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" |
8d8a24c8 | 49 | #include <ust/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, | |
1d498196 MD |
88 | struct lib_ring_buffer *buf, int cpu, |
89 | struct shm_handle *handle); | |
852c2936 MD |
90 | |
91 | /* | |
92 | * Must be called under cpu hotplug protection. | |
93 | */ | |
1d498196 MD |
94 | void lib_ring_buffer_free(struct lib_ring_buffer *buf, |
95 | struct 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 | */ | |
1d498196 MD |
115 | void lib_ring_buffer_reset(struct lib_ring_buffer *buf, |
116 | struct shm_handle *handle) | |
852c2936 | 117 | { |
1d498196 | 118 | struct channel *chan = shmp(handle, buf->backend.chan); |
193183fb | 119 | const struct 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 | */ | |
170 | int lib_ring_buffer_create(struct lib_ring_buffer *buf, | |
a6352fd4 | 171 | struct channel_backend *chanb, int cpu, |
1d498196 MD |
172 | struct shm_handle *handle, |
173 | struct shm_object *shmobj) | |
852c2936 | 174 | { |
193183fb | 175 | const struct lib_ring_buffer_config *config = &chanb->config; |
14641deb | 176 | struct channel *chan = caa_container_of(chanb, struct channel, backend); |
852c2936 MD |
177 | void *priv = chanb->priv; |
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 | { | |
245 | struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data; | |
1d498196 | 246 | struct channel *chan = shmp(handle, buf->backend.chan); |
193183fb | 247 | const struct 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 | |
1d498196 MD |
265 | static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer *buf, |
266 | struct shm_handle *handle) | |
852c2936 | 267 | { |
1d498196 | 268 | struct channel *chan = shmp(handle, buf->backend.chan); |
193183fb | 269 | const struct 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 | ||
1d498196 MD |
285 | static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer *buf, |
286 | struct 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 | { | |
304 | struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data; | |
1d498196 | 305 | struct channel *chan = shmp(handle, buf->backend.chan); |
193183fb | 306 | const struct 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 | |
1d498196 MD |
327 | static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer *buf, |
328 | struct shm_handle *handle) | |
852c2936 | 329 | { |
1d498196 | 330 | struct channel *chan = shmp(handle, buf->backend.chan); |
193183fb | 331 | const struct 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 | ||
1d498196 MD |
351 | static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer *buf, |
352 | struct shm_handle *handle) | |
852c2936 | 353 | { |
1d498196 | 354 | struct channel *chan = shmp(handle, buf->backend.chan); |
193183fb | 355 | const struct 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 MD |
376 | static void channel_unregister_notifiers(struct channel *chan, |
377 | struct shm_handle *handle) | |
852c2936 | 378 | { |
193183fb | 379 | const struct 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) { |
1d498196 | 384 | struct 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 { |
1d498196 | 390 | struct 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 | ||
824f40b8 MD |
398 | static void channel_free(struct channel *chan, struct shm_handle *handle, |
399 | int shadow) | |
852c2936 | 400 | { |
431d5cf0 MD |
401 | int ret; |
402 | ||
824f40b8 MD |
403 | if (!shadow) |
404 | channel_backend_free(&chan->backend, handle); | |
431d5cf0 | 405 | /* chan is freed by shm teardown */ |
1d498196 MD |
406 | shm_object_table_destroy(handle->table); |
407 | free(handle); | |
852c2936 MD |
408 | } |
409 | ||
410 | /** | |
411 | * channel_create - Create channel. | |
412 | * @config: ring buffer instance configuration | |
413 | * @name: name of the channel | |
414 | * @priv: ring buffer client private data | |
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 | */ | |
431d5cf0 | 428 | struct shm_handle *channel_create(const struct lib_ring_buffer_config *config, |
852c2936 MD |
429 | const char *name, void *priv, void *buf_addr, |
430 | size_t subbuf_size, | |
431 | size_t num_subbuf, unsigned int switch_timer_interval, | |
193183fb MD |
432 | unsigned int read_timer_interval, |
433 | int *shm_fd, int *wait_fd, uint64_t *memory_map_size) | |
852c2936 | 434 | { |
1d498196 MD |
435 | int ret, cpu; |
436 | size_t shmsize; | |
852c2936 | 437 | struct channel *chan; |
431d5cf0 | 438 | struct shm_handle *handle; |
1d498196 | 439 | struct shm_object *shmobj; |
193183fb | 440 | struct shm_ref *ref; |
852c2936 MD |
441 | |
442 | if (lib_ring_buffer_check_config(config, switch_timer_interval, | |
443 | read_timer_interval)) | |
444 | return NULL; | |
445 | ||
431d5cf0 MD |
446 | handle = zmalloc(sizeof(struct shm_handle)); |
447 | if (!handle) | |
448 | return NULL; | |
449 | ||
1d498196 MD |
450 | /* Allocate table for channel + per-cpu buffers */ |
451 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
452 | if (!handle->table) | |
453 | goto error_table_alloc; | |
852c2936 | 454 | |
1d498196 MD |
455 | /* Calculate the shm allocation layout */ |
456 | shmsize = sizeof(struct channel); | |
457 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
458 | shmsize += sizeof(struct lib_ring_buffer_shmp) * num_possible_cpus(); | |
459 | else | |
460 | shmsize += sizeof(struct lib_ring_buffer_shmp); | |
a6352fd4 | 461 | |
1d498196 | 462 | shmobj = shm_object_table_append(handle->table, shmsize); |
b5a14697 MD |
463 | if (!shmobj) |
464 | goto error_append; | |
1d498196 MD |
465 | set_shmp(handle->chan, zalloc_shm(shmobj, sizeof(struct channel))); |
466 | chan = shmp(handle, handle->chan); | |
a6352fd4 | 467 | if (!chan) |
1d498196 | 468 | goto error_append; |
a6352fd4 MD |
469 | |
470 | ret = channel_backend_init(&chan->backend, name, config, priv, | |
1d498196 | 471 | subbuf_size, num_subbuf, handle); |
