2 * ring_buffer_frontend.c
4 * (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
6 * Ring buffer wait-free buffer synchronization. Producer-consumer and flight
7 * recorder (overwrite) modes. See thesis:
9 * Desnoyers, Mathieu (2009), "Low-Impact Operating System Tracing", Ph.D.
10 * dissertation, Ecole Polytechnique de Montreal.
11 * http://www.lttng.org/pub/thesis/desnoyers-dissertation-2009-12.pdf
13 * - Algorithm presentation in Chapter 5:
14 * "Lockless Multi-Core High-Throughput Buffering".
15 * - Algorithm formal verification in Section 8.6:
16 * "Formal verification of LTTng"
19 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
21 * Inspired from LTT and RelayFS:
22 * Karim Yaghmour <karim@opersys.com>
23 * Tom Zanussi <zanussi@us.ibm.com>
24 * Bob Wisniewski <bob@watson.ibm.com>
26 * Bob Wisniewski <bob@watson.ibm.com>
28 * Buffer reader semantic :
31 * while buffer is not finalized and empty
33 * - if return value != 0, continue
34 * - splice one subbuffer worth of data to a pipe
35 * - splice the data from pipe to disk/network
38 * Dual LGPL v2.1/GPL v2 license.
48 * Internal structure representing offsets to use at a sub-buffer switch.
50 struct switch_offsets
{
51 unsigned long begin
, end
, old
;
52 size_t pre_header_padding
, size
;
53 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
64 static ATOMIC_NOTIFIER_HEAD(tick_nohz_notifier
);
65 #endif /* CONFIG_NO_HZ */
67 static DEFINE_PER_CPU(spinlock_t
, ring_buffer_nohz_lock
);
69 DEFINE_PER_CPU(unsigned int, lib_ring_buffer_nesting
);
70 EXPORT_PER_CPU_SYMBOL(lib_ring_buffer_nesting
);
73 void lib_ring_buffer_print_errors(struct channel
*chan
,
74 struct lib_ring_buffer
*buf
, int cpu
);
77 * Must be called under cpu hotplug protection.
79 void lib_ring_buffer_free(struct lib_ring_buffer
*buf
)
81 struct channel
*chan
= buf
->backend
.chan
;
83 lib_ring_buffer_print_errors(chan
, buf
, buf
->backend
.cpu
);
84 kfree(buf
->commit_hot
);
85 kfree(buf
->commit_cold
);
87 lib_ring_buffer_backend_free(&buf
->backend
);
91 * lib_ring_buffer_reset - Reset ring buffer to initial values.
94 * Effectively empty the ring buffer. Should be called when the buffer is not
95 * used for writing. The ring buffer can be opened for reading, but the reader
96 * should not be using the iterator concurrently with reset. The previous
97 * current iterator record is reset.
99 void lib_ring_buffer_reset(struct lib_ring_buffer
*buf
)
101 struct channel
*chan
= buf
->backend
.chan
;
102 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
106 * Reset iterator first. It will put the subbuffer if it currently holds
109 lib_ring_buffer_iterator_reset(buf
);
110 v_set(config
, &buf
->offset
, 0);
111 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
112 v_set(config
, &buf
->commit_hot
[i
].cc
, 0);
113 v_set(config
, &buf
->commit_hot
[i
].seq
, 0);
114 v_set(config
, &buf
->commit_cold
[i
].cc_sb
, 0);
116 atomic_long_set(&buf
->consumed
, 0);
117 atomic_set(&buf
->record_disabled
, 0);
118 v_set(config
, &buf
->last_tsc
, 0);
119 lib_ring_buffer_backend_reset(&buf
->backend
);
120 /* Don't reset number of active readers */
121 v_set(config
, &buf
->records_lost_full
, 0);
122 v_set(config
, &buf
->records_lost_wrap
, 0);
123 v_set(config
, &buf
->records_lost_big
, 0);
124 v_set(config
, &buf
->records_count
, 0);
125 v_set(config
, &buf
->records_overrun
, 0);
128 EXPORT_SYMBOL_GPL(lib_ring_buffer_reset
);
131 * channel_reset - Reset channel to initial values.
134 * Effectively empty the channel. Should be called when the channel is not used
135 * for writing. The channel can be opened for reading, but the reader should not
136 * be using the iterator concurrently with reset. The previous current iterator
139 void channel_reset(struct channel
*chan
)
142 * Reset iterators first. Will put the subbuffer if held for reading.
144 channel_iterator_reset(chan
);
145 atomic_set(&chan
->record_disabled
, 0);
146 /* Don't reset commit_count_mask, still valid */
147 channel_backend_reset(&chan
->backend
);
148 /* Don't reset switch/read timer interval */
149 /* Don't reset notifiers and notifier enable bits */
150 /* Don't reset reader reference count */
152 EXPORT_SYMBOL_GPL(channel_reset
);
155 * Must be called under cpu hotplug protection.
157 int lib_ring_buffer_create(struct lib_ring_buffer
*buf
,
158 struct channel_backend
*chanb
, int cpu
)
160 const struct lib_ring_buffer_config
*config
= chanb
->config
;
161 struct channel
*chan
= container_of(chanb
, struct channel
, backend
);
162 void *priv
= chanb
->priv
;
163 unsigned int num_subbuf
;
164 size_t subbuf_header_size
;
168 /* Test for cpu hotplug */
169 if (buf
->backend
.allocated
)
173 * Paranoia: per cpu dynamic allocation is not officially documented as
174 * zeroing the memory, so let's do it here too, just in case.
