1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
7 #include <linux/trace_recursion.h>
8 #include <linux/trace_events.h>
9 #include <linux/ring_buffer.h>
10 #include <linux/trace_clock.h>
11 #include <linux/sched/clock.h>
12 #include <linux/trace_seq.h>
13 #include <linux/spinlock.h>
14 #include <linux/irq_work.h>
15 #include <linux/security.h>
16 #include <linux/uaccess.h>
17 #include <linux/hardirq.h>
18 #include <linux/kthread.h> /* for self test */
19 #include <linux/module.h>
20 #include <linux/percpu.h>
21 #include <linux/mutex.h>
22 #include <linux/delay.h>
23 #include <linux/slab.h>
24 #include <linux/init.h>
25 #include <linux/hash.h>
26 #include <linux/list.h>
27 #include <linux/cpu.h>
28 #include <linux/oom.h>
30 #include <asm/local64.h>
31 #include <asm/local.h>
34 * The "absolute" timestamp in the buffer is only 59 bits.
35 * If a clock has the 5 MSBs set, it needs to be saved and
38 #define TS_MSB (0xf8ULL << 56)
39 #define ABS_TS_MASK (~TS_MSB)
41 static void update_pages_handler(struct work_struct *work);
44 * The ring buffer header is special. We must manually up keep it.
46 int ring_buffer_print_entry_header(struct trace_seq *s)
48 trace_seq_puts(s, "# compressed entry header\n");
49 trace_seq_puts(s, "\ttype_len : 5 bits\n");
50 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
51 trace_seq_puts(s, "\tarray : 32 bits\n");
52 trace_seq_putc(s, '\n');
53 trace_seq_printf(s, "\tpadding : type == %d\n",
54 RINGBUF_TYPE_PADDING);
55 trace_seq_printf(s, "\ttime_extend : type == %d\n",
56 RINGBUF_TYPE_TIME_EXTEND);
57 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
58 RINGBUF_TYPE_TIME_STAMP);
59 trace_seq_printf(s, "\tdata max type_len == %d\n",
60 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
62 return !trace_seq_has_overflowed(s);
66 * The ring buffer is made up of a list of pages. A separate list of pages is
67 * allocated for each CPU. A writer may only write to a buffer that is
68 * associated with the CPU it is currently executing on. A reader may read
69 * from any per cpu buffer.
71 * The reader is special. For each per cpu buffer, the reader has its own
72 * reader page. When a reader has read the entire reader page, this reader
73 * page is swapped with another page in the ring buffer.
75 * Now, as long as the writer is off the reader page, the reader can do what
76 * ever it wants with that page. The writer will never write to that page
77 * again (as long as it is out of the ring buffer).
79 * Here's some silly ASCII art.
82 * |reader| RING BUFFER
84 * +------+ +---+ +---+ +---+
93 * |reader| RING BUFFER
94 * |page |------------------v
95 * +------+ +---+ +---+ +---+
104 * |reader| RING BUFFER
105 * |page |------------------v
106 * +------+ +---+ +---+ +---+
108 * | +---+ +---+ +---+
111 * +------------------------------+
115 * |buffer| RING BUFFER
116 * |page |------------------v
117 * +------+ +---+ +---+ +---+
119 * | New +---+ +---+ +---+
122 * +------------------------------+
125 * After we make this swap, the reader can hand this page off to the splice
126 * code and be done with it. It can even allocate a new page if it needs to
127 * and swap that into the ring buffer.
129 * We will be using cmpxchg soon to make all this lockless.
133 /* Used for individual buffers (after the counter) */
134 #define RB_BUFFER_OFF (1 << 20)
136 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
138 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
139 #define RB_ALIGNMENT 4U
140 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
141 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
143 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
144 # define RB_FORCE_8BYTE_ALIGNMENT 0
145 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
147 # define RB_FORCE_8BYTE_ALIGNMENT 1
148 # define RB_ARCH_ALIGNMENT 8U
151 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
153 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
154 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
157 RB_LEN_TIME_EXTEND = 8,
158 RB_LEN_TIME_STAMP = 8,
161 #define skip_time_extend(event) \
162 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
164 #define extended_time(event) \
165 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
167 static inline bool rb_null_event(struct ring_buffer_event *event)
169 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
172 static void rb_event_set_padding(struct ring_buffer_event *event)
174 /* padding has a NULL time_delta */
175 event->type_len = RINGBUF_TYPE_PADDING;
176 event->time_delta = 0;
180 rb_event_data_length(struct ring_buffer_event *event)
185 length = event->type_len * RB_ALIGNMENT;
187 length = event->array[0];
188 return length + RB_EVNT_HDR_SIZE;
192 * Return the length of the given event. Will return
193 * the length of the time extend if the event is a
196 static inline unsigned
197 rb_event_length(struct ring_buffer_event *event)
199 switch (event->type_len) {
200 case RINGBUF_TYPE_PADDING:
201 if (rb_null_event(event))
204 return event->array[0] + RB_EVNT_HDR_SIZE;
206 case RINGBUF_TYPE_TIME_EXTEND:
207 return RB_LEN_TIME_EXTEND;
209 case RINGBUF_TYPE_TIME_STAMP:
210 return RB_LEN_TIME_STAMP;
212 case RINGBUF_TYPE_DATA:
213 return rb_event_data_length(event);
222 * Return total length of time extend and data,
223 * or just the event length for all other events.
225 static inline unsigned
226 rb_event_ts_length(struct ring_buffer_event *event)
230 if (extended_time(event)) {
231 /* time extends include the data event after it */
232 len = RB_LEN_TIME_EXTEND;
233 event = skip_time_extend(event);
235 return len + rb_event_length(event);
239 * ring_buffer_event_length - return the length of the event
240 * @event: the event to get the length of
242 * Returns the size of the data load of a data event.
243 * If the event is something other than a data event, it
244 * returns the size of the event itself. With the exception
245 * of a TIME EXTEND, where it still returns the size of the
246 * data load of the data event after it.
248 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
252 if (extended_time(event))
253 event = skip_time_extend(event);
255 length = rb_event_length(event);
256 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
258 length -= RB_EVNT_HDR_SIZE;
259 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
260 length -= sizeof(event->array[0]);
263 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
265 /* inline for ring buffer fast paths */
266 static __always_inline void *
267 rb_event_data(struct ring_buffer_event *event)
269 if (extended_time(event))
270 event = skip_time_extend(event);
271 WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
272 /* If length is in len field, then array[0] has the data */
274 return (void *)&event->array[0];
275 /* Otherwise length is in array[0] and array[1] has the data */
276 return (void *)&event->array[1];
280 * ring_buffer_event_data - return the data of the event
281 * @event: the event to get the data from
283 void *ring_buffer_event_data(struct ring_buffer_event *event)
285 return rb_event_data(event);
287 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
289 #define for_each_buffer_cpu(buffer, cpu) \
290 for_each_cpu(cpu, buffer->cpumask)
292 #define for_each_online_buffer_cpu(buffer, cpu) \
293 for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
296 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
297 #define TS_DELTA_TEST (~TS_MASK)
299 static u64 rb_event_time_stamp(struct ring_buffer_event *event)
303 ts = event->array[0];
305 ts += event->time_delta;
310 /* Flag when events were overwritten */
311 #define RB_MISSED_EVENTS (1 << 31)
312 /* Missed count stored at end */
313 #define RB_MISSED_STORED (1 << 30)
315 struct buffer_data_page {
316 u64 time_stamp; /* page time stamp */
317 local_t commit; /* write committed index */
318 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
321 struct buffer_data_read_page {
322 unsigned order; /* order of the page */
323 struct buffer_data_page *data; /* actual data, stored in this page */
327 * Note, the buffer_page list must be first. The buffer pages
328 * are allocated in cache lines, which means that each buffer
329 * page will be at the beginning of a cache line, and thus
330 * the least significant bits will be zero. We use this to
331 * add flags in the list struct pointers, to make the ring buffer
335 struct list_head list; /* list of buffer pages */
336 local_t write; /* index for next write */
337 unsigned read; /* index for next read */
338 local_t entries; /* entries on this page */
339 unsigned long real_end; /* real end of data */
340 unsigned order; /* order of the page */
341 struct buffer_data_page *page; /* Actual data page */
345 * The buffer page counters, write and entries, must be reset
346 * atomically when crossing page boundaries. To synchronize this
347 * update, two counters are inserted into the number. One is
348 * the actual counter for the write position or count on the page.
350 * The other is a counter of updaters. Before an update happens
351 * the update partition of the counter is incremented. This will
352 * allow the updater to update the counter atomically.
354 * The counter is 20 bits, and the state data is 12.
356 #define RB_WRITE_MASK 0xfffff
357 #define RB_WRITE_INTCNT (1 << 20)
359 static void rb_init_page(struct buffer_data_page *bpage)
361 local_set(&bpage->commit, 0);
364 static __always_inline unsigned int rb_page_commit(struct buffer_page *bpage)
366 return local_read(&bpage->page->commit);
369 static void free_buffer_page(struct buffer_page *bpage)
371 free_pages((unsigned long)bpage->page, bpage->order);
376 * We need to fit the time_stamp delta into 27 bits.
378 static inline bool test_time_stamp(u64 delta)
380 return !!(delta & TS_DELTA_TEST);
384 struct irq_work work;
385 wait_queue_head_t waiters;
386 wait_queue_head_t full_waiters;
388 bool waiters_pending;
389 bool full_waiters_pending;
394 * Structure to hold event state and handle nested events.
396 struct rb_event_info {
401 unsigned long length;
402 struct buffer_page *tail_page;
407 * Used for the add_timestamp
409 * EXTEND - wants a time extend
410 * ABSOLUTE - the buffer requests all events to have absolute time stamps
411 * FORCE - force a full time stamp.
414 RB_ADD_STAMP_NONE = 0,
415 RB_ADD_STAMP_EXTEND = BIT(1),
416 RB_ADD_STAMP_ABSOLUTE = BIT(2),
417 RB_ADD_STAMP_FORCE = BIT(3)
420 * Used for which event context the event is in.
427 * See trace_recursive_lock() comment below for more details.
438 struct rb_time_struct {
441 typedef struct rb_time_struct rb_time_t;
446 * head_page == tail_page && head == tail then buffer is empty.
448 struct ring_buffer_per_cpu {
450 atomic_t record_disabled;
451 atomic_t resize_disabled;
452 struct trace_buffer *buffer;
453 raw_spinlock_t reader_lock; /* serialize readers */
454 arch_spinlock_t lock;
455 struct lock_class_key lock_key;
456 struct buffer_data_page *free_page;
457 unsigned long nr_pages;
458 unsigned int current_context;
459 struct list_head *pages;
460 struct buffer_page *head_page; /* read from head */
461 struct buffer_page *tail_page; /* write to tail */
462 struct buffer_page *commit_page; /* committed pages */
463 struct buffer_page *reader_page;
464 unsigned long lost_events;
465 unsigned long last_overrun;
467 local_t entries_bytes;
470 local_t commit_overrun;
471 local_t dropped_events;
474 local_t pages_touched;
477 long last_pages_touch;
478 size_t shortest_full;
480 unsigned long read_bytes;
481 rb_time_t write_stamp;
482 rb_time_t before_stamp;
483 u64 event_stamp[MAX_NEST];
485 /* pages removed since last reset */
486 unsigned long pages_removed;
487 /* ring buffer pages to update, > 0 to add, < 0 to remove */
488 long nr_pages_to_update;
489 struct list_head new_pages; /* new pages to add */
490 struct work_struct update_pages_work;
491 struct completion update_done;
493 struct rb_irq_work irq_work;
496 struct trace_buffer {
499 atomic_t record_disabled;
501 cpumask_var_t cpumask;
503 struct lock_class_key *reader_lock_key;
507 struct ring_buffer_per_cpu **buffers;
509 struct hlist_node node;
512 struct rb_irq_work irq_work;
515 unsigned int subbuf_size;
516 unsigned int subbuf_order;
517 unsigned int max_data_size;
520 struct ring_buffer_iter {
521 struct ring_buffer_per_cpu *cpu_buffer;
523 unsigned long next_event;
524 struct buffer_page *head_page;
525 struct buffer_page *cache_reader_page;
526 unsigned long cache_read;
527 unsigned long cache_pages_removed;
530 struct ring_buffer_event *event;
535 int ring_buffer_print_page_header(struct trace_buffer *buffer, struct trace_seq *s)
537 struct buffer_data_page field;
539 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
540 "offset:0;\tsize:%u;\tsigned:%u;\n",
541 (unsigned int)sizeof(field.time_stamp),
542 (unsigned int)is_signed_type(u64));
544 trace_seq_printf(s, "\tfield: local_t commit;\t"
545 "offset:%u;\tsize:%u;\tsigned:%u;\n",
546 (unsigned int)offsetof(typeof(field), commit),
547 (unsigned int)sizeof(field.commit),
548 (unsigned int)is_signed_type(long));
550 trace_seq_printf(s, "\tfield: int overwrite;\t"
551 "offset:%u;\tsize:%u;\tsigned:%u;\n",
552 (unsigned int)offsetof(typeof(field), commit),
554 (unsigned int)is_signed_type(long));
556 trace_seq_printf(s, "\tfield: char data;\t"
557 "offset:%u;\tsize:%u;\tsigned:%u;\n",
558 (unsigned int)offsetof(typeof(field), data),
559 (unsigned int)buffer->subbuf_size,
560 (unsigned int)is_signed_type(char));
562 return !trace_seq_has_overflowed(s);
565 static inline void rb_time_read(rb_time_t *t, u64 *ret)
567 *ret = local64_read(&t->time);
569 static void rb_time_set(rb_time_t *t, u64 val)
571 local64_set(&t->time, val);
575 * Enable this to make sure that the event passed to
576 * ring_buffer_event_time_stamp() is not committed and also
577 * is on the buffer that it passed in.
579 //#define RB_VERIFY_EVENT
580 #ifdef RB_VERIFY_EVENT
581 static struct list_head *rb_list_head(struct list_head *list);
582 static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
585 struct buffer_page *page = cpu_buffer->commit_page;
586 struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
587 struct list_head *next;
589 unsigned long addr = (unsigned long)event;
593 /* Make sure the event exists and is not committed yet */
595 if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
597 commit = local_read(&page->page->commit);
598 write = local_read(&page->write);
599 if (addr >= (unsigned long)&page->page->data[commit] &&
600 addr < (unsigned long)&page->page->data[write])
603 next = rb_list_head(page->list.next);
604 page = list_entry(next, struct buffer_page, list);
609 static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
616 * The absolute time stamp drops the 5 MSBs and some clocks may
617 * require them. The rb_fix_abs_ts() will take a previous full
618 * time stamp, and add the 5 MSB of that time stamp on to the
619 * saved absolute time stamp. Then they are compared in case of
620 * the unlikely event that the latest time stamp incremented
623 static inline u64 rb_fix_abs_ts(u64 abs, u64 save_ts)
625 if (save_ts & TS_MSB) {
626 abs |= save_ts & TS_MSB;
627 /* Check for overflow */
628 if (unlikely(abs < save_ts))
634 static inline u64 rb_time_stamp(struct trace_buffer *buffer);
637 * ring_buffer_event_time_stamp - return the event's current time stamp
638 * @buffer: The buffer that the event is on
639 * @event: the event to get the time stamp of
641 * Note, this must be called after @event is reserved, and before it is
642 * committed to the ring buffer. And must be called from the same
643 * context where the event was reserved (normal, softirq, irq, etc).
645 * Returns the time stamp associated with the current event.
646 * If the event has an extended time stamp, then that is used as
647 * the time stamp to return.
648 * In the highly unlikely case that the event was nested more than
649 * the max nesting, then the write_stamp of the buffer is returned,
650 * otherwise current time is returned, but that really neither of
651 * the last two cases should ever happen.
653 u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
654 struct ring_buffer_event *event)
656 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
660 /* If the event includes an absolute time, then just use that */
661 if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
662 ts = rb_event_time_stamp(event);
663 return rb_fix_abs_ts(ts, cpu_buffer->tail_page->page->time_stamp);
666 nest = local_read(&cpu_buffer->committing);
667 verify_event(cpu_buffer, event);
668 if (WARN_ON_ONCE(!nest))
671 /* Read the current saved nesting level time stamp */
672 if (likely(--nest < MAX_NEST))
673 return cpu_buffer->event_stamp[nest];
675 /* Shouldn't happen, warn if it does */
676 WARN_ONCE(1, "nest (%d) greater than max", nest);
679 rb_time_read(&cpu_buffer->write_stamp, &ts);
685 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
686 * @buffer: The ring_buffer to get the number of pages from
687 * @cpu: The cpu of the ring_buffer to get the number of pages from
689 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
691 size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
693 return buffer->buffers[cpu]->nr_pages;
697 * ring_buffer_nr_dirty_pages - get the number of used pages in the ring buffer
698 * @buffer: The ring_buffer to get the number of pages from
699 * @cpu: The cpu of the ring_buffer to get the number of pages from
701 * Returns the number of pages that have content in the ring buffer.
703 size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
709 read = local_read(&buffer->buffers[cpu]->pages_read);
710 lost = local_read(&buffer->buffers[cpu]->pages_lost);
711 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
713 if (WARN_ON_ONCE(cnt < lost))
718 /* The reader can read an empty page, but not more than that */
720 WARN_ON_ONCE(read > cnt + 1);
727 static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
729 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
733 nr_pages = cpu_buffer->nr_pages;
734 if (!nr_pages || !full)
738 * Add one as dirty will never equal nr_pages, as the sub-buffer
739 * that the writer is on is not counted as dirty.
740 * This is needed if "buffer_percent" is set to 100.
742 dirty = ring_buffer_nr_dirty_pages(buffer, cpu) + 1;
744 return (dirty * 100) >= (full * nr_pages);
748 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
750 * Schedules a delayed work to wake up any task that is blocked on the
751 * ring buffer waiters queue.
753 static void rb_wake_up_waiters(struct irq_work *work)
755 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
757 wake_up_all(&rbwork->waiters);
758 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
759 rbwork->wakeup_full = false;
760 rbwork->full_waiters_pending = false;
761 wake_up_all(&rbwork->full_waiters);
766 * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
767 * @buffer: The ring buffer to wake waiters on
768 * @cpu: The CPU buffer to wake waiters on
770 * In the case of a file that represents a ring buffer is closing,
771 * it is prudent to wake up any waiters that are on this.
773 void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
775 struct ring_buffer_per_cpu *cpu_buffer;
776 struct rb_irq_work *rbwork;
781 if (cpu == RING_BUFFER_ALL_CPUS) {
783 /* Wake up individual ones too. One level recursion */
784 for_each_buffer_cpu(buffer, cpu)
785 ring_buffer_wake_waiters(buffer, cpu);
787 rbwork = &buffer->irq_work;
789 if (WARN_ON_ONCE(!buffer->buffers))
791 if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
794 cpu_buffer = buffer->buffers[cpu];
795 /* The CPU buffer may not have been initialized yet */
798 rbwork = &cpu_buffer->irq_work;
801 rbwork->wait_index++;
802 /* make sure the waiters see the new index */
805 /* This can be called in any context */
806 irq_work_queue(&rbwork->work);
810 * ring_buffer_wait - wait for input to the ring buffer
811 * @buffer: buffer to wait on
812 * @cpu: the cpu buffer to wait on
813 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
815 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
816 * as data is added to any of the @buffer's cpu buffers. Otherwise
817 * it will wait for data to be added to a specific cpu buffer.