852c2936 | 472 | if (ret) |
1d498196 | 473 | goto error_backend_init; |
852c2936 MD |
474 | |
475 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); | |
a6352fd4 MD |
476 | //TODO |
477 | //chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval); | |
478 | //chan->read_timer_interval = usecs_to_jiffies(read_timer_interval); | |
a6352fd4 MD |
479 | //TODO |
480 | //init_waitqueue_head(&chan->read_wait); | |
481 | //init_waitqueue_head(&chan->hp_wait); | |
852c2936 MD |
482 | |
483 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
852c2936 MD |
484 | /* |
485 | * In case of non-hotplug cpu, if the ring-buffer is allocated | |
486 | * in early initcall, it will not be notified of secondary cpus. | |
487 | * In that off case, we need to allocate for all possible cpus. | |
488 | */ | |
852c2936 | 489 | for_each_possible_cpu(cpu) { |
1d498196 MD |
490 | struct lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
491 | lib_ring_buffer_start_switch_timer(buf, handle); | |
492 | lib_ring_buffer_start_read_timer(buf, handle); | |
852c2936 | 493 | } |
852c2936 | 494 | } else { |
1d498196 | 495 | struct lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 496 | |
1d498196 MD |
497 | lib_ring_buffer_start_switch_timer(buf, handle); |
498 | lib_ring_buffer_start_read_timer(buf, handle); | |
852c2936 | 499 | } |
193183fb MD |
500 | ref = &handle->chan._ref; |
501 | shm_get_object_data(handle, ref, shm_fd, wait_fd, memory_map_size); | |
431d5cf0 | 502 | return handle; |
852c2936 | 503 | |
1d498196 MD |
504 | error_backend_init: |
505 | error_append: | |
506 | shm_object_table_destroy(handle->table); | |
507 | error_table_alloc: | |
431d5cf0 | 508 | free(handle); |
852c2936 MD |
509 | return NULL; |
510 | } | |
852c2936 | 511 | |
193183fb MD |
512 | struct shm_handle *channel_handle_create(int shm_fd, int wait_fd, |
513 | uint64_t memory_map_size) | |
514 | { | |
515 | struct shm_handle *handle; | |
516 | struct shm_object *object; | |
517 | ||
518 | handle = zmalloc(sizeof(struct shm_handle)); | |
519 | if (!handle) | |
520 | return NULL; | |
521 | ||
522 | /* Allocate table for channel + per-cpu buffers */ | |
523 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
524 | if (!handle->table) | |
525 | goto error_table_alloc; | |
526 | /* Add channel object */ | |
527 | object = shm_object_table_append_shadow(handle->table, | |
528 | shm_fd, wait_fd, memory_map_size); | |
529 | if (!object) | |
530 | goto error_table_object; | |
531 | ||
532 | return handle; | |
533 | ||
534 | error_table_object: | |
535 | shm_object_table_destroy(handle->table); | |
536 | error_table_alloc: | |
537 | free(handle); | |
538 | return NULL; | |
539 | } | |
540 | ||
541 | int channel_handle_add_stream(struct shm_handle *handle, | |
542 | int shm_fd, int wait_fd, uint64_t memory_map_size) | |
543 | { | |
544 | struct shm_object *object; | |
545 | ||
546 | /* Add stream object */ | |
547 | object = shm_object_table_append_shadow(handle->table, | |
548 | shm_fd, wait_fd, memory_map_size); | |
549 | if (!object) | |
550 | return -1; | |
551 | return 0; | |
552 | } | |
553 | ||
852c2936 | 554 | static |
824f40b8 MD |
555 | void channel_release(struct channel *chan, struct shm_handle *handle, |
556 | int shadow) | |
852c2936 | 557 | { |
824f40b8 | 558 | channel_free(chan, handle, shadow); |
852c2936 MD |
559 | } |
560 | ||
561 | /** | |
562 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
563 | * @chan: channel to destroy | |
564 | * | |
565 | * Holds cpu hotplug. | |
431d5cf0 MD |
566 | * Call "destroy" callback, finalize channels, decrement the channel |
567 | * reference count. Note that when readers have completed data | |
568 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
569 | * They should release their handle at that point. Returns the private | |
570 | * data pointer. | |
852c2936 | 571 | */ |
824f40b8 MD |
572 | void *channel_destroy(struct channel *chan, struct shm_handle *handle, |
573 | int shadow) | |
852c2936 | 574 | { |
193183fb | 575 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 | 576 | void *priv; |
431d5cf0 | 577 | int cpu; |
852c2936 | 578 | |
824f40b8 MD |
579 | if (shadow) { |
580 | channel_release(chan, handle, shadow); | |
581 | return NULL; | |
582 | } | |
583 | ||
1d498196 | 584 | channel_unregister_notifiers(chan, handle); |
852c2936 MD |
585 | |
586 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
852c2936 | 587 | for_each_channel_cpu(cpu, chan) { |
1d498196 | 588 | struct lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
852c2936 MD |
589 | |
590 | if (config->cb.buffer_finalize) | |
591 | config->cb.buffer_finalize(buf, | |
592 | chan->backend.priv, | |
1d498196 | 593 | cpu, handle); |
852c2936 | 594 | if (buf->backend.allocated) |
1d498196 MD |
595 | lib_ring_buffer_switch_slow(buf, SWITCH_FLUSH, |
596 | handle); | |
852c2936 MD |
597 | /* |
598 | * Perform flush before writing to finalized. | |
599 | */ | |
a6352fd4 | 600 | cmm_smp_wmb(); |
14641deb | 601 | CMM_ACCESS_ONCE(buf->finalized) = 1; |
a6352fd4 | 602 | //wake_up_interruptible(&buf->read_wait); |
852c2936 MD |
603 | } |
604 | } else { | |
1d498196 | 605 | struct lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 MD |
606 | |
607 | if (config->cb.buffer_finalize) | |
1d498196 | 608 | config->cb.buffer_finalize(buf, chan->backend.priv, -1, handle); |
852c2936 | 609 | if (buf->backend.allocated) |
1d498196 MD |
610 | lib_ring_buffer_switch_slow(buf, SWITCH_FLUSH, |
611 | handle); | |
852c2936 MD |
612 | /* |
613 | * Perform flush before writing to finalized. | |
614 | */ | |
a6352fd4 | 615 | cmm_smp_wmb(); |
14641deb | 616 | CMM_ACCESS_ONCE(buf->finalized) = 1; |
a6352fd4 | 617 | //wake_up_interruptible(&buf->read_wait); |
852c2936 | 618 | } |
14641deb | 619 | CMM_ACCESS_ONCE(chan->finalized) = 1; |
a6352fd4 MD |
620 | //wake_up_interruptible(&chan->hp_wait); |
621 | //wake_up_interruptible(&chan->read_wait); | |
431d5cf0 MD |
622 | /* |
