1067 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1068 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 | 1069 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, |
1d498196 | 1070 | commit_count, oldidx, handle); |
852c2936 MD |
1071 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
1072 | offsets->old, commit_count, | |
1d498196 | 1073 | padding_size, handle); |
852c2936 MD |
1074 | } |
1075 | ||
1076 | /* | |
1077 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1078 | * | |
1079 | * This code can be executed unordered : writers may already have written to the | |
1080 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1081 | * that this code is executed before the deliver of this sub-buffer. | |
1082 | */ | |
1083 | static | |
4cfec15c | 1084 | void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1085 | struct channel *chan, |
1086 | struct switch_offsets *offsets, | |
1d498196 | 1087 | u64 tsc, |
38fae1d3 | 1088 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1089 | { |
4cfec15c | 1090 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1091 | unsigned long beginidx = subbuf_index(offsets->begin, chan); |
1092 | unsigned long commit_count; | |
1093 | ||
1d498196 | 1094 | config->cb.buffer_begin(buf, tsc, beginidx, handle); |
852c2936 MD |
1095 | |
1096 | /* | |
1097 | * Order all writes to buffer before the commit count update that will | |
1098 | * determine that the subbuffer is full. | |
1099 | */ | |
a6352fd4 | 1100 | cmm_smp_wmb(); |
852c2936 | 1101 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1102 | &shmp_index(handle, buf->commit_hot, beginidx)->cc); |
1103 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, beginidx)->cc); | |
852c2936 MD |
1104 | /* Check if the written buffer has to be delivered */ |
1105 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
1d498196 | 1106 | commit_count, beginidx, handle); |
852c2936 MD |
1107 | lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx, |
1108 | offsets->begin, commit_count, | |
1d498196 MD |
1109 | config->cb.subbuffer_header_size(), |
1110 | handle); | |
852c2936 MD |
1111 | } |
1112 | ||
1113 | /* | |
1114 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1115 | * | |
1116 | * The only remaining threads could be the ones with pending commits. They will | |
1117 | * have to do the deliver themselves. | |
1118 | */ | |
1119 | static | |
4cfec15c | 1120 | void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 MD |
1121 | struct channel *chan, |
1122 | struct switch_offsets *offsets, | |
1123 | u64 tsc, | |
38fae1d3 | 1124 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1125 | { |
4cfec15c | 1126 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1127 | unsigned long endidx = subbuf_index(offsets->end - 1, chan); |
1128 | unsigned long commit_count, padding_size, data_size; | |
1129 | ||
1130 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; | |
1131 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1132 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size, |
1133 | handle); | |
852c2936 MD |
1134 | |
1135 | /* | |
1136 | * Order all writes to buffer before the commit count update that will | |
1137 | * determine that the subbuffer is full. | |
1138 | */ | |
a6352fd4 | 1139 | cmm_smp_wmb(); |
4746ae29 MD |
1140 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, endidx)->cc); |
1141 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc); | |
852c2936 | 1142 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->end - 1, |
1d498196 | 1143 | commit_count, endidx, handle); |
852c2936 MD |
1144 | lib_ring_buffer_write_commit_counter(config, buf, chan, endidx, |
1145 | offsets->end, commit_count, | |
1d498196 | 1146 | padding_size, handle); |
852c2936 MD |
1147 | } |
1148 | ||
1149 | /* | |
1150 | * Returns : | |
1151 | * 0 if ok | |
1152 | * !0 if execution must be aborted. | |
1153 | */ | |
1154 | static | |
1155 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
4cfec15c | 1156 | struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1157 | struct channel *chan, |
1158 | struct switch_offsets *offsets, | |
1159 | u64 *tsc) | |
1160 | { | |
4cfec15c | 1161 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1162 | unsigned long off; |
1163 | ||
1164 | offsets->begin = v_read(config, &buf->offset); | |
