2 * ring_buffer_frontend.c
4 * (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
6 * Ring buffer wait-free buffer synchronization. Producer-consumer and flight
7 * recorder (overwrite) modes. See thesis:
9 * Desnoyers, Mathieu (2009), "Low-Impact Operating System Tracing", Ph.D.
10 * dissertation, Ecole Polytechnique de Montreal.
11 * http://www.lttng.org/pub/thesis/desnoyers-dissertation-2009-12.pdf
13 * - Algorithm presentation in Chapter 5:
14 * "Lockless Multi-Core High-Throughput Buffering".
15 * - Algorithm formal verification in Section 8.6:
16 * "Formal verification of LTTng"
19 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
21 * Inspired from LTT and RelayFS:
22 * Karim Yaghmour <karim@opersys.com>
23 * Tom Zanussi <zanussi@us.ibm.com>
24 * Bob Wisniewski <bob@watson.ibm.com>
26 * Bob Wisniewski <bob@watson.ibm.com>
28 * Buffer reader semantic :
31 * while buffer is not finalized and empty
33 * - if return value != 0, continue
34 * - splice one subbuffer worth of data to a pipe
35 * - splice the data from pipe to disk/network
38 * Dual LGPL v2.1/GPL v2 license.
42 #include <sys/types.h>
46 #include <urcu/compiler.h>
51 #include <lttng/ringbuffer-config.h>
58 #define max(a, b) ((a) > (b) ? (a) : (b))
61 /* Print DBG() messages about events lost only every 1048576 hits */
62 #define DBG_PRINT_NR_LOST (1UL << 20)
65 * Use POSIX SHM: shm_open(3) and shm_unlink(3).
66 * close(2) to close the fd returned by shm_open.
67 * shm_unlink releases the shared memory object name.
68 * ftruncate(2) sets the size of the memory object.
69 * mmap/munmap maps the shared memory obj to a virtual address in the
70 * calling proceess (should be done both in libust and consumer).
71 * See shm_overview(7) for details.
72 * Pass file descriptor returned by shm_open(3) to ltt-sessiond through
75 * Since we don't need to access the object using its name, we can
76 * immediately shm_unlink(3) it, and only keep the handle with its file
81 * Internal structure representing offsets to use at a sub-buffer switch.
83 struct switch_offsets
{
84 unsigned long begin
, end
, old
;
85 size_t pre_header_padding
, size
;
86 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
90 __thread
unsigned int lib_ring_buffer_nesting
;
93 * TODO: this is unused. Errors are saved within the ring buffer.
94 * Eventually, allow consumerd to print these errors.
97 void lib_ring_buffer_print_errors(struct channel
*chan
,
98 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
99 struct lttng_ust_shm_handle
*handle
)
100 __attribute__((unused
));
103 * lib_ring_buffer_reset - Reset ring buffer to initial values.
106 * Effectively empty the ring buffer. Should be called when the buffer is not
107 * used for writing. The ring buffer can be opened for reading, but the reader
108 * should not be using the iterator concurrently with reset. The previous
109 * current iterator record is reset.
111 void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer
*buf
,
112 struct lttng_ust_shm_handle
*handle
)
114 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
115 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
119 * Reset iterator first. It will put the subbuffer if it currently holds
122 v_set(config
, &buf
->offset
, 0);
123 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
124 v_set(config
, &shmp_index(handle
, buf
->commit_hot
, i
)->cc
, 0);
125 v_set(config
, &shmp_index(handle
, buf
->commit_hot
, i
)->seq
, 0);
126 v_set(config
, &shmp_index(handle
, buf
->commit_cold
, i
)->cc_sb
, 0);
128 uatomic_set(&buf
->consumed
, 0);
129 uatomic_set(&buf
->record_disabled
, 0);
130 v_set(config
, &buf
->last_tsc
, 0);
131 lib_ring_buffer_backend_reset(&buf
->backend
, handle
);
132 /* Don't reset number of active readers */
133 v_set(config
, &buf
->records_lost_full
, 0);
134 v_set(config
, &buf
->records_lost_wrap
, 0);
135 v_set(config
, &buf
->records_lost_big
, 0);
136 v_set(config
, &buf
->records_count
, 0);
137 v_set(config
, &buf
->records_overrun
, 0);
142 * channel_reset - Reset channel to initial values.