819 int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
821 struct ring_buffer_per_cpu *cpu_buffer;
823 struct rb_irq_work *work;
828 * Depending on what the caller is waiting for, either any
829 * data in any cpu buffer, or a specific buffer, put the
830 * caller on the appropriate wait queue.
832 if (cpu == RING_BUFFER_ALL_CPUS) {
833 work = &buffer->irq_work;
834 /* Full only makes sense on per cpu reads */
837 if (!cpumask_test_cpu(cpu, buffer->cpumask))
839 cpu_buffer = buffer->buffers[cpu];
840 work = &cpu_buffer->irq_work;
843 wait_index = READ_ONCE(work->wait_index);
847 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
849 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
852 * The events can happen in critical sections where
853 * checking a work queue can cause deadlocks.
854 * After adding a task to the queue, this flag is set
855 * only to notify events to try to wake up the queue
858 * We don't clear it even if the buffer is no longer
859 * empty. The flag only causes the next event to run
860 * irq_work to do the work queue wake up. The worse
861 * that can happen if we race with !trace_empty() is that
862 * an event will cause an irq_work to try to wake up
865 * There's no reason to protect this flag either, as
866 * the work queue and irq_work logic will do the necessary
867 * synchronization for the wake ups. The only thing
868 * that is necessary is that the wake up happens after
869 * a task has been queued. It's OK for spurious wake ups.
872 work->full_waiters_pending = true;
874 work->waiters_pending = true;
876 if (signal_pending(current)) {
881 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
884 if (cpu != RING_BUFFER_ALL_CPUS &&
885 !ring_buffer_empty_cpu(buffer, cpu)) {
893 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
894 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
895 done = !pagebusy && full_hit(buffer, cpu, full);
897 if (!cpu_buffer->shortest_full ||
898 cpu_buffer->shortest_full > full)
899 cpu_buffer->shortest_full = full;
900 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
907 /* Make sure to see the new wait index */
909 if (wait_index != work->wait_index)
914 finish_wait(&work->full_waiters, &wait);
916 finish_wait(&work->waiters, &wait);
922 * ring_buffer_poll_wait - poll on buffer input
923 * @buffer: buffer to wait on
924 * @cpu: the cpu buffer to wait on
925 * @filp: the file descriptor
926 * @poll_table: The poll descriptor
927 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
929 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
930 * as data is added to any of the @buffer's cpu buffers. Otherwise
931 * it will wait for data to be added to a specific cpu buffer.
933 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
936 __poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
937 struct file *filp, poll_table *poll_table, int full)
939 struct ring_buffer_per_cpu *cpu_buffer;
940 struct rb_irq_work *work;
942 if (cpu == RING_BUFFER_ALL_CPUS) {
943 work = &buffer->irq_work;
946 if (!cpumask_test_cpu(cpu, buffer->cpumask))
949 cpu_buffer = buffer->buffers[cpu];
950 work = &cpu_buffer->irq_work;
954 poll_wait(filp, &work->full_waiters, poll_table);
955 work->full_waiters_pending = true;
956 if (!cpu_buffer->shortest_full ||
957 cpu_buffer->shortest_full > full)
958 cpu_buffer->shortest_full = full;
960 poll_wait(filp, &work->waiters, poll_table);
961 work->waiters_pending = true;
965 * There's a tight race between setting the waiters_pending and
966 * checking if the ring buffer is empty. Once the waiters_pending bit
967 * is set, the next event will wake the task up, but we can get stuck
968 * if there's only a single event in.
970 * FIXME: Ideally, we need a memory barrier on the writer side as well,
971 * but adding a memory barrier to all events will cause too much of a
972 * performance hit in the fast path. We only need a memory barrier when
973 * the buffer goes from empty to having content. But as this race is
974 * extremely small, and it's not a problem if another event comes in, we
980 return full_hit(buffer, cpu, full) ? EPOLLIN | EPOLLRDNORM : 0;
982 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
983 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
984 return EPOLLIN | EPOLLRDNORM;
988 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
989 #define RB_WARN_ON(b, cond) \
991 int _____ret = unlikely(cond); \
993 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
994 struct ring_buffer_per_cpu *__b = \
996 atomic_inc(&__b->buffer->record_disabled); \
998 atomic_inc(&b->record_disabled); \
1004 /* Up this if you want to test the TIME_EXTENTS and normalization */
1005 #define DEBUG_SHIFT 0
1007 static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1011 /* Skip retpolines :-( */
1012 if (IS_ENABLED(CONFIG_MITIGATION_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1013 ts = trace_clock_local();
1015 ts = buffer->clock();
1017 /* shift to debug/test normalization and TIME_EXTENTS */
1018 return ts << DEBUG_SHIFT;
1021 u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1025 preempt_disable_notrace();
1026 time = rb_time_stamp(buffer);
1027 preempt_enable_notrace();
1031 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1033 void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1036 /* Just stupid testing the normalize function and deltas */
1037 *ts >>= DEBUG_SHIFT;
1039 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1042 * Making the ring buffer lockless makes things tricky.
1043 * Although writes only happen on the CPU that they are on,
1044 * and they only need to worry about interrupts. Reads can
1045 * happen on any CPU.
1047 * The reader page is always off the ring buffer, but when the
1048 * reader finishes with a page, it needs to swap its page with
1049 * a new one from the buffer. The reader needs to take from
1050 * the head (writes go to the tail). But if a writer is in overwrite
1051 * mode and wraps, it must push the head page forward.
1053 * Here lies the problem.
1055 * The reader must be careful to replace only the head page, and
1056 * not another one. As described at the top of the file in the
1057 * ASCII art, the reader sets its old page to point to the next
1058 * page after head. It then sets the page after head to point to
1059 * the old reader page. But if the writer moves the head page
1060 * during this operation, the reader could end up with the tail.
1062 * We use cmpxchg to help prevent this race. We also do something
1063 * special with the page before head. We set the LSB to 1.
1065 * When the writer must push the page forward, it will clear the
1066 * bit that points to the head page, move the head, and then set
1067 * the bit that points to the new head page.
1069 * We also don't want an interrupt coming in and moving the head
1070 * page on another writer. Thus we use the second LSB to catch
1073 * head->list->prev->next bit 1 bit 0
1076 * Points to head page 0 1
1079 * Note we can not trust the prev pointer of the head page, because:
1081 * +----+ +-----+ +-----+
1082 * | |------>| T |---X--->| N |
1084 * +----+ +-----+ +-----+
1087 * +----------| R |----------+ |
1091 * Key: ---X--> HEAD flag set in pointer
1096 * (see __rb_reserve_next() to see where this happens)
1098 * What the above shows is that the reader just swapped out
1099 * the reader page with a page in the buffer, but before it
1100 * could make the new header point back to the new page added
1101 * it was preempted by a writer. The writer moved forward onto
1102 * the new page added by the reader and is about to move forward
1105 * You can see, it is legitimate for the previous pointer of
1106 * the head (or any page) not to point back to itself. But only
1110 #define RB_PAGE_NORMAL 0UL
1111 #define RB_PAGE_HEAD 1UL
1112 #define RB_PAGE_UPDATE 2UL
1115 #define RB_FLAG_MASK 3UL
1117 /* PAGE_MOVED is not part of the mask */
1118 #define RB_PAGE_MOVED 4UL
1121 * rb_list_head - remove any bit
1123 static struct list_head *rb_list_head(struct list_head *list)
1125 unsigned long val = (unsigned long)list;
1127 return (struct list_head *)(val & ~RB_FLAG_MASK);
1131 * rb_is_head_page - test if the given page is the head page
1133 * Because the reader may move the head_page pointer, we can
1134 * not trust what the head page is (it may be pointing to
1135 * the reader page). But if the next page is a header page,
1136 * its flags will be non zero.
1139 rb_is_head_page(struct buffer_page *page, struct list_head *list)
1143 val = (unsigned long)list->next;
1145 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1146 return RB_PAGE_MOVED;
1148 return val & RB_FLAG_MASK;
1154 * The unique thing about the reader page, is that, if the
1155 * writer is ever on it, the previous pointer never points
1156 * back to the reader page.
1158 static bool rb_is_reader_page(struct buffer_page *page)
1160 struct list_head *list = page->list.prev;
1162 return rb_list_head(list->next) != &page->list;
1166 * rb_set_list_to_head - set a list_head to be pointing to head.
1168 static void rb_set_list_to_head(struct list_head *list)
1172 ptr = (unsigned long *)&list->next;
1173 *ptr |= RB_PAGE_HEAD;
1174 *ptr &= ~RB_PAGE_UPDATE;
1178 * rb_head_page_activate - sets up head page
1180 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1182 struct buffer_page *head;
1184 head = cpu_buffer->head_page;
1189 * Set the previous list pointer to have the HEAD flag.
1191 rb_set_list_to_head(head->list.prev);
1194 static void rb_list_head_clear(struct list_head *list)
1196 unsigned long *ptr = (unsigned long *)&list->next;
1198 *ptr &= ~RB_FLAG_MASK;
1202 * rb_head_page_deactivate - clears head page ptr (for free list)
1205 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1207 struct list_head *hd;
1209 /* Go through the whole list and clear any pointers found. */
1210 rb_list_head_clear(cpu_buffer->pages);
1212 list_for_each(hd, cpu_buffer->pages)
1213 rb_list_head_clear(hd);
1216 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1217 struct buffer_page *head,
1218 struct buffer_page *prev,
1219 int old_flag, int new_flag)
1221 struct list_head *list;
1222 unsigned long val = (unsigned long)&head->list;
1227 val &= ~RB_FLAG_MASK;
1229 ret = cmpxchg((unsigned long *)&list->next,
1230 val | old_flag, val | new_flag);
1232 /* check if the reader took the page */
1233 if ((ret & ~RB_FLAG_MASK) != val)
1234 return RB_PAGE_MOVED;
1236 return ret & RB_FLAG_MASK;
1239 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1240 struct buffer_page *head,
1241 struct buffer_page *prev,
1244 return rb_head_page_set(cpu_buffer, head, prev,
1245 old_flag, RB_PAGE_UPDATE);
1248 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1249 struct buffer_page *head,
1250 struct buffer_page *prev,
1253 return rb_head_page_set(cpu_buffer, head, prev,
1254 old_flag, RB_PAGE_HEAD);
1257 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1258 struct buffer_page *head,
1259 struct buffer_page *prev,
1262 return rb_head_page_set(cpu_buffer, head, prev,
1263 old_flag, RB_PAGE_NORMAL);
1266 static inline void rb_inc_page(struct buffer_page **bpage)
1268 struct list_head *p = rb_list_head((*bpage)->list.next);
1270 *bpage = list_entry(p, struct buffer_page, list);
1273 static struct buffer_page *
1274 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1276 struct buffer_page *head;
1277 struct buffer_page *page;
1278 struct list_head *list;
1281 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1285 list = cpu_buffer->pages;
1286 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1289 page = head = cpu_buffer->head_page;
1291 * It is possible that the writer moves the header behind
1292 * where we started, and we miss in one loop.
1293 * A second loop should grab the header, but we'll do
1294 * three loops just because I'm paranoid.
1296 for (i = 0; i < 3; i++) {
1298 if (rb_is_head_page(page, page->list.prev)) {
1299 cpu_buffer->head_page = page;
1303 } while (page != head);
1306 RB_WARN_ON(cpu_buffer, 1);
1311 static bool rb_head_page_replace(struct buffer_page *old,
1312 struct buffer_page *new)
1314 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1317 val = *ptr & ~RB_FLAG_MASK;
1318 val |= RB_PAGE_HEAD;
1320 return try_cmpxchg(ptr, &val, (unsigned long)&new->list);
1324 * rb_tail_page_update - move the tail page forward
1326 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1327 struct buffer_page *tail_page,
1328 struct buffer_page *next_page)
1330 unsigned long old_entries;
1331 unsigned long old_write;
1334 * The tail page now needs to be moved forward.
1336 * We need to reset the tail page, but without messing
1337 * with possible erasing of data brought in by interrupts
1338 * that have moved the tail page and are currently on it.
1340 * We add a counter to the write field to denote this.
1342 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1343 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1345 local_inc(&cpu_buffer->pages_touched);
1347 * Just make sure we have seen our old_write and synchronize
1348 * with any interrupts that come in.
1353 * If the tail page is still the same as what we think
1354 * it is, then it is up to us to update the tail
1357 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1358 /* Zero the write counter */
1359 unsigned long val = old_write & ~RB_WRITE_MASK;
1360 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1363 * This will only succeed if an interrupt did
1364 * not come in and change it. In which case, we
1365 * do not want to modify it.
1367 * We add (void) to let the compiler know that we do not care
1368 * about the return value of these functions. We use the
1369 * cmpxchg to only update if an interrupt did not already
1370 * do it for us. If the cmpxchg fails, we don't care.
1372 (void)local_cmpxchg(&next_page->write, old_write, val);
1373 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1376 * No need to worry about races with clearing out the commit.
1377 * it only can increment when a commit takes place. But that
1378 * only happens in the outer most nested commit.
1380 local_set(&next_page->page->commit, 0);
1382 /* Again, either we update tail_page or an interrupt does */
1383 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1387 static void rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1388 struct buffer_page *bpage)
1390 unsigned long val = (unsigned long)bpage;
1392 RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK);
1396 * rb_check_pages - integrity check of buffer pages
1397 * @cpu_buffer: CPU buffer with pages to test
1399 * As a safety measure we check to make sure the data pages have not
1402 static void rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1404 struct list_head *head = rb_list_head(cpu_buffer->pages);
1405 struct list_head *tmp;
1407 if (RB_WARN_ON(cpu_buffer,
1408 rb_list_head(rb_list_head(head->next)->prev) != head))
1411 if (RB_WARN_ON(cpu_buffer,
1412 rb_list_head(rb_list_head(head->prev)->next) != head))
1415 for (tmp = rb_list_head(head->next); tmp != head; tmp = rb_list_head(tmp->next)) {
1416 if (RB_WARN_ON(cpu_buffer,
1417 rb_list_head(rb_list_head(tmp->next)->prev) != tmp))
1420 if (RB_WARN_ON(cpu_buffer,
1421 rb_list_head(rb_list_head(tmp->prev)->next) != tmp))
1426 static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1427 long nr_pages, struct list_head *pages)
1429 struct buffer_page *bpage, *tmp;
1430 bool user_thread = current->mm != NULL;
1435 * Check if the available memory is there first.
1436 * Note, si_mem_available() only gives us a rough estimate of available
1437 * memory. It may not be accurate. But we don't care, we just want
1438 * to prevent doing any allocation when it is obvious that it is
1439 * not going to succeed.
1441 i = si_mem_available();
1446 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1447 * gracefully without invoking oom-killer and the system is not
1450 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1453 * If a user thread allocates too much, and si_mem_available()
1454 * reports there's enough memory, even though there is not.
1455 * Make sure the OOM killer kills this thread. This can happen
1456 * even with RETRY_MAYFAIL because another task may be doing
1457 * an allocation after this task has taken all memory.
1458 * This is the task the OOM killer needs to take out during this
1459 * loop, even if it was triggered by an allocation somewhere else.
1462 set_current_oom_origin();
1463 for (i = 0; i < nr_pages; i++) {
1466 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1467 mflags, cpu_to_node(cpu_buffer->cpu));
1471 rb_check_bpage(cpu_buffer, bpage);
1473 list_add(&bpage->list, pages);
1475 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), mflags,
1476 cpu_buffer->buffer->subbuf_order);
1479 bpage->page = page_address(page);
1480 bpage->order = cpu_buffer->buffer->subbuf_order;
1481 rb_init_page(bpage->page);
1483 if (user_thread && fatal_signal_pending(current))
1487 clear_current_oom_origin();
1492 list_for_each_entry_safe(bpage, tmp, pages, list) {
1493 list_del_init(&bpage->list);
1494 free_buffer_page(bpage);
1497 clear_current_oom_origin();
1502 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1503 unsigned long nr_pages)
1509 if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
1513 * The ring buffer page list is a circular list that does not
1514 * start and end with a list head. All page list items point to
1517 cpu_buffer->pages = pages.next;
1520 cpu_buffer->nr_pages = nr_pages;
1522 rb_check_pages(cpu_buffer);
1527 static struct ring_buffer_per_cpu *
1528 rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1530 struct ring_buffer_per_cpu *cpu_buffer;
1531 struct buffer_page *bpage;
1535 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1536 GFP_KERNEL, cpu_to_node(cpu));
1540 cpu_buffer->cpu = cpu;
1541 cpu_buffer->buffer = buffer;
1542 raw_spin_lock_init(&cpu_buffer->reader_lock);
1543 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1544 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1545 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1546 init_completion(&cpu_buffer->update_done);
1547 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1548 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1549 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1551 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1552 GFP_KERNEL, cpu_to_node(cpu));
1554 goto fail_free_buffer;
1556 rb_check_bpage(cpu_buffer, bpage);
1558 cpu_buffer->reader_page = bpage;
1560 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, cpu_buffer->buffer->subbuf_order);
1562 goto fail_free_reader;
1563 bpage->page = page_address(page);
1564 rb_init_page(bpage->page);
1566 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1567 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1569 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1571 goto fail_free_reader;
1573 cpu_buffer->head_page
1574 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1575 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1577 rb_head_page_activate(cpu_buffer);
1582 free_buffer_page(cpu_buffer->reader_page);
1589 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1591 struct list_head *head = cpu_buffer->pages;
1592 struct buffer_page *bpage, *tmp;
1594 irq_work_sync(&cpu_buffer->irq_work.work);
1596 free_buffer_page(cpu_buffer->reader_page);
1599 rb_head_page_deactivate(cpu_buffer);
1601 list_for_each_entry_safe(bpage, tmp, head, list) {
1602 list_del_init(&bpage->list);
1603 free_buffer_page(bpage);
1605 bpage = list_entry(head, struct buffer_page, list);
1606 free_buffer_page(bpage);
1609 free_page((unsigned long)cpu_buffer->free_page);
1615 * __ring_buffer_alloc - allocate a new ring_buffer
1616 * @size: the size in bytes per cpu that is needed.
1617 * @flags: attributes to set for the ring buffer.
1618 * @key: ring buffer reader_lock_key.
1620 * Currently the only flag that is available is the RB_FL_OVERWRITE
1621 * flag. This flag means that the buffer will overwrite old data
1622 * when the buffer wraps. If this flag is not set, the buffer will
1623 * drop data when the tail hits the head.