1165 | offsets->old = offsets->begin; | |
1166 | offsets->switch_old_start = 0; | |
1167 | off = subbuf_offset(offsets->begin, chan); | |
1168 | ||
1169 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1170 | ||
1171 | /* | |
1172 | * Ensure we flush the header of an empty subbuffer when doing the | |
1173 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1174 | * total data gathering duration even if there were no records saved | |
1175 | * after the last buffer switch. | |
1176 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1177 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1178 | * subbuffer header as appropriate. | |
1179 | * The next record that reserves space will be responsible for | |
1180 | * populating the following subbuffer header. We choose not to populate | |
1181 | * the next subbuffer header here because we want to be able to use | |
a6352fd4 MD |
1182 | * SWITCH_ACTIVE for periodical buffer flush, which must |
1183 | * guarantee that all the buffer content (records and header | |
1184 | * timestamps) are visible to the reader. This is required for | |
1185 | * quiescence guarantees for the fusion merge. | |
852c2936 MD |
1186 | */ |
1187 | if (mode == SWITCH_FLUSH || off > 0) { | |
b5a3dfa5 | 1188 | if (caa_unlikely(off == 0)) { |
852c2936 MD |
1189 | /* |
1190 | * The client does not save any header information. | |
1191 | * Don't switch empty subbuffer on finalize, because it | |
1192 | * is invalid to deliver a completely empty subbuffer. | |
1193 | */ | |
1194 | if (!config->cb.subbuffer_header_size()) | |
1195 | return -1; | |
1196 | /* | |
1197 | * Need to write the subbuffer start header on finalize. | |
1198 | */ | |
1199 | offsets->switch_old_start = 1; | |
1200 | } | |
1201 | offsets->begin = subbuf_align(offsets->begin, chan); | |
1202 | } else | |
1203 | return -1; /* we do not have to switch : buffer is empty */ | |
1204 | /* Note: old points to the next subbuf at offset 0 */ | |
1205 | offsets->end = offsets->begin; | |
1206 | return 0; | |
1207 | } | |
1208 | ||
1209 | /* | |
1210 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
1211 | * called while tracing is active with absolutely no lock held. | |
1212 | * | |
1213 | * Note, however, that as a v_cmpxchg is used for some atomic | |
1214 | * operations, this function must be called from the CPU which owns the buffer | |
1215 | * for a ACTIVE flush. | |
1216 | */ | |
4cfec15c | 1217 | void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode, |
38fae1d3 | 1218 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1219 | { |
1d498196 | 1220 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 1221 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1222 | struct switch_offsets offsets; |
1223 | unsigned long oldidx; | |
1224 | u64 tsc; | |
1225 | ||
1226 | offsets.size = 0; | |
1227 | ||
1228 | /* | |
1229 | * Perform retryable operations. | |
1230 | */ | |
1231 | do { | |
1232 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
1233 | &tsc)) | |
1234 | return; /* Switch not needed */ | |
1235 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
1236 | != offsets.old); | |
1237 | ||
1238 | /* | |
1239 | * Atomically update last_tsc. This update races against concurrent | |
1240 | * atomic updates, but the race will always cause supplementary full TSC | |
1241 | * records, never the opposite (missing a full TSC record when it would | |
1242 | * be needed). | |
1243 | */ | |
1244 | save_last_tsc(config, buf, tsc); | |
1245 | ||
1246 | /* | |
1247 | * Push the reader if necessary | |
1248 | */ | |
1249 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
1250 | ||
1251 | oldidx = subbuf_index(offsets.old, chan); | |
1d498196 | 1252 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx, handle); |
852c2936 MD |
1253 | |
1254 | /* | |
1255 | * May need to populate header start on SWITCH_FLUSH. | |
1256 | */ | |
1257 | if (offsets.switch_old_start) { | |
1d498196 | 1258 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc, handle); |
852c2936 MD |
1259 | offsets.old += config->cb.subbuffer_header_size(); |
1260 | } | |
1261 | ||
1262 | /* | |
1263 | * Switch old subbuffer. | |
1264 | */ | |
1d498196 | 1265 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc, handle); |
852c2936 | 1266 | } |
852c2936 MD |
1267 | |
1268 | /* | |