176 memset(buf
, 0, sizeof(*buf
));
178 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
, cpu
);
183 kzalloc_node(ALIGN(sizeof(*buf
->commit_hot
)
184 * chan
->backend
.num_subbuf
,
185 1 << INTERNODE_CACHE_SHIFT
),
186 GFP_KERNEL
, cpu_to_node(max(cpu
, 0)));
187 if (!buf
->commit_hot
) {
193 kzalloc_node(ALIGN(sizeof(*buf
->commit_cold
)
194 * chan
->backend
.num_subbuf
,
195 1 << INTERNODE_CACHE_SHIFT
),
196 GFP_KERNEL
, cpu_to_node(max(cpu
, 0)));
197 if (!buf
->commit_cold
) {
202 num_subbuf
= chan
->backend
.num_subbuf
;
203 init_waitqueue_head(&buf
->read_wait
);
204 raw_spin_lock_init(&buf
->raw_tick_nohz_spinlock
);
207 * Write the subbuffer header for first subbuffer so we know the total
208 * duration of data gathering.
210 subbuf_header_size
= config
->cb
.subbuffer_header_size();
211 v_set(config
, &buf
->offset
, subbuf_header_size
);
212 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_wsb
[0].id
);
213 tsc
= config
->cb
.ring_buffer_clock_read(buf
->backend
.chan
);
214 config
->cb
.buffer_begin(buf
, tsc
, 0);
215 v_add(config
, subbuf_header_size
, &buf
->commit_hot
[0].cc
);
217 if (config
->cb
.buffer_create
) {
218 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
);
224 * Ensure the buffer is ready before setting it to allocated and setting
226 * Used for cpu hotplug vs cpumask iteration.
229 buf
->backend
.allocated
= 1;
231 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
232 CHAN_WARN_ON(chan
, cpumask_test_cpu(cpu
,
233 chan
->backend
.cpumask
));
234 cpumask_set_cpu(cpu
, chan
->backend
.cpumask
);
241 kfree(buf
->commit_cold
);
243 kfree(buf
->commit_hot
);
245 lib_ring_buffer_backend_free(&buf
->backend
);
249 static void switch_buffer_timer(unsigned long data
)
251 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
252 struct channel
*chan
= buf
->backend
.chan
;
253 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
256 * Only flush buffers periodically if readers are active.
258 if (atomic_long_read(&buf
->active_readers
))
259 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
261 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
262 mod_timer_pinned(&buf
->switch_timer
,
263 jiffies
+ chan
->switch_timer_interval
);
265 mod_timer(&buf
->switch_timer
,
266 jiffies
+ chan
->switch_timer_interval
);
270 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
272 static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer
*buf
)
274 struct channel
*chan
= buf
->backend
.chan
;
275 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
277 if (!chan
->switch_timer_interval
|| buf
->switch_timer_enabled
)
279 init_timer(&buf
->switch_timer
);
280 buf
->switch_timer
.function
= switch_buffer_timer
;
281 buf
->switch_timer
.expires
= jiffies
+ chan
->switch_timer_interval
;
282 buf
->switch_timer
.data
= (unsigned long)buf
;
283 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
284 add_timer_on(&buf
->switch_timer
, buf
->backend
.cpu
);
286 add_timer(&buf
->switch_timer
);
287 buf
->switch_timer_enabled
= 1;
291 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
293 static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer
*buf
)
295 struct channel
*chan
= buf
->backend
.chan
;
297 if (!chan
->switch_timer_interval
|| !buf
->switch_timer_enabled
)
300 del_timer_sync(&buf
->switch_timer
);
301 buf
->switch_timer_enabled
= 0;
305 * Polling timer to check the channels for data.
307 static void read_buffer_timer(unsigned long data
)
309 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
310 struct channel
*chan
= buf
->backend
.chan
;
311 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
313 CHAN_WARN_ON(chan
, !buf
->backend
.allocated
);
315 if (atomic_long_read(&buf
->active_readers
)
316 && lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
317 wake_up_interruptible(&buf
->read_wait
);
318 wake_up_interruptible(&chan
->read_wait
);
321 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
322 mod_timer_pinned(&buf
->read_timer
,
323 jiffies
+ chan
->read_timer_interval
);
325 mod_timer(&buf
->read_timer
,
326 jiffies
+ chan
->read_timer_interval
);
330 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
332 static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer
*buf
)
334 struct channel
*chan
= buf
->backend
.chan
;
335 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
337 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
338 || !chan
->read_timer_interval
339 || buf
->read_timer_enabled
)
342 init_timer(&buf
->read_timer
);
343 buf
->read_timer
.function
= read_buffer_timer
;
344 buf
->read_timer
.expires
= jiffies
+ chan
->read_timer_interval
;
345 buf
->read_timer
.data
= (unsigned long)buf
;
347 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
348 add_timer_on(&buf
->read_timer
, buf
->backend
.cpu
);
350 add_timer(&buf
->read_timer
);
351 buf
->read_timer_enabled
= 1;
355 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
357 static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer
*buf
)
359 struct channel
*chan
= buf
->backend
.chan
;
360 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
362 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
363 || !chan
->read_timer_interval
364 || !buf
->read_timer_enabled
)
367 del_timer_sync(&buf
->read_timer
);
369 * do one more check to catch data that has been written in the last
372 if (lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
373 wake_up_interruptible(&buf
->read_wait
);
374 wake_up_interruptible(&chan
->read_wait
);
376 buf
->read_timer_enabled
= 0;
379 #ifdef CONFIG_HOTPLUG_CPU
381 * lib_ring_buffer_cpu_hp_callback - CPU hotplug callback
382 * @nb: notifier block
383 * @action: hotplug action to take
386 * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
389 int __cpuinit
lib_ring_buffer_cpu_hp_callback(struct notifier_block
*nb
,
390 unsigned long action
,
393 unsigned int cpu
= (unsigned long)hcpu
;
394 struct channel
*chan
= container_of(nb
, struct channel
,
396 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
, cpu
);
397 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
399 if (!chan
->cpu_hp_enable
)
402 CHAN_WARN_ON(chan
, config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
);
405 case CPU_DOWN_FAILED
:
406 case CPU_DOWN_FAILED_FROZEN
:
408 case CPU_ONLINE_FROZEN
:
409 wake_up_interruptible(&chan
->hp_wait
);
410 lib_ring_buffer_start_switch_timer(buf
);
411 lib_ring_buffer_start_read_timer(buf
);
414 case CPU_DOWN_PREPARE
:
415 case CPU_DOWN_PREPARE_FROZEN
:
416 lib_ring_buffer_stop_switch_timer(buf
);
417 lib_ring_buffer_stop_read_timer(buf
);
421 case CPU_DEAD_FROZEN
:
423 * Performing a buffer switch on a remote CPU. Performed by
424 * the CPU responsible for doing the hotunplug after the target
425 * CPU stopped running completely. Ensures that all data
426 * from that remote CPU is flushed.
