1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
7 #include <linux/trace_events.h>
8 #include <linux/ring_buffer.h>
9 #include <linux/trace_clock.h>
10 #include <linux/sched/clock.h>
11 #include <linux/trace_seq.h>
12 #include <linux/spinlock.h>
13 #include <linux/irq_work.h>
14 #include <linux/uaccess.h>
15 #include <linux/hardirq.h>
16 #include <linux/kthread.h> /* for self test */
17 #include <linux/module.h>
18 #include <linux/percpu.h>
19 #include <linux/mutex.h>
20 #include <linux/delay.h>
21 #include <linux/slab.h>
22 #include <linux/init.h>
23 #include <linux/hash.h>
24 #include <linux/list.h>
25 #include <linux/cpu.h>
26 #include <linux/oom.h>
28 #include <asm/local.h>
30 static void update_pages_handler(struct work_struct *work);
33 * The ring buffer header is special. We must manually up keep it.
35 int ring_buffer_print_entry_header(struct trace_seq *s)
37 trace_seq_puts(s, "# compressed entry header\n");
38 trace_seq_puts(s, "\ttype_len : 5 bits\n");
39 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
40 trace_seq_puts(s, "\tarray : 32 bits\n");
41 trace_seq_putc(s, '\n');
42 trace_seq_printf(s, "\tpadding : type == %d\n",
43 RINGBUF_TYPE_PADDING);
44 trace_seq_printf(s, "\ttime_extend : type == %d\n",
45 RINGBUF_TYPE_TIME_EXTEND);
46 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
47 RINGBUF_TYPE_TIME_STAMP);
48 trace_seq_printf(s, "\tdata max type_len == %d\n",
49 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
51 return !trace_seq_has_overflowed(s);
55 * The ring buffer is made up of a list of pages. A separate list of pages is
56 * allocated for each CPU. A writer may only write to a buffer that is
57 * associated with the CPU it is currently executing on. A reader may read
58 * from any per cpu buffer.
60 * The reader is special. For each per cpu buffer, the reader has its own
61 * reader page. When a reader has read the entire reader page, this reader
62 * page is swapped with another page in the ring buffer.
64 * Now, as long as the writer is off the reader page, the reader can do what
65 * ever it wants with that page. The writer will never write to that page
66 * again (as long as it is out of the ring buffer).
68 * Here's some silly ASCII art.
71 * |reader| RING BUFFER
73 * +------+ +---+ +---+ +---+
82 * |reader| RING BUFFER
83 * |page |------------------v
84 * +------+ +---+ +---+ +---+
93 * |reader| RING BUFFER
94 * |page |------------------v
95 * +------+ +---+ +---+ +---+
100 * +------------------------------+
104 * |buffer| RING BUFFER
105 * |page |------------------v
106 * +------+ +---+ +---+ +---+
108 * | New +---+ +---+ +---+
111 * +------------------------------+
114 * After we make this swap, the reader can hand this page off to the splice
115 * code and be done with it. It can even allocate a new page if it needs to
116 * and swap that into the ring buffer.
118 * We will be using cmpxchg soon to make all this lockless.
122 /* Used for individual buffers (after the counter) */
123 #define RB_BUFFER_OFF (1 << 20)
125 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
127 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
128 #define RB_ALIGNMENT 4U
129 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
130 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
131 #define RB_ALIGN_DATA __aligned(RB_ALIGNMENT)
133 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
134 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
137 RB_LEN_TIME_EXTEND = 8,
138 RB_LEN_TIME_STAMP = 8,
141 #define skip_time_extend(event) \
142 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
144 #define extended_time(event) \
145 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
147 static inline int rb_null_event(struct ring_buffer_event *event)
149 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
152 static void rb_event_set_padding(struct ring_buffer_event *event)
154 /* padding has a NULL time_delta */
155 event->type_len = RINGBUF_TYPE_PADDING;
156 event->time_delta = 0;
160 rb_event_data_length(struct ring_buffer_event *event)
165 length = event->type_len * RB_ALIGNMENT;
167 length = event->array[0];
168 return length + RB_EVNT_HDR_SIZE;
172 * Return the length of the given event. Will return
173 * the length of the time extend if the event is a
176 static inline unsigned
177 rb_event_length(struct ring_buffer_event *event)
179 switch (event->type_len) {
180 case RINGBUF_TYPE_PADDING:
181 if (rb_null_event(event))
184 return event->array[0] + RB_EVNT_HDR_SIZE;
186 case RINGBUF_TYPE_TIME_EXTEND:
187 return RB_LEN_TIME_EXTEND;
189 case RINGBUF_TYPE_TIME_STAMP:
190 return RB_LEN_TIME_STAMP;
192 case RINGBUF_TYPE_DATA:
193 return rb_event_data_length(event);
202 * Return total length of time extend and data,
203 * or just the event length for all other events.
205 static inline unsigned
206 rb_event_ts_length(struct ring_buffer_event *event)
210 if (extended_time(event)) {
211 /* time extends include the data event after it */
212 len = RB_LEN_TIME_EXTEND;
213 event = skip_time_extend(event);
215 return len + rb_event_length(event);
219 * ring_buffer_event_length - return the length of the event
220 * @event: the event to get the length of
222 * Returns the size of the data load of a data event.
223 * If the event is something other than a data event, it
224 * returns the size of the event itself. With the exception
225 * of a TIME EXTEND, where it still returns the size of the
226 * data load of the data event after it.
228 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
232 if (extended_time(event))
233 event = skip_time_extend(event);
235 length = rb_event_length(event);
236 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
238 length -= RB_EVNT_HDR_SIZE;
239 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
240 length -= sizeof(event->array[0]);
243 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
245 /* inline for ring buffer fast paths */
246 static __always_inline void *
247 rb_event_data(struct ring_buffer_event *event)
249 if (extended_time(event))
250 event = skip_time_extend(event);
251 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
252 /* If length is in len field, then array[0] has the data */
254 return (void *)&event->array[0];
255 /* Otherwise length is in array[0] and array[1] has the data */
256 return (void *)&event->array[1];
260 * ring_buffer_event_data - return the data of the event
261 * @event: the event to get the data from
263 void *ring_buffer_event_data(struct ring_buffer_event *event)
265 return rb_event_data(event);
267 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
269 #define for_each_buffer_cpu(buffer, cpu) \
270 for_each_cpu(cpu, buffer->cpumask)
273 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
274 #define TS_DELTA_TEST (~TS_MASK)
277 * ring_buffer_event_time_stamp - return the event's extended timestamp
278 * @event: the event to get the timestamp of
280 * Returns the extended timestamp associated with a data event.
281 * An extended time_stamp is a 64-bit timestamp represented
282 * internally in a special way that makes the best use of space
283 * contained within a ring buffer event. This function decodes
284 * it and maps it to a straight u64 value.
286 u64 ring_buffer_event_time_stamp(struct ring_buffer_event *event)
290 ts = event->array[0];
292 ts += event->time_delta;
297 /* Flag when events were overwritten */
298 #define RB_MISSED_EVENTS (1 << 31)
299 /* Missed count stored at end */
300 #define RB_MISSED_STORED (1 << 30)
302 #define RB_MISSED_FLAGS (RB_MISSED_EVENTS|RB_MISSED_STORED)
304 struct buffer_data_page {
305 u64 time_stamp; /* page time stamp */
306 local_t commit; /* write committed index */
307 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
311 * Note, the buffer_page list must be first. The buffer pages
312 * are allocated in cache lines, which means that each buffer
313 * page will be at the beginning of a cache line, and thus
314 * the least significant bits will be zero. We use this to
315 * add flags in the list struct pointers, to make the ring buffer
319 struct list_head list; /* list of buffer pages */
320 local_t write; /* index for next write */
321 unsigned read; /* index for next read */
322 local_t entries; /* entries on this page */
323 unsigned long real_end; /* real end of data */
324 struct buffer_data_page *page; /* Actual data page */
328 * The buffer page counters, write and entries, must be reset
329 * atomically when crossing page boundaries. To synchronize this
330 * update, two counters are inserted into the number. One is
331 * the actual counter for the write position or count on the page.
333 * The other is a counter of updaters. Before an update happens
334 * the update partition of the counter is incremented. This will
335 * allow the updater to update the counter atomically.
337 * The counter is 20 bits, and the state data is 12.
339 #define RB_WRITE_MASK 0xfffff
340 #define RB_WRITE_INTCNT (1 << 20)
342 static void rb_init_page(struct buffer_data_page *bpage)
344 local_set(&bpage->commit, 0);
348 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
351 static void free_buffer_page(struct buffer_page *bpage)
353 free_page((unsigned long)bpage->page);
358 * We need to fit the time_stamp delta into 27 bits.
360 static inline int test_time_stamp(u64 delta)
362 if (delta & TS_DELTA_TEST)
367 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
369 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
370 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
372 int ring_buffer_print_page_header(struct trace_seq *s)
374 struct buffer_data_page field;
376 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
377 "offset:0;\tsize:%u;\tsigned:%u;\n",
378 (unsigned int)sizeof(field.time_stamp),
379 (unsigned int)is_signed_type(u64));
381 trace_seq_printf(s, "\tfield: local_t commit;\t"
382 "offset:%u;\tsize:%u;\tsigned:%u;\n",
383 (unsigned int)offsetof(typeof(field), commit),
384 (unsigned int)sizeof(field.commit),
385 (unsigned int)is_signed_type(long));
387 trace_seq_printf(s, "\tfield: int overwrite;\t"
388 "offset:%u;\tsize:%u;\tsigned:%u;\n",
389 (unsigned int)offsetof(typeof(field), commit),
391 (unsigned int)is_signed_type(long));
393 trace_seq_printf(s, "\tfield: char data;\t"
394 "offset:%u;\tsize:%u;\tsigned:%u;\n",
395 (unsigned int)offsetof(typeof(field), data),
396 (unsigned int)BUF_PAGE_SIZE,
397 (unsigned int)is_signed_type(char));
399 return !trace_seq_has_overflowed(s);
403 struct irq_work work;
404 wait_queue_head_t waiters;
405 wait_queue_head_t full_waiters;
406 bool waiters_pending;
407 bool full_waiters_pending;
412 * Structure to hold event state and handle nested events.
414 struct rb_event_info {
417 unsigned long length;
418 struct buffer_page *tail_page;
423 * Used for which event context the event is in.
429 * See trace_recursive_lock() comment below for more details.
440 * head_page == tail_page && head == tail then buffer is empty.
442 struct ring_buffer_per_cpu {
444 atomic_t record_disabled;
445 struct ring_buffer *buffer;
446 raw_spinlock_t reader_lock; /* serialize readers */
447 arch_spinlock_t lock;
448 struct lock_class_key lock_key;
449 struct buffer_data_page *free_page;
450 unsigned long nr_pages;
451 unsigned int current_context;
452 struct list_head *pages;
453 struct buffer_page *head_page; /* read from head */
454 struct buffer_page *tail_page; /* write to tail */
455 struct buffer_page *commit_page; /* committed pages */
456 struct buffer_page *reader_page;
457 unsigned long lost_events;
458 unsigned long last_overrun;
460 local_t entries_bytes;
463 local_t commit_overrun;
464 local_t dropped_events;
467 local_t pages_touched;
469 long last_pages_touch;
470 size_t shortest_full;
472 unsigned long read_bytes;
475 /* ring buffer pages to update, > 0 to add, < 0 to remove */
476 long nr_pages_to_update;
477 struct list_head new_pages; /* new pages to add */
478 struct work_struct update_pages_work;
479 struct completion update_done;
481 struct rb_irq_work irq_work;
487 atomic_t record_disabled;
488 atomic_t resize_disabled;
489 cpumask_var_t cpumask;
491 struct lock_class_key *reader_lock_key;
495 struct ring_buffer_per_cpu **buffers;
497 struct hlist_node node;
500 struct rb_irq_work irq_work;
504 struct ring_buffer_iter {
505 struct ring_buffer_per_cpu *cpu_buffer;
507 struct buffer_page *head_page;
508 struct buffer_page *cache_reader_page;
509 unsigned long cache_read;
514 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
515 * @buffer: The ring_buffer to get the number of pages from
516 * @cpu: The cpu of the ring_buffer to get the number of pages from
518 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
520 size_t ring_buffer_nr_pages(struct ring_buffer *buffer, int cpu)
522 return buffer->buffers[cpu]->nr_pages;
526 * ring_buffer_nr_pages_dirty - get the number of used pages in the ring buffer
527 * @buffer: The ring_buffer to get the number of pages from
528 * @cpu: The cpu of the ring_buffer to get the number of pages from
530 * Returns the number of pages that have content in the ring buffer.
532 size_t ring_buffer_nr_dirty_pages(struct ring_buffer *buffer, int cpu)
537 read = local_read(&buffer->buffers[cpu]->pages_read);
538 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
539 /* The reader can read an empty page, but not more than that */
541 WARN_ON_ONCE(read > cnt + 1);
549 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
551 * Schedules a delayed work to wake up any task that is blocked on the
552 * ring buffer waiters queue.
554 static void rb_wake_up_waiters(struct irq_work *work)
556 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
558 wake_up_all(&rbwork->waiters);
559 if (rbwork->wakeup_full) {
560 rbwork->wakeup_full = false;
561 wake_up_all(&rbwork->full_waiters);
566 * ring_buffer_wait - wait for input to the ring buffer
567 * @buffer: buffer to wait on
568 * @cpu: the cpu buffer to wait on
569 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
571 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
572 * as data is added to any of the @buffer's cpu buffers. Otherwise
573 * it will wait for data to be added to a specific cpu buffer.
575 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, int full)
577 struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
579 struct rb_irq_work *work;
583 * Depending on what the caller is waiting for, either any
584 * data in any cpu buffer, or a specific buffer, put the
585 * caller on the appropriate wait queue.
587 if (cpu == RING_BUFFER_ALL_CPUS) {
588 work = &buffer->irq_work;
589 /* Full only makes sense on per cpu reads */
592 if (!cpumask_test_cpu(cpu, buffer->cpumask))
594 cpu_buffer = buffer->buffers[cpu];
595 work = &cpu_buffer->irq_work;
601 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
603 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
606 * The events can happen in critical sections where
607 * checking a work queue can cause deadlocks.
608 * After adding a task to the queue, this flag is set
609 * only to notify events to try to wake up the queue
612 * We don't clear it even if the buffer is no longer
613 * empty. The flag only causes the next event to run
614 * irq_work to do the work queue wake up. The worse
615 * that can happen if we race with !trace_empty() is that
616 * an event will cause an irq_work to try to wake up
619 * There's no reason to protect this flag either, as
620 * the work queue and irq_work logic will do the necessary
621 * synchronization for the wake ups. The only thing
622 * that is necessary is that the wake up happens after
623 * a task has been queued. It's OK for spurious wake ups.
