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); |
852c2936 MD |
119 | const struct lib_ring_buffer_config *config = chan->backend.config; |
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 MD |
174 | { |
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); |
852c2936 MD |
247 | const struct lib_ring_buffer_config *config = chan->backend.config; |
248 | ||
249 | /* | |
250 | * Only flush buffers periodically if readers are active. | |
251 | */ | |
a6352fd4 | 252 | if (uatomic_read(&buf->active_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); |
852c2936 MD |
269 | const struct lib_ring_buffer_config *config = chan->backend.config; |
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); |
852c2936 MD |
306 | const struct lib_ring_buffer_config *config = chan->backend.config; |
307 | ||
308 | CHAN_WARN_ON(chan, !buf->backend.allocated); | |
309 | ||
a6352fd4 | 310 | if (uatomic_read(&buf->active_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); |
852c2936 MD |
331 | const struct lib_ring_buffer_config *config = chan->backend.config; |
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); |
852c2936 MD |
355 | const struct lib_ring_buffer_config *config = chan->backend.config; |
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 MD |
378 | { |
379 | const struct lib_ring_buffer_config *config = chan->backend.config; | |
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 | ||
1d498196 | 398 | static void channel_free(struct channel *chan, struct shm_handle *handle) |
852c2936 | 399 | { |
431d5cf0 MD |
400 | int ret; |
401 | ||
1d498196 | 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 | |
412 | * @priv: ring buffer client private data | |
413 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous | |
414 | * address mapping. It is used only by RING_BUFFER_STATIC | |
415 | * configuration. It can be set to NULL for other backends. | |
416 | * @subbuf_size: subbuffer size | |
417 | * @num_subbuf: number of subbuffers | |
418 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
419 | * padding to let readers get those sub-buffers. | |
420 | * Used for live streaming. | |
421 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
422 | * | |
423 | * Holds cpu hotplug. | |
424 | * Returns NULL on failure. | |
425 | */ | |
431d5cf0 | 426 | struct shm_handle *channel_create(const struct lib_ring_buffer_config *config, |
852c2936 MD |
427 | const char *name, void *priv, void *buf_addr, |
428 | size_t subbuf_size, | |
429 | size_t num_subbuf, unsigned int switch_timer_interval, | |
431d5cf0 | 430 | unsigned int read_timer_interval) |
852c2936 | 431 | { |
1d498196 MD |
432 | int ret, cpu; |
433 | size_t shmsize; | |
852c2936 | 434 | struct channel *chan; |
431d5cf0 | 435 | struct shm_handle *handle; |
1d498196 | 436 | struct shm_object *shmobj; |
852c2936 MD |
437 | |
438 | if (lib_ring_buffer_check_config(config, switch_timer_interval, | |
439 | read_timer_interval)) | |
440 | return NULL; | |
441 | ||
431d5cf0 MD |
442 | handle = zmalloc(sizeof(struct shm_handle)); |
443 | if (!handle) | |
444 | return NULL; | |
445 | ||
1d498196 MD |
446 | /* Allocate table for channel + per-cpu buffers */ |
447 | handle->table = shm_object_table_create(1 + num_possible_cpus()); | |
448 | if (!handle->table) | |
449 | goto error_table_alloc; | |
852c2936 | 450 | |
1d498196 MD |
451 | /* Calculate the shm allocation layout */ |
452 | shmsize = sizeof(struct channel); | |
453 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
454 | shmsize += sizeof(struct lib_ring_buffer_shmp) * num_possible_cpus(); | |
455 | else | |
456 | shmsize += sizeof(struct lib_ring_buffer_shmp); | |
a6352fd4 | 457 | |
1d498196 | 458 | shmobj = shm_object_table_append(handle->table, shmsize); |
b5a14697 MD |
459 | if (!shmobj) |
460 | goto error_append; | |
1d498196 MD |
461 | set_shmp(handle->chan, zalloc_shm(shmobj, sizeof(struct channel))); |
462 | chan = shmp(handle, handle->chan); | |
a6352fd4 | 463 | if (!chan) |
1d498196 | 464 | goto error_append; |
a6352fd4 MD |
465 | |
466 | ret = channel_backend_init(&chan->backend, name, config, priv, | |
1d498196 | 467 | subbuf_size, num_subbuf, handle); |
852c2936 | 468 | if (ret) |
1d498196 | 469 | goto error_backend_init; |
852c2936 MD |
470 | |
471 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); | |
a6352fd4 MD |
472 | //TODO |
473 | //chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval); | |
474 | //chan->read_timer_interval = usecs_to_jiffies(read_timer_interval); | |
a6352fd4 MD |
475 | //TODO |
476 | //init_waitqueue_head(&chan->read_wait); | |
477 | //init_waitqueue_head(&chan->hp_wait); | |
852c2936 MD |
478 | |
479 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
852c2936 MD |
480 | /* |
481 | * In case of non-hotplug cpu, if the ring-buffer is allocated | |
482 | * in early initcall, it will not be notified of secondary cpus. | |
483 | * In that off case, we need to allocate for all possible cpus. | |
484 | */ | |
852c2936 | 485 | for_each_possible_cpu(cpu) { |
1d498196 MD |
486 | struct lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
487 | lib_ring_buffer_start_switch_timer(buf, handle); | |
488 | lib_ring_buffer_start_read_timer(buf, handle); | |
852c2936 | 489 | } |
852c2936 | 490 | } else { |
1d498196 | 491 | struct lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 492 | |
1d498196 MD |
493 | lib_ring_buffer_start_switch_timer(buf, handle); |
494 | lib_ring_buffer_start_read_timer(buf, handle); | |
