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 */
78 #define max(a, b) ((a) > (b) ? (a) : (b))
81 /* Print DBG() messages about events lost only every 1048576 hits */
82 #define DBG_PRINT_NR_LOST (1UL << 20)
84 #define LTTNG_UST_RB_SIG_FLUSH SIGRTMIN
85 #define LTTNG_UST_RB_SIG_READ SIGRTMIN + 1
86 #define LTTNG_UST_RB_SIG_TEARDOWN SIGRTMIN + 2
87 #define CLOCKID CLOCK_MONOTONIC
88 #define LTTNG_UST_RING_BUFFER_GET_RETRY 10
89 #define LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS 10
92 * Non-static to ensure the compiler does not optimize away the xor.
94 uint8_t lttng_crash_magic_xor
[] = RB_CRASH_DUMP_ABI_MAGIC_XOR
;
97 * Use POSIX SHM: shm_open(3) and shm_unlink(3).
98 * close(2) to close the fd returned by shm_open.
99 * shm_unlink releases the shared memory object name.
100 * ftruncate(2) sets the size of the memory object.
101 * mmap/munmap maps the shared memory obj to a virtual address in the
102 * calling proceess (should be done both in libust and consumer).
103 * See shm_overview(7) for details.
104 * Pass file descriptor returned by shm_open(3) to ltt-sessiond through
107 * Since we don't need to access the object using its name, we can
108 * immediately shm_unlink(3) it, and only keep the handle with its file
113 * Internal structure representing offsets to use at a sub-buffer switch.
115 struct switch_offsets
{
116 unsigned long begin
, end
, old
;
117 size_t pre_header_padding
, size
;
118 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
122 DEFINE_URCU_TLS(unsigned int, lib_ring_buffer_nesting
);
125 * wakeup_fd_mutex protects wakeup fd use by timer from concurrent
128 static pthread_mutex_t wakeup_fd_mutex
= PTHREAD_MUTEX_INITIALIZER
;
131 void lib_ring_buffer_print_errors(struct channel
*chan
,
132 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
133 struct lttng_ust_shm_handle
*handle
);
136 * Handle timer teardown race wrt memory free of private data by
137 * ring buffer signals are handled by a single thread, which permits
138 * a synchronization point between handling of each signal.
139 * Protected by the lock within the structure.
141 struct timer_signal_data
{
142 pthread_t tid
; /* thread id managing signals */
145 pthread_mutex_t lock
;
148 static struct timer_signal_data timer_signal
= {
152 .lock
= PTHREAD_MUTEX_INITIALIZER
,
156 * lib_ring_buffer_reset - Reset ring buffer to initial values.
159 * Effectively empty the ring buffer. Should be called when the buffer is not
160 * used for writing. The ring buffer can be opened for reading, but the reader
161 * should not be using the iterator concurrently with reset. The previous
162 * current iterator record is reset.
164 void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer
*buf
,
165 struct lttng_ust_shm_handle
*handle
)
167 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
168 const struct lttng_ust_lib_ring_buffer_config
*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
);
403 if (uatomic_read(&buf
->active_readers
))
404 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
,
408 struct lttng_ust_lib_ring_buffer
*buf
=
409 shmp(handle
, chan
->backend
.buf
[0].shmp
);
411 if (uatomic_read(&buf
->active_readers
))
412 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
,
415 pthread_mutex_unlock(&wakeup_fd_mutex
);
420 void lib_ring_buffer_channel_do_read(struct channel
*chan
)
422 const struct lttng_ust_lib_ring_buffer_config
*config
;
423 struct lttng_ust_shm_handle
*handle
;
426 handle
= chan
->handle
;
427 config
= &chan
->backend
.config
;
430 * Only flush buffers periodically if readers are active.
432 pthread_mutex_lock(&wakeup_fd_mutex
);
433 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
434 for_each_possible_cpu(cpu
) {
435 struct lttng_ust_lib_ring_buffer
*buf
=
436 shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
438 if (uatomic_read(&buf
->active_readers
)
439 && lib_ring_buffer_poll_deliver(config
, buf
,
441 lib_ring_buffer_wakeup(buf
, handle
);
445 struct lttng_ust_lib_ring_buffer
*buf
=
446 shmp(handle
, chan
->backend
.buf
[0].shmp
);
448 if (uatomic_read(&buf
->active_readers
)
449 && lib_ring_buffer_poll_deliver(config
, buf
,
451 lib_ring_buffer_wakeup(buf
, handle
);
454 pthread_mutex_unlock(&wakeup_fd_mutex
);
458 void lib_ring_buffer_channel_read_timer(int sig
, siginfo_t
*si
, void *uc
)
460 struct channel
*chan
;
462 assert(CMM_LOAD_SHARED(timer_signal
.tid
) == pthread_self());
463 chan
= si
->si_value
.sival_ptr
;
464 DBG("Read timer for channel %p\n", chan
);
465 lib_ring_buffer_channel_do_read(chan
);
470 void rb_setmask(sigset_t
*mask
)
474 ret
= sigemptyset(mask
);
476 PERROR("sigemptyset");
478 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_FLUSH
);
482 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_READ
);
486 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_TEARDOWN
);
493 void *sig_thread(void *arg
)
499 /* Only self thread will receive signal mask. */
501 CMM_STORE_SHARED(timer_signal
.tid
, pthread_self());
504 signr
= sigwaitinfo(&mask
, &info
);
507 PERROR("sigwaitinfo");
510 if (signr
== LTTNG_UST_RB_SIG_FLUSH
) {
511 lib_ring_buffer_channel_switch_timer(info
.si_signo
,
513 } else if (signr
== LTTNG_UST_RB_SIG_READ
) {
514 lib_ring_buffer_channel_read_timer(info
.si_signo
,
516 } else if (signr
== LTTNG_UST_RB_SIG_TEARDOWN
) {
518 CMM_STORE_SHARED(timer_signal
.qs_done
, 1);
521 ERR("Unexptected signal %d\n", info
.si_signo
);
528 * Ensure only a single thread listens on the timer signal.
531 void lib_ring_buffer_setup_timer_thread(void)
536 pthread_mutex_lock(&timer_signal
.lock
);
537 if (timer_signal
.setup_done
)
540 ret
= pthread_create(&thread
, NULL
, &sig_thread
, NULL
);
543 PERROR("pthread_create");
545 ret
= pthread_detach(thread
);
548 PERROR("pthread_detach");
550 timer_signal
.setup_done
= 1;
552 pthread_mutex_unlock(&timer_signal
.lock
);
556 * Wait for signal-handling thread quiescent state.
559 void lib_ring_buffer_wait_signal_thread_qs(unsigned int signr
)
561 sigset_t pending_set
;
565 * We need to be the only thread interacting with the thread
566 * that manages signals for teardown synchronization.
568 pthread_mutex_lock(&timer_signal
.lock
);
571 * Ensure we don't have any signal queued for this channel.
574 ret
= sigemptyset(&pending_set
);
576 PERROR("sigemptyset");
578 ret
= sigpending(&pending_set
);
580 PERROR("sigpending");
582 if (!sigismember(&pending_set
, signr
))
588 * From this point, no new signal handler will be fired that
589 * would try to access "chan". However, we still need to wait
590 * for any currently executing handler to complete.
593 CMM_STORE_SHARED(timer_signal
.qs_done
, 0);
597 * Kill with LTTNG_UST_RB_SIG_TEARDOWN, so signal management
600 kill(getpid(), LTTNG_UST_RB_SIG_TEARDOWN
);
602 while (!CMM_LOAD_SHARED(timer_signal
.qs_done
))
606 pthread_mutex_unlock(&timer_signal
.lock
);
610 void lib_ring_buffer_channel_switch_timer_start(struct channel
*chan
)
613 struct itimerspec its
;
616 if (!chan
->switch_timer_interval
|| chan
->switch_timer_enabled
)
619 chan
->switch_timer_enabled
= 1;
621 lib_ring_buffer_setup_timer_thread();
623 sev
.sigev_notify
= SIGEV_SIGNAL
;
624 sev
.sigev_signo
= LTTNG_UST_RB_SIG_FLUSH
;
625 sev
.sigev_value
.sival_ptr
= chan
;
626 ret
= timer_create(CLOCKID
, &sev
, &chan
->switch_timer
);
628 PERROR("timer_create");
631 its
.it_value
.tv_sec
= chan
->switch_timer_interval
/ 1000000;
632 its
.it_value
.tv_nsec
= chan
->switch_timer_interval
% 1000000;
633 its
.it_interval
.tv_sec
= its
.it_value
.tv_sec
;
634 its
.it_interval
.tv_nsec
= its
.it_value
.tv_nsec
;
636 ret
= timer_settime(chan
->switch_timer
, 0, &its
, NULL
);
638 PERROR("timer_settime");
643 void lib_ring_buffer_channel_switch_timer_stop(struct channel
*chan
)
647 if (!chan
->switch_timer_interval
|| !chan
->switch_timer_enabled
)
650 ret
= timer_delete(chan
->switch_timer
);
652 PERROR("timer_delete");
655 lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_FLUSH
);
657 chan
->switch_timer
= 0;
658 chan
->switch_timer_enabled
= 0;
662 void lib_ring_buffer_channel_read_timer_start(struct channel
*chan
)
664 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
666 struct itimerspec its
;
669 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
670 || !chan
->read_timer_interval
|| chan
->read_timer_enabled
)
673 chan
->read_timer_enabled
= 1;
675 lib_ring_buffer_setup_timer_thread();
677 sev
.sigev_notify
= SIGEV_SIGNAL
;
678 sev
.sigev_signo
= LTTNG_UST_RB_SIG_READ
;
679 sev
.sigev_value
.sival_ptr
= chan
;
680 ret
= timer_create(CLOCKID
, &sev
, &chan
->read_timer
);
682 PERROR("timer_create");
685 its
.it_value
.tv_sec
= chan
->read_timer_interval
/ 1000000;
686 its
.it_value
.tv_nsec
= chan
->read_timer_interval
% 1000000;
687 its
.it_interval
.tv_sec
= its
.it_value
.tv_sec
;
688 its
.it_interval
.tv_nsec
= its
.it_value
.tv_nsec
;
690 ret
= timer_settime(chan
->read_timer
, 0, &its
, NULL
);
692 PERROR("timer_settime");
697 void lib_ring_buffer_channel_read_timer_stop(struct channel
*chan
)
699 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
702 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
703 || !chan
->read_timer_interval
|| !chan
->read_timer_enabled
)
706 ret
= timer_delete(chan
->read_timer
);
708 PERROR("timer_delete");
712 * do one more check to catch data that has been written in the last
715 lib_ring_buffer_channel_do_read(chan
);
717 lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_READ
);
719 chan
->read_timer
= 0;
720 chan
->read_timer_enabled
= 0;
723 static void channel_unregister_notifiers(struct channel
*chan
,
724 struct lttng_ust_shm_handle
*handle
)
726 lib_ring_buffer_channel_switch_timer_stop(chan
);
727 lib_ring_buffer_channel_read_timer_stop(chan
);
730 static void channel_print_errors(struct channel
*chan
,
731 struct lttng_ust_shm_handle
*handle
)
733 const struct lttng_ust_lib_ring_buffer_config
*config
=
734 &chan
->backend
.config
;
737 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
738 for_each_possible_cpu(cpu
) {
739 struct lttng_ust_lib_ring_buffer
*buf
=
740 shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
741 lib_ring_buffer_print_errors(chan
, buf
, cpu
, handle
);
744 struct lttng_ust_lib_ring_buffer
*buf
=
745 shmp(handle
, chan
->backend
.buf
[0].shmp
);
747 lib_ring_buffer_print_errors(chan
, buf
, -1, handle
);
751 static void channel_free(struct channel
*chan
,
752 struct lttng_ust_shm_handle
*handle
)
754 channel_backend_free(&chan
->backend
, handle
);
755 /* chan is freed by shm teardown */
756 shm_object_table_destroy(handle
->table
);
761 * channel_create - Create channel.
762 * @config: ring buffer instance configuration
763 * @name: name of the channel
764 * @priv_data: ring buffer client private data area pointer (output)
765 * @priv_data_size: length, in bytes, of the private data area.
766 * @priv_data_init: initialization data for private data.
767 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
768 * address mapping. It is used only by RING_BUFFER_STATIC
769 * configuration. It can be set to NULL for other backends.
770 * @subbuf_size: subbuffer size
771 * @num_subbuf: number of subbuffers
772 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
773 * padding to let readers get those sub-buffers.
774 * Used for live streaming.
775 * @read_timer_interval: Time interval (in us) to wake up pending readers.
776 * @shm_path: Shared memory files path.
779 * Returns NULL on failure.
781 struct lttng_ust_shm_handle
*channel_create(const struct lttng_ust_lib_ring_buffer_config
*config
,
784 size_t priv_data_align
,
785 size_t priv_data_size
,
786 void *priv_data_init
,
787 void *buf_addr
, size_t subbuf_size
,
788 size_t num_subbuf
, unsigned int switch_timer_interval
,
789 unsigned int read_timer_interval
,
790 const char *shm_path
)
793 size_t shmsize
, chansize
;
794 struct channel
*chan
;
795 struct lttng_ust_shm_handle
*handle
;
796 struct shm_object
*shmobj
;
797 unsigned int nr_streams
;
799 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
800 nr_streams
= num_possible_cpus();
804 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
805 read_timer_interval
))
808 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
812 /* Allocate table for channel + per-cpu buffers */
813 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
815 goto error_table_alloc
;
817 /* Calculate the shm allocation layout */
818 shmsize
= sizeof(struct channel
);
819 shmsize
+= offset_align(shmsize
, __alignof__(struct lttng_ust_lib_ring_buffer_shmp
));
820 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
) * nr_streams
;
823 shmsize
+= offset_align(shmsize
, priv_data_align
);
824 shmsize
+= priv_data_size
;
826 /* Allocate normal memory for channel (not shared) */
827 shmobj
= shm_object_table_alloc(handle
->table
, shmsize
, SHM_OBJECT_MEM
,
831 /* struct channel is at object 0, offset 0 (hardcoded) */
832 set_shmp(handle
->chan
, zalloc_shm(shmobj
, chansize
));
833 assert(handle
->chan
._ref
.index
== 0);
834 assert(handle
->chan
._ref
.offset
== 0);
835 chan
= shmp(handle
, handle
->chan
);
838 chan
->nr_streams
= nr_streams
;
840 /* space for private data */
841 if (priv_data_size
) {
842 DECLARE_SHMP(void, priv_data_alloc
);
844 align_shm(shmobj
, priv_data_align
);
845 chan
->priv_data_offset
= shmobj
->allocated_len
;
846 set_shmp(priv_data_alloc
, zalloc_shm(shmobj
, priv_data_size
));
847 if (!shmp(handle
, priv_data_alloc
))
849 *priv_data
= channel_get_private(chan
);
850 memcpy(*priv_data
, priv_data_init
, priv_data_size
);
852 chan
->priv_data_offset
= -1;
857 ret
= channel_backend_init(&chan
->backend
, name
, config
,
858 subbuf_size
, num_subbuf
, handle
,
861 goto error_backend_init
;
863 chan
->handle
= handle
;
864 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
866 chan
->switch_timer_interval
= switch_timer_interval
;
867 chan
->read_timer_interval
= read_timer_interval
;
868 lib_ring_buffer_channel_switch_timer_start(chan
);
869 lib_ring_buffer_channel_read_timer_start(chan
);
875 shm_object_table_destroy(handle
->table
);
881 struct lttng_ust_shm_handle
*channel_handle_create(void *data
,
882 uint64_t memory_map_size
,
885 struct lttng_ust_shm_handle
*handle
;
886 struct shm_object
*object
;
888 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
892 /* Allocate table for channel + per-cpu buffers */
893 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
895 goto error_table_alloc
;
896 /* Add channel object */
897 object
= shm_object_table_append_mem(handle
->table
, data
,
898 memory_map_size
, wakeup_fd
);
900 goto error_table_object
;
901 /* struct channel is at object 0, offset 0 (hardcoded) */
902 handle
->chan
._ref
.index
= 0;
903 handle
->chan
._ref
.offset
= 0;
907 shm_object_table_destroy(handle
->table
);
913 int channel_handle_add_stream(struct lttng_ust_shm_handle
*handle
,
914 int shm_fd
, int wakeup_fd
, uint32_t stream_nr
,
915 uint64_t memory_map_size
)
917 struct shm_object
*object
;
919 /* Add stream object */
920 object
= shm_object_table_append_shm(handle
->table
,
921 shm_fd
, wakeup_fd
, stream_nr
,
928 unsigned int channel_handle_get_nr_streams(struct lttng_ust_shm_handle
*handle
)
930 assert(handle
->table
);
931 return handle
->table
->allocated_len
- 1;
935 void channel_release(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
)
937 channel_free(chan
, handle
);
941 * channel_destroy - Finalize, wait for q.s. and destroy channel.
942 * @chan: channel to destroy
945 * Call "destroy" callback, finalize channels, decrement the channel
946 * reference count. Note that when readers have completed data
947 * consumption of finalized channels, get_subbuf() will return -ENODATA.
948 * They should release their handle at that point.
950 void channel_destroy(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
955 * Note: the consumer takes care of finalizing and
956 * switching the buffers.
958 channel_unregister_notifiers(chan
, handle
);
960 * The consumer prints errors.
962 channel_print_errors(chan
, handle
);
966 * sessiond/consumer are keeping a reference on the shm file
967 * descriptor directly. No need to refcount.
969 channel_release(chan
, handle
);
973 struct lttng_ust_lib_ring_buffer
*channel_get_ring_buffer(
974 const struct lttng_ust_lib_ring_buffer_config
*config
,
975 struct channel
*chan
, int cpu
,
976 struct lttng_ust_shm_handle
*handle
,
977 int *shm_fd
, int *wait_fd
,
979 uint64_t *memory_map_size
)
983 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
986 if (cpu
>= num_possible_cpus())
989 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
990 *shm_fd
= shm_get_shm_fd(handle
, ref
);
991 *wait_fd
= shm_get_wait_fd(handle
, ref
);
992 *wakeup_fd
= shm_get_wakeup_fd(handle
, ref
);
993 if (shm_get_shm_size(handle
, ref
, memory_map_size
))
995 return shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
998 int ring_buffer_channel_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
999 struct channel
*chan
,
1000 struct lttng_ust_shm_handle
*handle
)
1002 struct shm_ref
*ref
;
1004 ref
= &handle
->chan
._ref
;
1005 return shm_close_wait_fd(handle
, ref
);
1008 int ring_buffer_channel_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1009 struct channel
*chan
,
1010 struct lttng_ust_shm_handle
*handle
)
1012 struct shm_ref
*ref
;
1014 ref
= &handle
->chan
._ref
;
1015 return shm_close_wakeup_fd(handle
, ref
);
1018 int ring_buffer_stream_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1019 struct channel
*chan
,
1020 struct lttng_ust_shm_handle
*handle
,
1023 struct shm_ref
*ref
;
1025 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1028 if (cpu
>= num_possible_cpus())
1031 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1032 return shm_close_wait_fd(handle
, ref
);
1035 int ring_buffer_stream_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1036 struct channel
*chan
,
1037 struct lttng_ust_shm_handle
*handle
,
1040 struct shm_ref
*ref
;
1043 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1046 if (cpu
>= num_possible_cpus())
1049 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1050 pthread_mutex_lock(&wakeup_fd_mutex
);
1051 ret
= shm_close_wakeup_fd(handle
, ref
);
1052 pthread_mutex_unlock(&wakeup_fd_mutex
);
1056 int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer
*buf
,
1057 struct lttng_ust_shm_handle
*handle
)
1059 if (uatomic_cmpxchg(&buf
->active_readers
, 0, 1) != 0)
1065 void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer
*buf
,
1066 struct lttng_ust_shm_handle
*handle
)
1068 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1070 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1072 uatomic_dec(&buf
->active_readers
);
1076 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
1078 * @consumed: consumed count indicating the position where to read
1079 * @produced: produced count, indicates position when to stop reading
1081 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1082 * data to read at consumed position, or 0 if the get operation succeeds.
1085 int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer
*buf
,
1086 unsigned long *consumed
, unsigned long *produced
,
1087 struct lttng_ust_shm_handle
*handle
)
1089 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1090 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1091 unsigned long consumed_cur
, write_offset
;
1094 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
1096 * Read finalized before counters.
1099 consumed_cur
= uatomic_read(&buf
->consumed
);
1101 * No need to issue a memory barrier between consumed count read and
1102 * write offset read, because consumed count can only change
1103 * concurrently in overwrite mode, and we keep a sequence counter
1104 * identifier derived from the write offset to check we are getting
1105 * the same sub-buffer we are expecting (the sub-buffers are atomically
1106 * "tagged" upon writes, tags are checked upon read).
1108 write_offset
= v_read(config
, &buf
->offset
);
1111 * Check that we are not about to read the same subbuffer in
1112 * which the writer head is.
1114 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
1118 *consumed
= consumed_cur
;
1119 *produced
= subbuf_trunc(write_offset
, chan
);
1125 * The memory barriers __wait_event()/wake_up_interruptible() take care
1126 * of "raw_spin_is_locked" memory ordering.
1135 * lib_ring_buffer_move_consumer - move consumed counter forward
1137 * @consumed_new: new consumed count value
1139 void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer
*buf
,
1140 unsigned long consumed_new
,
1141 struct lttng_ust_shm_handle
*handle
)
1143 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1144 struct channel
*chan
= shmp(handle
, bufb
->chan
);
1145 unsigned long consumed
;
1147 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1150 * Only push the consumed value forward.
1151 * If the consumed cmpxchg fails, this is because we have been pushed by
1152 * the writer in flight recorder mode.
1154 consumed
= uatomic_read(&buf
->consumed
);
1155 while ((long) consumed
- (long) consumed_new
< 0)
1156 consumed
= uatomic_cmpxchg(&buf
->consumed
, consumed
,
1161 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
1163 * @consumed: consumed count indicating the position where to read
1165 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1166 * data to read at consumed position, or 0 if the get operation succeeds.
1168 int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
1169 unsigned long consumed
,
1170 struct lttng_ust_shm_handle
*handle
)
1172 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1173 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1174 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
1175 int ret
, finalized
, nr_retry
= LTTNG_UST_RING_BUFFER_GET_RETRY
;
1178 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
1180 * Read finalized before counters.
1183 consumed_cur
= uatomic_read(&buf
->consumed
);
1184 consumed_idx
= subbuf_index(consumed
, chan
);
1185 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, consumed_idx
)->cc_sb
);
1187 * Make sure we read the commit count before reading the buffer
1188 * data and the write offset. Correct consumed offset ordering
1189 * wrt commit count is insured by the use of cmpxchg to update
1190 * the consumed offset.
1193 * Local rmb to match the remote wmb to read the commit count
1194 * before the buffer data and the write offset.
1198 write_offset
= v_read(config
, &buf
->offset
);
1201 * Check that the buffer we are getting is after or at consumed_cur
1204 if ((long) subbuf_trunc(consumed
, chan
)
1205 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
1209 * Check that the subbuffer we are trying to consume has been
1210 * already fully committed. There are a few causes that can make
1211 * this unavailability situation occur:
1213 * Temporary (short-term) situation:
1214 * - Application is running on a different CPU, between reserve
1215 * and commit ring buffer operations,
1216 * - Application is preempted between reserve and commit ring
1217 * buffer operations,
1219 * Long-term situation:
1220 * - Application is stopped (SIGSTOP) between reserve and commit
1221 * ring buffer operations. Could eventually be resumed by
1223 * - Application is killed (SIGTERM, SIGINT, SIGKILL) between
1224 * reserve and commit ring buffer operation.
1226 * From a consumer perspective, handling short-term
1227 * unavailability situations is performed by retrying a few
1228 * times after a delay. Handling long-term unavailability
1229 * situations is handled by failing to get the sub-buffer.
1231 * In all of those situations, if the application is taking a
1232 * long time to perform its commit after ring buffer space
1233 * reservation, we can end up in a situation where the producer
1234 * will fill the ring buffer and try to write into the same
1235 * sub-buffer again (which has a missing commit). This is
1236 * handled by the producer in the sub-buffer switch handling
1237 * code of the reserve routine by detecting unbalanced
1238 * reserve/commit counters and discarding all further events
1239 * until the situation is resolved in those situations. Two
1240 * scenarios can occur:
1242 * 1) The application causing the reserve/commit counters to be
1243 * unbalanced has been terminated. In this situation, all
1244 * further events will be discarded in the buffers, and no
1245 * further buffer data will be readable by the consumer
1246 * daemon. Tearing down the UST tracing session and starting
1247 * anew is a work-around for those situations. Note that this
1248 * only affects per-UID tracing. In per-PID tracing, the
1249 * application vanishes with the termination, and therefore
1250 * no more data needs to be written to the buffers.
1251 * 2) The application causing the unbalance has been delayed for
1252 * a long time, but will eventually try to increment the
1253 * commit counter after eventually writing to the sub-buffer.
1254 * This situation can cause events to be discarded until the
1255 * application resumes its operations.
1257 if (((commit_count
- chan
->backend
.subbuf_size
)
1258 & chan
->commit_count_mask
)
1259 - (buf_trunc(consumed
, chan
)
1260 >> chan
->backend
.num_subbuf_order
)
1262 if (nr_retry
-- > 0) {
1263 if (nr_retry
<= (LTTNG_UST_RING_BUFFER_GET_RETRY
>> 1))
1264 (void) poll(NULL
, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS
);
1272 * Check that we are not about to read the same subbuffer in
1273 * which the writer head is.
1275 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed
, chan
)
1280 * Failure to get the subbuffer causes a busy-loop retry without going
1281 * to a wait queue. These are caused by short-lived race windows where
1282 * the writer is getting access to a subbuffer we were trying to get
1283 * access to. Also checks that the "consumed" buffer count we are
1284 * looking for matches the one contained in the subbuffer id.
1286 * The short-lived race window described here can be affected by
1287 * application signals and preemption, thus requiring to bound
1288 * the loop to a maximum number of retry.
1290 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1291 consumed_idx
, buf_trunc_val(consumed
, chan
),
1294 if (nr_retry
-- > 0) {
1295 if (nr_retry
<= (LTTNG_UST_RING_BUFFER_GET_RETRY
>> 1))
1296 (void) poll(NULL
, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS
);
1302 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1304 buf
->get_subbuf_consumed
= consumed
;
1305 buf
->get_subbuf
= 1;
1311 * The memory barriers __wait_event()/wake_up_interruptible() take care
1312 * of "raw_spin_is_locked" memory ordering.
1321 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1324 void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
1325 struct lttng_ust_shm_handle
*handle
)
1327 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1328 struct channel
*chan
= shmp(handle
, bufb
->chan
);
1329 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1330 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
1332 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1334 if (!buf
->get_subbuf
) {
1336 * Reader puts a subbuffer it did not get.
1338 CHAN_WARN_ON(chan
, 1);
1341 consumed
= buf
->get_subbuf_consumed
;
1342 buf
->get_subbuf
= 0;
1345 * Clear the records_unread counter. (overruns counter)
1346 * Can still be non-zero if a file reader simply grabbed the data
1347 * without using iterators.
1348 * Can be below zero if an iterator is used on a snapshot more than
1351 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1352 v_add(config
, v_read(config
,
1353 &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
),
1354 &bufb
->records_read
);
1355 v_set(config
, &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
, 0);
1356 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1357 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1358 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1361 * Exchange the reader subbuffer with the one we put in its place in the
1362 * writer subbuffer table. Expect the original consumed count. If
1363 * update_read_sb_index fails, this is because the writer updated the
1364 * subbuffer concurrently. We should therefore keep the subbuffer we
1365 * currently have: it has become invalid to try reading this sub-buffer
1366 * consumed count value anyway.
1368 consumed_idx
= subbuf_index(consumed
, chan
);
1369 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1370 consumed_idx
, buf_trunc_val(consumed
, chan
),
1373 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1374 * if the writer concurrently updated it.
1379 * cons_offset is an iterator on all subbuffer offsets between the reader
1380 * position and the writer position. (inclusive)
1383 void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
1384 struct channel
*chan
,
1385 unsigned long cons_offset
,
1387 struct lttng_ust_shm_handle
*handle
)
1389 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1390 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1392 cons_idx
= subbuf_index(cons_offset
, chan
);
1393 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, cons_idx
)->cc
);
1394 commit_count_sb
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, cons_idx
)->cc_sb
);
1396 if (subbuf_offset(commit_count
, chan
) != 0)
1397 DBG("ring buffer %s, cpu %d: "
1398 "commit count in subbuffer %lu,\n"
1399 "expecting multiples of %lu bytes\n"
1400 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1401 chan
->backend
.name
, cpu
, cons_idx
,
1402 chan
->backend
.subbuf_size
,
1403 commit_count
, commit_count_sb
);
1405 DBG("ring buffer: %s, cpu %d: %lu bytes committed\n",
1406 chan
->backend
.name
, cpu
, commit_count
);
1410 void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
1411 struct channel
*chan
,
1412 void *priv
, int cpu
,
1413 struct lttng_ust_shm_handle
*handle
)
1415 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1416 unsigned long write_offset
, cons_offset
;
1419 * No need to order commit_count, write_offset and cons_offset reads
1420 * because we execute at teardown when no more writer nor reader
1421 * references are left.
1423 write_offset
= v_read(config
, &buf
->offset
);
1424 cons_offset
= uatomic_read(&buf
->consumed
);
1425 if (write_offset
!= cons_offset
)
1426 DBG("ring buffer %s, cpu %d: "
1427 "non-consumed data\n"
1428 " [ %lu bytes written, %lu bytes read ]\n",
1429 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1431 for (cons_offset
= uatomic_read(&buf
->consumed
);
1432 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1435 cons_offset
= subbuf_align(cons_offset
, chan
))
1436 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1441 void lib_ring_buffer_print_errors(struct channel
*chan
,
1442 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
1443 struct lttng_ust_shm_handle
*handle
)
1445 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1446 void *priv
= channel_get_private(chan
);
1448 if (!strcmp(chan
->backend
.name
, "relay-metadata-mmap")) {
1449 DBG("ring buffer %s: %lu records written, "
1450 "%lu records overrun\n",
1452 v_read(config
, &buf
->records_count
),
1453 v_read(config
, &buf
->records_overrun
));
1455 DBG("ring buffer %s, cpu %d: %lu records written, "
1456 "%lu records overrun\n",
1457 chan
->backend
.name
, cpu
,
1458 v_read(config
, &buf
->records_count
),
1459 v_read(config
, &buf
->records_overrun
));
1461 if (v_read(config
, &buf
->records_lost_full
)
1462 || v_read(config
, &buf
->records_lost_wrap
)
1463 || v_read(config
, &buf
->records_lost_big
))
1464 DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n"
1465 " [ %lu buffer full, %lu nest buffer wrap-around, "
1466 "%lu event too big ]\n",
1467 chan
->backend
.name
, cpu
,
1468 v_read(config
, &buf
->records_lost_full
),
1469 v_read(config
, &buf
->records_lost_wrap
),
1470 v_read(config
, &buf
->records_lost_big
));
1472 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
, handle
);
1476 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1478 * Only executed when the buffer is finalized, in SWITCH_FLUSH.
1481 void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer
*buf
,
1482 struct channel
*chan
,
1483 struct switch_offsets
*offsets
,
1485 struct lttng_ust_shm_handle
*handle
)
1487 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1488 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1489 unsigned long commit_count
;
1491 config
->cb
.buffer_begin(buf
, tsc
, oldidx
, handle
);
1494 * Order all writes to buffer before the commit count update that will
1495 * determine that the subbuffer is full.
1498 v_add(config
, config
->cb
.subbuffer_header_size(),
1499 &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1500 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1501 /* Check if the written buffer has to be delivered */
1502 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1503 commit_count
, oldidx
, handle
, tsc
);
1504 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1505 offsets
->old
+ config
->cb
.subbuffer_header_size(),
1506 commit_count
, handle
);
1510 * lib_ring_buffer_switch_old_end: switch old subbuffer
1512 * Note : offset_old should never be 0 here. It is ok, because we never perform
1513 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1514 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1518 void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer
*buf
,
1519 struct channel
*chan
,
1520 struct switch_offsets
*offsets
,
1522 struct lttng_ust_shm_handle
*handle
)
1524 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1525 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1526 unsigned long commit_count
, padding_size
, data_size
;
1528 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1529 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1530 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
,
1534 * Order all writes to buffer before the commit count update that will
1535 * determine that the subbuffer is full.
1538 v_add(config
, padding_size
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1539 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1540 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1541 commit_count
, oldidx
, handle
, tsc
);
1542 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1543 offsets
->old
+ padding_size
, commit_count
, handle
);
1547 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1549 * This code can be executed unordered : writers may already have written to the
1550 * sub-buffer before this code gets executed, caution. The commit makes sure
1551 * that this code is executed before the deliver of this sub-buffer.
1554 void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer
*buf
,
1555 struct channel
*chan
,
1556 struct switch_offsets
*offsets
,
1558 struct lttng_ust_shm_handle
*handle
)
1560 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1561 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1562 unsigned long commit_count
;
1564 config
->cb
.buffer_begin(buf
, tsc
, beginidx
, handle
);
1567 * Order all writes to buffer before the commit count update that will
1568 * determine that the subbuffer is full.
1571 v_add(config
, config
->cb
.subbuffer_header_size(),
1572 &shmp_index(handle
, buf
->commit_hot
, beginidx
)->cc
);
1573 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, beginidx
)->cc
);
1574 /* Check if the written buffer has to be delivered */
1575 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1576 commit_count
, beginidx
, handle
, tsc
);
1577 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, beginidx
,
1578 offsets
->begin
+ config
->cb
.subbuffer_header_size(),
1579 commit_count
, handle
);
1583 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1585 * Calls subbuffer_set_data_size() to set the data size of the current
1586 * sub-buffer. We do not need to perform check_deliver nor commit here,
1587 * since this task will be done by the "commit" of the event for which
1588 * we are currently doing the space reservation.
1591 void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer
*buf
,
1592 struct channel
*chan
,
1593 struct switch_offsets
*offsets
,
1595 struct lttng_ust_shm_handle
*handle
)
1597 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1598 unsigned long endidx
, data_size
;
1600 endidx
= subbuf_index(offsets
->end
- 1, chan
);
1601 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1602 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
,
1609 * !0 if execution must be aborted.
1612 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1613 struct lttng_ust_lib_ring_buffer
*buf
,
1614 struct channel
*chan
,
1615 struct switch_offsets
*offsets
,
1617 struct lttng_ust_shm_handle
*handle
)
1619 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1620 unsigned long off
, reserve_commit_diff
;
1622 offsets
->begin
= v_read(config
, &buf
->offset
);
1623 offsets
->old
= offsets
->begin
;
1624 offsets
->switch_old_start
= 0;
1625 off
= subbuf_offset(offsets
->begin
, chan
);
1627 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1630 * Ensure we flush the header of an empty subbuffer when doing the
1631 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1632 * total data gathering duration even if there were no records saved
1633 * after the last buffer switch.
1634 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1635 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1636 * subbuffer header as appropriate.
1637 * The next record that reserves space will be responsible for
1638 * populating the following subbuffer header. We choose not to populate
1639 * the next subbuffer header here because we want to be able to use
1640 * SWITCH_ACTIVE for periodical buffer flush, which must
1641 * guarantee that all the buffer content (records and header
1642 * timestamps) are visible to the reader. This is required for
1643 * quiescence guarantees for the fusion merge.
1645 if (mode
!= SWITCH_FLUSH
&& !off
)
1646 return -1; /* we do not have to switch : buffer is empty */
1648 if (caa_unlikely(off
== 0)) {
1649 unsigned long sb_index
, commit_count
;
1652 * We are performing a SWITCH_FLUSH. At this stage, there are no
1653 * concurrent writes into the buffer.
1655 * The client does not save any header information. Don't
1656 * switch empty subbuffer on finalize, because it is invalid to
1657 * deliver a completely empty subbuffer.
1659 if (!config
->cb
.subbuffer_header_size())
1662 /* Test new buffer integrity */
1663 sb_index
= subbuf_index(offsets
->begin
, chan
);
1664 commit_count
= v_read(config
,
1665 &shmp_index(handle
, buf
->commit_cold
,
1667 reserve_commit_diff
=
1668 (buf_trunc(offsets
->begin
, chan
)
1669 >> chan
->backend
.num_subbuf_order
)
1670 - (commit_count
& chan
->commit_count_mask
);
1671 if (caa_likely(reserve_commit_diff
== 0)) {
1672 /* Next subbuffer not being written to. */
1673 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1674 subbuf_trunc(offsets
->begin
, chan
)
1675 - subbuf_trunc((unsigned long)
1676 uatomic_read(&buf
->consumed
), chan
)
1677 >= chan
->backend
.buf_size
)) {
1679 * We do not overwrite non consumed buffers
1680 * and we are full : don't switch.
1685 * Next subbuffer not being written to, and we
1686 * are either in overwrite mode or the buffer is
1687 * not full. It's safe to write in this new
1693 * Next subbuffer reserve offset does not match the
1694 * commit offset. Don't perform switch in
1695 * producer-consumer and overwrite mode. Caused by
1696 * either a writer OOPS or too many nested writes over a
1697 * reserve/commit pair.
1703 * Need to write the subbuffer start header on finalize.
1705 offsets
->switch_old_start
= 1;
1707 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1708 /* Note: old points to the next subbuf at offset 0 */
1709 offsets
->end
= offsets
->begin
;
1714 * Force a sub-buffer switch. This operation is completely reentrant : can be
1715 * called while tracing is active with absolutely no lock held.
1717 * Note, however, that as a v_cmpxchg is used for some atomic
1718 * operations, this function must be called from the CPU which owns the buffer
1719 * for a ACTIVE flush.
1721 void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer
*buf
, enum switch_mode mode
,
1722 struct lttng_ust_shm_handle
*handle
)
1724 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1725 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1726 struct switch_offsets offsets
;
1727 unsigned long oldidx
;
1733 * Perform retryable operations.
1736 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1738 return; /* Switch not needed */
1739 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1743 * Atomically update last_tsc. This update races against concurrent
1744 * atomic updates, but the race will always cause supplementary full TSC
1745 * records, never the opposite (missing a full TSC record when it would
1748 save_last_tsc(config
, buf
, tsc
);
1751 * Push the reader if necessary
1753 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1755 oldidx
= subbuf_index(offsets
.old
, chan
);
1756 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
, handle
);
1759 * May need to populate header start on SWITCH_FLUSH.
1761 if (offsets
.switch_old_start
) {
1762 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
, handle
);
1763 offsets
.old
+= config
->cb
.subbuffer_header_size();
1767 * Switch old subbuffer.
1769 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
, handle
);
1775 * -ENOSPC if event size is too large for packet.
1776 * -ENOBUFS if there is currently not enough space in buffer for the event.
1777 * -EIO if data cannot be written into the buffer for any other reason.
1780 int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer
*buf
,
1781 struct channel
*chan
,
1782 struct switch_offsets
*offsets
,
1783 struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
1785 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1786 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
1787 unsigned long reserve_commit_diff
, offset_cmp
;
1790 offsets
->begin
= offset_cmp
= v_read(config
, &buf
->offset
);
1791 offsets
->old
= offsets
->begin
;
1792 offsets
->switch_new_start
= 0;
1793 offsets
->switch_new_end
= 0;
1794 offsets
->switch_old_end
= 0;
1795 offsets
->pre_header_padding
= 0;
1797 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1798 if ((int64_t) ctx
->tsc
== -EIO
)
1801 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1802 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
1804 if (caa_unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1805 offsets
->switch_new_start
= 1; /* For offsets->begin */
1807 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1809 &offsets
->pre_header_padding
,
1812 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1815 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
) +
1816 offsets
->size
> chan
->backend
.subbuf_size
)) {
1817 offsets
->switch_old_end
= 1; /* For offsets->old */
1818 offsets
->switch_new_start
= 1; /* For offsets->begin */
1821 if (caa_unlikely(offsets
->switch_new_start
)) {
1822 unsigned long sb_index
, commit_count
;
1825 * We are typically not filling the previous buffer completely.
1827 if (caa_likely(offsets
->switch_old_end
))
1828 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1829 offsets
->begin
= offsets
->begin
1830 + config
->cb
.subbuffer_header_size();
1831 /* Test new buffer integrity */
1832 sb_index
= subbuf_index(offsets
->begin
, chan
);
1834 * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
1835 * lib_ring_buffer_check_deliver() has the matching
1836 * memory barriers required around commit_cold cc_sb
1837 * updates to ensure reserve and commit counter updates
1838 * are not seen reordered when updated by another CPU.
1841 commit_count
= v_read(config
,
1842 &shmp_index(handle
, buf
->commit_cold
,
1844 /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
1846 if (caa_unlikely(offset_cmp
!= v_read(config
, &buf
->offset
))) {
1848 * The reserve counter have been concurrently updated
1849 * while we read the commit counter. This means the
1850 * commit counter we read might not match buf->offset
1851 * due to concurrent update. We therefore need to retry.
1855 reserve_commit_diff
=
1856 (buf_trunc(offsets
->begin
, chan
)
1857 >> chan
->backend
.num_subbuf_order
)
1858 - (commit_count
& chan
->commit_count_mask
);
1859 if (caa_likely(reserve_commit_diff
== 0)) {
1860 /* Next subbuffer not being written to. */
1861 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1862 subbuf_trunc(offsets
->begin
, chan
)
1863 - subbuf_trunc((unsigned long)
1864 uatomic_read(&buf
->consumed
), chan
)
1865 >= chan
->backend
.buf_size
)) {
1866 unsigned long nr_lost
;
1869 * We do not overwrite non consumed buffers
1870 * and we are full : record is lost.
1872 nr_lost
= v_read(config
, &buf
->records_lost_full
);
1873 v_inc(config
, &buf
->records_lost_full
);
1874 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
1875 DBG("%lu or more records lost in (%s:%d) (buffer full)\n",
1876 nr_lost
+ 1, chan
->backend
.name
,
1882 * Next subbuffer not being written to, and we
1883 * are either in overwrite mode or the buffer is
1884 * not full. It's safe to write in this new
1889 unsigned long nr_lost
;
1892 * Next subbuffer reserve offset does not match the
1893 * commit offset, and this did not involve update to the
1894 * reserve counter. Drop record in producer-consumer and
1895 * overwrite mode. Caused by either a writer OOPS or too
1896 * many nested writes over a reserve/commit pair.
1898 nr_lost
= v_read(config
, &buf
->records_lost_wrap
);
1899 v_inc(config
, &buf
->records_lost_wrap
);
1900 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
1901 DBG("%lu or more records lost in (%s:%d) (wrap-around)\n",
1902 nr_lost
+ 1, chan
->backend
.name
,
1908 config
->cb
.record_header_size(config
, chan
,
1910 &offsets
->pre_header_padding
,
1913 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1916 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
)
1917 + offsets
->size
> chan
->backend
.subbuf_size
)) {
1918 unsigned long nr_lost
;
1921 * Record too big for subbuffers, report error, don't
1922 * complete the sub-buffer switch.
1924 nr_lost
= v_read(config
, &buf
->records_lost_big
);
1925 v_inc(config
, &buf
->records_lost_big
);
1926 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
1927 DBG("%lu or more records lost in (%s:%d) record size "
1928 " of %zu bytes is too large for buffer\n",
1929 nr_lost
+ 1, chan
->backend
.name
,
1930 buf
->backend
.cpu
, offsets
->size
);
1935 * We just made a successful buffer switch and the
1936 * record fits in the new subbuffer. Let's write.
1941 * Record fits in the current buffer and we are not on a switch
1942 * boundary. It's safe to write.
1945 offsets
->end
= offsets
->begin
+ offsets
->size
;
1947 if (caa_unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
1949 * The offset_end will fall at the very beginning of the next
1952 offsets
->switch_new_end
= 1; /* For offsets->begin */
1958 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1959 * @ctx: ring buffer context.
1961 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
1962 * -EIO for other errors, else returns 0.
1963 * It will take care of sub-buffer switching.
1965 int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
1967 struct channel
*chan
= ctx
->chan
;
1968 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
1969 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1970 struct lttng_ust_lib_ring_buffer
*buf
;
1971 struct switch_offsets offsets
;
1974 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
1975 buf
= shmp(handle
, chan
->backend
.buf
[ctx
->cpu
].shmp
);
1977 buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
1983 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
1985 if (caa_unlikely(ret
))
1987 } while (caa_unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
1992 * Atomically update last_tsc. This update races against concurrent
1993 * atomic updates, but the race will always cause supplementary full TSC
1994 * records, never the opposite (missing a full TSC record when it would
1997 save_last_tsc(config
, buf
, ctx
->tsc
);
2000 * Push the reader if necessary
2002 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
2005 * Clear noref flag for this subbuffer.
2007 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2008 subbuf_index(offsets
.end
- 1, chan
),
2012 * Switch old subbuffer if needed.
2014 if (caa_unlikely(offsets
.switch_old_end
)) {
2015 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2016 subbuf_index(offsets
.old
- 1, chan
),
2018 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2022 * Populate new subbuffer.
2024 if (caa_unlikely(offsets
.switch_new_start
))
2025 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2027 if (caa_unlikely(offsets
.switch_new_end
))
2028 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2030 ctx
->slot_size
= offsets
.size
;
2031 ctx
->pre_offset
= offsets
.begin
;
2032 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
2037 * Force a read (imply TLS fixup for dlopen) of TLS variables.
2039 void lttng_fixup_ringbuffer_tls(void)
2041 asm volatile ("" : : "m" (URCU_TLS(lib_ring_buffer_nesting
)));
2044 void lib_ringbuffer_signal_init(void)
2050 * Block signal for entire process, so only our thread processes
2054 ret
= pthread_sigmask(SIG_BLOCK
, &mask
, NULL
);
2057 PERROR("pthread_sigmask");