2 * ring_buffer_frontend.c
4 * (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
6 * Ring buffer wait-free buffer synchronization. Producer-consumer and flight
7 * recorder (overwrite) modes. See thesis:
9 * Desnoyers, Mathieu (2009), "Low-Impact Operating System Tracing", Ph.D.
10 * dissertation, Ecole Polytechnique de Montreal.
11 * http://www.lttng.org/pub/thesis/desnoyers-dissertation-2009-12.pdf
13 * - Algorithm presentation in Chapter 5:
14 * "Lockless Multi-Core High-Throughput Buffering".
15 * - Algorithm formal verification in Section 8.6:
16 * "Formal verification of LTTng"
19 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
21 * Inspired from LTT and RelayFS:
22 * Karim Yaghmour <karim@opersys.com>
23 * Tom Zanussi <zanussi@us.ibm.com>
24 * Bob Wisniewski <bob@watson.ibm.com>
26 * Bob Wisniewski <bob@watson.ibm.com>
28 * Buffer reader semantic :
31 * while buffer is not finalized and empty
33 * - if return value != 0, continue
34 * - splice one subbuffer worth of data to a pipe
35 * - splice the data from pipe to disk/network
38 * Dual LGPL v2.1/GPL v2 license.
41 #include <linux/delay.h>
42 #include <linux/module.h>
43 #include <linux/percpu.h>
45 #include "../../wrapper/ringbuffer/config.h"
46 #include "../../wrapper/ringbuffer/backend.h"
47 #include "../../wrapper/ringbuffer/frontend.h"
48 #include "../../wrapper/ringbuffer/iterator.h"
49 #include "../../wrapper/ringbuffer/nohz.h"
52 * Internal structure representing offsets to use at a sub-buffer switch.
54 struct switch_offsets {
55 unsigned long begin, end, old;
56 size_t pre_header_padding, size;
57 unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1,
68 static ATOMIC_NOTIFIER_HEAD(tick_nohz_notifier);
69 #endif /* CONFIG_NO_HZ */
71 static DEFINE_PER_CPU(spinlock_t, ring_buffer_nohz_lock);
73 DEFINE_PER_CPU(unsigned int, lib_ring_buffer_nesting);
74 EXPORT_PER_CPU_SYMBOL(lib_ring_buffer_nesting);
77 void lib_ring_buffer_print_errors(struct channel *chan,
78 struct lib_ring_buffer *buf, int cpu);
81 * Must be called under cpu hotplug protection.
83 void lib_ring_buffer_free(struct lib_ring_buffer *buf)
85 struct channel *chan = buf->backend.chan;
87 lib_ring_buffer_print_errors(chan, buf, buf->backend.cpu);
88 kfree(buf->commit_hot);
89 kfree(buf->commit_cold);
91 lib_ring_buffer_backend_free(&buf->backend);
95 * lib_ring_buffer_reset - Reset ring buffer to initial values.
98 * Effectively empty the ring buffer. Should be called when the buffer is not
99 * used for writing. The ring buffer can be opened for reading, but the reader
100 * should not be using the iterator concurrently with reset. The previous
101 * current iterator record is reset.
103 void lib_ring_buffer_reset(struct lib_ring_buffer *buf)
105 struct channel *chan = buf->backend.chan;
106 const struct lib_ring_buffer_config *config = chan->backend.config;
110 * Reset iterator first. It will put the subbuffer if it currently holds
113 lib_ring_buffer_iterator_reset(buf);
114 v_set(config, &buf->offset, 0);
115 for (i = 0; i < chan->backend.num_subbuf; i++) {
116 v_set(config, &buf->commit_hot[i].cc, 0);
117 v_set(config, &buf->commit_hot[i].seq, 0);
118 v_set(config, &buf->commit_cold[i].cc_sb, 0);
120 atomic_long_set(&buf->consumed, 0);
121 atomic_set(&buf->record_disabled, 0);
122 v_set(config, &buf->last_tsc, 0);
123 lib_ring_buffer_backend_reset(&buf->backend);
124 /* Don't reset number of active readers */
125 v_set(config, &buf->records_lost_full, 0);
126 v_set(config, &buf->records_lost_wrap, 0);
127 v_set(config, &buf->records_lost_big, 0);
128 v_set(config, &buf->records_count, 0);
129 v_set(config, &buf->records_overrun, 0);
132 EXPORT_SYMBOL_GPL(lib_ring_buffer_reset);
135 * channel_reset - Reset channel to initial values.
138 * Effectively empty the channel. Should be called when the channel is not used
139 * for writing. The channel can be opened for reading, but the reader should not
140 * be using the iterator concurrently with reset. The previous current iterator
143 void channel_reset(struct channel *chan)
146 * Reset iterators first. Will put the subbuffer if held for reading.
148 channel_iterator_reset(chan);
149 atomic_set(&chan->record_disabled, 0);
150 /* Don't reset commit_count_mask, still valid */
151 channel_backend_reset(&chan->backend);
152 /* Don't reset switch/read timer interval */
153 /* Don't reset notifiers and notifier enable bits */
154 /* Don't reset reader reference count */
156 EXPORT_SYMBOL_GPL(channel_reset);
159 * Must be called under cpu hotplug protection.
161 int lib_ring_buffer_create(struct lib_ring_buffer *buf,
162 struct channel_backend *chanb, int cpu)
164 const struct lib_ring_buffer_config *config = chanb->config;
165 struct channel *chan = container_of(chanb, struct channel, backend);
166 void *priv = chanb->priv;
167 size_t subbuf_header_size;
171 /* Test for cpu hotplug */
172 if (buf->backend.allocated)
176 * Paranoia: per cpu dynamic allocation is not officially documented as
177 * zeroing the memory, so let's do it here too, just in case.
179 memset(buf, 0, sizeof(*buf));
181 ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, cpu);
186 kzalloc_node(ALIGN(sizeof(*buf->commit_hot)
187 * chan->backend.num_subbuf,
188 1 << INTERNODE_CACHE_SHIFT),
189 GFP_KERNEL, cpu_to_node(max(cpu, 0)));
190 if (!buf->commit_hot) {
196 kzalloc_node(ALIGN(sizeof(*buf->commit_cold)
197 * chan->backend.num_subbuf,
198 1 << INTERNODE_CACHE_SHIFT),
199 GFP_KERNEL, cpu_to_node(max(cpu, 0)));
200 if (!buf->commit_cold) {
205 init_waitqueue_head(&buf->read_wait);
206 init_waitqueue_head(&buf->write_wait);
207 raw_spin_lock_init(&buf->raw_tick_nohz_spinlock);
210 * Write the subbuffer header for first subbuffer so we know the total
211 * duration of data gathering.
213 subbuf_header_size = config->cb.subbuffer_header_size();
214 v_set(config, &buf->offset, subbuf_header_size);
215 subbuffer_id_clear_noref(config, &buf->backend.buf_wsb[0].id);
216 tsc = config->cb.ring_buffer_clock_read(buf->backend.chan);
217 config->cb.buffer_begin(buf, tsc, 0);
218 v_add(config, subbuf_header_size, &buf->commit_hot[0].cc);
220 if (config->cb.buffer_create) {
221 ret = config->cb.buffer_create(buf, priv, cpu, chanb->name);
227 * Ensure the buffer is ready before setting it to allocated and setting
229 * Used for cpu hotplug vs cpumask iteration.
232 buf->backend.allocated = 1;
234 if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
235 CHAN_WARN_ON(chan, cpumask_test_cpu(cpu,
236 chan->backend.cpumask));
237 cpumask_set_cpu(cpu, chan->backend.cpumask);
244 kfree(buf->commit_cold);
246 kfree(buf->commit_hot);
248 lib_ring_buffer_backend_free(&buf->backend);
252 static void switch_buffer_timer(unsigned long data)
254 struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data;
255 struct channel *chan = buf->backend.chan;
256 const struct lib_ring_buffer_config *config = chan->backend.config;
259 * Only flush buffers periodically if readers are active.
261 if (atomic_long_read(&buf->active_readers))
262 lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE);
264 if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
265 mod_timer_pinned(&buf->switch_timer,
266 jiffies + chan->switch_timer_interval);
268 mod_timer(&buf->switch_timer,
269 jiffies + chan->switch_timer_interval);
273 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
275 static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer *buf)
277 struct channel *chan = buf->backend.chan;
278 const struct lib_ring_buffer_config *config = chan->backend.config;
280 if (!chan->switch_timer_interval || buf->switch_timer_enabled)
282 init_timer(&buf->switch_timer);
283 buf->switch_timer.function = switch_buffer_timer;
284 buf->switch_timer.expires = jiffies + chan->switch_timer_interval;
285 buf->switch_timer.data = (unsigned long)buf;
286 if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
287 add_timer_on(&buf->switch_timer, buf->backend.cpu);
289 add_timer(&buf->switch_timer);
290 buf->switch_timer_enabled = 1;
294 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
296 static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer *buf)
298 struct channel *chan = buf->backend.chan;
300 if (!chan->switch_timer_interval || !buf->switch_timer_enabled)
303 del_timer_sync(&buf->switch_timer);
304 buf->switch_timer_enabled = 0;
308 * Polling timer to check the channels for data.
310 static void read_buffer_timer(unsigned long data)
312 struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data;
313 struct channel *chan = buf->backend.chan;
314 const struct lib_ring_buffer_config *config = chan->backend.config;
316 CHAN_WARN_ON(chan, !buf->backend.allocated);
318 if (atomic_long_read(&buf->active_readers)
319 && lib_ring_buffer_poll_deliver(config, buf, chan)) {
320 wake_up_interruptible(&buf->read_wait);
321 wake_up_interruptible(&chan->read_wait);
324 if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
325 mod_timer_pinned(&buf->read_timer,
326 jiffies + chan->read_timer_interval);
328 mod_timer(&buf->read_timer,
329 jiffies + chan->read_timer_interval);
333 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
335 static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer *buf)
337 struct channel *chan = buf->backend.chan;
338 const struct lib_ring_buffer_config *config = chan->backend.config;
340 if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER
341 || !chan->read_timer_interval
342 || buf->read_timer_enabled)
345 init_timer(&buf->read_timer);
346 buf->read_timer.function = read_buffer_timer;
347 buf->read_timer.expires = jiffies + chan->read_timer_interval;
348 buf->read_timer.data = (unsigned long)buf;
350 if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
351 add_timer_on(&buf->read_timer, buf->backend.cpu);
353 add_timer(&buf->read_timer);
354 buf->read_timer_enabled = 1;
358 * Called with ring_buffer_nohz_lock held for per-cpu buffers.
360 static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer *buf)
362 struct channel *chan = buf->backend.chan;
363 const struct lib_ring_buffer_config *config = chan->backend.config;
365 if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER
366 || !chan->read_timer_interval
367 || !buf->read_timer_enabled)
370 del_timer_sync(&buf->read_timer);
372 * do one more check to catch data that has been written in the last
375 if (lib_ring_buffer_poll_deliver(config, buf, chan)) {
376 wake_up_interruptible(&buf->read_wait);
377 wake_up_interruptible(&chan->read_wait);
379 buf->read_timer_enabled = 0;
382 #ifdef CONFIG_HOTPLUG_CPU
384 * lib_ring_buffer_cpu_hp_callback - CPU hotplug callback
385 * @nb: notifier block
386 * @action: hotplug action to take
389 * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
392 int __cpuinit lib_ring_buffer_cpu_hp_callback(struct notifier_block *nb,
393 unsigned long action,
396 unsigned int cpu = (unsigned long)hcpu;
397 struct channel *chan = container_of(nb, struct channel,
399 struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu);
400 const struct lib_ring_buffer_config *config = chan->backend.config;
402 if (!chan->cpu_hp_enable)
405 CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL);
408 case CPU_DOWN_FAILED:
409 case CPU_DOWN_FAILED_FROZEN:
411 case CPU_ONLINE_FROZEN:
412 wake_up_interruptible(&chan->hp_wait);
413 lib_ring_buffer_start_switch_timer(buf);
414 lib_ring_buffer_start_read_timer(buf);
417 case CPU_DOWN_PREPARE:
418 case CPU_DOWN_PREPARE_FROZEN:
419 lib_ring_buffer_stop_switch_timer(buf);
420 lib_ring_buffer_stop_read_timer(buf);
424 case CPU_DEAD_FROZEN:
426 * Performing a buffer switch on a remote CPU. Performed by
427 * the CPU responsible for doing the hotunplug after the target
428 * CPU stopped running completely. Ensures that all data
429 * from that remote CPU is flushed.
431 lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE);
440 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
442 * For per-cpu buffers, call the reader wakeups before switching the buffer, so
443 * that wake-up-tracing generated events are flushed before going idle (in
444 * tick_nohz). We test if the spinlock is locked to deal with the race where
445 * readers try to sample the ring buffer before we perform the switch. We let
446 * the readers retry in that case. If there is data in the buffer, the wake up
447 * is going to forbid the CPU running the reader thread from going idle.
449 static int notrace ring_buffer_tick_nohz_callback(struct notifier_block *nb,
453 struct channel *chan = container_of(nb, struct channel,
455 const struct lib_ring_buffer_config *config = chan->backend.config;
456 struct lib_ring_buffer *buf;
457 int cpu = smp_processor_id();
459 if (config->alloc != RING_BUFFER_ALLOC_PER_CPU) {
461 * We don't support keeping the system idle with global buffers
462 * and streaming active. In order to do so, we would need to
463 * sample a non-nohz-cpumask racelessly with the nohz updates
464 * without adding synchronization overhead to nohz. Leave this
465 * use-case out for now.
470 buf = channel_get_ring_buffer(config, chan, cpu);
472 case TICK_NOHZ_FLUSH:
473 raw_spin_lock(&buf->raw_tick_nohz_spinlock);
474 if (config->wakeup == RING_BUFFER_WAKEUP_BY_TIMER
475 && chan->read_timer_interval
476 && atomic_long_read(&buf->active_readers)
477 && (lib_ring_buffer_poll_deliver(config, buf, chan)
478 || lib_ring_buffer_pending_data(config, buf, chan))) {
479 wake_up_interruptible(&buf->read_wait);
480 wake_up_interruptible(&chan->read_wait);
482 if (chan->switch_timer_interval)
483 lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE);
484 raw_spin_unlock(&buf->raw_tick_nohz_spinlock);
487 spin_lock(&__get_cpu_var(ring_buffer_nohz_lock));
488 lib_ring_buffer_stop_switch_timer(buf);
489 lib_ring_buffer_stop_read_timer(buf);
490 spin_unlock(&__get_cpu_var(ring_buffer_nohz_lock));
492 case TICK_NOHZ_RESTART:
493 spin_lock(&__get_cpu_var(ring_buffer_nohz_lock));
494 lib_ring_buffer_start_read_timer(buf);
495 lib_ring_buffer_start_switch_timer(buf);
496 spin_unlock(&__get_cpu_var(ring_buffer_nohz_lock));
503 void notrace lib_ring_buffer_tick_nohz_flush(void)
505 atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_FLUSH,
509 void notrace lib_ring_buffer_tick_nohz_stop(void)
511 atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_STOP,
515 void notrace lib_ring_buffer_tick_nohz_restart(void)
517 atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_RESTART,
520 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
525 static void channel_unregister_notifiers(struct channel *chan)
527 const struct lib_ring_buffer_config *config = chan->backend.config;
530 channel_iterator_unregister_notifiers(chan);
531 if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
534 * Remove the nohz notifier first, so we are certain we stop
537 atomic_notifier_chain_unregister(&tick_nohz_notifier,
538 &chan->tick_nohz_notifier);
540 * ring_buffer_nohz_lock will not be needed below, because
541 * we just removed the notifiers, which were the only source of
544 #endif /* CONFIG_NO_HZ */
545 #ifdef CONFIG_HOTPLUG_CPU
547 chan->cpu_hp_enable = 0;
548 for_each_online_cpu(cpu) {
549 struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf,
551 lib_ring_buffer_stop_switch_timer(buf);
552 lib_ring_buffer_stop_read_timer(buf);
555 unregister_cpu_notifier(&chan->cpu_hp_notifier);
557 for_each_possible_cpu(cpu) {
558 struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf,
560 lib_ring_buffer_stop_switch_timer(buf);
561 lib_ring_buffer_stop_read_timer(buf);
565 struct lib_ring_buffer *buf = chan->backend.buf;
567 lib_ring_buffer_stop_switch_timer(buf);
568 lib_ring_buffer_stop_read_timer(buf);
570 channel_backend_unregister_notifiers(&chan->backend);
573 static void channel_free(struct channel *chan)
575 channel_iterator_free(chan);
576 channel_backend_free(&chan->backend);
581 * channel_create - Create channel.
582 * @config: ring buffer instance configuration
583 * @name: name of the channel
584 * @priv: ring buffer client private data
585 * @buf_addr: pointer the the beginning of the preallocated buffer contiguous
586 * address mapping. It is used only by RING_BUFFER_STATIC
587 * configuration. It can be set to NULL for other backends.
588 * @subbuf_size: subbuffer size
589 * @num_subbuf: number of subbuffers
590 * @switch_timer_interval: Time interval (in us) to fill sub-buffers with
591 * padding to let readers get those sub-buffers.
592 * Used for live streaming.
593 * @read_timer_interval: Time interval (in us) to wake up pending readers.
596 * Returns NULL on failure.
598 struct channel *channel_create(const struct lib_ring_buffer_config *config,
599 const char *name, void *priv, void *buf_addr,
601 size_t num_subbuf, unsigned int switch_timer_interval,
602 unsigned int read_timer_interval)
605 struct channel *chan;
607 if (lib_ring_buffer_check_config(config, switch_timer_interval,
608 read_timer_interval))
611 chan = kzalloc(sizeof(struct channel), GFP_KERNEL);
615 ret = channel_backend_init(&chan->backend, name, config, priv,
616 subbuf_size, num_subbuf);
620 ret = channel_iterator_init(chan);
622 goto error_free_backend;
624 chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order);
625 chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval);
626 chan->read_timer_interval = usecs_to_jiffies(read_timer_interval);
627 kref_init(&chan->ref);
628 init_waitqueue_head(&chan->read_wait);
629 init_waitqueue_head(&chan->hp_wait);
631 if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
632 #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER)
633 /* Only benefit from NO_HZ idle with per-cpu buffers for now. */
634 chan->tick_nohz_notifier.notifier_call =
635 ring_buffer_tick_nohz_callback;
636 chan->tick_nohz_notifier.priority = ~0U;
637 atomic_notifier_chain_register(&tick_nohz_notifier,
638 &chan->tick_nohz_notifier);
639 #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */
642 * In case of non-hotplug cpu, if the ring-buffer is allocated
643 * in early initcall, it will not be notified of secondary cpus.
644 * In that off case, we need to allocate for all possible cpus.
646 #ifdef CONFIG_HOTPLUG_CPU
647 chan->cpu_hp_notifier.notifier_call =
648 lib_ring_buffer_cpu_hp_callback;
649 chan->cpu_hp_notifier.priority = 6;
650 register_cpu_notifier(&chan->cpu_hp_notifier);
653 for_each_online_cpu(cpu) {
654 struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf,
656 spin_lock(&per_cpu(ring_buffer_nohz_lock, cpu));
657 lib_ring_buffer_start_switch_timer(buf);
658 lib_ring_buffer_start_read_timer(buf);
659 spin_unlock(&per_cpu(ring_buffer_nohz_lock, cpu));
661 chan->cpu_hp_enable = 1;
664 for_each_possible_cpu(cpu) {
665 struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf,
667 spin_lock(&per_cpu(ring_buffer_nohz_lock, cpu));
668 lib_ring_buffer_start_switch_timer(buf);
669 lib_ring_buffer_start_read_timer(buf);
670 spin_unlock(&per_cpu(ring_buffer_nohz_lock, cpu));
674 struct lib_ring_buffer *buf = chan->backend.buf;
676 lib_ring_buffer_start_switch_timer(buf);
677 lib_ring_buffer_start_read_timer(buf);
683 channel_backend_free(&chan->backend);
688 EXPORT_SYMBOL_GPL(channel_create);
691 void channel_release(struct kref *kref)
693 struct channel *chan = container_of(kref, struct channel, ref);
698 * channel_destroy - Finalize, wait for q.s. and destroy channel.
699 * @chan: channel to destroy
702 * Call "destroy" callback, finalize channels, and then decrement the
703 * channel reference count. Note that when readers have completed data
704 * consumption of finalized channels, get_subbuf() will return -ENODATA.
705 * They should release their handle at that point. Returns the private
708 void *channel_destroy(struct channel *chan)
711 const struct lib_ring_buffer_config *config = chan->backend.config;
714 channel_unregister_notifiers(chan);
716 if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
718 * No need to hold cpu hotplug, because all notifiers have been
721 for_each_channel_cpu(cpu, chan) {
722 struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf,
725 if (config->cb.buffer_finalize)
726 config->cb.buffer_finalize(buf,
729 if (buf->backend.allocated)
730 lib_ring_buffer_switch_slow(buf, SWITCH_FLUSH);
732 * Perform flush before writing to finalized.
735 ACCESS_ONCE(buf->finalized) = 1;
736 wake_up_interruptible(&buf->read_wait);
739 struct lib_ring_buffer *buf = chan->backend.buf;
741 if (config->cb.buffer_finalize)
742 config->cb.buffer_finalize(buf, chan->backend.priv, -1);
743 if (buf->backend.allocated)
744 lib_ring_buffer_switch_slow(buf, SWITCH_FLUSH);
746 * Perform flush before writing to finalized.
749 ACCESS_ONCE(buf->finalized) = 1;
750 wake_up_interruptible(&buf->read_wait);
752 ACCESS_ONCE(chan->finalized) = 1;
753 wake_up_interruptible(&chan->hp_wait);
754 wake_up_interruptible(&chan->read_wait);
755 priv = chan->backend.priv;
756 kref_put(&chan->ref, channel_release);
759 EXPORT_SYMBOL_GPL(channel_destroy);
761 struct lib_ring_buffer *channel_get_ring_buffer(
762 const struct lib_ring_buffer_config *config,
763 struct channel *chan, int cpu)
765 if (config->alloc == RING_BUFFER_ALLOC_GLOBAL)
766 return chan->backend.buf;
768 return per_cpu_ptr(chan->backend.buf, cpu);
770 EXPORT_SYMBOL_GPL(channel_get_ring_buffer);
772 int lib_ring_buffer_open_read(struct lib_ring_buffer *buf)
774 struct channel *chan = buf->backend.chan;
776 if (!atomic_long_add_unless(&buf->active_readers, 1, 1))
778 kref_get(&chan->ref);
779 smp_mb__after_atomic_inc();
782 EXPORT_SYMBOL_GPL(lib_ring_buffer_open_read);
784 void lib_ring_buffer_release_read(struct lib_ring_buffer *buf)
786 struct channel *chan = buf->backend.chan;
788 CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1);
789 smp_mb__before_atomic_dec();
790 atomic_long_dec(&buf->active_readers);
791 kref_put(&chan->ref, channel_release);
793 EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read);
796 * Promote compiler barrier to a smp_mb().
797 * For the specific ring buffer case, this IPI call should be removed if the
798 * architecture does not reorder writes. This should eventually be provided by
799 * a separate architecture-specific infrastructure.
801 static void remote_mb(void *info)
807 * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read)
809 * @consumed: consumed count indicating the position where to read
810 * @produced: produced count, indicates position when to stop reading
812 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
813 * data to read at consumed position, or 0 if the get operation succeeds.
814 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
817 int lib_ring_buffer_snapshot(struct lib_ring_buffer *buf,
818 unsigned long *consumed, unsigned long *produced)
820 struct channel *chan = buf->backend.chan;
821 const struct lib_ring_buffer_config *config = chan->backend.config;
822 unsigned long consumed_cur, write_offset;
826 finalized = ACCESS_ONCE(buf->finalized);
828 * Read finalized before counters.
831 consumed_cur = atomic_long_read(&buf->consumed);
833 * No need to issue a memory barrier between consumed count read and
834 * write offset read, because consumed count can only change
835 * concurrently in overwrite mode, and we keep a sequence counter
836 * identifier derived from the write offset to check we are getting
837 * the same sub-buffer we are expecting (the sub-buffers are atomically
838 * "tagged" upon writes, tags are checked upon read).
840 write_offset = v_read(config, &buf->offset);
843 * Check that we are not about to read the same subbuffer in
844 * which the writer head is.
846 if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan)
850 *consumed = consumed_cur;
851 *produced = subbuf_trunc(write_offset, chan);
857 * The memory barriers __wait_event()/wake_up_interruptible() take care
858 * of "raw_spin_is_locked" memory ordering.
862 else if (raw_spin_is_locked(&buf->raw_tick_nohz_spinlock))
867 EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot);
870 * lib_ring_buffer_put_snapshot - move consumed counter forward
872 * Should only be called from consumer context.
874 * @consumed_new: new consumed count value
876 void lib_ring_buffer_move_consumer(struct lib_ring_buffer *buf,
877 unsigned long consumed_new)
879 struct lib_ring_buffer_backend *bufb = &buf->backend;
880 struct channel *chan = bufb->chan;
881 unsigned long consumed;
883 CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1);
886 * Only push the consumed value forward.
887 * If the consumed cmpxchg fails, this is because we have been pushed by
888 * the writer in flight recorder mode.
890 consumed = atomic_long_read(&buf->consumed);
891 while ((long) consumed - (long) consumed_new < 0)
892 consumed = atomic_long_cmpxchg(&buf->consumed, consumed,
894 /* Wake-up the metadata producer */
895 wake_up_interruptible(&buf->write_wait);
897 EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer);
900 * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading
902 * @consumed: consumed count indicating the position where to read
904 * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no
905 * data to read at consumed position, or 0 if the get operation succeeds.
906 * Busy-loop trying to get data if the tick_nohz sequence lock is held.
908 int lib_ring_buffer_get_subbuf(struct lib_ring_buffer *buf,
909 unsigned long consumed)
911 struct channel *chan = buf->backend.chan;
912 const struct lib_ring_buffer_config *config = chan->backend.config;
913 unsigned long consumed_cur, consumed_idx, commit_count, write_offset;
918 finalized = ACCESS_ONCE(buf->finalized);
920 * Read finalized before counters.
923 consumed_cur = atomic_long_read(&buf->consumed);
924 consumed_idx = subbuf_index(consumed, chan);
925 commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb);
927 * Make sure we read the commit count before reading the buffer
928 * data and the write offset. Correct consumed offset ordering
929 * wrt commit count is insured by the use of cmpxchg to update
930 * the consumed offset.
931 * smp_call_function_single can fail if the remote CPU is offline,
932 * this is OK because then there is no wmb to execute there.
933 * If our thread is executing on the same CPU as the on the buffers
934 * belongs to, we don't have to synchronize it at all. If we are
935 * migrated, the scheduler will take care of the memory barriers.
936 * Normally, smp_call_function_single() should ensure program order when
937 * executing the remote function, which implies that it surrounds the
938 * function execution with :
949 * However, smp_call_function_single() does not seem to clearly execute
950 * such barriers. It depends on spinlock semantic to provide the barrier
951 * before executing the IPI and, when busy-looping, csd_lock_wait only
952 * executes smp_mb() when it has to wait for the other CPU.
954 * I don't trust this code. Therefore, let's add the smp_mb() sequence
955 * required ourself, even if duplicated. It has no performance impact
958 * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs
959 * read and write vs write. They do not ensure core synchronization. We
960 * really have to ensure total order between the 3 barriers running on
963 if (config->ipi == RING_BUFFER_IPI_BARRIER) {
964 if (config->sync == RING_BUFFER_SYNC_PER_CPU
965 && config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
966 if (raw_smp_processor_id() != buf->backend.cpu) {
967 /* Total order with IPI handler smp_mb() */
969 smp_call_function_single(buf->backend.cpu,
971 /* Total order with IPI handler smp_mb() */
975 /* Total order with IPI handler smp_mb() */
977 smp_call_function(remote_mb, NULL, 1);
978 /* Total order with IPI handler smp_mb() */
983 * Local rmb to match the remote wmb to read the commit count
984 * before the buffer data and the write offset.
989 write_offset = v_read(config, &buf->offset);
992 * Check that the buffer we are getting is after or at consumed_cur
995 if ((long) subbuf_trunc(consumed, chan)
996 - (long) subbuf_trunc(consumed_cur, chan) < 0)
1000 * Check that the subbuffer we are trying to consume has been
1001 * already fully committed.
1003 if (((commit_count - chan->backend.subbuf_size)
1004 & chan->commit_count_mask)
1005 - (buf_trunc(consumed_cur, chan)
1006 >> chan->backend.num_subbuf_order)
1011 * Check that we are not about to read the same subbuffer in
1012 * which the writer head is.
1014 if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan)
1019 * Failure to get the subbuffer causes a busy-loop retry without going
1020 * to a wait queue. These are caused by short-lived race windows where
1021 * the writer is getting access to a subbuffer we were trying to get
1022 * access to. Also checks that the "consumed" buffer count we are
1023 * looking for matches the one contained in the subbuffer id.
1025 ret = update_read_sb_index(config, &buf->backend, &chan->backend,
1026 consumed_idx, buf_trunc_val(consumed, chan));
1029 subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id);
1031 buf->get_subbuf_consumed = consumed;
1032 buf->get_subbuf = 1;
1038 * The memory barriers __wait_event()/wake_up_interruptible() take care
1039 * of "raw_spin_is_locked" memory ordering.
1043 else if (raw_spin_is_locked(&buf->raw_tick_nohz_spinlock))
1048 EXPORT_SYMBOL_GPL(lib_ring_buffer_get_subbuf);
1051 * lib_ring_buffer_put_subbuf - release exclusive subbuffer access
1054 void lib_ring_buffer_put_subbuf(struct lib_ring_buffer *buf)
1056 struct lib_ring_buffer_backend *bufb = &buf->backend;
1057 struct channel *chan = bufb->chan;
1058 const struct lib_ring_buffer_config *config = chan->backend.config;
1059 unsigned long read_sb_bindex, consumed_idx, consumed;
1061 CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1);
1063 if (!buf->get_subbuf) {
1065 * Reader puts a subbuffer it did not get.
1067 CHAN_WARN_ON(chan, 1);
1070 consumed = buf->get_subbuf_consumed;
1071 buf->get_subbuf = 0;
1074 * Clear the records_unread counter. (overruns counter)
1075 * Can still be non-zero if a file reader simply grabbed the data
1076 * without using iterators.
1077 * Can be below zero if an iterator is used on a snapshot more than
1080 read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id);
1081 v_add(config, v_read(config,
1082 &bufb->array[read_sb_bindex]->records_unread),
1083 &bufb->records_read);
1084 v_set(config, &bufb->array[read_sb_bindex]->records_unread, 0);
1085 CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE
1086 && subbuffer_id_is_noref(config, bufb->buf_rsb.id));
1087 subbuffer_id_set_noref(config, &bufb->buf_rsb.id);
1090 * Exchange the reader subbuffer with the one we put in its place in the
1091 * writer subbuffer table. Expect the original consumed count. If
1092 * update_read_sb_index fails, this is because the writer updated the
1093 * subbuffer concurrently. We should therefore keep the subbuffer we
1094 * currently have: it has become invalid to try reading this sub-buffer
1095 * consumed count value anyway.
1097 consumed_idx = subbuf_index(consumed, chan);
1098 update_read_sb_index(config, &buf->backend, &chan->backend,
1099 consumed_idx, buf_trunc_val(consumed, chan));
1101 * update_read_sb_index return value ignored. Don't exchange sub-buffer
1102 * if the writer concurrently updated it.
1105 EXPORT_SYMBOL_GPL(lib_ring_buffer_put_subbuf);
1108 * cons_offset is an iterator on all subbuffer offsets between the reader
1109 * position and the writer position. (inclusive)
1112 void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer *buf,
1113 struct channel *chan,
1114 unsigned long cons_offset,
1117 const struct lib_ring_buffer_config *config = chan->backend.config;
1118 unsigned long cons_idx, commit_count, commit_count_sb;
1120 cons_idx = subbuf_index(cons_offset, chan);
1121 commit_count = v_read(config, &buf->commit_hot[cons_idx].cc);
1122 commit_count_sb = v_read(config, &buf->commit_cold[cons_idx].cc_sb);
1124 if (subbuf_offset(commit_count, chan) != 0)
1126 "ring buffer %s, cpu %d: "
1127 "commit count in subbuffer %lu,\n"
1128 "expecting multiples of %lu bytes\n"
1129 " [ %lu bytes committed, %lu bytes reader-visible ]\n",
1130 chan->backend.name, cpu, cons_idx,
1131 chan->backend.subbuf_size,
1132 commit_count, commit_count_sb);
1134 printk(KERN_DEBUG "ring buffer: %s, cpu %d: %lu bytes committed\n",
1135 chan->backend.name, cpu, commit_count);
1139 void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer *buf,
1140 struct channel *chan,
1141 void *priv, int cpu)
1143 const struct lib_ring_buffer_config *config = chan->backend.config;
1144 unsigned long write_offset, cons_offset;
1147 * Can be called in the error path of allocation when
1148 * trans_channel_data is not yet set.
1153 * No need to order commit_count, write_offset and cons_offset reads
1154 * because we execute at teardown when no more writer nor reader
1155 * references are left.
1157 write_offset = v_read(config, &buf->offset);
1158 cons_offset = atomic_long_read(&buf->consumed);
1159 if (write_offset != cons_offset)
1161 "ring buffer %s, cpu %d: "
1162 "non-consumed data\n"
1163 " [ %lu bytes written, %lu bytes read ]\n",
1164 chan->backend.name, cpu, write_offset, cons_offset);
1166 for (cons_offset = atomic_long_read(&buf->consumed);
1167 (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset),
1170 cons_offset = subbuf_align(cons_offset, chan))
1171 lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset,
1176 void lib_ring_buffer_print_errors(struct channel *chan,
1177 struct lib_ring_buffer *buf, int cpu)
1179 const struct lib_ring_buffer_config *config = chan->backend.config;
1180 void *priv = chan->backend.priv;
1182 printk(KERN_DEBUG "ring buffer %s, cpu %d: %lu records written, "
1183 "%lu records overrun\n",
1184 chan->backend.name, cpu,
1185 v_read(config, &buf->records_count),
1186 v_read(config, &buf->records_overrun));
1188 if (v_read(config, &buf->records_lost_full)
1189 || v_read(config, &buf->records_lost_wrap)
1190 || v_read(config, &buf->records_lost_big))
1192 "ring buffer %s, cpu %d: records were lost. Caused by:\n"
1193 " [ %lu buffer full, %lu nest buffer wrap-around, "
1194 "%lu event too big ]\n",
1195 chan->backend.name, cpu,
1196 v_read(config, &buf->records_lost_full),
1197 v_read(config, &buf->records_lost_wrap),
1198 v_read(config, &buf->records_lost_big));
1200 lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu);
1204 * lib_ring_buffer_switch_old_start: Populate old subbuffer header.
1206 * Only executed when the buffer is finalized, in SWITCH_FLUSH.
1209 void lib_ring_buffer_switch_old_start(struct lib_ring_buffer *buf,
1210 struct channel *chan,
1211 struct switch_offsets *offsets,
1214 const struct lib_ring_buffer_config *config = chan->backend.config;
1215 unsigned long oldidx = subbuf_index(offsets->old, chan);
1216 unsigned long commit_count;
1218 config->cb.buffer_begin(buf, tsc, oldidx);
1221 * Order all writes to buffer before the commit count update that will
1222 * determine that the subbuffer is full.
1224 if (config->ipi == RING_BUFFER_IPI_BARRIER) {
1226 * Must write slot data before incrementing commit count. This
1227 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1233 v_add(config, config->cb.subbuffer_header_size(),
1234 &buf->commit_hot[oldidx].cc);
1235 commit_count = v_read(config, &buf->commit_hot[oldidx].cc);
1236 /* Check if the written buffer has to be delivered */
1237 lib_ring_buffer_check_deliver(config, buf, chan, offsets->old,
1238 commit_count, oldidx);
1239 lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx,
1240 offsets->old, commit_count,
1241 config->cb.subbuffer_header_size());
1245 * lib_ring_buffer_switch_old_end: switch old subbuffer
1247 * Note : offset_old should never be 0 here. It is ok, because we never perform
1248 * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller
1249 * increments the offset_old value when doing a SWITCH_FLUSH on an empty
1253 void lib_ring_buffer_switch_old_end(struct lib_ring_buffer *buf,
1254 struct channel *chan,
1255 struct switch_offsets *offsets,
1258 const struct lib_ring_buffer_config *config = chan->backend.config;
1259 unsigned long oldidx = subbuf_index(offsets->old - 1, chan);
1260 unsigned long commit_count, padding_size, data_size;
1262 data_size = subbuf_offset(offsets->old - 1, chan) + 1;
1263 padding_size = chan->backend.subbuf_size - data_size;
1264 subbuffer_set_data_size(config, &buf->backend, oldidx, data_size);
1267 * Order all writes to buffer before the commit count update that will
1268 * determine that the subbuffer is full.
1270 if (config->ipi == RING_BUFFER_IPI_BARRIER) {
1272 * Must write slot data before incrementing commit count. This
1273 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1279 v_add(config, padding_size, &buf->commit_hot[oldidx].cc);
1280 commit_count = v_read(config, &buf->commit_hot[oldidx].cc);
1281 lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1,
1282 commit_count, oldidx);
1283 lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx,
1284 offsets->old, commit_count,
1289 * lib_ring_buffer_switch_new_start: Populate new subbuffer.
1291 * This code can be executed unordered : writers may already have written to the
1292 * sub-buffer before this code gets executed, caution. The commit makes sure
1293 * that this code is executed before the deliver of this sub-buffer.
1296 void lib_ring_buffer_switch_new_start(struct lib_ring_buffer *buf,
1297 struct channel *chan,
1298 struct switch_offsets *offsets,
1301 const struct lib_ring_buffer_config *config = chan->backend.config;
1302 unsigned long beginidx = subbuf_index(offsets->begin, chan);
1303 unsigned long commit_count;
1305 config->cb.buffer_begin(buf, tsc, beginidx);
1308 * Order all writes to buffer before the commit count update that will
1309 * determine that the subbuffer is full.
1311 if (config->ipi == RING_BUFFER_IPI_BARRIER) {
1313 * Must write slot data before incrementing commit count. This
1314 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1320 v_add(config, config->cb.subbuffer_header_size(),
1321 &buf->commit_hot[beginidx].cc);
1322 commit_count = v_read(config, &buf->commit_hot[beginidx].cc);
1323 /* Check if the written buffer has to be delivered */
1324 lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin,
1325 commit_count, beginidx);
1326 lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx,
1327 offsets->begin, commit_count,
1328 config->cb.subbuffer_header_size());
1332 * lib_ring_buffer_switch_new_end: finish switching current subbuffer
1334 * The only remaining threads could be the ones with pending commits. They will
1335 * have to do the deliver themselves.
1338 void lib_ring_buffer_switch_new_end(struct lib_ring_buffer *buf,
1339 struct channel *chan,
1340 struct switch_offsets *offsets,
1343 const struct lib_ring_buffer_config *config = chan->backend.config;
1344 unsigned long endidx = subbuf_index(offsets->end - 1, chan);
1345 unsigned long commit_count, padding_size, data_size;
1347 data_size = subbuf_offset(offsets->end - 1, chan) + 1;
1348 padding_size = chan->backend.subbuf_size - data_size;
1349 subbuffer_set_data_size(config, &buf->backend, endidx, data_size);
1352 * Order all writes to buffer before the commit count update that will
1353 * determine that the subbuffer is full.
1355 if (config->ipi == RING_BUFFER_IPI_BARRIER) {
1357 * Must write slot data before incrementing commit count. This
1358 * compiler barrier is upgraded into a smp_mb() by the IPI sent
1364 v_add(config, padding_size, &buf->commit_hot[endidx].cc);
1365 commit_count = v_read(config, &buf->commit_hot[endidx].cc);
1366 lib_ring_buffer_check_deliver(config, buf, chan, offsets->end - 1,
1367 commit_count, endidx);
1368 lib_ring_buffer_write_commit_counter(config, buf, chan, endidx,
1369 offsets->end, commit_count,
1376 * !0 if execution must be aborted.
1379 int lib_ring_buffer_try_switch_slow(enum switch_mode mode,
1380 struct lib_ring_buffer *buf,
1381 struct channel *chan,
1382 struct switch_offsets *offsets,
1385 const struct lib_ring_buffer_config *config = chan->backend.config;
1388 offsets->begin = v_read(config, &buf->offset);
1389 offsets->old = offsets->begin;
1390 offsets->switch_old_start = 0;
1391 off = subbuf_offset(offsets->begin, chan);
1393 *tsc = config->cb.ring_buffer_clock_read(chan);
1396 * Ensure we flush the header of an empty subbuffer when doing the
1397 * finalize (SWITCH_FLUSH). This ensures that we end up knowing the
1398 * total data gathering duration even if there were no records saved
1399 * after the last buffer switch.
1400 * In SWITCH_ACTIVE mode, switch the buffer when it contains events.
1401 * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of
1402 * subbuffer header as appropriate.
1403 * The next record that reserves space will be responsible for
1404 * populating the following subbuffer header. We choose not to populate
1405 * the next subbuffer header here because we want to be able to use
1406 * SWITCH_ACTIVE for periodical buffer flush and CPU tick_nohz stop
1407 * buffer flush, which must guarantee that all the buffer content
1408 * (records and header timestamps) are visible to the reader. This is
1409 * required for quiescence guarantees for the fusion merge.
1411 if (mode == SWITCH_FLUSH || off > 0) {
1412 if (unlikely(off == 0)) {
1414 * The client does not save any header information.
1415 * Don't switch empty subbuffer on finalize, because it
1416 * is invalid to deliver a completely empty subbuffer.
1418 if (!config->cb.subbuffer_header_size())
1421 * Need to write the subbuffer start header on finalize.
1423 offsets->switch_old_start = 1;
1425 offsets->begin = subbuf_align(offsets->begin, chan);
1427 return -1; /* we do not have to switch : buffer is empty */
1428 /* Note: old points to the next subbuf at offset 0 */
1429 offsets->end = offsets->begin;
1434 * Force a sub-buffer switch. This operation is completely reentrant : can be
1435 * called while tracing is active with absolutely no lock held.
1437 * Note, however, that as a v_cmpxchg is used for some atomic
1438 * operations, this function must be called from the CPU which owns the buffer
1439 * for a ACTIVE flush.
1441 void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, enum switch_mode mode)
1443 struct channel *chan = buf->backend.chan;
1444 const struct lib_ring_buffer_config *config = chan->backend.config;
1445 struct switch_offsets offsets;
1446 unsigned long oldidx;
1452 * Perform retryable operations.
1455 if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets,
1457 return; /* Switch not needed */
1458 } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end)
1462 * Atomically update last_tsc. This update races against concurrent
1463 * atomic updates, but the race will always cause supplementary full TSC
1464 * records, never the opposite (missing a full TSC record when it would
1467 save_last_tsc(config, buf, tsc);
1470 * Push the reader if necessary
1472 lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old);
1474 oldidx = subbuf_index(offsets.old, chan);
1475 lib_ring_buffer_clear_noref(config, &buf->backend, oldidx);
1478 * May need to populate header start on SWITCH_FLUSH.
1480 if (offsets.switch_old_start) {
1481 lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc);
1482 offsets.old += config->cb.subbuffer_header_size();
1486 * Switch old subbuffer.
1488 lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc);
1490 EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow);
1495 * -ENOSPC if event size is too large for packet.
1496 * -ENOBUFS if there is currently not enough space in buffer for the event.
1497 * -EIO if data cannot be written into the buffer for any other reason.
1500 int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf,
1501 struct channel *chan,
1502 struct switch_offsets *offsets,
1503 struct lib_ring_buffer_ctx *ctx)
1505 const struct lib_ring_buffer_config *config = chan->backend.config;
1506 unsigned long reserve_commit_diff;
1508 offsets->begin = v_read(config, &buf->offset);
1509 offsets->old = offsets->begin;
1510 offsets->switch_new_start = 0;
1511 offsets->switch_new_end = 0;
1512 offsets->switch_old_end = 0;
1513 offsets->pre_header_padding = 0;
1515 ctx->tsc = config->cb.ring_buffer_clock_read(chan);
1516 if ((int64_t) ctx->tsc == -EIO)
1519 if (last_tsc_overflow(config, buf, ctx->tsc))
1520 ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC;
1522 if (unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) {
1523 offsets->switch_new_start = 1; /* For offsets->begin */
1525 offsets->size = config->cb.record_header_size(config, chan,
1527 &offsets->pre_header_padding,
1530 lib_ring_buffer_align(offsets->begin + offsets->size,
1533 if (unlikely(subbuf_offset(offsets->begin, chan) +
1534 offsets->size > chan->backend.subbuf_size)) {
1535 offsets->switch_old_end = 1; /* For offsets->old */
1536 offsets->switch_new_start = 1; /* For offsets->begin */
1539 if (unlikely(offsets->switch_new_start)) {
1540 unsigned long sb_index;
1543 * We are typically not filling the previous buffer completely.
1545 if (likely(offsets->switch_old_end))
1546 offsets->begin = subbuf_align(offsets->begin, chan);
1547 offsets->begin = offsets->begin
1548 + config->cb.subbuffer_header_size();
1549 /* Test new buffer integrity */
1550 sb_index = subbuf_index(offsets->begin, chan);
1551 reserve_commit_diff =
1552 (buf_trunc(offsets->begin, chan)
1553 >> chan->backend.num_subbuf_order)
1554 - ((unsigned long) v_read(config,
1555 &buf->commit_cold[sb_index].cc_sb)
1556 & chan->commit_count_mask);
1557 if (likely(reserve_commit_diff == 0)) {
1558 /* Next subbuffer not being written to. */
1559 if (unlikely(config->mode != RING_BUFFER_OVERWRITE &&
1560 subbuf_trunc(offsets->begin, chan)
1561 - subbuf_trunc((unsigned long)
1562 atomic_long_read(&buf->consumed), chan)
1563 >= chan->backend.buf_size)) {
1565 * We do not overwrite non consumed buffers
1566 * and we are full : record is lost.
1568 v_inc(config, &buf->records_lost_full);
1572 * Next subbuffer not being written to, and we
1573 * are either in overwrite mode or the buffer is
1574 * not full. It's safe to write in this new
1580 * Next subbuffer reserve offset does not match the
1581 * commit offset. Drop record in producer-consumer and
1582 * overwrite mode. Caused by either a writer OOPS or too
1583 * many nested writes over a reserve/commit pair.
1585 v_inc(config, &buf->records_lost_wrap);
1589 config->cb.record_header_size(config, chan,
1591 &offsets->pre_header_padding,
1594 lib_ring_buffer_align(offsets->begin + offsets->size,
1597 if (unlikely(subbuf_offset(offsets->begin, chan)
1598 + offsets->size > chan->backend.subbuf_size)) {
1600 * Record too big for subbuffers, report error, don't
1601 * complete the sub-buffer switch.
1603 v_inc(config, &buf->records_lost_big);
1607 * We just made a successful buffer switch and the
1608 * record fits in the new subbuffer. Let's write.
1613 * Record fits in the current buffer and we are not on a switch
1614 * boundary. It's safe to write.
1617 offsets->end = offsets->begin + offsets->size;
1619 if (unlikely(subbuf_offset(offsets->end, chan) == 0)) {
1621 * The offset_end will fall at the very beginning of the next
1624 offsets->switch_new_end = 1; /* For offsets->begin */
1630 * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer.
1631 * @ctx: ring buffer context.
1633 * Return : -NOBUFS if not enough space, -ENOSPC if event size too large,
1634 * -EIO for other errors, else returns 0.
1635 * It will take care of sub-buffer switching.
1637 int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx)
1639 struct channel *chan = ctx->chan;
1640 const struct lib_ring_buffer_config *config = chan->backend.config;
1641 struct lib_ring_buffer *buf;
1642 struct switch_offsets offsets;
1645 if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
1646 buf = per_cpu_ptr(chan->backend.buf, ctx->cpu);
1648 buf = chan->backend.buf;
1654 ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets,
1658 } while (unlikely(v_cmpxchg(config, &buf->offset, offsets.old,
1663 * Atomically update last_tsc. This update races against concurrent
1664 * atomic updates, but the race will always cause supplementary full TSC
1665 * records, never the opposite (missing a full TSC record when it would
1668 save_last_tsc(config, buf, ctx->tsc);
1671 * Push the reader if necessary
1673 lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1);
1676 * Clear noref flag for this subbuffer.
1678 lib_ring_buffer_clear_noref(config, &buf->backend,
1679 subbuf_index(offsets.end - 1, chan));
1682 * Switch old subbuffer if needed.
1684 if (unlikely(offsets.switch_old_end)) {
1685 lib_ring_buffer_clear_noref(config, &buf->backend,
1686 subbuf_index(offsets.old - 1, chan));
1687 lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc);
1691 * Populate new subbuffer.
1693 if (unlikely(offsets.switch_new_start))
1694 lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc);
1696 if (unlikely(offsets.switch_new_end))
1697 lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc);
1699 ctx->slot_size = offsets.size;
1700 ctx->pre_offset = offsets.begin;
1701 ctx->buf_offset = offsets.begin + offsets.pre_header_padding;
1704 EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow);
1706 int __init init_lib_ring_buffer_frontend(void)
1710 for_each_possible_cpu(cpu)
1711 spin_lock_init(&per_cpu(ring_buffer_nohz_lock, cpu));
1715 module_init(init_lib_ring_buffer_frontend);
1717 void __exit exit_lib_ring_buffer_frontend(void)
1721 module_exit(exit_lib_ring_buffer_frontend);