852c2936 MD |
495 | } |
496 | ||
431d5cf0 | 497 | return handle; |
852c2936 | 498 | |
1d498196 MD |
499 | error_backend_init: |
500 | error_append: | |
501 | shm_object_table_destroy(handle->table); | |
502 | error_table_alloc: | |
431d5cf0 | 503 | free(handle); |
852c2936 MD |
504 | return NULL; |
505 | } | |
852c2936 MD |
506 | |
507 | static | |
1d498196 | 508 | void channel_release(struct channel *chan, struct shm_handle *handle) |
852c2936 | 509 | { |
1d498196 | 510 | channel_free(chan, handle); |
852c2936 MD |
511 | } |
512 | ||
513 | /** | |
514 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
515 | * @chan: channel to destroy | |
516 | * | |
517 | * Holds cpu hotplug. | |
431d5cf0 MD |
518 | * Call "destroy" callback, finalize channels, decrement the channel |
519 | * reference count. Note that when readers have completed data | |
520 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
521 | * They should release their handle at that point. Returns the private | |
522 | * data pointer. | |
852c2936 | 523 | */ |
1d498196 | 524 | void *channel_destroy(struct channel *chan, struct shm_handle *handle) |
852c2936 | 525 | { |
852c2936 MD |
526 | const struct lib_ring_buffer_config *config = chan->backend.config; |
527 | void *priv; | |
431d5cf0 | 528 | int cpu; |
852c2936 | 529 | |
1d498196 | 530 | channel_unregister_notifiers(chan, handle); |
852c2936 MD |
531 | |
532 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
852c2936 | 533 | for_each_channel_cpu(cpu, chan) { |
1d498196 | 534 | struct lib_ring_buffer *buf = shmp(handle, chan->backend.buf[cpu].shmp); |
852c2936 MD |
535 | |
536 | if (config->cb.buffer_finalize) | |
537 | config->cb.buffer_finalize(buf, | |
538 | chan->backend.priv, | |
1d498196 | 539 | cpu, handle); |
852c2936 | 540 | if (buf->backend.allocated) |
1d498196 MD |
541 | lib_ring_buffer_switch_slow(buf, SWITCH_FLUSH, |
542 | handle); | |
852c2936 MD |
543 | /* |
544 | * Perform flush before writing to finalized. | |
545 | */ | |
a6352fd4 | 546 | cmm_smp_wmb(); |
14641deb | 547 | CMM_ACCESS_ONCE(buf->finalized) = 1; |
a6352fd4 | 548 | //wake_up_interruptible(&buf->read_wait); |
852c2936 MD |
549 | } |
550 | } else { | |
1d498196 | 551 | struct lib_ring_buffer *buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 MD |
552 | |
553 | if (config->cb.buffer_finalize) | |
1d498196 | 554 | config->cb.buffer_finalize(buf, chan->backend.priv, -1, handle); |
852c2936 | 555 | if (buf->backend.allocated) |
1d498196 MD |
556 | lib_ring_buffer_switch_slow(buf, SWITCH_FLUSH, |
557 | handle); | |
852c2936 MD |
558 | /* |
559 | * Perform flush before writing to finalized. | |
560 | */ | |
a6352fd4 | 561 | cmm_smp_wmb(); |
14641deb | 562 | CMM_ACCESS_ONCE(buf->finalized) = 1; |
a6352fd4 | 563 | //wake_up_interruptible(&buf->read_wait); |
852c2936 | 564 | } |
14641deb | 565 | CMM_ACCESS_ONCE(chan->finalized) = 1; |
a6352fd4 MD |
566 | //wake_up_interruptible(&chan->hp_wait); |
567 | //wake_up_interruptible(&chan->read_wait); | |
431d5cf0 MD |
568 | /* |
569 | * sessiond/consumer are keeping a reference on the shm file | |
570 | * descriptor directly. No need to refcount. | |
571 | */ | |
852c2936 | 572 | priv = chan->backend.priv; |
1fc6c5c9 | 573 | channel_release(chan, handle); |
852c2936 MD |
574 | return priv; |
575 | } | |
852c2936 MD |
576 | |
577 | struct lib_ring_buffer *channel_get_ring_buffer( | |
578 | const struct lib_ring_buffer_config *config, | |
1d498196 MD |
579 | struct channel *chan, int cpu, |
580 | struct shm_handle *handle) | |
852c2936 MD |
581 | { |
582 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) | |
1d498196 | 583 | return shmp(handle, chan->backend.buf[0].shmp); |
852c2936 | 584 | else |
1d498196 | 585 | return shmp(handle, chan->backend.buf[cpu].shmp); |
852c2936 | 586 | } |
852c2936 | 587 | |
1d498196 MD |
588 | int lib_ring_buffer_open_read(struct lib_ring_buffer *buf, |
589 | struct shm_handle *handle) | |
852c2936 | 590 | { |
1d498196 | 591 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 | 592 | |
a6352fd4 | 593 | if (uatomic_cmpxchg(&buf->active_readers, 0, 1) != 0) |
852c2936 | 594 | return -EBUSY; |
a6352fd4 | 595 | cmm_smp_mb(); |
852c2936 MD |
596 | return 0; |
597 | } | |
852c2936 | 598 | |
1d498196 MD |
599 | void lib_ring_buffer_release_read(struct lib_ring_buffer *buf, |
600 | struct shm_handle *handle) | |
852c2936 | 601 | { |
1d498196 | 602 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 | 603 | |
a6352fd4 MD |
604 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
605 | cmm_smp_mb(); | |
606 | uatomic_dec(&buf->active_readers); | |
852c2936 MD |
607 | } |
608 | ||
609 | /** | |
610 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
611 | * @buf: ring buffer | |
612 | * @consumed: consumed count indicating the position where to read | |
613 | * @produced: produced count, indicates position when to stop reading | |
614 | * | |
615 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
616 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 MD |
617 | */ |
618 | ||
619 | int lib_ring_buffer_snapshot(struct lib_ring_buffer *buf, | |
1d498196 MD |
620 | unsigned long *consumed, unsigned long *produced, |
621 | struct shm_handle *handle) | |
852c2936 | 622 | { |
1d498196 | 623 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 MD |
624 | const struct lib_ring_buffer_config *config = chan->backend.config; |
625 | unsigned long consumed_cur, write_offset; | |
