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/kref.h"
65 #include "../../wrapper/percpu-defs.h"
68 * Internal structure representing offsets to use at a sub-buffer switch.
70 struct switch_offsets
{
71 unsigned long begin
, end
, old
;
72 size_t pre_header_padding
, size
;
73 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
84 static ATOMIC_NOTIFIER_HEAD(tick_nohz_notifier
);
85 #endif /* CONFIG_NO_HZ */
87 static DEFINE_PER_CPU(spinlock_t
, ring_buffer_nohz_lock
);
89 DEFINE_PER_CPU(unsigned int, lib_ring_buffer_nesting
);
90 EXPORT_PER_CPU_SYMBOL(lib_ring_buffer_nesting
);
93 void lib_ring_buffer_print_errors(struct channel
*chan
,
94 struct lib_ring_buffer
*buf
, int cpu
);
96 void _lib_ring_buffer_switch_remote(struct lib_ring_buffer
*buf
,
97 enum switch_mode mode
);
100 * Must be called under cpu hotplug protection.
102 void lib_ring_buffer_free(struct lib_ring_buffer
*buf
)
104 struct channel
*chan
= buf
->backend
.chan
;
106 lib_ring_buffer_print_errors(chan
, buf
, buf
->backend
.cpu
);
107 kfree(buf
->commit_hot
);
108 kfree(buf
->commit_cold
);
110 lib_ring_buffer_backend_free(&buf
->backend
);
114 * lib_ring_buffer_reset - Reset ring buffer to initial values.
117 * Effectively empty the ring buffer. Should be called when the buffer is not
118 * used for writing. The ring buffer can be opened for reading, but the reader
119 * should not be using the iterator concurrently with reset. The previous
120 * current iterator record is reset.
122 void lib_ring_buffer_reset(struct lib_ring_buffer
*buf
)
124 struct channel
*chan
= buf
->backend
.chan
;
125 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
129 * Reset iterator first. It will put the subbuffer if it currently holds
132 lib_ring_buffer_iterator_reset(buf
);
133 v_set(config
, &buf
->offset
, 0);
134 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
135 v_set(config
, &buf
->commit_hot
[i
].cc
, 0);
136 v_set(config
, &buf
->commit_hot
[i
].seq
, 0);
137 v_set(config
, &buf
->commit_cold
[i
].cc_sb
, 0);
139 atomic_long_set(&buf
->consumed
, 0);
140 atomic_set(&buf
->record_disabled
, 0);
141 v_set(config
, &buf
->last_tsc
, 0);
142 lib_ring_buffer_backend_reset(&buf
->backend
);
143 /* Don't reset number of active readers */
144 v_set(config
, &buf
->records_lost_full
, 0);
145 v_set(config
, &buf
->records_lost_wrap
, 0);
146 v_set(config
, &buf
->records_lost_big
, 0);
147 v_set(config
, &buf
->records_count
, 0);
148 v_set(config
, &buf
->records_overrun
, 0);
151 EXPORT_SYMBOL_GPL(lib_ring_buffer_reset
);
154 * channel_reset - Reset channel to initial values.
157 * Effectively empty the channel. Should be called when the channel is not used
158 * for writing. The channel can be opened for reading, but the reader should not
159 * be using the iterator concurrently with reset. The previous current iterator
162 void channel_reset(struct channel
*chan
)
165 * Reset iterators first. Will put the subbuffer if held for reading.
167 channel_iterator_reset(chan
);
168 atomic_set(&chan
->record_disabled
, 0);
169 /* Don't reset commit_count_mask, still valid */
170 channel_backend_reset(&chan
->backend
);
171 /* Don't reset switch/read timer interval */
172 /* Don't reset notifiers and notifier enable bits */
173 /* Don't reset reader reference count */
175 EXPORT_SYMBOL_GPL(channel_reset
);
178 * Must be called under cpu hotplug protection.
180 int lib_ring_buffer_create(struct lib_ring_buffer
*buf
,
181 struct channel_backend
*chanb
, int cpu
)
183 const struct lib_ring_buffer_config
*config
= &chanb
->config
;
184 struct channel
*chan
= container_of(chanb
, struct channel
, backend
);
185 void *priv
= chanb
->priv
;
186 size_t subbuf_header_size
;
190 /* Test for cpu hotplug */
191 if (buf
->backend
.allocated
)
195 * Paranoia: per cpu dynamic allocation is not officially documented as
196 * zeroing the memory, so let's do it here too, just in case.
198 memset(buf
, 0, sizeof(*buf
));
200 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
, cpu
);
205 kzalloc_node(ALIGN(sizeof(*buf
->commit_hot
)
206 * chan
->backend
.num_subbuf
,
207 1 << INTERNODE_CACHE_SHIFT
),
208 GFP_KERNEL
| __GFP_NOWARN
,
209 cpu_to_node(max(cpu
, 0)));
210 if (!buf
->commit_hot
) {
216 kzalloc_node(ALIGN(sizeof(*buf
->commit_cold
)
217 * chan
->backend
.num_subbuf
,
218 1 << INTERNODE_CACHE_SHIFT
),
219 GFP_KERNEL
| __GFP_NOWARN
,
220 cpu_to_node(max(cpu
, 0)));
221 if (!buf
->commit_cold
) {
226 init_waitqueue_head(&buf
->read_wait
);
227 init_waitqueue_head(&buf
->write_wait
);
228 raw_spin_lock_init(&buf
->raw_tick_nohz_spinlock
);
231 * Write the subbuffer header for first subbuffer so we know the total
232 * duration of data gathering.
234 subbuf_header_size
= config
->cb
.subbuffer_header_size();
235 v_set(config
, &buf
->offset
, subbuf_header_size
);
236 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_wsb
[0].id
);
237 tsc
= config
->cb
.ring_buffer_clock_read(buf
->backend
.chan
);
238 config
->cb
.buffer_begin(buf
, tsc
, 0);
239 v_add(config
, subbuf_header_size
, &buf
->commit_hot
[0].cc
);
241 if (config
->cb
.buffer_create
) {
242 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
);
248 * Ensure the buffer is ready before setting it to allocated and setting
250 * Used for cpu hotplug vs cpumask iteration.
253 buf
->backend
.allocated
= 1;
255 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
256 CHAN_WARN_ON(chan
, cpumask_test_cpu(cpu
,
257 chan
->backend
.cpumask
));
258 cpumask_set_cpu(cpu
, chan
->backend
.cpumask
);
265 kfree(buf
->commit_cold
);
267 kfree(buf
->commit_hot
);
269 lib_ring_buffer_backend_free(&buf
->backend
);
273 static void switch_buffer_timer(unsigned long data
)
275 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
276 struct channel
*chan
= buf
->backend
.chan
;
277 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
280 * Only flush buffers periodically if readers are active.