626 work->full_waiters_pending = true;
628 work->waiters_pending = true;
630 if (signal_pending(current)) {
635 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
638 if (cpu != RING_BUFFER_ALL_CPUS &&
639 !ring_buffer_empty_cpu(buffer, cpu)) {
648 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
649 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
650 nr_pages = cpu_buffer->nr_pages;
651 dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
652 if (!cpu_buffer->shortest_full ||
653 cpu_buffer->shortest_full < full)
654 cpu_buffer->shortest_full = full;
655 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
657 (!nr_pages || (dirty * 100) > full * nr_pages))
665 finish_wait(&work->full_waiters, &wait);
667 finish_wait(&work->waiters, &wait);
673 * ring_buffer_poll_wait - poll on buffer input
674 * @buffer: buffer to wait on
675 * @cpu: the cpu buffer to wait on
676 * @filp: the file descriptor
677 * @poll_table: The poll descriptor
679 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
680 * as data is added to any of the @buffer's cpu buffers. Otherwise
681 * it will wait for data to be added to a specific cpu buffer.
683 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
686 __poll_t ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
687 struct file *filp, poll_table *poll_table)
689 struct ring_buffer_per_cpu *cpu_buffer;
690 struct rb_irq_work *work;
692 if (cpu == RING_BUFFER_ALL_CPUS)
693 work = &buffer->irq_work;
695 if (!cpumask_test_cpu(cpu, buffer->cpumask))
698 cpu_buffer = buffer->buffers[cpu];
699 work = &cpu_buffer->irq_work;
702 poll_wait(filp, &work->waiters, poll_table);
703 work->waiters_pending = true;
705 * There's a tight race between setting the waiters_pending and
706 * checking if the ring buffer is empty. Once the waiters_pending bit
707 * is set, the next event will wake the task up, but we can get stuck
708 * if there's only a single event in.
710 * FIXME: Ideally, we need a memory barrier on the writer side as well,
711 * but adding a memory barrier to all events will cause too much of a
712 * performance hit in the fast path. We only need a memory barrier when
713 * the buffer goes from empty to having content. But as this race is
714 * extremely small, and it's not a problem if another event comes in, we
719 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
720 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
721 return EPOLLIN | EPOLLRDNORM;
725 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
726 #define RB_WARN_ON(b, cond) \
728 int _____ret = unlikely(cond); \
730 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
731 struct ring_buffer_per_cpu *__b = \
733 atomic_inc(&__b->buffer->record_disabled); \
735 atomic_inc(&b->record_disabled); \
741 /* Up this if you want to test the TIME_EXTENTS and normalization */
742 #define DEBUG_SHIFT 0
744 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
746 /* shift to debug/test normalization and TIME_EXTENTS */
747 return buffer->clock() << DEBUG_SHIFT;
750 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
754 preempt_disable_notrace();
755 time = rb_time_stamp(buffer);
756 preempt_enable_notrace();
760 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
762 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
765 /* Just stupid testing the normalize function and deltas */
768 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
771 * Making the ring buffer lockless makes things tricky.
772 * Although writes only happen on the CPU that they are on,
773 * and they only need to worry about interrupts. Reads can
776 * The reader page is always off the ring buffer, but when the
777 * reader finishes with a page, it needs to swap its page with
778 * a new one from the buffer. The reader needs to take from
779 * the head (writes go to the tail). But if a writer is in overwrite
780 * mode and wraps, it must push the head page forward.
782 * Here lies the problem.
784 * The reader must be careful to replace only the head page, and
785 * not another one. As described at the top of the file in the
786 * ASCII art, the reader sets its old page to point to the next
787 * page after head. It then sets the page after head to point to
788 * the old reader page. But if the writer moves the head page
789 * during this operation, the reader could end up with the tail.
791 * We use cmpxchg to help prevent this race. We also do something
792 * special with the page before head. We set the LSB to 1.
794 * When the writer must push the page forward, it will clear the
795 * bit that points to the head page, move the head, and then set
796 * the bit that points to the new head page.
798 * We also don't want an interrupt coming in and moving the head
799 * page on another writer. Thus we use the second LSB to catch
802 * head->list->prev->next bit 1 bit 0
805 * Points to head page 0 1
808 * Note we can not trust the prev pointer of the head page, because:
810 * +----+ +-----+ +-----+
811 * | |------>| T |---X--->| N |
813 * +----+ +-----+ +-----+
816 * +----------| R |----------+ |
820 * Key: ---X--> HEAD flag set in pointer
825 * (see __rb_reserve_next() to see where this happens)
827 * What the above shows is that the reader just swapped out
828 * the reader page with a page in the buffer, but before it
829 * could make the new header point back to the new page added
830 * it was preempted by a writer. The writer moved forward onto
831 * the new page added by the reader and is about to move forward
834 * You can see, it is legitimate for the previous pointer of
835 * the head (or any page) not to point back to itself. But only
839 #define RB_PAGE_NORMAL 0UL
840 #define RB_PAGE_HEAD 1UL
841 #define RB_PAGE_UPDATE 2UL
844 #define RB_FLAG_MASK 3UL
846 /* PAGE_MOVED is not part of the mask */
847 #define RB_PAGE_MOVED 4UL
850 * rb_list_head - remove any bit
852 static struct list_head *rb_list_head(struct list_head *list)
854 unsigned long val = (unsigned long)list;
856 return (struct list_head *)(val & ~RB_FLAG_MASK);
860 * rb_is_head_page - test if the given page is the head page
862 * Because the reader may move the head_page pointer, we can
863 * not trust what the head page is (it may be pointing to
864 * the reader page). But if the next page is a header page,
865 * its flags will be non zero.
868 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
869 struct buffer_page *page, struct list_head *list)
873 val = (unsigned long)list->next;
875 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
876 return RB_PAGE_MOVED;
878 return val & RB_FLAG_MASK;
884 * The unique thing about the reader page, is that, if the
885 * writer is ever on it, the previous pointer never points
886 * back to the reader page.
888 static bool rb_is_reader_page(struct buffer_page *page)
890 struct list_head *list = page->list.prev;
892 return rb_list_head(list->next) != &page->list;
896 * rb_set_list_to_head - set a list_head to be pointing to head.
898 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
899 struct list_head *list)
903 ptr = (unsigned long *)&list->next;
904 *ptr |= RB_PAGE_HEAD;
905 *ptr &= ~RB_PAGE_UPDATE;
909 * rb_head_page_activate - sets up head page
911 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
913 struct buffer_page *head;
915 head = cpu_buffer->head_page;
920 * Set the previous list pointer to have the HEAD flag.
922 rb_set_list_to_head(cpu_buffer, head->list.prev);
925 static void rb_list_head_clear(struct list_head *list)
927 unsigned long *ptr = (unsigned long *)&list->next;
929 *ptr &= ~RB_FLAG_MASK;
933 * rb_head_page_deactivate - clears head page ptr (for free list)
936 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
938 struct list_head *hd;
940 /* Go through the whole list and clear any pointers found. */
941 rb_list_head_clear(cpu_buffer->pages);
943 list_for_each(hd, cpu_buffer->pages)
944 rb_list_head_clear(hd);
947 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
948 struct buffer_page *head,
949 struct buffer_page *prev,
950 int old_flag, int new_flag)
952 struct list_head *list;
953 unsigned long val = (unsigned long)&head->list;
958 val &= ~RB_FLAG_MASK;
960 ret = cmpxchg((unsigned long *)&list->next,
961 val | old_flag, val | new_flag);
963 /* check if the reader took the page */
964 if ((ret & ~RB_FLAG_MASK) != val)
965 return RB_PAGE_MOVED;
967 return ret & RB_FLAG_MASK;
970 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
971 struct buffer_page *head,
972 struct buffer_page *prev,
975 return rb_head_page_set(cpu_buffer, head, prev,
976 old_flag, RB_PAGE_UPDATE);
979 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
980 struct buffer_page *head,
981 struct buffer_page *prev,
984 return rb_head_page_set(cpu_buffer, head, prev,
985 old_flag, RB_PAGE_HEAD);
988 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
989 struct buffer_page *head,
990 struct buffer_page *prev,
993 return rb_head_page_set(cpu_buffer, head, prev,
994 old_flag, RB_PAGE_NORMAL);
997 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
998 struct buffer_page **bpage)
1000 struct list_head *p = rb_list_head((*bpage)->list.next);
1002 *bpage = list_entry(p, struct buffer_page, list);
1005 static struct buffer_page *
1006 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1008 struct buffer_page *head;
1009 struct buffer_page *page;
1010 struct list_head *list;
1013 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1017 list = cpu_buffer->pages;
1018 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1021 page = head = cpu_buffer->head_page;
1023 * It is possible that the writer moves the header behind
1024 * where we started, and we miss in one loop.
1025 * A second loop should grab the header, but we'll do
1026 * three loops just because I'm paranoid.
1028 for (i = 0; i < 3; i++) {
1030 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
1031 cpu_buffer->head_page = page;
1034 rb_inc_page(cpu_buffer, &page);
1035 } while (page != head);
1038 RB_WARN_ON(cpu_buffer, 1);
1043 static int rb_head_page_replace(struct buffer_page *old,
1044 struct buffer_page *new)
1046 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1050 val = *ptr & ~RB_FLAG_MASK;
1051 val |= RB_PAGE_HEAD;
1053 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1059 * rb_tail_page_update - move the tail page forward
1061 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1062 struct buffer_page *tail_page,
1063 struct buffer_page *next_page)
1065 unsigned long old_entries;
1066 unsigned long old_write;
1069 * The tail page now needs to be moved forward.
1071 * We need to reset the tail page, but without messing
1072 * with possible erasing of data brought in by interrupts
1073 * that have moved the tail page and are currently on it.
1075 * We add a counter to the write field to denote this.
1077 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1078 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1080 local_inc(&cpu_buffer->pages_touched);
1082 * Just make sure we have seen our old_write and synchronize
1083 * with any interrupts that come in.
1088 * If the tail page is still the same as what we think
1089 * it is, then it is up to us to update the tail
1092 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1093 /* Zero the write counter */
1094 unsigned long val = old_write & ~RB_WRITE_MASK;
1095 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1098 * This will only succeed if an interrupt did
1099 * not come in and change it. In which case, we
1100 * do not want to modify it.
1102 * We add (void) to let the compiler know that we do not care
1103 * about the return value of these functions. We use the
1104 * cmpxchg to only update if an interrupt did not already
1105 * do it for us. If the cmpxchg fails, we don't care.
1107 (void)local_cmpxchg(&next_page->write, old_write, val);
1108 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1111 * No need to worry about races with clearing out the commit.
1112 * it only can increment when a commit takes place. But that
1113 * only happens in the outer most nested commit.
1115 local_set(&next_page->page->commit, 0);
1117 /* Again, either we update tail_page or an interrupt does */
1118 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1122 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1123 struct buffer_page *bpage)
1125 unsigned long val = (unsigned long)bpage;
1127 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1134 * rb_check_list - make sure a pointer to a list has the last bits zero
1136 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1137 struct list_head *list)
1139 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1141 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1147 * rb_check_pages - integrity check of buffer pages
1148 * @cpu_buffer: CPU buffer with pages to test
1150 * As a safety measure we check to make sure the data pages have not
1153 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1155 struct list_head *head = cpu_buffer->pages;
1156 struct buffer_page *bpage, *tmp;
1158 /* Reset the head page if it exists */
1159 if (cpu_buffer->head_page)
1160 rb_set_head_page(cpu_buffer);
1162 rb_head_page_deactivate(cpu_buffer);
1164 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1166 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1169 if (rb_check_list(cpu_buffer, head))
1172 list_for_each_entry_safe(bpage, tmp, head, list) {
1173 if (RB_WARN_ON(cpu_buffer,
1174 bpage->list.next->prev != &bpage->list))
1176 if (RB_WARN_ON(cpu_buffer,
1177 bpage->list.prev->next != &bpage->list))
1179 if (rb_check_list(cpu_buffer, &bpage->list))
1183 rb_head_page_activate(cpu_buffer);
1188 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1190 struct buffer_page *bpage, *tmp;
1191 bool user_thread = current->mm != NULL;
1196 * Check if the available memory is there first.
1197 * Note, si_mem_available() only gives us a rough estimate of available
1198 * memory. It may not be accurate. But we don't care, we just want
1199 * to prevent doing any allocation when it is obvious that it is
1200 * not going to succeed.
1202 i = si_mem_available();
1207 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1208 * gracefully without invoking oom-killer and the system is not
1211 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1214 * If a user thread allocates too much, and si_mem_available()
1215 * reports there's enough memory, even though there is not.
1216 * Make sure the OOM killer kills this thread. This can happen
1217 * even with RETRY_MAYFAIL because another task may be doing
1218 * an allocation after this task has taken all memory.
1219 * This is the task the OOM killer needs to take out during this
1220 * loop, even if it was triggered by an allocation somewhere else.
1223 set_current_oom_origin();
1224 for (i = 0; i < nr_pages; i++) {
1227 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1228 mflags, cpu_to_node(cpu));
1232 list_add(&bpage->list, pages);
1234 page = alloc_pages_node(cpu_to_node(cpu), mflags, 0);
1237 bpage->page = page_address(page);
1238 rb_init_page(bpage->page);
1240 if (user_thread && fatal_signal_pending(current))
1244 clear_current_oom_origin();
1249 list_for_each_entry_safe(bpage, tmp, pages, list) {
1250 list_del_init(&bpage->list);
1251 free_buffer_page(bpage);
1254 clear_current_oom_origin();
1259 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1260 unsigned long nr_pages)
1266 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1270 * The ring buffer page list is a circular list that does not
1271 * start and end with a list head. All page list items point to
1274 cpu_buffer->pages = pages.next;
1277 cpu_buffer->nr_pages = nr_pages;
1279 rb_check_pages(cpu_buffer);
1284 static struct ring_buffer_per_cpu *
1285 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1287 struct ring_buffer_per_cpu *cpu_buffer;
1288 struct buffer_page *bpage;
1292 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1293 GFP_KERNEL, cpu_to_node(cpu));
1297 cpu_buffer->cpu = cpu;
1298 cpu_buffer->buffer = buffer;
1299 raw_spin_lock_init(&cpu_buffer->reader_lock);
1300 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1301 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1302 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1303 init_completion(&cpu_buffer->update_done);
1304 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1305 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1306 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1308 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1309 GFP_KERNEL, cpu_to_node(cpu));
1311 goto fail_free_buffer;
1313 rb_check_bpage(cpu_buffer, bpage);
1315 cpu_buffer->reader_page = bpage;
1316 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1318 goto fail_free_reader;
1319 bpage->page = page_address(page);
1320 rb_init_page(bpage->page);
1322 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1323 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1325 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1327 goto fail_free_reader;
1329 cpu_buffer->head_page
1330 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1331 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1333 rb_head_page_activate(cpu_buffer);
1338 free_buffer_page(cpu_buffer->reader_page);
1345 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1347 struct list_head *head = cpu_buffer->pages;
1348 struct buffer_page *bpage, *tmp;
1350 free_buffer_page(cpu_buffer->reader_page);
1352 rb_head_page_deactivate(cpu_buffer);
1355 list_for_each_entry_safe(bpage, tmp, head, list) {
1356 list_del_init(&bpage->list);
1357 free_buffer_page(bpage);
1359 bpage = list_entry(head, struct buffer_page, list);
1360 free_buffer_page(bpage);
1367 * __ring_buffer_alloc - allocate a new ring_buffer
1368 * @size: the size in bytes per cpu that is needed.