626 | int finalized; | |
627 | ||
14641deb | 628 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
629 | /* |
630 | * Read finalized before counters. | |
631 | */ | |
a6352fd4 MD |
632 | cmm_smp_rmb(); |
633 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 MD |
634 | /* |
635 | * No need to issue a memory barrier between consumed count read and | |
636 | * write offset read, because consumed count can only change | |
637 | * concurrently in overwrite mode, and we keep a sequence counter | |
638 | * identifier derived from the write offset to check we are getting | |
639 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
640 | * "tagged" upon writes, tags are checked upon read). | |
641 | */ | |
642 | write_offset = v_read(config, &buf->offset); | |
643 | ||
644 | /* | |
645 | * Check that we are not about to read the same subbuffer in | |
646 | * which the writer head is. | |
647 | */ | |
648 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
649 | == 0) | |
650 | goto nodata; | |
651 | ||
652 | *consumed = consumed_cur; | |
653 | *produced = subbuf_trunc(write_offset, chan); | |
654 | ||
655 | return 0; | |
656 | ||
657 | nodata: | |
658 | /* | |
659 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
660 | * of "raw_spin_is_locked" memory ordering. | |
661 | */ | |
662 | if (finalized) | |
663 | return -ENODATA; | |
852c2936 MD |
664 | else |
665 | return -EAGAIN; | |
666 | } | |
852c2936 MD |
667 | |
668 | /** | |
669 | * lib_ring_buffer_put_snapshot - move consumed counter forward | |
670 | * @buf: ring buffer | |
671 | * @consumed_new: new consumed count value | |
672 | */ | |
673 | void lib_ring_buffer_move_consumer(struct lib_ring_buffer *buf, | |
1d498196 MD |
674 | unsigned long consumed_new, |
675 | struct shm_handle *handle) | |
852c2936 MD |
676 | { |
677 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
1d498196 | 678 | struct channel *chan = shmp(handle, bufb->chan); |
852c2936 MD |
679 | unsigned long consumed; |
680 | ||
a6352fd4 | 681 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
852c2936 MD |
682 | |
683 | /* | |
684 | * Only push the consumed value forward. | |
685 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
686 | * the writer in flight recorder mode. | |
687 | */ | |
a6352fd4 | 688 | consumed = uatomic_read(&buf->consumed); |
852c2936 | 689 | while ((long) consumed - (long) consumed_new < 0) |
a6352fd4 MD |
690 | consumed = uatomic_cmpxchg(&buf->consumed, consumed, |
691 | consumed_new); | |
852c2936 | 692 | } |
852c2936 MD |
693 | |
694 | /** | |
695 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
696 | * @buf: ring buffer | |
697 | * @consumed: consumed count indicating the position where to read | |
698 | * | |
699 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
700 | * data to read at consumed position, or 0 if the get operation succeeds. | |
852c2936 MD |
701 | */ |
702 | int lib_ring_buffer_get_subbuf(struct lib_ring_buffer *buf, | |
1d498196 MD |
703 | unsigned long consumed, |
704 | struct shm_handle *handle) | |
852c2936 | 705 | { |
1d498196 | 706 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 MD |
707 | const struct lib_ring_buffer_config *config = chan->backend.config; |
708 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; | |
709 | int ret; | |
710 | int finalized; | |
711 | ||
712 | retry: | |
14641deb | 713 | finalized = CMM_ACCESS_ONCE(buf->finalized); |
852c2936 MD |
714 | /* |
715 | * Read finalized before counters. | |
716 | */ | |
a6352fd4 MD |
717 | cmm_smp_rmb(); |
718 | consumed_cur = uatomic_read(&buf->consumed); | |
852c2936 | 719 | consumed_idx = subbuf_index(consumed, chan); |
4746ae29 | 720 | commit_count = v_read(config, &shmp_index(handle, buf->commit_cold, consumed_idx)->cc_sb); |
852c2936 MD |
721 | /* |
722 | * Make sure we read the commit count before reading the buffer | |
723 | * data and the write offset. Correct consumed offset ordering | |
724 | * wrt commit count is insured by the use of cmpxchg to update | |
725 | * the consumed offset. | |
852c2936 | 726 | */ |
a6352fd4 MD |
727 | /* |
728 | * Local rmb to match the remote wmb to read the commit count | |
729 | * before the buffer data and the write offset. | |
730 | */ | |
731 | cmm_smp_rmb(); | |
852c2936 MD |
732 | |
733 | write_offset = v_read(config, &buf->offset); | |
734 | ||
735 | /* | |
736 | * Check that the buffer we are getting is after or at consumed_cur | |
737 | * position. | |
738 | */ | |
739 | if ((long) subbuf_trunc(consumed, chan) | |
740 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
741 | goto nodata; | |
742 | ||
743 | /* | |
744 | * Check that the subbuffer we are trying to consume has been | |
745 | * already fully committed. | |
746 | */ | |
747 | if (((commit_count - chan->backend.subbuf_size) | |
748 | & chan->commit_count_mask) | |
749 | - (buf_trunc(consumed_cur, chan) | |
750 | >> chan->backend.num_subbuf_order) | |
751 | != 0) | |
752 | goto nodata; | |
753 | ||
754 | /* | |
755 | * Check that we are not about to read the same subbuffer in | |
756 | * which the writer head is. | |
757 | */ | |
758 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
759 | == 0) | |
760 | goto nodata; | |
761 | ||
762 | /* | |
763 | * Failure to get the subbuffer causes a busy-loop retry without going | |
764 | * to a wait queue. These are caused by short-lived race windows where | |
765 | * the writer is getting access to a subbuffer we were trying to get | |