282 if (atomic_long_read(&buf
->active_readers
))
283 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
285 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
286 mod_timer_pinned(&buf
->switch_timer
,
287 jiffies
+ chan
->switch_timer_interval
);
289 mod_timer(&buf
->switch_timer
,
290 jiffies
+ chan
->switch_timer_interval
);
294 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
296 static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer
*buf
)
298 struct channel
*chan
= buf
->backend
.chan
;
299 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
301 if (!chan
->switch_timer_interval
|| buf
->switch_timer_enabled
)
303 init_timer(&buf
->switch_timer
);
304 buf
->switch_timer
.function
= switch_buffer_timer
;
305 buf
->switch_timer
.expires
= jiffies
+ chan
->switch_timer_interval
;
306 buf
->switch_timer
.data
= (unsigned long)buf
;
307 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
308 add_timer_on(&buf
->switch_timer
, buf
->backend
.cpu
);
310 add_timer(&buf
->switch_timer
);
311 buf
->switch_timer_enabled
= 1;
315 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
317 static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer
*buf
)
319 struct channel
*chan
= buf
->backend
.chan
;
321 if (!chan
->switch_timer_interval
|| !buf
->switch_timer_enabled
)
324 del_timer_sync(&buf
->switch_timer
);
325 buf
->switch_timer_enabled
= 0;
329 * Polling timer to check the channels for data.
331 static void read_buffer_timer(unsigned long data
)
333 struct lib_ring_buffer
*buf
= (struct lib_ring_buffer
*)data
;
334 struct channel
*chan
= buf
->backend
.chan
;
335 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
337 CHAN_WARN_ON(chan
, !buf
->backend
.allocated
);
339 if (atomic_long_read(&buf
->active_readers
)
340 && lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
341 wake_up_interruptible(&buf
->read_wait
);
342 wake_up_interruptible(&chan
->read_wait
);
345 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
346 mod_timer_pinned(&buf
->read_timer
,
347 jiffies
+ chan
->read_timer_interval
);
349 mod_timer(&buf
->read_timer
,
350 jiffies
+ chan
->read_timer_interval
);
354 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
356 static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer
*buf
)
358 struct channel
*chan
= buf
->backend
.chan
;
359 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
361 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
362 || !chan
->read_timer_interval
363 || buf
->read_timer_enabled
)
366 init_timer(&buf
->read_timer
);
367 buf
->read_timer
.function
= read_buffer_timer
;
368 buf
->read_timer
.expires
= jiffies
+ chan
->read_timer_interval
;
369 buf
->read_timer
.data
= (unsigned long)buf
;
371 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
372 add_timer_on(&buf
->read_timer
, buf
->backend
.cpu
);
374 add_timer(&buf
->read_timer
);
375 buf
->read_timer_enabled
= 1;
379 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
381 static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer
*buf
)
383 struct channel
*chan
= buf
->backend
.chan
;
384 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
386 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
387 || !chan
->read_timer_interval
388 || !buf
->read_timer_enabled
)
391 del_timer_sync(&buf
->read_timer
);
393 * do one more check to catch data that has been written in the last
396 if (lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
397 wake_up_interruptible(&buf
->read_wait
);
398 wake_up_interruptible(&chan
->read_wait
);
400 buf
->read_timer_enabled
= 0;
403 #ifdef CONFIG_HOTPLUG_CPU
405 * lib_ring_buffer_cpu_hp_callback - CPU hotplug callback
406 * @nb: notifier block
407 * @action: hotplug action to take
410 * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
413 int lib_ring_buffer_cpu_hp_callback(struct notifier_block
*nb
,
414 unsigned long action
,
417 unsigned int cpu
= (unsigned long)hcpu
;
418 struct channel
*chan
= container_of(nb
, struct channel
,
420 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
, cpu
);
421 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
423 if (!chan
->cpu_hp_enable
)
426 CHAN_WARN_ON(chan
, config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
);
429 case CPU_DOWN_FAILED
:
430 case CPU_DOWN_FAILED_FROZEN
:
432 case CPU_ONLINE_FROZEN
:
433 wake_up_interruptible(&chan
->hp_wait
);
434 lib_ring_buffer_start_switch_timer(buf
);
435 lib_ring_buffer_start_read_timer(buf
);
438 case CPU_DOWN_PREPARE
:
439 case CPU_DOWN_PREPARE_FROZEN
:
440 lib_ring_buffer_stop_switch_timer(buf
);
441 lib_ring_buffer_stop_read_timer(buf
);
445 case CPU_DEAD_FROZEN
:
447 * Performing a buffer switch on a remote CPU. Performed by
448 * the CPU responsible for doing the hotunplug after the target
449 * CPU stopped running completely. Ensures that all data
450 * from that remote CPU is flushed.
452 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
461 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
463 * For per-cpu buffers, call the reader wakeups before switching the buffer, so
464 * that wake-up-tracing generated events are flushed before going idle (in
465 * tick_nohz). We test if the spinlock is locked to deal with the race where
466 * readers try to sample the ring buffer before we perform the switch. We let
467 * the readers retry in that case. If there is data in the buffer, the wake up
468 * is going to forbid the CPU running the reader thread from going idle.
470 static int notrace
ring_buffer_tick_nohz_callback(struct notifier_block
*nb
,
474 struct channel
*chan
= container_of(nb
, struct channel
,
476 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
477 struct lib_ring_buffer
*buf
;
478 int cpu
= smp_processor_id();
480 if (config
->alloc
!= RING_BUFFER_ALLOC_PER_CPU
) {
482 * We don't support keeping the system idle with global buffers
483 * and streaming active. In order to do so, we would need to
484 * sample a non-nohz-cpumask racelessly with the nohz updates
485 * without adding synchronization overhead to nohz. Leave this
486 * use-case out for now.