428 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
437 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
439 * For per-cpu buffers, call the reader wakeups before switching the buffer, so
440 * that wake-up-tracing generated events are flushed before going idle (in
441 * tick_nohz). We test if the spinlock is locked to deal with the race where
442 * readers try to sample the ring buffer before we perform the switch. We let
443 * the readers retry in that case. If there is data in the buffer, the wake up
444 * is going to forbid the CPU running the reader thread from going idle.
446 static int notrace
ring_buffer_tick_nohz_callback(struct notifier_block
*nb
,
450 struct channel
*chan
= container_of(nb
, struct channel
,
452 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
453 struct lib_ring_buffer
*buf
;
454 int cpu
= smp_processor_id();
456 if (config
->alloc
!= RING_BUFFER_ALLOC_PER_CPU
) {
458 * We don't support keeping the system idle with global buffers
459 * and streaming active. In order to do so, we would need to
460 * sample a non-nohz-cpumask racelessly with the nohz updates
461 * without adding synchronization overhead to nohz. Leave this
462 * use-case out for now.
467 buf
= channel_get_ring_buffer(config
, chan
, cpu
);
469 case TICK_NOHZ_FLUSH
:
470 raw_spin_lock(&buf
->raw_tick_nohz_spinlock
);
471 if (config
->wakeup
== RING_BUFFER_WAKEUP_BY_TIMER
472 && chan
->read_timer_interval
473 && atomic_long_read(&buf
->active_readers
)
474 && (lib_ring_buffer_poll_deliver(config
, buf
, chan
)
475 || lib_ring_buffer_pending_data(config
, buf
, chan
))) {
476 wake_up_interruptible(&buf
->read_wait
);
477 wake_up_interruptible(&chan
->read_wait
);
479 if (chan
->switch_timer_interval
)
480 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
481 raw_spin_unlock(&buf
->raw_tick_nohz_spinlock
);
484 spin_lock(&__get_cpu_var(ring_buffer_nohz_lock
));
485 lib_ring_buffer_stop_switch_timer(buf
);
486 lib_ring_buffer_stop_read_timer(buf
);
487 spin_unlock(&__get_cpu_var(ring_buffer_nohz_lock
));
489 case TICK_NOHZ_RESTART
:
490 spin_lock(&__get_cpu_var(ring_buffer_nohz_lock
));
491 lib_ring_buffer_start_read_timer(buf
);
492 lib_ring_buffer_start_switch_timer(buf
);
493 spin_unlock(&__get_cpu_var(ring_buffer_nohz_lock
));
500 void notrace
lib_ring_buffer_tick_nohz_flush(void)
502 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_FLUSH
,
506 void notrace
lib_ring_buffer_tick_nohz_stop(void)
508 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_STOP
,
512 void notrace
lib_ring_buffer_tick_nohz_restart(void)
514 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_RESTART
,
517 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
522 static void channel_unregister_notifiers(struct channel
*chan
)
524 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
527 channel_iterator_unregister_notifiers(chan
);
528 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
531 * Remove the nohz notifier first, so we are certain we stop
534 atomic_notifier_chain_unregister(&tick_nohz_notifier
,
535 &chan
->tick_nohz_notifier
);
537 * ring_buffer_nohz_lock will not be needed below, because
538 * we just removed the notifiers, which were the only source of
541 #endif /* CONFIG_NO_HZ */
542 #ifdef CONFIG_HOTPLUG_CPU
544 chan
->cpu_hp_enable
= 0;
545 for_each_online_cpu(cpu
) {
546 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
548 lib_ring_buffer_stop_switch_timer(buf
);
549 lib_ring_buffer_stop_read_timer(buf
);
552 unregister_cpu_notifier(&chan
->cpu_hp_notifier
);
554 for_each_possible_cpu(cpu
) {
555 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
557 lib_ring_buffer_stop_switch_timer(buf
);
558 lib_ring_buffer_stop_read_timer(buf
);
562 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
564 lib_ring_buffer_stop_switch_timer(buf
);
565 lib_ring_buffer_stop_read_timer(buf
);
567 channel_backend_unregister_notifiers(&chan
->backend
);
570 static void channel_free(struct channel
*chan
)
572 channel_iterator_free(chan
);
573 channel_backend_free(&chan
->backend
);
578 * channel_create - Create channel.
579 * @config: ring buffer instance configuration
580 * @name: name of the channel
581 * @priv: ring buffer client private data
582 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
583 * address mapping. It is used only by RING_BUFFER_STATIC
584 * configuration. It can be set to NULL for other backends.
585 * @subbuf_size: subbuffer size
586 * @num_subbuf: number of subbuffers
587 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
588 * padding to let readers get those sub-buffers.
589 * Used for live streaming.
590 * @read_timer_interval: Time interval (in us) to wake up pending readers.
593 * Returns NULL on failure.
