2 * ring_buffer_frontend.c
4 * (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
6 * Ring buffer wait-free buffer synchronization. Producer-consumer and flight
7 * recorder (overwrite) modes. See thesis:
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
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"
19 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
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>
26 * Bob Wisniewski <bob@watson.ibm.com>
28 * Buffer reader semantic :
31 * while buffer is not finalized and empty
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
38 * Dual LGPL v2.1/GPL v2 license.
41 #include <sys/types.h>
45 #include <urcu/compiler.h>
49 #include <ust/ringbuffer-config.h>
55 #define max(a, b) ((a) > (b) ? (a) : (b))
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
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
75 * Internal structure representing offsets to use at a sub-buffer switch.
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,
84 __thread
unsigned int lib_ring_buffer_nesting
;
87 void lib_ring_buffer_print_errors(struct channel
*chan
,
88 struct lib_ring_buffer
*buf
, int cpu
,
89 struct shm_handle
*handle
);
92 * Must be called under cpu hotplug protection.
94 void lib_ring_buffer_free(struct lib_ring_buffer
*buf
,
95 struct shm_handle
*handle
)
97 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
99 lib_ring_buffer_print_errors(chan
, buf
, buf
->backend
.cpu
, handle
);
100 /* buf->commit_hot will be freed by shm teardown */
101 /* buf->commit_cold will be freed by shm teardown */
103 lib_ring_buffer_backend_free(&buf
->backend
);
107 * lib_ring_buffer_reset - Reset ring buffer to initial values.
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.
115 void lib_ring_buffer_reset(struct lib_ring_buffer
*buf
,
116 struct shm_handle
*handle
)
118 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
119 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
123 * Reset iterator first. It will put the subbuffer if it currently holds
126 v_set(config
, &buf
->offset
, 0);
127 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
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);
132 uatomic_set(&buf
->consumed
, 0);
133 uatomic_set(&buf
->record_disabled
, 0);
134 v_set(config
, &buf
->last_tsc
, 0);
135 lib_ring_buffer_backend_reset(&buf
->backend
, handle
);
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);
146 * channel_reset - Reset channel to initial values.
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
154 void channel_reset(struct channel
*chan
)
157 * Reset iterators first. Will put the subbuffer if held for reading.
159 uatomic_set(&chan
->record_disabled
, 0);
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 */
168 * Must be called under cpu hotplug protection.
170 int lib_ring_buffer_create(struct lib_ring_buffer
*buf
,
171 struct channel_backend
*chanb
, int cpu
,
172 struct shm_handle
*handle
,
173 struct shm_object
*shmobj
)
175 const struct lib_ring_buffer_config
*config
= &chanb
->config
;
176 struct channel
*chan
= caa_container_of(chanb
, struct channel
, backend
);
177 void *priv
= chanb
->priv
;
178 unsigned int num_subbuf
;
179 size_t subbuf_header_size
;
183 /* Test for cpu hotplug */
184 if (buf
->backend
.allocated
)
187 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
,
188 cpu
, handle
, shmobj
);
192 align_shm(shmobj
, __alignof__(struct commit_counters_hot
));
193 set_shmp(buf
->commit_hot
,
195 sizeof(struct commit_counters_hot
) * chan
->backend
.num_subbuf
));
196 if (!shmp(handle
, buf
->commit_hot
)) {
201 align_shm(shmobj
, __alignof__(struct commit_counters_cold
));
202 set_shmp(buf
->commit_cold
,
204 sizeof(struct commit_counters_cold
) * chan
->backend
.num_subbuf
));
205 if (!shmp(handle
, buf
->commit_cold
)) {
210 num_subbuf
= chan
->backend
.num_subbuf
;
211 //init_waitqueue_head(&buf->read_wait);
214 * Write the subbuffer header for first subbuffer so we know the total
215 * duration of data gathering.
217 subbuf_header_size
= config
->cb
.subbuffer_header_size();
218 v_set(config
, &buf
->offset
, subbuf_header_size
);
219 subbuffer_id_clear_noref(config
, &shmp_index(handle
, buf
->backend
.buf_wsb
, 0)->id
);
220 tsc
= config
->cb
.ring_buffer_clock_read(shmp(handle
, buf
->backend
.chan
));
221 config
->cb
.buffer_begin(buf
, tsc
, 0, handle
);
222 v_add(config
, subbuf_header_size
, &shmp_index(handle
, buf
->commit_hot
, 0)->cc
);
224 if (config
->cb
.buffer_create
) {
225 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
, handle
);
229 buf
->backend
.allocated
= 1;
234 /* commit_cold will be freed by shm teardown */
236 /* commit_hot will be freed by shm teardown */
238 lib_ring_buffer_backend_free(&buf
->backend
);
243 static void switch_buffer_timer(unsigned long data
)
245 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
246 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
247 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
250 * Only flush buffers periodically if readers are active.
252 if (uatomic_read(&buf
->active_readers
) || uatomic_read(&buf
->active_shadow_readers
))
253 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
, handle
);
256 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
257 // mod_timer_pinned(&buf->switch_timer,
258 // jiffies + chan->switch_timer_interval);
260 // mod_timer(&buf->switch_timer,
261 // jiffies + chan->switch_timer_interval);
265 static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer
*buf
,
266 struct shm_handle
*handle
)
268 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
269 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
271 if (!chan
->switch_timer_interval
|| buf
->switch_timer_enabled
)
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);
281 // add_timer(&buf->switch_timer);
282 buf
->switch_timer_enabled
= 1;
285 static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer
*buf
,
286 struct shm_handle
*handle
)
288 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
290 if (!chan
->switch_timer_interval
|| !buf
->switch_timer_enabled
)
294 //del_timer_sync(&buf->switch_timer);
295 buf
->switch_timer_enabled
= 0;
300 * Polling timer to check the channels for data.
302 static void read_buffer_timer(unsigned long data
)
304 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
305 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
306 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
308 CHAN_WARN_ON(chan
, !buf
->backend
.allocated
);
310 if (uatomic_read(&buf
->active_readers
) || uatomic_read(&buf
->active_shadow_readers
))
311 && lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
313 //wake_up_interruptible(&buf->read_wait);
314 //wake_up_interruptible(&chan->read_wait);
318 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
319 // mod_timer_pinned(&buf->read_timer,
320 // jiffies + chan->read_timer_interval);
322 // mod_timer(&buf->read_timer,
323 // jiffies + chan->read_timer_interval);
327 static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer
*buf
,
328 struct shm_handle
*handle
)
330 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
331 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
333 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
334 || !chan
->read_timer_interval
335 || buf
->read_timer_enabled
)
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;
344 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
345 // add_timer_on(&buf->read_timer, buf->backend.cpu);
347 // add_timer(&buf->read_timer);
348 buf
->read_timer_enabled
= 1;
351 static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer
*buf
,
352 struct shm_handle
*handle
)
354 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
355 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
357 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
358 || !chan
->read_timer_interval
359 || !buf
->read_timer_enabled
)
363 //del_timer_sync(&buf->read_timer);
365 * do one more check to catch data that has been written in the last
368 if (lib_ring_buffer_poll_deliver(config
, buf
, chan
, handle
)) {
370 //wake_up_interruptible(&buf->read_wait);
371 //wake_up_interruptible(&chan->read_wait);
373 buf
->read_timer_enabled
= 0;
376 static void channel_unregister_notifiers(struct channel
*chan
,
377 struct shm_handle
*handle
)
379 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
382 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
383 for_each_possible_cpu(cpu
) {
384 struct lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
386 lib_ring_buffer_stop_switch_timer(buf
, handle
);
387 lib_ring_buffer_stop_read_timer(buf
, handle
);
390 struct lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
392 lib_ring_buffer_stop_switch_timer(buf
, handle
);
393 lib_ring_buffer_stop_read_timer(buf
, handle
);
395 //channel_backend_unregister_notifiers(&chan->backend);
398 static void channel_free(struct channel
*chan
, struct shm_handle
*handle
,
404 channel_backend_free(&chan
->backend
, handle
);
405 /* chan is freed by shm teardown */
406 shm_object_table_destroy(handle
->table
);
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.
426 * Returns NULL on failure.
428 struct shm_handle
*channel_create(const struct lib_ring_buffer_config
*config
,
429 const char *name
, void *priv
, void *buf_addr
,
431 size_t num_subbuf
, unsigned int switch_timer_interval
,
432 unsigned int read_timer_interval
,
433 int *shm_fd
, int *wait_fd
, uint64_t *memory_map_size
)
437 struct channel
*chan
;
438 struct shm_handle
*handle
;
439 struct shm_object
*shmobj
;
442 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
443 read_timer_interval
))
446 handle
= zmalloc(sizeof(struct shm_handle
));
450 /* Allocate table for channel + per-cpu buffers */
451 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
453 goto error_table_alloc
;
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();
460 shmsize
+= sizeof(struct lib_ring_buffer_shmp
);
462 shmobj
= shm_object_table_append(handle
->table
, shmsize
);
465 set_shmp(handle
->chan
, zalloc_shm(shmobj
, sizeof(struct channel
)));
466 chan
= shmp(handle
, handle
->chan
);
470 ret
= channel_backend_init(&chan
->backend
, name
, config
, priv
,
471 subbuf_size
, num_subbuf
, handle
);
473 goto error_backend_init
;
475 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
477 //chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval);
478 //chan->read_timer_interval = usecs_to_jiffies(read_timer_interval);
480 //init_waitqueue_head(&chan->read_wait);
481 //init_waitqueue_head(&chan->hp_wait);
483 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
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.
489 for_each_possible_cpu(cpu
) {
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
);
495 struct lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
497 lib_ring_buffer_start_switch_timer(buf
, handle
);
498 lib_ring_buffer_start_read_timer(buf
, handle
);
500 ref
= &handle
->chan
._ref
;
501 shm_get_object_data(handle
, ref
, shm_fd
, wait_fd
, memory_map_size
);
506 shm_object_table_destroy(handle
->table
);
512 struct shm_handle
*channel_handle_create(int shm_fd
, int wait_fd
,
513 uint64_t memory_map_size
)
515 struct shm_handle
*handle
;
516 struct shm_object
*object
;
518 handle
= zmalloc(sizeof(struct shm_handle
));
522 /* Allocate table for channel + per-cpu buffers */
523 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
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
);
530 goto error_table_object
;
535 shm_object_table_destroy(handle
->table
);
541 int channel_handle_add_stream(struct shm_handle
*handle
,
542 int shm_fd
, int wait_fd
, uint64_t memory_map_size
)
544 struct shm_object
*object
;
546 /* Add stream object */
547 object
= shm_object_table_append_shadow(handle
->table
,
548 shm_fd
, wait_fd
, memory_map_size
);
555 void channel_release(struct channel
*chan
, struct shm_handle
*handle
,
558 channel_free(chan
, handle
, shadow
);
562 * channel_destroy - Finalize, wait for q.s. and destroy channel.
563 * @chan: channel to destroy
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
572 void *channel_destroy(struct channel
*chan
, struct shm_handle
*handle
,
575 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
580 channel_release(chan
, handle
, shadow
);
584 channel_unregister_notifiers(chan
, handle
);
586 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
587 for_each_channel_cpu(cpu
, chan
) {
588 struct lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
590 if (config
->cb
.buffer_finalize
)
591 config
->cb
.buffer_finalize(buf
,
594 if (buf
->backend
.allocated
)
595 lib_ring_buffer_switch_slow(buf
, SWITCH_FLUSH
,
598 * Perform flush before writing to finalized.
601 CMM_ACCESS_ONCE(buf
->finalized
) = 1;
602 //wake_up_interruptible(&buf->read_wait);
605 struct lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
607 if (config
->cb
.buffer_finalize
)
608 config
->cb
.buffer_finalize(buf
, chan
->backend
.priv
, -1, handle
);
609 if (buf
->backend
.allocated
)
610 lib_ring_buffer_switch_slow(buf
, SWITCH_FLUSH
,
613 * Perform flush before writing to finalized.
616 CMM_ACCESS_ONCE(buf
->finalized
) = 1;
617 //wake_up_interruptible(&buf->read_wait);
619 CMM_ACCESS_ONCE(chan
->finalized
) = 1;
620 //wake_up_interruptible(&chan->hp_wait);
621 //wake_up_interruptible(&chan->read_wait);
623 * sessiond/consumer are keeping a reference on the shm file
624 * descriptor directly. No need to refcount.
626 priv
= chan
->backend
.priv
;
627 channel_release(chan
, handle
, shadow
);
631 struct lib_ring_buffer
*channel_get_ring_buffer(
632 const struct lib_ring_buffer_config
*config
,
633 struct channel
*chan
, int cpu
,
634 struct shm_handle
*handle
,
635 int *shm_fd
, int *wait_fd
,
636 uint64_t *memory_map_size
)
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
,
644 return shmp(handle
, chan
->backend
.buf
[0].shmp
);
646 if (cpu
>= num_possible_cpus())
648 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
649 shm_get_object_data(handle
, ref
, shm_fd
, wait_fd
,
651 return shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
655 int lib_ring_buffer_open_read(struct lib_ring_buffer
*buf
,
656 struct shm_handle
*handle
,
659 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
662 if (uatomic_cmpxchg(&buf
->active_shadow_readers
, 0, 1) != 0)
667 if (uatomic_cmpxchg(&buf
->active_readers
, 0, 1) != 0)
673 void lib_ring_buffer_release_read(struct lib_ring_buffer
*buf
,
674 struct shm_handle
*handle
,
677 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
680 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_shadow_readers
) != 1);
682 uatomic_dec(&buf
->active_shadow_readers
);
685 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
687 uatomic_dec(&buf
->active_readers
);
691 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
693 * @consumed: consumed count indicating the position where to read
694 * @produced: produced count, indicates position when to stop reading
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.
700 int lib_ring_buffer_snapshot(struct lib_ring_buffer
*buf
,
701 unsigned long *consumed
, unsigned long *produced
,
702 struct shm_handle
*handle
)
704 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
705 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
706 unsigned long consumed_cur
, write_offset
;
709 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
711 * Read finalized before counters.
714 consumed_cur
= uatomic_read(&buf
->consumed
);
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).
723 write_offset
= v_read(config
, &buf
->offset
);
726 * Check that we are not about to read the same subbuffer in
727 * which the writer head is.
729 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
733 *consumed
= consumed_cur
;
734 *produced
= subbuf_trunc(write_offset
, chan
);
740 * The memory barriers __wait_event()/wake_up_interruptible() take care
741 * of "raw_spin_is_locked" memory ordering.
750 * lib_ring_buffer_put_snapshot - move consumed counter forward
752 * @consumed_new: new consumed count value
754 void lib_ring_buffer_move_consumer(struct lib_ring_buffer
*buf
,
755 unsigned long consumed_new
,
756 struct shm_handle
*handle
)
758 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
759 struct channel
*chan
= shmp(handle
, bufb
->chan
);
760 unsigned long consumed
;
762 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1
763 && uatomic_read(&buf
->active_shadow_readers
) != 1);
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.
770 consumed
= uatomic_read(&buf
->consumed
);
771 while ((long) consumed
- (long) consumed_new
< 0)
772 consumed
= uatomic_cmpxchg(&buf
->consumed
, consumed
,
777 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
779 * @consumed: consumed count indicating the position where to read
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.
784 int lib_ring_buffer_get_subbuf(struct lib_ring_buffer
*buf
,
785 unsigned long consumed
,
786 struct shm_handle
*handle
)
788 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
789 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
790 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
795 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
797 * Read finalized before counters.
800 consumed_cur
= uatomic_read(&buf
->consumed
);
801 consumed_idx
= subbuf_index(consumed
, chan
);
802 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, consumed_idx
)->cc_sb
);
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.
810 * Local rmb to match the remote wmb to read the commit count
811 * before the buffer data and the write offset.
815 write_offset
= v_read(config
, &buf
->offset
);
818 * Check that the buffer we are getting is after or at consumed_cur
821 if ((long) subbuf_trunc(consumed
, chan
)
822 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
826 * Check that the subbuffer we are trying to consume has been
827 * already fully committed.
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
)
837 * Check that we are not about to read the same subbuffer in
838 * which the writer head is.
840 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
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.
851 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
852 consumed_idx
, buf_trunc_val(consumed
, chan
),
856 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
858 buf
->get_subbuf_consumed
= consumed
;
865 * The memory barriers __wait_event()/wake_up_interruptible() take care
866 * of "raw_spin_is_locked" memory ordering.
875 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
878 void lib_ring_buffer_put_subbuf(struct lib_ring_buffer
*buf
,
879 struct shm_handle
*handle
)
881 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
882 struct channel
*chan
= shmp(handle
, bufb
->chan
);
883 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
884 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
886 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1
887 && uatomic_read(&buf
->active_shadow_readers
) != 1);
889 if (!buf
->get_subbuf
) {
891 * Reader puts a subbuffer it did not get.
893 CHAN_WARN_ON(chan
, 1);
896 consumed
= buf
->get_subbuf_consumed
;
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
906 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
907 v_add(config
, v_read(config
,
908 &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
),
909 &bufb
->records_read
);
910 v_set(config
, &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
, 0);
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
);
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.
923 consumed_idx
= subbuf_index(consumed
, chan
);
924 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
925 consumed_idx
, buf_trunc_val(consumed
, chan
),
928 * update_read_sb_index return value ignored. Don't exchange sub-buffer
929 * if the writer concurrently updated it.
934 * cons_offset is an iterator on all subbuffer offsets between the reader
935 * position and the writer position. (inclusive)
938 void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer
*buf
,
939 struct channel
*chan
,
940 unsigned long cons_offset
,
942 struct shm_handle
*handle
)
944 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
945 unsigned long cons_idx
, commit_count
, commit_count_sb
;
947 cons_idx
= subbuf_index(cons_offset
, chan
);
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
);
951 if (subbuf_offset(commit_count
, chan
) != 0)
952 ERRMSG("ring buffer %s, cpu %d: "
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
);
960 ERRMSG("ring buffer: %s, cpu %d: %lu bytes committed\n",
961 chan
->backend
.name
, cpu
, commit_count
);
965 void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer
*buf
,
966 struct channel
*chan
,
968 struct shm_handle
*handle
)
970 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
971 unsigned long write_offset
, cons_offset
;
974 * Can be called in the error path of allocation when
975 * trans_channel_data is not yet set.
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.
984 write_offset
= v_read(config
, &buf
->offset
);
985 cons_offset
= uatomic_read(&buf
->consumed
);
986 if (write_offset
!= cons_offset
)
987 ERRMSG("ring buffer %s, cpu %d: "
988 "non-consumed data\n"
989 " [ %lu bytes written, %lu bytes read ]\n",
990 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
992 for (cons_offset
= uatomic_read(&buf
->consumed
);
993 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
996 cons_offset
= subbuf_align(cons_offset
, chan
))
997 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1002 void lib_ring_buffer_print_errors(struct channel
*chan
,
1003 struct lib_ring_buffer
*buf
, int cpu
,
1004 struct shm_handle
*handle
)
1006 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1007 void *priv
= chan
->backend
.priv
;
1009 ERRMSG("ring buffer %s, cpu %d: %lu records written, "
1010 "%lu records overrun\n",
1011 chan
->backend
.name
, cpu
,
1012 v_read(config
, &buf
->records_count
),
1013 v_read(config
, &buf
->records_overrun
));
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
))
1018 ERRMSG("ring buffer %s, cpu %d: records were lost. Caused by:\n"
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
));
1026 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
, handle
);
1030 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1032 * Only executed when the buffer is finalized, in SWITCH_FLUSH.
1035 void lib_ring_buffer_switch_old_start(struct lib_ring_buffer
*buf
,
1036 struct channel
*chan
,
1037 struct switch_offsets
*offsets
,
1039 struct shm_handle
*handle
)
1041 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1042 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1043 unsigned long commit_count
;
1045 config
->cb
.buffer_begin(buf
, tsc
, oldidx
, handle
);
1048 * Order all writes to buffer before the commit count update that will
1049 * determine that the subbuffer is full.
1052 v_add(config
, config
->cb
.subbuffer_header_size(),
1053 &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1054 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1055 /* Check if the written buffer has to be delivered */
1056 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1057 commit_count
, oldidx
, handle
);
1058 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1059 offsets
->old
, commit_count
,
1060 config
->cb
.subbuffer_header_size(),
1065 * lib_ring_buffer_switch_old_end: switch old subbuffer
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
1073 void lib_ring_buffer_switch_old_end(struct lib_ring_buffer
*buf
,
1074 struct channel
*chan
,
1075 struct switch_offsets
*offsets
,
1077 struct shm_handle
*handle
)
1079 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1080 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1081 unsigned long commit_count
, padding_size
, data_size
;
1083 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1084 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1085 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
,
1089 * Order all writes to buffer before the commit count update that will
1090 * determine that the subbuffer is full.
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
);
1095 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1096 commit_count
, oldidx
, handle
);
1097 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1098 offsets
->old
, commit_count
,
1099 padding_size
, handle
);
1103 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
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.
1110 void lib_ring_buffer_switch_new_start(struct lib_ring_buffer
*buf
,
1111 struct channel
*chan
,
1112 struct switch_offsets
*offsets
,
1114 struct shm_handle
*handle
)
1116 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1117 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1118 unsigned long commit_count
;
1120 config
->cb
.buffer_begin(buf
, tsc
, beginidx
, handle
);
1123 * Order all writes to buffer before the commit count update that will
1124 * determine that the subbuffer is full.
1127 v_add(config
, config
->cb
.subbuffer_header_size(),
1128 &shmp_index(handle
, buf
->commit_hot
, beginidx
)->cc
);
1129 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, beginidx
)->cc
);
1130 /* Check if the written buffer has to be delivered */
1131 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1132 commit_count
, beginidx
, handle
);
1133 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, beginidx
,
1134 offsets
->begin
, commit_count
,
1135 config
->cb
.subbuffer_header_size(),
1140 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1142 * The only remaining threads could be the ones with pending commits. They will
1143 * have to do the deliver themselves.
1146 void lib_ring_buffer_switch_new_end(struct lib_ring_buffer
*buf
,
1147 struct channel
*chan
,
1148 struct switch_offsets
*offsets
,
1150 struct shm_handle
*handle
)
1152 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1153 unsigned long endidx
= subbuf_index(offsets
->end
- 1, chan
);
1154 unsigned long commit_count
, padding_size
, data_size
;
1156 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1157 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1158 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
,
1162 * Order all writes to buffer before the commit count update that will
1163 * determine that the subbuffer is full.
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
);
1168 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->end
- 1,
1169 commit_count
, endidx
, handle
);
1170 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, endidx
,
1171 offsets
->end
, commit_count
,
1172 padding_size
, handle
);
1178 * !0 if execution must be aborted.
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
,
1187 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
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
);
1195 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
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
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.
1213 if (mode
== SWITCH_FLUSH
|| off
> 0) {
1214 if (unlikely(off
== 0)) {
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.
1220 if (!config
->cb
.subbuffer_header_size())
1223 * Need to write the subbuffer start header on finalize.
1225 offsets
->switch_old_start
= 1;
1227 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
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
;
1236 * Force a sub-buffer switch. This operation is completely reentrant : can be
1237 * called while tracing is active with absolutely no lock held.
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.
1243 void lib_ring_buffer_switch_slow(struct lib_ring_buffer
*buf
, enum switch_mode mode
,
1244 struct shm_handle
*handle
)
1246 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1247 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1248 struct switch_offsets offsets
;
1249 unsigned long oldidx
;
1255 * Perform retryable operations.
1258 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1260 return; /* Switch not needed */
1261 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
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
1270 save_last_tsc(config
, buf
, tsc
);
1273 * Push the reader if necessary
1275 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1277 oldidx
= subbuf_index(offsets
.old
, chan
);
1278 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
, handle
);
1281 * May need to populate header start on SWITCH_FLUSH.
1283 if (offsets
.switch_old_start
) {
1284 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
, handle
);
1285 offsets
.old
+= config
->cb
.subbuffer_header_size();
1289 * Switch old subbuffer.
1291 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
, handle
);
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.
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
)
1307 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1308 struct shm_handle
*handle
= ctx
->handle
;
1309 unsigned long reserve_commit_diff
;
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;
1318 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1319 if ((int64_t) ctx
->tsc
== -EIO
)
1322 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1323 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
1325 if (unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1326 offsets
->switch_new_start
= 1; /* For offsets->begin */
1328 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1330 &offsets
->pre_header_padding
,
1333 lib_ring_buffer_align(offsets
->begin
+ offsets
->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 */
1342 if (unlikely(offsets
->switch_new_start
)) {
1343 unsigned long sb_index
;
1346 * We are typically not filling the previous buffer completely.
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
,
1358 &shmp_index(handle
, buf
->commit_cold
, sb_index
)->cc_sb
)
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)
1365 uatomic_read(&buf
->consumed
), chan
)
1366 >= chan
->backend
.buf_size
)) {
1368 * We do not overwrite non consumed buffers
1369 * and we are full : record is lost.
1371 v_inc(config
, &buf
->records_lost_full
);
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
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.
1388 v_inc(config
, &buf
->records_lost_wrap
);
1392 config
->cb
.record_header_size(config
, chan
,
1394 &offsets
->pre_header_padding
,
1397 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1400 if (unlikely(subbuf_offset(offsets
->begin
, chan
)
1401 + offsets
->size
> chan
->backend
.subbuf_size
)) {
1403 * Record too big for subbuffers, report error, don't
1404 * complete the sub-buffer switch.
1406 v_inc(config
, &buf
->records_lost_big
);
1410 * We just made a successful buffer switch and the
1411 * record fits in the new subbuffer. Let's write.
1416 * Record fits in the current buffer and we are not on a switch
1417 * boundary. It's safe to write.
1420 offsets
->end
= offsets
->begin
+ offsets
->size
;
1422 if (unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
1424 * The offset_end will fall at the very beginning of the next
1427 offsets
->switch_new_end
= 1; /* For offsets->begin */
1433 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1434 * @ctx: ring buffer context.
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.
1440 int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx
*ctx
)
1442 struct channel
*chan
= ctx
->chan
;
1443 struct shm_handle
*handle
= ctx
->handle
;
1444 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1445 struct lib_ring_buffer
*buf
;
1446 struct switch_offsets offsets
;
1449 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
1450 buf
= shmp(handle
, chan
->backend
.buf
[ctx
->cpu
].shmp
);
1452 buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
1458 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
1462 } while (unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
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
1472 save_last_tsc(config
, buf
, ctx
->tsc
);
1475 * Push the reader if necessary
1477 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
1480 * Clear noref flag for this subbuffer.
1482 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1483 subbuf_index(offsets
.end
- 1, chan
),
1487 * Switch old subbuffer if needed.
1489 if (unlikely(offsets
.switch_old_end
)) {
1490 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1491 subbuf_index(offsets
.old
- 1, chan
),
1493 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1497 * Populate new subbuffer.
1499 if (unlikely(offsets
.switch_new_start
))
1500 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1502 if (unlikely(offsets
.switch_new_end
))
1503 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1505 ctx
->slot_size
= offsets
.size
;
1506 ctx
->pre_offset
= offsets
.begin
;
1507 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;