623 | * sessiond/consumer are keeping a reference on the shm file | |
624 | * descriptor directly. No need to refcount. | |
625 | */ | |
852c2936 | 626 | priv = chan->backend.priv; |
824f40b8 | 627 | channel_release(chan, handle, shadow); |
852c2936 MD |
628 | return priv; |
629 | } | |
852c2936 MD |
630 | |
631 | struct lib_ring_buffer *channel_get_ring_buffer( | |
632 | const struct lib_ring_buffer_config *config, | |
1d498196 | 633 | struct channel *chan, int cpu, |
381c0f1e MD |
634 | struct shm_handle *handle, |
635 | int *shm_fd, int *wait_fd, | |
636 | uint64_t *memory_map_size) | |
852c2936 | 637 | { |
381c0f1e MD |
638 | struct shm_ref *ref; |
639 | ||
640 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) { | |
641 | ref = &chan->backend.buf[0].shmp._ref; | |
642 | shm_get_object_data(handle, ref, shm_fd, wait_fd, | |
643 | memory_map_size); | |
1d498196 | 644 | return shmp(handle, chan->backend.buf[0].shmp); |
381c0f1e | 645 | } else { |
e095d803 MD |
646 | if (cpu >= num_possible_cpus()) |
647 | return NULL; | |
381c0f1e MD |
648 | ref = &chan->backend.buf[cpu].shmp._ref; |
649 | shm_get_object_data(handle, ref, shm_fd, wait_fd, | |
650 | memory_map_size); | |
1d498196 | 651 | return shmp(handle, chan->backend.buf[cpu].shmp); |
381c0f1e | 652 | } |
852c2936 | 653 | } |
852c2936 | 654 | |
1d498196 | 655 | int lib_ring_buffer_open_read(struct lib_ring_buffer *buf, |
824f40b8 MD |
656 | struct shm_handle *handle, |
657 | int shadow) | |
852c2936 | 658 | { |
1d498196 | 659 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 | 660 | |
824f40b8 MD |
661 | if (shadow) { |
662 | if (uatomic_cmpxchg(&buf->active_shadow_readers, 0, 1) != 0) | |
663 | return -EBUSY; | |
664 | cmm_smp_mb(); | |
665 | return 0; | |
666 | } | |
a6352fd4 | 667 | if (uatomic_cmpxchg(&buf->active_readers, 0, 1) != 0) |
852c2936 | 668 | return -EBUSY; |
a6352fd4 | 669 | cmm_smp_mb(); |
852c2936 MD |
670 | return 0; |
671 | } | |
852c2936 | 672 | |
1d498196 | 673 | void lib_ring_buffer_release_read(struct lib_ring_buffer *buf, |
824f40b8 MD |
674 | struct shm_handle *handle, |
675 | int shadow) | |
852c2936 | 676 | { |
1d498196 | 677 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 | 678 | |
824f40b8 MD |
679 | if (shadow) { |
680 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_shadow_readers) != 1); | |
681 | cmm_smp_mb(); | |
682 | uatomic_dec(&buf->active_shadow_readers); | |
683 | return; | |
684 | } | |
a6352fd4 MD |
685 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
686 | cmm_smp_mb(); | |
687 | uatomic_dec(&buf->active_readers); | |
852c2936 MD |
688 | } |
689 | ||
690 | /** | |
691 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
692 | * @buf: ring buffer | |
693 | * @consumed: consumed count indicating the position where to read | |
694 | * @produced: produced count, indicates position when to stop reading | |
695 | * | |
696 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
697 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 MD |
698 | */ |
699 | ||
700 | int lib_ring_buffer_snapshot(struct lib_ring_buffer *buf, | |
1d498196 MD |
701 | unsigned long *consumed, unsigned long *produced, |
702 | struct shm_handle *handle) | |
852c2936 | 703 | { |
1d498196 | 704 | struct channel *chan = shmp(handle, buf->backend.chan); |
193183fb | 705 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
706 | unsigned long consumed_cur, write_offset; |
707 | int finalized; | |
708 | ||
14641deb | 709 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
710 | /* |
711 | * Read finalized before counters. | |
712 | */ | |
a6352fd4 MD |
713 | cmm_smp_rmb(); |
714 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 MD |
715 | /* |
716 | * No need to issue a memory barrier between consumed count read and | |
717 | * write offset read, because consumed count can only change | |
718 | * concurrently in overwrite mode, and we keep a sequence counter | |
719 | * identifier derived from the write offset to check we are getting | |
720 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
721 | * "tagged" upon writes, tags are checked upon read). | |
722 | */ | |
723 | write_offset = v_read(config, &buf->offset); | |
724 | ||
725 | /* | |
726 | * Check that we are not about to read the same subbuffer in | |
727 | * which the writer head is. | |
728 | */ | |
729 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
730 | == 0) | |
731 | goto nodata; | |
732 | ||
733 | *consumed = consumed_cur; | |
734 | *produced = subbuf_trunc(write_offset, chan); | |
735 | ||
736 | return 0; | |
737 | ||
738 | nodata: | |
739 | /* | |
740 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
741 | * of "raw_spin_is_locked" memory ordering. | |
742 | */ | |
743 | if (finalized) | |
744 | return -ENODATA; | |
852c2936 MD |
745 | else |
746 | return -EAGAIN; | |
747 | } | |
852c2936 MD |
748 | |
749 | /** | |
750 | * lib_ring_buffer_put_snapshot - move consumed counter forward | |
751 | * @buf: ring buffer | |
752 | * @consumed_new: new consumed count value | |
753 | */ | |
754 | void lib_ring_buffer_move_consumer(struct lib_ring_buffer *buf, | |
1d498196 MD |
755 | unsigned long consumed_new, |
756 | struct shm_handle *handle) | |
852c2936 MD |
757 | { |
758 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
1d498196 | 759 | struct channel *chan = shmp(handle, bufb->chan); |
852c2936 MD |
760 | unsigned long consumed; |
761 | ||
824f40b8 MD |
762 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1 |
763 | && uatomic_read(&buf->active_shadow_readers) != 1); | |
852c2936 MD |
764 | |
765 | /* | |
766 | * Only push the consumed value forward. | |
767 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
768 | * the writer in flight recorder mode. | |
769 | */ | |
a6352fd4 | 770 | consumed = uatomic_read(&buf->consumed); |
852c2936 | 771 | while ((long) consumed - (long) consumed_new < 0) |
a6352fd4 MD |
772 | consumed = uatomic_cmpxchg(&buf->consumed, consumed, |