491 buf
= channel_get_ring_buffer(config
, chan
, cpu
);
493 case TICK_NOHZ_FLUSH
:
494 raw_spin_lock(&buf
->raw_tick_nohz_spinlock
);
495 if (config
->wakeup
== RING_BUFFER_WAKEUP_BY_TIMER
496 && chan
->read_timer_interval
497 && atomic_long_read(&buf
->active_readers
)
498 && (lib_ring_buffer_poll_deliver(config
, buf
, chan
)
499 || lib_ring_buffer_pending_data(config
, buf
, chan
))) {
500 wake_up_interruptible(&buf
->read_wait
);
501 wake_up_interruptible(&chan
->read_wait
);
503 if (chan
->switch_timer_interval
)
504 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
505 raw_spin_unlock(&buf
->raw_tick_nohz_spinlock
);
508 spin_lock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
509 lib_ring_buffer_stop_switch_timer(buf
);
510 lib_ring_buffer_stop_read_timer(buf
);
511 spin_unlock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
513 case TICK_NOHZ_RESTART
:
514 spin_lock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
515 lib_ring_buffer_start_read_timer(buf
);
516 lib_ring_buffer_start_switch_timer(buf
);
517 spin_unlock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
524 void notrace
lib_ring_buffer_tick_nohz_flush(void)
526 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_FLUSH
,
530 void notrace
lib_ring_buffer_tick_nohz_stop(void)
532 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_STOP
,
536 void notrace
lib_ring_buffer_tick_nohz_restart(void)
538 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_RESTART
,
541 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
546 static void channel_unregister_notifiers(struct channel
*chan
)
548 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
551 channel_iterator_unregister_notifiers(chan
);
552 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
555 * Remove the nohz notifier first, so we are certain we stop
558 atomic_notifier_chain_unregister(&tick_nohz_notifier
,
559 &chan
->tick_nohz_notifier
);
561 * ring_buffer_nohz_lock will not be needed below, because
562 * we just removed the notifiers, which were the only source of
565 #endif /* CONFIG_NO_HZ */
566 #ifdef CONFIG_HOTPLUG_CPU
568 chan
->cpu_hp_enable
= 0;
569 for_each_online_cpu(cpu
) {
570 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
572 lib_ring_buffer_stop_switch_timer(buf
);
573 lib_ring_buffer_stop_read_timer(buf
);
576 unregister_cpu_notifier(&chan
->cpu_hp_notifier
);
578 for_each_possible_cpu(cpu
) {
579 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
581 lib_ring_buffer_stop_switch_timer(buf
);
582 lib_ring_buffer_stop_read_timer(buf
);
586 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
588 lib_ring_buffer_stop_switch_timer(buf
);
589 lib_ring_buffer_stop_read_timer(buf
);
591 channel_backend_unregister_notifiers(&chan
->backend
);
594 static void lib_ring_buffer_set_quiescent(struct lib_ring_buffer
*buf
)
596 if (!buf
->quiescent
) {
597 buf
->quiescent
= true;
598 _lib_ring_buffer_switch_remote(buf
, SWITCH_FLUSH
);
602 static void lib_ring_buffer_clear_quiescent(struct lib_ring_buffer
*buf
)
604 buf
->quiescent
= false;
607 void lib_ring_buffer_set_quiescent_channel(struct channel
*chan
)
610 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
612 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
614 for_each_channel_cpu(cpu
, chan
) {
615 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
618 lib_ring_buffer_set_quiescent(buf
);
622 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
624 lib_ring_buffer_set_quiescent(buf
);
627 EXPORT_SYMBOL_GPL(lib_ring_buffer_set_quiescent_channel
);
629 void lib_ring_buffer_clear_quiescent_channel(struct channel
*chan
)
632 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
634 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
636 for_each_channel_cpu(cpu
, chan
) {
637 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
640 lib_ring_buffer_clear_quiescent(buf
);
644 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
646 lib_ring_buffer_clear_quiescent(buf
);
649 EXPORT_SYMBOL_GPL(lib_ring_buffer_clear_quiescent_channel
);
651 static void channel_free(struct channel
*chan
)
653 if (chan
->backend
.release_priv_ops
) {
654 chan
->backend
.release_priv_ops(chan
->backend
.priv_ops
);
656 channel_iterator_free(chan
);
657 channel_backend_free(&chan
->backend
);
662 * channel_create - Create channel.
663 * @config: ring buffer instance configuration
664 * @name: name of the channel
665 * @priv: ring buffer client private data
666 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
667 * address mapping. It is used only by RING_BUFFER_STATIC
668 * configuration. It can be set to NULL for other backends.
669 * @subbuf_size: subbuffer size
670 * @num_subbuf: number of subbuffers
671 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
672 * padding to let readers get those sub-buffers.
673 * Used for live streaming.
674 * @read_timer_interval: Time interval (in us) to wake up pending readers.
677 * Returns NULL on failure.
679 struct channel
*channel_create(const struct lib_ring_buffer_config
*config
,
680 const char *name
, void *priv
, void *buf_addr
,
682 size_t num_subbuf
, unsigned int switch_timer_interval
,
683 unsigned int read_timer_interval
)
686 struct channel
*chan
;
688 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
689 read_timer_interval
))
692 chan
= kzalloc(sizeof(struct channel
), GFP_KERNEL
);
696 ret
= channel_backend_init(&chan
->backend
, name
, config
, priv
,
697 subbuf_size
, num_subbuf
);
701 ret
= channel_iterator_init(chan
);
703 goto error_free_backend
;
705 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
706 chan
->switch_timer_interval
= usecs_to_jiffies(switch_timer_interval
);
707 chan
->read_timer_interval
= usecs_to_jiffies(read_timer_interval
);
708 kref_init(&chan
->ref
);
709 init_waitqueue_head(&chan
->read_wait
);
710 init_waitqueue_head(&chan
->hp_wait
);
712 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
713 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
714 /* Only benefit from NO_HZ idle with per-cpu buffers for now. */
715 chan
->tick_nohz_notifier
.notifier_call
=
716 ring_buffer_tick_nohz_callback
;
717 chan
->tick_nohz_notifier
.priority
= ~0U;
718 atomic_notifier_chain_register(&tick_nohz_notifier
,
719 &chan
->tick_nohz_notifier
);
720 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
723 * In case of non-hotplug cpu, if the ring-buffer is allocated
724 * in early initcall, it will not be notified of secondary cpus.
725 * In that off case, we need to allocate for all possible cpus.
727 #ifdef CONFIG_HOTPLUG_CPU
728 chan
->cpu_hp_notifier
.notifier_call
=
729 lib_ring_buffer_cpu_hp_callback
;
730 chan
->cpu_hp_notifier
.priority
= 6;
731 register_cpu_notifier(&chan
->cpu_hp_notifier
);
734 for_each_online_cpu(cpu
) {
735 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
737 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
738 lib_ring_buffer_start_switch_timer(buf
);
739 lib_ring_buffer_start_read_timer(buf
);
740 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
742 chan
->cpu_hp_enable
= 1;
745 for_each_possible_cpu(cpu
) {
746 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
748 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
749 lib_ring_buffer_start_switch_timer(buf
);
750 lib_ring_buffer_start_read_timer(buf
);
751 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
755 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
757 lib_ring_buffer_start_switch_timer(buf
);
758 lib_ring_buffer_start_read_timer(buf
);
764 channel_backend_free(&chan
->backend
);
769 EXPORT_SYMBOL_GPL(channel_create
);
772 void channel_release(struct kref
*kref
)
774 struct channel
*chan
= container_of(kref
, struct channel
, ref
);
779 * channel_destroy - Finalize, wait for q.s. and destroy channel.
780 * @chan: channel to destroy
783 * Call "destroy" callback, finalize channels, and then decrement the
784 * channel reference count. Note that when readers have completed data
785 * consumption of finalized channels, get_subbuf() will return -ENODATA.
786 * They should release their handle at that point. Returns the private
789 void *channel_destroy(struct channel
*chan
)
792 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
795 channel_unregister_notifiers(chan
);
797 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
799 * No need to hold cpu hotplug, because all notifiers have been
802 for_each_channel_cpu(cpu
, chan
) {
803 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
806 if (config
->cb
.buffer_finalize
)
807 config
->cb
.buffer_finalize(buf
,
810 if (buf
->backend
.allocated
)
811 lib_ring_buffer_set_quiescent(buf
);
813 * Perform flush before writing to finalized.
816 ACCESS_ONCE(buf
->finalized
) = 1;
817 wake_up_interruptible(&buf
->read_wait
);
820 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
822 if (config
->cb
.buffer_finalize
)
823 config
->cb
.buffer_finalize(buf
, chan
->backend
.priv
, -1);
824 if (buf
->backend
.allocated
)
825 lib_ring_buffer_set_quiescent(buf
);
827 * Perform flush before writing to finalized.
830 ACCESS_ONCE(buf
->finalized
) = 1;
831 wake_up_interruptible(&buf
->read_wait
);
833 ACCESS_ONCE(chan
->finalized
) = 1;
834 wake_up_interruptible(&chan
->hp_wait
);
835 wake_up_interruptible(&chan
->read_wait
);
836 priv
= chan
->backend
.priv
;
837 kref_put(&chan
->ref
, channel_release
);
840 EXPORT_SYMBOL_GPL(channel_destroy
);
842 struct lib_ring_buffer
*channel_get_ring_buffer(
843 const struct lib_ring_buffer_config
*config
,
844 struct channel
*chan
, int cpu
)
846 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
)
847 return chan
->backend
.buf
;
849 return per_cpu_ptr(chan
->backend
.buf
, cpu
);
851 EXPORT_SYMBOL_GPL(channel_get_ring_buffer
);
853 int lib_ring_buffer_open_read(struct lib_ring_buffer
*buf
)
855 struct channel
*chan
= buf
->backend
.chan
;
857 if (!atomic_long_add_unless(&buf
->active_readers
, 1, 1))
859 if (!lttng_kref_get(&chan
->ref
)) {
860 atomic_long_dec(&buf
->active_readers
);
863 lttng_smp_mb__after_atomic();
866 EXPORT_SYMBOL_GPL(lib_ring_buffer_open_read
);
868 void lib_ring_buffer_release_read(struct lib_ring_buffer
*buf
)
870 struct channel
*chan
= buf
->backend
.chan
;
872 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
873 lttng_smp_mb__before_atomic();
874 atomic_long_dec(&buf
->active_readers
);
875 kref_put(&chan
->ref
, channel_release
);
877 EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read
);
880 * Promote compiler barrier to a smp_mb().
881 * For the specific ring buffer case, this IPI call should be removed if the
882 * architecture does not reorder writes. This should eventually be provided by
883 * a separate architecture-specific infrastructure.
885 static void remote_mb(void *info
)
891 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
893 * @consumed: consumed count indicating the position where to read
894 * @produced: produced count, indicates position when to stop reading
896 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
897 * data to read at consumed position, or 0 if the get operation succeeds.
898 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
901 int lib_ring_buffer_snapshot(struct lib_ring_buffer
*buf
,
902 unsigned long *consumed
, unsigned long *produced
)
904 struct channel
*chan
= buf
->backend
.chan
;
905 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
906 unsigned long consumed_cur
, write_offset
;
910 finalized
= ACCESS_ONCE(buf
->finalized
);
912 * Read finalized before counters.
915 consumed_cur
= atomic_long_read(&buf
->consumed
);
917 * No need to issue a memory barrier between consumed count read and
918 * write offset read, because consumed count can only change
919 * concurrently in overwrite mode, and we keep a sequence counter
920 * identifier derived from the write offset to check we are getting
921 * the same sub-buffer we are expecting (the sub-buffers are atomically
922 * "tagged" upon writes, tags are checked upon read).
924 write_offset
= v_read(config
, &buf
->offset
);
927 * Check that we are not about to read the same subbuffer in
928 * which the writer head is.
930 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
934 *consumed
= consumed_cur
;
935 *produced
= subbuf_trunc(write_offset
, chan
);
941 * The memory barriers __wait_event()/wake_up_interruptible() take care
942 * of "raw_spin_is_locked" memory ordering.
946 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
951 EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot
);
954 * lib_ring_buffer_put_snapshot - move consumed counter forward
956 * Should only be called from consumer context.
958 * @consumed_new: new consumed count value
960 void lib_ring_buffer_move_consumer(struct lib_ring_buffer
*buf
,
961 unsigned long consumed_new
)
963 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
964 struct channel
*chan
= bufb
->chan
;
965 unsigned long consumed
;
967 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
970 * Only push the consumed value forward.
971 * If the consumed cmpxchg fails, this is because we have been pushed by
972 * the writer in flight recorder mode.
974 consumed
= atomic_long_read(&buf
->consumed
);
975 while ((long) consumed
- (long) consumed_new
< 0)
976 consumed
= atomic_long_cmpxchg(&buf
->consumed
, consumed
,
978 /* Wake-up the metadata producer */
979 wake_up_interruptible(&buf
->write_wait
);
981 EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer
);
984 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
986 * @consumed: consumed count indicating the position where to read
988 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
989 * data to read at consumed position, or 0 if the get operation succeeds.
990 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
992 int lib_ring_buffer_get_subbuf(struct lib_ring_buffer
*buf
,
993 unsigned long consumed
)
995 struct channel
*chan
= buf
->backend
.chan
;
996 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
997 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
1001 if (buf
->get_subbuf
) {
1003 * Reader is trying to get a subbuffer twice.
1005 CHAN_WARN_ON(chan
, 1);
1009 finalized
= ACCESS_ONCE(buf
->finalized
);
1011 * Read finalized before counters.
1014 consumed_cur
= atomic_long_read(&buf
->consumed
);
1015 consumed_idx
= subbuf_index(consumed
, chan
);
1016 commit_count
= v_read(config
, &buf
->commit_cold
[consumed_idx
].cc_sb
);
1018 * Make sure we read the commit count before reading the buffer
1019 * data and the write offset. Correct consumed offset ordering
1020 * wrt commit count is insured by the use of cmpxchg to update
1021 * the consumed offset.
1022 * smp_call_function_single can fail if the remote CPU is offline,
1023 * this is OK because then there is no wmb to execute there.
1024 * If our thread is executing on the same CPU as the on the buffers
1025 * belongs to, we don't have to synchronize it at all. If we are
1026 * migrated, the scheduler will take care of the memory barriers.
1027 * Normally, smp_call_function_single() should ensure program order when
1028 * executing the remote function, which implies that it surrounds the
1029 * function execution with :
1040 * However, smp_call_function_single() does not seem to clearly execute
1041 * such barriers. It depends on spinlock semantic to provide the barrier
1042 * before executing the IPI and, when busy-looping, csd_lock_wait only
1043 * executes smp_mb() when it has to wait for the other CPU.
1045 * I don't trust this code. Therefore, let's add the smp_mb() sequence
1046 * required ourself, even if duplicated. It has no performance impact
1049 * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs
1050 * read and write vs write. They do not ensure core synchronization. We
1051 * really have to ensure total order between the 3 barriers running on
1054 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1055 if (config
->sync
== RING_BUFFER_SYNC_PER_CPU
1056 && config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
1057 if (raw_smp_processor_id() != buf
->backend
.cpu
) {
1058 /* Total order with IPI handler smp_mb() */
1060 smp_call_function_single(buf
->backend
.cpu
,
1061 remote_mb
, NULL
, 1);
1062 /* Total order with IPI handler smp_mb() */
1066 /* Total order with IPI handler smp_mb() */
1068 smp_call_function(remote_mb
, NULL
, 1);
1069 /* Total order with IPI handler smp_mb() */
1074 * Local rmb to match the remote wmb to read the commit count
1075 * before the buffer data and the write offset.
1080 write_offset
= v_read(config
, &buf
->offset
);
1083 * Check that the buffer we are getting is after or at consumed_cur
1086 if ((long) subbuf_trunc(consumed
, chan
)
1087 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
1091 * Check that the subbuffer we are trying to consume has been
1092 * already fully committed.
1094 if (((commit_count
- chan
->backend
.subbuf_size
)
1095 & chan
->commit_count_mask
)
1096 - (buf_trunc(consumed
, chan
)
1097 >> chan
->backend
.num_subbuf_order
)
1102 * Check that we are not about to read the same subbuffer in
1103 * which the writer head is.
1105 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed
, chan
)
1110 * Failure to get the subbuffer causes a busy-loop retry without going
1111 * to a wait queue. These are caused by short-lived race windows where
1112 * the writer is getting access to a subbuffer we were trying to get
1113 * access to. Also checks that the "consumed" buffer count we are
1114 * looking for matches the one contained in the subbuffer id.
1116 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1117 consumed_idx
, buf_trunc_val(consumed
, chan
));
1120 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1122 buf
->get_subbuf_consumed
= consumed
;
1123 buf
->get_subbuf
= 1;
1129 * The memory barriers __wait_event()/wake_up_interruptible() take care
1130 * of "raw_spin_is_locked" memory ordering.
1134 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
1139 EXPORT_SYMBOL_GPL(lib_ring_buffer_get_subbuf
);
1142 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1145 void lib_ring_buffer_put_subbuf(struct lib_ring_buffer
*buf
)
1147 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1148 struct channel
*chan
= bufb
->chan
;
1149 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1150 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
1152 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
1154 if (!buf
->get_subbuf
) {
1156 * Reader puts a subbuffer it did not get.
1158 CHAN_WARN_ON(chan
, 1);
1161 consumed
= buf
->get_subbuf_consumed
;
1162 buf
->get_subbuf
= 0;
1165 * Clear the records_unread counter. (overruns counter)
1166 * Can still be non-zero if a file reader simply grabbed the data
1167 * without using iterators.
1168 * Can be below zero if an iterator is used on a snapshot more than
1171 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1172 v_add(config
, v_read(config
,
1173 &bufb
->array
[read_sb_bindex
]->records_unread
),
1174 &bufb
->records_read
);
1175 v_set(config
, &bufb
->array
[read_sb_bindex
]->records_unread
, 0);
1176 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1177 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1178 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1181 * Exchange the reader subbuffer with the one we put in its place in the
1182 * writer subbuffer table. Expect the original consumed count. If
1183 * update_read_sb_index fails, this is because the writer updated the
1184 * subbuffer concurrently. We should therefore keep the subbuffer we
1185 * currently have: it has become invalid to try reading this sub-buffer
1186 * consumed count value anyway.
1188 consumed_idx
= subbuf_index(consumed
, chan
);
1189 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1190 consumed_idx
, buf_trunc_val(consumed
, chan
));
1192 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1193 * if the writer concurrently updated it.
1196 EXPORT_SYMBOL_GPL(lib_ring_buffer_put_subbuf
);
1199 * cons_offset is an iterator on all subbuffer offsets between the reader
1200 * position and the writer position. (inclusive)
1203 void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer
*buf
,
1204 struct channel
*chan
,
1205 unsigned long cons_offset
,
1208 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1209 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1211 cons_idx
= subbuf_index(cons_offset
, chan
);
1212 commit_count
= v_read(config
, &buf
->commit_hot
[cons_idx
].cc
);
1213 commit_count_sb
= v_read(config
, &buf
->commit_cold
[cons_idx
].cc_sb
);
1215 if (subbuf_offset(commit_count
, chan
) != 0)
1217 "ring buffer %s, cpu %d: "
1218 "commit count in subbuffer %lu,\n"
1219 "expecting multiples of %lu bytes\n"
1220 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1221 chan
->backend
.name
, cpu
, cons_idx
,
1222 chan
->backend
.subbuf_size
,
1223 commit_count
, commit_count_sb
);
1225 printk(KERN_DEBUG
"ring buffer: %s, cpu %d: %lu bytes committed\n",
1226 chan
->backend
.name
, cpu
, commit_count
);
1230 void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer
*buf
,
1231 struct channel
*chan
,
1232 void *priv
, int cpu
)
1234 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1235 unsigned long write_offset
, cons_offset
;
1238 * No need to order commit_count, write_offset and cons_offset reads
1239 * because we execute at teardown when no more writer nor reader
1240 * references are left.
1242 write_offset
= v_read(config
, &buf
->offset
);
1243 cons_offset
= atomic_long_read(&buf
->consumed
);
1244 if (write_offset
!= cons_offset
)
1246 "ring buffer %s, cpu %d: "
1247 "non-consumed data\n"
1248 " [ %lu bytes written, %lu bytes read ]\n",
1249 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1251 for (cons_offset
= atomic_long_read(&buf
->consumed
);
1252 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1255 cons_offset
= subbuf_align(cons_offset
, chan
))
1256 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1261 void lib_ring_buffer_print_errors(struct channel
*chan
,
1262 struct lib_ring_buffer
*buf
, int cpu
)
1264 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1265 void *priv
= chan
->backend
.priv
;
1267 if (!strcmp(chan
->backend
.name
, "relay-metadata")) {
1268 printk(KERN_DEBUG
"ring buffer %s: %lu records written, "
1269 "%lu records overrun\n",
1271 v_read(config
, &buf
->records_count
),
1272 v_read(config
, &buf
->records_overrun
));
1274 printk(KERN_DEBUG
"ring buffer %s, cpu %d: %lu records written, "
1275 "%lu records overrun\n",
1276 chan
->backend
.name
, cpu
,
1277 v_read(config
, &buf
->records_count
),
1278 v_read(config
, &buf
->records_overrun
));
1280 if (v_read(config
, &buf
->records_lost_full
)
1281 || v_read(config
, &buf
->records_lost_wrap
)
1282 || v_read(config
, &buf
->records_lost_big
))
1284 "ring buffer %s, cpu %d: records were lost. Caused by:\n"
1285 " [ %lu buffer full, %lu nest buffer wrap-around, "
1286 "%lu event too big ]\n",
1287 chan
->backend
.name
, cpu
,
1288 v_read(config
, &buf
->records_lost_full
),
1289 v_read(config
, &buf
->records_lost_wrap
),
1290 v_read(config
, &buf
->records_lost_big
));
1292 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
);
1296 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1298 * Only executed by SWITCH_FLUSH, which can be issued while tracing is active
1299 * or at buffer finalization (destroy).
1302 void lib_ring_buffer_switch_old_start(struct lib_ring_buffer
*buf
,
1303 struct channel
*chan
,
1304 struct switch_offsets
*offsets
,
1307 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1308 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1309 unsigned long commit_count
;
1311 config
->cb
.buffer_begin(buf
, tsc
, oldidx
);
1314 * Order all writes to buffer before the commit count update that will
1315 * determine that the subbuffer is full.
1317 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1319 * Must write slot data before incrementing commit count. This
1320 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1326 v_add(config
, config
->cb
.subbuffer_header_size(),
1327 &buf
->commit_hot
[oldidx
].cc
);
1328 commit_count
= v_read(config
, &buf
->commit_hot
[oldidx
].cc
);
1329 /* Check if the written buffer has to be delivered */
1330 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1331 commit_count
, oldidx
, tsc
);
1332 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1333 offsets
->old
+ config
->cb
.subbuffer_header_size(),
1338 * lib_ring_buffer_switch_old_end: switch old subbuffer
1340 * Note : offset_old should never be 0 here. It is ok, because we never perform
1341 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1342 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1346 void lib_ring_buffer_switch_old_end(struct lib_ring_buffer
*buf
,
1347 struct channel
*chan
,
1348 struct switch_offsets
*offsets
,
1351 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1352 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1353 unsigned long commit_count
, padding_size
, data_size
;
1355 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1356 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1357 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
);
1360 * Order all writes to buffer before the commit count update that will
1361 * determine that the subbuffer is full.
1363 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1365 * Must write slot data before incrementing commit count. This
1366 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1372 v_add(config
, padding_size
, &buf
->commit_hot
[oldidx
].cc
);
1373 commit_count
= v_read(config
, &buf
->commit_hot
[oldidx
].cc
);
1374 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1375 commit_count
, oldidx
, tsc
);
1376 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, oldidx
,
1377 offsets
->old
+ padding_size
, commit_count
);
1381 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1383 * This code can be executed unordered : writers may already have written to the
1384 * sub-buffer before this code gets executed, caution. The commit makes sure
1385 * that this code is executed before the deliver of this sub-buffer.
1388 void lib_ring_buffer_switch_new_start(struct lib_ring_buffer
*buf
,
1389 struct channel
*chan
,
1390 struct switch_offsets
*offsets
,
1393 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1394 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1395 unsigned long commit_count
;
1397 config
->cb
.buffer_begin(buf
, tsc
, beginidx
);
1400 * Order all writes to buffer before the commit count update that will
1401 * determine that the subbuffer is full.
1403 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1405 * Must write slot data before incrementing commit count. This
1406 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1412 v_add(config
, config
->cb
.subbuffer_header_size(),
1413 &buf
->commit_hot
[beginidx
].cc
);
1414 commit_count
= v_read(config
, &buf
->commit_hot
[beginidx
].cc
);
1415 /* Check if the written buffer has to be delivered */
1416 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1417 commit_count
, beginidx
, tsc
);
1418 lib_ring_buffer_write_commit_counter(config
, buf
, chan
, beginidx
,
1419 offsets
->begin
+ config
->cb
.subbuffer_header_size(),
1424 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1426 * Calls subbuffer_set_data_size() to set the data size of the current
1427 * sub-buffer. We do not need to perform check_deliver nor commit here,
1428 * since this task will be done by the "commit" of the event for which
1429 * we are currently doing the space reservation.
1432 void lib_ring_buffer_switch_new_end(struct lib_ring_buffer
*buf
,
1433 struct channel
*chan
,
1434 struct switch_offsets
*offsets
,
1437 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1438 unsigned long endidx
, data_size
;
1440 endidx
= subbuf_index(offsets
->end
- 1, chan
);
1441 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1442 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
);
1448 * !0 if execution must be aborted.
1451 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1452 struct lib_ring_buffer
*buf
,
1453 struct channel
*chan
,
1454 struct switch_offsets
*offsets
,
1457 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1458 unsigned long off
, reserve_commit_diff
;
1460 offsets
->begin
= v_read(config
, &buf
->offset
);
1461 offsets
->old
= offsets
->begin
;
1462 offsets
->switch_old_start
= 0;
1463 off
= subbuf_offset(offsets
->begin
, chan
);
1465 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1468 * Ensure we flush the header of an empty subbuffer when doing the
1469 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1470 * total data gathering duration even if there were no records saved
1471 * after the last buffer switch.
1472 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1473 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1474 * subbuffer header as appropriate.
1475 * The next record that reserves space will be responsible for
1476 * populating the following subbuffer header. We choose not to populate
1477 * the next subbuffer header here because we want to be able to use
1478 * SWITCH_ACTIVE for periodical buffer flush and CPU tick_nohz stop
1479 * buffer flush, which must guarantee that all the buffer content
1480 * (records and header timestamps) are visible to the reader. This is
1481 * required for quiescence guarantees for the fusion merge.
1483 if (mode
!= SWITCH_FLUSH
&& !off
)
1484 return -1; /* we do not have to switch : buffer is empty */
1486 if (unlikely(off
== 0)) {
1487 unsigned long sb_index
, commit_count
;
1490 * We are performing a SWITCH_FLUSH. There may be concurrent
1491 * writes into the buffer if e.g. invoked while performing a
1492 * snapshot on an active trace.
1494 * If the client does not save any header information (sub-buffer
1495 * header size == 0), don't switch empty subbuffer on finalize,
1496 * because it is invalid to deliver a completely empty
1499 if (!config
->cb
.subbuffer_header_size())
1502 /* Test new buffer integrity */
1503 sb_index
= subbuf_index(offsets
->begin
, chan
);
1504 commit_count
= v_read(config
,
1505 &buf
->commit_cold
[sb_index
].cc_sb
);
1506 reserve_commit_diff
=
1507 (buf_trunc(offsets
->begin
, chan
)
1508 >> chan
->backend
.num_subbuf_order
)
1509 - (commit_count
& chan
->commit_count_mask
);
1510 if (likely(reserve_commit_diff
== 0)) {
1511 /* Next subbuffer not being written to. */
1512 if (unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1513 subbuf_trunc(offsets
->begin
, chan
)
1514 - subbuf_trunc((unsigned long)
1515 atomic_long_read(&buf
->consumed
), chan
)
1516 >= chan
->backend
.buf_size
)) {
1518 * We do not overwrite non consumed buffers
1519 * and we are full : don't switch.
1524 * Next subbuffer not being written to, and we
1525 * are either in overwrite mode or the buffer is
1526 * not full. It's safe to write in this new
1532 * Next subbuffer reserve offset does not match the
1533 * commit offset. Don't perform switch in
1534 * producer-consumer and overwrite mode. Caused by
1535 * either a writer OOPS or too many nested writes over a
1536 * reserve/commit pair.
1542 * Need to write the subbuffer start header on finalize.
1544 offsets
->switch_old_start
= 1;
1546 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1547 /* Note: old points to the next subbuf at offset 0 */
1548 offsets
->end
= offsets
->begin
;
1553 * Force a sub-buffer switch. This operation is completely reentrant : can be
1554 * called while tracing is active with absolutely no lock held.
1556 * Note, however, that as a v_cmpxchg is used for some atomic
1557 * operations, this function must be called from the CPU which owns the buffer
1558 * for a ACTIVE flush.
1560 void lib_ring_buffer_switch_slow(struct lib_ring_buffer
*buf
, enum switch_mode mode
)
1562 struct channel
*chan
= buf
->backend
.chan
;
1563 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1564 struct switch_offsets offsets
;
1565 unsigned long oldidx
;
1571 * Perform retryable operations.
1574 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1576 return; /* Switch not needed */
1577 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1581 * Atomically update last_tsc. This update races against concurrent
1582 * atomic updates, but the race will always cause supplementary full TSC
1583 * records, never the opposite (missing a full TSC record when it would
1586 save_last_tsc(config
, buf
, tsc
);
1589 * Push the reader if necessary
1591 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1593 oldidx
= subbuf_index(offsets
.old
, chan
);
1594 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
);
1597 * May need to populate header start on SWITCH_FLUSH.
1599 if (offsets
.switch_old_start
) {
1600 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
);
1601 offsets
.old
+= config
->cb
.subbuffer_header_size();
1605 * Switch old subbuffer.
1607 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
);
1609 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow
);
1611 struct switch_param
{
1612 struct lib_ring_buffer
*buf
;
1613 enum switch_mode mode
;
1616 static void remote_switch(void *info
)
1618 struct switch_param
*param
= info
;
1619 struct lib_ring_buffer
*buf
= param
->buf
;
1621 lib_ring_buffer_switch_slow(buf
, param
->mode
);
1624 static void _lib_ring_buffer_switch_remote(struct lib_ring_buffer
*buf
,
1625 enum switch_mode mode
)
1627 struct channel
*chan
= buf
->backend
.chan
;
1628 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1630 struct switch_param param
;
1633 * With global synchronization we don't need to use the IPI scheme.
1635 if (config
->sync
== RING_BUFFER_SYNC_GLOBAL
) {
1636 lib_ring_buffer_switch_slow(buf
, mode
);
1641 * Taking lock on CPU hotplug to ensure two things: first, that the
1642 * target cpu is not taken concurrently offline while we are within
1643 * smp_call_function_single() (I don't trust that get_cpu() on the
1644 * _local_ CPU actually inhibit CPU hotplug for the _remote_ CPU (to be
1645 * confirmed)). Secondly, if it happens that the CPU is not online, our
1646 * own call to lib_ring_buffer_switch_slow() needs to be protected from
1647 * CPU hotplug handlers, which can also perform a remote subbuffer
1653 ret
= smp_call_function_single(buf
->backend
.cpu
,
1654 remote_switch
, ¶m
, 1);
1656 /* Remote CPU is offline, do it ourself. */
1657 lib_ring_buffer_switch_slow(buf
, mode
);
1662 void lib_ring_buffer_switch_remote(struct lib_ring_buffer
*buf
)
1664 _lib_ring_buffer_switch_remote(buf
, SWITCH_ACTIVE
);
1666 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote
);
1668 /* Switch sub-buffer even if current sub-buffer is empty. */
1669 void lib_ring_buffer_switch_remote_empty(struct lib_ring_buffer
*buf
)
1671 _lib_ring_buffer_switch_remote(buf
, SWITCH_FLUSH
);
1673 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote_empty
);
1678 * -ENOSPC if event size is too large for packet.
1679 * -ENOBUFS if there is currently not enough space in buffer for the event.
1680 * -EIO if data cannot be written into the buffer for any other reason.
1683 int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer
*buf
,
1684 struct channel
*chan
,
1685 struct switch_offsets
*offsets
,
1686 struct lib_ring_buffer_ctx
*ctx
)
1688 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1689 unsigned long reserve_commit_diff
, offset_cmp
;
1692 offsets
->begin
= offset_cmp
= v_read(config
, &buf
->offset
);
1693 offsets
->old
= offsets
->begin
;
1694 offsets
->switch_new_start
= 0;
1695 offsets
->switch_new_end
= 0;
1696 offsets
->switch_old_end
= 0;
1697 offsets
->pre_header_padding
= 0;
1699 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1700 if ((int64_t) ctx
->tsc
== -EIO
)
1703 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1704 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
1706 if (unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1707 offsets
->switch_new_start
= 1; /* For offsets->begin */
1709 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1711 &offsets
->pre_header_padding
,
1714 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1717 if (unlikely(subbuf_offset(offsets
->begin
, chan
) +
1718 offsets
->size
> chan
->backend
.subbuf_size
)) {
1719 offsets
->switch_old_end
= 1; /* For offsets->old */
1720 offsets
->switch_new_start
= 1; /* For offsets->begin */
1723 if (unlikely(offsets
->switch_new_start
)) {
1724 unsigned long sb_index
, commit_count
;
1727 * We are typically not filling the previous buffer completely.
1729 if (likely(offsets
->switch_old_end
))
1730 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1731 offsets
->begin
= offsets
->begin
1732 + config
->cb
.subbuffer_header_size();
1733 /* Test new buffer integrity */
1734 sb_index
= subbuf_index(offsets
->begin
, chan
);
1736 * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
1737 * lib_ring_buffer_check_deliver() has the matching
1738 * memory barriers required around commit_cold cc_sb
1739 * updates to ensure reserve and commit counter updates
1740 * are not seen reordered when updated by another CPU.
1743 commit_count
= v_read(config
,
1744 &buf
->commit_cold
[sb_index
].cc_sb
);
1745 /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
1747 if (unlikely(offset_cmp
!= v_read(config
, &buf
->offset
))) {
1749 * The reserve counter have been concurrently updated
1750 * while we read the commit counter. This means the
1751 * commit counter we read might not match buf->offset
1752 * due to concurrent update. We therefore need to retry.
1756 reserve_commit_diff
=
1757 (buf_trunc(offsets
->begin
, chan
)
1758 >> chan
->backend
.num_subbuf_order
)
1759 - (commit_count
& chan
->commit_count_mask
);
1760 if (likely(reserve_commit_diff
== 0)) {
1761 /* Next subbuffer not being written to. */
1762 if (unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1763 subbuf_trunc(offsets
->begin
, chan
)
1764 - subbuf_trunc((unsigned long)
1765 atomic_long_read(&buf
->consumed
), chan
)
1766 >= chan
->backend
.buf_size
)) {
1768 * We do not overwrite non consumed buffers
1769 * and we are full : record is lost.
1771 v_inc(config
, &buf
->records_lost_full
);
1775 * Next subbuffer not being written to, and we
1776 * are either in overwrite mode or the buffer is
1777 * not full. It's safe to write in this new
1783 * Next subbuffer reserve offset does not match the
1784 * commit offset, and this did not involve update to the
1785 * reserve counter. Drop record in producer-consumer and
1786 * overwrite mode. Caused by either a writer OOPS or
1787 * too many nested writes over a reserve/commit pair.
1789 v_inc(config
, &buf
->records_lost_wrap
);
1793 config
->cb
.record_header_size(config
, chan
,
1795 &offsets
->pre_header_padding
,
1798 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1801 if (unlikely(subbuf_offset(offsets
->begin
, chan
)
1802 + offsets
->size
> chan
->backend
.subbuf_size
)) {
1804 * Record too big for subbuffers, report error, don't
1805 * complete the sub-buffer switch.
1807 v_inc(config
, &buf
->records_lost_big
);
1811 * We just made a successful buffer switch and the
1812 * record fits in the new subbuffer. Let's write.
1817 * Record fits in the current buffer and we are not on a switch
1818 * boundary. It's safe to write.
1821 offsets
->end
= offsets
->begin
+ offsets
->size
;
1823 if (unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
1825 * The offset_end will fall at the very beginning of the next
1828 offsets
->switch_new_end
= 1; /* For offsets->begin */
1834 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1835 * @ctx: ring buffer context.
1837 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
1838 * -EIO for other errors, else returns 0.
1839 * It will take care of sub-buffer switching.
1841 int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx
*ctx
)
1843 struct channel
*chan
= ctx
->chan
;
1844 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1845 struct lib_ring_buffer
*buf
;
1846 struct switch_offsets offsets
;
1849 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
1850 buf
= per_cpu_ptr(chan
->backend
.buf
, ctx
->cpu
);
1852 buf
= chan
->backend
.buf
;
1858 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
1862 } while (unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
1867 * Atomically update last_tsc. This update races against concurrent
1868 * atomic updates, but the race will always cause supplementary full TSC
1869 * records, never the opposite (missing a full TSC record when it would
1872 save_last_tsc(config
, buf
, ctx
->tsc
);
1875 * Push the reader if necessary
1877 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
1880 * Clear noref flag for this subbuffer.
1882 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1883 subbuf_index(offsets
.end
- 1, chan
));
1886 * Switch old subbuffer if needed.
1888 if (unlikely(offsets
.switch_old_end
)) {
1889 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
1890 subbuf_index(offsets
.old
- 1, chan
));
1891 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
);
1895 * Populate new subbuffer.
1897 if (unlikely(offsets
.switch_new_start
))
1898 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
);
1900 if (unlikely(offsets
.switch_new_end
))
1901 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
);
1903 ctx
->slot_size
= offsets
.size
;
1904 ctx
->pre_offset
= offsets
.begin
;
1905 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
1908 EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow
);
1910 int __init
init_lib_ring_buffer_frontend(void)
1914 for_each_possible_cpu(cpu
)
1915 spin_lock_init(&per_cpu(ring_buffer_nohz_lock
, cpu
));
1919 module_init(init_lib_ring_buffer_frontend
);
1921 void __exit
exit_lib_ring_buffer_frontend(void)
1925 module_exit(exit_lib_ring_buffer_frontend
);