Commit | Line | Data |
---|---|---|
852c2936 MD |
1 | /* |
2 | * ring_buffer_frontend.c | |
3 | * | |
4 | * (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
5 | * | |
6 | * Ring buffer wait-free buffer synchronization. Producer-consumer and flight | |
7 | * recorder (overwrite) modes. See thesis: | |
8 | * | |
9 | * Desnoyers, Mathieu (2009), "Low-Impact Operating System Tracing", Ph.D. | |
10 | * dissertation, Ecole Polytechnique de Montreal. | |
11 | * http://www.lttng.org/pub/thesis/desnoyers-dissertation-2009-12.pdf | |
12 | * | |
13 | * - Algorithm presentation in Chapter 5: | |
14 | * "Lockless Multi-Core High-Throughput Buffering". | |
15 | * - Algorithm formal verification in Section 8.6: | |
16 | * "Formal verification of LTTng" | |
17 | * | |
18 | * Author: | |
19 | * Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
20 | * | |
21 | * Inspired from LTT and RelayFS: | |
22 | * Karim Yaghmour <karim@opersys.com> | |
23 | * Tom Zanussi <zanussi@us.ibm.com> | |
24 | * Bob Wisniewski <bob@watson.ibm.com> | |
25 | * And from K42 : | |
26 | * Bob Wisniewski <bob@watson.ibm.com> | |
27 | * | |
28 | * Buffer reader semantic : | |
29 | * | |
30 | * - get_subbuf_size | |
31 | * while buffer is not finalized and empty | |
32 | * - get_subbuf | |
33 | * - if return value != 0, continue | |
34 | * - splice one subbuffer worth of data to a pipe | |
35 | * - splice the data from pipe to disk/network | |
36 | * - put_subbuf | |
37 | * | |
38 | * Dual LGPL v2.1/GPL v2 license. | |
39 | */ | |
40 | ||
a6352fd4 | 41 | #include <sys/types.h> |
431d5cf0 MD |
42 | #include <sys/mman.h> |
43 | #include <sys/stat.h> | |
44 | #include <fcntl.h> | |
14641deb | 45 | #include <urcu/compiler.h> |
a6352fd4 | 46 | #include <urcu/ref.h> |
14641deb | 47 | |
a6352fd4 | 48 | #include "smp.h" |
4318ae1b | 49 | #include <lttng/ringbuffer-config.h> |
4931a13e MD |
50 | #include "backend.h" |
51 | #include "frontend.h" | |
a6352fd4 | 52 | #include "shm.h" |
852c2936 | 53 | |
431d5cf0 MD |
54 | #ifndef max |
55 | #define max(a, b) ((a) > (b) ? (a) : (b)) | |
56 | #endif | |
57 | ||
2432c3c9 MD |
58 | /* |
59 | * Use POSIX SHM: shm_open(3) and shm_unlink(3). | |
60 | * close(2) to close the fd returned by shm_open. | |
61 | * shm_unlink releases the shared memory object name. | |
62 | * ftruncate(2) sets the size of the memory object. | |
63 | * mmap/munmap maps the shared memory obj to a virtual address in the | |
64 | * calling proceess (should be done both in libust and consumer). | |
65 | * See shm_overview(7) for details. | |
66 | * Pass file descriptor returned by shm_open(3) to ltt-sessiond through | |
67 | * a UNIX socket. | |
68 | * | |
69 | * Since we don't need to access the object using its name, we can | |
70 | * immediately shm_unlink(3) it, and only keep the handle with its file | |
71 | * descriptor. | |
72 | */ | |
73 | ||
852c2936 MD |
74 | /* |
75 | * Internal structure representing offsets to use at a sub-buffer switch. | |
76 | */ | |
77 | struct switch_offsets { | |
78 | unsigned long begin, end, old; | |
79 | size_t pre_header_padding, size; | |
80 | unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1, | |
81 | switch_old_end:1; | |
82 | }; | |
83 | ||
a6352fd4 | 84 | __thread unsigned int lib_ring_buffer_nesting; |
852c2936 MD |
85 | |
86 | static | |
87 | void lib_ring_buffer_print_errors(struct channel *chan, | |
4cfec15c | 88 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
38fae1d3 | 89 | struct lttng_ust_shm_handle *handle); |
852c2936 MD |
90 | |
91 | /* | |
92 | * Must be called under cpu hotplug protection. | |
93 | */ | |
4cfec15c | 94 | void lib_ring_buffer_free(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 95 | struct lttng_ust_shm_handle *handle) |
852c2936 | 96 | { |
1d498196 | 97 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 | 98 | |
1d498196 | 99 | lib_ring_buffer_print_errors(chan, buf, buf->backend.cpu, handle); |
431d5cf0 MD |
100 | /* buf->commit_hot will be freed by shm teardown */ |
101 | /* buf->commit_cold will be freed by shm teardown */ | |
852c2936 MD |
102 | |
103 | lib_ring_buffer_backend_free(&buf->backend); | |
104 | } | |
105 | ||
106 | /** | |
107 | * lib_ring_buffer_reset - Reset ring buffer to initial values. | |
108 | * @buf: Ring buffer. | |
109 | * | |
110 | * Effectively empty the ring buffer. Should be called when the buffer is not | |
111 | * used for writing. The ring buffer can be opened for reading, but the reader | |
112 | * should not be using the iterator concurrently with reset. The previous | |
113 | * current iterator record is reset. | |
114 | */ | |
4cfec15c | 115 | void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 116 | struct lttng_ust_shm_handle *handle) |
852c2936 | 117 | { |
1d498196 | 118 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 119 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
120 | unsigned int i; |
121 | ||
122 | /* | |
123 | * Reset iterator first. It will put the subbuffer if it currently holds | |
124 | * it. | |
125 | */ | |
852c2936 MD |
126 | v_set(config, &buf->offset, 0); |
127 | for (i = 0; i < chan->backend.num_subbuf; i++) { | |
4746ae29 MD |
128 | v_set(config, &shmp_index(handle, buf->commit_hot, i)->cc, 0); |
129 | v_set(config, &shmp_index(handle, buf->commit_hot, i)->seq, 0); | |
130 | v_set(config, &shmp_index(handle, buf->commit_cold, i)->cc_sb, 0); | |
852c2936 | 131 | } |
a6352fd4 MD |
132 | uatomic_set(&buf->consumed, 0); |
133 | uatomic_set(&buf->record_disabled, 0); | |
852c2936 | 134 | v_set(config, &buf->last_tsc, 0); |
1d498196 | 135 | lib_ring_buffer_backend_reset(&buf->backend, handle); |
852c2936 MD |
136 | /* Don't reset number of active readers */ |
137 | v_set(config, &buf->records_lost_full, 0); | |
138 | v_set(config, &buf->records_lost_wrap, 0); | |
139 | v_set(config, &buf->records_lost_big, 0); | |
140 | v_set(config, &buf->records_count, 0); | |
141 | v_set(config, &buf->records_overrun, 0); | |
142 | buf->finalized = 0; | |
143 | } | |
852c2936 MD |
144 | |
145 | /** | |
146 | * channel_reset - Reset channel to initial values. | |
147 | * @chan: Channel. | |
148 | * | |
149 | * Effectively empty the channel. Should be called when the channel is not used | |
150 | * for writing. The channel can be opened for reading, but the reader should not | |
151 | * be using the iterator concurrently with reset. The previous current iterator | |
152 | * record is reset. | |
153 | */ | |
154 | void channel_reset(struct channel *chan) | |
155 | { | |
156 | /* | |
157 | * Reset iterators first. Will put the subbuffer if held for reading. | |
158 | */ | |
a6352fd4 | 159 | uatomic_set(&chan->record_disabled, 0); |
852c2936 MD |
160 | /* Don't reset commit_count_mask, still valid */ |
161 | channel_backend_reset(&chan->backend); | |
162 | /* Don't reset switch/read timer interval */ | |
163 | /* Don't reset notifiers and notifier enable bits */ | |
164 | /* Don't reset reader reference count */ | |
165 | } | |
852c2936 MD |
166 | |
167 | /* | |
168 | * Must be called under cpu hotplug protection. | |
169 | */ | |
4cfec15c | 170 | int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf, |
a6352fd4 | 171 | struct channel_backend *chanb, int cpu, |
38fae1d3 | 172 | struct lttng_ust_shm_handle *handle, |
1d498196 | 173 | struct shm_object *shmobj) |
852c2936 | 174 | { |
4cfec15c | 175 | const struct lttng_ust_lib_ring_buffer_config *config = &chanb->config; |
14641deb | 176 | struct channel *chan = caa_container_of(chanb, struct channel, backend); |
a3f61e7f | 177 | void *priv = channel_get_private(chan); |
852c2936 MD |
178 | unsigned int num_subbuf; |
179 | size_t subbuf_header_size; | |
180 | u64 tsc; | |
181 | int ret; | |
182 | ||
183 | /* Test for cpu hotplug */ | |
184 | if (buf->backend.allocated) | |
185 | return 0; | |
186 | ||
a6352fd4 | 187 | ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, |
1d498196 | 188 | cpu, handle, shmobj); |
852c2936 MD |
189 | if (ret) |
190 | return ret; | |
191 | ||
1d498196 MD |
192 | align_shm(shmobj, __alignof__(struct commit_counters_hot)); |
193 | set_shmp(buf->commit_hot, | |
194 | zalloc_shm(shmobj, | |
195 | sizeof(struct commit_counters_hot) * chan->backend.num_subbuf)); | |
196 | if (!shmp(handle, buf->commit_hot)) { | |
852c2936 MD |
197 | ret = -ENOMEM; |
198 | goto free_chanbuf; | |
199 | } | |
200 | ||
1d498196 MD |
201 | align_shm(shmobj, __alignof__(struct commit_counters_cold)); |
202 | set_shmp(buf->commit_cold, | |
203 | zalloc_shm(shmobj, | |
204 | sizeof(struct commit_counters_cold) * chan->backend.num_subbuf)); | |
205 | if (!shmp(handle, buf->commit_cold)) { | |
852c2936 MD |
206 | ret = -ENOMEM; |
207 | goto free_commit; | |
208 | } | |
209 | ||
210 | num_subbuf = chan->backend.num_subbuf; | |
a6352fd4 | 211 | //init_waitqueue_head(&buf->read_wait); |
852c2936 MD |
212 | |
213 | /* | |
214 | * Write the subbuffer header for first subbuffer so we know the total | |
215 | * duration of data gathering. | |
216 | */ | |
217 | subbuf_header_size = config->cb.subbuffer_header_size(); | |
218 | v_set(config, &buf->offset, subbuf_header_size); | |
4746ae29 | 219 | subbuffer_id_clear_noref(config, &shmp_index(handle, buf->backend.buf_wsb, 0)->id); |
1d498196 MD |
220 | tsc = config->cb.ring_buffer_clock_read(shmp(handle, buf->backend.chan)); |
221 | config->cb.buffer_begin(buf, tsc, 0, handle); | |
4746ae29 | 222 | v_add(config, subbuf_header_size, &shmp_index(handle, buf->commit_hot, 0)->cc); |
852c2936 MD |
223 | |
224 | if (config->cb.buffer_create) { | |
1d498196 | 225 | ret = config->cb.buffer_create(buf, priv, cpu, chanb->name, handle); |
852c2936 MD |
226 | if (ret) |
227 | goto free_init; | |
228 | } | |
852c2936 | 229 | buf->backend.allocated = 1; |
852c2936 MD |
230 | return 0; |
231 | ||
232 | /* Error handling */ | |
233 | free_init: | |
a6352fd4 | 234 | /* commit_cold will be freed by shm teardown */ |
852c2936 | 235 | free_commit: |
a6352fd4 | 236 | /* commit_hot will be freed by shm teardown */ |
852c2936 MD |
237 | free_chanbuf: |
238 | lib_ring_buffer_backend_free(&buf->backend); | |
239 | return ret; | |
240 | } | |
241 | ||
1d498196 | 242 | #if 0 |
852c2936 MD |
243 | static void switch_buffer_timer(unsigned long data) |
244 | { | |
4cfec15c | 245 | struct lttng_ust_lib_ring_buffer *buf = (struct lttng_ust_lib_ring_buffer *)data; |
1d498196 | 246 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 247 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
248 | |
249 | /* | |
250 | * Only flush buffers periodically if readers are active. | |
251 | */ | |
824f40b8 | 252 | if (uatomic_read(&buf->active_readers) || uatomic_read(&buf->active_shadow_readers)) |
1d498196 | 253 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE, handle); |
852c2936 | 254 | |
a6352fd4 MD |
255 | //TODO timers |
256 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
257 | // mod_timer_pinned(&buf->switch_timer, | |
258 | // jiffies + chan->switch_timer_interval); | |
259 | //else | |
260 | // mod_timer(&buf->switch_timer, | |
261 | // jiffies + chan->switch_timer_interval); | |
852c2936 | 262 | } |
1d498196 | 263 | #endif //0 |
852c2936 | 264 | |
4cfec15c | 265 | static void lib_ring_buffer_start_switch_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 266 | struct lttng_ust_shm_handle *handle) |
852c2936 | 267 | { |
1d498196 | 268 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 269 | //const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
270 | |
271 | if (!chan->switch_timer_interval || buf->switch_timer_enabled) | |
272 | return; | |
a6352fd4 MD |
273 | //TODO |
274 | //init_timer(&buf->switch_timer); | |
275 | //buf->switch_timer.function = switch_buffer_timer; | |
276 | //buf->switch_timer.expires = jiffies + chan->switch_timer_interval; | |
277 | //buf->switch_timer.data = (unsigned long)buf; | |
278 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
279 | // add_timer_on(&buf->switch_timer, buf->backend.cpu); | |
280 | //else | |
281 | // add_timer(&buf->switch_timer); | |
852c2936 MD |
282 | buf->switch_timer_enabled = 1; |
283 | } | |
284 | ||
4cfec15c | 285 | static void lib_ring_buffer_stop_switch_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 286 | struct lttng_ust_shm_handle *handle) |
852c2936 | 287 | { |
1d498196 | 288 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 MD |
289 | |
290 | if (!chan->switch_timer_interval || !buf->switch_timer_enabled) | |
291 | return; | |
292 | ||
a6352fd4 MD |
293 | //TODO |
294 | //del_timer_sync(&buf->switch_timer); | |
852c2936 MD |
295 | buf->switch_timer_enabled = 0; |
296 | } | |
297 | ||
1d498196 | 298 | #if 0 |
852c2936 MD |
299 | /* |
300 | * Polling timer to check the channels for data. | |
301 | */ | |
302 | static void read_buffer_timer(unsigned long data) | |
303 | { | |
4cfec15c | 304 | struct lttng_ust_lib_ring_buffer *buf = (struct lttng_ust_lib_ring_buffer *)data; |
1d498196 | 305 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 306 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
307 | |
308 | CHAN_WARN_ON(chan, !buf->backend.allocated); | |
309 | ||
824f40b8 | 310 | if (uatomic_read(&buf->active_readers) || uatomic_read(&buf->active_shadow_readers)) |
852c2936 | 311 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { |
a6352fd4 MD |
312 | //TODO |
313 | //wake_up_interruptible(&buf->read_wait); | |
314 | //wake_up_interruptible(&chan->read_wait); | |
852c2936 MD |
315 | } |
316 | ||
a6352fd4 MD |
317 | //TODO |
318 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
319 | // mod_timer_pinned(&buf->read_timer, | |
320 | // jiffies + chan->read_timer_interval); | |
321 | //else | |
322 | // mod_timer(&buf->read_timer, | |
323 | // jiffies + chan->read_timer_interval); | |
852c2936 | 324 | } |
1d498196 | 325 | #endif //0 |
852c2936 | 326 | |
4cfec15c | 327 | static void lib_ring_buffer_start_read_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 328 | struct lttng_ust_shm_handle *handle) |
852c2936 | 329 | { |
1d498196 | 330 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 331 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
332 | |
333 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
334 | || !chan->read_timer_interval | |
335 | || buf->read_timer_enabled) | |
336 | return; | |
337 | ||
a6352fd4 MD |
338 | //TODO |
339 | //init_timer(&buf->read_timer); | |
340 | //buf->read_timer.function = read_buffer_timer; | |
341 | //buf->read_timer.expires = jiffies + chan->read_timer_interval; | |
342 | //buf->read_timer.data = (unsigned long)buf; | |
852c2936 | 343 | |
a6352fd4 MD |
344 | //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
345 | // add_timer_on(&buf->read_timer, buf->backend.cpu); | |
346 | //else | |
347 | // add_timer(&buf->read_timer); | |
852c2936 MD |
348 | buf->read_timer_enabled = 1; |
349 | } | |
350 | ||
4cfec15c | 351 | static void lib_ring_buffer_stop_read_timer(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 352 | struct lttng_ust_shm_handle *handle) |
852c2936 | 353 | { |
1d498196 | 354 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 355 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
356 | |
357 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
358 | || !chan->read_timer_interval | |
359 | || !buf->read_timer_enabled) | |
360 | return; | |
361 | ||
a6352fd4 MD |
362 | //TODO |
363 | //del_timer_sync(&buf->read_timer); | |
852c2936 MD |
364 | /* |
365 | * do one more check to catch data that has been written in the last | |
366 | * timer period. | |
367 | */ | |
1d498196 | 368 | if (lib_ring_buffer_poll_deliver(config, buf, chan, handle)) { |
a6352fd4 MD |
369 | //TODO |
370 | //wake_up_interruptible(&buf->read_wait); | |
371 | //wake_up_interruptible(&chan->read_wait); | |
852c2936 MD |
372 | } |
373 | buf->read_timer_enabled = 0; | |
374 | } | |
375 | ||
1d498196 | 376 | static void channel_unregister_notifiers(struct channel *chan, |
38fae1d3 | 377 | struct lttng_ust_shm_handle *handle) |
852c2936 | 378 | { |
4cfec15c | 379 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
380 | int cpu; |
381 | ||
852c2936 | 382 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
852c2936 | 383 | for_each_possible_cpu(cpu) { |
4cfec15c | 384 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
a6352fd4 | 385 | |
1d498196 MD |
386 | lib_ring_buffer_stop_switch_timer(buf, handle); |
387 | lib_ring_buffer_stop_read_timer(buf, handle); | |
852c2936 | 388 | } |
852c2936 | 389 | } else { |
4cfec15c | 390 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 391 | |
1d498196 MD |
392 | lib_ring_buffer_stop_switch_timer(buf, handle); |
393 | lib_ring_buffer_stop_read_timer(buf, handle); | |
852c2936 | 394 | } |
8d8a24c8 | 395 | //channel_backend_unregister_notifiers(&chan->backend); |
852c2936 MD |
396 | } |
397 | ||
38fae1d3 | 398 | static void channel_free(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 399 | int shadow) |
852c2936 | 400 | { |
824f40b8 MD |
401 | if (!shadow) |
402 | channel_backend_free(&chan->backend, handle); | |
431d5cf0 | 403 | /* chan is freed by shm teardown */ |
1d498196 MD |
404 | shm_object_table_destroy(handle->table); |
405 | free(handle); | |
852c2936 MD |
406 | } |
407 | ||
408 | /** | |
409 | * channel_create - Create channel. | |
410 | * @config: ring buffer instance configuration | |
411 | * @name: name of the channel | |
a3f61e7f MD |
412 | * @priv_data: ring buffer client private data area pointer (output) |
413 | * @priv_data_size: length, in bytes, of the private data area. | |
d028eddb | 414 | * @priv_data_init: initialization data for private data. |
852c2936 MD |
415 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous |
416 | * address mapping. It is used only by RING_BUFFER_STATIC | |
417 | * configuration. It can be set to NULL for other backends. | |
418 | * @subbuf_size: subbuffer size | |
419 | * @num_subbuf: number of subbuffers | |
420 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
421 | * padding to let readers get those sub-buffers. | |
422 | * Used for live streaming. | |
423 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
424 | * | |
425 | * Holds cpu hotplug. | |
426 | * Returns NULL on failure. | |
427 | */ | |
4cfec15c | 428 | struct lttng_ust_shm_handle *channel_create(const struct lttng_ust_lib_ring_buffer_config *config, |
a3f61e7f MD |
429 | const char *name, |
430 | void **priv_data, | |
431 | size_t priv_data_align, | |
432 | size_t priv_data_size, | |
d028eddb | 433 | void *priv_data_init, |
a3f61e7f | 434 | void *buf_addr, size_t subbuf_size, |
852c2936 | 435 | size_t num_subbuf, unsigned int switch_timer_interval, |
193183fb MD |
436 | unsigned int read_timer_interval, |
437 | int *shm_fd, int *wait_fd, uint64_t *memory_map_size) | |
852c2936 | 438 | { |
1d498196 | 439 | int ret, cpu; |
a3f61e7f | 440 | size_t shmsize, chansize; |
852c2936 | 441 | struct channel *chan; |
38fae1d3 | 442 | struct lttng_ust_shm_handle *handle; |
1d498196 | 443 | struct shm_object *shmobj; |
193183fb | 444 | struct shm_ref *ref; |
852c2936 MD |
445 | |
446 | if (lib_ring_buffer_check_config(config, switch_timer_interval, | |
447 | read_timer_interval)) | |
448 | return NULL; | |
449 | ||
38fae1d3 | 450 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
431d5cf0 MD |
451 | if (!handle) |
452 | return NULL; | |
453 | ||
1d498196 MD |
454 | /* Allocate table for channel + per-cpu buffers */ |
455 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
456 | if (!handle->table) | |
457 | goto error_table_alloc; | |
852c2936 | 458 | |
1d498196 MD |
459 | /* Calculate the shm allocation layout */ |
460 | shmsize = sizeof(struct channel); | |
c1fca457 | 461 | shmsize += offset_align(shmsize, __alignof__(struct lttng_ust_lib_ring_buffer_shmp)); |
1d498196 | 462 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
4cfec15c | 463 | shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp) * num_possible_cpus(); |
1d498196 | 464 | else |
4cfec15c | 465 | shmsize += sizeof(struct lttng_ust_lib_ring_buffer_shmp); |
a3f61e7f MD |
466 | chansize = shmsize; |
467 | shmsize += offset_align(shmsize, priv_data_align); | |
468 | shmsize += priv_data_size; | |
a6352fd4 | 469 | |
1d498196 | 470 | shmobj = shm_object_table_append(handle->table, shmsize); |
b5a14697 MD |
471 | if (!shmobj) |
472 | goto error_append; | |
57773204 | 473 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
a3f61e7f | 474 | set_shmp(handle->chan, zalloc_shm(shmobj, chansize)); |
57773204 MD |
475 | assert(handle->chan._ref.index == 0); |
476 | assert(handle->chan._ref.offset == 0); | |
1d498196 | 477 | chan = shmp(handle, handle->chan); |
a6352fd4 | 478 | if (!chan) |
1d498196 | 479 | goto error_append; |
a6352fd4 | 480 | |
a3f61e7f MD |
481 | /* space for private data */ |
482 | if (priv_data_size) { | |
483 | DECLARE_SHMP(void, priv_data_alloc); | |
484 | ||
485 | align_shm(shmobj, priv_data_align); | |
486 | chan->priv_data_offset = shmobj->allocated_len; | |
487 | set_shmp(priv_data_alloc, zalloc_shm(shmobj, priv_data_size)); | |
488 | if (!shmp(handle, priv_data_alloc)) | |
489 | goto error_append; | |
490 | *priv_data = channel_get_private(chan); | |
d028eddb | 491 | memcpy(*priv_data, priv_data_init, priv_data_size); |
a3f61e7f MD |
492 | } else { |
493 | chan->priv_data_offset = -1; | |
494 | *priv_data = NULL; | |
495 | } | |
496 | ||
497 | ret = channel_backend_init(&chan->backend, name, config, | |
1d498196 | 498 | subbuf_size, num_subbuf, handle); |
852c2936 | 499 | if (ret) |
1d498196 | 500 | goto error_backend_init; |
852c2936 MD |
501 | |
502 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); | |
a6352fd4 MD |
503 | //TODO |
504 | //chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval); | |
505 | //chan->read_timer_interval = usecs_to_jiffies(read_timer_interval); | |
a6352fd4 MD |
506 | //TODO |
507 | //init_waitqueue_head(&chan->read_wait); | |
508 | //init_waitqueue_head(&chan->hp_wait); | |
852c2936 MD |
509 | |
510 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
852c2936 MD |
511 | /* |
512 | * In case of non-hotplug cpu, if the ring-buffer is allocated | |
513 | * in early initcall, it will not be notified of secondary cpus. | |
514 | * In that off case, we need to allocate for all possible cpus. | |
515 | */ | |
852c2936 | 516 | for_each_possible_cpu(cpu) { |
4cfec15c | 517 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
1d498196 MD |
518 | lib_ring_buffer_start_switch_timer(buf, handle); |
519 | lib_ring_buffer_start_read_timer(buf, handle); | |
852c2936 | 520 | } |
852c2936 | 521 | } else { |
4cfec15c | 522 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 523 | |
1d498196 MD |
524 | lib_ring_buffer_start_switch_timer(buf, handle); |
525 | lib_ring_buffer_start_read_timer(buf, handle); | |
852c2936 | 526 | } |
193183fb MD |
527 | ref = &handle->chan._ref; |
528 | shm_get_object_data(handle, ref, shm_fd, wait_fd, memory_map_size); | |
431d5cf0 | 529 | return handle; |
852c2936 | 530 | |
1d498196 MD |
531 | error_backend_init: |
532 | error_append: | |
533 | shm_object_table_destroy(handle->table); | |
534 | error_table_alloc: | |
431d5cf0 | 535 | free(handle); |
852c2936 MD |
536 | return NULL; |
537 | } | |
852c2936 | 538 | |
38fae1d3 | 539 | struct lttng_ust_shm_handle *channel_handle_create(int shm_fd, int wait_fd, |
193183fb MD |
540 | uint64_t memory_map_size) |
541 | { | |
38fae1d3 | 542 | struct lttng_ust_shm_handle *handle; |
193183fb MD |
543 | struct shm_object *object; |
544 | ||
38fae1d3 | 545 | handle = zmalloc(sizeof(struct lttng_ust_shm_handle)); |
193183fb MD |
546 | if (!handle) |
547 | return NULL; | |
548 | ||
549 | /* Allocate table for channel + per-cpu buffers */ | |
550 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
551 | if (!handle->table) | |
552 | goto error_table_alloc; | |
553 | /* Add channel object */ | |
554 | object = shm_object_table_append_shadow(handle->table, | |
555 | shm_fd, wait_fd, memory_map_size); | |
556 | if (!object) | |
557 | goto error_table_object; | |
57773204 MD |
558 | /* struct channel is at object 0, offset 0 (hardcoded) */ |
559 | handle->chan._ref.index = 0; | |
560 | handle->chan._ref.offset = 0; | |
193183fb MD |
561 | return handle; |
562 | ||
563 | error_table_object: | |
564 | shm_object_table_destroy(handle->table); | |
565 | error_table_alloc: | |
566 | free(handle); | |
567 | return NULL; | |
568 | } | |
569 | ||
38fae1d3 | 570 | int channel_handle_add_stream(struct lttng_ust_shm_handle *handle, |
193183fb MD |
571 | int shm_fd, int wait_fd, uint64_t memory_map_size) |
572 | { | |
573 | struct shm_object *object; | |
574 | ||
575 | /* Add stream object */ | |
576 | object = shm_object_table_append_shadow(handle->table, | |
577 | shm_fd, wait_fd, memory_map_size); | |
578 | if (!object) | |
579 | return -1; | |
580 | return 0; | |
581 | } | |
582 | ||
852c2936 | 583 | static |
38fae1d3 | 584 | void channel_release(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 585 | int shadow) |
852c2936 | 586 | { |
824f40b8 | 587 | channel_free(chan, handle, shadow); |
852c2936 MD |
588 | } |
589 | ||
590 | /** | |
591 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
592 | * @chan: channel to destroy | |
593 | * | |
594 | * Holds cpu hotplug. | |
431d5cf0 MD |
595 | * Call "destroy" callback, finalize channels, decrement the channel |
596 | * reference count. Note that when readers have completed data | |
597 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
a3f61e7f | 598 | * They should release their handle at that point. |
852c2936 | 599 | */ |
a3f61e7f | 600 | void channel_destroy(struct channel *chan, struct lttng_ust_shm_handle *handle, |
824f40b8 | 601 | int shadow) |
852c2936 | 602 | { |
4cfec15c | 603 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
431d5cf0 | 604 | int cpu; |
852c2936 | 605 | |
824f40b8 MD |
606 | if (shadow) { |
607 | channel_release(chan, handle, shadow); | |
a3f61e7f | 608 | return; |
824f40b8 MD |
609 | } |
610 | ||
1d498196 | 611 | channel_unregister_notifiers(chan, handle); |
852c2936 MD |
612 | |
613 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
852c2936 | 614 | for_each_channel_cpu(cpu, chan) { |
4cfec15c | 615 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
852c2936 MD |
616 | |
617 | if (config->cb.buffer_finalize) | |
618 | config->cb.buffer_finalize(buf, | |
a3f61e7f | 619 | channel_get_private(chan), |
1d498196 | 620 | cpu, handle); |
852c2936 | 621 | if (buf->backend.allocated) |
1d498196 MD |
622 | lib_ring_buffer_switch_slow(buf, SWITCH_FLUSH, |
623 | handle); | |
852c2936 MD |
624 | /* |
625 | * Perform flush before writing to finalized. | |
626 | */ | |
a6352fd4 | 627 | cmm_smp_wmb(); |
14641deb | 628 | CMM_ACCESS_ONCE(buf->finalized) = 1; |
a6352fd4 | 629 | //wake_up_interruptible(&buf->read_wait); |
852c2936 MD |
630 | } |
631 | } else { | |
4cfec15c | 632 | struct lttng_ust_lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 MD |
633 | |
634 | if (config->cb.buffer_finalize) | |
a3f61e7f | 635 | config->cb.buffer_finalize(buf, channel_get_private(chan), -1, handle); |
852c2936 | 636 | if (buf->backend.allocated) |
1d498196 MD |
637 | lib_ring_buffer_switch_slow(buf, SWITCH_FLUSH, |
638 | handle); | |
852c2936 MD |
639 | /* |
640 | * Perform flush before writing to finalized. | |
641 | */ | |
a6352fd4 | 642 | cmm_smp_wmb(); |
14641deb | 643 | CMM_ACCESS_ONCE(buf->finalized) = 1; |
a6352fd4 | 644 | //wake_up_interruptible(&buf->read_wait); |
852c2936 | 645 | } |
14641deb | 646 | CMM_ACCESS_ONCE(chan->finalized) = 1; |
a6352fd4 MD |
647 | //wake_up_interruptible(&chan->hp_wait); |
648 | //wake_up_interruptible(&chan->read_wait); | |
431d5cf0 MD |
649 | /* |
650 | * sessiond/consumer are keeping a reference on the shm file | |
651 | * descriptor directly. No need to refcount. | |
652 | */ | |
824f40b8 | 653 | channel_release(chan, handle, shadow); |
a3f61e7f | 654 | return; |
852c2936 | 655 | } |
852c2936 | 656 | |
4cfec15c MD |
657 | struct lttng_ust_lib_ring_buffer *channel_get_ring_buffer( |
658 | const struct lttng_ust_lib_ring_buffer_config *config, | |
1d498196 | 659 | struct channel *chan, int cpu, |
38fae1d3 | 660 | struct lttng_ust_shm_handle *handle, |
381c0f1e MD |
661 | int *shm_fd, int *wait_fd, |
662 | uint64_t *memory_map_size) | |
852c2936 | 663 | { |
381c0f1e MD |
664 | struct shm_ref *ref; |
665 | ||
666 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
667 | ref = &chan->backend.buf[0].shmp._ref; | |
668 | shm_get_object_data(handle, ref, shm_fd, wait_fd, | |
669 | memory_map_size); | |
1d498196 | 670 | return shmp(handle, chan->backend.buf[0].shmp); |
381c0f1e | 671 | } else { |
e095d803 MD |
672 | if (cpu >= num_possible_cpus()) |
673 | return NULL; | |
381c0f1e MD |
674 | ref = &chan->backend.buf[cpu].shmp._ref; |
675 | shm_get_object_data(handle, ref, shm_fd, wait_fd, | |
676 | memory_map_size); | |
1d498196 | 677 | return shmp(handle, chan->backend.buf[cpu].shmp); |
381c0f1e | 678 | } |
852c2936 | 679 | } |
852c2936 | 680 | |
4cfec15c | 681 | int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 682 | struct lttng_ust_shm_handle *handle, |
824f40b8 | 683 | int shadow) |
852c2936 | 684 | { |
824f40b8 MD |
685 | if (shadow) { |
686 | if (uatomic_cmpxchg(&buf->active_shadow_readers, 0, 1) != 0) | |
687 | return -EBUSY; | |
688 | cmm_smp_mb(); | |
689 | return 0; | |
690 | } | |
a6352fd4 | 691 | if (uatomic_cmpxchg(&buf->active_readers, 0, 1) != 0) |
852c2936 | 692 | return -EBUSY; |
a6352fd4 | 693 | cmm_smp_mb(); |
852c2936 MD |
694 | return 0; |
695 | } | |
852c2936 | 696 | |
4cfec15c | 697 | void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 698 | struct lttng_ust_shm_handle *handle, |
824f40b8 | 699 | int shadow) |
852c2936 | 700 | { |
1d498196 | 701 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 | 702 | |
824f40b8 MD |
703 | if (shadow) { |
704 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_shadow_readers) != 1); | |
705 | cmm_smp_mb(); | |
706 | uatomic_dec(&buf->active_shadow_readers); | |
707 | return; | |
708 | } | |
a6352fd4 MD |
709 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
710 | cmm_smp_mb(); | |
711 | uatomic_dec(&buf->active_readers); | |
852c2936 MD |
712 | } |
713 | ||
714 | /** | |
715 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
716 | * @buf: ring buffer | |
717 | * @consumed: consumed count indicating the position where to read | |
718 | * @produced: produced count, indicates position when to stop reading | |
719 | * | |
720 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
721 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 MD |
722 | */ |
723 | ||
4cfec15c | 724 | int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 725 | unsigned long *consumed, unsigned long *produced, |
38fae1d3 | 726 | struct lttng_ust_shm_handle *handle) |
852c2936 | 727 | { |
1d498196 | 728 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 729 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
730 | unsigned long consumed_cur, write_offset; |
731 | int finalized; | |
732 | ||
14641deb | 733 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
734 | /* |
735 | * Read finalized before counters. | |
736 | */ | |
a6352fd4 MD |
737 | cmm_smp_rmb(); |
738 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 MD |
739 | /* |
740 | * No need to issue a memory barrier between consumed count read and | |
741 | * write offset read, because consumed count can only change | |
742 | * concurrently in overwrite mode, and we keep a sequence counter | |
743 | * identifier derived from the write offset to check we are getting | |
744 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
745 | * "tagged" upon writes, tags are checked upon read). | |
746 | */ | |
747 | write_offset = v_read(config, &buf->offset); | |
748 | ||
749 | /* | |
750 | * Check that we are not about to read the same subbuffer in | |
751 | * which the writer head is. | |
752 | */ | |
753 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
754 | == 0) | |
755 | goto nodata; | |
756 | ||
757 | *consumed = consumed_cur; | |
758 | *produced = subbuf_trunc(write_offset, chan); | |
759 | ||
760 | return 0; | |
761 | ||
762 | nodata: | |
763 | /* | |
764 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
765 | * of "raw_spin_is_locked" memory ordering. | |
766 | */ | |
767 | if (finalized) | |
768 | return -ENODATA; | |
852c2936 MD |
769 | else |
770 | return -EAGAIN; | |
771 | } | |
852c2936 MD |
772 | |
773 | /** | |
774 | * lib_ring_buffer_put_snapshot - move consumed counter forward | |
775 | * @buf: ring buffer | |
776 | * @consumed_new: new consumed count value | |
777 | */ | |
4cfec15c | 778 | void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 779 | unsigned long consumed_new, |
38fae1d3 | 780 | struct lttng_ust_shm_handle *handle) |
852c2936 | 781 | { |
4cfec15c | 782 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
1d498196 | 783 | struct channel *chan = shmp(handle, bufb->chan); |
852c2936 MD |
784 | unsigned long consumed; |
785 | ||
824f40b8 MD |
786 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1 |
787 | && uatomic_read(&buf->active_shadow_readers) != 1); | |
852c2936 MD |
788 | |
789 | /* | |
790 | * Only push the consumed value forward. | |
791 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
792 | * the writer in flight recorder mode. | |
793 | */ | |
a6352fd4 | 794 | consumed = uatomic_read(&buf->consumed); |
852c2936 | 795 | while ((long) consumed - (long) consumed_new < 0) |
a6352fd4 MD |
796 | consumed = uatomic_cmpxchg(&buf->consumed, consumed, |
797 | consumed_new); | |
852c2936 | 798 | } |
852c2936 MD |
799 | |
800 | /** | |
801 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
802 | * @buf: ring buffer | |
803 | * @consumed: consumed count indicating the position where to read | |
804 | * | |
805 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
806 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 | 807 | */ |
4cfec15c | 808 | int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 809 | unsigned long consumed, |
38fae1d3 | 810 | struct lttng_ust_shm_handle *handle) |
852c2936 | 811 | { |
1d498196 | 812 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 813 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
814 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
815 | int ret; | |
816 | int finalized; | |
817 | ||
818 | retry: | |
14641deb | 819 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
820 | /* |
821 | * Read finalized before counters. | |
822 | */ | |
a6352fd4 MD |
823 | cmm_smp_rmb(); |
824 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 | 825 | consumed_idx = subbuf_index(consumed, chan); |
4746ae29 | 826 | commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb); |
852c2936 MD |
827 | /* |
828 | * Make sure we read the commit count before reading the buffer | |
829 | * data and the write offset. Correct consumed offset ordering | |
830 | * wrt commit count is insured by the use of cmpxchg to update | |
831 | * the consumed offset. | |
852c2936 | 832 | */ |
a6352fd4 MD |
833 | /* |
834 | * Local rmb to match the remote wmb to read the commit count | |
835 | * before the buffer data and the write offset. | |
836 | */ | |
837 | cmm_smp_rmb(); | |
852c2936 MD |
838 | |
839 | write_offset = v_read(config, &buf->offset); | |
840 | ||
841 | /* | |
842 | * Check that the buffer we are getting is after or at consumed_cur | |
843 | * position. | |
844 | */ | |
845 | if ((long) subbuf_trunc(consumed, chan) | |
846 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
847 | goto nodata; | |
848 | ||
849 | /* | |
850 | * Check that the subbuffer we are trying to consume has been | |
851 | * already fully committed. | |
852 | */ | |
853 | if (((commit_count - chan->backend.subbuf_size) | |
854 | & chan->commit_count_mask) | |
855 | - (buf_trunc(consumed_cur, chan) | |
856 | >> chan->backend.num_subbuf_order) | |
857 | != 0) | |
858 | goto nodata; | |
859 | ||
860 | /* | |
861 | * Check that we are not about to read the same subbuffer in | |
862 | * which the writer head is. | |
863 | */ | |
864 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
865 | == 0) | |
866 | goto nodata; | |
867 | ||
868 | /* | |
869 | * Failure to get the subbuffer causes a busy-loop retry without going | |
870 | * to a wait queue. These are caused by short-lived race windows where | |
871 | * the writer is getting access to a subbuffer we were trying to get | |
872 | * access to. Also checks that the "consumed" buffer count we are | |
873 | * looking for matches the one contained in the subbuffer id. | |
874 | */ | |
875 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
876 | consumed_idx, buf_trunc_val(consumed, chan), |
877 | handle); | |
852c2936 MD |
878 | if (ret) |
879 | goto retry; | |
880 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); | |
881 | ||
882 | buf->get_subbuf_consumed = consumed; | |
883 | buf->get_subbuf = 1; | |
884 | ||
885 | return 0; | |
886 | ||
887 | nodata: | |
888 | /* | |
889 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
890 | * of "raw_spin_is_locked" memory ordering. | |
891 | */ | |
892 | if (finalized) | |
893 | return -ENODATA; | |
852c2936 MD |
894 | else |
895 | return -EAGAIN; | |
896 | } | |
852c2936 MD |
897 | |
898 | /** | |
899 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
900 | * @buf: ring buffer | |
901 | */ | |
4cfec15c | 902 | void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 903 | struct lttng_ust_shm_handle *handle) |
852c2936 | 904 | { |
4cfec15c | 905 | struct lttng_ust_lib_ring_buffer_backend *bufb = &buf->backend; |
1d498196 | 906 | struct channel *chan = shmp(handle, bufb->chan); |
4cfec15c | 907 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
908 | unsigned long read_sb_bindex, consumed_idx, consumed; |
909 | ||
824f40b8 MD |
910 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1 |
911 | && uatomic_read(&buf->active_shadow_readers) != 1); | |
852c2936 MD |
912 | |
913 | if (!buf->get_subbuf) { | |
914 | /* | |
915 | * Reader puts a subbuffer it did not get. | |
916 | */ | |
917 | CHAN_WARN_ON(chan, 1); | |
918 | return; | |
919 | } | |
920 | consumed = buf->get_subbuf_consumed; | |
921 | buf->get_subbuf = 0; | |
922 | ||
923 | /* | |
924 | * Clear the records_unread counter. (overruns counter) | |
925 | * Can still be non-zero if a file reader simply grabbed the data | |
926 | * without using iterators. | |
927 | * Can be below zero if an iterator is used on a snapshot more than | |
928 | * once. | |
929 | */ | |
930 | read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); | |
931 | v_add(config, v_read(config, | |
4746ae29 | 932 | &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread), |
852c2936 | 933 | &bufb->records_read); |
4746ae29 | 934 | v_set(config, &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread, 0); |
852c2936 MD |
935 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE |
936 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
937 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
938 | ||
939 | /* | |
940 | * Exchange the reader subbuffer with the one we put in its place in the | |
941 | * writer subbuffer table. Expect the original consumed count. If | |
942 | * update_read_sb_index fails, this is because the writer updated the | |
943 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
944 | * currently have: it has become invalid to try reading this sub-buffer | |
945 | * consumed count value anyway. | |
946 | */ | |
947 | consumed_idx = subbuf_index(consumed, chan); | |
948 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
949 | consumed_idx, buf_trunc_val(consumed, chan), |
950 | handle); | |
852c2936 MD |
951 | /* |
952 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
953 | * if the writer concurrently updated it. | |
954 | */ | |
955 | } | |
852c2936 MD |
956 | |
957 | /* | |
958 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
959 | * position and the writer position. (inclusive) | |
960 | */ | |
961 | static | |
4cfec15c | 962 | void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
963 | struct channel *chan, |
964 | unsigned long cons_offset, | |
1d498196 | 965 | int cpu, |
38fae1d3 | 966 | struct lttng_ust_shm_handle *handle) |
852c2936 | 967 | { |
4cfec15c | 968 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
969 | unsigned long cons_idx, commit_count, commit_count_sb; |
970 | ||
971 | cons_idx = subbuf_index(cons_offset, chan); | |
4746ae29 MD |
972 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, cons_idx)->cc); |
973 | commit_count_sb = v_read(config, &shmp_index(handle, buf->commit_cold, cons_idx)->cc_sb); | |
852c2936 MD |
974 | |
975 | if (subbuf_offset(commit_count, chan) != 0) | |
4d3c9523 | 976 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
977 | "commit count in subbuffer %lu,\n" |
978 | "expecting multiples of %lu bytes\n" | |
979 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
980 | chan->backend.name, cpu, cons_idx, | |
981 | chan->backend.subbuf_size, | |
982 | commit_count, commit_count_sb); | |
983 | ||
4d3c9523 | 984 | DBG("ring buffer: %s, cpu %d: %lu bytes committed\n", |
852c2936 MD |
985 | chan->backend.name, cpu, commit_count); |
986 | } | |
987 | ||
988 | static | |
4cfec15c | 989 | void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 | 990 | struct channel *chan, |
1d498196 | 991 | void *priv, int cpu, |
38fae1d3 | 992 | struct lttng_ust_shm_handle *handle) |
852c2936 | 993 | { |
4cfec15c | 994 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
995 | unsigned long write_offset, cons_offset; |
996 | ||
997 | /* | |
998 | * Can be called in the error path of allocation when | |
999 | * trans_channel_data is not yet set. | |
1000 | */ | |
1001 | if (!chan) | |
1002 | return; | |
1003 | /* | |
1004 | * No need to order commit_count, write_offset and cons_offset reads | |
1005 | * because we execute at teardown when no more writer nor reader | |
1006 | * references are left. | |
1007 | */ | |
1008 | write_offset = v_read(config, &buf->offset); | |
a6352fd4 | 1009 | cons_offset = uatomic_read(&buf->consumed); |
852c2936 | 1010 | if (write_offset != cons_offset) |
4d3c9523 | 1011 | DBG("ring buffer %s, cpu %d: " |
852c2936 MD |
1012 | "non-consumed data\n" |
1013 | " [ %lu bytes written, %lu bytes read ]\n", | |
1014 | chan->backend.name, cpu, write_offset, cons_offset); | |
1015 | ||
a6352fd4 | 1016 | for (cons_offset = uatomic_read(&buf->consumed); |
852c2936 MD |
1017 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), |
1018 | chan) | |
1019 | - cons_offset) > 0; | |
1020 | cons_offset = subbuf_align(cons_offset, chan)) | |
1021 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1d498196 | 1022 | cpu, handle); |
852c2936 MD |
1023 | } |
1024 | ||
1025 | static | |
1026 | void lib_ring_buffer_print_errors(struct channel *chan, | |
4cfec15c | 1027 | struct lttng_ust_lib_ring_buffer *buf, int cpu, |
38fae1d3 | 1028 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1029 | { |
4cfec15c | 1030 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
a3f61e7f | 1031 | void *priv = channel_get_private(chan); |
852c2936 | 1032 | |
4d3c9523 | 1033 | DBG("ring buffer %s, cpu %d: %lu records written, " |
852c2936 MD |
1034 | "%lu records overrun\n", |
1035 | chan->backend.name, cpu, | |
1036 | v_read(config, &buf->records_count), | |
1037 | v_read(config, &buf->records_overrun)); | |
1038 | ||
1039 | if (v_read(config, &buf->records_lost_full) | |
1040 | || v_read(config, &buf->records_lost_wrap) | |
1041 | || v_read(config, &buf->records_lost_big)) | |
4d3c9523 | 1042 | DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n" |
852c2936 MD |
1043 | " [ %lu buffer full, %lu nest buffer wrap-around, " |
1044 | "%lu event too big ]\n", | |
1045 | chan->backend.name, cpu, | |
1046 | v_read(config, &buf->records_lost_full), | |
1047 | v_read(config, &buf->records_lost_wrap), | |
1048 | v_read(config, &buf->records_lost_big)); | |
1049 | ||
1d498196 | 1050 | lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu, handle); |
852c2936 MD |
1051 | } |
1052 | ||
1053 | /* | |
1054 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
1055 | * | |
1056 | * Only executed when the buffer is finalized, in SWITCH_FLUSH. | |
1057 | */ | |
1058 | static | |
4cfec15c | 1059 | void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1060 | struct channel *chan, |
1061 | struct switch_offsets *offsets, | |
1d498196 | 1062 | u64 tsc, |
38fae1d3 | 1063 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1064 | { |
4cfec15c | 1065 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1066 | unsigned long oldidx = subbuf_index(offsets->old, chan); |
1067 | unsigned long commit_count; | |
1068 | ||
1d498196 | 1069 | config->cb.buffer_begin(buf, tsc, oldidx, handle); |
852c2936 MD |
1070 | |
1071 | /* | |
1072 | * Order all writes to buffer before the commit count update that will | |
1073 | * determine that the subbuffer is full. | |
1074 | */ | |
a6352fd4 | 1075 | cmm_smp_wmb(); |
852c2936 | 1076 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1077 | &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1078 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 MD |
1079 | /* Check if the written buffer has to be delivered */ |
1080 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
1d498196 | 1081 | commit_count, oldidx, handle); |
852c2936 MD |
1082 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
1083 | offsets->old, commit_count, | |
1d498196 MD |
1084 | config->cb.subbuffer_header_size(), |
1085 | handle); | |
852c2936 MD |
1086 | } |
1087 | ||
1088 | /* | |
1089 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
1090 | * | |
1091 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
1092 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
1093 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
1094 | * subbuffer. | |
1095 | */ | |
1096 | static | |
4cfec15c | 1097 | void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1098 | struct channel *chan, |
1099 | struct switch_offsets *offsets, | |
1d498196 | 1100 | u64 tsc, |
38fae1d3 | 1101 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1102 | { |
4cfec15c | 1103 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1104 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); |
1105 | unsigned long commit_count, padding_size, data_size; | |
1106 | ||
1107 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1108 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1109 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size, |
1110 | handle); | |
852c2936 MD |
1111 | |
1112 | /* | |
1113 | * Order all writes to buffer before the commit count update that will | |
1114 | * determine that the subbuffer is full. | |
1115 | */ | |
a6352fd4 | 1116 | cmm_smp_wmb(); |
4746ae29 MD |
1117 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1118 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 | 1119 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, |
1d498196 | 1120 | commit_count, oldidx, handle); |
852c2936 MD |
1121 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
1122 | offsets->old, commit_count, | |
1d498196 | 1123 | padding_size, handle); |
852c2936 MD |
1124 | } |
1125 | ||
1126 | /* | |
1127 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1128 | * | |
1129 | * This code can be executed unordered : writers may already have written to the | |
1130 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1131 | * that this code is executed before the deliver of this sub-buffer. | |
1132 | */ | |
1133 | static | |
4cfec15c | 1134 | void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1135 | struct channel *chan, |
1136 | struct switch_offsets *offsets, | |
1d498196 | 1137 | u64 tsc, |
38fae1d3 | 1138 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1139 | { |
4cfec15c | 1140 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1141 | unsigned long beginidx = subbuf_index(offsets->begin, chan); |
1142 | unsigned long commit_count; | |
1143 | ||
1d498196 | 1144 | config->cb.buffer_begin(buf, tsc, beginidx, handle); |
852c2936 MD |
1145 | |
1146 | /* | |
1147 | * Order all writes to buffer before the commit count update that will | |
1148 | * determine that the subbuffer is full. | |
1149 | */ | |
a6352fd4 | 1150 | cmm_smp_wmb(); |
852c2936 | 1151 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1152 | &shmp_index(handle, buf->commit_hot, beginidx)->cc); |
1153 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, beginidx)->cc); | |
852c2936 MD |
1154 | /* Check if the written buffer has to be delivered */ |
1155 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
1d498196 | 1156 | commit_count, beginidx, handle); |
852c2936 MD |
1157 | lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx, |
1158 | offsets->begin, commit_count, | |
1d498196 MD |
1159 | config->cb.subbuffer_header_size(), |
1160 | handle); | |
852c2936 MD |
1161 | } |
1162 | ||
1163 | /* | |
1164 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1165 | * | |
1166 | * The only remaining threads could be the ones with pending commits. They will | |
1167 | * have to do the deliver themselves. | |
1168 | */ | |
1169 | static | |
4cfec15c | 1170 | void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 MD |
1171 | struct channel *chan, |
1172 | struct switch_offsets *offsets, | |
1173 | u64 tsc, | |
38fae1d3 | 1174 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1175 | { |
4cfec15c | 1176 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1177 | unsigned long endidx = subbuf_index(offsets->end - 1, chan); |
1178 | unsigned long commit_count, padding_size, data_size; | |
1179 | ||
1180 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; | |
1181 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1182 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size, |
1183 | handle); | |
852c2936 MD |
1184 | |
1185 | /* | |
1186 | * Order all writes to buffer before the commit count update that will | |
1187 | * determine that the subbuffer is full. | |
1188 | */ | |
a6352fd4 | 1189 | cmm_smp_wmb(); |
4746ae29 MD |
1190 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, endidx)->cc); |
1191 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc); | |
852c2936 | 1192 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->end - 1, |
1d498196 | 1193 | commit_count, endidx, handle); |
852c2936 MD |
1194 | lib_ring_buffer_write_commit_counter(config, buf, chan, endidx, |
1195 | offsets->end, commit_count, | |
1d498196 | 1196 | padding_size, handle); |
852c2936 MD |
1197 | } |
1198 | ||
1199 | /* | |
1200 | * Returns : | |
1201 | * 0 if ok | |
1202 | * !0 if execution must be aborted. | |
1203 | */ | |
1204 | static | |
1205 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
4cfec15c | 1206 | struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1207 | struct channel *chan, |
1208 | struct switch_offsets *offsets, | |
1209 | u64 *tsc) | |
1210 | { | |
4cfec15c | 1211 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1212 | unsigned long off; |
1213 | ||
1214 | offsets->begin = v_read(config, &buf->offset); | |
1215 | offsets->old = offsets->begin; | |
1216 | offsets->switch_old_start = 0; | |
1217 | off = subbuf_offset(offsets->begin, chan); | |
1218 | ||
1219 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1220 | ||
1221 | /* | |
1222 | * Ensure we flush the header of an empty subbuffer when doing the | |
1223 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1224 | * total data gathering duration even if there were no records saved | |
1225 | * after the last buffer switch. | |
1226 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1227 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1228 | * subbuffer header as appropriate. | |
1229 | * The next record that reserves space will be responsible for | |
1230 | * populating the following subbuffer header. We choose not to populate | |
1231 | * the next subbuffer header here because we want to be able to use | |
a6352fd4 MD |
1232 | * SWITCH_ACTIVE for periodical buffer flush, which must |
1233 | * guarantee that all the buffer content (records and header | |
1234 | * timestamps) are visible to the reader. This is required for | |
1235 | * quiescence guarantees for the fusion merge. | |
852c2936 MD |
1236 | */ |
1237 | if (mode == SWITCH_FLUSH || off > 0) { | |
b5a3dfa5 | 1238 | if (caa_unlikely(off == 0)) { |
852c2936 MD |
1239 | /* |
1240 | * The client does not save any header information. | |
1241 | * Don't switch empty subbuffer on finalize, because it | |
1242 | * is invalid to deliver a completely empty subbuffer. | |
1243 | */ | |
1244 | if (!config->cb.subbuffer_header_size()) | |
1245 | return -1; | |
1246 | /* | |
1247 | * Need to write the subbuffer start header on finalize. | |
1248 | */ | |
1249 | offsets->switch_old_start = 1; | |
1250 | } | |
1251 | offsets->begin = subbuf_align(offsets->begin, chan); | |
1252 | } else | |
1253 | return -1; /* we do not have to switch : buffer is empty */ | |
1254 | /* Note: old points to the next subbuf at offset 0 */ | |
1255 | offsets->end = offsets->begin; | |
1256 | return 0; | |
1257 | } | |
1258 | ||
1259 | /* | |
1260 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
1261 | * called while tracing is active with absolutely no lock held. | |
1262 | * | |
1263 | * Note, however, that as a v_cmpxchg is used for some atomic | |
1264 | * operations, this function must be called from the CPU which owns the buffer | |
1265 | * for a ACTIVE flush. | |
1266 | */ | |
4cfec15c | 1267 | void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode, |
38fae1d3 | 1268 | struct lttng_ust_shm_handle *handle) |
852c2936 | 1269 | { |
1d498196 | 1270 | struct channel *chan = shmp(handle, buf->backend.chan); |
4cfec15c | 1271 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1272 | struct switch_offsets offsets; |
1273 | unsigned long oldidx; | |
1274 | u64 tsc; | |
1275 | ||
1276 | offsets.size = 0; | |
1277 | ||
1278 | /* | |
1279 | * Perform retryable operations. | |
1280 | */ | |
1281 | do { | |
1282 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
1283 | &tsc)) | |
1284 | return; /* Switch not needed */ | |
1285 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
1286 | != offsets.old); | |
1287 | ||
1288 | /* | |
1289 | * Atomically update last_tsc. This update races against concurrent | |
1290 | * atomic updates, but the race will always cause supplementary full TSC | |
1291 | * records, never the opposite (missing a full TSC record when it would | |
1292 | * be needed). | |
1293 | */ | |
1294 | save_last_tsc(config, buf, tsc); | |
1295 | ||
1296 | /* | |
1297 | * Push the reader if necessary | |
1298 | */ | |
1299 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
1300 | ||
1301 | oldidx = subbuf_index(offsets.old, chan); | |
1d498196 | 1302 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx, handle); |
852c2936 MD |
1303 | |
1304 | /* | |
1305 | * May need to populate header start on SWITCH_FLUSH. | |
1306 | */ | |
1307 | if (offsets.switch_old_start) { | |
1d498196 | 1308 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc, handle); |
852c2936 MD |
1309 | offsets.old += config->cb.subbuffer_header_size(); |
1310 | } | |
1311 | ||
1312 | /* | |
1313 | * Switch old subbuffer. | |
1314 | */ | |
1d498196 | 1315 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc, handle); |
852c2936 | 1316 | } |
852c2936 MD |
1317 | |
1318 | /* | |
1319 | * Returns : | |
1320 | * 0 if ok | |
1321 | * -ENOSPC if event size is too large for packet. | |
1322 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
1323 | * -EIO if data cannot be written into the buffer for any other reason. | |
1324 | */ | |
1325 | static | |
4cfec15c | 1326 | int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
1327 | struct channel *chan, |
1328 | struct switch_offsets *offsets, | |
4cfec15c | 1329 | struct lttng_ust_lib_ring_buffer_ctx *ctx) |
852c2936 | 1330 | { |
4cfec15c | 1331 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
38fae1d3 | 1332 | struct lttng_ust_shm_handle *handle = ctx->handle; |
852c2936 MD |
1333 | unsigned long reserve_commit_diff; |
1334 | ||
1335 | offsets->begin = v_read(config, &buf->offset); | |
1336 | offsets->old = offsets->begin; | |
1337 | offsets->switch_new_start = 0; | |
1338 | offsets->switch_new_end = 0; | |
1339 | offsets->switch_old_end = 0; | |
1340 | offsets->pre_header_padding = 0; | |
1341 | ||
1342 | ctx->tsc = config->cb.ring_buffer_clock_read(chan); | |
1343 | if ((int64_t) ctx->tsc == -EIO) | |
1344 | return -EIO; | |
1345 | ||
1346 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
1347 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
1348 | ||
b5a3dfa5 | 1349 | if (caa_unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) { |
852c2936 MD |
1350 | offsets->switch_new_start = 1; /* For offsets->begin */ |
1351 | } else { | |
1352 | offsets->size = config->cb.record_header_size(config, chan, | |
1353 | offsets->begin, | |
1354 | &offsets->pre_header_padding, | |
1355 | ctx); | |
1356 | offsets->size += | |
1357 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1358 | ctx->largest_align) | |
1359 | + ctx->data_size; | |
b5a3dfa5 | 1360 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) + |
852c2936 MD |
1361 | offsets->size > chan->backend.subbuf_size)) { |
1362 | offsets->switch_old_end = 1; /* For offsets->old */ | |
1363 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1364 | } | |
1365 | } | |
b5a3dfa5 | 1366 | if (caa_unlikely(offsets->switch_new_start)) { |
852c2936 MD |
1367 | unsigned long sb_index; |
1368 | ||
1369 | /* | |
1370 | * We are typically not filling the previous buffer completely. | |
1371 | */ | |
b5a3dfa5 | 1372 | if (caa_likely(offsets->switch_old_end)) |
852c2936 MD |
1373 | offsets->begin = subbuf_align(offsets->begin, chan); |
1374 | offsets->begin = offsets->begin | |
1375 | + config->cb.subbuffer_header_size(); | |
1376 | /* Test new buffer integrity */ | |
1377 | sb_index = subbuf_index(offsets->begin, chan); | |
1378 | reserve_commit_diff = | |
1379 | (buf_trunc(offsets->begin, chan) | |
1380 | >> chan->backend.num_subbuf_order) | |
1381 | - ((unsigned long) v_read(config, | |
4746ae29 | 1382 | &shmp_index(handle, buf->commit_cold, sb_index)->cc_sb) |
852c2936 | 1383 | & chan->commit_count_mask); |
b5a3dfa5 | 1384 | if (caa_likely(reserve_commit_diff == 0)) { |
852c2936 | 1385 | /* Next subbuffer not being written to. */ |
b5a3dfa5 | 1386 | if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE && |
852c2936 MD |
1387 | subbuf_trunc(offsets->begin, chan) |
1388 | - subbuf_trunc((unsigned long) | |
a6352fd4 | 1389 | uatomic_read(&buf->consumed), chan) |
852c2936 MD |
1390 | >= chan->backend.buf_size)) { |
1391 | /* | |
1392 | * We do not overwrite non consumed buffers | |
1393 | * and we are full : record is lost. | |
1394 | */ | |
1395 | v_inc(config, &buf->records_lost_full); | |
1396 | return -ENOBUFS; | |
1397 | } else { | |
1398 | /* | |
1399 | * Next subbuffer not being written to, and we | |
1400 | * are either in overwrite mode or the buffer is | |
1401 | * not full. It's safe to write in this new | |
1402 | * subbuffer. | |
1403 | */ | |
1404 | } | |
1405 | } else { | |
1406 | /* | |
1407 | * Next subbuffer reserve offset does not match the | |
1408 | * commit offset. Drop record in producer-consumer and | |
1409 | * overwrite mode. Caused by either a writer OOPS or too | |
1410 | * many nested writes over a reserve/commit pair. | |
1411 | */ | |
1412 | v_inc(config, &buf->records_lost_wrap); | |
1413 | return -EIO; | |
1414 | } | |
1415 | offsets->size = | |
1416 | config->cb.record_header_size(config, chan, | |
1417 | offsets->begin, | |
1418 | &offsets->pre_header_padding, | |
1419 | ctx); | |
1420 | offsets->size += | |
1421 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1422 | ctx->largest_align) | |
1423 | + ctx->data_size; | |
b5a3dfa5 | 1424 | if (caa_unlikely(subbuf_offset(offsets->begin, chan) |
852c2936 MD |
1425 | + offsets->size > chan->backend.subbuf_size)) { |
1426 | /* | |
1427 | * Record too big for subbuffers, report error, don't | |
1428 | * complete the sub-buffer switch. | |
1429 | */ | |
1430 | v_inc(config, &buf->records_lost_big); | |
1431 | return -ENOSPC; | |
1432 | } else { | |
1433 | /* | |
1434 | * We just made a successful buffer switch and the | |
1435 | * record fits in the new subbuffer. Let's write. | |
1436 | */ | |
1437 | } | |
1438 | } else { | |
1439 | /* | |
1440 | * Record fits in the current buffer and we are not on a switch | |
1441 | * boundary. It's safe to write. | |
1442 | */ | |
1443 | } | |
1444 | offsets->end = offsets->begin + offsets->size; | |
1445 | ||
b5a3dfa5 | 1446 | if (caa_unlikely(subbuf_offset(offsets->end, chan) == 0)) { |
852c2936 MD |
1447 | /* |
1448 | * The offset_end will fall at the very beginning of the next | |
1449 | * subbuffer. | |
1450 | */ | |
1451 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
1452 | } | |
1453 | return 0; | |
1454 | } | |
1455 | ||
1456 | /** | |
1457 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
1458 | * @ctx: ring buffer context. | |
1459 | * | |
1460 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, | |
1461 | * -EIO for other errors, else returns 0. | |
1462 | * It will take care of sub-buffer switching. | |
1463 | */ | |
4cfec15c | 1464 | int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx) |
852c2936 MD |
1465 | { |
1466 | struct channel *chan = ctx->chan; | |
38fae1d3 | 1467 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c MD |
1468 | const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config; |
1469 | struct lttng_ust_lib_ring_buffer *buf; | |
852c2936 MD |
1470 | struct switch_offsets offsets; |
1471 | int ret; | |
1472 | ||
1473 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1d498196 | 1474 | buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp); |
852c2936 | 1475 | else |
1d498196 | 1476 | buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 MD |
1477 | ctx->buf = buf; |
1478 | ||
1479 | offsets.size = 0; | |
1480 | ||
1481 | do { | |
1482 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, | |
1483 | ctx); | |
b5a3dfa5 | 1484 | if (caa_unlikely(ret)) |
852c2936 | 1485 | return ret; |
b5a3dfa5 | 1486 | } while (caa_unlikely(v_cmpxchg(config, &buf->offset, offsets.old, |
852c2936 MD |
1487 | offsets.end) |
1488 | != offsets.old)); | |
1489 | ||
1490 | /* | |
1491 | * Atomically update last_tsc. This update races against concurrent | |
1492 | * atomic updates, but the race will always cause supplementary full TSC | |
1493 | * records, never the opposite (missing a full TSC record when it would | |
1494 | * be needed). | |
1495 | */ | |
1496 | save_last_tsc(config, buf, ctx->tsc); | |
1497 | ||
1498 | /* | |
1499 | * Push the reader if necessary | |
1500 | */ | |
1501 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
1502 | ||
1503 | /* | |
1504 | * Clear noref flag for this subbuffer. | |
1505 | */ | |
1506 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1d498196 MD |
1507 | subbuf_index(offsets.end - 1, chan), |
1508 | handle); | |
852c2936 MD |
1509 | |
1510 | /* | |
1511 | * Switch old subbuffer if needed. | |
1512 | */ | |
b5a3dfa5 | 1513 | if (caa_unlikely(offsets.switch_old_end)) { |
852c2936 | 1514 | lib_ring_buffer_clear_noref(config, &buf->backend, |
1d498196 MD |
1515 | subbuf_index(offsets.old - 1, chan), |
1516 | handle); | |
1517 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc, handle); | |
852c2936 MD |
1518 | } |
1519 | ||
1520 | /* | |
1521 | * Populate new subbuffer. | |
1522 | */ | |
b5a3dfa5 | 1523 | if (caa_unlikely(offsets.switch_new_start)) |
1d498196 | 1524 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 | 1525 | |
b5a3dfa5 | 1526 | if (caa_unlikely(offsets.switch_new_end)) |
1d498196 | 1527 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 MD |
1528 | |
1529 | ctx->slot_size = offsets.size; | |
1530 | ctx->pre_offset = offsets.begin; | |
1531 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
1532 | return 0; | |
1533 | } |