1 /* SPDX-License-Identifier: (GPL-2.0 OR LGPL-2.1)
3 * ring_buffer_frontend.c
5 * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
7 * Ring buffer wait-free buffer synchronization. Producer-consumer and flight
8 * recorder (overwrite) modes. See thesis:
10 * Desnoyers, Mathieu (2009), "Low-Impact Operating System Tracing", Ph.D.
11 * dissertation, Ecole Polytechnique de Montreal.
12 * http://www.lttng.org/pub/thesis/desnoyers-dissertation-2009-12.pdf
14 * - Algorithm presentation in Chapter 5:
15 * "Lockless Multi-Core High-Throughput Buffering".
16 * - Algorithm formal verification in Section 8.6:
17 * "Formal verification of LTTng"
20 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
22 * Inspired from LTT and RelayFS:
23 * Karim Yaghmour <karim@opersys.com>
24 * Tom Zanussi <zanussi@us.ibm.com>
25 * Bob Wisniewski <bob@watson.ibm.com>
27 * Bob Wisniewski <bob@watson.ibm.com>
29 * Buffer reader semantic :
32 * while buffer is not finalized and empty
34 * - if return value != 0, continue
35 * - splice one subbuffer worth of data to a pipe
36 * - splice the data from pipe to disk/network
40 #include <linux/delay.h>
41 #include <linux/module.h>
42 #include <linux/percpu.h>
43 #include <asm/cacheflush.h>
45 #include <wrapper/ringbuffer/config.h>
46 #include <wrapper/ringbuffer/backend.h>
47 #include <wrapper/ringbuffer/frontend.h>
48 #include <wrapper/ringbuffer/iterator.h>
49 #include <wrapper/ringbuffer/nohz.h>
50 #include <wrapper/atomic.h>
51 #include <wrapper/kref.h>
52 #include <wrapper/percpu-defs.h>
53 #include <wrapper/timer.h>
54 #include <wrapper/vmalloc.h>
57 * Internal structure representing offsets to use at a sub-buffer switch.
59 struct switch_offsets
{
60 unsigned long begin
, end
, old
;
61 size_t pre_header_padding
, size
;
62 unsigned int switch_new_start
:1, switch_new_end
:1, switch_old_start
:1,
73 static ATOMIC_NOTIFIER_HEAD(tick_nohz_notifier
);
74 #endif /* CONFIG_NO_HZ */
76 static DEFINE_PER_CPU(spinlock_t
, ring_buffer_nohz_lock
);
78 DEFINE_PER_CPU(unsigned int, lib_ring_buffer_nesting
);
79 EXPORT_PER_CPU_SYMBOL(lib_ring_buffer_nesting
);
82 void lib_ring_buffer_print_errors(struct channel
*chan
,
83 struct lib_ring_buffer
*buf
, int cpu
);
85 void _lib_ring_buffer_switch_remote(struct lib_ring_buffer
*buf
,
86 enum switch_mode mode
);
89 int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config
*config
,
90 struct lib_ring_buffer
*buf
,
93 unsigned long consumed_old
, consumed_idx
, commit_count
, write_offset
;
95 consumed_old
= atomic_long_read(&buf
->consumed
);
96 consumed_idx
= subbuf_index(consumed_old
, chan
);
97 commit_count
= v_read(config
, &buf
->commit_cold
[consumed_idx
].cc_sb
);
99 * No memory barrier here, since we are only interested
100 * in a statistically correct polling result. The next poll will
101 * get the data is we are racing. The mb() that ensures correct
102 * memory order is in get_subbuf.
104 write_offset
= v_read(config
, &buf
->offset
);
107 * Check that the subbuffer we are trying to consume has been
108 * already fully committed.
111 if (((commit_count
- chan
->backend
.subbuf_size
)
112 & chan
->commit_count_mask
)
113 - (buf_trunc(consumed_old
, chan
)
114 >> chan
->backend
.num_subbuf_order
)
119 * Check that we are not about to read the same subbuffer in
120 * which the writer head is.
122 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_old
, chan
)
130 * Must be called under cpu hotplug protection.
132 void lib_ring_buffer_free(struct lib_ring_buffer
*buf
)
134 struct channel
*chan
= buf
->backend
.chan
;
136 lib_ring_buffer_print_errors(chan
, buf
, buf
->backend
.cpu
);
137 lttng_kvfree(buf
->commit_hot
);
138 lttng_kvfree(buf
->commit_cold
);
140 lib_ring_buffer_backend_free(&buf
->backend
);
144 * lib_ring_buffer_reset - Reset ring buffer to initial values.
147 * Effectively empty the ring buffer. Should be called when the buffer is not
148 * used for writing. The ring buffer can be opened for reading, but the reader
149 * should not be using the iterator concurrently with reset. The previous
150 * current iterator record is reset.
152 void lib_ring_buffer_reset(struct lib_ring_buffer
*buf
)
154 struct channel
*chan
= buf
->backend
.chan
;
155 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
159 * Reset iterator first. It will put the subbuffer if it currently holds
162 lib_ring_buffer_iterator_reset(buf
);
163 v_set(config
, &buf
->offset
, 0);
164 for (i
= 0; i
< chan
->backend
.num_subbuf
; i
++) {
165 v_set(config
, &buf
->commit_hot
[i
].cc
, 0);
166 v_set(config
, &buf
->commit_hot
[i
].seq
, 0);
167 v_set(config
, &buf
->commit_cold
[i
].cc_sb
, 0);
169 atomic_long_set(&buf
->consumed
, 0);
170 atomic_set(&buf
->record_disabled
, 0);
171 v_set(config
, &buf
->last_tsc
, 0);
172 lib_ring_buffer_backend_reset(&buf
->backend
);
173 /* Don't reset number of active readers */
174 v_set(config
, &buf
->records_lost_full
, 0);
175 v_set(config
, &buf
->records_lost_wrap
, 0);
176 v_set(config
, &buf
->records_lost_big
, 0);
177 v_set(config
, &buf
->records_count
, 0);
178 v_set(config
, &buf
->records_overrun
, 0);
181 EXPORT_SYMBOL_GPL(lib_ring_buffer_reset
);
184 * channel_reset - Reset channel to initial values.
187 * Effectively empty the channel. Should be called when the channel is not used
188 * for writing. The channel can be opened for reading, but the reader should not
189 * be using the iterator concurrently with reset. The previous current iterator
192 void channel_reset(struct channel
*chan
)
195 * Reset iterators first. Will put the subbuffer if held for reading.
197 channel_iterator_reset(chan
);
198 atomic_set(&chan
->record_disabled
, 0);
199 /* Don't reset commit_count_mask, still valid */
200 channel_backend_reset(&chan
->backend
);
201 /* Don't reset switch/read timer interval */
202 /* Don't reset notifiers and notifier enable bits */
203 /* Don't reset reader reference count */
205 EXPORT_SYMBOL_GPL(channel_reset
);
208 * Must be called under cpu hotplug protection.
210 int lib_ring_buffer_create(struct lib_ring_buffer
*buf
,
211 struct channel_backend
*chanb
, int cpu
)
213 const struct lib_ring_buffer_config
*config
= &chanb
->config
;
214 struct channel
*chan
= container_of(chanb
, struct channel
, backend
);
215 void *priv
= chanb
->priv
;
216 size_t subbuf_header_size
;
220 /* Test for cpu hotplug */
221 if (buf
->backend
.allocated
)
225 * Paranoia: per cpu dynamic allocation is not officially documented as
226 * zeroing the memory, so let's do it here too, just in case.
228 memset(buf
, 0, sizeof(*buf
));
230 ret
= lib_ring_buffer_backend_create(&buf
->backend
, &chan
->backend
, cpu
);
235 lttng_kvzalloc_node(ALIGN(sizeof(*buf
->commit_hot
)
236 * chan
->backend
.num_subbuf
,
237 1 << INTERNODE_CACHE_SHIFT
),
238 GFP_KERNEL
| __GFP_NOWARN
,
239 cpu_to_node(max(cpu
, 0)));
240 if (!buf
->commit_hot
) {
246 lttng_kvzalloc_node(ALIGN(sizeof(*buf
->commit_cold
)
247 * chan
->backend
.num_subbuf
,
248 1 << INTERNODE_CACHE_SHIFT
),
249 GFP_KERNEL
| __GFP_NOWARN
,
250 cpu_to_node(max(cpu
, 0)));
251 if (!buf
->commit_cold
) {
256 init_waitqueue_head(&buf
->read_wait
);
257 init_waitqueue_head(&buf
->write_wait
);
258 raw_spin_lock_init(&buf
->raw_tick_nohz_spinlock
);
261 * Write the subbuffer header for first subbuffer so we know the total
262 * duration of data gathering.
