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 unsigned int num_subbuf
;
168 size_t subbuf_header_size
;
172 /* Test for cpu hotplug */
173 if (buf
->backend
.allocated
)
177 * Paranoia: per cpu dynamic allocation is not officially documented as
178 * zeroing the memory, so let's do it here too, just in case.
180 memset(buf
, 0, sizeof(*buf
));
182 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
, cpu
);
187 kzalloc_node(ALIGN(sizeof(*buf
->commit_hot
)
188 * chan
->backend
.num_subbuf
,
189 1 << INTERNODE_CACHE_SHIFT
),
190 GFP_KERNEL
, cpu_to_node(max(cpu
, 0)));
191 if (!buf
->commit_hot
) {
197 kzalloc_node(ALIGN(sizeof(*buf
->commit_cold
)
198 * chan
->backend
.num_subbuf
,
199 1 << INTERNODE_CACHE_SHIFT
),
200 GFP_KERNEL
, cpu_to_node(max(cpu
, 0)));
201 if (!buf
->commit_cold
) {
206 num_subbuf
= chan
->backend
.num_subbuf
;
207 init_waitqueue_head(&buf
->read_wait
);
208 raw_spin_lock_init(&buf
->raw_tick_nohz_spinlock
);
211 * Write the subbuffer header for first subbuffer so we know the total
212 * duration of data gathering.
214 subbuf_header_size
= config
->cb
.subbuffer_header_size();
215 v_set(config
, &buf
->offset
, subbuf_header_size
);
216 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_wsb
[0].id
);
217 tsc
= config
->cb
.ring_buffer_clock_read(buf
->backend
.chan
);
218 config
->cb
.buffer_begin(buf
, tsc
, 0);
219 v_add(config
, subbuf_header_size
, &buf
->commit_hot
[0].cc
);
221 if (config
->cb
.buffer_create
) {
222 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
);
228 * Ensure the buffer is ready before setting it to allocated and setting
230 * Used for cpu hotplug vs cpumask iteration.
233 buf
->backend
.allocated
= 1;
235 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
236 CHAN_WARN_ON(chan
, cpumask_test_cpu(cpu
,
237 chan
->backend
.cpumask
));
238 cpumask_set_cpu(cpu
, chan
->backend
.cpumask
);
245 kfree(buf
->commit_cold
);
247 kfree(buf
->commit_hot
);
249 lib_ring_buffer_backend_free(&buf
->backend
);
253 static void switch_buffer_timer(unsigned long data
)
255 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
256 struct channel
*chan
= buf
->backend
.chan
;
257 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
260 * Only flush buffers periodically if readers are active.
262 if (atomic_long_read(&buf
->active_readers
))
263 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
265 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
266 mod_timer_pinned(&buf
->switch_timer
,
267 jiffies
+ chan
->switch_timer_interval
);
269 mod_timer(&buf
->switch_timer
,
270 jiffies
+ chan
->switch_timer_interval
);
274 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
276 static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer
*buf
)
278 struct channel
*chan
= buf
->backend
.chan
;
279 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
281 if (!chan
->switch_timer_interval
|| buf
->switch_timer_enabled
)
283 init_timer(&buf
->switch_timer
);
284 buf
->switch_timer
.function
= switch_buffer_timer
;
285 buf
->switch_timer
.expires
= jiffies
+ chan
->switch_timer_interval
;
286 buf
->switch_timer
.data
= (unsigned long)buf
;
287 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
288 add_timer_on(&buf
->switch_timer
, buf
->backend
.cpu
);
290 add_timer(&buf
->switch_timer
);
291 buf
->switch_timer_enabled
= 1;
295 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
297 static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer
*buf
)
299 struct channel
*chan
= buf
->backend
.chan
;
301 if (!chan
->switch_timer_interval
|| !buf
->switch_timer_enabled
)
304 del_timer_sync(&buf
->switch_timer
);
305 buf
->switch_timer_enabled
= 0;
309 * Polling timer to check the channels for data.
