2 * SPDX-License-Identifier: LGPL-2.1-only
4 * Copyright (C) 2005-2012 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
40 #include <sys/types.h>
49 #include <urcu/compiler.h>
51 #include <urcu/tls-compat.h>
53 #include <ust-helper.h>
55 #include <lttng/ust-align.h>
56 #include <lttng/ringbuffer-context.h>
59 #include "ringbuffer-config.h"
65 #include "../liblttng-ust/compat.h" /* For ENODATA */
67 /* Print DBG() messages about events lost only every 1048576 hits */
68 #define DBG_PRINT_NR_LOST (1UL << 20)
70 #define LTTNG_UST_RB_SIG_FLUSH SIGRTMIN
71 #define LTTNG_UST_RB_SIG_READ SIGRTMIN + 1
72 #define LTTNG_UST_RB_SIG_TEARDOWN SIGRTMIN + 2
73 #define CLOCKID CLOCK_MONOTONIC
74 #define LTTNG_UST_RING_BUFFER_GET_RETRY 10
75 #define LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS 10
76 #define RETRY_DELAY_MS 100 /* 100 ms. */
79 * Non-static to ensure the compiler does not optimize away the xor.
81 uint8_t lttng_crash_magic_xor
[] = RB_CRASH_DUMP_ABI_MAGIC_XOR
;
84 * Use POSIX SHM: shm_open(3) and shm_unlink(3).
85 * close(2) to close the fd returned by shm_open.
86 * shm_unlink releases the shared memory object name.
87 * ftruncate(2) sets the size of the memory object.
88 * mmap/munmap maps the shared memory obj to a virtual address in the
89 * calling proceess (should be done both in libust and consumer).
90 * See shm_overview(7) for details.
91 * Pass file descriptor returned by shm_open(3) to ltt-sessiond through
94 * Since we don't need to access the object using its name, we can
95 * immediately shm_unlink(3) it, and only keep the handle with its file
100 * Internal structure representing offsets to use at a sub-buffer switch.
102 struct switch_offsets
{
103 unsigned long begin
, end
, old
;
104 size_t pre_header_padding
, size
;
105 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
109 DEFINE_URCU_TLS(unsigned int, lib_ring_buffer_nesting
);
112 * wakeup_fd_mutex protects wakeup fd use by timer from concurrent
115 static pthread_mutex_t wakeup_fd_mutex
= PTHREAD_MUTEX_INITIALIZER
;
118 void lib_ring_buffer_print_errors(struct lttng_ust_lib_ring_buffer_channel
*chan
,
119 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
120 struct lttng_ust_shm_handle
*handle
);
123 * Handle timer teardown race wrt memory free of private data by
124 * ring buffer signals are handled by a single thread, which permits
125 * a synchronization point between handling of each signal.
126 * Protected by the lock within the structure.
128 struct timer_signal_data
{
129 pthread_t tid
; /* thread id managing signals */
132 pthread_mutex_t lock
;
135 static struct timer_signal_data timer_signal
= {
139 .lock
= PTHREAD_MUTEX_INITIALIZER
,
142 static bool lttng_ust_allow_blocking
;
144 void lttng_ust_ringbuffer_set_allow_blocking(void)
146 lttng_ust_allow_blocking
= true;
149 /* Get blocking timeout, in ms */
150 static int lttng_ust_ringbuffer_get_timeout(struct lttng_ust_lib_ring_buffer_channel
*chan
)
152 if (!lttng_ust_allow_blocking
)
154 return chan
->u
.s
.blocking_timeout_ms
;
158 * lib_ring_buffer_reset - Reset ring buffer to initial values.
161 * Effectively empty the ring buffer. Should be called when the buffer is not
162 * used for writing. The ring buffer can be opened for reading, but the reader
163 * should not be using the iterator concurrently with reset. The previous
164 * current iterator record is reset.
166 void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer
*buf
,
167 struct lttng_ust_shm_handle
*handle
)
169 struct lttng_ust_lib_ring_buffer_channel
*chan
;
170 const struct lttng_ust_lib_ring_buffer_config
*config
;
173 chan
= shmp(handle
, buf
->backend
.chan
);
176 config
= &chan
->backend
.config
;
178 * Reset iterator first. It will put the subbuffer if it currently holds
181 v_set(config
, &buf
->offset
, 0);
182 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
183 struct commit_counters_hot
*cc_hot
;
184 struct commit_counters_cold
*cc_cold
;
187 cc_hot
= shmp_index(handle
, buf
->commit_hot
, i
);
190 cc_cold
= shmp_index(handle
, buf
->commit_cold
, i
);
193 ts_end
= shmp_index(handle
, buf
->ts_end
, i
);
196 v_set(config
, &cc_hot
->cc
, 0);
197 v_set(config
, &cc_hot
->seq
, 0);
198 v_set(config
, &cc_cold
->cc_sb
, 0);
201 uatomic_set(&buf
->consumed
, 0);
202 uatomic_set(&buf
->record_disabled
, 0);
203 v_set(config
, &buf
->last_tsc
, 0);
204 lib_ring_buffer_backend_reset(&buf
->backend
, handle
);
205 /* Don't reset number of active readers */
206 v_set(config
, &buf
->records_lost_full
, 0);
207 v_set(config
, &buf
->records_lost_wrap
, 0);
208 v_set(config
, &buf
->records_lost_big
, 0);
209 v_set(config
, &buf
->records_count
, 0);
210 v_set(config
, &buf
->records_overrun
, 0);
215 * channel_reset - Reset channel to initial values.
218 * Effectively empty the channel. Should be called when the channel is not used
219 * for writing. The channel can be opened for reading, but the reader should not
220 * be using the iterator concurrently with reset. The previous current iterator
223 void channel_reset(struct lttng_ust_lib_ring_buffer_channel
*chan
)
226 * Reset iterators first. Will put the subbuffer if held for reading.
228 uatomic_set(&chan
->record_disabled
, 0);
229 /* Don't reset commit_count_mask, still valid */
230 channel_backend_reset(&chan
->backend
);
231 /* Don't reset switch/read timer interval */
232 /* Don't reset notifiers and notifier enable bits */
233 /* Don't reset reader reference count */
237 void init_crash_abi(const struct lttng_ust_lib_ring_buffer_config
*config
,
238 struct lttng_crash_abi
*crash_abi
,
239 struct lttng_ust_lib_ring_buffer
*buf
,
240 struct channel_backend
*chanb
,
241 struct shm_object
*shmobj
,
242 struct lttng_ust_shm_handle
*handle
)
246 for (i
= 0; i
< RB_CRASH_DUMP_ABI_MAGIC_LEN
; i
++)
247 crash_abi
->magic
[i
] = lttng_crash_magic_xor
[i
] ^ 0xFF;
248 crash_abi
->mmap_length
= shmobj
->memory_map_size
;
249 crash_abi
->endian
= RB_CRASH_ENDIAN
;
250 crash_abi
->major
= RB_CRASH_DUMP_ABI_MAJOR
;
251 crash_abi
->minor
= RB_CRASH_DUMP_ABI_MINOR
;
252 crash_abi
->word_size
= sizeof(unsigned long);
253 crash_abi
->layout_type
= LTTNG_CRASH_TYPE_UST
;
255 /* Offset of fields */
256 crash_abi
->offset
.prod_offset
=
257 (uint32_t) ((char *) &buf
->offset
- (char *) buf
);
258 crash_abi
->offset
.consumed_offset
=
259 (uint32_t) ((char *) &buf
->consumed
- (char *) buf
);
260 crash_abi
->offset
.commit_hot_array
=
261 (uint32_t) ((char *) shmp(handle
, buf
->commit_hot
) - (char *) buf
);
262 crash_abi
->offset
.commit_hot_seq
=
263 offsetof(struct commit_counters_hot
, seq
);
264 crash_abi
->offset
.buf_wsb_array
=
265 (uint32_t) ((char *) shmp(handle
, buf
->backend
.buf_wsb
) - (char *) buf
);
266 crash_abi
->offset
.buf_wsb_id
=
267 offsetof(struct lttng_ust_lib_ring_buffer_backend_subbuffer
, id
);
268 crash_abi
->offset
.sb_array
=
269 (uint32_t) ((char *) shmp(handle
, buf
->backend
.array
) - (char *) buf
);
270 crash_abi
->offset
.sb_array_shmp_offset
=
271 offsetof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp
,
273 crash_abi
->offset
.sb_backend_p_offset
=
274 offsetof(struct lttng_ust_lib_ring_buffer_backend_pages
,
278 crash_abi
->length
.prod_offset
= sizeof(buf
->offset
);
279 crash_abi
->length
.consumed_offset
= sizeof(buf
->consumed
);
280 crash_abi
->length
.commit_hot_seq
=
281 sizeof(((struct commit_counters_hot
*) NULL
)->seq
);
282 crash_abi
->length
.buf_wsb_id
=
283 sizeof(((struct lttng_ust_lib_ring_buffer_backend_subbuffer
*) NULL
)->id
);
284 crash_abi
->length
.sb_array_shmp_offset
=
285 sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*) NULL
)->shmp
._ref
.offset
);
286 crash_abi
->length
.sb_backend_p_offset
=
287 sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages
*) NULL
)->p
._ref
.offset
);
290 crash_abi
->stride
.commit_hot_array
=
291 sizeof(struct commit_counters_hot
);
292 crash_abi
->stride
.buf_wsb_array
=
293 sizeof(struct lttng_ust_lib_ring_buffer_backend_subbuffer
);
294 crash_abi
->stride
.sb_array
=
295 sizeof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp
);
297 /* Buffer constants */
298 crash_abi
->buf_size
= chanb
->buf_size
;
299 crash_abi
->subbuf_size
= chanb
->subbuf_size
;
300 crash_abi
->num_subbuf
= chanb
->num_subbuf
;
301 crash_abi
->mode
= (uint32_t) chanb
->config
.mode
;
303 if (config
->cb
.content_size_field
) {
304 size_t offset
, length
;
306 config
->cb
.content_size_field(config
, &offset
, &length
);
307 crash_abi
->offset
.content_size
= offset
;
308 crash_abi
->length
.content_size
= length
;
310 crash_abi
->offset
.content_size
= 0;
311 crash_abi
->length
.content_size
= 0;
313 if (config
->cb
.packet_size_field
) {
314 size_t offset
, length
;
316 config
->cb
.packet_size_field(config
, &offset
, &length
);
317 crash_abi
->offset
.packet_size
= offset
;
318 crash_abi
->length
.packet_size
= length
;
320 crash_abi
->offset
.packet_size
= 0;
321 crash_abi
->length
.packet_size
= 0;
326 * Must be called under cpu hotplug protection.