766 | * access to. Also checks that the "consumed" buffer count we are | |
767 | * looking for matches the one contained in the subbuffer id. | |
768 | */ | |
769 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
770 | consumed_idx, buf_trunc_val(consumed, chan), |
771 | handle); | |
852c2936 MD |
772 | if (ret) |
773 | goto retry; | |
774 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); | |
775 | ||
776 | buf->get_subbuf_consumed = consumed; | |
777 | buf->get_subbuf = 1; | |
778 | ||
779 | return 0; | |
780 | ||
781 | nodata: | |
782 | /* | |
783 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
784 | * of "raw_spin_is_locked" memory ordering. | |
785 | */ | |
786 | if (finalized) | |
787 | return -ENODATA; | |
852c2936 MD |
788 | else |
789 | return -EAGAIN; | |
790 | } | |
852c2936 MD |
791 | |
792 | /** | |
793 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
794 | * @buf: ring buffer | |
795 | */ | |
1d498196 MD |
796 | void lib_ring_buffer_put_subbuf(struct lib_ring_buffer *buf, |
797 | struct shm_handle *handle) | |
852c2936 MD |
798 | { |
799 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
1d498196 | 800 | struct channel *chan = shmp(handle, bufb->chan); |
852c2936 MD |
801 | const struct lib_ring_buffer_config *config = chan->backend.config; |
802 | unsigned long read_sb_bindex, consumed_idx, consumed; | |
803 | ||
a6352fd4 | 804 | CHAN_WARN_ON(chan, uatomic_read(&buf->active_readers) != 1); |
852c2936 MD |
805 | |
806 | if (!buf->get_subbuf) { | |
807 | /* | |
808 | * Reader puts a subbuffer it did not get. | |
809 | */ | |
810 | CHAN_WARN_ON(chan, 1); | |
811 | return; | |
812 | } | |
813 | consumed = buf->get_subbuf_consumed; | |
814 | buf->get_subbuf = 0; | |
815 | ||
816 | /* | |
817 | * Clear the records_unread counter. (overruns counter) | |
818 | * Can still be non-zero if a file reader simply grabbed the data | |
819 | * without using iterators. | |
820 | * Can be below zero if an iterator is used on a snapshot more than | |
821 | * once. | |
822 | */ | |
823 | read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); | |
824 | v_add(config, v_read(config, | |
4746ae29 | 825 | &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread), |
852c2936 | 826 | &bufb->records_read); |
4746ae29 | 827 | v_set(config, &shmp(handle, shmp_index(handle, bufb->array, read_sb_bindex)->shmp)->records_unread, 0); |
852c2936 MD |
828 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE |
829 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
830 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
831 | ||
832 | /* | |
833 | * Exchange the reader subbuffer with the one we put in its place in the | |
834 | * writer subbuffer table. Expect the original consumed count. If | |
835 | * update_read_sb_index fails, this is because the writer updated the | |
836 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
837 | * currently have: it has become invalid to try reading this sub-buffer | |
838 | * consumed count value anyway. | |
839 | */ | |
840 | consumed_idx = subbuf_index(consumed, chan); | |
841 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1d498196 MD |
842 | consumed_idx, buf_trunc_val(consumed, chan), |
843 | handle); | |
852c2936 MD |
844 | /* |
845 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
846 | * if the writer concurrently updated it. | |
847 | */ | |
848 | } | |
852c2936 MD |
849 | |
850 | /* | |
851 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
852 | * position and the writer position. (inclusive) | |
853 | */ | |
854 | static | |
855 | void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer *buf, | |
856 | struct channel *chan, | |
857 | unsigned long cons_offset, | |
1d498196 MD |
858 | int cpu, |
859 | struct shm_handle *handle) | |
852c2936 MD |
860 | { |
861 | const struct lib_ring_buffer_config *config = chan->backend.config; | |
862 | unsigned long cons_idx, commit_count, commit_count_sb; | |
863 | ||
864 | cons_idx = subbuf_index(cons_offset, chan); | |
4746ae29 MD |
865 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, cons_idx)->cc); |
866 | commit_count_sb = v_read(config, &shmp_index(handle, buf->commit_cold, cons_idx)->cc_sb); | |
852c2936 MD |
867 | |
868 | if (subbuf_offset(commit_count, chan) != 0) | |
a6352fd4 | 869 | ERRMSG("ring buffer %s, cpu %d: " |
852c2936 MD |
870 | "commit count in subbuffer %lu,\n" |
871 | "expecting multiples of %lu bytes\n" | |
872 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
873 | chan->backend.name, cpu, cons_idx, | |
874 | chan->backend.subbuf_size, | |
875 | commit_count, commit_count_sb); | |
876 | ||
a6352fd4 | 877 | ERRMSG("ring buffer: %s, cpu %d: %lu bytes committed\n", |
852c2936 MD |
878 | chan->backend.name, cpu, commit_count); |
879 | } | |
880 | ||
881 | static | |
882 | void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer *buf, | |
883 | struct channel *chan, | |
1d498196 MD |
884 | void *priv, int cpu, |
885 | struct shm_handle *handle) | |
852c2936 MD |
886 | { |
887 | const struct lib_ring_buffer_config *config = chan->backend.config; | |
888 | unsigned long write_offset, cons_offset; | |
889 | ||
890 | /* | |
891 | * Can be called in the error path of allocation when | |
892 | * trans_channel_data is not yet set. | |
893 | */ | |
894 | if (!chan) | |
895 | return; | |
896 | /* | |
897 | * No need to order commit_count, write_offset and cons_offset reads | |