311 static void read_buffer_timer(unsigned long data
)
313 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
314 struct channel
*chan
= buf
->backend
.chan
;
315 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
317 CHAN_WARN_ON(chan
, !buf
->backend
.allocated
);
319 if (atomic_long_read(&buf
->active_readers
)
320 && lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
321 wake_up_interruptible(&buf
->read_wait
);
322 wake_up_interruptible(&chan
->read_wait
);
325 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
326 mod_timer_pinned(&buf
->read_timer
,
327 jiffies
+ chan
->read_timer_interval
);
329 mod_timer(&buf
->read_timer
,
330 jiffies
+ chan
->read_timer_interval
);
334 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
336 static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer
*buf
)
338 struct channel
*chan
= buf
->backend
.chan
;
339 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
341 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
342 || !chan
->read_timer_interval
343 || buf
->read_timer_enabled
)
346 init_timer(&buf
->read_timer
);
347 buf
->read_timer
.function
= read_buffer_timer
;
348 buf
->read_timer
.expires
= jiffies
+ chan
->read_timer_interval
;
349 buf
->read_timer
.data
= (unsigned long)buf
;
351 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
352 add_timer_on(&buf
->read_timer
, buf
->backend
.cpu
);
354 add_timer(&buf
->read_timer
);
355 buf
->read_timer_enabled
= 1;
359 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
361 static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer
*buf
)
363 struct channel
*chan
= buf
->backend
.chan
;
364 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
366 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
367 || !chan
->read_timer_interval
368 || !buf
->read_timer_enabled
)
371 del_timer_sync(&buf
->read_timer
);
373 * do one more check to catch data that has been written in the last
376 if (lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
377 wake_up_interruptible(&buf
->read_wait
);
378 wake_up_interruptible(&chan
->read_wait
);
380 buf
->read_timer_enabled
= 0;
383 #ifdef CONFIG_HOTPLUG_CPU
385 * lib_ring_buffer_cpu_hp_callback - CPU hotplug callback
386 * @nb: notifier block
387 * @action: hotplug action to take
390 * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
393 int __cpuinit
lib_ring_buffer_cpu_hp_callback(struct notifier_block
*nb
,
394 unsigned long action
,
397 unsigned int cpu
= (unsigned long)hcpu
;
398 struct channel
*chan
= container_of(nb
, struct channel
,
400 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
, cpu
);
401 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
403 if (!chan
->cpu_hp_enable
)
406 CHAN_WARN_ON(chan
, config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
);
409 case CPU_DOWN_FAILED
:
410 case CPU_DOWN_FAILED_FROZEN
:
412 case CPU_ONLINE_FROZEN
:
413 lib_ring_buffer_start_switch_timer(buf
);
414 lib_ring_buffer_start_read_timer(buf
);
417 case CPU_DOWN_PREPARE
:
418 case CPU_DOWN_PREPARE_FROZEN
:
419 lib_ring_buffer_stop_switch_timer(buf
);
420 lib_ring_buffer_stop_read_timer(buf
);
424 case CPU_DEAD_FROZEN
:
426 * Performing a buffer switch on a remote CPU. Performed by
427 * the CPU responsible for doing the hotunplug after the target
428 * CPU stopped running completely. Ensures that all data
429 * from that remote CPU is flushed.
431 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
440 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
442 * For per-cpu buffers, call the reader wakeups before switching the buffer, so
443 * that wake-up-tracing generated events are flushed before going idle (in
444 * tick_nohz). We test if the spinlock is locked to deal with the race where
445 * readers try to sample the ring buffer before we perform the switch. We let
446 * the readers retry in that case. If there is data in the buffer, the wake up
447 * is going to forbid the CPU running the reader thread from going idle.
449 static int notrace
ring_buffer_tick_nohz_callback(struct notifier_block
*nb
,
453 struct channel
*chan
= container_of(nb
, struct channel
,
455 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
456 struct lib_ring_buffer
*buf
;
457 int cpu
= smp_processor_id();
459 if (config
->alloc
!= RING_BUFFER_ALLOC_PER_CPU
) {
461 * We don't support keeping the system idle with global buffers
462 * and streaming active. In order to do so, we would need to
463 * sample a non-nohz-cpumask racelessly with the nohz updates
464 * without adding synchronization overhead to nohz. Leave this
465 * use-case out for now.
470 buf
= channel_get_ring_buffer(config
, chan
, cpu
);
472 case TICK_NOHZ_FLUSH
:
473 raw_spin_lock(&buf
->raw_tick_nohz_spinlock
);
474 if (config
->wakeup
== RING_BUFFER_WAKEUP_BY_TIMER
475 && chan
->read_timer_interval
476 && atomic_long_read(&buf
->active_readers
)
477 && (lib_ring_buffer_poll_deliver(config
, buf
, chan
)
478 || lib_ring_buffer_pending_data(config
, buf
, chan
))) {
479 wake_up_interruptible(&buf
->read_wait
);
480 wake_up_interruptible(&chan
->read_wait
);
482 if (chan
->switch_timer_interval
)
483 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
484 raw_spin_unlock(&buf
->raw_tick_nohz_spinlock
);
487 spin_lock(&__get_cpu_var(ring_buffer_nohz_lock
));
488 lib_ring_buffer_stop_switch_timer(buf
);
489 lib_ring_buffer_stop_read_timer(buf
);
490 spin_unlock(&__get_cpu_var(ring_buffer_nohz_lock
));
492 case TICK_NOHZ_RESTART
:
493 spin_lock(&__get_cpu_var(ring_buffer_nohz_lock
));
494 lib_ring_buffer_start_read_timer(buf
);
495 lib_ring_buffer_start_switch_timer(buf
);
496 spin_unlock(&__get_cpu_var(ring_buffer_nohz_lock
));
503 void notrace
lib_ring_buffer_tick_nohz_flush(void)
505 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_FLUSH
,
509 void notrace
lib_ring_buffer_tick_nohz_stop(void)
511 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_STOP
,
515 void notrace
lib_ring_buffer_tick_nohz_restart(void)
517 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_RESTART
,
520 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
525 static void channel_unregister_notifiers(struct channel
*chan
)
527 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
530 channel_iterator_unregister_notifiers(chan
);
531 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
534 * Remove the nohz notifier first, so we are certain we stop
537 atomic_notifier_chain_unregister(&tick_nohz_notifier
,
538 &chan
->tick_nohz_notifier
);
540 * ring_buffer_nohz_lock will not be needed below, because
541 * we just removed the notifiers, which were the only source of
544 #endif /* CONFIG_NO_HZ */
545 #ifdef CONFIG_HOTPLUG_CPU
547 chan
->cpu_hp_enable
= 0;
548 for_each_online_cpu(cpu
) {
549 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
551 lib_ring_buffer_stop_switch_timer(buf
);
552 lib_ring_buffer_stop_read_timer(buf
);
555 unregister_cpu_notifier(&chan
->cpu_hp_notifier
);
557 for_each_possible_cpu(cpu
) {
558 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
560 lib_ring_buffer_stop_switch_timer(buf
);
561 lib_ring_buffer_stop_read_timer(buf
);
565 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
567 lib_ring_buffer_stop_switch_timer(buf
);
568 lib_ring_buffer_stop_read_timer(buf
);
570 channel_backend_unregister_notifiers(&chan
->backend
);
573 static void channel_free(struct channel
*chan
)
575 channel_iterator_free(chan
);
576 channel_backend_free(&chan
->backend
);
581 * channel_create - Create channel.
