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 | ||
5ad63a16 | 41 | #define _GNU_SOURCE |
a6352fd4 | 42 | #include <sys/types.h> |
431d5cf0 MD |
43 | #include <sys/mman.h> |
44 | #include <sys/stat.h> | |
45 | #include <fcntl.h> | |
14641deb | 46 | #include <urcu/compiler.h> |
a6352fd4 | 47 | #include <urcu/ref.h> |
35897f8b | 48 | #include <helper.h> |
14641deb | 49 | |
a6352fd4 | 50 | #include "smp.h" |
4318ae1b | 51 | #include <lttng/ringbuffer-config.h> |
2fed87ae | 52 | #include "vatomic.h" |
4931a13e MD |
53 | #include "backend.h" |
54 | #include "frontend.h" | |
a6352fd4 | 55 | #include "shm.h" |
852c2936 | 56 | |
431d5cf0 MD |
57 | #ifndef max |
58 | #define max(a, b) ((a) > (b) ? (a) : (b)) | |
59 | #endif | |
60 | ||
64493e4f MD |
61 | /* Print DBG() messages about events lost only every 1048576 hits */ |
62 | #define DBG_PRINT_NR_LOST (1UL << 20) | |
63 | ||
2432c3c9 MD |
64 | /* |
65 | * Use POSIX SHM: shm_open(3) and shm_unlink(3). | |
66 | * close(2) to close the fd returned by shm_open. | |
67 | * shm_unlink releases the shared memory object name. | |
68 | * ftruncate(2) sets the size of the memory object. | |
69 | * mmap/munmap maps the shared memory obj to a virtual address in the | |
70 | * calling proceess (should be done both in libust and consumer). | |
71 | * See shm_overview(7) for details. | |
72 | * Pass file descriptor returned by shm_open(3) to ltt-sessiond through | |
73 | * a UNIX socket. | |
74 | * | |
75 | * Since we don't need to access the object using its name, we can | |
76 | * immediately shm_unlink(3) it, and only keep the handle with its file | |
77 | * descriptor. | |
78 | */ | |
79 | ||
852c2936 MD |
80 | /* |
81 | * Internal structure representing offsets to use at a sub-buffer switch. | |
82 | */ | |
83 | struct switch_offsets { | |
84 | unsigned long begin, end, old; | |
85 | size_t pre_header_padding, size; | |
86 | unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1, | |
87 | switch_old_end:1; | |
88 | }; | |
89 | ||
a6352fd4 | 90 | __thread unsigned int lib_ring_buffer_nesting; |
852c2936 | 91 | |
45e9e699 MD |
92 | /* |
93 | * TODO: this is unused. Errors are saved within the ring buffer. | |
94 | * Eventually, allow consumerd to print these errors. | |
95 | */ | |
852c2936 MD |
96 | static |
97 | void lib_ring_buffer_print_errors(struct channel *chan, | |
4cfec15c | 98 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
b68d3dc0 MD |
99 | struct lttng_ust_shm_handle *handle) |
100 | __attribute__((unused)); | |
852c2936 | 101 | |
852c2936 MD |
102 | /** |
103 | * lib_ring_buffer_reset - Reset ring buffer to initial values. | |
104 | * @buf: Ring buffer. | |
105 | * | |
106 | * Effectively empty the ring buffer. Should be called when the buffer is not | |
107 | * used for writing. The ring buffer can be opened for reading, but the reader | |
108 | * should not be using the iterator concurrently with reset. The previous | |
109 | * current iterator record is reset. | |
110 | */ | |
4cfec15c | 111 | void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 112 | struct lttng_ust_shm_handle *handle) |
852c2936 | 113 | { |
1d498196 | 114 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 115 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
116 | unsigned int i; |
117 | ||
118 | /* | |
119 | * Reset iterator first. It will put the subbuffer if it currently holds | |
120 | * it. | |
121 | */ | |
852c2936 MD |
122 | v_set(config, &buf->offset, 0); |
123 | for (i = 0; i < chan->backend.num_subbuf; i++) { | |
4746ae29 MD |
124 | v_set(config, &shmp_index(handle, buf->commit_hot, i)->cc, 0); |
125 | v_set(config, &shmp_index(handle, buf->commit_hot, i)->seq, 0); | |
126 | v_set(config, &shmp_index(handle, buf->commit_cold, i)->cc_sb, 0); | |
852c2936 | 127 | } |
a6352fd4 MD |
128 | uatomic_set(&buf->consumed, 0); |
129 | uatomic_set(&buf->record_disabled, 0); | |
852c2936 | 130 | v_set(config, &buf->last_tsc, 0); |
1d498196 | 131 | lib_ring_buffer_backend_reset(&buf->backend, handle); |
852c2936 MD |
132 | /* Don't reset number of active readers */ |
133 | v_set(config, &buf->records_lost_full, 0); | |
134 | v_set(config, &buf->records_lost_wrap, 0); | |
135 | v_set(config, &buf->records_lost_big, 0); | |
136 | v_set(config, &buf->records_count, 0); | |
137 | v_set(config, &buf->records_overrun, 0); | |
138 | buf->finalized = 0; | |
139 | } | |
852c2936 MD |
140 | |
141 | /** | |
142 | * channel_reset - Reset channel to initial values. | |
143 | * @chan: Channel. | |
144 | * | |
145 | * Effectively empty the channel. Should be called when the channel is not used | |
146 | * for writing. The channel can be opened for reading, but the reader should not | |
147 | * be using the iterator concurrently with reset. The previous current iterator | |
148 | * record is reset. | |
149 | */ | |
150 | void channel_reset(struct channel *chan) | |
151 | { | |
152 | /* | |
153 | * Reset iterators first. Will put the subbuffer if held for reading. | |
154 | */ | |
a6352fd4 | 155 | uatomic_set(&chan->record_disabled, 0); |
852c2936 MD |
156 | /* Don't reset commit_count_mask, still valid */ |
157 | channel_backend_reset(&chan->backend); | |
158 | /* Don't reset switch/read timer interval */ | |
159 | /* Don't reset notifiers and notifier enable bits */ | |
160 | /* Don't reset reader reference count */ | |
161 | } | |
852c2936 MD |
162 | |
163 | /* | |
164 | * Must be called under cpu hotplug protection. | |
165 | */ | |
4cfec15c | 166 | int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf, |
a6352fd4 | 167 | struct channel_backend *chanb, int cpu, |