264 subbuf_header_size
= config
->cb
.subbuffer_header_size();
265 v_set(config
, &buf
->offset
, subbuf_header_size
);
266 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_wsb
[0].id
);
267 tsc
= config
->cb
.ring_buffer_clock_read(buf
->backend
.chan
);
268 config
->cb
.buffer_begin(buf
, tsc
, 0);
269 v_add(config
, subbuf_header_size
, &buf
->commit_hot
[0].cc
);
271 if (config
->cb
.buffer_create
) {
272 ret
= config
->cb
.buffer_create(buf
, priv
, cpu
, chanb
->name
);
278 * Ensure the buffer is ready before setting it to allocated and setting
280 * Used for cpu hotplug vs cpumask iteration.
283 buf
->backend
.allocated
= 1;
285 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
286 CHAN_WARN_ON(chan
, cpumask_test_cpu(cpu
,
287 chan
->backend
.cpumask
));
288 cpumask_set_cpu(cpu
, chan
->backend
.cpumask
);
295 lttng_kvfree(buf
->commit_cold
);
297 lttng_kvfree(buf
->commit_hot
);
299 lib_ring_buffer_backend_free(&buf
->backend
);
303 static void switch_buffer_timer(LTTNG_TIMER_FUNC_ARG_TYPE t
)
305 struct lib_ring_buffer
*buf
= lttng_from_timer(buf
, t
, switch_timer
);
306 struct channel
*chan
= buf
->backend
.chan
;
307 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
310 * Only flush buffers periodically if readers are active.
312 if (atomic_long_read(&buf
->active_readers
))
313 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
315 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
316 lttng_mod_timer_pinned(&buf
->switch_timer
,
317 jiffies
+ chan
->switch_timer_interval
);
319 mod_timer(&buf
->switch_timer
,
320 jiffies
+ chan
->switch_timer_interval
);
324 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
326 static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer
*buf
)
328 struct channel
*chan
= buf
->backend
.chan
;
329 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
330 unsigned int flags
= 0;
332 if (!chan
->switch_timer_interval
|| buf
->switch_timer_enabled
)
335 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
336 flags
= LTTNG_TIMER_PINNED
;
338 lttng_timer_setup(&buf
->switch_timer
, switch_buffer_timer
, flags
, buf
);
339 buf
->switch_timer
.expires
= jiffies
+ chan
->switch_timer_interval
;
341 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
342 add_timer_on(&buf
->switch_timer
, buf
->backend
.cpu
);
344 add_timer(&buf
->switch_timer
);
346 buf
->switch_timer_enabled
= 1;
350 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
352 static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer
*buf
)
354 struct channel
*chan
= buf
->backend
.chan
;
356 if (!chan
->switch_timer_interval
|| !buf
->switch_timer_enabled
)
359 del_timer_sync(&buf
->switch_timer
);
360 buf
->switch_timer_enabled
= 0;
364 * Polling timer to check the channels for data.
366 static void read_buffer_timer(LTTNG_TIMER_FUNC_ARG_TYPE t
)
368 struct lib_ring_buffer
*buf
= lttng_from_timer(buf
, t
, read_timer
);
369 struct channel
*chan
= buf
->backend
.chan
;
370 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
372 CHAN_WARN_ON(chan
, !buf
->backend
.allocated
);
374 if (atomic_long_read(&buf
->active_readers
)
375 && lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
376 wake_up_interruptible(&buf
->read_wait
);
377 wake_up_interruptible(&chan
->read_wait
);
380 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
381 lttng_mod_timer_pinned(&buf
->read_timer
,
382 jiffies
+ chan
->read_timer_interval
);
384 mod_timer(&buf
->read_timer
,
385 jiffies
+ chan
->read_timer_interval
);
389 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
391 static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer
*buf
)
393 struct channel
*chan
= buf
->backend
.chan
;
394 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
397 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
398 || !chan
->read_timer_interval
399 || buf
->read_timer_enabled
)
402 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
403 flags
= LTTNG_TIMER_PINNED
;
405 lttng_timer_setup(&buf
->read_timer
, read_buffer_timer
, flags
, buf
);
406 buf
->read_timer
.expires
= jiffies
+ chan
->read_timer_interval
;
408 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
409 add_timer_on(&buf
->read_timer
, buf
->backend
.cpu
);
411 add_timer(&buf
->read_timer
);
413 buf
->read_timer_enabled
= 1;
417 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
419 static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer
*buf
)
421 struct channel
*chan
= buf
->backend
.chan
;
422 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
424 if (config
->wakeup
!= RING_BUFFER_WAKEUP_BY_TIMER
425 || !chan
->read_timer_interval
426 || !buf
->read_timer_enabled
)
429 del_timer_sync(&buf
->read_timer
);
431 * do one more check to catch data that has been written in the last
434 if (lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
435 wake_up_interruptible(&buf
->read_wait
);
436 wake_up_interruptible(&chan
->read_wait
);
438 buf
->read_timer_enabled
= 0;
441 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0))
443 enum cpuhp_state lttng_rb_hp_prepare
;
444 enum cpuhp_state lttng_rb_hp_online
;
446 void lttng_rb_set_hp_prepare(enum cpuhp_state val
)
448 lttng_rb_hp_prepare
= val
;
450 EXPORT_SYMBOL_GPL(lttng_rb_set_hp_prepare
);
452 void lttng_rb_set_hp_online(enum cpuhp_state val
)
454 lttng_rb_hp_online
= val
;
456 EXPORT_SYMBOL_GPL(lttng_rb_set_hp_online
);
458 int lttng_cpuhp_rb_frontend_dead(unsigned int cpu
,
459 struct lttng_cpuhp_node
*node
)
461 struct channel
*chan
= container_of(node
, struct channel
,
463 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
, cpu
);
464 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
466 CHAN_WARN_ON(chan
, config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
);
469 * Performing a buffer switch on a remote CPU. Performed by
470 * the CPU responsible for doing the hotunplug after the target
471 * CPU stopped running completely. Ensures that all data
472 * from that remote CPU is flushed.
474 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
477 EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_dead
);
479 int lttng_cpuhp_rb_frontend_online(unsigned int cpu
,
480 struct lttng_cpuhp_node
*node
)
482 struct channel
*chan
= container_of(node
, struct channel
,
484 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
, cpu
);
485 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
487 CHAN_WARN_ON(chan
, config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
);
489 wake_up_interruptible(&chan
->hp_wait
);
490 lib_ring_buffer_start_switch_timer(buf
);
491 lib_ring_buffer_start_read_timer(buf
);
494 EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_online
);
496 int lttng_cpuhp_rb_frontend_offline(unsigned int cpu
,
497 struct lttng_cpuhp_node
*node
)
499 struct channel
*chan
= container_of(node
, struct channel
,
501 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
, cpu
);
502 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
504 CHAN_WARN_ON(chan
, config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
);
506 lib_ring_buffer_stop_switch_timer(buf
);
507 lib_ring_buffer_stop_read_timer(buf
);
510 EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_offline
);
512 #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */
514 #ifdef CONFIG_HOTPLUG_CPU
517 * lib_ring_buffer_cpu_hp_callback - CPU hotplug callback
518 * @nb: notifier block
519 * @action: hotplug action to take
522 * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
525 int lib_ring_buffer_cpu_hp_callback(struct notifier_block
*nb
,
526 unsigned long action
,
529 unsigned int cpu
= (unsigned long)hcpu
;
530 struct channel
*chan
= container_of(nb
, struct channel
,
532 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
, cpu
);
533 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
535 if (!chan
->cpu_hp_enable
)
538 CHAN_WARN_ON(chan
, config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
);
541 case CPU_DOWN_FAILED
:
542 case CPU_DOWN_FAILED_FROZEN
:
544 case CPU_ONLINE_FROZEN
:
545 wake_up_interruptible(&chan
->hp_wait
);
546 lib_ring_buffer_start_switch_timer(buf
);
547 lib_ring_buffer_start_read_timer(buf
);
550 case CPU_DOWN_PREPARE
:
551 case CPU_DOWN_PREPARE_FROZEN
:
552 lib_ring_buffer_stop_switch_timer(buf
);
553 lib_ring_buffer_stop_read_timer(buf
);
557 case CPU_DEAD_FROZEN
:
559 * Performing a buffer switch on a remote CPU. Performed by
560 * the CPU responsible for doing the hotunplug after the target
561 * CPU stopped running completely. Ensures that all data
562 * from that remote CPU is flushed.
564 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
574 #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */
576 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
578 * For per-cpu buffers, call the reader wakeups before switching the buffer, so
579 * that wake-up-tracing generated events are flushed before going idle (in
580 * tick_nohz). We test if the spinlock is locked to deal with the race where
581 * readers try to sample the ring buffer before we perform the switch. We let
582 * the readers retry in that case. If there is data in the buffer, the wake up
583 * is going to forbid the CPU running the reader thread from going idle.
585 static int notrace
ring_buffer_tick_nohz_callback(struct notifier_block
*nb
,
589 struct channel
*chan
= container_of(nb
, struct channel
,
591 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
592 struct lib_ring_buffer
*buf
;
593 int cpu
= smp_processor_id();
595 if (config
->alloc
!= RING_BUFFER_ALLOC_PER_CPU
) {
597 * We don't support keeping the system idle with global buffers
598 * and streaming active. In order to do so, we would need to
599 * sample a non-nohz-cpumask racelessly with the nohz updates
600 * without adding synchronization overhead to nohz. Leave this
601 * use-case out for now.