328 int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer
*buf
,
329 struct channel_backend
*chanb
, int cpu
,
330 struct lttng_ust_shm_handle
*handle
,
331 struct shm_object
*shmobj
)
333 const struct lttng_ust_lib_ring_buffer_config
*config
= &chanb
->config
;
334 struct lttng_ust_lib_ring_buffer_channel
*chan
= caa_container_of(chanb
,
335 struct lttng_ust_lib_ring_buffer_channel
, backend
);
336 struct lttng_ust_lib_ring_buffer_backend_subbuffer
*wsb
;
337 struct lttng_ust_lib_ring_buffer_channel
*shmp_chan
;
338 struct commit_counters_hot
*cc_hot
;
339 void *priv
= channel_get_private(chan
);
340 size_t subbuf_header_size
;
344 /* Test for cpu hotplug */
345 if (buf
->backend
.allocated
)
348 align_shm(shmobj
, __alignof__(struct commit_counters_hot
));
349 set_shmp(buf
->commit_hot
,
351 sizeof(struct commit_counters_hot
) * chan
->backend
.num_subbuf
));
352 if (!shmp(handle
, buf
->commit_hot
)) {
356 align_shm(shmobj
, __alignof__(struct commit_counters_cold
));
357 set_shmp(buf
->commit_cold
,
359 sizeof(struct commit_counters_cold
) * chan
->backend
.num_subbuf
));
360 if (!shmp(handle
, buf
->commit_cold
)) {
365 align_shm(shmobj
, __alignof__(uint64_t));
366 set_shmp(buf
->ts_end
,
368 sizeof(uint64_t) * chan
->backend
.num_subbuf
));
369 if (!shmp(handle
, buf
->ts_end
)) {
371 goto free_commit_cold
;
375 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
,
376 cpu
, handle
, shmobj
);
382 * Write the subbuffer header for first subbuffer so we know the total
383 * duration of data gathering.
385 subbuf_header_size
= config
->cb
.subbuffer_header_size();
386 v_set(config
, &buf
->offset
, subbuf_header_size
);
387 wsb
= shmp_index(handle
, buf
->backend
.buf_wsb
, 0);
392 subbuffer_id_clear_noref(config
, &wsb
->id
);
393 shmp_chan
= shmp(handle
, buf
->backend
.chan
);
398 tsc
= config
->cb
.ring_buffer_clock_read(shmp_chan
);
399 config
->cb
.buffer_begin(buf
, tsc
, 0, handle
);
400 cc_hot
= shmp_index(handle
, buf
->commit_hot
, 0);
405 v_add(config
, subbuf_header_size
, &cc_hot
->cc
);
406 v_add(config
, subbuf_header_size
, &cc_hot
->seq
);
408 if (config
->cb
.buffer_create
) {
409 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
, handle
);
414 init_crash_abi(config
, &buf
->crash_abi
, buf
, chanb
, shmobj
, handle
);
416 buf
->backend
.allocated
= 1;
421 /* ts_end will be freed by shm teardown */
423 /* commit_cold will be freed by shm teardown */
425 /* commit_hot will be freed by shm teardown */
431 void lib_ring_buffer_channel_switch_timer(int sig
, siginfo_t
*si
, void *uc
)
433 const struct lttng_ust_lib_ring_buffer_config
*config
;
434 struct lttng_ust_shm_handle
*handle
;
435 struct lttng_ust_lib_ring_buffer_channel
*chan
;
438 assert(CMM_LOAD_SHARED(timer_signal
.tid
) == pthread_self());
440 chan
= si
->si_value
.sival_ptr
;
441 handle
= chan
->handle
;
442 config
= &chan
->backend
.config
;
444 DBG("Switch timer for channel %p\n", chan
);
447 * Only flush buffers periodically if readers are active.
449 pthread_mutex_lock(&wakeup_fd_mutex
);
450 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
451 for_each_possible_cpu(cpu
) {
452 struct lttng_ust_lib_ring_buffer
*buf
=
453 shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
457 if (uatomic_read(&buf
->active_readers
))
458 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
,
462 struct lttng_ust_lib_ring_buffer
*buf
=
463 shmp(handle
, chan
->backend
.buf
[0].shmp
);
467 if (uatomic_read(&buf
->active_readers
))
468 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
,
472 pthread_mutex_unlock(&wakeup_fd_mutex
);
477 int lib_ring_buffer_poll_deliver(const struct lttng_ust_lib_ring_buffer_config
*config
,
478 struct lttng_ust_lib_ring_buffer
*buf
,
479 struct lttng_ust_lib_ring_buffer_channel
*chan
,
480 struct lttng_ust_shm_handle
*handle
)
482 unsigned long consumed_old
, consumed_idx
, commit_count
, write_offset
;
483 struct commit_counters_cold
*cc_cold
;
485 consumed_old
= uatomic_read(&buf
->consumed
);
486 consumed_idx
= subbuf_index(consumed_old
, chan
);
487 cc_cold
= shmp_index(handle
, buf
->commit_cold
, consumed_idx
);
490 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
492 * No memory barrier here, since we are only interested
493 * in a statistically correct polling result. The next poll will
494 * get the data is we are racing. The mb() that ensures correct
495 * memory order is in get_subbuf.
497 write_offset
= v_read(config
, &buf
->offset
);
500 * Check that the subbuffer we are trying to consume has been
501 * already fully committed.
504 if (((commit_count
- chan
->backend
.subbuf_size
)
505 & chan
->commit_count_mask
)
506 - (buf_trunc(consumed_old
, chan
)
507 >> chan
->backend
.num_subbuf_order
)
512 * Check that we are not about to read the same subbuffer in
513 * which the writer head is.
515 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_old
, chan
)
523 void lib_ring_buffer_wakeup(struct lttng_ust_lib_ring_buffer
*buf
,
524 struct lttng_ust_shm_handle
*handle
)
526 int wakeup_fd
= shm_get_wakeup_fd(handle
, &buf
->self
._ref
);
527 sigset_t sigpipe_set
, pending_set
, old_set
;
528 int ret
, sigpipe_was_pending
= 0;
534 * Wake-up the other end by writing a null byte in the pipe
535 * (non-blocking). Important note: Because writing into the
536 * pipe is non-blocking (and therefore we allow dropping wakeup
537 * data, as long as there is wakeup data present in the pipe
538 * buffer to wake up the consumer), the consumer should perform
539 * the following sequence for waiting:
540 * 1) empty the pipe (reads).
541 * 2) check if there is data in the buffer.
542 * 3) wait on the pipe (poll).
544 * Discard the SIGPIPE from write(), not disturbing any SIGPIPE
545 * that might be already pending. If a bogus SIGPIPE is sent to
546 * the entire process concurrently by a malicious user, it may
547 * be simply discarded.
549 ret
= sigemptyset(&pending_set
);
552 * sigpending returns the mask of signals that are _both_
553 * blocked for the thread _and_ pending for either the thread or
554 * the entire process.
556 ret
= sigpending(&pending_set
);
558 sigpipe_was_pending
= sigismember(&pending_set
, SIGPIPE
);
560 * If sigpipe was pending, it means it was already blocked, so
561 * no need to block it.
563 if (!sigpipe_was_pending
) {
564 ret
= sigemptyset(&sigpipe_set
);
566 ret
= sigaddset(&sigpipe_set
, SIGPIPE
);
568 ret
= pthread_sigmask(SIG_BLOCK
, &sigpipe_set
, &old_set
);
572 ret
= write(wakeup_fd
, "", 1);
573 } while (ret
== -1L && errno
== EINTR
);
574 if (ret
== -1L && errno
== EPIPE
&& !sigpipe_was_pending
) {
575 struct timespec timeout
= { 0, 0 };
577 ret
= sigtimedwait(&sigpipe_set
, NULL
,
579 } while (ret
== -1L && errno
== EINTR
);
581 if (!sigpipe_was_pending
) {
582 ret
= pthread_sigmask(SIG_SETMASK
, &old_set
, NULL
);
588 void lib_ring_buffer_channel_do_read(struct lttng_ust_lib_ring_buffer_channel
*chan
)
590 const struct lttng_ust_lib_ring_buffer_config
*config
;
591 struct lttng_ust_shm_handle
*handle
;
594 handle
= chan
->handle
;
595 config
= &chan
->backend
.config
;
598 * Only flush buffers periodically if readers are active.
600 pthread_mutex_lock(&wakeup_fd_mutex
);
601 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
602 for_each_possible_cpu(cpu
) {
603 struct lttng_ust_lib_ring_buffer
*buf
=
604 shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
608 if (uatomic_read(&buf
->active_readers
)
609 && lib_ring_buffer_poll_deliver(config
, buf
,
611 lib_ring_buffer_wakeup(buf
, handle
);
615 struct lttng_ust_lib_ring_buffer
*buf
=
616 shmp(handle
, chan
->backend
.buf
[0].shmp
);
620 if (uatomic_read(&buf
->active_readers
)
621 && lib_ring_buffer_poll_deliver(config
, buf
,
623 lib_ring_buffer_wakeup(buf
, handle
);
627 pthread_mutex_unlock(&wakeup_fd_mutex
);
631 void lib_ring_buffer_channel_read_timer(int sig
, siginfo_t
*si
, void *uc
)
633 struct lttng_ust_lib_ring_buffer_channel
*chan
;
635 assert(CMM_LOAD_SHARED(timer_signal
.tid
) == pthread_self());
636 chan
= si
->si_value
.sival_ptr
;
637 DBG("Read timer for channel %p\n", chan
);
638 lib_ring_buffer_channel_do_read(chan
);
643 void rb_setmask(sigset_t
*mask
)
647 ret
= sigemptyset(mask
);
649 PERROR("sigemptyset");
651 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_FLUSH
);
655 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_READ
);
659 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_TEARDOWN
);
666 void *sig_thread(void *arg
)
672 /* Only self thread will receive signal mask. */
674 CMM_STORE_SHARED(timer_signal
.tid
, pthread_self());
677 signr
= sigwaitinfo(&mask
, &info
);
680 PERROR("sigwaitinfo");
683 if (signr
== LTTNG_UST_RB_SIG_FLUSH
) {
684 lib_ring_buffer_channel_switch_timer(info
.si_signo
,
686 } else if (signr
== LTTNG_UST_RB_SIG_READ
) {
687 lib_ring_buffer_channel_read_timer(info
.si_signo
,
689 } else if (signr
== LTTNG_UST_RB_SIG_TEARDOWN
) {
691 CMM_STORE_SHARED(timer_signal
.qs_done
, 1);
694 ERR("Unexptected signal %d\n", info
.si_signo
);
701 * Ensure only a single thread listens on the timer signal.
704 void lib_ring_buffer_setup_timer_thread(void)
709 pthread_mutex_lock(&timer_signal
.lock
);
710 if (timer_signal
.setup_done
)
713 ret
= pthread_create(&thread
, NULL
, &sig_thread
, NULL
);
716 PERROR("pthread_create");
718 ret
= pthread_detach(thread
);
721 PERROR("pthread_detach");
723 timer_signal
.setup_done
= 1;
725 pthread_mutex_unlock(&timer_signal
.lock
);
729 * Wait for signal-handling thread quiescent state.
732 void lib_ring_buffer_wait_signal_thread_qs(unsigned int signr
)
734 sigset_t pending_set
;
738 * We need to be the only thread interacting with the thread
739 * that manages signals for teardown synchronization.
741 pthread_mutex_lock(&timer_signal
.lock
);
744 * Ensure we don't have any signal queued for this channel.
747 ret
= sigemptyset(&pending_set
);
749 PERROR("sigemptyset");
751 ret
= sigpending(&pending_set
);
753 PERROR("sigpending");
755 if (!sigismember(&pending_set
, signr
))
761 * From this point, no new signal handler will be fired that
762 * would try to access "chan". However, we still need to wait
763 * for any currently executing handler to complete.
766 CMM_STORE_SHARED(timer_signal
.qs_done
, 0);
770 * Kill with LTTNG_UST_RB_SIG_TEARDOWN, so signal management
773 kill(getpid(), LTTNG_UST_RB_SIG_TEARDOWN
);
775 while (!CMM_LOAD_SHARED(timer_signal
.qs_done
))
779 pthread_mutex_unlock(&timer_signal
.lock
);
783 void lib_ring_buffer_channel_switch_timer_start(struct lttng_ust_lib_ring_buffer_channel
*chan
)
786 struct itimerspec its
;
789 if (!chan
->switch_timer_interval
|| chan
->switch_timer_enabled
)
792 chan
->switch_timer_enabled
= 1;
794 lib_ring_buffer_setup_timer_thread();
796 memset(&sev
, 0, sizeof(sev
));
797 sev
.sigev_notify
= SIGEV_SIGNAL
;
798 sev
.sigev_signo
= LTTNG_UST_RB_SIG_FLUSH
;
799 sev
.sigev_value
.sival_ptr
= chan
;
800 ret
= timer_create(CLOCKID
, &sev
, &chan
->switch_timer
);
802 PERROR("timer_create");
805 its
.it_value
.tv_sec
= chan
->switch_timer_interval
/ 1000000;
806 its
.it_value
.tv_nsec
= (chan
->switch_timer_interval
% 1000000) * 1000;
807 its
.it_interval
.tv_sec
= its
.it_value
.tv_sec
;
808 its
.it_interval
.tv_nsec
= its
.it_value
.tv_nsec
;
810 ret
= timer_settime(chan
->switch_timer
, 0, &its
, NULL
);
812 PERROR("timer_settime");
817 void lib_ring_buffer_channel_switch_timer_stop(struct lttng_ust_lib_ring_buffer_channel
*chan
)
821 if (!chan
->switch_timer_interval
|| !chan
->switch_timer_enabled
)
824 ret
= timer_delete(chan
->switch_timer
);
826 PERROR("timer_delete");
829 lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_FLUSH
);
831 chan
->switch_timer
= 0;
832 chan
->switch_timer_enabled
= 0;
836 void lib_ring_buffer_channel_read_timer_start(struct lttng_ust_lib_ring_buffer_channel
*chan
)
838 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
840 struct itimerspec its
;
843 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
844 || !chan
->read_timer_interval
|| chan
->read_timer_enabled
)
847 chan
->read_timer_enabled
= 1;
849 lib_ring_buffer_setup_timer_thread();
851 sev
.sigev_notify
= SIGEV_SIGNAL
;
852 sev
.sigev_signo
= LTTNG_UST_RB_SIG_READ
;
853 sev
.sigev_value
.sival_ptr
= chan
;
854 ret
= timer_create(CLOCKID
, &sev
, &chan
->read_timer
);
856 PERROR("timer_create");
859 its
.it_value
.tv_sec
= chan
->read_timer_interval
/ 1000000;
860 its
.it_value
.tv_nsec
= (chan
->read_timer_interval
% 1000000) * 1000;
861 its
.it_interval
.tv_sec
= its
.it_value
.tv_sec
;
862 its
.it_interval
.tv_nsec
= its
.it_value
.tv_nsec
;
864 ret
= timer_settime(chan
->read_timer
, 0, &its
, NULL
);
866 PERROR("timer_settime");
871 void lib_ring_buffer_channel_read_timer_stop(struct lttng_ust_lib_ring_buffer_channel
*chan
)
873 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
876 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
877 || !chan
->read_timer_interval
|| !chan
->read_timer_enabled
)
880 ret
= timer_delete(chan
->read_timer
);
882 PERROR("timer_delete");
886 * do one more check to catch data that has been written in the last
889 lib_ring_buffer_channel_do_read(chan
);
891 lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_READ
);
893 chan
->read_timer
= 0;
894 chan
->read_timer_enabled
= 0;
897 static void channel_unregister_notifiers(struct lttng_ust_lib_ring_buffer_channel
*chan
,
898 struct lttng_ust_shm_handle
*handle
)
900 lib_ring_buffer_channel_switch_timer_stop(chan
);
901 lib_ring_buffer_channel_read_timer_stop(chan
);
904 static void channel_print_errors(struct lttng_ust_lib_ring_buffer_channel
*chan
,
905 struct lttng_ust_shm_handle
*handle
)
907 const struct lttng_ust_lib_ring_buffer_config
*config
=
908 &chan
->backend
.config
;
911 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
912 for_each_possible_cpu(cpu
) {
913 struct lttng_ust_lib_ring_buffer
*buf
=
914 shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
916 lib_ring_buffer_print_errors(chan
, buf
, cpu
, handle
);
919 struct lttng_ust_lib_ring_buffer
*buf
=
920 shmp(handle
, chan
->backend
.buf
[0].shmp
);
923 lib_ring_buffer_print_errors(chan
, buf
, -1, handle
);
927 static void channel_free(struct lttng_ust_lib_ring_buffer_channel
*chan
,
928 struct lttng_ust_shm_handle
*handle
,
931 channel_backend_free(&chan
->backend
, handle
);
932 /* chan is freed by shm teardown */
933 shm_object_table_destroy(handle
->table
, consumer
);
938 * channel_create - Create channel.
939 * @config: ring buffer instance configuration
940 * @name: name of the channel
941 * @priv_data: ring buffer client private data area pointer (output)
942 * @priv_data_size: length, in bytes, of the private data area.
943 * @priv_data_init: initialization data for private data.
944 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
945 * address mapping. It is used only by RING_BUFFER_STATIC
946 * configuration. It can be set to NULL for other backends.
947 * @subbuf_size: subbuffer size
948 * @num_subbuf: number of subbuffers
949 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
950 * padding to let readers get those sub-buffers.
