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.
41 #include <linux/delay.h>
42 #include <linux/module.h>
43 #include <linux/percpu.h>
45 #include "../../wrapper/ringbuffer/config.h"
46 #include "../../wrapper/ringbuffer/backend.h"
47 #include "../../wrapper/ringbuffer/frontend.h"
48 #include "../../wrapper/ringbuffer/iterator.h"
49 #include "../../wrapper/ringbuffer/nohz.h"
52 * Internal structure representing offsets to use at a sub-buffer switch.
54 struct switch_offsets
{
55 unsigned long begin
, end
, old
;
56 size_t pre_header_padding
, size
;
57 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
68 static ATOMIC_NOTIFIER_HEAD(tick_nohz_notifier
);
69 #endif /* CONFIG_NO_HZ */
71 static DEFINE_PER_CPU(spinlock_t
, ring_buffer_nohz_lock
);
73 DEFINE_PER_CPU(unsigned int, lib_ring_buffer_nesting
);
74 EXPORT_PER_CPU_SYMBOL(lib_ring_buffer_nesting
);
77 void lib_ring_buffer_print_errors(struct channel
*chan
,
78 struct lib_ring_buffer
*buf
, int cpu
);
81 * Must be called under cpu hotplug protection.
83 void lib_ring_buffer_free(struct lib_ring_buffer
*buf
)
85 struct channel
*chan
= buf
->backend
.chan
;
87 lib_ring_buffer_print_errors(chan
, buf
, buf
->backend
.cpu
);
88 kfree(buf
->commit_hot
);
89 kfree(buf
->commit_cold
);
91 lib_ring_buffer_backend_free(&buf
->backend
);
95 * lib_ring_buffer_reset - Reset ring buffer to initial values.
98 * Effectively empty the ring buffer. Should be called when the buffer is not
99 * used for writing. The ring buffer can be opened for reading, but the reader
100 * should not be using the iterator concurrently with reset. The previous
101 * current iterator record is reset.
103 void lib_ring_buffer_reset(struct lib_ring_buffer
*buf
)
105 struct channel
*chan
= buf
->backend
.chan
;
106 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
110 * Reset iterator first. It will put the subbuffer if it currently holds
113 lib_ring_buffer_iterator_reset(buf
);
114 v_set(config
, &buf
->offset
, 0);
115 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
116 v_set(config
, &buf
->commit_hot
[i
].cc
, 0);
117 v_set(config
, &buf
->commit_hot
[i
].seq
, 0);
118 v_set(config
, &buf
->commit_cold
[i
].cc_sb
, 0);
120 atomic_long_set(&buf
->consumed
, 0);
121 atomic_set(&buf
->record_disabled
, 0);
122 v_set(config
, &buf
->last_tsc
, 0);
123 lib_ring_buffer_backend_reset(&buf
->backend
);
124 /* Don't reset number of active readers */
125 v_set(config
, &buf
->records_lost_full
, 0);
126 v_set(config
, &buf
->records_lost_wrap
, 0);
127 v_set(config
, &buf
->records_lost_big
, 0);
128 v_set(config
, &buf
->records_count
, 0);
129 v_set(config
, &buf
->records_overrun
, 0);
132 EXPORT_SYMBOL_GPL(lib_ring_buffer_reset
);
135 * channel_reset - Reset channel to initial values.
138 * Effectively empty the channel. Should be called when the channel is not used
139 * for writing. The channel can be opened for reading, but the reader should not
140 * be using the iterator concurrently with reset. The previous current iterator
143 void channel_reset(struct channel
*chan
)
146 * Reset iterators first. Will put the subbuffer if held for reading.
148 channel_iterator_reset(chan
);
149 atomic_set(&chan
->record_disabled
, 0);
150 /* Don't reset commit_count_mask, still valid */
151 channel_backend_reset(&chan
->backend
);
152 /* Don't reset switch/read timer interval */
153 /* Don't reset notifiers and notifier enable bits */
154 /* Don't reset reader reference count */
156 EXPORT_SYMBOL_GPL(channel_reset
);
159 * Must be called under cpu hotplug protection.
161 int lib_ring_buffer_create(struct lib_ring_buffer
*buf
,
162 struct channel_backend
*chanb
, int cpu
)
164 const struct lib_ring_buffer_config
*config
= chanb
->config
;
165 struct channel
*chan
= container_of(chanb
, struct channel
, backend
);
166 void *priv
= chanb
->priv
;
167 size_t subbuf_header_size
;
171 /* Test for cpu hotplug */
172 if (buf
->backend
.allocated
)
176 * Paranoia: per cpu dynamic allocation is not officially documented as
177 * zeroing the memory, so let's do it here too, just in case.
179 memset(buf
, 0, sizeof(*buf
));
181 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
, cpu
);
186 kzalloc_node(ALIGN(sizeof(*buf
->commit_hot
)
187 * chan
->backend
.num_subbuf
,
188 1 << INTERNODE_CACHE_SHIFT
),
189 GFP_KERNEL
, cpu_to_node(max(cpu
, 0)));
190 if (!buf
->commit_hot
) {
196 kzalloc_node(ALIGN(sizeof(*buf
->commit_cold
)
197 * chan
->backend
.num_subbuf
,
198 1 << INTERNODE_CACHE_SHIFT
),
199 GFP_KERNEL
, cpu_to_node(max(cpu
, 0)));
200 if (!buf
->commit_cold
) {
205 init_waitqueue_head(&buf
->read_wait
);
206 raw_spin_lock_init(&buf
->raw_tick_nohz_spinlock
);
209 * Write the subbuffer header for first subbuffer so we know the total
210 * duration of data gathering.
