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>
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 lttng_ust_lib_ring_buffer
*buf
, int cpu
,
89 struct lttng_ust_shm_handle
*handle
);
92 * Must be called under cpu hotplug protection.
94 void lib_ring_buffer_free(struct lttng_ust_lib_ring_buffer
*buf
,
95 struct lttng_ust_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 lttng_ust_lib_ring_buffer
*buf
,
116 struct lttng_ust_shm_handle
*handle
)
118 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
119 const struct lttng_ust_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 lttng_ust_lib_ring_buffer
*buf
,
171 struct channel_backend
*chanb
, int cpu
,
172 struct lttng_ust_shm_handle
*handle
,
173 struct shm_object
*shmobj
)
175 const struct lttng_ust_lib_ring_buffer_config
*config
= &chanb
->config
;
176 struct channel
*chan
= caa_container_of(chanb
, struct channel
, backend
);
177 void *priv
= channel_get_private(chan
);
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 lttng_ust_lib_ring_buffer
*buf
= (struct lttng_ust_lib_ring_buffer
*)data
;
246 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
247 const struct lttng_ust_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 lttng_ust_lib_ring_buffer
*buf
,
266 struct lttng_ust_shm_handle
*handle
)
268 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
269 //const struct lttng_ust_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 lttng_ust_lib_ring_buffer
*buf
,
286 struct lttng_ust_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 lttng_ust_lib_ring_buffer
*buf
= (struct lttng_ust_lib_ring_buffer
*)data
;
305 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
306 const struct lttng_ust_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 lttng_ust_lib_ring_buffer
*buf
,
328 struct lttng_ust_shm_handle
*handle
)
330 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
331 const struct lttng_ust_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 lttng_ust_lib_ring_buffer
*buf
,
352 struct lttng_ust_shm_handle
*handle
)
354 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
355 const struct lttng_ust_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 lttng_ust_shm_handle
*handle
)
379 const struct lttng_ust_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 lttng_ust_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 lttng_ust_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 lttng_ust_shm_handle
*handle
,
402 channel_backend_free(&chan
->backend
, handle
);
403 /* chan is freed by shm teardown */
404 shm_object_table_destroy(handle
->table
);
409 * channel_create - Create channel.
410 * @config: ring buffer instance configuration
411 * @name: name of the channel
412 * @priv_data: ring buffer client private data area pointer (output)
413 * @priv_data_size: length, in bytes, of the private data area.
414 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
415 * address mapping. It is used only by RING_BUFFER_STATIC
416 * configuration. It can be set to NULL for other backends.
417 * @subbuf_size: subbuffer size
418 * @num_subbuf: number of subbuffers
419 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
420 * padding to let readers get those sub-buffers.
421 * Used for live streaming.
422 * @read_timer_interval: Time interval (in us) to wake up pending readers.
425 * Returns NULL on failure.
