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-utils.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_config(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_align: alignment, in bytes, of the private data area. (config)
942 * @priv_data_size: length, in bytes, of the private data area. (config)
943 * @priv_data_init: initialization data for private data. (config)
944 * @priv: local private data (memory owner by caller)
945 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
946 * address mapping. It is used only by RING_BUFFER_STATIC
947 * configuration. It can be set to NULL for other backends.
948 * @subbuf_size: subbuffer size
949 * @num_subbuf: number of subbuffers
950 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
951 * padding to let readers get those sub-buffers.
952 * Used for live streaming.
953 * @read_timer_interval: Time interval (in us) to wake up pending readers.
954 * @stream_fds: array of stream file descriptors.
955 * @nr_stream_fds: number of file descriptors in array.
958 * Returns NULL on failure.
960 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
,
966 void *buf_addr
, size_t subbuf_size
,
967 size_t num_subbuf
, unsigned int switch_timer_interval
,
968 unsigned int read_timer_interval
,
969 const int *stream_fds
, int nr_stream_fds
,
970 int64_t blocking_timeout
)
973 size_t shmsize
, chansize
;
974 struct lttng_ust_lib_ring_buffer_channel
*chan
;
975 struct lttng_ust_shm_handle
*handle
;
976 struct shm_object
*shmobj
;
977 unsigned int nr_streams
;
978 int64_t blocking_timeout_ms
;
980 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
981 nr_streams
= num_possible_cpus();
985 if (nr_stream_fds
!= nr_streams
)
988 if (blocking_timeout
< -1) {
992 if (blocking_timeout
== -1) {
993 blocking_timeout_ms
= -1;
995 blocking_timeout_ms
= blocking_timeout
/ 1000;
996 if (blocking_timeout_ms
!= (int32_t) blocking_timeout_ms
) {
1001 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
1002 read_timer_interval
))
1005 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
1009 /* Allocate table for channel + per-cpu buffers */
1010 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
1012 goto error_table_alloc
;
1014 /* Calculate the shm allocation layout */
1015 shmsize
= sizeof(struct lttng_ust_lib_ring_buffer_channel
);
1016 shmsize
+= lttng_ust_offset_align(shmsize
, __alignof__(struct lttng_ust_lib_ring_buffer_shmp
));
1017 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
) * nr_streams
;
1019 if (priv_data_align
)
1020 shmsize
+= lttng_ust_offset_align(shmsize
, priv_data_align
);
1021 shmsize
+= priv_data_size
;
1023 /* Allocate normal memory for channel (not shared) */
1024 shmobj
= shm_object_table_alloc(handle
->table
, shmsize
, SHM_OBJECT_MEM
,
1028 /* struct lttng_ust_lib_ring_buffer_channel is at object 0, offset 0 (hardcoded) */
1029 set_shmp(handle
->chan
, zalloc_shm(shmobj
, chansize
));
1030 assert(handle
->chan
._ref
.index
== 0);
1031 assert(handle
->chan
._ref
.offset
== 0);
1032 chan
= shmp(handle
, handle
->chan
);
1035 chan
->nr_streams
= nr_streams
;
1037 /* space for private data */
1038 if (priv_data_size
) {
1041 DECLARE_SHMP(void, priv_data_alloc
);
1043 align_shm(shmobj
, priv_data_align
);
1044 chan
->priv_data_offset
= shmobj
->allocated_len
;
1045 set_shmp(priv_data_alloc
, zalloc_shm(shmobj
, priv_data_size
));
1046 if (!shmp(handle
, priv_data_alloc
))
1048 priv_config
= channel_get_private_config(chan
);
1049 memcpy(priv_config
, priv_data_init
, priv_data_size
);
1051 chan
->priv_data_offset
= -1;
1054 chan
->u
.s
.blocking_timeout_ms
= (int32_t) blocking_timeout_ms
;
1056 channel_set_private(chan
, priv
);
1058 ret
= channel_backend_init(&chan
->backend
, name
, config
,
1059 subbuf_size
, num_subbuf
, handle
,
1062 goto error_backend_init
;
1064 chan
->handle
= handle
;
1065 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
1067 chan
->switch_timer_interval
= switch_timer_interval
;
1068 chan
->read_timer_interval
= read_timer_interval
;
1069 lib_ring_buffer_channel_switch_timer_start(chan
);
1070 lib_ring_buffer_channel_read_timer_start(chan
);
1076 shm_object_table_destroy(handle
->table
, 1);
1082 struct lttng_ust_shm_handle
*channel_handle_create(void *data
,
1083 uint64_t memory_map_size
,
1086 struct lttng_ust_shm_handle
*handle
;
1087 struct shm_object
*object
;
1089 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
1093 /* Allocate table for channel + per-cpu buffers */
1094 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
1096 goto error_table_alloc
;
1097 /* Add channel object */
1098 object
= shm_object_table_append_mem(handle
->table
, data
,
1099 memory_map_size
, wakeup_fd
);
1101 goto error_table_object
;
1102 /* struct lttng_ust_lib_ring_buffer_channel is at object 0, offset 0 (hardcoded) */
1103 handle
->chan
._ref
.index
= 0;
1104 handle
->chan
._ref
.offset
= 0;
1108 shm_object_table_destroy(handle
->table
, 0);
1114 int channel_handle_add_stream(struct lttng_ust_shm_handle
*handle
,
1115 int shm_fd
, int wakeup_fd
, uint32_t stream_nr
,
1116 uint64_t memory_map_size
)
1118 struct shm_object
*object
;
1120 /* Add stream object */
1121 object
= shm_object_table_append_shm(handle
->table
,
1122 shm_fd
, wakeup_fd
, stream_nr
,
1129 unsigned int channel_handle_get_nr_streams(struct lttng_ust_shm_handle
*handle
)
1131 assert(handle
->table
);
1132 return handle
->table
->allocated_len
- 1;
1136 void channel_release(struct lttng_ust_lib_ring_buffer_channel
*chan
, struct lttng_ust_shm_handle
*handle
,
1139 channel_free(chan
, handle
, consumer
);
1143 * channel_destroy - Finalize, wait for q.s. and destroy channel.
1144 * @chan: channel to destroy
1146 * Holds cpu hotplug.
1147 * Call "destroy" callback, finalize channels, decrement the channel
1148 * reference count. Note that when readers have completed data
1149 * consumption of finalized channels, get_subbuf() will return -ENODATA.
1150 * They should release their handle at that point.
1152 void channel_destroy(struct lttng_ust_lib_ring_buffer_channel
*chan
, struct lttng_ust_shm_handle
*handle
,
1157 * Note: the consumer takes care of finalizing and
1158 * switching the buffers.
1160 channel_unregister_notifiers(chan
, handle
);
1162 * The consumer prints errors.
1164 channel_print_errors(chan
, handle
);
1168 * sessiond/consumer are keeping a reference on the shm file
1169 * descriptor directly. No need to refcount.
1171 channel_release(chan
, handle
, consumer
);
1175 struct lttng_ust_lib_ring_buffer
*channel_get_ring_buffer(
1176 const struct lttng_ust_lib_ring_buffer_config
*config
,
1177 struct lttng_ust_lib_ring_buffer_channel
*chan
, int cpu
,
1178 struct lttng_ust_shm_handle
*handle
,
1179 int *shm_fd
, int *wait_fd
,
1181 uint64_t *memory_map_size
)
1183 struct shm_ref
*ref
;
1185 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1188 if (cpu
>= num_possible_cpus())
1191 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1192 *shm_fd
= shm_get_shm_fd(handle
, ref
);
1193 *wait_fd
= shm_get_wait_fd(handle
, ref
);
1194 *wakeup_fd
= shm_get_wakeup_fd(handle
, ref
);
1195 if (shm_get_shm_size(handle
, ref
, memory_map_size
))
1197 return shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
1200 int ring_buffer_channel_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1201 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1202 struct lttng_ust_shm_handle
*handle
)
1204 struct shm_ref
*ref
;
1206 ref
= &handle
->chan
._ref
;
1207 return shm_close_wait_fd(handle
, ref
);
1210 int ring_buffer_channel_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1211 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1212 struct lttng_ust_shm_handle
*handle
)
1214 struct shm_ref
*ref
;
1216 ref
= &handle
->chan
._ref
;
1217 return shm_close_wakeup_fd(handle
, ref
);
1220 int ring_buffer_stream_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1221 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1222 struct lttng_ust_shm_handle
*handle
,
1225 struct shm_ref
*ref
;
1227 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1230 if (cpu
>= num_possible_cpus())
1233 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1234 return shm_close_wait_fd(handle
, ref
);
1237 int ring_buffer_stream_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1238 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1239 struct lttng_ust_shm_handle
*handle
,
1242 struct shm_ref
*ref
;
1245 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1248 if (cpu
>= num_possible_cpus())
1251 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1252 pthread_mutex_lock(&wakeup_fd_mutex
);
1253 ret
= shm_close_wakeup_fd(handle
, ref
);
1254 pthread_mutex_unlock(&wakeup_fd_mutex
);
1258 int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer
*buf
,
1259 struct lttng_ust_shm_handle
*handle
)
1261 if (uatomic_cmpxchg(&buf
->active_readers
, 0, 1) != 0)
1267 void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer
*buf
,
1268 struct lttng_ust_shm_handle
*handle
)
1270 struct lttng_ust_lib_ring_buffer_channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1274 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1276 uatomic_dec(&buf
->active_readers
);
1280 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
1282 * @consumed: consumed count indicating the position where to read
1283 * @produced: produced count, indicates position when to stop reading
1285 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1286 * data to read at consumed position, or 0 if the get operation succeeds.
1289 int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer
*buf
,
1290 unsigned long *consumed
, unsigned long *produced
,
1291 struct lttng_ust_shm_handle
*handle
)
1293 struct lttng_ust_lib_ring_buffer_channel
*chan
;
1294 const struct lttng_ust_lib_ring_buffer_config
*config
;
1295 unsigned long consumed_cur
, write_offset
;
1298 chan
= shmp(handle
, buf
->backend
.chan
);
1301 config
= &chan
->backend
.config
;
1302 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
1304 * Read finalized before counters.
1307 consumed_cur
= uatomic_read(&buf
->consumed
);
1309 * No need to issue a memory barrier between consumed count read and
1310 * write offset read, because consumed count can only change
1311 * concurrently in overwrite mode, and we keep a sequence counter
1312 * identifier derived from the write offset to check we are getting
1313 * the same sub-buffer we are expecting (the sub-buffers are atomically
1314 * "tagged" upon writes, tags are checked upon read).
1316 write_offset
= v_read(config
, &buf
->offset
);
1319 * Check that we are not about to read the same subbuffer in
1320 * which the writer head is.
1322 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
1326 *consumed
= consumed_cur
;
1327 *produced
= subbuf_trunc(write_offset
, chan
);
1333 * The memory barriers __wait_event()/wake_up_interruptible() take care
1334 * of "raw_spin_is_locked" memory ordering.
1343 * Performs the same function as lib_ring_buffer_snapshot(), but the positions
1344 * are saved regardless of whether the consumed and produced positions are
1345 * in the same subbuffer.
1347 * @consumed: consumed byte count indicating the last position read
1348 * @produced: produced byte count indicating the last position written
1350 * This function is meant to provide information on the exact producer and
1351 * consumer positions without regard for the "snapshot" feature.
1353 int lib_ring_buffer_snapshot_sample_positions(
1354 struct lttng_ust_lib_ring_buffer
*buf
,
1355 unsigned long *consumed
, unsigned long *produced
,
1356 struct lttng_ust_shm_handle
*handle
)
1358 struct lttng_ust_lib_ring_buffer_channel
*chan
;
1359 const struct lttng_ust_lib_ring_buffer_config
*config
;
1361 chan
= shmp(handle
, buf
->backend
.chan
);
1364 config
= &chan
->backend
.config
;
1366 *consumed
= uatomic_read(&buf
->consumed
);
1368 * No need to issue a memory barrier between consumed count read and
1369 * write offset read, because consumed count can only change
1370 * concurrently in overwrite mode, and we keep a sequence counter
1371 * identifier derived from the write offset to check we are getting
1372 * the same sub-buffer we are expecting (the sub-buffers are atomically
1373 * "tagged" upon writes, tags are checked upon read).