1625 struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1626 struct lock_class_key *key)
1628 struct trace_buffer *buffer;
1634 /* keep it in its own cache line */
1635 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1640 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1641 goto fail_free_buffer;
1643 /* Default buffer page size - one system page */
1644 buffer->subbuf_order = 0;
1645 buffer->subbuf_size = PAGE_SIZE - BUF_PAGE_HDR_SIZE;
1647 /* Max payload is buffer page size - header (8bytes) */
1648 buffer->max_data_size = buffer->subbuf_size - (sizeof(u32) * 2);
1650 nr_pages = DIV_ROUND_UP(size, buffer->subbuf_size);
1651 buffer->flags = flags;
1652 buffer->clock = trace_clock_local;
1653 buffer->reader_lock_key = key;
1655 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1656 init_waitqueue_head(&buffer->irq_work.waiters);
1658 /* need at least two pages */
1662 buffer->cpus = nr_cpu_ids;
1664 bsize = sizeof(void *) * nr_cpu_ids;
1665 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1667 if (!buffer->buffers)
1668 goto fail_free_cpumask;
1670 cpu = raw_smp_processor_id();
1671 cpumask_set_cpu(cpu, buffer->cpumask);
1672 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1673 if (!buffer->buffers[cpu])
1674 goto fail_free_buffers;
1676 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1678 goto fail_free_buffers;
1680 mutex_init(&buffer->mutex);
1685 for_each_buffer_cpu(buffer, cpu) {
1686 if (buffer->buffers[cpu])
1687 rb_free_cpu_buffer(buffer->buffers[cpu]);
1689 kfree(buffer->buffers);
1692 free_cpumask_var(buffer->cpumask);
1698 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1701 * ring_buffer_free - free a ring buffer.
1702 * @buffer: the buffer to free.
1705 ring_buffer_free(struct trace_buffer *buffer)
1709 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1711 irq_work_sync(&buffer->irq_work.work);
1713 for_each_buffer_cpu(buffer, cpu)
1714 rb_free_cpu_buffer(buffer->buffers[cpu]);
1716 kfree(buffer->buffers);
1717 free_cpumask_var(buffer->cpumask);
1721 EXPORT_SYMBOL_GPL(ring_buffer_free);
1723 void ring_buffer_set_clock(struct trace_buffer *buffer,
1726 buffer->clock = clock;
1729 void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1731 buffer->time_stamp_abs = abs;
1734 bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1736 return buffer->time_stamp_abs;
1739 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1741 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1743 return local_read(&bpage->entries) & RB_WRITE_MASK;
1746 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1748 return local_read(&bpage->write) & RB_WRITE_MASK;
1752 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1754 struct list_head *tail_page, *to_remove, *next_page;
1755 struct buffer_page *to_remove_page, *tmp_iter_page;
1756 struct buffer_page *last_page, *first_page;
1757 unsigned long nr_removed;
1758 unsigned long head_bit;
1763 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1764 atomic_inc(&cpu_buffer->record_disabled);
1766 * We don't race with the readers since we have acquired the reader
1767 * lock. We also don't race with writers after disabling recording.
1768 * This makes it easy to figure out the first and the last page to be
1769 * removed from the list. We unlink all the pages in between including
1770 * the first and last pages. This is done in a busy loop so that we
1771 * lose the least number of traces.
1772 * The pages are freed after we restart recording and unlock readers.
1774 tail_page = &cpu_buffer->tail_page->list;
1777 * tail page might be on reader page, we remove the next page
1778 * from the ring buffer
1780 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1781 tail_page = rb_list_head(tail_page->next);
1782 to_remove = tail_page;
1784 /* start of pages to remove */
1785 first_page = list_entry(rb_list_head(to_remove->next),
1786 struct buffer_page, list);
1788 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1789 to_remove = rb_list_head(to_remove)->next;
1790 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1792 /* Read iterators need to reset themselves when some pages removed */
1793 cpu_buffer->pages_removed += nr_removed;
1795 next_page = rb_list_head(to_remove)->next;
1798 * Now we remove all pages between tail_page and next_page.
1799 * Make sure that we have head_bit value preserved for the
1802 tail_page->next = (struct list_head *)((unsigned long)next_page |
1804 next_page = rb_list_head(next_page);
1805 next_page->prev = tail_page;
1807 /* make sure pages points to a valid page in the ring buffer */
1808 cpu_buffer->pages = next_page;
1810 /* update head page */
1812 cpu_buffer->head_page = list_entry(next_page,
1813 struct buffer_page, list);
1815 /* pages are removed, resume tracing and then free the pages */
1816 atomic_dec(&cpu_buffer->record_disabled);
1817 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1819 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1821 /* last buffer page to remove */
1822 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1824 tmp_iter_page = first_page;
1829 to_remove_page = tmp_iter_page;
1830 rb_inc_page(&tmp_iter_page);
1832 /* update the counters */
1833 page_entries = rb_page_entries(to_remove_page);
1836 * If something was added to this page, it was full
1837 * since it is not the tail page. So we deduct the
1838 * bytes consumed in ring buffer from here.
1839 * Increment overrun to account for the lost events.
1841 local_add(page_entries, &cpu_buffer->overrun);
1842 local_sub(rb_page_commit(to_remove_page), &cpu_buffer->entries_bytes);
1843 local_inc(&cpu_buffer->pages_lost);
1847 * We have already removed references to this list item, just
1848 * free up the buffer_page and its page
1850 free_buffer_page(to_remove_page);
1853 } while (to_remove_page != last_page);
1855 RB_WARN_ON(cpu_buffer, nr_removed);
1857 return nr_removed == 0;
1861 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1863 struct list_head *pages = &cpu_buffer->new_pages;
1864 unsigned long flags;
1868 /* Can be called at early boot up, where interrupts must not been enabled */
1869 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1871 * We are holding the reader lock, so the reader page won't be swapped
1872 * in the ring buffer. Now we are racing with the writer trying to
1873 * move head page and the tail page.
1874 * We are going to adapt the reader page update process where:
1875 * 1. We first splice the start and end of list of new pages between
1876 * the head page and its previous page.
1877 * 2. We cmpxchg the prev_page->next to point from head page to the
1878 * start of new pages list.
1879 * 3. Finally, we update the head->prev to the end of new list.
1881 * We will try this process 10 times, to make sure that we don't keep
1887 struct list_head *head_page, *prev_page;
1888 struct list_head *last_page, *first_page;
1889 struct list_head *head_page_with_bit;
1890 struct buffer_page *hpage = rb_set_head_page(cpu_buffer);
1894 head_page = &hpage->list;
1895 prev_page = head_page->prev;
1897 first_page = pages->next;
1898 last_page = pages->prev;
1900 head_page_with_bit = (struct list_head *)
1901 ((unsigned long)head_page | RB_PAGE_HEAD);
1903 last_page->next = head_page_with_bit;
1904 first_page->prev = prev_page;
1906 /* caution: head_page_with_bit gets updated on cmpxchg failure */
1907 if (try_cmpxchg(&prev_page->next,
1908 &head_page_with_bit, first_page)) {
1910 * yay, we replaced the page pointer to our new list,
1911 * now, we just have to update to head page's prev
1912 * pointer to point to end of list
1914 head_page->prev = last_page;
1921 INIT_LIST_HEAD(pages);
1923 * If we weren't successful in adding in new pages, warn and stop
1926 RB_WARN_ON(cpu_buffer, !success);
1927 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1929 /* free pages if they weren't inserted */
1931 struct buffer_page *bpage, *tmp;
1932 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1934 list_del_init(&bpage->list);
1935 free_buffer_page(bpage);
1941 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1945 if (cpu_buffer->nr_pages_to_update > 0)
1946 success = rb_insert_pages(cpu_buffer);
1948 success = rb_remove_pages(cpu_buffer,
1949 -cpu_buffer->nr_pages_to_update);
1952 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1955 static void update_pages_handler(struct work_struct *work)
1957 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1958 struct ring_buffer_per_cpu, update_pages_work);
1959 rb_update_pages(cpu_buffer);
1960 complete(&cpu_buffer->update_done);
1964 * ring_buffer_resize - resize the ring buffer
1965 * @buffer: the buffer to resize.
1966 * @size: the new size.
1967 * @cpu_id: the cpu buffer to resize
1969 * Minimum size is 2 * buffer->subbuf_size.
1971 * Returns 0 on success and < 0 on failure.
1973 int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
1976 struct ring_buffer_per_cpu *cpu_buffer;
1977 unsigned long nr_pages;
1981 * Always succeed at resizing a non-existent buffer:
1986 /* Make sure the requested buffer exists */
1987 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1988 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1991 nr_pages = DIV_ROUND_UP(size, buffer->subbuf_size);
1993 /* we need a minimum of two pages */
1997 /* prevent another thread from changing buffer sizes */
1998 mutex_lock(&buffer->mutex);
1999 atomic_inc(&buffer->resizing);
2001 if (cpu_id == RING_BUFFER_ALL_CPUS) {
2003 * Don't succeed if resizing is disabled, as a reader might be
2004 * manipulating the ring buffer and is expecting a sane state while
2007 for_each_buffer_cpu(buffer, cpu) {
2008 cpu_buffer = buffer->buffers[cpu];
2009 if (atomic_read(&cpu_buffer->resize_disabled)) {
2011 goto out_err_unlock;
2015 /* calculate the pages to update */
2016 for_each_buffer_cpu(buffer, cpu) {
2017 cpu_buffer = buffer->buffers[cpu];
2019 cpu_buffer->nr_pages_to_update = nr_pages -
2020 cpu_buffer->nr_pages;
2022 * nothing more to do for removing pages or no update
2024 if (cpu_buffer->nr_pages_to_update <= 0)
2027 * to add pages, make sure all new pages can be
2028 * allocated without receiving ENOMEM
2030 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2031 if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2032 &cpu_buffer->new_pages)) {
2033 /* not enough memory for new pages */
2043 * Fire off all the required work handlers
2044 * We can't schedule on offline CPUs, but it's not necessary
2045 * since we can change their buffer sizes without any race.
2047 for_each_buffer_cpu(buffer, cpu) {
2048 cpu_buffer = buffer->buffers[cpu];
2049 if (!cpu_buffer->nr_pages_to_update)
2052 /* Can't run something on an offline CPU. */
2053 if (!cpu_online(cpu)) {
2054 rb_update_pages(cpu_buffer);
2055 cpu_buffer->nr_pages_to_update = 0;
2057 /* Run directly if possible. */
2059 if (cpu != smp_processor_id()) {
2061 schedule_work_on(cpu,
2062 &cpu_buffer->update_pages_work);
2064 update_pages_handler(&cpu_buffer->update_pages_work);
2070 /* wait for all the updates to complete */
2071 for_each_buffer_cpu(buffer, cpu) {
2072 cpu_buffer = buffer->buffers[cpu];
2073 if (!cpu_buffer->nr_pages_to_update)
2076 if (cpu_online(cpu))
2077 wait_for_completion(&cpu_buffer->update_done);
2078 cpu_buffer->nr_pages_to_update = 0;
2083 cpu_buffer = buffer->buffers[cpu_id];
2085 if (nr_pages == cpu_buffer->nr_pages)
2089 * Don't succeed if resizing is disabled, as a reader might be
2090 * manipulating the ring buffer and is expecting a sane state while
2093 if (atomic_read(&cpu_buffer->resize_disabled)) {
2095 goto out_err_unlock;
2098 cpu_buffer->nr_pages_to_update = nr_pages -
2099 cpu_buffer->nr_pages;
2101 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2102 if (cpu_buffer->nr_pages_to_update > 0 &&
2103 __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2104 &cpu_buffer->new_pages)) {
2111 /* Can't run something on an offline CPU. */
2112 if (!cpu_online(cpu_id))
2113 rb_update_pages(cpu_buffer);
2115 /* Run directly if possible. */
2117 if (cpu_id == smp_processor_id()) {
2118 rb_update_pages(cpu_buffer);
2122 schedule_work_on(cpu_id,
2123 &cpu_buffer->update_pages_work);
2124 wait_for_completion(&cpu_buffer->update_done);
2128 cpu_buffer->nr_pages_to_update = 0;
2134 * The ring buffer resize can happen with the ring buffer
2135 * enabled, so that the update disturbs the tracing as little
2136 * as possible. But if the buffer is disabled, we do not need
2137 * to worry about that, and we can take the time to verify
2138 * that the buffer is not corrupt.
2140 if (atomic_read(&buffer->record_disabled)) {
2141 atomic_inc(&buffer->record_disabled);
2143 * Even though the buffer was disabled, we must make sure
2144 * that it is truly disabled before calling rb_check_pages.
2145 * There could have been a race between checking
2146 * record_disable and incrementing it.
2149 for_each_buffer_cpu(buffer, cpu) {
2150 cpu_buffer = buffer->buffers[cpu];
2151 rb_check_pages(cpu_buffer);
2153 atomic_dec(&buffer->record_disabled);
2156 atomic_dec(&buffer->resizing);
2157 mutex_unlock(&buffer->mutex);
2161 for_each_buffer_cpu(buffer, cpu) {
2162 struct buffer_page *bpage, *tmp;
2164 cpu_buffer = buffer->buffers[cpu];
2165 cpu_buffer->nr_pages_to_update = 0;
2167 if (list_empty(&cpu_buffer->new_pages))
2170 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2172 list_del_init(&bpage->list);
2173 free_buffer_page(bpage);
2177 atomic_dec(&buffer->resizing);
2178 mutex_unlock(&buffer->mutex);
2181 EXPORT_SYMBOL_GPL(ring_buffer_resize);
2183 void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2185 mutex_lock(&buffer->mutex);
2187 buffer->flags |= RB_FL_OVERWRITE;
2189 buffer->flags &= ~RB_FL_OVERWRITE;
2190 mutex_unlock(&buffer->mutex);
2192 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2194 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2196 return bpage->page->data + index;
2199 static __always_inline struct ring_buffer_event *
2200 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2202 return __rb_page_index(cpu_buffer->reader_page,
2203 cpu_buffer->reader_page->read);
2206 static struct ring_buffer_event *
2207 rb_iter_head_event(struct ring_buffer_iter *iter)
2209 struct ring_buffer_event *event;
2210 struct buffer_page *iter_head_page = iter->head_page;
2211 unsigned long commit;
2214 if (iter->head != iter->next_event)
2218 * When the writer goes across pages, it issues a cmpxchg which
2219 * is a mb(), which will synchronize with the rmb here.
2220 * (see rb_tail_page_update() and __rb_reserve_next())
2222 commit = rb_page_commit(iter_head_page);
2225 /* An event needs to be at least 8 bytes in size */
2226 if (iter->head > commit - 8)
2229 event = __rb_page_index(iter_head_page, iter->head);
2230 length = rb_event_length(event);
2233 * READ_ONCE() doesn't work on functions and we don't want the
2234 * compiler doing any crazy optimizations with length.
2238 if ((iter->head + length) > commit || length > iter->event_size)
2239 /* Writer corrupted the read? */
2242 memcpy(iter->event, event, length);
2244 * If the page stamp is still the same after this rmb() then the
2245 * event was safely copied without the writer entering the page.
2249 /* Make sure the page didn't change since we read this */
2250 if (iter->page_stamp != iter_head_page->page->time_stamp ||
2251 commit > rb_page_commit(iter_head_page))
2254 iter->next_event = iter->head + length;
2257 /* Reset to the beginning */
2258 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2260 iter->next_event = 0;
2261 iter->missed_events = 1;
2265 /* Size is determined by what has been committed */
2266 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2268 return rb_page_commit(bpage);
2271 static __always_inline unsigned
2272 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2274 return rb_page_commit(cpu_buffer->commit_page);
2277 static __always_inline unsigned
2278 rb_event_index(struct ring_buffer_per_cpu *cpu_buffer, struct ring_buffer_event *event)
2280 unsigned long addr = (unsigned long)event;
2282 addr &= (PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1;
2284 return addr - BUF_PAGE_HDR_SIZE;
2287 static void rb_inc_iter(struct ring_buffer_iter *iter)
2289 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2292 * The iterator could be on the reader page (it starts there).
2293 * But the head could have moved, since the reader was
2294 * found. Check for this case and assign the iterator
2295 * to the head page instead of next.
2297 if (iter->head_page == cpu_buffer->reader_page)
2298 iter->head_page = rb_set_head_page(cpu_buffer);
2300 rb_inc_page(&iter->head_page);
2302 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2304 iter->next_event = 0;
2308 * rb_handle_head_page - writer hit the head page
2310 * Returns: +1 to retry page
2315 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2316 struct buffer_page *tail_page,
2317 struct buffer_page *next_page)
2319 struct buffer_page *new_head;
2324 entries = rb_page_entries(next_page);
2327 * The hard part is here. We need to move the head
2328 * forward, and protect against both readers on
2329 * other CPUs and writers coming in via interrupts.
2331 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2335 * type can be one of four:
2336 * NORMAL - an interrupt already moved it for us
2337 * HEAD - we are the first to get here.
2338 * UPDATE - we are the interrupt interrupting
2340 * MOVED - a reader on another CPU moved the next
2341 * pointer to its reader page. Give up
2348 * We changed the head to UPDATE, thus
2349 * it is our responsibility to update
2352 local_add(entries, &cpu_buffer->overrun);
2353 local_sub(rb_page_commit(next_page), &cpu_buffer->entries_bytes);
2354 local_inc(&cpu_buffer->pages_lost);
2357 * The entries will be zeroed out when we move the
2361 /* still more to do */
2364 case RB_PAGE_UPDATE:
2366 * This is an interrupt that interrupt the
2367 * previous update. Still more to do.
2370 case RB_PAGE_NORMAL:
2372 * An interrupt came in before the update
2373 * and processed this for us.
2374 * Nothing left to do.
2379 * The reader is on another CPU and just did
2380 * a swap with our next_page.
2385 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2390 * Now that we are here, the old head pointer is
2391 * set to UPDATE. This will keep the reader from
2392 * swapping the head page with the reader page.
2393 * The reader (on another CPU) will spin till
2396 * We just need to protect against interrupts
2397 * doing the job. We will set the next pointer
2398 * to HEAD. After that, we set the old pointer
2399 * to NORMAL, but only if it was HEAD before.
2400 * otherwise we are an interrupt, and only
2401 * want the outer most commit to reset it.
2403 new_head = next_page;
2404 rb_inc_page(&new_head);
2406 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2410 * Valid returns are:
2411 * HEAD - an interrupt came in and already set it.
2412 * NORMAL - One of two things:
2413 * 1) We really set it.
2414 * 2) A bunch of interrupts came in and moved
2415 * the page forward again.
2419 case RB_PAGE_NORMAL:
2423 RB_WARN_ON(cpu_buffer, 1);
2428 * It is possible that an interrupt came in,
2429 * set the head up, then more interrupts came in
2430 * and moved it again. When we get back here,
2431 * the page would have been set to NORMAL but we
2432 * just set it back to HEAD.
2434 * How do you detect this? Well, if that happened
2435 * the tail page would have moved.
2437 if (ret == RB_PAGE_NORMAL) {
2438 struct buffer_page *buffer_tail_page;
2440 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2442 * If the tail had moved passed next, then we need
2443 * to reset the pointer.
2445 if (buffer_tail_page != tail_page &&
2446 buffer_tail_page != next_page)
2447 rb_head_page_set_normal(cpu_buffer, new_head,
2453 * If this was the outer most commit (the one that
2454 * changed the original pointer from HEAD to UPDATE),
2455 * then it is up to us to reset it to NORMAL.
2457 if (type == RB_PAGE_HEAD) {
2458 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2461 if (RB_WARN_ON(cpu_buffer,
2462 ret != RB_PAGE_UPDATE))
2470 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2471 unsigned long tail, struct rb_event_info *info)
2473 unsigned long bsize = READ_ONCE(cpu_buffer->buffer->subbuf_size);
2474 struct buffer_page *tail_page = info->tail_page;
2475 struct ring_buffer_event *event;
2476 unsigned long length = info->length;
2479 * Only the event that crossed the page boundary
2480 * must fill the old tail_page with padding.
