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
55 #include <sys/types.h>
62 #include <urcu/compiler.h>
64 #include <urcu/tls-compat.h>
69 #include <lttng/ringbuffer-config.h>
75 #include "../liblttng-ust/compat.h" /* For ENODATA */
77 /* Print DBG() messages about events lost only every 1048576 hits */
78 #define DBG_PRINT_NR_LOST (1UL << 20)
80 #define LTTNG_UST_RB_SIG_FLUSH SIGRTMIN
81 #define LTTNG_UST_RB_SIG_READ SIGRTMIN + 1
82 #define LTTNG_UST_RB_SIG_TEARDOWN SIGRTMIN + 2
83 #define CLOCKID CLOCK_MONOTONIC
84 #define LTTNG_UST_RING_BUFFER_GET_RETRY 10
85 #define LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS 10
88 * Non-static to ensure the compiler does not optimize away the xor.
90 uint8_t lttng_crash_magic_xor
[] = RB_CRASH_DUMP_ABI_MAGIC_XOR
;
93 * Use POSIX SHM: shm_open(3) and shm_unlink(3).
94 * close(2) to close the fd returned by shm_open.
95 * shm_unlink releases the shared memory object name.
96 * ftruncate(2) sets the size of the memory object.
97 * mmap/munmap maps the shared memory obj to a virtual address in the
98 * calling proceess (should be done both in libust and consumer).
99 * See shm_overview(7) for details.
100 * Pass file descriptor returned by shm_open(3) to ltt-sessiond through
103 * Since we don't need to access the object using its name, we can
104 * immediately shm_unlink(3) it, and only keep the handle with its file
109 * Internal structure representing offsets to use at a sub-buffer switch.
111 struct switch_offsets
{
112 unsigned long begin
, end
, old
;
113 size_t pre_header_padding
, size
;
114 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
118 DEFINE_URCU_TLS(unsigned int, lib_ring_buffer_nesting
);
121 * wakeup_fd_mutex protects wakeup fd use by timer from concurrent
124 static pthread_mutex_t wakeup_fd_mutex
= PTHREAD_MUTEX_INITIALIZER
;
127 void lib_ring_buffer_print_errors(struct channel
*chan
,
128 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
129 struct lttng_ust_shm_handle
*handle
);
132 * Handle timer teardown race wrt memory free of private data by
133 * ring buffer signals are handled by a single thread, which permits
134 * a synchronization point between handling of each signal.
135 * Protected by the lock within the structure.
137 struct timer_signal_data
{
138 pthread_t tid
; /* thread id managing signals */
141 pthread_mutex_t lock
;
144 static struct timer_signal_data timer_signal
= {
148 .lock
= PTHREAD_MUTEX_INITIALIZER
,
152 * lib_ring_buffer_reset - Reset ring buffer to initial values.
155 * Effectively empty the ring buffer. Should be called when the buffer is not
156 * used for writing. The ring buffer can be opened for reading, but the reader
157 * should not be using the iterator concurrently with reset. The previous
158 * current iterator record is reset.
160 void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer
*buf
,
161 struct lttng_ust_shm_handle
*handle
)
163 struct channel
*chan
;
164 const struct lttng_ust_lib_ring_buffer_config
*config
;
167 chan
= shmp(handle
, buf
->backend
.chan
);
170 config
= &chan
->backend
.config
;
172 * Reset iterator first. It will put the subbuffer if it currently holds
175 v_set(config
, &buf
->offset
, 0);
176 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
177 v_set(config
, &shmp_index(handle
, buf
->commit_hot
, i
)->cc
, 0);
178 v_set(config
, &shmp_index(handle
, buf
->commit_hot
, i
)->seq
, 0);
179 v_set(config
, &shmp_index(handle
, buf
->commit_cold
, i
)->cc_sb
, 0);
181 uatomic_set(&buf
->consumed
, 0);
182 uatomic_set(&buf
->record_disabled
, 0);
183 v_set(config
, &buf
->last_tsc
, 0);
184 lib_ring_buffer_backend_reset(&buf
->backend
, handle
);
185 /* Don't reset number of active readers */
186 v_set(config
, &buf
->records_lost_full
, 0);
187 v_set(config
, &buf
->records_lost_wrap
, 0);
188 v_set(config
, &buf
->records_lost_big
, 0);
189 v_set(config
, &buf
->records_count
, 0);
190 v_set(config
, &buf
->records_overrun
, 0);
195 * channel_reset - Reset channel to initial values.
198 * Effectively empty the channel. Should be called when the channel is not used
199 * for writing. The channel can be opened for reading, but the reader should not
200 * be using the iterator concurrently with reset. The previous current iterator
203 void channel_reset(struct channel
*chan
)
206 * Reset iterators first. Will put the subbuffer if held for reading.
208 uatomic_set(&chan
->record_disabled
, 0);
209 /* Don't reset commit_count_mask, still valid */
210 channel_backend_reset(&chan
->backend
);
211 /* Don't reset switch/read timer interval */
212 /* Don't reset notifiers and notifier enable bits */
213 /* Don't reset reader reference count */
217 void init_crash_abi(const struct lttng_ust_lib_ring_buffer_config
*config
,
218 struct lttng_crash_abi
*crash_abi
,
219 struct lttng_ust_lib_ring_buffer
*buf
,
220 struct channel_backend
*chanb
,
221 struct shm_object
*shmobj
,
222 struct lttng_ust_shm_handle
*handle
)
226 for (i
= 0; i
< RB_CRASH_DUMP_ABI_MAGIC_LEN
; i
++)
227 crash_abi
->magic
[i
] = lttng_crash_magic_xor
[i
] ^ 0xFF;
228 crash_abi
->mmap_length
= shmobj
->memory_map_size
;
229 crash_abi
->endian
= RB_CRASH_ENDIAN
;
230 crash_abi
->major
= RB_CRASH_DUMP_ABI_MAJOR
;
231 crash_abi
->minor
= RB_CRASH_DUMP_ABI_MINOR
;
232 crash_abi
->word_size
= sizeof(unsigned long);
233 crash_abi
->layout_type
= LTTNG_CRASH_TYPE_UST
;
235 /* Offset of fields */
236 crash_abi
->offset
.prod_offset
=
237 (uint32_t) ((char *) &buf
->offset
- (char *) buf
);
238 crash_abi
->offset
.consumed_offset
=
239 (uint32_t) ((char *) &buf
->consumed
- (char *) buf
);
240 crash_abi
->offset
.commit_hot_array
=
241 (uint32_t) ((char *) shmp(handle
, buf
->commit_hot
) - (char *) buf
);
242 crash_abi
->offset
.commit_hot_seq
=
243 offsetof(struct commit_counters_hot
, seq
);
244 crash_abi
->offset
.buf_wsb_array
=
245 (uint32_t) ((char *) shmp(handle
, buf
->backend
.buf_wsb
) - (char *) buf
);
246 crash_abi
->offset
.buf_wsb_id
=
247 offsetof(struct lttng_ust_lib_ring_buffer_backend_subbuffer
, id
);
248 crash_abi
->offset
.sb_array
=
249 (uint32_t) ((char *) shmp(handle
, buf
->backend
.array
) - (char *) buf
);
250 crash_abi
->offset
.sb_array_shmp_offset
=
251 offsetof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp
,
253 crash_abi
->offset
.sb_backend_p_offset
=
254 offsetof(struct lttng_ust_lib_ring_buffer_backend_pages
,
258 crash_abi
->length
.prod_offset
= sizeof(buf
->offset
);
259 crash_abi
->length
.consumed_offset
= sizeof(buf
->consumed
);
260 crash_abi
->length
.commit_hot_seq
=
261 sizeof(((struct commit_counters_hot
*) NULL
)->seq
);
262 crash_abi
->length
.buf_wsb_id
=
263 sizeof(((struct lttng_ust_lib_ring_buffer_backend_subbuffer
*) NULL
)->id
);
264 crash_abi
->length
.sb_array_shmp_offset
=
265 sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*) NULL
)->shmp
._ref
.offset
);
266 crash_abi
->length
.sb_backend_p_offset
=
267 sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages
*) NULL
)->p
._ref
.offset
);
270 crash_abi
->stride
.commit_hot_array
=
271 sizeof(struct commit_counters_hot
);
272 crash_abi
->stride
.buf_wsb_array
=
273 sizeof(struct lttng_ust_lib_ring_buffer_backend_subbuffer
);
274 crash_abi
->stride
.sb_array
=
275 sizeof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp
);
277 /* Buffer constants */
278 crash_abi
->buf_size
= chanb
->buf_size
;
279 crash_abi
->subbuf_size
= chanb
->subbuf_size
;
280 crash_abi
->num_subbuf
= chanb
->num_subbuf
;
281 crash_abi
->mode
= (uint32_t) chanb
->config
.mode
;
283 if (config
->cb
.content_size_field
) {
284 size_t offset
, length
;
286 config
->cb
.content_size_field(config
, &offset
, &length
);
287 crash_abi
->offset
.content_size
= offset
;
288 crash_abi
->length
.content_size
= length
;
290 crash_abi
->offset
.content_size
= 0;
291 crash_abi
->length
.content_size
= 0;
293 if (config
->cb
.packet_size_field
) {
294 size_t offset
, length
;
296 config
->cb
.packet_size_field(config
, &offset
, &length
);
297 crash_abi
->offset
.packet_size
= offset
;
298 crash_abi
->length
.packet_size
= length
;
300 crash_abi
->offset
.packet_size
= 0;
301 crash_abi
->length
.packet_size
= 0;
306 * Must be called under cpu hotplug protection.
308 int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer
*buf
,
309 struct channel_backend
*chanb
, int cpu
,
310 struct lttng_ust_shm_handle
*handle
,
311 struct shm_object
*shmobj
)
313 const struct lttng_ust_lib_ring_buffer_config
*config
= &chanb
->config
;
314 struct channel
*chan
= caa_container_of(chanb
, struct channel
, backend
);
315 void *priv
= channel_get_private(chan
);
316 size_t subbuf_header_size
;
320 /* Test for cpu hotplug */
321 if (buf
->backend
.allocated
)
324 align_shm(shmobj
, __alignof__(struct commit_counters_hot
));
325 set_shmp(buf
->commit_hot
,
327 sizeof(struct commit_counters_hot
) * chan
->backend
.num_subbuf
));
328 if (!shmp(handle
, buf
->commit_hot
)) {
332 align_shm(shmobj
, __alignof__(struct commit_counters_cold
));
333 set_shmp(buf
->commit_cold
,
335 sizeof(struct commit_counters_cold
) * chan
->backend
.num_subbuf
));
336 if (!shmp(handle
, buf
->commit_cold
)) {
341 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
,
342 cpu
, handle
, shmobj
);
348 * Write the subbuffer header for first subbuffer so we know the total
349 * duration of data gathering.
351 subbuf_header_size
= config
->cb
.subbuffer_header_size();
352 v_set(config
, &buf
->offset
, subbuf_header_size
);
353 subbuffer_id_clear_noref(config
, &shmp_index(handle
, buf
->backend
.buf_wsb
, 0)->id
);
354 tsc
= config
->cb
.ring_buffer_clock_read(shmp(handle
, buf
->backend
.chan
));
355 config
->cb
.buffer_begin(buf
, tsc
, 0, handle
);
356 v_add(config
, subbuf_header_size
, &shmp_index(handle
, buf
->commit_hot
, 0)->cc
);
357 v_add(config
, subbuf_header_size
, &shmp_index(handle
, buf
->commit_hot
, 0)->seq
);
359 if (config
->cb
.buffer_create
) {
360 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
, handle
);
365 init_crash_abi(config
, &buf
->crash_abi
, buf
, chanb
, shmobj
, handle
);
367 buf
->backend
.allocated
= 1;
372 /* commit_cold will be freed by shm teardown */
374 /* commit_hot will be freed by shm teardown */
380 void lib_ring_buffer_channel_switch_timer(int sig
, siginfo_t
*si
, void *uc
)
382 const struct lttng_ust_lib_ring_buffer_config
*config
;
383 struct lttng_ust_shm_handle
*handle
;
384 struct channel
*chan
;
387 assert(CMM_LOAD_SHARED(timer_signal
.tid
) == pthread_self());
389 chan
= si
->si_value
.sival_ptr
;
390 handle
= chan
->handle
;
391 config
= &chan
->backend
.config
;
393 DBG("Switch timer for channel %p\n", chan
);
396 * Only flush buffers periodically if readers are active.
398 pthread_mutex_lock(&wakeup_fd_mutex
);
399 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
400 for_each_possible_cpu(cpu
) {
401 struct lttng_ust_lib_ring_buffer
*buf
=
402 shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
406 if (uatomic_read(&buf
->active_readers
))
407 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
,
411 struct lttng_ust_lib_ring_buffer
*buf
=
412 shmp(handle
, chan
->backend
.buf
[0].shmp
);
416 if (uatomic_read(&buf
->active_readers
))
417 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
,
420 pthread_mutex_unlock(&wakeup_fd_mutex
);
425 void lib_ring_buffer_channel_do_read(struct channel
*chan
)
427 const struct lttng_ust_lib_ring_buffer_config
*config
;
428 struct lttng_ust_shm_handle
*handle
;
431 handle
= chan
->handle
;
432 config
= &chan
->backend
.config
;
435 * Only flush buffers periodically if readers are active.
437 pthread_mutex_lock(&wakeup_fd_mutex
);
438 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
439 for_each_possible_cpu(cpu
) {
440 struct lttng_ust_lib_ring_buffer
*buf
=
441 shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
445 if (uatomic_read(&buf
->active_readers
)
446 && lib_ring_buffer_poll_deliver(config
, buf
,
448 lib_ring_buffer_wakeup(buf
, handle
);
452 struct lttng_ust_lib_ring_buffer
*buf
=
453 shmp(handle
, chan
->backend
.buf
[0].shmp
);
457 if (uatomic_read(&buf
->active_readers
)
458 && lib_ring_buffer_poll_deliver(config
, buf
,
460 lib_ring_buffer_wakeup(buf
, handle
);
463 pthread_mutex_unlock(&wakeup_fd_mutex
);
467 void lib_ring_buffer_channel_read_timer(int sig
, siginfo_t
*si
, void *uc
)
469 struct channel
*chan
;
471 assert(CMM_LOAD_SHARED(timer_signal
.tid
) == pthread_self());
472 chan
= si
->si_value
.sival_ptr
;
473 DBG("Read timer for channel %p\n", chan
);
474 lib_ring_buffer_channel_do_read(chan
);
479 void rb_setmask(sigset_t
*mask
)
483 ret
= sigemptyset(mask
);
485 PERROR("sigemptyset");
487 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_FLUSH
);
491 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_READ
);
495 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_TEARDOWN
);
502 void *sig_thread(void *arg
)
508 /* Only self thread will receive signal mask. */
510 CMM_STORE_SHARED(timer_signal
.tid
, pthread_self());
513 signr
= sigwaitinfo(&mask
, &info
);
516 PERROR("sigwaitinfo");
519 if (signr
== LTTNG_UST_RB_SIG_FLUSH
) {
520 lib_ring_buffer_channel_switch_timer(info
.si_signo
,
522 } else if (signr
== LTTNG_UST_RB_SIG_READ
) {
523 lib_ring_buffer_channel_read_timer(info
.si_signo
,
525 } else if (signr
== LTTNG_UST_RB_SIG_TEARDOWN
) {
527 CMM_STORE_SHARED(timer_signal
.qs_done
, 1);
530 ERR("Unexptected signal %d\n", info
.si_signo
);
537 * Ensure only a single thread listens on the timer signal.
540 void lib_ring_buffer_setup_timer_thread(void)
545 pthread_mutex_lock(&timer_signal
.lock
);
546 if (timer_signal
.setup_done
)
549 ret
= pthread_create(&thread
, NULL
, &sig_thread
, NULL
);
552 PERROR("pthread_create");
554 ret
= pthread_detach(thread
);
557 PERROR("pthread_detach");
559 timer_signal
.setup_done
= 1;
561 pthread_mutex_unlock(&timer_signal
.lock
);
565 * Wait for signal-handling thread quiescent state.
568 void lib_ring_buffer_wait_signal_thread_qs(unsigned int signr
)
570 sigset_t pending_set
;
574 * We need to be the only thread interacting with the thread
575 * that manages signals for teardown synchronization.
577 pthread_mutex_lock(&timer_signal
.lock
);
580 * Ensure we don't have any signal queued for this channel.
583 ret
= sigemptyset(&pending_set
);
585 PERROR("sigemptyset");
587 ret
= sigpending(&pending_set
);
589 PERROR("sigpending");
591 if (!sigismember(&pending_set
, signr
))
597 * From this point, no new signal handler will be fired that
598 * would try to access "chan". However, we still need to wait
599 * for any currently executing handler to complete.
602 CMM_STORE_SHARED(timer_signal
.qs_done
, 0);
606 * Kill with LTTNG_UST_RB_SIG_TEARDOWN, so signal management
609 kill(getpid(), LTTNG_UST_RB_SIG_TEARDOWN
);
611 while (!CMM_LOAD_SHARED(timer_signal
.qs_done
))
615 pthread_mutex_unlock(&timer_signal
.lock
);
619 void lib_ring_buffer_channel_switch_timer_start(struct channel
*chan
)
622 struct itimerspec its
;
625 if (!chan
->switch_timer_interval
|| chan
->switch_timer_enabled
)
628 chan
->switch_timer_enabled
= 1;
630 lib_ring_buffer_setup_timer_thread();
632 sev
.sigev_notify
= SIGEV_SIGNAL
;
633 sev
.sigev_signo
= LTTNG_UST_RB_SIG_FLUSH
;
634 sev
.sigev_value
.sival_ptr
= chan
;
635 ret
= timer_create(CLOCKID
, &sev
, &chan
->switch_timer
);
637 PERROR("timer_create");
640 its
.it_value
.tv_sec
= chan
->switch_timer_interval
/ 1000000;
641 its
.it_value
.tv_nsec
= (chan
->switch_timer_interval
% 1000000) * 1000;
642 its
.it_interval
.tv_sec
= its
.it_value
.tv_sec
;
643 its
.it_interval
.tv_nsec
= its
.it_value
.tv_nsec
;
645 ret
= timer_settime(chan
->switch_timer
, 0, &its
, NULL
);
647 PERROR("timer_settime");
652 void lib_ring_buffer_channel_switch_timer_stop(struct channel
*chan
)
656 if (!chan
->switch_timer_interval
|| !chan
->switch_timer_enabled
)
659 ret
= timer_delete(chan
->switch_timer
);
661 PERROR("timer_delete");
664 lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_FLUSH
);
666 chan
->switch_timer
= 0;
667 chan
->switch_timer_enabled
= 0;
671 void lib_ring_buffer_channel_read_timer_start(struct channel
*chan
)
673 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
675 struct itimerspec its
;
678 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
679 || !chan
->read_timer_interval
|| chan
->read_timer_enabled
)
682 chan
->read_timer_enabled
= 1;
684 lib_ring_buffer_setup_timer_thread();
686 sev
.sigev_notify
= SIGEV_SIGNAL
;
687 sev
.sigev_signo
= LTTNG_UST_RB_SIG_READ
;
688 sev
.sigev_value
.sival_ptr
= chan
;
689 ret
= timer_create(CLOCKID
, &sev
, &chan
->read_timer
);
691 PERROR("timer_create");
694 its
.it_value
.tv_sec
= chan
->read_timer_interval
/ 1000000;
695 its
.it_value
.tv_nsec
= (chan
->read_timer_interval
% 1000000) * 1000;
696 its
.it_interval
.tv_sec
= its
.it_value
.tv_sec
;
697 its
.it_interval
.tv_nsec
= its
.it_value
.tv_nsec
;
699 ret
= timer_settime(chan
->read_timer
, 0, &its
, NULL
);
701 PERROR("timer_settime");
706 void lib_ring_buffer_channel_read_timer_stop(struct channel
*chan
)
708 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
711 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
712 || !chan
->read_timer_interval
|| !chan
->read_timer_enabled
)
715 ret
= timer_delete(chan
->read_timer
);
717 PERROR("timer_delete");
721 * do one more check to catch data that has been written in the last
724 lib_ring_buffer_channel_do_read(chan
);
726 lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_READ
);
728 chan
->read_timer
= 0;
729 chan
->read_timer_enabled
= 0;
732 static void channel_unregister_notifiers(struct channel
*chan
,
733 struct lttng_ust_shm_handle
*handle
)
735 lib_ring_buffer_channel_switch_timer_stop(chan
);
736 lib_ring_buffer_channel_read_timer_stop(chan
);
739 static void channel_print_errors(struct channel
*chan
,
740 struct lttng_ust_shm_handle
*handle
)
742 const struct lttng_ust_lib_ring_buffer_config
*config
=
743 &chan
->backend
.config
;
746 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
747 for_each_possible_cpu(cpu
) {
748 struct lttng_ust_lib_ring_buffer
*buf
=
749 shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
750 lib_ring_buffer_print_errors(chan
, buf
, cpu
, handle
);
753 struct lttng_ust_lib_ring_buffer
*buf
=
754 shmp(handle
, chan
->backend
.buf
[0].shmp
);
756 lib_ring_buffer_print_errors(chan
, buf
, -1, handle
);
760 static void channel_free(struct channel
*chan
,
761 struct lttng_ust_shm_handle
*handle
)
763 channel_backend_free(&chan
->backend
, handle
);
764 /* chan is freed by shm teardown */
765 shm_object_table_destroy(handle
->table
);
770 * channel_create - Create channel.
771 * @config: ring buffer instance configuration
772 * @name: name of the channel
773 * @priv_data: ring buffer client private data area pointer (output)
774 * @priv_data_size: length, in bytes, of the private data area.
775 * @priv_data_init: initialization data for private data.
776 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
777 * address mapping. It is used only by RING_BUFFER_STATIC
778 * configuration. It can be set to NULL for other backends.
779 * @subbuf_size: subbuffer size
780 * @num_subbuf: number of subbuffers
781 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
782 * padding to let readers get those sub-buffers.
783 * Used for live streaming.
784 * @read_timer_interval: Time interval (in us) to wake up pending readers.
785 * @stream_fds: array of stream file descriptors.
786 * @nr_stream_fds: number of file descriptors in array.
789 * Returns NULL on failure.
791 struct lttng_ust_shm_handle
*channel_create(const struct lttng_ust_lib_ring_buffer_config
*config
,
794 size_t priv_data_align
,
795 size_t priv_data_size
,
796 void *priv_data_init
,
797 void *buf_addr
, size_t subbuf_size
,
798 size_t num_subbuf
, unsigned int switch_timer_interval
,
799 unsigned int read_timer_interval
,
800 const int *stream_fds
, int nr_stream_fds
)
803 size_t shmsize
, chansize
;
804 struct channel
*chan
;
805 struct lttng_ust_shm_handle
*handle
;
806 struct shm_object
*shmobj
;
807 unsigned int nr_streams
;
809 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
810 nr_streams
= num_possible_cpus();
814 if (nr_stream_fds
!= nr_streams
)
817 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
818 read_timer_interval
))
821 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
825 /* Allocate table for channel + per-cpu buffers */
826 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
828 goto error_table_alloc
;
830 /* Calculate the shm allocation layout */
831 shmsize
= sizeof(struct channel
);
832 shmsize
+= offset_align(shmsize
, __alignof__(struct lttng_ust_lib_ring_buffer_shmp
));
833 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
) * nr_streams
;
836 shmsize
+= offset_align(shmsize
, priv_data_align
);
837 shmsize
+= priv_data_size
;
839 /* Allocate normal memory for channel (not shared) */
840 shmobj
= shm_object_table_alloc(handle
->table
, shmsize
, SHM_OBJECT_MEM
,
844 /* struct channel is at object 0, offset 0 (hardcoded) */
845 set_shmp(handle
->chan
, zalloc_shm(shmobj
, chansize
));
846 assert(handle
->chan
._ref
.index
== 0);
847 assert(handle
->chan
._ref
.offset
== 0);
848 chan
= shmp(handle
, handle
->chan
);
851 chan
->nr_streams
= nr_streams
;
853 /* space for private data */
854 if (priv_data_size
) {
855 DECLARE_SHMP(void, priv_data_alloc
);
857 align_shm(shmobj
, priv_data_align
);
858 chan
->priv_data_offset
= shmobj
->allocated_len
;
859 set_shmp(priv_data_alloc
, zalloc_shm(shmobj
, priv_data_size
));
860 if (!shmp(handle
, priv_data_alloc
))
862 *priv_data
= channel_get_private(chan
);
863 memcpy(*priv_data
, priv_data_init
, priv_data_size
);
865 chan
->priv_data_offset
= -1;
870 ret
= channel_backend_init(&chan
->backend
, name
, config
,
871 subbuf_size
, num_subbuf
, handle
,
874 goto error_backend_init
;
876 chan
->handle
= handle
;
877 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
879 chan
->switch_timer_interval
= switch_timer_interval
;
880 chan
->read_timer_interval
= read_timer_interval
;
881 lib_ring_buffer_channel_switch_timer_start(chan
);
882 lib_ring_buffer_channel_read_timer_start(chan
);
888 shm_object_table_destroy(handle
->table
);
894 struct lttng_ust_shm_handle
*channel_handle_create(void *data
,
895 uint64_t memory_map_size
,
898 struct lttng_ust_shm_handle
*handle
;
899 struct shm_object
*object
;
901 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
905 /* Allocate table for channel + per-cpu buffers */
906 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
908 goto error_table_alloc
;
909 /* Add channel object */
910 object
= shm_object_table_append_mem(handle
->table
, data
,
911 memory_map_size
, wakeup_fd
);
913 goto error_table_object
;
914 /* struct channel is at object 0, offset 0 (hardcoded) */
915 handle
->chan
._ref
.index
= 0;
916 handle
->chan
._ref
.offset
= 0;
920 shm_object_table_destroy(handle
->table
);
926 int channel_handle_add_stream(struct lttng_ust_shm_handle
*handle
,
927 int shm_fd
, int wakeup_fd
, uint32_t stream_nr
,
928 uint64_t memory_map_size
)
930 struct shm_object
*object
;
932 /* Add stream object */
933 object
= shm_object_table_append_shm(handle
->table
,
934 shm_fd
, wakeup_fd
, stream_nr
,
941 unsigned int channel_handle_get_nr_streams(struct lttng_ust_shm_handle
*handle
)
943 assert(handle
->table
);
944 return handle
->table
->allocated_len
- 1;
948 void channel_release(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
)
950 channel_free(chan
, handle
);
954 * channel_destroy - Finalize, wait for q.s. and destroy channel.
955 * @chan: channel to destroy
958 * Call "destroy" callback, finalize channels, decrement the channel
959 * reference count. Note that when readers have completed data
960 * consumption of finalized channels, get_subbuf() will return -ENODATA.
961 * They should release their handle at that point.
963 void channel_destroy(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
968 * Note: the consumer takes care of finalizing and
969 * switching the buffers.
971 channel_unregister_notifiers(chan
, handle
);
973 * The consumer prints errors.
975 channel_print_errors(chan
, handle
);
979 * sessiond/consumer are keeping a reference on the shm file
980 * descriptor directly. No need to refcount.
982 channel_release(chan
, handle
);
986 struct lttng_ust_lib_ring_buffer
*channel_get_ring_buffer(
987 const struct lttng_ust_lib_ring_buffer_config
*config
,
988 struct channel
*chan
, int cpu
,
989 struct lttng_ust_shm_handle
*handle
,
990 int *shm_fd
, int *wait_fd
,
992 uint64_t *memory_map_size
)
996 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
999 if (cpu
>= num_possible_cpus())
1002 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1003 *shm_fd
= shm_get_shm_fd(handle
, ref
);
1004 *wait_fd
= shm_get_wait_fd(handle
, ref
);
1005 *wakeup_fd
= shm_get_wakeup_fd(handle
, ref
);
1006 if (shm_get_shm_size(handle
, ref
, memory_map_size
))
1008 return shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
1011 int ring_buffer_channel_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1012 struct channel
*chan
,
1013 struct lttng_ust_shm_handle
*handle
)
1015 struct shm_ref
*ref
;
1017 ref
= &handle
->chan
._ref
;
1018 return shm_close_wait_fd(handle
, ref
);
1021 int ring_buffer_channel_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1022 struct channel
*chan
,
1023 struct lttng_ust_shm_handle
*handle
)
1025 struct shm_ref
*ref
;
1027 ref
= &handle
->chan
._ref
;
1028 return shm_close_wakeup_fd(handle
, ref
);
1031 int ring_buffer_stream_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1032 struct channel
*chan
,
1033 struct lttng_ust_shm_handle
*handle
,
1036 struct shm_ref
*ref
;
1038 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1041 if (cpu
>= num_possible_cpus())
1044 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1045 return shm_close_wait_fd(handle
, ref
);
1048 int ring_buffer_stream_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1049 struct channel
*chan
,
1050 struct lttng_ust_shm_handle
*handle
,
1053 struct shm_ref
*ref
;
1056 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1059 if (cpu
>= num_possible_cpus())
1062 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1063 pthread_mutex_lock(&wakeup_fd_mutex
);
1064 ret
= shm_close_wakeup_fd(handle
, ref
);
1065 pthread_mutex_unlock(&wakeup_fd_mutex
);
1069 int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer
*buf
,
1070 struct lttng_ust_shm_handle
*handle
)
1072 if (uatomic_cmpxchg(&buf
->active_readers
, 0, 1) != 0)
1078 void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer
*buf
,
1079 struct lttng_ust_shm_handle
*handle
)
1081 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1083 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1085 uatomic_dec(&buf
->active_readers
);
1089 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
1091 * @consumed: consumed count indicating the position where to read
1092 * @produced: produced count, indicates position when to stop reading
1094 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1095 * data to read at consumed position, or 0 if the get operation succeeds.
1098 int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer
*buf
,
1099 unsigned long *consumed
, unsigned long *produced
,
1100 struct lttng_ust_shm_handle
*handle
)
1102 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1103 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1104 unsigned long consumed_cur
, write_offset
;
1107 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
1109 * Read finalized before counters.
1112 consumed_cur
= uatomic_read(&buf
->consumed
);
1114 * No need to issue a memory barrier between consumed count read and
1115 * write offset read, because consumed count can only change
1116 * concurrently in overwrite mode, and we keep a sequence counter
1117 * identifier derived from the write offset to check we are getting
1118 * the same sub-buffer we are expecting (the sub-buffers are atomically
1119 * "tagged" upon writes, tags are checked upon read).
1121 write_offset
= v_read(config
, &buf
->offset
);
1124 * Check that we are not about to read the same subbuffer in
1125 * which the writer head is.
1127 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
1131 *consumed
= consumed_cur
;
1132 *produced
= subbuf_trunc(write_offset
, chan
);
1138 * The memory barriers __wait_event()/wake_up_interruptible() take care
1139 * of "raw_spin_is_locked" memory ordering.
1148 * lib_ring_buffer_move_consumer - move consumed counter forward
1150 * @consumed_new: new consumed count value
1152 void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer
*buf
,
1153 unsigned long consumed_new
,
1154 struct lttng_ust_shm_handle
*handle
)
1156 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1157 struct channel
*chan
= shmp(handle
, bufb
->chan
);
1158 unsigned long consumed
;
1160 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1163 * Only push the consumed value forward.
1164 * If the consumed cmpxchg fails, this is because we have been pushed by
1165 * the writer in flight recorder mode.
1167 consumed
= uatomic_read(&buf
->consumed
);
1168 while ((long) consumed
- (long) consumed_new
< 0)
1169 consumed
= uatomic_cmpxchg(&buf
->consumed
, consumed
,
1174 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
1176 * @consumed: consumed count indicating the position where to read
1178 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1179 * data to read at consumed position, or 0 if the get operation succeeds.
1181 int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
1182 unsigned long consumed
,
1183 struct lttng_ust_shm_handle
*handle
)
1185 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1186 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1187 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
1188 int ret
, finalized
, nr_retry
= LTTNG_UST_RING_BUFFER_GET_RETRY
;
1191 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
1193 * Read finalized before counters.
1196 consumed_cur
= uatomic_read(&buf
->consumed
);
1197 consumed_idx
= subbuf_index(consumed
, chan
);
1198 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, consumed_idx
)->cc_sb
);
1200 * Make sure we read the commit count before reading the buffer
1201 * data and the write offset. Correct consumed offset ordering
1202 * wrt commit count is insured by the use of cmpxchg to update
1203 * the consumed offset.
1206 * Local rmb to match the remote wmb to read the commit count
1207 * before the buffer data and the write offset.
1211 write_offset
= v_read(config
, &buf
->offset
);
1214 * Check that the buffer we are getting is after or at consumed_cur
1217 if ((long) subbuf_trunc(consumed
, chan
)
1218 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
1222 * Check that the subbuffer we are trying to consume has been
1223 * already fully committed. There are a few causes that can make
1224 * this unavailability situation occur:
1226 * Temporary (short-term) situation:
1227 * - Application is running on a different CPU, between reserve
1228 * and commit ring buffer operations,
1229 * - Application is preempted between reserve and commit ring
1230 * buffer operations,
1232 * Long-term situation:
1233 * - Application is stopped (SIGSTOP) between reserve and commit
1234 * ring buffer operations. Could eventually be resumed by
1236 * - Application is killed (SIGTERM, SIGINT, SIGKILL) between
1237 * reserve and commit ring buffer operation.
1239 * From a consumer perspective, handling short-term
1240 * unavailability situations is performed by retrying a few
1241 * times after a delay. Handling long-term unavailability
1242 * situations is handled by failing to get the sub-buffer.
1244 * In all of those situations, if the application is taking a
1245 * long time to perform its commit after ring buffer space
1246 * reservation, we can end up in a situation where the producer
1247 * will fill the ring buffer and try to write into the same
1248 * sub-buffer again (which has a missing commit). This is
1249 * handled by the producer in the sub-buffer switch handling
1250 * code of the reserve routine by detecting unbalanced
1251 * reserve/commit counters and discarding all further events
1252 * until the situation is resolved in those situations. Two
1253 * scenarios can occur:
1255 * 1) The application causing the reserve/commit counters to be
1256 * unbalanced has been terminated. In this situation, all
1257 * further events will be discarded in the buffers, and no
1258 * further buffer data will be readable by the consumer
1259 * daemon. Tearing down the UST tracing session and starting
1260 * anew is a work-around for those situations. Note that this
1261 * only affects per-UID tracing. In per-PID tracing, the
1262 * application vanishes with the termination, and therefore
1263 * no more data needs to be written to the buffers.
1264 * 2) The application causing the unbalance has been delayed for
1265 * a long time, but will eventually try to increment the
1266 * commit counter after eventually writing to the sub-buffer.
1267 * This situation can cause events to be discarded until the
1268 * application resumes its operations.
1270 if (((commit_count
- chan
->backend
.subbuf_size
)
1271 & chan
->commit_count_mask
)
1272 - (buf_trunc(consumed
, chan
)
1273 >> chan
->backend
.num_subbuf_order
)
1275 if (nr_retry
-- > 0) {
1276 if (nr_retry
<= (LTTNG_UST_RING_BUFFER_GET_RETRY
>> 1))
1277 (void) poll(NULL
, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS
);
1285 * Check that we are not about to read the same subbuffer in
1286 * which the writer head is.
1288 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed
, chan
)
1293 * Failure to get the subbuffer causes a busy-loop retry without going
1294 * to a wait queue. These are caused by short-lived race windows where
1295 * the writer is getting access to a subbuffer we were trying to get
1296 * access to. Also checks that the "consumed" buffer count we are
1297 * looking for matches the one contained in the subbuffer id.
1299 * The short-lived race window described here can be affected by
1300 * application signals and preemption, thus requiring to bound
1301 * the loop to a maximum number of retry.
1303 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1304 consumed_idx
, buf_trunc_val(consumed
, chan
),
1307 if (nr_retry
-- > 0) {
1308 if (nr_retry
<= (LTTNG_UST_RING_BUFFER_GET_RETRY
>> 1))
1309 (void) poll(NULL
, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS
);
1315 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1317 buf
->get_subbuf_consumed
= consumed
;
1318 buf
->get_subbuf
= 1;
1324 * The memory barriers __wait_event()/wake_up_interruptible() take care
1325 * of "raw_spin_is_locked" memory ordering.
1334 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1337 void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
1338 struct lttng_ust_shm_handle
*handle
)
1340 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1341 struct channel
*chan
= shmp(handle
, bufb
->chan
);
1342 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1343 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
1345 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1347 if (!buf
->get_subbuf
) {
1349 * Reader puts a subbuffer it did not get.
1351 CHAN_WARN_ON(chan
, 1);
1354 consumed
= buf
->get_subbuf_consumed
;
1355 buf
->get_subbuf
= 0;
1358 * Clear the records_unread counter. (overruns counter)
1359 * Can still be non-zero if a file reader simply grabbed the data
1360 * without using iterators.
1361 * Can be below zero if an iterator is used on a snapshot more than
1364 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1365 v_add(config
, v_read(config
,
1366 &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
),
1367 &bufb
->records_read
);
1368 v_set(config
, &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
, 0);
1369 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1370 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1371 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1374 * Exchange the reader subbuffer with the one we put in its place in the
1375 * writer subbuffer table. Expect the original consumed count. If
1376 * update_read_sb_index fails, this is because the writer updated the
1377 * subbuffer concurrently. We should therefore keep the subbuffer we
1378 * currently have: it has become invalid to try reading this sub-buffer
1379 * consumed count value anyway.
1381 consumed_idx
= subbuf_index(consumed
, chan
);
1382 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1383 consumed_idx
, buf_trunc_val(consumed
, chan
),
1386 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1387 * if the writer concurrently updated it.
1392 * cons_offset is an iterator on all subbuffer offsets between the reader
1393 * position and the writer position. (inclusive)
1396 void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
1397 struct channel
*chan
,
1398 unsigned long cons_offset
,
1400 struct lttng_ust_shm_handle
*handle
)
1402 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1403 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1405 cons_idx
= subbuf_index(cons_offset
, chan
);
1406 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, cons_idx
)->cc
);
1407 commit_count_sb
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, cons_idx
)->cc_sb
);
1409 if (subbuf_offset(commit_count
, chan
) != 0)
1410 DBG("ring buffer %s, cpu %d: "
1411 "commit count in subbuffer %lu,\n"
1412 "expecting multiples of %lu bytes\n"
1413 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1414 chan
->backend
.name
, cpu
, cons_idx
,
1415 chan
->backend
.subbuf_size
,
1416 commit_count
, commit_count_sb
);
1418 DBG("ring buffer: %s, cpu %d: %lu bytes committed\n",
1419 chan
->backend
.name
, cpu
, commit_count
);
1423 void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
1424 struct channel
*chan
,
1425 void *priv
, int cpu
,
1426 struct lttng_ust_shm_handle
*handle
)
1428 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1429 unsigned long write_offset
, cons_offset
;
1432 * No need to order commit_count, write_offset and cons_offset reads
1433 * because we execute at teardown when no more writer nor reader
1434 * references are left.
1436 write_offset
= v_read(config
, &buf
->offset
);
1437 cons_offset
= uatomic_read(&buf
->consumed
);
1438 if (write_offset
!= cons_offset
)
1439 DBG("ring buffer %s, cpu %d: "
1440 "non-consumed data\n"
1441 " [ %lu bytes written, %lu bytes read ]\n",
1442 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1444 for (cons_offset
= uatomic_read(&buf
->consumed
);
1445 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1448 cons_offset
= subbuf_align(cons_offset
, chan
))
1449 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1454 void lib_ring_buffer_print_errors(struct channel
*chan
,
1455 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
1456 struct lttng_ust_shm_handle
*handle
)
1458 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1459 void *priv
= channel_get_private(chan
);
1461 if (!strcmp(chan
->backend
.name
, "relay-metadata-mmap")) {
1462 DBG("ring buffer %s: %lu records written, "
1463 "%lu records overrun\n",
1465 v_read(config
, &buf
->records_count
),
1466 v_read(config
, &buf
->records_overrun
));
1468 DBG("ring buffer %s, cpu %d: %lu records written, "
1469 "%lu records overrun\n",
1470 chan
->backend
.name
, cpu
,
1471 v_read(config
, &buf
->records_count
),
1472 v_read(config
, &buf
->records_overrun
));
1474 if (v_read(config
, &buf
->records_lost_full
)
1475 || v_read(config
, &buf
->records_lost_wrap
)
1476 || v_read(config
, &buf
->records_lost_big
))
1477 DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n"
1478 " [ %lu buffer full, %lu nest buffer wrap-around, "
1479 "%lu event too big ]\n",
1480 chan
->backend
.name
, cpu
,
1481 v_read(config
, &buf
->records_lost_full
),
1482 v_read(config
, &buf
->records_lost_wrap
),
1483 v_read(config
, &buf
->records_lost_big
));
1485 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
, handle
);
1489 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1491 * Only executed when the buffer is finalized, in SWITCH_FLUSH.
1494 void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer
*buf
,
1495 struct channel
*chan
,
1496 struct switch_offsets
*offsets
,
1498 struct lttng_ust_shm_handle
*handle
)
1500 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1501 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1502 unsigned long commit_count
;
1504 config
->cb
.buffer_begin(buf
, tsc
, oldidx
, handle
);
1507 * Order all writes to buffer before the commit count update that will
1508 * determine that the subbuffer is full.
1511 v_add(config
, config
->cb
.subbuffer_header_size(),
1512 &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1513 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1514 /* Check if the written buffer has to be delivered */
1515 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1516 commit_count
, oldidx
, handle
, tsc
);
1517 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1518 offsets
->old
+ config
->cb
.subbuffer_header_size(),
1519 commit_count
, handle
);
1523 * lib_ring_buffer_switch_old_end: switch old subbuffer
1525 * Note : offset_old should never be 0 here. It is ok, because we never perform
1526 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1527 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1531 void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer
*buf
,
1532 struct channel
*chan
,
1533 struct switch_offsets
*offsets
,
1535 struct lttng_ust_shm_handle
*handle
)
1537 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1538 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1539 unsigned long commit_count
, padding_size
, data_size
;
1541 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1542 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1543 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
,
1547 * Order all writes to buffer before the commit count update that will
1548 * determine that the subbuffer is full.
1551 v_add(config
, padding_size
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1552 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1553 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1554 commit_count
, oldidx
, handle
, tsc
);
1555 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1556 offsets
->old
+ padding_size
, commit_count
, handle
);
1560 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1562 * This code can be executed unordered : writers may already have written to the
1563 * sub-buffer before this code gets executed, caution. The commit makes sure
1564 * that this code is executed before the deliver of this sub-buffer.
1567 void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer
*buf
,
1568 struct channel
*chan
,
1569 struct switch_offsets
*offsets
,
1571 struct lttng_ust_shm_handle
*handle
)
1573 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1574 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1575 unsigned long commit_count
;
1577 config
->cb
.buffer_begin(buf
, tsc
, beginidx
, handle
);
1580 * Order all writes to buffer before the commit count update that will
1581 * determine that the subbuffer is full.
1584 v_add(config
, config
->cb
.subbuffer_header_size(),
1585 &shmp_index(handle
, buf
->commit_hot
, beginidx
)->cc
);
1586 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, beginidx
)->cc
);
1587 /* Check if the written buffer has to be delivered */
1588 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1589 commit_count
, beginidx
, handle
, tsc
);
1590 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, beginidx
,
1591 offsets
->begin
+ config
->cb
.subbuffer_header_size(),
1592 commit_count
, handle
);
1596 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1598 * Calls subbuffer_set_data_size() to set the data size of the current
1599 * sub-buffer. We do not need to perform check_deliver nor commit here,
1600 * since this task will be done by the "commit" of the event for which
1601 * we are currently doing the space reservation.
1604 void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer
*buf
,
1605 struct channel
*chan
,
1606 struct switch_offsets
*offsets
,
1608 struct lttng_ust_shm_handle
*handle
)
1610 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1611 unsigned long endidx
, data_size
;
1613 endidx
= subbuf_index(offsets
->end
- 1, chan
);
1614 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1615 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
,
1622 * !0 if execution must be aborted.
1625 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1626 struct lttng_ust_lib_ring_buffer
*buf
,
1627 struct channel
*chan
,
1628 struct switch_offsets
*offsets
,
1630 struct lttng_ust_shm_handle
*handle
)
1632 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1633 unsigned long off
, reserve_commit_diff
;
1635 offsets
->begin
= v_read(config
, &buf
->offset
);
1636 offsets
->old
= offsets
->begin
;
1637 offsets
->switch_old_start
= 0;
1638 off
= subbuf_offset(offsets
->begin
, chan
);
1640 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1643 * Ensure we flush the header of an empty subbuffer when doing the
1644 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1645 * total data gathering duration even if there were no records saved
1646 * after the last buffer switch.
1647 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1648 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1649 * subbuffer header as appropriate.
1650 * The next record that reserves space will be responsible for
1651 * populating the following subbuffer header. We choose not to populate
1652 * the next subbuffer header here because we want to be able to use
1653 * SWITCH_ACTIVE for periodical buffer flush, which must
1654 * guarantee that all the buffer content (records and header
1655 * timestamps) are visible to the reader. This is required for
1656 * quiescence guarantees for the fusion merge.
1658 if (mode
!= SWITCH_FLUSH
&& !off
)
1659 return -1; /* we do not have to switch : buffer is empty */
1661 if (caa_unlikely(off
== 0)) {
1662 unsigned long sb_index
, commit_count
;
1665 * We are performing a SWITCH_FLUSH. At this stage, there are no
1666 * concurrent writes into the buffer.
1668 * The client does not save any header information. Don't
1669 * switch empty subbuffer on finalize, because it is invalid to
1670 * deliver a completely empty subbuffer.
1672 if (!config
->cb
.subbuffer_header_size())
1675 /* Test new buffer integrity */
1676 sb_index
= subbuf_index(offsets
->begin
, chan
);
1677 commit_count
= v_read(config
,
1678 &shmp_index(handle
, buf
->commit_cold
,
1680 reserve_commit_diff
=
1681 (buf_trunc(offsets
->begin
, chan
)
1682 >> chan
->backend
.num_subbuf_order
)
1683 - (commit_count
& chan
->commit_count_mask
);
1684 if (caa_likely(reserve_commit_diff
== 0)) {
1685 /* Next subbuffer not being written to. */
1686 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1687 subbuf_trunc(offsets
->begin
, chan
)
1688 - subbuf_trunc((unsigned long)
1689 uatomic_read(&buf
->consumed
), chan
)
1690 >= chan
->backend
.buf_size
)) {
1692 * We do not overwrite non consumed buffers
1693 * and we are full : don't switch.
1698 * Next subbuffer not being written to, and we
1699 * are either in overwrite mode or the buffer is
1700 * not full. It's safe to write in this new
1706 * Next subbuffer reserve offset does not match the
1707 * commit offset. Don't perform switch in
1708 * producer-consumer and overwrite mode. Caused by
1709 * either a writer OOPS or too many nested writes over a
1710 * reserve/commit pair.
1716 * Need to write the subbuffer start header on finalize.
1718 offsets
->switch_old_start
= 1;
1720 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1721 /* Note: old points to the next subbuf at offset 0 */
1722 offsets
->end
= offsets
->begin
;
1727 * Force a sub-buffer switch. This operation is completely reentrant : can be
1728 * called while tracing is active with absolutely no lock held.
1730 * Note, however, that as a v_cmpxchg is used for some atomic
1731 * operations, this function must be called from the CPU which owns the buffer
1732 * for a ACTIVE flush.
1734 void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer
*buf
, enum switch_mode mode
,
1735 struct lttng_ust_shm_handle
*handle
)
1737 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1738 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1739 struct switch_offsets offsets
;
1740 unsigned long oldidx
;
1746 * Perform retryable operations.
1749 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1751 return; /* Switch not needed */
1752 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1756 * Atomically update last_tsc. This update races against concurrent
1757 * atomic updates, but the race will always cause supplementary full TSC
1758 * records, never the opposite (missing a full TSC record when it would
1761 save_last_tsc(config
, buf
, tsc
);
1764 * Push the reader if necessary
1766 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1768 oldidx
= subbuf_index(offsets
.old
, chan
);
1769 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
, handle
);
1772 * May need to populate header start on SWITCH_FLUSH.
1774 if (offsets
.switch_old_start
) {
1775 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
, handle
);
1776 offsets
.old
+= config
->cb
.subbuffer_header_size();
1780 * Switch old subbuffer.
1782 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
, handle
);
1788 * -ENOSPC if event size is too large for packet.
1789 * -ENOBUFS if there is currently not enough space in buffer for the event.
1790 * -EIO if data cannot be written into the buffer for any other reason.
1793 int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer
*buf
,
1794 struct channel
*chan
,
1795 struct switch_offsets
*offsets
,
1796 struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
1798 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1799 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
1800 unsigned long reserve_commit_diff
, offset_cmp
;
1803 offsets
->begin
= offset_cmp
= v_read(config
, &buf
->offset
);
1804 offsets
->old
= offsets
->begin
;
1805 offsets
->switch_new_start
= 0;
1806 offsets
->switch_new_end
= 0;
1807 offsets
->switch_old_end
= 0;
1808 offsets
->pre_header_padding
= 0;
1810 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1811 if ((int64_t) ctx
->tsc
== -EIO
)
1814 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1815 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
1817 if (caa_unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1818 offsets
->switch_new_start
= 1; /* For offsets->begin */
1820 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1822 &offsets
->pre_header_padding
,
1825 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1828 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
) +
1829 offsets
->size
> chan
->backend
.subbuf_size
)) {
1830 offsets
->switch_old_end
= 1; /* For offsets->old */
1831 offsets
->switch_new_start
= 1; /* For offsets->begin */
1834 if (caa_unlikely(offsets
->switch_new_start
)) {
1835 unsigned long sb_index
, commit_count
;
1838 * We are typically not filling the previous buffer completely.
1840 if (caa_likely(offsets
->switch_old_end
))
1841 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1842 offsets
->begin
= offsets
->begin
1843 + config
->cb
.subbuffer_header_size();
1844 /* Test new buffer integrity */
1845 sb_index
= subbuf_index(offsets
->begin
, chan
);
1847 * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
1848 * lib_ring_buffer_check_deliver() has the matching
1849 * memory barriers required around commit_cold cc_sb
1850 * updates to ensure reserve and commit counter updates
1851 * are not seen reordered when updated by another CPU.
1854 commit_count
= v_read(config
,
1855 &shmp_index(handle
, buf
->commit_cold
,
1857 /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
1859 if (caa_unlikely(offset_cmp
!= v_read(config
, &buf
->offset
))) {
1861 * The reserve counter have been concurrently updated
1862 * while we read the commit counter. This means the
1863 * commit counter we read might not match buf->offset
1864 * due to concurrent update. We therefore need to retry.
1868 reserve_commit_diff
=
1869 (buf_trunc(offsets
->begin
, chan
)
1870 >> chan
->backend
.num_subbuf_order
)
1871 - (commit_count
& chan
->commit_count_mask
);
1872 if (caa_likely(reserve_commit_diff
== 0)) {
1873 /* Next subbuffer not being written to. */
1874 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1875 subbuf_trunc(offsets
->begin
, chan
)
1876 - subbuf_trunc((unsigned long)
1877 uatomic_read(&buf
->consumed
), chan
)
1878 >= chan
->backend
.buf_size
)) {
1879 unsigned long nr_lost
;
1882 * We do not overwrite non consumed buffers
1883 * and we are full : record is lost.
1885 nr_lost
= v_read(config
, &buf
->records_lost_full
);
1886 v_inc(config
, &buf
->records_lost_full
);
1887 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
1888 DBG("%lu or more records lost in (%s:%d) (buffer full)\n",
1889 nr_lost
+ 1, chan
->backend
.name
,
1895 * Next subbuffer not being written to, and we
1896 * are either in overwrite mode or the buffer is
1897 * not full. It's safe to write in this new
1902 unsigned long nr_lost
;
1905 * Next subbuffer reserve offset does not match the
1906 * commit offset, and this did not involve update to the
1907 * reserve counter. Drop record in producer-consumer and
1908 * overwrite mode. Caused by either a writer OOPS or too
1909 * many nested writes over a reserve/commit pair.
1911 nr_lost
= v_read(config
, &buf
->records_lost_wrap
);
1912 v_inc(config
, &buf
->records_lost_wrap
);
1913 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
1914 DBG("%lu or more records lost in (%s:%d) (wrap-around)\n",
1915 nr_lost
+ 1, chan
->backend
.name
,
1921 config
->cb
.record_header_size(config
, chan
,
1923 &offsets
->pre_header_padding
,
1926 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1929 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
)
1930 + offsets
->size
> chan
->backend
.subbuf_size
)) {
1931 unsigned long nr_lost
;
1934 * Record too big for subbuffers, report error, don't
1935 * complete the sub-buffer switch.
1937 nr_lost
= v_read(config
, &buf
->records_lost_big
);
1938 v_inc(config
, &buf
->records_lost_big
);
1939 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
1940 DBG("%lu or more records lost in (%s:%d) record size "
1941 " of %zu bytes is too large for buffer\n",
1942 nr_lost
+ 1, chan
->backend
.name
,
1943 buf
->backend
.cpu
, offsets
->size
);
1948 * We just made a successful buffer switch and the
1949 * record fits in the new subbuffer. Let's write.
1954 * Record fits in the current buffer and we are not on a switch
1955 * boundary. It's safe to write.
1958 offsets
->end
= offsets
->begin
+ offsets
->size
;
1960 if (caa_unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
1962 * The offset_end will fall at the very beginning of the next
1965 offsets
->switch_new_end
= 1; /* For offsets->begin */
1971 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1972 * @ctx: ring buffer context.
1974 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
1975 * -EIO for other errors, else returns 0.
1976 * It will take care of sub-buffer switching.
1978 int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
1980 struct channel
*chan
= ctx
->chan
;
1981 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
1982 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1983 struct lttng_ust_lib_ring_buffer
*buf
;
1984 struct switch_offsets offsets
;
1987 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
1988 buf
= shmp(handle
, chan
->backend
.buf
[ctx
->cpu
].shmp
);
1990 buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
1996 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
1998 if (caa_unlikely(ret
))
2000 } while (caa_unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
2005 * Atomically update last_tsc. This update races against concurrent
2006 * atomic updates, but the race will always cause supplementary full TSC
2007 * records, never the opposite (missing a full TSC record when it would
2010 save_last_tsc(config
, buf
, ctx
->tsc
);
2013 * Push the reader if necessary
2015 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
2018 * Clear noref flag for this subbuffer.
2020 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2021 subbuf_index(offsets
.end
- 1, chan
),
2025 * Switch old subbuffer if needed.
2027 if (caa_unlikely(offsets
.switch_old_end
)) {
2028 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2029 subbuf_index(offsets
.old
- 1, chan
),
2031 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2035 * Populate new subbuffer.
2037 if (caa_unlikely(offsets
.switch_new_start
))
2038 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2040 if (caa_unlikely(offsets
.switch_new_end
))
2041 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2043 ctx
->slot_size
= offsets
.size
;
2044 ctx
->pre_offset
= offsets
.begin
;
2045 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
2050 * Force a read (imply TLS fixup for dlopen) of TLS variables.
2052 void lttng_fixup_ringbuffer_tls(void)
2054 asm volatile ("" : : "m" (URCU_TLS(lib_ring_buffer_nesting
)));
2057 void lib_ringbuffer_signal_init(void)
2063 * Block signal for entire process, so only our thread processes
2067 ret
= pthread_sigmask(SIG_BLOCK
, &mask
, NULL
);
2070 PERROR("pthread_sigmask");