1375 *produced
= v_read(config
, &buf
->offset
);
1380 * lib_ring_buffer_move_consumer - move consumed counter forward
1382 * @consumed_new: new consumed count value
1384 void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer
*buf
,
1385 unsigned long consumed_new
,
1386 struct lttng_ust_shm_handle
*handle
)
1388 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1389 struct lttng_ust_lib_ring_buffer_channel
*chan
;
1390 unsigned long consumed
;
1392 chan
= shmp(handle
, bufb
->chan
);
1395 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1398 * Only push the consumed value forward.
1399 * If the consumed cmpxchg fails, this is because we have been pushed by
1400 * the writer in flight recorder mode.
1402 consumed
= uatomic_read(&buf
->consumed
);
1403 while ((long) consumed
- (long) consumed_new
< 0)
1404 consumed
= uatomic_cmpxchg(&buf
->consumed
, consumed
,
1409 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
1411 * @consumed: consumed count indicating the position where to read
1413 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1414 * data to read at consumed position, or 0 if the get operation succeeds.
1416 int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
1417 unsigned long consumed
,
1418 struct lttng_ust_shm_handle
*handle
)
1420 struct lttng_ust_lib_ring_buffer_channel
*chan
;
1421 const struct lttng_ust_lib_ring_buffer_config
*config
;
1422 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
1423 int ret
, finalized
, nr_retry
= LTTNG_UST_RING_BUFFER_GET_RETRY
;
1424 struct commit_counters_cold
*cc_cold
;
1426 chan
= shmp(handle
, buf
->backend
.chan
);
1429 config
= &chan
->backend
.config
;
1431 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
1433 * Read finalized before counters.
1436 consumed_cur
= uatomic_read(&buf
->consumed
);
1437 consumed_idx
= subbuf_index(consumed
, chan
);
1438 cc_cold
= shmp_index(handle
, buf
->commit_cold
, consumed_idx
);
1441 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
1443 * Make sure we read the commit count before reading the buffer
1444 * data and the write offset. Correct consumed offset ordering
1445 * wrt commit count is insured by the use of cmpxchg to update
1446 * the consumed offset.
1449 * Local rmb to match the remote wmb to read the commit count
1450 * before the buffer data and the write offset.
1454 write_offset
= v_read(config
, &buf
->offset
);
1457 * Check that the buffer we are getting is after or at consumed_cur
1460 if ((long) subbuf_trunc(consumed
, chan
)
1461 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
1465 * Check that the subbuffer we are trying to consume has been
1466 * already fully committed. There are a few causes that can make
1467 * this unavailability situation occur:
1469 * Temporary (short-term) situation:
1470 * - Application is running on a different CPU, between reserve
1471 * and commit ring buffer operations,
1472 * - Application is preempted between reserve and commit ring
1473 * buffer operations,
1475 * Long-term situation:
1476 * - Application is stopped (SIGSTOP) between reserve and commit
1477 * ring buffer operations. Could eventually be resumed by
1479 * - Application is killed (SIGTERM, SIGINT, SIGKILL) between
1480 * reserve and commit ring buffer operation.
1482 * From a consumer perspective, handling short-term
1483 * unavailability situations is performed by retrying a few
1484 * times after a delay. Handling long-term unavailability
1485 * situations is handled by failing to get the sub-buffer.
1487 * In all of those situations, if the application is taking a
1488 * long time to perform its commit after ring buffer space
1489 * reservation, we can end up in a situation where the producer
1490 * will fill the ring buffer and try to write into the same
1491 * sub-buffer again (which has a missing commit). This is
1492 * handled by the producer in the sub-buffer switch handling
1493 * code of the reserve routine by detecting unbalanced
1494 * reserve/commit counters and discarding all further events
1495 * until the situation is resolved in those situations. Two
1496 * scenarios can occur:
1498 * 1) The application causing the reserve/commit counters to be
1499 * unbalanced has been terminated. In this situation, all
1500 * further events will be discarded in the buffers, and no
1501 * further buffer data will be readable by the consumer
1502 * daemon. Tearing down the UST tracing session and starting
1503 * anew is a work-around for those situations. Note that this
1504 * only affects per-UID tracing. In per-PID tracing, the
1505 * application vanishes with the termination, and therefore
1506 * no more data needs to be written to the buffers.
1507 * 2) The application causing the unbalance has been delayed for
1508 * a long time, but will eventually try to increment the
1509 * commit counter after eventually writing to the sub-buffer.
1510 * This situation can cause events to be discarded until the
1511 * application resumes its operations.
1513 if (((commit_count
- chan
->backend
.subbuf_size
)
1514 & chan
->commit_count_mask
)
1515 - (buf_trunc(consumed
, chan
)
1516 >> chan
->backend
.num_subbuf_order
)
1518 if (nr_retry
-- > 0) {
1519 if (nr_retry
<= (LTTNG_UST_RING_BUFFER_GET_RETRY
>> 1))
1520 (void) poll(NULL
, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS
);
1528 * Check that we are not about to read the same subbuffer in
1529 * which the writer head is.
1531 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed
, chan
)
1536 * Failure to get the subbuffer causes a busy-loop retry without going
1537 * to a wait queue. These are caused by short-lived race windows where
1538 * the writer is getting access to a subbuffer we were trying to get
1539 * access to. Also checks that the "consumed" buffer count we are
1540 * looking for matches the one contained in the subbuffer id.
1542 * The short-lived race window described here can be affected by
1543 * application signals and preemption, thus requiring to bound
1544 * the loop to a maximum number of retry.
1546 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1547 consumed_idx
, buf_trunc_val(consumed
, chan
),
1550 if (nr_retry
-- > 0) {
1551 if (nr_retry
<= (LTTNG_UST_RING_BUFFER_GET_RETRY
>> 1))
1552 (void) poll(NULL
, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS
);
1558 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1560 buf
->get_subbuf_consumed
= consumed
;
1561 buf
->get_subbuf
= 1;
1567 * The memory barriers __wait_event()/wake_up_interruptible() take care
1568 * of "raw_spin_is_locked" memory ordering.
1577 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1580 void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
1581 struct lttng_ust_shm_handle
*handle
)
1583 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1584 struct lttng_ust_lib_ring_buffer_channel
*chan
;
1585 const struct lttng_ust_lib_ring_buffer_config
*config
;
1586 unsigned long sb_bindex
, consumed_idx
, consumed
;
1587 struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*rpages
;
1588 struct lttng_ust_lib_ring_buffer_backend_pages
*backend_pages
;
1590 chan
= shmp(handle
, bufb
->chan
);
1593 config
= &chan
->backend
.config
;
1594 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1596 if (!buf
->get_subbuf
) {
1598 * Reader puts a subbuffer it did not get.
1600 CHAN_WARN_ON(chan
, 1);
1603 consumed
= buf
->get_subbuf_consumed
;
1604 buf
->get_subbuf
= 0;
1607 * Clear the records_unread counter. (overruns counter)
1608 * Can still be non-zero if a file reader simply grabbed the data
1609 * without using iterators.
1610 * Can be below zero if an iterator is used on a snapshot more than
1613 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1614 rpages
= shmp_index(handle
, bufb
->array
, sb_bindex
);
1617 backend_pages
= shmp(handle
, rpages
->shmp
);
1620 v_add(config
, v_read(config
, &backend_pages
->records_unread
),
1621 &bufb
->records_read
);
1622 v_set(config
, &backend_pages
->records_unread
, 0);
1623 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1624 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1625 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1628 * Exchange the reader subbuffer with the one we put in its place in the
1629 * writer subbuffer table. Expect the original consumed count. If
1630 * update_read_sb_index fails, this is because the writer updated the
1631 * subbuffer concurrently. We should therefore keep the subbuffer we
1632 * currently have: it has become invalid to try reading this sub-buffer
1633 * consumed count value anyway.
1635 consumed_idx
= subbuf_index(consumed
, chan
);
1636 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1637 consumed_idx
, buf_trunc_val(consumed
, chan
),
1640 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1641 * if the writer concurrently updated it.
1646 * cons_offset is an iterator on all subbuffer offsets between the reader
1647 * position and the writer position. (inclusive)
1650 void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
1651 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1652 unsigned long cons_offset
,
1654 struct lttng_ust_shm_handle
*handle
)
1656 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1657 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1658 struct commit_counters_hot
*cc_hot
;
1659 struct commit_counters_cold
*cc_cold
;
1661 cons_idx
= subbuf_index(cons_offset
, chan
);
1662 cc_hot
= shmp_index(handle
, buf
->commit_hot
, cons_idx
);
1665 cc_cold
= shmp_index(handle
, buf
->commit_cold
, cons_idx
);
1668 commit_count
= v_read(config
, &cc_hot
->cc
);
1669 commit_count_sb
= v_read(config
, &cc_cold
->cc_sb
);
1671 if (subbuf_offset(commit_count
, chan
) != 0)
1672 DBG("ring buffer %s, cpu %d: "
1673 "commit count in subbuffer %lu,\n"
1674 "expecting multiples of %lu bytes\n"
1675 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1676 chan
->backend
.name
, cpu
, cons_idx
,
1677 chan
->backend
.subbuf_size
,
1678 commit_count
, commit_count_sb
);
1680 DBG("ring buffer: %s, cpu %d: %lu bytes committed\n",
1681 chan
->backend
.name
, cpu
, commit_count
);
1685 void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
1686 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1687 int cpu
, struct lttng_ust_shm_handle
*handle
)
1689 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1690 unsigned long write_offset
, cons_offset
;
1693 * No need to order commit_count, write_offset and cons_offset reads
1694 * because we execute at teardown when no more writer nor reader
1695 * references are left.
1697 write_offset
= v_read(config
, &buf
->offset
);
1698 cons_offset
= uatomic_read(&buf
->consumed
);
1699 if (write_offset
!= cons_offset
)
1700 DBG("ring buffer %s, cpu %d: "
1701 "non-consumed data\n"
1702 " [ %lu bytes written, %lu bytes read ]\n",
1703 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1705 for (cons_offset
= uatomic_read(&buf
->consumed
);
1706 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1709 cons_offset
= subbuf_align(cons_offset
, chan
))
1710 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1715 void lib_ring_buffer_print_errors(struct lttng_ust_lib_ring_buffer_channel
*chan
,
1716 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
1717 struct lttng_ust_shm_handle
*handle
)
1719 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1721 if (!strcmp(chan
->backend
.name
, "relay-metadata-mmap")) {
1722 DBG("ring buffer %s: %lu records written, "
1723 "%lu records overrun\n",
1725 v_read(config
, &buf
->records_count
),
1726 v_read(config
, &buf
->records_overrun
));
1728 DBG("ring buffer %s, cpu %d: %lu records written, "
1729 "%lu records overrun\n",
1730 chan
->backend
.name
, cpu
,
1731 v_read(config
, &buf
->records_count
),
1732 v_read(config
, &buf
->records_overrun
));
1734 if (v_read(config
, &buf
->records_lost_full
)
1735 || v_read(config
, &buf
->records_lost_wrap
)
1736 || v_read(config
, &buf
->records_lost_big
))
1737 DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n"
1738 " [ %lu buffer full, %lu nest buffer wrap-around, "
1739 "%lu event too big ]\n",
1740 chan
->backend
.name
, cpu
,
1741 v_read(config
, &buf
->records_lost_full
),
1742 v_read(config
, &buf
->records_lost_wrap
),
1743 v_read(config
, &buf
->records_lost_big
));
1745 lib_ring_buffer_print_buffer_errors(buf
, chan
, cpu
, handle
);
1749 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1751 * Only executed by SWITCH_FLUSH, which can be issued while tracing is
1752 * active or at buffer finalization (destroy).
1755 void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer
*buf
,
1756 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1757 struct switch_offsets
*offsets
,
1759 struct lttng_ust_shm_handle
*handle
)
1761 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1762 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1763 unsigned long commit_count
;
1764 struct commit_counters_hot
*cc_hot
;
1766 config
->cb
.buffer_begin(buf
, tsc
, oldidx
, handle
);
1769 * Order all writes to buffer before the commit count update that will
1770 * determine that the subbuffer is full.
1773 cc_hot
= shmp_index(handle
, buf
->commit_hot
, oldidx
);
1776 v_add(config
, config
->cb
.subbuffer_header_size(),
1778 commit_count
= v_read(config
, &cc_hot
->cc
);
1779 /* Check if the written buffer has to be delivered */
1780 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1781 commit_count
, oldidx
, handle
, tsc
);
1782 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1783 offsets
->old
+ config
->cb
.subbuffer_header_size(),
1784 commit_count
, handle
, cc_hot
);
1788 * lib_ring_buffer_switch_old_end: switch old subbuffer
1790 * Note : offset_old should never be 0 here. It is ok, because we never perform
1791 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1792 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1796 void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer
*buf
,
1797 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1798 struct switch_offsets
*offsets
,
1800 struct lttng_ust_shm_handle
*handle
)
1802 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1803 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1804 unsigned long commit_count
, padding_size
, data_size
;
1805 struct commit_counters_hot
*cc_hot
;
1808 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1809 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1810 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
,
1813 ts_end
= shmp_index(handle
, buf
->ts_end
, oldidx
);
1817 * This is the last space reservation in that sub-buffer before
1818 * it gets delivered. This provides exclusive access to write to
1819 * this sub-buffer's ts_end. There are also no concurrent
1820 * readers of that ts_end because delivery of that sub-buffer is
1821 * postponed until the commit counter is incremented for the
1822 * current space reservation.
1827 * Order all writes to buffer and store to ts_end before the commit
1828 * count update that will determine that the subbuffer is full.
1831 cc_hot
= shmp_index(handle
, buf
->commit_hot
, oldidx
);
1834 v_add(config
, padding_size
, &cc_hot
->cc
);
1835 commit_count
= v_read(config
, &cc_hot
->cc
);
1836 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1837 commit_count
, oldidx
, handle
, tsc
);
1838 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1839 offsets
->old
+ padding_size
, commit_count
, handle
,
1844 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1846 * This code can be executed unordered : writers may already have written to the
1847 * sub-buffer before this code gets executed, caution. The commit makes sure
1848 * that this code is executed before the deliver of this sub-buffer.
1851 void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer
*buf
,
1852 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1853 struct switch_offsets
*offsets
,
1855 struct lttng_ust_shm_handle
*handle
)
1857 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1858 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1859 unsigned long commit_count
;
1860 struct commit_counters_hot
*cc_hot
;
1862 config
->cb
.buffer_begin(buf
, tsc
, beginidx
, handle
);
1865 * Order all writes to buffer before the commit count update that will
1866 * determine that the subbuffer is full.
1869 cc_hot
= shmp_index(handle
, buf
->commit_hot
, beginidx
);
1872 v_add(config
, config
->cb
.subbuffer_header_size(), &cc_hot
->cc
);
1873 commit_count
= v_read(config
, &cc_hot
->cc
);
1874 /* Check if the written buffer has to be delivered */
1875 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1876 commit_count
, beginidx
, handle
, tsc
);
1877 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1878 offsets
->begin
+ config
->cb
.subbuffer_header_size(),
1879 commit_count
, handle
, cc_hot
);
1883 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1885 * Calls subbuffer_set_data_size() to set the data size of the current
1886 * sub-buffer. We do not need to perform check_deliver nor commit here,
1887 * since this task will be done by the "commit" of the event for which
1888 * we are currently doing the space reservation.
1891 void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer
*buf
,
1892 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1893 struct switch_offsets
*offsets
,
1895 struct lttng_ust_shm_handle
*handle
)
1897 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1898 unsigned long endidx
, data_size
;
1901 endidx
= subbuf_index(offsets
->end
- 1, chan
);
1902 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1903 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
,
1905 ts_end
= shmp_index(handle
, buf
->ts_end
, endidx
);
1909 * This is the last space reservation in that sub-buffer before
1910 * it gets delivered. This provides exclusive access to write to
1911 * this sub-buffer's ts_end. There are also no concurrent
1912 * readers of that ts_end because delivery of that sub-buffer is
1913 * postponed until the commit counter is incremented for the
1914 * current space reservation.
1922 * !0 if execution must be aborted.
1925 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1926 struct lttng_ust_lib_ring_buffer
*buf
,
1927 struct lttng_ust_lib_ring_buffer_channel
*chan
,
1928 struct switch_offsets
*offsets
,
1930 struct lttng_ust_shm_handle
*handle
)
1932 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1933 unsigned long off
, reserve_commit_diff
;
1935 offsets
->begin
= v_read(config
, &buf
->offset
);
1936 offsets
->old
= offsets
->begin
;
1937 offsets
->switch_old_start
= 0;
1938 off
= subbuf_offset(offsets
->begin
, chan
);
1940 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1943 * Ensure we flush the header of an empty subbuffer when doing the
1944 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1945 * total data gathering duration even if there were no records saved
1946 * after the last buffer switch.
1947 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1948 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1949 * subbuffer header as appropriate.
1950 * The next record that reserves space will be responsible for
1951 * populating the following subbuffer header. We choose not to populate
1952 * the next subbuffer header here because we want to be able to use
1953 * SWITCH_ACTIVE for periodical buffer flush, which must
1954 * guarantee that all the buffer content (records and header
1955 * timestamps) are visible to the reader. This is required for
1956 * quiescence guarantees for the fusion merge.