212 subbuf_header_size
= config
->cb
.subbuffer_header_size();
213 v_set(config
, &buf
->offset
, subbuf_header_size
);
214 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_wsb
[0].id
);
215 tsc
= config
->cb
.ring_buffer_clock_read(buf
->backend
.chan
);
216 config
->cb
.buffer_begin(buf
, tsc
, 0);
217 v_add(config
, subbuf_header_size
, &buf
->commit_hot
[0].cc
);
219 if (config
->cb
.buffer_create
) {
220 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
);
226 * Ensure the buffer is ready before setting it to allocated and setting
228 * Used for cpu hotplug vs cpumask iteration.
231 buf
->backend
.allocated
= 1;
233 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
234 CHAN_WARN_ON(chan
, cpumask_test_cpu(cpu
,
235 chan
->backend
.cpumask
));
236 cpumask_set_cpu(cpu
, chan
->backend
.cpumask
);
243 kfree(buf
->commit_cold
);
245 kfree(buf
->commit_hot
);
247 lib_ring_buffer_backend_free(&buf
->backend
);
251 static void switch_buffer_timer(unsigned long data
)
253 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
254 struct channel
*chan
= buf
->backend
.chan
;
255 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
258 * Only flush buffers periodically if readers are active.
260 if (atomic_long_read(&buf
->active_readers
))
261 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
263 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
264 mod_timer_pinned(&buf
->switch_timer
,
265 jiffies
+ chan
->switch_timer_interval
);
267 mod_timer(&buf
->switch_timer
,
268 jiffies
+ chan
->switch_timer_interval
);
272 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
274 static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer
*buf
)
276 struct channel
*chan
= buf
->backend
.chan
;
277 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
279 if (!chan
->switch_timer_interval
|| buf
->switch_timer_enabled
)
281 init_timer(&buf
->switch_timer
);
282 buf
->switch_timer
.function
= switch_buffer_timer
;
283 buf
->switch_timer
.expires
= jiffies
+ chan
->switch_timer_interval
;
284 buf
->switch_timer
.data
= (unsigned long)buf
;
285 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
286 add_timer_on(&buf
->switch_timer
, buf
->backend
.cpu
);
288 add_timer(&buf
->switch_timer
);
289 buf
->switch_timer_enabled
= 1;
293 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
295 static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer
*buf
)
297 struct channel
*chan
= buf
->backend
.chan
;
299 if (!chan
->switch_timer_interval
|| !buf
->switch_timer_enabled
)
302 del_timer_sync(&buf
->switch_timer
);
303 buf
->switch_timer_enabled
= 0;
307 * Polling timer to check the channels for data.
309 static void read_buffer_timer(unsigned long data
)
311 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
312 struct channel
*chan
= buf
->backend
.chan
;
313 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
315 CHAN_WARN_ON(chan
, !buf
->backend
.allocated
);
317 if (atomic_long_read(&buf
->active_readers
)
318 && lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
319 wake_up_interruptible(&buf
->read_wait
);
320 wake_up_interruptible(&chan
->read_wait
);
323 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
324 mod_timer_pinned(&buf
->read_timer
,
325 jiffies
+ chan
->read_timer_interval
);
327 mod_timer(&buf
->read_timer
,
328 jiffies
+ chan
->read_timer_interval
);
332 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
334 static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer
*buf
)
336 struct channel
*chan
= buf
->backend
.chan
;
337 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
339 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
340 || !chan
->read_timer_interval
341 || buf
->read_timer_enabled
)
344 init_timer(&buf
->read_timer
);
345 buf
->read_timer
.function
= read_buffer_timer
;
346 buf
->read_timer
.expires
= jiffies
+ chan
->read_timer_interval
;
347 buf
->read_timer
.data
= (unsigned long)buf
;
349 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
350 add_timer_on(&buf
->read_timer
, buf
->backend
.cpu
);
352 add_timer(&buf
->read_timer
);
353 buf
->read_timer_enabled
= 1;
357 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
359 static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer
*buf
)
361 struct channel
*chan
= buf
->backend
.chan
;
362 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
364 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
365 || !chan
->read_timer_interval
366 || !buf
->read_timer_enabled
)
369 del_timer_sync(&buf
->read_timer
);
371 * do one more check to catch data that has been written in the last
374 if (lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
375 wake_up_interruptible(&buf
->read_wait
);
376 wake_up_interruptible(&chan
->read_wait
);
378 buf
->read_timer_enabled
= 0;
381 #ifdef CONFIG_HOTPLUG_CPU
383 * lib_ring_buffer_cpu_hp_callback - CPU hotplug callback
384 * @nb: notifier block
385 * @action: hotplug action to take
388 * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
391 int __cpuinit
lib_ring_buffer_cpu_hp_callback(struct notifier_block
*nb
,
392 unsigned long action
,
395 unsigned int cpu
= (unsigned long)hcpu
;
396 struct channel
*chan
= container_of(nb
, struct channel
,
398 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
, cpu
);
399 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
401 if (!chan
->cpu_hp_enable
)
404 CHAN_WARN_ON(chan
, config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
);
407 case CPU_DOWN_FAILED
:
408 case CPU_DOWN_FAILED_FROZEN
:
410 case CPU_ONLINE_FROZEN
:
411 wake_up_interruptible(&chan
->hp_wait
);
412 lib_ring_buffer_start_switch_timer(buf
);
413 lib_ring_buffer_start_read_timer(buf
);
416 case CPU_DOWN_PREPARE
:
417 case CPU_DOWN_PREPARE_FROZEN
:
418 lib_ring_buffer_stop_switch_timer(buf
);
419 lib_ring_buffer_stop_read_timer(buf
);
423 case CPU_DEAD_FROZEN
:
425 * Performing a buffer switch on a remote CPU. Performed by
426 * the CPU responsible for doing the hotunplug after the target
427 * CPU stopped running completely. Ensures that all data
428 * from that remote CPU is flushed.