898 | * because we execute at teardown when no more writer nor reader | |
899 | * references are left. | |
900 | */ | |
901 | write_offset = v_read(config, &buf->offset); | |
a6352fd4 | 902 | cons_offset = uatomic_read(&buf->consumed); |
852c2936 | 903 | if (write_offset != cons_offset) |
a6352fd4 | 904 | ERRMSG("ring buffer %s, cpu %d: " |
852c2936 MD |
905 | "non-consumed data\n" |
906 | " [ %lu bytes written, %lu bytes read ]\n", | |
907 | chan->backend.name, cpu, write_offset, cons_offset); | |
908 | ||
a6352fd4 | 909 | for (cons_offset = uatomic_read(&buf->consumed); |
852c2936 MD |
910 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), |
911 | chan) | |
912 | - cons_offset) > 0; | |
913 | cons_offset = subbuf_align(cons_offset, chan)) | |
914 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1d498196 | 915 | cpu, handle); |
852c2936 MD |
916 | } |
917 | ||
918 | static | |
919 | void lib_ring_buffer_print_errors(struct channel *chan, | |
1d498196 MD |
920 | struct lib_ring_buffer *buf, int cpu, |
921 | struct shm_handle *handle) | |
852c2936 MD |
922 | { |
923 | const struct lib_ring_buffer_config *config = chan->backend.config; | |
924 | void *priv = chan->backend.priv; | |
925 | ||
a6352fd4 | 926 | ERRMSG("ring buffer %s, cpu %d: %lu records written, " |
852c2936 MD |
927 | "%lu records overrun\n", |
928 | chan->backend.name, cpu, | |
929 | v_read(config, &buf->records_count), | |
930 | v_read(config, &buf->records_overrun)); | |
931 | ||
932 | if (v_read(config, &buf->records_lost_full) | |
933 | || v_read(config, &buf->records_lost_wrap) | |
934 | || v_read(config, &buf->records_lost_big)) | |
a6352fd4 | 935 | ERRMSG("ring buffer %s, cpu %d: records were lost. Caused by:\n" |
852c2936 MD |
936 | " [ %lu buffer full, %lu nest buffer wrap-around, " |
937 | "%lu event too big ]\n", | |
938 | chan->backend.name, cpu, | |
939 | v_read(config, &buf->records_lost_full), | |
940 | v_read(config, &buf->records_lost_wrap), | |
941 | v_read(config, &buf->records_lost_big)); | |
942 | ||
1d498196 | 943 | lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu, handle); |
852c2936 MD |
944 | } |
945 | ||
946 | /* | |
947 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
948 | * | |
949 | * Only executed when the buffer is finalized, in SWITCH_FLUSH. | |
950 | */ | |
951 | static | |
952 | void lib_ring_buffer_switch_old_start(struct lib_ring_buffer *buf, | |
953 | struct channel *chan, | |
954 | struct switch_offsets *offsets, | |
1d498196 MD |
955 | u64 tsc, |
956 | struct shm_handle *handle) | |
852c2936 MD |
957 | { |
958 | const struct lib_ring_buffer_config *config = chan->backend.config; | |
959 | unsigned long oldidx = subbuf_index(offsets->old, chan); | |
960 | unsigned long commit_count; | |
961 | ||
1d498196 | 962 | config->cb.buffer_begin(buf, tsc, oldidx, handle); |
852c2936 MD |
963 | |
964 | /* | |
965 | * Order all writes to buffer before the commit count update that will | |
966 | * determine that the subbuffer is full. | |
967 | */ | |
a6352fd4 | 968 | cmm_smp_wmb(); |
852c2936 | 969 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
970 | &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
971 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 MD |
972 | /* Check if the written buffer has to be delivered */ |
973 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
1d498196 | 974 | commit_count, oldidx, handle); |
852c2936 MD |
975 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
976 | offsets->old, commit_count, | |
1d498196 MD |
977 | config->cb.subbuffer_header_size(), |
978 | handle); | |
852c2936 MD |
979 | } |
980 | ||
981 | /* | |
982 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
983 | * | |
984 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
985 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
986 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
987 | * subbuffer. | |
988 | */ | |
989 | static | |
990 | void lib_ring_buffer_switch_old_end(struct lib_ring_buffer *buf, | |
991 | struct channel *chan, | |
992 | struct switch_offsets *offsets, | |
1d498196 MD |
993 | u64 tsc, |
994 | struct shm_handle *handle) | |
852c2936 MD |
995 | { |
996 | const struct lib_ring_buffer_config *config = chan->backend.config; | |
997 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); | |
998 | unsigned long commit_count, padding_size, data_size; | |
999 | ||
1000 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1001 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1002 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size, |
1003 | handle); | |
852c2936 MD |
1004 | |
1005 | /* | |
1006 | * Order all writes to buffer before the commit count update that will | |
1007 | * determine that the subbuffer is full. | |
1008 | */ | |
a6352fd4 | 1009 | cmm_smp_wmb(); |
4746ae29 MD |
1010 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, oldidx)->cc); |
1011 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, oldidx)->cc); | |
852c2936 | 1012 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, |
1d498196 | 1013 | commit_count, oldidx, handle); |
852c2936 MD |
1014 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
1015 | offsets->old, commit_count, | |
1d498196 | 1016 | padding_size, handle); |
852c2936 MD |
1017 | } |
1018 | ||
1019 | /* | |
1020 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1021 | * | |
1022 | * This code can be executed unordered : writers may already have written to the | |