145 * Effectively empty the channel. Should be called when the channel is not used
146 * for writing. The channel can be opened for reading, but the reader should not
147 * be using the iterator concurrently with reset. The previous current iterator
150 void channel_reset(struct channel
*chan
)
153 * Reset iterators first. Will put the subbuffer if held for reading.
155 uatomic_set(&chan
->record_disabled
, 0);
156 /* Don't reset commit_count_mask, still valid */
157 channel_backend_reset(&chan
->backend
);
158 /* Don't reset switch/read timer interval */
159 /* Don't reset notifiers and notifier enable bits */
160 /* Don't reset reader reference count */
164 * Must be called under cpu hotplug protection.
166 int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer
*buf
,
167 struct channel_backend
*chanb
, int cpu
,
168 struct lttng_ust_shm_handle
*handle
,
169 struct shm_object
*shmobj
)
171 const struct lttng_ust_lib_ring_buffer_config
*config
= &chanb
->config
;
172 struct channel
*chan
= caa_container_of(chanb
, struct channel
, backend
);
173 void *priv
= channel_get_private(chan
);
174 size_t subbuf_header_size
;
178 /* Test for cpu hotplug */
179 if (buf
->backend
.allocated
)
182 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
,
183 cpu
, handle
, shmobj
);
187 align_shm(shmobj
, __alignof__(struct commit_counters_hot
));
188 set_shmp(buf
->commit_hot
,
190 sizeof(struct commit_counters_hot
) * chan
->backend
.num_subbuf
));
191 if (!shmp(handle
, buf
->commit_hot
)) {
196 align_shm(shmobj
, __alignof__(struct commit_counters_cold
));
197 set_shmp(buf
->commit_cold
,
199 sizeof(struct commit_counters_cold
) * chan
->backend
.num_subbuf
));
200 if (!shmp(handle
, buf
->commit_cold
)) {
206 * Write the subbuffer header for first subbuffer so we know the total
207 * duration of data gathering.
209 subbuf_header_size
= config
->cb
.subbuffer_header_size();
210 v_set(config
, &buf
->offset
, subbuf_header_size
);
211 subbuffer_id_clear_noref(config
, &shmp_index(handle
, buf
->backend
.buf_wsb
, 0)->id
);
212 tsc
= config
->cb
.ring_buffer_clock_read(shmp(handle
, buf
->backend
.chan
));
213 config
->cb
.buffer_begin(buf
, tsc
, 0, handle
);
214 v_add(config
, subbuf_header_size
, &shmp_index(handle
, buf
->commit_hot
, 0)->cc
);
216 if (config
->cb
.buffer_create
) {
217 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
, handle
);
221 buf
->backend
.allocated
= 1;
226 /* commit_cold will be freed by shm teardown */
228 /* commit_hot will be freed by shm teardown */
234 static void switch_buffer_timer(unsigned long data
)
236 struct lttng_ust_lib_ring_buffer
*buf
= (struct lttng_ust_lib_ring_buffer
*)data
;
237 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
238 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
241 * Only flush buffers periodically if readers are active.
243 if (uatomic_read(&buf
->active_readers
) || uatomic_read(&buf
->active_shadow_readers
))
244 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
, handle
);
247 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
248 // mod_timer_pinned(&buf->switch_timer,
249 // jiffies + chan->switch_timer_interval);
251 // mod_timer(&buf->switch_timer,
252 // jiffies + chan->switch_timer_interval);
256 static void lib_ring_buffer_start_switch_timer(struct lttng_ust_lib_ring_buffer
*buf
,
257 struct lttng_ust_shm_handle
*handle
)
259 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
260 //const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
262 if (!chan
->switch_timer_interval
|| buf
->switch_timer_enabled
)
265 //init_timer(&buf->switch_timer);
266 //buf->switch_timer.function = switch_buffer_timer;
267 //buf->switch_timer.expires = jiffies + chan->switch_timer_interval;
268 //buf->switch_timer.data = (unsigned long)buf;
269 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
270 // add_timer_on(&buf->switch_timer, buf->backend.cpu);
272 // add_timer(&buf->switch_timer);
273 buf
->switch_timer_enabled
= 1;
276 static void lib_ring_buffer_stop_switch_timer(struct lttng_ust_lib_ring_buffer
*buf
,
277 struct lttng_ust_shm_handle
*handle
)
279 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
281 if (!chan
->switch_timer_interval
|| !buf
->switch_timer_enabled
)
285 //del_timer_sync(&buf->switch_timer);
286 buf
->switch_timer_enabled
= 0;
291 * Polling timer to check the channels for data.
293 static void read_buffer_timer(unsigned long data
)
295 struct lttng_ust_lib_ring_buffer
*buf
= (struct lttng_ust_lib_ring_buffer
*)data
;
296 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
297 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
299 CHAN_WARN_ON(chan
, !buf
->backend
.allocated
);
301 if (uatomic_read(&buf
->active_readers
) || uatomic_read(&buf
->active_shadow_readers
))
302 && lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
304 //wake_up_interruptible(&buf->read_wait);
305 //wake_up_interruptible(&chan->read_wait);
309 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
310 // mod_timer_pinned(&buf->read_timer,
311 // jiffies + chan->read_timer_interval);
313 // mod_timer(&buf->read_timer,
314 // jiffies + chan->read_timer_interval);
318 static void lib_ring_buffer_start_read_timer(struct lttng_ust_lib_ring_buffer
*buf
,
319 struct lttng_ust_shm_handle
*handle
)
321 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
322 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
324 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
325 || !chan
->read_timer_interval
326 || buf
->read_timer_enabled
)
330 //init_timer(&buf->read_timer);
331 //buf->read_timer.function = read_buffer_timer;
332 //buf->read_timer.expires = jiffies + chan->read_timer_interval;
333 //buf->read_timer.data = (unsigned long)buf;
335 //if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
336 // add_timer_on(&buf->read_timer, buf->backend.cpu);
338 // add_timer(&buf->read_timer);
339 buf
->read_timer_enabled
= 1;
342 static void lib_ring_buffer_stop_read_timer(struct lttng_ust_lib_ring_buffer
*buf
,
343 struct lttng_ust_shm_handle
*handle
)
345 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
346 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
348 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
349 || !chan
->read_timer_interval
350 || !buf
->read_timer_enabled
)
354 //del_timer_sync(&buf->read_timer);
356 * do one more check to catch data that has been written in the last
359 if (lib_ring_buffer_poll_deliver(config
, buf
, chan
, handle
)) {
361 //wake_up_interruptible(&buf->read_wait);
362 //wake_up_interruptible(&chan->read_wait);
364 buf
->read_timer_enabled
= 0;
367 static void channel_unregister_notifiers(struct channel
*chan
,
368 struct lttng_ust_shm_handle
*handle
)
370 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
373 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
374 for_each_possible_cpu(cpu
) {
375 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
377 lib_ring_buffer_stop_switch_timer(buf
, handle
);
378 lib_ring_buffer_stop_read_timer(buf
, handle
);
381 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
383 lib_ring_buffer_stop_switch_timer(buf
, handle
);
384 lib_ring_buffer_stop_read_timer(buf
, handle
);
386 //channel_backend_unregister_notifiers(&chan->backend);
389 static void channel_free(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
393 channel_backend_free(&chan
->backend
, handle
);
394 /* chan is freed by shm teardown */
395 shm_object_table_destroy(handle
->table
);
400 * channel_create - Create channel.
401 * @config: ring buffer instance configuration
402 * @name: name of the channel
403 * @priv_data: ring buffer client private data area pointer (output)
404 * @priv_data_size: length, in bytes, of the private data area.
405 * @priv_data_init: initialization data for private data.
406 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
407 * address mapping. It is used only by RING_BUFFER_STATIC
408 * configuration. It can be set to NULL for other backends.
409 * @subbuf_size: subbuffer size
410 * @num_subbuf: number of subbuffers
411 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
412 * padding to let readers get those sub-buffers.
413 * Used for live streaming.
414 * @read_timer_interval: Time interval (in us) to wake up pending readers.
417 * Returns NULL on failure.
419 struct lttng_ust_shm_handle
*channel_create(const struct lttng_ust_lib_ring_buffer_config
*config
,
422 size_t priv_data_align
,
423 size_t priv_data_size
,
424 void *priv_data_init
,
425 void *buf_addr
, size_t subbuf_size
,
426 size_t num_subbuf
, unsigned int switch_timer_interval
,
427 unsigned int read_timer_interval
,
428 int **shm_fd
, int **wait_fd
, uint64_t **memory_map_size
)
431 size_t shmsize
, chansize
;
432 struct channel
*chan
;
433 struct lttng_ust_shm_handle
*handle
;
434 struct shm_object
*shmobj
;
437 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
438 read_timer_interval
))
441 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
445 /* Allocate table for channel + per-cpu buffers */
446 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
448 goto error_table_alloc
;
450 /* Calculate the shm allocation layout */
451 shmsize
= sizeof(struct channel
);
452 shmsize
+= offset_align(shmsize
, __alignof__(struct lttng_ust_lib_ring_buffer_shmp
));
453 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
454 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
) * num_possible_cpus();
456 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
);
458 shmsize
+= offset_align(shmsize
, priv_data_align
);
459 shmsize
+= priv_data_size
;
461 shmobj
= shm_object_table_append(handle
->table
, shmsize
);
464 /* struct channel is at object 0, offset 0 (hardcoded) */
465 set_shmp(handle
->chan
, zalloc_shm(shmobj
, chansize
));
466 assert(handle
->chan
._ref
.index
== 0);
467 assert(handle
->chan
._ref
.offset
== 0);
468 chan
= shmp(handle
, handle
->chan
);
472 /* space for private data */
473 if (priv_data_size
) {
474 DECLARE_SHMP(void, priv_data_alloc
);
476 align_shm(shmobj
, priv_data_align
);
477 chan
->priv_data_offset
= shmobj
->allocated_len
;
478 set_shmp(priv_data_alloc
, zalloc_shm(shmobj
, priv_data_size
));
479 if (!shmp(handle
, priv_data_alloc
))
481 *priv_data
= channel_get_private(chan
);
482 memcpy(*priv_data
, priv_data_init
, priv_data_size
);
484 chan
->priv_data_offset
= -1;
488 ret
= channel_backend_init(&chan
->backend
, name
, config
,
489 subbuf_size
, num_subbuf
, handle
);
491 goto error_backend_init
;
493 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
495 //chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval);
496 //chan->read_timer_interval = usecs_to_jiffies(read_timer_interval);
498 //init_waitqueue_head(&chan->read_wait);
499 //init_waitqueue_head(&chan->hp_wait);
501 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
503 * In case of non-hotplug cpu, if the ring-buffer is allocated
504 * in early initcall, it will not be notified of secondary cpus.
505 * In that off case, we need to allocate for all possible cpus.
507 for_each_possible_cpu(cpu
) {
508 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
509 lib_ring_buffer_start_switch_timer(buf
, handle
);
510 lib_ring_buffer_start_read_timer(buf
, handle
);
513 struct lttng_ust_lib_ring_buffer
*buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
515 lib_ring_buffer_start_switch_timer(buf
, handle
);
516 lib_ring_buffer_start_read_timer(buf
, handle
);
518 ref
= &handle
->chan
._ref
;
519 shm_get_object_data(handle
, ref
, shm_fd
, wait_fd
, memory_map_size
);
524 shm_object_table_destroy(handle
->table
);
530 struct lttng_ust_shm_handle
*channel_handle_create(int shm_fd
, int wait_fd
,
531 uint64_t memory_map_size
)
533 struct lttng_ust_shm_handle
*handle
;
534 struct shm_object
*object
;
536 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
540 /* Allocate table for channel + per-cpu buffers */
541 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
543 goto error_table_alloc
;
544 /* Add channel object */
545 object
= shm_object_table_append_shadow(handle
->table
,
546 shm_fd
, wait_fd
, memory_map_size
);
548 goto error_table_object
;
549 /* struct channel is at object 0, offset 0 (hardcoded) */
550 handle
->chan
._ref
.index
= 0;
551 handle
->chan
._ref
.offset
= 0;
555 shm_object_table_destroy(handle
->table
);
561 int channel_handle_add_stream(struct lttng_ust_shm_handle
*handle
,
562 int shm_fd
, int wait_fd
, uint64_t memory_map_size
)
564 struct shm_object
*object
;
566 /* Add stream object */
567 object
= shm_object_table_append_shadow(handle
->table
,
568 shm_fd
, wait_fd
, memory_map_size
);
575 void channel_release(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
578 channel_free(chan
, handle
, shadow
);
582 * channel_destroy - Finalize, wait for q.s. and destroy channel.
583 * @chan: channel to destroy
586 * Call "destroy" callback, finalize channels, decrement the channel
587 * reference count. Note that when readers have completed data
588 * consumption of finalized channels, get_subbuf() will return -ENODATA.
589 * They should release their handle at that point.
591 void channel_destroy(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
595 channel_release(chan
, handle
, shadow
);
599 channel_unregister_notifiers(chan
, handle
);
602 * Note: the consumer takes care of finalizing and switching the
607 * sessiond/consumer are keeping a reference on the shm file
608 * descriptor directly. No need to refcount.
610 channel_release(chan
, handle
, shadow
);
614 struct lttng_ust_lib_ring_buffer
*channel_get_ring_buffer(
615 const struct lttng_ust_lib_ring_buffer_config
*config
,
616 struct channel
*chan
, int cpu
,
617 struct lttng_ust_shm_handle
*handle
,
618 int **shm_fd
, int **wait_fd
,
619 uint64_t **memory_map_size
)
623 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
624 ref
= &chan
->backend
.buf
[0].shmp
._ref
;
625 shm_get_object_data(handle
, ref
, shm_fd
, wait_fd
,
627 return shmp(handle
, chan
->backend
.buf
[0].shmp
);
629 if (cpu
>= num_possible_cpus())
631 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
632 shm_get_object_data(handle
, ref
, shm_fd
, wait_fd
,
634 return shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
638 int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer
*buf
,
639 struct lttng_ust_shm_handle
*handle
,
643 if (uatomic_cmpxchg(&buf
->active_shadow_readers
, 0, 1) != 0)
648 if (uatomic_cmpxchg(&buf
->active_readers
, 0, 1) != 0)
654 void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer
*buf
,
655 struct lttng_ust_shm_handle
*handle
,
658 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
661 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_shadow_readers
) != 1);
663 uatomic_dec(&buf
->active_shadow_readers
);
666 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
668 uatomic_dec(&buf
->active_readers
);
672 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
674 * @consumed: consumed count indicating the position where to read
675 * @produced: produced count, indicates position when to stop reading
677 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
678 * data to read at consumed position, or 0 if the get operation succeeds.
681 int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer
*buf
,
682 unsigned long *consumed
, unsigned long *produced
,
683 struct lttng_ust_shm_handle
*handle
)
685 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
686 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
687 unsigned long consumed_cur
, write_offset
;
690 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
692 * Read finalized before counters.
695 consumed_cur
= uatomic_read(&buf
->consumed
);
697 * No need to issue a memory barrier between consumed count read and
698 * write offset read, because consumed count can only change
699 * concurrently in overwrite mode, and we keep a sequence counter
700 * identifier derived from the write offset to check we are getting
701 * the same sub-buffer we are expecting (the sub-buffers are atomically
702 * "tagged" upon writes, tags are checked upon read).
704 write_offset
= v_read(config
, &buf
->offset
);
707 * Check that we are not about to read the same subbuffer in
708 * which the writer head is.
710 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
714 *consumed
= consumed_cur
;
715 *produced
= subbuf_trunc(write_offset
, chan
);
721 * The memory barriers __wait_event()/wake_up_interruptible() take care
722 * of "raw_spin_is_locked" memory ordering.
731 * lib_ring_buffer_put_snapshot - move consumed counter forward
733 * @consumed_new: new consumed count value
735 void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer
*buf
,
736 unsigned long consumed_new
,
737 struct lttng_ust_shm_handle
*handle
)
739 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
740 struct channel
*chan
= shmp(handle
, bufb
->chan
);
741 unsigned long consumed
;
743 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1
744 && uatomic_read(&buf
->active_shadow_readers
) != 1);
747 * Only push the consumed value forward.
748 * If the consumed cmpxchg fails, this is because we have been pushed by
749 * the writer in flight recorder mode.
751 consumed
= uatomic_read(&buf
->consumed
);
752 while ((long) consumed
- (long) consumed_new
< 0)
753 consumed
= uatomic_cmpxchg(&buf
->consumed
, consumed
,
758 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
760 * @consumed: consumed count indicating the position where to read
762 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
763 * data to read at consumed position, or 0 if the get operation succeeds.
765 int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
766 unsigned long consumed
,
767 struct lttng_ust_shm_handle
*handle
)
769 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
770 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
771 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
776 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
778 * Read finalized before counters.
781 consumed_cur
= uatomic_read(&buf
->consumed
);
782 consumed_idx
= subbuf_index(consumed
, chan
);
783 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, consumed_idx
)->cc_sb
);
785 * Make sure we read the commit count before reading the buffer
786 * data and the write offset. Correct consumed offset ordering
787 * wrt commit count is insured by the use of cmpxchg to update
788 * the consumed offset.
791 * Local rmb to match the remote wmb to read the commit count
792 * before the buffer data and the write offset.
796 write_offset
= v_read(config
, &buf
->offset
);
799 * Check that the buffer we are getting is after or at consumed_cur
802 if ((long) subbuf_trunc(consumed
, chan
)
803 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
807 * Check that the subbuffer we are trying to consume has been
808 * already fully committed.
810 if (((commit_count
- chan
->backend
.subbuf_size
)
811 & chan
->commit_count_mask
)
812 - (buf_trunc(consumed_cur
, chan
)
813 >> chan
->backend
.num_subbuf_order
)
818 * Check that we are not about to read the same subbuffer in
819 * which the writer head is.
821 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
826 * Failure to get the subbuffer causes a busy-loop retry without going
827 * to a wait queue. These are caused by short-lived race windows where
828 * the writer is getting access to a subbuffer we were trying to get
829 * access to. Also checks that the "consumed" buffer count we are
830 * looking for matches the one contained in the subbuffer id.
832 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
833 consumed_idx
, buf_trunc_val(consumed
, chan
),
837 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
839 buf
->get_subbuf_consumed
= consumed
;
846 * The memory barriers __wait_event()/wake_up_interruptible() take care
847 * of "raw_spin_is_locked" memory ordering.
856 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
859 void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
860 struct lttng_ust_shm_handle
*handle
)
862 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
863 struct channel
*chan
= shmp(handle
, bufb
->chan
);
864 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
865 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
867 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1
868 && uatomic_read(&buf
->active_shadow_readers
) != 1);
870 if (!buf
->get_subbuf
) {
872 * Reader puts a subbuffer it did not get.
874 CHAN_WARN_ON(chan
, 1);
877 consumed
= buf
->get_subbuf_consumed
;
881 * Clear the records_unread counter. (overruns counter)
882 * Can still be non-zero if a file reader simply grabbed the data
883 * without using iterators.
884 * Can be below zero if an iterator is used on a snapshot more than
887 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
888 v_add(config
, v_read(config
,
889 &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
),
890 &bufb
->records_read
);
891 v_set(config
, &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
, 0);
892 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
893 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
894 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
897 * Exchange the reader subbuffer with the one we put in its place in the
898 * writer subbuffer table. Expect the original consumed count. If
899 * update_read_sb_index fails, this is because the writer updated the
900 * subbuffer concurrently. We should therefore keep the subbuffer we
901 * currently have: it has become invalid to try reading this sub-buffer
902 * consumed count value anyway.
904 consumed_idx
= subbuf_index(consumed
, chan
);
905 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
906 consumed_idx
, buf_trunc_val(consumed
, chan
),
909 * update_read_sb_index return value ignored. Don't exchange sub-buffer
910 * if the writer concurrently updated it.
915 * cons_offset is an iterator on all subbuffer offsets between the reader
916 * position and the writer position. (inclusive)
919 void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
920 struct channel
*chan
,
921 unsigned long cons_offset
,
923 struct lttng_ust_shm_handle
*handle
)
925 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
926 unsigned long cons_idx
, commit_count
, commit_count_sb
;
928 cons_idx
= subbuf_index(cons_offset
, chan
);
929 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, cons_idx
)->cc
);
930 commit_count_sb
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, cons_idx
)->cc_sb
);
932 if (subbuf_offset(commit_count
, chan
) != 0)
933 DBG("ring buffer %s, cpu %d: "
934 "commit count in subbuffer %lu,\n"
935 "expecting multiples of %lu bytes\n"
936 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
937 chan
->backend
.name
, cpu
, cons_idx
,
938 chan
->backend
.subbuf_size
,
939 commit_count
, commit_count_sb
);
941 DBG("ring buffer: %s, cpu %d: %lu bytes committed\n",
942 chan
->backend
.name
, cpu
, commit_count
);
946 void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
947 struct channel
*chan
,
949 struct lttng_ust_shm_handle
*handle
)
951 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
952 unsigned long write_offset
, cons_offset
;
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
)) {
1342 unsigned long nr_lost
;
1345 * We do not overwrite non consumed buffers
1346 * and we are full : record is lost.
1348 nr_lost
= v_read(config
, &buf
->records_lost_full
);
1349 v_inc(config
, &buf
->records_lost_full
);
1350 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
1351 DBG("%lu or more records lost in (%s:%d) (buffer full)\n",
1352 nr_lost
+ 1, chan
->backend
.name
,
1358 * Next subbuffer not being written to, and we
1359 * are either in overwrite mode or the buffer is
1360 * not full. It's safe to write in this new
1365 unsigned long nr_lost
;
1368 * Next subbuffer reserve offset does not match the
1369 * commit offset. Drop record in producer-consumer and
1370 * overwrite mode. Caused by either a writer OOPS or too
1371 * many nested writes over a reserve/commit pair.
1373 nr_lost
= v_read(config
, &buf
->records_lost_wrap
);
1374 v_inc(config
, &buf
->records_lost_wrap
);
1375 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
1376 DBG("%lu or more records lost in (%s:%d) (wrap-around)\n",
1377 nr_lost
+ 1, chan
->backend
.name
,
1383 config
->cb
.record_header_size(config
, chan
,
1385 &offsets
->pre_header_padding
,
1388 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1391 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
)
1392 + offsets
->size
> chan
->backend
.subbuf_size
)) {
1393 unsigned long nr_lost
;
1396 * Record too big for subbuffers, report error, don't
1397 * complete the sub-buffer switch.
1399 nr_lost
= v_read(config
, &buf
->records_lost_big
);
1400 v_inc(config
, &buf
->records_lost_big
);
1401 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
1402 DBG("%lu or more records lost in (%s:%d) record size "
1403 " of %zu bytes is too large for buffer\n",
1404 nr_lost
+ 1, chan
->backend
.name
,
1405 buf
->backend
.cpu
, offsets
->size
);
1410 * We just made a successful buffer switch and the
1411 * record fits in the new subbuffer. Let's write.
1416 * Record fits in the current buffer and we are not on a switch
1417 * boundary. It's safe to write.
1420 offsets
->end
= offsets
->begin
+ offsets
->size
;
1422 if (caa_unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
1424 * The offset_end will fall at the very beginning of the next
1427 offsets
->switch_new_end
= 1; /* For offsets->begin */
1433 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1434 * @ctx: ring buffer context.
1436 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
1437 * -EIO for other errors, else returns 0.
1438 * It will take care of sub-buffer switching.
1440 int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
1442 struct channel
*chan
= ctx
->chan
;
1443 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
1444 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1445 struct lttng_ust_lib_ring_buffer
*buf
;
1446 struct switch_offsets offsets
;
1449 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
1450 buf
= shmp(handle
, chan
->backend
.buf
[ctx
->cpu
].shmp
);
1452 buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
1458 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
1460 if (caa_unlikely(ret
))
1462 } while (caa_unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
1467 * Atomically update last_tsc. This update races against concurrent
1468 * atomic updates, but the race will always cause supplementary full TSC
1469 * records, never the opposite (missing a full TSC record when it would
1472 save_last_tsc(config
, buf
, ctx
->tsc
);
1475 * Push the reader if necessary
1477 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
1480 * Clear noref flag for this subbuffer.
1482 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1483 subbuf_index(offsets
.end
- 1, chan
),
1487 * Switch old subbuffer if needed.
1489 if (caa_unlikely(offsets
.switch_old_end
)) {
1490 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1491 subbuf_index(offsets
.old
- 1, chan
),
1493 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1497 * Populate new subbuffer.
1499 if (caa_unlikely(offsets
.switch_new_start
))
1500 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1502 if (caa_unlikely(offsets
.switch_new_end
))
1503 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
1505 ctx
->slot_size
= offsets
.size
;
1506 ctx
->pre_offset
= offsets
.begin
;
1507 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;