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"
63 #include "../../wrapper/atomic.h"
64 #include "../../wrapper/percpu-defs.h"
67 * Internal structure representing offsets to use at a sub-buffer switch.
69 struct switch_offsets
{
70 unsigned long begin
, end
, old
;
71 size_t pre_header_padding
, size
;
72 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
83 static ATOMIC_NOTIFIER_HEAD(tick_nohz_notifier
);
84 #endif /* CONFIG_NO_HZ */
86 static DEFINE_PER_CPU(spinlock_t
, ring_buffer_nohz_lock
);
88 DEFINE_PER_CPU(unsigned int, lib_ring_buffer_nesting
);
89 EXPORT_PER_CPU_SYMBOL(lib_ring_buffer_nesting
);
92 void lib_ring_buffer_print_errors(struct channel
*chan
,
93 struct lib_ring_buffer
*buf
, int cpu
);
95 void _lib_ring_buffer_switch_remote(struct lib_ring_buffer
*buf
,
96 enum switch_mode mode
);
99 * Must be called under cpu hotplug protection.
101 void lib_ring_buffer_free(struct lib_ring_buffer
*buf
)
103 struct channel
*chan
= buf
->backend
.chan
;
105 lib_ring_buffer_print_errors(chan
, buf
, buf
->backend
.cpu
);
106 kfree(buf
->commit_hot
);
107 kfree(buf
->commit_cold
);
109 lib_ring_buffer_backend_free(&buf
->backend
);
113 * lib_ring_buffer_reset - Reset ring buffer to initial values.
116 * Effectively empty the ring buffer. Should be called when the buffer is not
117 * used for writing. The ring buffer can be opened for reading, but the reader
118 * should not be using the iterator concurrently with reset. The previous
119 * current iterator record is reset.
121 void lib_ring_buffer_reset(struct lib_ring_buffer
*buf
)
123 struct channel
*chan
= buf
->backend
.chan
;
124 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
128 * Reset iterator first. It will put the subbuffer if it currently holds
131 lib_ring_buffer_iterator_reset(buf
);
132 v_set(config
, &buf
->offset
, 0);
133 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
134 v_set(config
, &buf
->commit_hot
[i
].cc
, 0);
135 v_set(config
, &buf
->commit_hot
[i
].seq
, 0);
136 v_set(config
, &buf
->commit_cold
[i
].cc_sb
, 0);
138 atomic_long_set(&buf
->consumed
, 0);
139 atomic_set(&buf
->record_disabled
, 0);
140 v_set(config
, &buf
->last_tsc
, 0);
141 lib_ring_buffer_backend_reset(&buf
->backend
);
142 /* Don't reset number of active readers */
143 v_set(config
, &buf
->records_lost_full
, 0);
144 v_set(config
, &buf
->records_lost_wrap
, 0);
145 v_set(config
, &buf
->records_lost_big
, 0);
146 v_set(config
, &buf
->records_count
, 0);
147 v_set(config
, &buf
->records_overrun
, 0);
150 EXPORT_SYMBOL_GPL(lib_ring_buffer_reset
);
153 * channel_reset - Reset channel to initial values.
156 * Effectively empty the channel. Should be called when the channel is not used
157 * for writing. The channel can be opened for reading, but the reader should not
158 * be using the iterator concurrently with reset. The previous current iterator
161 void channel_reset(struct channel
*chan
)
164 * Reset iterators first. Will put the subbuffer if held for reading.
166 channel_iterator_reset(chan
);
167 atomic_set(&chan
->record_disabled
, 0);
168 /* Don't reset commit_count_mask, still valid */
169 channel_backend_reset(&chan
->backend
);
170 /* Don't reset switch/read timer interval */
171 /* Don't reset notifiers and notifier enable bits */
172 /* Don't reset reader reference count */
174 EXPORT_SYMBOL_GPL(channel_reset
);
177 * Must be called under cpu hotplug protection.
179 int lib_ring_buffer_create(struct lib_ring_buffer
*buf
,
180 struct channel_backend
*chanb
, int cpu
)
182 const struct lib_ring_buffer_config
*config
= &chanb
->config
;
183 struct channel
*chan
= container_of(chanb
, struct channel
, backend
);
184 void *priv
= chanb
->priv
;
185 size_t subbuf_header_size
;
189 /* Test for cpu hotplug */
190 if (buf
->backend
.allocated
)
194 * Paranoia: per cpu dynamic allocation is not officially documented as
195 * zeroing the memory, so let's do it here too, just in case.
197 memset(buf
, 0, sizeof(*buf
));
199 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
, cpu
);
204 kzalloc_node(ALIGN(sizeof(*buf
->commit_hot
)
205 * chan
->backend
.num_subbuf
,
206 1 << INTERNODE_CACHE_SHIFT
),
207 GFP_KERNEL
, cpu_to_node(max(cpu
, 0)));
208 if (!buf
->commit_hot
) {
214 kzalloc_node(ALIGN(sizeof(*buf
->commit_cold
)
215 * chan
->backend
.num_subbuf
,
216 1 << INTERNODE_CACHE_SHIFT
),
217 GFP_KERNEL
, cpu_to_node(max(cpu
, 0)));
218 if (!buf
->commit_cold
) {
223 init_waitqueue_head(&buf
->read_wait
);
224 init_waitqueue_head(&buf
->write_wait
);
225 raw_spin_lock_init(&buf
->raw_tick_nohz_spinlock
);
228 * Write the subbuffer header for first subbuffer so we know the total
229 * duration of data gathering.
231 subbuf_header_size
= config
->cb
.subbuffer_header_size();
232 v_set(config
, &buf
->offset
, subbuf_header_size
);
233 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_wsb
[0].id
);
234 tsc
= config
->cb
.ring_buffer_clock_read(buf
->backend
.chan
);
235 config
->cb
.buffer_begin(buf
, tsc
, 0);
236 v_add(config
, subbuf_header_size
, &buf
->commit_hot
[0].cc
);
238 if (config
->cb
.buffer_create
) {
239 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
);
245 * Ensure the buffer is ready before setting it to allocated and setting
247 * Used for cpu hotplug vs cpumask iteration.