430 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
439 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
441 * For per-cpu buffers, call the reader wakeups before switching the buffer, so
442 * that wake-up-tracing generated events are flushed before going idle (in
443 * tick_nohz). We test if the spinlock is locked to deal with the race where
444 * readers try to sample the ring buffer before we perform the switch. We let
445 * the readers retry in that case. If there is data in the buffer, the wake up
446 * is going to forbid the CPU running the reader thread from going idle.
448 static int notrace
ring_buffer_tick_nohz_callback(struct notifier_block
*nb
,
452 struct channel
*chan
= container_of(nb
, struct channel
,
454 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
455 struct lib_ring_buffer
*buf
;
456 int cpu
= smp_processor_id();
458 if (config
->alloc
!= RING_BUFFER_ALLOC_PER_CPU
) {
460 * We don't support keeping the system idle with global buffers
461 * and streaming active. In order to do so, we would need to
462 * sample a non-nohz-cpumask racelessly with the nohz updates
463 * without adding synchronization overhead to nohz. Leave this
464 * use-case out for now.
469 buf
= channel_get_ring_buffer(config
, chan
, cpu
);
471 case TICK_NOHZ_FLUSH
:
472 raw_spin_lock(&buf
->raw_tick_nohz_spinlock
);
473 if (config
->wakeup
== RING_BUFFER_WAKEUP_BY_TIMER
474 && chan
->read_timer_interval
475 && atomic_long_read(&buf
->active_readers
)
476 && (lib_ring_buffer_poll_deliver(config
, buf
, chan
)
477 || lib_ring_buffer_pending_data(config
, buf
, chan
))) {
478 wake_up_interruptible(&buf
->read_wait
);
479 wake_up_interruptible(&chan
->read_wait
);
481 if (chan
->switch_timer_interval
)
482 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
483 raw_spin_unlock(&buf
->raw_tick_nohz_spinlock
);
486 spin_lock(&__get_cpu_var(ring_buffer_nohz_lock
));
487 lib_ring_buffer_stop_switch_timer(buf
);
488 lib_ring_buffer_stop_read_timer(buf
);
489 spin_unlock(&__get_cpu_var(ring_buffer_nohz_lock
));
491 case TICK_NOHZ_RESTART
:
492 spin_lock(&__get_cpu_var(ring_buffer_nohz_lock
));
493 lib_ring_buffer_start_read_timer(buf
);
494 lib_ring_buffer_start_switch_timer(buf
);
495 spin_unlock(&__get_cpu_var(ring_buffer_nohz_lock
));
502 void notrace
lib_ring_buffer_tick_nohz_flush(void)
504 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_FLUSH
,
508 void notrace
lib_ring_buffer_tick_nohz_stop(void)
510 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_STOP
,
514 void notrace
lib_ring_buffer_tick_nohz_restart(void)
516 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_RESTART
,
519 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
524 static void channel_unregister_notifiers(struct channel
*chan
)
526 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
529 channel_iterator_unregister_notifiers(chan
);
530 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
533 * Remove the nohz notifier first, so we are certain we stop
536 atomic_notifier_chain_unregister(&tick_nohz_notifier
,
537 &chan
->tick_nohz_notifier
);
539 * ring_buffer_nohz_lock will not be needed below, because
540 * we just removed the notifiers, which were the only source of
543 #endif /* CONFIG_NO_HZ */
544 #ifdef CONFIG_HOTPLUG_CPU
546 chan
->cpu_hp_enable
= 0;
547 for_each_online_cpu(cpu
) {
548 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
550 lib_ring_buffer_stop_switch_timer(buf
);
551 lib_ring_buffer_stop_read_timer(buf
);
554 unregister_cpu_notifier(&chan
->cpu_hp_notifier
);
556 for_each_possible_cpu(cpu
) {
557 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
559 lib_ring_buffer_stop_switch_timer(buf
);
560 lib_ring_buffer_stop_read_timer(buf
);
564 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
566 lib_ring_buffer_stop_switch_timer(buf
);
567 lib_ring_buffer_stop_read_timer(buf
);
569 channel_backend_unregister_notifiers(&chan
->backend
);
572 static void channel_free(struct channel
*chan
)
574 channel_iterator_free(chan
);
575 channel_backend_free(&chan
->backend
);
580 * channel_create - Create channel.
581 * @config: ring buffer instance configuration
582 * @name: name of the channel
583 * @priv: ring buffer client private data
584 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
585 * address mapping. It is used only by RING_BUFFER_STATIC
586 * configuration. It can be set to NULL for other backends.
587 * @subbuf_size: subbuffer size
588 * @num_subbuf: number of subbuffers
589 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
590 * padding to let readers get those sub-buffers.
591 * Used for live streaming.
592 * @read_timer_interval: Time interval (in us) to wake up pending readers.
595 * Returns NULL on failure.
597 struct channel
*channel_create(const struct lib_ring_buffer_config
*config
,
598 const char *name
, void *priv
, void *buf_addr
,
600 size_t num_subbuf
, unsigned int switch_timer_interval
,
601 unsigned int read_timer_interval
)
604 struct channel
*chan
;
606 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
607 read_timer_interval
))
610 chan
= kzalloc(sizeof(struct channel
), GFP_KERNEL
);
614 ret
= channel_backend_init(&chan
->backend
, name
, config
, priv
,
615 subbuf_size
, num_subbuf
);
619 ret
= channel_iterator_init(chan
);
621 goto error_free_backend
;
623 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
624 chan
->switch_timer_interval
= usecs_to_jiffies(switch_timer_interval
);
625 chan
->read_timer_interval
= usecs_to_jiffies(read_timer_interval
);
626 kref_init(&chan
->ref
);
627 init_waitqueue_head(&chan
->read_wait
);
628 init_waitqueue_head(&chan
->hp_wait
);
630 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
631 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
632 /* Only benefit from NO_HZ idle with per-cpu buffers for now. */
633 chan
->tick_nohz_notifier
.notifier_call
=
634 ring_buffer_tick_nohz_callback
;
635 chan
->tick_nohz_notifier
.priority
= ~0U;
636 atomic_notifier_chain_register(&tick_nohz_notifier
,
637 &chan
->tick_nohz_notifier
);
638 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
641 * In case of non-hotplug cpu, if the ring-buffer is allocated
642 * in early initcall, it will not be notified of secondary cpus.