582 * @config: ring buffer instance configuration
583 * @name: name of the channel
584 * @priv: ring buffer client private data
585 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
586 * address mapping. It is used only by RING_BUFFER_STATIC
587 * configuration. It can be set to NULL for other backends.
588 * @subbuf_size: subbuffer size
589 * @num_subbuf: number of subbuffers
590 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
591 * padding to let readers get those sub-buffers.
592 * Used for live streaming.
593 * @read_timer_interval: Time interval (in us) to wake up pending readers.
596 * Returns NULL on failure.
598 struct channel
*channel_create(const struct lib_ring_buffer_config
*config
,
599 const char *name
, void *priv
, void *buf_addr
,
601 size_t num_subbuf
, unsigned int switch_timer_interval
,
602 unsigned int read_timer_interval
)
605 struct channel
*chan
;
607 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
608 read_timer_interval
))
611 chan
= kzalloc(sizeof(struct channel
), GFP_KERNEL
);
615 ret
= channel_backend_init(&chan
->backend
, name
, config
, priv
,
616 subbuf_size
, num_subbuf
);
620 ret
= channel_iterator_init(chan
);
622 goto error_free_backend
;
624 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
625 chan
->switch_timer_interval
= usecs_to_jiffies(switch_timer_interval
);
626 chan
->read_timer_interval
= usecs_to_jiffies(read_timer_interval
);
627 init_waitqueue_head(&chan
->read_wait
);
629 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
630 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
631 /* Only benefit from NO_HZ idle with per-cpu buffers for now. */
632 chan
->tick_nohz_notifier
.notifier_call
=
633 ring_buffer_tick_nohz_callback
;
634 chan
->tick_nohz_notifier
.priority
= ~0U;
635 atomic_notifier_chain_register(&tick_nohz_notifier
,
636 &chan
->tick_nohz_notifier
);
637 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
640 * In case of non-hotplug cpu, if the ring-buffer is allocated
641 * in early initcall, it will not be notified of secondary cpus.
642 * In that off case, we need to allocate for all possible cpus.
644 #ifdef CONFIG_HOTPLUG_CPU
645 chan
->cpu_hp_notifier
.notifier_call
=
646 lib_ring_buffer_cpu_hp_callback
;
647 chan
->cpu_hp_notifier
.priority
= 6;
648 register_cpu_notifier(&chan
->cpu_hp_notifier
);
651 for_each_online_cpu(cpu
) {
652 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
654 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
655 lib_ring_buffer_start_switch_timer(buf
);
656 lib_ring_buffer_start_read_timer(buf
);
657 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
659 chan
->cpu_hp_enable
= 1;
662 for_each_possible_cpu(cpu
) {
663 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
665 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
666 lib_ring_buffer_start_switch_timer(buf
);
667 lib_ring_buffer_start_read_timer(buf
);
668 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
672 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
674 lib_ring_buffer_start_switch_timer(buf
);
675 lib_ring_buffer_start_read_timer(buf
);
681 channel_backend_free(&chan
->backend
);
686 EXPORT_SYMBOL_GPL(channel_create
);
689 * channel_destroy - Finalize, wait for q.s. and destroy channel.
690 * @chan: channel to destroy
693 * Call "destroy" callback, finalize channels, wait for readers to release their
694 * reference, then destroy ring buffer data. Note that when readers have
695 * completed data consumption of finalized channels, get_subbuf() will return
696 * -ENODATA. They should release their handle at that point.
697 * Returns the private data pointer.
699 void *channel_destroy(struct channel
*chan
)
702 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
705 channel_unregister_notifiers(chan
);
707 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
709 * No need to hold cpu hotplug, because all notifiers have been
712 for_each_channel_cpu(cpu
, chan
) {
713 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
716 if (config
->cb
.buffer_finalize
)
717 config
->cb
.buffer_finalize(buf
,
720 if (buf
->backend
.allocated
)
721 lib_ring_buffer_switch_slow(buf
, SWITCH_FLUSH
);
723 * Perform flush before writing to finalized.