773 | consumed_new); | |
852c2936 | 774 | } |
852c2936 MD |
775 | |
776 | /** | |
777 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
778 | * @buf: ring buffer | |
779 | * @consumed: consumed count indicating the position where to read | |
780 | * | |
781 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
782 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 MD |
783 | */ |
784 | int lib_ring_buffer_get_subbuf(struct lib_ring_buffer *buf, | |
1d498196 MD |
785 | unsigned long consumed, |
786 | struct shm_handle *handle) | |
852c2936 | 787 | { |
1d498196 | 788 | struct channel *chan = shmp(handle, buf->backend.chan); |
193183fb | 789 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
790 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
791 | int ret; | |
792 | int finalized; | |
793 | ||
794 | retry: | |
14641deb | 795 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
796 | /* |
797 | * Read finalized before counters. | |
798 | */ | |
a6352fd4 MD |
799 | cmm_smp_rmb(); |
800 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 | 801 | consumed_idx = subbuf_index(consumed, chan); |
4746ae29 | 802 | commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb); |
852c2936 MD |
803 | /* |
804 | * Make sure we read the commit count before reading the buffer | |
805 | * data and the write offset. Correct consumed offset ordering | |
806 | * wrt commit count is insured by the use of cmpxchg to update | |
807 | * the consumed offset. | |
852c2936 | 808 | */ |
a6352fd4 MD |
809 | /* |
810 | * Local rmb to match the remote wmb to read the commit count | |
811 | * before the buffer data and the write offset. | |
812 | */ | |
813 | cmm_smp_rmb(); | |
852c2936 MD |
814 | |
815 | write_offset = v_read(config, &buf->offset); | |
816 | ||
817 | /* | |
818 | * Check that the buffer we are getting is after or at consumed_cur | |
819 | * position. | |
820 | */ | |
821 | if ((long) subbuf_trunc(consumed, chan) | |
822 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
823 | goto nodata; | |
824 | ||
825 | /* | |
826 | * Check that the subbuffer we are trying to consume has been | |
827 | * already fully committed. | |
828 | */ | |
829 | if (((commit_count - chan->backend.subbuf_size) | |
830 | & chan->commit_count_mask) | |
831 | - (buf_trunc(consumed_cur, chan) | |
832 | >> chan->backend.num_subbuf_order) | |
833 | != 0) | |
834 | goto nodata; | |
835 | ||
836 | /* | |
837 | * Check that we are not about to read the same subbuffer in | |
838 | * which the writer head is. | |
839 | */ | |
840 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
841 | == 0) | |
842 | goto nodata; | |
843 | ||
844 | /* | |
845 | * Failure to get the subbuffer causes a busy-loop retry without going | |
846 | * to a wait queue. These are caused by short-lived race windows where | |
847 | * the writer is getting access to a subbuffer we were trying to get | |
848 | * access to. Also checks that the "consumed" buffer count we are | |
849 | * looking for matches the one contained in the subbuffer id. | |
850 | */ | |
851 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
852 | consumed_idx, buf_trunc_val(consumed, chan), |
853 | handle); | |
852c2936 MD |
854 | if (ret) |
855 | goto retry; | |
856 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); | |
857 | ||
858 | buf->get_subbuf_consumed = consumed; | |
859 | buf->get_subbuf = 1; | |
860 | ||
861 | return 0; | |
862 | ||
863 | nodata: | |
864 | /* | |
865 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
866 | * of "raw_spin_is_locked" memory ordering. | |
867 | */ | |
868 | if (finalized) | |
869 | return -ENODATA; | |
852c2936 MD |
870 | else |
871 | return -EAGAIN; | |
872 | } | |
852c2936 MD |
873 | |
874 | /** | |
875 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
876 | * @buf: ring buffer | |
877 | */ | |
1d498196 MD |
878 | void lib_ring_buffer_put_subbuf(struct lib_ring_buffer *buf, |
879 | struct shm_handle *handle) | |
852c2936 MD |
880 | { |
881 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
1d498196 | 882 | struct channel *chan = shmp(handle, bufb->chan); |
193183fb | 883 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
884 | unsigned long read_sb_bindex, consumed_idx, consumed; |
885 | ||
824f40b8 MD |
886 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1 |
887 | && uatomic_read(&buf->active_shadow_readers) != 1); | |
852c2936 MD |
888 | |
889 | if (!buf->get_subbuf) { | |
890 | /* | |
891 | * Reader puts a subbuffer it did not get. | |
892 | */ | |
893 | CHAN_WARN_ON(chan, 1); | |
894 | return; | |
895 | } | |
896 | consumed = buf->get_subbuf_consumed; | |
897 | buf->get_subbuf = 0; | |
898 | ||
899 | /* | |
900 | * Clear the records_unread counter. (overruns counter) | |
901 | * Can still be non-zero if a file reader simply grabbed the data | |
902 | * without using iterators. | |
903 | * Can be below zero if an iterator is used on a snapshot more than | |
904 | * once. | |
905 | */ | |
906 | read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); | |
907 | v_add(config, v_read(config, | |
4746ae29 | 908 | &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread), |
852c2936 | 909 | &bufb->records_read); |
4746ae29 | 910 | v_set(config, &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread, 0); |
852c2936 MD |
911 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE |
912 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
913 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
914 | ||
915 | /* | |
916 | * Exchange the reader subbuffer with the one we put in its place in the | |
917 | * writer subbuffer table. Expect the original consumed count. If | |
918 | * update_read_sb_index fails, this is because the writer updated the | |
919 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
920 | * currently have: it has become invalid to try reading this sub-buffer | |
921 | * consumed count value anyway. | |