38fae1d3 | 168 | struct lttng_ust_shm_handle *handle, |
1d498196 | 169 | struct shm_object *shmobj) |
852c2936 | 170 | { |
4cfec15c | 171 | const struct lttng_ust_lib_ring_buffer_config *config = &chanb->config; |
14641deb | 172 | struct channel *chan = caa_container_of(chanb, struct channel, backend); |
a3f61e7f | 173 | void *priv = channel_get_private(chan); |
852c2936 | 174 | size_t subbuf_header_size; |
2fed87ae | 175 | uint64_t tsc; |
852c2936 MD |
176 | int ret; |
177 | ||
178 | /* Test for cpu hotplug */ | |
179 | if (buf->backend.allocated) | |
180 | return 0; | |
181 | ||
a6352fd4 | 182 | ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, |
1d498196 | 183 | cpu, handle, shmobj); |
852c2936 MD |
184 | if (ret) |
185 | return ret; | |
186 | ||
1d498196 MD |
187 | align_shm(shmobj, __alignof__(struct commit_counters_hot)); |
188 | set_shmp(buf->commit_hot, | |
189 | zalloc_shm(shmobj, | |
190 | sizeof(struct commit_counters_hot) * chan->backend.num_subbuf)); | |
191 | if (!shmp(handle, buf->commit_hot)) { | |
852c2936 MD |
192 | ret = -ENOMEM; |
193 | goto free_chanbuf; | |
194 | } | |
195 | ||
1d498196 MD |
196 | align_shm(shmobj, __alignof__(struct commit_counters_cold)); |
197 | set_shmp(buf->commit_cold, | |
198 | zalloc_shm(shmobj, | |
199 | sizeof(struct commit_counters_cold) * chan->backend.num_subbuf)); | |
200 | if (!shmp(handle, buf->commit_cold)) { | |
852c2936 MD |
201 | ret = -ENOMEM; |
202 | goto free_commit; | |
203 | } | |
204 | ||
852c2936 MD |
205 | /* |
206 | * Write the subbuffer header for first subbuffer so we know the total | |
207 | * duration of data gathering. | |
208 | */ | |
209 | subbuf_header_size = config->cb.subbuffer_header_size(); | |
210 | v_set(config, &buf->offset, subbuf_header_size); | |
4746ae29 | 211 | subbuffer_id_clear_noref(config, &shmp_index(handle, buf->backend.buf_wsb, 0)->id); |
1d498196 MD |
212 | tsc = config->cb.ring_buffer_clock_read(shmp(handle, buf->backend.chan)); |
213 | config->cb.buffer_begin(buf, tsc, 0, handle); | |
4746ae29 | 214 | v_add(config, subbuf_header_size, &shmp_index(handle, buf->commit_hot, 0)->cc); |
852c2936 MD |
215 | |
216 | if (config->cb.buffer_create) { | |
1d498196 | 217 | ret = config->cb.buffer_create(buf, priv, cpu, chanb->name, handle); |
852c2936 MD |
218 | if (ret) |
219 | goto free_init; | |
220 | } | |
852c2936 | 221 | buf->backend.allocated = 1; |
852c2936 MD |
222 | return 0; |
223 | ||
224 | /* Error handling */ | |
225 | free_init: | |
a6352fd4 | 226 | /* commit_cold will be freed by shm teardown */ |
852c2936 | 227 | free_commit: |
a6352fd4 | 228 | /* commit_hot will be freed by shm teardown */ |
852c2936 | 229 | free_chanbuf: |
852c2936 MD |
230 | return ret; |
231 | } | |
232 | ||
1d498196 | 233 | #if 0 |
852c2936 MD |
234 | static void switch_buffer_timer(unsigned long data) |
235 | { | |
4cfec15c | 236 | struct lttng_ust_lib_ring_buffer *buf = (struct lttng_ust_lib_ring_buffer *)data; |
1d498196 | 237 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 238 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
239 | |
240 | /* | |
241 | * Only flush buffers periodically if readers are active. | |
242 | */ | |
824f40b8 | 243 | if (uatomic_read(&buf->active_readers) || uatomic_read(&buf->active_shadow_readers)) |
1d498196 | 244 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, handle); |
852c2936 | 245 | |
a6352fd4 MD |
246 | //TODO timers |
247 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
248 | // mod_timer_pinned(&buf->switch_timer, | |
249 | // jiffies + chan->switch_timer_interval); | |
250 | //else | |
251 | // mod_timer(&buf->switch_timer, | |
252 | // jiffies + chan->switch_timer_interval); | |
852c2936 | 253 | } |
1d498196 | 254 | #endif //0 |
852c2936 | 255 | |
4cfec15c | 256 | static void lib_ring_buffer_start_switch_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 257 | struct lttng_ust_shm_handle *handle) |
852c2936 | 258 | { |
1d498196 | 259 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 260 | //const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
261 | |
262 | if (!chan->switch_timer_interval || buf->switch_timer_enabled) | |
263 | return; | |
a6352fd4 MD |
264 | //TODO |
265 | //init_timer(&buf->switch_timer); | |
266 | //buf->switch_timer.function = switch_buffer_timer; | |
267 | //buf->switch_timer.expires = jiffies + chan->switch_timer_interval; | |
268 | //buf->switch_timer.data = (unsigned long)buf; | |
269 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
270 | // add_timer_on(&buf->switch_timer, buf->backend.cpu); | |
271 | //else | |
272 | // add_timer(&buf->switch_timer); | |
852c2936 MD |
273 | buf->switch_timer_enabled = 1; |
274 | } | |
275 | ||
4cfec15c | 276 | static void lib_ring_buffer_stop_switch_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 277 | struct lttng_ust_shm_handle *handle) |
852c2936 | 278 | { |
1d498196 | 279 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 MD |
280 | |
281 | if (!chan->switch_timer_interval || !buf->switch_timer_enabled) | |
282 | return; | |
283 | ||
a6352fd4 MD |
284 | //TODO |
285 | //del_timer_sync(&buf->switch_timer); | |
852c2936 MD |
286 | buf->switch_timer_enabled = 0; |
287 | } | |
288 | ||
1d498196 | 289 | #if 0 |
852c2936 MD |
290 | /* |
291 | * Polling timer to check the channels for data. | |
292 | */ | |
293 | static void read_buffer_timer(unsigned long data) | |
294 | { | |
4cfec15c | 295 | struct lttng_ust_lib_ring_buffer *buf = (struct lttng_ust_lib_ring_buffer *)data; |
1d498196 | 296 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 297 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