643 * In that off case, we need to allocate for all possible cpus.
645 #ifdef CONFIG_HOTPLUG_CPU
646 chan
->cpu_hp_notifier
.notifier_call
=
647 lib_ring_buffer_cpu_hp_callback
;
648 chan
->cpu_hp_notifier
.priority
= 6;
649 register_cpu_notifier(&chan
->cpu_hp_notifier
);
652 for_each_online_cpu(cpu
) {
653 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
655 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
656 lib_ring_buffer_start_switch_timer(buf
);
657 lib_ring_buffer_start_read_timer(buf
);
658 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
660 chan
->cpu_hp_enable
= 1;
663 for_each_possible_cpu(cpu
) {
664 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
666 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
667 lib_ring_buffer_start_switch_timer(buf
);
668 lib_ring_buffer_start_read_timer(buf
);
669 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
673 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
675 lib_ring_buffer_start_switch_timer(buf
);
676 lib_ring_buffer_start_read_timer(buf
);
682 channel_backend_free(&chan
->backend
);
687 EXPORT_SYMBOL_GPL(channel_create
);
690 void channel_release(struct kref
*kref
)
692 struct channel
*chan
= container_of(kref
, struct channel
, ref
);
697 * channel_destroy - Finalize, wait for q.s. and destroy channel.
698 * @chan: channel to destroy
701 * Call "destroy" callback, finalize channels, and then decrement the
702 * channel reference count. Note that when readers have completed data
703 * consumption of finalized channels, get_subbuf() will return -ENODATA.
704 * They should release their handle at that point. Returns the private
707 void *channel_destroy(struct channel
*chan
)
710 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
713 channel_unregister_notifiers(chan
);
715 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
717 * No need to hold cpu hotplug, because all notifiers have been
720 for_each_channel_cpu(cpu
, chan
) {
721 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
724 if (config
->cb
.buffer_finalize
)
725 config
->cb
.buffer_finalize(buf
,
728 if (buf
->backend
.allocated
)
729 lib_ring_buffer_switch_slow(buf
, SWITCH_FLUSH
);
731 * Perform flush before writing to finalized.
734 ACCESS_ONCE(buf
->finalized
) = 1;
735 wake_up_interruptible(&buf
->read_wait
);
738 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
740 if (config
->cb
.buffer_finalize
)
741 config
->cb
.buffer_finalize(buf
, chan
->backend
.priv
, -1);
742 if (buf
->backend
.allocated
)
743 lib_ring_buffer_switch_slow(buf
, SWITCH_FLUSH
);
745 * Perform flush before writing to finalized.
748 ACCESS_ONCE(buf
->finalized
) = 1;
749 wake_up_interruptible(&buf
->read_wait
);
751 ACCESS_ONCE(chan
->finalized
) = 1;
752 wake_up_interruptible(&chan
->hp_wait
);
753 wake_up_interruptible(&chan
->read_wait
);
754 kref_put(&chan
->ref
, channel_release
);
755 priv
= chan
->backend
.priv
;
758 EXPORT_SYMBOL_GPL(channel_destroy
);
760 struct lib_ring_buffer
*channel_get_ring_buffer(
761 const struct lib_ring_buffer_config
*config
,
762 struct channel
*chan
, int cpu
)
764 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
)
765 return chan
->backend
.buf
;
767 return per_cpu_ptr(chan
->backend
.buf
, cpu
);
769 EXPORT_SYMBOL_GPL(channel_get_ring_buffer
);
771 int lib_ring_buffer_open_read(struct lib_ring_buffer
*buf
)
773 struct channel
*chan
= buf
->backend
.chan
;
775 if (!atomic_long_add_unless(&buf
->active_readers
, 1, 1))
777 kref_get(&chan
->ref
);
778 smp_mb__after_atomic_inc();
781 EXPORT_SYMBOL_GPL(lib_ring_buffer_open_read
);
783 void lib_ring_buffer_release_read(struct lib_ring_buffer
*buf
)
785 struct channel
*chan
= buf
->backend
.chan
;
787 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
788 smp_mb__before_atomic_dec();
789 atomic_long_dec(&buf
->active_readers
);
790 kref_put(&chan
->ref
, channel_release
);
792 EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read
);
795 * Promote compiler barrier to a smp_mb().
796 * For the specific ring buffer case, this IPI call should be removed if the
797 * architecture does not reorder writes. This should eventually be provided by
798 * a separate architecture-specific infrastructure.
800 static void remote_mb(void *info
)
806 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
808 * @consumed: consumed count indicating the position where to read
809 * @produced: produced count, indicates position when to stop reading
811 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
812 * data to read at consumed position, or 0 if the get operation succeeds.
813 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
816 int lib_ring_buffer_snapshot(struct lib_ring_buffer
*buf
,
817 unsigned long *consumed
, unsigned long *produced
)
819 struct channel
*chan
= buf
->backend
.chan
;
820 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
821 unsigned long consumed_cur
, write_offset
;
825 finalized
= ACCESS_ONCE(buf
->finalized
);
827 * Read finalized before counters.