726 ACCESS_ONCE(buf
->finalized
) = 1;
727 wake_up_interruptible(&buf
->read_wait
);
730 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
732 if (config
->cb
.buffer_finalize
)
733 config
->cb
.buffer_finalize(buf
, chan
->backend
.priv
, -1);
734 if (buf
->backend
.allocated
)
735 lib_ring_buffer_switch_slow(buf
, SWITCH_FLUSH
);
737 * Perform flush before writing to finalized.
740 ACCESS_ONCE(buf
->finalized
) = 1;
741 wake_up_interruptible(&buf
->read_wait
);
743 wake_up_interruptible(&chan
->read_wait
);
745 while (atomic_long_read(&chan
->read_ref
) > 0)
747 /* Finish waiting for refcount before free */
749 priv
= chan
->backend
.priv
;
753 EXPORT_SYMBOL_GPL(channel_destroy
);
755 struct lib_ring_buffer
*channel_get_ring_buffer(
756 const struct lib_ring_buffer_config
*config
,
757 struct channel
*chan
, int cpu
)
759 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
)
760 return chan
->backend
.buf
;
762 return per_cpu_ptr(chan
->backend
.buf
, cpu
);
764 EXPORT_SYMBOL_GPL(channel_get_ring_buffer
);
766 int lib_ring_buffer_open_read(struct lib_ring_buffer
*buf
)
768 struct channel
*chan
= buf
->backend
.chan
;
770 if (!atomic_long_add_unless(&buf
->active_readers
, 1, 1))
772 atomic_long_inc(&chan
->read_ref
);
773 smp_mb__after_atomic_inc();
776 EXPORT_SYMBOL_GPL(lib_ring_buffer_open_read
);
778 void lib_ring_buffer_release_read(struct lib_ring_buffer
*buf
)
780 struct channel
*chan
= buf
->backend
.chan
;
782 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
783 smp_mb__before_atomic_dec();
784 atomic_long_dec(&chan
->read_ref
);
785 atomic_long_dec(&buf
->active_readers
);
787 EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read
);
790 * Promote compiler barrier to a smp_mb().
791 * For the specific ring buffer case, this IPI call should be removed if the
792 * architecture does not reorder writes. This should eventually be provided by
793 * a separate architecture-specific infrastructure.
795 static void remote_mb(void *info
)
801 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
803 * @consumed: consumed count indicating the position where to read
804 * @produced: produced count, indicates position when to stop reading
806 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
807 * data to read at consumed position, or 0 if the get operation succeeds.
808 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
811 int lib_ring_buffer_snapshot(struct lib_ring_buffer
*buf
,
812 unsigned long *consumed
, unsigned long *produced
)
814 struct channel
*chan
= buf
->backend
.chan
;
815 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
816 unsigned long consumed_cur
, write_offset
;
820 finalized
= ACCESS_ONCE(buf
->finalized
);
822 * Read finalized before counters.
825 consumed_cur
= atomic_long_read(&buf
->consumed
);
827 * No need to issue a memory barrier between consumed count read and
828 * write offset read, because consumed count can only change
829 * concurrently in overwrite mode, and we keep a sequence counter
830 * identifier derived from the write offset to check we are getting
831 * the same sub-buffer we are expecting (the sub-buffers are atomically
832 * "tagged" upon writes, tags are checked upon read).
834 write_offset
= v_read(config
, &buf
->offset
);
837 * Check that we are not about to read the same subbuffer in
838 * which the writer head is.
840 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
844 *consumed
= consumed_cur
;
845 *produced
= subbuf_trunc(write_offset
, chan
);
851 * The memory barriers __wait_event()/wake_up_interruptible() take care
852 * of "raw_spin_is_locked" memory ordering.
856 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
861 EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot
);
864 * lib_ring_buffer_put_snapshot - move consumed counter forward
866 * @consumed_new: new consumed count value
868 void lib_ring_buffer_move_consumer(struct lib_ring_buffer
*buf
,
869 unsigned long consumed_new
)
871 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
872 struct channel
*chan
= bufb
->chan
;
873 unsigned long consumed
;
875 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
878 * Only push the consumed value forward.
879 * If the consumed cmpxchg fails, this is because we have been pushed by
880 * the writer in flight recorder mode.
882 consumed
= atomic_long_read(&buf
->consumed
);
883 while ((long) consumed
- (long) consumed_new
< 0)
884 consumed
= atomic_long_cmpxchg(&buf
->consumed
, consumed
,
887 EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer
);
890 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
892 * @consumed: consumed count indicating the position where to read
894 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
895 * data to read at consumed position, or 0 if the get operation succeeds.
896 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
898 int lib_ring_buffer_get_subbuf(struct lib_ring_buffer
*buf
,
899 unsigned long consumed
)
901 struct channel
*chan
= buf
->backend
.chan
;
902 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
903 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
908 finalized
= ACCESS_ONCE(buf
->finalized
);
910 * Read finalized before counters.