922 | */ | |
923 | consumed_idx = subbuf_index(consumed, chan); | |
924 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
925 | consumed_idx, buf_trunc_val(consumed, chan), |
926 | handle); | |
852c2936 MD |
927 | /* |
928 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
929 | * if the writer concurrently updated it. | |
930 | */ | |
931 | } | |
852c2936 MD |
932 | |
933 | /* | |
934 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
935 | * position and the writer position. (inclusive) | |
936 | */ | |
937 | static | |
938 | void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer *buf, | |
939 | struct channel *chan, | |
940 | unsigned long cons_offset, | |
1d498196 MD |
941 | int cpu, |
942 | struct shm_handle *handle) | |
852c2936 | 943 | { |
193183fb | 944 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
945 | unsigned long cons_idx, commit_count, commit_count_sb; |
946 | ||
947 | cons_idx = subbuf_index(cons_offset, chan); | |
4746ae29 MD |
948 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, cons_idx)->cc); |
949 | commit_count_sb = v_read(config, &shmp_index(handle, buf->commit_cold, cons_idx)->cc_sb); | |
852c2936 MD |
950 | |
951 | if (subbuf_offset(commit_count, chan) != 0) | |
a6352fd4 | 952 | ERRMSG("ring buffer %s, cpu %d: " |
852c2936 MD |
953 | "commit count in subbuffer %lu,\n" |
954 | "expecting multiples of %lu bytes\n" | |
955 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
956 | chan->backend.name, cpu, cons_idx, | |
957 | chan->backend.subbuf_size, | |
958 | commit_count, commit_count_sb); | |
959 | ||
a6352fd4 | 960 | ERRMSG("ring buffer: %s, cpu %d: %lu bytes committed\n", |
852c2936 MD |
961 | chan->backend.name, cpu, commit_count); |
962 | } | |
963 | ||
964 | static | |
965 | void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer *buf, | |
966 | struct channel *chan, | |
1d498196 MD |
967 | void *priv, int cpu, |
968 | struct shm_handle *handle) | |
852c2936 | 969 | { |
193183fb | 970 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
971 | unsigned long write_offset, cons_offset; |
972 | ||
973 | /* | |
974 | * Can be called in the error path of allocation when | |
975 | * trans_channel_data is not yet set. | |
976 | */ | |
977 | if (!chan) | |
978 | return; | |
979 | /* | |
980 | * No need to order commit_count, write_offset and cons_offset reads | |
981 | * because we execute at teardown when no more writer nor reader | |
982 | * references are left. | |
983 | */ | |
984 | write_offset = v_read(config, &buf->offset); | |
a6352fd4 | 985 | cons_offset = uatomic_read(&buf->consumed); |
852c2936 | 986 | if (write_offset != cons_offset) |
a6352fd4 | 987 | ERRMSG("ring buffer %s, cpu %d: " |
852c2936 MD |
988 | "non-consumed data\n" |
989 | " [ %lu bytes written, %lu bytes read ]\n", | |
990 | chan->backend.name, cpu, write_offset, cons_offset); | |
991 | ||
a6352fd4 | 992 | for (cons_offset = uatomic_read(&buf->consumed); |
852c2936 MD |
993 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), |
994 | chan) | |
995 | - cons_offset) > 0; | |
996 | cons_offset = subbuf_align(cons_offset, chan)) | |
997 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1d498196 | 998 | cpu, handle); |
852c2936 MD |
999 | } |
1000 | ||
1001 | static | |
1002 | void lib_ring_buffer_print_errors(struct channel *chan, | |
1d498196 MD |
1003 | struct lib_ring_buffer *buf, int cpu, |
1004 | struct shm_handle *handle) | |
852c2936 | 1005 | { |
193183fb | 1006 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1007 | void *priv = chan->backend.priv; |
1008 | ||
a6352fd4 | 1009 | ERRMSG("ring buffer %s, cpu %d: %lu records written, " |
852c2936 MD |
1010 | "%lu records overrun\n", |
1011 | chan->backend.name, cpu, | |
1012 | v_read(config, &buf->records_count), | |
1013 | v_read(config, &buf->records_overrun)); | |
1014 | ||
1015 | if (v_read(config, &buf->records_lost_full) | |
1016 | || v_read(config, &buf->records_lost_wrap) | |
1017 | || v_read(config, &buf->records_lost_big)) | |
a6352fd4 | 1018 | ERRMSG("ring buffer %s, cpu %d: records were lost. Caused by:\n" |
852c2936 MD |
1019 | " [ %lu buffer full, %lu nest buffer wrap-around, " |
1020 | "%lu event too big ]\n", | |
1021 | chan->backend.name, cpu, | |
1022 | v_read(config, &buf->records_lost_full), | |
1023 | v_read(config, &buf->records_lost_wrap), | |
1024 | v_read(config, &buf->records_lost_big)); | |
1025 | ||
1d498196 | 1026 | lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu, handle); |
852c2936 MD |
1027 | } |
1028 | ||
1029 | /* | |
1030 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
1031 | * | |
1032 | * Only executed when the buffer is finalized, in SWITCH_FLUSH. | |
1033 | */ | |
1034 | static | |
1035 | void lib_ring_buffer_switch_old_start(struct lib_ring_buffer *buf, | |
1036 | struct channel *chan, | |
1037 | struct switch_offsets *offsets, | |
1d498196 MD |
1038 | u64 tsc, |
1039 | struct shm_handle *handle) | |
852c2936 | 1040 | { |
193183fb | 1041 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1042 | unsigned long oldidx = subbuf_index(offsets->old, chan); |
1043 | unsigned long commit_count; | |
1044 | ||
1d498196 | 1045 | config->cb.buffer_begin(buf, tsc, oldidx, handle); |
852c2936 MD |
1046 | |
1047 | /* | |
1048 | * Order all writes to buffer before the commit count update that will | |
1049 | * determine that the subbuffer is full. | |
1050 | */ | |
a6352fd4 | 1051 | cmm_smp_wmb(); |
852c2936 | 1052 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1053 | &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1054 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 MD |
1055 | /* Check if the written buffer has to be delivered */ |
1056 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
1d498196 | 1057 | commit_count, oldidx, handle); |
852c2936 MD |
1058 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
1059 | offsets->old, commit_count, | |
1d498196 MD |
1060 | config->cb.subbuffer_header_size(), |
1061 | handle); | |
852c2936 MD |
1062 | } |
1063 | ||
1064 | /* | |
1065 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
1066 | * | |
1067 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
1068 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
1069 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
1070 | * subbuffer. | |
1071 | */ | |
1072 | static | |
1073 | void lib_ring_buffer_switch_old_end(struct lib_ring_buffer *buf, | |
1074 | struct channel *chan, | |
1075 | struct switch_offsets *offsets, | |
1d498196 MD |
1076 | u64 tsc, |
1077 | struct shm_handle *handle) | |
852c2936 | 1078 | { |
193183fb | 1079 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1080 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); |
1081 | unsigned long commit_count, padding_size, data_size; | |
1082 | ||
1083 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1084 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1085 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size, |
1086 | handle); | |
852c2936 MD |
1087 | |
1088 | /* | |
1089 | * Order all writes to buffer before the commit count update that will | |
1090 | * determine that the subbuffer is full. | |
1091 | */ | |
a6352fd4 | 1092 | cmm_smp_wmb(); |
4746ae29 MD |
1093 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1094 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 | 1095 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, |
1d498196 | 1096 | commit_count, oldidx, handle); |
852c2936 MD |
1097 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
1098 | offsets->old, commit_count, | |
1d498196 | 1099 | padding_size, handle); |
852c2936 MD |
1100 | } |
1101 | ||
1102 | /* | |
1103 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1104 | * | |
1105 | * This code can be executed unordered : writers may already have written to the | |
1106 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1107 | * that this code is executed before the deliver of this sub-buffer. | |
1108 | */ | |
1109 | static | |
1110 | void lib_ring_buffer_switch_new_start(struct lib_ring_buffer *buf, | |
1111 | struct channel *chan, | |
1112 | struct switch_offsets *offsets, | |
1d498196 MD |
1113 | u64 tsc, |
1114 | struct shm_handle *handle) | |
852c2936 | 1115 | { |
193183fb | 1116 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1117 | unsigned long beginidx = subbuf_index(offsets->begin, chan); |
1118 | unsigned long commit_count; | |
1119 | ||
1d498196 | 1120 | config->cb.buffer_begin(buf, tsc, beginidx, handle); |
852c2936 MD |
1121 | |
1122 | /* | |
1123 | * Order all writes to buffer before the commit count update that will | |
1124 | * determine that the subbuffer is full. | |
1125 | */ | |
a6352fd4 | 1126 | cmm_smp_wmb(); |
852c2936 | 1127 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1128 | &shmp_index(handle, buf->commit_hot, beginidx)->cc); |
1129 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, beginidx)->cc); | |
852c2936 MD |
1130 | /* Check if the written buffer has to be delivered */ |
1131 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
1d498196 | 1132 | commit_count, beginidx, handle); |
852c2936 MD |
1133 | lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx, |
1134 | offsets->begin, commit_count, | |
1d498196 MD |
1135 | config->cb.subbuffer_header_size(), |
1136 | handle); | |
852c2936 MD |
1137 | } |
1138 | ||
1139 | /* | |
1140 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1141 | * | |
1142 | * The only remaining threads could be the ones with pending commits. They will | |
1143 | * have to do the deliver themselves. | |
1144 | */ | |
1145 | static | |
1146 | void lib_ring_buffer_switch_new_end(struct lib_ring_buffer *buf, | |
1d498196 MD |
1147 | struct channel *chan, |
1148 | struct switch_offsets *offsets, | |
1149 | u64 tsc, | |
1150 | struct shm_handle *handle) | |
852c2936 | 1151 | { |
193183fb | 1152 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1153 | unsigned long endidx = subbuf_index(offsets->end - 1, chan); |
1154 | unsigned long commit_count, padding_size, data_size; | |
1155 | ||
1156 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; | |
1157 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1158 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size, |
1159 | handle); | |
852c2936 MD |
1160 | |
1161 | /* | |
1162 | * Order all writes to buffer before the commit count update that will | |
1163 | * determine that the subbuffer is full. | |
1164 | */ | |
a6352fd4 | 1165 | cmm_smp_wmb(); |
4746ae29 MD |
1166 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, endidx)->cc); |
1167 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc); | |
852c2936 | 1168 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->end - 1, |
1d498196 | 1169 | commit_count, endidx, handle); |
852c2936 MD |
1170 | lib_ring_buffer_write_commit_counter(config, buf, chan, endidx, |
1171 | offsets->end, commit_count, | |
1d498196 | 1172 | padding_size, handle); |
852c2936 MD |
1173 | } |
1174 | ||
1175 | /* | |
1176 | * Returns : | |
1177 | * 0 if ok | |
1178 | * !0 if execution must be aborted. | |
1179 | */ | |
1180 | static | |
1181 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
1182 | struct lib_ring_buffer *buf, | |
1183 | struct channel *chan, | |
1184 | struct switch_offsets *offsets, | |
1185 | u64 *tsc) | |
1186 | { | |
193183fb | 1187 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1188 | unsigned long off; |
1189 | ||
1190 | offsets->begin = v_read(config, &buf->offset); | |
1191 | offsets->old = offsets->begin; | |
1192 | offsets->switch_old_start = 0; | |
1193 | off = subbuf_offset(offsets->begin, chan); | |
1194 | ||
1195 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1196 | ||
1197 | /* | |
1198 | * Ensure we flush the header of an empty subbuffer when doing the | |
1199 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1200 | * total data gathering duration even if there were no records saved | |
1201 | * after the last buffer switch. | |
1202 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1203 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1204 | * subbuffer header as appropriate. | |
1205 | * The next record that reserves space will be responsible for | |
1206 | * populating the following subbuffer header. We choose not to populate | |
1207 | * the next subbuffer header here because we want to be able to use | |
a6352fd4 MD |
1208 | * SWITCH_ACTIVE for periodical buffer flush, which must |
1209 | * guarantee that all the buffer content (records and header | |
1210 | * timestamps) are visible to the reader. This is required for | |
1211 | * quiescence guarantees for the fusion merge. | |
852c2936 MD |
1212 | */ |
1213 | if (mode == SWITCH_FLUSH || off > 0) { | |
1214 | if (unlikely(off == 0)) { | |
1215 | /* | |
1216 | * The client does not save any header information. | |
1217 | * Don't switch empty subbuffer on finalize, because it | |
1218 | * is invalid to deliver a completely empty subbuffer. | |
1219 | */ | |
1220 | if (!config->cb.subbuffer_header_size()) | |
1221 | return -1; | |
1222 | /* | |
1223 | * Need to write the subbuffer start header on finalize. | |
1224 | */ | |
1225 | offsets->switch_old_start = 1; | |
1226 | } | |
1227 | offsets->begin = subbuf_align(offsets->begin, chan); | |
1228 | } else | |
1229 | return -1; /* we do not have to switch : buffer is empty */ | |
1230 | /* Note: old points to the next subbuf at offset 0 */ | |
1231 | offsets->end = offsets->begin; | |
1232 | return 0; | |
1233 | } | |
1234 | ||
1235 | /* | |
1236 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
1237 | * called while tracing is active with absolutely no lock held. | |
1238 | * | |
1239 | * Note, however, that as a v_cmpxchg is used for some atomic | |
1240 | * operations, this function must be called from the CPU which owns the buffer | |
1241 | * for a ACTIVE flush. | |
1242 | */ | |
1d498196 MD |
1243 | void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, enum switch_mode mode, |
1244 | struct shm_handle *handle) | |
852c2936 | 1245 | { |
1d498196 | 1246 | struct channel *chan = shmp(handle, buf->backend.chan); |
193183fb | 1247 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1248 | struct switch_offsets offsets; |
1249 | unsigned long oldidx; | |
1250 | u64 tsc; | |
1251 | ||
1252 | offsets.size = 0; | |
1253 | ||
1254 | /* | |
1255 | * Perform retryable operations. | |
1256 | */ | |
1257 | do { | |
1258 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
1259 | &tsc)) | |
1260 | return; /* Switch not needed */ | |
1261 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
1262 | != offsets.old); | |
1263 | ||
1264 | /* | |
1265 | * Atomically update last_tsc. This update races against concurrent | |
1266 | * atomic updates, but the race will always cause supplementary full TSC | |
1267 | * records, never the opposite (missing a full TSC record when it would | |
1268 | * be needed). | |
1269 | */ | |
1270 | save_last_tsc(config, buf, tsc); | |
1271 | ||
1272 | /* | |
1273 | * Push the reader if necessary | |
1274 | */ | |
1275 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
1276 | ||
1277 | oldidx = subbuf_index(offsets.old, chan); | |
1d498196 | 1278 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx, handle); |
852c2936 MD |
1279 | |
1280 | /* | |
1281 | * May need to populate header start on SWITCH_FLUSH. | |
1282 | */ | |
1283 | if (offsets.switch_old_start) { | |
1d498196 | 1284 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc, handle); |
852c2936 MD |
1285 | offsets.old += config->cb.subbuffer_header_size(); |
1286 | } | |
1287 | ||
1288 | /* | |
1289 | * Switch old subbuffer. | |
1290 | */ | |
1d498196 | 1291 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc, handle); |
852c2936 | 1292 | } |
852c2936 MD |
1293 | |
1294 | /* | |
1295 | * Returns : | |
1296 | * 0 if ok | |
1297 | * -ENOSPC if event size is too large for packet. | |
1298 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
1299 | * -EIO if data cannot be written into the buffer for any other reason. | |
1300 | */ | |
1301 | static | |
1302 | int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf, | |
1303 | struct channel *chan, | |
1304 | struct switch_offsets *offsets, | |
1305 | struct lib_ring_buffer_ctx *ctx) | |
1306 | { | |
193183fb | 1307 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
1d498196 | 1308 | struct shm_handle *handle = ctx->handle; |
852c2936 MD |
1309 | unsigned long reserve_commit_diff; |
1310 | ||
1311 | offsets->begin = v_read(config, &buf->offset); | |
1312 | offsets->old = offsets->begin; | |
1313 | offsets->switch_new_start = 0; | |
1314 | offsets->switch_new_end = 0; | |
1315 | offsets->switch_old_end = 0; | |
1316 | offsets->pre_header_padding = 0; | |
1317 | ||
1318 | ctx->tsc = config->cb.ring_buffer_clock_read(chan); | |
1319 | if ((int64_t) ctx->tsc == -EIO) | |
1320 | return -EIO; | |
1321 | ||
1322 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
1323 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
1324 | ||
1325 | if (unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) { | |
1326 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1327 | } else { | |
1328 | offsets->size = config->cb.record_header_size(config, chan, | |
1329 | offsets->begin, | |
1330 | &offsets->pre_header_padding, | |
1331 | ctx); | |
1332 | offsets->size += | |
1333 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1334 | ctx->largest_align) | |
1335 | + ctx->data_size; | |
1336 | if (unlikely(subbuf_offset(offsets->begin, chan) + | |
1337 | offsets->size > chan->backend.subbuf_size)) { | |
1338 | offsets->switch_old_end = 1; /* For offsets->old */ | |
1339 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1340 | } | |
1341 | } | |
1342 | if (unlikely(offsets->switch_new_start)) { | |
1343 | unsigned long sb_index; | |
1344 | ||
1345 | /* | |
1346 | * We are typically not filling the previous buffer completely. | |
1347 | */ | |
1348 | if (likely(offsets->switch_old_end)) | |
1349 | offsets->begin = subbuf_align(offsets->begin, chan); | |
1350 | offsets->begin = offsets->begin | |
1351 | + config->cb.subbuffer_header_size(); | |
1352 | /* Test new buffer integrity */ | |
1353 | sb_index = subbuf_index(offsets->begin, chan); | |
1354 | reserve_commit_diff = | |
1355 | (buf_trunc(offsets->begin, chan) | |
1356 | >> chan->backend.num_subbuf_order) | |
1357 | - ((unsigned long) v_read(config, | |
4746ae29 | 1358 | &shmp_index(handle, buf->commit_cold, sb_index)->cc_sb) |
852c2936 MD |
1359 | & chan->commit_count_mask); |
1360 | if (likely(reserve_commit_diff == 0)) { | |
1361 | /* Next subbuffer not being written to. */ | |
1362 | if (unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
1363 | subbuf_trunc(offsets->begin, chan) | |
1364 | - subbuf_trunc((unsigned long) | |
a6352fd4 | 1365 | uatomic_read(&buf->consumed), chan) |
852c2936 MD |
1366 | >= chan->backend.buf_size)) { |
1367 | /* | |
1368 | * We do not overwrite non consumed buffers | |
1369 | * and we are full : record is lost. | |
1370 | */ | |
1371 | v_inc(config, &buf->records_lost_full); | |
1372 | return -ENOBUFS; | |
1373 | } else { | |
1374 | /* | |
1375 | * Next subbuffer not being written to, and we | |
1376 | * are either in overwrite mode or the buffer is | |
1377 | * not full. It's safe to write in this new | |
1378 | * subbuffer. | |
1379 | */ | |
1380 | } | |
1381 | } else { | |
1382 | /* | |
1383 | * Next subbuffer reserve offset does not match the | |
1384 | * commit offset. Drop record in producer-consumer and | |
1385 | * overwrite mode. Caused by either a writer OOPS or too | |
1386 | * many nested writes over a reserve/commit pair. | |
1387 | */ | |
1388 | v_inc(config, &buf->records_lost_wrap); | |
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; | |
1400 | if (unlikely(subbuf_offset(offsets->begin, chan) | |
1401 | + offsets->size > chan->backend.subbuf_size)) { | |
1402 | /* | |
1403 | * Record too big for subbuffers, report error, don't | |
1404 | * complete the sub-buffer switch. | |
1405 | */ | |
1406 | v_inc(config, &buf->records_lost_big); | |
1407 | return -ENOSPC; | |
1408 | } else { | |
1409 | /* | |
1410 | * We just made a successful buffer switch and the | |
1411 | * record fits in the new subbuffer. Let's write. | |
1412 | */ | |
1413 | } | |
1414 | } else { | |
1415 | /* | |
1416 | * Record fits in the current buffer and we are not on a switch | |
1417 | * boundary. It's safe to write. | |
1418 | */ | |
1419 | } | |
1420 | offsets->end = offsets->begin + offsets->size; | |
1421 | ||
1422 | if (unlikely(subbuf_offset(offsets->end, chan) == 0)) { | |
1423 | /* | |
1424 | * The offset_end will fall at the very beginning of the next | |
1425 | * subbuffer. | |
1426 | */ | |
1427 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
1428 | } | |
1429 | return 0; | |
1430 | } | |
1431 | ||
1432 | /** | |
1433 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
1434 | * @ctx: ring buffer context. | |
1435 | * | |
1436 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, | |
1437 | * -EIO for other errors, else returns 0. | |
1438 | * It will take care of sub-buffer switching. | |
1439 | */ | |
1440 | int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx) | |
1441 | { | |
1442 | struct channel *chan = ctx->chan; | |
1d498196 | 1443 | struct shm_handle *handle = ctx->handle; |
193183fb | 1444 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
852c2936 MD |
1445 | struct lib_ring_buffer *buf; |
1446 | struct switch_offsets offsets; | |
1447 | int ret; | |
1448 | ||
1449 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1d498196 | 1450 | buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp); |
852c2936 | 1451 | else |
1d498196 | 1452 | buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 MD |
1453 | ctx->buf = buf; |
1454 | ||
1455 | offsets.size = 0; | |
1456 | ||
1457 | do { | |
1458 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, | |
1459 | ctx); | |
1460 | if (unlikely(ret)) | |
1461 | return ret; | |
1462 | } while (unlikely(v_cmpxchg(config, &buf->offset, offsets.old, | |
1463 | offsets.end) | |
1464 | != offsets.old)); | |
1465 | ||
1466 | /* | |
1467 | * Atomically update last_tsc. This update races against concurrent | |
1468 | * atomic updates, but the race will always cause supplementary full TSC | |
1469 | * records, never the opposite (missing a full TSC record when it would | |
1470 | * be needed). | |
1471 | */ | |
1472 | save_last_tsc(config, buf, ctx->tsc); | |
1473 | ||
1474 | /* | |
1475 | * Push the reader if necessary | |
1476 | */ | |
1477 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
1478 | ||
1479 | /* | |
1480 | * Clear noref flag for this subbuffer. | |
1481 | */ | |
1482 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1d498196 MD |
1483 | subbuf_index(offsets.end - 1, chan), |
1484 | handle); | |
852c2936 MD |
1485 | |
1486 | /* | |
1487 | * Switch old subbuffer if needed. | |
1488 | */ | |
1489 | if (unlikely(offsets.switch_old_end)) { | |
1490 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1d498196 MD |
1491 | subbuf_index(offsets.old - 1, chan), |
1492 | handle); | |
1493 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc, handle); | |
852c2936 MD |
1494 | } |
1495 | ||
1496 | /* | |
1497 | * Populate new subbuffer. | |
1498 | */ | |
1499 | if (unlikely(offsets.switch_new_start)) | |
1d498196 | 1500 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 MD |
1501 | |
1502 | if (unlikely(offsets.switch_new_end)) | |
1d498196 | 1503 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 MD |
1504 | |
1505 | ctx->slot_size = offsets.size; | |
1506 | ctx->pre_offset = offsets.begin; | |
1507 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
1508 | return 0; | |
1509 | } |