830 consumed_cur
= atomic_long_read(&buf
->consumed
);
832 * No need to issue a memory barrier between consumed count read and
833 * write offset read, because consumed count can only change
834 * concurrently in overwrite mode, and we keep a sequence counter
835 * identifier derived from the write offset to check we are getting
836 * the same sub-buffer we are expecting (the sub-buffers are atomically
837 * "tagged" upon writes, tags are checked upon read).
839 write_offset
= v_read(config
, &buf
->offset
);
842 * Check that we are not about to read the same subbuffer in
843 * which the writer head is.
845 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
849 *consumed
= consumed_cur
;
850 *produced
= subbuf_trunc(write_offset
, chan
);
856 * The memory barriers __wait_event()/wake_up_interruptible() take care
857 * of "raw_spin_is_locked" memory ordering.
861 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
866 EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot
);
869 * lib_ring_buffer_put_snapshot - move consumed counter forward
871 * @consumed_new: new consumed count value
873 void lib_ring_buffer_move_consumer(struct lib_ring_buffer
*buf
,
874 unsigned long consumed_new
)
876 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
877 struct channel
*chan
= bufb
->chan
;
878 unsigned long consumed
;
880 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
883 * Only push the consumed value forward.
884 * If the consumed cmpxchg fails, this is because we have been pushed by
885 * the writer in flight recorder mode.
887 consumed
= atomic_long_read(&buf
->consumed
);
888 while ((long) consumed
- (long) consumed_new
< 0)
889 consumed
= atomic_long_cmpxchg(&buf
->consumed
, consumed
,
892 EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer
);
895 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
897 * @consumed: consumed count indicating the position where to read
899 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
900 * data to read at consumed position, or 0 if the get operation succeeds.
901 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
903 int lib_ring_buffer_get_subbuf(struct lib_ring_buffer
*buf
,
904 unsigned long consumed
)
906 struct channel
*chan
= buf
->backend
.chan
;
907 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
908 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
913 finalized
= ACCESS_ONCE(buf
->finalized
);
915 * Read finalized before counters.
918 consumed_cur
= atomic_long_read(&buf
->consumed
);
919 consumed_idx
= subbuf_index(consumed
, chan
);
920 commit_count
= v_read(config
, &buf
->commit_cold
[consumed_idx
].cc_sb
);
922 * Make sure we read the commit count before reading the buffer
923 * data and the write offset. Correct consumed offset ordering
924 * wrt commit count is insured by the use of cmpxchg to update
925 * the consumed offset.
926 * smp_call_function_single can fail if the remote CPU is offline,
927 * this is OK because then there is no wmb to execute there.
928 * If our thread is executing on the same CPU as the on the buffers
929 * belongs to, we don't have to synchronize it at all. If we are
930 * migrated, the scheduler will take care of the memory barriers.
931 * Normally, smp_call_function_single() should ensure program order when
932 * executing the remote function, which implies that it surrounds the
933 * function execution with :
944 * However, smp_call_function_single() does not seem to clearly execute
945 * such barriers. It depends on spinlock semantic to provide the barrier
946 * before executing the IPI and, when busy-looping, csd_lock_wait only
947 * executes smp_mb() when it has to wait for the other CPU.
949 * I don't trust this code. Therefore, let's add the smp_mb() sequence
950 * required ourself, even if duplicated. It has no performance impact
953 * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs
954 * read and write vs write. They do not ensure core synchronization. We
955 * really have to ensure total order between the 3 barriers running on
958 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
959 if (config
->sync
== RING_BUFFER_SYNC_PER_CPU
960 && config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
961 if (raw_smp_processor_id() != buf
->backend
.cpu
) {
962 /* Total order with IPI handler smp_mb() */
964 smp_call_function_single(buf
->backend
.cpu
,
966 /* Total order with IPI handler smp_mb() */
970 /* Total order with IPI handler smp_mb() */
972 smp_call_function(remote_mb
, NULL
, 1);
973 /* Total order with IPI handler smp_mb() */
978 * Local rmb to match the remote wmb to read the commit count
979 * before the buffer data and the write offset.
984 write_offset
= v_read(config
, &buf
->offset
);
987 * Check that the buffer we are getting is after or at consumed_cur
990 if ((long) subbuf_trunc(consumed
, chan
)
991 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
995 * Check that the subbuffer we are trying to consume has been
996 * already fully committed.
998 if (((commit_count
- chan
->backend
.subbuf_size
)
999 & chan
->commit_count_mask
)
1000 - (buf_trunc(consumed_cur
, chan
)
1001 >> chan
->backend
.num_subbuf_order
)
1006 * Check that we are not about to read the same subbuffer in
1007 * which the writer head is.
1009 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
1014 * Failure to get the subbuffer causes a busy-loop retry without going
1015 * to a wait queue. These are caused by short-lived race windows where
1016 * the writer is getting access to a subbuffer we were trying to get
1017 * access to. Also checks that the "consumed" buffer count we are
1018 * looking for matches the one contained in the subbuffer id.
1020 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1021 consumed_idx
, buf_trunc_val(consumed
, chan
));
1024 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1026 buf
->get_subbuf_consumed
= consumed
;
1027 buf
->get_subbuf
= 1;
1033 * The memory barriers __wait_event()/wake_up_interruptible() take care
1034 * of "raw_spin_is_locked" memory ordering.
1038 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
1043 EXPORT_SYMBOL_GPL(lib_ring_buffer_get_subbuf
);
1046 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1049 void lib_ring_buffer_put_subbuf(struct lib_ring_buffer
*buf
)
1051 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1052 struct channel
*chan
= bufb
->chan
;
1053 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1054 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
1056 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
1058 if (!buf
->get_subbuf
) {
1060 * Reader puts a subbuffer it did not get.