298 | |
299 | CHAN_WARN_ON(chan, !buf->backend.allocated); | |
300 | ||
824f40b8 | 301 | if (uatomic_read(&buf->active_readers) || uatomic_read(&buf->active_shadow_readers)) |
852c2936 | 302 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { |
a6352fd4 MD |
303 | //TODO |
304 | //wake_up_interruptible(&buf->read_wait); | |
305 | //wake_up_interruptible(&chan->read_wait); | |
852c2936 MD |
306 | } |
307 | ||
a6352fd4 MD |
308 | //TODO |
309 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
310 | // mod_timer_pinned(&buf->read_timer, | |
311 | // jiffies + chan->read_timer_interval); | |
312 | //else | |
313 | // mod_timer(&buf->read_timer, | |
314 | // jiffies + chan->read_timer_interval); | |
852c2936 | 315 | } |
1d498196 | 316 | #endif //0 |
852c2936 | 317 | |
4cfec15c | 318 | static void lib_ring_buffer_start_read_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 319 | struct lttng_ust_shm_handle *handle) |
852c2936 | 320 | { |
1d498196 | 321 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 322 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
323 | |
324 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
325 | || !chan->read_timer_interval | |
326 | || buf->read_timer_enabled) | |
327 | return; | |
328 | ||
a6352fd4 MD |
329 | //TODO |
330 | //init_timer(&buf->read_timer); | |
331 | //buf->read_timer.function = read_buffer_timer; | |
332 | //buf->read_timer.expires = jiffies + chan->read_timer_interval; | |
333 | //buf->read_timer.data = (unsigned long)buf; | |
852c2936 | 334 | |
a6352fd4 MD |
335 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
336 | // add_timer_on(&buf->read_timer, buf->backend.cpu); | |
337 | //else | |
338 | // add_timer(&buf->read_timer); | |
852c2936 MD |
339 | buf->read_timer_enabled = 1; |
340 | } | |
341 | ||
4cfec15c | 342 | static void lib_ring_buffer_stop_read_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 343 | struct lttng_ust_shm_handle *handle) |
852c2936 | 344 | { |
1d498196 | 345 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 346 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
347 | |
348 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
349 | || !chan->read_timer_interval | |
350 | || !buf->read_timer_enabled) | |
351 | return; | |
352 | ||
a6352fd4 MD |
353 | //TODO |
354 | //del_timer_sync(&buf->read_timer); | |
852c2936 MD |
355 | /* |
356 | * do one more check to catch data that has been written in the last | |
357 | * timer period. | |
358 | */ | |
1d498196 | 359 | if (lib_ring_buffer_poll_deliver(config, buf, chan, handle)) { |
a6352fd4 MD |
360 | //TODO |
361 | //wake_up_interruptible(&buf->read_wait); | |
362 | //wake_up_interruptible(&chan->read_wait); | |
852c2936 MD |
363 | } |
364 | buf->read_timer_enabled = 0; | |
365 | } | |
366 | ||
1d498196 | 367 | static void channel_unregister_notifiers(struct channel *chan, |
38fae1d3 | 368 | struct lttng_ust_shm_handle *handle) |
852c2936 | 369 | { |
4cfec15c | 370 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
371 | int cpu; |
372 | ||
852c2936 | 373 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
852c2936 | 374 | for_each_possible_cpu(cpu) { |
4cfec15c | 375 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
a6352fd4 | 376 | |
1d498196 MD |
377 | lib_ring_buffer_stop_switch_timer(buf, handle); |
378 | lib_ring_buffer_stop_read_timer(buf, handle); | |
852c2936 | 379 | } |
852c2936 | 380 | } else { |
4cfec15c | 381 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 382 | |
1d498196 MD |
383 | lib_ring_buffer_stop_switch_timer(buf, handle); |
384 | lib_ring_buffer_stop_read_timer(buf, handle); | |
852c2936 | 385 | } |
8d8a24c8 | 386 | //channel_backend_unregister_notifiers(&chan->backend); |
852c2936 MD |
387 | } |
388 | ||
38fae1d3 | 389 | static void channel_free(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 390 | int shadow) |
852c2936 | 391 | { |
824f40b8 MD |
392 | if (!shadow) |
393 | channel_backend_free(&chan->backend, handle); | |
431d5cf0 | 394 | /* chan is freed by shm teardown */ |
1d498196 MD |
395 | shm_object_table_destroy(handle->table); |
396 | free(handle); | |
852c2936 MD |
397 | } |
398 | ||
399 | /** | |
400 | * channel_create - Create channel. | |
401 | * @config: ring buffer instance configuration | |
402 | * @name: name of the channel | |
a3f61e7f MD |
403 | * @priv_data: ring buffer client private data area pointer (output) |
404 | * @priv_data_size: length, in bytes, of the private data area. | |
d028eddb | 405 | * @priv_data_init: initialization data for private data. |
852c2936 MD |
406 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous |
407 | * address mapping. It is used only by RING_BUFFER_STATIC | |
408 | * configuration. It can be set to NULL for other backends. | |
409 | * @subbuf_size: subbuffer size | |
410 | * @num_subbuf: number of subbuffers | |
411 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
412 | * padding to let readers get those sub-buffers. | |
413 | * Used for live streaming. | |
414 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
415 | * | |
416 | * Holds cpu hotplug. | |
417 | * Returns NULL on failure. | |
418 | */ | |
4cfec15c | 419 | struct lttng_ust_shm_handle *channel_create(const struct lttng_ust_lib_ring_buffer_config *config, |
a3f61e7f MD |
420 | const char *name, |
421 | void **priv_data, | |
422 | size_t priv_data_align, | |
423 | size_t priv_data_size, | |
d028eddb | 424 | void *priv_data_init, |
a3f61e7f | 425 | void *buf_addr, size_t subbuf_size, |
852c2936 | 426 | size_t num_subbuf, unsigned int switch_timer_interval, |
