2 * ring_buffer_frontend.c
4 * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; only
9 * version 2.1 of the License.
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 * Ring buffer wait-free buffer synchronization. Producer-consumer and flight
22 * recorder (overwrite) modes. See thesis:
24 * Desnoyers, Mathieu (2009), "Low-Impact Operating System Tracing", Ph.D.
25 * dissertation, Ecole Polytechnique de Montreal.
26 * http://www.lttng.org/pub/thesis/desnoyers-dissertation-2009-12.pdf
28 * - Algorithm presentation in Chapter 5:
29 * "Lockless Multi-Core High-Throughput Buffering".
30 * - Algorithm formal verification in Section 8.6:
31 * "Formal verification of LTTng"
34 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
36 * Inspired from LTT and RelayFS:
37 * Karim Yaghmour <karim@opersys.com>
38 * Tom Zanussi <zanussi@us.ibm.com>
39 * Bob Wisniewski <bob@watson.ibm.com>
41 * Bob Wisniewski <bob@watson.ibm.com>
43 * Buffer reader semantic :
46 * while buffer is not finalized and empty
48 * - if return value != 0, continue
49 * - splice one subbuffer worth of data to a pipe
50 * - splice the data from pipe to disk/network
56 #include <sys/types.h>
63 #include <urcu/compiler.h>
65 #include <urcu/tls-compat.h>
70 #include <lttng/ringbuffer-config.h>
76 #include "../liblttng-ust/compat.h" /* For ENODATA */
78 /* Print DBG() messages about events lost only every 1048576 hits */
79 #define DBG_PRINT_NR_LOST (1UL << 20)
81 #define LTTNG_UST_RB_SIG_FLUSH SIGRTMIN
82 #define LTTNG_UST_RB_SIG_READ SIGRTMIN + 1
83 #define LTTNG_UST_RB_SIG_TEARDOWN SIGRTMIN + 2
84 #define CLOCKID CLOCK_MONOTONIC
85 #define LTTNG_UST_RING_BUFFER_GET_RETRY 10
86 #define LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS 10
89 * Non-static to ensure the compiler does not optimize away the xor.
91 uint8_t lttng_crash_magic_xor
[] = RB_CRASH_DUMP_ABI_MAGIC_XOR
;
94 * Use POSIX SHM: shm_open(3) and shm_unlink(3).
95 * close(2) to close the fd returned by shm_open.
96 * shm_unlink releases the shared memory object name.
97 * ftruncate(2) sets the size of the memory object.
98 * mmap/munmap maps the shared memory obj to a virtual address in the
99 * calling proceess (should be done both in libust and consumer).
100 * See shm_overview(7) for details.
101 * Pass file descriptor returned by shm_open(3) to ltt-sessiond through
104 * Since we don't need to access the object using its name, we can
105 * immediately shm_unlink(3) it, and only keep the handle with its file
110 * Internal structure representing offsets to use at a sub-buffer switch.
112 struct switch_offsets
{
113 unsigned long begin
, end
, old
;
114 size_t pre_header_padding
, size
;
115 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
119 DEFINE_URCU_TLS(unsigned int, lib_ring_buffer_nesting
);
122 * wakeup_fd_mutex protects wakeup fd use by timer from concurrent
125 static pthread_mutex_t wakeup_fd_mutex
= PTHREAD_MUTEX_INITIALIZER
;
128 void lib_ring_buffer_print_errors(struct channel
*chan
,
129 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
130 struct lttng_ust_shm_handle
*handle
);
133 * Handle timer teardown race wrt memory free of private data by
134 * ring buffer signals are handled by a single thread, which permits
135 * a synchronization point between handling of each signal.
136 * Protected by the lock within the structure.
138 struct timer_signal_data
{
139 pthread_t tid
; /* thread id managing signals */
142 pthread_mutex_t lock
;
145 static struct timer_signal_data timer_signal
= {
149 .lock
= PTHREAD_MUTEX_INITIALIZER
,
153 * lib_ring_buffer_reset - Reset ring buffer to initial values.
156 * Effectively empty the ring buffer. Should be called when the buffer is not
157 * used for writing. The ring buffer can be opened for reading, but the reader
158 * should not be using the iterator concurrently with reset. The previous
159 * current iterator record is reset.
161 void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer
*buf
,
162 struct lttng_ust_shm_handle
*handle
)
164 struct channel
*chan
;
165 const struct lttng_ust_lib_ring_buffer_config
*config
;
168 chan
= shmp(handle
, buf
->backend
.chan
);
171 config
= &chan
->backend
.config
;
173 * Reset iterator first. It will put the subbuffer if it currently holds
176 v_set(config
, &buf
->offset
, 0);
177 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
178 struct commit_counters_hot
*cc_hot
;
179 struct commit_counters_cold
*cc_cold
;
181 cc_hot
= shmp_index(handle
, buf
->commit_hot
, i
);
184 cc_cold
= shmp_index(handle
, buf
->commit_cold
, i
);
187 v_set(config
, &cc_hot
->cc
, 0);
189 v_set(config
, &cc_hot
->seq
, 0);
190 v_set(config
, &cc_cold
->cc_sb
, 0);
192 uatomic_set(&buf
->consumed
, 0);
193 uatomic_set(&buf
->record_disabled
, 0);
194 v_set(config
, &buf
->last_tsc
, 0);
195 lib_ring_buffer_backend_reset(&buf
->backend
, handle
);
196 /* Don't reset number of active readers */
197 v_set(config
, &buf
->records_lost_full
, 0);
198 v_set(config
, &buf
->records_lost_wrap
, 0);
199 v_set(config
, &buf
->records_lost_big
, 0);
200 v_set(config
, &buf
->records_count
, 0);
201 v_set(config
, &buf
->records_overrun
, 0);
206 * channel_reset - Reset channel to initial values.
209 * Effectively empty the channel. Should be called when the channel is not used
210 * for writing. The channel can be opened for reading, but the reader should not
211 * be using the iterator concurrently with reset. The previous current iterator
214 void channel_reset(struct channel
*chan
)
217 * Reset iterators first. Will put the subbuffer if held for reading.
219 uatomic_set(&chan
->record_disabled
, 0);
220 /* Don't reset commit_count_mask, still valid */
221 channel_backend_reset(&chan
->backend
);
222 /* Don't reset switch/read timer interval */
223 /* Don't reset notifiers and notifier enable bits */
224 /* Don't reset reader reference count */
228 void init_crash_abi(const struct lttng_ust_lib_ring_buffer_config
*config
,
229 struct lttng_crash_abi
*crash_abi
,
230 struct lttng_ust_lib_ring_buffer
*buf
,
231 struct channel_backend
*chanb
,
232 struct shm_object
*shmobj
,
233 struct lttng_ust_shm_handle
*handle
)
237 for (i
= 0; i
< RB_CRASH_DUMP_ABI_MAGIC_LEN
; i
++)
238 crash_abi
->magic
[i
] = lttng_crash_magic_xor
[i
] ^ 0xFF;
239 crash_abi
->mmap_length
= shmobj
->memory_map_size
;
240 crash_abi
->endian
= RB_CRASH_ENDIAN
;
241 crash_abi
->major
= RB_CRASH_DUMP_ABI_MAJOR
;
242 crash_abi
->minor
= RB_CRASH_DUMP_ABI_MINOR
;
243 crash_abi
->word_size
= sizeof(unsigned long);
244 crash_abi
->layout_type
= LTTNG_CRASH_TYPE_UST
;
246 /* Offset of fields */
247 crash_abi
->offset
.prod_offset
=
248 (uint32_t) ((char *) &buf
->offset
- (char *) buf
);
249 crash_abi
->offset
.consumed_offset
=
250 (uint32_t) ((char *) &buf
->consumed
- (char *) buf
);
251 crash_abi
->offset
.commit_hot_array
=
252 (uint32_t) ((char *) shmp(handle
, buf
->commit_hot
) - (char *) buf
);
253 crash_abi
->offset
.commit_hot_seq
=
254 offsetof(struct commit_counters_hot
, seq
);
255 crash_abi
->offset
.buf_wsb_array
=
256 (uint32_t) ((char *) shmp(handle
, buf
->backend
.buf_wsb
) - (char *) buf
);
257 crash_abi
->offset
.buf_wsb_id
=
258 offsetof(struct lttng_ust_lib_ring_buffer_backend_subbuffer
, id
);
259 crash_abi
->offset
.sb_array
=
260 (uint32_t) ((char *) shmp(handle
, buf
->backend
.array
) - (char *) buf
);
261 crash_abi
->offset
.sb_array_shmp_offset
=
262 offsetof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp
,
264 crash_abi
->offset
.sb_backend_p_offset
=
265 offsetof(struct lttng_ust_lib_ring_buffer_backend_pages
,
269 crash_abi
->length
.prod_offset
= sizeof(buf
->offset
);
270 crash_abi
->length
.consumed_offset
= sizeof(buf
->consumed
);
271 crash_abi
->length
.commit_hot_seq
=
272 sizeof(((struct commit_counters_hot
*) NULL
)->seq
);
273 crash_abi
->length
.buf_wsb_id
=
274 sizeof(((struct lttng_ust_lib_ring_buffer_backend_subbuffer
*) NULL
)->id
);
275 crash_abi
->length
.sb_array_shmp_offset
=
276 sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*) NULL
)->shmp
._ref
.offset
);
277 crash_abi
->length
.sb_backend_p_offset
=
278 sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages
*) NULL
)->p
._ref
.offset
);
281 crash_abi
->stride
.commit_hot_array
=
282 sizeof(struct commit_counters_hot
);
283 crash_abi
->stride
.buf_wsb_array
=
284 sizeof(struct lttng_ust_lib_ring_buffer_backend_subbuffer
);
285 crash_abi
->stride
.sb_array
=
286 sizeof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp
);
288 /* Buffer constants */
289 crash_abi
->buf_size
= chanb
->buf_size
;
290 crash_abi
->subbuf_size
= chanb
->subbuf_size
;
291 crash_abi
->num_subbuf
= chanb
->num_subbuf
;
292 crash_abi
->mode
= (uint32_t) chanb
->config
.mode
;
294 if (config
->cb
.content_size_field
) {
295 size_t offset
, length
;
297 config
->cb
.content_size_field(config
, &offset
, &length
);
298 crash_abi
->offset
.content_size
= offset
;
299 crash_abi
->length
.content_size
= length
;
301 crash_abi
->offset
.content_size
= 0;
302 crash_abi
->length
.content_size
= 0;
304 if (config
->cb
.packet_size_field
) {
305 size_t offset
, length
;
307 config
->cb
.packet_size_field(config
, &offset
, &length
);
308 crash_abi
->offset
.packet_size
= offset
;
309 crash_abi
->length
.packet_size
= length
;
311 crash_abi
->offset
.packet_size
= 0;
312 crash_abi
->length
.packet_size
= 0;
317 * Must be called under cpu hotplug protection.
319 int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer
*buf
,
320 struct channel_backend
*chanb
, int cpu
,
321 struct lttng_ust_shm_handle
*handle
,
322 struct shm_object
*shmobj
)
324 const struct lttng_ust_lib_ring_buffer_config
*config
= &chanb
->config
;
325 struct channel
*chan
= caa_container_of(chanb
, struct channel
, backend
);
326 struct lttng_ust_lib_ring_buffer_backend_subbuffer
*wsb
;
327 struct channel
*shmp_chan
;
328 struct commit_counters_hot
*cc_hot
;
329 void *priv
= channel_get_private(chan
);
330 size_t subbuf_header_size
;
334 /* Test for cpu hotplug */
335 if (buf
->backend
.allocated
)
338 align_shm(shmobj
, __alignof__(struct commit_counters_hot
));
339 set_shmp(buf
->commit_hot
,
341 sizeof(struct commit_counters_hot
) * chan
->backend
.num_subbuf
));
342 if (!shmp(handle
, buf
->commit_hot
)) {
346 align_shm(shmobj
, __alignof__(struct commit_counters_cold
));
347 set_shmp(buf
->commit_cold
,
349 sizeof(struct commit_counters_cold
) * chan
->backend
.num_subbuf
));
350 if (!shmp(handle
, buf
->commit_cold
)) {
355 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
,
356 cpu
, handle
, shmobj
);
362 * Write the subbuffer header for first subbuffer so we know the total
363 * duration of data gathering.
365 subbuf_header_size
= config
->cb
.subbuffer_header_size();
366 v_set(config
, &buf
->offset
, subbuf_header_size
);
367 wsb
= shmp_index(handle
, buf
->backend
.buf_wsb
, 0);
372 subbuffer_id_clear_noref(config
, &wsb
->id
);
373 shmp_chan
= shmp(handle
, buf
->backend
.chan
);
378 tsc
= config
->cb
.ring_buffer_clock_read(shmp_chan
);
379 config
->cb
.buffer_begin(buf
, tsc
, 0, handle
);
380 cc_hot
= shmp_index(handle
, buf
->commit_hot
, 0);
385 v_add(config
, subbuf_header_size
, &cc_hot
->cc
);
386 v_add(config
, subbuf_header_size
, &cc_hot
->seq
);
388 if (config
->cb
.buffer_create
) {
389 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
, handle
);
394 init_crash_abi(config
, &buf
->crash_abi
, buf
, chanb
, shmobj
, handle
);
396 buf
->backend
.allocated
= 1;
401 /* commit_cold will be freed by shm teardown */
403 /* commit_hot will be freed by shm teardown */
409 void lib_ring_buffer_channel_switch_timer(int sig
, siginfo_t
*si
, void *uc
)
411 const struct lttng_ust_lib_ring_buffer_config
*config
;
412 struct lttng_ust_shm_handle
*handle
;
413 struct channel
*chan
;
416 assert(CMM_LOAD_SHARED(timer_signal
.tid
) == pthread_self());
418 chan
= si
->si_value
.sival_ptr
;
419 handle
= chan
->handle
;
420 config
= &chan
->backend
.config
;
422 DBG("Switch timer for channel %p\n", chan
);
425 * Only flush buffers periodically if readers are active.
427 pthread_mutex_lock(&wakeup_fd_mutex
);
428 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
429 for_each_possible_cpu(cpu
) {
430 struct lttng_ust_lib_ring_buffer
*buf
=
431 shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
435 if (uatomic_read(&buf
->active_readers
))
436 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
,
440 struct lttng_ust_lib_ring_buffer
*buf
=
441 shmp(handle
, chan
->backend
.buf
[0].shmp
);
445 if (uatomic_read(&buf
->active_readers
))
446 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
,
450 pthread_mutex_unlock(&wakeup_fd_mutex
);
455 int lib_ring_buffer_poll_deliver(const struct lttng_ust_lib_ring_buffer_config
*config
,
456 struct lttng_ust_lib_ring_buffer
*buf
,
457 struct channel
*chan
,
458 struct lttng_ust_shm_handle
*handle
)
460 unsigned long consumed_old
, consumed_idx
, commit_count
, write_offset
;
461 struct commit_counters_cold
*cc_cold
;
463 consumed_old
= uatomic_read(&buf
->consumed
);
464 consumed_idx
= subbuf_index(consumed_old
, chan
);
465 cc_cold
= shmp_index(handle
, buf
->commit_cold
, consumed_idx
);
468 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
470 * No memory barrier here, since we are only interested
471 * in a statistically correct polling result. The next poll will
472 * get the data is we are racing. The mb() that ensures correct
473 * memory order is in get_subbuf.
475 write_offset
= v_read(config
, &buf
->offset
);
478 * Check that the subbuffer we are trying to consume has been
479 * already fully committed.
482 if (((commit_count
- chan
->backend
.subbuf_size
)
483 & chan
->commit_count_mask
)
484 - (buf_trunc(consumed_old
, chan
)
485 >> chan
->backend
.num_subbuf_order
)
490 * Check that we are not about to read the same subbuffer in
491 * which the writer head is.
493 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_old
, chan
)
501 void lib_ring_buffer_wakeup(struct lttng_ust_lib_ring_buffer
*buf
,
502 struct lttng_ust_shm_handle
*handle
)
504 int wakeup_fd
= shm_get_wakeup_fd(handle
, &buf
->self
._ref
);
505 sigset_t sigpipe_set
, pending_set
, old_set
;
506 int ret
, sigpipe_was_pending
= 0;
512 * Wake-up the other end by writing a null byte in the pipe
513 * (non-blocking). Important note: Because writing into the
514 * pipe is non-blocking (and therefore we allow dropping wakeup
515 * data, as long as there is wakeup data present in the pipe
516 * buffer to wake up the consumer), the consumer should perform
517 * the following sequence for waiting:
518 * 1) empty the pipe (reads).
519 * 2) check if there is data in the buffer.
520 * 3) wait on the pipe (poll).
522 * Discard the SIGPIPE from write(), not disturbing any SIGPIPE
523 * that might be already pending. If a bogus SIGPIPE is sent to
524 * the entire process concurrently by a malicious user, it may
525 * be simply discarded.
527 ret
= sigemptyset(&pending_set
);
530 * sigpending returns the mask of signals that are _both_
531 * blocked for the thread _and_ pending for either the thread or
532 * the entire process.
534 ret
= sigpending(&pending_set
);
536 sigpipe_was_pending
= sigismember(&pending_set
, SIGPIPE
);
538 * If sigpipe was pending, it means it was already blocked, so
539 * no need to block it.
541 if (!sigpipe_was_pending
) {
542 ret
= sigemptyset(&sigpipe_set
);
544 ret
= sigaddset(&sigpipe_set
, SIGPIPE
);
546 ret
= pthread_sigmask(SIG_BLOCK
, &sigpipe_set
, &old_set
);
550 ret
= write(wakeup_fd
, "", 1);
551 } while (ret
== -1L && errno
== EINTR
);
552 if (ret
== -1L && errno
== EPIPE
&& !sigpipe_was_pending
) {
553 struct timespec timeout
= { 0, 0 };
555 ret
= sigtimedwait(&sigpipe_set
, NULL
,
557 } while (ret
== -1L && errno
== EINTR
);
559 if (!sigpipe_was_pending
) {
560 ret
= pthread_sigmask(SIG_SETMASK
, &old_set
, NULL
);
566 void lib_ring_buffer_channel_do_read(struct channel
*chan
)
568 const struct lttng_ust_lib_ring_buffer_config
*config
;
569 struct lttng_ust_shm_handle
*handle
;
572 handle
= chan
->handle
;
573 config
= &chan
->backend
.config
;
576 * Only flush buffers periodically if readers are active.
578 pthread_mutex_lock(&wakeup_fd_mutex
);
579 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
580 for_each_possible_cpu(cpu
) {
581 struct lttng_ust_lib_ring_buffer
*buf
=
582 shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
586 if (uatomic_read(&buf
->active_readers
)
587 && lib_ring_buffer_poll_deliver(config
, buf
,
589 lib_ring_buffer_wakeup(buf
, handle
);
593 struct lttng_ust_lib_ring_buffer
*buf
=
594 shmp(handle
, chan
->backend
.buf
[0].shmp
);
598 if (uatomic_read(&buf
->active_readers
)
599 && lib_ring_buffer_poll_deliver(config
, buf
,
601 lib_ring_buffer_wakeup(buf
, handle
);
605 pthread_mutex_unlock(&wakeup_fd_mutex
);
609 void lib_ring_buffer_channel_read_timer(int sig
, siginfo_t
*si
, void *uc
)
611 struct channel
*chan
;
613 assert(CMM_LOAD_SHARED(timer_signal
.tid
) == pthread_self());
614 chan
= si
->si_value
.sival_ptr
;
615 DBG("Read timer for channel %p\n", chan
);
616 lib_ring_buffer_channel_do_read(chan
);
621 void rb_setmask(sigset_t
*mask
)
625 ret
= sigemptyset(mask
);
627 PERROR("sigemptyset");
629 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_FLUSH
);
633 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_READ
);
637 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_TEARDOWN
);
644 void *sig_thread(void *arg
)
650 /* Only self thread will receive signal mask. */
652 CMM_STORE_SHARED(timer_signal
.tid
, pthread_self());
655 signr
= sigwaitinfo(&mask
, &info
);
658 PERROR("sigwaitinfo");
661 if (signr
== LTTNG_UST_RB_SIG_FLUSH
) {
662 lib_ring_buffer_channel_switch_timer(info
.si_signo
,
664 } else if (signr
== LTTNG_UST_RB_SIG_READ
) {
665 lib_ring_buffer_channel_read_timer(info
.si_signo
,
667 } else if (signr
== LTTNG_UST_RB_SIG_TEARDOWN
) {
669 CMM_STORE_SHARED(timer_signal
.qs_done
, 1);
672 ERR("Unexptected signal %d\n", info
.si_signo
);
679 * Ensure only a single thread listens on the timer signal.
682 void lib_ring_buffer_setup_timer_thread(void)
687 pthread_mutex_lock(&timer_signal
.lock
);
688 if (timer_signal
.setup_done
)
691 ret
= pthread_create(&thread
, NULL
, &sig_thread
, NULL
);
694 PERROR("pthread_create");
696 ret
= pthread_detach(thread
);
699 PERROR("pthread_detach");
701 timer_signal
.setup_done
= 1;
703 pthread_mutex_unlock(&timer_signal
.lock
);
707 * Wait for signal-handling thread quiescent state.
710 void lib_ring_buffer_wait_signal_thread_qs(unsigned int signr
)
712 sigset_t pending_set
;
716 * We need to be the only thread interacting with the thread
717 * that manages signals for teardown synchronization.
719 pthread_mutex_lock(&timer_signal
.lock
);
722 * Ensure we don't have any signal queued for this channel.
725 ret
= sigemptyset(&pending_set
);
727 PERROR("sigemptyset");
729 ret
= sigpending(&pending_set
);
731 PERROR("sigpending");
733 if (!sigismember(&pending_set
, signr
))
739 * From this point, no new signal handler will be fired that
740 * would try to access "chan". However, we still need to wait
741 * for any currently executing handler to complete.
744 CMM_STORE_SHARED(timer_signal
.qs_done
, 0);
748 * Kill with LTTNG_UST_RB_SIG_TEARDOWN, so signal management
751 kill(getpid(), LTTNG_UST_RB_SIG_TEARDOWN
);
753 while (!CMM_LOAD_SHARED(timer_signal
.qs_done
))
757 pthread_mutex_unlock(&timer_signal
.lock
);
761 void lib_ring_buffer_channel_switch_timer_start(struct channel
*chan
)
764 struct itimerspec its
;
767 if (!chan
->switch_timer_interval
|| chan
->switch_timer_enabled
)
770 chan
->switch_timer_enabled
= 1;
772 lib_ring_buffer_setup_timer_thread();
774 sev
.sigev_notify
= SIGEV_SIGNAL
;
775 sev
.sigev_signo
= LTTNG_UST_RB_SIG_FLUSH
;
776 sev
.sigev_value
.sival_ptr
= chan
;
777 ret
= timer_create(CLOCKID
, &sev
, &chan
->switch_timer
);
779 PERROR("timer_create");
782 its
.it_value
.tv_sec
= chan
->switch_timer_interval
/ 1000000;
783 its
.it_value
.tv_nsec
= (chan
->switch_timer_interval
% 1000000) * 1000;
784 its
.it_interval
.tv_sec
= its
.it_value
.tv_sec
;
785 its
.it_interval
.tv_nsec
= its
.it_value
.tv_nsec
;
787 ret
= timer_settime(chan
->switch_timer
, 0, &its
, NULL
);
789 PERROR("timer_settime");
794 void lib_ring_buffer_channel_switch_timer_stop(struct channel
*chan
)
798 if (!chan
->switch_timer_interval
|| !chan
->switch_timer_enabled
)
801 ret
= timer_delete(chan
->switch_timer
);
803 PERROR("timer_delete");
806 lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_FLUSH
);
808 chan
->switch_timer
= 0;
809 chan
->switch_timer_enabled
= 0;
813 void lib_ring_buffer_channel_read_timer_start(struct channel
*chan
)
815 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
817 struct itimerspec its
;
820 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
821 || !chan
->read_timer_interval
|| chan
->read_timer_enabled
)
824 chan
->read_timer_enabled
= 1;
826 lib_ring_buffer_setup_timer_thread();
828 sev
.sigev_notify
= SIGEV_SIGNAL
;
829 sev
.sigev_signo
= LTTNG_UST_RB_SIG_READ
;
830 sev
.sigev_value
.sival_ptr
= chan
;
831 ret
= timer_create(CLOCKID
, &sev
, &chan
->read_timer
);
833 PERROR("timer_create");
836 its
.it_value
.tv_sec
= chan
->read_timer_interval
/ 1000000;
837 its
.it_value
.tv_nsec
= (chan
->read_timer_interval
% 1000000) * 1000;
838 its
.it_interval
.tv_sec
= its
.it_value
.tv_sec
;
839 its
.it_interval
.tv_nsec
= its
.it_value
.tv_nsec
;
841 ret
= timer_settime(chan
->read_timer
, 0, &its
, NULL
);
843 PERROR("timer_settime");
848 void lib_ring_buffer_channel_read_timer_stop(struct channel
*chan
)
850 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
853 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
854 || !chan
->read_timer_interval
|| !chan
->read_timer_enabled
)
857 ret
= timer_delete(chan
->read_timer
);
859 PERROR("timer_delete");
863 * do one more check to catch data that has been written in the last
866 lib_ring_buffer_channel_do_read(chan
);
868 lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_READ
);
870 chan
->read_timer
= 0;
871 chan
->read_timer_enabled
= 0;
874 static void channel_unregister_notifiers(struct channel
*chan
,
875 struct lttng_ust_shm_handle
*handle
)
877 lib_ring_buffer_channel_switch_timer_stop(chan
);
878 lib_ring_buffer_channel_read_timer_stop(chan
);
881 static void channel_print_errors(struct channel
*chan
,
882 struct lttng_ust_shm_handle
*handle
)
884 const struct lttng_ust_lib_ring_buffer_config
*config
=
885 &chan
->backend
.config
;
888 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
889 for_each_possible_cpu(cpu
) {
890 struct lttng_ust_lib_ring_buffer
*buf
=
891 shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
893 lib_ring_buffer_print_errors(chan
, buf
, cpu
, handle
);
896 struct lttng_ust_lib_ring_buffer
*buf
=
897 shmp(handle
, chan
->backend
.buf
[0].shmp
);
900 lib_ring_buffer_print_errors(chan
, buf
, -1, handle
);
904 static void channel_free(struct channel
*chan
,
905 struct lttng_ust_shm_handle
*handle
,
908 channel_backend_free(&chan
->backend
, handle
);
909 /* chan is freed by shm teardown */
910 shm_object_table_destroy(handle
->table
, consumer
);
915 * channel_create - Create channel.
916 * @config: ring buffer instance configuration
917 * @name: name of the channel
918 * @priv_data: ring buffer client private data area pointer (output)
919 * @priv_data_size: length, in bytes, of the private data area.
920 * @priv_data_init: initialization data for private data.
921 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
922 * address mapping. It is used only by RING_BUFFER_STATIC
923 * configuration. It can be set to NULL for other backends.
924 * @subbuf_size: subbuffer size
925 * @num_subbuf: number of subbuffers
926 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
927 * padding to let readers get those sub-buffers.
928 * Used for live streaming.
929 * @read_timer_interval: Time interval (in us) to wake up pending readers.
930 * @stream_fds: array of stream file descriptors.
931 * @nr_stream_fds: number of file descriptors in array.
934 * Returns NULL on failure.
936 struct lttng_ust_shm_handle
*channel_create(const struct lttng_ust_lib_ring_buffer_config
*config
,
939 size_t priv_data_align
,
940 size_t priv_data_size
,
941 void *priv_data_init
,
942 void *buf_addr
, size_t subbuf_size
,
943 size_t num_subbuf
, unsigned int switch_timer_interval
,
944 unsigned int read_timer_interval
,
945 const int *stream_fds
, int nr_stream_fds
)
948 size_t shmsize
, chansize
;
949 struct channel
*chan
;
950 struct lttng_ust_shm_handle
*handle
;
951 struct shm_object
*shmobj
;
952 unsigned int nr_streams
;
954 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
955 nr_streams
= num_possible_cpus();
959 if (nr_stream_fds
!= nr_streams
)
962 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
963 read_timer_interval
))
966 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
970 /* Allocate table for channel + per-cpu buffers */
971 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
973 goto error_table_alloc
;
975 /* Calculate the shm allocation layout */
976 shmsize
= sizeof(struct channel
);
977 shmsize
+= offset_align(shmsize
, __alignof__(struct lttng_ust_lib_ring_buffer_shmp
));
978 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
) * nr_streams
;
981 shmsize
+= offset_align(shmsize
, priv_data_align
);
982 shmsize
+= priv_data_size
;
984 /* Allocate normal memory for channel (not shared) */
985 shmobj
= shm_object_table_alloc(handle
->table
, shmsize
, SHM_OBJECT_MEM
,
989 /* struct channel is at object 0, offset 0 (hardcoded) */
990 set_shmp(handle
->chan
, zalloc_shm(shmobj
, chansize
));
991 assert(handle
->chan
._ref
.index
== 0);
992 assert(handle
->chan
._ref
.offset
== 0);
993 chan
= shmp(handle
, handle
->chan
);
996 chan
->nr_streams
= nr_streams
;
998 /* space for private data */
999 if (priv_data_size
) {
1000 DECLARE_SHMP(void, priv_data_alloc
);
1002 align_shm(shmobj
, priv_data_align
);
1003 chan
->priv_data_offset
= shmobj
->allocated_len
;
1004 set_shmp(priv_data_alloc
, zalloc_shm(shmobj
, priv_data_size
));
1005 if (!shmp(handle
, priv_data_alloc
))
1007 *priv_data
= channel_get_private(chan
);
1008 memcpy(*priv_data
, priv_data_init
, priv_data_size
);
1010 chan
->priv_data_offset
= -1;
1015 ret
= channel_backend_init(&chan
->backend
, name
, config
,
1016 subbuf_size
, num_subbuf
, handle
,
1019 goto error_backend_init
;
1021 chan
->handle
= handle
;
1022 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
1024 chan
->switch_timer_interval
= switch_timer_interval
;
1025 chan
->read_timer_interval
= read_timer_interval
;
1026 lib_ring_buffer_channel_switch_timer_start(chan
);
1027 lib_ring_buffer_channel_read_timer_start(chan
);
1033 shm_object_table_destroy(handle
->table
, 1);
1039 struct lttng_ust_shm_handle
*channel_handle_create(void *data
,
1040 uint64_t memory_map_size
,
1043 struct lttng_ust_shm_handle
*handle
;
1044 struct shm_object
*object
;
1046 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
1050 /* Allocate table for channel + per-cpu buffers */
1051 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
1053 goto error_table_alloc
;
1054 /* Add channel object */
1055 object
= shm_object_table_append_mem(handle
->table
, data
,
1056 memory_map_size
, wakeup_fd
);
1058 goto error_table_object
;
1059 /* struct channel is at object 0, offset 0 (hardcoded) */
1060 handle
->chan
._ref
.index
= 0;
1061 handle
->chan
._ref
.offset
= 0;
1065 shm_object_table_destroy(handle
->table
, 0);
1071 int channel_handle_add_stream(struct lttng_ust_shm_handle
*handle
,
1072 int shm_fd
, int wakeup_fd
, uint32_t stream_nr
,
1073 uint64_t memory_map_size
)
1075 struct shm_object
*object
;
1077 /* Add stream object */
1078 object
= shm_object_table_append_shm(handle
->table
,
1079 shm_fd
, wakeup_fd
, stream_nr
,
1086 unsigned int channel_handle_get_nr_streams(struct lttng_ust_shm_handle
*handle
)
1088 assert(handle
->table
);
1089 return handle
->table
->allocated_len
- 1;
1093 void channel_release(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
1096 channel_free(chan
, handle
, consumer
);
1100 * channel_destroy - Finalize, wait for q.s. and destroy channel.
1101 * @chan: channel to destroy
1103 * Holds cpu hotplug.
1104 * Call "destroy" callback, finalize channels, decrement the channel
1105 * reference count. Note that when readers have completed data
1106 * consumption of finalized channels, get_subbuf() will return -ENODATA.
1107 * They should release their handle at that point.
1109 void channel_destroy(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
1114 * Note: the consumer takes care of finalizing and
1115 * switching the buffers.
1117 channel_unregister_notifiers(chan
, handle
);
1119 * The consumer prints errors.
1121 channel_print_errors(chan
, handle
);
1125 * sessiond/consumer are keeping a reference on the shm file
1126 * descriptor directly. No need to refcount.
1128 channel_release(chan
, handle
, consumer
);
1132 struct lttng_ust_lib_ring_buffer
*channel_get_ring_buffer(
1133 const struct lttng_ust_lib_ring_buffer_config
*config
,
1134 struct channel
*chan
, int cpu
,
1135 struct lttng_ust_shm_handle
*handle
,
1136 int *shm_fd
, int *wait_fd
,
1138 uint64_t *memory_map_size
)
1140 struct shm_ref
*ref
;
1142 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1145 if (cpu
>= num_possible_cpus())
1148 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1149 *shm_fd
= shm_get_shm_fd(handle
, ref
);
1150 *wait_fd
= shm_get_wait_fd(handle
, ref
);
1151 *wakeup_fd
= shm_get_wakeup_fd(handle
, ref
);
1152 if (shm_get_shm_size(handle
, ref
, memory_map_size
))
1154 return shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
1157 int ring_buffer_channel_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1158 struct channel
*chan
,
1159 struct lttng_ust_shm_handle
*handle
)
1161 struct shm_ref
*ref
;
1163 ref
= &handle
->chan
._ref
;
1164 return shm_close_wait_fd(handle
, ref
);
1167 int ring_buffer_channel_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1168 struct channel
*chan
,
1169 struct lttng_ust_shm_handle
*handle
)
1171 struct shm_ref
*ref
;
1173 ref
= &handle
->chan
._ref
;
1174 return shm_close_wakeup_fd(handle
, ref
);
1177 int ring_buffer_stream_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1178 struct channel
*chan
,
1179 struct lttng_ust_shm_handle
*handle
,
1182 struct shm_ref
*ref
;
1184 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1187 if (cpu
>= num_possible_cpus())
1190 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1191 return shm_close_wait_fd(handle
, ref
);
1194 int ring_buffer_stream_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1195 struct channel
*chan
,
1196 struct lttng_ust_shm_handle
*handle
,
1199 struct shm_ref
*ref
;
1202 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1205 if (cpu
>= num_possible_cpus())
1208 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1209 pthread_mutex_lock(&wakeup_fd_mutex
);
1210 ret
= shm_close_wakeup_fd(handle
, ref
);
1211 pthread_mutex_unlock(&wakeup_fd_mutex
);
1215 int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer
*buf
,
1216 struct lttng_ust_shm_handle
*handle
)
1218 if (uatomic_cmpxchg(&buf
->active_readers
, 0, 1) != 0)
1224 void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer
*buf
,
1225 struct lttng_ust_shm_handle
*handle
)
1227 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1231 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1233 uatomic_dec(&buf
->active_readers
);
1237 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
1239 * @consumed: consumed count indicating the position where to read
1240 * @produced: produced count, indicates position when to stop reading
1242 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1243 * data to read at consumed position, or 0 if the get operation succeeds.
1246 int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer
*buf
,
1247 unsigned long *consumed
, unsigned long *produced
,
1248 struct lttng_ust_shm_handle
*handle
)
1250 struct channel
*chan
;
1251 const struct lttng_ust_lib_ring_buffer_config
*config
;
1252 unsigned long consumed_cur
, write_offset
;
1255 chan
= shmp(handle
, buf
->backend
.chan
);
1258 config
= &chan
->backend
.config
;
1259 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
1261 * Read finalized before counters.
1264 consumed_cur
= uatomic_read(&buf
->consumed
);
1266 * No need to issue a memory barrier between consumed count read and
1267 * write offset read, because consumed count can only change
1268 * concurrently in overwrite mode, and we keep a sequence counter
1269 * identifier derived from the write offset to check we are getting
1270 * the same sub-buffer we are expecting (the sub-buffers are atomically
1271 * "tagged" upon writes, tags are checked upon read).
1273 write_offset
= v_read(config
, &buf
->offset
);
1276 * Check that we are not about to read the same subbuffer in
1277 * which the writer head is.
1279 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
1283 *consumed
= consumed_cur
;
1284 *produced
= subbuf_trunc(write_offset
, chan
);
1290 * The memory barriers __wait_event()/wake_up_interruptible() take care
1291 * of "raw_spin_is_locked" memory ordering.
1300 * lib_ring_buffer_move_consumer - move consumed counter forward
1302 * @consumed_new: new consumed count value
1304 void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer
*buf
,
1305 unsigned long consumed_new
,
1306 struct lttng_ust_shm_handle
*handle
)
1308 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1309 struct channel
*chan
;
1310 unsigned long consumed
;
1312 chan
= shmp(handle
, bufb
->chan
);
1315 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1318 * Only push the consumed value forward.
1319 * If the consumed cmpxchg fails, this is because we have been pushed by
1320 * the writer in flight recorder mode.
1322 consumed
= uatomic_read(&buf
->consumed
);
1323 while ((long) consumed
- (long) consumed_new
< 0)
1324 consumed
= uatomic_cmpxchg(&buf
->consumed
, consumed
,
1329 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
1331 * @consumed: consumed count indicating the position where to read
1333 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1334 * data to read at consumed position, or 0 if the get operation succeeds.
1336 int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
1337 unsigned long consumed
,
1338 struct lttng_ust_shm_handle
*handle
)
1340 struct channel
*chan
;
1341 const struct lttng_ust_lib_ring_buffer_config
*config
;
1342 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
1343 int ret
, finalized
, nr_retry
= LTTNG_UST_RING_BUFFER_GET_RETRY
;
1344 struct commit_counters_cold
*cc_cold
;
1346 chan
= shmp(handle
, buf
->backend
.chan
);
1349 config
= &chan
->backend
.config
;
1351 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
1353 * Read finalized before counters.
1356 consumed_cur
= uatomic_read(&buf
->consumed
);
1357 consumed_idx
= subbuf_index(consumed
, chan
);
1358 cc_cold
= shmp_index(handle
, buf
->commit_cold
, consumed_idx
);
1361 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
1363 * Make sure we read the commit count before reading the buffer
1364 * data and the write offset. Correct consumed offset ordering
1365 * wrt commit count is insured by the use of cmpxchg to update
1366 * the consumed offset.
1369 * Local rmb to match the remote wmb to read the commit count
1370 * before the buffer data and the write offset.
1374 write_offset
= v_read(config
, &buf
->offset
);
1377 * Check that the buffer we are getting is after or at consumed_cur
1380 if ((long) subbuf_trunc(consumed
, chan
)
1381 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
1385 * Check that the subbuffer we are trying to consume has been
1386 * already fully committed. There are a few causes that can make
1387 * this unavailability situation occur:
1389 * Temporary (short-term) situation:
1390 * - Application is running on a different CPU, between reserve
1391 * and commit ring buffer operations,
1392 * - Application is preempted between reserve and commit ring
1393 * buffer operations,
1395 * Long-term situation:
1396 * - Application is stopped (SIGSTOP) between reserve and commit
1397 * ring buffer operations. Could eventually be resumed by
1399 * - Application is killed (SIGTERM, SIGINT, SIGKILL) between
1400 * reserve and commit ring buffer operation.
1402 * From a consumer perspective, handling short-term
1403 * unavailability situations is performed by retrying a few
1404 * times after a delay. Handling long-term unavailability
1405 * situations is handled by failing to get the sub-buffer.
1407 * In all of those situations, if the application is taking a
1408 * long time to perform its commit after ring buffer space
1409 * reservation, we can end up in a situation where the producer
1410 * will fill the ring buffer and try to write into the same
1411 * sub-buffer again (which has a missing commit). This is
1412 * handled by the producer in the sub-buffer switch handling
1413 * code of the reserve routine by detecting unbalanced
1414 * reserve/commit counters and discarding all further events
1415 * until the situation is resolved in those situations. Two
1416 * scenarios can occur:
1418 * 1) The application causing the reserve/commit counters to be
1419 * unbalanced has been terminated. In this situation, all
1420 * further events will be discarded in the buffers, and no
1421 * further buffer data will be readable by the consumer
1422 * daemon. Tearing down the UST tracing session and starting
1423 * anew is a work-around for those situations. Note that this
1424 * only affects per-UID tracing. In per-PID tracing, the
1425 * application vanishes with the termination, and therefore
1426 * no more data needs to be written to the buffers.
1427 * 2) The application causing the unbalance has been delayed for
1428 * a long time, but will eventually try to increment the
1429 * commit counter after eventually writing to the sub-buffer.
1430 * This situation can cause events to be discarded until the
1431 * application resumes its operations.
1433 if (((commit_count
- chan
->backend
.subbuf_size
)
1434 & chan
->commit_count_mask
)
1435 - (buf_trunc(consumed
, chan
)
1436 >> chan
->backend
.num_subbuf_order
)
1438 if (nr_retry
-- > 0) {
1439 if (nr_retry
<= (LTTNG_UST_RING_BUFFER_GET_RETRY
>> 1))
1440 (void) poll(NULL
, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS
);
1448 * Check that we are not about to read the same subbuffer in
1449 * which the writer head is.
1451 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed
, chan
)
1456 * Failure to get the subbuffer causes a busy-loop retry without going
1457 * to a wait queue. These are caused by short-lived race windows where
1458 * the writer is getting access to a subbuffer we were trying to get
1459 * access to. Also checks that the "consumed" buffer count we are
1460 * looking for matches the one contained in the subbuffer id.
1462 * The short-lived race window described here can be affected by
1463 * application signals and preemption, thus requiring to bound
1464 * the loop to a maximum number of retry.
1466 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1467 consumed_idx
, buf_trunc_val(consumed
, chan
),
1470 if (nr_retry
-- > 0) {
1471 if (nr_retry
<= (LTTNG_UST_RING_BUFFER_GET_RETRY
>> 1))
1472 (void) poll(NULL
, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS
);
1478 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1480 buf
->get_subbuf_consumed
= consumed
;
1481 buf
->get_subbuf
= 1;
1487 * The memory barriers __wait_event()/wake_up_interruptible() take care
1488 * of "raw_spin_is_locked" memory ordering.
1497 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1500 void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
1501 struct lttng_ust_shm_handle
*handle
)
1503 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1504 struct channel
*chan
;
1505 const struct lttng_ust_lib_ring_buffer_config
*config
;
1506 unsigned long sb_bindex
, consumed_idx
, consumed
;
1507 struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*rpages
;
1508 struct lttng_ust_lib_ring_buffer_backend_pages
*backend_pages
;
1510 chan
= shmp(handle
, bufb
->chan
);
1513 config
= &chan
->backend
.config
;
1514 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1516 if (!buf
->get_subbuf
) {
1518 * Reader puts a subbuffer it did not get.
1520 CHAN_WARN_ON(chan
, 1);
1523 consumed
= buf
->get_subbuf_consumed
;
1524 buf
->get_subbuf
= 0;
1527 * Clear the records_unread counter. (overruns counter)
1528 * Can still be non-zero if a file reader simply grabbed the data
1529 * without using iterators.
1530 * Can be below zero if an iterator is used on a snapshot more than
1533 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1534 rpages
= shmp_index(handle
, bufb
->array
, sb_bindex
);
1537 backend_pages
= shmp(handle
, rpages
->shmp
);
1540 v_add(config
, v_read(config
, &backend_pages
->records_unread
),
1541 &bufb
->records_read
);
1542 v_set(config
, &backend_pages
->records_unread
, 0);
1543 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1544 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1545 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1548 * Exchange the reader subbuffer with the one we put in its place in the
1549 * writer subbuffer table. Expect the original consumed count. If
1550 * update_read_sb_index fails, this is because the writer updated the
1551 * subbuffer concurrently. We should therefore keep the subbuffer we
1552 * currently have: it has become invalid to try reading this sub-buffer
1553 * consumed count value anyway.
1555 consumed_idx
= subbuf_index(consumed
, chan
);
1556 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1557 consumed_idx
, buf_trunc_val(consumed
, chan
),
1560 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1561 * if the writer concurrently updated it.
1566 * cons_offset is an iterator on all subbuffer offsets between the reader
1567 * position and the writer position. (inclusive)
1570 void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
1571 struct channel
*chan
,
1572 unsigned long cons_offset
,
1574 struct lttng_ust_shm_handle
*handle
)
1576 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1577 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1578 struct commit_counters_hot
*cc_hot
;
1579 struct commit_counters_cold
*cc_cold
;
1581 cons_idx
= subbuf_index(cons_offset
, chan
);
1582 cc_hot
= shmp_index(handle
, buf
->commit_hot
, cons_idx
);
1585 cc_cold
= shmp_index(handle
, buf
->commit_cold
, cons_idx
);
1588 commit_count
= v_read(config
, &cc_hot
->cc
);
1589 commit_count
+= cc_hot
->cc_rseq
;
1590 commit_count_sb
= v_read(config
, &cc_cold
->cc_sb
);
1592 if (subbuf_offset(commit_count
, chan
) != 0)
1593 DBG("ring buffer %s, cpu %d: "
1594 "commit count in subbuffer %lu,\n"
1595 "expecting multiples of %lu bytes\n"
1596 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1597 chan
->backend
.name
, cpu
, cons_idx
,
1598 chan
->backend
.subbuf_size
,
1599 commit_count
, commit_count_sb
);
1601 DBG("ring buffer: %s, cpu %d: %lu bytes committed\n",
1602 chan
->backend
.name
, cpu
, commit_count
);
1606 void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
1607 struct channel
*chan
,
1608 void *priv
, int cpu
,
1609 struct lttng_ust_shm_handle
*handle
)
1611 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1612 unsigned long write_offset
, cons_offset
;
1615 * No need to order commit_count, write_offset and cons_offset reads
1616 * because we execute at teardown when no more writer nor reader
1617 * references are left.
1619 write_offset
= v_read(config
, &buf
->offset
);
1620 cons_offset
= uatomic_read(&buf
->consumed
);
1621 if (write_offset
!= cons_offset
)
1622 DBG("ring buffer %s, cpu %d: "
1623 "non-consumed data\n"
1624 " [ %lu bytes written, %lu bytes read ]\n",
1625 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1627 for (cons_offset
= uatomic_read(&buf
->consumed
);
1628 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1631 cons_offset
= subbuf_align(cons_offset
, chan
))
1632 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1637 void lib_ring_buffer_print_errors(struct channel
*chan
,
1638 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
1639 struct lttng_ust_shm_handle
*handle
)
1641 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1642 void *priv
= channel_get_private(chan
);
1644 if (!strcmp(chan
->backend
.name
, "relay-metadata-mmap")) {
1645 DBG("ring buffer %s: %lu records written, "
1646 "%lu records overrun\n",
1648 v_read(config
, &buf
->records_count
),
1649 v_read(config
, &buf
->records_overrun
));
1651 DBG("ring buffer %s, cpu %d: %lu records written, "
1652 "%lu records overrun\n",
1653 chan
->backend
.name
, cpu
,
1654 v_read(config
, &buf
->records_count
),
1655 v_read(config
, &buf
->records_overrun
));
1657 if (v_read(config
, &buf
->records_lost_full
)
1658 || v_read(config
, &buf
->records_lost_wrap
)
1659 || v_read(config
, &buf
->records_lost_big
))
1660 DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n"
1661 " [ %lu buffer full, %lu nest buffer wrap-around, "
1662 "%lu event too big ]\n",
1663 chan
->backend
.name
, cpu
,
1664 v_read(config
, &buf
->records_lost_full
),
1665 v_read(config
, &buf
->records_lost_wrap
),
1666 v_read(config
, &buf
->records_lost_big
));
1668 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
, handle
);
1672 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1674 * Only executed by SWITCH_FLUSH, which can be issued while tracing is
1675 * active or at buffer finalization (destroy).
1678 void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer
*buf
,
1679 struct channel
*chan
,
1680 struct switch_offsets
*offsets
,
1682 struct lttng_ust_shm_handle
*handle
)
1684 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1685 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1686 unsigned long commit_count
;
1687 struct commit_counters_hot
*cc_hot
;
1689 config
->cb
.buffer_begin(buf
, tsc
, oldidx
, handle
);
1692 * Order all writes to buffer before the commit count update that will
1693 * determine that the subbuffer is full.
1696 cc_hot
= shmp_index(handle
, buf
->commit_hot
, oldidx
);
1699 v_add(config
, config
->cb
.subbuffer_header_size(),
1701 commit_count
= v_read(config
, &cc_hot
->cc
);
1702 commit_count
+= cc_hot
->cc_rseq
;
1703 /* Check if the written buffer has to be delivered */
1704 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1705 commit_count
, oldidx
, handle
, tsc
);
1706 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1707 offsets
->old
+ config
->cb
.subbuffer_header_size(),
1708 commit_count
, handle
, cc_hot
);
1712 * lib_ring_buffer_switch_old_end: switch old subbuffer
1714 * Note : offset_old should never be 0 here. It is ok, because we never perform
1715 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1716 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1720 void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer
*buf
,
1721 struct channel
*chan
,
1722 struct switch_offsets
*offsets
,
1724 struct lttng_ust_shm_handle
*handle
)
1726 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1727 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1728 unsigned long commit_count
, padding_size
, data_size
;
1729 struct commit_counters_hot
*cc_hot
;
1731 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1732 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1733 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
,
1737 * Order all writes to buffer before the commit count update that will
1738 * determine that the subbuffer is full.
1741 cc_hot
= shmp_index(handle
, buf
->commit_hot
, oldidx
);
1744 v_add(config
, padding_size
, &cc_hot
->cc
);
1745 commit_count
= v_read(config
, &cc_hot
->cc
);
1746 commit_count
+= cc_hot
->cc_rseq
;
1747 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1748 commit_count
, oldidx
, handle
, tsc
);
1749 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1750 offsets
->old
+ padding_size
, commit_count
, handle
,
1755 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1757 * This code can be executed unordered : writers may already have written to the
1758 * sub-buffer before this code gets executed, caution. The commit makes sure
1759 * that this code is executed before the deliver of this sub-buffer.
1762 void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer
*buf
,
1763 struct channel
*chan
,
1764 struct switch_offsets
*offsets
,
1766 struct lttng_ust_shm_handle
*handle
)
1768 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1769 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1770 unsigned long commit_count
;
1771 struct commit_counters_hot
*cc_hot
;
1773 config
->cb
.buffer_begin(buf
, tsc
, beginidx
, handle
);
1776 * Order all writes to buffer before the commit count update that will
1777 * determine that the subbuffer is full.
1780 cc_hot
= shmp_index(handle
, buf
->commit_hot
, beginidx
);
1783 v_add(config
, config
->cb
.subbuffer_header_size(), &cc_hot
->cc
);
1784 commit_count
= v_read(config
, &cc_hot
->cc
);
1785 commit_count
+= cc_hot
->cc_rseq
;
1786 /* Check if the written buffer has to be delivered */
1787 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1788 commit_count
, beginidx
, handle
, tsc
);
1789 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1790 offsets
->begin
+ config
->cb
.subbuffer_header_size(),
1791 commit_count
, handle
, cc_hot
);
1795 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1797 * Calls subbuffer_set_data_size() to set the data size of the current
1798 * sub-buffer. We do not need to perform check_deliver nor commit here,
1799 * since this task will be done by the "commit" of the event for which
1800 * we are currently doing the space reservation.
1803 void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer
*buf
,
1804 struct channel
*chan
,
1805 struct switch_offsets
*offsets
,
1807 struct lttng_ust_shm_handle
*handle
)
1809 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1810 unsigned long endidx
, data_size
;
1812 endidx
= subbuf_index(offsets
->end
- 1, chan
);
1813 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1814 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
,
1821 * !0 if execution must be aborted.
1824 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1825 struct lttng_ust_lib_ring_buffer
*buf
,
1826 struct channel
*chan
,
1827 struct switch_offsets
*offsets
,
1829 struct lttng_ust_shm_handle
*handle
)
1831 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1832 unsigned long off
, reserve_commit_diff
;
1834 offsets
->begin
= v_read(config
, &buf
->offset
);
1835 offsets
->old
= offsets
->begin
;
1836 offsets
->switch_old_start
= 0;
1837 off
= subbuf_offset(offsets
->begin
, chan
);
1839 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1842 * Ensure we flush the header of an empty subbuffer when doing the
1843 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1844 * total data gathering duration even if there were no records saved
1845 * after the last buffer switch.
1846 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1847 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1848 * subbuffer header as appropriate.
1849 * The next record that reserves space will be responsible for
1850 * populating the following subbuffer header. We choose not to populate
1851 * the next subbuffer header here because we want to be able to use
1852 * SWITCH_ACTIVE for periodical buffer flush, which must
1853 * guarantee that all the buffer content (records and header
1854 * timestamps) are visible to the reader. This is required for
1855 * quiescence guarantees for the fusion merge.
1857 if (mode
!= SWITCH_FLUSH
&& !off
)
1858 return -1; /* we do not have to switch : buffer is empty */
1860 if (caa_unlikely(off
== 0)) {
1861 unsigned long sb_index
, commit_count
;
1862 struct commit_counters_cold
*cc_cold
;
1865 * We are performing a SWITCH_FLUSH. There may be concurrent
1866 * writes into the buffer if e.g. invoked while performing a
1867 * snapshot on an active trace.
1869 * If the client does not save any header information
1870 * (sub-buffer header size == 0), don't switch empty subbuffer
1871 * on finalize, because it is invalid to deliver a completely
1874 if (!config
->cb
.subbuffer_header_size())
1877 /* Test new buffer integrity */
1878 sb_index
= subbuf_index(offsets
->begin
, chan
);
1879 cc_cold
= shmp_index(handle
, buf
->commit_cold
, sb_index
);
1882 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
1883 reserve_commit_diff
=
1884 (buf_trunc(offsets
->begin
, chan
)
1885 >> chan
->backend
.num_subbuf_order
)
1886 - (commit_count
& chan
->commit_count_mask
);
1887 if (caa_likely(reserve_commit_diff
== 0)) {
1888 /* Next subbuffer not being written to. */
1889 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1890 subbuf_trunc(offsets
->begin
, chan
)
1891 - subbuf_trunc((unsigned long)
1892 uatomic_read(&buf
->consumed
), chan
)
1893 >= chan
->backend
.buf_size
)) {
1895 * We do not overwrite non consumed buffers
1896 * and we are full : don't switch.
1901 * Next subbuffer not being written to, and we
1902 * are either in overwrite mode or the buffer is
1903 * not full. It's safe to write in this new
1909 * Next subbuffer reserve offset does not match the
1910 * commit offset. Don't perform switch in
1911 * producer-consumer and overwrite mode. Caused by
1912 * either a writer OOPS or too many nested writes over a
1913 * reserve/commit pair.
1919 * Need to write the subbuffer start header on finalize.
1921 offsets
->switch_old_start
= 1;
1923 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1924 /* Note: old points to the next subbuf at offset 0 */
1925 offsets
->end
= offsets
->begin
;
1930 * Force a sub-buffer switch. This operation is completely reentrant : can be
1931 * called while tracing is active with absolutely no lock held.
1933 * Note, however, that as a v_cmpxchg is used for some atomic
1934 * operations, this function must be called from the CPU which owns the buffer
1935 * for a ACTIVE flush.
1937 void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer
*buf
, enum switch_mode mode
,
1938 struct lttng_ust_shm_handle
*handle
)
1940 struct channel
*chan
;
1941 const struct lttng_ust_lib_ring_buffer_config
*config
;
1942 struct switch_offsets offsets
;
1943 unsigned long oldidx
;
1946 chan
= shmp(handle
, buf
->backend
.chan
);
1949 config
= &chan
->backend
.config
;
1954 * Perform retryable operations.
1957 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1959 return; /* Switch not needed */
1960 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1964 * Atomically update last_tsc. This update races against concurrent
1965 * atomic updates, but the race will always cause supplementary full TSC
1966 * records, never the opposite (missing a full TSC record when it would
1969 save_last_tsc(config
, buf
, tsc
);
1972 * Push the reader if necessary
1974 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1976 oldidx
= subbuf_index(offsets
.old
, chan
);
1977 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
, handle
);
1980 * May need to populate header start on SWITCH_FLUSH.
1982 if (offsets
.switch_old_start
) {
1983 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
, handle
);
1984 offsets
.old
+= config
->cb
.subbuffer_header_size();
1988 * Switch old subbuffer.
1990 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
, handle
);
1996 * -ENOSPC if event size is too large for packet.
1997 * -ENOBUFS if there is currently not enough space in buffer for the event.
1998 * -EIO if data cannot be written into the buffer for any other reason.
2001 int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer
*buf
,
2002 struct channel
*chan
,
2003 struct switch_offsets
*offsets
,
2004 struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
2006 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
2007 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
2008 unsigned long reserve_commit_diff
, offset_cmp
;
2011 offsets
->begin
= offset_cmp
= v_read(config
, &buf
->offset
);
2012 offsets
->old
= offsets
->begin
;
2013 offsets
->switch_new_start
= 0;
2014 offsets
->switch_new_end
= 0;
2015 offsets
->switch_old_end
= 0;
2016 offsets
->pre_header_padding
= 0;
2018 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
2019 if ((int64_t) ctx
->tsc
== -EIO
)
2022 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
2023 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
2025 if (caa_unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
2026 offsets
->switch_new_start
= 1; /* For offsets->begin */
2028 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
2030 &offsets
->pre_header_padding
,
2033 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
2036 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
) +
2037 offsets
->size
> chan
->backend
.subbuf_size
)) {
2038 offsets
->switch_old_end
= 1; /* For offsets->old */
2039 offsets
->switch_new_start
= 1; /* For offsets->begin */
2042 if (caa_unlikely(offsets
->switch_new_start
)) {
2043 unsigned long sb_index
, commit_count
;
2044 struct commit_counters_cold
*cc_cold
;
2047 * We are typically not filling the previous buffer completely.
2049 if (caa_likely(offsets
->switch_old_end
))
2050 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
2051 offsets
->begin
= offsets
->begin
2052 + config
->cb
.subbuffer_header_size();
2053 /* Test new buffer integrity */
2054 sb_index
= subbuf_index(offsets
->begin
, chan
);
2056 * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
2057 * lib_ring_buffer_check_deliver() has the matching
2058 * memory barriers required around commit_cold cc_sb
2059 * updates to ensure reserve and commit counter updates
2060 * are not seen reordered when updated by another CPU.
2063 cc_cold
= shmp_index(handle
, buf
->commit_cold
, sb_index
);
2066 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
2067 /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
2069 if (caa_unlikely(offset_cmp
!= v_read(config
, &buf
->offset
))) {
2071 * The reserve counter have been concurrently updated
2072 * while we read the commit counter. This means the
2073 * commit counter we read might not match buf->offset
2074 * due to concurrent update. We therefore need to retry.
2078 reserve_commit_diff
=
2079 (buf_trunc(offsets
->begin
, chan
)
2080 >> chan
->backend
.num_subbuf_order
)
2081 - (commit_count
& chan
->commit_count_mask
);
2082 if (caa_likely(reserve_commit_diff
== 0)) {
2083 /* Next subbuffer not being written to. */
2084 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
2085 subbuf_trunc(offsets
->begin
, chan
)
2086 - subbuf_trunc((unsigned long)
2087 uatomic_read(&buf
->consumed
), chan
)
2088 >= chan
->backend
.buf_size
)) {
2089 unsigned long nr_lost
;
2092 * We do not overwrite non consumed buffers
2093 * and we are full : record is lost.
2095 nr_lost
= v_read(config
, &buf
->records_lost_full
);
2096 v_inc(config
, &buf
->records_lost_full
);
2097 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
2098 DBG("%lu or more records lost in (%s:%d) (buffer full)\n",
2099 nr_lost
+ 1, chan
->backend
.name
,
2105 * Next subbuffer not being written to, and we
2106 * are either in overwrite mode or the buffer is
2107 * not full. It's safe to write in this new
2112 unsigned long nr_lost
;
2115 * Next subbuffer reserve offset does not match the
2116 * commit offset, and this did not involve update to the
2117 * reserve counter. Drop record in producer-consumer and
2118 * overwrite mode. Caused by either a writer OOPS or too
2119 * many nested writes over a reserve/commit pair.
2121 nr_lost
= v_read(config
, &buf
->records_lost_wrap
);
2122 v_inc(config
, &buf
->records_lost_wrap
);
2123 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
2124 DBG("%lu or more records lost in (%s:%d) (wrap-around)\n",
2125 nr_lost
+ 1, chan
->backend
.name
,
2131 config
->cb
.record_header_size(config
, chan
,
2133 &offsets
->pre_header_padding
,
2136 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
2139 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
)
2140 + offsets
->size
> chan
->backend
.subbuf_size
)) {
2141 unsigned long nr_lost
;
2144 * Record too big for subbuffers, report error, don't
2145 * complete the sub-buffer switch.
2147 nr_lost
= v_read(config
, &buf
->records_lost_big
);
2148 v_inc(config
, &buf
->records_lost_big
);
2149 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
2150 DBG("%lu or more records lost in (%s:%d) record size "
2151 " of %zu bytes is too large for buffer\n",
2152 nr_lost
+ 1, chan
->backend
.name
,
2153 buf
->backend
.cpu
, offsets
->size
);
2158 * We just made a successful buffer switch and the
2159 * record fits in the new subbuffer. Let's write.
2164 * Record fits in the current buffer and we are not on a switch
2165 * boundary. It's safe to write.
2168 offsets
->end
= offsets
->begin
+ offsets
->size
;
2170 if (caa_unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
2172 * The offset_end will fall at the very beginning of the next
2175 offsets
->switch_new_end
= 1; /* For offsets->begin */
2181 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
2182 * @ctx: ring buffer context.
2184 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
2185 * -EIO for other errors, else returns 0.
2186 * It will take care of sub-buffer switching.
2188 int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
2190 struct channel
*chan
= ctx
->chan
;
2191 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
2192 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
2193 struct lttng_ust_lib_ring_buffer
*buf
;
2194 struct switch_offsets offsets
;
2197 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
2198 buf
= shmp(handle
, chan
->backend
.buf
[ctx
->cpu
].shmp
);
2200 buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
2208 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
2210 if (caa_unlikely(ret
))
2212 } while (caa_unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
2217 * Atomically update last_tsc. This update races against concurrent
2218 * atomic updates, but the race will always cause supplementary full TSC
2219 * records, never the opposite (missing a full TSC record when it would
2222 save_last_tsc(config
, buf
, ctx
->tsc
);
2225 * Push the reader if necessary
2227 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
2230 * Clear noref flag for this subbuffer.
2232 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2233 subbuf_index(offsets
.end
- 1, chan
),
2237 * Switch old subbuffer if needed.
2239 if (caa_unlikely(offsets
.switch_old_end
)) {
2240 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2241 subbuf_index(offsets
.old
- 1, chan
),
2243 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2247 * Populate new subbuffer.
2249 if (caa_unlikely(offsets
.switch_new_start
))
2250 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2252 if (caa_unlikely(offsets
.switch_new_end
))
2253 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2255 ctx
->slot_size
= offsets
.size
;
2256 ctx
->pre_offset
= offsets
.begin
;
2257 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
2262 void lib_ring_buffer_vmcore_check_deliver(const struct lttng_ust_lib_ring_buffer_config
*config
,
2263 struct lttng_ust_lib_ring_buffer
*buf
,
2264 unsigned long commit_count
,
2266 struct lttng_ust_shm_handle
*handle
)
2268 struct commit_counters_hot
*cc_hot
;
2270 if (config
->oops
!= RING_BUFFER_OOPS_CONSISTENCY
)
2272 cc_hot
= shmp_index(handle
, buf
->commit_hot
, idx
);
2275 v_set(config
, &cc_hot
->seq
, commit_count
);
2279 * The ring buffer can count events recorded and overwritten per buffer,
2280 * but it is disabled by default due to its performance overhead.
2282 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
2284 void deliver_count_events(const struct lttng_ust_lib_ring_buffer_config
*config
,
2285 struct lttng_ust_lib_ring_buffer
*buf
,
2287 struct lttng_ust_shm_handle
*handle
)
2289 v_add(config
, subbuffer_get_records_count(config
,
2290 &buf
->backend
, idx
, handle
),
2291 &buf
->records_count
);
2292 v_add(config
, subbuffer_count_records_overrun(config
,
2293 &buf
->backend
, idx
, handle
),
2294 &buf
->records_overrun
);
2296 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
2298 void deliver_count_events(const struct lttng_ust_lib_ring_buffer_config
*config
,
2299 struct lttng_ust_lib_ring_buffer
*buf
,
2301 struct lttng_ust_shm_handle
*handle
)
2304 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
2306 void lib_ring_buffer_check_deliver_slow(const struct lttng_ust_lib_ring_buffer_config
*config
,
2307 struct lttng_ust_lib_ring_buffer
*buf
,
2308 struct channel
*chan
,
2309 unsigned long offset
,
2310 unsigned long commit_count
,
2312 struct lttng_ust_shm_handle
*handle
,
2315 unsigned long old_commit_count
= commit_count
2316 - chan
->backend
.subbuf_size
;
2317 struct commit_counters_cold
*cc_cold
;
2320 * If we succeeded at updating cc_sb below, we are the subbuffer
2321 * writer delivering the subbuffer. Deals with concurrent
2322 * updates of the "cc" value without adding a add_return atomic
2323 * operation to the fast path.
2325 * We are doing the delivery in two steps:
2326 * - First, we cmpxchg() cc_sb to the new value
2327 * old_commit_count + 1. This ensures that we are the only
2328 * subbuffer user successfully filling the subbuffer, but we
2329 * do _not_ set the cc_sb value to "commit_count" yet.
2330 * Therefore, other writers that would wrap around the ring
2331 * buffer and try to start writing to our subbuffer would
2332 * have to drop records, because it would appear as
2334 * We therefore have exclusive access to the subbuffer control
2335 * structures. This mutual exclusion with other writers is
2336 * crucially important to perform record overruns count in
2337 * flight recorder mode locklessly.
2338 * - When we are ready to release the subbuffer (either for
2339 * reading or for overrun by other writers), we simply set the
2340 * cc_sb value to "commit_count" and perform delivery.
2342 * The subbuffer size is least 2 bytes (minimum size: 1 page).
2343 * This guarantees that old_commit_count + 1 != commit_count.
2347 * Order prior updates to reserve count prior to the
2348 * commit_cold cc_sb update.
2351 cc_cold
= shmp_index(handle
, buf
->commit_cold
, idx
);
2354 if (caa_likely(v_cmpxchg(config
, &cc_cold
->cc_sb
,
2355 old_commit_count
, old_commit_count
+ 1)
2356 == old_commit_count
)) {
2358 * Start of exclusive subbuffer access. We are
2359 * guaranteed to be the last writer in this subbuffer
2360 * and any other writer trying to access this subbuffer
2361 * in this state is required to drop records.
2363 deliver_count_events(config
, buf
, idx
, handle
);
2364 config
->cb
.buffer_end(buf
, tsc
, idx
,
2365 lib_ring_buffer_get_data_size(config
,
2372 * Increment the packet counter while we have exclusive
2375 subbuffer_inc_packet_count(config
, &buf
->backend
, idx
, handle
);
2378 * Set noref flag and offset for this subbuffer id.
2379 * Contains a memory barrier that ensures counter stores
2380 * are ordered before set noref and offset.
2382 lib_ring_buffer_set_noref_offset(config
, &buf
->backend
, idx
,
2383 buf_trunc_val(offset
, chan
), handle
);
2386 * Order set_noref and record counter updates before the
2387 * end of subbuffer exclusive access. Orders with
2388 * respect to writers coming into the subbuffer after
2389 * wrap around, and also order wrt concurrent readers.
2392 /* End of exclusive subbuffer access */
2393 v_set(config
, &cc_cold
->cc_sb
, commit_count
);
2395 * Order later updates to reserve count after
2396 * the commit cold cc_sb update.
2399 lib_ring_buffer_vmcore_check_deliver(config
, buf
,
2400 commit_count
, idx
, handle
);
2403 * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
2405 if (config
->wakeup
== RING_BUFFER_WAKEUP_BY_WRITER
2406 && uatomic_read(&buf
->active_readers
)
2407 && lib_ring_buffer_poll_deliver(config
, buf
, chan
, handle
)) {
2408 lib_ring_buffer_wakeup(buf
, handle
);
2414 * Force a read (imply TLS fixup for dlopen) of TLS variables.
2416 void lttng_fixup_ringbuffer_tls(void)
2418 asm volatile ("" : : "m" (URCU_TLS(lib_ring_buffer_nesting
)));
2421 void lib_ringbuffer_signal_init(void)
2427 * Block signal for entire process, so only our thread processes
2431 ret
= pthread_sigmask(SIG_BLOCK
, &mask
, NULL
);
2434 PERROR("pthread_sigmask");