1269 | * Returns : | |
1270 | * 0 if ok | |
1271 | * -ENOSPC if event size is too large for packet. | |
1272 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
1273 | * -EIO if data cannot be written into the buffer for any other reason. | |
1274 | */ | |
1275 | static | |
4cfec15c | 1276 | int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1277 | struct channel *chan, |
1278 | struct switch_offsets *offsets, | |
4cfec15c | 1279 | struct lttng_ust_lib_ring_buffer_ctx *ctx) |
852c2936 | 1280 | { |
4cfec15c | 1281 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
38fae1d3 | 1282 | struct lttng_ust_shm_handle *handle = ctx->handle; |
852c2936 MD |
1283 | unsigned long reserve_commit_diff; |
1284 | ||
1285 | offsets->begin = v_read(config, &buf->offset); | |
1286 | offsets->old = offsets->begin; | |
1287 | offsets->switch_new_start = 0; | |
1288 | offsets->switch_new_end = 0; | |
1289 | offsets->switch_old_end = 0; | |
1290 | offsets->pre_header_padding = 0; | |
1291 | ||
1292 | ctx->tsc = config->cb.ring_buffer_clock_read(chan); | |
1293 | if ((int64_t) ctx->tsc == -EIO) | |
1294 | return -EIO; | |
1295 | ||
1296 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
1297 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
1298 | ||
b5a3dfa5 | 1299 | if (caa_unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) { |
852c2936 MD |
1300 | offsets->switch_new_start = 1; /* For offsets->begin */ |
1301 | } else { | |
1302 | offsets->size = config->cb.record_header_size(config, chan, | |
1303 | offsets->begin, | |
1304 | &offsets->pre_header_padding, | |
1305 | ctx); | |
1306 | offsets->size += | |
1307 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1308 | ctx->largest_align) | |
1309 | + ctx->data_size; | |
b5a3dfa5 | 1310 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) + |
852c2936 MD |
1311 | offsets->size > chan->backend.subbuf_size)) { |
1312 | offsets->switch_old_end = 1; /* For offsets->old */ | |
1313 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1314 | } | |
1315 | } | |
b5a3dfa5 | 1316 | if (caa_unlikely(offsets->switch_new_start)) { |
852c2936 MD |
1317 | unsigned long sb_index; |
1318 | ||
1319 | /* | |
1320 | * We are typically not filling the previous buffer completely. | |
1321 | */ | |
b5a3dfa5 | 1322 | if (caa_likely(offsets->switch_old_end)) |
852c2936 MD |
1323 | offsets->begin = subbuf_align(offsets->begin, chan); |
1324 | offsets->begin = offsets->begin | |
1325 | + config->cb.subbuffer_header_size(); | |
1326 | /* Test new buffer integrity */ | |
1327 | sb_index = subbuf_index(offsets->begin, chan); | |
1328 | reserve_commit_diff = | |
1329 | (buf_trunc(offsets->begin, chan) | |
1330 | >> chan->backend.num_subbuf_order) | |
1331 | - ((unsigned long) v_read(config, | |
4746ae29 | 1332 | &shmp_index(handle, buf->commit_cold, sb_index)->cc_sb) |
852c2936 | 1333 | & chan->commit_count_mask); |
b5a3dfa5 | 1334 | if (caa_likely(reserve_commit_diff == 0)) { |
852c2936 | 1335 | /* Next subbuffer not being written to. */ |
b5a3dfa5 | 1336 | if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE && |
852c2936 MD |
1337 | subbuf_trunc(offsets->begin, chan) |
1338 | - subbuf_trunc((unsigned long) | |
a6352fd4 | 1339 | uatomic_read(&buf->consumed), chan) |
852c2936 MD |
1340 | >= chan->backend.buf_size)) { |
1341 | /* | |
1342 | * We do not overwrite non consumed buffers | |
1343 | * and we are full : record is lost. | |
1344 | */ | |
1345 | v_inc(config, &buf->records_lost_full); | |
1346 | return -ENOBUFS; | |
1347 | } else { | |
1348 | /* | |
1349 | * Next subbuffer not being written to, and we | |
1350 | * are either in overwrite mode or the buffer is | |
1351 | * not full. It's safe to write in this new | |
1352 | * subbuffer. | |
1353 | */ | |
1354 | } | |
1355 | } else { | |
1356 | /* | |
1357 | * Next subbuffer reserve offset does not match the | |
1358 | * commit offset. Drop record in producer-consumer and | |
1359 | * overwrite mode. Caused by either a writer OOPS or too | |
1360 | * many nested writes over a reserve/commit pair. | |
1361 | */ | |
1362 | v_inc(config, &buf->records_lost_wrap); | |
1363 | return -EIO; | |
1364 | } | |
1365 | offsets->size = | |
1366 | config->cb.record_header_size(config, chan, | |
1367 | offsets->begin, | |
1368 | &offsets->pre_header_padding, | |