193183fb | 427 | unsigned int read_timer_interval, |
ef9ff354 | 428 | int **shm_fd, int **wait_fd, uint64_t **memory_map_size) |
852c2936 | 429 | { |
1d498196 | 430 | int ret, cpu; |
a3f61e7f | 431 | size_t shmsize, chansize; |
852c2936 | 432 | struct channel *chan; |
38fae1d3 | 433 | struct lttng_ust_shm_handle *handle; |
1d498196 | 434 | struct shm_object *shmobj; |
193183fb | 435 | struct shm_ref *ref; |
852c2936 MD |
436 | |
437 | if (lib_ring_buffer_check_config(config, switch_timer_interval, | |
438 | read_timer_interval)) | |
439 | return NULL; | |
440 | ||
38fae1d3 | 441 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
431d5cf0 MD |
442 | if (!handle) |
443 | return NULL; | |
444 | ||
1d498196 MD |
445 | /* Allocate table for channel + per-cpu buffers */ |
446 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
447 | if (!handle->table) | |
448 | goto error_table_alloc; | |
852c2936 | 449 | |
1d498196 MD |
450 | /* Calculate the shm allocation layout */ |
451 | shmsize = sizeof(struct channel); | |
c1fca457 | 452 | shmsize += offset_align(shmsize, __alignof__(struct lttng_ust_lib_ring_buffer_shmp)); |
1d498196 | 453 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
4cfec15c | 454 | shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp) * num_possible_cpus(); |
1d498196 | 455 | else |
4cfec15c | 456 | shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp); |
a3f61e7f MD |
457 | chansize = shmsize; |
458 | shmsize += offset_align(shmsize, priv_data_align); | |
459 | shmsize += priv_data_size; | |
a6352fd4 | 460 | |
1d498196 | 461 | shmobj = shm_object_table_append(handle->table, shmsize); |
b5a14697 MD |
462 | if (!shmobj) |
463 | goto error_append; | |
57773204 | 464 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
a3f61e7f | 465 | set_shmp(handle->chan, zalloc_shm(shmobj, chansize)); |
57773204 MD |
466 | assert(handle->chan._ref.index == 0); |
467 | assert(handle->chan._ref.offset == 0); | |
1d498196 | 468 | chan = shmp(handle, handle->chan); |
a6352fd4 | 469 | if (!chan) |
1d498196 | 470 | goto error_append; |
a6352fd4 | 471 | |
a3f61e7f MD |
472 | /* space for private data */ |
473 | if (priv_data_size) { | |
474 | DECLARE_SHMP(void, priv_data_alloc); | |
475 | ||
476 | align_shm(shmobj, priv_data_align); | |
477 | chan->priv_data_offset = shmobj->allocated_len; | |
478 | set_shmp(priv_data_alloc, zalloc_shm(shmobj, priv_data_size)); | |
479 | if (!shmp(handle, priv_data_alloc)) | |
480 | goto error_append; | |
481 | *priv_data = channel_get_private(chan); | |
d028eddb | 482 | memcpy(*priv_data, priv_data_init, priv_data_size); |
a3f61e7f MD |
483 | } else { |
484 | chan->priv_data_offset = -1; | |
485 | *priv_data = NULL; | |
486 | } | |
487 | ||
488 | ret = channel_backend_init(&chan->backend, name, config, | |
1d498196 | 489 | subbuf_size, num_subbuf, handle); |
852c2936 | 490 | if (ret) |
1d498196 | 491 | goto error_backend_init; |
852c2936 MD |
492 | |
493 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); | |
a6352fd4 MD |
494 | //TODO |
495 | //chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval); | |
496 | //chan->read_timer_interval = usecs_to_jiffies(read_timer_interval); | |
a6352fd4 MD |
497 | //TODO |
498 | //init_waitqueue_head(&chan->read_wait); | |
499 | //init_waitqueue_head(&chan->hp_wait); | |
852c2936 MD |
500 | |
501 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
852c2936 MD |
502 | /* |
503 | * In case of non-hotplug cpu, if the ring-buffer is allocated | |
504 | * in early initcall, it will not be notified of secondary cpus. | |
505 | * In that off case, we need to allocate for all possible cpus. | |
506 | */ | |
852c2936 | 507 | for_each_possible_cpu(cpu) { |
4cfec15c | 508 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
1d498196 MD |
509 | lib_ring_buffer_start_switch_timer(buf, handle); |
510 | lib_ring_buffer_start_read_timer(buf, handle); | |
852c2936 | 511 | } |
852c2936 | 512 | } else { |
4cfec15c | 513 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 514 | |
1d498196 MD |
515 | lib_ring_buffer_start_switch_timer(buf, handle); |
516 | lib_ring_buffer_start_read_timer(buf, handle); | |
852c2936 | 517 | } |
193183fb MD |
518 | ref = &handle->chan._ref; |
519 | shm_get_object_data(handle, ref, shm_fd, wait_fd, memory_map_size); | |
431d5cf0 | 520 | return handle; |
852c2936 | 521 | |
1d498196 MD |
522 | error_backend_init: |
523 | error_append: | |
524 | shm_object_table_destroy(handle->table); | |
525 | error_table_alloc: | |
431d5cf0 | 526 | free(handle); |
852c2936 MD |
527 | return NULL; |
528 | } | |
852c2936 | 529 | |
38fae1d3 | 530 | struct lttng_ust_shm_handle *channel_handle_create(int shm_fd, int wait_fd, |
193183fb MD |
531 | uint64_t memory_map_size) |
532 | { | |
38fae1d3 | 533 | struct lttng_ust_shm_handle *handle; |
193183fb MD |
534 | struct shm_object *object; |
535 | ||
38fae1d3 | 536 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
193183fb MD |
537 | if (!handle) |
538 | return NULL; | |
539 | ||
540 | /* Allocate table for channel + per-cpu buffers */ | |
541 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
542 | if (!handle->table) | |
543 | goto error_table_alloc; | |
544 | /* Add channel object */ | |
545 | object = shm_object_table_append_shadow(handle->table, | |
546 | shm_fd, wait_fd, memory_map_size); | |
547 | if (!object) | |
548 | goto error_table_object; | |
57773204 MD |
549 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
550 | handle->chan._ref.index = 0; | |
551 | handle->chan._ref.offset = 0; | |
193183fb MD |
552 | return handle; |