2482 if (tail >= bsize) {
2484 * If the page was filled, then we still need
2485 * to update the real_end. Reset it to zero
2486 * and the reader will ignore it.
2489 tail_page->real_end = 0;
2491 local_sub(length, &tail_page->write);
2495 event = __rb_page_index(tail_page, tail);
2498 * Save the original length to the meta data.
2499 * This will be used by the reader to add lost event
2502 tail_page->real_end = tail;
2505 * If this event is bigger than the minimum size, then
2506 * we need to be careful that we don't subtract the
2507 * write counter enough to allow another writer to slip
2509 * We put in a discarded commit instead, to make sure
2510 * that this space is not used again, and this space will
2511 * not be accounted into 'entries_bytes'.
2513 * If we are less than the minimum size, we don't need to
2516 if (tail > (bsize - RB_EVNT_MIN_SIZE)) {
2517 /* No room for any events */
2519 /* Mark the rest of the page with padding */
2520 rb_event_set_padding(event);
2522 /* Make sure the padding is visible before the write update */
2525 /* Set the write back to the previous setting */
2526 local_sub(length, &tail_page->write);
2530 /* Put in a discarded event */
2531 event->array[0] = (bsize - tail) - RB_EVNT_HDR_SIZE;
2532 event->type_len = RINGBUF_TYPE_PADDING;
2533 /* time delta must be non zero */
2534 event->time_delta = 1;
2536 /* account for padding bytes */
2537 local_add(bsize - tail, &cpu_buffer->entries_bytes);
2539 /* Make sure the padding is visible before the tail_page->write update */
2542 /* Set write to end of buffer */
2543 length = (tail + length) - bsize;
2544 local_sub(length, &tail_page->write);
2547 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2550 * This is the slow path, force gcc not to inline it.
2552 static noinline struct ring_buffer_event *
2553 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2554 unsigned long tail, struct rb_event_info *info)
2556 struct buffer_page *tail_page = info->tail_page;
2557 struct buffer_page *commit_page = cpu_buffer->commit_page;
2558 struct trace_buffer *buffer = cpu_buffer->buffer;
2559 struct buffer_page *next_page;
2562 next_page = tail_page;
2564 rb_inc_page(&next_page);
2567 * If for some reason, we had an interrupt storm that made
2568 * it all the way around the buffer, bail, and warn
2571 if (unlikely(next_page == commit_page)) {
2572 local_inc(&cpu_buffer->commit_overrun);
2577 * This is where the fun begins!
2579 * We are fighting against races between a reader that
2580 * could be on another CPU trying to swap its reader
2581 * page with the buffer head.
2583 * We are also fighting against interrupts coming in and
2584 * moving the head or tail on us as well.
2586 * If the next page is the head page then we have filled
2587 * the buffer, unless the commit page is still on the
2590 if (rb_is_head_page(next_page, &tail_page->list)) {
2593 * If the commit is not on the reader page, then
2594 * move the header page.
2596 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2598 * If we are not in overwrite mode,
2599 * this is easy, just stop here.
2601 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2602 local_inc(&cpu_buffer->dropped_events);
2606 ret = rb_handle_head_page(cpu_buffer,
2615 * We need to be careful here too. The
2616 * commit page could still be on the reader
2617 * page. We could have a small buffer, and
2618 * have filled up the buffer with events
2619 * from interrupts and such, and wrapped.
2621 * Note, if the tail page is also on the
2622 * reader_page, we let it move out.
2624 if (unlikely((cpu_buffer->commit_page !=
2625 cpu_buffer->tail_page) &&
2626 (cpu_buffer->commit_page ==
2627 cpu_buffer->reader_page))) {
2628 local_inc(&cpu_buffer->commit_overrun);
2634 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2638 rb_reset_tail(cpu_buffer, tail, info);
2640 /* Commit what we have for now. */
2641 rb_end_commit(cpu_buffer);
2642 /* rb_end_commit() decs committing */
2643 local_inc(&cpu_buffer->committing);
2645 /* fail and let the caller try again */
2646 return ERR_PTR(-EAGAIN);
2650 rb_reset_tail(cpu_buffer, tail, info);
2656 static struct ring_buffer_event *
2657 rb_add_time_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2658 struct ring_buffer_event *event, u64 delta, bool abs)
2661 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2663 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2665 /* Not the first event on the page, or not delta? */
2666 if (abs || rb_event_index(cpu_buffer, event)) {
2667 event->time_delta = delta & TS_MASK;
2668 event->array[0] = delta >> TS_SHIFT;
2670 /* nope, just zero it */
2671 event->time_delta = 0;
2672 event->array[0] = 0;
2675 return skip_time_extend(event);
2678 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2679 static inline bool sched_clock_stable(void)
2686 rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2687 struct rb_event_info *info)
2691 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2692 (unsigned long long)info->delta,
2693 (unsigned long long)info->ts,
2694 (unsigned long long)info->before,
2695 (unsigned long long)info->after,
2696 (unsigned long long)({rb_time_read(&cpu_buffer->write_stamp, &write_stamp); write_stamp;}),
2697 sched_clock_stable() ? "" :
2698 "If you just came from a suspend/resume,\n"
2699 "please switch to the trace global clock:\n"
2700 " echo global > /sys/kernel/tracing/trace_clock\n"
2701 "or add trace_clock=global to the kernel command line\n");
2704 static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2705 struct ring_buffer_event **event,
2706 struct rb_event_info *info,
2708 unsigned int *length)
2710 bool abs = info->add_timestamp &
2711 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2713 if (unlikely(info->delta > (1ULL << 59))) {
2715 * Some timers can use more than 59 bits, and when a timestamp
2716 * is added to the buffer, it will lose those bits.
2718 if (abs && (info->ts & TS_MSB)) {
2719 info->delta &= ABS_TS_MASK;
2721 /* did the clock go backwards */
2722 } else if (info->before == info->after && info->before > info->ts) {
2723 /* not interrupted */
2727 * This is possible with a recalibrating of the TSC.
2728 * Do not produce a call stack, but just report it.
2732 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2733 info->before, info->ts);
2736 rb_check_timestamp(cpu_buffer, info);
2740 *event = rb_add_time_stamp(cpu_buffer, *event, info->delta, abs);
2741 *length -= RB_LEN_TIME_EXTEND;
2746 * rb_update_event - update event type and data
2747 * @cpu_buffer: The per cpu buffer of the @event
2748 * @event: the event to update
2749 * @info: The info to update the @event with (contains length and delta)
2751 * Update the type and data fields of the @event. The length
2752 * is the actual size that is written to the ring buffer,
2753 * and with this, we can determine what to place into the
2757 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2758 struct ring_buffer_event *event,
2759 struct rb_event_info *info)
2761 unsigned length = info->length;
2762 u64 delta = info->delta;
2763 unsigned int nest = local_read(&cpu_buffer->committing) - 1;
2765 if (!WARN_ON_ONCE(nest >= MAX_NEST))
2766 cpu_buffer->event_stamp[nest] = info->ts;
2769 * If we need to add a timestamp, then we
2770 * add it to the start of the reserved space.
2772 if (unlikely(info->add_timestamp))
2773 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2775 event->time_delta = delta;
2776 length -= RB_EVNT_HDR_SIZE;
2777 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2778 event->type_len = 0;
2779 event->array[0] = length;
2781 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2784 static unsigned rb_calculate_event_length(unsigned length)
2786 struct ring_buffer_event event; /* Used only for sizeof array */
2788 /* zero length can cause confusions */
2792 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2793 length += sizeof(event.array[0]);
2795 length += RB_EVNT_HDR_SIZE;
2796 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2799 * In case the time delta is larger than the 27 bits for it
2800 * in the header, we need to add a timestamp. If another
2801 * event comes in when trying to discard this one to increase
2802 * the length, then the timestamp will be added in the allocated
2803 * space of this event. If length is bigger than the size needed
2804 * for the TIME_EXTEND, then padding has to be used. The events
2805 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2806 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2807 * As length is a multiple of 4, we only need to worry if it
2808 * is 12 (RB_LEN_TIME_EXTEND + 4).
2810 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2811 length += RB_ALIGNMENT;
2817 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2818 struct ring_buffer_event *event)
2820 unsigned long new_index, old_index;
2821 struct buffer_page *bpage;
2824 new_index = rb_event_index(cpu_buffer, event);
2825 old_index = new_index + rb_event_ts_length(event);
2826 addr = (unsigned long)event;
2827 addr &= ~((PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1);
2829 bpage = READ_ONCE(cpu_buffer->tail_page);
2832 * Make sure the tail_page is still the same and
2833 * the next write location is the end of this event
2835 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2836 unsigned long write_mask =
2837 local_read(&bpage->write) & ~RB_WRITE_MASK;
2838 unsigned long event_length = rb_event_length(event);
2841 * For the before_stamp to be different than the write_stamp
2842 * to make sure that the next event adds an absolute
2843 * value and does not rely on the saved write stamp, which
2844 * is now going to be bogus.
2846 * By setting the before_stamp to zero, the next event
2847 * is not going to use the write_stamp and will instead
2848 * create an absolute timestamp. This means there's no
2849 * reason to update the wirte_stamp!
2851 rb_time_set(&cpu_buffer->before_stamp, 0);
2854 * If an event were to come in now, it would see that the
2855 * write_stamp and the before_stamp are different, and assume
2856 * that this event just added itself before updating
2857 * the write stamp. The interrupting event will fix the
2858 * write stamp for us, and use an absolute timestamp.
2862 * This is on the tail page. It is possible that
2863 * a write could come in and move the tail page
2864 * and write to the next page. That is fine
2865 * because we just shorten what is on this page.
2867 old_index += write_mask;
2868 new_index += write_mask;
2870 /* caution: old_index gets updated on cmpxchg failure */
2871 if (local_try_cmpxchg(&bpage->write, &old_index, new_index)) {
2872 /* update counters */
2873 local_sub(event_length, &cpu_buffer->entries_bytes);
2878 /* could not discard */
2882 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2884 local_inc(&cpu_buffer->committing);
2885 local_inc(&cpu_buffer->commits);
2888 static __always_inline void
2889 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2891 unsigned long max_count;
2894 * We only race with interrupts and NMIs on this CPU.
2895 * If we own the commit event, then we can commit
2896 * all others that interrupted us, since the interruptions
2897 * are in stack format (they finish before they come
2898 * back to us). This allows us to do a simple loop to
2899 * assign the commit to the tail.
2902 max_count = cpu_buffer->nr_pages * 100;
2904 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2905 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2907 if (RB_WARN_ON(cpu_buffer,
2908 rb_is_reader_page(cpu_buffer->tail_page)))
2911 * No need for a memory barrier here, as the update
2912 * of the tail_page did it for this page.
2914 local_set(&cpu_buffer->commit_page->page->commit,
2915 rb_page_write(cpu_buffer->commit_page));
2916 rb_inc_page(&cpu_buffer->commit_page);
2917 /* add barrier to keep gcc from optimizing too much */
2920 while (rb_commit_index(cpu_buffer) !=
2921 rb_page_write(cpu_buffer->commit_page)) {
2923 /* Make sure the readers see the content of what is committed. */
2925 local_set(&cpu_buffer->commit_page->page->commit,
2926 rb_page_write(cpu_buffer->commit_page));
2927 RB_WARN_ON(cpu_buffer,
2928 local_read(&cpu_buffer->commit_page->page->commit) &
2933 /* again, keep gcc from optimizing */
2937 * If an interrupt came in just after the first while loop
2938 * and pushed the tail page forward, we will be left with
2939 * a dangling commit that will never go forward.
2941 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2945 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2947 unsigned long commits;
2949 if (RB_WARN_ON(cpu_buffer,
2950 !local_read(&cpu_buffer->committing)))
2954 commits = local_read(&cpu_buffer->commits);
2955 /* synchronize with interrupts */
2957 if (local_read(&cpu_buffer->committing) == 1)
2958 rb_set_commit_to_write(cpu_buffer);
2960 local_dec(&cpu_buffer->committing);
2962 /* synchronize with interrupts */
2966 * Need to account for interrupts coming in between the
2967 * updating of the commit page and the clearing of the
2968 * committing counter.
2970 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2971 !local_read(&cpu_buffer->committing)) {
2972 local_inc(&cpu_buffer->committing);
2977 static inline void rb_event_discard(struct ring_buffer_event *event)
2979 if (extended_time(event))
2980 event = skip_time_extend(event);
2982 /* array[0] holds the actual length for the discarded event */
2983 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2984 event->type_len = RINGBUF_TYPE_PADDING;
2985 /* time delta must be non zero */
2986 if (!event->time_delta)
2987 event->time_delta = 1;
2990 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer)
2992 local_inc(&cpu_buffer->entries);
2993 rb_end_commit(cpu_buffer);
2996 static __always_inline void
2997 rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2999 if (buffer->irq_work.waiters_pending) {
3000 buffer->irq_work.waiters_pending = false;
3001 /* irq_work_queue() supplies it's own memory barriers */
3002 irq_work_queue(&buffer->irq_work.work);
3005 if (cpu_buffer->irq_work.waiters_pending) {
3006 cpu_buffer->irq_work.waiters_pending = false;
3007 /* irq_work_queue() supplies it's own memory barriers */
3008 irq_work_queue(&cpu_buffer->irq_work.work);
3011 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3014 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3017 if (!cpu_buffer->irq_work.full_waiters_pending)
3020 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3022 if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
3025 cpu_buffer->irq_work.wakeup_full = true;
3026 cpu_buffer->irq_work.full_waiters_pending = false;
3027 /* irq_work_queue() supplies it's own memory barriers */
3028 irq_work_queue(&cpu_buffer->irq_work.work);
3031 #ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3032 # define do_ring_buffer_record_recursion() \
3033 do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3035 # define do_ring_buffer_record_recursion() do { } while (0)
3039 * The lock and unlock are done within a preempt disable section.
3040 * The current_context per_cpu variable can only be modified
3041 * by the current task between lock and unlock. But it can
3042 * be modified more than once via an interrupt. To pass this
3043 * information from the lock to the unlock without having to
3044 * access the 'in_interrupt()' functions again (which do show
3045 * a bit of overhead in something as critical as function tracing,
3046 * we use a bitmask trick.
3048 * bit 1 = NMI context
3049 * bit 2 = IRQ context
3050 * bit 3 = SoftIRQ context
3051 * bit 4 = normal context.
3053 * This works because this is the order of contexts that can
3054 * preempt other contexts. A SoftIRQ never preempts an IRQ
3057 * When the context is determined, the corresponding bit is
3058 * checked and set (if it was set, then a recursion of that context
3061 * On unlock, we need to clear this bit. To do so, just subtract
3062 * 1 from the current_context and AND it to itself.
3066 * 101 & 100 = 100 (clearing bit zero)
3069 * 1010 & 1001 = 1000 (clearing bit 1)
3071 * The least significant bit can be cleared this way, and it
3072 * just so happens that it is the same bit corresponding to
3073 * the current context.
3075 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3076 * is set when a recursion is detected at the current context, and if
3077 * the TRANSITION bit is already set, it will fail the recursion.
3078 * This is needed because there's a lag between the changing of
3079 * interrupt context and updating the preempt count. In this case,
3080 * a false positive will be found. To handle this, one extra recursion
3081 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3082 * bit is already set, then it is considered a recursion and the function
3083 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3085 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3086 * to be cleared. Even if it wasn't the context that set it. That is,
3087 * if an interrupt comes in while NORMAL bit is set and the ring buffer
3088 * is called before preempt_count() is updated, since the check will
3089 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3090 * NMI then comes in, it will set the NMI bit, but when the NMI code
3091 * does the trace_recursive_unlock() it will clear the TRANSITION bit
3092 * and leave the NMI bit set. But this is fine, because the interrupt
3093 * code that set the TRANSITION bit will then clear the NMI bit when it
3094 * calls trace_recursive_unlock(). If another NMI comes in, it will
3095 * set the TRANSITION bit and continue.
3097 * Note: The TRANSITION bit only handles a single transition between context.
3100 static __always_inline bool
3101 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3103 unsigned int val = cpu_buffer->current_context;
3104 int bit = interrupt_context_level();
3106 bit = RB_CTX_NORMAL - bit;
3108 if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3110 * It is possible that this was called by transitioning
3111 * between interrupt context, and preempt_count() has not
3112 * been updated yet. In this case, use the TRANSITION bit.
3114 bit = RB_CTX_TRANSITION;
3115 if (val & (1 << (bit + cpu_buffer->nest))) {
3116 do_ring_buffer_record_recursion();
3121 val |= (1 << (bit + cpu_buffer->nest));
3122 cpu_buffer->current_context = val;
3127 static __always_inline void
3128 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3130 cpu_buffer->current_context &=
3131 cpu_buffer->current_context - (1 << cpu_buffer->nest);
3134 /* The recursive locking above uses 5 bits */
3135 #define NESTED_BITS 5
3138 * ring_buffer_nest_start - Allow to trace while nested
3139 * @buffer: The ring buffer to modify
3141 * The ring buffer has a safety mechanism to prevent recursion.
3142 * But there may be a case where a trace needs to be done while
3143 * tracing something else. In this case, calling this function
3144 * will allow this function to nest within a currently active
3145 * ring_buffer_lock_reserve().
3147 * Call this function before calling another ring_buffer_lock_reserve() and
3148 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3150 void ring_buffer_nest_start(struct trace_buffer *buffer)
3152 struct ring_buffer_per_cpu *cpu_buffer;
3155 /* Enabled by ring_buffer_nest_end() */
3156 preempt_disable_notrace();
3157 cpu = raw_smp_processor_id();
3158 cpu_buffer = buffer->buffers[cpu];
3159 /* This is the shift value for the above recursive locking */
3160 cpu_buffer->nest += NESTED_BITS;
3164 * ring_buffer_nest_end - Allow to trace while nested
3165 * @buffer: The ring buffer to modify
3167 * Must be called after ring_buffer_nest_start() and after the
3168 * ring_buffer_unlock_commit().
3170 void ring_buffer_nest_end(struct trace_buffer *buffer)
3172 struct ring_buffer_per_cpu *cpu_buffer;
3175 /* disabled by ring_buffer_nest_start() */
3176 cpu = raw_smp_processor_id();
3177 cpu_buffer = buffer->buffers[cpu];
3178 /* This is the shift value for the above recursive locking */
3179 cpu_buffer->nest -= NESTED_BITS;
3180 preempt_enable_notrace();
3184 * ring_buffer_unlock_commit - commit a reserved
3185 * @buffer: The buffer to commit to
3187 * This commits the data to the ring buffer, and releases any locks held.
3189 * Must be paired with ring_buffer_lock_reserve.
