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
54 #include <linux/delay.h>
55 #include <linux/module.h>
56 #include <linux/percpu.h>
57 #include <asm/cacheflush.h>
59 #include <wrapper/ringbuffer/config.h>
60 #include <wrapper/ringbuffer/backend.h>
61 #include <wrapper/ringbuffer/frontend.h>
62 #include <wrapper/ringbuffer/iterator.h>
63 #include <wrapper/ringbuffer/nohz.h>
64 #include <wrapper/atomic.h>
65 #include <wrapper/kref.h>
66 #include <wrapper/percpu-defs.h>
67 #include <wrapper/timer.h>
68 #include <wrapper/vmalloc.h>
71 * Internal structure representing offsets to use at a sub-buffer switch.
73 struct switch_offsets
{
74 unsigned long begin
, end
, old
;
75 size_t pre_header_padding
, size
;
76 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
87 static ATOMIC_NOTIFIER_HEAD(tick_nohz_notifier
);
88 #endif /* CONFIG_NO_HZ */
90 static DEFINE_PER_CPU(spinlock_t
, ring_buffer_nohz_lock
);
92 DEFINE_PER_CPU(unsigned int, lib_ring_buffer_nesting
);
93 EXPORT_PER_CPU_SYMBOL(lib_ring_buffer_nesting
);
96 void lib_ring_buffer_print_errors(struct channel
*chan
,
97 struct lib_ring_buffer
*buf
, int cpu
);
99 void _lib_ring_buffer_switch_remote(struct lib_ring_buffer
*buf
,
100 enum switch_mode mode
);
103 int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config
*config
,
104 struct lib_ring_buffer
*buf
,
105 struct channel
*chan
)
107 unsigned long consumed_old
, consumed_idx
, commit_count
, write_offset
;
109 consumed_old
= atomic_long_read(&buf
->consumed
);
110 consumed_idx
= subbuf_index(consumed_old
, chan
);
111 commit_count
= v_read(config
, &buf
->commit_cold
[consumed_idx
].cc_sb
);
113 * No memory barrier here, since we are only interested
114 * in a statistically correct polling result. The next poll will
115 * get the data is we are racing. The mb() that ensures correct
116 * memory order is in get_subbuf.
118 write_offset
= v_read(config
, &buf
->offset
);
121 * Check that the subbuffer we are trying to consume has been
122 * already fully committed.
125 if (((commit_count
- chan
->backend
.subbuf_size
)
126 & chan
->commit_count_mask
)
127 - (buf_trunc(consumed_old
, chan
)
128 >> chan
->backend
.num_subbuf_order
)
133 * Check that we are not about to read the same subbuffer in
134 * which the writer head is.
136 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_old
, chan
)
144 * Must be called under cpu hotplug protection.
146 void lib_ring_buffer_free(struct lib_ring_buffer
*buf
)
148 struct channel
*chan
= buf
->backend
.chan
;
150 lib_ring_buffer_print_errors(chan
, buf
, buf
->backend
.cpu
);
151 lttng_kvfree(buf
->commit_hot
);
152 lttng_kvfree(buf
->commit_cold
);
154 lib_ring_buffer_backend_free(&buf
->backend
);
158 * lib_ring_buffer_reset - Reset ring buffer to initial values.
161 * Effectively empty the ring buffer. Should be called when the buffer is not
162 * used for writing. The ring buffer can be opened for reading, but the reader
163 * should not be using the iterator concurrently with reset. The previous
164 * current iterator record is reset.
166 void lib_ring_buffer_reset(struct lib_ring_buffer
*buf
)
168 struct channel
*chan
= buf
->backend
.chan
;
169 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
173 * Reset iterator first. It will put the subbuffer if it currently holds
176 lib_ring_buffer_iterator_reset(buf
);
177 v_set(config
, &buf
->offset
, 0);
178 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
179 v_set(config
, &buf
->commit_hot
[i
].cc
, 0);
180 v_set(config
, &buf
->commit_hot
[i
].seq
, 0);
181 v_set(config
, &buf
->commit_cold
[i
].cc_sb
, 0);
183 atomic_long_set(&buf
->consumed
, 0);
184 atomic_set(&buf
->record_disabled
, 0);
185 v_set(config
, &buf
->last_tsc
, 0);
186 lib_ring_buffer_backend_reset(&buf
->backend
);
187 /* Don't reset number of active readers */
188 v_set(config
, &buf
->records_lost_full
, 0);
189 v_set(config
, &buf
->records_lost_wrap
, 0);
190 v_set(config
, &buf
->records_lost_big
, 0);
191 v_set(config
, &buf
->records_count
, 0);
192 v_set(config
, &buf
->records_overrun
, 0);
195 EXPORT_SYMBOL_GPL(lib_ring_buffer_reset
);
198 * channel_reset - Reset channel to initial values.
201 * Effectively empty the channel. Should be called when the channel is not used
202 * for writing. The channel can be opened for reading, but the reader should not
203 * be using the iterator concurrently with reset. The previous current iterator
206 void channel_reset(struct channel
*chan
)
209 * Reset iterators first. Will put the subbuffer if held for reading.
211 channel_iterator_reset(chan
);
212 atomic_set(&chan
->record_disabled
, 0);
213 /* Don't reset commit_count_mask, still valid */
214 channel_backend_reset(&chan
->backend
);
215 /* Don't reset switch/read timer interval */
216 /* Don't reset notifiers and notifier enable bits */
217 /* Don't reset reader reference count */
219 EXPORT_SYMBOL_GPL(channel_reset
);
222 * Must be called under cpu hotplug protection.
224 int lib_ring_buffer_create(struct lib_ring_buffer
*buf
,
225 struct channel_backend
*chanb
, int cpu
)
227 const struct lib_ring_buffer_config
*config
= &chanb
->config
;
228 struct channel
*chan
= container_of(chanb
, struct channel
, backend
);
229 void *priv
= chanb
->priv
;
230 size_t subbuf_header_size
;
234 /* Test for cpu hotplug */
235 if (buf
->backend
.allocated
)
239 * Paranoia: per cpu dynamic allocation is not officially documented as
240 * zeroing the memory, so let's do it here too, just in case.
242 memset(buf
, 0, sizeof(*buf
));
244 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
, cpu
);
249 lttng_kvzalloc_node(ALIGN(sizeof(*buf
->commit_hot
)
250 * chan
->backend
.num_subbuf
,
251 1 << INTERNODE_CACHE_SHIFT
),
252 GFP_KERNEL
| __GFP_NOWARN
,
253 cpu_to_node(max(cpu
, 0)));
254 if (!buf
->commit_hot
) {
260 lttng_kvzalloc_node(ALIGN(sizeof(*buf
->commit_cold
)
261 * chan
->backend
.num_subbuf
,
262 1 << INTERNODE_CACHE_SHIFT
),
263 GFP_KERNEL
| __GFP_NOWARN
,
264 cpu_to_node(max(cpu
, 0)));
265 if (!buf
->commit_cold
) {
270 init_waitqueue_head(&buf
->read_wait
);
271 init_waitqueue_head(&buf
->write_wait
);
272 raw_spin_lock_init(&buf
->raw_tick_nohz_spinlock
);
275 * Write the subbuffer header for first subbuffer so we know the total
276 * duration of data gathering.
278 subbuf_header_size
= config
->cb
.subbuffer_header_size();
279 v_set(config
, &buf
->offset
, subbuf_header_size
);
280 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_wsb
[0].id
);
281 tsc
= config
->cb
.ring_buffer_clock_read(buf
->backend
.chan
);
282 config
->cb
.buffer_begin(buf
, tsc
, 0);
283 v_add(config
, subbuf_header_size
, &buf
->commit_hot
[0].cc
);
285 if (config
->cb
.buffer_create
) {
286 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
);
292 * Ensure the buffer is ready before setting it to allocated and setting
294 * Used for cpu hotplug vs cpumask iteration.
297 buf
->backend
.allocated
= 1;
299 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
300 CHAN_WARN_ON(chan
, cpumask_test_cpu(cpu
,
301 chan
->backend
.cpumask
));
302 cpumask_set_cpu(cpu
, chan
->backend
.cpumask
);
309 lttng_kvfree(buf
->commit_cold
);
311 lttng_kvfree(buf
->commit_hot
);
313 lib_ring_buffer_backend_free(&buf
->backend
);
317 static void switch_buffer_timer(unsigned long data
)
319 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
320 struct channel
*chan
= buf
->backend
.chan
;
321 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
324 * Only flush buffers periodically if readers are active.
326 if (atomic_long_read(&buf
->active_readers
))
327 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
329 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
330 lttng_mod_timer_pinned(&buf
->switch_timer
,
331 jiffies
+ chan
->switch_timer_interval
);
333 mod_timer(&buf
->switch_timer
,
334 jiffies
+ chan
->switch_timer_interval
);
338 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
340 static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer
*buf
)
342 struct channel
*chan
= buf
->backend
.chan
;
343 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
345 if (!chan
->switch_timer_interval
|| buf
->switch_timer_enabled
)
348 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
349 lttng_init_timer_pinned(&buf
->switch_timer
);
351 init_timer(&buf
->switch_timer
);
353 buf
->switch_timer
.function
= switch_buffer_timer
;
354 buf
->switch_timer
.expires
= jiffies
+ chan
->switch_timer_interval
;
355 buf
->switch_timer
.data
= (unsigned long)buf
;
356 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
357 add_timer_on(&buf
->switch_timer
, buf
->backend
.cpu
);
359 add_timer(&buf
->switch_timer
);
360 buf
->switch_timer_enabled
= 1;
364 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
366 static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer
*buf
)
368 struct channel
*chan
= buf
->backend
.chan
;
370 if (!chan
->switch_timer_interval
|| !buf
->switch_timer_enabled
)
373 del_timer_sync(&buf
->switch_timer
);
374 buf
->switch_timer_enabled
= 0;
378 * Polling timer to check the channels for data.
380 static void read_buffer_timer(unsigned long data
)
382 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
383 struct channel
*chan
= buf
->backend
.chan
;
384 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
386 CHAN_WARN_ON(chan
, !buf
->backend
.allocated
);
388 if (atomic_long_read(&buf
->active_readers
)
389 && lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
390 wake_up_interruptible(&buf
->read_wait
);
391 wake_up_interruptible(&chan
->read_wait
);
394 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
395 lttng_mod_timer_pinned(&buf
->read_timer
,
396 jiffies
+ chan
->read_timer_interval
);
398 mod_timer(&buf
->read_timer
,
399 jiffies
+ chan
->read_timer_interval
);
403 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
405 static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer
*buf
)
407 struct channel
*chan
= buf
->backend
.chan
;
408 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
410 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
411 || !chan
->read_timer_interval
412 || buf
->read_timer_enabled
)
415 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
416 lttng_init_timer_pinned(&buf
->read_timer
);
418 init_timer(&buf
->read_timer
);
420 buf
->read_timer
.function
= read_buffer_timer
;
421 buf
->read_timer
.expires
= jiffies
+ chan
->read_timer_interval
;
422 buf
->read_timer
.data
= (unsigned long)buf
;
424 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
425 add_timer_on(&buf
->read_timer
, buf
->backend
.cpu
);
427 add_timer(&buf
->read_timer
);
428 buf
->read_timer_enabled
= 1;
432 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
434 static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer
*buf
)
436 struct channel
*chan
= buf
->backend
.chan
;
437 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
439 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
440 || !chan
->read_timer_interval
441 || !buf
->read_timer_enabled
)
444 del_timer_sync(&buf
->read_timer
);
446 * do one more check to catch data that has been written in the last
449 if (lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
450 wake_up_interruptible(&buf
->read_wait
);
451 wake_up_interruptible(&chan
->read_wait
);
453 buf
->read_timer_enabled
= 0;
456 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0))
458 enum cpuhp_state lttng_rb_hp_prepare
;
459 enum cpuhp_state lttng_rb_hp_online
;
461 void lttng_rb_set_hp_prepare(enum cpuhp_state val
)
463 lttng_rb_hp_prepare
= val
;
465 EXPORT_SYMBOL_GPL(lttng_rb_set_hp_prepare
);
467 void lttng_rb_set_hp_online(enum cpuhp_state val
)
469 lttng_rb_hp_online
= val
;
471 EXPORT_SYMBOL_GPL(lttng_rb_set_hp_online
);
473 int lttng_cpuhp_rb_frontend_dead(unsigned int cpu
,
474 struct lttng_cpuhp_node
*node
)
476 struct channel
*chan
= container_of(node
, struct channel
,
478 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
, cpu
);
479 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
481 CHAN_WARN_ON(chan
, config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
);
484 * Performing a buffer switch on a remote CPU. Performed by
485 * the CPU responsible for doing the hotunplug after the target
486 * CPU stopped running completely. Ensures that all data
487 * from that remote CPU is flushed.
489 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
492 EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_dead
);
494 int lttng_cpuhp_rb_frontend_online(unsigned int cpu
,
495 struct lttng_cpuhp_node
*node
)
497 struct channel
*chan
= container_of(node
, struct channel
,
499 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
, cpu
);
500 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
502 CHAN_WARN_ON(chan
, config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
);
504 wake_up_interruptible(&chan
->hp_wait
);
505 lib_ring_buffer_start_switch_timer(buf
);
506 lib_ring_buffer_start_read_timer(buf
);
509 EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_online
);
511 int lttng_cpuhp_rb_frontend_offline(unsigned int cpu
,
512 struct lttng_cpuhp_node
*node
)
514 struct channel
*chan
= container_of(node
, struct channel
,
516 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
, cpu
);
517 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
519 CHAN_WARN_ON(chan
, config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
);
521 lib_ring_buffer_stop_switch_timer(buf
);
522 lib_ring_buffer_stop_read_timer(buf
);
525 EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_offline
);
527 #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */
529 #ifdef CONFIG_HOTPLUG_CPU
532 * lib_ring_buffer_cpu_hp_callback - CPU hotplug callback
533 * @nb: notifier block
534 * @action: hotplug action to take
537 * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
540 int lib_ring_buffer_cpu_hp_callback(struct notifier_block
*nb
,
541 unsigned long action
,
544 unsigned int cpu
= (unsigned long)hcpu
;
545 struct channel
*chan
= container_of(nb
, struct channel
,
547 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
, cpu
);
548 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
550 if (!chan
->cpu_hp_enable
)
553 CHAN_WARN_ON(chan
, config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
);
556 case CPU_DOWN_FAILED
:
557 case CPU_DOWN_FAILED_FROZEN
:
559 case CPU_ONLINE_FROZEN
:
560 wake_up_interruptible(&chan
->hp_wait
);
561 lib_ring_buffer_start_switch_timer(buf
);
562 lib_ring_buffer_start_read_timer(buf
);
565 case CPU_DOWN_PREPARE
:
566 case CPU_DOWN_PREPARE_FROZEN
:
567 lib_ring_buffer_stop_switch_timer(buf
);
568 lib_ring_buffer_stop_read_timer(buf
);
572 case CPU_DEAD_FROZEN
:
574 * Performing a buffer switch on a remote CPU. Performed by
575 * the CPU responsible for doing the hotunplug after the target
576 * CPU stopped running completely. Ensures that all data
577 * from that remote CPU is flushed.
579 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
589 #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */
591 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
593 * For per-cpu buffers, call the reader wakeups before switching the buffer, so
594 * that wake-up-tracing generated events are flushed before going idle (in
595 * tick_nohz). We test if the spinlock is locked to deal with the race where
596 * readers try to sample the ring buffer before we perform the switch. We let
597 * the readers retry in that case. If there is data in the buffer, the wake up
598 * is going to forbid the CPU running the reader thread from going idle.
600 static int notrace
ring_buffer_tick_nohz_callback(struct notifier_block
*nb
,
604 struct channel
*chan
= container_of(nb
, struct channel
,
606 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
607 struct lib_ring_buffer
*buf
;
608 int cpu
= smp_processor_id();
610 if (config
->alloc
!= RING_BUFFER_ALLOC_PER_CPU
) {
612 * We don't support keeping the system idle with global buffers
613 * and streaming active. In order to do so, we would need to
614 * sample a non-nohz-cpumask racelessly with the nohz updates
615 * without adding synchronization overhead to nohz. Leave this
616 * use-case out for now.
621 buf
= channel_get_ring_buffer(config
, chan
, cpu
);
623 case TICK_NOHZ_FLUSH
:
624 raw_spin_lock(&buf
->raw_tick_nohz_spinlock
);
625 if (config
->wakeup
== RING_BUFFER_WAKEUP_BY_TIMER
626 && chan
->read_timer_interval
627 && atomic_long_read(&buf
->active_readers
)
628 && (lib_ring_buffer_poll_deliver(config
, buf
, chan
)
629 || lib_ring_buffer_pending_data(config
, buf
, chan
))) {
630 wake_up_interruptible(&buf
->read_wait
);
631 wake_up_interruptible(&chan
->read_wait
);
633 if (chan
->switch_timer_interval
)
634 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
635 raw_spin_unlock(&buf
->raw_tick_nohz_spinlock
);
638 spin_lock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
639 lib_ring_buffer_stop_switch_timer(buf
);
640 lib_ring_buffer_stop_read_timer(buf
);
641 spin_unlock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
643 case TICK_NOHZ_RESTART
:
644 spin_lock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
645 lib_ring_buffer_start_read_timer(buf
);
646 lib_ring_buffer_start_switch_timer(buf
);
647 spin_unlock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
654 void notrace
lib_ring_buffer_tick_nohz_flush(void)
656 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_FLUSH
,
660 void notrace
lib_ring_buffer_tick_nohz_stop(void)
662 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_STOP
,
666 void notrace
lib_ring_buffer_tick_nohz_restart(void)
668 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_RESTART
,
671 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
676 static void channel_unregister_notifiers(struct channel
*chan
)
678 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
680 channel_iterator_unregister_notifiers(chan
);
681 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
684 * Remove the nohz notifier first, so we are certain we stop
687 atomic_notifier_chain_unregister(&tick_nohz_notifier
,
688 &chan
->tick_nohz_notifier
);
690 * ring_buffer_nohz_lock will not be needed below, because
691 * we just removed the notifiers, which were the only source of
694 #endif /* CONFIG_NO_HZ */
695 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0))
699 ret
= cpuhp_state_remove_instance(lttng_rb_hp_online
,
700 &chan
->cpuhp_online
.node
);
702 ret
= cpuhp_state_remove_instance_nocalls(lttng_rb_hp_prepare
,
703 &chan
->cpuhp_prepare
.node
);
706 #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */
710 #ifdef CONFIG_HOTPLUG_CPU
712 chan
->cpu_hp_enable
= 0;
713 for_each_online_cpu(cpu
) {
714 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
716 lib_ring_buffer_stop_switch_timer(buf
);
717 lib_ring_buffer_stop_read_timer(buf
);
720 unregister_cpu_notifier(&chan
->cpu_hp_notifier
);
722 for_each_possible_cpu(cpu
) {
723 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
725 lib_ring_buffer_stop_switch_timer(buf
);
726 lib_ring_buffer_stop_read_timer(buf
);
730 #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */
732 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
734 lib_ring_buffer_stop_switch_timer(buf
);
735 lib_ring_buffer_stop_read_timer(buf
);
737 channel_backend_unregister_notifiers(&chan
->backend
);
740 static void lib_ring_buffer_set_quiescent(struct lib_ring_buffer
*buf
)
742 if (!buf
->quiescent
) {
743 buf
->quiescent
= true;
744 _lib_ring_buffer_switch_remote(buf
, SWITCH_FLUSH
);
748 static void lib_ring_buffer_clear_quiescent(struct lib_ring_buffer
*buf
)
750 buf
->quiescent
= false;
753 void lib_ring_buffer_set_quiescent_channel(struct channel
*chan
)
756 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
758 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
760 for_each_channel_cpu(cpu
, chan
) {
761 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
764 lib_ring_buffer_set_quiescent(buf
);
768 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
770 lib_ring_buffer_set_quiescent(buf
);
773 EXPORT_SYMBOL_GPL(lib_ring_buffer_set_quiescent_channel
);
775 void lib_ring_buffer_clear_quiescent_channel(struct channel
*chan
)
778 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
780 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
782 for_each_channel_cpu(cpu
, chan
) {
783 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
786 lib_ring_buffer_clear_quiescent(buf
);
790 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
792 lib_ring_buffer_clear_quiescent(buf
);
795 EXPORT_SYMBOL_GPL(lib_ring_buffer_clear_quiescent_channel
);
797 static void channel_free(struct channel
*chan
)
799 if (chan
->backend
.release_priv_ops
) {
800 chan
->backend
.release_priv_ops(chan
->backend
.priv_ops
);
802 channel_iterator_free(chan
);
803 channel_backend_free(&chan
->backend
);
808 * channel_create - Create channel.
809 * @config: ring buffer instance configuration
810 * @name: name of the channel
811 * @priv: ring buffer client private data
812 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
813 * address mapping. It is used only by RING_BUFFER_STATIC
814 * configuration. It can be set to NULL for other backends.
815 * @subbuf_size: subbuffer size
816 * @num_subbuf: number of subbuffers
817 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
818 * padding to let readers get those sub-buffers.
819 * Used for live streaming.
820 * @read_timer_interval: Time interval (in us) to wake up pending readers.
823 * Returns NULL on failure.
825 struct channel
*channel_create(const struct lib_ring_buffer_config
*config
,
826 const char *name
, void *priv
, void *buf_addr
,
828 size_t num_subbuf
, unsigned int switch_timer_interval
,
829 unsigned int read_timer_interval
)
832 struct channel
*chan
;
834 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
835 read_timer_interval
))
838 chan
= kzalloc(sizeof(struct channel
), GFP_KERNEL
);
842 ret
= channel_backend_init(&chan
->backend
, name
, config
, priv
,
843 subbuf_size
, num_subbuf
);
847 ret
= channel_iterator_init(chan
);
849 goto error_free_backend
;
851 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
852 chan
->switch_timer_interval
= usecs_to_jiffies(switch_timer_interval
);
853 chan
->read_timer_interval
= usecs_to_jiffies(read_timer_interval
);
854 kref_init(&chan
->ref
);
855 init_waitqueue_head(&chan
->read_wait
);
856 init_waitqueue_head(&chan
->hp_wait
);
858 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
859 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0))
860 chan
->cpuhp_prepare
.component
= LTTNG_RING_BUFFER_FRONTEND
;
861 ret
= cpuhp_state_add_instance_nocalls(lttng_rb_hp_prepare
,
862 &chan
->cpuhp_prepare
.node
);
864 goto cpuhp_prepare_error
;
866 chan
->cpuhp_online
.component
= LTTNG_RING_BUFFER_FRONTEND
;
867 ret
= cpuhp_state_add_instance(lttng_rb_hp_online
,
868 &chan
->cpuhp_online
.node
);
870 goto cpuhp_online_error
;
871 #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */
875 * In case of non-hotplug cpu, if the ring-buffer is allocated
876 * in early initcall, it will not be notified of secondary cpus.
877 * In that off case, we need to allocate for all possible cpus.
879 #ifdef CONFIG_HOTPLUG_CPU
880 chan
->cpu_hp_notifier
.notifier_call
=
881 lib_ring_buffer_cpu_hp_callback
;
882 chan
->cpu_hp_notifier
.priority
= 6;
883 register_cpu_notifier(&chan
->cpu_hp_notifier
);
886 for_each_online_cpu(cpu
) {
887 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
889 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
890 lib_ring_buffer_start_switch_timer(buf
);
891 lib_ring_buffer_start_read_timer(buf
);
892 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
894 chan
->cpu_hp_enable
= 1;
897 for_each_possible_cpu(cpu
) {
898 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
900 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
901 lib_ring_buffer_start_switch_timer(buf
);
902 lib_ring_buffer_start_read_timer(buf
);
903 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
907 #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */
909 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
910 /* Only benefit from NO_HZ idle with per-cpu buffers for now. */
911 chan
->tick_nohz_notifier
.notifier_call
=
912 ring_buffer_tick_nohz_callback
;
913 chan
->tick_nohz_notifier
.priority
= ~0U;
914 atomic_notifier_chain_register(&tick_nohz_notifier
,
915 &chan
->tick_nohz_notifier
);
916 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
919 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
921 lib_ring_buffer_start_switch_timer(buf
);
922 lib_ring_buffer_start_read_timer(buf
);
927 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0))
929 ret
= cpuhp_state_remove_instance_nocalls(lttng_rb_hp_prepare
,
930 &chan
->cpuhp_prepare
.node
);
933 #endif /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */
935 channel_backend_free(&chan
->backend
);
940 EXPORT_SYMBOL_GPL(channel_create
);
943 void channel_release(struct kref
*kref
)
945 struct channel
*chan
= container_of(kref
, struct channel
, ref
);
950 * channel_destroy - Finalize, wait for q.s. and destroy channel.
951 * @chan: channel to destroy
954 * Call "destroy" callback, finalize channels, and then decrement the
955 * channel reference count. Note that when readers have completed data
956 * consumption of finalized channels, get_subbuf() will return -ENODATA.
957 * They should release their handle at that point. Returns the private
960 void *channel_destroy(struct channel
*chan
)
963 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
966 channel_unregister_notifiers(chan
);
968 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
970 * No need to hold cpu hotplug, because all notifiers have been
973 for_each_channel_cpu(cpu
, chan
) {
974 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
977 if (config
->cb
.buffer_finalize
)
978 config
->cb
.buffer_finalize(buf
,
981 if (buf
->backend
.allocated
)
982 lib_ring_buffer_set_quiescent(buf
);
984 * Perform flush before writing to finalized.
987 ACCESS_ONCE(buf
->finalized
) = 1;
988 wake_up_interruptible(&buf
->read_wait
);
991 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
993 if (config
->cb
.buffer_finalize
)
994 config
->cb
.buffer_finalize(buf
, chan
->backend
.priv
, -1);
995 if (buf
->backend
.allocated
)
996 lib_ring_buffer_set_quiescent(buf
);
998 * Perform flush before writing to finalized.
1001 ACCESS_ONCE(buf
->finalized
) = 1;
1002 wake_up_interruptible(&buf
->read_wait
);
1004 ACCESS_ONCE(chan
->finalized
) = 1;
1005 wake_up_interruptible(&chan
->hp_wait
);
1006 wake_up_interruptible(&chan
->read_wait
);
1007 priv
= chan
->backend
.priv
;
1008 kref_put(&chan
->ref
, channel_release
);
1011 EXPORT_SYMBOL_GPL(channel_destroy
);
1013 struct lib_ring_buffer
*channel_get_ring_buffer(
1014 const struct lib_ring_buffer_config
*config
,
1015 struct channel
*chan
, int cpu
)
1017 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
)
1018 return chan
->backend
.buf
;
1020 return per_cpu_ptr(chan
->backend
.buf
, cpu
);
1022 EXPORT_SYMBOL_GPL(channel_get_ring_buffer
);
1024 int lib_ring_buffer_open_read(struct lib_ring_buffer
*buf
)
1026 struct channel
*chan
= buf
->backend
.chan
;
1028 if (!atomic_long_add_unless(&buf
->active_readers
, 1, 1))
1030 if (!lttng_kref_get(&chan
->ref
)) {
1031 atomic_long_dec(&buf
->active_readers
);
1034 lttng_smp_mb__after_atomic();
1037 EXPORT_SYMBOL_GPL(lib_ring_buffer_open_read
);
1039 void lib_ring_buffer_release_read(struct lib_ring_buffer
*buf
)
1041 struct channel
*chan
= buf
->backend
.chan
;
1043 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
1044 lttng_smp_mb__before_atomic();
1045 atomic_long_dec(&buf
->active_readers
);
1046 kref_put(&chan
->ref
, channel_release
);
1048 EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read
);
1051 * Promote compiler barrier to a smp_mb().
1052 * For the specific ring buffer case, this IPI call should be removed if the
1053 * architecture does not reorder writes. This should eventually be provided by
1054 * a separate architecture-specific infrastructure.
1056 static void remote_mb(void *info
)
1062 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
1064 * @consumed: consumed count indicating the position where to read
1065 * @produced: produced count, indicates position when to stop reading
1067 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1068 * data to read at consumed position, or 0 if the get operation succeeds.
1069 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
1072 int lib_ring_buffer_snapshot(struct lib_ring_buffer
*buf
,
1073 unsigned long *consumed
, unsigned long *produced
)
1075 struct channel
*chan
= buf
->backend
.chan
;
1076 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1077 unsigned long consumed_cur
, write_offset
;
1081 finalized
= ACCESS_ONCE(buf
->finalized
);
1083 * Read finalized before counters.
1086 consumed_cur
= atomic_long_read(&buf
->consumed
);
1088 * No need to issue a memory barrier between consumed count read and
1089 * write offset read, because consumed count can only change
1090 * concurrently in overwrite mode, and we keep a sequence counter
1091 * identifier derived from the write offset to check we are getting
1092 * the same sub-buffer we are expecting (the sub-buffers are atomically
1093 * "tagged" upon writes, tags are checked upon read).
1095 write_offset
= v_read(config
, &buf
->offset
);
1098 * Check that we are not about to read the same subbuffer in
1099 * which the writer head is.
1101 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
1105 *consumed
= consumed_cur
;
1106 *produced
= subbuf_trunc(write_offset
, chan
);
1112 * The memory barriers __wait_event()/wake_up_interruptible() take care
1113 * of "raw_spin_is_locked" memory ordering.
1117 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
1122 EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot
);
1125 * Performs the same function as lib_ring_buffer_snapshot(), but the positions
1126 * are saved regardless of whether the consumed and produced positions are
1127 * in the same subbuffer.
1129 * @consumed: consumed byte count indicating the last position read
1130 * @produced: produced byte count indicating the last position written
1132 * This function is meant to provide information on the exact producer and
1133 * consumer positions without regard for the "snapshot" feature.
1135 int lib_ring_buffer_snapshot_sample_positions(struct lib_ring_buffer
*buf
,
1136 unsigned long *consumed
, unsigned long *produced
)
1138 struct channel
*chan
= buf
->backend
.chan
;
1139 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1142 *consumed
= atomic_long_read(&buf
->consumed
);
1144 * No need to issue a memory barrier between consumed count read and
1145 * write offset read, because consumed count can only change
1146 * concurrently in overwrite mode, and we keep a sequence counter
1147 * identifier derived from the write offset to check we are getting
1148 * the same sub-buffer we are expecting (the sub-buffers are atomically
1149 * "tagged" upon writes, tags are checked upon read).
1151 *produced
= v_read(config
, &buf
->offset
);
1156 * lib_ring_buffer_put_snapshot - move consumed counter forward
1158 * Should only be called from consumer context.
1160 * @consumed_new: new consumed count value
1162 void lib_ring_buffer_move_consumer(struct lib_ring_buffer
*buf
,
1163 unsigned long consumed_new
)
1165 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1166 struct channel
*chan
= bufb
->chan
;
1167 unsigned long consumed
;
1169 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
1172 * Only push the consumed value forward.
1173 * If the consumed cmpxchg fails, this is because we have been pushed by
1174 * the writer in flight recorder mode.
1176 consumed
= atomic_long_read(&buf
->consumed
);
1177 while ((long) consumed
- (long) consumed_new
< 0)
1178 consumed
= atomic_long_cmpxchg(&buf
->consumed
, consumed
,
1180 /* Wake-up the metadata producer */
1181 wake_up_interruptible(&buf
->write_wait
);
1183 EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer
);
1185 #if ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE
1186 static void lib_ring_buffer_flush_read_subbuf_dcache(
1187 const struct lib_ring_buffer_config
*config
,
1188 struct channel
*chan
,
1189 struct lib_ring_buffer
*buf
)
1191 struct lib_ring_buffer_backend_pages
*pages
;
1192 unsigned long sb_bindex
, id
, i
, nr_pages
;
1194 if (config
->output
!= RING_BUFFER_MMAP
)
1198 * Architectures with caches aliased on virtual addresses may
1199 * use different cache lines for the linear mapping vs
1200 * user-space memory mapping. Given that the ring buffer is
1201 * based on the kernel linear mapping, aligning it with the
1202 * user-space mapping is not straightforward, and would require
1203 * extra TLB entries. Therefore, simply flush the dcache for the
1204 * entire sub-buffer before reading it.
1206 id
= buf
->backend
.buf_rsb
.id
;
1207 sb_bindex
= subbuffer_id_get_index(config
, id
);
1208 pages
= buf
->backend
.array
[sb_bindex
];
1209 nr_pages
= buf
->backend
.num_pages_per_subbuf
;
1210 for (i
= 0; i
< nr_pages
; i
++) {
1211 struct lib_ring_buffer_backend_page
*backend_page
;
1213 backend_page
= &pages
->p
[i
];
1214 flush_dcache_page(pfn_to_page(backend_page
->pfn
));
1218 static void lib_ring_buffer_flush_read_subbuf_dcache(
1219 const struct lib_ring_buffer_config
*config
,
1220 struct channel
*chan
,
1221 struct lib_ring_buffer
*buf
)
1227 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
1229 * @consumed: consumed count indicating the position where to read
1231 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1232 * data to read at consumed position, or 0 if the get operation succeeds.
1233 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
1235 int lib_ring_buffer_get_subbuf(struct lib_ring_buffer
*buf
,
1236 unsigned long consumed
)
1238 struct channel
*chan
= buf
->backend
.chan
;
1239 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1240 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
1244 if (buf
->get_subbuf
) {
1246 * Reader is trying to get a subbuffer twice.
1248 CHAN_WARN_ON(chan
, 1);
1252 finalized
= ACCESS_ONCE(buf
->finalized
);
1254 * Read finalized before counters.
1257 consumed_cur
= atomic_long_read(&buf
->consumed
);
1258 consumed_idx
= subbuf_index(consumed
, chan
);
1259 commit_count
= v_read(config
, &buf
->commit_cold
[consumed_idx
].cc_sb
);
1261 * Make sure we read the commit count before reading the buffer
1262 * data and the write offset. Correct consumed offset ordering
1263 * wrt commit count is insured by the use of cmpxchg to update
1264 * the consumed offset.
1265 * smp_call_function_single can fail if the remote CPU is offline,
1266 * this is OK because then there is no wmb to execute there.
1267 * If our thread is executing on the same CPU as the on the buffers
1268 * belongs to, we don't have to synchronize it at all. If we are
1269 * migrated, the scheduler will take care of the memory barriers.
1270 * Normally, smp_call_function_single() should ensure program order when
1271 * executing the remote function, which implies that it surrounds the
1272 * function execution with :
1283 * However, smp_call_function_single() does not seem to clearly execute
1284 * such barriers. It depends on spinlock semantic to provide the barrier
1285 * before executing the IPI and, when busy-looping, csd_lock_wait only
1286 * executes smp_mb() when it has to wait for the other CPU.
1288 * I don't trust this code. Therefore, let's add the smp_mb() sequence
1289 * required ourself, even if duplicated. It has no performance impact
1292 * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs
1293 * read and write vs write. They do not ensure core synchronization. We
1294 * really have to ensure total order between the 3 barriers running on
1297 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1298 if (config
->sync
== RING_BUFFER_SYNC_PER_CPU
1299 && config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
1300 if (raw_smp_processor_id() != buf
->backend
.cpu
) {
1301 /* Total order with IPI handler smp_mb() */
1303 smp_call_function_single(buf
->backend
.cpu
,
1304 remote_mb
, NULL
, 1);
1305 /* Total order with IPI handler smp_mb() */
1309 /* Total order with IPI handler smp_mb() */
1311 smp_call_function(remote_mb
, NULL
, 1);
1312 /* Total order with IPI handler smp_mb() */
1317 * Local rmb to match the remote wmb to read the commit count
1318 * before the buffer data and the write offset.
1323 write_offset
= v_read(config
, &buf
->offset
);
1326 * Check that the buffer we are getting is after or at consumed_cur
1329 if ((long) subbuf_trunc(consumed
, chan
)
1330 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
1334 * Check that the subbuffer we are trying to consume has been
1335 * already fully committed.
1337 if (((commit_count
- chan
->backend
.subbuf_size
)
1338 & chan
->commit_count_mask
)
1339 - (buf_trunc(consumed
, chan
)
1340 >> chan
->backend
.num_subbuf_order
)
1345 * Check that we are not about to read the same subbuffer in
1346 * which the writer head is.
1348 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed
, chan
)
1353 * Failure to get the subbuffer causes a busy-loop retry without going
1354 * to a wait queue. These are caused by short-lived race windows where
1355 * the writer is getting access to a subbuffer we were trying to get
1356 * access to. Also checks that the "consumed" buffer count we are
1357 * looking for matches the one contained in the subbuffer id.
1359 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1360 consumed_idx
, buf_trunc_val(consumed
, chan
));
1363 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1365 buf
->get_subbuf_consumed
= consumed
;
1366 buf
->get_subbuf
= 1;
1368 lib_ring_buffer_flush_read_subbuf_dcache(config
, chan
, buf
);
1374 * The memory barriers __wait_event()/wake_up_interruptible() take care
1375 * of "raw_spin_is_locked" memory ordering.
1379 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
1384 EXPORT_SYMBOL_GPL(lib_ring_buffer_get_subbuf
);
1387 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1390 void lib_ring_buffer_put_subbuf(struct lib_ring_buffer
*buf
)
1392 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1393 struct channel
*chan
= bufb
->chan
;
1394 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1395 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
1397 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
1399 if (!buf
->get_subbuf
) {
1401 * Reader puts a subbuffer it did not get.
1403 CHAN_WARN_ON(chan
, 1);
1406 consumed
= buf
->get_subbuf_consumed
;
1407 buf
->get_subbuf
= 0;
1410 * Clear the records_unread counter. (overruns counter)
1411 * Can still be non-zero if a file reader simply grabbed the data
1412 * without using iterators.
1413 * Can be below zero if an iterator is used on a snapshot more than
1416 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1417 v_add(config
, v_read(config
,
1418 &bufb
->array
[read_sb_bindex
]->records_unread
),
1419 &bufb
->records_read
);
1420 v_set(config
, &bufb
->array
[read_sb_bindex
]->records_unread
, 0);
1421 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1422 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1423 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1426 * Exchange the reader subbuffer with the one we put in its place in the
1427 * writer subbuffer table. Expect the original consumed count. If
1428 * update_read_sb_index fails, this is because the writer updated the
1429 * subbuffer concurrently. We should therefore keep the subbuffer we
1430 * currently have: it has become invalid to try reading this sub-buffer
1431 * consumed count value anyway.
1433 consumed_idx
= subbuf_index(consumed
, chan
);
1434 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1435 consumed_idx
, buf_trunc_val(consumed
, chan
));
1437 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1438 * if the writer concurrently updated it.
1441 EXPORT_SYMBOL_GPL(lib_ring_buffer_put_subbuf
);
1444 * cons_offset is an iterator on all subbuffer offsets between the reader
1445 * position and the writer position. (inclusive)
1448 void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer
*buf
,
1449 struct channel
*chan
,
1450 unsigned long cons_offset
,
1453 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1454 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1456 cons_idx
= subbuf_index(cons_offset
, chan
);
1457 commit_count
= v_read(config
, &buf
->commit_hot
[cons_idx
].cc
);
1458 commit_count_sb
= v_read(config
, &buf
->commit_cold
[cons_idx
].cc_sb
);
1460 if (subbuf_offset(commit_count
, chan
) != 0)
1462 "ring buffer %s, cpu %d: "
1463 "commit count in subbuffer %lu,\n"
1464 "expecting multiples of %lu bytes\n"
1465 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1466 chan
->backend
.name
, cpu
, cons_idx
,
1467 chan
->backend
.subbuf_size
,
1468 commit_count
, commit_count_sb
);
1470 printk(KERN_DEBUG
"ring buffer: %s, cpu %d: %lu bytes committed\n",
1471 chan
->backend
.name
, cpu
, commit_count
);
1475 void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer
*buf
,
1476 struct channel
*chan
,
1477 void *priv
, int cpu
)
1479 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1480 unsigned long write_offset
, cons_offset
;
1483 * No need to order commit_count, write_offset and cons_offset reads
1484 * because we execute at teardown when no more writer nor reader
1485 * references are left.
1487 write_offset
= v_read(config
, &buf
->offset
);
1488 cons_offset
= atomic_long_read(&buf
->consumed
);
1489 if (write_offset
!= cons_offset
)
1491 "ring buffer %s, cpu %d: "
1492 "non-consumed data\n"
1493 " [ %lu bytes written, %lu bytes read ]\n",
1494 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1496 for (cons_offset
= atomic_long_read(&buf
->consumed
);
1497 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1500 cons_offset
= subbuf_align(cons_offset
, chan
))
1501 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1506 void lib_ring_buffer_print_errors(struct channel
*chan
,
1507 struct lib_ring_buffer
*buf
, int cpu
)
1509 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1510 void *priv
= chan
->backend
.priv
;
1512 if (!strcmp(chan
->backend
.name
, "relay-metadata")) {
1513 printk(KERN_DEBUG
"ring buffer %s: %lu records written, "
1514 "%lu records overrun\n",
1516 v_read(config
, &buf
->records_count
),
1517 v_read(config
, &buf
->records_overrun
));
1519 printk(KERN_DEBUG
"ring buffer %s, cpu %d: %lu records written, "
1520 "%lu records overrun\n",
1521 chan
->backend
.name
, cpu
,
1522 v_read(config
, &buf
->records_count
),
1523 v_read(config
, &buf
->records_overrun
));
1525 if (v_read(config
, &buf
->records_lost_full
)
1526 || v_read(config
, &buf
->records_lost_wrap
)
1527 || v_read(config
, &buf
->records_lost_big
))
1529 "ring buffer %s, cpu %d: records were lost. Caused by:\n"
1530 " [ %lu buffer full, %lu nest buffer wrap-around, "
1531 "%lu event too big ]\n",
1532 chan
->backend
.name
, cpu
,
1533 v_read(config
, &buf
->records_lost_full
),
1534 v_read(config
, &buf
->records_lost_wrap
),
1535 v_read(config
, &buf
->records_lost_big
));
1537 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
);
1541 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1543 * Only executed when the buffer is finalized, in SWITCH_FLUSH.
1546 void lib_ring_buffer_switch_old_start(struct lib_ring_buffer
*buf
,
1547 struct channel
*chan
,
1548 struct switch_offsets
*offsets
,
1551 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1552 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1553 unsigned long commit_count
;
1554 struct commit_counters_hot
*cc_hot
;
1556 config
->cb
.buffer_begin(buf
, tsc
, oldidx
);
1559 * Order all writes to buffer before the commit count update that will
1560 * determine that the subbuffer is full.
1562 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1564 * Must write slot data before incrementing commit count. This
1565 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1571 cc_hot
= &buf
->commit_hot
[oldidx
];
1572 v_add(config
, config
->cb
.subbuffer_header_size(), &cc_hot
->cc
);
1573 commit_count
= v_read(config
, &cc_hot
->cc
);
1574 /* Check if the written buffer has to be delivered */
1575 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1576 commit_count
, oldidx
, tsc
);
1577 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1578 offsets
->old
+ config
->cb
.subbuffer_header_size(),
1579 commit_count
, cc_hot
);
1583 * lib_ring_buffer_switch_old_end: switch old subbuffer
1585 * Note : offset_old should never be 0 here. It is ok, because we never perform
1586 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1587 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1591 void lib_ring_buffer_switch_old_end(struct lib_ring_buffer
*buf
,
1592 struct channel
*chan
,
1593 struct switch_offsets
*offsets
,
1596 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1597 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1598 unsigned long commit_count
, padding_size
, data_size
;
1599 struct commit_counters_hot
*cc_hot
;
1601 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1602 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1603 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
);
1606 * Order all writes to buffer before the commit count update that will
1607 * determine that the subbuffer is full.
1609 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1611 * Must write slot data before incrementing commit count. This
1612 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1618 cc_hot
= &buf
->commit_hot
[oldidx
];
1619 v_add(config
, padding_size
, &cc_hot
->cc
);
1620 commit_count
= v_read(config
, &cc_hot
->cc
);
1621 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1622 commit_count
, oldidx
, tsc
);
1623 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1624 offsets
->old
+ padding_size
, commit_count
,
1629 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1631 * This code can be executed unordered : writers may already have written to the
1632 * sub-buffer before this code gets executed, caution. The commit makes sure
1633 * that this code is executed before the deliver of this sub-buffer.
1636 void lib_ring_buffer_switch_new_start(struct lib_ring_buffer
*buf
,
1637 struct channel
*chan
,
1638 struct switch_offsets
*offsets
,
1641 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1642 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1643 unsigned long commit_count
;
1644 struct commit_counters_hot
*cc_hot
;
1646 config
->cb
.buffer_begin(buf
, tsc
, beginidx
);
1649 * Order all writes to buffer before the commit count update that will
1650 * determine that the subbuffer is full.
1652 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1654 * Must write slot data before incrementing commit count. This
1655 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1661 cc_hot
= &buf
->commit_hot
[beginidx
];
1662 v_add(config
, config
->cb
.subbuffer_header_size(), &cc_hot
->cc
);
1663 commit_count
= v_read(config
, &cc_hot
->cc
);
1664 /* Check if the written buffer has to be delivered */
1665 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1666 commit_count
, beginidx
, tsc
);
1667 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1668 offsets
->begin
+ config
->cb
.subbuffer_header_size(),
1669 commit_count
, cc_hot
);
1673 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1675 * Calls subbuffer_set_data_size() to set the data size of the current
1676 * sub-buffer. We do not need to perform check_deliver nor commit here,
1677 * since this task will be done by the "commit" of the event for which
1678 * we are currently doing the space reservation.
1681 void lib_ring_buffer_switch_new_end(struct lib_ring_buffer
*buf
,
1682 struct channel
*chan
,
1683 struct switch_offsets
*offsets
,
1686 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1687 unsigned long endidx
, data_size
;
1689 endidx
= subbuf_index(offsets
->end
- 1, chan
);
1690 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1691 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
);
1697 * !0 if execution must be aborted.
1700 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1701 struct lib_ring_buffer
*buf
,
1702 struct channel
*chan
,
1703 struct switch_offsets
*offsets
,
1706 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1707 unsigned long off
, reserve_commit_diff
;
1709 offsets
->begin
= v_read(config
, &buf
->offset
);
1710 offsets
->old
= offsets
->begin
;
1711 offsets
->switch_old_start
= 0;
1712 off
= subbuf_offset(offsets
->begin
, chan
);
1714 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1717 * Ensure we flush the header of an empty subbuffer when doing the
1718 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1719 * total data gathering duration even if there were no records saved
1720 * after the last buffer switch.
1721 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1722 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1723 * subbuffer header as appropriate.
1724 * The next record that reserves space will be responsible for
1725 * populating the following subbuffer header. We choose not to populate
1726 * the next subbuffer header here because we want to be able to use
1727 * SWITCH_ACTIVE for periodical buffer flush and CPU tick_nohz stop
1728 * buffer flush, which must guarantee that all the buffer content
1729 * (records and header timestamps) are visible to the reader. This is
1730 * required for quiescence guarantees for the fusion merge.
1732 if (mode
!= SWITCH_FLUSH
&& !off
)
1733 return -1; /* we do not have to switch : buffer is empty */
1735 if (unlikely(off
== 0)) {
1736 unsigned long sb_index
, commit_count
;
1739 * We are performing a SWITCH_FLUSH. At this stage, there are no
1740 * concurrent writes into the buffer.
1742 * The client does not save any header information. Don't
1743 * switch empty subbuffer on finalize, because it is invalid to
1744 * deliver a completely empty subbuffer.
1746 if (!config
->cb
.subbuffer_header_size())
1749 /* Test new buffer integrity */
1750 sb_index
= subbuf_index(offsets
->begin
, chan
);
1751 commit_count
= v_read(config
,
1752 &buf
->commit_cold
[sb_index
].cc_sb
);
1753 reserve_commit_diff
=
1754 (buf_trunc(offsets
->begin
, chan
)
1755 >> chan
->backend
.num_subbuf_order
)
1756 - (commit_count
& chan
->commit_count_mask
);
1757 if (likely(reserve_commit_diff
== 0)) {
1758 /* Next subbuffer not being written to. */
1759 if (unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1760 subbuf_trunc(offsets
->begin
, chan
)
1761 - subbuf_trunc((unsigned long)
1762 atomic_long_read(&buf
->consumed
), chan
)
1763 >= chan
->backend
.buf_size
)) {
1765 * We do not overwrite non consumed buffers
1766 * and we are full : don't switch.
1771 * Next subbuffer not being written to, and we
1772 * are either in overwrite mode or the buffer is
1773 * not full. It's safe to write in this new
1779 * Next subbuffer reserve offset does not match the
1780 * commit offset. Don't perform switch in
1781 * producer-consumer and overwrite mode. Caused by
1782 * either a writer OOPS or too many nested writes over a
1783 * reserve/commit pair.
1789 * Need to write the subbuffer start header on finalize.
1791 offsets
->switch_old_start
= 1;
1793 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1794 /* Note: old points to the next subbuf at offset 0 */
1795 offsets
->end
= offsets
->begin
;
1800 * Force a sub-buffer switch. This operation is completely reentrant : can be
1801 * called while tracing is active with absolutely no lock held.
1803 * Note, however, that as a v_cmpxchg is used for some atomic
1804 * operations, this function must be called from the CPU which owns the buffer
1805 * for a ACTIVE flush.
1807 void lib_ring_buffer_switch_slow(struct lib_ring_buffer
*buf
, enum switch_mode mode
)
1809 struct channel
*chan
= buf
->backend
.chan
;
1810 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1811 struct switch_offsets offsets
;
1812 unsigned long oldidx
;
1818 * Perform retryable operations.
1821 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1823 return; /* Switch not needed */
1824 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1828 * Atomically update last_tsc. This update races against concurrent
1829 * atomic updates, but the race will always cause supplementary full TSC
1830 * records, never the opposite (missing a full TSC record when it would
1833 save_last_tsc(config
, buf
, tsc
);
1836 * Push the reader if necessary
1838 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1840 oldidx
= subbuf_index(offsets
.old
, chan
);
1841 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
);
1844 * May need to populate header start on SWITCH_FLUSH.
1846 if (offsets
.switch_old_start
) {
1847 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
);
1848 offsets
.old
+= config
->cb
.subbuffer_header_size();
1852 * Switch old subbuffer.
1854 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
);
1856 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow
);
1858 struct switch_param
{
1859 struct lib_ring_buffer
*buf
;
1860 enum switch_mode mode
;
1863 static void remote_switch(void *info
)
1865 struct switch_param
*param
= info
;
1866 struct lib_ring_buffer
*buf
= param
->buf
;
1868 lib_ring_buffer_switch_slow(buf
, param
->mode
);
1871 static void _lib_ring_buffer_switch_remote(struct lib_ring_buffer
*buf
,
1872 enum switch_mode mode
)
1874 struct channel
*chan
= buf
->backend
.chan
;
1875 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1877 struct switch_param param
;
1880 * With global synchronization we don't need to use the IPI scheme.
1882 if (config
->sync
== RING_BUFFER_SYNC_GLOBAL
) {
1883 lib_ring_buffer_switch_slow(buf
, mode
);
1888 * Taking lock on CPU hotplug to ensure two things: first, that the
1889 * target cpu is not taken concurrently offline while we are within
1890 * smp_call_function_single() (I don't trust that get_cpu() on the
1891 * _local_ CPU actually inhibit CPU hotplug for the _remote_ CPU (to be
1892 * confirmed)). Secondly, if it happens that the CPU is not online, our
1893 * own call to lib_ring_buffer_switch_slow() needs to be protected from
1894 * CPU hotplug handlers, which can also perform a remote subbuffer
1900 ret
= smp_call_function_single(buf
->backend
.cpu
,
1901 remote_switch
, ¶m
, 1);
1903 /* Remote CPU is offline, do it ourself. */
1904 lib_ring_buffer_switch_slow(buf
, mode
);
1909 /* Switch sub-buffer if current sub-buffer is non-empty. */
1910 void lib_ring_buffer_switch_remote(struct lib_ring_buffer
*buf
)
1912 _lib_ring_buffer_switch_remote(buf
, SWITCH_ACTIVE
);
1914 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote
);
1916 /* Switch sub-buffer even if current sub-buffer is empty. */
1917 void lib_ring_buffer_switch_remote_empty(struct lib_ring_buffer
*buf
)
1919 _lib_ring_buffer_switch_remote(buf
, SWITCH_FLUSH
);
1921 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote_empty
);
1926 * -ENOSPC if event size is too large for packet.
1927 * -ENOBUFS if there is currently not enough space in buffer for the event.
1928 * -EIO if data cannot be written into the buffer for any other reason.
1931 int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer
*buf
,
1932 struct channel
*chan
,
1933 struct switch_offsets
*offsets
,
1934 struct lib_ring_buffer_ctx
*ctx
)
1936 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1937 unsigned long reserve_commit_diff
, offset_cmp
;
1940 offsets
->begin
= offset_cmp
= v_read(config
, &buf
->offset
);
1941 offsets
->old
= offsets
->begin
;
1942 offsets
->switch_new_start
= 0;
1943 offsets
->switch_new_end
= 0;
1944 offsets
->switch_old_end
= 0;
1945 offsets
->pre_header_padding
= 0;
1947 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1948 if ((int64_t) ctx
->tsc
== -EIO
)
1951 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1952 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
1954 if (unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1955 offsets
->switch_new_start
= 1; /* For offsets->begin */
1957 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1959 &offsets
->pre_header_padding
,
1962 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1965 if (unlikely(subbuf_offset(offsets
->begin
, chan
) +
1966 offsets
->size
> chan
->backend
.subbuf_size
)) {
1967 offsets
->switch_old_end
= 1; /* For offsets->old */
1968 offsets
->switch_new_start
= 1; /* For offsets->begin */
1971 if (unlikely(offsets
->switch_new_start
)) {
1972 unsigned long sb_index
, commit_count
;
1975 * We are typically not filling the previous buffer completely.
1977 if (likely(offsets
->switch_old_end
))
1978 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1979 offsets
->begin
= offsets
->begin
1980 + config
->cb
.subbuffer_header_size();
1981 /* Test new buffer integrity */
1982 sb_index
= subbuf_index(offsets
->begin
, chan
);
1984 * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
1985 * lib_ring_buffer_check_deliver() has the matching
1986 * memory barriers required around commit_cold cc_sb
1987 * updates to ensure reserve and commit counter updates
1988 * are not seen reordered when updated by another CPU.
1991 commit_count
= v_read(config
,
1992 &buf
->commit_cold
[sb_index
].cc_sb
);
1993 /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
1995 if (unlikely(offset_cmp
!= v_read(config
, &buf
->offset
))) {
1997 * The reserve counter have been concurrently updated
1998 * while we read the commit counter. This means the
1999 * commit counter we read might not match buf->offset
2000 * due to concurrent update. We therefore need to retry.
2004 reserve_commit_diff
=
2005 (buf_trunc(offsets
->begin
, chan
)
2006 >> chan
->backend
.num_subbuf_order
)
2007 - (commit_count
& chan
->commit_count_mask
);
2008 if (likely(reserve_commit_diff
== 0)) {
2009 /* Next subbuffer not being written to. */
2010 if (unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
2011 subbuf_trunc(offsets
->begin
, chan
)
2012 - subbuf_trunc((unsigned long)
2013 atomic_long_read(&buf
->consumed
), chan
)
2014 >= chan
->backend
.buf_size
)) {
2016 * We do not overwrite non consumed buffers
2017 * and we are full : record is lost.
2019 v_inc(config
, &buf
->records_lost_full
);
2023 * Next subbuffer not being written to, and we
2024 * are either in overwrite mode or the buffer is
2025 * not full. It's safe to write in this new
2031 * Next subbuffer reserve offset does not match the
2032 * commit offset, and this did not involve update to the
2033 * reserve counter. Drop record in producer-consumer and
2034 * overwrite mode. Caused by either a writer OOPS or
2035 * too many nested writes over a reserve/commit pair.
2037 v_inc(config
, &buf
->records_lost_wrap
);
2041 config
->cb
.record_header_size(config
, chan
,
2043 &offsets
->pre_header_padding
,
2046 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
2049 if (unlikely(subbuf_offset(offsets
->begin
, chan
)
2050 + offsets
->size
> chan
->backend
.subbuf_size
)) {
2052 * Record too big for subbuffers, report error, don't
2053 * complete the sub-buffer switch.
2055 v_inc(config
, &buf
->records_lost_big
);
2059 * We just made a successful buffer switch and the
2060 * record fits in the new subbuffer. Let's write.
2065 * Record fits in the current buffer and we are not on a switch
2066 * boundary. It's safe to write.
2069 offsets
->end
= offsets
->begin
+ offsets
->size
;
2071 if (unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
2073 * The offset_end will fall at the very beginning of the next
2076 offsets
->switch_new_end
= 1; /* For offsets->begin */
2081 static struct lib_ring_buffer
*get_current_buf(struct channel
*chan
, int cpu
)
2083 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
2085 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
2086 return per_cpu_ptr(chan
->backend
.buf
, cpu
);
2088 return chan
->backend
.buf
;
2091 void lib_ring_buffer_lost_event_too_big(struct channel
*chan
)
2093 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
2094 struct lib_ring_buffer
*buf
= get_current_buf(chan
, smp_processor_id());
2096 v_inc(config
, &buf
->records_lost_big
);
2098 EXPORT_SYMBOL_GPL(lib_ring_buffer_lost_event_too_big
);
2101 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
2102 * @ctx: ring buffer context.
2104 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
2105 * -EIO for other errors, else returns 0.
2106 * It will take care of sub-buffer switching.
2108 int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx
*ctx
)
2110 struct channel
*chan
= ctx
->chan
;
2111 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
2112 struct lib_ring_buffer
*buf
;
2113 struct switch_offsets offsets
;
2116 ctx
->buf
= buf
= get_current_buf(chan
, ctx
->cpu
);
2120 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
2124 } while (unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
2129 * Atomically update last_tsc. This update races against concurrent
2130 * atomic updates, but the race will always cause supplementary full TSC
2131 * records, never the opposite (missing a full TSC record when it would
2134 save_last_tsc(config
, buf
, ctx
->tsc
);
2137 * Push the reader if necessary
2139 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
2142 * Clear noref flag for this subbuffer.
2144 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2145 subbuf_index(offsets
.end
- 1, chan
));
2148 * Switch old subbuffer if needed.
2150 if (unlikely(offsets
.switch_old_end
)) {
2151 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2152 subbuf_index(offsets
.old
- 1, chan
));
2153 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
);
2157 * Populate new subbuffer.
2159 if (unlikely(offsets
.switch_new_start
))
2160 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
);
2162 if (unlikely(offsets
.switch_new_end
))
2163 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
);
2165 ctx
->slot_size
= offsets
.size
;
2166 ctx
->pre_offset
= offsets
.begin
;
2167 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
2170 EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow
);
2173 void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config
*config
,
2174 struct lib_ring_buffer
*buf
,
2175 unsigned long commit_count
,
2178 if (config
->oops
== RING_BUFFER_OOPS_CONSISTENCY
)
2179 v_set(config
, &buf
->commit_hot
[idx
].seq
, commit_count
);
2183 * The ring buffer can count events recorded and overwritten per buffer,
2184 * but it is disabled by default due to its performance overhead.
2186 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
2188 void deliver_count_events(const struct lib_ring_buffer_config
*config
,
2189 struct lib_ring_buffer
*buf
,
2192 v_add(config
, subbuffer_get_records_count(config
,
2193 &buf
->backend
, idx
),
2194 &buf
->records_count
);
2195 v_add(config
, subbuffer_count_records_overrun(config
,
2196 &buf
->backend
, idx
),
2197 &buf
->records_overrun
);
2199 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
2201 void deliver_count_events(const struct lib_ring_buffer_config
*config
,
2202 struct lib_ring_buffer
*buf
,
2206 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
2209 void lib_ring_buffer_check_deliver_slow(const struct lib_ring_buffer_config
*config
,
2210 struct lib_ring_buffer
*buf
,
2211 struct channel
*chan
,
2212 unsigned long offset
,
2213 unsigned long commit_count
,
2217 unsigned long old_commit_count
= commit_count
2218 - chan
->backend
.subbuf_size
;
2221 * If we succeeded at updating cc_sb below, we are the subbuffer
2222 * writer delivering the subbuffer. Deals with concurrent
2223 * updates of the "cc" value without adding a add_return atomic
2224 * operation to the fast path.
2226 * We are doing the delivery in two steps:
2227 * - First, we cmpxchg() cc_sb to the new value
2228 * old_commit_count + 1. This ensures that we are the only
2229 * subbuffer user successfully filling the subbuffer, but we
2230 * do _not_ set the cc_sb value to "commit_count" yet.
2231 * Therefore, other writers that would wrap around the ring
2232 * buffer and try to start writing to our subbuffer would
2233 * have to drop records, because it would appear as
2235 * We therefore have exclusive access to the subbuffer control
2236 * structures. This mutual exclusion with other writers is
2237 * crucially important to perform record overruns count in
2238 * flight recorder mode locklessly.
2239 * - When we are ready to release the subbuffer (either for
2240 * reading or for overrun by other writers), we simply set the
2241 * cc_sb value to "commit_count" and perform delivery.
2243 * The subbuffer size is least 2 bytes (minimum size: 1 page).
2244 * This guarantees that old_commit_count + 1 != commit_count.
2248 * Order prior updates to reserve count prior to the
2249 * commit_cold cc_sb update.
2252 if (likely(v_cmpxchg(config
, &buf
->commit_cold
[idx
].cc_sb
,
2253 old_commit_count
, old_commit_count
+ 1)
2254 == old_commit_count
)) {
2256 * Start of exclusive subbuffer access. We are
2257 * guaranteed to be the last writer in this subbuffer
2258 * and any other writer trying to access this subbuffer
2259 * in this state is required to drop records.
2261 deliver_count_events(config
, buf
, idx
);
2262 config
->cb
.buffer_end(buf
, tsc
, idx
,
2263 lib_ring_buffer_get_data_size(config
,
2268 * Increment the packet counter while we have exclusive
2271 subbuffer_inc_packet_count(config
, &buf
->backend
, idx
);
2274 * Set noref flag and offset for this subbuffer id.
2275 * Contains a memory barrier that ensures counter stores
2276 * are ordered before set noref and offset.
2278 lib_ring_buffer_set_noref_offset(config
, &buf
->backend
, idx
,
2279 buf_trunc_val(offset
, chan
));
2282 * Order set_noref and record counter updates before the
2283 * end of subbuffer exclusive access. Orders with
2284 * respect to writers coming into the subbuffer after
2285 * wrap around, and also order wrt concurrent readers.
2288 /* End of exclusive subbuffer access */
2289 v_set(config
, &buf
->commit_cold
[idx
].cc_sb
,
2292 * Order later updates to reserve count after
2293 * the commit_cold cc_sb update.
2296 lib_ring_buffer_vmcore_check_deliver(config
, buf
,
2300 * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
2302 if (config
->wakeup
== RING_BUFFER_WAKEUP_BY_WRITER
2303 && atomic_long_read(&buf
->active_readers
)
2304 && lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
2305 wake_up_interruptible(&buf
->read_wait
);
2306 wake_up_interruptible(&chan
->read_wait
);
2311 EXPORT_SYMBOL_GPL(lib_ring_buffer_check_deliver_slow
);
2313 int __init
init_lib_ring_buffer_frontend(void)
2317 for_each_possible_cpu(cpu
)
2318 spin_lock_init(&per_cpu(ring_buffer_nohz_lock
, cpu
));
2322 module_init(init_lib_ring_buffer_frontend
);
2324 void __exit
exit_lib_ring_buffer_frontend(void)
2328 module_exit(exit_lib_ring_buffer_frontend
);