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.
42 #include <sys/types.h>
46 #include <urcu/compiler.h>
51 #include <lttng/ringbuffer-config.h>
56 #include "../liblttng-ust/compat.h" /* For ENODATA */
59 #define max(a, b) ((a) > (b) ? (a) : (b))
63 * Use POSIX SHM: shm_open(3) and shm_unlink(3).
64 * close(2) to close the fd returned by shm_open.
65 * shm_unlink releases the shared memory object name.
66 * ftruncate(2) sets the size of the memory object.
67 * mmap/munmap maps the shared memory obj to a virtual address in the
68 * calling proceess (should be done both in libust and consumer).
69 * See shm_overview(7) for details.
70 * Pass file descriptor returned by shm_open(3) to ltt-sessiond through
73 * Since we don't need to access the object using its name, we can
74 * immediately shm_unlink(3) it, and only keep the handle with its file
79 * Internal structure representing offsets to use at a sub-buffer switch.
81 struct switch_offsets
{
82 unsigned long begin
, end
, old
;
83 size_t pre_header_padding
, size
;
84 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
88 __thread
unsigned int lib_ring_buffer_nesting
;
91 * TODO: this is unused. Errors are saved within the ring buffer.
92 * Eventually, allow consumerd to print these errors.
95 void lib_ring_buffer_print_errors(struct channel
*chan
,
96 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
97 struct lttng_ust_shm_handle
*handle
)
98 __attribute__((unused
));
101 * lib_ring_buffer_reset - Reset ring buffer to initial values.
104 * Effectively empty the ring buffer. Should be called when the buffer is not
105 * used for writing. The ring buffer can be opened for reading, but the reader
106 * should not be using the iterator concurrently with reset. The previous
107 * current iterator record is reset.
109 void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer
*buf
,
110 struct lttng_ust_shm_handle
*handle
)
112 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
113 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
117 * Reset iterator first. It will put the subbuffer if it currently holds
120 v_set(config
, &buf
->offset
, 0);
121 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
122 v_set(config
, &shmp_index(handle
, buf
->commit_hot
, i
)->cc
, 0);
123 v_set(config
, &shmp_index(handle
, buf
->commit_hot
, i
)->seq
, 0);
124 v_set(config
, &shmp_index(handle
, buf
->commit_cold
, i
)->cc_sb
, 0);
126 uatomic_set(&buf
->consumed
, 0);
127 uatomic_set(&buf
->record_disabled
, 0);
128 v_set(config
, &buf
->last_tsc
, 0);
129 lib_ring_buffer_backend_reset(&buf
->backend
, handle
);
130 /* Don't reset number of active readers */
131 v_set(config
, &buf
->records_lost_full
, 0);
132 v_set(config
, &buf
->records_lost_wrap
, 0);
133 v_set(config
, &buf
->records_lost_big
, 0);
134 v_set(config
, &buf
->records_count
, 0);
135 v_set(config
, &buf
->records_overrun
, 0);
140 * channel_reset - Reset channel to initial values.
143 * Effectively empty the channel. Should be called when the channel is not used
144 * for writing. The channel can be opened for reading, but the reader should not
145 * be using the iterator concurrently with reset. The previous current iterator
148 void channel_reset(struct channel
*chan
)
151 * Reset iterators first. Will put the subbuffer if held for reading.
153 uatomic_set(&chan
->record_disabled
, 0);
154 /* Don't reset commit_count_mask, still valid */
155 channel_backend_reset(&chan
->backend
);
156 /* Don't reset switch/read timer interval */
157 /* Don't reset notifiers and notifier enable bits */
158 /* Don't reset reader reference count */
162 * Must be called under cpu hotplug protection.
164 int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer
*buf
,
165 struct channel_backend
*chanb
, int cpu
,
166 struct lttng_ust_shm_handle
*handle
,
167 struct shm_object
*shmobj
)
169 const struct lttng_ust_lib_ring_buffer_config
*config
= &chanb
->config
;
170 struct channel
*chan
= caa_container_of(chanb
, struct channel
, backend
);
171 void *priv
= channel_get_private(chan
);
172 size_t subbuf_header_size
;
176 /* Test for cpu hotplug */
177 if (buf
->backend
.allocated
)
180 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
,
181 cpu
, handle
, shmobj
);
185 align_shm(shmobj
, __alignof__(struct commit_counters_hot
));
186 set_shmp(buf
->commit_hot
,
188 sizeof(struct commit_counters_hot
) * chan
->backend
.num_subbuf
));
189 if (!shmp(handle
, buf
->commit_hot
)) {
194 align_shm(shmobj
, __alignof__(struct commit_counters_cold
));
195 set_shmp(buf
->commit_cold
,
197 sizeof(struct commit_counters_cold
) * chan
->backend
.num_subbuf
));
198 if (!shmp(handle
, buf
->commit_cold
)) {
204 * Write the subbuffer header for first subbuffer so we know the total
205 * duration of data gathering.