606 buf
= channel_get_ring_buffer(config
, chan
, cpu
);
608 case TICK_NOHZ_FLUSH
:
609 raw_spin_lock(&buf
->raw_tick_nohz_spinlock
);
610 if (config
->wakeup
== RING_BUFFER_WAKEUP_BY_TIMER
611 && chan
->read_timer_interval
612 && atomic_long_read(&buf
->active_readers
)
613 && (lib_ring_buffer_poll_deliver(config
, buf
, chan
)
614 || lib_ring_buffer_pending_data(config
, buf
, chan
))) {
615 wake_up_interruptible(&buf
->read_wait
);
616 wake_up_interruptible(&chan
->read_wait
);
618 if (chan
->switch_timer_interval
)
619 lib_ring_buffer_switch_slow(buf
, SWITCH_ACTIVE
);
620 raw_spin_unlock(&buf
->raw_tick_nohz_spinlock
);
623 spin_lock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
624 lib_ring_buffer_stop_switch_timer(buf
);
625 lib_ring_buffer_stop_read_timer(buf
);
626 spin_unlock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
628 case TICK_NOHZ_RESTART
:
629 spin_lock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
630 lib_ring_buffer_start_read_timer(buf
);
631 lib_ring_buffer_start_switch_timer(buf
);
632 spin_unlock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock
));
639 void notrace
lib_ring_buffer_tick_nohz_flush(void)
641 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_FLUSH
,
645 void notrace
lib_ring_buffer_tick_nohz_stop(void)
647 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_STOP
,
651 void notrace
lib_ring_buffer_tick_nohz_restart(void)
653 atomic_notifier_call_chain(&tick_nohz_notifier
, TICK_NOHZ_RESTART
,
656 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
661 static void channel_unregister_notifiers(struct channel
*chan
)
663 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
665 channel_iterator_unregister_notifiers(chan
);
666 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
669 * Remove the nohz notifier first, so we are certain we stop
672 atomic_notifier_chain_unregister(&tick_nohz_notifier
,
673 &chan
->tick_nohz_notifier
);
675 * ring_buffer_nohz_lock will not be needed below, because
676 * we just removed the notifiers, which were the only source of
679 #endif /* CONFIG_NO_HZ */
680 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0))
684 ret
= cpuhp_state_remove_instance(lttng_rb_hp_online
,
685 &chan
->cpuhp_online
.node
);
687 ret
= cpuhp_state_remove_instance_nocalls(lttng_rb_hp_prepare
,
688 &chan
->cpuhp_prepare
.node
);
691 #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */
695 #ifdef CONFIG_HOTPLUG_CPU
697 chan
->cpu_hp_enable
= 0;
698 for_each_online_cpu(cpu
) {
699 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
701 lib_ring_buffer_stop_switch_timer(buf
);
702 lib_ring_buffer_stop_read_timer(buf
);
705 unregister_cpu_notifier(&chan
->cpu_hp_notifier
);
707 for_each_possible_cpu(cpu
) {
708 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
710 lib_ring_buffer_stop_switch_timer(buf
);
711 lib_ring_buffer_stop_read_timer(buf
);
715 #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */
717 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
719 lib_ring_buffer_stop_switch_timer(buf
);
720 lib_ring_buffer_stop_read_timer(buf
);
722 channel_backend_unregister_notifiers(&chan
->backend
);
725 static void lib_ring_buffer_set_quiescent(struct lib_ring_buffer
*buf
)
727 if (!buf
->quiescent
) {
728 buf
->quiescent
= true;
729 _lib_ring_buffer_switch_remote(buf
, SWITCH_FLUSH
);
733 static void lib_ring_buffer_clear_quiescent(struct lib_ring_buffer
*buf
)
735 buf
->quiescent
= false;
738 void lib_ring_buffer_set_quiescent_channel(struct channel
*chan
)
741 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
743 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
745 for_each_channel_cpu(cpu
, chan
) {
746 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
749 lib_ring_buffer_set_quiescent(buf
);
753 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
755 lib_ring_buffer_set_quiescent(buf
);
758 EXPORT_SYMBOL_GPL(lib_ring_buffer_set_quiescent_channel
);
760 void lib_ring_buffer_clear_quiescent_channel(struct channel
*chan
)
763 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
765 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
767 for_each_channel_cpu(cpu
, chan
) {
768 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
771 lib_ring_buffer_clear_quiescent(buf
);
775 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
777 lib_ring_buffer_clear_quiescent(buf
);
780 EXPORT_SYMBOL_GPL(lib_ring_buffer_clear_quiescent_channel
);
782 static void channel_free(struct channel
*chan
)
784 if (chan
->backend
.release_priv_ops
) {
785 chan
->backend
.release_priv_ops(chan
->backend
.priv_ops
);
787 channel_iterator_free(chan
);
788 channel_backend_free(&chan
->backend
);
793 * channel_create - Create channel.
794 * @config: ring buffer instance configuration
795 * @name: name of the channel
796 * @priv: ring buffer client private data
797 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
798 * address mapping. It is used only by RING_BUFFER_STATIC
799 * configuration. It can be set to NULL for other backends.
800 * @subbuf_size: subbuffer size
801 * @num_subbuf: number of subbuffers
802 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
803 * padding to let readers get those sub-buffers.
804 * Used for live streaming.
805 * @read_timer_interval: Time interval (in us) to wake up pending readers.
808 * Returns NULL on failure.
810 struct channel
*channel_create(const struct lib_ring_buffer_config
*config
,
811 const char *name
, void *priv
, void *buf_addr
,
813 size_t num_subbuf
, unsigned int switch_timer_interval
,
814 unsigned int read_timer_interval
)
817 struct channel
*chan
;
819 if (lib_ring_buffer_check_config(config
, switch_timer_interval
,
820 read_timer_interval
))
823 chan
= kzalloc(sizeof(struct channel
), GFP_KERNEL
);
827 ret
= channel_backend_init(&chan
->backend
, name
, config
, priv
,
828 subbuf_size
, num_subbuf
);
832 ret
= channel_iterator_init(chan
);
834 goto error_free_backend
;
836 chan
->commit_count_mask
= (~0UL >> chan
->backend
.num_subbuf_order
);
837 chan
->switch_timer_interval
= usecs_to_jiffies(switch_timer_interval
);
838 chan
->read_timer_interval
= usecs_to_jiffies(read_timer_interval
);
839 kref_init(&chan
->ref
);
840 init_waitqueue_head(&chan
->read_wait
);
841 init_waitqueue_head(&chan
->hp_wait
);
843 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
844 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0))
845 chan
->cpuhp_prepare
.component
= LTTNG_RING_BUFFER_FRONTEND
;
846 ret
= cpuhp_state_add_instance_nocalls(lttng_rb_hp_prepare
,
847 &chan
->cpuhp_prepare
.node
);
849 goto cpuhp_prepare_error
;
851 chan
->cpuhp_online
.component
= LTTNG_RING_BUFFER_FRONTEND
;
852 ret
= cpuhp_state_add_instance(lttng_rb_hp_online
,
853 &chan
->cpuhp_online
.node
);
855 goto cpuhp_online_error
;
856 #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */
860 * In case of non-hotplug cpu, if the ring-buffer is allocated
861 * in early initcall, it will not be notified of secondary cpus.
862 * In that off case, we need to allocate for all possible cpus.
864 #ifdef CONFIG_HOTPLUG_CPU
865 chan
->cpu_hp_notifier
.notifier_call
=
866 lib_ring_buffer_cpu_hp_callback
;
867 chan
->cpu_hp_notifier
.priority
= 6;
868 register_cpu_notifier(&chan
->cpu_hp_notifier
);
871 for_each_online_cpu(cpu
) {
872 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
874 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
875 lib_ring_buffer_start_switch_timer(buf
);
876 lib_ring_buffer_start_read_timer(buf
);
877 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
879 chan
->cpu_hp_enable
= 1;
882 for_each_possible_cpu(cpu
) {
883 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
885 spin_lock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
886 lib_ring_buffer_start_switch_timer(buf
);
887 lib_ring_buffer_start_read_timer(buf
);
888 spin_unlock(&per_cpu(ring_buffer_nohz_lock
, cpu
));
892 #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */
894 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
895 /* Only benefit from NO_HZ idle with per-cpu buffers for now. */
896 chan
->tick_nohz_notifier
.notifier_call
=
897 ring_buffer_tick_nohz_callback
;
898 chan
->tick_nohz_notifier
.priority
= ~0U;
899 atomic_notifier_chain_register(&tick_nohz_notifier
,
900 &chan
->tick_nohz_notifier
);
901 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
904 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
906 lib_ring_buffer_start_switch_timer(buf
);
907 lib_ring_buffer_start_read_timer(buf
);
912 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0))
914 ret
= cpuhp_state_remove_instance_nocalls(lttng_rb_hp_prepare
,
915 &chan
->cpuhp_prepare
.node
);
918 #endif /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */
920 channel_backend_free(&chan
->backend
);
925 EXPORT_SYMBOL_GPL(channel_create
);
928 void channel_release(struct kref
*kref
)
930 struct channel
*chan
= container_of(kref
, struct channel
, ref
);
935 * channel_destroy - Finalize, wait for q.s. and destroy channel.
