2 * ring_buffer_frontend.c
4 * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; only
9 * version 2.1 of the License.
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 * Ring buffer wait-free buffer synchronization. Producer-consumer and flight
22 * recorder (overwrite) modes. See thesis:
24 * Desnoyers, Mathieu (2009), "Low-Impact Operating System Tracing", Ph.D.
25 * dissertation, Ecole Polytechnique de Montreal.
26 * http://www.lttng.org/pub/thesis/desnoyers-dissertation-2009-12.pdf
28 * - Algorithm presentation in Chapter 5:
29 * "Lockless Multi-Core High-Throughput Buffering".
30 * - Algorithm formal verification in Section 8.6:
31 * "Formal verification of LTTng"
34 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
36 * Inspired from LTT and RelayFS:
37 * Karim Yaghmour <karim@opersys.com>
38 * Tom Zanussi <zanussi@us.ibm.com>
39 * Bob Wisniewski <bob@watson.ibm.com>
41 * Bob Wisniewski <bob@watson.ibm.com>
43 * Buffer reader semantic :
46 * while buffer is not finalized and empty
48 * - if return value != 0, continue
49 * - splice one subbuffer worth of data to a pipe
50 * - splice the data from pipe to disk/network
54 #include <linux/delay.h>
55 #include <linux/module.h>
56 #include <linux/percpu.h>
58 #include "../../wrapper/ringbuffer/config.h"
59 #include "../../wrapper/ringbuffer/backend.h"
60 #include "../../wrapper/ringbuffer/frontend.h"
61 #include "../../wrapper/ringbuffer/iterator.h"
62 #include "../../wrapper/ringbuffer/nohz.h"
65 * Internal structure representing offsets to use at a sub-buffer switch.
67 struct switch_offsets
{
68 unsigned long begin
, end
, old
;
69 size_t pre_header_padding
, size
;
70 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
81 static ATOMIC_NOTIFIER_HEAD(tick_nohz_notifier
);
82 #endif /* CONFIG_NO_HZ */
84 static DEFINE_PER_CPU(spinlock_t
, ring_buffer_nohz_lock
);
86 DEFINE_PER_CPU(unsigned int, lib_ring_buffer_nesting
);
87 EXPORT_PER_CPU_SYMBOL(lib_ring_buffer_nesting
);
90 void lib_ring_buffer_print_errors(struct channel
*chan
,
91 struct lib_ring_buffer
*buf
, int cpu
);
94 * Must be called under cpu hotplug protection.
96 void lib_ring_buffer_free(struct lib_ring_buffer
*buf
)
98 struct channel
*chan
= buf
->backend
.chan
;
100 lib_ring_buffer_print_errors(chan
, buf
, buf
->backend
.cpu
);
101 kfree(buf
->commit_hot
);
102 kfree(buf
->commit_cold
);
104 lib_ring_buffer_backend_free(&buf
->backend
);
108 * lib_ring_buffer_reset - Reset ring buffer to initial values.
111 * Effectively empty the ring buffer. Should be called when the buffer is not
112 * used for writing. The ring buffer can be opened for reading, but the reader
113 * should not be using the iterator concurrently with reset. The previous
114 * current iterator record is reset.
116 void lib_ring_buffer_reset(struct lib_ring_buffer
*buf
)
118 struct channel
*chan
= buf
->backend
.chan
;
119 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
123 * Reset iterator first. It will put the subbuffer if it currently holds
126 lib_ring_buffer_iterator_reset(buf
);
127 v_set(config
, &buf
->offset
, 0);
128 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
129 v_set(config
, &buf
->commit_hot
[i
].cc
, 0);
130 v_set(config
, &buf
->commit_hot
[i
].seq
, 0);
131 v_set(config
, &buf
->commit_cold
[i
].cc_sb
, 0);
133 atomic_long_set(&buf
->consumed
, 0);
134 atomic_set(&buf
->record_disabled
, 0);
135 v_set(config
, &buf
->last_tsc
, 0);
136 lib_ring_buffer_backend_reset(&buf
->backend
);
137 /* Don't reset number of active readers */
138 v_set(config
, &buf
->records_lost_full
, 0);
139 v_set(config
, &buf
->records_lost_wrap
, 0);
140 v_set(config
, &buf
->records_lost_big
, 0);
141 v_set(config
, &buf
->records_count
, 0);
142 v_set(config
, &buf
->records_overrun
, 0);
145 EXPORT_SYMBOL_GPL(lib_ring_buffer_reset
);
148 * channel_reset - Reset channel to initial values.
151 * Effectively empty the channel. Should be called when the channel is not used
152 * for writing. The channel can be opened for reading, but the reader should not
153 * be using the iterator concurrently with reset. The previous current iterator
156 void channel_reset(struct channel
*chan
)
159 * Reset iterators first. Will put the subbuffer if held for reading.
161 channel_iterator_reset(chan
);
162 atomic_set(&chan
->record_disabled
, 0);
163 /* Don't reset commit_count_mask, still valid */
164 channel_backend_reset(&chan
->backend
);
165 /* Don't reset switch/read timer interval */
166 /* Don't reset notifiers and notifier enable bits */
167 /* Don't reset reader reference count */
169 EXPORT_SYMBOL_GPL(channel_reset
);
172 * Must be called under cpu hotplug protection.
174 int lib_ring_buffer_create(struct lib_ring_buffer
*buf
,
175 struct channel_backend
*chanb
, int cpu
)
177 const struct lib_ring_buffer_config
*config
= &chanb
->config
;
178 struct channel
*chan
= container_of(chanb
, struct channel
, backend
);
179 void *priv
= chanb
->priv
;
180 size_t subbuf_header_size
;
184 /* Test for cpu hotplug */
185 if (buf
->backend
.allocated
)
189 * Paranoia: per cpu dynamic allocation is not officially documented as
190 * zeroing the memory, so let's do it here too, just in case.