207 subbuf_header_size
= config
->cb
.subbuffer_header_size();
208 v_set(config
, &buf
->offset
, subbuf_header_size
);
209 subbuffer_id_clear_noref(config
, &shmp_index(handle
, buf
->backend
.buf_wsb
, 0)->id
);
210 tsc
= config
->cb
.ring_buffer_clock_read(shmp(handle
, buf
->backend
.chan
));
211 config
->cb
.buffer_begin(buf
, tsc
, 0, handle
);
212 v_add(config
, subbuf_header_size
, &shmp_index(handle
, buf
->commit_hot
, 0)->cc
);
214 if (config
->cb
.buffer_create
) {
215 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
, handle
);
219 buf
->backend
.allocated
= 1;
224 /* commit_cold will be freed by shm teardown */
226 /* commit_hot will be freed by shm teardown */
232 static void switch_buffer_timer(unsigned long data
)
234 struct lttng_ust_lib_ring_buffer
*buf
= (struct lttng_ust_lib_ring_buffer
*)data
;
235 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
236 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
239 * Only flush buffers periodically if readers are active.
241 if (uatomic_read(&buf
->active_readers
) || uatomic_read(&buf
->active_shadow_readers
))
242 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
, handle
);
245 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
246 // mod_timer_pinned(&buf->switch_timer,
247 // jiffies + chan->switch_timer_interval);
249 // mod_timer(&buf->switch_timer,
250 // jiffies + chan->switch_timer_interval);
254 static void lib_ring_buffer_start_switch_timer(struct lttng_ust_lib_ring_buffer
*buf
,
255 struct lttng_ust_shm_handle
*handle
)
257 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
258 //const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
260 if (!chan
->switch_timer_interval
|| buf
->switch_timer_enabled
)
263 //init_timer(&buf->switch_timer);
264 //buf->switch_timer.function = switch_buffer_timer;
265 //buf->switch_timer.expires = jiffies + chan->switch_timer_interval;
266 //buf->switch_timer.data = (unsigned long)buf;
267 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
268 // add_timer_on(&buf->switch_timer, buf->backend.cpu);
270 // add_timer(&buf->switch_timer);
271 buf
->switch_timer_enabled
= 1;
274 static void lib_ring_buffer_stop_switch_timer(struct lttng_ust_lib_ring_buffer
*buf
,
275 struct lttng_ust_shm_handle
*handle
)
277 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
279 if (!chan
->switch_timer_interval
|| !buf
->switch_timer_enabled
)
283 //del_timer_sync(&buf->switch_timer);
284 buf
->switch_timer_enabled
= 0;
289 * Polling timer to check the channels for data.
291 static void read_buffer_timer(unsigned long data
)
293 struct lttng_ust_lib_ring_buffer
*buf
= (struct lttng_ust_lib_ring_buffer
*)data
;
294 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
295 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
297 CHAN_WARN_ON(chan
, !buf
->backend
.allocated
);
299 if (uatomic_read(&buf
->active_readers
) || uatomic_read(&buf
->active_shadow_readers
))
300 && lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
302 //wake_up_interruptible(&buf->read_wait);
303 //wake_up_interruptible(&chan->read_wait);
307 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
308 // mod_timer_pinned(&buf->read_timer,
309 // jiffies + chan->read_timer_interval);
311 // mod_timer(&buf->read_timer,
312 // jiffies + chan->read_timer_interval);
316 static void lib_ring_buffer_start_read_timer(struct lttng_ust_lib_ring_buffer
*buf
,
317 struct lttng_ust_shm_handle
*handle
)
319 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
320 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
322 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
323 || !chan
->read_timer_interval
324 || buf
->read_timer_enabled
)
328 //init_timer(&buf->read_timer);
329 //buf->read_timer.function = read_buffer_timer;
330 //buf->read_timer.expires = jiffies + chan->read_timer_interval;
331 //buf->read_timer.data = (unsigned long)buf;
333 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
334 // add_timer_on(&buf->read_timer, buf->backend.cpu);
336 // add_timer(&buf->read_timer);
337 buf
->read_timer_enabled
= 1;
340 static void lib_ring_buffer_stop_read_timer(struct lttng_ust_lib_ring_buffer
*buf
,
341 struct lttng_ust_shm_handle
*handle
)
343 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
344 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
346 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
347 || !chan
->read_timer_interval
348 || !buf
->read_timer_enabled
)
352 //del_timer_sync(&buf->read_timer);
354 * do one more check to catch data that has been written in the last
357 if (lib_ring_buffer_poll_deliver(config
, buf
, chan
, handle
)) {
359 //wake_up_interruptible(&buf->read_wait);
360 //wake_up_interruptible(&chan->read_wait);
362 buf
->read_timer_enabled
= 0;
365 static void channel_unregister_notifiers(struct channel
*chan
,
366 struct lttng_ust_shm_handle
*handle
)
368 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
371 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
372 for_each_possible_cpu(cpu
) {
373 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
375 lib_ring_buffer_stop_switch_timer(buf
, handle
);
376 lib_ring_buffer_stop_read_timer(buf
, handle
);
379 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
381 lib_ring_buffer_stop_switch_timer(buf
, handle
);
382 lib_ring_buffer_stop_read_timer(buf
, handle
);
384 //channel_backend_unregister_notifiers(&chan->backend);
387 static void channel_free(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
391 channel_backend_free(&chan
->backend
, handle
);
392 /* chan is freed by shm teardown */
393 shm_object_table_destroy(handle
->table
);
398 * channel_create - Create channel.
399 * @config: ring buffer instance configuration
400 * @name: name of the channel
401 * @priv_data: ring buffer client private data area pointer (output)
402 * @priv_data_size: length, in bytes, of the private data area.
403 * @priv_data_init: initialization data for private data.
404 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
405 * address mapping. It is used only by RING_BUFFER_STATIC
406 * configuration. It can be set to NULL for other backends.
407 * @subbuf_size: subbuffer size
408 * @num_subbuf: number of subbuffers
409 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
410 * padding to let readers get those sub-buffers.
411 * Used for live streaming.
412 * @read_timer_interval: Time interval (in us) to wake up pending readers.
415 * Returns NULL on failure.
417 struct lttng_ust_shm_handle
*channel_create(const struct lttng_ust_lib_ring_buffer_config
*config
,
420 size_t priv_data_align
,
421 size_t priv_data_size
,
422 void *priv_data_init
,
423 void *buf_addr
, size_t subbuf_size
,
424 size_t num_subbuf
, unsigned int switch_timer_interval
,
425 unsigned int read_timer_interval
,
426 int **shm_fd
, int **wait_fd
, uint64_t **memory_map_size
)
429 size_t shmsize
, chansize
;
430 struct channel
*chan
;
431 struct lttng_ust_shm_handle
*handle
;
432 struct shm_object
*shmobj
;
435 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
436 read_timer_interval
))
439 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
443 /* Allocate table for channel + per-cpu buffers */
444 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
446 goto error_table_alloc
;
448 /* Calculate the shm allocation layout */
449 shmsize
= sizeof(struct channel
);
450 shmsize
+= offset_align(shmsize
, __alignof__(struct lttng_ust_lib_ring_buffer_shmp
));
451 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
452 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
) * num_possible_cpus();
454 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
);
456 shmsize
+= offset_align(shmsize
, priv_data_align
);
457 shmsize
+= priv_data_size
;
459 shmobj
= shm_object_table_append(handle
->table
, shmsize
);
462 /* struct channel is at object 0, offset 0 (hardcoded) */
463 set_shmp(handle
->chan
, zalloc_shm(shmobj
, chansize
));
464 assert(handle
->chan
._ref
.index
== 0);
465 assert(handle
->chan
._ref
.offset
== 0);
466 chan
= shmp(handle
, handle
->chan
);
470 /* space for private data */
471 if (priv_data_size
) {
472 DECLARE_SHMP(void, priv_data_alloc
);
474 align_shm(shmobj
, priv_data_align
);
475 chan
->priv_data_offset
= shmobj
->allocated_len
;
476 set_shmp(priv_data_alloc
, zalloc_shm(shmobj
, priv_data_size
));
477 if (!shmp(handle
, priv_data_alloc
))
479 *priv_data
= channel_get_private(chan
);
480 memcpy(*priv_data
, priv_data_init
, priv_data_size
);
482 chan
->priv_data_offset
= -1;
486 ret
= channel_backend_init(&chan
->backend
, name
, config
,
487 subbuf_size
, num_subbuf
, handle
);
489 goto error_backend_init
;
491 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
493 //chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval);
494 //chan->read_timer_interval = usecs_to_jiffies(read_timer_interval);
496 //init_waitqueue_head(&chan->read_wait);
497 //init_waitqueue_head(&chan->hp_wait);
499 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
501 * In case of non-hotplug cpu, if the ring-buffer is allocated
502 * in early initcall, it will not be notified of secondary cpus.
503 * In that off case, we need to allocate for all possible cpus.
505 for_each_possible_cpu(cpu
) {
506 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
507 lib_ring_buffer_start_switch_timer(buf
, handle
);
508 lib_ring_buffer_start_read_timer(buf
, handle
);
511 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
513 lib_ring_buffer_start_switch_timer(buf
, handle
);
514 lib_ring_buffer_start_read_timer(buf
, handle
);
516 ref
= &handle
->chan
._ref
;
517 shm_get_object_data(handle
, ref
, shm_fd
, wait_fd
, memory_map_size
);
522 shm_object_table_destroy(handle
->table
);
528 struct lttng_ust_shm_handle
*channel_handle_create(int shm_fd
, int wait_fd
,
529 uint64_t memory_map_size
)
531 struct lttng_ust_shm_handle
*handle
;
532 struct shm_object
*object
;
534 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
538 /* Allocate table for channel + per-cpu buffers */
539 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
541 goto error_table_alloc
;
542 /* Add channel object */
543 object
= shm_object_table_append_shadow(handle
->table
,
544 shm_fd
, wait_fd
, memory_map_size
);
546 goto error_table_object
;
547 /* struct channel is at object 0, offset 0 (hardcoded) */
548 handle
->chan
._ref
.index
= 0;
549 handle
->chan
._ref
.offset
= 0;
553 shm_object_table_destroy(handle
->table
);
559 int channel_handle_add_stream(struct lttng_ust_shm_handle
*handle
,
560 int shm_fd
, int wait_fd
, uint64_t memory_map_size
)
562 struct shm_object
*object
;
564 /* Add stream object */
565 object
= shm_object_table_append_shadow(handle
->table
,
566 shm_fd
, wait_fd
, memory_map_size
);
573 void channel_release(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
576 channel_free(chan
, handle
, shadow
);
580 * channel_destroy - Finalize, wait for q.s. and destroy channel.
581 * @chan: channel to destroy
584 * Call "destroy" callback, finalize channels, decrement the channel
585 * reference count. Note that when readers have completed data
586 * consumption of finalized channels, get_subbuf() will return -ENODATA.
587 * They should release their handle at that point.
589 void channel_destroy(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
593 channel_release(chan
, handle
, shadow
);
597 channel_unregister_notifiers(chan
, handle
);
600 * Note: the consumer takes care of finalizing and switching the
605 * sessiond/consumer are keeping a reference on the shm file
606 * descriptor directly. No need to refcount.
608 channel_release(chan
, handle
, shadow
);
612 struct lttng_ust_lib_ring_buffer
*channel_get_ring_buffer(
613 const struct lttng_ust_lib_ring_buffer_config
*config
,
614 struct channel
*chan
, int cpu
,
615 struct lttng_ust_shm_handle
*handle
,
616 int **shm_fd
, int **wait_fd
,
617 uint64_t **memory_map_size
)
621 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
622 ref
= &chan
->backend
.buf
[0].shmp
._ref
;
623 shm_get_object_data(handle
, ref
, shm_fd
, wait_fd
,
625 return shmp(handle
, chan
->backend
.buf
[0].shmp
);
627 if (cpu
>= num_possible_cpus())
629 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
630 shm_get_object_data(handle
, ref
, shm_fd
, wait_fd
,
632 return shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
636 int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer
*buf
,
637 struct lttng_ust_shm_handle
*handle
,
641 if (uatomic_cmpxchg(&buf
->active_shadow_readers
, 0, 1) != 0)
646 if (uatomic_cmpxchg(&buf
->active_readers
, 0, 1) != 0)
652 void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer
*buf
,
653 struct lttng_ust_shm_handle
*handle
,
656 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
659 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_shadow_readers
) != 1);
661 uatomic_dec(&buf
->active_shadow_readers
);
664 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
666 uatomic_dec(&buf
->active_readers
);
670 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
672 * @consumed: consumed count indicating the position where to read
673 * @produced: produced count, indicates position when to stop reading
675 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
676 * data to read at consumed position, or 0 if the get operation succeeds.
679 int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer
*buf
,
680 unsigned long *consumed
, unsigned long *produced
,
681 struct lttng_ust_shm_handle
*handle
)
683 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
684 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
685 unsigned long consumed_cur
, write_offset
;
688 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
690 * Read finalized before counters.
693 consumed_cur
= uatomic_read(&buf
->consumed
);
695 * No need to issue a memory barrier between consumed count read and
696 * write offset read, because consumed count can only change
697 * concurrently in overwrite mode, and we keep a sequence counter
698 * identifier derived from the write offset to check we are getting
699 * the same sub-buffer we are expecting (the sub-buffers are atomically
700 * "tagged" upon writes, tags are checked upon read).
702 write_offset
= v_read(config
, &buf
->offset
);
705 * Check that we are not about to read the same subbuffer in
706 * which the writer head is.
708 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
712 *consumed
= consumed_cur
;
713 *produced
= subbuf_trunc(write_offset
, chan
);
719 * The memory barriers __wait_event()/wake_up_interruptible() take care
720 * of "raw_spin_is_locked" memory ordering.
729 * lib_ring_buffer_put_snapshot - move consumed counter forward
731 * @consumed_new: new consumed count value
733 void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer
*buf
,
734 unsigned long consumed_new
,
735 struct lttng_ust_shm_handle
*handle
)
737 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
738 struct channel
*chan
= shmp(handle
, bufb
->chan
);
739 unsigned long consumed
;
741 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1
742 && uatomic_read(&buf
->active_shadow_readers
) != 1);
745 * Only push the consumed value forward.
746 * If the consumed cmpxchg fails, this is because we have been pushed by
747 * the writer in flight recorder mode.
749 consumed
= uatomic_read(&buf
->consumed
);
750 while ((long) consumed
- (long) consumed_new
< 0)
751 consumed
= uatomic_cmpxchg(&buf
->consumed
, consumed
,
756 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
758 * @consumed: consumed count indicating the position where to read
760 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
761 * data to read at consumed position, or 0 if the get operation succeeds.
763 int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
764 unsigned long consumed
,
765 struct lttng_ust_shm_handle
*handle
)
767 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
768 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
769 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
774 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
776 * Read finalized before counters.
779 consumed_cur
= uatomic_read(&buf
->consumed
);
780 consumed_idx
= subbuf_index(consumed
, chan
);
781 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, consumed_idx
)->cc_sb
);
783 * Make sure we read the commit count before reading the buffer
784 * data and the write offset. Correct consumed offset ordering
785 * wrt commit count is insured by the use of cmpxchg to update
786 * the consumed offset.
789 * Local rmb to match the remote wmb to read the commit count
790 * before the buffer data and the write offset.
794 write_offset
= v_read(config
, &buf
->offset
);
797 * Check that the buffer we are getting is after or at consumed_cur
800 if ((long) subbuf_trunc(consumed
, chan
)
801 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
805 * Check that the subbuffer we are trying to consume has been
806 * already fully committed.
808 if (((commit_count
- chan
->backend
.subbuf_size
)
809 & chan
->commit_count_mask
)
810 - (buf_trunc(consumed_cur
, chan
)
811 >> chan
->backend
.num_subbuf_order
)
816 * Check that we are not about to read the same subbuffer in
817 * which the writer head is.
819 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
824 * Failure to get the subbuffer causes a busy-loop retry without going
825 * to a wait queue. These are caused by short-lived race windows where
826 * the writer is getting access to a subbuffer we were trying to get
827 * access to. Also checks that the "consumed" buffer count we are
828 * looking for matches the one contained in the subbuffer id.
830 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
831 consumed_idx
, buf_trunc_val(consumed
, chan
),
835 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
837 buf
->get_subbuf_consumed
= consumed
;
844 * The memory barriers __wait_event()/wake_up_interruptible() take care
845 * of "raw_spin_is_locked" memory ordering.
854 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
857 void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
858 struct lttng_ust_shm_handle
*handle
)
860 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
861 struct channel
*chan
= shmp(handle
, bufb
->chan
);
862 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
863 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
865 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1
866 && uatomic_read(&buf
->active_shadow_readers
) != 1);
868 if (!buf
->get_subbuf
) {
870 * Reader puts a subbuffer it did not get.
872 CHAN_WARN_ON(chan
, 1);
875 consumed
= buf
->get_subbuf_consumed
;
879 * Clear the records_unread counter. (overruns counter)
880 * Can still be non-zero if a file reader simply grabbed the data
881 * without using iterators.
882 * Can be below zero if an iterator is used on a snapshot more than
885 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
886 v_add(config
, v_read(config
,
887 &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
),
888 &bufb
->records_read
);
889 v_set(config
, &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
, 0);
890 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
891 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
892 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
895 * Exchange the reader subbuffer with the one we put in its place in the
896 * writer subbuffer table. Expect the original consumed count. If
897 * update_read_sb_index fails, this is because the writer updated the
898 * subbuffer concurrently. We should therefore keep the subbuffer we
899 * currently have: it has become invalid to try reading this sub-buffer
900 * consumed count value anyway.
902 consumed_idx
= subbuf_index(consumed
, chan
);
903 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
904 consumed_idx
, buf_trunc_val(consumed
, chan
),
907 * update_read_sb_index return value ignored. Don't exchange sub-buffer
908 * if the writer concurrently updated it.
913 * cons_offset is an iterator on all subbuffer offsets between the reader
914 * position and the writer position. (inclusive)
917 void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
918 struct channel
*chan
,
919 unsigned long cons_offset
,
921 struct lttng_ust_shm_handle
*handle
)
923 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
924 unsigned long cons_idx
, commit_count
, commit_count_sb
;
926 cons_idx
= subbuf_index(cons_offset
, chan
);
927 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, cons_idx
)->cc
);
928 commit_count_sb
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, cons_idx
)->cc_sb
);
930 if (subbuf_offset(commit_count
, chan
) != 0)
931 DBG("ring buffer %s, cpu %d: "
932 "commit count in subbuffer %lu,\n"
933 "expecting multiples of %lu bytes\n"
934 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
935 chan
->backend
.name
, cpu
, cons_idx
,
936 chan
->backend
.subbuf_size
,
937 commit_count
, commit_count_sb
);
939 DBG("ring buffer: %s, cpu %d: %lu bytes committed\n",
940 chan
->backend
.name
, cpu
, commit_count
);
944 void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
945 struct channel
*chan
,
947 struct lttng_ust_shm_handle
*handle
)
949 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
950 unsigned long write_offset
, cons_offset
;
953 * No need to order commit_count, write_offset and cons_offset reads
954 * because we execute at teardown when no more writer nor reader
955 * references are left.
957 write_offset
= v_read(config
, &buf
->offset
);
958 cons_offset
= uatomic_read(&buf
->consumed
);
959 if (write_offset
!= cons_offset
)
960 DBG("ring buffer %s, cpu %d: "
961 "non-consumed data\n"
962 " [ %lu bytes written, %lu bytes read ]\n",
963 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
965 for (cons_offset
= uatomic_read(&buf
->consumed
);
966 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
969 cons_offset
= subbuf_align(cons_offset
, chan
))
970 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
975 void lib_ring_buffer_print_errors(struct channel
*chan
,
976 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
977 struct lttng_ust_shm_handle
*handle
)
979 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
980 void *priv
= channel_get_private(chan
);
982 DBG("ring buffer %s, cpu %d: %lu records written, "
983 "%lu records overrun\n",
984 chan
->backend
.name
, cpu
,
985 v_read(config
, &buf
->records_count
),
986 v_read(config
, &buf
->records_overrun
));
988 if (v_read(config
, &buf
->records_lost_full
)
989 || v_read(config
, &buf
->records_lost_wrap
)
990 || v_read(config
, &buf
->records_lost_big
))
991 DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n"
992 " [ %lu buffer full, %lu nest buffer wrap-around, "
993 "%lu event too big ]\n",
994 chan
->backend
.name
, cpu
,
995 v_read(config
, &buf
->records_lost_full
),
996 v_read(config
, &buf
->records_lost_wrap
),
997 v_read(config
, &buf
->records_lost_big
));
999 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
, handle
);
1003 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1005 * Only executed when the buffer is finalized, in SWITCH_FLUSH.
1008 void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer
*buf
,
1009 struct channel
*chan
,
1010 struct switch_offsets
*offsets
,
1012 struct lttng_ust_shm_handle
*handle
)
1014 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1015 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1016 unsigned long commit_count
;
1018 config
->cb
.buffer_begin(buf
, tsc
, oldidx
, handle
);
1021 * Order all writes to buffer before the commit count update that will
1022 * determine that the subbuffer is full.
1025 v_add(config
, config
->cb
.subbuffer_header_size(),
1026 &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1027 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1028 /* Check if the written buffer has to be delivered */
1029 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1030 commit_count
, oldidx
, handle
);
1031 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1032 offsets
->old
, commit_count
,
1033 config
->cb
.subbuffer_header_size(),
1038 * lib_ring_buffer_switch_old_end: switch old subbuffer
1040 * Note : offset_old should never be 0 here. It is ok, because we never perform
1041 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1042 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1046 void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer
*buf
,
1047 struct channel
*chan
,
1048 struct switch_offsets
*offsets
,
1050 struct lttng_ust_shm_handle
*handle
)
1052 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1053 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1054 unsigned long commit_count
, padding_size
, data_size
;
1056 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1057 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1058 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
,
1062 * Order all writes to buffer before the commit count update that will
1063 * determine that the subbuffer is full.
1066 v_add(config
, padding_size
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1067 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1068 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1069 commit_count
, oldidx
, handle
);
1070 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1071 offsets
->old
, commit_count
,
1072 padding_size
, handle
);
1076 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1078 * This code can be executed unordered : writers may already have written to the
1079 * sub-buffer before this code gets executed, caution. The commit makes sure
1080 * that this code is executed before the deliver of this sub-buffer.
1083 void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer
*buf
,
1084 struct channel
*chan
,
1085 struct switch_offsets
*offsets
,
1087 struct lttng_ust_shm_handle
*handle
)
1089 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1090 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1091 unsigned long commit_count
;
1093 config
->cb
.buffer_begin(buf
, tsc
, beginidx
, handle
);
1096 * Order all writes to buffer before the commit count update that will
1097 * determine that the subbuffer is full.
1100 v_add(config
, config
->cb
.subbuffer_header_size(),
1101 &shmp_index(handle
, buf
->commit_hot
, beginidx
)->cc
);
1102 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, beginidx
)->cc
);
1103 /* Check if the written buffer has to be delivered */
1104 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1105 commit_count
, beginidx
, handle
);
1106 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, beginidx
,
1107 offsets
->begin
, commit_count
,
1108 config
->cb
.subbuffer_header_size(),
1113 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1115 * The only remaining threads could be the ones with pending commits. They will
1116 * have to do the deliver themselves.
1119 void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer
*buf
,
1120 struct channel
*chan
,
1121 struct switch_offsets
*offsets
,
1123 struct lttng_ust_shm_handle
*handle
)
1125 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1126 unsigned long endidx
= subbuf_index(offsets
->end
- 1, chan
);
1127 unsigned long commit_count
, padding_size
, data_size
;
1129 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1130 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1131 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
,
1135 * Order all writes to buffer before the commit count update that will
1136 * determine that the subbuffer is full.
1139 v_add(config
, padding_size
, &shmp_index(handle
, buf
->commit_hot
, endidx
)->cc
);
1140 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, endidx
)->cc
);
1141 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->end
- 1,
1142 commit_count
, endidx
, handle
);
1143 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, endidx
,
1144 offsets
->end
, commit_count
,
1145 padding_size
, handle
);
1151 * !0 if execution must be aborted.
1154 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1155 struct lttng_ust_lib_ring_buffer
*buf
,
1156 struct channel
*chan
,
1157 struct switch_offsets
*offsets
,
1160 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1163 offsets
->begin
= v_read(config
, &buf
->offset
);
1164 offsets
->old
= offsets
->begin
;
1165 offsets
->switch_old_start
= 0;
1166 off
= subbuf_offset(offsets
->begin
, chan
);
1168 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1171 * Ensure we flush the header of an empty subbuffer when doing the
1172 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1173 * total data gathering duration even if there were no records saved
1174 * after the last buffer switch.
1175 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1176 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1177 * subbuffer header as appropriate.
1178 * The next record that reserves space will be responsible for
1179 * populating the following subbuffer header. We choose not to populate
1180 * the next subbuffer header here because we want to be able to use
1181 * SWITCH_ACTIVE for periodical buffer flush, which must
1182 * guarantee that all the buffer content (records and header
1183 * timestamps) are visible to the reader. This is required for
1184 * quiescence guarantees for the fusion merge.
1186 if (mode
== SWITCH_FLUSH
|| off
> 0) {
1187 if (caa_unlikely(off
== 0)) {
1189 * The client does not save any header information.
1190 * Don't switch empty subbuffer on finalize, because it
1191 * is invalid to deliver a completely empty subbuffer.
1193 if (!config
->cb
.subbuffer_header_size())
1196 * Need to write the subbuffer start header on finalize.
1198 offsets
->switch_old_start
= 1;
1200 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1202 return -1; /* we do not have to switch : buffer is empty */
1203 /* Note: old points to the next subbuf at offset 0 */
1204 offsets
->end
= offsets
->begin
;
1209 * Force a sub-buffer switch. This operation is completely reentrant : can be
1210 * called while tracing is active with absolutely no lock held.
1212 * Note, however, that as a v_cmpxchg is used for some atomic
1213 * operations, this function must be called from the CPU which owns the buffer
1214 * for a ACTIVE flush.
1216 void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer
*buf
, enum switch_mode mode
,
1217 struct lttng_ust_shm_handle
*handle
)
1219 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1220 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1221 struct switch_offsets offsets
;
1222 unsigned long oldidx
;
1228 * Perform retryable operations.
1231 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1233 return; /* Switch not needed */
1234 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1238 * Atomically update last_tsc. This update races against concurrent
1239 * atomic updates, but the race will always cause supplementary full TSC
1240 * records, never the opposite (missing a full TSC record when it would
1243 save_last_tsc(config
, buf
, tsc
);
1246 * Push the reader if necessary
1248 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1250 oldidx
= subbuf_index(offsets
.old
, chan
);
1251 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
, handle
);
1254 * May need to populate header start on SWITCH_FLUSH.
1256 if (offsets
.switch_old_start
) {
1257 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
, handle
);
1258 offsets
.old
+= config
->cb
.subbuffer_header_size();
1262 * Switch old subbuffer.
1264 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
, handle
);
1270 * -ENOSPC if event size is too large for packet.
1271 * -ENOBUFS if there is currently not enough space in buffer for the event.
1272 * -EIO if data cannot be written into the buffer for any other reason.