1369 * @flags: attributes to set for the ring buffer.
1371 * Currently the only flag that is available is the RB_FL_OVERWRITE
1372 * flag. This flag means that the buffer will overwrite old data
1373 * when the buffer wraps. If this flag is not set, the buffer will
1374 * drop data when the tail hits the head.
1376 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1377 struct lock_class_key *key)
1379 struct ring_buffer *buffer;
1385 /* keep it in its own cache line */
1386 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1391 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1392 goto fail_free_buffer;
1394 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1395 buffer->flags = flags;
1396 buffer->clock = trace_clock_local;
1397 buffer->reader_lock_key = key;
1399 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1400 init_waitqueue_head(&buffer->irq_work.waiters);
1402 /* need at least two pages */
1406 buffer->cpus = nr_cpu_ids;
1408 bsize = sizeof(void *) * nr_cpu_ids;
1409 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1411 if (!buffer->buffers)
1412 goto fail_free_cpumask;
1414 cpu = raw_smp_processor_id();
1415 cpumask_set_cpu(cpu, buffer->cpumask);
1416 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1417 if (!buffer->buffers[cpu])
1418 goto fail_free_buffers;
1420 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1422 goto fail_free_buffers;
1424 mutex_init(&buffer->mutex);
1429 for_each_buffer_cpu(buffer, cpu) {
1430 if (buffer->buffers[cpu])
1431 rb_free_cpu_buffer(buffer->buffers[cpu]);
1433 kfree(buffer->buffers);
1436 free_cpumask_var(buffer->cpumask);
1442 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1445 * ring_buffer_free - free a ring buffer.
1446 * @buffer: the buffer to free.
1449 ring_buffer_free(struct ring_buffer *buffer)
1453 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1455 for_each_buffer_cpu(buffer, cpu)
1456 rb_free_cpu_buffer(buffer->buffers[cpu]);
1458 kfree(buffer->buffers);
1459 free_cpumask_var(buffer->cpumask);
1463 EXPORT_SYMBOL_GPL(ring_buffer_free);
1465 void ring_buffer_set_clock(struct ring_buffer *buffer,
1468 buffer->clock = clock;
1471 void ring_buffer_set_time_stamp_abs(struct ring_buffer *buffer, bool abs)
1473 buffer->time_stamp_abs = abs;
1476 bool ring_buffer_time_stamp_abs(struct ring_buffer *buffer)
1478 return buffer->time_stamp_abs;
1481 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1483 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1485 return local_read(&bpage->entries) & RB_WRITE_MASK;
1488 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1490 return local_read(&bpage->write) & RB_WRITE_MASK;
1494 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1496 struct list_head *tail_page, *to_remove, *next_page;
1497 struct buffer_page *to_remove_page, *tmp_iter_page;
1498 struct buffer_page *last_page, *first_page;
1499 unsigned long nr_removed;
1500 unsigned long head_bit;
1505 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1506 atomic_inc(&cpu_buffer->record_disabled);
1508 * We don't race with the readers since we have acquired the reader
1509 * lock. We also don't race with writers after disabling recording.
1510 * This makes it easy to figure out the first and the last page to be
1511 * removed from the list. We unlink all the pages in between including
1512 * the first and last pages. This is done in a busy loop so that we
1513 * lose the least number of traces.
1514 * The pages are freed after we restart recording and unlock readers.
1516 tail_page = &cpu_buffer->tail_page->list;
1519 * tail page might be on reader page, we remove the next page
1520 * from the ring buffer
1522 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1523 tail_page = rb_list_head(tail_page->next);
1524 to_remove = tail_page;
1526 /* start of pages to remove */
1527 first_page = list_entry(rb_list_head(to_remove->next),
1528 struct buffer_page, list);
1530 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1531 to_remove = rb_list_head(to_remove)->next;
1532 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1535 next_page = rb_list_head(to_remove)->next;
1538 * Now we remove all pages between tail_page and next_page.
1539 * Make sure that we have head_bit value preserved for the
1542 tail_page->next = (struct list_head *)((unsigned long)next_page |
1544 next_page = rb_list_head(next_page);
1545 next_page->prev = tail_page;
1547 /* make sure pages points to a valid page in the ring buffer */
1548 cpu_buffer->pages = next_page;
1550 /* update head page */
1552 cpu_buffer->head_page = list_entry(next_page,
1553 struct buffer_page, list);
1556 * change read pointer to make sure any read iterators reset
1559 cpu_buffer->read = 0;
1561 /* pages are removed, resume tracing and then free the pages */
1562 atomic_dec(&cpu_buffer->record_disabled);
1563 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1565 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1567 /* last buffer page to remove */
1568 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1570 tmp_iter_page = first_page;
1575 to_remove_page = tmp_iter_page;
1576 rb_inc_page(cpu_buffer, &tmp_iter_page);
1578 /* update the counters */
1579 page_entries = rb_page_entries(to_remove_page);
1582 * If something was added to this page, it was full
1583 * since it is not the tail page. So we deduct the
1584 * bytes consumed in ring buffer from here.
1585 * Increment overrun to account for the lost events.
1587 local_add(page_entries, &cpu_buffer->overrun);
1588 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1592 * We have already removed references to this list item, just
1593 * free up the buffer_page and its page
1595 free_buffer_page(to_remove_page);
1598 } while (to_remove_page != last_page);
1600 RB_WARN_ON(cpu_buffer, nr_removed);
1602 return nr_removed == 0;
1606 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1608 struct list_head *pages = &cpu_buffer->new_pages;
1609 int retries, success;
1611 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1613 * We are holding the reader lock, so the reader page won't be swapped
1614 * in the ring buffer. Now we are racing with the writer trying to
1615 * move head page and the tail page.
1616 * We are going to adapt the reader page update process where:
1617 * 1. We first splice the start and end of list of new pages between
1618 * the head page and its previous page.
1619 * 2. We cmpxchg the prev_page->next to point from head page to the
1620 * start of new pages list.
1621 * 3. Finally, we update the head->prev to the end of new list.
1623 * We will try this process 10 times, to make sure that we don't keep
1629 struct list_head *head_page, *prev_page, *r;
1630 struct list_head *last_page, *first_page;
1631 struct list_head *head_page_with_bit;
1633 head_page = &rb_set_head_page(cpu_buffer)->list;
1636 prev_page = head_page->prev;
1638 first_page = pages->next;
1639 last_page = pages->prev;
1641 head_page_with_bit = (struct list_head *)
1642 ((unsigned long)head_page | RB_PAGE_HEAD);
1644 last_page->next = head_page_with_bit;
1645 first_page->prev = prev_page;
1647 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1649 if (r == head_page_with_bit) {
1651 * yay, we replaced the page pointer to our new list,
1652 * now, we just have to update to head page's prev
1653 * pointer to point to end of list
1655 head_page->prev = last_page;
1662 INIT_LIST_HEAD(pages);
1664 * If we weren't successful in adding in new pages, warn and stop
1667 RB_WARN_ON(cpu_buffer, !success);
1668 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1670 /* free pages if they weren't inserted */
1672 struct buffer_page *bpage, *tmp;
1673 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1675 list_del_init(&bpage->list);
1676 free_buffer_page(bpage);
1682 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1686 if (cpu_buffer->nr_pages_to_update > 0)
1687 success = rb_insert_pages(cpu_buffer);
1689 success = rb_remove_pages(cpu_buffer,
1690 -cpu_buffer->nr_pages_to_update);
1693 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1696 static void update_pages_handler(struct work_struct *work)
1698 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1699 struct ring_buffer_per_cpu, update_pages_work);
1700 rb_update_pages(cpu_buffer);
1701 complete(&cpu_buffer->update_done);
1705 * ring_buffer_resize - resize the ring buffer
1706 * @buffer: the buffer to resize.
1707 * @size: the new size.
1708 * @cpu_id: the cpu buffer to resize
1710 * Minimum size is 2 * BUF_PAGE_SIZE.
1712 * Returns 0 on success and < 0 on failure.
1714 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1717 struct ring_buffer_per_cpu *cpu_buffer;
1718 unsigned long nr_pages;
1722 * Always succeed at resizing a non-existent buffer:
1727 /* Make sure the requested buffer exists */
1728 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1729 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1732 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1734 /* we need a minimum of two pages */
1738 size = nr_pages * BUF_PAGE_SIZE;
1741 * Don't succeed if resizing is disabled, as a reader might be
1742 * manipulating the ring buffer and is expecting a sane state while
1745 if (atomic_read(&buffer->resize_disabled))
1748 /* prevent another thread from changing buffer sizes */
1749 mutex_lock(&buffer->mutex);
1751 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1752 /* calculate the pages to update */
1753 for_each_buffer_cpu(buffer, cpu) {
1754 cpu_buffer = buffer->buffers[cpu];
1756 cpu_buffer->nr_pages_to_update = nr_pages -
1757 cpu_buffer->nr_pages;
1759 * nothing more to do for removing pages or no update
1761 if (cpu_buffer->nr_pages_to_update <= 0)
1764 * to add pages, make sure all new pages can be
1765 * allocated without receiving ENOMEM
1767 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1768 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1769 &cpu_buffer->new_pages, cpu)) {
1770 /* not enough memory for new pages */
1778 * Fire off all the required work handlers
1779 * We can't schedule on offline CPUs, but it's not necessary
1780 * since we can change their buffer sizes without any race.
1782 for_each_buffer_cpu(buffer, cpu) {
1783 cpu_buffer = buffer->buffers[cpu];
1784 if (!cpu_buffer->nr_pages_to_update)
1787 /* Can't run something on an offline CPU. */
1788 if (!cpu_online(cpu)) {
1789 rb_update_pages(cpu_buffer);
1790 cpu_buffer->nr_pages_to_update = 0;
1792 schedule_work_on(cpu,
1793 &cpu_buffer->update_pages_work);
1797 /* wait for all the updates to complete */
1798 for_each_buffer_cpu(buffer, cpu) {
1799 cpu_buffer = buffer->buffers[cpu];
1800 if (!cpu_buffer->nr_pages_to_update)
1803 if (cpu_online(cpu))
1804 wait_for_completion(&cpu_buffer->update_done);
1805 cpu_buffer->nr_pages_to_update = 0;
1810 /* Make sure this CPU has been initialized */
1811 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1814 cpu_buffer = buffer->buffers[cpu_id];
1816 if (nr_pages == cpu_buffer->nr_pages)
1819 cpu_buffer->nr_pages_to_update = nr_pages -
1820 cpu_buffer->nr_pages;
1822 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1823 if (cpu_buffer->nr_pages_to_update > 0 &&
1824 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1825 &cpu_buffer->new_pages, cpu_id)) {
1832 /* Can't run something on an offline CPU. */
1833 if (!cpu_online(cpu_id))
1834 rb_update_pages(cpu_buffer);
1836 schedule_work_on(cpu_id,
1837 &cpu_buffer->update_pages_work);
1838 wait_for_completion(&cpu_buffer->update_done);
1841 cpu_buffer->nr_pages_to_update = 0;
1847 * The ring buffer resize can happen with the ring buffer
1848 * enabled, so that the update disturbs the tracing as little
1849 * as possible. But if the buffer is disabled, we do not need
1850 * to worry about that, and we can take the time to verify
1851 * that the buffer is not corrupt.
1853 if (atomic_read(&buffer->record_disabled)) {
1854 atomic_inc(&buffer->record_disabled);
1856 * Even though the buffer was disabled, we must make sure
1857 * that it is truly disabled before calling rb_check_pages.
1858 * There could have been a race between checking
1859 * record_disable and incrementing it.
1862 for_each_buffer_cpu(buffer, cpu) {
1863 cpu_buffer = buffer->buffers[cpu];
1864 rb_check_pages(cpu_buffer);
1866 atomic_dec(&buffer->record_disabled);
1869 mutex_unlock(&buffer->mutex);
1873 for_each_buffer_cpu(buffer, cpu) {
1874 struct buffer_page *bpage, *tmp;
1876 cpu_buffer = buffer->buffers[cpu];
1877 cpu_buffer->nr_pages_to_update = 0;
1879 if (list_empty(&cpu_buffer->new_pages))
1882 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1884 list_del_init(&bpage->list);
1885 free_buffer_page(bpage);
1888 mutex_unlock(&buffer->mutex);
1891 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1893 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1895 mutex_lock(&buffer->mutex);
1897 buffer->flags |= RB_FL_OVERWRITE;
1899 buffer->flags &= ~RB_FL_OVERWRITE;
1900 mutex_unlock(&buffer->mutex);
1902 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1904 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1906 return bpage->page->data + index;
1909 static __always_inline struct ring_buffer_event *
1910 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1912 return __rb_page_index(cpu_buffer->reader_page,
1913 cpu_buffer->reader_page->read);
1916 static __always_inline struct ring_buffer_event *
1917 rb_iter_head_event(struct ring_buffer_iter *iter)
1919 return __rb_page_index(iter->head_page, iter->head);
1922 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1924 return local_read(&bpage->page->commit);
1927 /* Size is determined by what has been committed */
1928 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1930 return rb_page_commit(bpage);
1933 static __always_inline unsigned
1934 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1936 return rb_page_commit(cpu_buffer->commit_page);
1939 static __always_inline unsigned
1940 rb_event_index(struct ring_buffer_event *event)
1942 unsigned long addr = (unsigned long)event;
1944 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1947 static void rb_inc_iter(struct ring_buffer_iter *iter)
1949 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1952 * The iterator could be on the reader page (it starts there).
1953 * But the head could have moved, since the reader was
1954 * found. Check for this case and assign the iterator
1955 * to the head page instead of next.
1957 if (iter->head_page == cpu_buffer->reader_page)
1958 iter->head_page = rb_set_head_page(cpu_buffer);
1960 rb_inc_page(cpu_buffer, &iter->head_page);
1962 iter->read_stamp = iter->head_page->page->time_stamp;
1967 * rb_handle_head_page - writer hit the head page
1969 * Returns: +1 to retry page
1974 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1975 struct buffer_page *tail_page,
1976 struct buffer_page *next_page)
1978 struct buffer_page *new_head;
1983 entries = rb_page_entries(next_page);
1986 * The hard part is here. We need to move the head
1987 * forward, and protect against both readers on
1988 * other CPUs and writers coming in via interrupts.
1990 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1994 * type can be one of four:
1995 * NORMAL - an interrupt already moved it for us
1996 * HEAD - we are the first to get here.
1997 * UPDATE - we are the interrupt interrupting
1999 * MOVED - a reader on another CPU moved the next
2000 * pointer to its reader page. Give up
2007 * We changed the head to UPDATE, thus
2008 * it is our responsibility to update
2011 local_add(entries, &cpu_buffer->overrun);
2012 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2015 * The entries will be zeroed out when we move the
2019 /* still more to do */
2022 case RB_PAGE_UPDATE:
2024 * This is an interrupt that interrupt the
2025 * previous update. Still more to do.
2028 case RB_PAGE_NORMAL:
2030 * An interrupt came in before the update
2031 * and processed this for us.
2032 * Nothing left to do.
2037 * The reader is on another CPU and just did
2038 * a swap with our next_page.