951 * Used for live streaming.
952 * @read_timer_interval: Time interval (in us) to wake up pending readers.
953 * @stream_fds: array of stream file descriptors.
954 * @nr_stream_fds: number of file descriptors in array.
957 * Returns NULL on failure.
959 struct lttng_ust_shm_handle
*channel_create(const struct lttng_ust_lib_ring_buffer_config
*config
,
962 size_t priv_data_align
,
963 size_t priv_data_size
,
964 void *priv_data_init
,
965 void *buf_addr
, size_t subbuf_size
,
966 size_t num_subbuf
, unsigned int switch_timer_interval
,
967 unsigned int read_timer_interval
,
968 const int *stream_fds
, int nr_stream_fds
,
969 int64_t blocking_timeout
)
972 size_t shmsize
, chansize
;
973 struct lttng_ust_lib_ring_buffer_channel
*chan
;
974 struct lttng_ust_shm_handle
*handle
;
975 struct shm_object
*shmobj
;
976 unsigned int nr_streams
;
977 int64_t blocking_timeout_ms
;
979 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
980 nr_streams
= num_possible_cpus();
984 if (nr_stream_fds
!= nr_streams
)
987 if (blocking_timeout
< -1) {
991 if (blocking_timeout
== -1) {
992 blocking_timeout_ms
= -1;
994 blocking_timeout_ms
= blocking_timeout
/ 1000;
995 if (blocking_timeout_ms
!= (int32_t) blocking_timeout_ms
) {
1000 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
1001 read_timer_interval
))
1004 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
1008 /* Allocate table for channel + per-cpu buffers */
1009 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
1011 goto error_table_alloc
;
1013 /* Calculate the shm allocation layout */
1014 shmsize
= sizeof(struct lttng_ust_lib_ring_buffer_channel
);
1015 shmsize
+= lttng_ust_offset_align(shmsize
, __alignof__(struct lttng_ust_lib_ring_buffer_shmp
));
1016 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
) * nr_streams
;
1018 if (priv_data_align
)
1019 shmsize
+= lttng_ust_offset_align(shmsize
, priv_data_align
);
1020 shmsize
+= priv_data_size
;
1022 /* Allocate normal memory for channel (not shared) */
1023 shmobj
= shm_object_table_alloc(handle
->table
, shmsize
, SHM_OBJECT_MEM
,
1027 /* struct lttng_ust_lib_ring_buffer_channel is at object 0, offset 0 (hardcoded) */
1028 set_shmp(handle
->chan
, zalloc_shm(shmobj
, chansize
));
1029 assert(handle
->chan
._ref
.index
== 0);
1030 assert(handle
->chan
._ref
.offset
== 0);
1031 chan
= shmp(handle
, handle
->chan
);
1034 chan
->nr_streams
= nr_streams
;
1036 /* space for private data */
1037 if (priv_data_size
) {
1038 DECLARE_SHMP(void, priv_data_alloc
);
1040 align_shm(shmobj
, priv_data_align
);
1041 chan
->priv_data_offset
= shmobj
->allocated_len
;
1042 set_shmp(priv_data_alloc
, zalloc_shm(shmobj
, priv_data_size
));
1043 if (!shmp(handle
, priv_data_alloc
))
1045 *priv_data
= channel_get_private(chan
);
1046 memcpy(*priv_data
, priv_data_init
, priv_data_size
);
1048 chan
->priv_data_offset
= -1;
1053 chan
->u
.s
.blocking_timeout_ms
= (int32_t) blocking_timeout_ms
;
1055 ret
= channel_backend_init(&chan
->backend
, name
, config
,
1056 subbuf_size
, num_subbuf
, handle
,
1059 goto error_backend_init
;
1061 chan
->handle
= handle
;
1062 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
1064 chan
->switch_timer_interval
= switch_timer_interval
;
1065 chan
->read_timer_interval
= read_timer_interval
;
1066 lib_ring_buffer_channel_switch_timer_start(chan
);
1067 lib_ring_buffer_channel_read_timer_start(chan
);
1073 shm_object_table_destroy(handle
->table
, 1);
1079 struct lttng_ust_shm_handle
*channel_handle_create(void *data
,
1080 uint64_t memory_map_size
,
1083 struct lttng_ust_shm_handle
*handle
;
1084 struct shm_object
*object
;
1086 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
1090 /* Allocate table for channel + per-cpu buffers */
1091 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
1093 goto error_table_alloc
;
1094 /* Add channel object */
1095 object
= shm_object_table_append_mem(handle
->table
, data
,
1096 memory_map_size
, wakeup_fd
);
1098 goto error_table_object
;
1099 /* struct lttng_ust_lib_ring_buffer_channel is at object 0, offset 0 (hardcoded) */
1100 handle
->chan
._ref
.index
= 0;
1101 handle
->chan
._ref
.offset
= 0;
1105 shm_object_table_destroy(handle
->table
, 0);
1111 int channel_handle_add_stream(struct lttng_ust_shm_handle
*handle
,
1112 int shm_fd
, int wakeup_fd
, uint32_t stream_nr
,
1113 uint64_t memory_map_size
)
1115 struct shm_object
*object
;
1117 /* Add stream object */
1118 object
= shm_object_table_append_shm(handle
->table
,
1119 shm_fd
, wakeup_fd
, stream_nr
,
1126 unsigned int channel_handle_get_nr_streams(struct lttng_ust_shm_handle
*handle
)
1128 assert(handle
->table
);
1129 return handle
->table
->allocated_len
- 1;
1133 void channel_release(struct lttng_ust_lib_ring_buffer_channel
*chan
, struct lttng_ust_shm_handle
*handle
,
1136 channel_free(chan
, handle
, consumer
);
1140 * channel_destroy - Finalize, wait for q.s. and destroy channel.
1141 * @chan: channel to destroy
1143 * Holds cpu hotplug.
1144 * Call "destroy" callback, finalize channels, decrement the channel
1145 * reference count. Note that when readers have completed data
1146 * consumption of finalized channels, get_subbuf() will return -ENODATA.
1147 * They should release their handle at that point.
1149 void channel_destroy(struct lttng_ust_lib_ring_buffer_channel
*chan
, struct lttng_ust_shm_handle
*handle
,
1154 * Note: the consumer takes care of finalizing and
1155 * switching the buffers.
1157 channel_unregister_notifiers(chan
, handle
);
1159 * The consumer prints errors.
1161 channel_print_errors(chan
, handle
);
1165 * sessiond/consumer are keeping a reference on the shm file
1166 * descriptor directly. No need to refcount.
1168 channel_release(chan
, handle
, consumer
);
1172 struct lttng_ust_lib_ring_buffer
*channel_get_ring_buffer(
1173 const struct lttng_ust_lib_ring_buffer_config
*config
,
1174 struct lttng_ust_lib_ring_buffer_channel
*chan
, int cpu
,
1175 struct lttng_ust_shm_handle
*handle
,
1176 int *shm_fd
, int *wait_fd
,
1178 uint64_t *memory_map_size
)
1180 struct shm_ref
*ref
;
1182 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1185 if (cpu
>= num_possible_cpus())
1188 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1189 *shm_fd
= shm_get_shm_fd(handle
, ref
);
1190 *wait_fd
= shm_get_wait_fd(handle
, ref
);
1191 *wakeup_fd
= shm_get_wakeup_fd(handle
, ref
);
1192 if (shm_get_shm_size(handle
, ref
, memory_map_size
))
1194 return shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
1197 int ring_buffer_channel_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1198 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1199 struct lttng_ust_shm_handle
*handle
)
1201 struct shm_ref
*ref
;
1203 ref
= &handle
->chan
._ref
;
1204 return shm_close_wait_fd(handle
, ref
);
1207 int ring_buffer_channel_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1208 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1209 struct lttng_ust_shm_handle
*handle
)
1211 struct shm_ref
*ref
;
1213 ref
= &handle
->chan
._ref
;
1214 return shm_close_wakeup_fd(handle
, ref
);
1217 int ring_buffer_stream_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1218 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1219 struct lttng_ust_shm_handle
*handle
,
1222 struct shm_ref
*ref
;
1224 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1227 if (cpu
>= num_possible_cpus())
1230 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1231 return shm_close_wait_fd(handle
, ref
);
1234 int ring_buffer_stream_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1235 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1236 struct lttng_ust_shm_handle
*handle
,
1239 struct shm_ref
*ref
;
1242 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1245 if (cpu
>= num_possible_cpus())
1248 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1249 pthread_mutex_lock(&wakeup_fd_mutex
);
1250 ret
= shm_close_wakeup_fd(handle
, ref
);
1251 pthread_mutex_unlock(&wakeup_fd_mutex
);
1255 int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer
*buf
,
1256 struct lttng_ust_shm_handle
*handle
)
1258 if (uatomic_cmpxchg(&buf
->active_readers
, 0, 1) != 0)
1264 void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer
*buf
,
1265 struct lttng_ust_shm_handle
*handle
)
1267 struct lttng_ust_lib_ring_buffer_channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1271 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1273 uatomic_dec(&buf
->active_readers
);
1277 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
1279 * @consumed: consumed count indicating the position where to read
1280 * @produced: produced count, indicates position when to stop reading
1282 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1283 * data to read at consumed position, or 0 if the get operation succeeds.