913 consumed_cur
= atomic_long_read(&buf
->consumed
);
914 consumed_idx
= subbuf_index(consumed
, chan
);
915 commit_count
= v_read(config
, &buf
->commit_cold
[consumed_idx
].cc_sb
);
917 * Make sure we read the commit count before reading the buffer
918 * data and the write offset. Correct consumed offset ordering
919 * wrt commit count is insured by the use of cmpxchg to update
920 * the consumed offset.
921 * smp_call_function_single can fail if the remote CPU is offline,
922 * this is OK because then there is no wmb to execute there.
923 * If our thread is executing on the same CPU as the on the buffers
924 * belongs to, we don't have to synchronize it at all. If we are
925 * migrated, the scheduler will take care of the memory barriers.
926 * Normally, smp_call_function_single() should ensure program order when
927 * executing the remote function, which implies that it surrounds the
928 * function execution with :
939 * However, smp_call_function_single() does not seem to clearly execute
940 * such barriers. It depends on spinlock semantic to provide the barrier
941 * before executing the IPI and, when busy-looping, csd_lock_wait only
942 * executes smp_mb() when it has to wait for the other CPU.
944 * I don't trust this code. Therefore, let's add the smp_mb() sequence
945 * required ourself, even if duplicated. It has no performance impact
948 * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs
949 * read and write vs write. They do not ensure core synchronization. We
950 * really have to ensure total order between the 3 barriers running on
953 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
954 if (config
->sync
== RING_BUFFER_SYNC_PER_CPU
955 && config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
956 if (raw_smp_processor_id() != buf
->backend
.cpu
) {
957 /* Total order with IPI handler smp_mb() */
959 smp_call_function_single(buf
->backend
.cpu
,
961 /* Total order with IPI handler smp_mb() */
965 /* Total order with IPI handler smp_mb() */
967 smp_call_function(remote_mb
, NULL
, 1);
968 /* Total order with IPI handler smp_mb() */
973 * Local rmb to match the remote wmb to read the commit count
974 * before the buffer data and the write offset.
979 write_offset
= v_read(config
, &buf
->offset
);
982 * Check that the buffer we are getting is after or at consumed_cur
985 if ((long) subbuf_trunc(consumed
, chan
)
986 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
990 * Check that the subbuffer we are trying to consume has been
991 * already fully committed.
993 if (((commit_count
- chan
->backend
.subbuf_size
)
994 & chan
->commit_count_mask
)
995 - (buf_trunc(consumed_cur
, chan
)
996 >> chan
->backend
.num_subbuf_order
)
1001 * Check that we are not about to read the same subbuffer in
1002 * which the writer head is.
1004 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
1009 * Failure to get the subbuffer causes a busy-loop retry without going
1010 * to a wait queue. These are caused by short-lived race windows where
1011 * the writer is getting access to a subbuffer we were trying to get
1012 * access to. Also checks that the "consumed" buffer count we are
1013 * looking for matches the one contained in the subbuffer id.
1015 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1016 consumed_idx
, buf_trunc_val(consumed
, chan
));
1019 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1021 buf
->get_subbuf_consumed
= consumed
;
1022 buf
->get_subbuf
= 1;
1028 * The memory barriers __wait_event()/wake_up_interruptible() take care
1029 * of "raw_spin_is_locked" memory ordering.
1033 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
1038 EXPORT_SYMBOL_GPL(lib_ring_buffer_get_subbuf
);
1041 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1044 void lib_ring_buffer_put_subbuf(struct lib_ring_buffer
*buf
)
1046 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1047 struct channel
*chan
= bufb
->chan
;
1048 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1049 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
1051 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
1053 if (!buf
->get_subbuf
) {
1055 * Reader puts a subbuffer it did not get.
1057 CHAN_WARN_ON(chan
, 1);
1060 consumed
= buf
->get_subbuf_consumed
;
1061 buf
->get_subbuf
= 0;
1064 * Clear the records_unread counter. (overruns counter)
1065 * Can still be non-zero if a file reader simply grabbed the data
1066 * without using iterators.
1067 * Can be below zero if an iterator is used on a snapshot more than
1070 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1071 v_add(config
, v_read(config
,
1072 &bufb
->array
[read_sb_bindex
]->records_unread
),
1073 &bufb
->records_read
);
1074 v_set(config
, &bufb
->array
[read_sb_bindex
]->records_unread
, 0);
1075 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1076 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1077 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1080 * Exchange the reader subbuffer with the one we put in its place in the
1081 * writer subbuffer table. Expect the original consumed count. If
1082 * update_read_sb_index fails, this is because the writer updated the
1083 * subbuffer concurrently. We should therefore keep the subbuffer we
1084 * currently have: it has become invalid to try reading this sub-buffer
1085 * consumed count value anyway.
1087 consumed_idx
= subbuf_index(consumed
, chan
);
1088 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1089 consumed_idx
, buf_trunc_val(consumed
, chan
));
1091 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1092 * if the writer concurrently updated it.