250 buf
->backend
.allocated
= 1;
252 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
253 CHAN_WARN_ON(chan
, cpumask_test_cpu(cpu
,
254 chan
->backend
.cpumask
));
255 cpumask_set_cpu(cpu
, chan
->backend
.cpumask
);
262 kfree(buf
->commit_cold
);
264 kfree(buf
->commit_hot
);
266 lib_ring_buffer_backend_free(&buf
->backend
);
270 static void switch_buffer_timer(unsigned long data
)
272 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
273 struct channel
*chan
= buf
->backend
.chan
;
274 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
277 * Only flush buffers periodically if readers are active.
279 if (atomic_long_read(&buf
->active_readers
))
280 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
282 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
283 mod_timer_pinned(&buf
->switch_timer
,
284 jiffies
+ chan
->switch_timer_interval
);
286 mod_timer(&buf
->switch_timer
,
287 jiffies
+ chan
->switch_timer_interval
);
291 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
293 static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer
*buf
)
295 struct channel
*chan
= buf
->backend
.chan
;
296 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
298 if (!chan
->switch_timer_interval
|| buf
->switch_timer_enabled
)
300 init_timer(&buf
->switch_timer
);
301 buf
->switch_timer
.function
= switch_buffer_timer
;
302 buf
->switch_timer
.expires
= jiffies
+ chan
->switch_timer_interval
;
303 buf
->switch_timer
.data
= (unsigned long)buf
;
304 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
305 add_timer_on(&buf
->switch_timer
, buf
->backend
.cpu
);
307 add_timer(&buf
->switch_timer
);
308 buf
->switch_timer_enabled
= 1;
312 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
314 static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer
*buf
)
316 struct channel
*chan
= buf
->backend
.chan
;
318 if (!chan
->switch_timer_interval
|| !buf
->switch_timer_enabled
)
321 del_timer_sync(&buf
->switch_timer
);
322 buf
->switch_timer_enabled
= 0;
326 * Polling timer to check the channels for data.
328 static void read_buffer_timer(unsigned long data
)
330 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
331 struct channel
*chan
= buf
->backend
.chan
;
332 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
334 CHAN_WARN_ON(chan
, !buf
->backend
.allocated
);
336 if (atomic_long_read(&buf
->active_readers
)
337 && lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
338 wake_up_interruptible(&buf
->read_wait
);
339 wake_up_interruptible(&chan
->read_wait
);
342 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
343 mod_timer_pinned(&buf
->read_timer
,
344 jiffies
+ chan
->read_timer_interval
);
346 mod_timer(&buf
->read_timer
,
347 jiffies
+ chan
->read_timer_interval
);
351 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
353 static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer
*buf
)
355 struct channel
*chan
= buf
->backend
.chan
;
356 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
358 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
359 || !chan
->read_timer_interval
360 || buf
->read_timer_enabled
)
363 init_timer(&buf
->read_timer
);
364 buf
->read_timer
.function
= read_buffer_timer
;
365 buf
->read_timer
.expires
= jiffies
+ chan
->read_timer_interval
;
366 buf
->read_timer
.data
= (unsigned long)buf
;
368 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
369 add_timer_on(&buf
->read_timer
, buf
->backend
.cpu
);
371 add_timer(&buf
->read_timer
);
372 buf
->read_timer_enabled
= 1;
376 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
378 static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer
*buf
)
380 struct channel
*chan
= buf
->backend
.chan
;
381 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
383 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
384 || !chan
->read_timer_interval
385 || !buf
->read_timer_enabled
)
388 del_timer_sync(&buf
->read_timer
);
390 * do one more check to catch data that has been written in the last
393 if (lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
394 wake_up_interruptible(&buf
->read_wait
);
395 wake_up_interruptible(&chan
->read_wait
);
397 buf
->read_timer_enabled
= 0;
400 #ifdef CONFIG_HOTPLUG_CPU
402 * lib_ring_buffer_cpu_hp_callback - CPU hotplug callback
403 * @nb: notifier block
404 * @action: hotplug action to take
407 * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
410 int lib_ring_buffer_cpu_hp_callback(struct notifier_block
*nb
,
411 unsigned long action
,
414 unsigned int cpu
= (unsigned long)hcpu
;
415 struct channel
*chan
= container_of(nb
, struct channel
,
417 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
, cpu
);
418 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
420 if (!chan
->cpu_hp_enable
)
423 CHAN_WARN_ON(chan
, config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
);
426 case CPU_DOWN_FAILED
:
427 case CPU_DOWN_FAILED_FROZEN
:
429 case CPU_ONLINE_FROZEN
:
430 wake_up_interruptible(&chan
->hp_wait
);
431 lib_ring_buffer_start_switch_timer(buf
);
432 lib_ring_buffer_start_read_timer(buf
);
435 case CPU_DOWN_PREPARE
:
436 case CPU_DOWN_PREPARE_FROZEN
:
437 lib_ring_buffer_stop_switch_timer(buf
);
438 lib_ring_buffer_stop_read_timer(buf
);
442 case CPU_DEAD_FROZEN
:
444 * Performing a buffer switch on a remote CPU. Performed by
445 * the CPU responsible for doing the hotunplug after the target
446 * CPU stopped running completely. Ensures that all data
447 * from that remote CPU is flushed.
449 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
458 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
460 * For per-cpu buffers, call the reader wakeups before switching the buffer, so
461 * that wake-up-tracing generated events are flushed before going idle (in
462 * tick_nohz). We test if the spinlock is locked to deal with the race where
463 * readers try to sample the ring buffer before we perform the switch. We let
464 * the readers retry in that case. If there is data in the buffer, the wake up
465 * is going to forbid the CPU running the reader thread from going idle.
467 static int notrace
ring_buffer_tick_nohz_callback(struct notifier_block
*nb
,
471 struct channel
*chan
= container_of(nb
, struct channel
,
473 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
474 struct lib_ring_buffer
*buf
;
475 int cpu
= smp_processor_id();
477 if (config
->alloc
!= RING_BUFFER_ALLOC_PER_CPU
) {
479 * We don't support keeping the system idle with global buffers
480 * and streaming active. In order to do so, we would need to
481 * sample a non-nohz-cpumask racelessly with the nohz updates
482 * without adding synchronization overhead to nohz. Leave this
483 * use-case out for now.