1369 | ctx); | |
1370 | offsets->size += | |
1371 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1372 | ctx->largest_align) | |
1373 | + ctx->data_size; | |
b5a3dfa5 | 1374 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) |
852c2936 MD |
1375 | + offsets->size > chan->backend.subbuf_size)) { |
1376 | /* | |
1377 | * Record too big for subbuffers, report error, don't | |
1378 | * complete the sub-buffer switch. | |
1379 | */ | |
1380 | v_inc(config, &buf->records_lost_big); | |
1381 | return -ENOSPC; | |
1382 | } else { | |
1383 | /* | |
1384 | * We just made a successful buffer switch and the | |
1385 | * record fits in the new subbuffer. Let's write. | |
1386 | */ | |
1387 | } | |
1388 | } else { | |
1389 | /* | |
1390 | * Record fits in the current buffer and we are not on a switch | |
1391 | * boundary. It's safe to write. | |
1392 | */ | |
1393 | } | |
1394 | offsets->end = offsets->begin + offsets->size; | |
1395 | ||
b5a3dfa5 | 1396 | if (caa_unlikely(subbuf_offset(offsets->end, chan) == 0)) { |
852c2936 MD |
1397 | /* |
1398 | * The offset_end will fall at the very beginning of the next | |
1399 | * subbuffer. | |
1400 | */ | |
1401 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
1402 | } | |
1403 | return 0; | |
1404 | } | |
1405 | ||
1406 | /** | |
1407 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
1408 | * @ctx: ring buffer context. | |
1409 | * | |
1410 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, | |
1411 | * -EIO for other errors, else returns 0. | |
1412 | * It will take care of sub-buffer switching. | |
1413 | */ | |
4cfec15c | 1414 | int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx) |
852c2936 MD |
1415 | { |
1416 | struct channel *chan = ctx->chan; | |
38fae1d3 | 1417 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c MD |
1418 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
1419 | struct lttng_ust_lib_ring_buffer *buf; | |
852c2936 MD |
1420 | struct switch_offsets offsets; |
1421 | int ret; | |
1422 | ||
1423 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1d498196 | 1424 | buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp); |
852c2936 | 1425 | else |
1d498196 | 1426 | buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 MD |
1427 | ctx->buf = buf; |
1428 | ||
1429 | offsets.size = 0; | |
1430 | ||
1431 | do { | |
1432 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, | |
1433 | ctx); | |
b5a3dfa5 | 1434 | if (caa_unlikely(ret)) |
852c2936 | 1435 | return ret; |
b5a3dfa5 | 1436 | } while (caa_unlikely(v_cmpxchg(config, &buf->offset, offsets.old, |
852c2936 MD |
1437 | offsets.end) |
1438 | != offsets.old)); | |
1439 | ||
1440 | /* | |
1441 | * Atomically update last_tsc. This update races against concurrent | |
1442 | * atomic updates, but the race will always cause supplementary full TSC | |
1443 | * records, never the opposite (missing a full TSC record when it would | |
1444 | * be needed). | |
1445 | */ | |
1446 | save_last_tsc(config, buf, ctx->tsc); | |
1447 | ||
1448 | /* | |
1449 | * Push the reader if necessary | |
1450 | */ | |
1451 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
1452 | ||
1453 | /* | |
1454 | * Clear noref flag for this subbuffer. | |
1455 | */ | |
1456 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1d498196 MD |
1457 | subbuf_index(offsets.end - 1, chan), |
1458 | handle); | |
852c2936 MD |
1459 | |
1460 | /* | |
1461 | * Switch old subbuffer if needed. | |
1462 | */ | |
b5a3dfa5 | 1463 | if (caa_unlikely(offsets.switch_old_end)) { |
852c2936 | 1464 | lib_ring_buffer_clear_noref(config, &buf->backend, |
1d498196 MD |
1465 | subbuf_index(offsets.old - 1, chan), |
1466 | handle); | |
1467 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc, handle); | |
852c2936 MD |
1468 | } |
1469 | ||
1470 | /* | |
1471 | * Populate new subbuffer. | |
1472 | */ | |
b5a3dfa5 | 1473 | if (caa_unlikely(offsets.switch_new_start)) |
1d498196 | 1474 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 | 1475 | |
b5a3dfa5 | 1476 | if (caa_unlikely(offsets.switch_new_end)) |
1d498196 | 1477 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 MD |
1478 | |
1479 | ctx->slot_size = offsets.size; | |
1480 | ctx->pre_offset = offsets.begin; | |
1481 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
1482 | return 0; | |
1483 | } |