1275 int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer
*buf
,
1276 struct channel
*chan
,
1277 struct switch_offsets
*offsets
,
1278 struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
1280 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1281 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
1282 unsigned long reserve_commit_diff
;
1284 offsets
->begin
= v_read(config
, &buf
->offset
);
1285 offsets
->old
= offsets
->begin
;
1286 offsets
->switch_new_start
= 0;
1287 offsets
->switch_new_end
= 0;
1288 offsets
->switch_old_end
= 0;
1289 offsets
->pre_header_padding
= 0;
1291 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1292 if ((int64_t) ctx
->tsc
== -EIO
)
1295 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1296 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
1298 if (caa_unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1299 offsets
->switch_new_start
= 1; /* For offsets->begin */
1301 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1303 &offsets
->pre_header_padding
,
1306 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1309 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
) +
1310 offsets
->size
> chan
->backend
.subbuf_size
)) {
1311 offsets
->switch_old_end
= 1; /* For offsets->old */
1312 offsets
->switch_new_start
= 1; /* For offsets->begin */
1315 if (caa_unlikely(offsets
->switch_new_start
)) {
1316 unsigned long sb_index
;
1319 * We are typically not filling the previous buffer completely.
1321 if (caa_likely(offsets
->switch_old_end
))
1322 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1323 offsets
->begin
= offsets
->begin
1324 + config
->cb
.subbuffer_header_size();
1325 /* Test new buffer integrity */
1326 sb_index
= subbuf_index(offsets
->begin
, chan
);
1327 reserve_commit_diff
=
1328 (buf_trunc(offsets
->begin
, chan
)
1329 >> chan
->backend
.num_subbuf_order
)
1330 - ((unsigned long) v_read(config
,
1331 &shmp_index(handle
, buf
->commit_cold
, sb_index
)->cc_sb
)
1332 & chan
->commit_count_mask
);
1333 if (caa_likely(reserve_commit_diff
== 0)) {
1334 /* Next subbuffer not being written to. */
1335 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1336 subbuf_trunc(offsets
->begin
, chan
)
1337 - subbuf_trunc((unsigned long)
1338 uatomic_read(&buf
->consumed
), chan
)
1339 >= chan
->backend
.buf_size
)) {
1341 * We do not overwrite non consumed buffers
1342 * and we are full : record is lost.
1344 v_inc(config
, &buf
->records_lost_full
);
1348 * Next subbuffer not being written to, and we
1349 * are either in overwrite mode or the buffer is
1350 * not full. It's safe to write in this new
1356 * Next subbuffer reserve offset does not match the
1357 * commit offset. Drop record in producer-consumer and
1358 * overwrite mode. Caused by either a writer OOPS or too
1359 * many nested writes over a reserve/commit pair.
1361 v_inc(config
, &buf
->records_lost_wrap
);
1365 config
->cb
.record_header_size(config
, chan
,
1367 &offsets
->pre_header_padding
,
1370 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1373 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
)
1374 + offsets
->size
> chan
->backend
.subbuf_size
)) {
1376 * Record too big for subbuffers, report error, don't
1377 * complete the sub-buffer switch.
1379 v_inc(config
, &buf
->records_lost_big
);
1383 * We just made a successful buffer switch and the
1384 * record fits in the new subbuffer. Let's write.
1389 * Record fits in the current buffer and we are not on a switch
1390 * boundary. It's safe to write.
1393 offsets
->end
= offsets
->begin
+ offsets
->size
;
1395 if (caa_unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
1397 * The offset_end will fall at the very beginning of the next
1400 offsets
->switch_new_end
= 1; /* For offsets->begin */
1406 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1407 * @ctx: ring buffer context.
1409 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
1410 * -EIO for other errors, else returns 0.
1411 * It will take care of sub-buffer switching.
1413 int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
1415 struct channel
*chan
= ctx
->chan
;
1416 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
1417 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1418 struct lttng_ust_lib_ring_buffer
*buf
;
1419 struct switch_offsets offsets
;
1422 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
1423 buf
= shmp(handle
, chan
->backend
.buf
[ctx
->cpu
].shmp
);
1425 buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
1431 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
1433 if (caa_unlikely(ret
))
1435 } while (caa_unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
1440 * Atomically update last_tsc. This update races against concurrent
1441 * atomic updates, but the race will always cause supplementary full TSC
1442 * records, never the opposite (missing a full TSC record when it would
1445 save_last_tsc(config
, buf
, ctx
->tsc
);
1448 * Push the reader if necessary
1450 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
1453 * Clear noref flag for this subbuffer.
1455 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1456 subbuf_index(offsets
.end
- 1, chan
),
1460 * Switch old subbuffer if needed.
1462 if (caa_unlikely(offsets
.switch_old_end
)) {
1463 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1464 subbuf_index(offsets
.old
- 1, chan
),
1466 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1470 * Populate new subbuffer.
1472 if (caa_unlikely(offsets
.switch_new_start
))
1473 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1475 if (caa_unlikely(offsets
.switch_new_end
))
1476 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1478 ctx
->slot_size
= offsets
.size
;
1479 ctx
->pre_offset
= offsets
.begin
;
1480 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;