2 * ring_buffer_frontend.c
4 * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; only
9 * version 2.1 of the License.
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 * Ring buffer wait-free buffer synchronization. Producer-consumer and flight
22 * recorder (overwrite) modes. See thesis:
24 * Desnoyers, Mathieu (2009), "Low-Impact Operating System Tracing", Ph.D.
25 * dissertation, Ecole Polytechnique de Montreal.
26 * http://www.lttng.org/pub/thesis/desnoyers-dissertation-2009-12.pdf
28 * - Algorithm presentation in Chapter 5:
29 * "Lockless Multi-Core High-Throughput Buffering".
30 * - Algorithm formal verification in Section 8.6:
31 * "Formal verification of LTTng"
34 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
36 * Inspired from LTT and RelayFS:
37 * Karim Yaghmour <karim@opersys.com>
38 * Tom Zanussi <zanussi@us.ibm.com>
39 * Bob Wisniewski <bob@watson.ibm.com>
41 * Bob Wisniewski <bob@watson.ibm.com>
43 * Buffer reader semantic :
46 * while buffer is not finalized and empty
48 * - if return value != 0, continue
49 * - splice one subbuffer worth of data to a pipe
50 * - splice the data from pipe to disk/network
56 #include <sys/types.h>
63 #include <urcu/compiler.h>
65 #include <urcu/tls-compat.h>
70 #include <lttng/ringbuffer-config.h>
76 #include "../liblttng-ust/compat.h" /* For ENODATA */
78 /* Print DBG() messages about events lost only every 1048576 hits */
79 #define DBG_PRINT_NR_LOST (1UL << 20)
81 #define LTTNG_UST_RB_SIG_FLUSH SIGRTMIN
82 #define LTTNG_UST_RB_SIG_READ SIGRTMIN + 1
83 #define LTTNG_UST_RB_SIG_TEARDOWN SIGRTMIN + 2
84 #define CLOCKID CLOCK_MONOTONIC
85 #define LTTNG_UST_RING_BUFFER_GET_RETRY 10
86 #define LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS 10
89 * Non-static to ensure the compiler does not optimize away the xor.
91 uint8_t lttng_crash_magic_xor
[] = RB_CRASH_DUMP_ABI_MAGIC_XOR
;
94 * Use POSIX SHM: shm_open(3) and shm_unlink(3).
95 * close(2) to close the fd returned by shm_open.
96 * shm_unlink releases the shared memory object name.
97 * ftruncate(2) sets the size of the memory object.
98 * mmap/munmap maps the shared memory obj to a virtual address in the
99 * calling proceess (should be done both in libust and consumer).
100 * See shm_overview(7) for details.
101 * Pass file descriptor returned by shm_open(3) to ltt-sessiond through
104 * Since we don't need to access the object using its name, we can
105 * immediately shm_unlink(3) it, and only keep the handle with its file
110 * Internal structure representing offsets to use at a sub-buffer switch.
112 struct switch_offsets
{
113 unsigned long begin
, end
, old
;
114 size_t pre_header_padding
, size
;
115 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
119 DEFINE_URCU_TLS(unsigned int, lib_ring_buffer_nesting
);
122 * wakeup_fd_mutex protects wakeup fd use by timer from concurrent
125 static pthread_mutex_t wakeup_fd_mutex
= PTHREAD_MUTEX_INITIALIZER
;
128 void lib_ring_buffer_print_errors(struct channel
*chan
,
129 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
130 struct lttng_ust_shm_handle
*handle
);
133 * Handle timer teardown race wrt memory free of private data by
134 * ring buffer signals are handled by a single thread, which permits
135 * a synchronization point between handling of each signal.
136 * Protected by the lock within the structure.
138 struct timer_signal_data
{
139 pthread_t tid
; /* thread id managing signals */
142 pthread_mutex_t lock
;
145 static struct timer_signal_data timer_signal
= {
149 .lock
= PTHREAD_MUTEX_INITIALIZER
,
153 * lib_ring_buffer_reset - Reset ring buffer to initial values.
156 * Effectively empty the ring buffer. Should be called when the buffer is not
157 * used for writing. The ring buffer can be opened for reading, but the reader
158 * should not be using the iterator concurrently with reset. The previous
159 * current iterator record is reset.
161 void lib_ring_buffer_reset(struct lttng_ust_lib_ring_buffer
*buf
,
162 struct lttng_ust_shm_handle
*handle
)
164 struct channel
*chan
;
165 const struct lttng_ust_lib_ring_buffer_config
*config
;
168 chan
= shmp(handle
, buf
->backend
.chan
);
171 config
= &chan
->backend
.config
;
173 * Reset iterator first. It will put the subbuffer if it currently holds
176 v_set(config
, &buf
->offset
, 0);
177 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
178 struct commit_counters_hot
*cc_hot
;
179 struct commit_counters_cold
*cc_cold
;
181 cc_hot
= shmp_index(handle
, buf
->commit_hot
, i
);
184 cc_cold
= shmp_index(handle
, buf
->commit_cold
, i
);
187 v_set(config
, &cc_hot
->cc
, 0);
188 v_set(config
, &cc_hot
->seq
, 0);
189 v_set(config
, &cc_cold
->cc_sb
, 0);
191 uatomic_set(&buf
->consumed
, 0);
192 uatomic_set(&buf
->record_disabled
, 0);
193 v_set(config
, &buf
->last_tsc
, 0);
194 lib_ring_buffer_backend_reset(&buf
->backend
, handle
);
195 /* Don't reset number of active readers */
196 v_set(config
, &buf
->records_lost_full
, 0);
197 v_set(config
, &buf
->records_lost_wrap
, 0);
198 v_set(config
, &buf
->records_lost_big
, 0);
199 v_set(config
, &buf
->records_count
, 0);
200 v_set(config
, &buf
->records_overrun
, 0);
205 * channel_reset - Reset channel to initial values.
208 * Effectively empty the channel. Should be called when the channel is not used
209 * for writing. The channel can be opened for reading, but the reader should not
210 * be using the iterator concurrently with reset. The previous current iterator
213 void channel_reset(struct channel
*chan
)
216 * Reset iterators first. Will put the subbuffer if held for reading.
218 uatomic_set(&chan
->record_disabled
, 0);
219 /* Don't reset commit_count_mask, still valid */
220 channel_backend_reset(&chan
->backend
);
221 /* Don't reset switch/read timer interval */
222 /* Don't reset notifiers and notifier enable bits */
223 /* Don't reset reader reference count */
227 void init_crash_abi(const struct lttng_ust_lib_ring_buffer_config
*config
,
228 struct lttng_crash_abi
*crash_abi
,
229 struct lttng_ust_lib_ring_buffer
*buf
,
230 struct channel_backend
*chanb
,
231 struct shm_object
*shmobj
,
232 struct lttng_ust_shm_handle
*handle
)
236 for (i
= 0; i
< RB_CRASH_DUMP_ABI_MAGIC_LEN
; i
++)
237 crash_abi
->magic
[i
] = lttng_crash_magic_xor
[i
] ^ 0xFF;
238 crash_abi
->mmap_length
= shmobj
->memory_map_size
;
239 crash_abi
->endian
= RB_CRASH_ENDIAN
;
240 crash_abi
->major
= RB_CRASH_DUMP_ABI_MAJOR
;
241 crash_abi
->minor
= RB_CRASH_DUMP_ABI_MINOR
;
242 crash_abi
->word_size
= sizeof(unsigned long);
243 crash_abi
->layout_type
= LTTNG_CRASH_TYPE_UST
;
245 /* Offset of fields */
246 crash_abi
->offset
.prod_offset
=
247 (uint32_t) ((char *) &buf
->offset
- (char *) buf
);
248 crash_abi
->offset
.consumed_offset
=
249 (uint32_t) ((char *) &buf
->consumed
- (char *) buf
);
250 crash_abi
->offset
.commit_hot_array
=
251 (uint32_t) ((char *) shmp(handle
, buf
->commit_hot
) - (char *) buf
);
252 crash_abi
->offset
.commit_hot_seq
=
253 offsetof(struct commit_counters_hot
, seq
);
254 crash_abi
->offset
.buf_wsb_array
=
255 (uint32_t) ((char *) shmp(handle
, buf
->backend
.buf_wsb
) - (char *) buf
);
256 crash_abi
->offset
.buf_wsb_id
=
257 offsetof(struct lttng_ust_lib_ring_buffer_backend_subbuffer
, id
);
258 crash_abi
->offset
.sb_array
=
259 (uint32_t) ((char *) shmp(handle
, buf
->backend
.array
) - (char *) buf
);
260 crash_abi
->offset
.sb_array_shmp_offset
=
261 offsetof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp
,
263 crash_abi
->offset
.sb_backend_p_offset
=
264 offsetof(struct lttng_ust_lib_ring_buffer_backend_pages
,
268 crash_abi
->length
.prod_offset
= sizeof(buf
->offset
);
269 crash_abi
->length
.consumed_offset
= sizeof(buf
->consumed
);
270 crash_abi
->length
.commit_hot_seq
=
271 sizeof(((struct commit_counters_hot
*) NULL
)->seq
);
272 crash_abi
->length
.buf_wsb_id
=
273 sizeof(((struct lttng_ust_lib_ring_buffer_backend_subbuffer
*) NULL
)->id
);
274 crash_abi
->length
.sb_array_shmp_offset
=
275 sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*) NULL
)->shmp
._ref
.offset
);
276 crash_abi
->length
.sb_backend_p_offset
=
277 sizeof(((struct lttng_ust_lib_ring_buffer_backend_pages
*) NULL
)->p
._ref
.offset
);
280 crash_abi
->stride
.commit_hot_array
=
281 sizeof(struct commit_counters_hot
);
282 crash_abi
->stride
.buf_wsb_array
=
283 sizeof(struct lttng_ust_lib_ring_buffer_backend_subbuffer
);
284 crash_abi
->stride
.sb_array
=
285 sizeof(struct lttng_ust_lib_ring_buffer_backend_pages_shmp
);
287 /* Buffer constants */
288 crash_abi
->buf_size
= chanb
->buf_size
;
289 crash_abi
->subbuf_size
= chanb
->subbuf_size
;
290 crash_abi
->num_subbuf
= chanb
->num_subbuf
;
291 crash_abi
->mode
= (uint32_t) chanb
->config
.mode
;
293 if (config
->cb
.content_size_field
) {
294 size_t offset
, length
;
296 config
->cb
.content_size_field(config
, &offset
, &length
);
297 crash_abi
->offset
.content_size
= offset
;
298 crash_abi
->length
.content_size
= length
;
300 crash_abi
->offset
.content_size
= 0;
301 crash_abi
->length
.content_size
= 0;
303 if (config
->cb
.packet_size_field
) {
304 size_t offset
, length
;
306 config
->cb
.packet_size_field(config
, &offset
, &length
);
307 crash_abi
->offset
.packet_size
= offset
;
308 crash_abi
->length
.packet_size
= length
;
310 crash_abi
->offset
.packet_size
= 0;
311 crash_abi
->length
.packet_size
= 0;
316 * Must be called under cpu hotplug protection.
318 int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer
*buf
,
319 struct channel_backend
*chanb
, int cpu
,
320 struct lttng_ust_shm_handle
*handle
,
321 struct shm_object
*shmobj
)
323 const struct lttng_ust_lib_ring_buffer_config
*config
= &chanb
->config
;
324 struct channel
*chan
= caa_container_of(chanb
, struct channel
, backend
);
325 struct lttng_ust_lib_ring_buffer_backend_subbuffer
*wsb
;
326 struct channel
*shmp_chan
;
327 struct commit_counters_hot
*cc_hot
;
328 void *priv
= channel_get_private(chan
);
329 size_t subbuf_header_size
;
333 /* Test for cpu hotplug */
334 if (buf
->backend
.allocated
)
337 align_shm(shmobj
, __alignof__(struct commit_counters_hot
));
338 set_shmp(buf
->commit_hot
,
340 sizeof(struct commit_counters_hot
) * chan
->backend
.num_subbuf
));
341 if (!shmp(handle
, buf
->commit_hot
)) {
345 align_shm(shmobj
, __alignof__(struct commit_counters_cold
));
346 set_shmp(buf
->commit_cold
,
348 sizeof(struct commit_counters_cold
) * chan
->backend
.num_subbuf
));
349 if (!shmp(handle
, buf
->commit_cold
)) {
354 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
,
355 cpu
, handle
, shmobj
);
361 * Write the subbuffer header for first subbuffer so we know the total
362 * duration of data gathering.
364 subbuf_header_size
= config
->cb
.subbuffer_header_size();
365 v_set(config
, &buf
->offset
, subbuf_header_size
);
366 wsb
= shmp_index(handle
, buf
->backend
.buf_wsb
, 0);
371 subbuffer_id_clear_noref(config
, &wsb
->id
);
372 shmp_chan
= shmp(handle
, buf
->backend
.chan
);
377 tsc
= config
->cb
.ring_buffer_clock_read(shmp_chan
);
378 config
->cb
.buffer_begin(buf
, tsc
, 0, handle
);
379 cc_hot
= shmp_index(handle
, buf
->commit_hot
, 0);
384 v_add(config
, subbuf_header_size
, &cc_hot
->cc
);
385 v_add(config
, subbuf_header_size
, &cc_hot
->seq
);
387 if (config
->cb
.buffer_create
) {
388 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
, handle
);
393 init_crash_abi(config
, &buf
->crash_abi
, buf
, chanb
, shmobj
, handle
);
395 buf
->backend
.allocated
= 1;
400 /* commit_cold will be freed by shm teardown */
402 /* commit_hot will be freed by shm teardown */
408 void lib_ring_buffer_channel_switch_timer(int sig
, siginfo_t
*si
, void *uc
)
410 const struct lttng_ust_lib_ring_buffer_config
*config
;
411 struct lttng_ust_shm_handle
*handle
;
412 struct channel
*chan
;
415 assert(CMM_LOAD_SHARED(timer_signal
.tid
) == pthread_self());
417 chan
= si
->si_value
.sival_ptr
;
418 handle
= chan
->handle
;
419 config
= &chan
->backend
.config
;
421 DBG("Switch timer for channel %p\n", chan
);
424 * Only flush buffers periodically if readers are active.
426 pthread_mutex_lock(&wakeup_fd_mutex
);
427 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
428 for_each_possible_cpu(cpu
) {
429 struct lttng_ust_lib_ring_buffer
*buf
=
430 shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
434 if (uatomic_read(&buf
->active_readers
))
435 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
,
439 struct lttng_ust_lib_ring_buffer
*buf
=
440 shmp(handle
, chan
->backend
.buf
[0].shmp
);
444 if (uatomic_read(&buf
->active_readers
))
445 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
,
449 pthread_mutex_unlock(&wakeup_fd_mutex
);
454 int lib_ring_buffer_poll_deliver(const struct lttng_ust_lib_ring_buffer_config
*config
,
455 struct lttng_ust_lib_ring_buffer
*buf
,
456 struct channel
*chan
,
457 struct lttng_ust_shm_handle
*handle
)
459 unsigned long consumed_old
, consumed_idx
, commit_count
, write_offset
;
460 struct commit_counters_cold
*cc_cold
;
462 consumed_old
= uatomic_read(&buf
->consumed
);
463 consumed_idx
= subbuf_index(consumed_old
, chan
);
464 cc_cold
= shmp_index(handle
, buf
->commit_cold
, consumed_idx
);
467 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
469 * No memory barrier here, since we are only interested
470 * in a statistically correct polling result. The next poll will
471 * get the data is we are racing. The mb() that ensures correct
472 * memory order is in get_subbuf.
474 write_offset
= v_read(config
, &buf
->offset
);
477 * Check that the subbuffer we are trying to consume has been
478 * already fully committed.
481 if (((commit_count
- chan
->backend
.subbuf_size
)
482 & chan
->commit_count_mask
)
483 - (buf_trunc(consumed_old
, chan
)
484 >> chan
->backend
.num_subbuf_order
)
489 * Check that we are not about to read the same subbuffer in
490 * which the writer head is.
492 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_old
, chan
)
500 void lib_ring_buffer_wakeup(struct lttng_ust_lib_ring_buffer
*buf
,
501 struct lttng_ust_shm_handle
*handle
)
503 int wakeup_fd
= shm_get_wakeup_fd(handle
, &buf
->self
._ref
);
504 sigset_t sigpipe_set
, pending_set
, old_set
;
505 int ret
, sigpipe_was_pending
= 0;
511 * Wake-up the other end by writing a null byte in the pipe
512 * (non-blocking). Important note: Because writing into the
513 * pipe is non-blocking (and therefore we allow dropping wakeup
514 * data, as long as there is wakeup data present in the pipe
515 * buffer to wake up the consumer), the consumer should perform
516 * the following sequence for waiting:
517 * 1) empty the pipe (reads).
518 * 2) check if there is data in the buffer.
519 * 3) wait on the pipe (poll).
521 * Discard the SIGPIPE from write(), not disturbing any SIGPIPE
522 * that might be already pending. If a bogus SIGPIPE is sent to
523 * the entire process concurrently by a malicious user, it may
524 * be simply discarded.
526 ret
= sigemptyset(&pending_set
);
529 * sigpending returns the mask of signals that are _both_
530 * blocked for the thread _and_ pending for either the thread or
531 * the entire process.
533 ret
= sigpending(&pending_set
);
535 sigpipe_was_pending
= sigismember(&pending_set
, SIGPIPE
);
537 * If sigpipe was pending, it means it was already blocked, so
538 * no need to block it.
540 if (!sigpipe_was_pending
) {
541 ret
= sigemptyset(&sigpipe_set
);
543 ret
= sigaddset(&sigpipe_set
, SIGPIPE
);
545 ret
= pthread_sigmask(SIG_BLOCK
, &sigpipe_set
, &old_set
);
549 ret
= write(wakeup_fd
, "", 1);
550 } while (ret
== -1L && errno
== EINTR
);
551 if (ret
== -1L && errno
== EPIPE
&& !sigpipe_was_pending
) {
552 struct timespec timeout
= { 0, 0 };
554 ret
= sigtimedwait(&sigpipe_set
, NULL
,
556 } while (ret
== -1L && errno
== EINTR
);
558 if (!sigpipe_was_pending
) {
559 ret
= pthread_sigmask(SIG_SETMASK
, &old_set
, NULL
);
565 void lib_ring_buffer_channel_do_read(struct channel
*chan
)
567 const struct lttng_ust_lib_ring_buffer_config
*config
;
568 struct lttng_ust_shm_handle
*handle
;
571 handle
= chan
->handle
;
572 config
= &chan
->backend
.config
;
575 * Only flush buffers periodically if readers are active.
577 pthread_mutex_lock(&wakeup_fd_mutex
);
578 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
579 for_each_possible_cpu(cpu
) {
580 struct lttng_ust_lib_ring_buffer
*buf
=
581 shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
585 if (uatomic_read(&buf
->active_readers
)
586 && lib_ring_buffer_poll_deliver(config
, buf
,
588 lib_ring_buffer_wakeup(buf
, handle
);
592 struct lttng_ust_lib_ring_buffer
*buf
=
593 shmp(handle
, chan
->backend
.buf
[0].shmp
);
597 if (uatomic_read(&buf
->active_readers
)
598 && lib_ring_buffer_poll_deliver(config
, buf
,
600 lib_ring_buffer_wakeup(buf
, handle
);
604 pthread_mutex_unlock(&wakeup_fd_mutex
);
608 void lib_ring_buffer_channel_read_timer(int sig
, siginfo_t
*si
, void *uc
)
610 struct channel
*chan
;
612 assert(CMM_LOAD_SHARED(timer_signal
.tid
) == pthread_self());
613 chan
= si
->si_value
.sival_ptr
;
614 DBG("Read timer for channel %p\n", chan
);
615 lib_ring_buffer_channel_do_read(chan
);
620 void rb_setmask(sigset_t
*mask
)
624 ret
= sigemptyset(mask
);
626 PERROR("sigemptyset");
628 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_FLUSH
);
632 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_READ
);
636 ret
= sigaddset(mask
, LTTNG_UST_RB_SIG_TEARDOWN
);
643 void *sig_thread(void *arg
)
649 /* Only self thread will receive signal mask. */
651 CMM_STORE_SHARED(timer_signal
.tid
, pthread_self());
654 signr
= sigwaitinfo(&mask
, &info
);
657 PERROR("sigwaitinfo");
660 if (signr
== LTTNG_UST_RB_SIG_FLUSH
) {
661 lib_ring_buffer_channel_switch_timer(info
.si_signo
,
663 } else if (signr
== LTTNG_UST_RB_SIG_READ
) {
664 lib_ring_buffer_channel_read_timer(info
.si_signo
,
666 } else if (signr
== LTTNG_UST_RB_SIG_TEARDOWN
) {
668 CMM_STORE_SHARED(timer_signal
.qs_done
, 1);
671 ERR("Unexptected signal %d\n", info
.si_signo
);
678 * Ensure only a single thread listens on the timer signal.
681 void lib_ring_buffer_setup_timer_thread(void)
686 pthread_mutex_lock(&timer_signal
.lock
);
687 if (timer_signal
.setup_done
)
690 ret
= pthread_create(&thread
, NULL
, &sig_thread
, NULL
);
693 PERROR("pthread_create");
695 ret
= pthread_detach(thread
);
698 PERROR("pthread_detach");
700 timer_signal
.setup_done
= 1;
702 pthread_mutex_unlock(&timer_signal
.lock
);
706 * Wait for signal-handling thread quiescent state.
709 void lib_ring_buffer_wait_signal_thread_qs(unsigned int signr
)
711 sigset_t pending_set
;
715 * We need to be the only thread interacting with the thread
716 * that manages signals for teardown synchronization.
718 pthread_mutex_lock(&timer_signal
.lock
);
721 * Ensure we don't have any signal queued for this channel.
724 ret
= sigemptyset(&pending_set
);
726 PERROR("sigemptyset");
728 ret
= sigpending(&pending_set
);
730 PERROR("sigpending");
732 if (!sigismember(&pending_set
, signr
))
738 * From this point, no new signal handler will be fired that
739 * would try to access "chan". However, we still need to wait
740 * for any currently executing handler to complete.
743 CMM_STORE_SHARED(timer_signal
.qs_done
, 0);
747 * Kill with LTTNG_UST_RB_SIG_TEARDOWN, so signal management
750 kill(getpid(), LTTNG_UST_RB_SIG_TEARDOWN
);
752 while (!CMM_LOAD_SHARED(timer_signal
.qs_done
))
756 pthread_mutex_unlock(&timer_signal
.lock
);
760 void lib_ring_buffer_channel_switch_timer_start(struct channel
*chan
)
763 struct itimerspec its
;
766 if (!chan
->switch_timer_interval
|| chan
->switch_timer_enabled
)
769 chan
->switch_timer_enabled
= 1;
771 lib_ring_buffer_setup_timer_thread();
773 sev
.sigev_notify
= SIGEV_SIGNAL
;
774 sev
.sigev_signo
= LTTNG_UST_RB_SIG_FLUSH
;
775 sev
.sigev_value
.sival_ptr
= chan
;
776 ret
= timer_create(CLOCKID
, &sev
, &chan
->switch_timer
);
778 PERROR("timer_create");
781 its
.it_value
.tv_sec
= chan
->switch_timer_interval
/ 1000000;
782 its
.it_value
.tv_nsec
= (chan
->switch_timer_interval
% 1000000) * 1000;
783 its
.it_interval
.tv_sec
= its
.it_value
.tv_sec
;
784 its
.it_interval
.tv_nsec
= its
.it_value
.tv_nsec
;
786 ret
= timer_settime(chan
->switch_timer
, 0, &its
, NULL
);
788 PERROR("timer_settime");
793 void lib_ring_buffer_channel_switch_timer_stop(struct channel
*chan
)
797 if (!chan
->switch_timer_interval
|| !chan
->switch_timer_enabled
)
800 ret
= timer_delete(chan
->switch_timer
);
802 PERROR("timer_delete");
805 lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_FLUSH
);
807 chan
->switch_timer
= 0;
808 chan
->switch_timer_enabled
= 0;
812 void lib_ring_buffer_channel_read_timer_start(struct channel
*chan
)
814 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
816 struct itimerspec its
;
819 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
820 || !chan
->read_timer_interval
|| chan
->read_timer_enabled
)
823 chan
->read_timer_enabled
= 1;
825 lib_ring_buffer_setup_timer_thread();
827 sev
.sigev_notify
= SIGEV_SIGNAL
;
828 sev
.sigev_signo
= LTTNG_UST_RB_SIG_READ
;
829 sev
.sigev_value
.sival_ptr
= chan
;
830 ret
= timer_create(CLOCKID
, &sev
, &chan
->read_timer
);
832 PERROR("timer_create");
835 its
.it_value
.tv_sec
= chan
->read_timer_interval
/ 1000000;
836 its
.it_value
.tv_nsec
= (chan
->read_timer_interval
% 1000000) * 1000;
837 its
.it_interval
.tv_sec
= its
.it_value
.tv_sec
;
838 its
.it_interval
.tv_nsec
= its
.it_value
.tv_nsec
;
840 ret
= timer_settime(chan
->read_timer
, 0, &its
, NULL
);
842 PERROR("timer_settime");
847 void lib_ring_buffer_channel_read_timer_stop(struct channel
*chan
)
849 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
852 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
853 || !chan
->read_timer_interval
|| !chan
->read_timer_enabled
)
856 ret
= timer_delete(chan
->read_timer
);
858 PERROR("timer_delete");
862 * do one more check to catch data that has been written in the last
865 lib_ring_buffer_channel_do_read(chan
);
867 lib_ring_buffer_wait_signal_thread_qs(LTTNG_UST_RB_SIG_READ
);
869 chan
->read_timer
= 0;
870 chan
->read_timer_enabled
= 0;
873 static void channel_unregister_notifiers(struct channel
*chan
,
874 struct lttng_ust_shm_handle
*handle
)
876 lib_ring_buffer_channel_switch_timer_stop(chan
);
877 lib_ring_buffer_channel_read_timer_stop(chan
);
880 static void channel_print_errors(struct channel
*chan
,
881 struct lttng_ust_shm_handle
*handle
)
883 const struct lttng_ust_lib_ring_buffer_config
*config
=
884 &chan
->backend
.config
;
887 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
888 for_each_possible_cpu(cpu
) {
889 struct lttng_ust_lib_ring_buffer
*buf
=
890 shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
892 lib_ring_buffer_print_errors(chan
, buf
, cpu
, handle
);
895 struct lttng_ust_lib_ring_buffer
*buf
=
896 shmp(handle
, chan
->backend
.buf
[0].shmp
);
899 lib_ring_buffer_print_errors(chan
, buf
, -1, handle
);
903 static void channel_free(struct channel
*chan
,
904 struct lttng_ust_shm_handle
*handle
,
907 channel_backend_free(&chan
->backend
, handle
);
908 /* chan is freed by shm teardown */
909 shm_object_table_destroy(handle
->table
, consumer
);
914 * channel_create - Create channel.
915 * @config: ring buffer instance configuration
916 * @name: name of the channel
917 * @priv_data: ring buffer client private data area pointer (output)
918 * @priv_data_size: length, in bytes, of the private data area.
919 * @priv_data_init: initialization data for private data.
920 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
921 * address mapping. It is used only by RING_BUFFER_STATIC
922 * configuration. It can be set to NULL for other backends.
923 * @subbuf_size: subbuffer size
924 * @num_subbuf: number of subbuffers
925 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
926 * padding to let readers get those sub-buffers.
927 * Used for live streaming.
928 * @read_timer_interval: Time interval (in us) to wake up pending readers.
929 * @stream_fds: array of stream file descriptors.
930 * @nr_stream_fds: number of file descriptors in array.
933 * Returns NULL on failure.
935 struct lttng_ust_shm_handle
*channel_create(const struct lttng_ust_lib_ring_buffer_config
*config
,
938 size_t priv_data_align
,
939 size_t priv_data_size
,
940 void *priv_data_init
,
941 void *buf_addr
, size_t subbuf_size
,
942 size_t num_subbuf
, unsigned int switch_timer_interval
,
943 unsigned int read_timer_interval
,
944 const int *stream_fds
, int nr_stream_fds
)
947 size_t shmsize
, chansize
;
948 struct channel
*chan
;
949 struct lttng_ust_shm_handle
*handle
;
950 struct shm_object
*shmobj
;
951 unsigned int nr_streams
;
953 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
954 nr_streams
= num_possible_cpus();
958 if (nr_stream_fds
!= nr_streams
)
961 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
962 read_timer_interval
))
965 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
969 /* Allocate table for channel + per-cpu buffers */
970 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
972 goto error_table_alloc
;
974 /* Calculate the shm allocation layout */
975 shmsize
= sizeof(struct channel
);
976 shmsize
+= offset_align(shmsize
, __alignof__(struct lttng_ust_lib_ring_buffer_shmp
));
977 shmsize
+= sizeof(struct lttng_ust_lib_ring_buffer_shmp
) * nr_streams
;
980 shmsize
+= offset_align(shmsize
, priv_data_align
);
981 shmsize
+= priv_data_size
;
983 /* Allocate normal memory for channel (not shared) */
984 shmobj
= shm_object_table_alloc(handle
->table
, shmsize
, SHM_OBJECT_MEM
,
988 /* struct channel is at object 0, offset 0 (hardcoded) */
989 set_shmp(handle
->chan
, zalloc_shm(shmobj
, chansize
));
990 assert(handle
->chan
._ref
.index
== 0);
991 assert(handle
->chan
._ref
.offset
== 0);
992 chan
= shmp(handle
, handle
->chan
);
995 chan
->nr_streams
= nr_streams
;
997 /* space for private data */
998 if (priv_data_size
) {
999 DECLARE_SHMP(void, priv_data_alloc
);
1001 align_shm(shmobj
, priv_data_align
);
1002 chan
->priv_data_offset
= shmobj
->allocated_len
;
1003 set_shmp(priv_data_alloc
, zalloc_shm(shmobj
, priv_data_size
));
1004 if (!shmp(handle
, priv_data_alloc
))
1006 *priv_data
= channel_get_private(chan
);
1007 memcpy(*priv_data
, priv_data_init
, priv_data_size
);
1009 chan
->priv_data_offset
= -1;
1014 ret
= channel_backend_init(&chan
->backend
, name
, config
,
1015 subbuf_size
, num_subbuf
, handle
,
1018 goto error_backend_init
;
1020 chan
->handle
= handle
;
1021 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
1023 chan
->switch_timer_interval
= switch_timer_interval
;
1024 chan
->read_timer_interval
= read_timer_interval
;
1025 lib_ring_buffer_channel_switch_timer_start(chan
);
1026 lib_ring_buffer_channel_read_timer_start(chan
);
1032 shm_object_table_destroy(handle
->table
, 1);
1038 struct lttng_ust_shm_handle
*channel_handle_create(void *data
,
1039 uint64_t memory_map_size
,
1042 struct lttng_ust_shm_handle
*handle
;
1043 struct shm_object
*object
;
1045 handle
= zmalloc(sizeof(struct lttng_ust_shm_handle
));
1049 /* Allocate table for channel + per-cpu buffers */
1050 handle
->table
= shm_object_table_create(1 + num_possible_cpus());
1052 goto error_table_alloc
;
1053 /* Add channel object */
1054 object
= shm_object_table_append_mem(handle
->table
, data
,
1055 memory_map_size
, wakeup_fd
);
1057 goto error_table_object
;
1058 /* struct channel is at object 0, offset 0 (hardcoded) */
1059 handle
->chan
._ref
.index
= 0;
1060 handle
->chan
._ref
.offset
= 0;
1064 shm_object_table_destroy(handle
->table
, 0);
1070 int channel_handle_add_stream(struct lttng_ust_shm_handle
*handle
,
1071 int shm_fd
, int wakeup_fd
, uint32_t stream_nr
,
1072 uint64_t memory_map_size
)
1074 struct shm_object
*object
;
1076 /* Add stream object */
1077 object
= shm_object_table_append_shm(handle
->table
,
1078 shm_fd
, wakeup_fd
, stream_nr
,
1085 unsigned int channel_handle_get_nr_streams(struct lttng_ust_shm_handle
*handle
)
1087 assert(handle
->table
);
1088 return handle
->table
->allocated_len
- 1;
1092 void channel_release(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
1095 channel_free(chan
, handle
, consumer
);
1099 * channel_destroy - Finalize, wait for q.s. and destroy channel.
1100 * @chan: channel to destroy
1102 * Holds cpu hotplug.
1103 * Call "destroy" callback, finalize channels, decrement the channel
1104 * reference count. Note that when readers have completed data
1105 * consumption of finalized channels, get_subbuf() will return -ENODATA.
1106 * They should release their handle at that point.
1108 void channel_destroy(struct channel
*chan
, struct lttng_ust_shm_handle
*handle
,
1113 * Note: the consumer takes care of finalizing and
1114 * switching the buffers.
1116 channel_unregister_notifiers(chan
, handle
);
1118 * The consumer prints errors.
1120 channel_print_errors(chan
, handle
);
1124 * sessiond/consumer are keeping a reference on the shm file
1125 * descriptor directly. No need to refcount.
1127 channel_release(chan
, handle
, consumer
);
1131 struct lttng_ust_lib_ring_buffer
*channel_get_ring_buffer(
1132 const struct lttng_ust_lib_ring_buffer_config
*config
,
1133 struct channel
*chan
, int cpu
,
1134 struct lttng_ust_shm_handle
*handle
,
1135 int *shm_fd
, int *wait_fd
,
1137 uint64_t *memory_map_size
)
1139 struct shm_ref
*ref
;
1141 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1144 if (cpu
>= num_possible_cpus())
1147 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1148 *shm_fd
= shm_get_shm_fd(handle
, ref
);
1149 *wait_fd
= shm_get_wait_fd(handle
, ref
);
1150 *wakeup_fd
= shm_get_wakeup_fd(handle
, ref
);
1151 if (shm_get_shm_size(handle
, ref
, memory_map_size
))
1153 return shmp(handle
, chan
->backend
.buf
[cpu
].shmp
);
1156 int ring_buffer_channel_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1157 struct channel
*chan
,
1158 struct lttng_ust_shm_handle
*handle
)
1160 struct shm_ref
*ref
;
1162 ref
= &handle
->chan
._ref
;
1163 return shm_close_wait_fd(handle
, ref
);
1166 int ring_buffer_channel_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1167 struct channel
*chan
,
1168 struct lttng_ust_shm_handle
*handle
)
1170 struct shm_ref
*ref
;
1172 ref
= &handle
->chan
._ref
;
1173 return shm_close_wakeup_fd(handle
, ref
);
1176 int ring_buffer_stream_close_wait_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1177 struct channel
*chan
,
1178 struct lttng_ust_shm_handle
*handle
,
1181 struct shm_ref
*ref
;
1183 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1186 if (cpu
>= num_possible_cpus())
1189 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1190 return shm_close_wait_fd(handle
, ref
);
1193 int ring_buffer_stream_close_wakeup_fd(const struct lttng_ust_lib_ring_buffer_config
*config
,
1194 struct channel
*chan
,
1195 struct lttng_ust_shm_handle
*handle
,
1198 struct shm_ref
*ref
;
1201 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
) {
1204 if (cpu
>= num_possible_cpus())
1207 ref
= &chan
->backend
.buf
[cpu
].shmp
._ref
;
1208 pthread_mutex_lock(&wakeup_fd_mutex
);
1209 ret
= shm_close_wakeup_fd(handle
, ref
);
1210 pthread_mutex_unlock(&wakeup_fd_mutex
);
1214 int lib_ring_buffer_open_read(struct lttng_ust_lib_ring_buffer
*buf
,
1215 struct lttng_ust_shm_handle
*handle
)
1217 if (uatomic_cmpxchg(&buf
->active_readers
, 0, 1) != 0)
1223 void lib_ring_buffer_release_read(struct lttng_ust_lib_ring_buffer
*buf
,
1224 struct lttng_ust_shm_handle
*handle
)
1226 struct channel
*chan
= shmp(handle
, buf
->backend
.chan
);
1230 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1232 uatomic_dec(&buf
->active_readers
);
1236 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
1238 * @consumed: consumed count indicating the position where to read
1239 * @produced: produced count, indicates position when to stop reading
1241 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1242 * data to read at consumed position, or 0 if the get operation succeeds.
1245 int lib_ring_buffer_snapshot(struct lttng_ust_lib_ring_buffer
*buf
,
1246 unsigned long *consumed
, unsigned long *produced
,
1247 struct lttng_ust_shm_handle
*handle
)
1249 struct channel
*chan
;
1250 const struct lttng_ust_lib_ring_buffer_config
*config
;
1251 unsigned long consumed_cur
, write_offset
;
1254 chan
= shmp(handle
, buf
->backend
.chan
);
1257 config
= &chan
->backend
.config
;
1258 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
1260 * Read finalized before counters.
1263 consumed_cur
= uatomic_read(&buf
->consumed
);
1265 * No need to issue a memory barrier between consumed count read and
1266 * write offset read, because consumed count can only change
1267 * concurrently in overwrite mode, and we keep a sequence counter
1268 * identifier derived from the write offset to check we are getting
1269 * the same sub-buffer we are expecting (the sub-buffers are atomically
1270 * "tagged" upon writes, tags are checked upon read).
1272 write_offset
= v_read(config
, &buf
->offset
);
1275 * Check that we are not about to read the same subbuffer in
1276 * which the writer head is.
1278 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
1282 *consumed
= consumed_cur
;
1283 *produced
= subbuf_trunc(write_offset
, chan
);
1289 * The memory barriers __wait_event()/wake_up_interruptible() take care
1290 * of "raw_spin_is_locked" memory ordering.
1299 * lib_ring_buffer_move_consumer - move consumed counter forward
1301 * @consumed_new: new consumed count value
1303 void lib_ring_buffer_move_consumer(struct lttng_ust_lib_ring_buffer
*buf
,
1304 unsigned long consumed_new
,
1305 struct lttng_ust_shm_handle
*handle
)
1307 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1308 struct channel
*chan
;
1309 unsigned long consumed
;
1311 chan
= shmp(handle
, bufb
->chan
);
1314 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1317 * Only push the consumed value forward.
1318 * If the consumed cmpxchg fails, this is because we have been pushed by
1319 * the writer in flight recorder mode.
1321 consumed
= uatomic_read(&buf
->consumed
);
1322 while ((long) consumed
- (long) consumed_new
< 0)
1323 consumed
= uatomic_cmpxchg(&buf
->consumed
, consumed
,
1328 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
1330 * @consumed: consumed count indicating the position where to read
1332 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1333 * data to read at consumed position, or 0 if the get operation succeeds.
1335 int lib_ring_buffer_get_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
1336 unsigned long consumed
,
1337 struct lttng_ust_shm_handle
*handle
)
1339 struct channel
*chan
;
1340 const struct lttng_ust_lib_ring_buffer_config
*config
;
1341 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
1342 int ret
, finalized
, nr_retry
= LTTNG_UST_RING_BUFFER_GET_RETRY
;
1343 struct commit_counters_cold
*cc_cold
;
1345 chan
= shmp(handle
, buf
->backend
.chan
);
1348 config
= &chan
->backend
.config
;
1350 finalized
= CMM_ACCESS_ONCE(buf
->finalized
);
1352 * Read finalized before counters.
1355 consumed_cur
= uatomic_read(&buf
->consumed
);
1356 consumed_idx
= subbuf_index(consumed
, chan
);
1357 cc_cold
= shmp_index(handle
, buf
->commit_cold
, consumed_idx
);
1360 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
1362 * Make sure we read the commit count before reading the buffer
1363 * data and the write offset. Correct consumed offset ordering
1364 * wrt commit count is insured by the use of cmpxchg to update
1365 * the consumed offset.
1368 * Local rmb to match the remote wmb to read the commit count
1369 * before the buffer data and the write offset.
1373 write_offset
= v_read(config
, &buf
->offset
);
1376 * Check that the buffer we are getting is after or at consumed_cur
1379 if ((long) subbuf_trunc(consumed
, chan
)
1380 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
1384 * Check that the subbuffer we are trying to consume has been
1385 * already fully committed. There are a few causes that can make
1386 * this unavailability situation occur:
1388 * Temporary (short-term) situation:
1389 * - Application is running on a different CPU, between reserve
1390 * and commit ring buffer operations,
1391 * - Application is preempted between reserve and commit ring
1392 * buffer operations,
1394 * Long-term situation:
1395 * - Application is stopped (SIGSTOP) between reserve and commit
1396 * ring buffer operations. Could eventually be resumed by
1398 * - Application is killed (SIGTERM, SIGINT, SIGKILL) between
1399 * reserve and commit ring buffer operation.
1401 * From a consumer perspective, handling short-term
1402 * unavailability situations is performed by retrying a few
1403 * times after a delay. Handling long-term unavailability
1404 * situations is handled by failing to get the sub-buffer.
1406 * In all of those situations, if the application is taking a
1407 * long time to perform its commit after ring buffer space
1408 * reservation, we can end up in a situation where the producer
1409 * will fill the ring buffer and try to write into the same
1410 * sub-buffer again (which has a missing commit). This is
1411 * handled by the producer in the sub-buffer switch handling
1412 * code of the reserve routine by detecting unbalanced
1413 * reserve/commit counters and discarding all further events
1414 * until the situation is resolved in those situations. Two
1415 * scenarios can occur:
1417 * 1) The application causing the reserve/commit counters to be
1418 * unbalanced has been terminated. In this situation, all
1419 * further events will be discarded in the buffers, and no
1420 * further buffer data will be readable by the consumer
1421 * daemon. Tearing down the UST tracing session and starting
1422 * anew is a work-around for those situations. Note that this
1423 * only affects per-UID tracing. In per-PID tracing, the
1424 * application vanishes with the termination, and therefore
1425 * no more data needs to be written to the buffers.
1426 * 2) The application causing the unbalance has been delayed for
1427 * a long time, but will eventually try to increment the
1428 * commit counter after eventually writing to the sub-buffer.
1429 * This situation can cause events to be discarded until the
1430 * application resumes its operations.
1432 if (((commit_count
- chan
->backend
.subbuf_size
)
1433 & chan
->commit_count_mask
)
1434 - (buf_trunc(consumed
, chan
)
1435 >> chan
->backend
.num_subbuf_order
)
1437 if (nr_retry
-- > 0) {
1438 if (nr_retry
<= (LTTNG_UST_RING_BUFFER_GET_RETRY
>> 1))
1439 (void) poll(NULL
, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS
);
1447 * Check that we are not about to read the same subbuffer in
1448 * which the writer head is.
1450 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed
, chan
)
1455 * Failure to get the subbuffer causes a busy-loop retry without going
1456 * to a wait queue. These are caused by short-lived race windows where
1457 * the writer is getting access to a subbuffer we were trying to get
1458 * access to. Also checks that the "consumed" buffer count we are
1459 * looking for matches the one contained in the subbuffer id.
1461 * The short-lived race window described here can be affected by
1462 * application signals and preemption, thus requiring to bound
1463 * the loop to a maximum number of retry.
1465 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1466 consumed_idx
, buf_trunc_val(consumed
, chan
),
1469 if (nr_retry
-- > 0) {
1470 if (nr_retry
<= (LTTNG_UST_RING_BUFFER_GET_RETRY
>> 1))
1471 (void) poll(NULL
, 0, LTTNG_UST_RING_BUFFER_RETRY_DELAY_MS
);
1477 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1479 buf
->get_subbuf_consumed
= consumed
;
1480 buf
->get_subbuf
= 1;
1486 * The memory barriers __wait_event()/wake_up_interruptible() take care
1487 * of "raw_spin_is_locked" memory ordering.
1496 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1499 void lib_ring_buffer_put_subbuf(struct lttng_ust_lib_ring_buffer
*buf
,
1500 struct lttng_ust_shm_handle
*handle
)
1502 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1503 struct channel
*chan
;
1504 const struct lttng_ust_lib_ring_buffer_config
*config
;
1505 unsigned long sb_bindex
, consumed_idx
, consumed
;
1506 struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*rpages
;
1507 struct lttng_ust_lib_ring_buffer_backend_pages
*backend_pages
;
1509 chan
= shmp(handle
, bufb
->chan
);
1512 config
= &chan
->backend
.config
;
1513 CHAN_WARN_ON(chan
, uatomic_read(&buf
->active_readers
) != 1);
1515 if (!buf
->get_subbuf
) {
1517 * Reader puts a subbuffer it did not get.
1519 CHAN_WARN_ON(chan
, 1);
1522 consumed
= buf
->get_subbuf_consumed
;
1523 buf
->get_subbuf
= 0;
1526 * Clear the records_unread counter. (overruns counter)
1527 * Can still be non-zero if a file reader simply grabbed the data
1528 * without using iterators.
1529 * Can be below zero if an iterator is used on a snapshot more than
1532 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1533 rpages
= shmp_index(handle
, bufb
->array
, sb_bindex
);
1536 backend_pages
= shmp(handle
, rpages
->shmp
);
1539 v_add(config
, v_read(config
, &backend_pages
->records_unread
),
1540 &bufb
->records_read
);
1541 v_set(config
, &backend_pages
->records_unread
, 0);
1542 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1543 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1544 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1547 * Exchange the reader subbuffer with the one we put in its place in the
1548 * writer subbuffer table. Expect the original consumed count. If
1549 * update_read_sb_index fails, this is because the writer updated the
1550 * subbuffer concurrently. We should therefore keep the subbuffer we
1551 * currently have: it has become invalid to try reading this sub-buffer
1552 * consumed count value anyway.
1554 consumed_idx
= subbuf_index(consumed
, chan
);
1555 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1556 consumed_idx
, buf_trunc_val(consumed
, chan
),
1559 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1560 * if the writer concurrently updated it.
1565 * cons_offset is an iterator on all subbuffer offsets between the reader
1566 * position and the writer position. (inclusive)
1569 void lib_ring_buffer_print_subbuffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
1570 struct channel
*chan
,
1571 unsigned long cons_offset
,
1573 struct lttng_ust_shm_handle
*handle
)
1575 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1576 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1577 struct commit_counters_hot
*cc_hot
;
1578 struct commit_counters_cold
*cc_cold
;
1580 cons_idx
= subbuf_index(cons_offset
, chan
);
1581 cc_hot
= shmp_index(handle
, buf
->commit_hot
, cons_idx
);
1584 cc_cold
= shmp_index(handle
, buf
->commit_cold
, cons_idx
);
1587 commit_count
= v_read(config
, &cc_hot
->cc
);
1588 commit_count_sb
= v_read(config
, &cc_cold
->cc_sb
);
1590 if (subbuf_offset(commit_count
, chan
) != 0)
1591 DBG("ring buffer %s, cpu %d: "
1592 "commit count in subbuffer %lu,\n"
1593 "expecting multiples of %lu bytes\n"
1594 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1595 chan
->backend
.name
, cpu
, cons_idx
,
1596 chan
->backend
.subbuf_size
,
1597 commit_count
, commit_count_sb
);
1599 DBG("ring buffer: %s, cpu %d: %lu bytes committed\n",
1600 chan
->backend
.name
, cpu
, commit_count
);
1604 void lib_ring_buffer_print_buffer_errors(struct lttng_ust_lib_ring_buffer
*buf
,
1605 struct channel
*chan
,
1606 void *priv
, int cpu
,
1607 struct lttng_ust_shm_handle
*handle
)
1609 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1610 unsigned long write_offset
, cons_offset
;
1613 * No need to order commit_count, write_offset and cons_offset reads
1614 * because we execute at teardown when no more writer nor reader
1615 * references are left.
1617 write_offset
= v_read(config
, &buf
->offset
);
1618 cons_offset
= uatomic_read(&buf
->consumed
);
1619 if (write_offset
!= cons_offset
)
1620 DBG("ring buffer %s, cpu %d: "
1621 "non-consumed data\n"
1622 " [ %lu bytes written, %lu bytes read ]\n",
1623 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1625 for (cons_offset
= uatomic_read(&buf
->consumed
);
1626 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1629 cons_offset
= subbuf_align(cons_offset
, chan
))
1630 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1635 void lib_ring_buffer_print_errors(struct channel
*chan
,
1636 struct lttng_ust_lib_ring_buffer
*buf
, int cpu
,
1637 struct lttng_ust_shm_handle
*handle
)
1639 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1640 void *priv
= channel_get_private(chan
);
1642 if (!strcmp(chan
->backend
.name
, "relay-metadata-mmap")) {
1643 DBG("ring buffer %s: %lu records written, "
1644 "%lu records overrun\n",
1646 v_read(config
, &buf
->records_count
),
1647 v_read(config
, &buf
->records_overrun
));
1649 DBG("ring buffer %s, cpu %d: %lu records written, "
1650 "%lu records overrun\n",
1651 chan
->backend
.name
, cpu
,
1652 v_read(config
, &buf
->records_count
),
1653 v_read(config
, &buf
->records_overrun
));
1655 if (v_read(config
, &buf
->records_lost_full
)
1656 || v_read(config
, &buf
->records_lost_wrap
)
1657 || v_read(config
, &buf
->records_lost_big
))
1658 DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n"
1659 " [ %lu buffer full, %lu nest buffer wrap-around, "
1660 "%lu event too big ]\n",
1661 chan
->backend
.name
, cpu
,
1662 v_read(config
, &buf
->records_lost_full
),
1663 v_read(config
, &buf
->records_lost_wrap
),
1664 v_read(config
, &buf
->records_lost_big
));
1666 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
, handle
);
1670 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1672 * Only executed by SWITCH_FLUSH, which can be issued while tracing is
1673 * active or at buffer finalization (destroy).
1676 void lib_ring_buffer_switch_old_start(struct lttng_ust_lib_ring_buffer
*buf
,
1677 struct channel
*chan
,
1678 struct switch_offsets
*offsets
,
1680 struct lttng_ust_shm_handle
*handle
)
1682 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1683 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1684 unsigned long commit_count
;
1685 struct commit_counters_hot
*cc_hot
;
1687 config
->cb
.buffer_begin(buf
, tsc
, oldidx
, handle
);
1690 * Order all writes to buffer before the commit count update that will
1691 * determine that the subbuffer is full.
1694 cc_hot
= shmp_index(handle
, buf
->commit_hot
, oldidx
);
1697 v_add(config
, config
->cb
.subbuffer_header_size(),
1699 commit_count
= v_read(config
, &cc_hot
->cc
);
1700 /* Check if the written buffer has to be delivered */
1701 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1702 commit_count
, oldidx
, handle
, tsc
);
1703 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1704 offsets
->old
+ config
->cb
.subbuffer_header_size(),
1705 commit_count
, handle
, cc_hot
);
1709 * lib_ring_buffer_switch_old_end: switch old subbuffer
1711 * Note : offset_old should never be 0 here. It is ok, because we never perform
1712 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1713 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1717 void lib_ring_buffer_switch_old_end(struct lttng_ust_lib_ring_buffer
*buf
,
1718 struct channel
*chan
,
1719 struct switch_offsets
*offsets
,
1721 struct lttng_ust_shm_handle
*handle
)
1723 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1724 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1725 unsigned long commit_count
, padding_size
, data_size
;
1726 struct commit_counters_hot
*cc_hot
;
1728 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1729 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1730 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
,
1734 * Order all writes to buffer before the commit count update that will
1735 * determine that the subbuffer is full.
1738 cc_hot
= shmp_index(handle
, buf
->commit_hot
, oldidx
);
1741 v_add(config
, padding_size
, &cc_hot
->cc
);
1742 commit_count
= v_read(config
, &cc_hot
->cc
);
1743 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1744 commit_count
, oldidx
, handle
, tsc
);
1745 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1746 offsets
->old
+ padding_size
, commit_count
, handle
,
1751 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1753 * This code can be executed unordered : writers may already have written to the
1754 * sub-buffer before this code gets executed, caution. The commit makes sure
1755 * that this code is executed before the deliver of this sub-buffer.
1758 void lib_ring_buffer_switch_new_start(struct lttng_ust_lib_ring_buffer
*buf
,
1759 struct channel
*chan
,
1760 struct switch_offsets
*offsets
,
1762 struct lttng_ust_shm_handle
*handle
)
1764 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1765 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1766 unsigned long commit_count
;
1767 struct commit_counters_hot
*cc_hot
;
1769 config
->cb
.buffer_begin(buf
, tsc
, beginidx
, handle
);
1772 * Order all writes to buffer before the commit count update that will
1773 * determine that the subbuffer is full.
1776 cc_hot
= shmp_index(handle
, buf
->commit_hot
, beginidx
);
1779 v_add(config
, config
->cb
.subbuffer_header_size(), &cc_hot
->cc
);
1780 commit_count
= v_read(config
, &cc_hot
->cc
);
1781 /* Check if the written buffer has to be delivered */
1782 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1783 commit_count
, beginidx
, handle
, tsc
);
1784 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1785 offsets
->begin
+ config
->cb
.subbuffer_header_size(),
1786 commit_count
, handle
, cc_hot
);
1790 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1792 * Calls subbuffer_set_data_size() to set the data size of the current
1793 * sub-buffer. We do not need to perform check_deliver nor commit here,
1794 * since this task will be done by the "commit" of the event for which
1795 * we are currently doing the space reservation.
1798 void lib_ring_buffer_switch_new_end(struct lttng_ust_lib_ring_buffer
*buf
,
1799 struct channel
*chan
,
1800 struct switch_offsets
*offsets
,
1802 struct lttng_ust_shm_handle
*handle
)
1804 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1805 unsigned long endidx
, data_size
;
1807 endidx
= subbuf_index(offsets
->end
- 1, chan
);
1808 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1809 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
,
1816 * !0 if execution must be aborted.
1819 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1820 struct lttng_ust_lib_ring_buffer
*buf
,
1821 struct channel
*chan
,
1822 struct switch_offsets
*offsets
,
1824 struct lttng_ust_shm_handle
*handle
)
1826 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1827 unsigned long off
, reserve_commit_diff
;
1829 offsets
->begin
= v_read(config
, &buf
->offset
);
1830 offsets
->old
= offsets
->begin
;
1831 offsets
->switch_old_start
= 0;
1832 off
= subbuf_offset(offsets
->begin
, chan
);
1834 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1837 * Ensure we flush the header of an empty subbuffer when doing the
1838 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1839 * total data gathering duration even if there were no records saved
1840 * after the last buffer switch.
1841 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1842 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1843 * subbuffer header as appropriate.
1844 * The next record that reserves space will be responsible for
1845 * populating the following subbuffer header. We choose not to populate
1846 * the next subbuffer header here because we want to be able to use
1847 * SWITCH_ACTIVE for periodical buffer flush, which must
1848 * guarantee that all the buffer content (records and header
1849 * timestamps) are visible to the reader. This is required for
1850 * quiescence guarantees for the fusion merge.
1852 if (mode
!= SWITCH_FLUSH
&& !off
)
1853 return -1; /* we do not have to switch : buffer is empty */
1855 if (caa_unlikely(off
== 0)) {
1856 unsigned long sb_index
, commit_count
;
1857 struct commit_counters_cold
*cc_cold
;
1860 * We are performing a SWITCH_FLUSH. There may be concurrent
1861 * writes into the buffer if e.g. invoked while performing a
1862 * snapshot on an active trace.
1864 * If the client does not save any header information
1865 * (sub-buffer header size == 0), don't switch empty subbuffer
1866 * on finalize, because it is invalid to deliver a completely
1869 if (!config
->cb
.subbuffer_header_size())
1872 /* Test new buffer integrity */
1873 sb_index
= subbuf_index(offsets
->begin
, chan
);
1874 cc_cold
= shmp_index(handle
, buf
->commit_cold
, sb_index
);
1877 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
1878 reserve_commit_diff
=
1879 (buf_trunc(offsets
->begin
, chan
)
1880 >> chan
->backend
.num_subbuf_order
)
1881 - (commit_count
& chan
->commit_count_mask
);
1882 if (caa_likely(reserve_commit_diff
== 0)) {
1883 /* Next subbuffer not being written to. */
1884 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1885 subbuf_trunc(offsets
->begin
, chan
)
1886 - subbuf_trunc((unsigned long)
1887 uatomic_read(&buf
->consumed
), chan
)
1888 >= chan
->backend
.buf_size
)) {
1890 * We do not overwrite non consumed buffers
1891 * and we are full : don't switch.
1896 * Next subbuffer not being written to, and we
1897 * are either in overwrite mode or the buffer is
1898 * not full. It's safe to write in this new
1904 * Next subbuffer reserve offset does not match the
1905 * commit offset. Don't perform switch in
1906 * producer-consumer and overwrite mode. Caused by
1907 * either a writer OOPS or too many nested writes over a
1908 * reserve/commit pair.
1914 * Need to write the subbuffer start header on finalize.
1916 offsets
->switch_old_start
= 1;
1918 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1919 /* Note: old points to the next subbuf at offset 0 */
1920 offsets
->end
= offsets
->begin
;
1925 * Force a sub-buffer switch. This operation is completely reentrant : can be
1926 * called while tracing is active with absolutely no lock held.
1928 * For RING_BUFFER_SYNC_PER_CPU ring buffers, as a v_cmpxchg is used for
1929 * some atomic operations, this function must be called from the CPU
1930 * which owns the buffer for a ACTIVE flush. However, for
1931 * RING_BUFFER_SYNC_GLOBAL ring buffers, this function can be called
1934 void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer
*buf
, enum switch_mode mode
,
1935 struct lttng_ust_shm_handle
*handle
)
1937 struct channel
*chan
;
1938 const struct lttng_ust_lib_ring_buffer_config
*config
;
1939 struct switch_offsets offsets
;
1940 unsigned long oldidx
;
1943 chan
= shmp(handle
, buf
->backend
.chan
);
1946 config
= &chan
->backend
.config
;
1951 * Perform retryable operations.
1954 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1956 return; /* Switch not needed */
1957 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1961 * Atomically update last_tsc. This update races against concurrent
1962 * atomic updates, but the race will always cause supplementary full TSC
1963 * records, never the opposite (missing a full TSC record when it would
1966 save_last_tsc(config
, buf
, tsc
);
1969 * Push the reader if necessary
1971 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1973 oldidx
= subbuf_index(offsets
.old
, chan
);
1974 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
, handle
);
1977 * May need to populate header start on SWITCH_FLUSH.
1979 if (offsets
.switch_old_start
) {
1980 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
, handle
);
1981 offsets
.old
+= config
->cb
.subbuffer_header_size();
1985 * Switch old subbuffer.
1987 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
, handle
);
1993 * -ENOSPC if event size is too large for packet.
1994 * -ENOBUFS if there is currently not enough space in buffer for the event.
1995 * -EIO if data cannot be written into the buffer for any other reason.
1998 int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer
*buf
,
1999 struct channel
*chan
,
2000 struct switch_offsets
*offsets
,
2001 struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
2003 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
2004 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
2005 unsigned long reserve_commit_diff
, offset_cmp
;
2008 offsets
->begin
= offset_cmp
= v_read(config
, &buf
->offset
);
2009 offsets
->old
= offsets
->begin
;
2010 offsets
->switch_new_start
= 0;
2011 offsets
->switch_new_end
= 0;
2012 offsets
->switch_old_end
= 0;
2013 offsets
->pre_header_padding
= 0;
2015 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
2016 if ((int64_t) ctx
->tsc
== -EIO
)
2019 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
2020 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
2022 if (caa_unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
2023 offsets
->switch_new_start
= 1; /* For offsets->begin */
2025 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
2027 &offsets
->pre_header_padding
,
2030 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
2033 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
) +
2034 offsets
->size
> chan
->backend
.subbuf_size
)) {
2035 offsets
->switch_old_end
= 1; /* For offsets->old */
2036 offsets
->switch_new_start
= 1; /* For offsets->begin */
2039 if (caa_unlikely(offsets
->switch_new_start
)) {
2040 unsigned long sb_index
, commit_count
;
2041 struct commit_counters_cold
*cc_cold
;
2044 * We are typically not filling the previous buffer completely.
2046 if (caa_likely(offsets
->switch_old_end
))
2047 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
2048 offsets
->begin
= offsets
->begin
2049 + config
->cb
.subbuffer_header_size();
2050 /* Test new buffer integrity */
2051 sb_index
= subbuf_index(offsets
->begin
, chan
);
2053 * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
2054 * lib_ring_buffer_check_deliver() has the matching
2055 * memory barriers required around commit_cold cc_sb
2056 * updates to ensure reserve and commit counter updates
2057 * are not seen reordered when updated by another CPU.
2060 cc_cold
= shmp_index(handle
, buf
->commit_cold
, sb_index
);
2063 commit_count
= v_read(config
, &cc_cold
->cc_sb
);
2064 /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
2066 if (caa_unlikely(offset_cmp
!= v_read(config
, &buf
->offset
))) {
2068 * The reserve counter have been concurrently updated
2069 * while we read the commit counter. This means the
2070 * commit counter we read might not match buf->offset
2071 * due to concurrent update. We therefore need to retry.
2075 reserve_commit_diff
=
2076 (buf_trunc(offsets
->begin
, chan
)
2077 >> chan
->backend
.num_subbuf_order
)
2078 - (commit_count
& chan
->commit_count_mask
);
2079 if (caa_likely(reserve_commit_diff
== 0)) {
2080 /* Next subbuffer not being written to. */
2081 if (caa_unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
2082 subbuf_trunc(offsets
->begin
, chan
)
2083 - subbuf_trunc((unsigned long)
2084 uatomic_read(&buf
->consumed
), chan
)
2085 >= chan
->backend
.buf_size
)) {
2086 unsigned long nr_lost
;
2089 * We do not overwrite non consumed buffers
2090 * and we are full : record is lost.
2092 nr_lost
= v_read(config
, &buf
->records_lost_full
);
2093 v_inc(config
, &buf
->records_lost_full
);
2094 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
2095 DBG("%lu or more records lost in (%s:%d) (buffer full)\n",
2096 nr_lost
+ 1, chan
->backend
.name
,
2102 * Next subbuffer not being written to, and we
2103 * are either in overwrite mode or the buffer is
2104 * not full. It's safe to write in this new
2109 unsigned long nr_lost
;
2112 * Next subbuffer reserve offset does not match the
2113 * commit offset, and this did not involve update to the
2114 * reserve counter. Drop record in producer-consumer and
2115 * overwrite mode. Caused by either a writer OOPS or too
2116 * many nested writes over a reserve/commit pair.
2118 nr_lost
= v_read(config
, &buf
->records_lost_wrap
);
2119 v_inc(config
, &buf
->records_lost_wrap
);
2120 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
2121 DBG("%lu or more records lost in (%s:%d) (wrap-around)\n",
2122 nr_lost
+ 1, chan
->backend
.name
,
2128 config
->cb
.record_header_size(config
, chan
,
2130 &offsets
->pre_header_padding
,
2133 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
2136 if (caa_unlikely(subbuf_offset(offsets
->begin
, chan
)
2137 + offsets
->size
> chan
->backend
.subbuf_size
)) {
2138 unsigned long nr_lost
;
2141 * Record too big for subbuffers, report error, don't
2142 * complete the sub-buffer switch.
2144 nr_lost
= v_read(config
, &buf
->records_lost_big
);
2145 v_inc(config
, &buf
->records_lost_big
);
2146 if ((nr_lost
& (DBG_PRINT_NR_LOST
- 1)) == 0) {
2147 DBG("%lu or more records lost in (%s:%d) record size "
2148 " of %zu bytes is too large for buffer\n",
2149 nr_lost
+ 1, chan
->backend
.name
,
2150 buf
->backend
.cpu
, offsets
->size
);
2155 * We just made a successful buffer switch and the
2156 * record fits in the new subbuffer. Let's write.
2161 * Record fits in the current buffer and we are not on a switch
2162 * boundary. It's safe to write.
2165 offsets
->end
= offsets
->begin
+ offsets
->size
;
2167 if (caa_unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
2169 * The offset_end will fall at the very beginning of the next
2172 offsets
->switch_new_end
= 1; /* For offsets->begin */
2178 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
2179 * @ctx: ring buffer context.
2181 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
2182 * -EIO for other errors, else returns 0.
2183 * It will take care of sub-buffer switching.
2185 int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
2187 struct channel
*chan
= ctx
->chan
;
2188 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
2189 const struct lttng_ust_lib_ring_buffer_config
*config
= &chan
->backend
.config
;
2190 struct lttng_ust_lib_ring_buffer
*buf
;
2191 struct switch_offsets offsets
;
2194 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
2195 buf
= shmp(handle
, chan
->backend
.buf
[ctx
->cpu
].shmp
);
2197 buf
= shmp(handle
, chan
->backend
.buf
[0].shmp
);
2205 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
2207 if (caa_unlikely(ret
))
2209 } while (caa_unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
2214 * Atomically update last_tsc. This update races against concurrent
2215 * atomic updates, but the race will always cause supplementary full TSC
2216 * records, never the opposite (missing a full TSC record when it would
2219 save_last_tsc(config
, buf
, ctx
->tsc
);
2222 * Push the reader if necessary
2224 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
2227 * Clear noref flag for this subbuffer.
2229 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2230 subbuf_index(offsets
.end
- 1, chan
),
2234 * Switch old subbuffer if needed.
2236 if (caa_unlikely(offsets
.switch_old_end
)) {
2237 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2238 subbuf_index(offsets
.old
- 1, chan
),
2240 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2244 * Populate new subbuffer.
2246 if (caa_unlikely(offsets
.switch_new_start
))
2247 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2249 if (caa_unlikely(offsets
.switch_new_end
))
2250 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
, handle
);
2252 ctx
->slot_size
= offsets
.size
;
2253 ctx
->pre_offset
= offsets
.begin
;
2254 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
2259 void lib_ring_buffer_vmcore_check_deliver(const struct lttng_ust_lib_ring_buffer_config
*config
,
2260 struct lttng_ust_lib_ring_buffer
*buf
,
2261 unsigned long commit_count
,
2263 struct lttng_ust_shm_handle
*handle
)
2265 struct commit_counters_hot
*cc_hot
;
2267 if (config
->oops
!= RING_BUFFER_OOPS_CONSISTENCY
)
2269 cc_hot
= shmp_index(handle
, buf
->commit_hot
, idx
);
2272 v_set(config
, &cc_hot
->seq
, commit_count
);
2276 * The ring buffer can count events recorded and overwritten per buffer,
2277 * but it is disabled by default due to its performance overhead.
2279 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
2281 void deliver_count_events(const struct lttng_ust_lib_ring_buffer_config
*config
,
2282 struct lttng_ust_lib_ring_buffer
*buf
,
2284 struct lttng_ust_shm_handle
*handle
)
2286 v_add(config
, subbuffer_get_records_count(config
,
2287 &buf
->backend
, idx
, handle
),
2288 &buf
->records_count
);
2289 v_add(config
, subbuffer_count_records_overrun(config
,
2290 &buf
->backend
, idx
, handle
),
2291 &buf
->records_overrun
);
2293 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
2295 void deliver_count_events(const struct lttng_ust_lib_ring_buffer_config
*config
,
2296 struct lttng_ust_lib_ring_buffer
*buf
,
2298 struct lttng_ust_shm_handle
*handle
)
2301 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
2303 void lib_ring_buffer_check_deliver_slow(const struct lttng_ust_lib_ring_buffer_config
*config
,
2304 struct lttng_ust_lib_ring_buffer
*buf
,
2305 struct channel
*chan
,
2306 unsigned long offset
,
2307 unsigned long commit_count
,
2309 struct lttng_ust_shm_handle
*handle
,
2312 unsigned long old_commit_count
= commit_count
2313 - chan
->backend
.subbuf_size
;
2314 struct commit_counters_cold
*cc_cold
;
2317 * If we succeeded at updating cc_sb below, we are the subbuffer
2318 * writer delivering the subbuffer. Deals with concurrent
2319 * updates of the "cc" value without adding a add_return atomic
2320 * operation to the fast path.
2322 * We are doing the delivery in two steps:
2323 * - First, we cmpxchg() cc_sb to the new value
2324 * old_commit_count + 1. This ensures that we are the only
2325 * subbuffer user successfully filling the subbuffer, but we
2326 * do _not_ set the cc_sb value to "commit_count" yet.
2327 * Therefore, other writers that would wrap around the ring
2328 * buffer and try to start writing to our subbuffer would
2329 * have to drop records, because it would appear as
2331 * We therefore have exclusive access to the subbuffer control
2332 * structures. This mutual exclusion with other writers is
2333 * crucially important to perform record overruns count in
2334 * flight recorder mode locklessly.
2335 * - When we are ready to release the subbuffer (either for
2336 * reading or for overrun by other writers), we simply set the
2337 * cc_sb value to "commit_count" and perform delivery.
2339 * The subbuffer size is least 2 bytes (minimum size: 1 page).
2340 * This guarantees that old_commit_count + 1 != commit_count.
2344 * Order prior updates to reserve count prior to the
2345 * commit_cold cc_sb update.
2348 cc_cold
= shmp_index(handle
, buf
->commit_cold
, idx
);
2351 if (caa_likely(v_cmpxchg(config
, &cc_cold
->cc_sb
,
2352 old_commit_count
, old_commit_count
+ 1)
2353 == old_commit_count
)) {
2355 * Start of exclusive subbuffer access. We are
2356 * guaranteed to be the last writer in this subbuffer
2357 * and any other writer trying to access this subbuffer
2358 * in this state is required to drop records.
2360 deliver_count_events(config
, buf
, idx
, handle
);
2361 config
->cb
.buffer_end(buf
, tsc
, idx
,
2362 lib_ring_buffer_get_data_size(config
,
2369 * Increment the packet counter while we have exclusive
2372 subbuffer_inc_packet_count(config
, &buf
->backend
, idx
, handle
);
2375 * Set noref flag and offset for this subbuffer id.
2376 * Contains a memory barrier that ensures counter stores
2377 * are ordered before set noref and offset.
2379 lib_ring_buffer_set_noref_offset(config
, &buf
->backend
, idx
,
2380 buf_trunc_val(offset
, chan
), handle
);
2383 * Order set_noref and record counter updates before the
2384 * end of subbuffer exclusive access. Orders with
2385 * respect to writers coming into the subbuffer after
2386 * wrap around, and also order wrt concurrent readers.
2389 /* End of exclusive subbuffer access */
2390 v_set(config
, &cc_cold
->cc_sb
, commit_count
);
2392 * Order later updates to reserve count after
2393 * the commit cold cc_sb update.
2396 lib_ring_buffer_vmcore_check_deliver(config
, buf
,
2397 commit_count
, idx
, handle
);
2400 * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
2402 if (config
->wakeup
== RING_BUFFER_WAKEUP_BY_WRITER
2403 && uatomic_read(&buf
->active_readers
)
2404 && lib_ring_buffer_poll_deliver(config
, buf
, chan
, handle
)) {
2405 lib_ring_buffer_wakeup(buf
, handle
);
2411 * Force a read (imply TLS fixup for dlopen) of TLS variables.
2413 void lttng_fixup_ringbuffer_tls(void)
2415 asm volatile ("" : : "m" (URCU_TLS(lib_ring_buffer_nesting
)));
2418 void lib_ringbuffer_signal_init(void)
2424 * Block signal for entire process, so only our thread processes
2428 ret
= pthread_sigmask(SIG_BLOCK
, &mask
, NULL
);
2431 PERROR("pthread_sigmask");