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 * @stream_fds: array of stream file descriptors.
777 * @nr_stream_fds: number of file descriptors in array.
780 * Returns NULL on failure.
782 struct lttng_ust_shm_handle
*channel_create(const struct lttng_ust_lib_ring_buffer_config
*config
,
785 size_t priv_data_align
,
786 size_t priv_data_size
,
787 void *priv_data_init
,
788 void *buf_addr
, size_t subbuf_size
,
789 size_t num_subbuf
, unsigned int switch_timer_interval
,
790 unsigned int read_timer_interval
,
791 const int *stream_fds
, int nr_stream_fds
)
794 size_t shmsize
, chansize
;
795 struct channel
*chan
;
796 struct lttng_ust_shm_handle
*handle
;
797 struct shm_object
*shmobj
;
798 unsigned int nr_streams
;
800 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
801 nr_streams
= num_possible_cpus();
805 if (nr_stream_fds
!= nr_streams
)
808 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
809 read_timer_interval
))
812 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
816 /* Allocate table for channel + per-cpu buffers */
817 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
819 goto error_table_alloc
;
821 /* Calculate the shm allocation layout */
822 shmsize
= sizeof(struct channel
);
823 shmsize
+= offset_align(shmsize
, __alignof__(struct lttng_ust_lib_ring_buffer_shmp
));
824 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
) * nr_streams
;
827 shmsize
+= offset_align(shmsize
, priv_data_align
);
828 shmsize
+= priv_data_size
;
830 /* Allocate normal memory for channel (not shared) */
831 shmobj
= shm_object_table_alloc(handle
->table
, shmsize
, SHM_OBJECT_MEM
,
835 /* struct channel is at object 0, offset 0 (hardcoded) */
836 set_shmp(handle
->chan
, zalloc_shm(shmobj
, chansize
));
837 assert(handle
->chan
._ref
.index
== 0);
838 assert(handle
->chan
._ref
.offset
== 0);
839 chan
= shmp(handle
, handle
->chan
);
842 chan
->nr_streams
= nr_streams
;
844 /* space for private data */
845 if (priv_data_size
) {
846 DECLARE_SHMP(void, priv_data_alloc
);
848 align_shm(shmobj
, priv_data_align
);
849 chan
->priv_data_offset
= shmobj
->allocated_len
;
850 set_shmp(priv_data_alloc
, zalloc_shm(shmobj
, priv_data_size
));
851 if (!shmp(handle
, priv_data_alloc
))
853 *priv_data
= channel_get_private(chan
);
854 memcpy(*priv_data
, priv_data_init
, priv_data_size
);
856 chan
->priv_data_offset
= -1;
861 ret
= channel_backend_init(&chan
->backend
, name
, config
,
862 subbuf_size
, num_subbuf
, handle
,
865 goto error_backend_init
;
867 chan
->handle
= handle
;
868 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
870 chan
->switch_timer_interval
= switch_timer_interval
;
871 chan
->read_timer_interval
= read_timer_interval
;
872 lib_ring_buffer_channel_switch_timer_start(chan
);
873 lib_ring_buffer_channel_read_timer_start(chan
);
879 shm_object_table_destroy(handle
->table
);
885 struct lttng_ust_shm_handle
*channel_handle_create(void *data
,
886 uint64_t memory_map_size
,
889 struct lttng_ust_shm_handle
*handle
;
890 struct shm_object
*object
;
892 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
896 /* Allocate table for channel + per-cpu buffers */
897 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
899 goto error_table_alloc
;
900 /* Add channel object */
901 object
= shm_object_table_append_mem(handle
->table
, data
,
902 memory_map_size
, wakeup_fd
);
904 goto error_table_object
;
905 /* struct channel is at object 0, offset 0 (hardcoded) */
906 handle
->chan
._ref
.index
= 0;
907 handle
->chan
._ref
.offset
= 0;
911 shm_object_table_destroy(handle
->table
);
917 int channel_handle_add_stream(struct lttng_ust_shm_handle
*handle
,
918 int shm_fd
, int wakeup_fd
, uint32_t stream_nr
,
919 uint64_t memory_map_size
)
921 struct shm_object
*object
;
923 /* Add stream object */
924 object
= shm_object_table_append_shm(handle
->table
,
925 shm_fd
, wakeup_fd
, stream_nr
,
932 unsigned int channel_handle_get_nr_streams(struct lttng_ust_shm_handle
*handle
)
934 assert(handle
->table
);
935 return handle
->table
->allocated_len
- 1;
939 void channel_release(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
)
941 channel_free(chan
, handle
);
945 * channel_destroy - Finalize, wait for q.s. and destroy channel.