553 | ||
554 | error_table_object: | |
555 | shm_object_table_destroy(handle->table); | |
556 | error_table_alloc: | |
557 | free(handle); | |
558 | return NULL; | |
559 | } | |
560 | ||
38fae1d3 | 561 | int channel_handle_add_stream(struct lttng_ust_shm_handle *handle, |
193183fb MD |
562 | int shm_fd, int wait_fd, uint64_t memory_map_size) |
563 | { | |
564 | struct shm_object *object; | |
565 | ||
566 | /* Add stream object */ | |
567 | object = shm_object_table_append_shadow(handle->table, | |
568 | shm_fd, wait_fd, memory_map_size); | |
569 | if (!object) | |
570 | return -1; | |
571 | return 0; | |
572 | } | |
573 | ||
852c2936 | 574 | static |
38fae1d3 | 575 | void channel_release(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 576 | int shadow) |
852c2936 | 577 | { |
824f40b8 | 578 | channel_free(chan, handle, shadow); |
852c2936 MD |
579 | } |
580 | ||
581 | /** | |
582 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
583 | * @chan: channel to destroy | |
584 | * | |
585 | * Holds cpu hotplug. | |
431d5cf0 MD |
586 | * Call "destroy" callback, finalize channels, decrement the channel |
587 | * reference count. Note that when readers have completed data | |
588 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
a3f61e7f | 589 | * They should release their handle at that point. |
852c2936 | 590 | */ |
a3f61e7f | 591 | void channel_destroy(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 592 | int shadow) |
852c2936 | 593 | { |
824f40b8 MD |
594 | if (shadow) { |
595 | channel_release(chan, handle, shadow); | |
a3f61e7f | 596 | return; |
824f40b8 MD |
597 | } |
598 | ||
1d498196 | 599 | channel_unregister_notifiers(chan, handle); |
852c2936 | 600 | |
45e9e699 MD |
601 | /* |
602 | * Note: the consumer takes care of finalizing and switching the | |
603 | * buffers. | |
604 | */ | |
852c2936 | 605 | |
431d5cf0 MD |
606 | /* |
607 | * sessiond/consumer are keeping a reference on the shm file | |
608 | * descriptor directly. No need to refcount. | |
609 | */ | |
824f40b8 | 610 | channel_release(chan, handle, shadow); |
a3f61e7f | 611 | return; |
852c2936 | 612 | } |
852c2936 | 613 | |
4cfec15c MD |
614 | struct lttng_ust_lib_ring_buffer *channel_get_ring_buffer( |
615 | const struct lttng_ust_lib_ring_buffer_config *config, | |
1d498196 | 616 | struct channel *chan, int cpu, |
38fae1d3 | 617 | struct lttng_ust_shm_handle *handle, |
ef9ff354 MD |
618 | int **shm_fd, int **wait_fd, |
619 | uint64_t **memory_map_size) | |
852c2936 | 620 | { |
381c0f1e MD |
621 | struct shm_ref *ref; |
622 | ||
623 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
624 | ref = &chan->backend.buf[0].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[0].shmp); |
381c0f1e | 628 | } else { |
e095d803 MD |
629 | if (cpu >= num_possible_cpus()) |
630 | return NULL; | |
381c0f1e MD |
631 | ref = &chan->backend.buf[cpu].shmp._ref; |
632 | shm_get_object_data(handle, ref, shm_fd, wait_fd, | |
633 | memory_map_size); | |
1d498196 | 634 | return shmp(handle, chan->backend.buf[cpu].shmp); |
381c0f1e | 635 | } |
852c2936 | 636 | } |
852c2936 | 637 | |
4cfec15c | 638 | int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 639 | struct lttng_ust_shm_handle *handle, |
824f40b8 | 640 | int shadow) |
852c2936 | 641 | { |
824f40b8 MD |
642 | if (shadow) { |
643 | if (uatomic_cmpxchg(&buf->active_shadow_readers, 0, 1) != 0) | |
644 | return -EBUSY; | |
645 | cmm_smp_mb(); | |
646 | return 0; | |
647 | } | |
a6352fd4 | 648 | if (uatomic_cmpxchg(&buf->active_readers, 0, 1) != 0) |
852c2936 | 649 | return -EBUSY; |
a6352fd4 | 650 | cmm_smp_mb(); |
852c2936 MD |
651 | return 0; |
652 | } | |
852c2936 | 653 | |
4cfec15c | 654 | void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 655 | struct lttng_ust_shm_handle *handle, |
824f40b8 | 656 | int shadow) |
852c2936 | 657 | { |
1d498196 | 658 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 | 659 | |
824f40b8 MD |
660 | if (shadow) { |
661 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_shadow_readers) != 1); | |
662 | cmm_smp_mb(); | |
663 | uatomic_dec(&buf->active_shadow_readers); | |
664 | return; | |
665 | } | |
a6352fd4 MD |
666 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
667 | cmm_smp_mb(); | |
668 | uatomic_dec(&buf->active_readers); | |
852c2936 MD |
669 | } |
670 | ||
671 | /** | |
672 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
673 | * @buf: ring buffer | |
674 | * @consumed: consumed count indicating the position where to read | |
675 | * @produced: produced count, indicates position when to stop reading | |
676 | * | |
677 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
678 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 MD |
679 | */ |
680 | ||
4cfec15c | 681 | int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 682 | unsigned long *consumed, unsigned long *produced, |
38fae1d3 | 683 | struct lttng_ust_shm_handle *handle) |
852c2936 | 684 | { |
1d498196 | 685 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 686 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
687 | unsigned long consumed_cur, write_offset; |
688 | int finalized; | |
689 | ||
14641deb | 690 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
691 | /* |
692 | * Read finalized before counters. | |
693 | */ | |
a6352fd4 MD |
694 | cmm_smp_rmb(); |
695 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 MD |
696 | /* |
697 | * No need to issue a memory barrier between consumed count read and | |
698 | * write offset read, because consumed count can only change | |
699 | * concurrently in overwrite mode, and we keep a sequence counter | |
700 | * identifier derived from the write offset to check we are getting | |
701 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
702 | * "tagged" upon writes, tags are checked upon read). | |
703 | */ | |
704 | write_offset = v_read(config, &buf->offset); | |
705 | ||
706 | /* | |
707 | * Check that we are not about to read the same subbuffer in | |
708 | * which the writer head is. | |
709 | */ | |
710 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
711 | == 0) | |
712 | goto nodata; | |
713 | ||
714 | *consumed = consumed_cur; | |
715 | *produced = subbuf_trunc(write_offset, chan); | |
716 | ||
717 | return 0; | |
718 | ||
719 | nodata: | |
720 | /* | |
721 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
722 | * of "raw_spin_is_locked" memory ordering. | |
723 | */ | |
724 | if (finalized) | |
725 | return -ENODATA; | |
852c2936 MD |
726 | else |
727 | return -EAGAIN; | |
728 | } | |
852c2936 MD |
729 | |
730 | /** | |
731 | * lib_ring_buffer_put_snapshot - move consumed counter forward | |
732 | * @buf: ring buffer | |
733 | * @consumed_new: new consumed count value | |
734 | */ | |
4cfec15c | 735 | void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 736 | unsigned long consumed_new, |
38fae1d3 | 737 | struct lttng_ust_shm_handle *handle) |
852c2936 | 738 | { |
4cfec15c | 739 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
1d498196 | 740 | struct channel *chan = shmp(handle, bufb->chan); |
852c2936 MD |
741 | unsigned long consumed; |
742 | ||
824f40b8 MD |
743 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1 |
744 | && uatomic_read(&buf->active_shadow_readers) != 1); | |
852c2936 MD |
745 | |
746 | /* | |
747 | * Only push the consumed value forward. | |
748 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
749 | * the writer in flight recorder mode. | |
750 | */ | |
a6352fd4 | 751 | consumed = uatomic_read(&buf->consumed); |
852c2936 | 752 | while ((long) consumed - (long) consumed_new < 0) |
a6352fd4 MD |
753 | consumed = uatomic_cmpxchg(&buf->consumed, consumed, |
754 | consumed_new); | |
852c2936 | 755 | } |
852c2936 MD |
756 | |
757 | /** | |
758 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
759 | * @buf: ring buffer | |
760 | * @consumed: consumed count indicating the position where to read | |
761 | * | |
762 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
763 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 | 764 | */ |
4cfec15c | 765 | int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 766 | unsigned long consumed, |
38fae1d3 | 767 | struct lttng_ust_shm_handle *handle) |
852c2936 | 768 | { |
1d498196 | 769 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 770 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
771 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
772 | int ret; | |
773 | int finalized; | |
774 | ||
775 | retry: | |
14641deb | 776 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
777 | /* |
778 | * Read finalized before counters. | |
779 | */ | |
a6352fd4 MD |
780 | cmm_smp_rmb(); |
781 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 | 782 | consumed_idx = subbuf_index(consumed, chan); |
4746ae29 | 783 | commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb); |
852c2936 MD |
784 | /* |
785 | * Make sure we read the commit count before reading the buffer | |
786 | * data and the write offset. Correct consumed offset ordering | |
787 | * wrt commit count is insured by the use of cmpxchg to update | |
788 | * the consumed offset. | |
852c2936 | 789 | */ |
a6352fd4 MD |
790 | /* |
791 | * Local rmb to match the remote wmb to read the commit count | |
792 | * before the buffer data and the write offset. | |
793 | */ | |
794 | cmm_smp_rmb(); | |
852c2936 MD |
795 | |
796 | write_offset = v_read(config, &buf->offset); | |
797 | ||
798 | /* | |
799 | * Check that the buffer we are getting is after or at consumed_cur | |
800 | * position. | |
801 | */ | |
802 | if ((long) subbuf_trunc(consumed, chan) | |
803 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
804 | goto nodata; | |
805 | ||
806 | /* | |
807 | * Check that the subbuffer we are trying to consume has been | |
808 | * already fully committed. | |
809 | */ | |
810 | if (((commit_count - chan->backend.subbuf_size) | |
811 | & chan->commit_count_mask) | |
812 | - (buf_trunc(consumed_cur, chan) | |
813 | >> chan->backend.num_subbuf_order) | |
814 | != 0) | |
815 | goto nodata; | |
816 | ||
817 | /* | |
818 | * Check that we are not about to read the same subbuffer in | |
819 | * which the writer head is. | |
820 | */ | |
821 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
822 | == 0) | |
823 | goto nodata; | |
824 | ||
825 | /* | |
826 | * Failure to get the subbuffer causes a busy-loop retry without going | |
827 | * to a wait queue. These are caused by short-lived race windows where | |
828 | * the writer is getting access to a subbuffer we were trying to get | |
829 | * access to. Also checks that the "consumed" buffer count we are | |
830 | * looking for matches the one contained in the subbuffer id. | |
831 | */ | |
832 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
833 | consumed_idx, buf_trunc_val(consumed, chan), |
834 | handle); | |
852c2936 MD |
835 | if (ret) |
836 | goto retry; | |
837 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); | |
838 | ||
839 | buf->get_subbuf_consumed = consumed; | |
840 | buf->get_subbuf = 1; | |
841 | ||
842 | return 0; | |
843 | ||
844 | nodata: | |
845 | /* | |
846 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
847 | * of "raw_spin_is_locked" memory ordering. | |
848 | */ | |
849 | if (finalized) | |
850 | return -ENODATA; | |
852c2936 MD |
851 | else |
852 | return -EAGAIN; | |
853 | } | |
852c2936 MD |
854 | |
855 | /** | |
856 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
857 | * @buf: ring buffer | |
858 | */ | |
4cfec15c | 859 | void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 860 | struct lttng_ust_shm_handle *handle) |
852c2936 | 861 | { |
4cfec15c | 862 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
1d498196 | 863 | struct channel *chan = shmp(handle, bufb->chan); |
4cfec15c | 864 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
865 | unsigned long read_sb_bindex, consumed_idx, consumed; |
866 | ||
824f40b8 MD |
867 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1 |
868 | && uatomic_read(&buf->active_shadow_readers) != 1); | |
852c2936 MD |
869 | |
870 | if (!buf->get_subbuf) { | |
871 | /* | |
872 | * Reader puts a subbuffer it did not get. | |
873 | */ | |
874 | CHAN_WARN_ON(chan, 1); | |
875 | return; | |
876 | } | |
877 | consumed = buf->get_subbuf_consumed; | |
878 | buf->get_subbuf = 0; | |
879 | ||
880 | /* | |
881 | * Clear the records_unread counter. (overruns counter) | |
882 | * Can still be non-zero if a file reader simply grabbed the data | |
883 | * without using iterators. | |
884 | * Can be below zero if an iterator is used on a snapshot more than | |
885 | * once. | |
886 | */ | |
887 | read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); | |
888 | v_add(config, v_read(config, | |
4746ae29 | 889 | &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread), |
852c2936 | 890 | &bufb->records_read); |
4746ae29 | 891 | v_set(config, &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread, 0); |
852c2936 MD |
892 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE |
893 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
894 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
895 | ||
896 | /* | |
897 | * Exchange the reader subbuffer with the one we put in its place in the | |
898 | * writer subbuffer table. Expect the original consumed count. If | |
899 | * update_read_sb_index fails, this is because the writer updated the | |
900 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
901 | * currently have: it has become invalid to try reading this sub-buffer | |
902 | * consumed count value anyway. | |
903 | */ | |
904 | consumed_idx = subbuf_index(consumed, chan); | |
905 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
906 | consumed_idx, buf_trunc_val(consumed, chan), |
907 | handle); | |
852c2936 MD |
908 | /* |
909 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
910 | * if the writer concurrently updated it. | |
911 | */ | |
912 | } | |
852c2936 MD |
913 | |
914 | /* | |
915 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
916 | * position and the writer position. (inclusive) | |
917 | */ | |
918 | static | |
4cfec15c | 919 | void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
920 | struct channel *chan, |
921 | unsigned long cons_offset, | |
1d498196 | 922 | int cpu, |
38fae1d3 | 923 | struct lttng_ust_shm_handle *handle) |
852c2936 | 924 | { |
4cfec15c | 925 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
926 | unsigned long cons_idx, commit_count, commit_count_sb; |
927 | ||
928 | cons_idx = subbuf_index(cons_offset, chan); | |
4746ae29 MD |
929 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, cons_idx)->cc); |
930 | commit_count_sb = v_read(config, &shmp_index(handle, buf->commit_cold, cons_idx)->cc_sb); | |
852c2936 MD |
931 | |
932 | if (subbuf_offset(commit_count, chan) != 0) | |
4d3c9523 | 933 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
934 | "commit count in subbuffer %lu,\n" |
935 | "expecting multiples of %lu bytes\n" | |
936 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
937 | chan->backend.name, cpu, cons_idx, | |
938 | chan->backend.subbuf_size, | |
939 | commit_count, commit_count_sb); | |
940 | ||
4d3c9523 | 941 | DBG("ring buffer: %s, cpu %d: %lu bytes committed\n", |
852c2936 MD |
942 | chan->backend.name, cpu, commit_count); |
943 | } | |
944 | ||
945 | static | |
4cfec15c | 946 | void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 | 947 | struct channel *chan, |
1d498196 | 948 | void *priv, int cpu, |
38fae1d3 | 949 | struct lttng_ust_shm_handle *handle) |
852c2936 | 950 | { |
4cfec15c | 951 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
952 | unsigned long write_offset, cons_offset; |
953 | ||
852c2936 MD |
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 | 1341 | >= chan->backend.buf_size)) { |
64493e4f MD |
1342 | unsigned long nr_lost; |
1343 | ||
852c2936 MD |
1344 | /* |
1345 | * We do not overwrite non consumed buffers | |
1346 | * and we are full : record is lost. | |
1347 | */ | |
64493e4f | 1348 | nr_lost = v_read(config, &buf->records_lost_full); |
852c2936 | 1349 | v_inc(config, &buf->records_lost_full); |
64493e4f MD |
1350 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
1351 | DBG("%lu or more records lost in (%s:%d) (buffer full)\n", | |
1352 | nr_lost + 1, chan->backend.name, | |
1353 | buf->backend.cpu); | |
1354 | } | |
852c2936 MD |
1355 | return -ENOBUFS; |
1356 | } else { | |
1357 | /* | |
1358 | * Next subbuffer not being written to, and we | |
1359 | * are either in overwrite mode or the buffer is | |
1360 | * not full. It's safe to write in this new | |
1361 | * subbuffer. | |
1362 | */ | |
1363 | } | |
1364 | } else { | |
64493e4f MD |
1365 | unsigned long nr_lost; |
1366 | ||
852c2936 MD |
1367 | /* |
1368 | * Next subbuffer reserve offset does not match the | |
1369 | * commit offset. Drop record in producer-consumer and | |
1370 | * overwrite mode. Caused by either a writer OOPS or too | |
1371 | * many nested writes over a reserve/commit pair. | |
1372 | */ | |
64493e4f | 1373 | nr_lost = v_read(config, &buf->records_lost_wrap); |
852c2936 | 1374 | v_inc(config, &buf->records_lost_wrap); |
64493e4f MD |
1375 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
1376 | DBG("%lu or more records lost in (%s:%d) (wrap-around)\n", | |
1377 | nr_lost + 1, chan->backend.name, | |
1378 | buf->backend.cpu); | |
1379 | } | |
852c2936 MD |
1380 | return -EIO; |
1381 | } | |
1382 | offsets->size = | |
1383 | config->cb.record_header_size(config, chan, | |
1384 | offsets->begin, | |
1385 | &offsets->pre_header_padding, | |
1386 | ctx); | |
1387 | offsets->size += | |
1388 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1389 | ctx->largest_align) | |
1390 | + ctx->data_size; | |
b5a3dfa5 | 1391 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) |
852c2936 | 1392 | + offsets->size > chan->backend.subbuf_size)) { |
64493e4f MD |
1393 | unsigned long nr_lost; |
1394 | ||
852c2936 MD |
1395 | /* |
1396 | * Record too big for subbuffers, report error, don't | |
1397 | * complete the sub-buffer switch. | |
1398 | */ | |
64493e4f | 1399 | nr_lost = v_read(config, &buf->records_lost_big); |
852c2936 | 1400 | v_inc(config, &buf->records_lost_big); |
64493e4f MD |
1401 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
1402 | DBG("%lu or more records lost in (%s:%d) record size " | |
1403 | " of %zu bytes is too large for buffer\n", | |
1404 | nr_lost + 1, chan->backend.name, | |
1405 | buf->backend.cpu, offsets->size); | |
1406 | } | |
852c2936 MD |
1407 | return -ENOSPC; |
1408 | } else { | |
1409 | /* | |
1410 | * We just made a successful buffer switch and the | |
1411 | * record fits in the new subbuffer. Let's write. | |
1412 | */ | |
1413 | } | |
1414 | } else { | |
1415 | /* | |
1416 | * Record fits in the current buffer and we are not on a switch | |
1417 | * boundary. It's safe to write. | |
1418 | */ | |
1419 | } | |
1420 | offsets->end = offsets->begin + offsets->size; | |
1421 | ||
b5a3dfa5 | 1422 | if (caa_unlikely(subbuf_offset(offsets->end, chan) == 0)) { |
852c2936 MD |
1423 | /* |
1424 | * The offset_end will fall at the very beginning of the next | |
1425 | * subbuffer. | |
1426 | */ | |
1427 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
1428 | } | |
1429 | return 0; | |
1430 | } | |
1431 | ||
1432 | /** | |
1433 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
1434 | * @ctx: ring buffer context. | |
1435 | * | |
1436 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, | |
1437 | * -EIO for other errors, else returns 0. | |
1438 | * It will take care of sub-buffer switching. | |
1439 | */ | |
4cfec15c | 1440 | int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx) |
852c2936 MD |
1441 | { |
1442 | struct channel *chan = ctx->chan; | |
38fae1d3 | 1443 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c MD |
1444 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
1445 | struct lttng_ust_lib_ring_buffer *buf; | |
852c2936 MD |
1446 | struct switch_offsets offsets; |
1447 | int ret; | |
1448 | ||
1449 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1d498196 | 1450 | buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp); |
852c2936 | 1451 | else |
1d498196 | 1452 | buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 MD |
1453 | ctx->buf = buf; |
1454 | ||
1455 | offsets.size = 0; | |
1456 | ||
1457 | do { | |
1458 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, | |
1459 | ctx); | |
b5a3dfa5 | 1460 | if (caa_unlikely(ret)) |
852c2936 | 1461 | return ret; |
b5a3dfa5 | 1462 | } while (caa_unlikely(v_cmpxchg(config, &buf->offset, offsets.old, |
852c2936 MD |
1463 | offsets.end) |
1464 | != offsets.old)); | |
1465 | ||
1466 | /* | |
1467 | * Atomically update last_tsc. This update races against concurrent | |
1468 | * atomic updates, but the race will always cause supplementary full TSC | |
1469 | * records, never the opposite (missing a full TSC record when it would | |
1470 | * be needed). | |
1471 | */ | |
1472 | save_last_tsc(config, buf, ctx->tsc); | |
1473 | ||
1474 | /* | |
1475 | * Push the reader if necessary | |
1476 | */ | |
1477 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
1478 | ||
1479 | /* | |
1480 | * Clear noref flag for this subbuffer. | |
1481 | */ | |
1482 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1d498196 MD |
1483 | subbuf_index(offsets.end - 1, chan), |
1484 | handle); | |
852c2936 MD |
1485 | |
1486 | /* | |
1487 | * Switch old subbuffer if needed. | |
1488 | */ | |
b5a3dfa5 | 1489 | if (caa_unlikely(offsets.switch_old_end)) { |
852c2936 | 1490 | lib_ring_buffer_clear_noref(config, &buf->backend, |
1d498196 MD |
1491 | subbuf_index(offsets.old - 1, chan), |
1492 | handle); | |
1493 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc, handle); | |
852c2936 MD |
1494 | } |
1495 | ||
1496 | /* | |
1497 | * Populate new subbuffer. | |
1498 | */ | |
b5a3dfa5 | 1499 | if (caa_unlikely(offsets.switch_new_start)) |
1d498196 | 1500 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 | 1501 | |
b5a3dfa5 | 1502 | if (caa_unlikely(offsets.switch_new_end)) |
1d498196 | 1503 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 MD |
1504 | |
1505 | ctx->slot_size = offsets.size; | |
1506 | ctx->pre_offset = offsets.begin; | |
1507 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
1508 | return 0; | |
1509 | } |