192 memset(buf
, 0, sizeof(*buf
));
194 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
, cpu
);
199 kzalloc_node(ALIGN(sizeof(*buf
->commit_hot
)
200 * chan
->backend
.num_subbuf
,
201 1 << INTERNODE_CACHE_SHIFT
),
202 GFP_KERNEL
, cpu_to_node(max(cpu
, 0)));
203 if (!buf
->commit_hot
) {
209 kzalloc_node(ALIGN(sizeof(*buf
->commit_cold
)
210 * chan
->backend
.num_subbuf
,
211 1 << INTERNODE_CACHE_SHIFT
),
212 GFP_KERNEL
, cpu_to_node(max(cpu
, 0)));
213 if (!buf
->commit_cold
) {
218 init_waitqueue_head(&buf
->read_wait
);
219 init_waitqueue_head(&buf
->write_wait
);
220 raw_spin_lock_init(&buf
->raw_tick_nohz_spinlock
);
223 * Write the subbuffer header for first subbuffer so we know the total
224 * duration of data gathering.
226 subbuf_header_size
= config
->cb
.subbuffer_header_size();
227 v_set(config
, &buf
->offset
, subbuf_header_size
);
228 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_wsb
[0].id
);
229 tsc
= config
->cb
.ring_buffer_clock_read(buf
->backend
.chan
);
230 config
->cb
.buffer_begin(buf
, tsc
, 0);
231 v_add(config
, subbuf_header_size
, &buf
->commit_hot
[0].cc
);
233 if (config
->cb
.buffer_create
) {
234 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
);
240 * Ensure the buffer is ready before setting it to allocated and setting
242 * Used for cpu hotplug vs cpumask iteration.
245 buf
->backend
.allocated
= 1;
247 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
248 CHAN_WARN_ON(chan
, cpumask_test_cpu(cpu
,
249 chan
->backend
.cpumask
));
250 cpumask_set_cpu(cpu
, chan
->backend
.cpumask
);
257 kfree(buf
->commit_cold
);
259 kfree(buf
->commit_hot
);
261 lib_ring_buffer_backend_free(&buf
->backend
);
265 static void switch_buffer_timer(unsigned long data
)
267 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
268 struct channel
*chan
= buf
->backend
.chan
;
269 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
272 * Only flush buffers periodically if readers are active.
274 if (atomic_long_read(&buf
->active_readers
))
275 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
277 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
278 mod_timer_pinned(&buf
->switch_timer
,
279 jiffies
+ chan
->switch_timer_interval
);
281 mod_timer(&buf
->switch_timer
,
282 jiffies
+ chan
->switch_timer_interval
);
286 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
288 static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer
*buf
)
290 struct channel
*chan
= buf
->backend
.chan
;
291 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
293 if (!chan
->switch_timer_interval
|| buf
->switch_timer_enabled
)
295 init_timer(&buf
->switch_timer
);
296 buf
->switch_timer
.function
= switch_buffer_timer
;
297 buf
->switch_timer
.expires
= jiffies
+ chan
->switch_timer_interval
;
298 buf
->switch_timer
.data
= (unsigned long)buf
;
299 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
300 add_timer_on(&buf
->switch_timer
, buf
->backend
.cpu
);
302 add_timer(&buf
->switch_timer
);
303 buf
->switch_timer_enabled
= 1;
307 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
309 static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer
*buf
)
311 struct channel
*chan
= buf
->backend
.chan
;
313 if (!chan
->switch_timer_interval
|| !buf
->switch_timer_enabled
)
316 del_timer_sync(&buf
->switch_timer
);
317 buf
->switch_timer_enabled
= 0;
321 * Polling timer to check the channels for data.
323 static void read_buffer_timer(unsigned long data
)
325 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
326 struct channel
*chan
= buf
->backend
.chan
;
327 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
329 CHAN_WARN_ON(chan
, !buf
->backend
.allocated
);
331 if (atomic_long_read(&buf
->active_readers
)
332 && lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
333 wake_up_interruptible(&buf
->read_wait
);
334 wake_up_interruptible(&chan
->read_wait
);
337 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
338 mod_timer_pinned(&buf
->read_timer
,
339 jiffies
+ chan
->read_timer_interval
);
341 mod_timer(&buf
->read_timer
,
342 jiffies
+ chan
->read_timer_interval
);
346 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
348 static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer
*buf
)
350 struct channel
*chan
= buf
->backend
.chan
;
351 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
353 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
354 || !chan
->read_timer_interval
355 || buf
->read_timer_enabled
)
358 init_timer(&buf
->read_timer
);
359 buf
->read_timer
.function
= read_buffer_timer
;
360 buf
->read_timer
.expires
= jiffies
+ chan
->read_timer_interval
;
361 buf
->read_timer
.data
= (unsigned long)buf
;
363 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
364 add_timer_on(&buf
->read_timer
, buf
->backend
.cpu
);
366 add_timer(&buf
->read_timer
);
367 buf
->read_timer_enabled
= 1;
371 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
373 static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer
*buf
)
375 struct channel
*chan
= buf
->backend
.chan
;
376 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
378 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
379 || !chan
->read_timer_interval
380 || !buf
->read_timer_enabled
)
383 del_timer_sync(&buf
->read_timer
);
385 * do one more check to catch data that has been written in the last
388 if (lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
389 wake_up_interruptible(&buf
->read_wait
);
390 wake_up_interruptible(&chan
->read_wait
);
392 buf
->read_timer_enabled
= 0;
395 #ifdef CONFIG_HOTPLUG_CPU
397 * lib_ring_buffer_cpu_hp_callback - CPU hotplug callback
398 * @nb: notifier block
399 * @action: hotplug action to take
402 * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
405 int lib_ring_buffer_cpu_hp_callback(struct notifier_block
*nb
,
406 unsigned long action
,
409 unsigned int cpu
= (unsigned long)hcpu
;
410 struct channel
*chan
= container_of(nb
, struct channel
,
412 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
, cpu
);
413 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
415 if (!chan
->cpu_hp_enable
)
418 CHAN_WARN_ON(chan
, config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
);
421 case CPU_DOWN_FAILED
:
422 case CPU_DOWN_FAILED_FROZEN
:
424 case CPU_ONLINE_FROZEN
:
425 wake_up_interruptible(&chan
->hp_wait
);
426 lib_ring_buffer_start_switch_timer(buf
);
427 lib_ring_buffer_start_read_timer(buf
);
430 case CPU_DOWN_PREPARE
:
431 case CPU_DOWN_PREPARE_FROZEN
:
432 lib_ring_buffer_stop_switch_timer(buf
);
433 lib_ring_buffer_stop_read_timer(buf
);
437 case CPU_DEAD_FROZEN
:
439 * Performing a buffer switch on a remote CPU. Performed by
440 * the CPU responsible for doing the hotunplug after the target
441 * CPU stopped running completely. Ensures that all data
442 * from that remote CPU is flushed.
444 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
453 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
455 * For per-cpu buffers, call the reader wakeups before switching the buffer, so
456 * that wake-up-tracing generated events are flushed before going idle (in
457 * tick_nohz). We test if the spinlock is locked to deal with the race where
458 * readers try to sample the ring buffer before we perform the switch. We let
459 * the readers retry in that case. If there is data in the buffer, the wake up
460 * is going to forbid the CPU running the reader thread from going idle.
462 static int notrace
ring_buffer_tick_nohz_callback(struct notifier_block
*nb
,
466 struct channel
*chan
= container_of(nb
, struct channel
,
468 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
469 struct lib_ring_buffer
*buf
;
470 int cpu
= smp_processor_id();
472 if (config
->alloc
!= RING_BUFFER_ALLOC_PER_CPU
) {
474 * We don't support keeping the system idle with global buffers
475 * and streaming active. In order to do so, we would need to
476 * sample a non-nohz-cpumask racelessly with the nohz updates
477 * without adding synchronization overhead to nohz. Leave this
478 * use-case out for now.
483 buf
= channel_get_ring_buffer(config
, chan
, cpu
);
485 case TICK_NOHZ_FLUSH
:
486 raw_spin_lock(&buf
->raw_tick_nohz_spinlock
);
487 if (config
->wakeup
== RING_BUFFER_WAKEUP_BY_TIMER
488 && chan
->read_timer_interval
489 && atomic_long_read(&buf
->active_readers
)
490 && (lib_ring_buffer_poll_deliver(config
, buf
, chan
)
491 || lib_ring_buffer_pending_data(config
, buf
, chan
))) {
492 wake_up_interruptible(&buf
->read_wait
);
493 wake_up_interruptible(&chan
->read_wait
);
495 if (chan
->switch_timer_interval
)
496 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
497 raw_spin_unlock(&buf
->raw_tick_nohz_spinlock
);
500 spin_lock(&__get_cpu_var(ring_buffer_nohz_lock
));
501 lib_ring_buffer_stop_switch_timer(buf
);
502 lib_ring_buffer_stop_read_timer(buf
);
503 spin_unlock(&__get_cpu_var(ring_buffer_nohz_lock
));
505 case TICK_NOHZ_RESTART
:
506 spin_lock(&__get_cpu_var(ring_buffer_nohz_lock
));
507 lib_ring_buffer_start_read_timer(buf
);
508 lib_ring_buffer_start_switch_timer(buf
);
509 spin_unlock(&__get_cpu_var(ring_buffer_nohz_lock
));
516 void notrace
lib_ring_buffer_tick_nohz_flush(void)
518 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_FLUSH
,
522 void notrace
lib_ring_buffer_tick_nohz_stop(void)
524 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_STOP
,
528 void notrace
lib_ring_buffer_tick_nohz_restart(void)
530 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_RESTART
,
533 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
538 static void channel_unregister_notifiers(struct channel
*chan
)
540 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
543 channel_iterator_unregister_notifiers(chan
);
544 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
547 * Remove the nohz notifier first, so we are certain we stop
550 atomic_notifier_chain_unregister(&tick_nohz_notifier
,
551 &chan
->tick_nohz_notifier
);
553 * ring_buffer_nohz_lock will not be needed below, because
554 * we just removed the notifiers, which were the only source of
557 #endif /* CONFIG_NO_HZ */
558 #ifdef CONFIG_HOTPLUG_CPU
560 chan
->cpu_hp_enable
= 0;
561 for_each_online_cpu(cpu
) {
562 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
564 lib_ring_buffer_stop_switch_timer(buf
);
565 lib_ring_buffer_stop_read_timer(buf
);
568 unregister_cpu_notifier(&chan
->cpu_hp_notifier
);
570 for_each_possible_cpu(cpu
) {
571 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
573 lib_ring_buffer_stop_switch_timer(buf
);
574 lib_ring_buffer_stop_read_timer(buf
);
578 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
580 lib_ring_buffer_stop_switch_timer(buf
);
581 lib_ring_buffer_stop_read_timer(buf
);
583 channel_backend_unregister_notifiers(&chan
->backend
);
586 static void channel_free(struct channel
*chan
)
588 channel_iterator_free(chan
);
589 channel_backend_free(&chan
->backend
);
594 * channel_create - Create channel.
595 * @config: ring buffer instance configuration
596 * @name: name of the channel
597 * @priv: ring buffer client private data
598 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
599 * address mapping. It is used only by RING_BUFFER_STATIC
600 * configuration. It can be set to NULL for other backends.
601 * @subbuf_size: subbuffer size
602 * @num_subbuf: number of subbuffers
603 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
604 * padding to let readers get those sub-buffers.
605 * Used for live streaming.
606 * @read_timer_interval: Time interval (in us) to wake up pending readers.
609 * Returns NULL on failure.
611 struct channel
*channel_create(const struct lib_ring_buffer_config
*config
,
612 const char *name
, void *priv
, void *buf_addr
,
614 size_t num_subbuf
, unsigned int switch_timer_interval
,
615 unsigned int read_timer_interval
)
618 struct channel
*chan
;
620 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
621 read_timer_interval
))
624 chan
= kzalloc(sizeof(struct channel
), GFP_KERNEL
);
628 ret
= channel_backend_init(&chan
->backend
, name
, config
, priv
,
629 subbuf_size
, num_subbuf
);
633 ret
= channel_iterator_init(chan
);
635 goto error_free_backend
;
637 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
638 chan
->switch_timer_interval
= usecs_to_jiffies(switch_timer_interval
);
639 chan
->read_timer_interval
= usecs_to_jiffies(read_timer_interval
);
640 kref_init(&chan
->ref
);
641 init_waitqueue_head(&chan
->read_wait
);
642 init_waitqueue_head(&chan
->hp_wait
);
644 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
645 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
646 /* Only benefit from NO_HZ idle with per-cpu buffers for now. */
647 chan
->tick_nohz_notifier
.notifier_call
=
648 ring_buffer_tick_nohz_callback
;
649 chan
->tick_nohz_notifier
.priority
= ~0U;
650 atomic_notifier_chain_register(&tick_nohz_notifier
,
651 &chan
->tick_nohz_notifier
);
652 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
655 * In case of non-hotplug cpu, if the ring-buffer is allocated
656 * in early initcall, it will not be notified of secondary cpus.
657 * In that off case, we need to allocate for all possible cpus.
659 #ifdef CONFIG_HOTPLUG_CPU
660 chan
->cpu_hp_notifier
.notifier_call
=
661 lib_ring_buffer_cpu_hp_callback
;
662 chan
->cpu_hp_notifier
.priority
= 6;
663 register_cpu_notifier(&chan
->cpu_hp_notifier
);
666 for_each_online_cpu(cpu
) {
667 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
669 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
670 lib_ring_buffer_start_switch_timer(buf
);
671 lib_ring_buffer_start_read_timer(buf
);
672 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
674 chan
->cpu_hp_enable
= 1;
677 for_each_possible_cpu(cpu
) {
678 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
680 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
681 lib_ring_buffer_start_switch_timer(buf
);
682 lib_ring_buffer_start_read_timer(buf
);
683 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
687 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
689 lib_ring_buffer_start_switch_timer(buf
);
690 lib_ring_buffer_start_read_timer(buf
);
696 channel_backend_free(&chan
->backend
);
701 EXPORT_SYMBOL_GPL(channel_create
);
704 void channel_release(struct kref
*kref
)
706 struct channel
*chan
= container_of(kref
, struct channel
, ref
);
711 * channel_destroy - Finalize, wait for q.s. and destroy channel.
712 * @chan: channel to destroy
715 * Call "destroy" callback, finalize channels, and then decrement the
716 * channel reference count. Note that when readers have completed data
717 * consumption of finalized channels, get_subbuf() will return -ENODATA.
718 * They should release their handle at that point. Returns the private
721 void *channel_destroy(struct channel
*chan
)
724 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
727 channel_unregister_notifiers(chan
);
729 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
731 * No need to hold cpu hotplug, because all notifiers have been
734 for_each_channel_cpu(cpu
, chan
) {
735 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
738 if (config
->cb
.buffer_finalize
)
739 config
->cb
.buffer_finalize(buf
,
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
);
752 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
754 if (config
->cb
.buffer_finalize
)
755 config
->cb
.buffer_finalize(buf
, chan
->backend
.priv
, -1);
756 if (buf
->backend
.allocated
)
757 lib_ring_buffer_switch_slow(buf
, SWITCH_FLUSH
);
759 * Perform flush before writing to finalized.
762 ACCESS_ONCE(buf
->finalized
) = 1;
763 wake_up_interruptible(&buf
->read_wait
);
765 ACCESS_ONCE(chan
->finalized
) = 1;
766 wake_up_interruptible(&chan
->hp_wait
);
767 wake_up_interruptible(&chan
->read_wait
);
768 priv
= chan
->backend
.priv
;
769 kref_put(&chan
->ref
, channel_release
);
772 EXPORT_SYMBOL_GPL(channel_destroy
);
774 struct lib_ring_buffer
*channel_get_ring_buffer(
775 const struct lib_ring_buffer_config
*config
,
776 struct channel
*chan
, int cpu
)
778 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
)
779 return chan
->backend
.buf
;
781 return per_cpu_ptr(chan
->backend
.buf
, cpu
);
783 EXPORT_SYMBOL_GPL(channel_get_ring_buffer
);
785 int lib_ring_buffer_open_read(struct lib_ring_buffer
*buf
)
787 struct channel
*chan
= buf
->backend
.chan
;
789 if (!atomic_long_add_unless(&buf
->active_readers
, 1, 1))
791 kref_get(&chan
->ref
);
792 smp_mb__after_atomic_inc();
795 EXPORT_SYMBOL_GPL(lib_ring_buffer_open_read
);
797 void lib_ring_buffer_release_read(struct lib_ring_buffer
*buf
)
799 struct channel
*chan
= buf
->backend
.chan
;
801 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
802 smp_mb__before_atomic_dec();
803 atomic_long_dec(&buf
->active_readers
);
804 kref_put(&chan
->ref
, channel_release
);
806 EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read
);
809 * Promote compiler barrier to a smp_mb().
810 * For the specific ring buffer case, this IPI call should be removed if the
811 * architecture does not reorder writes. This should eventually be provided by
812 * a separate architecture-specific infrastructure.
814 static void remote_mb(void *info
)
820 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
822 * @consumed: consumed count indicating the position where to read
823 * @produced: produced count, indicates position when to stop reading
825 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
826 * data to read at consumed position, or 0 if the get operation succeeds.
827 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
830 int lib_ring_buffer_snapshot(struct lib_ring_buffer
*buf
,
831 unsigned long *consumed
, unsigned long *produced
)
833 struct channel
*chan
= buf
->backend
.chan
;
834 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
835 unsigned long consumed_cur
, write_offset
;
839 finalized
= ACCESS_ONCE(buf
->finalized
);
841 * Read finalized before counters.
844 consumed_cur
= atomic_long_read(&buf
->consumed
);
846 * No need to issue a memory barrier between consumed count read and
847 * write offset read, because consumed count can only change
848 * concurrently in overwrite mode, and we keep a sequence counter
849 * identifier derived from the write offset to check we are getting
850 * the same sub-buffer we are expecting (the sub-buffers are atomically
851 * "tagged" upon writes, tags are checked upon read).
853 write_offset
= v_read(config
, &buf
->offset
);
856 * Check that we are not about to read the same subbuffer in
857 * which the writer head is.
859 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
863 *consumed
= consumed_cur
;
864 *produced
= subbuf_trunc(write_offset
, chan
);
870 * The memory barriers __wait_event()/wake_up_interruptible() take care
871 * of "raw_spin_is_locked" memory ordering.
875 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
880 EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot
);
883 * lib_ring_buffer_put_snapshot - move consumed counter forward
885 * Should only be called from consumer context.
887 * @consumed_new: new consumed count value
889 void lib_ring_buffer_move_consumer(struct lib_ring_buffer
*buf
,
890 unsigned long consumed_new
)
892 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
893 struct channel
*chan
= bufb
->chan
;
894 unsigned long consumed
;
896 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
899 * Only push the consumed value forward.
900 * If the consumed cmpxchg fails, this is because we have been pushed by
901 * the writer in flight recorder mode.
903 consumed
= atomic_long_read(&buf
->consumed
);
904 while ((long) consumed
- (long) consumed_new
< 0)
905 consumed
= atomic_long_cmpxchg(&buf
->consumed
, consumed
,
907 /* Wake-up the metadata producer */
908 wake_up_interruptible(&buf
->write_wait
);
910 EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer
);
913 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
915 * @consumed: consumed count indicating the position where to read
917 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
918 * data to read at consumed position, or 0 if the get operation succeeds.
919 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
921 int lib_ring_buffer_get_subbuf(struct lib_ring_buffer
*buf
,
922 unsigned long consumed
)
924 struct channel
*chan
= buf
->backend
.chan
;
925 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
926 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
930 if (buf
->get_subbuf
) {
932 * Reader is trying to get a subbuffer twice.
934 CHAN_WARN_ON(chan
, 1);
938 finalized
= ACCESS_ONCE(buf
->finalized
);
940 * Read finalized before counters.
943 consumed_cur
= atomic_long_read(&buf
->consumed
);
944 consumed_idx
= subbuf_index(consumed
, chan
);
945 commit_count
= v_read(config
, &buf
->commit_cold
[consumed_idx
].cc_sb
);
947 * Make sure we read the commit count before reading the buffer
948 * data and the write offset. Correct consumed offset ordering
949 * wrt commit count is insured by the use of cmpxchg to update
950 * the consumed offset.
951 * smp_call_function_single can fail if the remote CPU is offline,
952 * this is OK because then there is no wmb to execute there.
953 * If our thread is executing on the same CPU as the on the buffers
954 * belongs to, we don't have to synchronize it at all. If we are
955 * migrated, the scheduler will take care of the memory barriers.
956 * Normally, smp_call_function_single() should ensure program order when
957 * executing the remote function, which implies that it surrounds the
958 * function execution with :
969 * However, smp_call_function_single() does not seem to clearly execute
970 * such barriers. It depends on spinlock semantic to provide the barrier
971 * before executing the IPI and, when busy-looping, csd_lock_wait only
972 * executes smp_mb() when it has to wait for the other CPU.
974 * I don't trust this code. Therefore, let's add the smp_mb() sequence
975 * required ourself, even if duplicated. It has no performance impact
978 * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs
979 * read and write vs write. They do not ensure core synchronization. We
980 * really have to ensure total order between the 3 barriers running on
983 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
984 if (config
->sync
== RING_BUFFER_SYNC_PER_CPU
985 && config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
986 if (raw_smp_processor_id() != buf
->backend
.cpu
) {
987 /* Total order with IPI handler smp_mb() */
989 smp_call_function_single(buf
->backend
.cpu
,
991 /* Total order with IPI handler smp_mb() */
995 /* Total order with IPI handler smp_mb() */
997 smp_call_function(remote_mb
, NULL
, 1);
998 /* Total order with IPI handler smp_mb() */
1003 * Local rmb to match the remote wmb to read the commit count
1004 * before the buffer data and the write offset.
1009 write_offset
= v_read(config
, &buf
->offset
);
1012 * Check that the buffer we are getting is after or at consumed_cur
1015 if ((long) subbuf_trunc(consumed
, chan
)
1016 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
1020 * Check that the subbuffer we are trying to consume has been
1021 * already fully committed.
1023 if (((commit_count
- chan
->backend
.subbuf_size
)
1024 & chan
->commit_count_mask
)
1025 - (buf_trunc(consumed_cur
, chan
)
1026 >> chan
->backend
.num_subbuf_order
)
1031 * Check that we are not about to read the same subbuffer in
1032 * which the writer head is.
1034 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
1039 * Failure to get the subbuffer causes a busy-loop retry without going
1040 * to a wait queue. These are caused by short-lived race windows where
1041 * the writer is getting access to a subbuffer we were trying to get
1042 * access to. Also checks that the "consumed" buffer count we are
1043 * looking for matches the one contained in the subbuffer id.
1045 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1046 consumed_idx
, buf_trunc_val(consumed
, chan
));
1049 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1051 buf
->get_subbuf_consumed
= consumed
;
1052 buf
->get_subbuf
= 1;
1058 * The memory barriers __wait_event()/wake_up_interruptible() take care
1059 * of "raw_spin_is_locked" memory ordering.
1063 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
1068 EXPORT_SYMBOL_GPL(lib_ring_buffer_get_subbuf
);
1071 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1074 void lib_ring_buffer_put_subbuf(struct lib_ring_buffer
*buf
)
1076 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1077 struct channel
*chan
= bufb
->chan
;
1078 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1079 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
1081 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
1083 if (!buf
->get_subbuf
) {
1085 * Reader puts a subbuffer it did not get.
1087 CHAN_WARN_ON(chan
, 1);
1090 consumed
= buf
->get_subbuf_consumed
;
1091 buf
->get_subbuf
= 0;
1094 * Clear the records_unread counter. (overruns counter)
1095 * Can still be non-zero if a file reader simply grabbed the data
1096 * without using iterators.
1097 * Can be below zero if an iterator is used on a snapshot more than
1100 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1101 v_add(config
, v_read(config
,
1102 &bufb
->array
[read_sb_bindex
]->records_unread
),
1103 &bufb
->records_read
);
1104 v_set(config
, &bufb
->array
[read_sb_bindex
]->records_unread
, 0);
1105 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1106 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1107 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1110 * Exchange the reader subbuffer with the one we put in its place in the
1111 * writer subbuffer table. Expect the original consumed count. If
1112 * update_read_sb_index fails, this is because the writer updated the
1113 * subbuffer concurrently. We should therefore keep the subbuffer we
1114 * currently have: it has become invalid to try reading this sub-buffer
1115 * consumed count value anyway.
1117 consumed_idx
= subbuf_index(consumed
, chan
);
1118 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1119 consumed_idx
, buf_trunc_val(consumed
, chan
));
1121 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1122 * if the writer concurrently updated it.
1125 EXPORT_SYMBOL_GPL(lib_ring_buffer_put_subbuf
);
1128 * cons_offset is an iterator on all subbuffer offsets between the reader
1129 * position and the writer position. (inclusive)
1132 void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer
*buf
,
1133 struct channel
*chan
,
1134 unsigned long cons_offset
,
1137 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1138 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1140 cons_idx
= subbuf_index(cons_offset
, chan
);
1141 commit_count
= v_read(config
, &buf
->commit_hot
[cons_idx
].cc
);
1142 commit_count_sb
= v_read(config
, &buf
->commit_cold
[cons_idx
].cc_sb
);
1144 if (subbuf_offset(commit_count
, chan
) != 0)
1146 "ring buffer %s, cpu %d: "
1147 "commit count in subbuffer %lu,\n"
1148 "expecting multiples of %lu bytes\n"
1149 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1150 chan
->backend
.name
, cpu
, cons_idx
,
1151 chan
->backend
.subbuf_size
,
1152 commit_count
, commit_count_sb
);
1154 printk(KERN_DEBUG
"ring buffer: %s, cpu %d: %lu bytes committed\n",
1155 chan
->backend
.name
, cpu
, commit_count
);
1159 void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer
*buf
,
1160 struct channel
*chan
,
1161 void *priv
, int cpu
)
1163 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1164 unsigned long write_offset
, cons_offset
;
1167 * No need to order commit_count, write_offset and cons_offset reads
1168 * because we execute at teardown when no more writer nor reader
1169 * references are left.
1171 write_offset
= v_read(config
, &buf
->offset
);
1172 cons_offset
= atomic_long_read(&buf
->consumed
);
1173 if (write_offset
!= cons_offset
)
1175 "ring buffer %s, cpu %d: "
1176 "non-consumed data\n"
1177 " [ %lu bytes written, %lu bytes read ]\n",
1178 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1180 for (cons_offset
= atomic_long_read(&buf
->consumed
);
1181 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1184 cons_offset
= subbuf_align(cons_offset
, chan
))
1185 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1190 void lib_ring_buffer_print_errors(struct channel
*chan
,
1191 struct lib_ring_buffer
*buf
, int cpu
)
1193 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1194 void *priv
= chan
->backend
.priv
;
1196 if (!strcmp(chan
->backend
.name
, "relay-metadata")) {
1197 printk(KERN_DEBUG
"ring buffer %s: %lu records written, "
1198 "%lu records overrun\n",
1200 v_read(config
, &buf
->records_count
),
1201 v_read(config
, &buf
->records_overrun
));
1203 printk(KERN_DEBUG
"ring buffer %s, cpu %d: %lu records written, "
1204 "%lu records overrun\n",
1205 chan
->backend
.name
, cpu
,
1206 v_read(config
, &buf
->records_count
),
1207 v_read(config
, &buf
->records_overrun
));
1209 if (v_read(config
, &buf
->records_lost_full
)
1210 || v_read(config
, &buf
->records_lost_wrap
)
1211 || v_read(config
, &buf
->records_lost_big
))
1213 "ring buffer %s, cpu %d: records were lost. Caused by:\n"
1214 " [ %lu buffer full, %lu nest buffer wrap-around, "
1215 "%lu event too big ]\n",
1216 chan
->backend
.name
, cpu
,
1217 v_read(config
, &buf
->records_lost_full
),
1218 v_read(config
, &buf
->records_lost_wrap
),
1219 v_read(config
, &buf
->records_lost_big
));
1221 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
);
1225 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1227 * Only executed when the buffer is finalized, in SWITCH_FLUSH.
1230 void lib_ring_buffer_switch_old_start(struct lib_ring_buffer
*buf
,
1231 struct channel
*chan
,
1232 struct switch_offsets
*offsets
,
1235 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1236 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1237 unsigned long commit_count
;
1239 config
->cb
.buffer_begin(buf
, tsc
, oldidx
);
1242 * Order all writes to buffer before the commit count update that will
1243 * determine that the subbuffer is full.
1245 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1247 * Must write slot data before incrementing commit count. This
1248 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1254 v_add(config
, config
->cb
.subbuffer_header_size(),
1255 &buf
->commit_hot
[oldidx
].cc
);
1256 commit_count
= v_read(config
, &buf
->commit_hot
[oldidx
].cc
);
1257 /* Check if the written buffer has to be delivered */
1258 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1259 commit_count
, oldidx
);
1260 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1261 offsets
->old
, commit_count
,
1262 config
->cb
.subbuffer_header_size());
1266 * lib_ring_buffer_switch_old_end: switch old subbuffer
1268 * Note : offset_old should never be 0 here. It is ok, because we never perform
1269 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1270 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1274 void lib_ring_buffer_switch_old_end(struct lib_ring_buffer
*buf
,
1275 struct channel
*chan
,
1276 struct switch_offsets
*offsets
,
1279 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1280 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1281 unsigned long commit_count
, padding_size
, data_size
;
1283 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1284 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1285 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
);
1288 * Order all writes to buffer before the commit count update that will
1289 * determine that the subbuffer is full.
1291 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1293 * Must write slot data before incrementing commit count. This
1294 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1300 v_add(config
, padding_size
, &buf
->commit_hot
[oldidx
].cc
);
1301 commit_count
= v_read(config
, &buf
->commit_hot
[oldidx
].cc
);
1302 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1303 commit_count
, oldidx
);
1304 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1305 offsets
->old
, commit_count
,
1310 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1312 * This code can be executed unordered : writers may already have written to the
1313 * sub-buffer before this code gets executed, caution. The commit makes sure
1314 * that this code is executed before the deliver of this sub-buffer.
1317 void lib_ring_buffer_switch_new_start(struct lib_ring_buffer
*buf
,
1318 struct channel
*chan
,
1319 struct switch_offsets
*offsets
,
1322 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1323 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1324 unsigned long commit_count
;
1326 config
->cb
.buffer_begin(buf
, tsc
, beginidx
);
1329 * Order all writes to buffer before the commit count update that will
1330 * determine that the subbuffer is full.
1332 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1334 * Must write slot data before incrementing commit count. This
1335 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1341 v_add(config
, config
->cb
.subbuffer_header_size(),
1342 &buf
->commit_hot
[beginidx
].cc
);
1343 commit_count
= v_read(config
, &buf
->commit_hot
[beginidx
].cc
);
1344 /* Check if the written buffer has to be delivered */
1345 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1346 commit_count
, beginidx
);
1347 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, beginidx
,
1348 offsets
->begin
, commit_count
,
1349 config
->cb
.subbuffer_header_size());
1353 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1355 * Calls subbuffer_set_data_size() to set the data size of the current
1356 * sub-buffer. We do not need to perform check_deliver nor commit here,
1357 * since this task will be done by the "commit" of the event for which
1358 * we are currently doing the space reservation.
1361 void lib_ring_buffer_switch_new_end(struct lib_ring_buffer
*buf
,
1362 struct channel
*chan
,
1363 struct switch_offsets
*offsets
,
1366 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1367 unsigned long endidx
, data_size
;
1369 endidx
= subbuf_index(offsets
->end
- 1, chan
);
1370 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1371 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
);
1377 * !0 if execution must be aborted.
1380 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1381 struct lib_ring_buffer
*buf
,
1382 struct channel
*chan
,
1383 struct switch_offsets
*offsets
,
1386 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1387 unsigned long off
, reserve_commit_diff
;
1389 offsets
->begin
= v_read(config
, &buf
->offset
);
1390 offsets
->old
= offsets
->begin
;
1391 offsets
->switch_old_start
= 0;
1392 off
= subbuf_offset(offsets
->begin
, chan
);
1394 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1397 * Ensure we flush the header of an empty subbuffer when doing the
1398 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1399 * total data gathering duration even if there were no records saved
1400 * after the last buffer switch.
1401 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1402 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1403 * subbuffer header as appropriate.
1404 * The next record that reserves space will be responsible for
1405 * populating the following subbuffer header. We choose not to populate
1406 * the next subbuffer header here because we want to be able to use
1407 * SWITCH_ACTIVE for periodical buffer flush and CPU tick_nohz stop
1408 * buffer flush, which must guarantee that all the buffer content
1409 * (records and header timestamps) are visible to the reader. This is
1410 * required for quiescence guarantees for the fusion merge.
1412 if (mode
!= SWITCH_FLUSH
&& !off
)
1413 return -1; /* we do not have to switch : buffer is empty */
1415 if (unlikely(off
== 0)) {
1416 unsigned long sb_index
, commit_count
;
1419 * We are performing a SWITCH_FLUSH. At this stage, there are no
1420 * concurrent writes into the buffer.
1422 * The client does not save any header information. Don't
1423 * switch empty subbuffer on finalize, because it is invalid to
1424 * deliver a completely empty subbuffer.
1426 if (!config
->cb
.subbuffer_header_size())
1429 /* Test new buffer integrity */
1430 sb_index
= subbuf_index(offsets
->begin
, chan
);
1431 commit_count
= v_read(config
,
1432 &buf
->commit_cold
[sb_index
].cc_sb
);
1433 reserve_commit_diff
=
1434 (buf_trunc(offsets
->begin
, chan
)
1435 >> chan
->backend
.num_subbuf_order
)
1436 - (commit_count
& chan
->commit_count_mask
);
1437 if (likely(reserve_commit_diff
== 0)) {
1438 /* Next subbuffer not being written to. */
1439 if (unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1440 subbuf_trunc(offsets
->begin
, chan
)
1441 - subbuf_trunc((unsigned long)
1442 atomic_long_read(&buf
->consumed
), chan
)
1443 >= chan
->backend
.buf_size
)) {
1445 * We do not overwrite non consumed buffers
1446 * and we are full : don't switch.
1451 * Next subbuffer not being written to, and we
1452 * are either in overwrite mode or the buffer is
1453 * not full. It's safe to write in this new
1459 * Next subbuffer reserve offset does not match the
1460 * commit offset. Don't perform switch in
1461 * producer-consumer and overwrite mode. Caused by
1462 * either a writer OOPS or too many nested writes over a
1463 * reserve/commit pair.
1469 * Need to write the subbuffer start header on finalize.
1471 offsets
->switch_old_start
= 1;
1473 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1474 /* Note: old points to the next subbuf at offset 0 */
1475 offsets
->end
= offsets
->begin
;
1480 * Force a sub-buffer switch. This operation is completely reentrant : can be
1481 * called while tracing is active with absolutely no lock held.
1483 * Note, however, that as a v_cmpxchg is used for some atomic
1484 * operations, this function must be called from the CPU which owns the buffer
1485 * for a ACTIVE flush.
1487 void lib_ring_buffer_switch_slow(struct lib_ring_buffer
*buf
, enum switch_mode mode
)
1489 struct channel
*chan
= buf
->backend
.chan
;
1490 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1491 struct switch_offsets offsets
;
1492 unsigned long oldidx
;
1498 * Perform retryable operations.
1501 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1503 return; /* Switch not needed */
1504 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1508 * Atomically update last_tsc. This update races against concurrent
1509 * atomic updates, but the race will always cause supplementary full TSC
1510 * records, never the opposite (missing a full TSC record when it would
1513 save_last_tsc(config
, buf
, tsc
);
1516 * Push the reader if necessary
1518 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1520 oldidx
= subbuf_index(offsets
.old
, chan
);
1521 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
);
1524 * May need to populate header start on SWITCH_FLUSH.
1526 if (offsets
.switch_old_start
) {
1527 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
);
1528 offsets
.old
+= config
->cb
.subbuffer_header_size();
1532 * Switch old subbuffer.
1534 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
);
1536 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow
);
1538 static void remote_switch(void *info
)
1540 struct lib_ring_buffer
*buf
= info
;
1542 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
1545 void lib_ring_buffer_switch_remote(struct lib_ring_buffer
*buf
)
1547 struct channel
*chan
= buf
->backend
.chan
;
1548 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1552 * With global synchronization we don't need to use the IPI scheme.
1554 if (config
->sync
== RING_BUFFER_SYNC_GLOBAL
) {
1555 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
1560 * Taking lock on CPU hotplug to ensure two things: first, that the
1561 * target cpu is not taken concurrently offline while we are within
1562 * smp_call_function_single() (I don't trust that get_cpu() on the
1563 * _local_ CPU actually inhibit CPU hotplug for the _remote_ CPU (to be
1564 * confirmed)). Secondly, if it happens that the CPU is not online, our
1565 * own call to lib_ring_buffer_switch_slow() needs to be protected from
1566 * CPU hotplug handlers, which can also perform a remote subbuffer
1570 ret
= smp_call_function_single(buf
->backend
.cpu
,
1571 remote_switch
, buf
, 1);
1573 /* Remote CPU is offline, do it ourself. */
1574 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
1578 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote
);
1583 * -ENOSPC if event size is too large for packet.
1584 * -ENOBUFS if there is currently not enough space in buffer for the event.
1585 * -EIO if data cannot be written into the buffer for any other reason.
1588 int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer
*buf
,
1589 struct channel
*chan
,
1590 struct switch_offsets
*offsets
,
1591 struct lib_ring_buffer_ctx
*ctx
)
1593 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1594 unsigned long reserve_commit_diff
, offset_cmp
;
1597 offsets
->begin
= offset_cmp
= v_read(config
, &buf
->offset
);
1598 offsets
->old
= offsets
->begin
;
1599 offsets
->switch_new_start
= 0;
1600 offsets
->switch_new_end
= 0;
1601 offsets
->switch_old_end
= 0;
1602 offsets
->pre_header_padding
= 0;
1604 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1605 if ((int64_t) ctx
->tsc
== -EIO
)
1608 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1609 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
1611 if (unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1612 offsets
->switch_new_start
= 1; /* For offsets->begin */
1614 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1616 &offsets
->pre_header_padding
,
1619 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1622 if (unlikely(subbuf_offset(offsets
->begin
, chan
) +
1623 offsets
->size
> chan
->backend
.subbuf_size
)) {
1624 offsets
->switch_old_end
= 1; /* For offsets->old */
1625 offsets
->switch_new_start
= 1; /* For offsets->begin */
1628 if (unlikely(offsets
->switch_new_start
)) {
1629 unsigned long sb_index
, commit_count
;
1632 * We are typically not filling the previous buffer completely.
1634 if (likely(offsets
->switch_old_end
))
1635 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1636 offsets
->begin
= offsets
->begin
1637 + config
->cb
.subbuffer_header_size();
1638 /* Test new buffer integrity */
1639 sb_index
= subbuf_index(offsets
->begin
, chan
);
1641 * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
1642 * lib_ring_buffer_check_deliver() has the matching
1643 * memory barriers required around commit_cold cc_sb
1644 * updates to ensure reserve and commit counter updates
1645 * are not seen reordered when updated by another CPU.
1648 commit_count
= v_read(config
,
1649 &buf
->commit_cold
[sb_index
].cc_sb
);
1650 /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
1652 if (unlikely(offset_cmp
!= v_read(config
, &buf
->offset
))) {
1654 * The reserve counter have been concurrently updated
1655 * while we read the commit counter. This means the
1656 * commit counter we read might not match buf->offset
1657 * due to concurrent update. We therefore need to retry.
1661 reserve_commit_diff
=
1662 (buf_trunc(offsets
->begin
, chan
)
1663 >> chan
->backend
.num_subbuf_order
)
1664 - (commit_count
& chan
->commit_count_mask
);
1665 if (likely(reserve_commit_diff
== 0)) {
1666 /* Next subbuffer not being written to. */
1667 if (unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1668 subbuf_trunc(offsets
->begin
, chan
)
1669 - subbuf_trunc((unsigned long)
1670 atomic_long_read(&buf
->consumed
), chan
)
1671 >= chan
->backend
.buf_size
)) {
1673 * We do not overwrite non consumed buffers
1674 * and we are full : record is lost.
1676 v_inc(config
, &buf
->records_lost_full
);
1680 * Next subbuffer not being written to, and we
1681 * are either in overwrite mode or the buffer is
1682 * not full. It's safe to write in this new
1688 * Next subbuffer reserve offset does not match the
1689 * commit offset, and this did not involve update to the
1690 * reserve counter. Drop record in producer-consumer and
1691 * overwrite mode. Caused by either a writer OOPS or
1692 * too many nested writes over a reserve/commit pair.
1694 v_inc(config
, &buf
->records_lost_wrap
);
1698 config
->cb
.record_header_size(config
, chan
,
1700 &offsets
->pre_header_padding
,
1703 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1706 if (unlikely(subbuf_offset(offsets
->begin
, chan
)
1707 + offsets
->size
> chan
->backend
.subbuf_size
)) {
1709 * Record too big for subbuffers, report error, don't
1710 * complete the sub-buffer switch.
1712 v_inc(config
, &buf
->records_lost_big
);
1716 * We just made a successful buffer switch and the
1717 * record fits in the new subbuffer. Let's write.
1722 * Record fits in the current buffer and we are not on a switch
1723 * boundary. It's safe to write.
1726 offsets
->end
= offsets
->begin
+ offsets
->size
;
1728 if (unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
1730 * The offset_end will fall at the very beginning of the next
1733 offsets
->switch_new_end
= 1; /* For offsets->begin */
1739 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1740 * @ctx: ring buffer context.
1742 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
1743 * -EIO for other errors, else returns 0.
1744 * It will take care of sub-buffer switching.
1746 int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx
*ctx
)
1748 struct channel
*chan
= ctx
->chan
;
1749 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1750 struct lib_ring_buffer
*buf
;
1751 struct switch_offsets offsets
;
1754 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
1755 buf
= per_cpu_ptr(chan
->backend
.buf
, ctx
->cpu
);
1757 buf
= chan
->backend
.buf
;
1763 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
1767 } while (unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
1772 * Atomically update last_tsc. This update races against concurrent
1773 * atomic updates, but the race will always cause supplementary full TSC
1774 * records, never the opposite (missing a full TSC record when it would
1777 save_last_tsc(config
, buf
, ctx
->tsc
);
1780 * Push the reader if necessary
1782 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
1785 * Clear noref flag for this subbuffer.
1787 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1788 subbuf_index(offsets
.end
- 1, chan
));
1791 * Switch old subbuffer if needed.
1793 if (unlikely(offsets
.switch_old_end
)) {
1794 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1795 subbuf_index(offsets
.old
- 1, chan
));
1796 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
);
1800 * Populate new subbuffer.
1802 if (unlikely(offsets
.switch_new_start
))
1803 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
);
1805 if (unlikely(offsets
.switch_new_end
))
1806 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
);
1808 ctx
->slot_size
= offsets
.size
;
1809 ctx
->pre_offset
= offsets
.begin
;
1810 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
1813 EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow
);
1815 int __init
init_lib_ring_buffer_frontend(void)
1819 for_each_possible_cpu(cpu
)
1820 spin_lock_init(&per_cpu(ring_buffer_nohz_lock
, cpu
));
1824 module_init(init_lib_ring_buffer_frontend
);
1826 void __exit
exit_lib_ring_buffer_frontend(void)
1830 module_exit(exit_lib_ring_buffer_frontend
);