2043 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2048 * Now that we are here, the old head pointer is
2049 * set to UPDATE. This will keep the reader from
2050 * swapping the head page with the reader page.
2051 * The reader (on another CPU) will spin till
2054 * We just need to protect against interrupts
2055 * doing the job. We will set the next pointer
2056 * to HEAD. After that, we set the old pointer
2057 * to NORMAL, but only if it was HEAD before.
2058 * otherwise we are an interrupt, and only
2059 * want the outer most commit to reset it.
2061 new_head = next_page;
2062 rb_inc_page(cpu_buffer, &new_head);
2064 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2068 * Valid returns are:
2069 * HEAD - an interrupt came in and already set it.
2070 * NORMAL - One of two things:
2071 * 1) We really set it.
2072 * 2) A bunch of interrupts came in and moved
2073 * the page forward again.
2077 case RB_PAGE_NORMAL:
2081 RB_WARN_ON(cpu_buffer, 1);
2086 * It is possible that an interrupt came in,
2087 * set the head up, then more interrupts came in
2088 * and moved it again. When we get back here,
2089 * the page would have been set to NORMAL but we
2090 * just set it back to HEAD.
2092 * How do you detect this? Well, if that happened
2093 * the tail page would have moved.
2095 if (ret == RB_PAGE_NORMAL) {
2096 struct buffer_page *buffer_tail_page;
2098 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2100 * If the tail had moved passed next, then we need
2101 * to reset the pointer.
2103 if (buffer_tail_page != tail_page &&
2104 buffer_tail_page != next_page)
2105 rb_head_page_set_normal(cpu_buffer, new_head,
2111 * If this was the outer most commit (the one that
2112 * changed the original pointer from HEAD to UPDATE),
2113 * then it is up to us to reset it to NORMAL.
2115 if (type == RB_PAGE_HEAD) {
2116 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2119 if (RB_WARN_ON(cpu_buffer,
2120 ret != RB_PAGE_UPDATE))
2128 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2129 unsigned long tail, struct rb_event_info *info)
2131 struct buffer_page *tail_page = info->tail_page;
2132 struct ring_buffer_event *event;
2133 unsigned long length = info->length;
2136 * Only the event that crossed the page boundary
2137 * must fill the old tail_page with padding.
2139 if (tail >= BUF_PAGE_SIZE) {
2141 * If the page was filled, then we still need
2142 * to update the real_end. Reset it to zero
2143 * and the reader will ignore it.
2145 if (tail == BUF_PAGE_SIZE)
2146 tail_page->real_end = 0;
2148 local_sub(length, &tail_page->write);
2152 event = __rb_page_index(tail_page, tail);
2154 /* account for padding bytes */
2155 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2158 * Save the original length to the meta data.
2159 * This will be used by the reader to add lost event
2162 tail_page->real_end = tail;
2165 * If this event is bigger than the minimum size, then
2166 * we need to be careful that we don't subtract the
2167 * write counter enough to allow another writer to slip
2169 * We put in a discarded commit instead, to make sure
2170 * that this space is not used again.
2172 * If we are less than the minimum size, we don't need to
2175 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2176 /* No room for any events */
2178 /* Mark the rest of the page with padding */
2179 rb_event_set_padding(event);
2181 /* Set the write back to the previous setting */
2182 local_sub(length, &tail_page->write);
2186 /* Put in a discarded event */
2187 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2188 event->type_len = RINGBUF_TYPE_PADDING;
2189 /* time delta must be non zero */
2190 event->time_delta = 1;
2192 /* Set write to end of buffer */
2193 length = (tail + length) - BUF_PAGE_SIZE;
2194 local_sub(length, &tail_page->write);
2197 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2200 * This is the slow path, force gcc not to inline it.
2202 static noinline struct ring_buffer_event *
2203 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2204 unsigned long tail, struct rb_event_info *info)
2206 struct buffer_page *tail_page = info->tail_page;
2207 struct buffer_page *commit_page = cpu_buffer->commit_page;
2208 struct ring_buffer *buffer = cpu_buffer->buffer;
2209 struct buffer_page *next_page;
2212 next_page = tail_page;
2214 rb_inc_page(cpu_buffer, &next_page);
2217 * If for some reason, we had an interrupt storm that made
2218 * it all the way around the buffer, bail, and warn
2221 if (unlikely(next_page == commit_page)) {
2222 local_inc(&cpu_buffer->commit_overrun);
2227 * This is where the fun begins!
2229 * We are fighting against races between a reader that
2230 * could be on another CPU trying to swap its reader
2231 * page with the buffer head.
2233 * We are also fighting against interrupts coming in and
2234 * moving the head or tail on us as well.
2236 * If the next page is the head page then we have filled
2237 * the buffer, unless the commit page is still on the
2240 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2243 * If the commit is not on the reader page, then
2244 * move the header page.
2246 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2248 * If we are not in overwrite mode,
2249 * this is easy, just stop here.
2251 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2252 local_inc(&cpu_buffer->dropped_events);
2256 ret = rb_handle_head_page(cpu_buffer,
2265 * We need to be careful here too. The
2266 * commit page could still be on the reader
2267 * page. We could have a small buffer, and
2268 * have filled up the buffer with events
2269 * from interrupts and such, and wrapped.
2271 * Note, if the tail page is also the on the
2272 * reader_page, we let it move out.
2274 if (unlikely((cpu_buffer->commit_page !=
2275 cpu_buffer->tail_page) &&
2276 (cpu_buffer->commit_page ==
2277 cpu_buffer->reader_page))) {
2278 local_inc(&cpu_buffer->commit_overrun);
2284 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2288 rb_reset_tail(cpu_buffer, tail, info);
2290 /* Commit what we have for now. */
2291 rb_end_commit(cpu_buffer);
2292 /* rb_end_commit() decs committing */
2293 local_inc(&cpu_buffer->committing);
2295 /* fail and let the caller try again */
2296 return ERR_PTR(-EAGAIN);
2300 rb_reset_tail(cpu_buffer, tail, info);
2305 /* Slow path, do not inline */
2306 static noinline struct ring_buffer_event *
2307 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2310 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2312 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2314 /* Not the first event on the page, or not delta? */
2315 if (abs || rb_event_index(event)) {
2316 event->time_delta = delta & TS_MASK;
2317 event->array[0] = delta >> TS_SHIFT;
2319 /* nope, just zero it */
2320 event->time_delta = 0;
2321 event->array[0] = 0;
2324 return skip_time_extend(event);
2327 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2328 struct ring_buffer_event *event);
2331 * rb_update_event - update event type and data
2332 * @event: the event to update
2333 * @type: the type of event
2334 * @length: the size of the event field in the ring buffer
2336 * Update the type and data fields of the event. The length
2337 * is the actual size that is written to the ring buffer,
2338 * and with this, we can determine what to place into the
2342 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2343 struct ring_buffer_event *event,
2344 struct rb_event_info *info)
2346 unsigned length = info->length;
2347 u64 delta = info->delta;
2349 /* Only a commit updates the timestamp */
2350 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2354 * If we need to add a timestamp, then we
2355 * add it to the start of the reserved space.
2357 if (unlikely(info->add_timestamp)) {
2358 bool abs = ring_buffer_time_stamp_abs(cpu_buffer->buffer);
2360 event = rb_add_time_stamp(event, info->delta, abs);
2361 length -= RB_LEN_TIME_EXTEND;
2365 event->time_delta = delta;
2366 length -= RB_EVNT_HDR_SIZE;
2367 if (length > RB_MAX_SMALL_DATA) {
2368 event->type_len = 0;
2369 event->array[0] = length;
2371 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2374 static unsigned rb_calculate_event_length(unsigned length)
2376 struct ring_buffer_event event; /* Used only for sizeof array */
2378 /* zero length can cause confusions */
2382 if (length > RB_MAX_SMALL_DATA)
2383 length += sizeof(event.array[0]);
2385 length += RB_EVNT_HDR_SIZE;
2386 length = ALIGN(length, RB_ALIGNMENT);
2389 * In case the time delta is larger than the 27 bits for it
2390 * in the header, we need to add a timestamp. If another
2391 * event comes in when trying to discard this one to increase
2392 * the length, then the timestamp will be added in the allocated
2393 * space of this event. If length is bigger than the size needed
2394 * for the TIME_EXTEND, then padding has to be used. The events
2395 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2396 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2397 * As length is a multiple of 4, we only need to worry if it
2398 * is 12 (RB_LEN_TIME_EXTEND + 4).
2400 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2401 length += RB_ALIGNMENT;
2406 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2407 static inline bool sched_clock_stable(void)
2414 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2415 struct ring_buffer_event *event)
2417 unsigned long new_index, old_index;
2418 struct buffer_page *bpage;
2419 unsigned long index;
2422 new_index = rb_event_index(event);
2423 old_index = new_index + rb_event_ts_length(event);
2424 addr = (unsigned long)event;
2427 bpage = READ_ONCE(cpu_buffer->tail_page);
2429 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2430 unsigned long write_mask =
2431 local_read(&bpage->write) & ~RB_WRITE_MASK;
2432 unsigned long event_length = rb_event_length(event);
2434 * This is on the tail page. It is possible that
2435 * a write could come in and move the tail page
2436 * and write to the next page. That is fine
2437 * because we just shorten what is on this page.
2439 old_index += write_mask;
2440 new_index += write_mask;
2441 index = local_cmpxchg(&bpage->write, old_index, new_index);
2442 if (index == old_index) {
2443 /* update counters */
2444 local_sub(event_length, &cpu_buffer->entries_bytes);
2449 /* could not discard */
2453 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2455 local_inc(&cpu_buffer->committing);
2456 local_inc(&cpu_buffer->commits);
2459 static __always_inline void
2460 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2462 unsigned long max_count;
2465 * We only race with interrupts and NMIs on this CPU.
2466 * If we own the commit event, then we can commit
2467 * all others that interrupted us, since the interruptions
2468 * are in stack format (they finish before they come
2469 * back to us). This allows us to do a simple loop to
2470 * assign the commit to the tail.
2473 max_count = cpu_buffer->nr_pages * 100;
2475 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2476 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2478 if (RB_WARN_ON(cpu_buffer,
2479 rb_is_reader_page(cpu_buffer->tail_page)))
2481 local_set(&cpu_buffer->commit_page->page->commit,
2482 rb_page_write(cpu_buffer->commit_page));
2483 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2484 /* Only update the write stamp if the page has an event */
2485 if (rb_page_write(cpu_buffer->commit_page))
2486 cpu_buffer->write_stamp =
2487 cpu_buffer->commit_page->page->time_stamp;
2488 /* add barrier to keep gcc from optimizing too much */
2491 while (rb_commit_index(cpu_buffer) !=
2492 rb_page_write(cpu_buffer->commit_page)) {
2494 local_set(&cpu_buffer->commit_page->page->commit,
2495 rb_page_write(cpu_buffer->commit_page));
2496 RB_WARN_ON(cpu_buffer,
2497 local_read(&cpu_buffer->commit_page->page->commit) &
2502 /* again, keep gcc from optimizing */
2506 * If an interrupt came in just after the first while loop
2507 * and pushed the tail page forward, we will be left with
2508 * a dangling commit that will never go forward.
2510 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2514 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2516 unsigned long commits;
2518 if (RB_WARN_ON(cpu_buffer,
2519 !local_read(&cpu_buffer->committing)))
2523 commits = local_read(&cpu_buffer->commits);
2524 /* synchronize with interrupts */
2526 if (local_read(&cpu_buffer->committing) == 1)
2527 rb_set_commit_to_write(cpu_buffer);
2529 local_dec(&cpu_buffer->committing);
2531 /* synchronize with interrupts */
2535 * Need to account for interrupts coming in between the
2536 * updating of the commit page and the clearing of the
2537 * committing counter.
2539 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2540 !local_read(&cpu_buffer->committing)) {
2541 local_inc(&cpu_buffer->committing);
2546 static inline void rb_event_discard(struct ring_buffer_event *event)
2548 if (extended_time(event))
2549 event = skip_time_extend(event);
2551 /* array[0] holds the actual length for the discarded event */
2552 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2553 event->type_len = RINGBUF_TYPE_PADDING;
2554 /* time delta must be non zero */
2555 if (!event->time_delta)
2556 event->time_delta = 1;
2559 static __always_inline bool
2560 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2561 struct ring_buffer_event *event)
2563 unsigned long addr = (unsigned long)event;
2564 unsigned long index;
2566 index = rb_event_index(event);
2569 return cpu_buffer->commit_page->page == (void *)addr &&
2570 rb_commit_index(cpu_buffer) == index;
2573 static __always_inline void
2574 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2575 struct ring_buffer_event *event)
2580 * The event first in the commit queue updates the
2583 if (rb_event_is_commit(cpu_buffer, event)) {
2585 * A commit event that is first on a page
2586 * updates the write timestamp with the page stamp
2588 if (!rb_event_index(event))
2589 cpu_buffer->write_stamp =
2590 cpu_buffer->commit_page->page->time_stamp;
2591 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2592 delta = ring_buffer_event_time_stamp(event);
2593 cpu_buffer->write_stamp += delta;
2594 } else if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
2595 delta = ring_buffer_event_time_stamp(event);
2596 cpu_buffer->write_stamp = delta;
2598 cpu_buffer->write_stamp += event->time_delta;
2602 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2603 struct ring_buffer_event *event)
2605 local_inc(&cpu_buffer->entries);
2606 rb_update_write_stamp(cpu_buffer, event);
2607 rb_end_commit(cpu_buffer);
2610 static __always_inline void
2611 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2617 if (buffer->irq_work.waiters_pending) {
2618 buffer->irq_work.waiters_pending = false;
2619 /* irq_work_queue() supplies it's own memory barriers */
2620 irq_work_queue(&buffer->irq_work.work);
2623 if (cpu_buffer->irq_work.waiters_pending) {
2624 cpu_buffer->irq_work.waiters_pending = false;
2625 /* irq_work_queue() supplies it's own memory barriers */
2626 irq_work_queue(&cpu_buffer->irq_work.work);
2629 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
2632 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
2635 if (!cpu_buffer->irq_work.full_waiters_pending)
2638 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
2640 full = cpu_buffer->shortest_full;
2641 nr_pages = cpu_buffer->nr_pages;
2642 dirty = ring_buffer_nr_dirty_pages(buffer, cpu_buffer->cpu);
2643 if (full && nr_pages && (dirty * 100) <= full * nr_pages)
2646 cpu_buffer->irq_work.wakeup_full = true;
2647 cpu_buffer->irq_work.full_waiters_pending = false;
2648 /* irq_work_queue() supplies it's own memory barriers */
2649 irq_work_queue(&cpu_buffer->irq_work.work);
2653 * The lock and unlock are done within a preempt disable section.
2654 * The current_context per_cpu variable can only be modified
2655 * by the current task between lock and unlock. But it can
2656 * be modified more than once via an interrupt. To pass this
2657 * information from the lock to the unlock without having to
2658 * access the 'in_interrupt()' functions again (which do show
2659 * a bit of overhead in something as critical as function tracing,
2660 * we use a bitmask trick.