1023 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1024 | * that this code is executed before the deliver of this sub-buffer. | |
1025 | */ | |
1026 | static | |
1027 | void lib_ring_buffer_switch_new_start(struct lib_ring_buffer *buf, | |
1028 | struct channel *chan, | |
1029 | struct switch_offsets *offsets, | |
1d498196 MD |
1030 | u64 tsc, |
1031 | struct shm_handle *handle) | |
852c2936 MD |
1032 | { |
1033 | const struct lib_ring_buffer_config *config = chan->backend.config; | |
1034 | unsigned long beginidx = subbuf_index(offsets->begin, chan); | |
1035 | unsigned long commit_count; | |
1036 | ||
1d498196 | 1037 | config->cb.buffer_begin(buf, tsc, beginidx, handle); |
852c2936 MD |
1038 | |
1039 | /* | |
1040 | * Order all writes to buffer before the commit count update that will | |
1041 | * determine that the subbuffer is full. | |
1042 | */ | |
a6352fd4 | 1043 | cmm_smp_wmb(); |
852c2936 | 1044 | v_add(config, config->cb.subbuffer_header_size(), |
4746ae29 MD |
1045 | &shmp_index(handle, buf->commit_hot, beginidx)->cc); |
1046 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, beginidx)->cc); | |
852c2936 MD |
1047 | /* Check if the written buffer has to be delivered */ |
1048 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
1d498196 | 1049 | commit_count, beginidx, handle); |
852c2936 MD |
1050 | lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx, |
1051 | offsets->begin, commit_count, | |
1d498196 MD |
1052 | config->cb.subbuffer_header_size(), |
1053 | handle); | |
852c2936 MD |
1054 | } |
1055 | ||
1056 | /* | |
1057 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1058 | * | |
1059 | * The only remaining threads could be the ones with pending commits. They will | |
1060 | * have to do the deliver themselves. | |
1061 | */ | |
1062 | static | |
1063 | void lib_ring_buffer_switch_new_end(struct lib_ring_buffer *buf, | |
1d498196 MD |
1064 | struct channel *chan, |
1065 | struct switch_offsets *offsets, | |
1066 | u64 tsc, | |
1067 | struct shm_handle *handle) | |
852c2936 MD |
1068 | { |
1069 | const struct lib_ring_buffer_config *config = chan->backend.config; | |
1070 | unsigned long endidx = subbuf_index(offsets->end - 1, chan); | |
1071 | unsigned long commit_count, padding_size, data_size; | |
1072 | ||
1073 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; | |
1074 | padding_size = chan->backend.subbuf_size - data_size; | |
1d498196 MD |
1075 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size, |
1076 | handle); | |
852c2936 MD |
1077 | |
1078 | /* | |
1079 | * Order all writes to buffer before the commit count update that will | |
1080 | * determine that the subbuffer is full. | |
1081 | */ | |
a6352fd4 | 1082 | cmm_smp_wmb(); |
4746ae29 MD |
1083 | v_add(config, padding_size, &shmp_index(handle, buf->commit_hot, endidx)->cc); |
1084 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc); | |
852c2936 | 1085 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->end - 1, |
1d498196 | 1086 | commit_count, endidx, handle); |
852c2936 MD |
1087 | lib_ring_buffer_write_commit_counter(config, buf, chan, endidx, |
1088 | offsets->end, commit_count, | |
1d498196 | 1089 | padding_size, handle); |
852c2936 MD |
1090 | } |
1091 | ||
1092 | /* | |
1093 | * Returns : | |
1094 | * 0 if ok | |
1095 | * !0 if execution must be aborted. | |
1096 | */ | |
1097 | static | |
1098 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
1099 | struct lib_ring_buffer *buf, | |
1100 | struct channel *chan, | |
1101 | struct switch_offsets *offsets, | |
1102 | u64 *tsc) | |
1103 | { | |
1104 | const struct lib_ring_buffer_config *config = chan->backend.config; | |
1105 | unsigned long off; | |
1106 | ||
1107 | offsets->begin = v_read(config, &buf->offset); | |
1108 | offsets->old = offsets->begin; | |
1109 | offsets->switch_old_start = 0; | |
1110 | off = subbuf_offset(offsets->begin, chan); | |
1111 | ||
1112 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1113 | ||
1114 | /* | |
1115 | * Ensure we flush the header of an empty subbuffer when doing the | |
1116 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1117 | * total data gathering duration even if there were no records saved | |
1118 | * after the last buffer switch. | |
1119 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1120 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1121 | * subbuffer header as appropriate. | |
1122 | * The next record that reserves space will be responsible for | |
1123 | * populating the following subbuffer header. We choose not to populate | |
1124 | * the next subbuffer header here because we want to be able to use | |
a6352fd4 MD |
1125 | * SWITCH_ACTIVE for periodical buffer flush, which must |
1126 | * guarantee that all the buffer content (records and header | |
1127 | * timestamps) are visible to the reader. This is required for | |
1128 | * quiescence guarantees for the fusion merge. | |
852c2936 MD |
1129 | */ |
1130 | if (mode == SWITCH_FLUSH || off > 0) { | |
1131 | if (unlikely(off == 0)) { | |
1132 | /* | |
1133 | * The client does not save any header information. | |
1134 | * Don't switch empty subbuffer on finalize, because it | |
1135 | * is invalid to deliver a completely empty subbuffer. | |
1136 | */ | |
1137 | if (!config->cb.subbuffer_header_size()) | |
1138 | return -1; | |
1139 | /* | |
1140 | * Need to write the subbuffer start header on finalize. | |
1141 | */ | |
1142 | offsets->switch_old_start = 1; | |
1143 | } | |
1144 | offsets->begin = subbuf_align(offsets->begin, chan); | |
1145 | } else | |
1146 | return -1; /* we do not have to switch : buffer is empty */ | |
1147 | /* Note: old points to the next subbuf at offset 0 */ | |
1148 | offsets->end = offsets->begin; | |
1149 | return 0; | |
1150 | } | |
1151 | ||
1152 | /* | |
1153 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
1154 | * called while tracing is active with absolutely no lock held. | |
1155 | * | |
1156 | * Note, however, that as a v_cmpxchg is used for some atomic | |
1157 | * operations, this function must be called from the CPU which owns the buffer | |
1158 | * for a ACTIVE flush. | |
1159 | */ | |
1d498196 MD |
1160 | void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, enum switch_mode mode, |
1161 | struct shm_handle *handle) | |
852c2936 | 1162 | { |
1d498196 | 1163 | struct channel *chan = shmp(handle, buf->backend.chan); |
852c2936 MD |
1164 | const struct lib_ring_buffer_config *config = chan->backend.config; |
1165 | struct switch_offsets offsets; | |
1166 | unsigned long oldidx; | |
1167 | u64 tsc; | |
1168 | ||
1169 | offsets.size = 0; | |
1170 | ||
1171 | /* | |
1172 | * Perform retryable operations. | |
1173 | */ | |
1174 | do { | |
1175 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
1176 | &tsc)) | |
1177 | return; /* Switch not needed */ | |
1178 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
1179 | != offsets.old); | |
1180 | ||
1181 | /* | |
1182 | * Atomically update last_tsc. This update races against concurrent | |
1183 | * atomic updates, but the race will always cause supplementary full TSC | |
1184 | * records, never the opposite (missing a full TSC record when it would | |
1185 | * be needed). | |
1186 | */ | |
1187 | save_last_tsc(config, buf, tsc); | |
1188 | ||
1189 | /* | |
1190 | * Push the reader if necessary | |
1191 | */ | |
1192 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
1193 | ||
1194 | oldidx = subbuf_index(offsets.old, chan); | |
1d498196 | 1195 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx, handle); |
852c2936 MD |
1196 | |
1197 | /* | |
1198 | * May need to populate header start on SWITCH_FLUSH. | |
1199 | */ | |
1200 | if (offsets.switch_old_start) { | |
1d498196 | 1201 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc, handle); |
852c2936 MD |
1202 | offsets.old += config->cb.subbuffer_header_size(); |
1203 | } | |
1204 | ||
1205 | /* | |
1206 | * Switch old subbuffer. | |
1207 | */ | |
1d498196 | 1208 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc, handle); |
852c2936 | 1209 | } |
852c2936 MD |
1210 | |
1211 | /* | |
1212 | * Returns : | |
1213 | * 0 if ok | |
1214 | * -ENOSPC if event size is too large for packet. | |
1215 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
1216 | * -EIO if data cannot be written into the buffer for any other reason. | |
1217 | */ | |
1218 | static | |
1219 | int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf, | |
1220 | struct channel *chan, | |
1221 | struct switch_offsets *offsets, | |
1222 | struct lib_ring_buffer_ctx *ctx) | |
1223 | { | |
1224 | const struct lib_ring_buffer_config *config = chan->backend.config; | |
1d498196 | 1225 | struct shm_handle *handle = ctx->handle; |
852c2936 MD |
1226 | unsigned long reserve_commit_diff; |
1227 | ||
1228 | offsets->begin = v_read(config, &buf->offset); | |
1229 | offsets->old = offsets->begin; | |
1230 | offsets->switch_new_start = 0; | |
1231 | offsets->switch_new_end = 0; | |
1232 | offsets->switch_old_end = 0; | |
1233 | offsets->pre_header_padding = 0; | |
1234 | ||
1235 | ctx->tsc = config->cb.ring_buffer_clock_read(chan); | |
1236 | if ((int64_t) ctx->tsc == -EIO) | |
1237 | return -EIO; | |
1238 | ||
1239 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
1240 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
1241 | ||
1242 | if (unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) { | |
1243 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1244 | } else { | |
1245 | offsets->size = config->cb.record_header_size(config, chan, | |
1246 | offsets->begin, | |
1247 | &offsets->pre_header_padding, | |
1248 | ctx); | |
1249 | offsets->size += | |
1250 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1251 | ctx->largest_align) | |
1252 | + ctx->data_size; | |
1253 | if (unlikely(subbuf_offset(offsets->begin, chan) + | |
1254 | offsets->size > chan->backend.subbuf_size)) { | |
1255 | offsets->switch_old_end = 1; /* For offsets->old */ | |
1256 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1257 | } | |
1258 | } | |
1259 | if (unlikely(offsets->switch_new_start)) { | |
1260 | unsigned long sb_index; | |
1261 | ||
1262 | /* | |
1263 | * We are typically not filling the previous buffer completely. | |
1264 | */ | |
1265 | if (likely(offsets->switch_old_end)) | |
1266 | offsets->begin = subbuf_align(offsets->begin, chan); | |
1267 | offsets->begin = offsets->begin | |
1268 | + config->cb.subbuffer_header_size(); | |
1269 | /* Test new buffer integrity */ | |
1270 | sb_index = subbuf_index(offsets->begin, chan); | |
1271 | reserve_commit_diff = | |
1272 | (buf_trunc(offsets->begin, chan) | |
1273 | >> chan->backend.num_subbuf_order) | |
1274 | - ((unsigned long) v_read(config, | |
4746ae29 | 1275 | &shmp_index(handle, buf->commit_cold, sb_index)->cc_sb) |
852c2936 MD |
1276 | & chan->commit_count_mask); |
1277 | if (likely(reserve_commit_diff == 0)) { | |
1278 | /* Next subbuffer not being written to. */ | |
1279 | if (unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
1280 | subbuf_trunc(offsets->begin, chan) | |
1281 | - subbuf_trunc((unsigned long) | |
a6352fd4 | 1282 | uatomic_read(&buf->consumed), chan) |
852c2936 MD |
1283 | >= chan->backend.buf_size)) { |
1284 | /* | |
1285 | * We do not overwrite non consumed buffers | |
1286 | * and we are full : record is lost. | |
1287 | */ | |
1288 | v_inc(config, &buf->records_lost_full); | |
1289 | return -ENOBUFS; | |
1290 | } else { | |
1291 | /* | |
1292 | * Next subbuffer not being written to, and we | |
1293 | * are either in overwrite mode or the buffer is | |
1294 | * not full. It's safe to write in this new | |
1295 | * subbuffer. | |
1296 | */ | |
1297 | } | |
1298 | } else { | |
1299 | /* | |
1300 | * Next subbuffer reserve offset does not match the | |
1301 | * commit offset. Drop record in producer-consumer and | |
1302 | * overwrite mode. Caused by either a writer OOPS or too | |
1303 | * many nested writes over a reserve/commit pair. | |
1304 | */ | |
1305 | v_inc(config, &buf->records_lost_wrap); | |
1306 | return -EIO; | |
1307 | } | |
1308 | offsets->size = | |
1309 | config->cb.record_header_size(config, chan, | |
1310 | offsets->begin, | |
1311 | &offsets->pre_header_padding, | |
1312 | ctx); | |
1313 | offsets->size += | |
1314 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1315 | ctx->largest_align) | |
1316 | + ctx->data_size; | |
1317 | if (unlikely(subbuf_offset(offsets->begin, chan) | |
1318 | + offsets->size > chan->backend.subbuf_size)) { | |
1319 | /* | |
1320 | * Record too big for subbuffers, report error, don't | |
1321 | * complete the sub-buffer switch. | |
1322 | */ | |
1323 | v_inc(config, &buf->records_lost_big); | |
1324 | return -ENOSPC; | |
1325 | } else { | |
1326 | /* | |
1327 | * We just made a successful buffer switch and the | |
1328 | * record fits in the new subbuffer. Let's write. | |
1329 | */ | |
1330 | } | |
1331 | } else { | |
1332 | /* | |
1333 | * Record fits in the current buffer and we are not on a switch | |
1334 | * boundary. It's safe to write. | |
1335 | */ | |
1336 | } | |
1337 | offsets->end = offsets->begin + offsets->size; | |
1338 | ||
1339 | if (unlikely(subbuf_offset(offsets->end, chan) == 0)) { | |
1340 | /* | |
1341 | * The offset_end will fall at the very beginning of the next | |
1342 | * subbuffer. | |
1343 | */ | |
1344 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
1345 | } | |
1346 | return 0; | |
1347 | } | |
1348 | ||
1349 | /** | |
1350 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
1351 | * @ctx: ring buffer context. | |
1352 | * | |
1353 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, | |
1354 | * -EIO for other errors, else returns 0. | |
1355 | * It will take care of sub-buffer switching. | |
1356 | */ | |
1357 | int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx) | |
1358 | { | |
1359 | struct channel *chan = ctx->chan; | |
1d498196 | 1360 | struct shm_handle *handle = ctx->handle; |
852c2936 MD |
1361 | const struct lib_ring_buffer_config *config = chan->backend.config; |
1362 | struct lib_ring_buffer *buf; | |
1363 | struct switch_offsets offsets; | |
1364 | int ret; | |
1365 | ||
1366 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1d498196 | 1367 | buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp); |
852c2936 | 1368 | else |
1d498196 | 1369 | buf = shmp(handle, chan->backend.buf[0].shmp); |
852c2936 MD |
1370 | ctx->buf = buf; |
1371 | ||
1372 | offsets.size = 0; | |
1373 | ||
1374 | do { | |
1375 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, | |
1376 | ctx); | |
1377 | if (unlikely(ret)) | |
1378 | return ret; | |
1379 | } while (unlikely(v_cmpxchg(config, &buf->offset, offsets.old, | |
1380 | offsets.end) | |
1381 | != offsets.old)); | |
1382 | ||
1383 | /* | |
1384 | * Atomically update last_tsc. This update races against concurrent | |
1385 | * atomic updates, but the race will always cause supplementary full TSC | |
1386 | * records, never the opposite (missing a full TSC record when it would | |
1387 | * be needed). | |
1388 | */ | |
1389 | save_last_tsc(config, buf, ctx->tsc); | |
1390 | ||
1391 | /* | |
1392 | * Push the reader if necessary | |
1393 | */ | |
1394 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
1395 | ||
1396 | /* | |
1397 | * Clear noref flag for this subbuffer. | |
1398 | */ | |
1399 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1d498196 MD |
1400 | subbuf_index(offsets.end - 1, chan), |
1401 | handle); | |
852c2936 MD |
1402 | |
1403 | /* | |
1404 | * Switch old subbuffer if needed. | |
1405 | */ | |
1406 | if (unlikely(offsets.switch_old_end)) { | |
1407 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1d498196 MD |
1408 | subbuf_index(offsets.old - 1, chan), |
1409 | handle); | |
1410 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc, handle); | |
852c2936 MD |
1411 | } |
1412 | ||
1413 | /* | |
1414 | * Populate new subbuffer. | |
1415 | */ | |
1416 | if (unlikely(offsets.switch_new_start)) | |
1d498196 | 1417 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 MD |
1418 | |
1419 | if (unlikely(offsets.switch_new_end)) | |
1d498196 | 1420 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc, handle); |
852c2936 MD |
1421 | |
1422 | ctx->slot_size = offsets.size; | |
1423 | ctx->pre_offset = offsets.begin; | |
1424 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
1425 | return 0; | |
1426 | } |