3191 int ring_buffer_unlock_commit(struct trace_buffer *buffer)
3193 struct ring_buffer_per_cpu *cpu_buffer;
3194 int cpu = raw_smp_processor_id();
3196 cpu_buffer = buffer->buffers[cpu];
3198 rb_commit(cpu_buffer);
3200 rb_wakeups(buffer, cpu_buffer);
3202 trace_recursive_unlock(cpu_buffer);
3204 preempt_enable_notrace();
3208 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3210 /* Special value to validate all deltas on a page. */
3211 #define CHECK_FULL_PAGE 1L
3213 #ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
3215 static const char *show_irq_str(int bits)
3217 const char *type[] = {
3231 /* Assume this is an trace event */
3232 static const char *show_flags(struct ring_buffer_event *event)
3234 struct trace_entry *entry;
3237 if (rb_event_data_length(event) - RB_EVNT_HDR_SIZE < sizeof(*entry))
3240 entry = ring_buffer_event_data(event);
3242 if (entry->flags & TRACE_FLAG_SOFTIRQ)
3245 if (entry->flags & TRACE_FLAG_HARDIRQ)
3248 if (entry->flags & TRACE_FLAG_NMI)
3251 return show_irq_str(bits);
3254 static const char *show_irq(struct ring_buffer_event *event)
3256 struct trace_entry *entry;
3258 if (rb_event_data_length(event) - RB_EVNT_HDR_SIZE < sizeof(*entry))
3261 entry = ring_buffer_event_data(event);
3262 if (entry->flags & TRACE_FLAG_IRQS_OFF)
3267 static const char *show_interrupt_level(void)
3269 unsigned long pc = preempt_count();
3270 unsigned char level = 0;
3272 if (pc & SOFTIRQ_OFFSET)
3275 if (pc & HARDIRQ_MASK)
3281 return show_irq_str(level);
3284 static void dump_buffer_page(struct buffer_data_page *bpage,
3285 struct rb_event_info *info,
3288 struct ring_buffer_event *event;
3292 ts = bpage->time_stamp;
3293 pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
3295 for (e = 0; e < tail; e += rb_event_length(event)) {
3297 event = (struct ring_buffer_event *)(bpage->data + e);
3299 switch (event->type_len) {
3301 case RINGBUF_TYPE_TIME_EXTEND:
3302 delta = rb_event_time_stamp(event);
3304 pr_warn(" 0x%x: [%lld] delta:%lld TIME EXTEND\n",
3308 case RINGBUF_TYPE_TIME_STAMP:
3309 delta = rb_event_time_stamp(event);
3310 ts = rb_fix_abs_ts(delta, ts);
3311 pr_warn(" 0x%x: [%lld] absolute:%lld TIME STAMP\n",
3315 case RINGBUF_TYPE_PADDING:
3316 ts += event->time_delta;
3317 pr_warn(" 0x%x: [%lld] delta:%d PADDING\n",
3318 e, ts, event->time_delta);
3321 case RINGBUF_TYPE_DATA:
3322 ts += event->time_delta;
3323 pr_warn(" 0x%x: [%lld] delta:%d %s%s\n",
3324 e, ts, event->time_delta,
3325 show_flags(event), show_irq(event));
3332 pr_warn("expected end:0x%lx last event actually ended at:0x%x\n", tail, e);
3335 static DEFINE_PER_CPU(atomic_t, checking);
3336 static atomic_t ts_dump;
3338 #define buffer_warn_return(fmt, ...) \
3340 /* If another report is happening, ignore this one */ \
3341 if (atomic_inc_return(&ts_dump) != 1) { \
3342 atomic_dec(&ts_dump); \
3345 atomic_inc(&cpu_buffer->record_disabled); \
3346 pr_warn(fmt, ##__VA_ARGS__); \
3347 dump_buffer_page(bpage, info, tail); \
3348 atomic_dec(&ts_dump); \
3349 /* There's some cases in boot up that this can happen */ \
3350 if (WARN_ON_ONCE(system_state != SYSTEM_BOOTING)) \
3351 /* Do not re-enable checking */ \
3356 * Check if the current event time stamp matches the deltas on
3359 static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3360 struct rb_event_info *info,
3363 struct ring_buffer_event *event;
3364 struct buffer_data_page *bpage;
3369 bpage = info->tail_page->page;
3371 if (tail == CHECK_FULL_PAGE) {
3373 tail = local_read(&bpage->commit);
3374 } else if (info->add_timestamp &
3375 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
3376 /* Ignore events with absolute time stamps */
3381 * Do not check the first event (skip possible extends too).
3382 * Also do not check if previous events have not been committed.
3384 if (tail <= 8 || tail > local_read(&bpage->commit))
3388 * If this interrupted another event,
3390 if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
3393 ts = bpage->time_stamp;
3395 for (e = 0; e < tail; e += rb_event_length(event)) {
3397 event = (struct ring_buffer_event *)(bpage->data + e);
3399 switch (event->type_len) {
3401 case RINGBUF_TYPE_TIME_EXTEND:
3402 delta = rb_event_time_stamp(event);
3406 case RINGBUF_TYPE_TIME_STAMP:
3407 delta = rb_event_time_stamp(event);
3408 delta = rb_fix_abs_ts(delta, ts);
3410 buffer_warn_return("[CPU: %d]ABSOLUTE TIME WENT BACKWARDS: last ts: %lld absolute ts: %lld\n",
3411 cpu_buffer->cpu, ts, delta);
3416 case RINGBUF_TYPE_PADDING:
3417 if (event->time_delta == 1)
3420 case RINGBUF_TYPE_DATA:
3421 ts += event->time_delta;
3425 RB_WARN_ON(cpu_buffer, 1);
3428 if ((full && ts > info->ts) ||
3429 (!full && ts + info->delta != info->ts)) {
3430 buffer_warn_return("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s context:%s\n",
3432 ts + info->delta, info->ts, info->delta,
3433 info->before, info->after,
3434 full ? " (full)" : "", show_interrupt_level());
3437 atomic_dec(this_cpu_ptr(&checking));
3440 static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3441 struct rb_event_info *info,
3445 #endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
3447 static struct ring_buffer_event *
3448 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3449 struct rb_event_info *info)
3451 struct ring_buffer_event *event;
3452 struct buffer_page *tail_page;
3453 unsigned long tail, write, w;
3455 /* Don't let the compiler play games with cpu_buffer->tail_page */
3456 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3458 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
3460 rb_time_read(&cpu_buffer->before_stamp, &info->before);
3461 rb_time_read(&cpu_buffer->write_stamp, &info->after);
3463 info->ts = rb_time_stamp(cpu_buffer->buffer);
3465 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3466 info->delta = info->ts;
3469 * If interrupting an event time update, we may need an
3470 * absolute timestamp.
3471 * Don't bother if this is the start of a new page (w == 0).
3474 /* Use the sub-buffer timestamp */
3476 } else if (unlikely(info->before != info->after)) {
3477 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3478 info->length += RB_LEN_TIME_EXTEND;
3480 info->delta = info->ts - info->after;
3481 if (unlikely(test_time_stamp(info->delta))) {
3482 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3483 info->length += RB_LEN_TIME_EXTEND;
3488 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
3490 /*C*/ write = local_add_return(info->length, &tail_page->write);
3492 /* set write to only the index of the write */
3493 write &= RB_WRITE_MASK;
3495 tail = write - info->length;
3497 /* See if we shot pass the end of this buffer page */
3498 if (unlikely(write > cpu_buffer->buffer->subbuf_size)) {
3499 check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
3500 return rb_move_tail(cpu_buffer, tail, info);
3503 if (likely(tail == w)) {
3504 /* Nothing interrupted us between A and C */
3505 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
3507 * If something came in between C and D, the write stamp
3508 * may now not be in sync. But that's fine as the before_stamp
3509 * will be different and then next event will just be forced
3510 * to use an absolute timestamp.
3512 if (likely(!(info->add_timestamp &
3513 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3514 /* This did not interrupt any time update */
3515 info->delta = info->ts - info->after;
3517 /* Just use full timestamp for interrupting event */
3518 info->delta = info->ts;
3519 check_buffer(cpu_buffer, info, tail);
3522 /* SLOW PATH - Interrupted between A and C */
3524 /* Save the old before_stamp */
3525 rb_time_read(&cpu_buffer->before_stamp, &info->before);
3528 * Read a new timestamp and update the before_stamp to make
3529 * the next event after this one force using an absolute
3530 * timestamp. This is in case an interrupt were to come in
3533 ts = rb_time_stamp(cpu_buffer->buffer);
3534 rb_time_set(&cpu_buffer->before_stamp, ts);
3537 /*E*/ rb_time_read(&cpu_buffer->write_stamp, &info->after);
3539 /*F*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3540 info->after == info->before && info->after < ts) {
3542 * Nothing came after this event between C and F, it is
3543 * safe to use info->after for the delta as it
3544 * matched info->before and is still valid.
3546 info->delta = ts - info->after;
3549 * Interrupted between C and F:
3550 * Lost the previous events time stamp. Just set the
3551 * delta to zero, and this will be the same time as
3552 * the event this event interrupted. And the events that
3553 * came after this will still be correct (as they would
3554 * have built their delta on the previous event.
3559 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3563 * If this is the first commit on the page, then it has the same
3564 * timestamp as the page itself.
3566 if (unlikely(!tail && !(info->add_timestamp &
3567 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3570 /* We reserved something on the buffer */
3572 event = __rb_page_index(tail_page, tail);
3573 rb_update_event(cpu_buffer, event, info);
3575 local_inc(&tail_page->entries);
3578 * If this is the first commit on the page, then update
3581 if (unlikely(!tail))
3582 tail_page->page->time_stamp = info->ts;
3584 /* account for these added bytes */
3585 local_add(info->length, &cpu_buffer->entries_bytes);
3590 static __always_inline struct ring_buffer_event *
3591 rb_reserve_next_event(struct trace_buffer *buffer,
3592 struct ring_buffer_per_cpu *cpu_buffer,
3593 unsigned long length)
3595 struct ring_buffer_event *event;
3596 struct rb_event_info info;
3600 /* ring buffer does cmpxchg, make sure it is safe in NMI context */
3601 if (!IS_ENABLED(CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG) &&
3602 (unlikely(in_nmi()))) {
3606 rb_start_commit(cpu_buffer);
3607 /* The commit page can not change after this */
3609 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3611 * Due to the ability to swap a cpu buffer from a buffer
3612 * it is possible it was swapped before we committed.
3613 * (committing stops a swap). We check for it here and
3614 * if it happened, we have to fail the write.
3617 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3618 local_dec(&cpu_buffer->committing);
3619 local_dec(&cpu_buffer->commits);
3624 info.length = rb_calculate_event_length(length);
3626 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3627 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3628 info.length += RB_LEN_TIME_EXTEND;
3629 if (info.length > cpu_buffer->buffer->max_data_size)
3632 add_ts_default = RB_ADD_STAMP_NONE;
3636 info.add_timestamp = add_ts_default;
3640 * We allow for interrupts to reenter here and do a trace.
3641 * If one does, it will cause this original code to loop
3642 * back here. Even with heavy interrupts happening, this
3643 * should only happen a few times in a row. If this happens
3644 * 1000 times in a row, there must be either an interrupt
3645 * storm or we have something buggy.
3648 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3651 event = __rb_reserve_next(cpu_buffer, &info);
3653 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3654 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3655 info.length -= RB_LEN_TIME_EXTEND;
3662 rb_end_commit(cpu_buffer);
3667 * ring_buffer_lock_reserve - reserve a part of the buffer
3668 * @buffer: the ring buffer to reserve from
3669 * @length: the length of the data to reserve (excluding event header)
3671 * Returns a reserved event on the ring buffer to copy directly to.
3672 * The user of this interface will need to get the body to write into
3673 * and can use the ring_buffer_event_data() interface.
3675 * The length is the length of the data needed, not the event length
3676 * which also includes the event header.
3678 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3679 * If NULL is returned, then nothing has been allocated or locked.
3681 struct ring_buffer_event *
3682 ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3684 struct ring_buffer_per_cpu *cpu_buffer;
3685 struct ring_buffer_event *event;
3688 /* If we are tracing schedule, we don't want to recurse */
3689 preempt_disable_notrace();
3691 if (unlikely(atomic_read(&buffer->record_disabled)))
3694 cpu = raw_smp_processor_id();
3696 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3699 cpu_buffer = buffer->buffers[cpu];
3701 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3704 if (unlikely(length > buffer->max_data_size))
3707 if (unlikely(trace_recursive_lock(cpu_buffer)))
3710 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3717 trace_recursive_unlock(cpu_buffer);
3719 preempt_enable_notrace();
3722 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3725 * Decrement the entries to the page that an event is on.
3726 * The event does not even need to exist, only the pointer
3727 * to the page it is on. This may only be called before the commit
3731 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3732 struct ring_buffer_event *event)
3734 unsigned long addr = (unsigned long)event;
3735 struct buffer_page *bpage = cpu_buffer->commit_page;
3736 struct buffer_page *start;
3738 addr &= ~((PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1);
3740 /* Do the likely case first */
3741 if (likely(bpage->page == (void *)addr)) {
3742 local_dec(&bpage->entries);
3747 * Because the commit page may be on the reader page we
3748 * start with the next page and check the end loop there.
3750 rb_inc_page(&bpage);
3753 if (bpage->page == (void *)addr) {
3754 local_dec(&bpage->entries);
3757 rb_inc_page(&bpage);
3758 } while (bpage != start);
3760 /* commit not part of this buffer?? */
3761 RB_WARN_ON(cpu_buffer, 1);
3765 * ring_buffer_discard_commit - discard an event that has not been committed
3766 * @buffer: the ring buffer
3767 * @event: non committed event to discard
3769 * Sometimes an event that is in the ring buffer needs to be ignored.
3770 * This function lets the user discard an event in the ring buffer
3771 * and then that event will not be read later.
3773 * This function only works if it is called before the item has been
3774 * committed. It will try to free the event from the ring buffer
3775 * if another event has not been added behind it.
3777 * If another event has been added behind it, it will set the event
3778 * up as discarded, and perform the commit.
3780 * If this function is called, do not call ring_buffer_unlock_commit on
3783 void ring_buffer_discard_commit(struct trace_buffer *buffer,
3784 struct ring_buffer_event *event)
3786 struct ring_buffer_per_cpu *cpu_buffer;
3789 /* The event is discarded regardless */
3790 rb_event_discard(event);
3792 cpu = smp_processor_id();
3793 cpu_buffer = buffer->buffers[cpu];
3796 * This must only be called if the event has not been
3797 * committed yet. Thus we can assume that preemption
3798 * is still disabled.
3800 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3802 rb_decrement_entry(cpu_buffer, event);
3803 if (rb_try_to_discard(cpu_buffer, event))
3807 rb_end_commit(cpu_buffer);
3809 trace_recursive_unlock(cpu_buffer);
3811 preempt_enable_notrace();
3814 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3817 * ring_buffer_write - write data to the buffer without reserving
3818 * @buffer: The ring buffer to write to.
3819 * @length: The length of the data being written (excluding the event header)
3820 * @data: The data to write to the buffer.
3822 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3823 * one function. If you already have the data to write to the buffer, it
3824 * may be easier to simply call this function.
3826 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3827 * and not the length of the event which would hold the header.
3829 int ring_buffer_write(struct trace_buffer *buffer,
3830 unsigned long length,
3833 struct ring_buffer_per_cpu *cpu_buffer;
3834 struct ring_buffer_event *event;
3839 preempt_disable_notrace();
3841 if (atomic_read(&buffer->record_disabled))
3844 cpu = raw_smp_processor_id();
3846 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3849 cpu_buffer = buffer->buffers[cpu];
3851 if (atomic_read(&cpu_buffer->record_disabled))
3854 if (length > buffer->max_data_size)
3857 if (unlikely(trace_recursive_lock(cpu_buffer)))
3860 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3864 body = rb_event_data(event);
3866 memcpy(body, data, length);
3868 rb_commit(cpu_buffer);
3870 rb_wakeups(buffer, cpu_buffer);
3875 trace_recursive_unlock(cpu_buffer);
3878 preempt_enable_notrace();
3882 EXPORT_SYMBOL_GPL(ring_buffer_write);
3884 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3886 struct buffer_page *reader = cpu_buffer->reader_page;
3887 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3888 struct buffer_page *commit = cpu_buffer->commit_page;
3890 /* In case of error, head will be NULL */
3891 if (unlikely(!head))
3894 /* Reader should exhaust content in reader page */
3895 if (reader->read != rb_page_commit(reader))
3899 * If writers are committing on the reader page, knowing all
3900 * committed content has been read, the ring buffer is empty.
3902 if (commit == reader)
3906 * If writers are committing on a page other than reader page
3907 * and head page, there should always be content to read.
3913 * Writers are committing on the head page, we just need
3914 * to care about there're committed data, and the reader will
3915 * swap reader page with head page when it is to read data.
3917 return rb_page_commit(commit) == 0;
3921 * ring_buffer_record_disable - stop all writes into the buffer
3922 * @buffer: The ring buffer to stop writes to.
3924 * This prevents all writes to the buffer. Any attempt to write
3925 * to the buffer after this will fail and return NULL.
3927 * The caller should call synchronize_rcu() after this.
3929 void ring_buffer_record_disable(struct trace_buffer *buffer)
3931 atomic_inc(&buffer->record_disabled);
3933 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3936 * ring_buffer_record_enable - enable writes to the buffer
3937 * @buffer: The ring buffer to enable writes
3939 * Note, multiple disables will need the same number of enables
3940 * to truly enable the writing (much like preempt_disable).
3942 void ring_buffer_record_enable(struct trace_buffer *buffer)
3944 atomic_dec(&buffer->record_disabled);
3946 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3949 * ring_buffer_record_off - stop all writes into the buffer
3950 * @buffer: The ring buffer to stop writes to.
3952 * This prevents all writes to the buffer. Any attempt to write
3953 * to the buffer after this will fail and return NULL.
3955 * This is different than ring_buffer_record_disable() as
3956 * it works like an on/off switch, where as the disable() version
3957 * must be paired with a enable().
3959 void ring_buffer_record_off(struct trace_buffer *buffer)
3962 unsigned int new_rd;
3964 rd = atomic_read(&buffer->record_disabled);
3966 new_rd = rd | RB_BUFFER_OFF;
3967 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
3969 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3972 * ring_buffer_record_on - restart writes into the buffer
3973 * @buffer: The ring buffer to start writes to.
3975 * This enables all writes to the buffer that was disabled by
3976 * ring_buffer_record_off().
3978 * This is different than ring_buffer_record_enable() as
3979 * it works like an on/off switch, where as the enable() version
3980 * must be paired with a disable().
3982 void ring_buffer_record_on(struct trace_buffer *buffer)
3985 unsigned int new_rd;
3987 rd = atomic_read(&buffer->record_disabled);
3989 new_rd = rd & ~RB_BUFFER_OFF;
3990 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
3992 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3995 * ring_buffer_record_is_on - return true if the ring buffer can write
3996 * @buffer: The ring buffer to see if write is enabled
3998 * Returns true if the ring buffer is in a state that it accepts writes.
4000 bool ring_buffer_record_is_on(struct trace_buffer *buffer)
4002 return !atomic_read(&buffer->record_disabled);
4006 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
4007 * @buffer: The ring buffer to see if write is set enabled
4009 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
4010 * Note that this does NOT mean it is in a writable state.
4012 * It may return true when the ring buffer has been disabled by
4013 * ring_buffer_record_disable(), as that is a temporary disabling of
4016 bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4018 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4022 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4023 * @buffer: The ring buffer to stop writes to.
4024 * @cpu: The CPU buffer to stop
4026 * This prevents all writes to the buffer. Any attempt to write
4027 * to the buffer after this will fail and return NULL.