1958 if (mode
!= SWITCH_FLUSH
&& !off
)
1959 return -1; /* we do not have to switch : buffer is empty */
1961 if (caa_unlikely(off
== 0)) {
1962 unsigned long sb_index
, commit_count
;
1963 struct commit_counters_cold
*cc_cold
;
1966 * We are performing a SWITCH_FLUSH. There may be concurrent
1967 * writes into the buffer if e.g. invoked while performing a
1968 * snapshot on an active trace.
1970 * If the client does not save any header information
1971 * (sub-buffer header size == 0), don't switch empty subbuffer
1972 * on finalize, because it is invalid to deliver a completely
1975 if (!config
->cb
.subbuffer_header_size())
1978 /* Test new buffer integrity */
1979 sb_index
= subbuf_index(offsets
->begin
, chan
);
1980 cc_cold
= shmp_index(handle
, buf
->commit_cold
, sb_index
);
1983 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
1984 reserve_commit_diff
=
1985 (buf_trunc(offsets
->begin
, chan
)
1986 >> chan
->backend
.num_subbuf_order
)
1987 - (commit_count
& chan
->commit_count_mask
);
1988 if (caa_likely(reserve_commit_diff
== 0)) {
1989 /* Next subbuffer not being written to. */
1990 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1991 subbuf_trunc(offsets
->begin
, chan
)
1992 - subbuf_trunc((unsigned long)
1993 uatomic_read(&buf
->consumed
), chan
)
1994 >= chan
->backend
.buf_size
)) {
1996 * We do not overwrite non consumed buffers
1997 * and we are full : don't switch.
2002 * Next subbuffer not being written to, and we
2003 * are either in overwrite mode or the buffer is
2004 * not full. It's safe to write in this new
2010 * Next subbuffer reserve offset does not match the
2011 * commit offset. Don't perform switch in
2012 * producer-consumer and overwrite mode. Caused by
2013 * either a writer OOPS or too many nested writes over a
2014 * reserve/commit pair.
2020 * Need to write the subbuffer start header on finalize.
2022 offsets
->switch_old_start
= 1;
2024 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
2025 /* Note: old points to the next subbuf at offset 0 */
2026 offsets
->end
= offsets
->begin
;
2031 * Force a sub-buffer switch. This operation is completely reentrant : can be
2032 * called while tracing is active with absolutely no lock held.
2034 * For RING_BUFFER_SYNC_PER_CPU ring buffers, as a v_cmpxchg is used for
2035 * some atomic operations, this function must be called from the CPU
2036 * which owns the buffer for a ACTIVE flush. However, for
2037 * RING_BUFFER_SYNC_GLOBAL ring buffers, this function can be called
2040 void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer
*buf
, enum switch_mode mode
,
2041 struct lttng_ust_shm_handle
*handle
)
2043 struct lttng_ust_lib_ring_buffer_channel
*chan
;
2044 const struct lttng_ust_lib_ring_buffer_config
*config
;
2045 struct switch_offsets offsets
;
2046 unsigned long oldidx
;
2049 chan
= shmp(handle
, buf
->backend
.chan
);
2052 config
= &chan
->backend
.config
;
2057 * Perform retryable operations.
2060 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
2062 return; /* Switch not needed */
2063 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
2067 * Atomically update last_tsc. This update races against concurrent
2068 * atomic updates, but the race will always cause supplementary full TSC
2069 * records, never the opposite (missing a full TSC record when it would
2072 save_last_tsc(config
, buf
, tsc
);
2075 * Push the reader if necessary
2077 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
2079 oldidx
= subbuf_index(offsets
.old
, chan
);
2080 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
, handle
);
2083 * May need to populate header start on SWITCH_FLUSH.
2085 if (offsets
.switch_old_start
) {
2086 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
, handle
);
2087 offsets
.old
+= config
->cb
.subbuffer_header_size();
2091 * Switch old subbuffer.
2093 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
, handle
);
2097 bool handle_blocking_retry(int *timeout_left_ms
)
2099 int timeout
= *timeout_left_ms
, delay
;
2101 if (caa_likely(!timeout
))
2102 return false; /* Do not retry, discard event. */
2103 if (timeout
< 0) /* Wait forever. */
2104 delay
= RETRY_DELAY_MS
;
2106 delay
= min_t(int, timeout
, RETRY_DELAY_MS
);
2107 (void) poll(NULL
, 0, delay
);
2109 *timeout_left_ms
-= delay
;
2110 return true; /* Retry. */
2116 * -ENOSPC if event size is too large for packet.
2117 * -ENOBUFS if there is currently not enough space in buffer for the event.
2118 * -EIO if data cannot be written into the buffer for any other reason.
2121 int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer
*buf
,
2122 struct lttng_ust_lib_ring_buffer_channel
*chan
,
2123 struct switch_offsets
*offsets
,
2124 struct lttng_ust_lib_ring_buffer_ctx
*ctx
,
2127 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
2128 struct lttng_ust_shm_handle
*handle
= ctx
->chan
->handle
;
2129 unsigned long reserve_commit_diff
, offset_cmp
;
2130 int timeout_left_ms
= lttng_ust_ringbuffer_get_timeout(chan
);
2133 offsets
->begin
= offset_cmp
= v_read(config
, &buf
->offset
);
2134 offsets
->old
= offsets
->begin
;
2135 offsets
->switch_new_start
= 0;
2136 offsets
->switch_new_end
= 0;
2137 offsets
->switch_old_end
= 0;
2138 offsets
->pre_header_padding
= 0;
2140 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
2141 if ((int64_t) ctx
->tsc
== -EIO
)
2144 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
2145 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
2147 if (caa_unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
2148 offsets
->switch_new_start
= 1; /* For offsets->begin */
2150 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
2152 &offsets
->pre_header_padding
,
2155 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
2158 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
) +
2159 offsets
->size
> chan
->backend
.subbuf_size
)) {
2160 offsets
->switch_old_end
= 1; /* For offsets->old */
2161 offsets
->switch_new_start
= 1; /* For offsets->begin */
2164 if (caa_unlikely(offsets
->switch_new_start
)) {
2165 unsigned long sb_index
, commit_count
;
2166 struct commit_counters_cold
*cc_cold
;
2169 * We are typically not filling the previous buffer completely.
2171 if (caa_likely(offsets
->switch_old_end
))
2172 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
2173 offsets
->begin
= offsets
->begin
2174 + config
->cb
.subbuffer_header_size();
2175 /* Test new buffer integrity */
2176 sb_index
= subbuf_index(offsets
->begin
, chan
);
2178 * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
2179 * lib_ring_buffer_check_deliver() has the matching
2180 * memory barriers required around commit_cold cc_sb
2181 * updates to ensure reserve and commit counter updates
2182 * are not seen reordered when updated by another CPU.
2185 cc_cold
= shmp_index(handle
, buf
->commit_cold
, sb_index
);
2188 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
2189 /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
2191 if (caa_unlikely(offset_cmp
!= v_read(config
, &buf
->offset
))) {
2193 * The reserve counter have been concurrently updated
2194 * while we read the commit counter. This means the
2195 * commit counter we read might not match buf->offset
2196 * due to concurrent update. We therefore need to retry.
2200 reserve_commit_diff
=
2201 (buf_trunc(offsets
->begin
, chan
)
2202 >> chan
->backend
.num_subbuf_order
)
2203 - (commit_count
& chan
->commit_count_mask
);
2204 if (caa_likely(reserve_commit_diff
== 0)) {
2205 /* Next subbuffer not being written to. */
2206 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
2207 subbuf_trunc(offsets
->begin
, chan
)
2208 - subbuf_trunc((unsigned long)
2209 uatomic_read(&buf
->consumed
), chan
)
2210 >= chan
->backend
.buf_size
)) {
2211 unsigned long nr_lost
;
2213 if (handle_blocking_retry(&timeout_left_ms
))
2217 * We do not overwrite non consumed buffers
2218 * and we are full : record is lost.
2220 nr_lost
= v_read(config
, &buf
->records_lost_full
);
2221 v_inc(config
, &buf
->records_lost_full
);
2222 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
2223 DBG("%lu or more records lost in (%s:%d) (buffer full)\n",
2224 nr_lost
+ 1, chan
->backend
.name
,
2230 * Next subbuffer not being written to, and we
2231 * are either in overwrite mode or the buffer is
2232 * not full. It's safe to write in this new
2237 unsigned long nr_lost
;
2240 * Next subbuffer reserve offset does not match the
2241 * commit offset, and this did not involve update to the
2242 * reserve counter. Drop record in producer-consumer and
2243 * overwrite mode. Caused by either a writer OOPS or too
2244 * many nested writes over a reserve/commit pair.
2246 nr_lost
= v_read(config
, &buf
->records_lost_wrap
);
2247 v_inc(config
, &buf
->records_lost_wrap
);
2248 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
2249 DBG("%lu or more records lost in (%s:%d) (wrap-around)\n",
2250 nr_lost
+ 1, chan
->backend
.name
,
2256 config
->cb
.record_header_size(config
, chan
,
2258 &offsets
->pre_header_padding
,
2261 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
2264 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
)
2265 + offsets
->size
> chan
->backend
.subbuf_size
)) {
2266 unsigned long nr_lost
;
2269 * Record too big for subbuffers, report error, don't
2270 * complete the sub-buffer switch.
2272 nr_lost
= v_read(config
, &buf
->records_lost_big
);
2273 v_inc(config
, &buf
->records_lost_big
);
2274 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
2275 DBG("%lu or more records lost in (%s:%d) record size "
2276 " of %zu bytes is too large for buffer\n",
2277 nr_lost
+ 1, chan
->backend
.name
,
2278 buf
->backend
.cpu
, offsets
->size
);
2283 * We just made a successful buffer switch and the
2284 * record fits in the new subbuffer. Let's write.
2289 * Record fits in the current buffer and we are not on a switch
2290 * boundary. It's safe to write.
2293 offsets
->end
= offsets
->begin
+ offsets
->size
;
2295 if (caa_unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
2297 * The offset_end will fall at the very beginning of the next
2300 offsets
->switch_new_end
= 1; /* For offsets->begin */
2306 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
2307 * @ctx: ring buffer context.
2309 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
2310 * -EIO for other errors, else returns 0.
2311 * It will take care of sub-buffer switching.
2313 int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx
*ctx
,
2316 struct lttng_ust_lib_ring_buffer_channel
*chan
= ctx
->chan
;
2317 struct lttng_ust_shm_handle
*handle
= ctx
->chan
->handle
;
2318 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
2319 struct lttng_ust_lib_ring_buffer
*buf
;
2320 struct switch_offsets offsets
;
2323 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
2324 buf
= shmp(handle
, chan
->backend
.buf
[ctx
->reserve_cpu
].shmp
);
2326 buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
2334 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
2336 if (caa_unlikely(ret
))
2338 } while (caa_unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
2343 * Atomically update last_tsc. This update races against concurrent
2344 * atomic updates, but the race will always cause supplementary full TSC
2345 * records, never the opposite (missing a full TSC record when it would
2348 save_last_tsc(config
, buf
, ctx
->tsc
);
2351 * Push the reader if necessary
2353 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
2356 * Clear noref flag for this subbuffer.
2358 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2359 subbuf_index(offsets
.end
- 1, chan
),
2363 * Switch old subbuffer if needed.
2365 if (caa_unlikely(offsets
.switch_old_end
)) {
2366 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2367 subbuf_index(offsets
.old
- 1, chan
),
2369 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2373 * Populate new subbuffer.
2375 if (caa_unlikely(offsets
.switch_new_start
))
2376 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2378 if (caa_unlikely(offsets
.switch_new_end
))
2379 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2381 ctx
->slot_size
= offsets
.size
;
2382 ctx
->pre_offset
= offsets
.begin
;
2383 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
2388 void lib_ring_buffer_vmcore_check_deliver(const struct lttng_ust_lib_ring_buffer_config
*config
,
2389 struct lttng_ust_lib_ring_buffer
*buf
,
2390 unsigned long commit_count
,
2392 struct lttng_ust_shm_handle
*handle
)
2394 struct commit_counters_hot
*cc_hot
;
2396 if (config
->oops
!= RING_BUFFER_OOPS_CONSISTENCY
)
2398 cc_hot
= shmp_index(handle
, buf
->commit_hot
, idx
);
2401 v_set(config
, &cc_hot
->seq
, commit_count
);
2405 * The ring buffer can count events recorded and overwritten per buffer,
2406 * but it is disabled by default due to its performance overhead.
2408 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
2410 void deliver_count_events(const struct lttng_ust_lib_ring_buffer_config
*config
,
2411 struct lttng_ust_lib_ring_buffer
*buf
,
2413 struct lttng_ust_shm_handle
*handle
)
2415 v_add(config
, subbuffer_get_records_count(config
,
2416 &buf
->backend
, idx
, handle
),
2417 &buf
->records_count
);
2418 v_add(config
, subbuffer_count_records_overrun(config
,
2419 &buf
->backend
, idx
, handle
),
2420 &buf
->records_overrun
);
2422 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
2424 void deliver_count_events(const struct lttng_ust_lib_ring_buffer_config
*config
,
2425 struct lttng_ust_lib_ring_buffer
*buf
,
2427 struct lttng_ust_shm_handle
*handle
)
2430 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
2432 void lib_ring_buffer_check_deliver_slow(const struct lttng_ust_lib_ring_buffer_config
*config
,
2433 struct lttng_ust_lib_ring_buffer
*buf
,
2434 struct lttng_ust_lib_ring_buffer_channel
*chan
,
2435 unsigned long offset
,
2436 unsigned long commit_count
,
2438 struct lttng_ust_shm_handle
*handle
,
2441 unsigned long old_commit_count
= commit_count
2442 - chan
->backend
.subbuf_size
;
2443 struct commit_counters_cold
*cc_cold
;
2446 * If we succeeded at updating cc_sb below, we are the subbuffer
2447 * writer delivering the subbuffer. Deals with concurrent
2448 * updates of the "cc" value without adding a add_return atomic
2449 * operation to the fast path.
2451 * We are doing the delivery in two steps:
2452 * - First, we cmpxchg() cc_sb to the new value
2453 * old_commit_count + 1. This ensures that we are the only
2454 * subbuffer user successfully filling the subbuffer, but we
2455 * do _not_ set the cc_sb value to "commit_count" yet.
2456 * Therefore, other writers that would wrap around the ring
2457 * buffer and try to start writing to our subbuffer would
2458 * have to drop records, because it would appear as
2460 * We therefore have exclusive access to the subbuffer control
2461 * structures. This mutual exclusion with other writers is
2462 * crucially important to perform record overruns count in
2463 * flight recorder mode locklessly.
2464 * - When we are ready to release the subbuffer (either for
2465 * reading or for overrun by other writers), we simply set the
2466 * cc_sb value to "commit_count" and perform delivery.
2468 * The subbuffer size is least 2 bytes (minimum size: 1 page).
2469 * This guarantees that old_commit_count + 1 != commit_count.
2473 * Order prior updates to reserve count prior to the
2474 * commit_cold cc_sb update.
2477 cc_cold
= shmp_index(handle
, buf
->commit_cold
, idx
);
2480 if (caa_likely(v_cmpxchg(config
, &cc_cold
->cc_sb
,
2481 old_commit_count
, old_commit_count
+ 1)
2482 == old_commit_count
)) {
2486 * Start of exclusive subbuffer access. We are
2487 * guaranteed to be the last writer in this subbuffer
2488 * and any other writer trying to access this subbuffer
2489 * in this state is required to drop records.
2491 * We can read the ts_end for the current sub-buffer
2492 * which has been saved by the very last space
2493 * reservation for the current sub-buffer.
2495 * Order increment of commit counter before reading ts_end.
2498 ts_end
= shmp_index(handle
, buf
->ts_end
, idx
);
2501 deliver_count_events(config
, buf
, idx
, handle
);
2502 config
->cb
.buffer_end(buf
, *ts_end
, idx
,
2503 lib_ring_buffer_get_data_size(config
,
2510 * Increment the packet counter while we have exclusive
2513 subbuffer_inc_packet_count(config
, &buf
->backend
, idx
, handle
);
2516 * Set noref flag and offset for this subbuffer id.
2517 * Contains a memory barrier that ensures counter stores
2518 * are ordered before set noref and offset.
2520 lib_ring_buffer_set_noref_offset(config
, &buf
->backend
, idx
,
2521 buf_trunc_val(offset
, chan
), handle
);
2524 * Order set_noref and record counter updates before the
2525 * end of subbuffer exclusive access. Orders with
2526 * respect to writers coming into the subbuffer after
2527 * wrap around, and also order wrt concurrent readers.
2530 /* End of exclusive subbuffer access */
2531 v_set(config
, &cc_cold
->cc_sb
, commit_count
);
2533 * Order later updates to reserve count after
2534 * the commit cold cc_sb update.
2537 lib_ring_buffer_vmcore_check_deliver(config
, buf
,
2538 commit_count
, idx
, handle
);
2541 * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
2543 if (config
->wakeup
== RING_BUFFER_WAKEUP_BY_WRITER
2544 && uatomic_read(&buf
->active_readers
)
2545 && lib_ring_buffer_poll_deliver(config
, buf
, chan
, handle
)) {
2546 lib_ring_buffer_wakeup(buf
, handle
);
2552 * Force a read (imply TLS fixup for dlopen) of TLS variables.
2554 void lttng_fixup_ringbuffer_tls(void)
2556 asm volatile ("" : : "m" (URCU_TLS(lib_ring_buffer_nesting
)));
2559 void lib_ringbuffer_signal_init(void)
2565 * Block signal for entire process, so only our thread processes
2569 ret
= pthread_sigmask(SIG_BLOCK
, &mask
, NULL
);
2572 PERROR("pthread_sigmask");