427 struct lttng_ust_shm_handle
*channel_create(const struct lttng_ust_lib_ring_buffer_config
*config
,
430 size_t priv_data_align
,
431 size_t priv_data_size
,
432 void *buf_addr
, size_t subbuf_size
,
433 size_t num_subbuf
, unsigned int switch_timer_interval
,
434 unsigned int read_timer_interval
,
435 int *shm_fd
, int *wait_fd
, uint64_t *memory_map_size
)
438 size_t shmsize
, chansize
;
439 struct channel
*chan
;
440 struct lttng_ust_shm_handle
*handle
;
441 struct shm_object
*shmobj
;
444 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
445 read_timer_interval
))
448 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
452 /* Allocate table for channel + per-cpu buffers */
453 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
455 goto error_table_alloc
;
457 /* Calculate the shm allocation layout */
458 shmsize
= sizeof(struct channel
);
459 shmsize
+= offset_align(shmsize
, __alignof__(struct lttng_ust_lib_ring_buffer_shmp
));
460 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
461 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
) * num_possible_cpus();
463 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
);
465 shmsize
+= offset_align(shmsize
, priv_data_align
);
466 shmsize
+= priv_data_size
;
468 shmobj
= shm_object_table_append(handle
->table
, shmsize
);
471 /* struct channel is at object 0, offset 0 (hardcoded) */
472 set_shmp(handle
->chan
, zalloc_shm(shmobj
, chansize
));
473 assert(handle
->chan
._ref
.index
== 0);
474 assert(handle
->chan
._ref
.offset
== 0);
475 chan
= shmp(handle
, handle
->chan
);
479 /* space for private data */
480 if (priv_data_size
) {
481 DECLARE_SHMP(void, priv_data_alloc
);
483 align_shm(shmobj
, priv_data_align
);
484 chan
->priv_data_offset
= shmobj
->allocated_len
;
485 set_shmp(priv_data_alloc
, zalloc_shm(shmobj
, priv_data_size
));
486 if (!shmp(handle
, priv_data_alloc
))
488 *priv_data
= channel_get_private(chan
);
490 chan
->priv_data_offset
= -1;
494 ret
= channel_backend_init(&chan
->backend
, name
, config
,
495 subbuf_size
, num_subbuf
, handle
);
497 goto error_backend_init
;
499 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
501 //chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval);
502 //chan->read_timer_interval = usecs_to_jiffies(read_timer_interval);
504 //init_waitqueue_head(&chan->read_wait);
505 //init_waitqueue_head(&chan->hp_wait);
507 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
509 * In case of non-hotplug cpu, if the ring-buffer is allocated
510 * in early initcall, it will not be notified of secondary cpus.
511 * In that off case, we need to allocate for all possible cpus.
513 for_each_possible_cpu(cpu
) {
514 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
515 lib_ring_buffer_start_switch_timer(buf
, handle
);
516 lib_ring_buffer_start_read_timer(buf
, handle
);
519 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
521 lib_ring_buffer_start_switch_timer(buf
, handle
);
522 lib_ring_buffer_start_read_timer(buf
, handle
);
524 ref
= &handle
->chan
._ref
;
525 shm_get_object_data(handle
, ref
, shm_fd
, wait_fd
, memory_map_size
);
530 shm_object_table_destroy(handle
->table
);
536 struct lttng_ust_shm_handle
*channel_handle_create(int shm_fd
, int wait_fd
,
537 uint64_t memory_map_size
)
539 struct lttng_ust_shm_handle
*handle
;
540 struct shm_object
*object
;
542 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
546 /* Allocate table for channel + per-cpu buffers */
547 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
549 goto error_table_alloc
;
550 /* Add channel object */
551 object
= shm_object_table_append_shadow(handle
->table
,
552 shm_fd
, wait_fd
, memory_map_size
);
554 goto error_table_object
;
555 /* struct channel is at object 0, offset 0 (hardcoded) */
556 handle
->chan
._ref
.index
= 0;
557 handle
->chan
._ref
.offset
= 0;
561 shm_object_table_destroy(handle
->table
);
567 int channel_handle_add_stream(struct lttng_ust_shm_handle
*handle
,
568 int shm_fd
, int wait_fd
, uint64_t memory_map_size
)
570 struct shm_object
*object
;
572 /* Add stream object */
573 object
= shm_object_table_append_shadow(handle
->table
,
574 shm_fd
, wait_fd
, memory_map_size
);
581 void channel_release(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
584 channel_free(chan
, handle
, shadow
);
588 * channel_destroy - Finalize, wait for q.s. and destroy channel.
589 * @chan: channel to destroy
592 * Call "destroy" callback, finalize channels, decrement the channel
593 * reference count. Note that when readers have completed data
594 * consumption of finalized channels, get_subbuf() will return -ENODATA.
595 * They should release their handle at that point.
597 void channel_destroy(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
600 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
604 channel_release(chan
, handle
, shadow
);
608 channel_unregister_notifiers(chan
, handle
);
610 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
611 for_each_channel_cpu(cpu
, chan
) {
612 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
614 if (config
->cb
.buffer_finalize
)
615 config
->cb
.buffer_finalize(buf
,
616 channel_get_private(chan
),
618 if (buf
->backend
.allocated
)
619 lib_ring_buffer_switch_slow(buf
, SWITCH_FLUSH
,
622 * Perform flush before writing to finalized.
625 CMM_ACCESS_ONCE(buf
->finalized
) = 1;
626 //wake_up_interruptible(&buf->read_wait);
629 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
631 if (config
->cb
.buffer_finalize
)
632 config
->cb
.buffer_finalize(buf
, channel_get_private(chan
), -1, handle
);
633 if (buf
->backend
.allocated
)
634 lib_ring_buffer_switch_slow(buf
, SWITCH_FLUSH
,
637 * Perform flush before writing to finalized.
640 CMM_ACCESS_ONCE(buf
->finalized
) = 1;
641 //wake_up_interruptible(&buf->read_wait);
643 CMM_ACCESS_ONCE(chan
->finalized
) = 1;
644 //wake_up_interruptible(&chan->hp_wait);
645 //wake_up_interruptible(&chan->read_wait);
647 * sessiond/consumer are keeping a reference on the shm file
648 * descriptor directly. No need to refcount.
650 channel_release(chan
, handle
, shadow
);
654 struct lttng_ust_lib_ring_buffer
*channel_get_ring_buffer(
655 const struct lttng_ust_lib_ring_buffer_config
*config
,
656 struct channel
*chan
, int cpu
,
657 struct lttng_ust_shm_handle
*handle
,
658 int *shm_fd
, int *wait_fd
,
659 uint64_t *memory_map_size
)
663 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
664 ref
= &chan
->backend
.buf
[0].shmp
._ref
;
665 shm_get_object_data(handle
, ref
, shm_fd
, wait_fd
,
667 return shmp(handle
, chan
->backend
.buf
[0].shmp
);
669 if (cpu
>= num_possible_cpus())
671 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
672 shm_get_object_data(handle
, ref
, shm_fd
, wait_fd
,
674 return shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
678 int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer
*buf
,
679 struct lttng_ust_shm_handle
*handle
,
683 if (uatomic_cmpxchg(&buf
->active_shadow_readers
, 0, 1) != 0)
688 if (uatomic_cmpxchg(&buf
->active_readers
, 0, 1) != 0)
694 void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer
*buf
,
695 struct lttng_ust_shm_handle
*handle
,
698 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
701 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_shadow_readers
) != 1);
703 uatomic_dec(&buf
->active_shadow_readers
);
706 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
708 uatomic_dec(&buf
->active_readers
);
712 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
714 * @consumed: consumed count indicating the position where to read
715 * @produced: produced count, indicates position when to stop reading
717 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
718 * data to read at consumed position, or 0 if the get operation succeeds.
721 int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer
*buf
,
722 unsigned long *consumed
, unsigned long *produced
,
723 struct lttng_ust_shm_handle
*handle
)
725 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
726 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
727 unsigned long consumed_cur
, write_offset
;
730 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
732 * Read finalized before counters.
735 consumed_cur
= uatomic_read(&buf
->consumed
);
737 * No need to issue a memory barrier between consumed count read and
738 * write offset read, because consumed count can only change
739 * concurrently in overwrite mode, and we keep a sequence counter
740 * identifier derived from the write offset to check we are getting
741 * the same sub-buffer we are expecting (the sub-buffers are atomically
742 * "tagged" upon writes, tags are checked upon read).
744 write_offset
= v_read(config
, &buf
->offset
);
747 * Check that we are not about to read the same subbuffer in
748 * which the writer head is.
750 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
754 *consumed
= consumed_cur
;
755 *produced
= subbuf_trunc(write_offset
, chan
);
761 * The memory barriers __wait_event()/wake_up_interruptible() take care
762 * of "raw_spin_is_locked" memory ordering.
771 * lib_ring_buffer_put_snapshot - move consumed counter forward
773 * @consumed_new: new consumed count value
775 void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer
*buf
,
776 unsigned long consumed_new
,
777 struct lttng_ust_shm_handle
*handle
)
779 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
780 struct channel
*chan
= shmp(handle
, bufb
->chan
);
781 unsigned long consumed
;
783 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1
784 && uatomic_read(&buf
->active_shadow_readers
) != 1);
787 * Only push the consumed value forward.
788 * If the consumed cmpxchg fails, this is because we have been pushed by
789 * the writer in flight recorder mode.
791 consumed
= uatomic_read(&buf
->consumed
);
792 while ((long) consumed
- (long) consumed_new
< 0)
793 consumed
= uatomic_cmpxchg(&buf
->consumed
, consumed
,
798 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
800 * @consumed: consumed count indicating the position where to read
802 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
803 * data to read at consumed position, or 0 if the get operation succeeds.
805 int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
806 unsigned long consumed
,
807 struct lttng_ust_shm_handle
*handle
)
809 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
810 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
811 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
816 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
818 * Read finalized before counters.
821 consumed_cur
= uatomic_read(&buf
->consumed
);
822 consumed_idx
= subbuf_index(consumed
, chan
);
823 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, consumed_idx
)->cc_sb
);
825 * Make sure we read the commit count before reading the buffer
826 * data and the write offset. Correct consumed offset ordering
827 * wrt commit count is insured by the use of cmpxchg to update
828 * the consumed offset.
831 * Local rmb to match the remote wmb to read the commit count
832 * before the buffer data and the write offset.
836 write_offset
= v_read(config
, &buf
->offset
);
839 * Check that the buffer we are getting is after or at consumed_cur
842 if ((long) subbuf_trunc(consumed
, chan
)
843 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
847 * Check that the subbuffer we are trying to consume has been
848 * already fully committed.
850 if (((commit_count
- chan
->backend
.subbuf_size
)
851 & chan
->commit_count_mask
)
852 - (buf_trunc(consumed_cur
, chan
)
853 >> chan
->backend
.num_subbuf_order
)
858 * Check that we are not about to read the same subbuffer in
859 * which the writer head is.
861 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
866 * Failure to get the subbuffer causes a busy-loop retry without going
867 * to a wait queue. These are caused by short-lived race windows where
868 * the writer is getting access to a subbuffer we were trying to get
869 * access to. Also checks that the "consumed" buffer count we are
870 * looking for matches the one contained in the subbuffer id.
872 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
873 consumed_idx
, buf_trunc_val(consumed
, chan
),
877 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
879 buf
->get_subbuf_consumed
= consumed
;
886 * The memory barriers __wait_event()/wake_up_interruptible() take care
887 * of "raw_spin_is_locked" memory ordering.
896 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
899 void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
900 struct lttng_ust_shm_handle
*handle
)
902 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
903 struct channel
*chan
= shmp(handle
, bufb
->chan
);
904 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
905 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
907 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1
908 && uatomic_read(&buf
->active_shadow_readers
) != 1);
910 if (!buf
->get_subbuf
) {
912 * Reader puts a subbuffer it did not get.
914 CHAN_WARN_ON(chan
, 1);
917 consumed
= buf
->get_subbuf_consumed
;
921 * Clear the records_unread counter. (overruns counter)
922 * Can still be non-zero if a file reader simply grabbed the data
923 * without using iterators.
924 * Can be below zero if an iterator is used on a snapshot more than
927 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
928 v_add(config
, v_read(config
,
929 &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
),
930 &bufb
->records_read
);
931 v_set(config
, &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
, 0);
932 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
933 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
934 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
937 * Exchange the reader subbuffer with the one we put in its place in the
938 * writer subbuffer table. Expect the original consumed count. If
939 * update_read_sb_index fails, this is because the writer updated the
940 * subbuffer concurrently. We should therefore keep the subbuffer we
941 * currently have: it has become invalid to try reading this sub-buffer
942 * consumed count value anyway.
944 consumed_idx
= subbuf_index(consumed
, chan
);
945 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
946 consumed_idx
, buf_trunc_val(consumed
, chan
),
949 * update_read_sb_index return value ignored. Don't exchange sub-buffer
950 * if the writer concurrently updated it.
955 * cons_offset is an iterator on all subbuffer offsets between the reader
956 * position and the writer position. (inclusive)
959 void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
960 struct channel
*chan
,
961 unsigned long cons_offset
,
963 struct lttng_ust_shm_handle
*handle
)
965 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
966 unsigned long cons_idx
, commit_count
, commit_count_sb
;
968 cons_idx
= subbuf_index(cons_offset
, chan
);
969 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, cons_idx
)->cc
);
970 commit_count_sb
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, cons_idx
)->cc_sb
);
972 if (subbuf_offset(commit_count
, chan
) != 0)
973 ERRMSG("ring buffer %s, cpu %d: "
974 "commit count in subbuffer %lu,\n"
975 "expecting multiples of %lu bytes\n"
976 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
977 chan
->backend
.name
, cpu
, cons_idx
,
978 chan
->backend
.subbuf_size
,
979 commit_count
, commit_count_sb
);
981 ERRMSG("ring buffer: %s, cpu %d: %lu bytes committed\n",
982 chan
->backend
.name
, cpu
, commit_count
);
986 void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
987 struct channel
*chan
,
989 struct lttng_ust_shm_handle
*handle
)
991 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
992 unsigned long write_offset
, cons_offset
;
995 * Can be called in the error path of allocation when
996 * trans_channel_data is not yet set.
1001 * No need to order commit_count, write_offset and cons_offset reads
1002 * because we execute at teardown when no more writer nor reader
1003 * references are left.
1005 write_offset
= v_read(config
, &buf
->offset
);
1006 cons_offset
= uatomic_read(&buf
->consumed
);
1007 if (write_offset
!= cons_offset
)
1008 ERRMSG("ring buffer %s, cpu %d: "
1009 "non-consumed data\n"
1010 " [ %lu bytes written, %lu bytes read ]\n",
1011 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1013 for (cons_offset
= uatomic_read(&buf
->consumed
);
1014 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1017 cons_offset
= subbuf_align(cons_offset
, chan
))
1018 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1023 void lib_ring_buffer_print_errors(struct channel
*chan
,
1024 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
1025 struct lttng_ust_shm_handle
*handle
)
1027 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1028 void *priv
= channel_get_private(chan
);
1030 ERRMSG("ring buffer %s, cpu %d: %lu records written, "
1031 "%lu records overrun\n",
1032 chan
->backend
.name
, cpu
,
1033 v_read(config
, &buf
->records_count
),
1034 v_read(config
, &buf
->records_overrun
));
1036 if (v_read(config
, &buf
->records_lost_full
)
1037 || v_read(config
, &buf
->records_lost_wrap
)
1038 || v_read(config
, &buf
->records_lost_big
))
1039 ERRMSG("ring buffer %s, cpu %d: records were lost. Caused by:\n"
1040 " [ %lu buffer full, %lu nest buffer wrap-around, "
1041 "%lu event too big ]\n",
1042 chan
->backend
.name
, cpu
,
1043 v_read(config
, &buf
->records_lost_full
),
1044 v_read(config
, &buf
->records_lost_wrap
),
1045 v_read(config
, &buf
->records_lost_big
));
1047 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
, handle
);
1051 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1053 * Only executed when the buffer is finalized, in SWITCH_FLUSH.
1056 void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer
*buf
,
1057 struct channel
*chan
,
1058 struct switch_offsets
*offsets
,
1060 struct lttng_ust_shm_handle
*handle
)
1062 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1063 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1064 unsigned long commit_count
;
1066 config
->cb
.buffer_begin(buf
, tsc
, oldidx
, handle
);
1069 * Order all writes to buffer before the commit count update that will
1070 * determine that the subbuffer is full.
1073 v_add(config
, config
->cb
.subbuffer_header_size(),
1074 &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1075 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1076 /* Check if the written buffer has to be delivered */
1077 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1078 commit_count
, oldidx
, handle
);
1079 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1080 offsets
->old
, commit_count
,
1081 config
->cb
.subbuffer_header_size(),
1086 * lib_ring_buffer_switch_old_end: switch old subbuffer
1088 * Note : offset_old should never be 0 here. It is ok, because we never perform
1089 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1090 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1094 void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer
*buf
,
1095 struct channel
*chan
,
1096 struct switch_offsets
*offsets
,
1098 struct lttng_ust_shm_handle
*handle
)
1100 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1101 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1102 unsigned long commit_count
, padding_size
, data_size
;
1104 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1105 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1106 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
,
1110 * Order all writes to buffer before the commit count update that will
1111 * determine that the subbuffer is full.
1114 v_add(config
, padding_size
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1115 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1116 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1117 commit_count
, oldidx
, handle
);
1118 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1119 offsets
->old
, commit_count
,
1120 padding_size
, handle
);
1124 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1126 * This code can be executed unordered : writers may already have written to the
1127 * sub-buffer before this code gets executed, caution. The commit makes sure
1128 * that this code is executed before the deliver of this sub-buffer.
1131 void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer
*buf
,
1132 struct channel
*chan
,
1133 struct switch_offsets
*offsets
,
1135 struct lttng_ust_shm_handle
*handle
)
1137 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1138 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1139 unsigned long commit_count
;
1141 config
->cb
.buffer_begin(buf
, tsc
, beginidx
, handle
);
1144 * Order all writes to buffer before the commit count update that will
1145 * determine that the subbuffer is full.
1148 v_add(config
, config
->cb
.subbuffer_header_size(),
1149 &shmp_index(handle
, buf
->commit_hot
, beginidx
)->cc
);
1150 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, beginidx
)->cc
);
1151 /* Check if the written buffer has to be delivered */
1152 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1153 commit_count
, beginidx
, handle
);
1154 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, beginidx
,
1155 offsets
->begin
, commit_count
,
1156 config
->cb
.subbuffer_header_size(),
1161 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1163 * The only remaining threads could be the ones with pending commits. They will
1164 * have to do the deliver themselves.
1167 void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer
*buf
,
1168 struct channel
*chan
,
1169 struct switch_offsets
*offsets
,
1171 struct lttng_ust_shm_handle
*handle
)
1173 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1174 unsigned long endidx
= subbuf_index(offsets
->end
- 1, chan
);
1175 unsigned long commit_count
, padding_size
, data_size
;
1177 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1178 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1179 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
,
1183 * Order all writes to buffer before the commit count update that will
1184 * determine that the subbuffer is full.
1187 v_add(config
, padding_size
, &shmp_index(handle
, buf
->commit_hot
, endidx
)->cc
);
1188 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, endidx
)->cc
);
1189 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->end
- 1,
1190 commit_count
, endidx
, handle
);
1191 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, endidx
,
1192 offsets
->end
, commit_count
,
1193 padding_size
, handle
);
1199 * !0 if execution must be aborted.
1202 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1203 struct lttng_ust_lib_ring_buffer
*buf
,
1204 struct channel
*chan
,
1205 struct switch_offsets
*offsets
,
1208 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1211 offsets
->begin
= v_read(config
, &buf
->offset
);
1212 offsets
->old
= offsets
->begin
;
1213 offsets
->switch_old_start
= 0;
1214 off
= subbuf_offset(offsets
->begin
, chan
);
1216 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1219 * Ensure we flush the header of an empty subbuffer when doing the
1220 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1221 * total data gathering duration even if there were no records saved
1222 * after the last buffer switch.
1223 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1224 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1225 * subbuffer header as appropriate.
1226 * The next record that reserves space will be responsible for
1227 * populating the following subbuffer header. We choose not to populate
1228 * the next subbuffer header here because we want to be able to use
1229 * SWITCH_ACTIVE for periodical buffer flush, which must
1230 * guarantee that all the buffer content (records and header
1231 * timestamps) are visible to the reader. This is required for
1232 * quiescence guarantees for the fusion merge.
1234 if (mode
== SWITCH_FLUSH
|| off
> 0) {
1235 if (caa_unlikely(off
== 0)) {
1237 * The client does not save any header information.
1238 * Don't switch empty subbuffer on finalize, because it
1239 * is invalid to deliver a completely empty subbuffer.
1241 if (!config
->cb
.subbuffer_header_size())
1244 * Need to write the subbuffer start header on finalize.
1246 offsets
->switch_old_start
= 1;
1248 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1250 return -1; /* we do not have to switch : buffer is empty */
1251 /* Note: old points to the next subbuf at offset 0 */
1252 offsets
->end
= offsets
->begin
;
1257 * Force a sub-buffer switch. This operation is completely reentrant : can be
1258 * called while tracing is active with absolutely no lock held.
1260 * Note, however, that as a v_cmpxchg is used for some atomic
1261 * operations, this function must be called from the CPU which owns the buffer
1262 * for a ACTIVE flush.
1264 void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer
*buf
, enum switch_mode mode
,
1265 struct lttng_ust_shm_handle
*handle
)
1267 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1268 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1269 struct switch_offsets offsets
;
1270 unsigned long oldidx
;
1276 * Perform retryable operations.
1279 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1281 return; /* Switch not needed */
1282 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1286 * Atomically update last_tsc. This update races against concurrent
1287 * atomic updates, but the race will always cause supplementary full TSC
1288 * records, never the opposite (missing a full TSC record when it would
1291 save_last_tsc(config
, buf
, tsc
);
1294 * Push the reader if necessary
1296 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1298 oldidx
= subbuf_index(offsets
.old
, chan
);
1299 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
, handle
);
1302 * May need to populate header start on SWITCH_FLUSH.
1304 if (offsets
.switch_old_start
) {
1305 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
, handle
);
1306 offsets
.old
+= config
->cb
.subbuffer_header_size();
1310 * Switch old subbuffer.
1312 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
, handle
);
1318 * -ENOSPC if event size is too large for packet.
1319 * -ENOBUFS if there is currently not enough space in buffer for the event.
1320 * -EIO if data cannot be written into the buffer for any other reason.
1323 int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer
*buf
,
1324 struct channel
*chan
,
1325 struct switch_offsets
*offsets
,
1326 struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
1328 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1329 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
1330 unsigned long reserve_commit_diff
;
1332 offsets
->begin
= v_read(config
, &buf
->offset
);
1333 offsets
->old
= offsets
->begin
;
1334 offsets
->switch_new_start
= 0;
1335 offsets
->switch_new_end
= 0;
1336 offsets
->switch_old_end
= 0;
1337 offsets
->pre_header_padding
= 0;
1339 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1340 if ((int64_t) ctx
->tsc
== -EIO
)
1343 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1344 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
1346 if (caa_unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1347 offsets
->switch_new_start
= 1; /* For offsets->begin */
1349 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1351 &offsets
->pre_header_padding
,
1354 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1357 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
) +
1358 offsets
->size
> chan
->backend
.subbuf_size
)) {
1359 offsets
->switch_old_end
= 1; /* For offsets->old */
1360 offsets
->switch_new_start
= 1; /* For offsets->begin */
1363 if (caa_unlikely(offsets
->switch_new_start
)) {
1364 unsigned long sb_index
;
1367 * We are typically not filling the previous buffer completely.
1369 if (caa_likely(offsets
->switch_old_end
))
1370 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1371 offsets
->begin
= offsets
->begin
1372 + config
->cb
.subbuffer_header_size();
1373 /* Test new buffer integrity */
1374 sb_index
= subbuf_index(offsets
->begin
, chan
);
1375 reserve_commit_diff
=
1376 (buf_trunc(offsets
->begin
, chan
)
1377 >> chan
->backend
.num_subbuf_order
)
1378 - ((unsigned long) v_read(config
,
1379 &shmp_index(handle
, buf
->commit_cold
, sb_index
)->cc_sb
)
1380 & chan
->commit_count_mask
);
1381 if (caa_likely(reserve_commit_diff
== 0)) {
1382 /* Next subbuffer not being written to. */
1383 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1384 subbuf_trunc(offsets
->begin
, chan
)
1385 - subbuf_trunc((unsigned long)
1386 uatomic_read(&buf
->consumed
), chan
)
1387 >= chan
->backend
.buf_size
)) {
1389 * We do not overwrite non consumed buffers
1390 * and we are full : record is lost.
1392 v_inc(config
, &buf
->records_lost_full
);
1396 * Next subbuffer not being written to, and we
1397 * are either in overwrite mode or the buffer is
1398 * not full. It's safe to write in this new
1404 * Next subbuffer reserve offset does not match the
1405 * commit offset. Drop record in producer-consumer and
1406 * overwrite mode. Caused by either a writer OOPS or too
1407 * many nested writes over a reserve/commit pair.
1409 v_inc(config
, &buf
->records_lost_wrap
);
1413 config
->cb
.record_header_size(config
, chan
,
1415 &offsets
->pre_header_padding
,
1418 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1421 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
)
1422 + offsets
->size
> chan
->backend
.subbuf_size
)) {
1424 * Record too big for subbuffers, report error, don't
1425 * complete the sub-buffer switch.
1427 v_inc(config
, &buf
->records_lost_big
);
1431 * We just made a successful buffer switch and the
1432 * record fits in the new subbuffer. Let's write.
1437 * Record fits in the current buffer and we are not on a switch
1438 * boundary. It's safe to write.
1441 offsets
->end
= offsets
->begin
+ offsets
->size
;
1443 if (caa_unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
1445 * The offset_end will fall at the very beginning of the next
1448 offsets
->switch_new_end
= 1; /* For offsets->begin */
1454 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1455 * @ctx: ring buffer context.
1457 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
1458 * -EIO for other errors, else returns 0.
1459 * It will take care of sub-buffer switching.
1461 int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
1463 struct channel
*chan
= ctx
->chan
;
1464 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
1465 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1466 struct lttng_ust_lib_ring_buffer
*buf
;
1467 struct switch_offsets offsets
;
1470 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
1471 buf
= shmp(handle
, chan
->backend
.buf
[ctx
->cpu
].shmp
);
1473 buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
1479 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
1481 if (caa_unlikely(ret
))
1483 } while (caa_unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
1488 * Atomically update last_tsc. This update races against concurrent
1489 * atomic updates, but the race will always cause supplementary full TSC
1490 * records, never the opposite (missing a full TSC record when it would
1493 save_last_tsc(config
, buf
, ctx
->tsc
);
1496 * Push the reader if necessary
1498 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
1501 * Clear noref flag for this subbuffer.
1503 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1504 subbuf_index(offsets
.end
- 1, chan
),
1508 * Switch old subbuffer if needed.
1510 if (caa_unlikely(offsets
.switch_old_end
)) {
1511 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1512 subbuf_index(offsets
.old
- 1, chan
),
1514 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1518 * Populate new subbuffer.
1520 if (caa_unlikely(offsets
.switch_new_start
))
1521 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1523 if (caa_unlikely(offsets
.switch_new_end
))
1524 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1526 ctx
->slot_size
= offsets
.size
;
1527 ctx
->pre_offset
= offsets
.begin
;
1528 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;