4029 * The caller should call synchronize_rcu() after this.
4031 void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4033 struct ring_buffer_per_cpu *cpu_buffer;
4035 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4038 cpu_buffer = buffer->buffers[cpu];
4039 atomic_inc(&cpu_buffer->record_disabled);
4041 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4044 * ring_buffer_record_enable_cpu - enable writes to the buffer
4045 * @buffer: The ring buffer to enable writes
4046 * @cpu: The CPU to enable.
4048 * Note, multiple disables will need the same number of enables
4049 * to truly enable the writing (much like preempt_disable).
4051 void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4053 struct ring_buffer_per_cpu *cpu_buffer;
4055 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4058 cpu_buffer = buffer->buffers[cpu];
4059 atomic_dec(&cpu_buffer->record_disabled);
4061 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4064 * The total entries in the ring buffer is the running counter
4065 * of entries entered into the ring buffer, minus the sum of
4066 * the entries read from the ring buffer and the number of
4067 * entries that were overwritten.
4069 static inline unsigned long
4070 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4072 return local_read(&cpu_buffer->entries) -
4073 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4077 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4078 * @buffer: The ring buffer
4079 * @cpu: The per CPU buffer to read from.
4081 u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4083 unsigned long flags;
4084 struct ring_buffer_per_cpu *cpu_buffer;
4085 struct buffer_page *bpage;
4088 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4091 cpu_buffer = buffer->buffers[cpu];
4092 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4094 * if the tail is on reader_page, oldest time stamp is on the reader
4097 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4098 bpage = cpu_buffer->reader_page;
4100 bpage = rb_set_head_page(cpu_buffer);
4102 ret = bpage->page->time_stamp;
4103 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4107 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4110 * ring_buffer_bytes_cpu - get the number of bytes unconsumed in a cpu buffer
4111 * @buffer: The ring buffer
4112 * @cpu: The per CPU buffer to read from.
4114 unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4116 struct ring_buffer_per_cpu *cpu_buffer;
4119 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4122 cpu_buffer = buffer->buffers[cpu];
4123 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4127 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4130 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4131 * @buffer: The ring buffer
4132 * @cpu: The per CPU buffer to get the entries from.
4134 unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4136 struct ring_buffer_per_cpu *cpu_buffer;
4138 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4141 cpu_buffer = buffer->buffers[cpu];
4143 return rb_num_of_entries(cpu_buffer);
4145 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4148 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4149 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4150 * @buffer: The ring buffer
4151 * @cpu: The per CPU buffer to get the number of overruns from
4153 unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4155 struct ring_buffer_per_cpu *cpu_buffer;
4158 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4161 cpu_buffer = buffer->buffers[cpu];
4162 ret = local_read(&cpu_buffer->overrun);
4166 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4169 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4170 * commits failing due to the buffer wrapping around while there are uncommitted
4171 * events, such as during an interrupt storm.
4172 * @buffer: The ring buffer
4173 * @cpu: The per CPU buffer to get the number of overruns from
4176 ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4178 struct ring_buffer_per_cpu *cpu_buffer;
4181 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4184 cpu_buffer = buffer->buffers[cpu];
4185 ret = local_read(&cpu_buffer->commit_overrun);
4189 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4192 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4193 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4194 * @buffer: The ring buffer
4195 * @cpu: The per CPU buffer to get the number of overruns from
4198 ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4200 struct ring_buffer_per_cpu *cpu_buffer;
4203 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4206 cpu_buffer = buffer->buffers[cpu];
4207 ret = local_read(&cpu_buffer->dropped_events);
4211 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4214 * ring_buffer_read_events_cpu - get the number of events successfully read
4215 * @buffer: The ring buffer
4216 * @cpu: The per CPU buffer to get the number of events read
4219 ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4221 struct ring_buffer_per_cpu *cpu_buffer;
4223 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4226 cpu_buffer = buffer->buffers[cpu];
4227 return cpu_buffer->read;
4229 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4232 * ring_buffer_entries - get the number of entries in a buffer
4233 * @buffer: The ring buffer
4235 * Returns the total number of entries in the ring buffer
4238 unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4240 struct ring_buffer_per_cpu *cpu_buffer;
4241 unsigned long entries = 0;
4244 /* if you care about this being correct, lock the buffer */
4245 for_each_buffer_cpu(buffer, cpu) {
4246 cpu_buffer = buffer->buffers[cpu];
4247 entries += rb_num_of_entries(cpu_buffer);
4252 EXPORT_SYMBOL_GPL(ring_buffer_entries);
4255 * ring_buffer_overruns - get the number of overruns in buffer
4256 * @buffer: The ring buffer
4258 * Returns the total number of overruns in the ring buffer
4261 unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4263 struct ring_buffer_per_cpu *cpu_buffer;
4264 unsigned long overruns = 0;
4267 /* if you care about this being correct, lock the buffer */
4268 for_each_buffer_cpu(buffer, cpu) {
4269 cpu_buffer = buffer->buffers[cpu];
4270 overruns += local_read(&cpu_buffer->overrun);
4275 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4277 static void rb_iter_reset(struct ring_buffer_iter *iter)
4279 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4281 /* Iterator usage is expected to have record disabled */
4282 iter->head_page = cpu_buffer->reader_page;
4283 iter->head = cpu_buffer->reader_page->read;
4284 iter->next_event = iter->head;
4286 iter->cache_reader_page = iter->head_page;
4287 iter->cache_read = cpu_buffer->read;
4288 iter->cache_pages_removed = cpu_buffer->pages_removed;
4291 iter->read_stamp = cpu_buffer->read_stamp;
4292 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4294 iter->read_stamp = iter->head_page->page->time_stamp;
4295 iter->page_stamp = iter->read_stamp;
4300 * ring_buffer_iter_reset - reset an iterator
4301 * @iter: The iterator to reset
4303 * Resets the iterator, so that it will start from the beginning
4306 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4308 struct ring_buffer_per_cpu *cpu_buffer;
4309 unsigned long flags;
4314 cpu_buffer = iter->cpu_buffer;
4316 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4317 rb_iter_reset(iter);
4318 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4320 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4323 * ring_buffer_iter_empty - check if an iterator has no more to read
4324 * @iter: The iterator to check
4326 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4328 struct ring_buffer_per_cpu *cpu_buffer;
4329 struct buffer_page *reader;
4330 struct buffer_page *head_page;
4331 struct buffer_page *commit_page;
4332 struct buffer_page *curr_commit_page;
4337 cpu_buffer = iter->cpu_buffer;
4338 reader = cpu_buffer->reader_page;
4339 head_page = cpu_buffer->head_page;
4340 commit_page = cpu_buffer->commit_page;
4341 commit_ts = commit_page->page->time_stamp;
4344 * When the writer goes across pages, it issues a cmpxchg which
4345 * is a mb(), which will synchronize with the rmb here.
4346 * (see rb_tail_page_update())
4349 commit = rb_page_commit(commit_page);
4350 /* We want to make sure that the commit page doesn't change */
4353 /* Make sure commit page didn't change */
4354 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4355 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4357 /* If the commit page changed, then there's more data */
4358 if (curr_commit_page != commit_page ||
4359 curr_commit_ts != commit_ts)
4362 /* Still racy, as it may return a false positive, but that's OK */
4363 return ((iter->head_page == commit_page && iter->head >= commit) ||
4364 (iter->head_page == reader && commit_page == head_page &&
4365 head_page->read == commit &&
4366 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4368 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4371 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4372 struct ring_buffer_event *event)
4376 switch (event->type_len) {
4377 case RINGBUF_TYPE_PADDING:
4380 case RINGBUF_TYPE_TIME_EXTEND:
4381 delta = rb_event_time_stamp(event);
4382 cpu_buffer->read_stamp += delta;
4385 case RINGBUF_TYPE_TIME_STAMP:
4386 delta = rb_event_time_stamp(event);
4387 delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
4388 cpu_buffer->read_stamp = delta;
4391 case RINGBUF_TYPE_DATA:
4392 cpu_buffer->read_stamp += event->time_delta;
4396 RB_WARN_ON(cpu_buffer, 1);
4401 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4402 struct ring_buffer_event *event)
4406 switch (event->type_len) {
4407 case RINGBUF_TYPE_PADDING:
4410 case RINGBUF_TYPE_TIME_EXTEND:
4411 delta = rb_event_time_stamp(event);
4412 iter->read_stamp += delta;
4415 case RINGBUF_TYPE_TIME_STAMP:
4416 delta = rb_event_time_stamp(event);
4417 delta = rb_fix_abs_ts(delta, iter->read_stamp);
4418 iter->read_stamp = delta;
4421 case RINGBUF_TYPE_DATA:
4422 iter->read_stamp += event->time_delta;
4426 RB_WARN_ON(iter->cpu_buffer, 1);
4430 static struct buffer_page *
4431 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4433 struct buffer_page *reader = NULL;
4434 unsigned long bsize = READ_ONCE(cpu_buffer->buffer->subbuf_size);
4435 unsigned long overwrite;
4436 unsigned long flags;
4440 local_irq_save(flags);
4441 arch_spin_lock(&cpu_buffer->lock);
4445 * This should normally only loop twice. But because the
4446 * start of the reader inserts an empty page, it causes
4447 * a case where we will loop three times. There should be no
4448 * reason to loop four times (that I know of).
4450 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4455 reader = cpu_buffer->reader_page;
4457 /* If there's more to read, return this page */
4458 if (cpu_buffer->reader_page->read < rb_page_size(reader))
4461 /* Never should we have an index greater than the size */
4462 if (RB_WARN_ON(cpu_buffer,
4463 cpu_buffer->reader_page->read > rb_page_size(reader)))
4466 /* check if we caught up to the tail */
4468 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4471 /* Don't bother swapping if the ring buffer is empty */
4472 if (rb_num_of_entries(cpu_buffer) == 0)
4476 * Reset the reader page to size zero.
4478 local_set(&cpu_buffer->reader_page->write, 0);
4479 local_set(&cpu_buffer->reader_page->entries, 0);
4480 local_set(&cpu_buffer->reader_page->page->commit, 0);
4481 cpu_buffer->reader_page->real_end = 0;
4485 * Splice the empty reader page into the list around the head.
4487 reader = rb_set_head_page(cpu_buffer);
4490 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4491 cpu_buffer->reader_page->list.prev = reader->list.prev;
4494 * cpu_buffer->pages just needs to point to the buffer, it
4495 * has no specific buffer page to point to. Lets move it out
4496 * of our way so we don't accidentally swap it.
4498 cpu_buffer->pages = reader->list.prev;
4500 /* The reader page will be pointing to the new head */
4501 rb_set_list_to_head(&cpu_buffer->reader_page->list);
4504 * We want to make sure we read the overruns after we set up our
4505 * pointers to the next object. The writer side does a
4506 * cmpxchg to cross pages which acts as the mb on the writer
4507 * side. Note, the reader will constantly fail the swap
4508 * while the writer is updating the pointers, so this
4509 * guarantees that the overwrite recorded here is the one we
4510 * want to compare with the last_overrun.
4513 overwrite = local_read(&(cpu_buffer->overrun));
4516 * Here's the tricky part.
4518 * We need to move the pointer past the header page.
4519 * But we can only do that if a writer is not currently
4520 * moving it. The page before the header page has the
4521 * flag bit '1' set if it is pointing to the page we want.
4522 * but if the writer is in the process of moving it
4523 * than it will be '2' or already moved '0'.
4526 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4529 * If we did not convert it, then we must try again.
4535 * Yay! We succeeded in replacing the page.
4537 * Now make the new head point back to the reader page.
4539 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4540 rb_inc_page(&cpu_buffer->head_page);
4542 local_inc(&cpu_buffer->pages_read);
4544 /* Finally update the reader page to the new head */
4545 cpu_buffer->reader_page = reader;
4546 cpu_buffer->reader_page->read = 0;
4548 if (overwrite != cpu_buffer->last_overrun) {
4549 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4550 cpu_buffer->last_overrun = overwrite;
4556 /* Update the read_stamp on the first event */
4557 if (reader && reader->read == 0)
4558 cpu_buffer->read_stamp = reader->page->time_stamp;
4560 arch_spin_unlock(&cpu_buffer->lock);
4561 local_irq_restore(flags);
4564 * The writer has preempt disable, wait for it. But not forever
4565 * Although, 1 second is pretty much "forever"
4567 #define USECS_WAIT 1000000
4568 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
4569 /* If the write is past the end of page, a writer is still updating it */
4570 if (likely(!reader || rb_page_write(reader) <= bsize))
4575 /* Get the latest version of the reader write value */
4579 /* The writer is not moving forward? Something is wrong */
4580 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
4584 * Make sure we see any padding after the write update
4585 * (see rb_reset_tail()).
4587 * In addition, a writer may be writing on the reader page
4588 * if the page has not been fully filled, so the read barrier
4589 * is also needed to make sure we see the content of what is
4590 * committed by the writer (see rb_set_commit_to_write()).
4598 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4600 struct ring_buffer_event *event;
4601 struct buffer_page *reader;
4604 reader = rb_get_reader_page(cpu_buffer);
4606 /* This function should not be called when buffer is empty */
4607 if (RB_WARN_ON(cpu_buffer, !reader))
4610 event = rb_reader_event(cpu_buffer);
4612 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4615 rb_update_read_stamp(cpu_buffer, event);
4617 length = rb_event_length(event);
4618 cpu_buffer->reader_page->read += length;
4619 cpu_buffer->read_bytes += length;
4622 static void rb_advance_iter(struct ring_buffer_iter *iter)
4624 struct ring_buffer_per_cpu *cpu_buffer;
4626 cpu_buffer = iter->cpu_buffer;
4628 /* If head == next_event then we need to jump to the next event */
4629 if (iter->head == iter->next_event) {
4630 /* If the event gets overwritten again, there's nothing to do */
4631 if (rb_iter_head_event(iter) == NULL)
4635 iter->head = iter->next_event;
4638 * Check if we are at the end of the buffer.
4640 if (iter->next_event >= rb_page_size(iter->head_page)) {
4641 /* discarded commits can make the page empty */
4642 if (iter->head_page == cpu_buffer->commit_page)
4648 rb_update_iter_read_stamp(iter, iter->event);
4651 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4653 return cpu_buffer->lost_events;
4656 static struct ring_buffer_event *
4657 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4658 unsigned long *lost_events)
4660 struct ring_buffer_event *event;
4661 struct buffer_page *reader;
4668 * We repeat when a time extend is encountered.
4669 * Since the time extend is always attached to a data event,
4670 * we should never loop more than once.
4671 * (We never hit the following condition more than twice).
4673 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4676 reader = rb_get_reader_page(cpu_buffer);
4680 event = rb_reader_event(cpu_buffer);
4682 switch (event->type_len) {
4683 case RINGBUF_TYPE_PADDING:
4684 if (rb_null_event(event))
4685 RB_WARN_ON(cpu_buffer, 1);
4687 * Because the writer could be discarding every
4688 * event it creates (which would probably be bad)
4689 * if we were to go back to "again" then we may never
4690 * catch up, and will trigger the warn on, or lock
4691 * the box. Return the padding, and we will release
4692 * the current locks, and try again.
4696 case RINGBUF_TYPE_TIME_EXTEND:
4697 /* Internal data, OK to advance */
4698 rb_advance_reader(cpu_buffer);
4701 case RINGBUF_TYPE_TIME_STAMP:
4703 *ts = rb_event_time_stamp(event);
4704 *ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
4705 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4706 cpu_buffer->cpu, ts);
4708 /* Internal data, OK to advance */
4709 rb_advance_reader(cpu_buffer);
4712 case RINGBUF_TYPE_DATA:
4714 *ts = cpu_buffer->read_stamp + event->time_delta;
4715 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4716 cpu_buffer->cpu, ts);
4719 *lost_events = rb_lost_events(cpu_buffer);
4723 RB_WARN_ON(cpu_buffer, 1);
4728 EXPORT_SYMBOL_GPL(ring_buffer_peek);
4730 static struct ring_buffer_event *
4731 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4733 struct trace_buffer *buffer;
4734 struct ring_buffer_per_cpu *cpu_buffer;
4735 struct ring_buffer_event *event;
4741 cpu_buffer = iter->cpu_buffer;
4742 buffer = cpu_buffer->buffer;
4745 * Check if someone performed a consuming read to the buffer
4746 * or removed some pages from the buffer. In these cases,
4747 * iterator was invalidated and we need to reset it.
4749 if (unlikely(iter->cache_read != cpu_buffer->read ||
4750 iter->cache_reader_page != cpu_buffer->reader_page ||
4751 iter->cache_pages_removed != cpu_buffer->pages_removed))
4752 rb_iter_reset(iter);
4755 if (ring_buffer_iter_empty(iter))
4759 * As the writer can mess with what the iterator is trying
4760 * to read, just give up if we fail to get an event after
4761 * three tries. The iterator is not as reliable when reading
4762 * the ring buffer with an active write as the consumer is.
4763 * Do not warn if the three failures is reached.
4768 if (rb_per_cpu_empty(cpu_buffer))
4771 if (iter->head >= rb_page_size(iter->head_page)) {
4776 event = rb_iter_head_event(iter);
4780 switch (event->type_len) {
4781 case RINGBUF_TYPE_PADDING:
4782 if (rb_null_event(event)) {
4786 rb_advance_iter(iter);
4789 case RINGBUF_TYPE_TIME_EXTEND:
4790 /* Internal data, OK to advance */
4791 rb_advance_iter(iter);
4794 case RINGBUF_TYPE_TIME_STAMP:
4796 *ts = rb_event_time_stamp(event);
4797 *ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
4798 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4799 cpu_buffer->cpu, ts);
4801 /* Internal data, OK to advance */
4802 rb_advance_iter(iter);
4805 case RINGBUF_TYPE_DATA:
4807 *ts = iter->read_stamp + event->time_delta;
4808 ring_buffer_normalize_time_stamp(buffer,
4809 cpu_buffer->cpu, ts);
4814 RB_WARN_ON(cpu_buffer, 1);
4819 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4821 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4823 if (likely(!in_nmi())) {
4824 raw_spin_lock(&cpu_buffer->reader_lock);
4829 * If an NMI die dumps out the content of the ring buffer
4830 * trylock must be used to prevent a deadlock if the NMI
4831 * preempted a task that holds the ring buffer locks. If
4832 * we get the lock then all is fine, if not, then continue
4833 * to do the read, but this can corrupt the ring buffer,
4834 * so it must be permanently disabled from future writes.
4835 * Reading from NMI is a oneshot deal.
4837 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4840 /* Continue without locking, but disable the ring buffer */
4841 atomic_inc(&cpu_buffer->record_disabled);
4846 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4849 raw_spin_unlock(&cpu_buffer->reader_lock);
4853 * ring_buffer_peek - peek at the next event to be read
4854 * @buffer: The ring buffer to read
4855 * @cpu: The cpu to peak at
4856 * @ts: The timestamp counter of this event.
4857 * @lost_events: a variable to store if events were lost (may be NULL)
4859 * This will return the event that will be read next, but does
4860 * not consume the data.
