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 <lttng/ringbuffer-config.h>
56 #define max(a, b) ((a) > (b) ? (a) : (b))
60 * Use POSIX SHM: shm_open(3) and shm_unlink(3).
61 * close(2) to close the fd returned by shm_open.
62 * shm_unlink releases the shared memory object name.
63 * ftruncate(2) sets the size of the memory object.
64 * mmap/munmap maps the shared memory obj to a virtual address in the
65 * calling proceess (should be done both in libust and consumer).
66 * See shm_overview(7) for details.
67 * Pass file descriptor returned by shm_open(3) to ltt-sessiond through
70 * Since we don't need to access the object using its name, we can
71 * immediately shm_unlink(3) it, and only keep the handle with its file
76 * Internal structure representing offsets to use at a sub-buffer switch.
78 struct switch_offsets
{
79 unsigned long begin
, end
, old
;
80 size_t pre_header_padding
, size
;
81 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
85 __thread
unsigned int lib_ring_buffer_nesting
;
88 * TODO: this is unused. Errors are saved within the ring buffer.
89 * Eventually, allow consumerd to print these errors.
92 void lib_ring_buffer_print_errors(struct channel
*chan
,
93 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
94 struct lttng_ust_shm_handle
*handle
);
97 * lib_ring_buffer_reset - Reset ring buffer to initial values.
100 * Effectively empty the ring buffer. Should be called when the buffer is not
101 * used for writing. The ring buffer can be opened for reading, but the reader
102 * should not be using the iterator concurrently with reset. The previous
103 * current iterator record is reset.
105 void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer
*buf
,
106 struct lttng_ust_shm_handle
*handle
)
108 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
109 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
113 * Reset iterator first. It will put the subbuffer if it currently holds
116 v_set(config
, &buf
->offset
, 0);
117 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
118 v_set(config
, &shmp_index(handle
, buf
->commit_hot
, i
)->cc
, 0);
119 v_set(config
, &shmp_index(handle
, buf
->commit_hot
, i
)->seq
, 0);
120 v_set(config
, &shmp_index(handle
, buf
->commit_cold
, i
)->cc_sb
, 0);
122 uatomic_set(&buf
->consumed
, 0);
123 uatomic_set(&buf
->record_disabled
, 0);
124 v_set(config
, &buf
->last_tsc
, 0);
125 lib_ring_buffer_backend_reset(&buf
->backend
, handle
);
126 /* Don't reset number of active readers */
127 v_set(config
, &buf
->records_lost_full
, 0);
128 v_set(config
, &buf
->records_lost_wrap
, 0);
129 v_set(config
, &buf
->records_lost_big
, 0);
130 v_set(config
, &buf
->records_count
, 0);
131 v_set(config
, &buf
->records_overrun
, 0);
136 * channel_reset - Reset channel to initial values.
139 * Effectively empty the channel. Should be called when the channel is not used
140 * for writing. The channel can be opened for reading, but the reader should not
141 * be using the iterator concurrently with reset. The previous current iterator
144 void channel_reset(struct channel
*chan
)
147 * Reset iterators first. Will put the subbuffer if held for reading.
149 uatomic_set(&chan
->record_disabled
, 0);
150 /* Don't reset commit_count_mask, still valid */
151 channel_backend_reset(&chan
->backend
);
152 /* Don't reset switch/read timer interval */
153 /* Don't reset notifiers and notifier enable bits */
154 /* Don't reset reader reference count */
158 * Must be called under cpu hotplug protection.
160 int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer
*buf
,
161 struct channel_backend
*chanb
, int cpu
,
162 struct lttng_ust_shm_handle
*handle
,
163 struct shm_object
*shmobj
)
165 const struct lttng_ust_lib_ring_buffer_config
*config
= &chanb
->config
;
166 struct channel
*chan
= caa_container_of(chanb
, struct channel
, backend
);
167 void *priv
= channel_get_private(chan
);
168 size_t subbuf_header_size
;
172 /* Test for cpu hotplug */
173 if (buf
->backend
.allocated
)
176 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
,
177 cpu
, handle
, shmobj
);
181 align_shm(shmobj
, __alignof__(struct commit_counters_hot
));
182 set_shmp(buf
->commit_hot
,
184 sizeof(struct commit_counters_hot
) * chan
->backend
.num_subbuf
));
185 if (!shmp(handle
, buf
->commit_hot
)) {
190 align_shm(shmobj
, __alignof__(struct commit_counters_cold
));
191 set_shmp(buf
->commit_cold
,
193 sizeof(struct commit_counters_cold
) * chan
->backend
.num_subbuf
));
194 if (!shmp(handle
, buf
->commit_cold
)) {
200 * Write the subbuffer header for first subbuffer so we know the total
201 * duration of data gathering.
203 subbuf_header_size
= config
->cb
.subbuffer_header_size();
204 v_set(config
, &buf
->offset
, subbuf_header_size
);
205 subbuffer_id_clear_noref(config
, &shmp_index(handle
, buf
->backend
.buf_wsb
, 0)->id
);
206 tsc
= config
->cb
.ring_buffer_clock_read(shmp(handle
, buf
->backend
.chan
));
207 config
->cb
.buffer_begin(buf
, tsc
, 0, handle
);
208 v_add(config
, subbuf_header_size
, &shmp_index(handle
, buf
->commit_hot
, 0)->cc
);
210 if (config
->cb
.buffer_create
) {
211 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
, handle
);
215 buf
->backend
.allocated
= 1;
220 /* commit_cold will be freed by shm teardown */
222 /* commit_hot will be freed by shm teardown */
228 static void switch_buffer_timer(unsigned long data
)
230 struct lttng_ust_lib_ring_buffer
*buf
= (struct lttng_ust_lib_ring_buffer
*)data
;
231 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
232 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
235 * Only flush buffers periodically if readers are active.
237 if (uatomic_read(&buf
->active_readers
) || uatomic_read(&buf
->active_shadow_readers
))
238 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
, handle
);
241 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
242 // mod_timer_pinned(&buf->switch_timer,
243 // jiffies + chan->switch_timer_interval);
245 // mod_timer(&buf->switch_timer,
246 // jiffies + chan->switch_timer_interval);
250 static void lib_ring_buffer_start_switch_timer(struct lttng_ust_lib_ring_buffer
*buf
,
251 struct lttng_ust_shm_handle
*handle
)
253 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
254 //const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
256 if (!chan
->switch_timer_interval
|| buf
->switch_timer_enabled
)
259 //init_timer(&buf->switch_timer);
260 //buf->switch_timer.function = switch_buffer_timer;
261 //buf->switch_timer.expires = jiffies + chan->switch_timer_interval;
262 //buf->switch_timer.data = (unsigned long)buf;
263 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
264 // add_timer_on(&buf->switch_timer, buf->backend.cpu);
266 // add_timer(&buf->switch_timer);
267 buf
->switch_timer_enabled
= 1;
270 static void lib_ring_buffer_stop_switch_timer(struct lttng_ust_lib_ring_buffer
*buf
,
271 struct lttng_ust_shm_handle
*handle
)
273 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
275 if (!chan
->switch_timer_interval
|| !buf
->switch_timer_enabled
)
279 //del_timer_sync(&buf->switch_timer);
280 buf
->switch_timer_enabled
= 0;
285 * Polling timer to check the channels for data.
287 static void read_buffer_timer(unsigned long data
)
289 struct lttng_ust_lib_ring_buffer
*buf
= (struct lttng_ust_lib_ring_buffer
*)data
;
290 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
291 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
293 CHAN_WARN_ON(chan
, !buf
->backend
.allocated
);
295 if (uatomic_read(&buf
->active_readers
) || uatomic_read(&buf
->active_shadow_readers
))
296 && lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
298 //wake_up_interruptible(&buf->read_wait);
299 //wake_up_interruptible(&chan->read_wait);
303 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
304 // mod_timer_pinned(&buf->read_timer,
305 // jiffies + chan->read_timer_interval);
307 // mod_timer(&buf->read_timer,
308 // jiffies + chan->read_timer_interval);
312 static void lib_ring_buffer_start_read_timer(struct lttng_ust_lib_ring_buffer
*buf
,
313 struct lttng_ust_shm_handle
*handle
)
315 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
316 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
318 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
319 || !chan
->read_timer_interval
320 || buf
->read_timer_enabled
)
324 //init_timer(&buf->read_timer);
325 //buf->read_timer.function = read_buffer_timer;
326 //buf->read_timer.expires = jiffies + chan->read_timer_interval;
327 //buf->read_timer.data = (unsigned long)buf;
329 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
330 // add_timer_on(&buf->read_timer, buf->backend.cpu);
332 // add_timer(&buf->read_timer);
333 buf
->read_timer_enabled
= 1;
336 static void lib_ring_buffer_stop_read_timer(struct lttng_ust_lib_ring_buffer
*buf
,
337 struct lttng_ust_shm_handle
*handle
)
339 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
340 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
342 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
343 || !chan
->read_timer_interval
344 || !buf
->read_timer_enabled
)
348 //del_timer_sync(&buf->read_timer);
350 * do one more check to catch data that has been written in the last
353 if (lib_ring_buffer_poll_deliver(config
, buf
, chan
, handle
)) {
355 //wake_up_interruptible(&buf->read_wait);
356 //wake_up_interruptible(&chan->read_wait);
358 buf
->read_timer_enabled
= 0;
361 static void channel_unregister_notifiers(struct channel
*chan
,
362 struct lttng_ust_shm_handle
*handle
)
364 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
367 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
368 for_each_possible_cpu(cpu
) {
369 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
371 lib_ring_buffer_stop_switch_timer(buf
, handle
);
372 lib_ring_buffer_stop_read_timer(buf
, handle
);
375 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
377 lib_ring_buffer_stop_switch_timer(buf
, handle
);
378 lib_ring_buffer_stop_read_timer(buf
, handle
);
380 //channel_backend_unregister_notifiers(&chan->backend);
383 static void channel_free(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
387 channel_backend_free(&chan
->backend
, handle
);
388 /* chan is freed by shm teardown */
389 shm_object_table_destroy(handle
->table
);
394 * channel_create - Create channel.
395 * @config: ring buffer instance configuration
396 * @name: name of the channel
397 * @priv_data: ring buffer client private data area pointer (output)
398 * @priv_data_size: length, in bytes, of the private data area.
399 * @priv_data_init: initialization data for private data.
400 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
401 * address mapping. It is used only by RING_BUFFER_STATIC
402 * configuration. It can be set to NULL for other backends.
403 * @subbuf_size: subbuffer size
404 * @num_subbuf: number of subbuffers
405 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
406 * padding to let readers get those sub-buffers.
407 * Used for live streaming.
408 * @read_timer_interval: Time interval (in us) to wake up pending readers.
411 * Returns NULL on failure.
413 struct lttng_ust_shm_handle
*channel_create(const struct lttng_ust_lib_ring_buffer_config
*config
,
416 size_t priv_data_align
,
417 size_t priv_data_size
,
418 void *priv_data_init
,
419 void *buf_addr
, size_t subbuf_size
,
420 size_t num_subbuf
, unsigned int switch_timer_interval
,
421 unsigned int read_timer_interval
,
422 int *shm_fd
, int *wait_fd
, uint64_t *memory_map_size
)
425 size_t shmsize
, chansize
;
426 struct channel
*chan
;
427 struct lttng_ust_shm_handle
*handle
;
428 struct shm_object
*shmobj
;
431 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
432 read_timer_interval
))
435 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
439 /* Allocate table for channel + per-cpu buffers */
440 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
442 goto error_table_alloc
;
444 /* Calculate the shm allocation layout */
445 shmsize
= sizeof(struct channel
);
446 shmsize
+= offset_align(shmsize
, __alignof__(struct lttng_ust_lib_ring_buffer_shmp
));
447 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
448 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
) * num_possible_cpus();
450 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
);
452 shmsize
+= offset_align(shmsize
, priv_data_align
);
453 shmsize
+= priv_data_size
;
455 shmobj
= shm_object_table_append(handle
->table
, shmsize
);
458 /* struct channel is at object 0, offset 0 (hardcoded) */
459 set_shmp(handle
->chan
, zalloc_shm(shmobj
, chansize
));
460 assert(handle
->chan
._ref
.index
== 0);
461 assert(handle
->chan
._ref
.offset
== 0);
462 chan
= shmp(handle
, handle
->chan
);
466 /* space for private data */
467 if (priv_data_size
) {
468 DECLARE_SHMP(void, priv_data_alloc
);
470 align_shm(shmobj
, priv_data_align
);
471 chan
->priv_data_offset
= shmobj
->allocated_len
;
472 set_shmp(priv_data_alloc
, zalloc_shm(shmobj
, priv_data_size
));
473 if (!shmp(handle
, priv_data_alloc
))
475 *priv_data
= channel_get_private(chan
);
476 memcpy(*priv_data
, priv_data_init
, priv_data_size
);
478 chan
->priv_data_offset
= -1;
482 ret
= channel_backend_init(&chan
->backend
, name
, config
,
483 subbuf_size
, num_subbuf
, handle
);
485 goto error_backend_init
;
487 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
489 //chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval);
490 //chan->read_timer_interval = usecs_to_jiffies(read_timer_interval);
492 //init_waitqueue_head(&chan->read_wait);
493 //init_waitqueue_head(&chan->hp_wait);
495 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
497 * In case of non-hotplug cpu, if the ring-buffer is allocated
498 * in early initcall, it will not be notified of secondary cpus.
499 * In that off case, we need to allocate for all possible cpus.
501 for_each_possible_cpu(cpu
) {
502 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
503 lib_ring_buffer_start_switch_timer(buf
, handle
);
504 lib_ring_buffer_start_read_timer(buf
, handle
);
507 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
509 lib_ring_buffer_start_switch_timer(buf
, handle
);
510 lib_ring_buffer_start_read_timer(buf
, handle
);
512 ref
= &handle
->chan
._ref
;
513 shm_get_object_data(handle
, ref
, shm_fd
, wait_fd
, memory_map_size
);
518 shm_object_table_destroy(handle
->table
);
524 struct lttng_ust_shm_handle
*channel_handle_create(int shm_fd
, int wait_fd
,
525 uint64_t memory_map_size
)
527 struct lttng_ust_shm_handle
*handle
;
528 struct shm_object
*object
;
530 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
534 /* Allocate table for channel + per-cpu buffers */
535 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
537 goto error_table_alloc
;
538 /* Add channel object */
539 object
= shm_object_table_append_shadow(handle
->table
,
540 shm_fd
, wait_fd
, memory_map_size
);
542 goto error_table_object
;
543 /* struct channel is at object 0, offset 0 (hardcoded) */
544 handle
->chan
._ref
.index
= 0;
545 handle
->chan
._ref
.offset
= 0;
549 shm_object_table_destroy(handle
->table
);
555 int channel_handle_add_stream(struct lttng_ust_shm_handle
*handle
,
556 int shm_fd
, int wait_fd
, uint64_t memory_map_size
)
558 struct shm_object
*object
;
560 /* Add stream object */
561 object
= shm_object_table_append_shadow(handle
->table
,
562 shm_fd
, wait_fd
, memory_map_size
);
569 void channel_release(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
572 channel_free(chan
, handle
, shadow
);
576 * channel_destroy - Finalize, wait for q.s. and destroy channel.
577 * @chan: channel to destroy
580 * Call "destroy" callback, finalize channels, decrement the channel
581 * reference count. Note that when readers have completed data
582 * consumption of finalized channels, get_subbuf() will return -ENODATA.
583 * They should release their handle at that point.
585 void channel_destroy(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
589 channel_release(chan
, handle
, shadow
);
593 channel_unregister_notifiers(chan
, handle
);
596 * Note: the consumer takes care of finalizing and switching the
601 * sessiond/consumer are keeping a reference on the shm file
602 * descriptor directly. No need to refcount.
604 channel_release(chan
, handle
, shadow
);
608 struct lttng_ust_lib_ring_buffer
*channel_get_ring_buffer(
609 const struct lttng_ust_lib_ring_buffer_config
*config
,
610 struct channel
*chan
, int cpu
,
611 struct lttng_ust_shm_handle
*handle
,
612 int *shm_fd
, int *wait_fd
,
613 uint64_t *memory_map_size
)
617 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
618 ref
= &chan
->backend
.buf
[0].shmp
._ref
;
619 shm_get_object_data(handle
, ref
, shm_fd
, wait_fd
,
621 return shmp(handle
, chan
->backend
.buf
[0].shmp
);
623 if (cpu
>= num_possible_cpus())
625 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
626 shm_get_object_data(handle
, ref
, shm_fd
, wait_fd
,
628 return shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
632 int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer
*buf
,
633 struct lttng_ust_shm_handle
*handle
,
637 if (uatomic_cmpxchg(&buf
->active_shadow_readers
, 0, 1) != 0)
642 if (uatomic_cmpxchg(&buf
->active_readers
, 0, 1) != 0)
648 void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer
*buf
,
649 struct lttng_ust_shm_handle
*handle
,
652 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
655 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_shadow_readers
) != 1);
657 uatomic_dec(&buf
->active_shadow_readers
);
660 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
662 uatomic_dec(&buf
->active_readers
);
666 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
668 * @consumed: consumed count indicating the position where to read
669 * @produced: produced count, indicates position when to stop reading
671 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
672 * data to read at consumed position, or 0 if the get operation succeeds.
675 int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer
*buf
,
676 unsigned long *consumed
, unsigned long *produced
,
677 struct lttng_ust_shm_handle
*handle
)
679 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
680 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
681 unsigned long consumed_cur
, write_offset
;
684 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
686 * Read finalized before counters.
689 consumed_cur
= uatomic_read(&buf
->consumed
);
691 * No need to issue a memory barrier between consumed count read and
692 * write offset read, because consumed count can only change
693 * concurrently in overwrite mode, and we keep a sequence counter
694 * identifier derived from the write offset to check we are getting
695 * the same sub-buffer we are expecting (the sub-buffers are atomically
696 * "tagged" upon writes, tags are checked upon read).
698 write_offset
= v_read(config
, &buf
->offset
);
701 * Check that we are not about to read the same subbuffer in
702 * which the writer head is.
704 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
708 *consumed
= consumed_cur
;
709 *produced
= subbuf_trunc(write_offset
, chan
);
715 * The memory barriers __wait_event()/wake_up_interruptible() take care
716 * of "raw_spin_is_locked" memory ordering.
725 * lib_ring_buffer_put_snapshot - move consumed counter forward
727 * @consumed_new: new consumed count value
729 void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer
*buf
,
730 unsigned long consumed_new
,
731 struct lttng_ust_shm_handle
*handle
)
733 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
734 struct channel
*chan
= shmp(handle
, bufb
->chan
);
735 unsigned long consumed
;
737 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1
738 && uatomic_read(&buf
->active_shadow_readers
) != 1);
741 * Only push the consumed value forward.
742 * If the consumed cmpxchg fails, this is because we have been pushed by
743 * the writer in flight recorder mode.
745 consumed
= uatomic_read(&buf
->consumed
);
746 while ((long) consumed
- (long) consumed_new
< 0)
747 consumed
= uatomic_cmpxchg(&buf
->consumed
, consumed
,
752 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
754 * @consumed: consumed count indicating the position where to read
756 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
757 * data to read at consumed position, or 0 if the get operation succeeds.
759 int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
760 unsigned long consumed
,
761 struct lttng_ust_shm_handle
*handle
)
763 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
764 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
765 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
770 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
772 * Read finalized before counters.
775 consumed_cur
= uatomic_read(&buf
->consumed
);
776 consumed_idx
= subbuf_index(consumed
, chan
);
777 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, consumed_idx
)->cc_sb
);
779 * Make sure we read the commit count before reading the buffer
780 * data and the write offset. Correct consumed offset ordering
781 * wrt commit count is insured by the use of cmpxchg to update
782 * the consumed offset.
785 * Local rmb to match the remote wmb to read the commit count
786 * before the buffer data and the write offset.
790 write_offset
= v_read(config
, &buf
->offset
);
793 * Check that the buffer we are getting is after or at consumed_cur
796 if ((long) subbuf_trunc(consumed
, chan
)
797 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
801 * Check that the subbuffer we are trying to consume has been
802 * already fully committed.
804 if (((commit_count
- chan
->backend
.subbuf_size
)
805 & chan
->commit_count_mask
)
806 - (buf_trunc(consumed_cur
, chan
)
807 >> chan
->backend
.num_subbuf_order
)
812 * Check that we are not about to read the same subbuffer in
813 * which the writer head is.
815 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
820 * Failure to get the subbuffer causes a busy-loop retry without going
821 * to a wait queue. These are caused by short-lived race windows where
822 * the writer is getting access to a subbuffer we were trying to get
823 * access to. Also checks that the "consumed" buffer count we are
824 * looking for matches the one contained in the subbuffer id.
826 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
827 consumed_idx
, buf_trunc_val(consumed
, chan
),
831 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
833 buf
->get_subbuf_consumed
= consumed
;
840 * The memory barriers __wait_event()/wake_up_interruptible() take care
841 * of "raw_spin_is_locked" memory ordering.
850 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
853 void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
854 struct lttng_ust_shm_handle
*handle
)
856 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
857 struct channel
*chan
= shmp(handle
, bufb
->chan
);
858 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
859 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
861 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1
862 && uatomic_read(&buf
->active_shadow_readers
) != 1);
864 if (!buf
->get_subbuf
) {
866 * Reader puts a subbuffer it did not get.
868 CHAN_WARN_ON(chan
, 1);
871 consumed
= buf
->get_subbuf_consumed
;
875 * Clear the records_unread counter. (overruns counter)
876 * Can still be non-zero if a file reader simply grabbed the data
877 * without using iterators.
878 * Can be below zero if an iterator is used on a snapshot more than
881 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
882 v_add(config
, v_read(config
,
883 &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
),
884 &bufb
->records_read
);
885 v_set(config
, &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
, 0);
886 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
887 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
888 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
891 * Exchange the reader subbuffer with the one we put in its place in the
892 * writer subbuffer table. Expect the original consumed count. If
893 * update_read_sb_index fails, this is because the writer updated the
894 * subbuffer concurrently. We should therefore keep the subbuffer we
895 * currently have: it has become invalid to try reading this sub-buffer
896 * consumed count value anyway.
898 consumed_idx
= subbuf_index(consumed
, chan
);
899 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
900 consumed_idx
, buf_trunc_val(consumed
, chan
),
903 * update_read_sb_index return value ignored. Don't exchange sub-buffer
904 * if the writer concurrently updated it.
909 * cons_offset is an iterator on all subbuffer offsets between the reader
910 * position and the writer position. (inclusive)
913 void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
914 struct channel
*chan
,
915 unsigned long cons_offset
,
917 struct lttng_ust_shm_handle
*handle
)
919 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
920 unsigned long cons_idx
, commit_count
, commit_count_sb
;
922 cons_idx
= subbuf_index(cons_offset
, chan
);
923 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, cons_idx
)->cc
);
924 commit_count_sb
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, cons_idx
)->cc_sb
);
926 if (subbuf_offset(commit_count
, chan
) != 0)
927 DBG("ring buffer %s, cpu %d: "
928 "commit count in subbuffer %lu,\n"
929 "expecting multiples of %lu bytes\n"
930 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
931 chan
->backend
.name
, cpu
, cons_idx
,
932 chan
->backend
.subbuf_size
,
933 commit_count
, commit_count_sb
);
935 DBG("ring buffer: %s, cpu %d: %lu bytes committed\n",
936 chan
->backend
.name
, cpu
, commit_count
);
940 void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
941 struct channel
*chan
,
943 struct lttng_ust_shm_handle
*handle
)
945 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
946 unsigned long write_offset
, cons_offset
;
949 * Can be called in the error path of allocation when
950 * trans_channel_data is not yet set.
955 * No need to order commit_count, write_offset and cons_offset reads
956 * because we execute at teardown when no more writer nor reader
957 * references are left.
959 write_offset
= v_read(config
, &buf
->offset
);
960 cons_offset
= uatomic_read(&buf
->consumed
);
961 if (write_offset
!= cons_offset
)
962 DBG("ring buffer %s, cpu %d: "
963 "non-consumed data\n"
964 " [ %lu bytes written, %lu bytes read ]\n",
965 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
967 for (cons_offset
= uatomic_read(&buf
->consumed
);
968 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
971 cons_offset
= subbuf_align(cons_offset
, chan
))
972 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
977 void lib_ring_buffer_print_errors(struct channel
*chan
,
978 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
979 struct lttng_ust_shm_handle
*handle
)
981 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
982 void *priv
= channel_get_private(chan
);
984 DBG("ring buffer %s, cpu %d: %lu records written, "
985 "%lu records overrun\n",
986 chan
->backend
.name
, cpu
,
987 v_read(config
, &buf
->records_count
),
988 v_read(config
, &buf
->records_overrun
));
990 if (v_read(config
, &buf
->records_lost_full
)
991 || v_read(config
, &buf
->records_lost_wrap
)
992 || v_read(config
, &buf
->records_lost_big
))
993 DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n"
994 " [ %lu buffer full, %lu nest buffer wrap-around, "
995 "%lu event too big ]\n",
996 chan
->backend
.name
, cpu
,
997 v_read(config
, &buf
->records_lost_full
),
998 v_read(config
, &buf
->records_lost_wrap
),
999 v_read(config
, &buf
->records_lost_big
));
1001 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
, handle
);
1005 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1007 * Only executed when the buffer is finalized, in SWITCH_FLUSH.
1010 void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer
*buf
,
1011 struct channel
*chan
,
1012 struct switch_offsets
*offsets
,
1014 struct lttng_ust_shm_handle
*handle
)
1016 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1017 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1018 unsigned long commit_count
;
1020 config
->cb
.buffer_begin(buf
, tsc
, oldidx
, handle
);
1023 * Order all writes to buffer before the commit count update that will
1024 * determine that the subbuffer is full.
1027 v_add(config
, config
->cb
.subbuffer_header_size(),
1028 &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1029 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1030 /* Check if the written buffer has to be delivered */
1031 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1032 commit_count
, oldidx
, handle
);
1033 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1034 offsets
->old
, commit_count
,
1035 config
->cb
.subbuffer_header_size(),
1040 * lib_ring_buffer_switch_old_end: switch old subbuffer
1042 * Note : offset_old should never be 0 here. It is ok, because we never perform
1043 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1044 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1048 void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer
*buf
,
1049 struct channel
*chan
,
1050 struct switch_offsets
*offsets
,
1052 struct lttng_ust_shm_handle
*handle
)
1054 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1055 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1056 unsigned long commit_count
, padding_size
, data_size
;
1058 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1059 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1060 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
,
1064 * Order all writes to buffer before the commit count update that will
1065 * determine that the subbuffer is full.
1068 v_add(config
, padding_size
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1069 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1070 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1071 commit_count
, oldidx
, handle
);
1072 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1073 offsets
->old
, commit_count
,
1074 padding_size
, handle
);
1078 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1080 * This code can be executed unordered : writers may already have written to the
1081 * sub-buffer before this code gets executed, caution. The commit makes sure
1082 * that this code is executed before the deliver of this sub-buffer.
1085 void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer
*buf
,
1086 struct channel
*chan
,
1087 struct switch_offsets
*offsets
,
1089 struct lttng_ust_shm_handle
*handle
)
1091 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1092 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1093 unsigned long commit_count
;
1095 config
->cb
.buffer_begin(buf
, tsc
, beginidx
, handle
);
1098 * Order all writes to buffer before the commit count update that will
1099 * determine that the subbuffer is full.
1102 v_add(config
, config
->cb
.subbuffer_header_size(),
1103 &shmp_index(handle
, buf
->commit_hot
, beginidx
)->cc
);
1104 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, beginidx
)->cc
);
1105 /* Check if the written buffer has to be delivered */
1106 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1107 commit_count
, beginidx
, handle
);
1108 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, beginidx
,
1109 offsets
->begin
, commit_count
,
1110 config
->cb
.subbuffer_header_size(),
1115 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1117 * The only remaining threads could be the ones with pending commits. They will
1118 * have to do the deliver themselves.
1121 void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer
*buf
,
1122 struct channel
*chan
,
1123 struct switch_offsets
*offsets
,
1125 struct lttng_ust_shm_handle
*handle
)
1127 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1128 unsigned long endidx
= subbuf_index(offsets
->end
- 1, chan
);
1129 unsigned long commit_count
, padding_size
, data_size
;
1131 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1132 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1133 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
,
1137 * Order all writes to buffer before the commit count update that will
1138 * determine that the subbuffer is full.
1141 v_add(config
, padding_size
, &shmp_index(handle
, buf
->commit_hot
, endidx
)->cc
);
1142 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, endidx
)->cc
);
1143 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->end
- 1,
1144 commit_count
, endidx
, handle
);
1145 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, endidx
,
1146 offsets
->end
, commit_count
,
1147 padding_size
, handle
);
1153 * !0 if execution must be aborted.
1156 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1157 struct lttng_ust_lib_ring_buffer
*buf
,
1158 struct channel
*chan
,
1159 struct switch_offsets
*offsets
,
1162 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1165 offsets
->begin
= v_read(config
, &buf
->offset
);
1166 offsets
->old
= offsets
->begin
;
1167 offsets
->switch_old_start
= 0;
1168 off
= subbuf_offset(offsets
->begin
, chan
);
1170 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1173 * Ensure we flush the header of an empty subbuffer when doing the
1174 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1175 * total data gathering duration even if there were no records saved
1176 * after the last buffer switch.
1177 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1178 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1179 * subbuffer header as appropriate.
1180 * The next record that reserves space will be responsible for
1181 * populating the following subbuffer header. We choose not to populate
1182 * the next subbuffer header here because we want to be able to use
1183 * SWITCH_ACTIVE for periodical buffer flush, which must
1184 * guarantee that all the buffer content (records and header
1185 * timestamps) are visible to the reader. This is required for
1186 * quiescence guarantees for the fusion merge.
1188 if (mode
== SWITCH_FLUSH
|| off
> 0) {
1189 if (caa_unlikely(off
== 0)) {
1191 * The client does not save any header information.
1192 * Don't switch empty subbuffer on finalize, because it
1193 * is invalid to deliver a completely empty subbuffer.
1195 if (!config
->cb
.subbuffer_header_size())
1198 * Need to write the subbuffer start header on finalize.
1200 offsets
->switch_old_start
= 1;
1202 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1204 return -1; /* we do not have to switch : buffer is empty */
1205 /* Note: old points to the next subbuf at offset 0 */
1206 offsets
->end
= offsets
->begin
;
1211 * Force a sub-buffer switch. This operation is completely reentrant : can be
1212 * called while tracing is active with absolutely no lock held.
1214 * Note, however, that as a v_cmpxchg is used for some atomic
1215 * operations, this function must be called from the CPU which owns the buffer
1216 * for a ACTIVE flush.
1218 void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer
*buf
, enum switch_mode mode
,
1219 struct lttng_ust_shm_handle
*handle
)
1221 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1222 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1223 struct switch_offsets offsets
;
1224 unsigned long oldidx
;
1230 * Perform retryable operations.
1233 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1235 return; /* Switch not needed */
1236 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1240 * Atomically update last_tsc. This update races against concurrent
1241 * atomic updates, but the race will always cause supplementary full TSC
1242 * records, never the opposite (missing a full TSC record when it would
1245 save_last_tsc(config
, buf
, tsc
);
1248 * Push the reader if necessary
1250 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1252 oldidx
= subbuf_index(offsets
.old
, chan
);
1253 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
, handle
);
1256 * May need to populate header start on SWITCH_FLUSH.
1258 if (offsets
.switch_old_start
) {
1259 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
, handle
);
1260 offsets
.old
+= config
->cb
.subbuffer_header_size();
1264 * Switch old subbuffer.
1266 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
, handle
);
1272 * -ENOSPC if event size is too large for packet.
1273 * -ENOBUFS if there is currently not enough space in buffer for the event.
1274 * -EIO if data cannot be written into the buffer for any other reason.
1277 int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer
*buf
,
1278 struct channel
*chan
,
1279 struct switch_offsets
*offsets
,
1280 struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
1282 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1283 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
1284 unsigned long reserve_commit_diff
;
1286 offsets
->begin
= v_read(config
, &buf
->offset
);
1287 offsets
->old
= offsets
->begin
;
1288 offsets
->switch_new_start
= 0;
1289 offsets
->switch_new_end
= 0;
1290 offsets
->switch_old_end
= 0;
1291 offsets
->pre_header_padding
= 0;
1293 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1294 if ((int64_t) ctx
->tsc
== -EIO
)
1297 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1298 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
1300 if (caa_unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1301 offsets
->switch_new_start
= 1; /* For offsets->begin */
1303 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1305 &offsets
->pre_header_padding
,
1308 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1311 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
) +
1312 offsets
->size
> chan
->backend
.subbuf_size
)) {
1313 offsets
->switch_old_end
= 1; /* For offsets->old */
1314 offsets
->switch_new_start
= 1; /* For offsets->begin */
1317 if (caa_unlikely(offsets
->switch_new_start
)) {
1318 unsigned long sb_index
;
1321 * We are typically not filling the previous buffer completely.
1323 if (caa_likely(offsets
->switch_old_end
))
1324 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1325 offsets
->begin
= offsets
->begin
1326 + config
->cb
.subbuffer_header_size();
1327 /* Test new buffer integrity */
1328 sb_index
= subbuf_index(offsets
->begin
, chan
);
1329 reserve_commit_diff
=
1330 (buf_trunc(offsets
->begin
, chan
)
1331 >> chan
->backend
.num_subbuf_order
)
1332 - ((unsigned long) v_read(config
,
1333 &shmp_index(handle
, buf
->commit_cold
, sb_index
)->cc_sb
)
1334 & chan
->commit_count_mask
);
1335 if (caa_likely(reserve_commit_diff
== 0)) {
1336 /* Next subbuffer not being written to. */
1337 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1338 subbuf_trunc(offsets
->begin
, chan
)
1339 - subbuf_trunc((unsigned long)
1340 uatomic_read(&buf
->consumed
), chan
)
1341 >= chan
->backend
.buf_size
)) {
1343 * We do not overwrite non consumed buffers
1344 * and we are full : record is lost.
1346 v_inc(config
, &buf
->records_lost_full
);
1350 * Next subbuffer not being written to, and we
1351 * are either in overwrite mode or the buffer is
1352 * not full. It's safe to write in this new
1358 * Next subbuffer reserve offset does not match the
1359 * commit offset. Drop record in producer-consumer and
1360 * overwrite mode. Caused by either a writer OOPS or too
1361 * many nested writes over a reserve/commit pair.
1363 v_inc(config
, &buf
->records_lost_wrap
);
1367 config
->cb
.record_header_size(config
, chan
,
1369 &offsets
->pre_header_padding
,
1372 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1375 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
)
1376 + offsets
->size
> chan
->backend
.subbuf_size
)) {
1378 * Record too big for subbuffers, report error, don't
1379 * complete the sub-buffer switch.
1381 v_inc(config
, &buf
->records_lost_big
);
1385 * We just made a successful buffer switch and the
1386 * record fits in the new subbuffer. Let's write.
1391 * Record fits in the current buffer and we are not on a switch
1392 * boundary. It's safe to write.
1395 offsets
->end
= offsets
->begin
+ offsets
->size
;
1397 if (caa_unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
1399 * The offset_end will fall at the very beginning of the next
1402 offsets
->switch_new_end
= 1; /* For offsets->begin */
1408 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1409 * @ctx: ring buffer context.
1411 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
1412 * -EIO for other errors, else returns 0.
1413 * It will take care of sub-buffer switching.
1415 int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
1417 struct channel
*chan
= ctx
->chan
;
1418 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
1419 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1420 struct lttng_ust_lib_ring_buffer
*buf
;
1421 struct switch_offsets offsets
;
1424 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
1425 buf
= shmp(handle
, chan
->backend
.buf
[ctx
->cpu
].shmp
);
1427 buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
1433 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
1435 if (caa_unlikely(ret
))
1437 } while (caa_unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
1442 * Atomically update last_tsc. This update races against concurrent
1443 * atomic updates, but the race will always cause supplementary full TSC
1444 * records, never the opposite (missing a full TSC record when it would
1447 save_last_tsc(config
, buf
, ctx
->tsc
);
1450 * Push the reader if necessary
1452 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
1455 * Clear noref flag for this subbuffer.
1457 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1458 subbuf_index(offsets
.end
- 1, chan
),
1462 * Switch old subbuffer if needed.
1464 if (caa_unlikely(offsets
.switch_old_end
)) {
1465 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1466 subbuf_index(offsets
.old
- 1, chan
),
1468 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1472 * Populate new subbuffer.
1474 if (caa_unlikely(offsets
.switch_new_start
))
1475 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1477 if (caa_unlikely(offsets
.switch_new_end
))
1478 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1480 ctx
->slot_size
= offsets
.size
;
1481 ctx
->pre_offset
= offsets
.begin
;
1482 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;