595 struct channel
*channel_create(const struct lib_ring_buffer_config
*config
,
596 const char *name
, void *priv
, void *buf_addr
,
598 size_t num_subbuf
, unsigned int switch_timer_interval
,
599 unsigned int read_timer_interval
)
602 struct channel
*chan
;
604 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
605 read_timer_interval
))
608 chan
= kzalloc(sizeof(struct channel
), GFP_KERNEL
);
612 ret
= channel_backend_init(&chan
->backend
, name
, config
, priv
,
613 subbuf_size
, num_subbuf
);
617 ret
= channel_iterator_init(chan
);
619 goto error_free_backend
;
621 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
622 chan
->switch_timer_interval
= usecs_to_jiffies(switch_timer_interval
);
623 chan
->read_timer_interval
= usecs_to_jiffies(read_timer_interval
);
624 kref_init(&chan
->ref
);
625 init_waitqueue_head(&chan
->read_wait
);
626 init_waitqueue_head(&chan
->hp_wait
);
628 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
629 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
630 /* Only benefit from NO_HZ idle with per-cpu buffers for now. */
631 chan
->tick_nohz_notifier
.notifier_call
=
632 ring_buffer_tick_nohz_callback
;
633 chan
->tick_nohz_notifier
.priority
= ~0U;
634 atomic_notifier_chain_register(&tick_nohz_notifier
,
635 &chan
->tick_nohz_notifier
);
636 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
639 * In case of non-hotplug cpu, if the ring-buffer is allocated
640 * in early initcall, it will not be notified of secondary cpus.
641 * In that off case, we need to allocate for all possible cpus.
643 #ifdef CONFIG_HOTPLUG_CPU
644 chan
->cpu_hp_notifier
.notifier_call
=
645 lib_ring_buffer_cpu_hp_callback
;
646 chan
->cpu_hp_notifier
.priority
= 6;
647 register_cpu_notifier(&chan
->cpu_hp_notifier
);
650 for_each_online_cpu(cpu
) {
651 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
653 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
654 lib_ring_buffer_start_switch_timer(buf
);
655 lib_ring_buffer_start_read_timer(buf
);
656 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
658 chan
->cpu_hp_enable
= 1;
661 for_each_possible_cpu(cpu
) {
662 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
664 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
665 lib_ring_buffer_start_switch_timer(buf
);
666 lib_ring_buffer_start_read_timer(buf
);
667 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
671 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
673 lib_ring_buffer_start_switch_timer(buf
);
674 lib_ring_buffer_start_read_timer(buf
);
680 channel_backend_free(&chan
->backend
);
685 EXPORT_SYMBOL_GPL(channel_create
);
688 void channel_release(struct kref
*kref
)
690 struct channel
*chan
= container_of(kref
, struct channel
, ref
);
695 * channel_destroy - Finalize, wait for q.s. and destroy channel.
696 * @chan: channel to destroy
699 * Call "destroy" callback, finalize channels, wait for readers to release their
700 * reference, then destroy ring buffer data. Note that when readers have
701 * completed data consumption of finalized channels, get_subbuf() will return
702 * -ENODATA. They should release their handle at that point.
703 * Returns the private data pointer.
705 void *channel_destroy(struct channel
*chan
)
708 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
711 channel_unregister_notifiers(chan
);
713 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
715 * No need to hold cpu hotplug, because all notifiers have been
718 for_each_channel_cpu(cpu
, chan
) {
719 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
722 if (config
->cb
.buffer_finalize
)
723 config
->cb
.buffer_finalize(buf
,
726 if (buf
->backend
.allocated
)
727 lib_ring_buffer_switch_slow(buf
, SWITCH_FLUSH
);
729 * Perform flush before writing to finalized.
732 ACCESS_ONCE(buf
->finalized
) = 1;
733 wake_up_interruptible(&buf
->read_wait
);
736 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
738 if (config
->cb
.buffer_finalize
)
739 config
->cb
.buffer_finalize(buf
, chan
->backend
.priv
, -1);
740 if (buf
->backend
.allocated
)
741 lib_ring_buffer_switch_slow(buf
, SWITCH_FLUSH
);
743 * Perform flush before writing to finalized.
746 ACCESS_ONCE(buf
->finalized
) = 1;
747 wake_up_interruptible(&buf
->read_wait
);
749 ACCESS_ONCE(chan
->finalized
) = 1;
750 wake_up_interruptible(&chan
->hp_wait
);
751 wake_up_interruptible(&chan
->read_wait
);
752 kref_put(&chan
->ref
, channel_release
);
753 priv
= chan
->backend
.priv
;
756 EXPORT_SYMBOL_GPL(channel_destroy
);
758 struct lib_ring_buffer
*channel_get_ring_buffer(
759 const struct lib_ring_buffer_config
*config
,
760 struct channel
*chan
, int cpu
)
762 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
)
763 return chan
->backend
.buf
;
765 return per_cpu_ptr(chan
->backend
.buf
, cpu
);
767 EXPORT_SYMBOL_GPL(channel_get_ring_buffer
);
769 int lib_ring_buffer_open_read(struct lib_ring_buffer
*buf
)
771 struct channel
*chan
= buf
->backend
.chan
;
773 if (!atomic_long_add_unless(&buf
->active_readers
, 1, 1))
775 kref_get(&chan
->ref
);
776 smp_mb__after_atomic_inc();
779 EXPORT_SYMBOL_GPL(lib_ring_buffer_open_read
);
781 void lib_ring_buffer_release_read(struct lib_ring_buffer
*buf
)
783 struct channel
*chan
= buf
->backend
.chan
;
785 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
786 smp_mb__before_atomic_dec();
787 atomic_long_dec(&buf
->active_readers
);
788 kref_put(&chan
->ref
, channel_release
);
790 EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read
);
793 * Promote compiler barrier to a smp_mb().
794 * For the specific ring buffer case, this IPI call should be removed if the
795 * architecture does not reorder writes. This should eventually be provided by
796 * a separate architecture-specific infrastructure.