4862 struct ring_buffer_event *
4863 ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4864 unsigned long *lost_events)
4866 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4867 struct ring_buffer_event *event;
4868 unsigned long flags;
4871 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4875 local_irq_save(flags);
4876 dolock = rb_reader_lock(cpu_buffer);
4877 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4878 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4879 rb_advance_reader(cpu_buffer);
4880 rb_reader_unlock(cpu_buffer, dolock);
4881 local_irq_restore(flags);
4883 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4889 /** ring_buffer_iter_dropped - report if there are dropped events
4890 * @iter: The ring buffer iterator
4892 * Returns true if there was dropped events since the last peek.
4894 bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4896 bool ret = iter->missed_events != 0;
4898 iter->missed_events = 0;
4901 EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
4904 * ring_buffer_iter_peek - peek at the next event to be read
4905 * @iter: The ring buffer iterator
4906 * @ts: The timestamp counter of this event.
4908 * This will return the event that will be read next, but does
4909 * not increment the iterator.
4911 struct ring_buffer_event *
4912 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4914 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4915 struct ring_buffer_event *event;
4916 unsigned long flags;
4919 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4920 event = rb_iter_peek(iter, ts);
4921 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4923 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4930 * ring_buffer_consume - return an event and consume it
4931 * @buffer: The ring buffer to get the next event from
4932 * @cpu: the cpu to read the buffer from
4933 * @ts: a variable to store the timestamp (may be NULL)
4934 * @lost_events: a variable to store if events were lost (may be NULL)
4936 * Returns the next event in the ring buffer, and that event is consumed.
4937 * Meaning, that sequential reads will keep returning a different event,
4938 * and eventually empty the ring buffer if the producer is slower.
4940 struct ring_buffer_event *
4941 ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
4942 unsigned long *lost_events)
4944 struct ring_buffer_per_cpu *cpu_buffer;
4945 struct ring_buffer_event *event = NULL;
4946 unsigned long flags;
4950 /* might be called in atomic */
4953 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4956 cpu_buffer = buffer->buffers[cpu];
4957 local_irq_save(flags);
4958 dolock = rb_reader_lock(cpu_buffer);
4960 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4962 cpu_buffer->lost_events = 0;
4963 rb_advance_reader(cpu_buffer);
4966 rb_reader_unlock(cpu_buffer, dolock);
4967 local_irq_restore(flags);
4972 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4977 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4980 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4981 * @buffer: The ring buffer to read from
4982 * @cpu: The cpu buffer to iterate over
4983 * @flags: gfp flags to use for memory allocation
4985 * This performs the initial preparations necessary to iterate
4986 * through the buffer. Memory is allocated, buffer recording
4987 * is disabled, and the iterator pointer is returned to the caller.
4989 * Disabling buffer recording prevents the reading from being
4990 * corrupted. This is not a consuming read, so a producer is not
4993 * After a sequence of ring_buffer_read_prepare calls, the user is
4994 * expected to make at least one call to ring_buffer_read_prepare_sync.
4995 * Afterwards, ring_buffer_read_start is invoked to get things going
4998 * This overall must be paired with ring_buffer_read_finish.
5000 struct ring_buffer_iter *
5001 ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
5003 struct ring_buffer_per_cpu *cpu_buffer;
5004 struct ring_buffer_iter *iter;
5006 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5009 iter = kzalloc(sizeof(*iter), flags);
5013 /* Holds the entire event: data and meta data */
5014 iter->event_size = buffer->subbuf_size;
5015 iter->event = kmalloc(iter->event_size, flags);
5021 cpu_buffer = buffer->buffers[cpu];
5023 iter->cpu_buffer = cpu_buffer;
5025 atomic_inc(&cpu_buffer->resize_disabled);
5029 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
5032 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
5034 * All previously invoked ring_buffer_read_prepare calls to prepare
5035 * iterators will be synchronized. Afterwards, read_buffer_read_start
5036 * calls on those iterators are allowed.
5039 ring_buffer_read_prepare_sync(void)
5043 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
5046 * ring_buffer_read_start - start a non consuming read of the buffer
5047 * @iter: The iterator returned by ring_buffer_read_prepare
5049 * This finalizes the startup of an iteration through the buffer.
5050 * The iterator comes from a call to ring_buffer_read_prepare and
5051 * an intervening ring_buffer_read_prepare_sync must have been
5054 * Must be paired with ring_buffer_read_finish.
5057 ring_buffer_read_start(struct ring_buffer_iter *iter)
5059 struct ring_buffer_per_cpu *cpu_buffer;
5060 unsigned long flags;
5065 cpu_buffer = iter->cpu_buffer;
5067 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5068 arch_spin_lock(&cpu_buffer->lock);
5069 rb_iter_reset(iter);
5070 arch_spin_unlock(&cpu_buffer->lock);
5071 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5073 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5076 * ring_buffer_read_finish - finish reading the iterator of the buffer
5077 * @iter: The iterator retrieved by ring_buffer_start
5079 * This re-enables the recording to the buffer, and frees the
5083 ring_buffer_read_finish(struct ring_buffer_iter *iter)
5085 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5086 unsigned long flags;
5089 * Ring buffer is disabled from recording, here's a good place
5090 * to check the integrity of the ring buffer.
5091 * Must prevent readers from trying to read, as the check
5092 * clears the HEAD page and readers require it.
5094 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5095 rb_check_pages(cpu_buffer);
5096 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5098 atomic_dec(&cpu_buffer->resize_disabled);
5102 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5105 * ring_buffer_iter_advance - advance the iterator to the next location
5106 * @iter: The ring buffer iterator
5108 * Move the location of the iterator such that the next read will
5109 * be the next location of the iterator.
5111 void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5113 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5114 unsigned long flags;
5116 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5118 rb_advance_iter(iter);
5120 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5122 EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5125 * ring_buffer_size - return the size of the ring buffer (in bytes)
5126 * @buffer: The ring buffer.
5127 * @cpu: The CPU to get ring buffer size from.
5129 unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5131 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5134 return buffer->subbuf_size * buffer->buffers[cpu]->nr_pages;
5136 EXPORT_SYMBOL_GPL(ring_buffer_size);
5139 * ring_buffer_max_event_size - return the max data size of an event
5140 * @buffer: The ring buffer.
5142 * Returns the maximum size an event can be.
5144 unsigned long ring_buffer_max_event_size(struct trace_buffer *buffer)
5146 /* If abs timestamp is requested, events have a timestamp too */
5147 if (ring_buffer_time_stamp_abs(buffer))
5148 return buffer->max_data_size - RB_LEN_TIME_EXTEND;
5149 return buffer->max_data_size;
5151 EXPORT_SYMBOL_GPL(ring_buffer_max_event_size);
5153 static void rb_clear_buffer_page(struct buffer_page *page)
5155 local_set(&page->write, 0);
5156 local_set(&page->entries, 0);
5157 rb_init_page(page->page);
5162 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5164 struct buffer_page *page;
5166 rb_head_page_deactivate(cpu_buffer);
5168 cpu_buffer->head_page
5169 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5170 rb_clear_buffer_page(cpu_buffer->head_page);
5171 list_for_each_entry(page, cpu_buffer->pages, list) {
5172 rb_clear_buffer_page(page);
5175 cpu_buffer->tail_page = cpu_buffer->head_page;
5176 cpu_buffer->commit_page = cpu_buffer->head_page;
5178 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5179 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5180 rb_clear_buffer_page(cpu_buffer->reader_page);
5182 local_set(&cpu_buffer->entries_bytes, 0);
5183 local_set(&cpu_buffer->overrun, 0);
5184 local_set(&cpu_buffer->commit_overrun, 0);
5185 local_set(&cpu_buffer->dropped_events, 0);
5186 local_set(&cpu_buffer->entries, 0);
5187 local_set(&cpu_buffer->committing, 0);
5188 local_set(&cpu_buffer->commits, 0);
5189 local_set(&cpu_buffer->pages_touched, 0);
5190 local_set(&cpu_buffer->pages_lost, 0);
5191 local_set(&cpu_buffer->pages_read, 0);
5192 cpu_buffer->last_pages_touch = 0;
5193 cpu_buffer->shortest_full = 0;
5194 cpu_buffer->read = 0;
5195 cpu_buffer->read_bytes = 0;
5197 rb_time_set(&cpu_buffer->write_stamp, 0);
5198 rb_time_set(&cpu_buffer->before_stamp, 0);
5200 memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5202 cpu_buffer->lost_events = 0;
5203 cpu_buffer->last_overrun = 0;
5205 rb_head_page_activate(cpu_buffer);
5206 cpu_buffer->pages_removed = 0;
5209 /* Must have disabled the cpu buffer then done a synchronize_rcu */
5210 static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5212 unsigned long flags;
5214 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5216 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5219 arch_spin_lock(&cpu_buffer->lock);
5221 rb_reset_cpu(cpu_buffer);
5223 arch_spin_unlock(&cpu_buffer->lock);
5226 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5230 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5231 * @buffer: The ring buffer to reset a per cpu buffer of
5232 * @cpu: The CPU buffer to be reset
5234 void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5236 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5238 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5241 /* prevent another thread from changing buffer sizes */
5242 mutex_lock(&buffer->mutex);
5244 atomic_inc(&cpu_buffer->resize_disabled);
5245 atomic_inc(&cpu_buffer->record_disabled);
5247 /* Make sure all commits have finished */
5250 reset_disabled_cpu_buffer(cpu_buffer);
5252 atomic_dec(&cpu_buffer->record_disabled);
5253 atomic_dec(&cpu_buffer->resize_disabled);
5255 mutex_unlock(&buffer->mutex);
5257 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5259 /* Flag to ensure proper resetting of atomic variables */
5260 #define RESET_BIT (1 << 30)
5263 * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
5264 * @buffer: The ring buffer to reset a per cpu buffer of
5266 void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5268 struct ring_buffer_per_cpu *cpu_buffer;
5271 /* prevent another thread from changing buffer sizes */
5272 mutex_lock(&buffer->mutex);
5274 for_each_online_buffer_cpu(buffer, cpu) {
5275 cpu_buffer = buffer->buffers[cpu];
5277 atomic_add(RESET_BIT, &cpu_buffer->resize_disabled);
5278 atomic_inc(&cpu_buffer->record_disabled);
5281 /* Make sure all commits have finished */
5284 for_each_buffer_cpu(buffer, cpu) {
5285 cpu_buffer = buffer->buffers[cpu];
5288 * If a CPU came online during the synchronize_rcu(), then
5291 if (!(atomic_read(&cpu_buffer->resize_disabled) & RESET_BIT))
5294 reset_disabled_cpu_buffer(cpu_buffer);
5296 atomic_dec(&cpu_buffer->record_disabled);
5297 atomic_sub(RESET_BIT, &cpu_buffer->resize_disabled);
5300 mutex_unlock(&buffer->mutex);
5304 * ring_buffer_reset - reset a ring buffer
5305 * @buffer: The ring buffer to reset all cpu buffers
5307 void ring_buffer_reset(struct trace_buffer *buffer)
5309 struct ring_buffer_per_cpu *cpu_buffer;
5312 /* prevent another thread from changing buffer sizes */
5313 mutex_lock(&buffer->mutex);
5315 for_each_buffer_cpu(buffer, cpu) {
5316 cpu_buffer = buffer->buffers[cpu];
5318 atomic_inc(&cpu_buffer->resize_disabled);
5319 atomic_inc(&cpu_buffer->record_disabled);
5322 /* Make sure all commits have finished */
5325 for_each_buffer_cpu(buffer, cpu) {
5326 cpu_buffer = buffer->buffers[cpu];
5328 reset_disabled_cpu_buffer(cpu_buffer);
5330 atomic_dec(&cpu_buffer->record_disabled);
5331 atomic_dec(&cpu_buffer->resize_disabled);
5334 mutex_unlock(&buffer->mutex);
5336 EXPORT_SYMBOL_GPL(ring_buffer_reset);
5339 * ring_buffer_empty - is the ring buffer empty?
5340 * @buffer: The ring buffer to test
5342 bool ring_buffer_empty(struct trace_buffer *buffer)
5344 struct ring_buffer_per_cpu *cpu_buffer;
5345 unsigned long flags;
5350 /* yes this is racy, but if you don't like the race, lock the buffer */
5351 for_each_buffer_cpu(buffer, cpu) {
5352 cpu_buffer = buffer->buffers[cpu];
5353 local_irq_save(flags);
5354 dolock = rb_reader_lock(cpu_buffer);
5355 ret = rb_per_cpu_empty(cpu_buffer);
5356 rb_reader_unlock(cpu_buffer, dolock);
5357 local_irq_restore(flags);
5365 EXPORT_SYMBOL_GPL(ring_buffer_empty);
5368 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5369 * @buffer: The ring buffer
5370 * @cpu: The CPU buffer to test
5372 bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5374 struct ring_buffer_per_cpu *cpu_buffer;
5375 unsigned long flags;
5379 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5382 cpu_buffer = buffer->buffers[cpu];
5383 local_irq_save(flags);
5384 dolock = rb_reader_lock(cpu_buffer);
5385 ret = rb_per_cpu_empty(cpu_buffer);
5386 rb_reader_unlock(cpu_buffer, dolock);
5387 local_irq_restore(flags);
5391 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5393 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5395 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5396 * @buffer_a: One buffer to swap with
5397 * @buffer_b: The other buffer to swap with
5398 * @cpu: the CPU of the buffers to swap
5400 * This function is useful for tracers that want to take a "snapshot"
5401 * of a CPU buffer and has another back up buffer lying around.
5402 * it is expected that the tracer handles the cpu buffer not being
5403 * used at the moment.
5405 int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5406 struct trace_buffer *buffer_b, int cpu)
5408 struct ring_buffer_per_cpu *cpu_buffer_a;
5409 struct ring_buffer_per_cpu *cpu_buffer_b;
5412 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5413 !cpumask_test_cpu(cpu, buffer_b->cpumask))
5416 cpu_buffer_a = buffer_a->buffers[cpu];
5417 cpu_buffer_b = buffer_b->buffers[cpu];
5419 /* At least make sure the two buffers are somewhat the same */
5420 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5423 if (buffer_a->subbuf_order != buffer_b->subbuf_order)
5428 if (atomic_read(&buffer_a->record_disabled))
5431 if (atomic_read(&buffer_b->record_disabled))
5434 if (atomic_read(&cpu_buffer_a->record_disabled))
5437 if (atomic_read(&cpu_buffer_b->record_disabled))
5441 * We can't do a synchronize_rcu here because this
5442 * function can be called in atomic context.
5443 * Normally this will be called from the same CPU as cpu.
5444 * If not it's up to the caller to protect this.
5446 atomic_inc(&cpu_buffer_a->record_disabled);
5447 atomic_inc(&cpu_buffer_b->record_disabled);
5450 if (local_read(&cpu_buffer_a->committing))
5452 if (local_read(&cpu_buffer_b->committing))
5456 * When resize is in progress, we cannot swap it because
5457 * it will mess the state of the cpu buffer.
5459 if (atomic_read(&buffer_a->resizing))
5461 if (atomic_read(&buffer_b->resizing))
5464 buffer_a->buffers[cpu] = cpu_buffer_b;
5465 buffer_b->buffers[cpu] = cpu_buffer_a;
5467 cpu_buffer_b->buffer = buffer_a;
5468 cpu_buffer_a->buffer = buffer_b;
5473 atomic_dec(&cpu_buffer_a->record_disabled);
5474 atomic_dec(&cpu_buffer_b->record_disabled);
5478 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5479 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5482 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5483 * @buffer: the buffer to allocate for.
5484 * @cpu: the cpu buffer to allocate.
5486 * This function is used in conjunction with ring_buffer_read_page.
5487 * When reading a full page from the ring buffer, these functions
5488 * can be used to speed up the process. The calling function should
5489 * allocate a few pages first with this function. Then when it
5490 * needs to get pages from the ring buffer, it passes the result
5491 * of this function into ring_buffer_read_page, which will swap
5492 * the page that was allocated, with the read page of the buffer.
5495 * The page allocated, or ERR_PTR
5497 struct buffer_data_read_page *
5498 ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5500 struct ring_buffer_per_cpu *cpu_buffer;
5501 struct buffer_data_read_page *bpage = NULL;
5502 unsigned long flags;
5505 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5506 return ERR_PTR(-ENODEV);
5508 bpage = kzalloc(sizeof(*bpage), GFP_KERNEL);
5510 return ERR_PTR(-ENOMEM);
5512 bpage->order = buffer->subbuf_order;
5513 cpu_buffer = buffer->buffers[cpu];
5514 local_irq_save(flags);
5515 arch_spin_lock(&cpu_buffer->lock);
5517 if (cpu_buffer->free_page) {
5518 bpage->data = cpu_buffer->free_page;
5519 cpu_buffer->free_page = NULL;
5522 arch_spin_unlock(&cpu_buffer->lock);
5523 local_irq_restore(flags);
5528 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL | __GFP_NORETRY,
5529 cpu_buffer->buffer->subbuf_order);
5532 return ERR_PTR(-ENOMEM);
5535 bpage->data = page_address(page);
5538 rb_init_page(bpage->data);
5542 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5545 * ring_buffer_free_read_page - free an allocated read page
5546 * @buffer: the buffer the page was allocate for
5547 * @cpu: the cpu buffer the page came from
5548 * @data_page: the page to free
5550 * Free a page allocated from ring_buffer_alloc_read_page.
5552 void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu,
5553 struct buffer_data_read_page *data_page)
5555 struct ring_buffer_per_cpu *cpu_buffer;
5556 struct buffer_data_page *bpage = data_page->data;
5557 struct page *page = virt_to_page(bpage);
5558 unsigned long flags;
5560 if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
5563 cpu_buffer = buffer->buffers[cpu];
5566 * If the page is still in use someplace else, or order of the page
5567 * is different from the subbuffer order of the buffer -
5570 if (page_ref_count(page) > 1 || data_page->order != buffer->subbuf_order)
5573 local_irq_save(flags);
5574 arch_spin_lock(&cpu_buffer->lock);
5576 if (!cpu_buffer->free_page) {
5577 cpu_buffer->free_page = bpage;
5581 arch_spin_unlock(&cpu_buffer->lock);
5582 local_irq_restore(flags);
5585 free_pages((unsigned long)bpage, data_page->order);
5588 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5591 * ring_buffer_read_page - extract a page from the ring buffer
5592 * @buffer: buffer to extract from
5593 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5594 * @len: amount to extract
5595 * @cpu: the cpu of the buffer to extract
5596 * @full: should the extraction only happen when the page is full.
5598 * This function will pull out a page from the ring buffer and consume it.
5599 * @data_page must be the address of the variable that was returned
5600 * from ring_buffer_alloc_read_page. This is because the page might be used
5601 * to swap with a page in the ring buffer.
5604 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
5605 * if (IS_ERR(rpage))
5606 * return PTR_ERR(rpage);
5607 * ret = ring_buffer_read_page(buffer, rpage, len, cpu, 0);
5609 * process_page(ring_buffer_read_page_data(rpage), ret);
5610 * ring_buffer_free_read_page(buffer, cpu, rpage);
5612 * When @full is set, the function will not return true unless
5613 * the writer is off the reader page.
5615 * Note: it is up to the calling functions to handle sleeps and wakeups.
5616 * The ring buffer can be used anywhere in the kernel and can not
5617 * blindly call wake_up. The layer that uses the ring buffer must be
5618 * responsible for that.