1286 int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer
*buf
,
1287 unsigned long *consumed
, unsigned long *produced
,
1288 struct lttng_ust_shm_handle
*handle
)
1290 struct lttng_ust_lib_ring_buffer_channel
*chan
;
1291 const struct lttng_ust_lib_ring_buffer_config
*config
;
1292 unsigned long consumed_cur
, write_offset
;
1295 chan
= shmp(handle
, buf
->backend
.chan
);
1298 config
= &chan
->backend
.config
;
1299 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
1301 * Read finalized before counters.
1304 consumed_cur
= uatomic_read(&buf
->consumed
);
1306 * No need to issue a memory barrier between consumed count read and
1307 * write offset read, because consumed count can only change
1308 * concurrently in overwrite mode, and we keep a sequence counter
1309 * identifier derived from the write offset to check we are getting
1310 * the same sub-buffer we are expecting (the sub-buffers are atomically
1311 * "tagged" upon writes, tags are checked upon read).
1313 write_offset
= v_read(config
, &buf
->offset
);
1316 * Check that we are not about to read the same subbuffer in
1317 * which the writer head is.
1319 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
1323 *consumed
= consumed_cur
;
1324 *produced
= subbuf_trunc(write_offset
, chan
);
1330 * The memory barriers __wait_event()/wake_up_interruptible() take care
1331 * of "raw_spin_is_locked" memory ordering.
1340 * Performs the same function as lib_ring_buffer_snapshot(), but the positions
1341 * are saved regardless of whether the consumed and produced positions are
1342 * in the same subbuffer.
1344 * @consumed: consumed byte count indicating the last position read
1345 * @produced: produced byte count indicating the last position written
1347 * This function is meant to provide information on the exact producer and
1348 * consumer positions without regard for the "snapshot" feature.
1350 int lib_ring_buffer_snapshot_sample_positions(
1351 struct lttng_ust_lib_ring_buffer
*buf
,
1352 unsigned long *consumed
, unsigned long *produced
,
1353 struct lttng_ust_shm_handle
*handle
)
1355 struct lttng_ust_lib_ring_buffer_channel
*chan
;
1356 const struct lttng_ust_lib_ring_buffer_config
*config
;
1358 chan
= shmp(handle
, buf
->backend
.chan
);
1361 config
= &chan
->backend
.config
;
1363 *consumed
= uatomic_read(&buf
->consumed
);
1365 * No need to issue a memory barrier between consumed count read and
1366 * write offset read, because consumed count can only change
1367 * concurrently in overwrite mode, and we keep a sequence counter
1368 * identifier derived from the write offset to check we are getting
1369 * the same sub-buffer we are expecting (the sub-buffers are atomically
1370 * "tagged" upon writes, tags are checked upon read).
1372 *produced
= v_read(config
, &buf
->offset
);
1377 * lib_ring_buffer_move_consumer - move consumed counter forward
1379 * @consumed_new: new consumed count value
1381 void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer
*buf
,
1382 unsigned long consumed_new
,
1383 struct lttng_ust_shm_handle
*handle
)
1385 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1386 struct lttng_ust_lib_ring_buffer_channel
*chan
;
1387 unsigned long consumed
;
1389 chan
= shmp(handle
, bufb
->chan
);
1392 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1395 * Only push the consumed value forward.
1396 * If the consumed cmpxchg fails, this is because we have been pushed by
1397 * the writer in flight recorder mode.
1399 consumed
= uatomic_read(&buf
->consumed
);
1400 while ((long) consumed
- (long) consumed_new
< 0)
1401 consumed
= uatomic_cmpxchg(&buf
->consumed
, consumed
,
1406 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
1408 * @consumed: consumed count indicating the position where to read
1410 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1411 * data to read at consumed position, or 0 if the get operation succeeds.
1413 int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
1414 unsigned long consumed
,
1415 struct lttng_ust_shm_handle
*handle
)
1417 struct lttng_ust_lib_ring_buffer_channel
*chan
;
1418 const struct lttng_ust_lib_ring_buffer_config
*config
;
1419 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
1420 int ret
, finalized
, nr_retry
= LTTNG_UST_RING_BUFFER_GET_RETRY
;
1421 struct commit_counters_cold
*cc_cold
;
1423 chan
= shmp(handle
, buf
->backend
.chan
);
1426 config
= &chan
->backend
.config
;
1428 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
1430 * Read finalized before counters.
1433 consumed_cur
= uatomic_read(&buf
->consumed
);
1434 consumed_idx
= subbuf_index(consumed
, chan
);
1435 cc_cold
= shmp_index(handle
, buf
->commit_cold
, consumed_idx
);
1438 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
1440 * Make sure we read the commit count before reading the buffer
1441 * data and the write offset. Correct consumed offset ordering
1442 * wrt commit count is insured by the use of cmpxchg to update
1443 * the consumed offset.
1446 * Local rmb to match the remote wmb to read the commit count
1447 * before the buffer data and the write offset.
1451 write_offset
= v_read(config
, &buf
->offset
);
1454 * Check that the buffer we are getting is after or at consumed_cur
1457 if ((long) subbuf_trunc(consumed
, chan
)
1458 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
1462 * Check that the subbuffer we are trying to consume has been
1463 * already fully committed. There are a few causes that can make
1464 * this unavailability situation occur:
1466 * Temporary (short-term) situation:
1467 * - Application is running on a different CPU, between reserve
1468 * and commit ring buffer operations,
1469 * - Application is preempted between reserve and commit ring
1470 * buffer operations,
1472 * Long-term situation:
1473 * - Application is stopped (SIGSTOP) between reserve and commit
1474 * ring buffer operations. Could eventually be resumed by
1476 * - Application is killed (SIGTERM, SIGINT, SIGKILL) between
1477 * reserve and commit ring buffer operation.
1479 * From a consumer perspective, handling short-term
1480 * unavailability situations is performed by retrying a few
1481 * times after a delay. Handling long-term unavailability
1482 * situations is handled by failing to get the sub-buffer.
1484 * In all of those situations, if the application is taking a
1485 * long time to perform its commit after ring buffer space
1486 * reservation, we can end up in a situation where the producer
1487 * will fill the ring buffer and try to write into the same
1488 * sub-buffer again (which has a missing commit). This is
1489 * handled by the producer in the sub-buffer switch handling
1490 * code of the reserve routine by detecting unbalanced
1491 * reserve/commit counters and discarding all further events
1492 * until the situation is resolved in those situations. Two
1493 * scenarios can occur:
1495 * 1) The application causing the reserve/commit counters to be
1496 * unbalanced has been terminated. In this situation, all
1497 * further events will be discarded in the buffers, and no
1498 * further buffer data will be readable by the consumer
1499 * daemon. Tearing down the UST tracing session and starting
1500 * anew is a work-around for those situations. Note that this
1501 * only affects per-UID tracing. In per-PID tracing, the
1502 * application vanishes with the termination, and therefore
1503 * no more data needs to be written to the buffers.
1504 * 2) The application causing the unbalance has been delayed for
1505 * a long time, but will eventually try to increment the
1506 * commit counter after eventually writing to the sub-buffer.
1507 * This situation can cause events to be discarded until the
1508 * application resumes its operations.
1510 if (((commit_count
- chan
->backend
.subbuf_size
)
1511 & chan
->commit_count_mask
)
1512 - (buf_trunc(consumed
, chan
)
1513 >> chan
->backend
.num_subbuf_order
)
1515 if (nr_retry
-- > 0) {
1516 if (nr_retry
<= (LTTNG_UST_RING_BUFFER_GET_RETRY
>> 1))
1517 (void) poll(NULL
, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS
);
1525 * Check that we are not about to read the same subbuffer in
1526 * which the writer head is.
1528 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed
, chan
)
1533 * Failure to get the subbuffer causes a busy-loop retry without going
1534 * to a wait queue. These are caused by short-lived race windows where
1535 * the writer is getting access to a subbuffer we were trying to get
1536 * access to. Also checks that the "consumed" buffer count we are
1537 * looking for matches the one contained in the subbuffer id.
1539 * The short-lived race window described here can be affected by
1540 * application signals and preemption, thus requiring to bound
1541 * the loop to a maximum number of retry.
1543 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1544 consumed_idx
, buf_trunc_val(consumed
, chan
),
1547 if (nr_retry
-- > 0) {
1548 if (nr_retry
<= (LTTNG_UST_RING_BUFFER_GET_RETRY
>> 1))
1549 (void) poll(NULL
, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS
);
1555 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1557 buf
->get_subbuf_consumed
= consumed
;
1558 buf
->get_subbuf
= 1;
1564 * The memory barriers __wait_event()/wake_up_interruptible() take care
1565 * of "raw_spin_is_locked" memory ordering.
1574 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1577 void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
1578 struct lttng_ust_shm_handle
*handle
)
1580 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1581 struct lttng_ust_lib_ring_buffer_channel
*chan
;
1582 const struct lttng_ust_lib_ring_buffer_config
*config
;
1583 unsigned long sb_bindex
, consumed_idx
, consumed
;
1584 struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*rpages
;
1585 struct lttng_ust_lib_ring_buffer_backend_pages
*backend_pages
;
1587 chan
= shmp(handle
, bufb
->chan
);
1590 config
= &chan
->backend
.config
;
1591 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1593 if (!buf
->get_subbuf
) {
1595 * Reader puts a subbuffer it did not get.
1597 CHAN_WARN_ON(chan
, 1);
1600 consumed
= buf
->get_subbuf_consumed
;
1601 buf
->get_subbuf
= 0;
1604 * Clear the records_unread counter. (overruns counter)
1605 * Can still be non-zero if a file reader simply grabbed the data
1606 * without using iterators.