936 * @chan: channel to destroy
939 * Call "destroy" callback, finalize channels, and then decrement the
940 * channel reference count. Note that when readers have completed data
941 * consumption of finalized channels, get_subbuf() will return -ENODATA.
942 * They should release their handle at that point. Returns the private
945 void *channel_destroy(struct channel
*chan
)
948 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
951 channel_unregister_notifiers(chan
);
953 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
955 * No need to hold cpu hotplug, because all notifiers have been
958 for_each_channel_cpu(cpu
, chan
) {
959 struct lib_ring_buffer
*buf
= per_cpu_ptr(chan
->backend
.buf
,
962 if (config
->cb
.buffer_finalize
)
963 config
->cb
.buffer_finalize(buf
,
966 if (buf
->backend
.allocated
)
967 lib_ring_buffer_set_quiescent(buf
);
969 * Perform flush before writing to finalized.
972 WRITE_ONCE(buf
->finalized
, 1);
973 wake_up_interruptible(&buf
->read_wait
);
976 struct lib_ring_buffer
*buf
= chan
->backend
.buf
;
978 if (config
->cb
.buffer_finalize
)
979 config
->cb
.buffer_finalize(buf
, chan
->backend
.priv
, -1);
980 if (buf
->backend
.allocated
)
981 lib_ring_buffer_set_quiescent(buf
);
983 * Perform flush before writing to finalized.
986 WRITE_ONCE(buf
->finalized
, 1);
987 wake_up_interruptible(&buf
->read_wait
);
989 WRITE_ONCE(chan
->finalized
, 1);
990 wake_up_interruptible(&chan
->hp_wait
);
991 wake_up_interruptible(&chan
->read_wait
);
992 priv
= chan
->backend
.priv
;
993 kref_put(&chan
->ref
, channel_release
);
996 EXPORT_SYMBOL_GPL(channel_destroy
);
998 struct lib_ring_buffer
*channel_get_ring_buffer(
999 const struct lib_ring_buffer_config
*config
,
1000 struct channel
*chan
, int cpu
)
1002 if (config
->alloc
== RING_BUFFER_ALLOC_GLOBAL
)
1003 return chan
->backend
.buf
;
1005 return per_cpu_ptr(chan
->backend
.buf
, cpu
);
1007 EXPORT_SYMBOL_GPL(channel_get_ring_buffer
);
1009 int lib_ring_buffer_open_read(struct lib_ring_buffer
*buf
)
1011 struct channel
*chan
= buf
->backend
.chan
;
1013 if (!atomic_long_add_unless(&buf
->active_readers
, 1, 1))
1015 if (!lttng_kref_get(&chan
->ref
)) {
1016 atomic_long_dec(&buf
->active_readers
);
1019 lttng_smp_mb__after_atomic();
1022 EXPORT_SYMBOL_GPL(lib_ring_buffer_open_read
);
1024 void lib_ring_buffer_release_read(struct lib_ring_buffer
*buf
)
1026 struct channel
*chan
= buf
->backend
.chan
;
1028 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
1029 lttng_smp_mb__before_atomic();
1030 atomic_long_dec(&buf
->active_readers
);
1031 kref_put(&chan
->ref
, channel_release
);
1033 EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read
);
1036 * Promote compiler barrier to a smp_mb().
1037 * For the specific ring buffer case, this IPI call should be removed if the
1038 * architecture does not reorder writes. This should eventually be provided by
1039 * a separate architecture-specific infrastructure.
1041 static void remote_mb(void *info
)
1047 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
1049 * @consumed: consumed count indicating the position where to read
1050 * @produced: produced count, indicates position when to stop reading
1052 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1053 * data to read at consumed position, or 0 if the get operation succeeds.
1054 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
1057 int lib_ring_buffer_snapshot(struct lib_ring_buffer
*buf
,
1058 unsigned long *consumed
, unsigned long *produced
)
1060 struct channel
*chan
= buf
->backend
.chan
;
1061 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1062 unsigned long consumed_cur
, write_offset
;
1066 finalized
= READ_ONCE(buf
->finalized
);
1068 * Read finalized before counters.
1071 consumed_cur
= atomic_long_read(&buf
->consumed
);
1073 * No need to issue a memory barrier between consumed count read and
1074 * write offset read, because consumed count can only change
1075 * concurrently in overwrite mode, and we keep a sequence counter
1076 * identifier derived from the write offset to check we are getting
1077 * the same sub-buffer we are expecting (the sub-buffers are atomically
1078 * "tagged" upon writes, tags are checked upon read).
1080 write_offset
= v_read(config
, &buf
->offset
);
1083 * Check that we are not about to read the same subbuffer in
1084 * which the writer head is.
1086 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed_cur
, chan
)
1090 *consumed
= consumed_cur
;
1091 *produced
= subbuf_trunc(write_offset
, chan
);
1097 * The memory barriers __wait_event()/wake_up_interruptible() take care
1098 * of "raw_spin_is_locked" memory ordering.
1102 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
1107 EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot
);
1110 * Performs the same function as lib_ring_buffer_snapshot(), but the positions
1111 * are saved regardless of whether the consumed and produced positions are
1112 * in the same subbuffer.
1114 * @consumed: consumed byte count indicating the last position read
1115 * @produced: produced byte count indicating the last position written
1117 * This function is meant to provide information on the exact producer and
1118 * consumer positions without regard for the "snapshot" feature.
1120 int lib_ring_buffer_snapshot_sample_positions(struct lib_ring_buffer
*buf
,
1121 unsigned long *consumed
, unsigned long *produced
)
1123 struct channel
*chan
= buf
->backend
.chan
;
1124 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1127 *consumed
= atomic_long_read(&buf
->consumed
);
1129 * No need to issue a memory barrier between consumed count read and
1130 * write offset read, because consumed count can only change
1131 * concurrently in overwrite mode, and we keep a sequence counter
1132 * identifier derived from the write offset to check we are getting
1133 * the same sub-buffer we are expecting (the sub-buffers are atomically
1134 * "tagged" upon writes, tags are checked upon read).
1136 *produced
= v_read(config
, &buf
->offset
);
1141 * lib_ring_buffer_put_snapshot - move consumed counter forward
1143 * Should only be called from consumer context.
1145 * @consumed_new: new consumed count value
1147 void lib_ring_buffer_move_consumer(struct lib_ring_buffer
*buf
,
1148 unsigned long consumed_new
)
1150 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1151 struct channel
*chan
= bufb
->chan
;
1152 unsigned long consumed
;
1154 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
1157 * Only push the consumed value forward.
1158 * If the consumed cmpxchg fails, this is because we have been pushed by
1159 * the writer in flight recorder mode.
1161 consumed
= atomic_long_read(&buf
->consumed
);
1162 while ((long) consumed
- (long) consumed_new
< 0)
1163 consumed
= atomic_long_cmpxchg(&buf
->consumed
, consumed
,
1165 /* Wake-up the metadata producer */
1166 wake_up_interruptible(&buf
->write_wait
);
1168 EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer
);
1170 #if ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE
1171 static void lib_ring_buffer_flush_read_subbuf_dcache(
1172 const struct lib_ring_buffer_config
*config
,
1173 struct channel
*chan
,
1174 struct lib_ring_buffer
*buf
)
1176 struct lib_ring_buffer_backend_pages
*pages
;
1177 unsigned long sb_bindex
, id
, i
, nr_pages
;
1179 if (config
->output
!= RING_BUFFER_MMAP
)
1183 * Architectures with caches aliased on virtual addresses may
1184 * use different cache lines for the linear mapping vs
1185 * user-space memory mapping. Given that the ring buffer is
1186 * based on the kernel linear mapping, aligning it with the
1187 * user-space mapping is not straightforward, and would require
1188 * extra TLB entries. Therefore, simply flush the dcache for the
1189 * entire sub-buffer before reading it.
1191 id
= buf
->backend
.buf_rsb
.id
;
1192 sb_bindex
= subbuffer_id_get_index(config
, id
);
1193 pages
= buf
->backend
.array
[sb_bindex
];
1194 nr_pages
= buf
->backend
.num_pages_per_subbuf
;
1195 for (i
= 0; i
< nr_pages
; i
++) {
1196 struct lib_ring_buffer_backend_page
*backend_page
;
1198 backend_page
= &pages
->p
[i
];
1199 flush_dcache_page(pfn_to_page(backend_page
->pfn
));
1203 static void lib_ring_buffer_flush_read_subbuf_dcache(
1204 const struct lib_ring_buffer_config
*config
,
1205 struct channel
*chan
,
1206 struct lib_ring_buffer
*buf
)
1212 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
1214 * @consumed: consumed count indicating the position where to read
1216 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
1217 * data to read at consumed position, or 0 if the get operation succeeds.
1218 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
1220 int lib_ring_buffer_get_subbuf(struct lib_ring_buffer
*buf
,
1221 unsigned long consumed
)
1223 struct channel
*chan
= buf
->backend
.chan
;
1224 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1225 unsigned long consumed_cur
, consumed_idx
, commit_count
, write_offset
;
1229 if (buf
->get_subbuf
) {
1231 * Reader is trying to get a subbuffer twice.
1233 CHAN_WARN_ON(chan
, 1);
1237 finalized
= READ_ONCE(buf
->finalized
);
1239 * Read finalized before counters.
1242 consumed_cur
= atomic_long_read(&buf
->consumed
);
1243 consumed_idx
= subbuf_index(consumed
, chan
);
1244 commit_count
= v_read(config
, &buf
->commit_cold
[consumed_idx
].cc_sb
);
1246 * Make sure we read the commit count before reading the buffer
1247 * data and the write offset. Correct consumed offset ordering
1248 * wrt commit count is insured by the use of cmpxchg to update
1249 * the consumed offset.
1250 * smp_call_function_single can fail if the remote CPU is offline,
1251 * this is OK because then there is no wmb to execute there.
1252 * If our thread is executing on the same CPU as the on the buffers
1253 * belongs to, we don't have to synchronize it at all. If we are
1254 * migrated, the scheduler will take care of the memory barriers.
1255 * Normally, smp_call_function_single() should ensure program order when
1256 * executing the remote function, which implies that it surrounds the
1257 * function execution with :
1268 * However, smp_call_function_single() does not seem to clearly execute
1269 * such barriers. It depends on spinlock semantic to provide the barrier
1270 * before executing the IPI and, when busy-looping, csd_lock_wait only
1271 * executes smp_mb() when it has to wait for the other CPU.
1273 * I don't trust this code. Therefore, let's add the smp_mb() sequence
1274 * required ourself, even if duplicated. It has no performance impact
1277 * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs
1278 * read and write vs write. They do not ensure core synchronization. We
1279 * really have to ensure total order between the 3 barriers running on
1282 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1283 if (config
->sync
== RING_BUFFER_SYNC_PER_CPU
1284 && config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
) {
1285 if (raw_smp_processor_id() != buf
->backend
.cpu
) {
1286 /* Total order with IPI handler smp_mb() */
1288 smp_call_function_single(buf
->backend
.cpu
,
1289 remote_mb
, NULL
, 1);
1290 /* Total order with IPI handler smp_mb() */
1294 /* Total order with IPI handler smp_mb() */
1296 smp_call_function(remote_mb
, NULL
, 1);
1297 /* Total order with IPI handler smp_mb() */
1302 * Local rmb to match the remote wmb to read the commit count
1303 * before the buffer data and the write offset.
1308 write_offset
= v_read(config
, &buf
->offset
);
1311 * Check that the buffer we are getting is after or at consumed_cur
1314 if ((long) subbuf_trunc(consumed
, chan
)
1315 - (long) subbuf_trunc(consumed_cur
, chan
) < 0)
1319 * Check that the subbuffer we are trying to consume has been
1320 * already fully committed.
1322 if (((commit_count
- chan
->backend
.subbuf_size
)
1323 & chan
->commit_count_mask
)
1324 - (buf_trunc(consumed
, chan
)
1325 >> chan
->backend
.num_subbuf_order
)
1330 * Check that we are not about to read the same subbuffer in
1331 * which the writer head is.
1333 if (subbuf_trunc(write_offset
, chan
) - subbuf_trunc(consumed
, chan
)
1338 * Failure to get the subbuffer causes a busy-loop retry without going
1339 * to a wait queue. These are caused by short-lived race windows where
1340 * the writer is getting access to a subbuffer we were trying to get
1341 * access to. Also checks that the "consumed" buffer count we are
1342 * looking for matches the one contained in the subbuffer id.
1344 ret
= update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1345 consumed_idx
, buf_trunc_val(consumed
, chan
));
1348 subbuffer_id_clear_noref(config
, &buf
->backend
.buf_rsb
.id
);
1350 buf
->get_subbuf_consumed
= consumed
;
1351 buf
->get_subbuf
= 1;
1353 lib_ring_buffer_flush_read_subbuf_dcache(config
, chan
, buf
);
1359 * The memory barriers __wait_event()/wake_up_interruptible() take care
1360 * of "raw_spin_is_locked" memory ordering.
1364 else if (raw_spin_is_locked(&buf
->raw_tick_nohz_spinlock
))
1369 EXPORT_SYMBOL_GPL(lib_ring_buffer_get_subbuf
);
1372 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1375 void lib_ring_buffer_put_subbuf(struct lib_ring_buffer
*buf
)
1377 struct lib_ring_buffer_backend
*bufb
= &buf
->backend
;
1378 struct channel
*chan
= bufb
->chan
;
1379 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1380 unsigned long read_sb_bindex
, consumed_idx
, consumed
;
1382 CHAN_WARN_ON(chan
, atomic_long_read(&buf
->active_readers
) != 1);
1384 if (!buf
->get_subbuf
) {
1386 * Reader puts a subbuffer it did not get.
1388 CHAN_WARN_ON(chan
, 1);
1391 consumed
= buf
->get_subbuf_consumed
;
1392 buf
->get_subbuf
= 0;
1395 * Clear the records_unread counter. (overruns counter)
1396 * Can still be non-zero if a file reader simply grabbed the data
1397 * without using iterators.
1398 * Can be below zero if an iterator is used on a snapshot more than
1401 read_sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
1402 v_add(config
, v_read(config
,
1403 &bufb
->array
[read_sb_bindex
]->records_unread
),
1404 &bufb
->records_read
);
1405 v_set(config
, &bufb
->array
[read_sb_bindex
]->records_unread
, 0);
1406 CHAN_WARN_ON(chan
, config
->mode
== RING_BUFFER_OVERWRITE
1407 && subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
1408 subbuffer_id_set_noref(config
, &bufb
->buf_rsb
.id
);
1411 * Exchange the reader subbuffer with the one we put in its place in the
1412 * writer subbuffer table. Expect the original consumed count. If
1413 * update_read_sb_index fails, this is because the writer updated the
1414 * subbuffer concurrently. We should therefore keep the subbuffer we
1415 * currently have: it has become invalid to try reading this sub-buffer
1416 * consumed count value anyway.
1418 consumed_idx
= subbuf_index(consumed
, chan
);
1419 update_read_sb_index(config
, &buf
->backend
, &chan
->backend
,
1420 consumed_idx
, buf_trunc_val(consumed
, chan
));
1422 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1423 * if the writer concurrently updated it.
1426 EXPORT_SYMBOL_GPL(lib_ring_buffer_put_subbuf
);
1429 * cons_offset is an iterator on all subbuffer offsets between the reader
1430 * position and the writer position. (inclusive)
1433 void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer
*buf
,
1434 struct channel
*chan
,
1435 unsigned long cons_offset
,
1438 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1439 unsigned long cons_idx
, commit_count
, commit_count_sb
;
1441 cons_idx
= subbuf_index(cons_offset
, chan
);
1442 commit_count
= v_read(config
, &buf
->commit_hot
[cons_idx
].cc
);
1443 commit_count_sb
= v_read(config
, &buf
->commit_cold
[cons_idx
].cc_sb
);
1445 if (subbuf_offset(commit_count
, chan
) != 0)
1447 "ring buffer %s, cpu %d: "
1448 "commit count in subbuffer %lu,\n"
1449 "expecting multiples of %lu bytes\n"
1450 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1451 chan
->backend
.name
, cpu
, cons_idx
,
1452 chan
->backend
.subbuf_size
,
1453 commit_count
, commit_count_sb
);
1455 printk(KERN_DEBUG
"ring buffer: %s, cpu %d: %lu bytes committed\n",
1456 chan
->backend
.name
, cpu
, commit_count
);
1460 void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer
*buf
,
1461 struct channel
*chan
,
1462 void *priv
, int cpu
)
1464 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1465 unsigned long write_offset
, cons_offset
;
1468 * No need to order commit_count, write_offset and cons_offset reads
1469 * because we execute at teardown when no more writer nor reader
1470 * references are left.
1472 write_offset
= v_read(config
, &buf
->offset
);
1473 cons_offset
= atomic_long_read(&buf
->consumed
);
1474 if (write_offset
!= cons_offset
)
1476 "ring buffer %s, cpu %d: "
1477 "non-consumed data\n"
1478 " [ %lu bytes written, %lu bytes read ]\n",
1479 chan
->backend
.name
, cpu
, write_offset
, cons_offset
);
1481 for (cons_offset
= atomic_long_read(&buf
->consumed
);
1482 (long) (subbuf_trunc((unsigned long) v_read(config
, &buf
->offset
),
1485 cons_offset
= subbuf_align(cons_offset
, chan
))
1486 lib_ring_buffer_print_subbuffer_errors(buf
, chan
, cons_offset
,
1491 void lib_ring_buffer_print_errors(struct channel
*chan
,
1492 struct lib_ring_buffer
*buf
, int cpu
)
1494 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1495 void *priv
= chan
->backend
.priv
;
1497 if (!strcmp(chan
->backend
.name
, "relay-metadata")) {
1498 printk(KERN_DEBUG
"ring buffer %s: %lu records written, "
1499 "%lu records overrun\n",
1501 v_read(config
, &buf
->records_count
),
1502 v_read(config
, &buf
->records_overrun
));
1504 printk(KERN_DEBUG
"ring buffer %s, cpu %d: %lu records written, "
1505 "%lu records overrun\n",
1506 chan
->backend
.name
, cpu
,
1507 v_read(config
, &buf
->records_count
),
1508 v_read(config
, &buf
->records_overrun
));
1510 if (v_read(config
, &buf
->records_lost_full
)
1511 || v_read(config
, &buf
->records_lost_wrap
)
1512 || v_read(config
, &buf
->records_lost_big
))
1514 "ring buffer %s, cpu %d: records were lost. Caused by:\n"
1515 " [ %lu buffer full, %lu nest buffer wrap-around, "
1516 "%lu event too big ]\n",
1517 chan
->backend
.name
, cpu
,
1518 v_read(config
, &buf
->records_lost_full
),
1519 v_read(config
, &buf
->records_lost_wrap
),
1520 v_read(config
, &buf
->records_lost_big
));
1522 lib_ring_buffer_print_buffer_errors(buf
, chan
, priv
, cpu
);
1526 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1528 * Only executed when the buffer is finalized, in SWITCH_FLUSH.
1531 void lib_ring_buffer_switch_old_start(struct lib_ring_buffer
*buf
,
1532 struct channel
*chan
,
1533 struct switch_offsets
*offsets
,
1536 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1537 unsigned long oldidx
= subbuf_index(offsets
->old
, chan
);
1538 unsigned long commit_count
;
1539 struct commit_counters_hot
*cc_hot
;
1541 config
->cb
.buffer_begin(buf
, tsc
, oldidx
);
1544 * Order all writes to buffer before the commit count update that will
1545 * determine that the subbuffer is full.
1547 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1549 * Must write slot data before incrementing commit count. This
1550 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1556 cc_hot
= &buf
->commit_hot
[oldidx
];
1557 v_add(config
, config
->cb
.subbuffer_header_size(), &cc_hot
->cc
);
1558 commit_count
= v_read(config
, &cc_hot
->cc
);
1559 /* Check if the written buffer has to be delivered */
1560 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
,
1561 commit_count
, oldidx
, tsc
);
1562 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1563 offsets
->old
+ config
->cb
.subbuffer_header_size(),
1564 commit_count
, cc_hot
);
1568 * lib_ring_buffer_switch_old_end: switch old subbuffer
1570 * Note : offset_old should never be 0 here. It is ok, because we never perform
1571 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1572 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1576 void lib_ring_buffer_switch_old_end(struct lib_ring_buffer
*buf
,
1577 struct channel
*chan
,
1578 struct switch_offsets
*offsets
,
1581 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1582 unsigned long oldidx
= subbuf_index(offsets
->old
- 1, chan
);
1583 unsigned long commit_count
, padding_size
, data_size
;
1584 struct commit_counters_hot
*cc_hot
;
1586 data_size
= subbuf_offset(offsets
->old
- 1, chan
) + 1;
1587 padding_size
= chan
->backend
.subbuf_size
- data_size
;
1588 subbuffer_set_data_size(config
, &buf
->backend
, oldidx
, data_size
);
1591 * Order all writes to buffer before the commit count update that will
1592 * determine that the subbuffer is full.
1594 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1596 * Must write slot data before incrementing commit count. This
1597 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1603 cc_hot
= &buf
->commit_hot
[oldidx
];
1604 v_add(config
, padding_size
, &cc_hot
->cc
);
1605 commit_count
= v_read(config
, &cc_hot
->cc
);
1606 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->old
- 1,
1607 commit_count
, oldidx
, tsc
);
1608 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1609 offsets
->old
+ padding_size
, commit_count
,
1614 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1616 * This code can be executed unordered : writers may already have written to the
1617 * sub-buffer before this code gets executed, caution. The commit makes sure
1618 * that this code is executed before the deliver of this sub-buffer.
1621 void lib_ring_buffer_switch_new_start(struct lib_ring_buffer
*buf
,
1622 struct channel
*chan
,
1623 struct switch_offsets
*offsets
,
1626 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1627 unsigned long beginidx
= subbuf_index(offsets
->begin
, chan
);
1628 unsigned long commit_count
;
1629 struct commit_counters_hot
*cc_hot
;
1631 config
->cb
.buffer_begin(buf
, tsc
, beginidx
);
1634 * Order all writes to buffer before the commit count update that will
1635 * determine that the subbuffer is full.
1637 if (config
->ipi
== RING_BUFFER_IPI_BARRIER
) {
1639 * Must write slot data before incrementing commit count. This
1640 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1646 cc_hot
= &buf
->commit_hot
[beginidx
];
1647 v_add(config
, config
->cb
.subbuffer_header_size(), &cc_hot
->cc
);
1648 commit_count
= v_read(config
, &cc_hot
->cc
);
1649 /* Check if the written buffer has to be delivered */
1650 lib_ring_buffer_check_deliver(config
, buf
, chan
, offsets
->begin
,
1651 commit_count
, beginidx
, tsc
);
1652 lib_ring_buffer_write_commit_counter(config
, buf
, chan
,
1653 offsets
->begin
+ config
->cb
.subbuffer_header_size(),
1654 commit_count
, cc_hot
);
1658 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1660 * Calls subbuffer_set_data_size() to set the data size of the current
1661 * sub-buffer. We do not need to perform check_deliver nor commit here,
1662 * since this task will be done by the "commit" of the event for which
1663 * we are currently doing the space reservation.
1666 void lib_ring_buffer_switch_new_end(struct lib_ring_buffer
*buf
,
1667 struct channel
*chan
,
1668 struct switch_offsets
*offsets
,
1671 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1672 unsigned long endidx
, data_size
;
1674 endidx
= subbuf_index(offsets
->end
- 1, chan
);
1675 data_size
= subbuf_offset(offsets
->end
- 1, chan
) + 1;
1676 subbuffer_set_data_size(config
, &buf
->backend
, endidx
, data_size
);
1682 * !0 if execution must be aborted.
1685 int lib_ring_buffer_try_switch_slow(enum switch_mode mode
,
1686 struct lib_ring_buffer
*buf
,
1687 struct channel
*chan
,
1688 struct switch_offsets
*offsets
,
1691 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1692 unsigned long off
, reserve_commit_diff
;
1694 offsets
->begin
= v_read(config
, &buf
->offset
);
1695 offsets
->old
= offsets
->begin
;
1696 offsets
->switch_old_start
= 0;
1697 off
= subbuf_offset(offsets
->begin
, chan
);
1699 *tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1702 * Ensure we flush the header of an empty subbuffer when doing the
1703 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1704 * total data gathering duration even if there were no records saved
1705 * after the last buffer switch.
1706 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1707 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1708 * subbuffer header as appropriate.
1709 * The next record that reserves space will be responsible for
1710 * populating the following subbuffer header. We choose not to populate
1711 * the next subbuffer header here because we want to be able to use
1712 * SWITCH_ACTIVE for periodical buffer flush and CPU tick_nohz stop
1713 * buffer flush, which must guarantee that all the buffer content
1714 * (records and header timestamps) are visible to the reader. This is
1715 * required for quiescence guarantees for the fusion merge.
1717 if (mode
!= SWITCH_FLUSH
&& !off
)
1718 return -1; /* we do not have to switch : buffer is empty */
1720 if (unlikely(off
== 0)) {
1721 unsigned long sb_index
, commit_count
;
1724 * We are performing a SWITCH_FLUSH. At this stage, there are no
1725 * concurrent writes into the buffer.
1727 * The client does not save any header information. Don't
1728 * switch empty subbuffer on finalize, because it is invalid to
1729 * deliver a completely empty subbuffer.
1731 if (!config
->cb
.subbuffer_header_size())
1734 /* Test new buffer integrity */
1735 sb_index
= subbuf_index(offsets
->begin
, chan
);
1736 commit_count
= v_read(config
,
1737 &buf
->commit_cold
[sb_index
].cc_sb
);
1738 reserve_commit_diff
=
1739 (buf_trunc(offsets
->begin
, chan
)
1740 >> chan
->backend
.num_subbuf_order
)
1741 - (commit_count
& chan
->commit_count_mask
);
1742 if (likely(reserve_commit_diff
== 0)) {
1743 /* Next subbuffer not being written to. */
1744 if (unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1745 subbuf_trunc(offsets
->begin
, chan
)
1746 - subbuf_trunc((unsigned long)
1747 atomic_long_read(&buf
->consumed
), chan
)
1748 >= chan
->backend
.buf_size
)) {
1750 * We do not overwrite non consumed buffers
1751 * and we are full : don't switch.
1756 * Next subbuffer not being written to, and we
1757 * are either in overwrite mode or the buffer is
1758 * not full. It's safe to write in this new
1764 * Next subbuffer reserve offset does not match the
1765 * commit offset. Don't perform switch in
1766 * producer-consumer and overwrite mode. Caused by
1767 * either a writer OOPS or too many nested writes over a
1768 * reserve/commit pair.
1774 * Need to write the subbuffer start header on finalize.
1776 offsets
->switch_old_start
= 1;
1778 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1779 /* Note: old points to the next subbuf at offset 0 */
1780 offsets
->end
= offsets
->begin
;
1785 * Force a sub-buffer switch. This operation is completely reentrant : can be
1786 * called while tracing is active with absolutely no lock held.
1788 * Note, however, that as a v_cmpxchg is used for some atomic
1789 * operations, this function must be called from the CPU which owns the buffer
1790 * for a ACTIVE flush.
1792 void lib_ring_buffer_switch_slow(struct lib_ring_buffer
*buf
, enum switch_mode mode
)
1794 struct channel
*chan
= buf
->backend
.chan
;
1795 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1796 struct switch_offsets offsets
;
1797 unsigned long oldidx
;
1803 * Perform retryable operations.
1806 if (lib_ring_buffer_try_switch_slow(mode
, buf
, chan
, &offsets
,
1808 return; /* Switch not needed */
1809 } while (v_cmpxchg(config
, &buf
->offset
, offsets
.old
, offsets
.end
)
1813 * Atomically update last_tsc. This update races against concurrent
1814 * atomic updates, but the race will always cause supplementary full TSC
1815 * records, never the opposite (missing a full TSC record when it would
1818 save_last_tsc(config
, buf
, tsc
);
1821 * Push the reader if necessary
1823 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.old
);
1825 oldidx
= subbuf_index(offsets
.old
, chan
);
1826 lib_ring_buffer_clear_noref(config
, &buf
->backend
, oldidx
);
1829 * May need to populate header start on SWITCH_FLUSH.
1831 if (offsets
.switch_old_start
) {
1832 lib_ring_buffer_switch_old_start(buf
, chan
, &offsets
, tsc
);
1833 offsets
.old
+= config
->cb
.subbuffer_header_size();
1837 * Switch old subbuffer.
1839 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, tsc
);
1841 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow
);
1843 struct switch_param
{
1844 struct lib_ring_buffer
*buf
;
1845 enum switch_mode mode
;
1848 static void remote_switch(void *info
)
1850 struct switch_param
*param
= info
;
1851 struct lib_ring_buffer
*buf
= param
->buf
;
1853 lib_ring_buffer_switch_slow(buf
, param
->mode
);
1856 static void _lib_ring_buffer_switch_remote(struct lib_ring_buffer
*buf
,
1857 enum switch_mode mode
)
1859 struct channel
*chan
= buf
->backend
.chan
;
1860 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1862 struct switch_param param
;
1865 * With global synchronization we don't need to use the IPI scheme.
1867 if (config
->sync
== RING_BUFFER_SYNC_GLOBAL
) {
1868 lib_ring_buffer_switch_slow(buf
, mode
);
1873 * Disabling preemption ensures two things: first, that the
1874 * target cpu is not taken concurrently offline while we are within
1875 * smp_call_function_single(). Secondly, if it happens that the
1876 * CPU is not online, our own call to lib_ring_buffer_switch_slow()
1877 * needs to be protected from CPU hotplug handlers, which can
1878 * also perform a remote subbuffer switch.
1883 ret
= smp_call_function_single(buf
->backend
.cpu
,
1884 remote_switch
, ¶m
, 1);
1886 /* Remote CPU is offline, do it ourself. */
1887 lib_ring_buffer_switch_slow(buf
, mode
);
1892 /* Switch sub-buffer if current sub-buffer is non-empty. */
1893 void lib_ring_buffer_switch_remote(struct lib_ring_buffer
*buf
)
1895 _lib_ring_buffer_switch_remote(buf
, SWITCH_ACTIVE
);
1897 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote
);
1899 /* Switch sub-buffer even if current sub-buffer is empty. */
1900 void lib_ring_buffer_switch_remote_empty(struct lib_ring_buffer
*buf
)
1902 _lib_ring_buffer_switch_remote(buf
, SWITCH_FLUSH
);
1904 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote_empty
);
1909 * -ENOSPC if event size is too large for packet.
1910 * -ENOBUFS if there is currently not enough space in buffer for the event.
1911 * -EIO if data cannot be written into the buffer for any other reason.
1914 int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer
*buf
,
1915 struct channel
*chan
,
1916 struct switch_offsets
*offsets
,
1917 struct lib_ring_buffer_ctx
*ctx
,
1920 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
1921 unsigned long reserve_commit_diff
, offset_cmp
;
1924 offsets
->begin
= offset_cmp
= v_read(config
, &buf
->offset
);
1925 offsets
->old
= offsets
->begin
;
1926 offsets
->switch_new_start
= 0;
1927 offsets
->switch_new_end
= 0;
1928 offsets
->switch_old_end
= 0;
1929 offsets
->pre_header_padding
= 0;
1931 ctx
->tsc
= config
->cb
.ring_buffer_clock_read(chan
);
1932 if ((int64_t) ctx
->tsc
== -EIO
)
1935 if (last_tsc_overflow(config
, buf
, ctx
->tsc
))
1936 ctx
->rflags
|= RING_BUFFER_RFLAG_FULL_TSC
;
1938 if (unlikely(subbuf_offset(offsets
->begin
, ctx
->chan
) == 0)) {
1939 offsets
->switch_new_start
= 1; /* For offsets->begin */
1941 offsets
->size
= config
->cb
.record_header_size(config
, chan
,
1943 &offsets
->pre_header_padding
,
1946 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
1949 if (unlikely(subbuf_offset(offsets
->begin
, chan
) +
1950 offsets
->size
> chan
->backend
.subbuf_size
)) {
1951 offsets
->switch_old_end
= 1; /* For offsets->old */
1952 offsets
->switch_new_start
= 1; /* For offsets->begin */
1955 if (unlikely(offsets
->switch_new_start
)) {
1956 unsigned long sb_index
, commit_count
;
1959 * We are typically not filling the previous buffer completely.
1961 if (likely(offsets
->switch_old_end
))
1962 offsets
->begin
= subbuf_align(offsets
->begin
, chan
);
1963 offsets
->begin
= offsets
->begin
1964 + config
->cb
.subbuffer_header_size();
1965 /* Test new buffer integrity */
1966 sb_index
= subbuf_index(offsets
->begin
, chan
);
1968 * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
1969 * lib_ring_buffer_check_deliver() has the matching
1970 * memory barriers required around commit_cold cc_sb
1971 * updates to ensure reserve and commit counter updates
1972 * are not seen reordered when updated by another CPU.
1975 commit_count
= v_read(config
,
1976 &buf
->commit_cold
[sb_index
].cc_sb
);
1977 /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
1979 if (unlikely(offset_cmp
!= v_read(config
, &buf
->offset
))) {
1981 * The reserve counter have been concurrently updated
1982 * while we read the commit counter. This means the
1983 * commit counter we read might not match buf->offset
1984 * due to concurrent update. We therefore need to retry.
1988 reserve_commit_diff
=
1989 (buf_trunc(offsets
->begin
, chan
)
1990 >> chan
->backend
.num_subbuf_order
)
1991 - (commit_count
& chan
->commit_count_mask
);
1992 if (likely(reserve_commit_diff
== 0)) {
1993 /* Next subbuffer not being written to. */
1994 if (unlikely(config
->mode
!= RING_BUFFER_OVERWRITE
&&
1995 subbuf_trunc(offsets
->begin
, chan
)
1996 - subbuf_trunc((unsigned long)
1997 atomic_long_read(&buf
->consumed
), chan
)
1998 >= chan
->backend
.buf_size
)) {
2000 * We do not overwrite non consumed buffers
2001 * and we are full : record is lost.
2003 v_inc(config
, &buf
->records_lost_full
);
2007 * Next subbuffer not being written to, and we
2008 * are either in overwrite mode or the buffer is
2009 * not full. It's safe to write in this new
2015 * Next subbuffer reserve offset does not match the
2016 * commit offset, and this did not involve update to the
2017 * reserve counter. Drop record in producer-consumer and
2018 * overwrite mode. Caused by either a writer OOPS or
2019 * too many nested writes over a reserve/commit pair.
2021 v_inc(config
, &buf
->records_lost_wrap
);
2025 config
->cb
.record_header_size(config
, chan
,
2027 &offsets
->pre_header_padding
,
2030 lib_ring_buffer_align(offsets
->begin
+ offsets
->size
,
2033 if (unlikely(subbuf_offset(offsets
->begin
, chan
)
2034 + offsets
->size
> chan
->backend
.subbuf_size
)) {
2036 * Record too big for subbuffers, report error, don't
2037 * complete the sub-buffer switch.
2039 v_inc(config
, &buf
->records_lost_big
);
2043 * We just made a successful buffer switch and the
2044 * record fits in the new subbuffer. Let's write.
2049 * Record fits in the current buffer and we are not on a switch
2050 * boundary. It's safe to write.
2053 offsets
->end
= offsets
->begin
+ offsets
->size
;
2055 if (unlikely(subbuf_offset(offsets
->end
, chan
) == 0)) {
2057 * The offset_end will fall at the very beginning of the next
2060 offsets
->switch_new_end
= 1; /* For offsets->begin */
2065 static struct lib_ring_buffer
*get_current_buf(struct channel
*chan
, int cpu
)
2067 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
2069 if (config
->alloc
== RING_BUFFER_ALLOC_PER_CPU
)
2070 return per_cpu_ptr(chan
->backend
.buf
, cpu
);
2072 return chan
->backend
.buf
;
2075 void lib_ring_buffer_lost_event_too_big(struct channel
*chan
)
2077 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
2078 struct lib_ring_buffer
*buf
= get_current_buf(chan
, smp_processor_id());
2080 v_inc(config
, &buf
->records_lost_big
);
2082 EXPORT_SYMBOL_GPL(lib_ring_buffer_lost_event_too_big
);
2085 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
2086 * @ctx: ring buffer context.
2088 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
2089 * -EIO for other errors, else returns 0.
2090 * It will take care of sub-buffer switching.
2092 int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx
*ctx
,
2095 struct channel
*chan
= ctx
->chan
;
2096 const struct lib_ring_buffer_config
*config
= &chan
->backend
.config
;
2097 struct lib_ring_buffer
*buf
;
2098 struct switch_offsets offsets
;
2101 ctx
->buf
= buf
= get_current_buf(chan
, ctx
->cpu
);
2105 ret
= lib_ring_buffer_try_reserve_slow(buf
, chan
, &offsets
,
2109 } while (unlikely(v_cmpxchg(config
, &buf
->offset
, offsets
.old
,
2114 * Atomically update last_tsc. This update races against concurrent
2115 * atomic updates, but the race will always cause supplementary full TSC
2116 * records, never the opposite (missing a full TSC record when it would
2119 save_last_tsc(config
, buf
, ctx
->tsc
);
2122 * Push the reader if necessary
2124 lib_ring_buffer_reserve_push_reader(buf
, chan
, offsets
.end
- 1);
2127 * Clear noref flag for this subbuffer.
2129 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2130 subbuf_index(offsets
.end
- 1, chan
));
2133 * Switch old subbuffer if needed.
2135 if (unlikely(offsets
.switch_old_end
)) {
2136 lib_ring_buffer_clear_noref(config
, &buf
->backend
,
2137 subbuf_index(offsets
.old
- 1, chan
));
2138 lib_ring_buffer_switch_old_end(buf
, chan
, &offsets
, ctx
->tsc
);
2142 * Populate new subbuffer.
2144 if (unlikely(offsets
.switch_new_start
))
2145 lib_ring_buffer_switch_new_start(buf
, chan
, &offsets
, ctx
->tsc
);
2147 if (unlikely(offsets
.switch_new_end
))
2148 lib_ring_buffer_switch_new_end(buf
, chan
, &offsets
, ctx
->tsc
);
2150 ctx
->slot_size
= offsets
.size
;
2151 ctx
->pre_offset
= offsets
.begin
;
2152 ctx
->buf_offset
= offsets
.begin
+ offsets
.pre_header_padding
;
2155 EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow
);
2158 void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config
*config
,
2159 struct lib_ring_buffer
*buf
,
2160 unsigned long commit_count
,
2163 if (config
->oops
== RING_BUFFER_OOPS_CONSISTENCY
)
2164 v_set(config
, &buf
->commit_hot
[idx
].seq
, commit_count
);
2168 * The ring buffer can count events recorded and overwritten per buffer,
2169 * but it is disabled by default due to its performance overhead.
2171 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
2173 void deliver_count_events(const struct lib_ring_buffer_config
*config
,
2174 struct lib_ring_buffer
*buf
,
2177 v_add(config
, subbuffer_get_records_count(config
,
2178 &buf
->backend
, idx
),
2179 &buf
->records_count
);
2180 v_add(config
, subbuffer_count_records_overrun(config
,
2181 &buf
->backend
, idx
),
2182 &buf
->records_overrun
);
2184 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
2186 void deliver_count_events(const struct lib_ring_buffer_config
*config
,
2187 struct lib_ring_buffer
*buf
,
2191 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
2194 void lib_ring_buffer_check_deliver_slow(const struct lib_ring_buffer_config
*config
,
2195 struct lib_ring_buffer
*buf
,
2196 struct channel
*chan
,
2197 unsigned long offset
,
2198 unsigned long commit_count
,
2202 unsigned long old_commit_count
= commit_count
2203 - chan
->backend
.subbuf_size
;
2206 * If we succeeded at updating cc_sb below, we are the subbuffer
2207 * writer delivering the subbuffer. Deals with concurrent
2208 * updates of the "cc" value without adding a add_return atomic
2209 * operation to the fast path.
2211 * We are doing the delivery in two steps:
2212 * - First, we cmpxchg() cc_sb to the new value
2213 * old_commit_count + 1. This ensures that we are the only
2214 * subbuffer user successfully filling the subbuffer, but we
2215 * do _not_ set the cc_sb value to "commit_count" yet.
2216 * Therefore, other writers that would wrap around the ring
2217 * buffer and try to start writing to our subbuffer would
2218 * have to drop records, because it would appear as
2220 * We therefore have exclusive access to the subbuffer control
2221 * structures. This mutual exclusion with other writers is
2222 * crucially important to perform record overruns count in
2223 * flight recorder mode locklessly.
2224 * - When we are ready to release the subbuffer (either for
2225 * reading or for overrun by other writers), we simply set the
2226 * cc_sb value to "commit_count" and perform delivery.
2228 * The subbuffer size is least 2 bytes (minimum size: 1 page).
2229 * This guarantees that old_commit_count + 1 != commit_count.
2233 * Order prior updates to reserve count prior to the
2234 * commit_cold cc_sb update.
2237 if (likely(v_cmpxchg(config
, &buf
->commit_cold
[idx
].cc_sb
,
2238 old_commit_count
, old_commit_count
+ 1)
2239 == old_commit_count
)) {
2241 * Start of exclusive subbuffer access. We are
2242 * guaranteed to be the last writer in this subbuffer
2243 * and any other writer trying to access this subbuffer
2244 * in this state is required to drop records.
2246 deliver_count_events(config
, buf
, idx
);
2247 config
->cb
.buffer_end(buf
, tsc
, idx
,
2248 lib_ring_buffer_get_data_size(config
,
2253 * Increment the packet counter while we have exclusive
2256 subbuffer_inc_packet_count(config
, &buf
->backend
, idx
);
2259 * Set noref flag and offset for this subbuffer id.
2260 * Contains a memory barrier that ensures counter stores
2261 * are ordered before set noref and offset.
2263 lib_ring_buffer_set_noref_offset(config
, &buf
->backend
, idx
,
2264 buf_trunc_val(offset
, chan
));
2267 * Order set_noref and record counter updates before the
2268 * end of subbuffer exclusive access. Orders with
2269 * respect to writers coming into the subbuffer after
2270 * wrap around, and also order wrt concurrent readers.
2273 /* End of exclusive subbuffer access */
2274 v_set(config
, &buf
->commit_cold
[idx
].cc_sb
,
2277 * Order later updates to reserve count after
2278 * the commit_cold cc_sb update.
2281 lib_ring_buffer_vmcore_check_deliver(config
, buf
,
2285 * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free.
2287 if (config
->wakeup
== RING_BUFFER_WAKEUP_BY_WRITER
2288 && atomic_long_read(&buf
->active_readers
)
2289 && lib_ring_buffer_poll_deliver(config
, buf
, chan
)) {
2290 wake_up_interruptible(&buf
->read_wait
);
2291 wake_up_interruptible(&chan
->read_wait
);
2296 EXPORT_SYMBOL_GPL(lib_ring_buffer_check_deliver_slow
);
2298 int __init
init_lib_ring_buffer_frontend(void)
2302 for_each_possible_cpu(cpu
)
2303 spin_lock_init(&per_cpu(ring_buffer_nohz_lock
, cpu
));
2307 module_init(init_lib_ring_buffer_frontend
);
2309 void __exit
exit_lib_ring_buffer_frontend(void)
2313 module_exit(exit_lib_ring_buffer_frontend
);