798 static void remote_mb(void *info
)
804 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
806 * @consumed: consumed count indicating the position where to read
807 * @produced: produced count, indicates position when to stop reading
809 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
810 * data to read at consumed position, or 0 if the get operation succeeds.
811 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
814 int lib_ring_buffer_snapshot(struct lib_ring_buffer
*buf
,
815 unsigned long *consumed
, unsigned long *produced
)
817 struct channel
*chan
= buf
->backend
.chan
;
818 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
819 unsigned long consumed_cur
, write_offset
;
823 finalized
= ACCESS_ONCE(buf
->finalized
);
825 * Read finalized before counters.
828 consumed_cur
= atomic_long_read(&buf
->consumed
);
830 * No need to issue a memory barrier between consumed count read and
831 * write offset read, because consumed count can only change
832 * concurrently in overwrite mode, and we keep a sequence counter
833 * identifier derived from the write offset to check we are getting
834 * the same sub-buffer we are expecting (the sub-buffers are atomically
835 * "tagged" upon writes, tags are checked upon read).
837 write_offset
= v_read(config
, &buf
->offset
);
840 * Check that we are not about to read the same subbuffer in
841 * which the writer head is.
843 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
847 *consumed
= consumed_cur
;
848 *produced
= subbuf_trunc(write_offset
, chan
);
854 * The memory barriers __wait_event()/wake_up_interruptible() take care
855 * of "raw_spin_is_locked" memory ordering.
859 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
864 EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot
);
867 * lib_ring_buffer_put_snapshot - move consumed counter forward
869 * @consumed_new: new consumed count value
871 void lib_ring_buffer_move_consumer(struct lib_ring_buffer
*buf
,
872 unsigned long consumed_new
)
874 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
875 struct channel
*chan
= bufb
->chan
;
876 unsigned long consumed
;
878 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
881 * Only push the consumed value forward.
882 * If the consumed cmpxchg fails, this is because we have been pushed by
883 * the writer in flight recorder mode.
885 consumed
= atomic_long_read(&buf
->consumed
);
886 while ((long) consumed
- (long) consumed_new
< 0)
887 consumed
= atomic_long_cmpxchg(&buf
->consumed
, consumed
,
890 EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer
);
893 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
895 * @consumed: consumed count indicating the position where to read
897 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
898 * data to read at consumed position, or 0 if the get operation succeeds.
899 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
901 int lib_ring_buffer_get_subbuf(struct lib_ring_buffer
*buf
,
902 unsigned long consumed
)
904 struct channel
*chan
= buf
->backend
.chan
;
905 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
906 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
911 finalized
= ACCESS_ONCE(buf
->finalized
);
913 * Read finalized before counters.
916 consumed_cur
= atomic_long_read(&buf
->consumed
);
917 consumed_idx
= subbuf_index(consumed
, chan
);
918 commit_count
= v_read(config
, &buf
->commit_cold
[consumed_idx
].cc_sb
);
920 * Make sure we read the commit count before reading the buffer
921 * data and the write offset. Correct consumed offset ordering
922 * wrt commit count is insured by the use of cmpxchg to update
923 * the consumed offset.
924 * smp_call_function_single can fail if the remote CPU is offline,
925 * this is OK because then there is no wmb to execute there.
926 * If our thread is executing on the same CPU as the on the buffers
927 * belongs to, we don't have to synchronize it at all. If we are
928 * migrated, the scheduler will take care of the memory barriers.
929 * Normally, smp_call_function_single() should ensure program order when
930 * executing the remote function, which implies that it surrounds the
931 * function execution with :
942 * However, smp_call_function_single() does not seem to clearly execute
943 * such barriers. It depends on spinlock semantic to provide the barrier
944 * before executing the IPI and, when busy-looping, csd_lock_wait only
945 * executes smp_mb() when it has to wait for the other CPU.
947 * I don't trust this code. Therefore, let's add the smp_mb() sequence
948 * required ourself, even if duplicated. It has no performance impact
951 * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs
952 * read and write vs write. They do not ensure core synchronization. We
953 * really have to ensure total order between the 3 barriers running on
956 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
957 if (config
->sync
== RING_BUFFER_SYNC_PER_CPU
958 && config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
959 if (raw_smp_processor_id() != buf
->backend
.cpu
) {
960 /* Total order with IPI handler smp_mb() */
962 smp_call_function_single(buf
->backend
.cpu
,
964 /* Total order with IPI handler smp_mb() */
968 /* Total order with IPI handler smp_mb() */
970 smp_call_function(remote_mb
, NULL
, 1);
971 /* Total order with IPI handler smp_mb() */
976 * Local rmb to match the remote wmb to read the commit count
977 * before the buffer data and the write offset.
982 write_offset
= v_read(config
, &buf
->offset
);
985 * Check that the buffer we are getting is after or at consumed_cur
988 if ((long) subbuf_trunc(consumed
, chan
)
989 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
993 * Check that the subbuffer we are trying to consume has been
994 * already fully committed.
996 if (((commit_count
- chan
->backend
.subbuf_size
)
997 & chan
->commit_count_mask
)
998 - (buf_trunc(consumed_cur
, chan
)
999 >> chan
->backend
.num_subbuf_order
)
1004 * Check that we are not about to read the same subbuffer in
1005 * which the writer head is.
1007 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
1012 * Failure to get the subbuffer causes a busy-loop retry without going
1013 * to a wait queue. These are caused by short-lived race windows where
1014 * the writer is getting access to a subbuffer we were trying to get
1015 * access to. Also checks that the "consumed" buffer count we are
1016 * looking for matches the one contained in the subbuffer id.
1018 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1019 consumed_idx
, buf_trunc_val(consumed
, chan
));
1022 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1024 buf
->get_subbuf_consumed
= consumed
;
1025 buf
->get_subbuf
= 1;
1031 * The memory barriers __wait_event()/wake_up_interruptible() take care
1032 * of "raw_spin_is_locked" memory ordering.