2662 * bit 0 = NMI context
2663 * bit 1 = IRQ context
2664 * bit 2 = SoftIRQ context
2665 * bit 3 = normal context.
2667 * This works because this is the order of contexts that can
2668 * preempt other contexts. A SoftIRQ never preempts an IRQ
2671 * When the context is determined, the corresponding bit is
2672 * checked and set (if it was set, then a recursion of that context
2675 * On unlock, we need to clear this bit. To do so, just subtract
2676 * 1 from the current_context and AND it to itself.
2680 * 101 & 100 = 100 (clearing bit zero)
2683 * 1010 & 1001 = 1000 (clearing bit 1)
2685 * The least significant bit can be cleared this way, and it
2686 * just so happens that it is the same bit corresponding to
2687 * the current context.
2690 static __always_inline int
2691 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2693 unsigned int val = cpu_buffer->current_context;
2694 unsigned long pc = preempt_count();
2697 if (!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
2698 bit = RB_CTX_NORMAL;
2700 bit = pc & NMI_MASK ? RB_CTX_NMI :
2701 pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
2703 if (unlikely(val & (1 << (bit + cpu_buffer->nest))))
2706 val |= (1 << (bit + cpu_buffer->nest));
2707 cpu_buffer->current_context = val;
2712 static __always_inline void
2713 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2715 cpu_buffer->current_context &=
2716 cpu_buffer->current_context - (1 << cpu_buffer->nest);
2719 /* The recursive locking above uses 4 bits */
2720 #define NESTED_BITS 4
2723 * ring_buffer_nest_start - Allow to trace while nested
2724 * @buffer: The ring buffer to modify
2726 * The ring buffer has a safety mechanism to prevent recursion.
2727 * But there may be a case where a trace needs to be done while
2728 * tracing something else. In this case, calling this function
2729 * will allow this function to nest within a currently active
2730 * ring_buffer_lock_reserve().
2732 * Call this function before calling another ring_buffer_lock_reserve() and
2733 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
2735 void ring_buffer_nest_start(struct ring_buffer *buffer)
2737 struct ring_buffer_per_cpu *cpu_buffer;
2740 /* Enabled by ring_buffer_nest_end() */
2741 preempt_disable_notrace();
2742 cpu = raw_smp_processor_id();
2743 cpu_buffer = buffer->buffers[cpu];
2744 /* This is the shift value for the above recursive locking */
2745 cpu_buffer->nest += NESTED_BITS;
2749 * ring_buffer_nest_end - Allow to trace while nested
2750 * @buffer: The ring buffer to modify
2752 * Must be called after ring_buffer_nest_start() and after the
2753 * ring_buffer_unlock_commit().
2755 void ring_buffer_nest_end(struct ring_buffer *buffer)
2757 struct ring_buffer_per_cpu *cpu_buffer;
2760 /* disabled by ring_buffer_nest_start() */
2761 cpu = raw_smp_processor_id();
2762 cpu_buffer = buffer->buffers[cpu];
2763 /* This is the shift value for the above recursive locking */
2764 cpu_buffer->nest -= NESTED_BITS;
2765 preempt_enable_notrace();
2769 * ring_buffer_unlock_commit - commit a reserved
2770 * @buffer: The buffer to commit to
2771 * @event: The event pointer to commit.
2773 * This commits the data to the ring buffer, and releases any locks held.
2775 * Must be paired with ring_buffer_lock_reserve.
2777 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2778 struct ring_buffer_event *event)
2780 struct ring_buffer_per_cpu *cpu_buffer;
2781 int cpu = raw_smp_processor_id();
2783 cpu_buffer = buffer->buffers[cpu];
2785 rb_commit(cpu_buffer, event);
2787 rb_wakeups(buffer, cpu_buffer);
2789 trace_recursive_unlock(cpu_buffer);
2791 preempt_enable_notrace();
2795 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2797 static noinline void
2798 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2799 struct rb_event_info *info)
2801 WARN_ONCE(info->delta > (1ULL << 59),
2802 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2803 (unsigned long long)info->delta,
2804 (unsigned long long)info->ts,
2805 (unsigned long long)cpu_buffer->write_stamp,
2806 sched_clock_stable() ? "" :
2807 "If you just came from a suspend/resume,\n"
2808 "please switch to the trace global clock:\n"
2809 " echo global > /sys/kernel/debug/tracing/trace_clock\n"
2810 "or add trace_clock=global to the kernel command line\n");
2811 info->add_timestamp = 1;
2814 static struct ring_buffer_event *
2815 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2816 struct rb_event_info *info)
2818 struct ring_buffer_event *event;
2819 struct buffer_page *tail_page;
2820 unsigned long tail, write;
2823 * If the time delta since the last event is too big to
2824 * hold in the time field of the event, then we append a
2825 * TIME EXTEND event ahead of the data event.
2827 if (unlikely(info->add_timestamp))
2828 info->length += RB_LEN_TIME_EXTEND;
2830 /* Don't let the compiler play games with cpu_buffer->tail_page */
2831 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2832 write = local_add_return(info->length, &tail_page->write);
2834 /* set write to only the index of the write */
2835 write &= RB_WRITE_MASK;
2836 tail = write - info->length;
2839 * If this is the first commit on the page, then it has the same
2840 * timestamp as the page itself.
2842 if (!tail && !ring_buffer_time_stamp_abs(cpu_buffer->buffer))
2845 /* See if we shot pass the end of this buffer page */
2846 if (unlikely(write > BUF_PAGE_SIZE))
2847 return rb_move_tail(cpu_buffer, tail, info);
2849 /* We reserved something on the buffer */
2851 event = __rb_page_index(tail_page, tail);
2852 rb_update_event(cpu_buffer, event, info);
2854 local_inc(&tail_page->entries);
2857 * If this is the first commit on the page, then update
2861 tail_page->page->time_stamp = info->ts;
2863 /* account for these added bytes */
2864 local_add(info->length, &cpu_buffer->entries_bytes);
2869 static __always_inline struct ring_buffer_event *
2870 rb_reserve_next_event(struct ring_buffer *buffer,
2871 struct ring_buffer_per_cpu *cpu_buffer,
2872 unsigned long length)
2874 struct ring_buffer_event *event;
2875 struct rb_event_info info;
2879 rb_start_commit(cpu_buffer);
2881 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2883 * Due to the ability to swap a cpu buffer from a buffer
2884 * it is possible it was swapped before we committed.
2885 * (committing stops a swap). We check for it here and
2886 * if it happened, we have to fail the write.
2889 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
2890 local_dec(&cpu_buffer->committing);
2891 local_dec(&cpu_buffer->commits);
2896 info.length = rb_calculate_event_length(length);
2898 info.add_timestamp = 0;
2902 * We allow for interrupts to reenter here and do a trace.
2903 * If one does, it will cause this original code to loop
2904 * back here. Even with heavy interrupts happening, this
2905 * should only happen a few times in a row. If this happens
2906 * 1000 times in a row, there must be either an interrupt
2907 * storm or we have something buggy.
2910 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2913 info.ts = rb_time_stamp(cpu_buffer->buffer);
2914 diff = info.ts - cpu_buffer->write_stamp;
2916 /* make sure this diff is calculated here */
2919 if (ring_buffer_time_stamp_abs(buffer)) {
2920 info.delta = info.ts;
2921 rb_handle_timestamp(cpu_buffer, &info);
2922 } else /* Did the write stamp get updated already? */
2923 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2925 if (unlikely(test_time_stamp(info.delta)))
2926 rb_handle_timestamp(cpu_buffer, &info);
2929 event = __rb_reserve_next(cpu_buffer, &info);
2931 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2932 if (info.add_timestamp)
2933 info.length -= RB_LEN_TIME_EXTEND;
2943 rb_end_commit(cpu_buffer);
2948 * ring_buffer_lock_reserve - reserve a part of the buffer
2949 * @buffer: the ring buffer to reserve from
2950 * @length: the length of the data to reserve (excluding event header)
2952 * Returns a reserved event on the ring buffer to copy directly to.
2953 * The user of this interface will need to get the body to write into
2954 * and can use the ring_buffer_event_data() interface.
2956 * The length is the length of the data needed, not the event length
2957 * which also includes the event header.
2959 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2960 * If NULL is returned, then nothing has been allocated or locked.
2962 struct ring_buffer_event *
2963 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2965 struct ring_buffer_per_cpu *cpu_buffer;
2966 struct ring_buffer_event *event;
2969 /* If we are tracing schedule, we don't want to recurse */
2970 preempt_disable_notrace();
2972 if (unlikely(atomic_read(&buffer->record_disabled)))
2975 cpu = raw_smp_processor_id();
2977 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2980 cpu_buffer = buffer->buffers[cpu];
2982 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2985 if (unlikely(length > BUF_MAX_DATA_SIZE))
2988 if (unlikely(trace_recursive_lock(cpu_buffer)))
2991 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2998 trace_recursive_unlock(cpu_buffer);
3000 preempt_enable_notrace();
3003 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3006 * Decrement the entries to the page that an event is on.
3007 * The event does not even need to exist, only the pointer
3008 * to the page it is on. This may only be called before the commit
3012 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3013 struct ring_buffer_event *event)
3015 unsigned long addr = (unsigned long)event;
3016 struct buffer_page *bpage = cpu_buffer->commit_page;
3017 struct buffer_page *start;
3021 /* Do the likely case first */
3022 if (likely(bpage->page == (void *)addr)) {
3023 local_dec(&bpage->entries);
3028 * Because the commit page may be on the reader page we
3029 * start with the next page and check the end loop there.
3031 rb_inc_page(cpu_buffer, &bpage);
3034 if (bpage->page == (void *)addr) {
3035 local_dec(&bpage->entries);
3038 rb_inc_page(cpu_buffer, &bpage);
3039 } while (bpage != start);
3041 /* commit not part of this buffer?? */
3042 RB_WARN_ON(cpu_buffer, 1);
3046 * ring_buffer_commit_discard - discard an event that has not been committed
3047 * @buffer: the ring buffer
3048 * @event: non committed event to discard
3050 * Sometimes an event that is in the ring buffer needs to be ignored.
3051 * This function lets the user discard an event in the ring buffer
3052 * and then that event will not be read later.
3054 * This function only works if it is called before the item has been
3055 * committed. It will try to free the event from the ring buffer
3056 * if another event has not been added behind it.
3058 * If another event has been added behind it, it will set the event
3059 * up as discarded, and perform the commit.
3061 * If this function is called, do not call ring_buffer_unlock_commit on
3064 void ring_buffer_discard_commit(struct ring_buffer *buffer,
3065 struct ring_buffer_event *event)
3067 struct ring_buffer_per_cpu *cpu_buffer;
3070 /* The event is discarded regardless */
3071 rb_event_discard(event);
3073 cpu = smp_processor_id();
3074 cpu_buffer = buffer->buffers[cpu];
3077 * This must only be called if the event has not been
3078 * committed yet. Thus we can assume that preemption
3079 * is still disabled.
3081 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3083 rb_decrement_entry(cpu_buffer, event);
3084 if (rb_try_to_discard(cpu_buffer, event))
3088 * The commit is still visible by the reader, so we
3089 * must still update the timestamp.
3091 rb_update_write_stamp(cpu_buffer, event);
3093 rb_end_commit(cpu_buffer);
3095 trace_recursive_unlock(cpu_buffer);
3097 preempt_enable_notrace();
3100 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3103 * ring_buffer_write - write data to the buffer without reserving
3104 * @buffer: The ring buffer to write to.
3105 * @length: The length of the data being written (excluding the event header)
3106 * @data: The data to write to the buffer.
3108 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3109 * one function. If you already have the data to write to the buffer, it
3110 * may be easier to simply call this function.
3112 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3113 * and not the length of the event which would hold the header.
3115 int ring_buffer_write(struct ring_buffer *buffer,
3116 unsigned long length,
3119 struct ring_buffer_per_cpu *cpu_buffer;
3120 struct ring_buffer_event *event;
3125 preempt_disable_notrace();
3127 if (atomic_read(&buffer->record_disabled))
3130 cpu = raw_smp_processor_id();
3132 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3135 cpu_buffer = buffer->buffers[cpu];
3137 if (atomic_read(&cpu_buffer->record_disabled))
3140 if (length > BUF_MAX_DATA_SIZE)
3143 if (unlikely(trace_recursive_lock(cpu_buffer)))
3146 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3150 body = rb_event_data(event);
3152 memcpy(body, data, length);
3154 rb_commit(cpu_buffer, event);
3156 rb_wakeups(buffer, cpu_buffer);
3161 trace_recursive_unlock(cpu_buffer);
3164 preempt_enable_notrace();
3168 EXPORT_SYMBOL_GPL(ring_buffer_write);
3170 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3172 struct buffer_page *reader = cpu_buffer->reader_page;
3173 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3174 struct buffer_page *commit = cpu_buffer->commit_page;
3176 /* In case of error, head will be NULL */
3177 if (unlikely(!head))
3180 return reader->read == rb_page_commit(reader) &&
3181 (commit == reader ||
3183 head->read == rb_page_commit(commit)));
3187 * ring_buffer_record_disable - stop all writes into the buffer
3188 * @buffer: The ring buffer to stop writes to.
3190 * This prevents all writes to the buffer. Any attempt to write
3191 * to the buffer after this will fail and return NULL.
3193 * The caller should call synchronize_rcu() after this.
3195 void ring_buffer_record_disable(struct ring_buffer *buffer)
3197 atomic_inc(&buffer->record_disabled);
3199 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3202 * ring_buffer_record_enable - enable writes to the buffer
3203 * @buffer: The ring buffer to enable writes
3205 * Note, multiple disables will need the same number of enables
3206 * to truly enable the writing (much like preempt_disable).
3208 void ring_buffer_record_enable(struct ring_buffer *buffer)
3210 atomic_dec(&buffer->record_disabled);
3212 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3215 * ring_buffer_record_off - stop all writes into the buffer
3216 * @buffer: The ring buffer to stop writes to.
3218 * This prevents all writes to the buffer. Any attempt to write
3219 * to the buffer after this will fail and return NULL.
3221 * This is different than ring_buffer_record_disable() as
3222 * it works like an on/off switch, where as the disable() version
3223 * must be paired with a enable().
3225 void ring_buffer_record_off(struct ring_buffer *buffer)
3228 unsigned int new_rd;
3231 rd = atomic_read(&buffer->record_disabled);
3232 new_rd = rd | RB_BUFFER_OFF;
3233 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3235 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3238 * ring_buffer_record_on - restart writes into the buffer
3239 * @buffer: The ring buffer to start writes to.
3241 * This enables all writes to the buffer that was disabled by
3242 * ring_buffer_record_off().
3244 * This is different than ring_buffer_record_enable() as
3245 * it works like an on/off switch, where as the enable() version
3246 * must be paired with a disable().
3248 void ring_buffer_record_on(struct ring_buffer *buffer)
3251 unsigned int new_rd;
3254 rd = atomic_read(&buffer->record_disabled);
3255 new_rd = rd & ~RB_BUFFER_OFF;
3256 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3258 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3261 * ring_buffer_record_is_on - return true if the ring buffer can write
3262 * @buffer: The ring buffer to see if write is enabled
3264 * Returns true if the ring buffer is in a state that it accepts writes.