1062 CHAN_WARN_ON(chan
, 1);
1065 consumed
= buf
->get_subbuf_consumed
;
1066 buf
->get_subbuf
= 0;
1069 * Clear the records_unread counter. (overruns counter)
1070 * Can still be non-zero if a file reader simply grabbed the data
1071 * without using iterators.
1072 * Can be below zero if an iterator is used on a snapshot more than
1075 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1076 v_add(config
, v_read(config
,
1077 &bufb
->array
[read_sb_bindex
]->records_unread
),
1078 &bufb
->records_read
);
1079 v_set(config
, &bufb
->array
[read_sb_bindex
]->records_unread
, 0);
1080 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1081 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1082 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1085 * Exchange the reader subbuffer with the one we put in its place in the
1086 * writer subbuffer table. Expect the original consumed count. If
1087 * update_read_sb_index fails, this is because the writer updated the
1088 * subbuffer concurrently. We should therefore keep the subbuffer we
1089 * currently have: it has become invalid to try reading this sub-buffer
1090 * consumed count value anyway.
1092 consumed_idx
= subbuf_index(consumed
, chan
);
1093 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1094 consumed_idx
, buf_trunc_val(consumed
, chan
));
1096 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1097 * if the writer concurrently updated it.
1100 EXPORT_SYMBOL_GPL(lib_ring_buffer_put_subbuf
);
1103 * cons_offset is an iterator on all subbuffer offsets between the reader
1104 * position and the writer position. (inclusive)
1107 void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer
*buf
,
1108 struct channel
*chan
,
1109 unsigned long cons_offset
,
1112 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1113 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1115 cons_idx
= subbuf_index(cons_offset
, chan
);
1116 commit_count
= v_read(config
, &buf
->commit_hot
[cons_idx
].cc
);
1117 commit_count_sb
= v_read(config
, &buf
->commit_cold
[cons_idx
].cc_sb
);
1119 if (subbuf_offset(commit_count
, chan
) != 0)
1121 "ring buffer %s, cpu %d: "
1122 "commit count in subbuffer %lu,\n"
1123 "expecting multiples of %lu bytes\n"
1124 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1125 chan
->backend
.name
, cpu
, cons_idx
,
1126 chan
->backend
.subbuf_size
,
1127 commit_count
, commit_count_sb
);
1129 printk(KERN_DEBUG
"ring buffer: %s, cpu %d: %lu bytes committed\n",
1130 chan
->backend
.name
, cpu
, commit_count
);
1134 void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer
*buf
,
1135 struct channel
*chan
,
1136 void *priv
, int cpu
)
1138 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1139 unsigned long write_offset
, cons_offset
;
1142 * Can be called in the error path of allocation when
1143 * trans_channel_data is not yet set.
1148 * No need to order commit_count, write_offset and cons_offset reads
1149 * because we execute at teardown when no more writer nor reader
1150 * references are left.
1152 write_offset
= v_read(config
, &buf
->offset
);
1153 cons_offset
= atomic_long_read(&buf
->consumed
);
1154 if (write_offset
!= cons_offset
)
1156 "ring buffer %s, cpu %d: "
1157 "non-consumed data\n"
1158 " [ %lu bytes written, %lu bytes read ]\n",
1159 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1161 for (cons_offset
= atomic_long_read(&buf
->consumed
);
1162 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1165 cons_offset
= subbuf_align(cons_offset
, chan
))
1166 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1171 void lib_ring_buffer_print_errors(struct channel
*chan
,
1172 struct lib_ring_buffer
*buf
, int cpu
)
1174 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1175 void *priv
= chan
->backend
.priv
;
1177 printk(KERN_DEBUG
"ring buffer %s, cpu %d: %lu records written, "
1178 "%lu records overrun\n",
1179 chan
->backend
.name
, cpu
,
1180 v_read(config
, &buf
->records_count
),
1181 v_read(config
, &buf
->records_overrun
));
1183 if (v_read(config
, &buf
->records_lost_full
)
1184 || v_read(config
, &buf
->records_lost_wrap
)
1185 || v_read(config
, &buf
->records_lost_big
))
1187 "ring buffer %s, cpu %d: records were lost. Caused by:\n"
1188 " [ %lu buffer full, %lu nest buffer wrap-around, "
1189 "%lu event too big ]\n",
1190 chan
->backend
.name
, cpu
,
1191 v_read(config
, &buf
->records_lost_full
),
1192 v_read(config
, &buf
->records_lost_wrap
),
1193 v_read(config
, &buf
->records_lost_big
));
1195 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
);
1199 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1201 * Only executed when the buffer is finalized, in SWITCH_FLUSH.
1204 void lib_ring_buffer_switch_old_start(struct lib_ring_buffer
*buf
,
1205 struct channel
*chan
,
1206 struct switch_offsets
*offsets
,
1209 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1210 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1211 unsigned long commit_count
;
1213 config
->cb
.buffer_begin(buf
, tsc
, oldidx
);
1216 * Order all writes to buffer before the commit count update that will
1217 * determine that the subbuffer is full.
1219 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1221 * Must write slot data before incrementing commit count. This
1222 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1228 v_add(config
, config
->cb
.subbuffer_header_size(),
1229 &buf
->commit_hot
[oldidx
].cc
);
1230 commit_count
= v_read(config
, &buf
->commit_hot
[oldidx
].cc
);
1231 /* Check if the written buffer has to be delivered */
1232 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1233 commit_count
, oldidx
);
1234 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1235 offsets
->old
, commit_count
,
1236 config
->cb
.subbuffer_header_size());
1240 * lib_ring_buffer_switch_old_end: switch old subbuffer
1242 * Note : offset_old should never be 0 here. It is ok, because we never perform
1243 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1244 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1248 void lib_ring_buffer_switch_old_end(struct lib_ring_buffer
*buf
,
1249 struct channel
*chan
,
1250 struct switch_offsets
*offsets
,
1253 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1254 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1255 unsigned long commit_count
, padding_size
, data_size
;
1257 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1258 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1259 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
);
1262 * Order all writes to buffer before the commit count update that will
1263 * determine that the subbuffer is full.