1036 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
1041 EXPORT_SYMBOL_GPL(lib_ring_buffer_get_subbuf
);
1044 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1047 void lib_ring_buffer_put_subbuf(struct lib_ring_buffer
*buf
)
1049 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1050 struct channel
*chan
= bufb
->chan
;
1051 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1052 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
1054 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
1056 if (!buf
->get_subbuf
) {
1058 * Reader puts a subbuffer it did not get.
1060 CHAN_WARN_ON(chan
, 1);
1063 consumed
= buf
->get_subbuf_consumed
;
1064 buf
->get_subbuf
= 0;
1067 * Clear the records_unread counter. (overruns counter)
1068 * Can still be non-zero if a file reader simply grabbed the data
1069 * without using iterators.
1070 * Can be below zero if an iterator is used on a snapshot more than
1073 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1074 v_add(config
, v_read(config
,
1075 &bufb
->array
[read_sb_bindex
]->records_unread
),
1076 &bufb
->records_read
);
1077 v_set(config
, &bufb
->array
[read_sb_bindex
]->records_unread
, 0);
1078 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1079 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1080 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1083 * Exchange the reader subbuffer with the one we put in its place in the
1084 * writer subbuffer table. Expect the original consumed count. If
1085 * update_read_sb_index fails, this is because the writer updated the
1086 * subbuffer concurrently. We should therefore keep the subbuffer we
1087 * currently have: it has become invalid to try reading this sub-buffer
1088 * consumed count value anyway.
1090 consumed_idx
= subbuf_index(consumed
, chan
);
1091 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1092 consumed_idx
, buf_trunc_val(consumed
, chan
));
1094 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1095 * if the writer concurrently updated it.
1098 EXPORT_SYMBOL_GPL(lib_ring_buffer_put_subbuf
);
1101 * cons_offset is an iterator on all subbuffer offsets between the reader
1102 * position and the writer position. (inclusive)
1105 void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer
*buf
,
1106 struct channel
*chan
,
1107 unsigned long cons_offset
,
1110 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1111 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1113 cons_idx
= subbuf_index(cons_offset
, chan
);
1114 commit_count
= v_read(config
, &buf
->commit_hot
[cons_idx
].cc
);
1115 commit_count_sb
= v_read(config
, &buf
->commit_cold
[cons_idx
].cc_sb
);
1117 if (subbuf_offset(commit_count
, chan
) != 0)
1119 "ring buffer %s, cpu %d: "
1120 "commit count in subbuffer %lu,\n"
1121 "expecting multiples of %lu bytes\n"
1122 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1123 chan
->backend
.name
, cpu
, cons_idx
,
1124 chan
->backend
.subbuf_size
,
1125 commit_count
, commit_count_sb
);
1127 printk(KERN_DEBUG
"ring buffer: %s, cpu %d: %lu bytes committed\n",
1128 chan
->backend
.name
, cpu
, commit_count
);
1132 void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer
*buf
,
1133 struct channel
*chan
,
1134 void *priv
, int cpu
)
1136 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1137 unsigned long write_offset
, cons_offset
;
1140 * Can be called in the error path of allocation when
1141 * trans_channel_data is not yet set.
1146 * No need to order commit_count, write_offset and cons_offset reads
1147 * because we execute at teardown when no more writer nor reader
1148 * references are left.
1150 write_offset
= v_read(config
, &buf
->offset
);
1151 cons_offset
= atomic_long_read(&buf
->consumed
);
1152 if (write_offset
!= cons_offset
)
1154 "ring buffer %s, cpu %d: "
1155 "non-consumed data\n"
1156 " [ %lu bytes written, %lu bytes read ]\n",
1157 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1159 for (cons_offset
= atomic_long_read(&buf
->consumed
);
1160 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1163 cons_offset
= subbuf_align(cons_offset
, chan
))
1164 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1169 void lib_ring_buffer_print_errors(struct channel
*chan
,
1170 struct lib_ring_buffer
*buf
, int cpu
)
1172 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1173 void *priv
= chan
->backend
.priv
;
1175 printk(KERN_DEBUG
"ring buffer %s, cpu %d: %lu records written, "
1176 "%lu records overrun\n",
1177 chan
->backend
.name
, cpu
,
1178 v_read(config
, &buf
->records_count
),
1179 v_read(config
, &buf
->records_overrun
));
1181 if (v_read(config
, &buf
->records_lost_full
)
1182 || v_read(config
, &buf
->records_lost_wrap
)
1183 || v_read(config
, &buf
->records_lost_big
))
1185 "ring buffer %s, cpu %d: records were lost. Caused by:\n"
1186 " [ %lu buffer full, %lu nest buffer wrap-around, "
1187 "%lu event too big ]\n",
1188 chan
->backend
.name
, cpu
,
1189 v_read(config
, &buf
->records_lost_full
),
1190 v_read(config
, &buf
->records_lost_wrap
),
1191 v_read(config
, &buf
->records_lost_big
));
1193 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
);
1197 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1199 * Only executed when the buffer is finalized, in SWITCH_FLUSH.
1202 void lib_ring_buffer_switch_old_start(struct lib_ring_buffer
*buf
,
1203 struct channel
*chan
,
1204 struct switch_offsets
*offsets
,
1207 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1208 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1209 unsigned long commit_count
;
1211 config
->cb
.buffer_begin(buf
, tsc
, oldidx
);
1214 * Order all writes to buffer before the commit count update that will
1215 * determine that the subbuffer is full.
1217 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1219 * Must write slot data before incrementing commit count. This
1220 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1226 v_add(config
, config
->cb
.subbuffer_header_size(),
1227 &buf
->commit_hot
[oldidx
].cc
);
1228 commit_count
= v_read(config
, &buf
->commit_hot
[oldidx
].cc
);
1229 /* Check if the written buffer has to be delivered */
1230 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1231 commit_count
, oldidx
);
1232 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1233 offsets
->old
, commit_count
,
1234 config
->cb
.subbuffer_header_size());
1238 * lib_ring_buffer_switch_old_end: switch old subbuffer
1240 * Note : offset_old should never be 0 here. It is ok, because we never perform
1241 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1242 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1246 void lib_ring_buffer_switch_old_end(struct lib_ring_buffer
*buf
,
1247 struct channel
*chan
,
1248 struct switch_offsets
*offsets
,
1251 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1252 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1253 unsigned long commit_count
, padding_size
, data_size
;
1255 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1256 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1257 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
);
1260 * Order all writes to buffer before the commit count update that will
1261 * determine that the subbuffer is full.
1263 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1265 * Must write slot data before incrementing commit count. This
1266 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1272 v_add(config
, padding_size
, &buf
->commit_hot
[oldidx
].cc
);
1273 commit_count
= v_read(config
, &buf
->commit_hot
[oldidx
].cc
);
1274 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1275 commit_count
, oldidx
);
1276 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1277 offsets
->old
, commit_count
,
1282 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1284 * This code can be executed unordered : writers may already have written to the
1285 * sub-buffer before this code gets executed, caution. The commit makes sure
1286 * that this code is executed before the deliver of this sub-buffer.
1289 void lib_ring_buffer_switch_new_start(struct lib_ring_buffer
*buf
,
1290 struct channel
*chan
,
1291 struct switch_offsets
*offsets
,
1294 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1295 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1296 unsigned long commit_count
;
1298 config
->cb
.buffer_begin(buf
, tsc
, beginidx
);
1301 * Order all writes to buffer before the commit count update that will
1302 * determine that the subbuffer is full.
1304 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1306 * Must write slot data before incrementing commit count. This
1307 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1313 v_add(config
, config
->cb
.subbuffer_header_size(),
1314 &buf
->commit_hot
[beginidx
].cc
);
1315 commit_count
= v_read(config
, &buf
->commit_hot
[beginidx
].cc
);
1316 /* Check if the written buffer has to be delivered */
1317 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1318 commit_count
, beginidx
);
1319 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, beginidx
,
1320 offsets
->begin
, commit_count
,
1321 config
->cb
.subbuffer_header_size());
1325 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1327 * The only remaining threads could be the ones with pending commits. They will
1328 * have to do the deliver themselves.
1331 void lib_ring_buffer_switch_new_end(struct lib_ring_buffer
*buf
,
1332 struct channel
*chan
,
1333 struct switch_offsets
*offsets
,
1336 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1337 unsigned long endidx
= subbuf_index(offsets
->end
- 1, chan
);
1338 unsigned long commit_count
, padding_size
, data_size
;
1340 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1341 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1342 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
);
1345 * Order all writes to buffer before the commit count update that will
1346 * determine that the subbuffer is full.
1348 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1350 * Must write slot data before incrementing commit count. This
1351 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1357 v_add(config
, padding_size
, &buf
->commit_hot
[endidx
].cc
);
1358 commit_count
= v_read(config
, &buf
->commit_hot
[endidx
].cc
);
1359 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->end
- 1,
1360 commit_count
, endidx
);
1361 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, endidx
,
1362 offsets
->end
, commit_count
,
1369 * !0 if execution must be aborted.
1372 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1373 struct lib_ring_buffer
*buf
,
1374 struct channel
*chan
,
1375 struct switch_offsets
*offsets
,
1378 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1381 offsets
->begin
= v_read(config
, &buf
->offset
);
1382 offsets
->old
= offsets
->begin
;
1383 offsets
->switch_old_start
= 0;
1384 off
= subbuf_offset(offsets
->begin
, chan
);
1386 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1389 * Ensure we flush the header of an empty subbuffer when doing the
1390 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1391 * total data gathering duration even if there were no records saved
1392 * after the last buffer switch.
1393 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1394 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1395 * subbuffer header as appropriate.
1396 * The next record that reserves space will be responsible for
1397 * populating the following subbuffer header. We choose not to populate
1398 * the next subbuffer header here because we want to be able to use
1399 * SWITCH_ACTIVE for periodical buffer flush and CPU tick_nohz stop
1400 * buffer flush, which must guarantee that all the buffer content
1401 * (records and header timestamps) are visible to the reader. This is
1402 * required for quiescence guarantees for the fusion merge.
1404 if (mode
== SWITCH_FLUSH
|| off
> 0) {
1405 if (unlikely(off
== 0)) {
1407 * The client does not save any header information.
1408 * Don't switch empty subbuffer on finalize, because it
1409 * is invalid to deliver a completely empty subbuffer.
1411 if (!config
->cb
.subbuffer_header_size())
1414 * Need to write the subbuffer start header on finalize.
1416 offsets
->switch_old_start
= 1;
1418 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1420 return -1; /* we do not have to switch : buffer is empty */
1421 /* Note: old points to the next subbuf at offset 0 */
1422 offsets
->end
= offsets
->begin
;
1427 * Force a sub-buffer switch. This operation is completely reentrant : can be
1428 * called while tracing is active with absolutely no lock held.
1430 * Note, however, that as a v_cmpxchg is used for some atomic
1431 * operations, this function must be called from the CPU which owns the buffer
1432 * for a ACTIVE flush.
1434 void lib_ring_buffer_switch_slow(struct lib_ring_buffer
*buf
, enum switch_mode mode
)
1436 struct channel
*chan
= buf
->backend
.chan
;
1437 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1438 struct switch_offsets offsets
;
1439 unsigned long oldidx
;
1445 * Perform retryable operations.
1448 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1450 return; /* Switch not needed */
1451 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1455 * Atomically update last_tsc. This update races against concurrent
1456 * atomic updates, but the race will always cause supplementary full TSC
1457 * records, never the opposite (missing a full TSC record when it would
1460 save_last_tsc(config
, buf
, tsc
);
1463 * Push the reader if necessary
1465 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1467 oldidx
= subbuf_index(offsets
.old
, chan
);
1468 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
);
1471 * May need to populate header start on SWITCH_FLUSH.