1607 * Can be below zero if an iterator is used on a snapshot more than
1610 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1611 rpages
= shmp_index(handle
, bufb
->array
, sb_bindex
);
1614 backend_pages
= shmp(handle
, rpages
->shmp
);
1617 v_add(config
, v_read(config
, &backend_pages
->records_unread
),
1618 &bufb
->records_read
);
1619 v_set(config
, &backend_pages
->records_unread
, 0);
1620 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1621 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1622 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1625 * Exchange the reader subbuffer with the one we put in its place in the
1626 * writer subbuffer table. Expect the original consumed count. If
1627 * update_read_sb_index fails, this is because the writer updated the
1628 * subbuffer concurrently. We should therefore keep the subbuffer we
1629 * currently have: it has become invalid to try reading this sub-buffer
1630 * consumed count value anyway.
1632 consumed_idx
= subbuf_index(consumed
, chan
);
1633 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1634 consumed_idx
, buf_trunc_val(consumed
, chan
),
1637 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1638 * if the writer concurrently updated it.
1643 * cons_offset is an iterator on all subbuffer offsets between the reader
1644 * position and the writer position. (inclusive)
1647 void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
1648 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1649 unsigned long cons_offset
,
1651 struct lttng_ust_shm_handle
*handle
)
1653 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1654 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1655 struct commit_counters_hot
*cc_hot
;
1656 struct commit_counters_cold
*cc_cold
;
1658 cons_idx
= subbuf_index(cons_offset
, chan
);
1659 cc_hot
= shmp_index(handle
, buf
->commit_hot
, cons_idx
);
1662 cc_cold
= shmp_index(handle
, buf
->commit_cold
, cons_idx
);
1665 commit_count
= v_read(config
, &cc_hot
->cc
);
1666 commit_count_sb
= v_read(config
, &cc_cold
->cc_sb
);
1668 if (subbuf_offset(commit_count
, chan
) != 0)
1669 DBG("ring buffer %s, cpu %d: "
1670 "commit count in subbuffer %lu,\n"
1671 "expecting multiples of %lu bytes\n"
1672 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1673 chan
->backend
.name
, cpu
, cons_idx
,
1674 chan
->backend
.subbuf_size
,
1675 commit_count
, commit_count_sb
);
1677 DBG("ring buffer: %s, cpu %d: %lu bytes committed\n",
1678 chan
->backend
.name
, cpu
, commit_count
);
1682 void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
1683 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1684 void *priv
, int cpu
,
1685 struct lttng_ust_shm_handle
*handle
)
1687 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1688 unsigned long write_offset
, cons_offset
;
1691 * No need to order commit_count, write_offset and cons_offset reads
1692 * because we execute at teardown when no more writer nor reader
1693 * references are left.
1695 write_offset
= v_read(config
, &buf
->offset
);
1696 cons_offset
= uatomic_read(&buf
->consumed
);
1697 if (write_offset
!= cons_offset
)
1698 DBG("ring buffer %s, cpu %d: "
1699 "non-consumed data\n"
1700 " [ %lu bytes written, %lu bytes read ]\n",
1701 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1703 for (cons_offset
= uatomic_read(&buf
->consumed
);
1704 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1707 cons_offset
= subbuf_align(cons_offset
, chan
))
1708 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1713 void lib_ring_buffer_print_errors(struct lttng_ust_lib_ring_buffer_channel
*chan
,
1714 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
1715 struct lttng_ust_shm_handle
*handle
)
1717 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1718 void *priv
= channel_get_private(chan
);
1720 if (!strcmp(chan
->backend
.name
, "relay-metadata-mmap")) {
1721 DBG("ring buffer %s: %lu records written, "
1722 "%lu records overrun\n",
1724 v_read(config
, &buf
->records_count
),
1725 v_read(config
, &buf
->records_overrun
));
1727 DBG("ring buffer %s, cpu %d: %lu records written, "
1728 "%lu records overrun\n",
1729 chan
->backend
.name
, cpu
,
1730 v_read(config
, &buf
->records_count
),
1731 v_read(config
, &buf
->records_overrun
));
1733 if (v_read(config
, &buf
->records_lost_full
)
1734 || v_read(config
, &buf
->records_lost_wrap
)
1735 || v_read(config
, &buf
->records_lost_big
))
1736 DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n"
1737 " [ %lu buffer full, %lu nest buffer wrap-around, "
1738 "%lu event too big ]\n",
1739 chan
->backend
.name
, cpu
,
1740 v_read(config
, &buf
->records_lost_full
),
1741 v_read(config
, &buf
->records_lost_wrap
),
1742 v_read(config
, &buf
->records_lost_big
));
1744 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
, handle
);
1748 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1750 * Only executed by SWITCH_FLUSH, which can be issued while tracing is
1751 * active or at buffer finalization (destroy).
1754 void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer
*buf
,
1755 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1756 struct switch_offsets
*offsets
,
1758 struct lttng_ust_shm_handle
*handle
)
1760 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1761 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1762 unsigned long commit_count
;
1763 struct commit_counters_hot
*cc_hot
;
1765 config
->cb
.buffer_begin(buf
, tsc
, oldidx
, handle
);
1768 * Order all writes to buffer before the commit count update that will
1769 * determine that the subbuffer is full.
1772 cc_hot
= shmp_index(handle
, buf
->commit_hot
, oldidx
);
1775 v_add(config
, config
->cb
.subbuffer_header_size(),
1777 commit_count
= v_read(config
, &cc_hot
->cc
);
1778 /* Check if the written buffer has to be delivered */
1779 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1780 commit_count
, oldidx
, handle
, tsc
);
1781 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1782 offsets
->old
+ config
->cb
.subbuffer_header_size(),
1783 commit_count
, handle
, cc_hot
);
1787 * lib_ring_buffer_switch_old_end: switch old subbuffer
1789 * Note : offset_old should never be 0 here. It is ok, because we never perform
1790 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1791 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1795 void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer
*buf
,
1796 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1797 struct switch_offsets
*offsets
,
1799 struct lttng_ust_shm_handle
*handle
)
1801 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1802 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1803 unsigned long commit_count
, padding_size
, data_size
;
1804 struct commit_counters_hot
*cc_hot
;
1807 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1808 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1809 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
,
1812 ts_end
= shmp_index(handle
, buf
->ts_end
, oldidx
);
1816 * This is the last space reservation in that sub-buffer before
1817 * it gets delivered. This provides exclusive access to write to
1818 * this sub-buffer's ts_end. There are also no concurrent
1819 * readers of that ts_end because delivery of that sub-buffer is
1820 * postponed until the commit counter is incremented for the
1821 * current space reservation.
1826 * Order all writes to buffer and store to ts_end before the commit
1827 * count update that will determine that the subbuffer is full.
1830 cc_hot
= shmp_index(handle
, buf
->commit_hot
, oldidx
);
1833 v_add(config
, padding_size
, &cc_hot
->cc
);
1834 commit_count
= v_read(config
, &cc_hot
->cc
);
1835 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1836 commit_count
, oldidx
, handle
, tsc
);
1837 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1838 offsets
->old
+ padding_size
, commit_count
, handle
,
1843 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1845 * This code can be executed unordered : writers may already have written to the
1846 * sub-buffer before this code gets executed, caution. The commit makes sure
1847 * that this code is executed before the deliver of this sub-buffer.
1850 void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer
*buf
,
1851 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1852 struct switch_offsets
*offsets
,
1854 struct lttng_ust_shm_handle
*handle
)
1856 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1857 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1858 unsigned long commit_count
;
1859 struct commit_counters_hot
*cc_hot
;
1861 config
->cb
.buffer_begin(buf
, tsc
, beginidx
, handle
);
1864 * Order all writes to buffer before the commit count update that will
1865 * determine that the subbuffer is full.
1868 cc_hot
= shmp_index(handle
, buf
->commit_hot
, beginidx
);
1871 v_add(config
, config
->cb
.subbuffer_header_size(), &cc_hot
->cc
);
1872 commit_count
= v_read(config
, &cc_hot
->cc
);
1873 /* Check if the written buffer has to be delivered */
1874 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1875 commit_count
, beginidx
, handle
, tsc
);
1876 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1877 offsets
->begin
+ config
->cb
.subbuffer_header_size(),
1878 commit_count
, handle
, cc_hot
);
1882 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1884 * Calls subbuffer_set_data_size() to set the data size of the current
1885 * sub-buffer. We do not need to perform check_deliver nor commit here,
1886 * since this task will be done by the "commit" of the event for which
1887 * we are currently doing the space reservation.
1890 void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer
*buf
,
1891 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1892 struct switch_offsets
*offsets
,
1894 struct lttng_ust_shm_handle
*handle
)
1896 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1897 unsigned long endidx
, data_size
;
1900 endidx
= subbuf_index(offsets
->end
- 1, chan
);
1901 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1902 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
,
1904 ts_end
= shmp_index(handle
, buf
->ts_end
, endidx
);
1908 * This is the last space reservation in that sub-buffer before
1909 * it gets delivered. This provides exclusive access to write to
1910 * this sub-buffer's ts_end. There are also no concurrent
1911 * readers of that ts_end because delivery of that sub-buffer is
1912 * postponed until the commit counter is incremented for the
1913 * current space reservation.
1921 * !0 if execution must be aborted.
1924 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1925 struct lttng_ust_lib_ring_buffer
*buf
,
1926 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1927 struct switch_offsets
*offsets
,
1929 struct lttng_ust_shm_handle
*handle
)
1931 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1932 unsigned long off
, reserve_commit_diff
;
1934 offsets
->begin
= v_read(config
, &buf
->offset
);
1935 offsets
->old
= offsets
->begin
;
1936 offsets
->switch_old_start
= 0;
1937 off
= subbuf_offset(offsets
->begin
, chan
);
1939 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1942 * Ensure we flush the header of an empty subbuffer when doing the
1943 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1944 * total data gathering duration even if there were no records saved
1945 * after the last buffer switch.
1946 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1947 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1948 * subbuffer header as appropriate.
1949 * The next record that reserves space will be responsible for
1950 * populating the following subbuffer header. We choose not to populate
1951 * the next subbuffer header here because we want to be able to use
1952 * SWITCH_ACTIVE for periodical buffer flush, which must
1953 * guarantee that all the buffer content (records and header
1954 * timestamps) are visible to the reader. This is required for
1955 * quiescence guarantees for the fusion merge.
1957 if (mode
!= SWITCH_FLUSH
&& !off
)
1958 return -1; /* we do not have to switch : buffer is empty */
1960 if (caa_unlikely(off
== 0)) {
1961 unsigned long sb_index
, commit_count
;
1962 struct commit_counters_cold
*cc_cold
;
1965 * We are performing a SWITCH_FLUSH. There may be concurrent
1966 * writes into the buffer if e.g. invoked while performing a
1967 * snapshot on an active trace.
1969 * If the client does not save any header information
1970 * (sub-buffer header size == 0), don't switch empty subbuffer
1971 * on finalize, because it is invalid to deliver a completely
1974 if (!config
->cb
.subbuffer_header_size())
1977 /* Test new buffer integrity */
1978 sb_index
= subbuf_index(offsets
->begin
, chan
);
1979 cc_cold
= shmp_index(handle
, buf
->commit_cold
, sb_index
);
1982 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
1983 reserve_commit_diff
=
1984 (buf_trunc(offsets
->begin
, chan
)
1985 >> chan
->backend
.num_subbuf_order
)
1986 - (commit_count
& chan
->commit_count_mask
);
1987 if (caa_likely(reserve_commit_diff
== 0)) {
1988 /* Next subbuffer not being written to. */
1989 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1990 subbuf_trunc(offsets
->begin
, chan
)
1991 - subbuf_trunc((unsigned long)
1992 uatomic_read(&buf
->consumed
), chan
)
1993 >= chan
->backend
.buf_size
)) {
1995 * We do not overwrite non consumed buffers
1996 * and we are full : don't switch.
2001 * Next subbuffer not being written to, and we
2002 * are either in overwrite mode or the buffer is
2003 * not full. It's safe to write in this new
2009 * Next subbuffer reserve offset does not match the
2010 * commit offset. Don't perform switch in
2011 * producer-consumer and overwrite mode. Caused by
2012 * either a writer OOPS or too many nested writes over a
2013 * reserve/commit pair.
2019 * Need to write the subbuffer start header on finalize.
2021 offsets
->switch_old_start
= 1;
2023 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
2024 /* Note: old points to the next subbuf at offset 0 */
2025 offsets
->end
= offsets
->begin
;
2030 * Force a sub-buffer switch. This operation is completely reentrant : can be
2031 * called while tracing is active with absolutely no lock held.
2033 * For RING_BUFFER_SYNC_PER_CPU ring buffers, as a v_cmpxchg is used for
2034 * some atomic operations, this function must be called from the CPU
2035 * which owns the buffer for a ACTIVE flush. However, for
2036 * RING_BUFFER_SYNC_GLOBAL ring buffers, this function can be called
2039 void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer
*buf
, enum switch_mode mode
,
2040 struct lttng_ust_shm_handle
*handle
)
2042 struct lttng_ust_lib_ring_buffer_channel
*chan
;
2043 const struct lttng_ust_lib_ring_buffer_config
*config
;
2044 struct switch_offsets offsets
;
2045 unsigned long oldidx
;
2048 chan
= shmp(handle
, buf
->backend
.chan
);
2051 config
= &chan
->backend
.config
;
2056 * Perform retryable operations.
2059 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
2061 return; /* Switch not needed */
2062 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
2066 * Atomically update last_tsc. This update races against concurrent
2067 * atomic updates, but the race will always cause supplementary full TSC
2068 * records, never the opposite (missing a full TSC record when it would
2071 save_last_tsc(config
, buf
, tsc
);
2074 * Push the reader if necessary
2076 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
2078 oldidx
= subbuf_index(offsets
.old
, chan
);
2079 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
, handle
);
2082 * May need to populate header start on SWITCH_FLUSH.
2084 if (offsets
.switch_old_start
) {
2085 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
, handle
);
2086 offsets
.old
+= config
->cb
.subbuffer_header_size();
2090 * Switch old subbuffer.
2092 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
, handle
);
2096 bool handle_blocking_retry(int *timeout_left_ms
)
2098 int timeout
= *timeout_left_ms
, delay
;
2100 if (caa_likely(!timeout
))
2101 return false; /* Do not retry, discard event. */
2102 if (timeout
< 0) /* Wait forever. */
2103 delay
= RETRY_DELAY_MS
;
2105 delay
= min_t(int, timeout
, RETRY_DELAY_MS
);
2106 (void) poll(NULL
, 0, delay
);
2108 *timeout_left_ms
-= delay
;
2109 return true; /* Retry. */
2115 * -ENOSPC if event size is too large for packet.
2116 * -ENOBUFS if there is currently not enough space in buffer for the event.
2117 * -EIO if data cannot be written into the buffer for any other reason.
2120 int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer
*buf
,
2121 struct lttng_ust_lib_ring_buffer_channel
*chan
,
2122 struct switch_offsets
*offsets
,
2123 struct lttng_ust_lib_ring_buffer_ctx
*ctx
,
2126 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
2127 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
2128 unsigned long reserve_commit_diff
, offset_cmp
;
2129 int timeout_left_ms
= lttng_ust_ringbuffer_get_timeout(chan
);
2132 offsets
->begin
= offset_cmp
= v_read(config
, &buf
->offset
);
2133 offsets
->old
= offsets
->begin
;
2134 offsets
->switch_new_start
= 0;
2135 offsets
->switch_new_end
= 0;
2136 offsets
->switch_old_end
= 0;
2137 offsets
->pre_header_padding
= 0;
2139 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
2140 if ((int64_t) ctx
->tsc
== -EIO
)
2143 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
2144 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
2146 if (caa_unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
2147 offsets
->switch_new_start
= 1; /* For offsets->begin */
2149 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
2151 &offsets
->pre_header_padding
,
2154 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
2157 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
) +
2158 offsets
->size
> chan
->backend
.subbuf_size
)) {
2159 offsets
->switch_old_end
= 1; /* For offsets->old */
2160 offsets
->switch_new_start
= 1; /* For offsets->begin */
2163 if (caa_unlikely(offsets
->switch_new_start
)) {
2164 unsigned long sb_index
, commit_count
;
2165 struct commit_counters_cold
*cc_cold
;
2168 * We are typically not filling the previous buffer completely.
2170 if (caa_likely(offsets
->switch_old_end
))
2171 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
2172 offsets
->begin
= offsets
->begin
2173 + config
->cb
.subbuffer_header_size();
2174 /* Test new buffer integrity */
2175 sb_index
= subbuf_index(offsets
->begin
, chan
);
2177 * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
2178 * lib_ring_buffer_check_deliver() has the matching
2179 * memory barriers required around commit_cold cc_sb
2180 * updates to ensure reserve and commit counter updates
2181 * are not seen reordered when updated by another CPU.
2184 cc_cold
= shmp_index(handle
, buf
->commit_cold
, sb_index
);
2187 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
2188 /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
2190 if (caa_unlikely(offset_cmp
!= v_read(config
, &buf
->offset
))) {
2192 * The reserve counter have been concurrently updated
2193 * while we read the commit counter. This means the
2194 * commit counter we read might not match buf->offset
2195 * due to concurrent update. We therefore need to retry.
2199 reserve_commit_diff
=
2200 (buf_trunc(offsets
->begin
, chan
)
2201 >> chan
->backend
.num_subbuf_order
)
2202 - (commit_count
& chan
->commit_count_mask
);
2203 if (caa_likely(reserve_commit_diff
== 0)) {
2204 /* Next subbuffer not being written to. */
2205 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
2206 subbuf_trunc(offsets
->begin
, chan
)
2207 - subbuf_trunc((unsigned long)
2208 uatomic_read(&buf
->consumed
), chan
)
2209 >= chan
->backend
.buf_size
)) {
2210 unsigned long nr_lost
;
2212 if (handle_blocking_retry(&timeout_left_ms
))
2216 * We do not overwrite non consumed buffers
2217 * and we are full : record is lost.
2219 nr_lost
= v_read(config
, &buf
->records_lost_full
);
2220 v_inc(config
, &buf
->records_lost_full
);
2221 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
2222 DBG("%lu or more records lost in (%s:%d) (buffer full)\n",
2223 nr_lost
+ 1, chan
->backend
.name
,
2229 * Next subbuffer not being written to, and we
2230 * are either in overwrite mode or the buffer is
2231 * not full. It's safe to write in this new
2236 unsigned long nr_lost
;
2239 * Next subbuffer reserve offset does not match the
2240 * commit offset, and this did not involve update to the
2241 * reserve counter. Drop record in producer-consumer and
2242 * overwrite mode. Caused by either a writer OOPS or too
2243 * many nested writes over a reserve/commit pair.