946 * @chan: channel to destroy
949 * Call "destroy" callback, finalize channels, decrement the channel
950 * reference count. Note that when readers have completed data
951 * consumption of finalized channels, get_subbuf() will return -ENODATA.
952 * They should release their handle at that point.
954 void channel_destroy(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
959 * Note: the consumer takes care of finalizing and
960 * switching the buffers.
962 channel_unregister_notifiers(chan
, handle
);
964 * The consumer prints errors.
966 channel_print_errors(chan
, handle
);
970 * sessiond/consumer are keeping a reference on the shm file
971 * descriptor directly. No need to refcount.
973 channel_release(chan
, handle
);
977 struct lttng_ust_lib_ring_buffer
*channel_get_ring_buffer(
978 const struct lttng_ust_lib_ring_buffer_config
*config
,
979 struct channel
*chan
, int cpu
,
980 struct lttng_ust_shm_handle
*handle
,
981 int *shm_fd
, int *wait_fd
,
983 uint64_t *memory_map_size
)
987 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
990 if (cpu
>= num_possible_cpus())
993 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
994 *shm_fd
= shm_get_shm_fd(handle
, ref
);
995 *wait_fd
= shm_get_wait_fd(handle
, ref
);
996 *wakeup_fd
= shm_get_wakeup_fd(handle
, ref
);
997 if (shm_get_shm_size(handle
, ref
, memory_map_size
))
999 return shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
1002 int ring_buffer_channel_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1003 struct channel
*chan
,
1004 struct lttng_ust_shm_handle
*handle
)
1006 struct shm_ref
*ref
;
1008 ref
= &handle
->chan
._ref
;
1009 return shm_close_wait_fd(handle
, ref
);
1012 int ring_buffer_channel_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1013 struct channel
*chan
,
1014 struct lttng_ust_shm_handle
*handle
)
1016 struct shm_ref
*ref
;
1018 ref
= &handle
->chan
._ref
;
1019 return shm_close_wakeup_fd(handle
, ref
);
1022 int ring_buffer_stream_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1023 struct channel
*chan
,
1024 struct lttng_ust_shm_handle
*handle
,
1027 struct shm_ref
*ref
;
1029 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1032 if (cpu
>= num_possible_cpus())
1035 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1036 return shm_close_wait_fd(handle
, ref
);
1039 int ring_buffer_stream_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1040 struct channel
*chan
,
1041 struct lttng_ust_shm_handle
*handle
,
1044 struct shm_ref
*ref
;
1047 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1050 if (cpu
>= num_possible_cpus())
1053 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1054 pthread_mutex_lock(&wakeup_fd_mutex
);
1055 ret
= shm_close_wakeup_fd(handle
, ref
);
1056 pthread_mutex_unlock(&wakeup_fd_mutex
);
1060 int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer
*buf
,
1061 struct lttng_ust_shm_handle
*handle
)
1063 if (uatomic_cmpxchg(&buf
->active_readers
, 0, 1) != 0)
1069 void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer
*buf
,
1070 struct lttng_ust_shm_handle
*handle
)
1072 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1074 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1076 uatomic_dec(&buf
->active_readers
);
1080 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
1082 * @consumed: consumed count indicating the position where to read
1083 * @produced: produced count, indicates position when to stop reading
1085 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1086 * data to read at consumed position, or 0 if the get operation succeeds.
1089 int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer
*buf
,
1090 unsigned long *consumed
, unsigned long *produced
,
1091 struct lttng_ust_shm_handle
*handle
)
1093 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1094 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1095 unsigned long consumed_cur
, write_offset
;
1098 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
1100 * Read finalized before counters.
1103 consumed_cur
= uatomic_read(&buf
->consumed
);
1105 * No need to issue a memory barrier between consumed count read and
1106 * write offset read, because consumed count can only change
1107 * concurrently in overwrite mode, and we keep a sequence counter
1108 * identifier derived from the write offset to check we are getting
1109 * the same sub-buffer we are expecting (the sub-buffers are atomically
1110 * "tagged" upon writes, tags are checked upon read).
1112 write_offset
= v_read(config
, &buf
->offset
);
1115 * Check that we are not about to read the same subbuffer in
1116 * which the writer head is.
1118 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
1122 *consumed
= consumed_cur
;
1123 *produced
= subbuf_trunc(write_offset
, chan
);
1129 * The memory barriers __wait_event()/wake_up_interruptible() take care
1130 * of "raw_spin_is_locked" memory ordering.
1139 * lib_ring_buffer_move_consumer - move consumed counter forward
1141 * @consumed_new: new consumed count value
1143 void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer
*buf
,
1144 unsigned long consumed_new
,
1145 struct lttng_ust_shm_handle
*handle
)
1147 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1148 struct channel
*chan
= shmp(handle
, bufb
->chan
);
1149 unsigned long consumed
;
1151 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1154 * Only push the consumed value forward.
1155 * If the consumed cmpxchg fails, this is because we have been pushed by
1156 * the writer in flight recorder mode.
1158 consumed
= uatomic_read(&buf
->consumed
);
1159 while ((long) consumed
- (long) consumed_new
< 0)
1160 consumed
= uatomic_cmpxchg(&buf
->consumed
, consumed
,
1165 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
1167 * @consumed: consumed count indicating the position where to read
1169 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1170 * data to read at consumed position, or 0 if the get operation succeeds.
1172 int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
1173 unsigned long consumed
,
1174 struct lttng_ust_shm_handle
*handle
)
1176 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1177 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1178 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
1179 int ret
, finalized
, nr_retry
= LTTNG_UST_RING_BUFFER_GET_RETRY
;
1182 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
1184 * Read finalized before counters.
1187 consumed_cur
= uatomic_read(&buf
->consumed
);
1188 consumed_idx
= subbuf_index(consumed
, chan
);
1189 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, consumed_idx
)->cc_sb
);
1191 * Make sure we read the commit count before reading the buffer
1192 * data and the write offset. Correct consumed offset ordering
1193 * wrt commit count is insured by the use of cmpxchg to update
1194 * the consumed offset.
1197 * Local rmb to match the remote wmb to read the commit count
1198 * before the buffer data and the write offset.
1202 write_offset
= v_read(config
, &buf
->offset
);
1205 * Check that the buffer we are getting is after or at consumed_cur
1208 if ((long) subbuf_trunc(consumed
, chan
)
1209 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
1213 * Check that the subbuffer we are trying to consume has been
1214 * already fully committed. There are a few causes that can make
1215 * this unavailability situation occur:
1217 * Temporary (short-term) situation:
1218 * - Application is running on a different CPU, between reserve
1219 * and commit ring buffer operations,
1220 * - Application is preempted between reserve and commit ring
1221 * buffer operations,
1223 * Long-term situation:
1224 * - Application is stopped (SIGSTOP) between reserve and commit
1225 * ring buffer operations. Could eventually be resumed by
1227 * - Application is killed (SIGTERM, SIGINT, SIGKILL) between
1228 * reserve and commit ring buffer operation.
1230 * From a consumer perspective, handling short-term
1231 * unavailability situations is performed by retrying a few
1232 * times after a delay. Handling long-term unavailability
1233 * situations is handled by failing to get the sub-buffer.
1235 * In all of those situations, if the application is taking a
1236 * long time to perform its commit after ring buffer space
1237 * reservation, we can end up in a situation where the producer
1238 * will fill the ring buffer and try to write into the same
1239 * sub-buffer again (which has a missing commit). This is
1240 * handled by the producer in the sub-buffer switch handling
1241 * code of the reserve routine by detecting unbalanced
1242 * reserve/commit counters and discarding all further events
1243 * until the situation is resolved in those situations. Two
1244 * scenarios can occur:
1246 * 1) The application causing the reserve/commit counters to be
1247 * unbalanced has been terminated. In this situation, all
1248 * further events will be discarded in the buffers, and no
1249 * further buffer data will be readable by the consumer
1250 * daemon. Tearing down the UST tracing session and starting
1251 * anew is a work-around for those situations. Note that this
1252 * only affects per-UID tracing. In per-PID tracing, the
1253 * application vanishes with the termination, and therefore
1254 * no more data needs to be written to the buffers.
1255 * 2) The application causing the unbalance has been delayed for
1256 * a long time, but will eventually try to increment the
1257 * commit counter after eventually writing to the sub-buffer.
1258 * This situation can cause events to be discarded until the
1259 * application resumes its operations.
1261 if (((commit_count
- chan
->backend
.subbuf_size
)
1262 & chan
->commit_count_mask
)
1263 - (buf_trunc(consumed
, chan
)
1264 >> chan
->backend
.num_subbuf_order
)
1266 if (nr_retry
-- > 0) {
1267 if (nr_retry
<= (LTTNG_UST_RING_BUFFER_GET_RETRY
>> 1))
1268 (void) poll(NULL
, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS
);
1276 * Check that we are not about to read the same subbuffer in
1277 * which the writer head is.
1279 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed
, chan
)
1284 * Failure to get the subbuffer causes a busy-loop retry without going
1285 * to a wait queue. These are caused by short-lived race windows where
1286 * the writer is getting access to a subbuffer we were trying to get
1287 * access to. Also checks that the "consumed" buffer count we are
1288 * looking for matches the one contained in the subbuffer id.
1290 * The short-lived race window described here can be affected by
1291 * application signals and preemption, thus requiring to bound
1292 * the loop to a maximum number of retry.
1294 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1295 consumed_idx
, buf_trunc_val(consumed
, chan
),
1298 if (nr_retry
-- > 0) {
1299 if (nr_retry
<= (LTTNG_UST_RING_BUFFER_GET_RETRY
>> 1))
1300 (void) poll(NULL
, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS
);
1306 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1308 buf
->get_subbuf_consumed
= consumed
;
1309 buf
->get_subbuf
= 1;
1315 * The memory barriers __wait_event()/wake_up_interruptible() take care
1316 * of "raw_spin_is_locked" memory ordering.
1325 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1328 void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
1329 struct lttng_ust_shm_handle
*handle
)
1331 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1332 struct channel
*chan
= shmp(handle
, bufb
->chan
);
1333 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1334 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
1336 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1338 if (!buf
->get_subbuf
) {
1340 * Reader puts a subbuffer it did not get.
1342 CHAN_WARN_ON(chan
, 1);
1345 consumed
= buf
->get_subbuf_consumed
;
1346 buf
->get_subbuf
= 0;
1349 * Clear the records_unread counter. (overruns counter)
1350 * Can still be non-zero if a file reader simply grabbed the data
1351 * without using iterators.
1352 * Can be below zero if an iterator is used on a snapshot more than
1355 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1356 v_add(config
, v_read(config
,
1357 &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
),
1358 &bufb
->records_read
);
1359 v_set(config
, &shmp(handle
, shmp_index(handle
, bufb
->array
, read_sb_bindex
)->shmp
)->records_unread
, 0);
1360 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1361 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1362 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1365 * Exchange the reader subbuffer with the one we put in its place in the
1366 * writer subbuffer table. Expect the original consumed count. If
1367 * update_read_sb_index fails, this is because the writer updated the
1368 * subbuffer concurrently. We should therefore keep the subbuffer we
1369 * currently have: it has become invalid to try reading this sub-buffer
1370 * consumed count value anyway.
1372 consumed_idx
= subbuf_index(consumed
, chan
);
1373 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1374 consumed_idx
, buf_trunc_val(consumed
, chan
),
1377 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1378 * if the writer concurrently updated it.
1383 * cons_offset is an iterator on all subbuffer offsets between the reader
1384 * position and the writer position. (inclusive)
1387 void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
1388 struct channel
*chan
,
1389 unsigned long cons_offset
,
1391 struct lttng_ust_shm_handle
*handle
)
1393 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1394 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1396 cons_idx
= subbuf_index(cons_offset
, chan
);
1397 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, cons_idx
)->cc
);
1398 commit_count_sb
= v_read(config
, &shmp_index(handle
, buf
->commit_cold
, cons_idx
)->cc_sb
);
1400 if (subbuf_offset(commit_count
, chan
) != 0)
1401 DBG("ring buffer %s, cpu %d: "
1402 "commit count in subbuffer %lu,\n"
1403 "expecting multiples of %lu bytes\n"
1404 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1405 chan
->backend
.name
, cpu
, cons_idx
,
1406 chan
->backend
.subbuf_size
,
1407 commit_count
, commit_count_sb
);
1409 DBG("ring buffer: %s, cpu %d: %lu bytes committed\n",
1410 chan
->backend
.name
, cpu
, commit_count
);
1414 void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
1415 struct channel
*chan
,
1416 void *priv
, int cpu
,
1417 struct lttng_ust_shm_handle
*handle
)
1419 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1420 unsigned long write_offset
, cons_offset
;
1423 * No need to order commit_count, write_offset and cons_offset reads
1424 * because we execute at teardown when no more writer nor reader
1425 * references are left.
1427 write_offset
= v_read(config
, &buf
->offset
);
1428 cons_offset
= uatomic_read(&buf
->consumed
);
1429 if (write_offset
!= cons_offset
)
1430 DBG("ring buffer %s, cpu %d: "
1431 "non-consumed data\n"
1432 " [ %lu bytes written, %lu bytes read ]\n",
1433 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1435 for (cons_offset
= uatomic_read(&buf
->consumed
);
1436 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1439 cons_offset
= subbuf_align(cons_offset
, chan
))
1440 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1445 void lib_ring_buffer_print_errors(struct channel
*chan
,
1446 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
1447 struct lttng_ust_shm_handle
*handle
)
1449 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1450 void *priv
= channel_get_private(chan
);
1452 if (!strcmp(chan
->backend
.name
, "relay-metadata-mmap")) {
1453 DBG("ring buffer %s: %lu records written, "
1454 "%lu records overrun\n",
1456 v_read(config
, &buf
->records_count
),
1457 v_read(config
, &buf
->records_overrun
));
1459 DBG("ring buffer %s, cpu %d: %lu records written, "
1460 "%lu records overrun\n",
1461 chan
->backend
.name
, cpu
,
1462 v_read(config
, &buf
->records_count
),
1463 v_read(config
, &buf
->records_overrun
));
1465 if (v_read(config
, &buf
->records_lost_full
)
1466 || v_read(config
, &buf
->records_lost_wrap
)
1467 || v_read(config
, &buf
->records_lost_big
))
1468 DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n"
1469 " [ %lu buffer full, %lu nest buffer wrap-around, "
1470 "%lu event too big ]\n",
1471 chan
->backend
.name
, cpu
,
1472 v_read(config
, &buf
->records_lost_full
),
1473 v_read(config
, &buf
->records_lost_wrap
),
1474 v_read(config
, &buf
->records_lost_big
));
1476 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
, handle
);
1480 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1482 * Only executed when the buffer is finalized, in SWITCH_FLUSH.
1485 void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer
*buf
,
1486 struct channel
*chan
,
1487 struct switch_offsets
*offsets
,
1489 struct lttng_ust_shm_handle
*handle
)
1491 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1492 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1493 unsigned long commit_count
;
1495 config
->cb
.buffer_begin(buf
, tsc
, oldidx
, handle
);
1498 * Order all writes to buffer before the commit count update that will
1499 * determine that the subbuffer is full.
1502 v_add(config
, config
->cb
.subbuffer_header_size(),
1503 &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1504 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1505 /* Check if the written buffer has to be delivered */
1506 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1507 commit_count
, oldidx
, handle
, tsc
);
1508 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1509 offsets
->old
+ config
->cb
.subbuffer_header_size(),
1510 commit_count
, handle
);
1514 * lib_ring_buffer_switch_old_end: switch old subbuffer
1516 * Note : offset_old should never be 0 here. It is ok, because we never perform
1517 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1518 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1522 void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer
*buf
,
1523 struct channel
*chan
,
1524 struct switch_offsets
*offsets
,
1526 struct lttng_ust_shm_handle
*handle
)
1528 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1529 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1530 unsigned long commit_count
, padding_size
, data_size
;
1532 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1533 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1534 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
,
1538 * Order all writes to buffer before the commit count update that will
1539 * determine that the subbuffer is full.
1542 v_add(config
, padding_size
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1543 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, oldidx
)->cc
);
1544 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1545 commit_count
, oldidx
, handle
, tsc
);
1546 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1547 offsets
->old
+ padding_size
, commit_count
, handle
);
1551 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1553 * This code can be executed unordered : writers may already have written to the
1554 * sub-buffer before this code gets executed, caution. The commit makes sure
1555 * that this code is executed before the deliver of this sub-buffer.
1558 void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer
*buf
,
1559 struct channel
*chan
,
1560 struct switch_offsets
*offsets
,
1562 struct lttng_ust_shm_handle
*handle
)
1564 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1565 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1566 unsigned long commit_count
;
1568 config
->cb
.buffer_begin(buf
, tsc
, beginidx
, handle
);
1571 * Order all writes to buffer before the commit count update that will
1572 * determine that the subbuffer is full.
1575 v_add(config
, config
->cb
.subbuffer_header_size(),
1576 &shmp_index(handle
, buf
->commit_hot
, beginidx
)->cc
);
1577 commit_count
= v_read(config
, &shmp_index(handle
, buf
->commit_hot
, beginidx
)->cc
);
1578 /* Check if the written buffer has to be delivered */
1579 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1580 commit_count
, beginidx
, handle
, tsc
);
1581 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, beginidx
,
1582 offsets
->begin
+ config
->cb
.subbuffer_header_size(),
1583 commit_count
, handle
);
1587 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1589 * Calls subbuffer_set_data_size() to set the data size of the current
1590 * sub-buffer. We do not need to perform check_deliver nor commit here,
1591 * since this task will be done by the "commit" of the event for which
1592 * we are currently doing the space reservation.
1595 void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer
*buf
,
1596 struct channel
*chan
,
1597 struct switch_offsets
*offsets
,
1599 struct lttng_ust_shm_handle
*handle
)
1601 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1602 unsigned long endidx
, data_size
;
1604 endidx
= subbuf_index(offsets
->end
- 1, chan
);
1605 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1606 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
,
1613 * !0 if execution must be aborted.
1616 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1617 struct lttng_ust_lib_ring_buffer
*buf
,
1618 struct channel
*chan
,
1619 struct switch_offsets
*offsets
,
1621 struct lttng_ust_shm_handle
*handle
)
1623 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1624 unsigned long off
, reserve_commit_diff
;
1626 offsets
->begin
= v_read(config
, &buf
->offset
);
1627 offsets
->old
= offsets
->begin
;
1628 offsets
->switch_old_start
= 0;
1629 off
= subbuf_offset(offsets
->begin
, chan
);
1631 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1634 * Ensure we flush the header of an empty subbuffer when doing the
1635 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1636 * total data gathering duration even if there were no records saved
1637 * after the last buffer switch.
1638 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1639 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1640 * subbuffer header as appropriate.
1641 * The next record that reserves space will be responsible for
1642 * populating the following subbuffer header. We choose not to populate
1643 * the next subbuffer header here because we want to be able to use
1644 * SWITCH_ACTIVE for periodical buffer flush, which must
1645 * guarantee that all the buffer content (records and header
1646 * timestamps) are visible to the reader. This is required for
1647 * quiescence guarantees for the fusion merge.
1649 if (mode
!= SWITCH_FLUSH
&& !off
)
1650 return -1; /* we do not have to switch : buffer is empty */
1652 if (caa_unlikely(off
== 0)) {
1653 unsigned long sb_index
, commit_count
;
1656 * We are performing a SWITCH_FLUSH. At this stage, there are no
1657 * concurrent writes into the buffer.
1659 * The client does not save any header information. Don't
1660 * switch empty subbuffer on finalize, because it is invalid to
1661 * deliver a completely empty subbuffer.
1663 if (!config
->cb
.subbuffer_header_size())
1666 /* Test new buffer integrity */
1667 sb_index
= subbuf_index(offsets
->begin
, chan
);
1668 commit_count
= v_read(config
,
1669 &shmp_index(handle
, buf
->commit_cold
,
1671 reserve_commit_diff
=
1672 (buf_trunc(offsets
->begin
, chan
)
1673 >> chan
->backend
.num_subbuf_order
)
1674 - (commit_count
& chan
->commit_count_mask
);
1675 if (caa_likely(reserve_commit_diff
== 0)) {
1676 /* Next subbuffer not being written to. */
1677 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1678 subbuf_trunc(offsets
->begin
, chan
)
1679 - subbuf_trunc((unsigned long)
1680 uatomic_read(&buf
->consumed
), chan
)
1681 >= chan
->backend
.buf_size
)) {
1683 * We do not overwrite non consumed buffers
1684 * and we are full : don't switch.
1689 * Next subbuffer not being written to, and we
1690 * are either in overwrite mode or the buffer is
1691 * not full. It's safe to write in this new
1697 * Next subbuffer reserve offset does not match the
1698 * commit offset. Don't perform switch in
1699 * producer-consumer and overwrite mode. Caused by
1700 * either a writer OOPS or too many nested writes over a
1701 * reserve/commit pair.
1707 * Need to write the subbuffer start header on finalize.
1709 offsets
->switch_old_start
= 1;
1711 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1712 /* Note: old points to the next subbuf at offset 0 */
1713 offsets
->end
= offsets
->begin
;
1718 * Force a sub-buffer switch. This operation is completely reentrant : can be
1719 * called while tracing is active with absolutely no lock held.
1721 * Note, however, that as a v_cmpxchg is used for some atomic
1722 * operations, this function must be called from the CPU which owns the buffer
1723 * for a ACTIVE flush.
1725 void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer
*buf
, enum switch_mode mode
,
1726 struct lttng_ust_shm_handle
*handle
)
1728 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1729 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1730 struct switch_offsets offsets
;
1731 unsigned long oldidx
;
1737 * Perform retryable operations.
1740 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1742 return; /* Switch not needed */
1743 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1747 * Atomically update last_tsc. This update races against concurrent
1748 * atomic updates, but the race will always cause supplementary full TSC
1749 * records, never the opposite (missing a full TSC record when it would
1752 save_last_tsc(config
, buf
, tsc
);
1755 * Push the reader if necessary
1757 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1759 oldidx
= subbuf_index(offsets
.old
, chan
);
1760 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
, handle
);
1763 * May need to populate header start on SWITCH_FLUSH.
1765 if (offsets
.switch_old_start
) {
1766 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
, handle
);
1767 offsets
.old
+= config
->cb
.subbuffer_header_size();
1771 * Switch old subbuffer.
1773 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
, handle
);
1779 * -ENOSPC if event size is too large for packet.
1780 * -ENOBUFS if there is currently not enough space in buffer for the event.
1781 * -EIO if data cannot be written into the buffer for any other reason.
1784 int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer
*buf
,
1785 struct channel
*chan
,
1786 struct switch_offsets
*offsets
,
1787 struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
1789 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1790 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
1791 unsigned long reserve_commit_diff
, offset_cmp
;
1794 offsets
->begin
= offset_cmp
= v_read(config
, &buf
->offset
);
1795 offsets
->old
= offsets
->begin
;
1796 offsets
->switch_new_start
= 0;
1797 offsets
->switch_new_end
= 0;
1798 offsets
->switch_old_end
= 0;
1799 offsets
->pre_header_padding
= 0;
1801 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1802 if ((int64_t) ctx
->tsc
== -EIO
)
1805 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1806 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
1808 if (caa_unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1809 offsets
->switch_new_start
= 1; /* For offsets->begin */
1811 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1813 &offsets
->pre_header_padding
,
1816 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1819 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
) +
1820 offsets
->size
> chan
->backend
.subbuf_size
)) {
1821 offsets
->switch_old_end
= 1; /* For offsets->old */
1822 offsets
->switch_new_start
= 1; /* For offsets->begin */
1825 if (caa_unlikely(offsets
->switch_new_start
)) {
1826 unsigned long sb_index
, commit_count
;
1829 * We are typically not filling the previous buffer completely.
1831 if (caa_likely(offsets
->switch_old_end
))
1832 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1833 offsets
->begin
= offsets
->begin
1834 + config
->cb
.subbuffer_header_size();
1835 /* Test new buffer integrity */
1836 sb_index
= subbuf_index(offsets
->begin
, chan
);
1838 * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
1839 * lib_ring_buffer_check_deliver() has the matching
1840 * memory barriers required around commit_cold cc_sb
1841 * updates to ensure reserve and commit counter updates
1842 * are not seen reordered when updated by another CPU.
1845 commit_count
= v_read(config
,
1846 &shmp_index(handle
, buf
->commit_cold
,
1848 /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
1850 if (caa_unlikely(offset_cmp
!= v_read(config
, &buf
->offset
))) {
1852 * The reserve counter have been concurrently updated
1853 * while we read the commit counter. This means the
1854 * commit counter we read might not match buf->offset
1855 * due to concurrent update. We therefore need to retry.
1859 reserve_commit_diff
=
1860 (buf_trunc(offsets
->begin
, chan
)
1861 >> chan
->backend
.num_subbuf_order
)
1862 - (commit_count
& chan
->commit_count_mask
);
1863 if (caa_likely(reserve_commit_diff
== 0)) {
1864 /* Next subbuffer not being written to. */
1865 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1866 subbuf_trunc(offsets
->begin
, chan
)
1867 - subbuf_trunc((unsigned long)
1868 uatomic_read(&buf
->consumed
), chan
)
1869 >= chan
->backend
.buf_size
)) {
1870 unsigned long nr_lost
;
1873 * We do not overwrite non consumed buffers
1874 * and we are full : record is lost.
1876 nr_lost
= v_read(config
, &buf
->records_lost_full
);
1877 v_inc(config
, &buf
->records_lost_full
);
1878 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
1879 DBG("%lu or more records lost in (%s:%d) (buffer full)\n",
1880 nr_lost
+ 1, chan
->backend
.name
,
1886 * Next subbuffer not being written to, and we
1887 * are either in overwrite mode or the buffer is
1888 * not full. It's safe to write in this new
1893 unsigned long nr_lost
;
1896 * Next subbuffer reserve offset does not match the
1897 * commit offset, and this did not involve update to the
1898 * reserve counter. Drop record in producer-consumer and
1899 * overwrite mode. Caused by either a writer OOPS or too
1900 * many nested writes over a reserve/commit pair.
1902 nr_lost
= v_read(config
, &buf
->records_lost_wrap
);
1903 v_inc(config
, &buf
->records_lost_wrap
);
1904 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
1905 DBG("%lu or more records lost in (%s:%d) (wrap-around)\n",
1906 nr_lost
+ 1, chan
->backend
.name
,
1912 config
->cb
.record_header_size(config
, chan
,
1914 &offsets
->pre_header_padding
,
1917 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1920 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
)
1921 + offsets
->size
> chan
->backend
.subbuf_size
)) {
1922 unsigned long nr_lost
;
1925 * Record too big for subbuffers, report error, don't
1926 * complete the sub-buffer switch.
1928 nr_lost
= v_read(config
, &buf
->records_lost_big
);
1929 v_inc(config
, &buf
->records_lost_big
);
1930 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
1931 DBG("%lu or more records lost in (%s:%d) record size "
1932 " of %zu bytes is too large for buffer\n",
1933 nr_lost
+ 1, chan
->backend
.name
,
1934 buf
->backend
.cpu
, offsets
->size
);
1939 * We just made a successful buffer switch and the
1940 * record fits in the new subbuffer. Let's write.
1945 * Record fits in the current buffer and we are not on a switch
1946 * boundary. It's safe to write.
1949 offsets
->end
= offsets
->begin
+ offsets
->size
;
1951 if (caa_unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
1953 * The offset_end will fall at the very beginning of the next
1956 offsets
->switch_new_end
= 1; /* For offsets->begin */
1962 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1963 * @ctx: ring buffer context.
1965 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
1966 * -EIO for other errors, else returns 0.
1967 * It will take care of sub-buffer switching.
1969 int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
1971 struct channel
*chan
= ctx
->chan
;
1972 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
1973 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1974 struct lttng_ust_lib_ring_buffer
*buf
;
1975 struct switch_offsets offsets
;
1978 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
1979 buf
= shmp(handle
, chan
->backend
.buf
[ctx
->cpu
].shmp
);
1981 buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
1987 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
1989 if (caa_unlikely(ret
))
1991 } while (caa_unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
1996 * Atomically update last_tsc. This update races against concurrent
1997 * atomic updates, but the race will always cause supplementary full TSC
1998 * records, never the opposite (missing a full TSC record when it would
2001 save_last_tsc(config
, buf
, ctx
->tsc
);
2004 * Push the reader if necessary
2006 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
2009 * Clear noref flag for this subbuffer.
2011 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2012 subbuf_index(offsets
.end
- 1, chan
),
2016 * Switch old subbuffer if needed.
2018 if (caa_unlikely(offsets
.switch_old_end
)) {
2019 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2020 subbuf_index(offsets
.old
- 1, chan
),
2022 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2026 * Populate new subbuffer.
2028 if (caa_unlikely(offsets
.switch_new_start
))
2029 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2031 if (caa_unlikely(offsets
.switch_new_end
))
2032 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2034 ctx
->slot_size
= offsets
.size
;
2035 ctx
->pre_offset
= offsets
.begin
;
2036 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
2041 * Force a read (imply TLS fixup for dlopen) of TLS variables.
2043 void lttng_fixup_ringbuffer_tls(void)
2045 asm volatile ("" : : "m" (URCU_TLS(lib_ring_buffer_nesting
)));
2048 void lib_ringbuffer_signal_init(void)
2054 * Block signal for entire process, so only our thread processes
2058 ret
= pthread_sigmask(SIG_BLOCK
, &mask
, NULL
);
2061 PERROR("pthread_sigmask");