1473 if (offsets
.switch_old_start
) {
1474 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
);
1475 offsets
.old
+= config
->cb
.subbuffer_header_size();
1479 * Switch old subbuffer.
1481 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
);
1483 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow
);
1488 * -ENOSPC if event size is too large for packet.
1489 * -ENOBUFS if there is currently not enough space in buffer for the event.
1490 * -EIO if data cannot be written into the buffer for any other reason.
1493 int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer
*buf
,
1494 struct channel
*chan
,
1495 struct switch_offsets
*offsets
,
1496 struct lib_ring_buffer_ctx
*ctx
)
1498 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1499 unsigned long reserve_commit_diff
;
1501 offsets
->begin
= v_read(config
, &buf
->offset
);
1502 offsets
->old
= offsets
->begin
;
1503 offsets
->switch_new_start
= 0;
1504 offsets
->switch_new_end
= 0;
1505 offsets
->switch_old_end
= 0;
1506 offsets
->pre_header_padding
= 0;
1508 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1509 if ((int64_t) ctx
->tsc
== -EIO
)
1512 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1513 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
1515 if (unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1516 offsets
->switch_new_start
= 1; /* For offsets->begin */
1518 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1520 &offsets
->pre_header_padding
,
1523 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1526 if (unlikely(subbuf_offset(offsets
->begin
, chan
) +
1527 offsets
->size
> chan
->backend
.subbuf_size
)) {
1528 offsets
->switch_old_end
= 1; /* For offsets->old */
1529 offsets
->switch_new_start
= 1; /* For offsets->begin */
1532 if (unlikely(offsets
->switch_new_start
)) {
1533 unsigned long sb_index
;
1536 * We are typically not filling the previous buffer completely.
1538 if (likely(offsets
->switch_old_end
))
1539 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1540 offsets
->begin
= offsets
->begin
1541 + config
->cb
.subbuffer_header_size();
1542 /* Test new buffer integrity */
1543 sb_index
= subbuf_index(offsets
->begin
, chan
);
1544 reserve_commit_diff
=
1545 (buf_trunc(offsets
->begin
, chan
)
1546 >> chan
->backend
.num_subbuf_order
)
1547 - ((unsigned long) v_read(config
,
1548 &buf
->commit_cold
[sb_index
].cc_sb
)
1549 & chan
->commit_count_mask
);
1550 if (likely(reserve_commit_diff
== 0)) {
1551 /* Next subbuffer not being written to. */
1552 if (unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1553 subbuf_trunc(offsets
->begin
, chan
)
1554 - subbuf_trunc((unsigned long)
1555 atomic_long_read(&buf
->consumed
), chan
)
1556 >= chan
->backend
.buf_size
)) {
1558 * We do not overwrite non consumed buffers
1559 * and we are full : record is lost.
1561 v_inc(config
, &buf
->records_lost_full
);
1565 * Next subbuffer not being written to, and we
1566 * are either in overwrite mode or the buffer is
1567 * not full. It's safe to write in this new
1573 * Next subbuffer reserve offset does not match the
1574 * commit offset. Drop record in producer-consumer and
1575 * overwrite mode. Caused by either a writer OOPS or too
1576 * many nested writes over a reserve/commit pair.
1578 v_inc(config
, &buf
->records_lost_wrap
);
1582 config
->cb
.record_header_size(config
, chan
,
1584 &offsets
->pre_header_padding
,
1587 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1590 if (unlikely(subbuf_offset(offsets
->begin
, chan
)
1591 + offsets
->size
> chan
->backend
.subbuf_size
)) {
1593 * Record too big for subbuffers, report error, don't
1594 * complete the sub-buffer switch.
1596 v_inc(config
, &buf
->records_lost_big
);
1600 * We just made a successful buffer switch and the
1601 * record fits in the new subbuffer. Let's write.
1606 * Record fits in the current buffer and we are not on a switch
1607 * boundary. It's safe to write.
1610 offsets
->end
= offsets
->begin
+ offsets
->size
;
1612 if (unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
1614 * The offset_end will fall at the very beginning of the next
1617 offsets
->switch_new_end
= 1; /* For offsets->begin */
1623 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1624 * @ctx: ring buffer context.
1626 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
1627 * -EIO for other errors, else returns 0.
1628 * It will take care of sub-buffer switching.
1630 int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx
*ctx
)
1632 struct channel
*chan
= ctx
->chan
;
1633 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1634 struct lib_ring_buffer
*buf
;
1635 struct switch_offsets offsets
;
1638 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
1639 buf
= per_cpu_ptr(chan
->backend
.buf
, ctx
->cpu
);
1641 buf
= chan
->backend
.buf
;
1647 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
1651 } while (unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
1656 * Atomically update last_tsc. This update races against concurrent
1657 * atomic updates, but the race will always cause supplementary full TSC
1658 * records, never the opposite (missing a full TSC record when it would
1661 save_last_tsc(config
, buf
, ctx
->tsc
);
1664 * Push the reader if necessary
1666 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
1669 * Clear noref flag for this subbuffer.
1671 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1672 subbuf_index(offsets
.end
- 1, chan
));
1675 * Switch old subbuffer if needed.
1677 if (unlikely(offsets
.switch_old_end
)) {
1678 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1679 subbuf_index(offsets
.old
- 1, chan
));
1680 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
);
1684 * Populate new subbuffer.
1686 if (unlikely(offsets
.switch_new_start
))
1687 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
);
1689 if (unlikely(offsets
.switch_new_end
))
1690 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
);
1692 ctx
->slot_size
= offsets
.size
;
1693 ctx
->pre_offset
= offsets
.begin
;
1694 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
1697 EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow
);