1095 EXPORT_SYMBOL_GPL(lib_ring_buffer_put_subbuf
);
1098 * cons_offset is an iterator on all subbuffer offsets between the reader
1099 * position and the writer position. (inclusive)
1102 void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer
*buf
,
1103 struct channel
*chan
,
1104 unsigned long cons_offset
,
1107 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1108 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1110 cons_idx
= subbuf_index(cons_offset
, chan
);
1111 commit_count
= v_read(config
, &buf
->commit_hot
[cons_idx
].cc
);
1112 commit_count_sb
= v_read(config
, &buf
->commit_cold
[cons_idx
].cc_sb
);
1114 if (subbuf_offset(commit_count
, chan
) != 0)
1116 "ring buffer %s, cpu %d: "
1117 "commit count in subbuffer %lu,\n"
1118 "expecting multiples of %lu bytes\n"
1119 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1120 chan
->backend
.name
, cpu
, cons_idx
,
1121 chan
->backend
.subbuf_size
,
1122 commit_count
, commit_count_sb
);
1124 printk(KERN_DEBUG
"ring buffer: %s, cpu %d: %lu bytes committed\n",
1125 chan
->backend
.name
, cpu
, commit_count
);
1129 void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer
*buf
,
1130 struct channel
*chan
,
1131 void *priv
, int cpu
)
1133 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1134 unsigned long write_offset
, cons_offset
;
1137 * Can be called in the error path of allocation when
1138 * trans_channel_data is not yet set.
1143 * No need to order commit_count, write_offset and cons_offset reads
1144 * because we execute at teardown when no more writer nor reader
1145 * references are left.
1147 write_offset
= v_read(config
, &buf
->offset
);
1148 cons_offset
= atomic_long_read(&buf
->consumed
);
1149 if (write_offset
!= cons_offset
)
1151 "ring buffer %s, cpu %d: "
1152 "non-consumed data\n"
1153 " [ %lu bytes written, %lu bytes read ]\n",
1154 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1156 for (cons_offset
= atomic_long_read(&buf
->consumed
);
1157 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1160 cons_offset
= subbuf_align(cons_offset
, chan
))
1161 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1166 void lib_ring_buffer_print_errors(struct channel
*chan
,
1167 struct lib_ring_buffer
*buf
, int cpu
)
1169 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1170 void *priv
= chan
->backend
.priv
;
1172 printk(KERN_DEBUG
"ring buffer %s, cpu %d: %lu records written, "
1173 "%lu records overrun\n",
1174 chan
->backend
.name
, cpu
,
1175 v_read(config
, &buf
->records_count
),
1176 v_read(config
, &buf
->records_overrun
));
1178 if (v_read(config
, &buf
->records_lost_full
)
1179 || v_read(config
, &buf
->records_lost_wrap
)
1180 || v_read(config
, &buf
->records_lost_big
))
1182 "ring buffer %s, cpu %d: records were lost. Caused by:\n"
1183 " [ %lu buffer full, %lu nest buffer wrap-around, "
1184 "%lu event too big ]\n",
1185 chan
->backend
.name
, cpu
,
1186 v_read(config
, &buf
->records_lost_full
),
1187 v_read(config
, &buf
->records_lost_wrap
),
1188 v_read(config
, &buf
->records_lost_big
));
1190 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
);
1194 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1196 * Only executed when the buffer is finalized, in SWITCH_FLUSH.
1199 void lib_ring_buffer_switch_old_start(struct lib_ring_buffer
*buf
,
1200 struct channel
*chan
,
1201 struct switch_offsets
*offsets
,
1204 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1205 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1206 unsigned long commit_count
;
1208 config
->cb
.buffer_begin(buf
, tsc
, oldidx
);
1211 * Order all writes to buffer before the commit count update that will
1212 * determine that the subbuffer is full.
1214 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1216 * Must write slot data before incrementing commit count. This
1217 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1223 v_add(config
, config
->cb
.subbuffer_header_size(),
1224 &buf
->commit_hot
[oldidx
].cc
);
1225 commit_count
= v_read(config
, &buf
->commit_hot
[oldidx
].cc
);
1226 /* Check if the written buffer has to be delivered */
1227 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1228 commit_count
, oldidx
);
1229 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1230 offsets
->old
, commit_count
,
1231 config
->cb
.subbuffer_header_size());
1235 * lib_ring_buffer_switch_old_end: switch old subbuffer
1237 * Note : offset_old should never be 0 here. It is ok, because we never perform
1238 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1239 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1243 void lib_ring_buffer_switch_old_end(struct lib_ring_buffer
*buf
,
1244 struct channel
*chan
,
1245 struct switch_offsets
*offsets
,
1248 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1249 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1250 unsigned long commit_count
, padding_size
, data_size
;
1252 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1253 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1254 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
);
1257 * Order all writes to buffer before the commit count update that will
1258 * determine that the subbuffer is full.
1260 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1262 * Must write slot data before incrementing commit count. This
1263 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1269 v_add(config
, padding_size
, &buf
->commit_hot
[oldidx
].cc
);
1270 commit_count
= v_read(config
, &buf
->commit_hot
[oldidx
].cc
);
1271 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1272 commit_count
, oldidx
);
1273 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1274 offsets
->old
, commit_count
,
1279 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1281 * This code can be executed unordered : writers may already have written to the
1282 * sub-buffer before this code gets executed, caution. The commit makes sure
1283 * that this code is executed before the deliver of this sub-buffer.