1265 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1267 * Must write slot data before incrementing commit count. This
1268 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1274 v_add(config
, padding_size
, &buf
->commit_hot
[oldidx
].cc
);
1275 commit_count
= v_read(config
, &buf
->commit_hot
[oldidx
].cc
);
1276 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1277 commit_count
, oldidx
);
1278 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1279 offsets
->old
, commit_count
,
1284 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1286 * This code can be executed unordered : writers may already have written to the
1287 * sub-buffer before this code gets executed, caution. The commit makes sure
1288 * that this code is executed before the deliver of this sub-buffer.
1291 void lib_ring_buffer_switch_new_start(struct lib_ring_buffer
*buf
,
1292 struct channel
*chan
,
1293 struct switch_offsets
*offsets
,
1296 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1297 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1298 unsigned long commit_count
;
1300 config
->cb
.buffer_begin(buf
, tsc
, beginidx
);
1303 * Order all writes to buffer before the commit count update that will
1304 * determine that the subbuffer is full.
1306 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1308 * Must write slot data before incrementing commit count. This
1309 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1315 v_add(config
, config
->cb
.subbuffer_header_size(),
1316 &buf
->commit_hot
[beginidx
].cc
);
1317 commit_count
= v_read(config
, &buf
->commit_hot
[beginidx
].cc
);
1318 /* Check if the written buffer has to be delivered */
1319 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1320 commit_count
, beginidx
);
1321 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, beginidx
,
1322 offsets
->begin
, commit_count
,
1323 config
->cb
.subbuffer_header_size());
1327 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1329 * The only remaining threads could be the ones with pending commits. They will
1330 * have to do the deliver themselves.
1333 void lib_ring_buffer_switch_new_end(struct lib_ring_buffer
*buf
,
1334 struct channel
*chan
,
1335 struct switch_offsets
*offsets
,
1338 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1339 unsigned long endidx
= subbuf_index(offsets
->end
- 1, chan
);
1340 unsigned long commit_count
, padding_size
, data_size
;
1342 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1343 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1344 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
);
1347 * Order all writes to buffer before the commit count update that will
1348 * determine that the subbuffer is full.
1350 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1352 * Must write slot data before incrementing commit count. This
1353 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1359 v_add(config
, padding_size
, &buf
->commit_hot
[endidx
].cc
);
1360 commit_count
= v_read(config
, &buf
->commit_hot
[endidx
].cc
);
1361 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->end
- 1,
1362 commit_count
, endidx
);
1363 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, endidx
,
1364 offsets
->end
, commit_count
,
1371 * !0 if execution must be aborted.
1374 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1375 struct lib_ring_buffer
*buf
,
1376 struct channel
*chan
,
1377 struct switch_offsets
*offsets
,
1380 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1383 offsets
->begin
= v_read(config
, &buf
->offset
);
1384 offsets
->old
= offsets
->begin
;
1385 offsets
->switch_old_start
= 0;
1386 off
= subbuf_offset(offsets
->begin
, chan
);
1388 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1391 * Ensure we flush the header of an empty subbuffer when doing the
1392 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1393 * total data gathering duration even if there were no records saved
1394 * after the last buffer switch.
1395 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1396 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1397 * subbuffer header as appropriate.
1398 * The next record that reserves space will be responsible for
1399 * populating the following subbuffer header. We choose not to populate
1400 * the next subbuffer header here because we want to be able to use
1401 * SWITCH_ACTIVE for periodical buffer flush and CPU tick_nohz stop
1402 * buffer flush, which must guarantee that all the buffer content
1403 * (records and header timestamps) are visible to the reader. This is
1404 * required for quiescence guarantees for the fusion merge.
1406 if (mode
== SWITCH_FLUSH
|| off
> 0) {
1407 if (unlikely(off
== 0)) {
1409 * The client does not save any header information.
1410 * Don't switch empty subbuffer on finalize, because it
1411 * is invalid to deliver a completely empty subbuffer.
1413 if (!config
->cb
.subbuffer_header_size())
1416 * Need to write the subbuffer start header on finalize.
1418 offsets
->switch_old_start
= 1;
1420 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1422 return -1; /* we do not have to switch : buffer is empty */
1423 /* Note: old points to the next subbuf at offset 0 */
1424 offsets
->end
= offsets
->begin
;
1429 * Force a sub-buffer switch. This operation is completely reentrant : can be
1430 * called while tracing is active with absolutely no lock held.
1432 * Note, however, that as a v_cmpxchg is used for some atomic
1433 * operations, this function must be called from the CPU which owns the buffer
1434 * for a ACTIVE flush.
1436 void lib_ring_buffer_switch_slow(struct lib_ring_buffer
*buf
, enum switch_mode mode
)
1438 struct channel
*chan
= buf
->backend
.chan
;
1439 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1440 struct switch_offsets offsets
;
1441 unsigned long oldidx
;
1447 * Perform retryable operations.
1450 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1452 return; /* Switch not needed */
1453 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1457 * Atomically update last_tsc. This update races against concurrent
1458 * atomic updates, but the race will always cause supplementary full TSC
1459 * records, never the opposite (missing a full TSC record when it would
1462 save_last_tsc(config
, buf
, tsc
);
1465 * Push the reader if necessary
1467 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1469 oldidx
= subbuf_index(offsets
.old
, chan
);
1470 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
);
1473 * May need to populate header start on SWITCH_FLUSH.
1475 if (offsets
.switch_old_start
) {
1476 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
);
1477 offsets
.old
+= config
->cb
.subbuffer_header_size();
1481 * Switch old subbuffer.
1483 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
);
1485 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow
);
1490 * -ENOSPC if event size is too large for packet.
1491 * -ENOBUFS if there is currently not enough space in buffer for the event.
1492 * -EIO if data cannot be written into the buffer for any other reason.
1495 int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer
*buf
,
1496 struct channel
*chan
,
1497 struct switch_offsets
*offsets
,
1498 struct lib_ring_buffer_ctx
*ctx
)
1500 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1501 unsigned long reserve_commit_diff
;
1503 offsets
->begin
= v_read(config
, &buf
->offset
);
1504 offsets
->old
= offsets
->begin
;
1505 offsets
->switch_new_start
= 0;
1506 offsets
->switch_new_end
= 0;
1507 offsets
->switch_old_end
= 0;
1508 offsets
->pre_header_padding
= 0;
1510 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1511 if ((int64_t) ctx
->tsc
== -EIO
)
1514 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1515 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
1517 if (unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1518 offsets
->switch_new_start
= 1; /* For offsets->begin */
1520 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1522 &offsets
->pre_header_padding
,
1525 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1528 if (unlikely(subbuf_offset(offsets
->begin
, chan
) +
1529 offsets
->size
> chan
->backend
.subbuf_size
)) {
1530 offsets
->switch_old_end
= 1; /* For offsets->old */
1531 offsets
->switch_new_start
= 1; /* For offsets->begin */
1534 if (unlikely(offsets
->switch_new_start
)) {
1535 unsigned long sb_index
;
1538 * We are typically not filling the previous buffer completely.
1540 if (likely(offsets
->switch_old_end
))
1541 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1542 offsets
->begin
= offsets
->begin
1543 + config
->cb
.subbuffer_header_size();
1544 /* Test new buffer integrity */
1545 sb_index
= subbuf_index(offsets
->begin
, chan
);
1546 reserve_commit_diff
=
1547 (buf_trunc(offsets
->begin
, chan
)
1548 >> chan
->backend
.num_subbuf_order
)
1549 - ((unsigned long) v_read(config
,
1550 &buf
->commit_cold
[sb_index
].cc_sb
)
1551 & chan
->commit_count_mask
);
1552 if (likely(reserve_commit_diff
== 0)) {
1553 /* Next subbuffer not being written to. */
1554 if (unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1555 subbuf_trunc(offsets
->begin
, chan
)
1556 - subbuf_trunc((unsigned long)
1557 atomic_long_read(&buf
->consumed
), chan
)
1558 >= chan
->backend
.buf_size
)) {
1560 * We do not overwrite non consumed buffers
1561 * and we are full : record is lost.
1563 v_inc(config
, &buf
->records_lost_full
);
1567 * Next subbuffer not being written to, and we
1568 * are either in overwrite mode or the buffer is
1569 * not full. It's safe to write in this new
1575 * Next subbuffer reserve offset does not match the
1576 * commit offset. Drop record in producer-consumer and
1577 * overwrite mode. Caused by either a writer OOPS or too
1578 * many nested writes over a reserve/commit pair.
1580 v_inc(config
, &buf
->records_lost_wrap
);
1584 config
->cb
.record_header_size(config
, chan
,
1586 &offsets
->pre_header_padding
,
1589 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1592 if (unlikely(subbuf_offset(offsets
->begin
, chan
)
1593 + offsets
->size
> chan
->backend
.subbuf_size
)) {
1595 * Record too big for subbuffers, report error, don't
1596 * complete the sub-buffer switch.
1598 v_inc(config
, &buf
->records_lost_big
);
1602 * We just made a successful buffer switch and the
1603 * record fits in the new subbuffer. Let's write.
1608 * Record fits in the current buffer and we are not on a switch
1609 * boundary. It's safe to write.
1612 offsets
->end
= offsets
->begin
+ offsets
->size
;
1614 if (unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
1616 * The offset_end will fall at the very beginning of the next
1619 offsets
->switch_new_end
= 1; /* For offsets->begin */
1625 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1626 * @ctx: ring buffer context.
1628 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
1629 * -EIO for other errors, else returns 0.
1630 * It will take care of sub-buffer switching.
1632 int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx
*ctx
)
1634 struct channel
*chan
= ctx
->chan
;
1635 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1636 struct lib_ring_buffer
*buf
;
1637 struct switch_offsets offsets
;
1640 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
1641 buf
= per_cpu_ptr(chan
->backend
.buf
, ctx
->cpu
);
1643 buf
= chan
->backend
.buf
;
1649 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
1653 } while (unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
1658 * Atomically update last_tsc. This update races against concurrent
1659 * atomic updates, but the race will always cause supplementary full TSC
1660 * records, never the opposite (missing a full TSC record when it would
1663 save_last_tsc(config
, buf
, ctx
->tsc
);
1666 * Push the reader if necessary
1668 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
1671 * Clear noref flag for this subbuffer.
1673 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1674 subbuf_index(offsets
.end
- 1, chan
));
1677 * Switch old subbuffer if needed.
1679 if (unlikely(offsets
.switch_old_end
)) {
1680 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1681 subbuf_index(offsets
.old
- 1, chan
));
1682 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
);
1686 * Populate new subbuffer.
1688 if (unlikely(offsets
.switch_new_start
))
1689 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
);
1691 if (unlikely(offsets
.switch_new_end
))
1692 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
);
1694 ctx
->slot_size
= offsets
.size
;
1695 ctx
->pre_offset
= offsets
.begin
;
1696 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
1699 EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow
);