2245 nr_lost
= v_read(config
, &buf
->records_lost_wrap
);
2246 v_inc(config
, &buf
->records_lost_wrap
);
2247 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
2248 DBG("%lu or more records lost in (%s:%d) (wrap-around)\n",
2249 nr_lost
+ 1, chan
->backend
.name
,
2255 config
->cb
.record_header_size(config
, chan
,
2257 &offsets
->pre_header_padding
,
2260 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
2263 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
)
2264 + offsets
->size
> chan
->backend
.subbuf_size
)) {
2265 unsigned long nr_lost
;
2268 * Record too big for subbuffers, report error, don't
2269 * complete the sub-buffer switch.
2271 nr_lost
= v_read(config
, &buf
->records_lost_big
);
2272 v_inc(config
, &buf
->records_lost_big
);
2273 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
2274 DBG("%lu or more records lost in (%s:%d) record size "
2275 " of %zu bytes is too large for buffer\n",
2276 nr_lost
+ 1, chan
->backend
.name
,
2277 buf
->backend
.cpu
, offsets
->size
);
2282 * We just made a successful buffer switch and the
2283 * record fits in the new subbuffer. Let's write.
2288 * Record fits in the current buffer and we are not on a switch
2289 * boundary. It's safe to write.
2292 offsets
->end
= offsets
->begin
+ offsets
->size
;
2294 if (caa_unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
2296 * The offset_end will fall at the very beginning of the next
2299 offsets
->switch_new_end
= 1; /* For offsets->begin */
2305 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
2306 * @ctx: ring buffer context.
2308 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
2309 * -EIO for other errors, else returns 0.
2310 * It will take care of sub-buffer switching.
2312 int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx
*ctx
,
2315 struct lttng_ust_lib_ring_buffer_channel
*chan
= ctx
->chan
;
2316 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
2317 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
2318 struct lttng_ust_lib_ring_buffer
*buf
;
2319 struct switch_offsets offsets
;
2322 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
2323 buf
= shmp(handle
, chan
->backend
.buf
[ctx
->cpu
].shmp
);
2325 buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
2333 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
2335 if (caa_unlikely(ret
))
2337 } while (caa_unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
2342 * Atomically update last_tsc. This update races against concurrent
2343 * atomic updates, but the race will always cause supplementary full TSC
2344 * records, never the opposite (missing a full TSC record when it would
2347 save_last_tsc(config
, buf
, ctx
->tsc
);
2350 * Push the reader if necessary
2352 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
2355 * Clear noref flag for this subbuffer.
2357 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2358 subbuf_index(offsets
.end
- 1, chan
),
2362 * Switch old subbuffer if needed.
2364 if (caa_unlikely(offsets
.switch_old_end
)) {
2365 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2366 subbuf_index(offsets
.old
- 1, chan
),
2368 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2372 * Populate new subbuffer.
2374 if (caa_unlikely(offsets
.switch_new_start
))
2375 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2377 if (caa_unlikely(offsets
.switch_new_end
))
2378 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2380 ctx
->slot_size
= offsets
.size
;
2381 ctx
->pre_offset
= offsets
.begin
;
2382 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
2387 void lib_ring_buffer_vmcore_check_deliver(const struct lttng_ust_lib_ring_buffer_config
*config
,
2388 struct lttng_ust_lib_ring_buffer
*buf
,
2389 unsigned long commit_count
,
2391 struct lttng_ust_shm_handle
*handle
)
2393 struct commit_counters_hot
*cc_hot
;
2395 if (config
->oops
!= RING_BUFFER_OOPS_CONSISTENCY
)
2397 cc_hot
= shmp_index(handle
, buf
->commit_hot
, idx
);
2400 v_set(config
, &cc_hot
->seq
, commit_count
);
2404 * The ring buffer can count events recorded and overwritten per buffer,
2405 * but it is disabled by default due to its performance overhead.
2407 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
2409 void deliver_count_events(const struct lttng_ust_lib_ring_buffer_config
*config
,
2410 struct lttng_ust_lib_ring_buffer
*buf
,
2412 struct lttng_ust_shm_handle
*handle
)
2414 v_add(config
, subbuffer_get_records_count(config
,
2415 &buf
->backend
, idx
, handle
),
2416 &buf
->records_count
);
2417 v_add(config
, subbuffer_count_records_overrun(config
,
2418 &buf
->backend
, idx
, handle
),
2419 &buf
->records_overrun
);
2421 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
2423 void deliver_count_events(const struct lttng_ust_lib_ring_buffer_config
*config
,
2424 struct lttng_ust_lib_ring_buffer
*buf
,
2426 struct lttng_ust_shm_handle
*handle
)
2429 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
2431 void lib_ring_buffer_check_deliver_slow(const struct lttng_ust_lib_ring_buffer_config
*config
,
2432 struct lttng_ust_lib_ring_buffer
*buf
,
2433 struct lttng_ust_lib_ring_buffer_channel
*chan
,
2434 unsigned long offset
,
2435 unsigned long commit_count
,
2437 struct lttng_ust_shm_handle
*handle
,
2440 unsigned long old_commit_count
= commit_count
2441 - chan
->backend
.subbuf_size
;
2442 struct commit_counters_cold
*cc_cold
;
2445 * If we succeeded at updating cc_sb below, we are the subbuffer
2446 * writer delivering the subbuffer. Deals with concurrent
2447 * updates of the "cc" value without adding a add_return atomic
2448 * operation to the fast path.
2450 * We are doing the delivery in two steps:
2451 * - First, we cmpxchg() cc_sb to the new value
2452 * old_commit_count + 1. This ensures that we are the only
2453 * subbuffer user successfully filling the subbuffer, but we
2454 * do _not_ set the cc_sb value to "commit_count" yet.
2455 * Therefore, other writers that would wrap around the ring
2456 * buffer and try to start writing to our subbuffer would
2457 * have to drop records, because it would appear as
2459 * We therefore have exclusive access to the subbuffer control
2460 * structures. This mutual exclusion with other writers is
2461 * crucially important to perform record overruns count in
2462 * flight recorder mode locklessly.
2463 * - When we are ready to release the subbuffer (either for
2464 * reading or for overrun by other writers), we simply set the
2465 * cc_sb value to "commit_count" and perform delivery.
2467 * The subbuffer size is least 2 bytes (minimum size: 1 page).
2468 * This guarantees that old_commit_count + 1 != commit_count.
2472 * Order prior updates to reserve count prior to the
2473 * commit_cold cc_sb update.
2476 cc_cold
= shmp_index(handle
, buf
->commit_cold
, idx
);
2479 if (caa_likely(v_cmpxchg(config
, &cc_cold
->cc_sb
,
2480 old_commit_count
, old_commit_count
+ 1)
2481 == old_commit_count
)) {
2485 * Start of exclusive subbuffer access. We are
2486 * guaranteed to be the last writer in this subbuffer
2487 * and any other writer trying to access this subbuffer
2488 * in this state is required to drop records.
2490 * We can read the ts_end for the current sub-buffer
2491 * which has been saved by the very last space
2492 * reservation for the current sub-buffer.
2494 * Order increment of commit counter before reading ts_end.
2497 ts_end
= shmp_index(handle
, buf
->ts_end
, idx
);
2500 deliver_count_events(config
, buf
, idx
, handle
);
2501 config
->cb
.buffer_end(buf
, *ts_end
, idx
,
2502 lib_ring_buffer_get_data_size(config
,
2509 * Increment the packet counter while we have exclusive
2512 subbuffer_inc_packet_count(config
, &buf
->backend
, idx
, handle
);
2515 * Set noref flag and offset for this subbuffer id.
2516 * Contains a memory barrier that ensures counter stores
2517 * are ordered before set noref and offset.
2519 lib_ring_buffer_set_noref_offset(config
, &buf
->backend
, idx
,
2520 buf_trunc_val(offset
, chan
), handle
);
2523 * Order set_noref and record counter updates before the
2524 * end of subbuffer exclusive access. Orders with
2525 * respect to writers coming into the subbuffer after
2526 * wrap around, and also order wrt concurrent readers.
2529 /* End of exclusive subbuffer access */
2530 v_set(config
, &cc_cold
->cc_sb
, commit_count
);
2532 * Order later updates to reserve count after
2533 * the commit cold cc_sb update.
2536 lib_ring_buffer_vmcore_check_deliver(config
, buf
,
2537 commit_count
, idx
, handle
);
2540 * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
2542 if (config
->wakeup
== RING_BUFFER_WAKEUP_BY_WRITER
2543 && uatomic_read(&buf
->active_readers
)
2544 && lib_ring_buffer_poll_deliver(config
, buf
, chan
, handle
)) {
2545 lib_ring_buffer_wakeup(buf
, handle
);
2551 * Force a read (imply TLS fixup for dlopen) of TLS variables.
2553 void lttng_fixup_ringbuffer_tls(void)
2555 asm volatile ("" : : "m" (URCU_TLS(lib_ring_buffer_nesting
)));
2558 void lib_ringbuffer_signal_init(void)
2564 * Block signal for entire process, so only our thread processes
2568 ret
= pthread_sigmask(SIG_BLOCK
, &mask
, NULL
);
2571 PERROR("pthread_sigmask");