3266 bool ring_buffer_record_is_on(struct ring_buffer *buffer)
3268 return !atomic_read(&buffer->record_disabled);
3272 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3273 * @buffer: The ring buffer to see if write is set enabled
3275 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3276 * Note that this does NOT mean it is in a writable state.
3278 * It may return true when the ring buffer has been disabled by
3279 * ring_buffer_record_disable(), as that is a temporary disabling of
3282 bool ring_buffer_record_is_set_on(struct ring_buffer *buffer)
3284 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3288 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3289 * @buffer: The ring buffer to stop writes to.
3290 * @cpu: The CPU buffer to stop
3292 * This prevents all writes to the buffer. Any attempt to write
3293 * to the buffer after this will fail and return NULL.
3295 * The caller should call synchronize_rcu() after this.
3297 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3299 struct ring_buffer_per_cpu *cpu_buffer;
3301 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3304 cpu_buffer = buffer->buffers[cpu];
3305 atomic_inc(&cpu_buffer->record_disabled);
3307 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3310 * ring_buffer_record_enable_cpu - enable writes to the buffer
3311 * @buffer: The ring buffer to enable writes
3312 * @cpu: The CPU to enable.
3314 * Note, multiple disables will need the same number of enables
3315 * to truly enable the writing (much like preempt_disable).
3317 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3319 struct ring_buffer_per_cpu *cpu_buffer;
3321 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3324 cpu_buffer = buffer->buffers[cpu];
3325 atomic_dec(&cpu_buffer->record_disabled);
3327 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3330 * The total entries in the ring buffer is the running counter
3331 * of entries entered into the ring buffer, minus the sum of
3332 * the entries read from the ring buffer and the number of
3333 * entries that were overwritten.
3335 static inline unsigned long
3336 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3338 return local_read(&cpu_buffer->entries) -
3339 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3343 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3344 * @buffer: The ring buffer
3345 * @cpu: The per CPU buffer to read from.
3347 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3349 unsigned long flags;
3350 struct ring_buffer_per_cpu *cpu_buffer;
3351 struct buffer_page *bpage;
3354 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3357 cpu_buffer = buffer->buffers[cpu];
3358 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3360 * if the tail is on reader_page, oldest time stamp is on the reader
3363 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3364 bpage = cpu_buffer->reader_page;
3366 bpage = rb_set_head_page(cpu_buffer);
3368 ret = bpage->page->time_stamp;
3369 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3373 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3376 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3377 * @buffer: The ring buffer
3378 * @cpu: The per CPU buffer to read from.
3380 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3382 struct ring_buffer_per_cpu *cpu_buffer;
3385 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3388 cpu_buffer = buffer->buffers[cpu];
3389 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3393 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3396 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3397 * @buffer: The ring buffer
3398 * @cpu: The per CPU buffer to get the entries from.
3400 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3402 struct ring_buffer_per_cpu *cpu_buffer;
3404 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3407 cpu_buffer = buffer->buffers[cpu];
3409 return rb_num_of_entries(cpu_buffer);
3411 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3414 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3415 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3416 * @buffer: The ring buffer
3417 * @cpu: The per CPU buffer to get the number of overruns from
3419 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3421 struct ring_buffer_per_cpu *cpu_buffer;
3424 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3427 cpu_buffer = buffer->buffers[cpu];
3428 ret = local_read(&cpu_buffer->overrun);
3432 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3435 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3436 * commits failing due to the buffer wrapping around while there are uncommitted
3437 * events, such as during an interrupt storm.
3438 * @buffer: The ring buffer
3439 * @cpu: The per CPU buffer to get the number of overruns from
3442 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3444 struct ring_buffer_per_cpu *cpu_buffer;
3447 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3450 cpu_buffer = buffer->buffers[cpu];
3451 ret = local_read(&cpu_buffer->commit_overrun);
3455 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3458 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3459 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3460 * @buffer: The ring buffer
3461 * @cpu: The per CPU buffer to get the number of overruns from
3464 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3466 struct ring_buffer_per_cpu *cpu_buffer;
3469 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3472 cpu_buffer = buffer->buffers[cpu];
3473 ret = local_read(&cpu_buffer->dropped_events);
3477 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3480 * ring_buffer_read_events_cpu - get the number of events successfully read
3481 * @buffer: The ring buffer
3482 * @cpu: The per CPU buffer to get the number of events read
3485 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3487 struct ring_buffer_per_cpu *cpu_buffer;
3489 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3492 cpu_buffer = buffer->buffers[cpu];
3493 return cpu_buffer->read;
3495 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3498 * ring_buffer_entries - get the number of entries in a buffer
3499 * @buffer: The ring buffer
3501 * Returns the total number of entries in the ring buffer
3504 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3506 struct ring_buffer_per_cpu *cpu_buffer;
3507 unsigned long entries = 0;
3510 /* if you care about this being correct, lock the buffer */
3511 for_each_buffer_cpu(buffer, cpu) {
3512 cpu_buffer = buffer->buffers[cpu];
3513 entries += rb_num_of_entries(cpu_buffer);
3518 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3521 * ring_buffer_overruns - get the number of overruns in buffer
3522 * @buffer: The ring buffer
3524 * Returns the total number of overruns in the ring buffer
3527 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3529 struct ring_buffer_per_cpu *cpu_buffer;
3530 unsigned long overruns = 0;
3533 /* if you care about this being correct, lock the buffer */
3534 for_each_buffer_cpu(buffer, cpu) {
3535 cpu_buffer = buffer->buffers[cpu];
3536 overruns += local_read(&cpu_buffer->overrun);
3541 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3543 static void rb_iter_reset(struct ring_buffer_iter *iter)
3545 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3547 /* Iterator usage is expected to have record disabled */
3548 iter->head_page = cpu_buffer->reader_page;
3549 iter->head = cpu_buffer->reader_page->read;
3551 iter->cache_reader_page = iter->head_page;
3552 iter->cache_read = cpu_buffer->read;
3555 iter->read_stamp = cpu_buffer->read_stamp;
3557 iter->read_stamp = iter->head_page->page->time_stamp;
3561 * ring_buffer_iter_reset - reset an iterator
3562 * @iter: The iterator to reset
3564 * Resets the iterator, so that it will start from the beginning
3567 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3569 struct ring_buffer_per_cpu *cpu_buffer;
3570 unsigned long flags;
3575 cpu_buffer = iter->cpu_buffer;
3577 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3578 rb_iter_reset(iter);
3579 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3581 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3584 * ring_buffer_iter_empty - check if an iterator has no more to read
3585 * @iter: The iterator to check
3587 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3589 struct ring_buffer_per_cpu *cpu_buffer;
3590 struct buffer_page *reader;
3591 struct buffer_page *head_page;
3592 struct buffer_page *commit_page;
3595 cpu_buffer = iter->cpu_buffer;
3597 /* Remember, trace recording is off when iterator is in use */
3598 reader = cpu_buffer->reader_page;
3599 head_page = cpu_buffer->head_page;
3600 commit_page = cpu_buffer->commit_page;
3601 commit = rb_page_commit(commit_page);
3603 return ((iter->head_page == commit_page && iter->head == commit) ||
3604 (iter->head_page == reader && commit_page == head_page &&
3605 head_page->read == commit &&
3606 iter->head == rb_page_commit(cpu_buffer->reader_page)));
3608 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3611 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3612 struct ring_buffer_event *event)
3616 switch (event->type_len) {
3617 case RINGBUF_TYPE_PADDING:
3620 case RINGBUF_TYPE_TIME_EXTEND:
3621 delta = ring_buffer_event_time_stamp(event);
3622 cpu_buffer->read_stamp += delta;
3625 case RINGBUF_TYPE_TIME_STAMP:
3626 delta = ring_buffer_event_time_stamp(event);
3627 cpu_buffer->read_stamp = delta;
3630 case RINGBUF_TYPE_DATA:
3631 cpu_buffer->read_stamp += event->time_delta;
3641 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3642 struct ring_buffer_event *event)
3646 switch (event->type_len) {
3647 case RINGBUF_TYPE_PADDING:
3650 case RINGBUF_TYPE_TIME_EXTEND:
3651 delta = ring_buffer_event_time_stamp(event);
3652 iter->read_stamp += delta;
3655 case RINGBUF_TYPE_TIME_STAMP:
3656 delta = ring_buffer_event_time_stamp(event);
3657 iter->read_stamp = delta;
3660 case RINGBUF_TYPE_DATA:
3661 iter->read_stamp += event->time_delta;
3670 static struct buffer_page *
3671 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3673 struct buffer_page *reader = NULL;
3674 unsigned long overwrite;
3675 unsigned long flags;
3679 local_irq_save(flags);
3680 arch_spin_lock(&cpu_buffer->lock);
3684 * This should normally only loop twice. But because the
3685 * start of the reader inserts an empty page, it causes
3686 * a case where we will loop three times. There should be no
3687 * reason to loop four times (that I know of).
3689 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3694 reader = cpu_buffer->reader_page;
3696 /* If there's more to read, return this page */
3697 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3700 /* Never should we have an index greater than the size */
3701 if (RB_WARN_ON(cpu_buffer,
3702 cpu_buffer->reader_page->read > rb_page_size(reader)))
3705 /* check if we caught up to the tail */
3707 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3710 /* Don't bother swapping if the ring buffer is empty */
3711 if (rb_num_of_entries(cpu_buffer) == 0)
3715 * Reset the reader page to size zero.
3717 local_set(&cpu_buffer->reader_page->write, 0);
3718 local_set(&cpu_buffer->reader_page->entries, 0);
3719 local_set(&cpu_buffer->reader_page->page->commit, 0);
3720 cpu_buffer->reader_page->real_end = 0;
3724 * Splice the empty reader page into the list around the head.
3726 reader = rb_set_head_page(cpu_buffer);
3729 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3730 cpu_buffer->reader_page->list.prev = reader->list.prev;
3733 * cpu_buffer->pages just needs to point to the buffer, it
3734 * has no specific buffer page to point to. Lets move it out
3735 * of our way so we don't accidentally swap it.
3737 cpu_buffer->pages = reader->list.prev;
3739 /* The reader page will be pointing to the new head */
3740 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3743 * We want to make sure we read the overruns after we set up our
3744 * pointers to the next object. The writer side does a
3745 * cmpxchg to cross pages which acts as the mb on the writer
3746 * side. Note, the reader will constantly fail the swap
3747 * while the writer is updating the pointers, so this
3748 * guarantees that the overwrite recorded here is the one we
3749 * want to compare with the last_overrun.
3752 overwrite = local_read(&(cpu_buffer->overrun));
3755 * Here's the tricky part.
3757 * We need to move the pointer past the header page.
3758 * But we can only do that if a writer is not currently
3759 * moving it. The page before the header page has the
3760 * flag bit '1' set if it is pointing to the page we want.
3761 * but if the writer is in the process of moving it
3762 * than it will be '2' or already moved '0'.
3765 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3768 * If we did not convert it, then we must try again.
3774 * Yay! We succeeded in replacing the page.
3776 * Now make the new head point back to the reader page.
3778 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3779 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3781 local_inc(&cpu_buffer->pages_read);
3783 /* Finally update the reader page to the new head */
3784 cpu_buffer->reader_page = reader;
3785 cpu_buffer->reader_page->read = 0;
3787 if (overwrite != cpu_buffer->last_overrun) {
3788 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3789 cpu_buffer->last_overrun = overwrite;
3795 /* Update the read_stamp on the first event */
3796 if (reader && reader->read == 0)
3797 cpu_buffer->read_stamp = reader->page->time_stamp;
3799 arch_spin_unlock(&cpu_buffer->lock);
3800 local_irq_restore(flags);
3805 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3807 struct ring_buffer_event *event;
3808 struct buffer_page *reader;
3811 reader = rb_get_reader_page(cpu_buffer);
3813 /* This function should not be called when buffer is empty */
3814 if (RB_WARN_ON(cpu_buffer, !reader))
3817 event = rb_reader_event(cpu_buffer);
3819 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3822 rb_update_read_stamp(cpu_buffer, event);
3824 length = rb_event_length(event);
3825 cpu_buffer->reader_page->read += length;
3828 static void rb_advance_iter(struct ring_buffer_iter *iter)
3830 struct ring_buffer_per_cpu *cpu_buffer;
3831 struct ring_buffer_event *event;
3834 cpu_buffer = iter->cpu_buffer;
3837 * Check if we are at the end of the buffer.
3839 if (iter->head >= rb_page_size(iter->head_page)) {
3840 /* discarded commits can make the page empty */
3841 if (iter->head_page == cpu_buffer->commit_page)
3847 event = rb_iter_head_event(iter);
3849 length = rb_event_length(event);
3852 * This should not be called to advance the header if we are
3853 * at the tail of the buffer.
3855 if (RB_WARN_ON(cpu_buffer,
3856 (iter->head_page == cpu_buffer->commit_page) &&
3857 (iter->head + length > rb_commit_index(cpu_buffer))))
3860 rb_update_iter_read_stamp(iter, event);
3862 iter->head += length;
3864 /* check for end of page padding */
3865 if ((iter->head >= rb_page_size(iter->head_page)) &&
3866 (iter->head_page != cpu_buffer->commit_page))
3870 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3872 return cpu_buffer->lost_events;
3875 static struct ring_buffer_event *
3876 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3877 unsigned long *lost_events)
3879 struct ring_buffer_event *event;
3880 struct buffer_page *reader;
3887 * We repeat when a time extend is encountered.
3888 * Since the time extend is always attached to a data event,
3889 * we should never loop more than once.
3890 * (We never hit the following condition more than twice).
3892 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3895 reader = rb_get_reader_page(cpu_buffer);
3899 event = rb_reader_event(cpu_buffer);
3901 switch (event->type_len) {
3902 case RINGBUF_TYPE_PADDING:
3903 if (rb_null_event(event))
3904 RB_WARN_ON(cpu_buffer, 1);
3906 * Because the writer could be discarding every
3907 * event it creates (which would probably be bad)
3908 * if we were to go back to "again" then we may never
3909 * catch up, and will trigger the warn on, or lock
3910 * the box. Return the padding, and we will release
3911 * the current locks, and try again.
3915 case RINGBUF_TYPE_TIME_EXTEND:
3916 /* Internal data, OK to advance */
3917 rb_advance_reader(cpu_buffer);
3920 case RINGBUF_TYPE_TIME_STAMP:
3922 *ts = ring_buffer_event_time_stamp(event);
3923 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3924 cpu_buffer->cpu, ts);
3926 /* Internal data, OK to advance */
3927 rb_advance_reader(cpu_buffer);
3930 case RINGBUF_TYPE_DATA:
3932 *ts = cpu_buffer->read_stamp + event->time_delta;
3933 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3934 cpu_buffer->cpu, ts);
3937 *lost_events = rb_lost_events(cpu_buffer);
3946 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3948 static struct ring_buffer_event *
3949 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3951 struct ring_buffer *buffer;
3952 struct ring_buffer_per_cpu *cpu_buffer;
3953 struct ring_buffer_event *event;
3959 cpu_buffer = iter->cpu_buffer;
3960 buffer = cpu_buffer->buffer;
3963 * Check if someone performed a consuming read to
3964 * the buffer. A consuming read invalidates the iterator
3965 * and we need to reset the iterator in this case.