488 buf
= channel_get_ring_buffer(config
, chan
, cpu
);
490 case TICK_NOHZ_FLUSH
:
491 raw_spin_lock(&buf
->raw_tick_nohz_spinlock
);
492 if (config
->wakeup
== RING_BUFFER_WAKEUP_BY_TIMER
493 && chan
->read_timer_interval
494 && atomic_long_read(&buf
->active_readers
)
495 && (lib_ring_buffer_poll_deliver(config
, buf
, chan
)
496 || lib_ring_buffer_pending_data(config
, buf
, chan
))) {
497 wake_up_interruptible(&buf
->read_wait
);
498 wake_up_interruptible(&chan
->read_wait
);
500 if (chan
->switch_timer_interval
)
501 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
502 raw_spin_unlock(&buf
->raw_tick_nohz_spinlock
);
505 spin_lock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
506 lib_ring_buffer_stop_switch_timer(buf
);
507 lib_ring_buffer_stop_read_timer(buf
);
508 spin_unlock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
510 case TICK_NOHZ_RESTART
:
511 spin_lock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
512 lib_ring_buffer_start_read_timer(buf
);
513 lib_ring_buffer_start_switch_timer(buf
);
514 spin_unlock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
521 void notrace
lib_ring_buffer_tick_nohz_flush(void)
523 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_FLUSH
,
527 void notrace
lib_ring_buffer_tick_nohz_stop(void)
529 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_STOP
,
533 void notrace
lib_ring_buffer_tick_nohz_restart(void)
535 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_RESTART
,
538 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
543 static void channel_unregister_notifiers(struct channel
*chan
)
545 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
548 channel_iterator_unregister_notifiers(chan
);
549 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
552 * Remove the nohz notifier first, so we are certain we stop
555 atomic_notifier_chain_unregister(&tick_nohz_notifier
,
556 &chan
->tick_nohz_notifier
);
558 * ring_buffer_nohz_lock will not be needed below, because
559 * we just removed the notifiers, which were the only source of
562 #endif /* CONFIG_NO_HZ */
563 #ifdef CONFIG_HOTPLUG_CPU
565 chan
->cpu_hp_enable
= 0;
566 for_each_online_cpu(cpu
) {
567 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
569 lib_ring_buffer_stop_switch_timer(buf
);
570 lib_ring_buffer_stop_read_timer(buf
);
573 unregister_cpu_notifier(&chan
->cpu_hp_notifier
);
575 for_each_possible_cpu(cpu
) {
576 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
578 lib_ring_buffer_stop_switch_timer(buf
);
579 lib_ring_buffer_stop_read_timer(buf
);
583 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
585 lib_ring_buffer_stop_switch_timer(buf
);
586 lib_ring_buffer_stop_read_timer(buf
);
588 channel_backend_unregister_notifiers(&chan
->backend
);
591 static void lib_ring_buffer_set_quiescent(struct lib_ring_buffer
*buf
)
593 if (!buf
->quiescent
) {
594 buf
->quiescent
= true;
595 _lib_ring_buffer_switch_remote(buf
, SWITCH_FLUSH
);
599 static void lib_ring_buffer_clear_quiescent(struct lib_ring_buffer
*buf
)
601 buf
->quiescent
= false;
604 void lib_ring_buffer_set_quiescent_channel(struct channel
*chan
)
607 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
609 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
611 for_each_channel_cpu(cpu
, chan
) {
612 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
615 lib_ring_buffer_set_quiescent(buf
);
619 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
621 lib_ring_buffer_set_quiescent(buf
);
624 EXPORT_SYMBOL_GPL(lib_ring_buffer_set_quiescent_channel
);
626 void lib_ring_buffer_clear_quiescent_channel(struct channel
*chan
)
629 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
631 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
633 for_each_channel_cpu(cpu
, chan
) {
634 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
637 lib_ring_buffer_clear_quiescent(buf
);
641 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
643 lib_ring_buffer_clear_quiescent(buf
);
646 EXPORT_SYMBOL_GPL(lib_ring_buffer_clear_quiescent_channel
);
648 static void channel_free(struct channel
*chan
)
650 if (chan
->backend
.release_priv_ops
) {
651 chan
->backend
.release_priv_ops(chan
->backend
.priv_ops
);
653 channel_iterator_free(chan
);
654 channel_backend_free(&chan
->backend
);
659 * channel_create - Create channel.
660 * @config: ring buffer instance configuration
661 * @name: name of the channel
662 * @priv: ring buffer client private data
663 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
664 * address mapping. It is used only by RING_BUFFER_STATIC
665 * configuration. It can be set to NULL for other backends.
666 * @subbuf_size: subbuffer size
667 * @num_subbuf: number of subbuffers
668 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
669 * padding to let readers get those sub-buffers.
670 * Used for live streaming.
671 * @read_timer_interval: Time interval (in us) to wake up pending readers.
674 * Returns NULL on failure.
676 struct channel
*channel_create(const struct lib_ring_buffer_config
*config
,
677 const char *name
, void *priv
, void *buf_addr
,
679 size_t num_subbuf
, unsigned int switch_timer_interval
,
680 unsigned int read_timer_interval
)
683 struct channel
*chan
;
685 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
686 read_timer_interval
))
689 chan
= kzalloc(sizeof(struct channel
), GFP_KERNEL
);
693 ret
= channel_backend_init(&chan
->backend
, name
, config
, priv
,
694 subbuf_size
, num_subbuf
);
698 ret
= channel_iterator_init(chan
);
700 goto error_free_backend
;
702 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
703 chan
->switch_timer_interval
= usecs_to_jiffies(switch_timer_interval
);
704 chan
->read_timer_interval
= usecs_to_jiffies(read_timer_interval
);
705 kref_init(&chan
->ref
);
706 init_waitqueue_head(&chan
->read_wait
);
707 init_waitqueue_head(&chan
->hp_wait
);
709 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
710 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
711 /* Only benefit from NO_HZ idle with per-cpu buffers for now. */
712 chan
->tick_nohz_notifier
.notifier_call
=
713 ring_buffer_tick_nohz_callback
;
714 chan
->tick_nohz_notifier
.priority
= ~0U;
715 atomic_notifier_chain_register(&tick_nohz_notifier
,
716 &chan
->tick_nohz_notifier
);
717 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
720 * In case of non-hotplug cpu, if the ring-buffer is allocated
721 * in early initcall, it will not be notified of secondary cpus.
722 * In that off case, we need to allocate for all possible cpus.
724 #ifdef CONFIG_HOTPLUG_CPU
725 chan
->cpu_hp_notifier
.notifier_call
=
726 lib_ring_buffer_cpu_hp_callback
;
727 chan
->cpu_hp_notifier
.priority
= 6;
728 register_cpu_notifier(&chan
->cpu_hp_notifier
);
731 for_each_online_cpu(cpu
) {
732 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
734 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
735 lib_ring_buffer_start_switch_timer(buf
);
736 lib_ring_buffer_start_read_timer(buf
);
737 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
739 chan
->cpu_hp_enable
= 1;
742 for_each_possible_cpu(cpu
) {
743 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
745 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
746 lib_ring_buffer_start_switch_timer(buf
);
747 lib_ring_buffer_start_read_timer(buf
);
748 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
752 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
754 lib_ring_buffer_start_switch_timer(buf
);
755 lib_ring_buffer_start_read_timer(buf
);
761 channel_backend_free(&chan
->backend
);
766 EXPORT_SYMBOL_GPL(channel_create
);
769 void channel_release(struct kref
*kref
)
771 struct channel
*chan
= container_of(kref
, struct channel
, ref
);
776 * channel_destroy - Finalize, wait for q.s. and destroy channel.
777 * @chan: channel to destroy
780 * Call "destroy" callback, finalize channels, and then decrement the
781 * channel reference count. Note that when readers have completed data
782 * consumption of finalized channels, get_subbuf() will return -ENODATA.
783 * They should release their handle at that point. Returns the private
786 void *channel_destroy(struct channel
*chan
)
789 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
792 channel_unregister_notifiers(chan
);
794 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
796 * No need to hold cpu hotplug, because all notifiers have been
799 for_each_channel_cpu(cpu
, chan
) {
800 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
803 if (config
->cb
.buffer_finalize
)
804 config
->cb
.buffer_finalize(buf
,
807 if (buf
->backend
.allocated
)
808 lib_ring_buffer_set_quiescent(buf
);
810 * Perform flush before writing to finalized.
813 ACCESS_ONCE(buf
->finalized
) = 1;
814 wake_up_interruptible(&buf
->read_wait
);
817 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
819 if (config
->cb
.buffer_finalize
)
820 config
->cb
.buffer_finalize(buf
, chan
->backend
.priv
, -1);
821 if (buf
->backend
.allocated
)
822 lib_ring_buffer_set_quiescent(buf
);
824 * Perform flush before writing to finalized.
827 ACCESS_ONCE(buf
->finalized
) = 1;
828 wake_up_interruptible(&buf
->read_wait
);
830 ACCESS_ONCE(chan
->finalized
) = 1;
831 wake_up_interruptible(&chan
->hp_wait
);
832 wake_up_interruptible(&chan
->read_wait
);
833 priv
= chan
->backend
.priv
;
834 kref_put(&chan
->ref
, channel_release
);
837 EXPORT_SYMBOL_GPL(channel_destroy
);
839 struct lib_ring_buffer
*channel_get_ring_buffer(
840 const struct lib_ring_buffer_config
*config
,
841 struct channel
*chan
, int cpu
)
843 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
)
844 return chan
->backend
.buf
;
846 return per_cpu_ptr(chan
->backend
.buf
, cpu
);
848 EXPORT_SYMBOL_GPL(channel_get_ring_buffer
);
850 int lib_ring_buffer_open_read(struct lib_ring_buffer
*buf
)
852 struct channel
*chan
= buf
->backend
.chan
;
854 if (!atomic_long_add_unless(&buf
->active_readers
, 1, 1))
856 kref_get(&chan
->ref
);
857 lttng_smp_mb__after_atomic();
860 EXPORT_SYMBOL_GPL(lib_ring_buffer_open_read
);
862 void lib_ring_buffer_release_read(struct lib_ring_buffer
*buf
)
864 struct channel
*chan
= buf
->backend
.chan
;
866 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
867 lttng_smp_mb__before_atomic();
868 atomic_long_dec(&buf
->active_readers
);
869 kref_put(&chan
->ref
, channel_release
);
871 EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read
);
874 * Promote compiler barrier to a smp_mb().
875 * For the specific ring buffer case, this IPI call should be removed if the
876 * architecture does not reorder writes. This should eventually be provided by
877 * a separate architecture-specific infrastructure.
879 static void remote_mb(void *info
)
885 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
887 * @consumed: consumed count indicating the position where to read
888 * @produced: produced count, indicates position when to stop reading
890 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
891 * data to read at consumed position, or 0 if the get operation succeeds.
892 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
895 int lib_ring_buffer_snapshot(struct lib_ring_buffer
*buf
,
896 unsigned long *consumed
, unsigned long *produced
)
898 struct channel
*chan
= buf
->backend
.chan
;
899 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
900 unsigned long consumed_cur
, write_offset
;
904 * First, ensure we perform a "final" flush onto the stream. This will
905 * ensure we create a packet of padding if we encounter an empty
906 * packet. This ensures the time-stamps right before the snapshot is
907 * used as end of packet timestamp.
910 _lib_ring_buffer_switch_remote(buf
, SWITCH_FLUSH
);
913 finalized
= ACCESS_ONCE(buf
->finalized
);
915 * Read finalized before counters.
918 consumed_cur
= atomic_long_read(&buf
->consumed
);
920 * No need to issue a memory barrier between consumed count read and
921 * write offset read, because consumed count can only change
922 * concurrently in overwrite mode, and we keep a sequence counter
923 * identifier derived from the write offset to check we are getting
924 * the same sub-buffer we are expecting (the sub-buffers are atomically
925 * "tagged" upon writes, tags are checked upon read).
927 write_offset
= v_read(config
, &buf
->offset
);
930 * Check that we are not about to read the same subbuffer in
931 * which the writer head is.
933 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
937 *consumed
= consumed_cur
;
938 *produced
= subbuf_trunc(write_offset
, chan
);
944 * The memory barriers __wait_event()/wake_up_interruptible() take care
945 * of "raw_spin_is_locked" memory ordering.
949 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
954 EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot
);
957 * lib_ring_buffer_put_snapshot - move consumed counter forward
959 * Should only be called from consumer context.
961 * @consumed_new: new consumed count value
963 void lib_ring_buffer_move_consumer(struct lib_ring_buffer
*buf
,
964 unsigned long consumed_new
)
966 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
967 struct channel
*chan
= bufb
->chan
;
968 unsigned long consumed
;
970 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
973 * Only push the consumed value forward.
974 * If the consumed cmpxchg fails, this is because we have been pushed by
975 * the writer in flight recorder mode.
977 consumed
= atomic_long_read(&buf
->consumed
);
978 while ((long) consumed
- (long) consumed_new
< 0)
979 consumed
= atomic_long_cmpxchg(&buf
->consumed
, consumed
,
981 /* Wake-up the metadata producer */
982 wake_up_interruptible(&buf
->write_wait
);
984 EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer
);
987 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
989 * @consumed: consumed count indicating the position where to read
991 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
992 * data to read at consumed position, or 0 if the get operation succeeds.
993 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
995 int lib_ring_buffer_get_subbuf(struct lib_ring_buffer
*buf
,
996 unsigned long consumed
)
998 struct channel
*chan
= buf
->backend
.chan
;
999 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1000 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
1004 if (buf
->get_subbuf
) {
1006 * Reader is trying to get a subbuffer twice.
1008 CHAN_WARN_ON(chan
, 1);
1012 finalized
= ACCESS_ONCE(buf
->finalized
);
1014 * Read finalized before counters.
1017 consumed_cur
= atomic_long_read(&buf
->consumed
);
1018 consumed_idx
= subbuf_index(consumed
, chan
);
1019 commit_count
= v_read(config
, &buf
->commit_cold
[consumed_idx
].cc_sb
);
1021 * Make sure we read the commit count before reading the buffer
1022 * data and the write offset. Correct consumed offset ordering
1023 * wrt commit count is insured by the use of cmpxchg to update
1024 * the consumed offset.
1025 * smp_call_function_single can fail if the remote CPU is offline,
1026 * this is OK because then there is no wmb to execute there.
1027 * If our thread is executing on the same CPU as the on the buffers
1028 * belongs to, we don't have to synchronize it at all. If we are
1029 * migrated, the scheduler will take care of the memory barriers.
1030 * Normally, smp_call_function_single() should ensure program order when
1031 * executing the remote function, which implies that it surrounds the
1032 * function execution with :
1043 * However, smp_call_function_single() does not seem to clearly execute
1044 * such barriers. It depends on spinlock semantic to provide the barrier
1045 * before executing the IPI and, when busy-looping, csd_lock_wait only
1046 * executes smp_mb() when it has to wait for the other CPU.
1048 * I don't trust this code. Therefore, let's add the smp_mb() sequence
1049 * required ourself, even if duplicated. It has no performance impact
1052 * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs
1053 * read and write vs write. They do not ensure core synchronization. We
1054 * really have to ensure total order between the 3 barriers running on
1057 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1058 if (config
->sync
== RING_BUFFER_SYNC_PER_CPU
1059 && config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
1060 if (raw_smp_processor_id() != buf
->backend
.cpu
) {
1061 /* Total order with IPI handler smp_mb() */
1063 smp_call_function_single(buf
->backend
.cpu
,
1064 remote_mb
, NULL
, 1);
1065 /* Total order with IPI handler smp_mb() */
1069 /* Total order with IPI handler smp_mb() */
1071 smp_call_function(remote_mb
, NULL
, 1);
1072 /* Total order with IPI handler smp_mb() */
1077 * Local rmb to match the remote wmb to read the commit count
1078 * before the buffer data and the write offset.
1083 write_offset
= v_read(config
, &buf
->offset
);
1086 * Check that the buffer we are getting is after or at consumed_cur
1089 if ((long) subbuf_trunc(consumed
, chan
)
1090 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
1094 * Check that the subbuffer we are trying to consume has been
1095 * already fully committed.
1097 if (((commit_count
- chan
->backend
.subbuf_size
)
1098 & chan
->commit_count_mask
)
1099 - (buf_trunc(consumed
, chan
)
1100 >> chan
->backend
.num_subbuf_order
)
1105 * Check that we are not about to read the same subbuffer in
1106 * which the writer head is.
1108 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed
, chan
)
1113 * Failure to get the subbuffer causes a busy-loop retry without going
1114 * to a wait queue. These are caused by short-lived race windows where
1115 * the writer is getting access to a subbuffer we were trying to get
1116 * access to. Also checks that the "consumed" buffer count we are
1117 * looking for matches the one contained in the subbuffer id.
1119 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1120 consumed_idx
, buf_trunc_val(consumed
, chan
));
1123 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1125 buf
->get_subbuf_consumed
= consumed
;
1126 buf
->get_subbuf
= 1;
1132 * The memory barriers __wait_event()/wake_up_interruptible() take care
1133 * of "raw_spin_is_locked" memory ordering.
1137 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
1142 EXPORT_SYMBOL_GPL(lib_ring_buffer_get_subbuf
);
1145 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1148 void lib_ring_buffer_put_subbuf(struct lib_ring_buffer
*buf
)
1150 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1151 struct channel
*chan
= bufb
->chan
;
1152 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1153 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
1155 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
1157 if (!buf
->get_subbuf
) {
1159 * Reader puts a subbuffer it did not get.
1161 CHAN_WARN_ON(chan
, 1);
1164 consumed
= buf
->get_subbuf_consumed
;
1165 buf
->get_subbuf
= 0;
1168 * Clear the records_unread counter. (overruns counter)
1169 * Can still be non-zero if a file reader simply grabbed the data
1170 * without using iterators.
1171 * Can be below zero if an iterator is used on a snapshot more than
1174 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1175 v_add(config
, v_read(config
,
1176 &bufb
->array
[read_sb_bindex
]->records_unread
),
1177 &bufb
->records_read
);
1178 v_set(config
, &bufb
->array
[read_sb_bindex
]->records_unread
, 0);
1179 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1180 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1181 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1184 * Exchange the reader subbuffer with the one we put in its place in the
1185 * writer subbuffer table. Expect the original consumed count. If
1186 * update_read_sb_index fails, this is because the writer updated the
1187 * subbuffer concurrently. We should therefore keep the subbuffer we
1188 * currently have: it has become invalid to try reading this sub-buffer
1189 * consumed count value anyway.
1191 consumed_idx
= subbuf_index(consumed
, chan
);
1192 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1193 consumed_idx
, buf_trunc_val(consumed
, chan
));
1195 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1196 * if the writer concurrently updated it.
1199 EXPORT_SYMBOL_GPL(lib_ring_buffer_put_subbuf
);
1202 * cons_offset is an iterator on all subbuffer offsets between the reader
1203 * position and the writer position. (inclusive)
1206 void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer
*buf
,
1207 struct channel
*chan
,
1208 unsigned long cons_offset
,
1211 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1212 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1214 cons_idx
= subbuf_index(cons_offset
, chan
);
1215 commit_count
= v_read(config
, &buf
->commit_hot
[cons_idx
].cc
);
1216 commit_count_sb
= v_read(config
, &buf
->commit_cold
[cons_idx
].cc_sb
);
1218 if (subbuf_offset(commit_count
, chan
) != 0)
1220 "ring buffer %s, cpu %d: "
1221 "commit count in subbuffer %lu,\n"
1222 "expecting multiples of %lu bytes\n"
1223 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1224 chan
->backend
.name
, cpu
, cons_idx
,
1225 chan
->backend
.subbuf_size
,
1226 commit_count
, commit_count_sb
);
1228 printk(KERN_DEBUG
"ring buffer: %s, cpu %d: %lu bytes committed\n",
1229 chan
->backend
.name
, cpu
, commit_count
);
1233 void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer
*buf
,
1234 struct channel
*chan
,
1235 void *priv
, int cpu
)
1237 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1238 unsigned long write_offset
, cons_offset
;
1241 * No need to order commit_count, write_offset and cons_offset reads
1242 * because we execute at teardown when no more writer nor reader
1243 * references are left.
1245 write_offset
= v_read(config
, &buf
->offset
);
1246 cons_offset
= atomic_long_read(&buf
->consumed
);
1247 if (write_offset
!= cons_offset
)
1249 "ring buffer %s, cpu %d: "
1250 "non-consumed data\n"
1251 " [ %lu bytes written, %lu bytes read ]\n",
1252 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1254 for (cons_offset
= atomic_long_read(&buf
->consumed
);
1255 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1258 cons_offset
= subbuf_align(cons_offset
, chan
))
1259 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1264 void lib_ring_buffer_print_errors(struct channel
*chan
,
1265 struct lib_ring_buffer
*buf
, int cpu
)
1267 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1268 void *priv
= chan
->backend
.priv
;
1270 if (!strcmp(chan
->backend
.name
, "relay-metadata")) {
1271 printk(KERN_DEBUG
"ring buffer %s: %lu records written, "
1272 "%lu records overrun\n",
1274 v_read(config
, &buf
->records_count
),
1275 v_read(config
, &buf
->records_overrun
));
1277 printk(KERN_DEBUG
"ring buffer %s, cpu %d: %lu records written, "
1278 "%lu records overrun\n",
1279 chan
->backend
.name
, cpu
,
1280 v_read(config
, &buf
->records_count
),
1281 v_read(config
, &buf
->records_overrun
));
1283 if (v_read(config
, &buf
->records_lost_full
)
1284 || v_read(config
, &buf
->records_lost_wrap
)
1285 || v_read(config
, &buf
->records_lost_big
))
1287 "ring buffer %s, cpu %d: records were lost. Caused by:\n"
1288 " [ %lu buffer full, %lu nest buffer wrap-around, "
1289 "%lu event too big ]\n",
1290 chan
->backend
.name
, cpu
,
1291 v_read(config
, &buf
->records_lost_full
),
1292 v_read(config
, &buf
->records_lost_wrap
),
1293 v_read(config
, &buf
->records_lost_big
));
1295 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
);
1299 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1301 * Only executed by SWITCH_FLUSH, which can be issued while tracing is active
1302 * or at buffer finalization (destroy).
1305 void lib_ring_buffer_switch_old_start(struct lib_ring_buffer
*buf
,
1306 struct channel
*chan
,
1307 struct switch_offsets
*offsets
,
1310 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1311 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1312 unsigned long commit_count
;
1314 config
->cb
.buffer_begin(buf
, tsc
, oldidx
);
1317 * Order all writes to buffer before the commit count update that will
1318 * determine that the subbuffer is full.
1320 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1322 * Must write slot data before incrementing commit count. This
1323 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1329 v_add(config
, config
->cb
.subbuffer_header_size(),
1330 &buf
->commit_hot
[oldidx
].cc
);
1331 commit_count
= v_read(config
, &buf
->commit_hot
[oldidx
].cc
);
1332 /* Check if the written buffer has to be delivered */
1333 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1334 commit_count
, oldidx
, tsc
);
1335 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1336 offsets
->old
+ config
->cb
.subbuffer_header_size(),
1341 * lib_ring_buffer_switch_old_end: switch old subbuffer
1343 * Note : offset_old should never be 0 here. It is ok, because we never perform
1344 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1345 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1349 void lib_ring_buffer_switch_old_end(struct lib_ring_buffer
*buf
,
1350 struct channel
*chan
,
1351 struct switch_offsets
*offsets
,
1354 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1355 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1356 unsigned long commit_count
, padding_size
, data_size
;
1358 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1359 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1360 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
);
1363 * Order all writes to buffer before the commit count update that will
1364 * determine that the subbuffer is full.
1366 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1368 * Must write slot data before incrementing commit count. This
1369 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1375 v_add(config
, padding_size
, &buf
->commit_hot
[oldidx
].cc
);
1376 commit_count
= v_read(config
, &buf
->commit_hot
[oldidx
].cc
);
1377 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1378 commit_count
, oldidx
, tsc
);
1379 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1380 offsets
->old
+ padding_size
, commit_count
);
1384 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1386 * This code can be executed unordered : writers may already have written to the
1387 * sub-buffer before this code gets executed, caution. The commit makes sure
1388 * that this code is executed before the deliver of this sub-buffer.
1391 void lib_ring_buffer_switch_new_start(struct lib_ring_buffer
*buf
,
1392 struct channel
*chan
,
1393 struct switch_offsets
*offsets
,
1396 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1397 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1398 unsigned long commit_count
;
1400 config
->cb
.buffer_begin(buf
, tsc
, beginidx
);
1403 * Order all writes to buffer before the commit count update that will
1404 * determine that the subbuffer is full.
1406 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1408 * Must write slot data before incrementing commit count. This
1409 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1415 v_add(config
, config
->cb
.subbuffer_header_size(),
1416 &buf
->commit_hot
[beginidx
].cc
);
1417 commit_count
= v_read(config
, &buf
->commit_hot
[beginidx
].cc
);
1418 /* Check if the written buffer has to be delivered */
1419 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1420 commit_count
, beginidx
, tsc
);
1421 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, beginidx
,
1422 offsets
->begin
+ config
->cb
.subbuffer_header_size(),
1427 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1429 * Calls subbuffer_set_data_size() to set the data size of the current
1430 * sub-buffer. We do not need to perform check_deliver nor commit here,
1431 * since this task will be done by the "commit" of the event for which
1432 * we are currently doing the space reservation.
1435 void lib_ring_buffer_switch_new_end(struct lib_ring_buffer
*buf
,
1436 struct channel
*chan
,
1437 struct switch_offsets
*offsets
,
1440 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1441 unsigned long endidx
, data_size
;
1443 endidx
= subbuf_index(offsets
->end
- 1, chan
);
1444 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1445 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
);
1451 * !0 if execution must be aborted.
1454 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1455 struct lib_ring_buffer
*buf
,
1456 struct channel
*chan
,
1457 struct switch_offsets
*offsets
,
1460 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1461 unsigned long off
, reserve_commit_diff
;
1463 offsets
->begin
= v_read(config
, &buf
->offset
);
1464 offsets
->old
= offsets
->begin
;
1465 offsets
->switch_old_start
= 0;
1466 off
= subbuf_offset(offsets
->begin
, chan
);
1468 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1471 * Ensure we flush the header of an empty subbuffer when doing the
1472 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1473 * total data gathering duration even if there were no records saved
1474 * after the last buffer switch.
1475 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1476 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1477 * subbuffer header as appropriate.
1478 * The next record that reserves space will be responsible for
1479 * populating the following subbuffer header. We choose not to populate
1480 * the next subbuffer header here because we want to be able to use
1481 * SWITCH_ACTIVE for periodical buffer flush and CPU tick_nohz stop
1482 * buffer flush, which must guarantee that all the buffer content
1483 * (records and header timestamps) are visible to the reader. This is
1484 * required for quiescence guarantees for the fusion merge.
1486 if (mode
!= SWITCH_FLUSH
&& !off
)
1487 return -1; /* we do not have to switch : buffer is empty */
1489 if (unlikely(off
== 0)) {
1490 unsigned long sb_index
, commit_count
;
1493 * We are performing a SWITCH_FLUSH. There may be concurrent
1494 * writes into the buffer if e.g. invoked while performing a
1495 * snapshot on an active trace.
1497 * If the client does not save any header information (sub-buffer
1498 * header size == 0), don't switch empty subbuffer on finalize,
1499 * because it is invalid to deliver a completely empty
1502 if (!config
->cb
.subbuffer_header_size())
1505 /* Test new buffer integrity */
1506 sb_index
= subbuf_index(offsets
->begin
, chan
);
1507 commit_count
= v_read(config
,
1508 &buf
->commit_cold
[sb_index
].cc_sb
);
1509 reserve_commit_diff
=
1510 (buf_trunc(offsets
->begin
, chan
)
1511 >> chan
->backend
.num_subbuf_order
)
1512 - (commit_count
& chan
->commit_count_mask
);
1513 if (likely(reserve_commit_diff
== 0)) {
1514 /* Next subbuffer not being written to. */
1515 if (unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1516 subbuf_trunc(offsets
->begin
, chan
)
1517 - subbuf_trunc((unsigned long)
1518 atomic_long_read(&buf
->consumed
), chan
)
1519 >= chan
->backend
.buf_size
)) {
1521 * We do not overwrite non consumed buffers
1522 * and we are full : don't switch.
1527 * Next subbuffer not being written to, and we
1528 * are either in overwrite mode or the buffer is
1529 * not full. It's safe to write in this new
1535 * Next subbuffer reserve offset does not match the
1536 * commit offset. Don't perform switch in
1537 * producer-consumer and overwrite mode. Caused by
1538 * either a writer OOPS or too many nested writes over a
1539 * reserve/commit pair.
1545 * Need to write the subbuffer start header on finalize.
1547 offsets
->switch_old_start
= 1;
1549 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1550 /* Note: old points to the next subbuf at offset 0 */
1551 offsets
->end
= offsets
->begin
;
1556 * Force a sub-buffer switch. This operation is completely reentrant : can be
1557 * called while tracing is active with absolutely no lock held.
1559 * Note, however, that as a v_cmpxchg is used for some atomic
1560 * operations, this function must be called from the CPU which owns the buffer
1561 * for a ACTIVE flush.
1563 void lib_ring_buffer_switch_slow(struct lib_ring_buffer
*buf
, enum switch_mode mode
)
1565 struct channel
*chan
= buf
->backend
.chan
;
1566 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1567 struct switch_offsets offsets
;
1568 unsigned long oldidx
;
1574 * Perform retryable operations.
1577 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1579 return; /* Switch not needed */
1580 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1584 * Atomically update last_tsc. This update races against concurrent
1585 * atomic updates, but the race will always cause supplementary full TSC
1586 * records, never the opposite (missing a full TSC record when it would
1589 save_last_tsc(config
, buf
, tsc
);
1592 * Push the reader if necessary
1594 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1596 oldidx
= subbuf_index(offsets
.old
, chan
);
1597 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
);
1600 * May need to populate header start on SWITCH_FLUSH.
1602 if (offsets
.switch_old_start
) {
1603 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
);
1604 offsets
.old
+= config
->cb
.subbuffer_header_size();
1608 * Switch old subbuffer.
1610 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
);
1612 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow
);
1614 static void remote_switch(void *info
)
1616 struct lib_ring_buffer
*buf
= info
;
1618 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
1621 static void _lib_ring_buffer_switch_remote(struct lib_ring_buffer
*buf
,
1622 enum switch_mode mode
)
1624 struct channel
*chan
= buf
->backend
.chan
;
1625 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1629 * With global synchronization we don't need to use the IPI scheme.
1631 if (config
->sync
== RING_BUFFER_SYNC_GLOBAL
) {
1632 lib_ring_buffer_switch_slow(buf
, mode
);
1637 * Taking lock on CPU hotplug to ensure two things: first, that the
1638 * target cpu is not taken concurrently offline while we are within
1639 * smp_call_function_single() (I don't trust that get_cpu() on the
1640 * _local_ CPU actually inhibit CPU hotplug for the _remote_ CPU (to be
1641 * confirmed)). Secondly, if it happens that the CPU is not online, our
1642 * own call to lib_ring_buffer_switch_slow() needs to be protected from
1643 * CPU hotplug handlers, which can also perform a remote subbuffer
1647 ret
= smp_call_function_single(buf
->backend
.cpu
,
1648 remote_switch
, buf
, 1);
1650 /* Remote CPU is offline, do it ourself. */
1651 lib_ring_buffer_switch_slow(buf
, mode
);
1656 void lib_ring_buffer_switch_remote(struct lib_ring_buffer
*buf
)
1658 _lib_ring_buffer_switch_remote(buf
, SWITCH_ACTIVE
);
1660 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote
);
1665 * -ENOSPC if event size is too large for packet.
1666 * -ENOBUFS if there is currently not enough space in buffer for the event.
1667 * -EIO if data cannot be written into the buffer for any other reason.
1670 int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer
*buf
,
1671 struct channel
*chan
,
1672 struct switch_offsets
*offsets
,
1673 struct lib_ring_buffer_ctx
*ctx
)
1675 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1676 unsigned long reserve_commit_diff
, offset_cmp
;
1679 offsets
->begin
= offset_cmp
= v_read(config
, &buf
->offset
);
1680 offsets
->old
= offsets
->begin
;
1681 offsets
->switch_new_start
= 0;
1682 offsets
->switch_new_end
= 0;
1683 offsets
->switch_old_end
= 0;
1684 offsets
->pre_header_padding
= 0;
1686 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1687 if ((int64_t) ctx
->tsc
== -EIO
)
1690 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1691 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
1693 if (unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1694 offsets
->switch_new_start
= 1; /* For offsets->begin */
1696 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1698 &offsets
->pre_header_padding
,
1701 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1704 if (unlikely(subbuf_offset(offsets
->begin
, chan
) +
1705 offsets
->size
> chan
->backend
.subbuf_size
)) {
1706 offsets
->switch_old_end
= 1; /* For offsets->old */
1707 offsets
->switch_new_start
= 1; /* For offsets->begin */
1710 if (unlikely(offsets
->switch_new_start
)) {
1711 unsigned long sb_index
, commit_count
;
1714 * We are typically not filling the previous buffer completely.
1716 if (likely(offsets
->switch_old_end
))
1717 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1718 offsets
->begin
= offsets
->begin
1719 + config
->cb
.subbuffer_header_size();
1720 /* Test new buffer integrity */
1721 sb_index
= subbuf_index(offsets
->begin
, chan
);
1723 * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
1724 * lib_ring_buffer_check_deliver() has the matching
1725 * memory barriers required around commit_cold cc_sb
1726 * updates to ensure reserve and commit counter updates
1727 * are not seen reordered when updated by another CPU.
1730 commit_count
= v_read(config
,
1731 &buf
->commit_cold
[sb_index
].cc_sb
);
1732 /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
1734 if (unlikely(offset_cmp
!= v_read(config
, &buf
->offset
))) {
1736 * The reserve counter have been concurrently updated
1737 * while we read the commit counter. This means the
1738 * commit counter we read might not match buf->offset
1739 * due to concurrent update. We therefore need to retry.
1743 reserve_commit_diff
=
1744 (buf_trunc(offsets
->begin
, chan
)
1745 >> chan
->backend
.num_subbuf_order
)
1746 - (commit_count
& chan
->commit_count_mask
);
1747 if (likely(reserve_commit_diff
== 0)) {
1748 /* Next subbuffer not being written to. */
1749 if (unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1750 subbuf_trunc(offsets
->begin
, chan
)
1751 - subbuf_trunc((unsigned long)
1752 atomic_long_read(&buf
->consumed
), chan
)
1753 >= chan
->backend
.buf_size
)) {
1755 * We do not overwrite non consumed buffers
1756 * and we are full : record is lost.
1758 v_inc(config
, &buf
->records_lost_full
);
1762 * Next subbuffer not being written to, and we
1763 * are either in overwrite mode or the buffer is
1764 * not full. It's safe to write in this new
1770 * Next subbuffer reserve offset does not match the
1771 * commit offset, and this did not involve update to the
1772 * reserve counter. Drop record in producer-consumer and
1773 * overwrite mode. Caused by either a writer OOPS or
1774 * too many nested writes over a reserve/commit pair.
1776 v_inc(config
, &buf
->records_lost_wrap
);
1780 config
->cb
.record_header_size(config
, chan
,
1782 &offsets
->pre_header_padding
,
1785 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1788 if (unlikely(subbuf_offset(offsets
->begin
, chan
)
1789 + offsets
->size
> chan
->backend
.subbuf_size
)) {
1791 * Record too big for subbuffers, report error, don't
1792 * complete the sub-buffer switch.
1794 v_inc(config
, &buf
->records_lost_big
);
1798 * We just made a successful buffer switch and the
1799 * record fits in the new subbuffer. Let's write.
1804 * Record fits in the current buffer and we are not on a switch
1805 * boundary. It's safe to write.
1808 offsets
->end
= offsets
->begin
+ offsets
->size
;
1810 if (unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
1812 * The offset_end will fall at the very beginning of the next
1815 offsets
->switch_new_end
= 1; /* For offsets->begin */
1821 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1822 * @ctx: ring buffer context.
1824 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
1825 * -EIO for other errors, else returns 0.
1826 * It will take care of sub-buffer switching.
1828 int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx
*ctx
)
1830 struct channel
*chan
= ctx
->chan
;
1831 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1832 struct lib_ring_buffer
*buf
;
1833 struct switch_offsets offsets
;
1836 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
1837 buf
= per_cpu_ptr(chan
->backend
.buf
, ctx
->cpu
);
1839 buf
= chan
->backend
.buf
;
1845 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
1849 } while (unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
1854 * Atomically update last_tsc. This update races against concurrent
1855 * atomic updates, but the race will always cause supplementary full TSC
1856 * records, never the opposite (missing a full TSC record when it would
1859 save_last_tsc(config
, buf
, ctx
->tsc
);
1862 * Push the reader if necessary
1864 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
1867 * Clear noref flag for this subbuffer.
1869 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1870 subbuf_index(offsets
.end
- 1, chan
));
1873 * Switch old subbuffer if needed.
1875 if (unlikely(offsets
.switch_old_end
)) {
1876 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1877 subbuf_index(offsets
.old
- 1, chan
));
1878 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
);
1882 * Populate new subbuffer.
1884 if (unlikely(offsets
.switch_new_start
))
1885 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
);
1887 if (unlikely(offsets
.switch_new_end
))
1888 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
);
1890 ctx
->slot_size
= offsets
.size
;
1891 ctx
->pre_offset
= offsets
.begin
;
1892 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
1895 EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow
);
1897 int __init
init_lib_ring_buffer_frontend(void)
1901 for_each_possible_cpu(cpu
)
1902 spin_lock_init(&per_cpu(ring_buffer_nohz_lock
, cpu
));
1906 module_init(init_lib_ring_buffer_frontend
);
1908 void __exit
exit_lib_ring_buffer_frontend(void)
1912 module_exit(exit_lib_ring_buffer_frontend
);