1286 void lib_ring_buffer_switch_new_start(struct lib_ring_buffer
*buf
,
1287 struct channel
*chan
,
1288 struct switch_offsets
*offsets
,
1291 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1292 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1293 unsigned long commit_count
;
1295 config
->cb
.buffer_begin(buf
, tsc
, beginidx
);
1298 * Order all writes to buffer before the commit count update that will
1299 * determine that the subbuffer is full.
1301 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1303 * Must write slot data before incrementing commit count. This
1304 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1310 v_add(config
, config
->cb
.subbuffer_header_size(),
1311 &buf
->commit_hot
[beginidx
].cc
);
1312 commit_count
= v_read(config
, &buf
->commit_hot
[beginidx
].cc
);
1313 /* Check if the written buffer has to be delivered */
1314 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1315 commit_count
, beginidx
);
1316 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, beginidx
,
1317 offsets
->begin
, commit_count
,
1318 config
->cb
.subbuffer_header_size());
1322 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1324 * The only remaining threads could be the ones with pending commits. They will
1325 * have to do the deliver themselves.
1328 void lib_ring_buffer_switch_new_end(struct lib_ring_buffer
*buf
,
1329 struct channel
*chan
,
1330 struct switch_offsets
*offsets
,
1333 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1334 unsigned long endidx
= subbuf_index(offsets
->end
- 1, chan
);
1335 unsigned long commit_count
, padding_size
, data_size
;
1337 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1338 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1339 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
);
1342 * Order all writes to buffer before the commit count update that will
1343 * determine that the subbuffer is full.
1345 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1347 * Must write slot data before incrementing commit count. This
1348 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1354 v_add(config
, padding_size
, &buf
->commit_hot
[endidx
].cc
);
1355 commit_count
= v_read(config
, &buf
->commit_hot
[endidx
].cc
);
1356 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->end
- 1,
1357 commit_count
, endidx
);
1358 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, endidx
,
1359 offsets
->end
, commit_count
,
1366 * !0 if execution must be aborted.
1369 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1370 struct lib_ring_buffer
*buf
,
1371 struct channel
*chan
,
1372 struct switch_offsets
*offsets
,
1375 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1378 offsets
->begin
= v_read(config
, &buf
->offset
);
1379 offsets
->old
= offsets
->begin
;
1380 offsets
->switch_old_start
= 0;
1381 off
= subbuf_offset(offsets
->begin
, chan
);
1383 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1386 * Ensure we flush the header of an empty subbuffer when doing the
1387 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1388 * total data gathering duration even if there were no records saved
1389 * after the last buffer switch.
1390 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1391 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1392 * subbuffer header as appropriate.
1393 * The next record that reserves space will be responsible for
1394 * populating the following subbuffer header. We choose not to populate
1395 * the next subbuffer header here because we want to be able to use
1396 * SWITCH_ACTIVE for periodical buffer flush and CPU tick_nohz stop
1397 * buffer flush, which must guarantee that all the buffer content
1398 * (records and header timestamps) are visible to the reader. This is
1399 * required for quiescence guarantees for the fusion merge.
1401 if (mode
== SWITCH_FLUSH
|| off
> 0) {
1402 if (unlikely(off
== 0)) {
1404 * The client does not save any header information.
1405 * Don't switch empty subbuffer on finalize, because it
1406 * is invalid to deliver a completely empty subbuffer.
1408 if (!config
->cb
.subbuffer_header_size())
1411 * Need to write the subbuffer start header on finalize.
1413 offsets
->switch_old_start
= 1;
1415 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1417 return -1; /* we do not have to switch : buffer is empty */
1418 /* Note: old points to the next subbuf at offset 0 */
1419 offsets
->end
= offsets
->begin
;
1424 * Force a sub-buffer switch. This operation is completely reentrant : can be
1425 * called while tracing is active with absolutely no lock held.
1427 * Note, however, that as a v_cmpxchg is used for some atomic
1428 * operations, this function must be called from the CPU which owns the buffer
1429 * for a ACTIVE flush.
1431 void lib_ring_buffer_switch_slow(struct lib_ring_buffer
*buf
, enum switch_mode mode
)
1433 struct channel
*chan
= buf
->backend
.chan
;
1434 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1435 struct switch_offsets offsets
;
1436 unsigned long oldidx
;
1442 * Perform retryable operations.
1445 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1447 return; /* Switch not needed */
1448 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1452 * Atomically update last_tsc. This update races against concurrent
1453 * atomic updates, but the race will always cause supplementary full TSC
1454 * records, never the opposite (missing a full TSC record when it would
1457 save_last_tsc(config
, buf
, tsc
);
1460 * Push the reader if necessary
1462 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1464 oldidx
= subbuf_index(offsets
.old
, chan
);
1465 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
);
1468 * May need to populate header start on SWITCH_FLUSH.
1470 if (offsets
.switch_old_start
) {
1471 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
);
1472 offsets
.old
+= config
->cb
.subbuffer_header_size();
1476 * Switch old subbuffer.
1478 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
);
1480 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow
);
1485 * !0 if execution must be aborted.
1488 int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer
*buf
,
1489 struct channel
*chan
,
1490 struct switch_offsets
*offsets
,
1491 struct lib_ring_buffer_ctx
*ctx
)
1493 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1494 unsigned long reserve_commit_diff
;
1496 offsets
->begin
= v_read(config
, &buf
->offset
);
1497 offsets
->old
= offsets
->begin
;
1498 offsets
->switch_new_start
= 0;
1499 offsets
->switch_new_end
= 0;
1500 offsets
->switch_old_end
= 0;
1501 offsets
->pre_header_padding
= 0;
1503 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1505 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1506 ctx
->rflags
= RING_BUFFER_RFLAG_FULL_TSC
;
1508 if (unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1509 offsets
->switch_new_start
= 1; /* For offsets->begin */
1511 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1514 &offsets
->pre_header_padding
,
1517 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1520 if (unlikely(subbuf_offset(offsets
->begin
, chan
) +
1521 offsets
->size
> chan
->backend
.subbuf_size
)) {
1522 offsets
->switch_old_end
= 1; /* For offsets->old */
1523 offsets
->switch_new_start
= 1; /* For offsets->begin */
1526 if (unlikely(offsets
->switch_new_start
)) {
1527 unsigned long sb_index
;
1530 * We are typically not filling the previous buffer completely.
1532 if (likely(offsets
->switch_old_end
))
1533 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1534 offsets
->begin
= offsets
->begin
1535 + config
->cb
.subbuffer_header_size();
1536 /* Test new buffer integrity */
1537 sb_index
= subbuf_index(offsets
->begin
, chan
);
1538 reserve_commit_diff
=
1539 (buf_trunc(offsets
->begin
, chan
)
1540 >> chan
->backend
.num_subbuf_order
)
1541 - ((unsigned long) v_read(config
,
1542 &buf
->commit_cold
[sb_index
].cc_sb
)
1543 & chan
->commit_count_mask
);
1544 if (likely(reserve_commit_diff
== 0)) {
1545 /* Next subbuffer not being written to. */
1546 if (unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1547 subbuf_trunc(offsets
->begin
, chan
)
1548 - subbuf_trunc((unsigned long)
1549 atomic_long_read(&buf
->consumed
), chan
)
1550 >= chan
->backend
.buf_size
)) {
1552 * We do not overwrite non consumed buffers
1553 * and we are full : record is lost.
1555 v_inc(config
, &buf
->records_lost_full
);
1559 * Next subbuffer not being written to, and we
1560 * are either in overwrite mode or the buffer is
1561 * not full. It's safe to write in this new
1567 * Next subbuffer reserve offset does not match the
1568 * commit offset. Drop record in producer-consumer and
1569 * overwrite mode. Caused by either a writer OOPS or too
1570 * many nested writes over a reserve/commit pair.
1572 v_inc(config
, &buf
->records_lost_wrap
);
1576 config
->cb
.record_header_size(config
, chan
,
1579 &offsets
->pre_header_padding
,
1582 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1585 if (unlikely(subbuf_offset(offsets
->begin
, chan
)
1586 + offsets
->size
> chan
->backend
.subbuf_size
)) {
1588 * Record too big for subbuffers, report error, don't
1589 * complete the sub-buffer switch.
1591 v_inc(config
, &buf
->records_lost_big
);
1595 * We just made a successful buffer switch and the
1596 * record fits in the new subbuffer. Let's write.
1601 * Record fits in the current buffer and we are not on a switch
1602 * boundary. It's safe to write.
1605 offsets
->end
= offsets
->begin
+ offsets
->size
;
1607 if (unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
1609 * The offset_end will fall at the very beginning of the next
1612 offsets
->switch_new_end
= 1; /* For offsets->begin */
1618 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1619 * @ctx: ring buffer context.
1621 * Return : -ENOSPC if not enough space, else returns 0.
1622 * It will take care of sub-buffer switching.
1624 int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx
*ctx
)
1626 struct channel
*chan
= ctx
->chan
;
1627 const struct lib_ring_buffer_config
*config
= chan
->backend
.config
;
1628 struct lib_ring_buffer
*buf
;
1629 struct switch_offsets offsets
;
1631 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
1632 buf
= per_cpu_ptr(chan
->backend
.buf
, ctx
->cpu
);
1634 buf
= chan
->backend
.buf
;
1640 if (unlikely(lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
1643 } while (unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
1648 * Atomically update last_tsc. This update races against concurrent
1649 * atomic updates, but the race will always cause supplementary full TSC
1650 * records, never the opposite (missing a full TSC record when it would
1653 save_last_tsc(config
, buf
, ctx
->tsc
);
1656 * Push the reader if necessary
1658 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
1661 * Clear noref flag for this subbuffer.
1663 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1664 subbuf_index(offsets
.end
- 1, chan
));
1667 * Switch old subbuffer if needed.
1669 if (unlikely(offsets
.switch_old_end
)) {
1670 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1671 subbuf_index(offsets
.old
- 1, chan
));
1672 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
);
1676 * Populate new subbuffer.
1678 if (unlikely(offsets
.switch_new_start
))
1679 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
);
1681 if (unlikely(offsets
.switch_new_end
))
1682 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
);
1684 ctx
->slot_size
= offsets
.size
;
1685 ctx
->pre_offset
= offsets
.begin
;
1686 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
1689 EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow
);