5621 * >=0 if data has been transferred, returns the offset of consumed data.
5622 * <0 if no data has been transferred.
5624 int ring_buffer_read_page(struct trace_buffer *buffer,
5625 struct buffer_data_read_page *data_page,
5626 size_t len, int cpu, int full)
5628 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5629 struct ring_buffer_event *event;
5630 struct buffer_data_page *bpage;
5631 struct buffer_page *reader;
5632 unsigned long missed_events;
5633 unsigned long flags;
5634 unsigned int commit;
5639 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5643 * If len is not big enough to hold the page header, then
5644 * we can not copy anything.
5646 if (len <= BUF_PAGE_HDR_SIZE)
5649 len -= BUF_PAGE_HDR_SIZE;
5651 if (!data_page || !data_page->data)
5653 if (data_page->order != buffer->subbuf_order)
5656 bpage = data_page->data;
5660 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5662 reader = rb_get_reader_page(cpu_buffer);
5666 event = rb_reader_event(cpu_buffer);
5668 read = reader->read;
5669 commit = rb_page_commit(reader);
5671 /* Check if any events were dropped */
5672 missed_events = cpu_buffer->lost_events;
5675 * If this page has been partially read or
5676 * if len is not big enough to read the rest of the page or
5677 * a writer is still on the page, then
5678 * we must copy the data from the page to the buffer.
5679 * Otherwise, we can simply swap the page with the one passed in.
5681 if (read || (len < (commit - read)) ||
5682 cpu_buffer->reader_page == cpu_buffer->commit_page) {
5683 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5684 unsigned int rpos = read;
5685 unsigned int pos = 0;
5689 * If a full page is expected, this can still be returned
5690 * if there's been a previous partial read and the
5691 * rest of the page can be read and the commit page is off
5695 (!read || (len < (commit - read)) ||
5696 cpu_buffer->reader_page == cpu_buffer->commit_page))
5699 if (len > (commit - read))
5700 len = (commit - read);
5702 /* Always keep the time extend and data together */
5703 size = rb_event_ts_length(event);
5708 /* save the current timestamp, since the user will need it */
5709 save_timestamp = cpu_buffer->read_stamp;
5711 /* Need to copy one event at a time */
5713 /* We need the size of one event, because
5714 * rb_advance_reader only advances by one event,
5715 * whereas rb_event_ts_length may include the size of
5716 * one or two events.
5717 * We have already ensured there's enough space if this
5718 * is a time extend. */
5719 size = rb_event_length(event);
5720 memcpy(bpage->data + pos, rpage->data + rpos, size);
5724 rb_advance_reader(cpu_buffer);
5725 rpos = reader->read;
5731 event = rb_reader_event(cpu_buffer);
5732 /* Always keep the time extend and data together */
5733 size = rb_event_ts_length(event);
5734 } while (len >= size);
5737 local_set(&bpage->commit, pos);
5738 bpage->time_stamp = save_timestamp;
5740 /* we copied everything to the beginning */
5743 /* update the entry counter */
5744 cpu_buffer->read += rb_page_entries(reader);
5745 cpu_buffer->read_bytes += rb_page_commit(reader);
5747 /* swap the pages */
5748 rb_init_page(bpage);
5749 bpage = reader->page;
5750 reader->page = data_page->data;
5751 local_set(&reader->write, 0);
5752 local_set(&reader->entries, 0);
5754 data_page->data = bpage;
5757 * Use the real_end for the data size,
5758 * This gives us a chance to store the lost events
5761 if (reader->real_end)
5762 local_set(&bpage->commit, reader->real_end);
5766 cpu_buffer->lost_events = 0;
5768 commit = local_read(&bpage->commit);
5770 * Set a flag in the commit field if we lost events
5772 if (missed_events) {
5773 /* If there is room at the end of the page to save the
5774 * missed events, then record it there.
5776 if (buffer->subbuf_size - commit >= sizeof(missed_events)) {
5777 memcpy(&bpage->data[commit], &missed_events,
5778 sizeof(missed_events));
5779 local_add(RB_MISSED_STORED, &bpage->commit);
5780 commit += sizeof(missed_events);
5782 local_add(RB_MISSED_EVENTS, &bpage->commit);
5786 * This page may be off to user land. Zero it out here.
5788 if (commit < buffer->subbuf_size)
5789 memset(&bpage->data[commit], 0, buffer->subbuf_size - commit);
5792 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5797 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5800 * ring_buffer_read_page_data - get pointer to the data in the page.
5801 * @page: the page to get the data from
5803 * Returns pointer to the actual data in this page.
5805 void *ring_buffer_read_page_data(struct buffer_data_read_page *page)
5809 EXPORT_SYMBOL_GPL(ring_buffer_read_page_data);
5812 * ring_buffer_subbuf_size_get - get size of the sub buffer.
5813 * @buffer: the buffer to get the sub buffer size from
5815 * Returns size of the sub buffer, in bytes.
5817 int ring_buffer_subbuf_size_get(struct trace_buffer *buffer)
5819 return buffer->subbuf_size + BUF_PAGE_HDR_SIZE;
5821 EXPORT_SYMBOL_GPL(ring_buffer_subbuf_size_get);
5824 * ring_buffer_subbuf_order_get - get order of system sub pages in one buffer page.
5825 * @buffer: The ring_buffer to get the system sub page order from
5827 * By default, one ring buffer sub page equals to one system page. This parameter
5828 * is configurable, per ring buffer. The size of the ring buffer sub page can be
5829 * extended, but must be an order of system page size.
5831 * Returns the order of buffer sub page size, in system pages:
5832 * 0 means the sub buffer size is 1 system page and so forth.
5833 * In case of an error < 0 is returned.
5835 int ring_buffer_subbuf_order_get(struct trace_buffer *buffer)
5840 return buffer->subbuf_order;
5842 EXPORT_SYMBOL_GPL(ring_buffer_subbuf_order_get);
5845 * ring_buffer_subbuf_order_set - set the size of ring buffer sub page.
5846 * @buffer: The ring_buffer to set the new page size.
5847 * @order: Order of the system pages in one sub buffer page
5849 * By default, one ring buffer pages equals to one system page. This API can be
5850 * used to set new size of the ring buffer page. The size must be order of
5851 * system page size, that's why the input parameter @order is the order of
5852 * system pages that are allocated for one ring buffer page:
5854 * 1 - 2 system pages
5855 * 3 - 4 system pages
5858 * Returns 0 on success or < 0 in case of an error.
5860 int ring_buffer_subbuf_order_set(struct trace_buffer *buffer, int order)
5862 struct ring_buffer_per_cpu *cpu_buffer;
5863 struct buffer_page *bpage, *tmp;
5864 int old_order, old_size;
5870 if (!buffer || order < 0)
5873 if (buffer->subbuf_order == order)
5876 psize = (1 << order) * PAGE_SIZE;
5877 if (psize <= BUF_PAGE_HDR_SIZE)
5880 old_order = buffer->subbuf_order;
5881 old_size = buffer->subbuf_size;
5883 /* prevent another thread from changing buffer sizes */
5884 mutex_lock(&buffer->mutex);
5885 atomic_inc(&buffer->record_disabled);
5887 /* Make sure all commits have finished */
5890 buffer->subbuf_order = order;
5891 buffer->subbuf_size = psize - BUF_PAGE_HDR_SIZE;
5893 /* Make sure all new buffers are allocated, before deleting the old ones */
5894 for_each_buffer_cpu(buffer, cpu) {
5896 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5899 cpu_buffer = buffer->buffers[cpu];
5901 /* Update the number of pages to match the new size */
5902 nr_pages = old_size * buffer->buffers[cpu]->nr_pages;
5903 nr_pages = DIV_ROUND_UP(nr_pages, buffer->subbuf_size);
5905 /* we need a minimum of two pages */
5909 cpu_buffer->nr_pages_to_update = nr_pages;
5911 /* Include the reader page */
5914 /* Allocate the new size buffer */
5915 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5916 if (__rb_allocate_pages(cpu_buffer, nr_pages,
5917 &cpu_buffer->new_pages)) {
5918 /* not enough memory for new pages */
5924 for_each_buffer_cpu(buffer, cpu) {
5926 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5929 cpu_buffer = buffer->buffers[cpu];
5931 /* Clear the head bit to make the link list normal to read */
5932 rb_head_page_deactivate(cpu_buffer);
5934 /* Now walk the list and free all the old sub buffers */
5935 list_for_each_entry_safe(bpage, tmp, cpu_buffer->pages, list) {
5936 list_del_init(&bpage->list);
5937 free_buffer_page(bpage);
5939 /* The above loop stopped an the last page needing to be freed */
5940 bpage = list_entry(cpu_buffer->pages, struct buffer_page, list);
5941 free_buffer_page(bpage);
5943 /* Free the current reader page */
5944 free_buffer_page(cpu_buffer->reader_page);
5946 /* One page was allocated for the reader page */
5947 cpu_buffer->reader_page = list_entry(cpu_buffer->new_pages.next,
5948 struct buffer_page, list);
5949 list_del_init(&cpu_buffer->reader_page->list);
5951 /* The cpu_buffer pages are a link list with no head */
5952 cpu_buffer->pages = cpu_buffer->new_pages.next;
5953 cpu_buffer->new_pages.next->prev = cpu_buffer->new_pages.prev;
5954 cpu_buffer->new_pages.prev->next = cpu_buffer->new_pages.next;
5956 /* Clear the new_pages list */
5957 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5959 cpu_buffer->head_page
5960 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5961 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
5963 cpu_buffer->nr_pages = cpu_buffer->nr_pages_to_update;
5964 cpu_buffer->nr_pages_to_update = 0;
5966 free_pages((unsigned long)cpu_buffer->free_page, old_order);
5967 cpu_buffer->free_page = NULL;
5969 rb_head_page_activate(cpu_buffer);
5971 rb_check_pages(cpu_buffer);
5974 atomic_dec(&buffer->record_disabled);
5975 mutex_unlock(&buffer->mutex);
5980 buffer->subbuf_order = old_order;
5981 buffer->subbuf_size = old_size;
5983 atomic_dec(&buffer->record_disabled);
5984 mutex_unlock(&buffer->mutex);
5986 for_each_buffer_cpu(buffer, cpu) {
5987 cpu_buffer = buffer->buffers[cpu];
5989 if (!cpu_buffer->nr_pages_to_update)
5992 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages, list) {
5993 list_del_init(&bpage->list);
5994 free_buffer_page(bpage);
6000 EXPORT_SYMBOL_GPL(ring_buffer_subbuf_order_set);
6003 * We only allocate new buffers, never free them if the CPU goes down.
6004 * If we were to free the buffer, then the user would lose any trace that was in
6007 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
6009 struct trace_buffer *buffer;
6012 unsigned long nr_pages;
6014 buffer = container_of(node, struct trace_buffer, node);
6015 if (cpumask_test_cpu(cpu, buffer->cpumask))
6020 /* check if all cpu sizes are same */
6021 for_each_buffer_cpu(buffer, cpu_i) {
6022 /* fill in the size from first enabled cpu */
6024 nr_pages = buffer->buffers[cpu_i]->nr_pages;
6025 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
6030 /* allocate minimum pages, user can later expand it */
6033 buffer->buffers[cpu] =
6034 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
6035 if (!buffer->buffers[cpu]) {
6036 WARN(1, "failed to allocate ring buffer on CPU %u\n",
6041 cpumask_set_cpu(cpu, buffer->cpumask);
6045 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
6047 * This is a basic integrity check of the ring buffer.
6048 * Late in the boot cycle this test will run when configured in.
6049 * It will kick off a thread per CPU that will go into a loop
6050 * writing to the per cpu ring buffer various sizes of data.
6051 * Some of the data will be large items, some small.
6053 * Another thread is created that goes into a spin, sending out
6054 * IPIs to the other CPUs to also write into the ring buffer.
6055 * this is to test the nesting ability of the buffer.
6057 * Basic stats are recorded and reported. If something in the
6058 * ring buffer should happen that's not expected, a big warning
6059 * is displayed and all ring buffers are disabled.
6061 static struct task_struct *rb_threads[NR_CPUS] __initdata;
6063 struct rb_test_data {
6064 struct trace_buffer *buffer;
6065 unsigned long events;
6066 unsigned long bytes_written;
6067 unsigned long bytes_alloc;
6068 unsigned long bytes_dropped;
6069 unsigned long events_nested;
6070 unsigned long bytes_written_nested;
6071 unsigned long bytes_alloc_nested;
6072 unsigned long bytes_dropped_nested;
6073 int min_size_nested;
6074 int max_size_nested;
6081 static struct rb_test_data rb_data[NR_CPUS] __initdata;
6084 #define RB_TEST_BUFFER_SIZE 1048576
6086 static char rb_string[] __initdata =
6087 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
6088 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
6089 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
6091 static bool rb_test_started __initdata;
6098 static __init int rb_write_something(struct rb_test_data *data, bool nested)
6100 struct ring_buffer_event *event;
6101 struct rb_item *item;
6108 /* Have nested writes different that what is written */
6109 cnt = data->cnt + (nested ? 27 : 0);
6111 /* Multiply cnt by ~e, to make some unique increment */
6112 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
6114 len = size + sizeof(struct rb_item);
6116 started = rb_test_started;
6117 /* read rb_test_started before checking buffer enabled */
6120 event = ring_buffer_lock_reserve(data->buffer, len);
6122 /* Ignore dropped events before test starts. */
6125 data->bytes_dropped += len;
6127 data->bytes_dropped_nested += len;
6132 event_len = ring_buffer_event_length(event);
6134 if (RB_WARN_ON(data->buffer, event_len < len))
6137 item = ring_buffer_event_data(event);
6139 memcpy(item->str, rb_string, size);
6142 data->bytes_alloc_nested += event_len;
6143 data->bytes_written_nested += len;
6144 data->events_nested++;
6145 if (!data->min_size_nested || len < data->min_size_nested)
6146 data->min_size_nested = len;
6147 if (len > data->max_size_nested)
6148 data->max_size_nested = len;
6150 data->bytes_alloc += event_len;
6151 data->bytes_written += len;
6153 if (!data->min_size || len < data->min_size)
6154 data->max_size = len;
6155 if (len > data->max_size)
6156 data->max_size = len;
6160 ring_buffer_unlock_commit(data->buffer);
6165 static __init int rb_test(void *arg)
6167 struct rb_test_data *data = arg;
6169 while (!kthread_should_stop()) {
6170 rb_write_something(data, false);
6173 set_current_state(TASK_INTERRUPTIBLE);
6174 /* Now sleep between a min of 100-300us and a max of 1ms */
6175 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
6181 static __init void rb_ipi(void *ignore)
6183 struct rb_test_data *data;
6184 int cpu = smp_processor_id();
6186 data = &rb_data[cpu];
6187 rb_write_something(data, true);
6190 static __init int rb_hammer_test(void *arg)
6192 while (!kthread_should_stop()) {
6194 /* Send an IPI to all cpus to write data! */
6195 smp_call_function(rb_ipi, NULL, 1);
6196 /* No sleep, but for non preempt, let others run */
6203 static __init int test_ringbuffer(void)
6205 struct task_struct *rb_hammer;
6206 struct trace_buffer *buffer;
6210 if (security_locked_down(LOCKDOWN_TRACEFS)) {
6211 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
6215 pr_info("Running ring buffer tests...\n");
6217 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
6218 if (WARN_ON(!buffer))
6221 /* Disable buffer so that threads can't write to it yet */
6222 ring_buffer_record_off(buffer);
6224 for_each_online_cpu(cpu) {
6225 rb_data[cpu].buffer = buffer;
6226 rb_data[cpu].cpu = cpu;
6227 rb_data[cpu].cnt = cpu;
6228 rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
6229 cpu, "rbtester/%u");
6230 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
6231 pr_cont("FAILED\n");
6232 ret = PTR_ERR(rb_threads[cpu]);
6237 /* Now create the rb hammer! */
6238 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
6239 if (WARN_ON(IS_ERR(rb_hammer))) {
6240 pr_cont("FAILED\n");
6241 ret = PTR_ERR(rb_hammer);
6245 ring_buffer_record_on(buffer);
6247 * Show buffer is enabled before setting rb_test_started.
6248 * Yes there's a small race window where events could be
6249 * dropped and the thread wont catch it. But when a ring
6250 * buffer gets enabled, there will always be some kind of
6251 * delay before other CPUs see it. Thus, we don't care about
6252 * those dropped events. We care about events dropped after
6253 * the threads see that the buffer is active.
6256 rb_test_started = true;
6258 set_current_state(TASK_INTERRUPTIBLE);
6259 /* Just run for 10 seconds */;
6260 schedule_timeout(10 * HZ);
6262 kthread_stop(rb_hammer);
6265 for_each_online_cpu(cpu) {
6266 if (!rb_threads[cpu])
6268 kthread_stop(rb_threads[cpu]);
6271 ring_buffer_free(buffer);
6276 pr_info("finished\n");
6277 for_each_online_cpu(cpu) {
6278 struct ring_buffer_event *event;
6279 struct rb_test_data *data = &rb_data[cpu];
6280 struct rb_item *item;
6281 unsigned long total_events;
6282 unsigned long total_dropped;
6283 unsigned long total_written;
6284 unsigned long total_alloc;
6285 unsigned long total_read = 0;
6286 unsigned long total_size = 0;
6287 unsigned long total_len = 0;
6288 unsigned long total_lost = 0;
6291 int small_event_size;
6295 total_events = data->events + data->events_nested;
6296 total_written = data->bytes_written + data->bytes_written_nested;
6297 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
6298 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
6300 big_event_size = data->max_size + data->max_size_nested;
6301 small_event_size = data->min_size + data->min_size_nested;
6303 pr_info("CPU %d:\n", cpu);
6304 pr_info(" events: %ld\n", total_events);
6305 pr_info(" dropped bytes: %ld\n", total_dropped);
6306 pr_info(" alloced bytes: %ld\n", total_alloc);
6307 pr_info(" written bytes: %ld\n", total_written);
6308 pr_info(" biggest event: %d\n", big_event_size);
6309 pr_info(" smallest event: %d\n", small_event_size);
6311 if (RB_WARN_ON(buffer, total_dropped))
6316 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
6318 item = ring_buffer_event_data(event);
6319 total_len += ring_buffer_event_length(event);
6320 total_size += item->size + sizeof(struct rb_item);
6321 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
6322 pr_info("FAILED!\n");
6323 pr_info("buffer had: %.*s\n", item->size, item->str);
6324 pr_info("expected: %.*s\n", item->size, rb_string);
6325 RB_WARN_ON(buffer, 1);
6336 pr_info(" read events: %ld\n", total_read);
6337 pr_info(" lost events: %ld\n", total_lost);
6338 pr_info(" total events: %ld\n", total_lost + total_read);
6339 pr_info(" recorded len bytes: %ld\n", total_len);
6340 pr_info(" recorded size bytes: %ld\n", total_size);
6342 pr_info(" With dropped events, record len and size may not match\n"
6343 " alloced and written from above\n");
6345 if (RB_WARN_ON(buffer, total_len != total_alloc ||
6346 total_size != total_written))
6349 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
6355 pr_info("Ring buffer PASSED!\n");
6357 ring_buffer_free(buffer);
6361 late_initcall(test_ringbuffer);
6362 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */