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" |
d51652f7 | 56 | #include "tlsfixup.h" |
bdcf8d82 | 57 | #include "../liblttng-ust/compat.h" /* For ENODATA */ |
852c2936 | 58 | |
431d5cf0 MD |
59 | #ifndef max |
60 | #define max(a, b) ((a) > (b) ? (a) : (b)) | |
61 | #endif | |
62 | ||
64493e4f MD |
63 | /* Print DBG() messages about events lost only every 1048576 hits */ |
64 | #define DBG_PRINT_NR_LOST (1UL << 20) | |
65 | ||
2432c3c9 MD |
66 | /* |
67 | * Use POSIX SHM: shm_open(3) and shm_unlink(3). | |
68 | * close(2) to close the fd returned by shm_open. | |
69 | * shm_unlink releases the shared memory object name. | |
70 | * ftruncate(2) sets the size of the memory object. | |
71 | * mmap/munmap maps the shared memory obj to a virtual address in the | |
72 | * calling proceess (should be done both in libust and consumer). | |
73 | * See shm_overview(7) for details. | |
74 | * Pass file descriptor returned by shm_open(3) to ltt-sessiond through | |
75 | * a UNIX socket. | |
76 | * | |
77 | * Since we don't need to access the object using its name, we can | |
78 | * immediately shm_unlink(3) it, and only keep the handle with its file | |
79 | * descriptor. | |
80 | */ | |
81 | ||
852c2936 MD |
82 | /* |
83 | * Internal structure representing offsets to use at a sub-buffer switch. | |
84 | */ | |
85 | struct switch_offsets { | |
86 | unsigned long begin, end, old; | |
87 | size_t pre_header_padding, size; | |
88 | unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1, | |
89 | switch_old_end:1; | |
90 | }; | |
91 | ||
a6352fd4 | 92 | __thread unsigned int lib_ring_buffer_nesting; |
852c2936 | 93 | |
45e9e699 MD |
94 | /* |
95 | * TODO: this is unused. Errors are saved within the ring buffer. | |
96 | * Eventually, allow consumerd to print these errors. | |
97 | */ | |
852c2936 MD |
98 | static |
99 | void lib_ring_buffer_print_errors(struct channel *chan, | |
4cfec15c | 100 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
b68d3dc0 MD |
101 | struct lttng_ust_shm_handle *handle) |
102 | __attribute__((unused)); | |
852c2936 | 103 | |
852c2936 MD |
104 | /** |
105 | * lib_ring_buffer_reset - Reset ring buffer to initial values. | |
106 | * @buf: Ring buffer. | |
107 | * | |
108 | * Effectively empty the ring buffer. Should be called when the buffer is not | |
109 | * used for writing. The ring buffer can be opened for reading, but the reader | |
110 | * should not be using the iterator concurrently with reset. The previous | |
111 | * current iterator record is reset. | |
112 | */ | |
4cfec15c | 113 | void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 114 | struct lttng_ust_shm_handle *handle) |
852c2936 | 115 | { |
1d498196 | 116 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 117 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
118 | unsigned int i; |
119 | ||
120 | /* | |
121 | * Reset iterator first. It will put the subbuffer if it currently holds | |
122 | * it. | |
123 | */ | |
852c2936 MD |
124 | v_set(config, &buf->offset, 0); |
125 | for (i = 0; i < chan->backend.num_subbuf; i++) { | |
4746ae29 MD |
126 | v_set(config, &shmp_index(handle, buf->commit_hot, i)->cc, 0); |
127 | v_set(config, &shmp_index(handle, buf->commit_hot, i)->seq, 0); | |
128 | v_set(config, &shmp_index(handle, buf->commit_cold, i)->cc_sb, 0); | |
852c2936 | 129 | } |
a6352fd4 MD |
130 | uatomic_set(&buf->consumed, 0); |
131 | uatomic_set(&buf->record_disabled, 0); | |
852c2936 | 132 | v_set(config, &buf->last_tsc, 0); |
1d498196 | 133 | lib_ring_buffer_backend_reset(&buf->backend, handle); |
852c2936 MD |
134 | /* Don't reset number of active readers */ |
135 | v_set(config, &buf->records_lost_full, 0); | |
136 | v_set(config, &buf->records_lost_wrap, 0); | |
137 | v_set(config, &buf->records_lost_big, 0); | |
138 | v_set(config, &buf->records_count, 0); | |
139 | v_set(config, &buf->records_overrun, 0); | |
140 | buf->finalized = 0; | |
141 | } | |
852c2936 MD |
142 | |
143 | /** | |
144 | * channel_reset - Reset channel to initial values. | |
145 | * @chan: Channel. | |
146 | * | |
147 | * Effectively empty the channel. Should be called when the channel is not used | |
148 | * for writing. The channel can be opened for reading, but the reader should not | |
149 | * be using the iterator concurrently with reset. The previous current iterator | |
150 | * record is reset. | |
151 | */ | |
152 | void channel_reset(struct channel *chan) | |
153 | { | |
154 | /* | |
155 | * Reset iterators first. Will put the subbuffer if held for reading. | |
156 | */ | |
a6352fd4 | 157 | uatomic_set(&chan->record_disabled, 0); |
852c2936 MD |
158 | /* Don't reset commit_count_mask, still valid */ |
159 | channel_backend_reset(&chan->backend); | |
160 | /* Don't reset switch/read timer interval */ | |
161 | /* Don't reset notifiers and notifier enable bits */ | |
162 | /* Don't reset reader reference count */ | |
163 | } | |
852c2936 MD |
164 | |
165 | /* | |
166 | * Must be called under cpu hotplug protection. | |
167 | */ | |
4cfec15c | 168 | int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf, |
a6352fd4 | 169 | struct channel_backend *chanb, int cpu, |
38fae1d3 | 170 | struct lttng_ust_shm_handle *handle, |
1d498196 | 171 | struct shm_object *shmobj) |
852c2936 | 172 | { |
4cfec15c | 173 | const struct lttng_ust_lib_ring_buffer_config *config = &chanb->config; |
14641deb | 174 | struct channel *chan = caa_container_of(chanb, struct channel, backend); |
a3f61e7f | 175 | void *priv = channel_get_private(chan); |
852c2936 | 176 | size_t subbuf_header_size; |
2fed87ae | 177 | uint64_t tsc; |
852c2936 MD |
178 | int ret; |
179 | ||
180 | /* Test for cpu hotplug */ | |
181 | if (buf->backend.allocated) | |
182 | return 0; | |
183 | ||
a6352fd4 | 184 | ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, |
1d498196 | 185 | cpu, handle, shmobj); |
852c2936 MD |
186 | if (ret) |
187 | return ret; | |
188 | ||
1d498196 MD |
189 | align_shm(shmobj, __alignof__(struct commit_counters_hot)); |
190 | set_shmp(buf->commit_hot, | |
191 | zalloc_shm(shmobj, | |
192 | sizeof(struct commit_counters_hot) * chan->backend.num_subbuf)); | |
193 | if (!shmp(handle, buf->commit_hot)) { | |
852c2936 MD |
194 | ret = -ENOMEM; |
195 | goto free_chanbuf; | |
196 | } | |
197 | ||
1d498196 MD |
198 | align_shm(shmobj, __alignof__(struct commit_counters_cold)); |
199 | set_shmp(buf->commit_cold, | |
200 | zalloc_shm(shmobj, | |
201 | sizeof(struct commit_counters_cold) * chan->backend.num_subbuf)); | |
202 | if (!shmp(handle, buf->commit_cold)) { | |
852c2936 MD |
203 | ret = -ENOMEM; |
204 | goto free_commit; | |
205 | } | |
206 | ||
852c2936 MD |
207 | /* |
208 | * Write the subbuffer header for first subbuffer so we know the total | |
209 | * duration of data gathering. | |
210 | */ | |
211 | subbuf_header_size = config->cb.subbuffer_header_size(); | |
212 | v_set(config, &buf->offset, subbuf_header_size); | |
4746ae29 | 213 | subbuffer_id_clear_noref(config, &shmp_index(handle, buf->backend.buf_wsb, 0)->id); |
1d498196 MD |
214 | tsc = config->cb.ring_buffer_clock_read(shmp(handle, buf->backend.chan)); |
215 | config->cb.buffer_begin(buf, tsc, 0, handle); | |
4746ae29 | 216 | v_add(config, subbuf_header_size, &shmp_index(handle, buf->commit_hot, 0)->cc); |
852c2936 MD |
217 | |
218 | if (config->cb.buffer_create) { | |
1d498196 | 219 | ret = config->cb.buffer_create(buf, priv, cpu, chanb->name, handle); |
852c2936 MD |
220 | if (ret) |
221 | goto free_init; | |
222 | } | |
852c2936 | 223 | buf->backend.allocated = 1; |
852c2936 MD |
224 | return 0; |
225 | ||
226 | /* Error handling */ | |
227 | free_init: | |
a6352fd4 | 228 | /* commit_cold will be freed by shm teardown */ |
852c2936 | 229 | free_commit: |
a6352fd4 | 230 | /* commit_hot will be freed by shm teardown */ |
852c2936 | 231 | free_chanbuf: |
852c2936 MD |
232 | return ret; |
233 | } | |
234 | ||
1d498196 | 235 | #if 0 |
852c2936 MD |
236 | static void switch_buffer_timer(unsigned long data) |
237 | { | |
4cfec15c | 238 | struct lttng_ust_lib_ring_buffer *buf = (struct lttng_ust_lib_ring_buffer *)data; |
1d498196 | 239 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 240 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
241 | |
242 | /* | |
243 | * Only flush buffers periodically if readers are active. | |
244 | */ | |
824f40b8 | 245 | if (uatomic_read(&buf->active_readers) || uatomic_read(&buf->active_shadow_readers)) |
1d498196 | 246 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, handle); |
852c2936 | 247 | |
a6352fd4 MD |
248 | //TODO timers |
249 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
250 | // mod_timer_pinned(&buf->switch_timer, | |
251 | // jiffies + chan->switch_timer_interval); | |
252 | //else | |
253 | // mod_timer(&buf->switch_timer, | |
254 | // jiffies + chan->switch_timer_interval); | |
852c2936 | 255 | } |
1d498196 | 256 | #endif //0 |
852c2936 | 257 | |
4cfec15c | 258 | static void lib_ring_buffer_start_switch_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 259 | struct lttng_ust_shm_handle *handle) |
852c2936 | 260 | { |
1d498196 | 261 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 262 | //const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
263 | |
264 | if (!chan->switch_timer_interval || buf->switch_timer_enabled) | |
265 | return; | |
a6352fd4 MD |
266 | //TODO |
267 | //init_timer(&buf->switch_timer); | |
268 | //buf->switch_timer.function = switch_buffer_timer; | |
269 | //buf->switch_timer.expires = jiffies + chan->switch_timer_interval; | |
270 | //buf->switch_timer.data = (unsigned long)buf; | |
271 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
272 | // add_timer_on(&buf->switch_timer, buf->backend.cpu); | |
273 | //else | |
274 | // add_timer(&buf->switch_timer); | |
852c2936 MD |
275 | buf->switch_timer_enabled = 1; |
276 | } | |
277 | ||
4cfec15c | 278 | static void lib_ring_buffer_stop_switch_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 279 | struct lttng_ust_shm_handle *handle) |
852c2936 | 280 | { |
1d498196 | 281 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 MD |
282 | |
283 | if (!chan->switch_timer_interval || !buf->switch_timer_enabled) | |
284 | return; | |
285 | ||
a6352fd4 MD |
286 | //TODO |
287 | //del_timer_sync(&buf->switch_timer); | |
852c2936 MD |
288 | buf->switch_timer_enabled = 0; |
289 | } | |
290 | ||
1d498196 | 291 | #if 0 |
852c2936 MD |
292 | /* |
293 | * Polling timer to check the channels for data. | |
294 | */ | |
295 | static void read_buffer_timer(unsigned long data) | |
296 | { | |
4cfec15c | 297 | struct lttng_ust_lib_ring_buffer *buf = (struct lttng_ust_lib_ring_buffer *)data; |
1d498196 | 298 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 299 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
300 | |
301 | CHAN_WARN_ON(chan, !buf->backend.allocated); | |
302 | ||
824f40b8 | 303 | if (uatomic_read(&buf->active_readers) || uatomic_read(&buf->active_shadow_readers)) |
852c2936 | 304 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { |
a6352fd4 MD |
305 | //TODO |
306 | //wake_up_interruptible(&buf->read_wait); | |
307 | //wake_up_interruptible(&chan->read_wait); | |
852c2936 MD |
308 | } |
309 | ||
a6352fd4 MD |
310 | //TODO |
311 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
312 | // mod_timer_pinned(&buf->read_timer, | |
313 | // jiffies + chan->read_timer_interval); | |
314 | //else | |
315 | // mod_timer(&buf->read_timer, | |
316 | // jiffies + chan->read_timer_interval); | |
852c2936 | 317 | } |
1d498196 | 318 | #endif //0 |
852c2936 | 319 | |
4cfec15c | 320 | static void lib_ring_buffer_start_read_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 321 | struct lttng_ust_shm_handle *handle) |
852c2936 | 322 | { |
1d498196 | 323 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 324 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
325 | |
326 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
327 | || !chan->read_timer_interval | |
328 | || buf->read_timer_enabled) | |
329 | return; | |
330 | ||
a6352fd4 MD |
331 | //TODO |
332 | //init_timer(&buf->read_timer); | |
333 | //buf->read_timer.function = read_buffer_timer; | |
334 | //buf->read_timer.expires = jiffies + chan->read_timer_interval; | |
335 | //buf->read_timer.data = (unsigned long)buf; | |
852c2936 | 336 | |
a6352fd4 MD |
337 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
338 | // add_timer_on(&buf->read_timer, buf->backend.cpu); | |
339 | //else | |
340 | // add_timer(&buf->read_timer); | |
852c2936 MD |
341 | buf->read_timer_enabled = 1; |
342 | } | |
343 | ||
4cfec15c | 344 | static void lib_ring_buffer_stop_read_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 345 | struct lttng_ust_shm_handle *handle) |
852c2936 | 346 | { |
1d498196 | 347 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 348 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
349 | |
350 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
351 | || !chan->read_timer_interval | |
352 | || !buf->read_timer_enabled) | |
353 | return; | |
354 | ||
a6352fd4 MD |
355 | //TODO |
356 | //del_timer_sync(&buf->read_timer); | |
852c2936 MD |
357 | /* |
358 | * do one more check to catch data that has been written in the last | |
359 | * timer period. | |
360 | */ | |
1d498196 | 361 | if (lib_ring_buffer_poll_deliver(config, buf, chan, handle)) { |
a6352fd4 MD |
362 | //TODO |
363 | //wake_up_interruptible(&buf->read_wait); | |
364 | //wake_up_interruptible(&chan->read_wait); | |
852c2936 MD |
365 | } |
366 | buf->read_timer_enabled = 0; | |
367 | } | |
368 | ||
1d498196 | 369 | static void channel_unregister_notifiers(struct channel *chan, |
38fae1d3 | 370 | struct lttng_ust_shm_handle *handle) |
852c2936 | 371 | { |
4cfec15c | 372 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
373 | int cpu; |
374 | ||
852c2936 | 375 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
852c2936 | 376 | for_each_possible_cpu(cpu) { |
4cfec15c | 377 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
a6352fd4 | 378 | |
1d498196 MD |
379 | lib_ring_buffer_stop_switch_timer(buf, handle); |
380 | lib_ring_buffer_stop_read_timer(buf, handle); | |
852c2936 | 381 | } |
852c2936 | 382 | } else { |
4cfec15c | 383 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 384 | |
1d498196 MD |
385 | lib_ring_buffer_stop_switch_timer(buf, handle); |
386 | lib_ring_buffer_stop_read_timer(buf, handle); | |
852c2936 | 387 | } |
8d8a24c8 | 388 | //channel_backend_unregister_notifiers(&chan->backend); |
852c2936 MD |
389 | } |
390 | ||
38fae1d3 | 391 | static void channel_free(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 392 | int shadow) |
852c2936 | 393 | { |
824f40b8 MD |
394 | if (!shadow) |
395 | channel_backend_free(&chan->backend, handle); | |
431d5cf0 | 396 | /* chan is freed by shm teardown */ |
1d498196 MD |
397 | shm_object_table_destroy(handle->table); |
398 | free(handle); | |
852c2936 MD |
399 | } |
400 | ||
401 | /** | |
402 | * channel_create - Create channel. | |
403 | * @config: ring buffer instance configuration | |
404 | * @name: name of the channel | |
a3f61e7f MD |
405 | * @priv_data: ring buffer client private data area pointer (output) |
406 | * @priv_data_size: length, in bytes, of the private data area. | |
d028eddb | 407 | * @priv_data_init: initialization data for private data. |
852c2936 MD |
408 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous |
409 | * address mapping. It is used only by RING_BUFFER_STATIC | |
410 | * configuration. It can be set to NULL for other backends. | |
411 | * @subbuf_size: subbuffer size | |
412 | * @num_subbuf: number of subbuffers | |
413 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
414 | * padding to let readers get those sub-buffers. | |
415 | * Used for live streaming. | |
416 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
417 | * | |
418 | * Holds cpu hotplug. | |
419 | * Returns NULL on failure. | |
420 | */ | |
4cfec15c | 421 | struct lttng_ust_shm_handle *channel_create(const struct lttng_ust_lib_ring_buffer_config *config, |
a3f61e7f MD |
422 | const char *name, |
423 | void **priv_data, | |
424 | size_t priv_data_align, | |
425 | size_t priv_data_size, | |
d028eddb | 426 | void *priv_data_init, |
a3f61e7f | 427 | void *buf_addr, size_t subbuf_size, |
852c2936 | 428 | size_t num_subbuf, unsigned int switch_timer_interval, |
193183fb | 429 | unsigned int read_timer_interval, |
ef9ff354 | 430 | int **shm_fd, int **wait_fd, uint64_t **memory_map_size) |
852c2936 | 431 | { |
1d498196 | 432 | int ret, cpu; |
a3f61e7f | 433 | size_t shmsize, chansize; |
852c2936 | 434 | struct channel *chan; |
38fae1d3 | 435 | struct lttng_ust_shm_handle *handle; |
1d498196 | 436 | struct shm_object *shmobj; |
193183fb | 437 | struct shm_ref *ref; |
852c2936 MD |
438 | |
439 | if (lib_ring_buffer_check_config(config, switch_timer_interval, | |
440 | read_timer_interval)) | |
441 | return NULL; | |
442 | ||
38fae1d3 | 443 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
431d5cf0 MD |
444 | if (!handle) |
445 | return NULL; | |
446 | ||
1d498196 MD |
447 | /* Allocate table for channel + per-cpu buffers */ |
448 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
449 | if (!handle->table) | |
450 | goto error_table_alloc; | |
852c2936 | 451 | |
1d498196 MD |
452 | /* Calculate the shm allocation layout */ |
453 | shmsize = sizeof(struct channel); | |
c1fca457 | 454 | shmsize += offset_align(shmsize, __alignof__(struct lttng_ust_lib_ring_buffer_shmp)); |
1d498196 | 455 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
4cfec15c | 456 | shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp) * num_possible_cpus(); |
1d498196 | 457 | else |
4cfec15c | 458 | shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp); |
a3f61e7f MD |
459 | chansize = shmsize; |
460 | shmsize += offset_align(shmsize, priv_data_align); | |
461 | shmsize += priv_data_size; | |
a6352fd4 | 462 | |
1d498196 | 463 | shmobj = shm_object_table_append(handle->table, shmsize); |
b5a14697 MD |
464 | if (!shmobj) |
465 | goto error_append; | |
57773204 | 466 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
a3f61e7f | 467 | set_shmp(handle->chan, zalloc_shm(shmobj, chansize)); |
57773204 MD |
468 | assert(handle->chan._ref.index == 0); |
469 | assert(handle->chan._ref.offset == 0); | |
1d498196 | 470 | chan = shmp(handle, handle->chan); |
a6352fd4 | 471 | if (!chan) |
1d498196 | 472 | goto error_append; |
a6352fd4 | 473 | |
a3f61e7f MD |
474 | /* space for private data */ |
475 | if (priv_data_size) { | |
476 | DECLARE_SHMP(void, priv_data_alloc); | |
477 | ||
478 | align_shm(shmobj, priv_data_align); | |
479 | chan->priv_data_offset = shmobj->allocated_len; | |
480 | set_shmp(priv_data_alloc, zalloc_shm(shmobj, priv_data_size)); | |
481 | if (!shmp(handle, priv_data_alloc)) | |
482 | goto error_append; | |
483 | *priv_data = channel_get_private(chan); | |
d028eddb | 484 | memcpy(*priv_data, priv_data_init, priv_data_size); |
a3f61e7f MD |
485 | } else { |
486 | chan->priv_data_offset = -1; | |
487 | *priv_data = NULL; | |
488 | } | |
489 | ||
490 | ret = channel_backend_init(&chan->backend, name, config, | |
1d498196 | 491 | subbuf_size, num_subbuf, handle); |
852c2936 | 492 | if (ret) |
1d498196 | 493 | goto error_backend_init; |
852c2936 MD |
494 | |
495 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); | |
a6352fd4 MD |
496 | //TODO |
497 | //chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval); | |
498 | //chan->read_timer_interval = usecs_to_jiffies(read_timer_interval); | |
a6352fd4 MD |
499 | //TODO |
500 | //init_waitqueue_head(&chan->read_wait); | |
501 | //init_waitqueue_head(&chan->hp_wait); | |
852c2936 MD |
502 | |
503 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
852c2936 MD |
504 | /* |
505 | * In case of non-hotplug cpu, if the ring-buffer is allocated | |
506 | * in early initcall, it will not be notified of secondary cpus. | |
507 | * In that off case, we need to allocate for all possible cpus. | |
508 | */ | |
852c2936 | 509 | for_each_possible_cpu(cpu) { |
4cfec15c | 510 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
1d498196 MD |
511 | lib_ring_buffer_start_switch_timer(buf, handle); |
512 | lib_ring_buffer_start_read_timer(buf, handle); | |
852c2936 | 513 | } |
852c2936 | 514 | } else { |
4cfec15c | 515 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 516 | |
1d498196 MD |
517 | lib_ring_buffer_start_switch_timer(buf, handle); |
518 | lib_ring_buffer_start_read_timer(buf, handle); | |
852c2936 | 519 | } |
193183fb MD |
520 | ref = &handle->chan._ref; |
521 | shm_get_object_data(handle, ref, shm_fd, wait_fd, memory_map_size); | |
431d5cf0 | 522 | return handle; |
852c2936 | 523 | |
1d498196 MD |
524 | error_backend_init: |
525 | error_append: | |
526 | shm_object_table_destroy(handle->table); | |
527 | error_table_alloc: | |
431d5cf0 | 528 | free(handle); |
852c2936 MD |
529 | return NULL; |
530 | } | |
852c2936 | 531 | |
38fae1d3 | 532 | struct lttng_ust_shm_handle *channel_handle_create(int shm_fd, int wait_fd, |
193183fb MD |
533 | uint64_t memory_map_size) |
534 | { | |
38fae1d3 | 535 | struct lttng_ust_shm_handle *handle; |
193183fb MD |
536 | struct shm_object *object; |
537 | ||
38fae1d3 | 538 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
193183fb MD |
539 | if (!handle) |
540 | return NULL; | |
541 | ||
542 | /* Allocate table for channel + per-cpu buffers */ | |
543 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
544 | if (!handle->table) | |
545 | goto error_table_alloc; | |
546 | /* Add channel object */ | |
547 | object = shm_object_table_append_shadow(handle->table, | |
548 | shm_fd, wait_fd, memory_map_size); | |
549 | if (!object) | |
550 | goto error_table_object; | |
57773204 MD |
551 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
552 | handle->chan._ref.index = 0; | |
553 | handle->chan._ref.offset = 0; | |
193183fb MD |
554 | return handle; |
555 | ||
556 | error_table_object: | |
557 | shm_object_table_destroy(handle->table); | |
558 | error_table_alloc: | |
559 | free(handle); | |
560 | return NULL; | |
561 | } | |
562 | ||
38fae1d3 | 563 | int channel_handle_add_stream(struct lttng_ust_shm_handle *handle, |
193183fb MD |
564 | int shm_fd, int wait_fd, uint64_t memory_map_size) |
565 | { | |
566 | struct shm_object *object; | |
567 | ||
568 | /* Add stream object */ | |
569 | object = shm_object_table_append_shadow(handle->table, | |
570 | shm_fd, wait_fd, memory_map_size); | |
571 | if (!object) | |
572 | return -1; | |
573 | return 0; | |
574 | } | |
575 | ||
852c2936 | 576 | static |
38fae1d3 | 577 | void channel_release(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 578 | int shadow) |
852c2936 | 579 | { |
824f40b8 | 580 | channel_free(chan, handle, shadow); |
852c2936 MD |
581 | } |
582 | ||
583 | /** | |
584 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
585 | * @chan: channel to destroy | |
586 | * | |
587 | * Holds cpu hotplug. | |
431d5cf0 MD |
588 | * Call "destroy" callback, finalize channels, decrement the channel |
589 | * reference count. Note that when readers have completed data | |
590 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
a3f61e7f | 591 | * They should release their handle at that point. |
852c2936 | 592 | */ |
a3f61e7f | 593 | void channel_destroy(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 594 | int shadow) |
852c2936 | 595 | { |
824f40b8 MD |
596 | if (shadow) { |
597 | channel_release(chan, handle, shadow); | |
a3f61e7f | 598 | return; |
824f40b8 MD |
599 | } |
600 | ||
1d498196 | 601 | channel_unregister_notifiers(chan, handle); |
852c2936 | 602 | |
45e9e699 MD |
603 | /* |
604 | * Note: the consumer takes care of finalizing and switching the | |
605 | * buffers. | |
606 | */ | |
852c2936 | 607 | |
431d5cf0 MD |
608 | /* |
609 | * sessiond/consumer are keeping a reference on the shm file | |
610 | * descriptor directly. No need to refcount. | |
611 | */ | |
824f40b8 | 612 | channel_release(chan, handle, shadow); |
a3f61e7f | 613 | return; |
852c2936 | 614 | } |
852c2936 | 615 | |
4cfec15c MD |
616 | struct lttng_ust_lib_ring_buffer *channel_get_ring_buffer( |
617 | const struct lttng_ust_lib_ring_buffer_config *config, | |
1d498196 | 618 | struct channel *chan, int cpu, |
38fae1d3 | 619 | struct lttng_ust_shm_handle *handle, |
ef9ff354 MD |
620 | int **shm_fd, int **wait_fd, |
621 | uint64_t **memory_map_size) | |
852c2936 | 622 | { |
381c0f1e MD |
623 | struct shm_ref *ref; |
624 | ||
625 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
626 | ref = &chan->backend.buf[0].shmp._ref; | |
627 | shm_get_object_data(handle, ref, shm_fd, wait_fd, | |
628 | memory_map_size); | |
1d498196 | 629 | return shmp(handle, chan->backend.buf[0].shmp); |
381c0f1e | 630 | } else { |
e095d803 MD |
631 | if (cpu >= num_possible_cpus()) |
632 | return NULL; | |
381c0f1e MD |
633 | ref = &chan->backend.buf[cpu].shmp._ref; |
634 | shm_get_object_data(handle, ref, shm_fd, wait_fd, | |
635 | memory_map_size); | |
1d498196 | 636 | return shmp(handle, chan->backend.buf[cpu].shmp); |
381c0f1e | 637 | } |
852c2936 | 638 | } |
852c2936 | 639 | |
4cfec15c | 640 | int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 641 | struct lttng_ust_shm_handle *handle, |
824f40b8 | 642 | int shadow) |
852c2936 | 643 | { |
824f40b8 MD |
644 | if (shadow) { |
645 | if (uatomic_cmpxchg(&buf->active_shadow_readers, 0, 1) != 0) | |
646 | return -EBUSY; | |
647 | cmm_smp_mb(); | |
648 | return 0; | |
649 | } | |
a6352fd4 | 650 | if (uatomic_cmpxchg(&buf->active_readers, 0, 1) != 0) |
852c2936 | 651 | return -EBUSY; |
a6352fd4 | 652 | cmm_smp_mb(); |
852c2936 MD |
653 | return 0; |
654 | } | |
852c2936 | 655 | |
4cfec15c | 656 | void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 657 | struct lttng_ust_shm_handle *handle, |
824f40b8 | 658 | int shadow) |
852c2936 | 659 | { |
1d498196 | 660 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 | 661 | |
824f40b8 MD |
662 | if (shadow) { |
663 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_shadow_readers) != 1); | |
664 | cmm_smp_mb(); | |
665 | uatomic_dec(&buf->active_shadow_readers); | |
666 | return; | |
667 | } | |
a6352fd4 MD |
668 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
669 | cmm_smp_mb(); | |
670 | uatomic_dec(&buf->active_readers); | |
852c2936 MD |
671 | } |
672 | ||
673 | /** | |
674 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
675 | * @buf: ring buffer | |
676 | * @consumed: consumed count indicating the position where to read | |
677 | * @produced: produced count, indicates position when to stop reading | |
678 | * | |
679 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
680 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 MD |
681 | */ |
682 | ||
4cfec15c | 683 | int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 684 | unsigned long *consumed, unsigned long *produced, |
38fae1d3 | 685 | struct lttng_ust_shm_handle *handle) |
852c2936 | 686 | { |
1d498196 | 687 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 688 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
689 | unsigned long consumed_cur, write_offset; |
690 | int finalized; | |
691 | ||
14641deb | 692 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
693 | /* |
694 | * Read finalized before counters. | |
695 | */ | |
a6352fd4 MD |
696 | cmm_smp_rmb(); |
697 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 MD |
698 | /* |
699 | * No need to issue a memory barrier between consumed count read and | |
700 | * write offset read, because consumed count can only change | |
701 | * concurrently in overwrite mode, and we keep a sequence counter | |
702 | * identifier derived from the write offset to check we are getting | |
703 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
704 | * "tagged" upon writes, tags are checked upon read). | |
705 | */ | |
706 | write_offset = v_read(config, &buf->offset); | |
707 | ||
708 | /* | |
709 | * Check that we are not about to read the same subbuffer in | |
710 | * which the writer head is. | |
711 | */ | |
712 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
713 | == 0) | |
714 | goto nodata; | |
715 | ||
716 | *consumed = consumed_cur; | |
717 | *produced = subbuf_trunc(write_offset, chan); | |
718 | ||
719 | return 0; | |
720 | ||
721 | nodata: | |
722 | /* | |
723 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
724 | * of "raw_spin_is_locked" memory ordering. | |
725 | */ | |
726 | if (finalized) | |
727 | return -ENODATA; | |
852c2936 MD |
728 | else |
729 | return -EAGAIN; | |
730 | } | |
852c2936 MD |
731 | |
732 | /** | |
733 | * lib_ring_buffer_put_snapshot - move consumed counter forward | |
734 | * @buf: ring buffer | |
735 | * @consumed_new: new consumed count value | |
736 | */ | |
4cfec15c | 737 | void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 738 | unsigned long consumed_new, |
38fae1d3 | 739 | struct lttng_ust_shm_handle *handle) |
852c2936 | 740 | { |
4cfec15c | 741 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
1d498196 | 742 | struct channel *chan = shmp(handle, bufb->chan); |
852c2936 MD |
743 | unsigned long consumed; |
744 | ||
824f40b8 MD |
745 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1 |
746 | && uatomic_read(&buf->active_shadow_readers) != 1); | |
852c2936 MD |
747 | |
748 | /* | |
749 | * Only push the consumed value forward. | |
750 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
751 | * the writer in flight recorder mode. | |
752 | */ | |
a6352fd4 | 753 | consumed = uatomic_read(&buf->consumed); |
852c2936 | 754 | while ((long) consumed - (long) consumed_new < 0) |
a6352fd4 MD |
755 | consumed = uatomic_cmpxchg(&buf->consumed, consumed, |
756 | consumed_new); | |
852c2936 | 757 | } |
852c2936 MD |
758 | |
759 | /** | |
760 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
761 | * @buf: ring buffer | |
762 | * @consumed: consumed count indicating the position where to read | |
763 | * | |
764 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
765 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 | 766 | */ |
4cfec15c | 767 | int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 768 | unsigned long consumed, |
38fae1d3 | 769 | struct lttng_ust_shm_handle *handle) |
852c2936 | 770 | { |
1d498196 | 771 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 772 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
773 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
774 | int ret; | |
775 | int finalized; | |
776 | ||
777 | retry: | |
14641deb | 778 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
779 | /* |
780 | * Read finalized before counters. | |
781 | */ | |
a6352fd4 MD |
782 | cmm_smp_rmb(); |
783 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 | 784 | consumed_idx = subbuf_index(consumed, chan); |
4746ae29 | 785 | commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb); |
852c2936 MD |
786 | /* |
787 | * Make sure we read the commit count before reading the buffer | |
788 | * data and the write offset. Correct consumed offset ordering | |
789 | * wrt commit count is insured by the use of cmpxchg to update | |
790 | * the consumed offset. | |
852c2936 | 791 | */ |
a6352fd4 MD |
792 | /* |
793 | * Local rmb to match the remote wmb to read the commit count | |
794 | * before the buffer data and the write offset. | |
795 | */ | |
796 | cmm_smp_rmb(); | |
852c2936 MD |
797 | |
798 | write_offset = v_read(config, &buf->offset); | |
799 | ||
800 | /* | |
801 | * Check that the buffer we are getting is after or at consumed_cur | |
802 | * position. | |
803 | */ | |
804 | if ((long) subbuf_trunc(consumed, chan) | |
805 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
806 | goto nodata; | |
807 | ||
808 | /* | |
809 | * Check that the subbuffer we are trying to consume has been | |
810 | * already fully committed. | |
811 | */ | |
812 | if (((commit_count - chan->backend.subbuf_size) | |
813 | & chan->commit_count_mask) | |
814 | - (buf_trunc(consumed_cur, chan) | |
815 | >> chan->backend.num_subbuf_order) | |
816 | != 0) | |
817 | goto nodata; | |
818 | ||
819 | /* | |
820 | * Check that we are not about to read the same subbuffer in | |
821 | * which the writer head is. | |
822 | */ | |
823 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
824 | == 0) | |
825 | goto nodata; | |
826 | ||
827 | /* | |
828 | * Failure to get the subbuffer causes a busy-loop retry without going | |
829 | * to a wait queue. These are caused by short-lived race windows where | |
830 | * the writer is getting access to a subbuffer we were trying to get | |
831 | * access to. Also checks that the "consumed" buffer count we are | |
832 | * looking for matches the one contained in the subbuffer id. | |
833 | */ | |
834 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
835 | consumed_idx, buf_trunc_val(consumed, chan), |
836 | handle); | |
852c2936 MD |
837 | if (ret) |
838 | goto retry; | |
839 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); | |
840 | ||
841 | buf->get_subbuf_consumed = consumed; | |
842 | buf->get_subbuf = 1; | |
843 | ||
844 | return 0; | |
845 | ||
846 | nodata: | |
847 | /* | |
848 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
849 | * of "raw_spin_is_locked" memory ordering. | |
850 | */ | |
851 | if (finalized) | |
852 | return -ENODATA; | |
852c2936 MD |
853 | else |
854 | return -EAGAIN; | |
855 | } | |
852c2936 MD |
856 | |
857 | /** | |
858 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
859 | * @buf: ring buffer | |
860 | */ | |
4cfec15c | 861 | void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 862 | struct lttng_ust_shm_handle *handle) |
852c2936 | 863 | { |
4cfec15c | 864 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
1d498196 | 865 | struct channel *chan = shmp(handle, bufb->chan); |
4cfec15c | 866 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
867 | unsigned long read_sb_bindex, consumed_idx, consumed; |
868 | ||
824f40b8 MD |
869 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1 |
870 | && uatomic_read(&buf->active_shadow_readers) != 1); | |
852c2936 MD |
871 | |
872 | if (!buf->get_subbuf) { | |
873 | /* | |
874 | * Reader puts a subbuffer it did not get. | |
875 | */ | |
876 | CHAN_WARN_ON(chan, 1); | |
877 | return; | |
878 | } | |
879 | consumed = buf->get_subbuf_consumed; | |
880 | buf->get_subbuf = 0; | |
881 | ||
882 | /* | |
883 | * Clear the records_unread counter. (overruns counter) | |
884 | * Can still be non-zero if a file reader simply grabbed the data | |
885 | * without using iterators. | |
886 | * Can be below zero if an iterator is used on a snapshot more than | |
887 | * once. | |
888 | */ | |
889 | read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); | |
890 | v_add(config, v_read(config, | |
4746ae29 | 891 | &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread), |
852c2936 | 892 | &bufb->records_read); |
4746ae29 | 893 | v_set(config, &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread, 0); |
852c2936 MD |
894 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE |
895 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
896 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
897 | ||
898 | /* | |
899 | * Exchange the reader subbuffer with the one we put in its place in the | |
900 | * writer subbuffer table. Expect the original consumed count. If | |
901 | * update_read_sb_index fails, this is because the writer updated the | |
902 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
903 | * currently have: it has become invalid to try reading this sub-buffer | |
904 | * consumed count value anyway. | |
905 | */ | |
906 | consumed_idx = subbuf_index(consumed, chan); | |
907 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
908 | consumed_idx, buf_trunc_val(consumed, chan), |
909 | handle); | |
852c2936 MD |
910 | /* |
911 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
912 | * if the writer concurrently updated it. | |
913 | */ | |
914 | } | |
852c2936 MD |
915 | |
916 | /* | |
917 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
918 | * position and the writer position. (inclusive) | |
919 | */ | |
920 | static | |
4cfec15c | 921 | void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
922 | struct channel *chan, |
923 | unsigned long cons_offset, | |
1d498196 | 924 | int cpu, |
38fae1d3 | 925 | struct lttng_ust_shm_handle *handle) |
852c2936 | 926 | { |
4cfec15c | 927 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
928 | unsigned long cons_idx, commit_count, commit_count_sb; |
929 | ||
930 | cons_idx = subbuf_index(cons_offset, chan); | |
4746ae29 MD |
931 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, cons_idx)->cc); |
932 | commit_count_sb = v_read(config, &shmp_index(handle, buf->commit_cold, cons_idx)->cc_sb); | |
852c2936 MD |
933 | |
934 | if (subbuf_offset(commit_count, chan) != 0) | |
4d3c9523 | 935 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
936 | "commit count in subbuffer %lu,\n" |
937 | "expecting multiples of %lu bytes\n" | |
938 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
939 | chan->backend.name, cpu, cons_idx, | |
940 | chan->backend.subbuf_size, | |
941 | commit_count, commit_count_sb); | |
942 | ||
4d3c9523 | 943 | DBG("ring buffer: %s, cpu %d: %lu bytes committed\n", |
852c2936 MD |
944 | chan->backend.name, cpu, commit_count); |
945 | } | |
946 | ||
947 | static | |
4cfec15c | 948 | void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 | 949 | struct channel *chan, |
1d498196 | 950 | void *priv, int cpu, |
38fae1d3 | 951 | struct lttng_ust_shm_handle *handle) |
852c2936 | 952 | { |
4cfec15c | 953 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
954 | unsigned long write_offset, cons_offset; |
955 | ||
852c2936 MD |
956 | /* |
957 | * No need to order commit_count, write_offset and cons_offset reads | |
958 | * because we execute at teardown when no more writer nor reader | |
959 | * references are left. | |
960 | */ | |
961 | write_offset = v_read(config, &buf->offset); | |
a6352fd4 | 962 | cons_offset = uatomic_read(&buf->consumed); |
852c2936 | 963 | if (write_offset != cons_offset) |
4d3c9523 | 964 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
965 | "non-consumed data\n" |
966 | " [ %lu bytes written, %lu bytes read ]\n", | |
967 | chan->backend.name, cpu, write_offset, cons_offset); | |
968 | ||
a6352fd4 | 969 | for (cons_offset = uatomic_read(&buf->consumed); |
852c2936 MD |
970 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), |
971 | chan) | |
972 | - cons_offset) > 0; | |
973 | cons_offset = subbuf_align(cons_offset, chan)) | |
974 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1d498196 | 975 | cpu, handle); |
852c2936 MD |
976 | } |
977 | ||
978 | static | |
979 | void lib_ring_buffer_print_errors(struct channel *chan, | |
4cfec15c | 980 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
38fae1d3 | 981 | struct lttng_ust_shm_handle *handle) |
852c2936 | 982 | { |
4cfec15c | 983 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
a3f61e7f | 984 | void *priv = channel_get_private(chan); |
852c2936 | 985 | |
a1360615 MD |
986 | if (!strcmp(chan->backend.name, "relay-metadata-mmap")) { |
987 | DBG("ring buffer %s: %lu records written, " | |
988 | "%lu records overrun\n", | |
989 | chan->backend.name, | |
990 | v_read(config, &buf->records_count), | |
991 | v_read(config, &buf->records_overrun)); | |
992 | } else { | |
993 | DBG("ring buffer %s, cpu %d: %lu records written, " | |
994 | "%lu records overrun\n", | |
995 | chan->backend.name, cpu, | |
996 | v_read(config, &buf->records_count), | |
997 | v_read(config, &buf->records_overrun)); | |
998 | ||
999 | if (v_read(config, &buf->records_lost_full) | |
1000 | || v_read(config, &buf->records_lost_wrap) | |
1001 | || v_read(config, &buf->records_lost_big)) | |
1002 | DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n" | |
1003 | " [ %lu buffer full, %lu nest buffer wrap-around, " | |
1004 | "%lu event too big ]\n", | |
1005 | chan->backend.name, cpu, | |
1006 | v_read(config, &buf->records_lost_full), | |
1007 | v_read(config, &buf->records_lost_wrap), | |
1008 | v_read(config, &buf->records_lost_big)); | |
1009 | } | |
1d498196 | 1010 | lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu, handle); |
852c2936 MD |
1011 | } |
1012 | ||
1013 | /* | |
1014 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
1015 | * | |
1016 | * Only executed when the buffer is finalized, in SWITCH_FLUSH. | |
1017 | */ | |
1018 | static | |
4cfec15c | 1019 | void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1020 | struct channel *chan, |
1021 | struct switch_offsets *offsets, | |
2fed87ae | 1022 | uint64_t tsc, |
38fae1d3 | 1023 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1024 | { |
4cfec15c | 1025 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1026 | unsigned long oldidx = subbuf_index(offsets->old, chan); |
1027 | unsigned long commit_count; | |
1028 | ||
1d498196 | 1029 | config->cb.buffer_begin(buf, tsc, oldidx, handle); |
852c2936 MD |
1030 | |
1031 | /* | |
1032 | * Order all writes to buffer before the commit count update that will | |
1033 | * determine that the subbuffer is full. | |
1034 | */ | |
a6352fd4 | 1035 | cmm_smp_wmb(); |
852c2936 | 1036 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1037 | &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1038 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 MD |
1039 | /* Check if the written buffer has to be delivered */ |
1040 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
1d498196 | 1041 | commit_count, oldidx, handle); |
852c2936 MD |
1042 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
1043 | offsets->old, commit_count, | |
1d498196 MD |
1044 | config->cb.subbuffer_header_size(), |
1045 | handle); | |
852c2936 MD |
1046 | } |
1047 | ||
1048 | /* | |
1049 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
1050 | * | |
1051 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
1052 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
1053 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
1054 | * subbuffer. | |
1055 | */ | |
1056 | static | |
4cfec15c | 1057 | void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1058 | struct channel *chan, |
1059 | struct switch_offsets *offsets, | |
2fed87ae | 1060 | uint64_t tsc, |
38fae1d3 | 1061 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1062 | { |
4cfec15c | 1063 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1064 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); |
1065 | unsigned long commit_count, padding_size, data_size; | |
1066 | ||
1067 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1068 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1069 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size, |
1070 | handle); | |
852c2936 MD |
1071 | |
1072 | /* | |
1073 | * Order all writes to buffer before the commit count update that will | |
1074 | * determine that the subbuffer is full. | |
1075 | */ | |
a6352fd4 | 1076 | cmm_smp_wmb(); |
4746ae29 MD |
1077 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1078 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 | 1079 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, |
1d498196 | 1080 | commit_count, oldidx, handle); |
852c2936 MD |
1081 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
1082 | offsets->old, commit_count, | |
1d498196 | 1083 | padding_size, handle); |
852c2936 MD |
1084 | } |
1085 | ||
1086 | /* | |
1087 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1088 | * | |
1089 | * This code can be executed unordered : writers may already have written to the | |
1090 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1091 | * that this code is executed before the deliver of this sub-buffer. | |
1092 | */ | |
1093 | static | |
4cfec15c | 1094 | void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1095 | struct channel *chan, |
1096 | struct switch_offsets *offsets, | |
2fed87ae | 1097 | uint64_t tsc, |
38fae1d3 | 1098 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1099 | { |
4cfec15c | 1100 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1101 | unsigned long beginidx = subbuf_index(offsets->begin, chan); |
1102 | unsigned long commit_count; | |
1103 | ||
1d498196 | 1104 | config->cb.buffer_begin(buf, tsc, beginidx, handle); |
852c2936 MD |
1105 | |
1106 | /* | |
1107 | * Order all writes to buffer before the commit count update that will | |
1108 | * determine that the subbuffer is full. | |
1109 | */ | |
a6352fd4 | 1110 | cmm_smp_wmb(); |
852c2936 | 1111 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1112 | &shmp_index(handle, buf->commit_hot, beginidx)->cc); |
1113 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, beginidx)->cc); | |
852c2936 MD |
1114 | /* Check if the written buffer has to be delivered */ |
1115 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
1d498196 | 1116 | commit_count, beginidx, handle); |
852c2936 MD |
1117 | lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx, |
1118 | offsets->begin, commit_count, | |
1d498196 MD |
1119 | config->cb.subbuffer_header_size(), |
1120 | handle); | |
852c2936 MD |
1121 | } |
1122 | ||
1123 | /* | |
1124 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1125 | * | |
1126 | * The only remaining threads could be the ones with pending commits. They will | |
1127 | * have to do the deliver themselves. | |
1128 | */ | |
1129 | static | |
4cfec15c | 1130 | void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 MD |
1131 | struct channel *chan, |
1132 | struct switch_offsets *offsets, | |
2fed87ae | 1133 | uint64_t tsc, |
38fae1d3 | 1134 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1135 | { |
4cfec15c | 1136 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1137 | unsigned long endidx = subbuf_index(offsets->end - 1, chan); |
1138 | unsigned long commit_count, padding_size, data_size; | |
1139 | ||
1140 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; | |
1141 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1142 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size, |
1143 | handle); | |
852c2936 MD |
1144 | |
1145 | /* | |
1146 | * Order all writes to buffer before the commit count update that will | |
1147 | * determine that the subbuffer is full. | |
1148 | */ | |
a6352fd4 | 1149 | cmm_smp_wmb(); |
4746ae29 MD |
1150 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, endidx)->cc); |
1151 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc); | |
852c2936 | 1152 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->end - 1, |
1d498196 | 1153 | commit_count, endidx, handle); |
852c2936 MD |
1154 | lib_ring_buffer_write_commit_counter(config, buf, chan, endidx, |
1155 | offsets->end, commit_count, | |
1d498196 | 1156 | padding_size, handle); |
852c2936 MD |
1157 | } |
1158 | ||
1159 | /* | |
1160 | * Returns : | |
1161 | * 0 if ok | |
1162 | * !0 if execution must be aborted. | |
1163 | */ | |
1164 | static | |
1165 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
4cfec15c | 1166 | struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1167 | struct channel *chan, |
1168 | struct switch_offsets *offsets, | |
2fed87ae | 1169 | uint64_t *tsc) |
852c2936 | 1170 | { |
4cfec15c | 1171 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1172 | unsigned long off; |
1173 | ||
1174 | offsets->begin = v_read(config, &buf->offset); | |
1175 | offsets->old = offsets->begin; | |
1176 | offsets->switch_old_start = 0; | |
1177 | off = subbuf_offset(offsets->begin, chan); | |
1178 | ||
1179 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1180 | ||
1181 | /* | |
1182 | * Ensure we flush the header of an empty subbuffer when doing the | |
1183 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1184 | * total data gathering duration even if there were no records saved | |
1185 | * after the last buffer switch. | |
1186 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1187 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1188 | * subbuffer header as appropriate. | |
1189 | * The next record that reserves space will be responsible for | |
1190 | * populating the following subbuffer header. We choose not to populate | |
1191 | * the next subbuffer header here because we want to be able to use | |
a6352fd4 MD |
1192 | * SWITCH_ACTIVE for periodical buffer flush, which must |
1193 | * guarantee that all the buffer content (records and header | |
1194 | * timestamps) are visible to the reader. This is required for | |
1195 | * quiescence guarantees for the fusion merge. | |
852c2936 MD |
1196 | */ |
1197 | if (mode == SWITCH_FLUSH || off > 0) { | |
b5a3dfa5 | 1198 | if (caa_unlikely(off == 0)) { |
852c2936 MD |
1199 | /* |
1200 | * The client does not save any header information. | |
1201 | * Don't switch empty subbuffer on finalize, because it | |
1202 | * is invalid to deliver a completely empty subbuffer. | |
1203 | */ | |
1204 | if (!config->cb.subbuffer_header_size()) | |
1205 | return -1; | |
1206 | /* | |
1207 | * Need to write the subbuffer start header on finalize. | |
1208 | */ | |
1209 | offsets->switch_old_start = 1; | |
1210 | } | |
1211 | offsets->begin = subbuf_align(offsets->begin, chan); | |
1212 | } else | |
1213 | return -1; /* we do not have to switch : buffer is empty */ | |
1214 | /* Note: old points to the next subbuf at offset 0 */ | |
1215 | offsets->end = offsets->begin; | |
1216 | return 0; | |
1217 | } | |
1218 | ||
1219 | /* | |
1220 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
1221 | * called while tracing is active with absolutely no lock held. | |
1222 | * | |
1223 | * Note, however, that as a v_cmpxchg is used for some atomic | |
1224 | * operations, this function must be called from the CPU which owns the buffer | |
1225 | * for a ACTIVE flush. | |
1226 | */ | |
4cfec15c | 1227 | void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode, |
38fae1d3 | 1228 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1229 | { |
1d498196 | 1230 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 1231 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1232 | struct switch_offsets offsets; |
1233 | unsigned long oldidx; | |
2fed87ae | 1234 | uint64_t tsc; |
852c2936 MD |
1235 | |
1236 | offsets.size = 0; | |
1237 | ||
1238 | /* | |
1239 | * Perform retryable operations. | |
1240 | */ | |
1241 | do { | |
1242 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
1243 | &tsc)) | |
1244 | return; /* Switch not needed */ | |
1245 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
1246 | != offsets.old); | |
1247 | ||
1248 | /* | |
1249 | * Atomically update last_tsc. This update races against concurrent | |
1250 | * atomic updates, but the race will always cause supplementary full TSC | |
1251 | * records, never the opposite (missing a full TSC record when it would | |
1252 | * be needed). | |
1253 | */ | |
1254 | save_last_tsc(config, buf, tsc); | |
1255 | ||
1256 | /* | |
1257 | * Push the reader if necessary | |
1258 | */ | |
1259 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
1260 | ||
1261 | oldidx = subbuf_index(offsets.old, chan); | |
1d498196 | 1262 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx, handle); |
852c2936 MD |
1263 | |
1264 | /* | |
1265 | * May need to populate header start on SWITCH_FLUSH. | |
1266 | */ | |
1267 | if (offsets.switch_old_start) { | |
1d498196 | 1268 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc, handle); |
852c2936 MD |
1269 | offsets.old += config->cb.subbuffer_header_size(); |
1270 | } | |
1271 | ||
1272 | /* | |
1273 | * Switch old subbuffer. | |
1274 | */ | |
1d498196 | 1275 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc, handle); |
852c2936 | 1276 | } |
852c2936 MD |
1277 | |
1278 | /* | |
1279 | * Returns : | |
1280 | * 0 if ok | |
1281 | * -ENOSPC if event size is too large for packet. | |
1282 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
1283 | * -EIO if data cannot be written into the buffer for any other reason. | |
1284 | */ | |
1285 | static | |
4cfec15c | 1286 | int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1287 | struct channel *chan, |
1288 | struct switch_offsets *offsets, | |
4cfec15c | 1289 | struct lttng_ust_lib_ring_buffer_ctx *ctx) |
852c2936 | 1290 | { |
4cfec15c | 1291 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
38fae1d3 | 1292 | struct lttng_ust_shm_handle *handle = ctx->handle; |
852c2936 MD |
1293 | unsigned long reserve_commit_diff; |
1294 | ||
1295 | offsets->begin = v_read(config, &buf->offset); | |
1296 | offsets->old = offsets->begin; | |
1297 | offsets->switch_new_start = 0; | |
1298 | offsets->switch_new_end = 0; | |
1299 | offsets->switch_old_end = 0; | |
1300 | offsets->pre_header_padding = 0; | |
1301 | ||
1302 | ctx->tsc = config->cb.ring_buffer_clock_read(chan); | |
1303 | if ((int64_t) ctx->tsc == -EIO) | |
1304 | return -EIO; | |
1305 | ||
1306 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
1307 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
1308 | ||
b5a3dfa5 | 1309 | if (caa_unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) { |
852c2936 MD |
1310 | offsets->switch_new_start = 1; /* For offsets->begin */ |
1311 | } else { | |
1312 | offsets->size = config->cb.record_header_size(config, chan, | |
1313 | offsets->begin, | |
1314 | &offsets->pre_header_padding, | |
1315 | ctx); | |
1316 | offsets->size += | |
1317 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1318 | ctx->largest_align) | |
1319 | + ctx->data_size; | |
b5a3dfa5 | 1320 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) + |
852c2936 MD |
1321 | offsets->size > chan->backend.subbuf_size)) { |
1322 | offsets->switch_old_end = 1; /* For offsets->old */ | |
1323 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1324 | } | |
1325 | } | |
b5a3dfa5 | 1326 | if (caa_unlikely(offsets->switch_new_start)) { |
852c2936 MD |
1327 | unsigned long sb_index; |
1328 | ||
1329 | /* | |
1330 | * We are typically not filling the previous buffer completely. | |
1331 | */ | |
b5a3dfa5 | 1332 | if (caa_likely(offsets->switch_old_end)) |
852c2936 MD |
1333 | offsets->begin = subbuf_align(offsets->begin, chan); |
1334 | offsets->begin = offsets->begin | |
1335 | + config->cb.subbuffer_header_size(); | |
1336 | /* Test new buffer integrity */ | |
1337 | sb_index = subbuf_index(offsets->begin, chan); | |
1338 | reserve_commit_diff = | |
1339 | (buf_trunc(offsets->begin, chan) | |
1340 | >> chan->backend.num_subbuf_order) | |
1341 | - ((unsigned long) v_read(config, | |
4746ae29 | 1342 | &shmp_index(handle, buf->commit_cold, sb_index)->cc_sb) |
852c2936 | 1343 | & chan->commit_count_mask); |
b5a3dfa5 | 1344 | if (caa_likely(reserve_commit_diff == 0)) { |
852c2936 | 1345 | /* Next subbuffer not being written to. */ |
b5a3dfa5 | 1346 | if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE && |
852c2936 MD |
1347 | subbuf_trunc(offsets->begin, chan) |
1348 | - subbuf_trunc((unsigned long) | |
a6352fd4 | 1349 | uatomic_read(&buf->consumed), chan) |
852c2936 | 1350 | >= chan->backend.buf_size)) { |
64493e4f MD |
1351 | unsigned long nr_lost; |
1352 | ||
852c2936 MD |
1353 | /* |
1354 | * We do not overwrite non consumed buffers | |
1355 | * and we are full : record is lost. | |
1356 | */ | |
64493e4f | 1357 | nr_lost = v_read(config, &buf->records_lost_full); |
852c2936 | 1358 | v_inc(config, &buf->records_lost_full); |
64493e4f MD |
1359 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
1360 | DBG("%lu or more records lost in (%s:%d) (buffer full)\n", | |
1361 | nr_lost + 1, chan->backend.name, | |
1362 | buf->backend.cpu); | |
1363 | } | |
852c2936 MD |
1364 | return -ENOBUFS; |
1365 | } else { | |
1366 | /* | |
1367 | * Next subbuffer not being written to, and we | |
1368 | * are either in overwrite mode or the buffer is | |
1369 | * not full. It's safe to write in this new | |
1370 | * subbuffer. | |
1371 | */ | |
1372 | } | |
1373 | } else { | |
64493e4f MD |
1374 | unsigned long nr_lost; |
1375 | ||
852c2936 MD |
1376 | /* |
1377 | * Next subbuffer reserve offset does not match the | |
1378 | * commit offset. Drop record in producer-consumer and | |
1379 | * overwrite mode. Caused by either a writer OOPS or too | |
1380 | * many nested writes over a reserve/commit pair. | |
1381 | */ | |
64493e4f | 1382 | nr_lost = v_read(config, &buf->records_lost_wrap); |
852c2936 | 1383 | v_inc(config, &buf->records_lost_wrap); |
64493e4f MD |
1384 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
1385 | DBG("%lu or more records lost in (%s:%d) (wrap-around)\n", | |
1386 | nr_lost + 1, chan->backend.name, | |
1387 | buf->backend.cpu); | |
1388 | } | |
852c2936 MD |
1389 | return -EIO; |
1390 | } | |
1391 | offsets->size = | |
1392 | config->cb.record_header_size(config, chan, | |
1393 | offsets->begin, | |
1394 | &offsets->pre_header_padding, | |
1395 | ctx); | |
1396 | offsets->size += | |
1397 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1398 | ctx->largest_align) | |
1399 | + ctx->data_size; | |
b5a3dfa5 | 1400 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) |
852c2936 | 1401 | + offsets->size > chan->backend.subbuf_size)) { |
64493e4f MD |
1402 | unsigned long nr_lost; |
1403 | ||
852c2936 MD |
1404 | /* |
1405 | * Record too big for subbuffers, report error, don't | |
1406 | * complete the sub-buffer switch. | |
1407 | */ | |
64493e4f | 1408 | nr_lost = v_read(config, &buf->records_lost_big); |
852c2936 | 1409 | v_inc(config, &buf->records_lost_big); |
64493e4f MD |
1410 | if ((nr_lost & (DBG_PRINT_NR_LOST - 1)) == 0) { |
1411 | DBG("%lu or more records lost in (%s:%d) record size " | |
1412 | " of %zu bytes is too large for buffer\n", | |
1413 | nr_lost + 1, chan->backend.name, | |
1414 | buf->backend.cpu, offsets->size); | |
1415 | } | |
852c2936 MD |
1416 | return -ENOSPC; |
1417 | } else { | |
1418 | /* | |
1419 | * We just made a successful buffer switch and the | |
1420 | * record fits in the new subbuffer. Let's write. | |
1421 | */ | |
1422 | } | |
1423 | } else { | |
1424 | /* | |
1425 | * Record fits in the current buffer and we are not on a switch | |
1426 | * boundary. It's safe to write. | |
1427 | */ | |
1428 | } | |
1429 | offsets->end = offsets->begin + offsets->size; | |
1430 | ||
b5a3dfa5 | 1431 | if (caa_unlikely(subbuf_offset(offsets->end, chan) == 0)) { |
852c2936 MD |
1432 | /* |
1433 | * The offset_end will fall at the very beginning of the next | |
1434 | * subbuffer. | |
1435 | */ | |
1436 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
1437 | } | |
1438 | return 0; | |
1439 | } | |
1440 | ||
1441 | /** | |
1442 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
1443 | * @ctx: ring buffer context. | |
1444 | * | |
1445 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, | |
1446 | * -EIO for other errors, else returns 0. | |
1447 | * It will take care of sub-buffer switching. | |
1448 | */ | |
4cfec15c | 1449 | int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx) |
852c2936 MD |
1450 | { |
1451 | struct channel *chan = ctx->chan; | |
38fae1d3 | 1452 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c MD |
1453 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
1454 | struct lttng_ust_lib_ring_buffer *buf; | |
852c2936 MD |
1455 | struct switch_offsets offsets; |
1456 | int ret; | |
1457 | ||
1458 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1d498196 | 1459 | buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp); |
852c2936 | 1460 | else |
1d498196 | 1461 | buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 MD |
1462 | ctx->buf = buf; |
1463 | ||
1464 | offsets.size = 0; | |
1465 | ||
1466 | do { | |
1467 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, | |
1468 | ctx); | |
b5a3dfa5 | 1469 | if (caa_unlikely(ret)) |
852c2936 | 1470 | return ret; |
b5a3dfa5 | 1471 | } while (caa_unlikely(v_cmpxchg(config, &buf->offset, offsets.old, |
852c2936 MD |
1472 | offsets.end) |
1473 | != offsets.old)); | |
1474 | ||
1475 | /* | |
1476 | * Atomically update last_tsc. This update races against concurrent | |
1477 | * atomic updates, but the race will always cause supplementary full TSC | |
1478 | * records, never the opposite (missing a full TSC record when it would | |
1479 | * be needed). | |
1480 | */ | |
1481 | save_last_tsc(config, buf, ctx->tsc); | |
1482 | ||
1483 | /* | |
1484 | * Push the reader if necessary | |
1485 | */ | |
1486 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
1487 | ||
1488 | /* | |
1489 | * Clear noref flag for this subbuffer. | |
1490 | */ | |
1491 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1d498196 MD |
1492 | subbuf_index(offsets.end - 1, chan), |
1493 | handle); | |
852c2936 MD |
1494 | |
1495 | /* | |
1496 | * Switch old subbuffer if needed. | |
1497 | */ | |
b5a3dfa5 | 1498 | if (caa_unlikely(offsets.switch_old_end)) { |
852c2936 | 1499 | lib_ring_buffer_clear_noref(config, &buf->backend, |
1d498196 MD |
1500 | subbuf_index(offsets.old - 1, chan), |
1501 | handle); | |
1502 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc, handle); | |
852c2936 MD |
1503 | } |
1504 | ||
1505 | /* | |
1506 | * Populate new subbuffer. | |
1507 | */ | |
b5a3dfa5 | 1508 | if (caa_unlikely(offsets.switch_new_start)) |
1d498196 | 1509 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 | 1510 | |
b5a3dfa5 | 1511 | if (caa_unlikely(offsets.switch_new_end)) |
1d498196 | 1512 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 MD |
1513 | |
1514 | ctx->slot_size = offsets.size; | |
1515 | ctx->pre_offset = offsets.begin; | |
1516 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
1517 | return 0; | |
1518 | } | |
d51652f7 MD |
1519 | |
1520 | /* | |
1521 | * Force a read (imply TLS fixup for dlopen) of TLS variables. | |
1522 | */ | |
1523 | void lttng_fixup_ringbuffer_tls(void) | |
1524 | { | |
1525 | asm volatile ("" : : "m" (lib_ring_buffer_nesting)); | |
1526 | } |