3967 if (unlikely(iter->cache_read != cpu_buffer->read ||
3968 iter->cache_reader_page != cpu_buffer->reader_page))
3969 rb_iter_reset(iter);
3972 if (ring_buffer_iter_empty(iter))
3976 * We repeat when a time extend is encountered or we hit
3977 * the end of the page. Since the time extend is always attached
3978 * to a data event, we should never loop more than three times.
3979 * Once for going to next page, once on time extend, and
3980 * finally once to get the event.
3981 * (We never hit the following condition more than thrice).
3983 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3986 if (rb_per_cpu_empty(cpu_buffer))
3989 if (iter->head >= rb_page_size(iter->head_page)) {
3994 event = rb_iter_head_event(iter);
3996 switch (event->type_len) {
3997 case RINGBUF_TYPE_PADDING:
3998 if (rb_null_event(event)) {
4002 rb_advance_iter(iter);
4005 case RINGBUF_TYPE_TIME_EXTEND:
4006 /* Internal data, OK to advance */
4007 rb_advance_iter(iter);
4010 case RINGBUF_TYPE_TIME_STAMP:
4012 *ts = ring_buffer_event_time_stamp(event);
4013 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4014 cpu_buffer->cpu, ts);
4016 /* Internal data, OK to advance */
4017 rb_advance_iter(iter);
4020 case RINGBUF_TYPE_DATA:
4022 *ts = iter->read_stamp + event->time_delta;
4023 ring_buffer_normalize_time_stamp(buffer,
4024 cpu_buffer->cpu, ts);
4034 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4036 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4038 if (likely(!in_nmi())) {
4039 raw_spin_lock(&cpu_buffer->reader_lock);
4044 * If an NMI die dumps out the content of the ring buffer
4045 * trylock must be used to prevent a deadlock if the NMI
4046 * preempted a task that holds the ring buffer locks. If
4047 * we get the lock then all is fine, if not, then continue
4048 * to do the read, but this can corrupt the ring buffer,
4049 * so it must be permanently disabled from future writes.
4050 * Reading from NMI is a oneshot deal.
4052 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4055 /* Continue without locking, but disable the ring buffer */
4056 atomic_inc(&cpu_buffer->record_disabled);
4061 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4064 raw_spin_unlock(&cpu_buffer->reader_lock);
4069 * ring_buffer_peek - peek at the next event to be read
4070 * @buffer: The ring buffer to read
4071 * @cpu: The cpu to peak at
4072 * @ts: The timestamp counter of this event.
4073 * @lost_events: a variable to store if events were lost (may be NULL)
4075 * This will return the event that will be read next, but does
4076 * not consume the data.
4078 struct ring_buffer_event *
4079 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
4080 unsigned long *lost_events)
4082 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4083 struct ring_buffer_event *event;
4084 unsigned long flags;
4087 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4091 local_irq_save(flags);
4092 dolock = rb_reader_lock(cpu_buffer);
4093 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4094 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4095 rb_advance_reader(cpu_buffer);
4096 rb_reader_unlock(cpu_buffer, dolock);
4097 local_irq_restore(flags);
4099 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4106 * ring_buffer_iter_peek - peek at the next event to be read
4107 * @iter: The ring buffer iterator
4108 * @ts: The timestamp counter of this event.
4110 * This will return the event that will be read next, but does
4111 * not increment the iterator.
4113 struct ring_buffer_event *
4114 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4116 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4117 struct ring_buffer_event *event;
4118 unsigned long flags;
4121 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4122 event = rb_iter_peek(iter, ts);
4123 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4125 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4132 * ring_buffer_consume - return an event and consume it
4133 * @buffer: The ring buffer to get the next event from
4134 * @cpu: the cpu to read the buffer from
4135 * @ts: a variable to store the timestamp (may be NULL)
4136 * @lost_events: a variable to store if events were lost (may be NULL)
4138 * Returns the next event in the ring buffer, and that event is consumed.
4139 * Meaning, that sequential reads will keep returning a different event,
4140 * and eventually empty the ring buffer if the producer is slower.
4142 struct ring_buffer_event *
4143 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
4144 unsigned long *lost_events)
4146 struct ring_buffer_per_cpu *cpu_buffer;
4147 struct ring_buffer_event *event = NULL;
4148 unsigned long flags;
4152 /* might be called in atomic */
4155 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4158 cpu_buffer = buffer->buffers[cpu];
4159 local_irq_save(flags);
4160 dolock = rb_reader_lock(cpu_buffer);
4162 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4164 cpu_buffer->lost_events = 0;
4165 rb_advance_reader(cpu_buffer);
4168 rb_reader_unlock(cpu_buffer, dolock);
4169 local_irq_restore(flags);
4174 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4179 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4182 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4183 * @buffer: The ring buffer to read from
4184 * @cpu: The cpu buffer to iterate over
4185 * @flags: gfp flags to use for memory allocation
4187 * This performs the initial preparations necessary to iterate
4188 * through the buffer. Memory is allocated, buffer recording
4189 * is disabled, and the iterator pointer is returned to the caller.
4191 * Disabling buffer recording prevents the reading from being
4192 * corrupted. This is not a consuming read, so a producer is not
4195 * After a sequence of ring_buffer_read_prepare calls, the user is
4196 * expected to make at least one call to ring_buffer_read_prepare_sync.
4197 * Afterwards, ring_buffer_read_start is invoked to get things going
4200 * This overall must be paired with ring_buffer_read_finish.
4202 struct ring_buffer_iter *
4203 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu, gfp_t flags)
4205 struct ring_buffer_per_cpu *cpu_buffer;
4206 struct ring_buffer_iter *iter;
4208 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4211 iter = kmalloc(sizeof(*iter), flags);
4215 cpu_buffer = buffer->buffers[cpu];
4217 iter->cpu_buffer = cpu_buffer;
4219 atomic_inc(&buffer->resize_disabled);
4220 atomic_inc(&cpu_buffer->record_disabled);
4224 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4227 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4229 * All previously invoked ring_buffer_read_prepare calls to prepare
4230 * iterators will be synchronized. Afterwards, read_buffer_read_start
4231 * calls on those iterators are allowed.
4234 ring_buffer_read_prepare_sync(void)
4238 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4241 * ring_buffer_read_start - start a non consuming read of the buffer
4242 * @iter: The iterator returned by ring_buffer_read_prepare
4244 * This finalizes the startup of an iteration through the buffer.
4245 * The iterator comes from a call to ring_buffer_read_prepare and
4246 * an intervening ring_buffer_read_prepare_sync must have been
4249 * Must be paired with ring_buffer_read_finish.
4252 ring_buffer_read_start(struct ring_buffer_iter *iter)
4254 struct ring_buffer_per_cpu *cpu_buffer;
4255 unsigned long flags;
4260 cpu_buffer = iter->cpu_buffer;
4262 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4263 arch_spin_lock(&cpu_buffer->lock);
4264 rb_iter_reset(iter);
4265 arch_spin_unlock(&cpu_buffer->lock);
4266 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4268 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4271 * ring_buffer_read_finish - finish reading the iterator of the buffer
4272 * @iter: The iterator retrieved by ring_buffer_start
4274 * This re-enables the recording to the buffer, and frees the
4278 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4280 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4281 unsigned long flags;
4284 * Ring buffer is disabled from recording, here's a good place
4285 * to check the integrity of the ring buffer.
4286 * Must prevent readers from trying to read, as the check
4287 * clears the HEAD page and readers require it.
4289 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4290 rb_check_pages(cpu_buffer);
4291 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4293 atomic_dec(&cpu_buffer->record_disabled);
4294 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4297 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4300 * ring_buffer_read - read the next item in the ring buffer by the iterator
4301 * @iter: The ring buffer iterator
4302 * @ts: The time stamp of the event read.
4304 * This reads the next event in the ring buffer and increments the iterator.
4306 struct ring_buffer_event *
4307 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4309 struct ring_buffer_event *event;
4310 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4311 unsigned long flags;
4313 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4315 event = rb_iter_peek(iter, ts);
4319 if (event->type_len == RINGBUF_TYPE_PADDING)
4322 rb_advance_iter(iter);
4324 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4328 EXPORT_SYMBOL_GPL(ring_buffer_read);
4331 * ring_buffer_size - return the size of the ring buffer (in bytes)
4332 * @buffer: The ring buffer.
4334 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4337 * Earlier, this method returned
4338 * BUF_PAGE_SIZE * buffer->nr_pages
4339 * Since the nr_pages field is now removed, we have converted this to
4340 * return the per cpu buffer value.
4342 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4345 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4347 EXPORT_SYMBOL_GPL(ring_buffer_size);
4350 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4352 rb_head_page_deactivate(cpu_buffer);
4354 cpu_buffer->head_page
4355 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4356 local_set(&cpu_buffer->head_page->write, 0);
4357 local_set(&cpu_buffer->head_page->entries, 0);
4358 local_set(&cpu_buffer->head_page->page->commit, 0);
4360 cpu_buffer->head_page->read = 0;
4362 cpu_buffer->tail_page = cpu_buffer->head_page;
4363 cpu_buffer->commit_page = cpu_buffer->head_page;
4365 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4366 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4367 local_set(&cpu_buffer->reader_page->write, 0);
4368 local_set(&cpu_buffer->reader_page->entries, 0);
4369 local_set(&cpu_buffer->reader_page->page->commit, 0);
4370 cpu_buffer->reader_page->read = 0;
4372 local_set(&cpu_buffer->entries_bytes, 0);
4373 local_set(&cpu_buffer->overrun, 0);
4374 local_set(&cpu_buffer->commit_overrun, 0);
4375 local_set(&cpu_buffer->dropped_events, 0);
4376 local_set(&cpu_buffer->entries, 0);
4377 local_set(&cpu_buffer->committing, 0);
4378 local_set(&cpu_buffer->commits, 0);
4379 local_set(&cpu_buffer->pages_touched, 0);
4380 local_set(&cpu_buffer->pages_read, 0);
4381 cpu_buffer->last_pages_touch = 0;
4382 cpu_buffer->shortest_full = 0;
4383 cpu_buffer->read = 0;
4384 cpu_buffer->read_bytes = 0;
4386 cpu_buffer->write_stamp = 0;
4387 cpu_buffer->read_stamp = 0;
4389 cpu_buffer->lost_events = 0;
4390 cpu_buffer->last_overrun = 0;
4392 rb_head_page_activate(cpu_buffer);
4396 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4397 * @buffer: The ring buffer to reset a per cpu buffer of
4398 * @cpu: The CPU buffer to be reset
4400 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4402 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4403 unsigned long flags;
4405 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4408 atomic_inc(&buffer->resize_disabled);
4409 atomic_inc(&cpu_buffer->record_disabled);
4411 /* Make sure all commits have finished */
4414 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4416 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4419 arch_spin_lock(&cpu_buffer->lock);
4421 rb_reset_cpu(cpu_buffer);
4423 arch_spin_unlock(&cpu_buffer->lock);
4426 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4428 atomic_dec(&cpu_buffer->record_disabled);
4429 atomic_dec(&buffer->resize_disabled);
4431 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4434 * ring_buffer_reset - reset a ring buffer
4435 * @buffer: The ring buffer to reset all cpu buffers
4437 void ring_buffer_reset(struct ring_buffer *buffer)
4441 for_each_buffer_cpu(buffer, cpu)
4442 ring_buffer_reset_cpu(buffer, cpu);
4444 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4447 * rind_buffer_empty - is the ring buffer empty?
4448 * @buffer: The ring buffer to test
4450 bool ring_buffer_empty(struct ring_buffer *buffer)
4452 struct ring_buffer_per_cpu *cpu_buffer;
4453 unsigned long flags;
4458 /* yes this is racy, but if you don't like the race, lock the buffer */
4459 for_each_buffer_cpu(buffer, cpu) {
4460 cpu_buffer = buffer->buffers[cpu];
4461 local_irq_save(flags);
4462 dolock = rb_reader_lock(cpu_buffer);
4463 ret = rb_per_cpu_empty(cpu_buffer);
4464 rb_reader_unlock(cpu_buffer, dolock);
4465 local_irq_restore(flags);
4473 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4476 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4477 * @buffer: The ring buffer
4478 * @cpu: The CPU buffer to test
4480 bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4482 struct ring_buffer_per_cpu *cpu_buffer;
4483 unsigned long flags;
4487 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4490 cpu_buffer = buffer->buffers[cpu];
4491 local_irq_save(flags);
4492 dolock = rb_reader_lock(cpu_buffer);
4493 ret = rb_per_cpu_empty(cpu_buffer);
4494 rb_reader_unlock(cpu_buffer, dolock);
4495 local_irq_restore(flags);
4499 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4501 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4503 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4504 * @buffer_a: One buffer to swap with
4505 * @buffer_b: The other buffer to swap with
4507 * This function is useful for tracers that want to take a "snapshot"
4508 * of a CPU buffer and has another back up buffer lying around.
4509 * it is expected that the tracer handles the cpu buffer not being
4510 * used at the moment.
4512 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4513 struct ring_buffer *buffer_b, int cpu)
4515 struct ring_buffer_per_cpu *cpu_buffer_a;
4516 struct ring_buffer_per_cpu *cpu_buffer_b;
4519 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4520 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4523 cpu_buffer_a = buffer_a->buffers[cpu];
4524 cpu_buffer_b = buffer_b->buffers[cpu];
4526 /* At least make sure the two buffers are somewhat the same */
4527 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4532 if (atomic_read(&buffer_a->record_disabled))
4535 if (atomic_read(&buffer_b->record_disabled))
4538 if (atomic_read(&cpu_buffer_a->record_disabled))
4541 if (atomic_read(&cpu_buffer_b->record_disabled))
4545 * We can't do a synchronize_rcu here because this
4546 * function can be called in atomic context.
4547 * Normally this will be called from the same CPU as cpu.
4548 * If not it's up to the caller to protect this.
4550 atomic_inc(&cpu_buffer_a->record_disabled);
4551 atomic_inc(&cpu_buffer_b->record_disabled);
4554 if (local_read(&cpu_buffer_a->committing))
4556 if (local_read(&cpu_buffer_b->committing))
4559 buffer_a->buffers[cpu] = cpu_buffer_b;
4560 buffer_b->buffers[cpu] = cpu_buffer_a;
4562 cpu_buffer_b->buffer = buffer_a;
4563 cpu_buffer_a->buffer = buffer_b;
4568 atomic_dec(&cpu_buffer_a->record_disabled);
4569 atomic_dec(&cpu_buffer_b->record_disabled);
4573 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4574 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4577 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4578 * @buffer: the buffer to allocate for.
4579 * @cpu: the cpu buffer to allocate.
4581 * This function is used in conjunction with ring_buffer_read_page.
4582 * When reading a full page from the ring buffer, these functions
4583 * can be used to speed up the process. The calling function should
4584 * allocate a few pages first with this function. Then when it
4585 * needs to get pages from the ring buffer, it passes the result
4586 * of this function into ring_buffer_read_page, which will swap
4587 * the page that was allocated, with the read page of the buffer.
4590 * The page allocated, or ERR_PTR
4592 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4594 struct ring_buffer_per_cpu *cpu_buffer;
4595 struct buffer_data_page *bpage = NULL;
4596 unsigned long flags;
4599 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4600 return ERR_PTR(-ENODEV);
4602 cpu_buffer = buffer->buffers[cpu];
4603 local_irq_save(flags);
4604 arch_spin_lock(&cpu_buffer->lock);
4606 if (cpu_buffer->free_page) {
4607 bpage = cpu_buffer->free_page;
4608 cpu_buffer->free_page = NULL;
4611 arch_spin_unlock(&cpu_buffer->lock);
4612 local_irq_restore(flags);
4617 page = alloc_pages_node(cpu_to_node(cpu),
4618 GFP_KERNEL | __GFP_NORETRY, 0);
4620 return ERR_PTR(-ENOMEM);
4622 bpage = page_address(page);
4625 rb_init_page(bpage);
4629 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4632 * ring_buffer_free_read_page - free an allocated read page
4633 * @buffer: the buffer the page was allocate for
4634 * @cpu: the cpu buffer the page came from
4635 * @data: the page to free
4637 * Free a page allocated from ring_buffer_alloc_read_page.
4639 void ring_buffer_free_read_page(struct ring_buffer *buffer, int cpu, void *data)
4641 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4642 struct buffer_data_page *bpage = data;
4643 struct page *page = virt_to_page(bpage);
4644 unsigned long flags;
4646 /* If the page is still in use someplace else, we can't reuse it */
4647 if (page_ref_count(page) > 1)
4650 local_irq_save(flags);
4651 arch_spin_lock(&cpu_buffer->lock);
4653 if (!cpu_buffer->free_page) {
4654 cpu_buffer->free_page = bpage;
4658 arch_spin_unlock(&cpu_buffer->lock);
4659 local_irq_restore(flags);
4662 free_page((unsigned long)bpage);
4664 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4667 * ring_buffer_read_page - extract a page from the ring buffer
4668 * @buffer: buffer to extract from
4669 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4670 * @len: amount to extract
4671 * @cpu: the cpu of the buffer to extract
4672 * @full: should the extraction only happen when the page is full.
4674 * This function will pull out a page from the ring buffer and consume it.
4675 * @data_page must be the address of the variable that was returned
4676 * from ring_buffer_alloc_read_page. This is because the page might be used
4677 * to swap with a page in the ring buffer.
4680 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4681 * if (IS_ERR(rpage))
4682 * return PTR_ERR(rpage);
4683 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4685 * process_page(rpage, ret);
4687 * When @full is set, the function will not return true unless
4688 * the writer is off the reader page.
4690 * Note: it is up to the calling functions to handle sleeps and wakeups.
4691 * The ring buffer can be used anywhere in the kernel and can not
4692 * blindly call wake_up. The layer that uses the ring buffer must be
4693 * responsible for that.
4696 * >=0 if data has been transferred, returns the offset of consumed data.
4697 * <0 if no data has been transferred.
4699 int ring_buffer_read_page(struct ring_buffer *buffer,
4700 void **data_page, size_t len, int cpu, int full)
4702 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4703 struct ring_buffer_event *event;
4704 struct buffer_data_page *bpage;
4705 struct buffer_page *reader;
4706 unsigned long missed_events;
4707 unsigned long flags;
4708 unsigned int commit;
4713 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4717 * If len is not big enough to hold the page header, then
4718 * we can not copy anything.
4720 if (len <= BUF_PAGE_HDR_SIZE)
4723 len -= BUF_PAGE_HDR_SIZE;
4732 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4734 reader = rb_get_reader_page(cpu_buffer);
4738 event = rb_reader_event(cpu_buffer);
4740 read = reader->read;
4741 commit = rb_page_commit(reader);
4743 /* Check if any events were dropped */
4744 missed_events = cpu_buffer->lost_events;
4747 * If this page has been partially read or
4748 * if len is not big enough to read the rest of the page or
4749 * a writer is still on the page, then
4750 * we must copy the data from the page to the buffer.
4751 * Otherwise, we can simply swap the page with the one passed in.
4753 if (read || (len < (commit - read)) ||
4754 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4755 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4756 unsigned int rpos = read;
4757 unsigned int pos = 0;
4763 if (len > (commit - read))
4764 len = (commit - read);
4766 /* Always keep the time extend and data together */
4767 size = rb_event_ts_length(event);
4772 /* save the current timestamp, since the user will need it */
4773 save_timestamp = cpu_buffer->read_stamp;
4775 /* Need to copy one event at a time */
4777 /* We need the size of one event, because
4778 * rb_advance_reader only advances by one event,
4779 * whereas rb_event_ts_length may include the size of
4780 * one or two events.
4781 * We have already ensured there's enough space if this
4782 * is a time extend. */
4783 size = rb_event_length(event);
4784 memcpy(bpage->data + pos, rpage->data + rpos, size);
4788 rb_advance_reader(cpu_buffer);
4789 rpos = reader->read;
4795 event = rb_reader_event(cpu_buffer);
4796 /* Always keep the time extend and data together */
4797 size = rb_event_ts_length(event);
4798 } while (len >= size);
4801 local_set(&bpage->commit, pos);
4802 bpage->time_stamp = save_timestamp;
4804 /* we copied everything to the beginning */
4807 /* update the entry counter */
4808 cpu_buffer->read += rb_page_entries(reader);
4809 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4811 /* swap the pages */
4812 rb_init_page(bpage);
4813 bpage = reader->page;
4814 reader->page = *data_page;
4815 local_set(&reader->write, 0);
4816 local_set(&reader->entries, 0);
4821 * Use the real_end for the data size,
4822 * This gives us a chance to store the lost events
4825 if (reader->real_end)
4826 local_set(&bpage->commit, reader->real_end);
4830 cpu_buffer->lost_events = 0;
4832 commit = local_read(&bpage->commit);
4834 * Set a flag in the commit field if we lost events
4836 if (missed_events) {
4837 /* If there is room at the end of the page to save the
4838 * missed events, then record it there.
4840 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4841 memcpy(&bpage->data[commit], &missed_events,
4842 sizeof(missed_events));
4843 local_add(RB_MISSED_STORED, &bpage->commit);
4844 commit += sizeof(missed_events);
4846 local_add(RB_MISSED_EVENTS, &bpage->commit);
4850 * This page may be off to user land. Zero it out here.
4852 if (commit < BUF_PAGE_SIZE)
4853 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4856 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4861 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4864 * We only allocate new buffers, never free them if the CPU goes down.
4865 * If we were to free the buffer, then the user would lose any trace that was in
4868 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4870 struct ring_buffer *buffer;
4873 unsigned long nr_pages;
4875 buffer = container_of(node, struct ring_buffer, node);
4876 if (cpumask_test_cpu(cpu, buffer->cpumask))
4881 /* check if all cpu sizes are same */
4882 for_each_buffer_cpu(buffer, cpu_i) {
4883 /* fill in the size from first enabled cpu */
4885 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4886 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4891 /* allocate minimum pages, user can later expand it */
4894 buffer->buffers[cpu] =
4895 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4896 if (!buffer->buffers[cpu]) {
4897 WARN(1, "failed to allocate ring buffer on CPU %u\n",
4902 cpumask_set_cpu(cpu, buffer->cpumask);
4906 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4908 * This is a basic integrity check of the ring buffer.
4909 * Late in the boot cycle this test will run when configured in.
4910 * It will kick off a thread per CPU that will go into a loop
4911 * writing to the per cpu ring buffer various sizes of data.
4912 * Some of the data will be large items, some small.
4914 * Another thread is created that goes into a spin, sending out
4915 * IPIs to the other CPUs to also write into the ring buffer.
4916 * this is to test the nesting ability of the buffer.
4918 * Basic stats are recorded and reported. If something in the
4919 * ring buffer should happen that's not expected, a big warning
4920 * is displayed and all ring buffers are disabled.
4922 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4924 struct rb_test_data {
4925 struct ring_buffer *buffer;
4926 unsigned long events;
4927 unsigned long bytes_written;
4928 unsigned long bytes_alloc;
4929 unsigned long bytes_dropped;
4930 unsigned long events_nested;
4931 unsigned long bytes_written_nested;
4932 unsigned long bytes_alloc_nested;
4933 unsigned long bytes_dropped_nested;
4934 int min_size_nested;
4935 int max_size_nested;
4942 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4945 #define RB_TEST_BUFFER_SIZE 1048576
4947 static char rb_string[] __initdata =
4948 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4949 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4950 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4952 static bool rb_test_started __initdata;
4959 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4961 struct ring_buffer_event *event;
4962 struct rb_item *item;
4969 /* Have nested writes different that what is written */
4970 cnt = data->cnt + (nested ? 27 : 0);
4972 /* Multiply cnt by ~e, to make some unique increment */
4973 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
4975 len = size + sizeof(struct rb_item);
4977 started = rb_test_started;
4978 /* read rb_test_started before checking buffer enabled */
4981 event = ring_buffer_lock_reserve(data->buffer, len);
4983 /* Ignore dropped events before test starts. */
4986 data->bytes_dropped += len;
4988 data->bytes_dropped_nested += len;
4993 event_len = ring_buffer_event_length(event);
4995 if (RB_WARN_ON(data->buffer, event_len < len))
4998 item = ring_buffer_event_data(event);
5000 memcpy(item->str, rb_string, size);
5003 data->bytes_alloc_nested += event_len;
5004 data->bytes_written_nested += len;
5005 data->events_nested++;
5006 if (!data->min_size_nested || len < data->min_size_nested)
5007 data->min_size_nested = len;
5008 if (len > data->max_size_nested)
5009 data->max_size_nested = len;
5011 data->bytes_alloc += event_len;
5012 data->bytes_written += len;
5014 if (!data->min_size || len < data->min_size)
5015 data->max_size = len;
5016 if (len > data->max_size)
5017 data->max_size = len;
5021 ring_buffer_unlock_commit(data->buffer, event);
5026 static __init int rb_test(void *arg)
5028 struct rb_test_data *data = arg;
5030 while (!kthread_should_stop()) {
5031 rb_write_something(data, false);
5034 set_current_state(TASK_INTERRUPTIBLE);
5035 /* Now sleep between a min of 100-300us and a max of 1ms */
5036 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
5042 static __init void rb_ipi(void *ignore)
5044 struct rb_test_data *data;
5045 int cpu = smp_processor_id();
5047 data = &rb_data[cpu];
5048 rb_write_something(data, true);
5051 static __init int rb_hammer_test(void *arg)
5053 while (!kthread_should_stop()) {
5055 /* Send an IPI to all cpus to write data! */
5056 smp_call_function(rb_ipi, NULL, 1);
5057 /* No sleep, but for non preempt, let others run */
5064 static __init int test_ringbuffer(void)
5066 struct task_struct *rb_hammer;
5067 struct ring_buffer *buffer;
5071 pr_info("Running ring buffer tests...\n");
5073 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5074 if (WARN_ON(!buffer))
5077 /* Disable buffer so that threads can't write to it yet */
5078 ring_buffer_record_off(buffer);
5080 for_each_online_cpu(cpu) {
5081 rb_data[cpu].buffer = buffer;
5082 rb_data[cpu].cpu = cpu;
5083 rb_data[cpu].cnt = cpu;
5084 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5085 "rbtester/%d", cpu);
5086 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5087 pr_cont("FAILED\n");
5088 ret = PTR_ERR(rb_threads[cpu]);
5092 kthread_bind(rb_threads[cpu], cpu);
5093 wake_up_process(rb_threads[cpu]);
5096 /* Now create the rb hammer! */
5097 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5098 if (WARN_ON(IS_ERR(rb_hammer))) {
5099 pr_cont("FAILED\n");
5100 ret = PTR_ERR(rb_hammer);
5104 ring_buffer_record_on(buffer);
5106 * Show buffer is enabled before setting rb_test_started.
5107 * Yes there's a small race window where events could be
5108 * dropped and the thread wont catch it. But when a ring
5109 * buffer gets enabled, there will always be some kind of
5110 * delay before other CPUs see it. Thus, we don't care about
5111 * those dropped events. We care about events dropped after
5112 * the threads see that the buffer is active.
5115 rb_test_started = true;
5117 set_current_state(TASK_INTERRUPTIBLE);
5118 /* Just run for 10 seconds */;
5119 schedule_timeout(10 * HZ);
5121 kthread_stop(rb_hammer);
5124 for_each_online_cpu(cpu) {
5125 if (!rb_threads[cpu])
5127 kthread_stop(rb_threads[cpu]);
5130 ring_buffer_free(buffer);
5135 pr_info("finished\n");
5136 for_each_online_cpu(cpu) {
5137 struct ring_buffer_event *event;
5138 struct rb_test_data *data = &rb_data[cpu];
5139 struct rb_item *item;
5140 unsigned long total_events;
5141 unsigned long total_dropped;
5142 unsigned long total_written;
5143 unsigned long total_alloc;
5144 unsigned long total_read = 0;
5145 unsigned long total_size = 0;
5146 unsigned long total_len = 0;
5147 unsigned long total_lost = 0;
5150 int small_event_size;
5154 total_events = data->events + data->events_nested;
5155 total_written = data->bytes_written + data->bytes_written_nested;
5156 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5157 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5159 big_event_size = data->max_size + data->max_size_nested;
5160 small_event_size = data->min_size + data->min_size_nested;
5162 pr_info("CPU %d:\n", cpu);
5163 pr_info(" events: %ld\n", total_events);
5164 pr_info(" dropped bytes: %ld\n", total_dropped);
5165 pr_info(" alloced bytes: %ld\n", total_alloc);
5166 pr_info(" written bytes: %ld\n", total_written);
5167 pr_info(" biggest event: %d\n", big_event_size);
5168 pr_info(" smallest event: %d\n", small_event_size);
5170 if (RB_WARN_ON(buffer, total_dropped))
5175 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5177 item = ring_buffer_event_data(event);
5178 total_len += ring_buffer_event_length(event);
5179 total_size += item->size + sizeof(struct rb_item);
5180 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5181 pr_info("FAILED!\n");
5182 pr_info("buffer had: %.*s\n", item->size, item->str);
5183 pr_info("expected: %.*s\n", item->size, rb_string);
5184 RB_WARN_ON(buffer, 1);
5195 pr_info(" read events: %ld\n", total_read);
5196 pr_info(" lost events: %ld\n", total_lost);
5197 pr_info(" total events: %ld\n", total_lost + total_read);
5198 pr_info(" recorded len bytes: %ld\n", total_len);
5199 pr_info(" recorded size bytes: %ld\n", total_size);
5201 pr_info(" With dropped events, record len and size may not match\n"
5202 " alloced and written from above\n");
5204 if (RB_WARN_ON(buffer, total_len != total_alloc ||
5205 total_size != total_written))
5208 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5214 pr_info("Ring buffer PASSED!\n");
5216 ring_buffer_free(buffer);
5220 late_initcall(test_ringbuffer);
5221 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */