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 /* For waiters waiting for the first wake up */
758 (void)atomic_fetch_inc_release(&rbwork->seq);
760 wake_up_all(&rbwork->waiters);
761 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
762 /* Only cpu_buffer sets the above flags */
763 struct ring_buffer_per_cpu *cpu_buffer =
764 container_of(rbwork, struct ring_buffer_per_cpu, irq_work);
766 /* Called from interrupt context */
767 raw_spin_lock(&cpu_buffer->reader_lock);
768 rbwork->wakeup_full = false;
769 rbwork->full_waiters_pending = false;
771 /* Waking up all waiters, they will reset the shortest full */
772 cpu_buffer->shortest_full = 0;
773 raw_spin_unlock(&cpu_buffer->reader_lock);
775 wake_up_all(&rbwork->full_waiters);
780 * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
781 * @buffer: The ring buffer to wake waiters on
782 * @cpu: The CPU buffer to wake waiters on
784 * In the case of a file that represents a ring buffer is closing,
785 * it is prudent to wake up any waiters that are on this.
787 void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
789 struct ring_buffer_per_cpu *cpu_buffer;
790 struct rb_irq_work *rbwork;
795 if (cpu == RING_BUFFER_ALL_CPUS) {
797 /* Wake up individual ones too. One level recursion */
798 for_each_buffer_cpu(buffer, cpu)
799 ring_buffer_wake_waiters(buffer, cpu);
801 rbwork = &buffer->irq_work;
803 if (WARN_ON_ONCE(!buffer->buffers))
805 if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
808 cpu_buffer = buffer->buffers[cpu];
809 /* The CPU buffer may not have been initialized yet */
812 rbwork = &cpu_buffer->irq_work;
815 /* This can be called in any context */
816 irq_work_queue(&rbwork->work);
819 static bool rb_watermark_hit(struct trace_buffer *buffer, int cpu, int full)
821 struct ring_buffer_per_cpu *cpu_buffer;
824 /* Reads of all CPUs always waits for any data */
825 if (cpu == RING_BUFFER_ALL_CPUS)
826 return !ring_buffer_empty(buffer);
828 cpu_buffer = buffer->buffers[cpu];
830 if (!ring_buffer_empty_cpu(buffer, cpu)) {
837 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
838 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
839 ret = !pagebusy && full_hit(buffer, cpu, full);
841 if (!ret && (!cpu_buffer->shortest_full ||
842 cpu_buffer->shortest_full > full)) {
843 cpu_buffer->shortest_full = full;
845 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
851 rb_wait_cond(struct rb_irq_work *rbwork, struct trace_buffer *buffer,
852 int cpu, int full, ring_buffer_cond_fn cond, void *data)
854 if (rb_watermark_hit(buffer, cpu, full))
861 * The events can happen in critical sections where
862 * checking a work queue can cause deadlocks.
863 * After adding a task to the queue, this flag is set
864 * only to notify events to try to wake up the queue
867 * We don't clear it even if the buffer is no longer
868 * empty. The flag only causes the next event to run
869 * irq_work to do the work queue wake up. The worse
870 * that can happen if we race with !trace_empty() is that
871 * an event will cause an irq_work to try to wake up
874 * There's no reason to protect this flag either, as
875 * the work queue and irq_work logic will do the necessary
876 * synchronization for the wake ups. The only thing
877 * that is necessary is that the wake up happens after
878 * a task has been queued. It's OK for spurious wake ups.
881 rbwork->full_waiters_pending = true;
883 rbwork->waiters_pending = true;
888 struct rb_wait_data {
889 struct rb_irq_work *irq_work;
894 * The default wait condition for ring_buffer_wait() is to just to exit the
895 * wait loop the first time it is woken up.
897 static bool rb_wait_once(void *data)
899 struct rb_wait_data *rdata = data;
900 struct rb_irq_work *rbwork = rdata->irq_work;
902 return atomic_read_acquire(&rbwork->seq) != rdata->seq;
906 * ring_buffer_wait - wait for input to the ring buffer
907 * @buffer: buffer to wait on
908 * @cpu: the cpu buffer to wait on
909 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
910 * @cond: condition function to break out of wait (NULL to run once)
911 * @data: the data to pass to @cond.
913 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
914 * as data is added to any of the @buffer's cpu buffers. Otherwise
915 * it will wait for data to be added to a specific cpu buffer.
917 int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full,
918 ring_buffer_cond_fn cond, void *data)
920 struct ring_buffer_per_cpu *cpu_buffer;
921 struct wait_queue_head *waitq;
922 struct rb_irq_work *rbwork;
923 struct rb_wait_data rdata;
927 * Depending on what the caller is waiting for, either any
928 * data in any cpu buffer, or a specific buffer, put the
929 * caller on the appropriate wait queue.
931 if (cpu == RING_BUFFER_ALL_CPUS) {
932 rbwork = &buffer->irq_work;
933 /* Full only makes sense on per cpu reads */
936 if (!cpumask_test_cpu(cpu, buffer->cpumask))
938 cpu_buffer = buffer->buffers[cpu];
939 rbwork = &cpu_buffer->irq_work;
943 waitq = &rbwork->full_waiters;
945 waitq = &rbwork->waiters;
947 /* Set up to exit loop as soon as it is woken */
950 rdata.irq_work = rbwork;
951 rdata.seq = atomic_read_acquire(&rbwork->seq);
955 ret = wait_event_interruptible((*waitq),
956 rb_wait_cond(rbwork, buffer, cpu, full, cond, data));
962 * ring_buffer_poll_wait - poll on buffer input
963 * @buffer: buffer to wait on
964 * @cpu: the cpu buffer to wait on
965 * @filp: the file descriptor
966 * @poll_table: The poll descriptor
967 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
969 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
970 * as data is added to any of the @buffer's cpu buffers. Otherwise
971 * it will wait for data to be added to a specific cpu buffer.
973 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
976 __poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
977 struct file *filp, poll_table *poll_table, int full)
979 struct ring_buffer_per_cpu *cpu_buffer;
980 struct rb_irq_work *rbwork;
982 if (cpu == RING_BUFFER_ALL_CPUS) {
983 rbwork = &buffer->irq_work;
986 if (!cpumask_test_cpu(cpu, buffer->cpumask))
989 cpu_buffer = buffer->buffers[cpu];
990 rbwork = &cpu_buffer->irq_work;
994 poll_wait(filp, &rbwork->full_waiters, poll_table);
996 if (rb_watermark_hit(buffer, cpu, full))
997 return EPOLLIN | EPOLLRDNORM;
999 * Only allow full_waiters_pending update to be seen after
1000 * the shortest_full is set (in rb_watermark_hit). If the
1001 * writer sees the full_waiters_pending flag set, it will
1002 * compare the amount in the ring buffer to shortest_full.
1003 * If the amount in the ring buffer is greater than the
1004 * shortest_full percent, it will call the irq_work handler
1005 * to wake up this list. The irq_handler will reset shortest_full
1006 * back to zero. That's done under the reader_lock, but
1007 * the below smp_mb() makes sure that the update to
1008 * full_waiters_pending doesn't leak up into the above.
1011 rbwork->full_waiters_pending = true;
1015 poll_wait(filp, &rbwork->waiters, poll_table);
1016 rbwork->waiters_pending = true;
1019 * There's a tight race between setting the waiters_pending and
1020 * checking if the ring buffer is empty. Once the waiters_pending bit
1021 * is set, the next event will wake the task up, but we can get stuck
1022 * if there's only a single event in.
1024 * FIXME: Ideally, we need a memory barrier on the writer side as well,
1025 * but adding a memory barrier to all events will cause too much of a
1026 * performance hit in the fast path. We only need a memory barrier when
1027 * the buffer goes from empty to having content. But as this race is
1028 * extremely small, and it's not a problem if another event comes in, we
1029 * will fix it later.
1033 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1034 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1035 return EPOLLIN | EPOLLRDNORM;
1039 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
1040 #define RB_WARN_ON(b, cond) \
1042 int _____ret = unlikely(cond); \
1044 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1045 struct ring_buffer_per_cpu *__b = \
1047 atomic_inc(&__b->buffer->record_disabled); \
1049 atomic_inc(&b->record_disabled); \
1055 /* Up this if you want to test the TIME_EXTENTS and normalization */
1056 #define DEBUG_SHIFT 0
1058 static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1062 /* Skip retpolines :-( */
1063 if (IS_ENABLED(CONFIG_MITIGATION_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1064 ts = trace_clock_local();
1066 ts = buffer->clock();
1068 /* shift to debug/test normalization and TIME_EXTENTS */
1069 return ts << DEBUG_SHIFT;
1072 u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1076 preempt_disable_notrace();
1077 time = rb_time_stamp(buffer);
1078 preempt_enable_notrace();
1082 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1084 void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1087 /* Just stupid testing the normalize function and deltas */
1088 *ts >>= DEBUG_SHIFT;
1090 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1093 * Making the ring buffer lockless makes things tricky.
1094 * Although writes only happen on the CPU that they are on,
1095 * and they only need to worry about interrupts. Reads can
1096 * happen on any CPU.
1098 * The reader page is always off the ring buffer, but when the
1099 * reader finishes with a page, it needs to swap its page with
1100 * a new one from the buffer. The reader needs to take from
1101 * the head (writes go to the tail). But if a writer is in overwrite
1102 * mode and wraps, it must push the head page forward.
1104 * Here lies the problem.
1106 * The reader must be careful to replace only the head page, and
1107 * not another one. As described at the top of the file in the
1108 * ASCII art, the reader sets its old page to point to the next
1109 * page after head. It then sets the page after head to point to
1110 * the old reader page. But if the writer moves the head page
1111 * during this operation, the reader could end up with the tail.
1113 * We use cmpxchg to help prevent this race. We also do something
1114 * special with the page before head. We set the LSB to 1.
1116 * When the writer must push the page forward, it will clear the
1117 * bit that points to the head page, move the head, and then set
1118 * the bit that points to the new head page.
1120 * We also don't want an interrupt coming in and moving the head
1121 * page on another writer. Thus we use the second LSB to catch
1124 * head->list->prev->next bit 1 bit 0
1127 * Points to head page 0 1
1130 * Note we can not trust the prev pointer of the head page, because:
1132 * +----+ +-----+ +-----+
1133 * | |------>| T |---X--->| N |
1135 * +----+ +-----+ +-----+
1138 * +----------| R |----------+ |
1142 * Key: ---X--> HEAD flag set in pointer
1147 * (see __rb_reserve_next() to see where this happens)
1149 * What the above shows is that the reader just swapped out
1150 * the reader page with a page in the buffer, but before it
1151 * could make the new header point back to the new page added
1152 * it was preempted by a writer. The writer moved forward onto
1153 * the new page added by the reader and is about to move forward
1156 * You can see, it is legitimate for the previous pointer of
1157 * the head (or any page) not to point back to itself. But only
1161 #define RB_PAGE_NORMAL 0UL
1162 #define RB_PAGE_HEAD 1UL
1163 #define RB_PAGE_UPDATE 2UL
1166 #define RB_FLAG_MASK 3UL
1168 /* PAGE_MOVED is not part of the mask */
1169 #define RB_PAGE_MOVED 4UL
1172 * rb_list_head - remove any bit
1174 static struct list_head *rb_list_head(struct list_head *list)
1176 unsigned long val = (unsigned long)list;
1178 return (struct list_head *)(val & ~RB_FLAG_MASK);
1182 * rb_is_head_page - test if the given page is the head page
1184 * Because the reader may move the head_page pointer, we can
1185 * not trust what the head page is (it may be pointing to
1186 * the reader page). But if the next page is a header page,
1187 * its flags will be non zero.
1190 rb_is_head_page(struct buffer_page *page, struct list_head *list)
1194 val = (unsigned long)list->next;
1196 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1197 return RB_PAGE_MOVED;
1199 return val & RB_FLAG_MASK;
1205 * The unique thing about the reader page, is that, if the
1206 * writer is ever on it, the previous pointer never points
1207 * back to the reader page.
1209 static bool rb_is_reader_page(struct buffer_page *page)
1211 struct list_head *list = page->list.prev;
1213 return rb_list_head(list->next) != &page->list;
1217 * rb_set_list_to_head - set a list_head to be pointing to head.
1219 static void rb_set_list_to_head(struct list_head *list)
1223 ptr = (unsigned long *)&list->next;
1224 *ptr |= RB_PAGE_HEAD;
1225 *ptr &= ~RB_PAGE_UPDATE;
1229 * rb_head_page_activate - sets up head page
1231 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1233 struct buffer_page *head;
1235 head = cpu_buffer->head_page;
1240 * Set the previous list pointer to have the HEAD flag.
1242 rb_set_list_to_head(head->list.prev);
1245 static void rb_list_head_clear(struct list_head *list)
1247 unsigned long *ptr = (unsigned long *)&list->next;
1249 *ptr &= ~RB_FLAG_MASK;
1253 * rb_head_page_deactivate - clears head page ptr (for free list)
1256 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1258 struct list_head *hd;
1260 /* Go through the whole list and clear any pointers found. */
1261 rb_list_head_clear(cpu_buffer->pages);
1263 list_for_each(hd, cpu_buffer->pages)
1264 rb_list_head_clear(hd);
1267 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1268 struct buffer_page *head,
1269 struct buffer_page *prev,
1270 int old_flag, int new_flag)
1272 struct list_head *list;
1273 unsigned long val = (unsigned long)&head->list;
1278 val &= ~RB_FLAG_MASK;
1280 ret = cmpxchg((unsigned long *)&list->next,
1281 val | old_flag, val | new_flag);
1283 /* check if the reader took the page */
1284 if ((ret & ~RB_FLAG_MASK) != val)
1285 return RB_PAGE_MOVED;
1287 return ret & RB_FLAG_MASK;
1290 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1291 struct buffer_page *head,
1292 struct buffer_page *prev,
1295 return rb_head_page_set(cpu_buffer, head, prev,
1296 old_flag, RB_PAGE_UPDATE);
1299 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1300 struct buffer_page *head,
1301 struct buffer_page *prev,
1304 return rb_head_page_set(cpu_buffer, head, prev,
1305 old_flag, RB_PAGE_HEAD);
1308 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1309 struct buffer_page *head,
1310 struct buffer_page *prev,
1313 return rb_head_page_set(cpu_buffer, head, prev,
1314 old_flag, RB_PAGE_NORMAL);
1317 static inline void rb_inc_page(struct buffer_page **bpage)
1319 struct list_head *p = rb_list_head((*bpage)->list.next);
1321 *bpage = list_entry(p, struct buffer_page, list);
1324 static struct buffer_page *
1325 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1327 struct buffer_page *head;
1328 struct buffer_page *page;
1329 struct list_head *list;
1332 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1336 list = cpu_buffer->pages;
1337 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1340 page = head = cpu_buffer->head_page;
1342 * It is possible that the writer moves the header behind
1343 * where we started, and we miss in one loop.
1344 * A second loop should grab the header, but we'll do
1345 * three loops just because I'm paranoid.
1347 for (i = 0; i < 3; i++) {
1349 if (rb_is_head_page(page, page->list.prev)) {
1350 cpu_buffer->head_page = page;
1354 } while (page != head);
1357 RB_WARN_ON(cpu_buffer, 1);
1362 static bool rb_head_page_replace(struct buffer_page *old,
1363 struct buffer_page *new)
1365 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1368 val = *ptr & ~RB_FLAG_MASK;
1369 val |= RB_PAGE_HEAD;
1371 return try_cmpxchg(ptr, &val, (unsigned long)&new->list);
1375 * rb_tail_page_update - move the tail page forward
1377 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1378 struct buffer_page *tail_page,
1379 struct buffer_page *next_page)
1381 unsigned long old_entries;
1382 unsigned long old_write;
1385 * The tail page now needs to be moved forward.
1387 * We need to reset the tail page, but without messing
1388 * with possible erasing of data brought in by interrupts
1389 * that have moved the tail page and are currently on it.
1391 * We add a counter to the write field to denote this.
1393 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1394 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1396 local_inc(&cpu_buffer->pages_touched);
1398 * Just make sure we have seen our old_write and synchronize
1399 * with any interrupts that come in.
1404 * If the tail page is still the same as what we think
1405 * it is, then it is up to us to update the tail
1408 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1409 /* Zero the write counter */
1410 unsigned long val = old_write & ~RB_WRITE_MASK;
1411 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1414 * This will only succeed if an interrupt did
1415 * not come in and change it. In which case, we
1416 * do not want to modify it.
1418 * We add (void) to let the compiler know that we do not care
1419 * about the return value of these functions. We use the
1420 * cmpxchg to only update if an interrupt did not already
1421 * do it for us. If the cmpxchg fails, we don't care.
1423 (void)local_cmpxchg(&next_page->write, old_write, val);
1424 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1427 * No need to worry about races with clearing out the commit.
1428 * it only can increment when a commit takes place. But that
1429 * only happens in the outer most nested commit.
1431 local_set(&next_page->page->commit, 0);
1433 /* Again, either we update tail_page or an interrupt does */
1434 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1438 static void rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1439 struct buffer_page *bpage)
1441 unsigned long val = (unsigned long)bpage;
1443 RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK);
1447 * rb_check_pages - integrity check of buffer pages
1448 * @cpu_buffer: CPU buffer with pages to test
1450 * As a safety measure we check to make sure the data pages have not
1453 static void rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1455 struct list_head *head = rb_list_head(cpu_buffer->pages);
1456 struct list_head *tmp;
1458 if (RB_WARN_ON(cpu_buffer,
1459 rb_list_head(rb_list_head(head->next)->prev) != head))
1462 if (RB_WARN_ON(cpu_buffer,
1463 rb_list_head(rb_list_head(head->prev)->next) != head))
1466 for (tmp = rb_list_head(head->next); tmp != head; tmp = rb_list_head(tmp->next)) {
1467 if (RB_WARN_ON(cpu_buffer,
1468 rb_list_head(rb_list_head(tmp->next)->prev) != tmp))
1471 if (RB_WARN_ON(cpu_buffer,
1472 rb_list_head(rb_list_head(tmp->prev)->next) != tmp))
1477 static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1478 long nr_pages, struct list_head *pages)
1480 struct buffer_page *bpage, *tmp;
1481 bool user_thread = current->mm != NULL;
1486 * Check if the available memory is there first.
1487 * Note, si_mem_available() only gives us a rough estimate of available
1488 * memory. It may not be accurate. But we don't care, we just want
1489 * to prevent doing any allocation when it is obvious that it is
1490 * not going to succeed.
1492 i = si_mem_available();
1497 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1498 * gracefully without invoking oom-killer and the system is not
1501 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1504 * If a user thread allocates too much, and si_mem_available()
1505 * reports there's enough memory, even though there is not.
1506 * Make sure the OOM killer kills this thread. This can happen
1507 * even with RETRY_MAYFAIL because another task may be doing
1508 * an allocation after this task has taken all memory.
1509 * This is the task the OOM killer needs to take out during this
1510 * loop, even if it was triggered by an allocation somewhere else.
1513 set_current_oom_origin();
1514 for (i = 0; i < nr_pages; i++) {
1517 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1518 mflags, cpu_to_node(cpu_buffer->cpu));
1522 rb_check_bpage(cpu_buffer, bpage);
1524 list_add(&bpage->list, pages);
1526 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu),
1527 mflags | __GFP_ZERO,
1528 cpu_buffer->buffer->subbuf_order);
1531 bpage->page = page_address(page);
1532 bpage->order = cpu_buffer->buffer->subbuf_order;
1533 rb_init_page(bpage->page);
1535 if (user_thread && fatal_signal_pending(current))
1539 clear_current_oom_origin();
1544 list_for_each_entry_safe(bpage, tmp, pages, list) {
1545 list_del_init(&bpage->list);
1546 free_buffer_page(bpage);
1549 clear_current_oom_origin();
1554 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1555 unsigned long nr_pages)
1561 if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
1565 * The ring buffer page list is a circular list that does not
1566 * start and end with a list head. All page list items point to
1569 cpu_buffer->pages = pages.next;
1572 cpu_buffer->nr_pages = nr_pages;
1574 rb_check_pages(cpu_buffer);
1579 static struct ring_buffer_per_cpu *
1580 rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1582 struct ring_buffer_per_cpu *cpu_buffer;
1583 struct buffer_page *bpage;
1587 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1588 GFP_KERNEL, cpu_to_node(cpu));
1592 cpu_buffer->cpu = cpu;
1593 cpu_buffer->buffer = buffer;
1594 raw_spin_lock_init(&cpu_buffer->reader_lock);
1595 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1596 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1597 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1598 init_completion(&cpu_buffer->update_done);
1599 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1600 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1601 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1603 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1604 GFP_KERNEL, cpu_to_node(cpu));
1606 goto fail_free_buffer;
1608 rb_check_bpage(cpu_buffer, bpage);
1610 cpu_buffer->reader_page = bpage;
1612 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL | __GFP_ZERO,
1613 cpu_buffer->buffer->subbuf_order);
1615 goto fail_free_reader;
1616 bpage->page = page_address(page);
1617 rb_init_page(bpage->page);
1619 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1620 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1622 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1624 goto fail_free_reader;
1626 cpu_buffer->head_page
1627 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1628 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1630 rb_head_page_activate(cpu_buffer);
1635 free_buffer_page(cpu_buffer->reader_page);
1642 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1644 struct list_head *head = cpu_buffer->pages;
1645 struct buffer_page *bpage, *tmp;
1647 irq_work_sync(&cpu_buffer->irq_work.work);
1649 free_buffer_page(cpu_buffer->reader_page);
1652 rb_head_page_deactivate(cpu_buffer);
1654 list_for_each_entry_safe(bpage, tmp, head, list) {
1655 list_del_init(&bpage->list);
1656 free_buffer_page(bpage);
1658 bpage = list_entry(head, struct buffer_page, list);
1659 free_buffer_page(bpage);
1662 free_page((unsigned long)cpu_buffer->free_page);
1668 * __ring_buffer_alloc - allocate a new ring_buffer
1669 * @size: the size in bytes per cpu that is needed.
1670 * @flags: attributes to set for the ring buffer.
1671 * @key: ring buffer reader_lock_key.
1673 * Currently the only flag that is available is the RB_FL_OVERWRITE
1674 * flag. This flag means that the buffer will overwrite old data
1675 * when the buffer wraps. If this flag is not set, the buffer will
1676 * drop data when the tail hits the head.
1678 struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1679 struct lock_class_key *key)
1681 struct trace_buffer *buffer;
1687 /* keep it in its own cache line */
1688 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1693 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1694 goto fail_free_buffer;
1696 /* Default buffer page size - one system page */
1697 buffer->subbuf_order = 0;
1698 buffer->subbuf_size = PAGE_SIZE - BUF_PAGE_HDR_SIZE;
1700 /* Max payload is buffer page size - header (8bytes) */
1701 buffer->max_data_size = buffer->subbuf_size - (sizeof(u32) * 2);
1703 nr_pages = DIV_ROUND_UP(size, buffer->subbuf_size);
1704 buffer->flags = flags;
1705 buffer->clock = trace_clock_local;
1706 buffer->reader_lock_key = key;
1708 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1709 init_waitqueue_head(&buffer->irq_work.waiters);
1711 /* need at least two pages */
1715 buffer->cpus = nr_cpu_ids;
1717 bsize = sizeof(void *) * nr_cpu_ids;
1718 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1720 if (!buffer->buffers)
1721 goto fail_free_cpumask;
1723 cpu = raw_smp_processor_id();
1724 cpumask_set_cpu(cpu, buffer->cpumask);
1725 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1726 if (!buffer->buffers[cpu])
1727 goto fail_free_buffers;
1729 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1731 goto fail_free_buffers;
1733 mutex_init(&buffer->mutex);
1738 for_each_buffer_cpu(buffer, cpu) {
1739 if (buffer->buffers[cpu])
1740 rb_free_cpu_buffer(buffer->buffers[cpu]);
1742 kfree(buffer->buffers);
1745 free_cpumask_var(buffer->cpumask);
1751 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1754 * ring_buffer_free - free a ring buffer.
1755 * @buffer: the buffer to free.
1758 ring_buffer_free(struct trace_buffer *buffer)
1762 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1764 irq_work_sync(&buffer->irq_work.work);
1766 for_each_buffer_cpu(buffer, cpu)
1767 rb_free_cpu_buffer(buffer->buffers[cpu]);
1769 kfree(buffer->buffers);
1770 free_cpumask_var(buffer->cpumask);
1774 EXPORT_SYMBOL_GPL(ring_buffer_free);
1776 void ring_buffer_set_clock(struct trace_buffer *buffer,
1779 buffer->clock = clock;
1782 void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1784 buffer->time_stamp_abs = abs;
1787 bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1789 return buffer->time_stamp_abs;
1792 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1794 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1796 return local_read(&bpage->entries) & RB_WRITE_MASK;
1799 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1801 return local_read(&bpage->write) & RB_WRITE_MASK;
1805 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1807 struct list_head *tail_page, *to_remove, *next_page;
1808 struct buffer_page *to_remove_page, *tmp_iter_page;
1809 struct buffer_page *last_page, *first_page;
1810 unsigned long nr_removed;
1811 unsigned long head_bit;
1816 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1817 atomic_inc(&cpu_buffer->record_disabled);
1819 * We don't race with the readers since we have acquired the reader
1820 * lock. We also don't race with writers after disabling recording.
1821 * This makes it easy to figure out the first and the last page to be
1822 * removed from the list. We unlink all the pages in between including
1823 * the first and last pages. This is done in a busy loop so that we
1824 * lose the least number of traces.
1825 * The pages are freed after we restart recording and unlock readers.
1827 tail_page = &cpu_buffer->tail_page->list;
1830 * tail page might be on reader page, we remove the next page
1831 * from the ring buffer
1833 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1834 tail_page = rb_list_head(tail_page->next);
1835 to_remove = tail_page;
1837 /* start of pages to remove */
1838 first_page = list_entry(rb_list_head(to_remove->next),
1839 struct buffer_page, list);
1841 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1842 to_remove = rb_list_head(to_remove)->next;
1843 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1845 /* Read iterators need to reset themselves when some pages removed */
1846 cpu_buffer->pages_removed += nr_removed;
1848 next_page = rb_list_head(to_remove)->next;
1851 * Now we remove all pages between tail_page and next_page.
1852 * Make sure that we have head_bit value preserved for the
1855 tail_page->next = (struct list_head *)((unsigned long)next_page |
1857 next_page = rb_list_head(next_page);
1858 next_page->prev = tail_page;
1860 /* make sure pages points to a valid page in the ring buffer */
1861 cpu_buffer->pages = next_page;
1863 /* update head page */
1865 cpu_buffer->head_page = list_entry(next_page,
1866 struct buffer_page, list);
1868 /* pages are removed, resume tracing and then free the pages */
1869 atomic_dec(&cpu_buffer->record_disabled);
1870 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1872 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1874 /* last buffer page to remove */
1875 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1877 tmp_iter_page = first_page;
1882 to_remove_page = tmp_iter_page;
1883 rb_inc_page(&tmp_iter_page);
1885 /* update the counters */
1886 page_entries = rb_page_entries(to_remove_page);
1889 * If something was added to this page, it was full
1890 * since it is not the tail page. So we deduct the
1891 * bytes consumed in ring buffer from here.
1892 * Increment overrun to account for the lost events.
1894 local_add(page_entries, &cpu_buffer->overrun);
1895 local_sub(rb_page_commit(to_remove_page), &cpu_buffer->entries_bytes);
1896 local_inc(&cpu_buffer->pages_lost);
1900 * We have already removed references to this list item, just
1901 * free up the buffer_page and its page
1903 free_buffer_page(to_remove_page);
1906 } while (to_remove_page != last_page);
1908 RB_WARN_ON(cpu_buffer, nr_removed);
1910 return nr_removed == 0;
1914 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1916 struct list_head *pages = &cpu_buffer->new_pages;
1917 unsigned long flags;
1921 /* Can be called at early boot up, where interrupts must not been enabled */
1922 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1924 * We are holding the reader lock, so the reader page won't be swapped
1925 * in the ring buffer. Now we are racing with the writer trying to
1926 * move head page and the tail page.
1927 * We are going to adapt the reader page update process where:
1928 * 1. We first splice the start and end of list of new pages between
1929 * the head page and its previous page.
1930 * 2. We cmpxchg the prev_page->next to point from head page to the
1931 * start of new pages list.
1932 * 3. Finally, we update the head->prev to the end of new list.
1934 * We will try this process 10 times, to make sure that we don't keep
1940 struct list_head *head_page, *prev_page;
1941 struct list_head *last_page, *first_page;
1942 struct list_head *head_page_with_bit;
1943 struct buffer_page *hpage = rb_set_head_page(cpu_buffer);
1947 head_page = &hpage->list;
1948 prev_page = head_page->prev;
1950 first_page = pages->next;
1951 last_page = pages->prev;
1953 head_page_with_bit = (struct list_head *)
1954 ((unsigned long)head_page | RB_PAGE_HEAD);
1956 last_page->next = head_page_with_bit;
1957 first_page->prev = prev_page;
1959 /* caution: head_page_with_bit gets updated on cmpxchg failure */
1960 if (try_cmpxchg(&prev_page->next,
1961 &head_page_with_bit, first_page)) {
1963 * yay, we replaced the page pointer to our new list,
1964 * now, we just have to update to head page's prev
1965 * pointer to point to end of list
1967 head_page->prev = last_page;
1974 INIT_LIST_HEAD(pages);
1976 * If we weren't successful in adding in new pages, warn and stop
1979 RB_WARN_ON(cpu_buffer, !success);
1980 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1982 /* free pages if they weren't inserted */
1984 struct buffer_page *bpage, *tmp;
1985 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1987 list_del_init(&bpage->list);
1988 free_buffer_page(bpage);
1994 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1998 if (cpu_buffer->nr_pages_to_update > 0)
1999 success = rb_insert_pages(cpu_buffer);
2001 success = rb_remove_pages(cpu_buffer,
2002 -cpu_buffer->nr_pages_to_update);
2005 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
2008 static void update_pages_handler(struct work_struct *work)
2010 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
2011 struct ring_buffer_per_cpu, update_pages_work);
2012 rb_update_pages(cpu_buffer);
2013 complete(&cpu_buffer->update_done);
2017 * ring_buffer_resize - resize the ring buffer
2018 * @buffer: the buffer to resize.
2019 * @size: the new size.
2020 * @cpu_id: the cpu buffer to resize
2022 * Minimum size is 2 * buffer->subbuf_size.
2024 * Returns 0 on success and < 0 on failure.
2026 int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
2029 struct ring_buffer_per_cpu *cpu_buffer;
2030 unsigned long nr_pages;
2034 * Always succeed at resizing a non-existent buffer:
2039 /* Make sure the requested buffer exists */
2040 if (cpu_id != RING_BUFFER_ALL_CPUS &&
2041 !cpumask_test_cpu(cpu_id, buffer->cpumask))
2044 nr_pages = DIV_ROUND_UP(size, buffer->subbuf_size);
2046 /* we need a minimum of two pages */
2050 /* prevent another thread from changing buffer sizes */
2051 mutex_lock(&buffer->mutex);
2052 atomic_inc(&buffer->resizing);
2054 if (cpu_id == RING_BUFFER_ALL_CPUS) {
2056 * Don't succeed if resizing is disabled, as a reader might be
2057 * manipulating the ring buffer and is expecting a sane state while
2060 for_each_buffer_cpu(buffer, cpu) {
2061 cpu_buffer = buffer->buffers[cpu];
2062 if (atomic_read(&cpu_buffer->resize_disabled)) {
2064 goto out_err_unlock;
2068 /* calculate the pages to update */
2069 for_each_buffer_cpu(buffer, cpu) {
2070 cpu_buffer = buffer->buffers[cpu];
2072 cpu_buffer->nr_pages_to_update = nr_pages -
2073 cpu_buffer->nr_pages;
2075 * nothing more to do for removing pages or no update
2077 if (cpu_buffer->nr_pages_to_update <= 0)
2080 * to add pages, make sure all new pages can be
2081 * allocated without receiving ENOMEM
2083 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2084 if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2085 &cpu_buffer->new_pages)) {
2086 /* not enough memory for new pages */
2096 * Fire off all the required work handlers
2097 * We can't schedule on offline CPUs, but it's not necessary
2098 * since we can change their buffer sizes without any race.
2100 for_each_buffer_cpu(buffer, cpu) {
2101 cpu_buffer = buffer->buffers[cpu];
2102 if (!cpu_buffer->nr_pages_to_update)
2105 /* Can't run something on an offline CPU. */
2106 if (!cpu_online(cpu)) {
2107 rb_update_pages(cpu_buffer);
2108 cpu_buffer->nr_pages_to_update = 0;
2110 /* Run directly if possible. */
2112 if (cpu != smp_processor_id()) {
2114 schedule_work_on(cpu,
2115 &cpu_buffer->update_pages_work);
2117 update_pages_handler(&cpu_buffer->update_pages_work);
2123 /* wait for all the updates to complete */
2124 for_each_buffer_cpu(buffer, cpu) {
2125 cpu_buffer = buffer->buffers[cpu];
2126 if (!cpu_buffer->nr_pages_to_update)
2129 if (cpu_online(cpu))
2130 wait_for_completion(&cpu_buffer->update_done);
2131 cpu_buffer->nr_pages_to_update = 0;
2136 cpu_buffer = buffer->buffers[cpu_id];
2138 if (nr_pages == cpu_buffer->nr_pages)
2142 * Don't succeed if resizing is disabled, as a reader might be
2143 * manipulating the ring buffer and is expecting a sane state while
2146 if (atomic_read(&cpu_buffer->resize_disabled)) {
2148 goto out_err_unlock;
2151 cpu_buffer->nr_pages_to_update = nr_pages -
2152 cpu_buffer->nr_pages;
2154 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2155 if (cpu_buffer->nr_pages_to_update > 0 &&
2156 __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2157 &cpu_buffer->new_pages)) {
2164 /* Can't run something on an offline CPU. */
2165 if (!cpu_online(cpu_id))
2166 rb_update_pages(cpu_buffer);
2168 /* Run directly if possible. */
2170 if (cpu_id == smp_processor_id()) {
2171 rb_update_pages(cpu_buffer);
2175 schedule_work_on(cpu_id,
2176 &cpu_buffer->update_pages_work);
2177 wait_for_completion(&cpu_buffer->update_done);
2181 cpu_buffer->nr_pages_to_update = 0;
2187 * The ring buffer resize can happen with the ring buffer
2188 * enabled, so that the update disturbs the tracing as little
2189 * as possible. But if the buffer is disabled, we do not need
2190 * to worry about that, and we can take the time to verify
2191 * that the buffer is not corrupt.
2193 if (atomic_read(&buffer->record_disabled)) {
2194 atomic_inc(&buffer->record_disabled);
2196 * Even though the buffer was disabled, we must make sure
2197 * that it is truly disabled before calling rb_check_pages.
2198 * There could have been a race between checking
2199 * record_disable and incrementing it.
2202 for_each_buffer_cpu(buffer, cpu) {
2203 cpu_buffer = buffer->buffers[cpu];
2204 rb_check_pages(cpu_buffer);
2206 atomic_dec(&buffer->record_disabled);
2209 atomic_dec(&buffer->resizing);
2210 mutex_unlock(&buffer->mutex);
2214 for_each_buffer_cpu(buffer, cpu) {
2215 struct buffer_page *bpage, *tmp;
2217 cpu_buffer = buffer->buffers[cpu];
2218 cpu_buffer->nr_pages_to_update = 0;
2220 if (list_empty(&cpu_buffer->new_pages))
2223 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2225 list_del_init(&bpage->list);
2226 free_buffer_page(bpage);
2230 atomic_dec(&buffer->resizing);
2231 mutex_unlock(&buffer->mutex);
2234 EXPORT_SYMBOL_GPL(ring_buffer_resize);
2236 void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2238 mutex_lock(&buffer->mutex);
2240 buffer->flags |= RB_FL_OVERWRITE;
2242 buffer->flags &= ~RB_FL_OVERWRITE;
2243 mutex_unlock(&buffer->mutex);
2245 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2247 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2249 return bpage->page->data + index;
2252 static __always_inline struct ring_buffer_event *
2253 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2255 return __rb_page_index(cpu_buffer->reader_page,
2256 cpu_buffer->reader_page->read);
2259 static struct ring_buffer_event *
2260 rb_iter_head_event(struct ring_buffer_iter *iter)
2262 struct ring_buffer_event *event;
2263 struct buffer_page *iter_head_page = iter->head_page;
2264 unsigned long commit;
2267 if (iter->head != iter->next_event)
2271 * When the writer goes across pages, it issues a cmpxchg which
2272 * is a mb(), which will synchronize with the rmb here.
2273 * (see rb_tail_page_update() and __rb_reserve_next())
2275 commit = rb_page_commit(iter_head_page);
2278 /* An event needs to be at least 8 bytes in size */
2279 if (iter->head > commit - 8)
2282 event = __rb_page_index(iter_head_page, iter->head);
2283 length = rb_event_length(event);
2286 * READ_ONCE() doesn't work on functions and we don't want the
2287 * compiler doing any crazy optimizations with length.
2291 if ((iter->head + length) > commit || length > iter->event_size)
2292 /* Writer corrupted the read? */
2295 memcpy(iter->event, event, length);
2297 * If the page stamp is still the same after this rmb() then the
2298 * event was safely copied without the writer entering the page.
2302 /* Make sure the page didn't change since we read this */
2303 if (iter->page_stamp != iter_head_page->page->time_stamp ||
2304 commit > rb_page_commit(iter_head_page))
2307 iter->next_event = iter->head + length;
2310 /* Reset to the beginning */
2311 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2313 iter->next_event = 0;
2314 iter->missed_events = 1;
2318 /* Size is determined by what has been committed */
2319 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2321 return rb_page_commit(bpage);
2324 static __always_inline unsigned
2325 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2327 return rb_page_commit(cpu_buffer->commit_page);
2330 static __always_inline unsigned
2331 rb_event_index(struct ring_buffer_per_cpu *cpu_buffer, struct ring_buffer_event *event)
2333 unsigned long addr = (unsigned long)event;
2335 addr &= (PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1;
2337 return addr - BUF_PAGE_HDR_SIZE;
2340 static void rb_inc_iter(struct ring_buffer_iter *iter)
2342 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2345 * The iterator could be on the reader page (it starts there).
2346 * But the head could have moved, since the reader was
2347 * found. Check for this case and assign the iterator
2348 * to the head page instead of next.
2350 if (iter->head_page == cpu_buffer->reader_page)
2351 iter->head_page = rb_set_head_page(cpu_buffer);
2353 rb_inc_page(&iter->head_page);
2355 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2357 iter->next_event = 0;
2361 * rb_handle_head_page - writer hit the head page
2363 * Returns: +1 to retry page
2368 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2369 struct buffer_page *tail_page,
2370 struct buffer_page *next_page)
2372 struct buffer_page *new_head;
2377 entries = rb_page_entries(next_page);
2380 * The hard part is here. We need to move the head
2381 * forward, and protect against both readers on
2382 * other CPUs and writers coming in via interrupts.
2384 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2388 * type can be one of four:
2389 * NORMAL - an interrupt already moved it for us
2390 * HEAD - we are the first to get here.
2391 * UPDATE - we are the interrupt interrupting
2393 * MOVED - a reader on another CPU moved the next
2394 * pointer to its reader page. Give up
2401 * We changed the head to UPDATE, thus
2402 * it is our responsibility to update
2405 local_add(entries, &cpu_buffer->overrun);
2406 local_sub(rb_page_commit(next_page), &cpu_buffer->entries_bytes);
2407 local_inc(&cpu_buffer->pages_lost);
2410 * The entries will be zeroed out when we move the
2414 /* still more to do */
2417 case RB_PAGE_UPDATE:
2419 * This is an interrupt that interrupt the
2420 * previous update. Still more to do.
2423 case RB_PAGE_NORMAL:
2425 * An interrupt came in before the update
2426 * and processed this for us.
2427 * Nothing left to do.
2432 * The reader is on another CPU and just did
2433 * a swap with our next_page.
2438 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2443 * Now that we are here, the old head pointer is
2444 * set to UPDATE. This will keep the reader from
2445 * swapping the head page with the reader page.
2446 * The reader (on another CPU) will spin till
2449 * We just need to protect against interrupts
2450 * doing the job. We will set the next pointer
2451 * to HEAD. After that, we set the old pointer
2452 * to NORMAL, but only if it was HEAD before.
2453 * otherwise we are an interrupt, and only
2454 * want the outer most commit to reset it.
2456 new_head = next_page;
2457 rb_inc_page(&new_head);
2459 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2463 * Valid returns are:
2464 * HEAD - an interrupt came in and already set it.
2465 * NORMAL - One of two things:
2466 * 1) We really set it.
2467 * 2) A bunch of interrupts came in and moved
2468 * the page forward again.
2472 case RB_PAGE_NORMAL:
2476 RB_WARN_ON(cpu_buffer, 1);
2481 * It is possible that an interrupt came in,
2482 * set the head up, then more interrupts came in
2483 * and moved it again. When we get back here,
2484 * the page would have been set to NORMAL but we
2485 * just set it back to HEAD.
2487 * How do you detect this? Well, if that happened
2488 * the tail page would have moved.
2490 if (ret == RB_PAGE_NORMAL) {
2491 struct buffer_page *buffer_tail_page;
2493 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2495 * If the tail had moved passed next, then we need
2496 * to reset the pointer.
2498 if (buffer_tail_page != tail_page &&
2499 buffer_tail_page != next_page)
2500 rb_head_page_set_normal(cpu_buffer, new_head,
2506 * If this was the outer most commit (the one that
2507 * changed the original pointer from HEAD to UPDATE),
2508 * then it is up to us to reset it to NORMAL.
2510 if (type == RB_PAGE_HEAD) {
2511 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2514 if (RB_WARN_ON(cpu_buffer,
2515 ret != RB_PAGE_UPDATE))
2523 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2524 unsigned long tail, struct rb_event_info *info)
2526 unsigned long bsize = READ_ONCE(cpu_buffer->buffer->subbuf_size);
2527 struct buffer_page *tail_page = info->tail_page;
2528 struct ring_buffer_event *event;
2529 unsigned long length = info->length;
2532 * Only the event that crossed the page boundary
2533 * must fill the old tail_page with padding.
2535 if (tail >= bsize) {
2537 * If the page was filled, then we still need
2538 * to update the real_end. Reset it to zero
2539 * and the reader will ignore it.
2542 tail_page->real_end = 0;
2544 local_sub(length, &tail_page->write);
2548 event = __rb_page_index(tail_page, tail);
2551 * Save the original length to the meta data.
2552 * This will be used by the reader to add lost event
2555 tail_page->real_end = tail;
2558 * If this event is bigger than the minimum size, then
2559 * we need to be careful that we don't subtract the
2560 * write counter enough to allow another writer to slip
2562 * We put in a discarded commit instead, to make sure
2563 * that this space is not used again, and this space will
2564 * not be accounted into 'entries_bytes'.
2566 * If we are less than the minimum size, we don't need to
2569 if (tail > (bsize - RB_EVNT_MIN_SIZE)) {
2570 /* No room for any events */
2572 /* Mark the rest of the page with padding */
2573 rb_event_set_padding(event);
2575 /* Make sure the padding is visible before the write update */
2578 /* Set the write back to the previous setting */
2579 local_sub(length, &tail_page->write);
2583 /* Put in a discarded event */
2584 event->array[0] = (bsize - tail) - RB_EVNT_HDR_SIZE;
2585 event->type_len = RINGBUF_TYPE_PADDING;
2586 /* time delta must be non zero */
2587 event->time_delta = 1;
2589 /* account for padding bytes */
2590 local_add(bsize - tail, &cpu_buffer->entries_bytes);
2592 /* Make sure the padding is visible before the tail_page->write update */
2595 /* Set write to end of buffer */
2596 length = (tail + length) - bsize;
2597 local_sub(length, &tail_page->write);
2600 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2603 * This is the slow path, force gcc not to inline it.
2605 static noinline struct ring_buffer_event *
2606 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2607 unsigned long tail, struct rb_event_info *info)
2609 struct buffer_page *tail_page = info->tail_page;
2610 struct buffer_page *commit_page = cpu_buffer->commit_page;
2611 struct trace_buffer *buffer = cpu_buffer->buffer;
2612 struct buffer_page *next_page;
2615 next_page = tail_page;
2617 rb_inc_page(&next_page);
2620 * If for some reason, we had an interrupt storm that made
2621 * it all the way around the buffer, bail, and warn
2624 if (unlikely(next_page == commit_page)) {
2625 local_inc(&cpu_buffer->commit_overrun);
2630 * This is where the fun begins!
2632 * We are fighting against races between a reader that
2633 * could be on another CPU trying to swap its reader
2634 * page with the buffer head.
2636 * We are also fighting against interrupts coming in and
2637 * moving the head or tail on us as well.
2639 * If the next page is the head page then we have filled
2640 * the buffer, unless the commit page is still on the
2643 if (rb_is_head_page(next_page, &tail_page->list)) {
2646 * If the commit is not on the reader page, then
2647 * move the header page.
2649 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2651 * If we are not in overwrite mode,
2652 * this is easy, just stop here.
2654 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2655 local_inc(&cpu_buffer->dropped_events);
2659 ret = rb_handle_head_page(cpu_buffer,
2668 * We need to be careful here too. The
2669 * commit page could still be on the reader
2670 * page. We could have a small buffer, and
2671 * have filled up the buffer with events
2672 * from interrupts and such, and wrapped.
2674 * Note, if the tail page is also on the
2675 * reader_page, we let it move out.
2677 if (unlikely((cpu_buffer->commit_page !=
2678 cpu_buffer->tail_page) &&
2679 (cpu_buffer->commit_page ==
2680 cpu_buffer->reader_page))) {
2681 local_inc(&cpu_buffer->commit_overrun);
2687 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2691 rb_reset_tail(cpu_buffer, tail, info);
2693 /* Commit what we have for now. */
2694 rb_end_commit(cpu_buffer);
2695 /* rb_end_commit() decs committing */
2696 local_inc(&cpu_buffer->committing);
2698 /* fail and let the caller try again */
2699 return ERR_PTR(-EAGAIN);
2703 rb_reset_tail(cpu_buffer, tail, info);
2709 static struct ring_buffer_event *
2710 rb_add_time_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2711 struct ring_buffer_event *event, u64 delta, bool abs)
2714 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2716 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2718 /* Not the first event on the page, or not delta? */
2719 if (abs || rb_event_index(cpu_buffer, event)) {
2720 event->time_delta = delta & TS_MASK;
2721 event->array[0] = delta >> TS_SHIFT;
2723 /* nope, just zero it */
2724 event->time_delta = 0;
2725 event->array[0] = 0;
2728 return skip_time_extend(event);
2731 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2732 static inline bool sched_clock_stable(void)
2739 rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2740 struct rb_event_info *info)
2744 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2745 (unsigned long long)info->delta,
2746 (unsigned long long)info->ts,
2747 (unsigned long long)info->before,
2748 (unsigned long long)info->after,
2749 (unsigned long long)({rb_time_read(&cpu_buffer->write_stamp, &write_stamp); write_stamp;}),
2750 sched_clock_stable() ? "" :
2751 "If you just came from a suspend/resume,\n"
2752 "please switch to the trace global clock:\n"
2753 " echo global > /sys/kernel/tracing/trace_clock\n"
2754 "or add trace_clock=global to the kernel command line\n");
2757 static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2758 struct ring_buffer_event **event,
2759 struct rb_event_info *info,
2761 unsigned int *length)
2763 bool abs = info->add_timestamp &
2764 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2766 if (unlikely(info->delta > (1ULL << 59))) {
2768 * Some timers can use more than 59 bits, and when a timestamp
2769 * is added to the buffer, it will lose those bits.
2771 if (abs && (info->ts & TS_MSB)) {
2772 info->delta &= ABS_TS_MASK;
2774 /* did the clock go backwards */
2775 } else if (info->before == info->after && info->before > info->ts) {
2776 /* not interrupted */
2780 * This is possible with a recalibrating of the TSC.
2781 * Do not produce a call stack, but just report it.
2785 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2786 info->before, info->ts);
2789 rb_check_timestamp(cpu_buffer, info);
2793 *event = rb_add_time_stamp(cpu_buffer, *event, info->delta, abs);
2794 *length -= RB_LEN_TIME_EXTEND;
2799 * rb_update_event - update event type and data
2800 * @cpu_buffer: The per cpu buffer of the @event
2801 * @event: the event to update
2802 * @info: The info to update the @event with (contains length and delta)
2804 * Update the type and data fields of the @event. The length
2805 * is the actual size that is written to the ring buffer,
2806 * and with this, we can determine what to place into the
2810 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2811 struct ring_buffer_event *event,
2812 struct rb_event_info *info)
2814 unsigned length = info->length;
2815 u64 delta = info->delta;
2816 unsigned int nest = local_read(&cpu_buffer->committing) - 1;
2818 if (!WARN_ON_ONCE(nest >= MAX_NEST))
2819 cpu_buffer->event_stamp[nest] = info->ts;
2822 * If we need to add a timestamp, then we
2823 * add it to the start of the reserved space.
2825 if (unlikely(info->add_timestamp))
2826 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2828 event->time_delta = delta;
2829 length -= RB_EVNT_HDR_SIZE;
2830 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2831 event->type_len = 0;
2832 event->array[0] = length;
2834 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2837 static unsigned rb_calculate_event_length(unsigned length)
2839 struct ring_buffer_event event; /* Used only for sizeof array */
2841 /* zero length can cause confusions */
2845 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2846 length += sizeof(event.array[0]);
2848 length += RB_EVNT_HDR_SIZE;
2849 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2852 * In case the time delta is larger than the 27 bits for it
2853 * in the header, we need to add a timestamp. If another
2854 * event comes in when trying to discard this one to increase
2855 * the length, then the timestamp will be added in the allocated
2856 * space of this event. If length is bigger than the size needed
2857 * for the TIME_EXTEND, then padding has to be used. The events
2858 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2859 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2860 * As length is a multiple of 4, we only need to worry if it
2861 * is 12 (RB_LEN_TIME_EXTEND + 4).
2863 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2864 length += RB_ALIGNMENT;
2870 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2871 struct ring_buffer_event *event)
2873 unsigned long new_index, old_index;
2874 struct buffer_page *bpage;
2877 new_index = rb_event_index(cpu_buffer, event);
2878 old_index = new_index + rb_event_ts_length(event);
2879 addr = (unsigned long)event;
2880 addr &= ~((PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1);
2882 bpage = READ_ONCE(cpu_buffer->tail_page);
2885 * Make sure the tail_page is still the same and
2886 * the next write location is the end of this event
2888 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2889 unsigned long write_mask =
2890 local_read(&bpage->write) & ~RB_WRITE_MASK;
2891 unsigned long event_length = rb_event_length(event);
2894 * For the before_stamp to be different than the write_stamp
2895 * to make sure that the next event adds an absolute
2896 * value and does not rely on the saved write stamp, which
2897 * is now going to be bogus.
2899 * By setting the before_stamp to zero, the next event
2900 * is not going to use the write_stamp and will instead
2901 * create an absolute timestamp. This means there's no
2902 * reason to update the wirte_stamp!
2904 rb_time_set(&cpu_buffer->before_stamp, 0);
2907 * If an event were to come in now, it would see that the
2908 * write_stamp and the before_stamp are different, and assume
2909 * that this event just added itself before updating
2910 * the write stamp. The interrupting event will fix the
2911 * write stamp for us, and use an absolute timestamp.
2915 * This is on the tail page. It is possible that
2916 * a write could come in and move the tail page
2917 * and write to the next page. That is fine
2918 * because we just shorten what is on this page.
2920 old_index += write_mask;
2921 new_index += write_mask;
2923 /* caution: old_index gets updated on cmpxchg failure */
2924 if (local_try_cmpxchg(&bpage->write, &old_index, new_index)) {
2925 /* update counters */
2926 local_sub(event_length, &cpu_buffer->entries_bytes);
2931 /* could not discard */
2935 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2937 local_inc(&cpu_buffer->committing);
2938 local_inc(&cpu_buffer->commits);
2941 static __always_inline void
2942 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2944 unsigned long max_count;
2947 * We only race with interrupts and NMIs on this CPU.
2948 * If we own the commit event, then we can commit
2949 * all others that interrupted us, since the interruptions
2950 * are in stack format (they finish before they come
2951 * back to us). This allows us to do a simple loop to
2952 * assign the commit to the tail.
2955 max_count = cpu_buffer->nr_pages * 100;
2957 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2958 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2960 if (RB_WARN_ON(cpu_buffer,
2961 rb_is_reader_page(cpu_buffer->tail_page)))
2964 * No need for a memory barrier here, as the update
2965 * of the tail_page did it for this page.
2967 local_set(&cpu_buffer->commit_page->page->commit,
2968 rb_page_write(cpu_buffer->commit_page));
2969 rb_inc_page(&cpu_buffer->commit_page);
2970 /* add barrier to keep gcc from optimizing too much */
2973 while (rb_commit_index(cpu_buffer) !=
2974 rb_page_write(cpu_buffer->commit_page)) {
2976 /* Make sure the readers see the content of what is committed. */
2978 local_set(&cpu_buffer->commit_page->page->commit,
2979 rb_page_write(cpu_buffer->commit_page));
2980 RB_WARN_ON(cpu_buffer,
2981 local_read(&cpu_buffer->commit_page->page->commit) &
2986 /* again, keep gcc from optimizing */
2990 * If an interrupt came in just after the first while loop
2991 * and pushed the tail page forward, we will be left with
2992 * a dangling commit that will never go forward.
2994 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2998 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
3000 unsigned long commits;
3002 if (RB_WARN_ON(cpu_buffer,
3003 !local_read(&cpu_buffer->committing)))
3007 commits = local_read(&cpu_buffer->commits);
3008 /* synchronize with interrupts */
3010 if (local_read(&cpu_buffer->committing) == 1)
3011 rb_set_commit_to_write(cpu_buffer);
3013 local_dec(&cpu_buffer->committing);
3015 /* synchronize with interrupts */
3019 * Need to account for interrupts coming in between the
3020 * updating of the commit page and the clearing of the
3021 * committing counter.
3023 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
3024 !local_read(&cpu_buffer->committing)) {
3025 local_inc(&cpu_buffer->committing);
3030 static inline void rb_event_discard(struct ring_buffer_event *event)
3032 if (extended_time(event))
3033 event = skip_time_extend(event);
3035 /* array[0] holds the actual length for the discarded event */
3036 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
3037 event->type_len = RINGBUF_TYPE_PADDING;
3038 /* time delta must be non zero */
3039 if (!event->time_delta)
3040 event->time_delta = 1;
3043 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer)
3045 local_inc(&cpu_buffer->entries);
3046 rb_end_commit(cpu_buffer);
3049 static __always_inline void
3050 rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3052 if (buffer->irq_work.waiters_pending) {
3053 buffer->irq_work.waiters_pending = false;
3054 /* irq_work_queue() supplies it's own memory barriers */
3055 irq_work_queue(&buffer->irq_work.work);
3058 if (cpu_buffer->irq_work.waiters_pending) {
3059 cpu_buffer->irq_work.waiters_pending = false;
3060 /* irq_work_queue() supplies it's own memory barriers */
3061 irq_work_queue(&cpu_buffer->irq_work.work);
3064 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3067 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3070 if (!cpu_buffer->irq_work.full_waiters_pending)
3073 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3075 if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
3078 cpu_buffer->irq_work.wakeup_full = true;
3079 cpu_buffer->irq_work.full_waiters_pending = false;
3080 /* irq_work_queue() supplies it's own memory barriers */
3081 irq_work_queue(&cpu_buffer->irq_work.work);
3084 #ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3085 # define do_ring_buffer_record_recursion() \
3086 do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3088 # define do_ring_buffer_record_recursion() do { } while (0)
3092 * The lock and unlock are done within a preempt disable section.
3093 * The current_context per_cpu variable can only be modified
3094 * by the current task between lock and unlock. But it can
3095 * be modified more than once via an interrupt. To pass this
3096 * information from the lock to the unlock without having to
3097 * access the 'in_interrupt()' functions again (which do show
3098 * a bit of overhead in something as critical as function tracing,
3099 * we use a bitmask trick.
3101 * bit 1 = NMI context
3102 * bit 2 = IRQ context
3103 * bit 3 = SoftIRQ context
3104 * bit 4 = normal context.
3106 * This works because this is the order of contexts that can
3107 * preempt other contexts. A SoftIRQ never preempts an IRQ
3110 * When the context is determined, the corresponding bit is
3111 * checked and set (if it was set, then a recursion of that context
3114 * On unlock, we need to clear this bit. To do so, just subtract
3115 * 1 from the current_context and AND it to itself.
3119 * 101 & 100 = 100 (clearing bit zero)
3122 * 1010 & 1001 = 1000 (clearing bit 1)
3124 * The least significant bit can be cleared this way, and it
3125 * just so happens that it is the same bit corresponding to
3126 * the current context.
3128 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3129 * is set when a recursion is detected at the current context, and if
3130 * the TRANSITION bit is already set, it will fail the recursion.
3131 * This is needed because there's a lag between the changing of
3132 * interrupt context and updating the preempt count. In this case,
3133 * a false positive will be found. To handle this, one extra recursion
3134 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3135 * bit is already set, then it is considered a recursion and the function
3136 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3138 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3139 * to be cleared. Even if it wasn't the context that set it. That is,
3140 * if an interrupt comes in while NORMAL bit is set and the ring buffer
3141 * is called before preempt_count() is updated, since the check will
3142 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3143 * NMI then comes in, it will set the NMI bit, but when the NMI code
3144 * does the trace_recursive_unlock() it will clear the TRANSITION bit
3145 * and leave the NMI bit set. But this is fine, because the interrupt
3146 * code that set the TRANSITION bit will then clear the NMI bit when it
3147 * calls trace_recursive_unlock(). If another NMI comes in, it will
3148 * set the TRANSITION bit and continue.
3150 * Note: The TRANSITION bit only handles a single transition between context.
3153 static __always_inline bool
3154 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3156 unsigned int val = cpu_buffer->current_context;
3157 int bit = interrupt_context_level();
3159 bit = RB_CTX_NORMAL - bit;
3161 if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3163 * It is possible that this was called by transitioning
3164 * between interrupt context, and preempt_count() has not
3165 * been updated yet. In this case, use the TRANSITION bit.
3167 bit = RB_CTX_TRANSITION;
3168 if (val & (1 << (bit + cpu_buffer->nest))) {
3169 do_ring_buffer_record_recursion();
3174 val |= (1 << (bit + cpu_buffer->nest));
3175 cpu_buffer->current_context = val;
3180 static __always_inline void
3181 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3183 cpu_buffer->current_context &=
3184 cpu_buffer->current_context - (1 << cpu_buffer->nest);
3187 /* The recursive locking above uses 5 bits */
3188 #define NESTED_BITS 5
3191 * ring_buffer_nest_start - Allow to trace while nested
3192 * @buffer: The ring buffer to modify
3194 * The ring buffer has a safety mechanism to prevent recursion.
3195 * But there may be a case where a trace needs to be done while
3196 * tracing something else. In this case, calling this function
3197 * will allow this function to nest within a currently active
3198 * ring_buffer_lock_reserve().
3200 * Call this function before calling another ring_buffer_lock_reserve() and
3201 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3203 void ring_buffer_nest_start(struct trace_buffer *buffer)
3205 struct ring_buffer_per_cpu *cpu_buffer;
3208 /* Enabled by ring_buffer_nest_end() */
3209 preempt_disable_notrace();
3210 cpu = raw_smp_processor_id();
3211 cpu_buffer = buffer->buffers[cpu];
3212 /* This is the shift value for the above recursive locking */
3213 cpu_buffer->nest += NESTED_BITS;
3217 * ring_buffer_nest_end - Allow to trace while nested
3218 * @buffer: The ring buffer to modify
3220 * Must be called after ring_buffer_nest_start() and after the
3221 * ring_buffer_unlock_commit().
3223 void ring_buffer_nest_end(struct trace_buffer *buffer)
3225 struct ring_buffer_per_cpu *cpu_buffer;
3228 /* disabled by ring_buffer_nest_start() */
3229 cpu = raw_smp_processor_id();
3230 cpu_buffer = buffer->buffers[cpu];
3231 /* This is the shift value for the above recursive locking */
3232 cpu_buffer->nest -= NESTED_BITS;
3233 preempt_enable_notrace();
3237 * ring_buffer_unlock_commit - commit a reserved
3238 * @buffer: The buffer to commit to
3240 * This commits the data to the ring buffer, and releases any locks held.
3242 * Must be paired with ring_buffer_lock_reserve.
3244 int ring_buffer_unlock_commit(struct trace_buffer *buffer)
3246 struct ring_buffer_per_cpu *cpu_buffer;
3247 int cpu = raw_smp_processor_id();
3249 cpu_buffer = buffer->buffers[cpu];
3251 rb_commit(cpu_buffer);
3253 rb_wakeups(buffer, cpu_buffer);
3255 trace_recursive_unlock(cpu_buffer);
3257 preempt_enable_notrace();
3261 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3263 /* Special value to validate all deltas on a page. */
3264 #define CHECK_FULL_PAGE 1L
3266 #ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
3268 static const char *show_irq_str(int bits)
3270 const char *type[] = {
3284 /* Assume this is an trace event */
3285 static const char *show_flags(struct ring_buffer_event *event)
3287 struct trace_entry *entry;
3290 if (rb_event_data_length(event) - RB_EVNT_HDR_SIZE < sizeof(*entry))
3293 entry = ring_buffer_event_data(event);
3295 if (entry->flags & TRACE_FLAG_SOFTIRQ)
3298 if (entry->flags & TRACE_FLAG_HARDIRQ)
3301 if (entry->flags & TRACE_FLAG_NMI)
3304 return show_irq_str(bits);
3307 static const char *show_irq(struct ring_buffer_event *event)
3309 struct trace_entry *entry;
3311 if (rb_event_data_length(event) - RB_EVNT_HDR_SIZE < sizeof(*entry))
3314 entry = ring_buffer_event_data(event);
3315 if (entry->flags & TRACE_FLAG_IRQS_OFF)
3320 static const char *show_interrupt_level(void)
3322 unsigned long pc = preempt_count();
3323 unsigned char level = 0;
3325 if (pc & SOFTIRQ_OFFSET)
3328 if (pc & HARDIRQ_MASK)
3334 return show_irq_str(level);
3337 static void dump_buffer_page(struct buffer_data_page *bpage,
3338 struct rb_event_info *info,
3341 struct ring_buffer_event *event;
3345 ts = bpage->time_stamp;
3346 pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
3348 for (e = 0; e < tail; e += rb_event_length(event)) {
3350 event = (struct ring_buffer_event *)(bpage->data + e);
3352 switch (event->type_len) {
3354 case RINGBUF_TYPE_TIME_EXTEND:
3355 delta = rb_event_time_stamp(event);
3357 pr_warn(" 0x%x: [%lld] delta:%lld TIME EXTEND\n",
3361 case RINGBUF_TYPE_TIME_STAMP:
3362 delta = rb_event_time_stamp(event);
3363 ts = rb_fix_abs_ts(delta, ts);
3364 pr_warn(" 0x%x: [%lld] absolute:%lld TIME STAMP\n",
3368 case RINGBUF_TYPE_PADDING:
3369 ts += event->time_delta;
3370 pr_warn(" 0x%x: [%lld] delta:%d PADDING\n",
3371 e, ts, event->time_delta);
3374 case RINGBUF_TYPE_DATA:
3375 ts += event->time_delta;
3376 pr_warn(" 0x%x: [%lld] delta:%d %s%s\n",
3377 e, ts, event->time_delta,
3378 show_flags(event), show_irq(event));
3385 pr_warn("expected end:0x%lx last event actually ended at:0x%x\n", tail, e);
3388 static DEFINE_PER_CPU(atomic_t, checking);
3389 static atomic_t ts_dump;
3391 #define buffer_warn_return(fmt, ...) \
3393 /* If another report is happening, ignore this one */ \
3394 if (atomic_inc_return(&ts_dump) != 1) { \
3395 atomic_dec(&ts_dump); \
3398 atomic_inc(&cpu_buffer->record_disabled); \
3399 pr_warn(fmt, ##__VA_ARGS__); \
3400 dump_buffer_page(bpage, info, tail); \
3401 atomic_dec(&ts_dump); \
3402 /* There's some cases in boot up that this can happen */ \
3403 if (WARN_ON_ONCE(system_state != SYSTEM_BOOTING)) \
3404 /* Do not re-enable checking */ \
3409 * Check if the current event time stamp matches the deltas on
3412 static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3413 struct rb_event_info *info,
3416 struct ring_buffer_event *event;
3417 struct buffer_data_page *bpage;
3422 bpage = info->tail_page->page;
3424 if (tail == CHECK_FULL_PAGE) {
3426 tail = local_read(&bpage->commit);
3427 } else if (info->add_timestamp &
3428 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
3429 /* Ignore events with absolute time stamps */
3434 * Do not check the first event (skip possible extends too).
3435 * Also do not check if previous events have not been committed.
3437 if (tail <= 8 || tail > local_read(&bpage->commit))
3441 * If this interrupted another event,
3443 if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
3446 ts = bpage->time_stamp;
3448 for (e = 0; e < tail; e += rb_event_length(event)) {
3450 event = (struct ring_buffer_event *)(bpage->data + e);
3452 switch (event->type_len) {
3454 case RINGBUF_TYPE_TIME_EXTEND:
3455 delta = rb_event_time_stamp(event);
3459 case RINGBUF_TYPE_TIME_STAMP:
3460 delta = rb_event_time_stamp(event);
3461 delta = rb_fix_abs_ts(delta, ts);
3463 buffer_warn_return("[CPU: %d]ABSOLUTE TIME WENT BACKWARDS: last ts: %lld absolute ts: %lld\n",
3464 cpu_buffer->cpu, ts, delta);
3469 case RINGBUF_TYPE_PADDING:
3470 if (event->time_delta == 1)
3473 case RINGBUF_TYPE_DATA:
3474 ts += event->time_delta;
3478 RB_WARN_ON(cpu_buffer, 1);
3481 if ((full && ts > info->ts) ||
3482 (!full && ts + info->delta != info->ts)) {
3483 buffer_warn_return("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s context:%s\n",
3485 ts + info->delta, info->ts, info->delta,
3486 info->before, info->after,
3487 full ? " (full)" : "", show_interrupt_level());
3490 atomic_dec(this_cpu_ptr(&checking));
3493 static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3494 struct rb_event_info *info,
3498 #endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
3500 static struct ring_buffer_event *
3501 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3502 struct rb_event_info *info)
3504 struct ring_buffer_event *event;
3505 struct buffer_page *tail_page;
3506 unsigned long tail, write, w;
3508 /* Don't let the compiler play games with cpu_buffer->tail_page */
3509 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3511 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
3513 rb_time_read(&cpu_buffer->before_stamp, &info->before);
3514 rb_time_read(&cpu_buffer->write_stamp, &info->after);
3516 info->ts = rb_time_stamp(cpu_buffer->buffer);
3518 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3519 info->delta = info->ts;
3522 * If interrupting an event time update, we may need an
3523 * absolute timestamp.
3524 * Don't bother if this is the start of a new page (w == 0).
3527 /* Use the sub-buffer timestamp */
3529 } else if (unlikely(info->before != info->after)) {
3530 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3531 info->length += RB_LEN_TIME_EXTEND;
3533 info->delta = info->ts - info->after;
3534 if (unlikely(test_time_stamp(info->delta))) {
3535 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3536 info->length += RB_LEN_TIME_EXTEND;
3541 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
3543 /*C*/ write = local_add_return(info->length, &tail_page->write);
3545 /* set write to only the index of the write */
3546 write &= RB_WRITE_MASK;
3548 tail = write - info->length;
3550 /* See if we shot pass the end of this buffer page */
3551 if (unlikely(write > cpu_buffer->buffer->subbuf_size)) {
3552 check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
3553 return rb_move_tail(cpu_buffer, tail, info);
3556 if (likely(tail == w)) {
3557 /* Nothing interrupted us between A and C */
3558 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
3560 * If something came in between C and D, the write stamp
3561 * may now not be in sync. But that's fine as the before_stamp
3562 * will be different and then next event will just be forced
3563 * to use an absolute timestamp.
3565 if (likely(!(info->add_timestamp &
3566 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3567 /* This did not interrupt any time update */
3568 info->delta = info->ts - info->after;
3570 /* Just use full timestamp for interrupting event */
3571 info->delta = info->ts;
3572 check_buffer(cpu_buffer, info, tail);
3575 /* SLOW PATH - Interrupted between A and C */
3577 /* Save the old before_stamp */
3578 rb_time_read(&cpu_buffer->before_stamp, &info->before);
3581 * Read a new timestamp and update the before_stamp to make
3582 * the next event after this one force using an absolute
3583 * timestamp. This is in case an interrupt were to come in
3586 ts = rb_time_stamp(cpu_buffer->buffer);
3587 rb_time_set(&cpu_buffer->before_stamp, ts);
3590 /*E*/ rb_time_read(&cpu_buffer->write_stamp, &info->after);
3592 /*F*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3593 info->after == info->before && info->after < ts) {
3595 * Nothing came after this event between C and F, it is
3596 * safe to use info->after for the delta as it
3597 * matched info->before and is still valid.
3599 info->delta = ts - info->after;
3602 * Interrupted between C and F:
3603 * Lost the previous events time stamp. Just set the
3604 * delta to zero, and this will be the same time as
3605 * the event this event interrupted. And the events that
3606 * came after this will still be correct (as they would
3607 * have built their delta on the previous event.
3612 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3616 * If this is the first commit on the page, then it has the same
3617 * timestamp as the page itself.
3619 if (unlikely(!tail && !(info->add_timestamp &
3620 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3623 /* We reserved something on the buffer */
3625 event = __rb_page_index(tail_page, tail);
3626 rb_update_event(cpu_buffer, event, info);
3628 local_inc(&tail_page->entries);
3631 * If this is the first commit on the page, then update
3634 if (unlikely(!tail))
3635 tail_page->page->time_stamp = info->ts;
3637 /* account for these added bytes */
3638 local_add(info->length, &cpu_buffer->entries_bytes);
3643 static __always_inline struct ring_buffer_event *
3644 rb_reserve_next_event(struct trace_buffer *buffer,
3645 struct ring_buffer_per_cpu *cpu_buffer,
3646 unsigned long length)
3648 struct ring_buffer_event *event;
3649 struct rb_event_info info;
3653 /* ring buffer does cmpxchg, make sure it is safe in NMI context */
3654 if (!IS_ENABLED(CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG) &&
3655 (unlikely(in_nmi()))) {
3659 rb_start_commit(cpu_buffer);
3660 /* The commit page can not change after this */
3662 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3664 * Due to the ability to swap a cpu buffer from a buffer
3665 * it is possible it was swapped before we committed.
3666 * (committing stops a swap). We check for it here and
3667 * if it happened, we have to fail the write.
3670 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3671 local_dec(&cpu_buffer->committing);
3672 local_dec(&cpu_buffer->commits);
3677 info.length = rb_calculate_event_length(length);
3679 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3680 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3681 info.length += RB_LEN_TIME_EXTEND;
3682 if (info.length > cpu_buffer->buffer->max_data_size)
3685 add_ts_default = RB_ADD_STAMP_NONE;
3689 info.add_timestamp = add_ts_default;
3693 * We allow for interrupts to reenter here and do a trace.
3694 * If one does, it will cause this original code to loop
3695 * back here. Even with heavy interrupts happening, this
3696 * should only happen a few times in a row. If this happens
3697 * 1000 times in a row, there must be either an interrupt
3698 * storm or we have something buggy.
3701 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3704 event = __rb_reserve_next(cpu_buffer, &info);
3706 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3707 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3708 info.length -= RB_LEN_TIME_EXTEND;
3715 rb_end_commit(cpu_buffer);
3720 * ring_buffer_lock_reserve - reserve a part of the buffer
3721 * @buffer: the ring buffer to reserve from
3722 * @length: the length of the data to reserve (excluding event header)
3724 * Returns a reserved event on the ring buffer to copy directly to.
3725 * The user of this interface will need to get the body to write into
3726 * and can use the ring_buffer_event_data() interface.
3728 * The length is the length of the data needed, not the event length
3729 * which also includes the event header.
3731 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3732 * If NULL is returned, then nothing has been allocated or locked.
3734 struct ring_buffer_event *
3735 ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3737 struct ring_buffer_per_cpu *cpu_buffer;
3738 struct ring_buffer_event *event;
3741 /* If we are tracing schedule, we don't want to recurse */
3742 preempt_disable_notrace();
3744 if (unlikely(atomic_read(&buffer->record_disabled)))
3747 cpu = raw_smp_processor_id();
3749 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3752 cpu_buffer = buffer->buffers[cpu];
3754 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3757 if (unlikely(length > buffer->max_data_size))
3760 if (unlikely(trace_recursive_lock(cpu_buffer)))
3763 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3770 trace_recursive_unlock(cpu_buffer);
3772 preempt_enable_notrace();
3775 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3778 * Decrement the entries to the page that an event is on.
3779 * The event does not even need to exist, only the pointer
3780 * to the page it is on. This may only be called before the commit
3784 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3785 struct ring_buffer_event *event)
3787 unsigned long addr = (unsigned long)event;
3788 struct buffer_page *bpage = cpu_buffer->commit_page;
3789 struct buffer_page *start;
3791 addr &= ~((PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1);
3793 /* Do the likely case first */
3794 if (likely(bpage->page == (void *)addr)) {
3795 local_dec(&bpage->entries);
3800 * Because the commit page may be on the reader page we
3801 * start with the next page and check the end loop there.
3803 rb_inc_page(&bpage);
3806 if (bpage->page == (void *)addr) {
3807 local_dec(&bpage->entries);
3810 rb_inc_page(&bpage);
3811 } while (bpage != start);
3813 /* commit not part of this buffer?? */
3814 RB_WARN_ON(cpu_buffer, 1);
3818 * ring_buffer_discard_commit - discard an event that has not been committed
3819 * @buffer: the ring buffer
3820 * @event: non committed event to discard
3822 * Sometimes an event that is in the ring buffer needs to be ignored.
3823 * This function lets the user discard an event in the ring buffer
3824 * and then that event will not be read later.
3826 * This function only works if it is called before the item has been
3827 * committed. It will try to free the event from the ring buffer
3828 * if another event has not been added behind it.
3830 * If another event has been added behind it, it will set the event
3831 * up as discarded, and perform the commit.
3833 * If this function is called, do not call ring_buffer_unlock_commit on
3836 void ring_buffer_discard_commit(struct trace_buffer *buffer,
3837 struct ring_buffer_event *event)
3839 struct ring_buffer_per_cpu *cpu_buffer;
3842 /* The event is discarded regardless */
3843 rb_event_discard(event);
3845 cpu = smp_processor_id();
3846 cpu_buffer = buffer->buffers[cpu];
3849 * This must only be called if the event has not been
3850 * committed yet. Thus we can assume that preemption
3851 * is still disabled.
3853 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3855 rb_decrement_entry(cpu_buffer, event);
3856 if (rb_try_to_discard(cpu_buffer, event))
3860 rb_end_commit(cpu_buffer);
3862 trace_recursive_unlock(cpu_buffer);
3864 preempt_enable_notrace();
3867 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3870 * ring_buffer_write - write data to the buffer without reserving
3871 * @buffer: The ring buffer to write to.
3872 * @length: The length of the data being written (excluding the event header)
3873 * @data: The data to write to the buffer.
3875 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3876 * one function. If you already have the data to write to the buffer, it
3877 * may be easier to simply call this function.
3879 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3880 * and not the length of the event which would hold the header.
3882 int ring_buffer_write(struct trace_buffer *buffer,
3883 unsigned long length,
3886 struct ring_buffer_per_cpu *cpu_buffer;
3887 struct ring_buffer_event *event;
3892 preempt_disable_notrace();
3894 if (atomic_read(&buffer->record_disabled))
3897 cpu = raw_smp_processor_id();
3899 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3902 cpu_buffer = buffer->buffers[cpu];
3904 if (atomic_read(&cpu_buffer->record_disabled))
3907 if (length > buffer->max_data_size)
3910 if (unlikely(trace_recursive_lock(cpu_buffer)))
3913 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3917 body = rb_event_data(event);
3919 memcpy(body, data, length);
3921 rb_commit(cpu_buffer);
3923 rb_wakeups(buffer, cpu_buffer);
3928 trace_recursive_unlock(cpu_buffer);
3931 preempt_enable_notrace();
3935 EXPORT_SYMBOL_GPL(ring_buffer_write);
3937 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3939 struct buffer_page *reader = cpu_buffer->reader_page;
3940 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3941 struct buffer_page *commit = cpu_buffer->commit_page;
3943 /* In case of error, head will be NULL */
3944 if (unlikely(!head))
3947 /* Reader should exhaust content in reader page */
3948 if (reader->read != rb_page_commit(reader))
3952 * If writers are committing on the reader page, knowing all
3953 * committed content has been read, the ring buffer is empty.
3955 if (commit == reader)
3959 * If writers are committing on a page other than reader page
3960 * and head page, there should always be content to read.
3966 * Writers are committing on the head page, we just need
3967 * to care about there're committed data, and the reader will
3968 * swap reader page with head page when it is to read data.
3970 return rb_page_commit(commit) == 0;
3974 * ring_buffer_record_disable - stop all writes into the buffer
3975 * @buffer: The ring buffer to stop writes to.
3977 * This prevents all writes to the buffer. Any attempt to write
3978 * to the buffer after this will fail and return NULL.
3980 * The caller should call synchronize_rcu() after this.
3982 void ring_buffer_record_disable(struct trace_buffer *buffer)
3984 atomic_inc(&buffer->record_disabled);
3986 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3989 * ring_buffer_record_enable - enable writes to the buffer
3990 * @buffer: The ring buffer to enable writes
3992 * Note, multiple disables will need the same number of enables
3993 * to truly enable the writing (much like preempt_disable).
3995 void ring_buffer_record_enable(struct trace_buffer *buffer)
3997 atomic_dec(&buffer->record_disabled);
3999 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
4002 * ring_buffer_record_off - stop all writes into the buffer
4003 * @buffer: The ring buffer to stop writes to.
4005 * This prevents all writes to the buffer. Any attempt to write
4006 * to the buffer after this will fail and return NULL.
4008 * This is different than ring_buffer_record_disable() as
4009 * it works like an on/off switch, where as the disable() version
4010 * must be paired with a enable().
4012 void ring_buffer_record_off(struct trace_buffer *buffer)
4015 unsigned int new_rd;
4017 rd = atomic_read(&buffer->record_disabled);
4019 new_rd = rd | RB_BUFFER_OFF;
4020 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4022 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
4025 * ring_buffer_record_on - restart writes into the buffer
4026 * @buffer: The ring buffer to start writes to.
4028 * This enables all writes to the buffer that was disabled by
4029 * ring_buffer_record_off().
4031 * This is different than ring_buffer_record_enable() as
4032 * it works like an on/off switch, where as the enable() version
4033 * must be paired with a disable().
4035 void ring_buffer_record_on(struct trace_buffer *buffer)
4038 unsigned int new_rd;
4040 rd = atomic_read(&buffer->record_disabled);
4042 new_rd = rd & ~RB_BUFFER_OFF;
4043 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4045 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
4048 * ring_buffer_record_is_on - return true if the ring buffer can write
4049 * @buffer: The ring buffer to see if write is enabled
4051 * Returns true if the ring buffer is in a state that it accepts writes.
4053 bool ring_buffer_record_is_on(struct trace_buffer *buffer)
4055 return !atomic_read(&buffer->record_disabled);
4059 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
4060 * @buffer: The ring buffer to see if write is set enabled
4062 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
4063 * Note that this does NOT mean it is in a writable state.
4065 * It may return true when the ring buffer has been disabled by
4066 * ring_buffer_record_disable(), as that is a temporary disabling of
4069 bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4071 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4075 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4076 * @buffer: The ring buffer to stop writes to.
4077 * @cpu: The CPU buffer to stop
4079 * This prevents all writes to the buffer. Any attempt to write
4080 * to the buffer after this will fail and return NULL.
4082 * The caller should call synchronize_rcu() after this.
4084 void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4086 struct ring_buffer_per_cpu *cpu_buffer;
4088 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4091 cpu_buffer = buffer->buffers[cpu];
4092 atomic_inc(&cpu_buffer->record_disabled);
4094 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4097 * ring_buffer_record_enable_cpu - enable writes to the buffer
4098 * @buffer: The ring buffer to enable writes
4099 * @cpu: The CPU to enable.
4101 * Note, multiple disables will need the same number of enables
4102 * to truly enable the writing (much like preempt_disable).
4104 void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4106 struct ring_buffer_per_cpu *cpu_buffer;
4108 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4111 cpu_buffer = buffer->buffers[cpu];
4112 atomic_dec(&cpu_buffer->record_disabled);
4114 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4117 * The total entries in the ring buffer is the running counter
4118 * of entries entered into the ring buffer, minus the sum of
4119 * the entries read from the ring buffer and the number of
4120 * entries that were overwritten.
4122 static inline unsigned long
4123 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4125 return local_read(&cpu_buffer->entries) -
4126 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4130 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4131 * @buffer: The ring buffer
4132 * @cpu: The per CPU buffer to read from.
4134 u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4136 unsigned long flags;
4137 struct ring_buffer_per_cpu *cpu_buffer;
4138 struct buffer_page *bpage;
4141 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4144 cpu_buffer = buffer->buffers[cpu];
4145 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4147 * if the tail is on reader_page, oldest time stamp is on the reader
4150 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4151 bpage = cpu_buffer->reader_page;
4153 bpage = rb_set_head_page(cpu_buffer);
4155 ret = bpage->page->time_stamp;
4156 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4160 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4163 * ring_buffer_bytes_cpu - get the number of bytes unconsumed in a cpu buffer
4164 * @buffer: The ring buffer
4165 * @cpu: The per CPU buffer to read from.
4167 unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4169 struct ring_buffer_per_cpu *cpu_buffer;
4172 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4175 cpu_buffer = buffer->buffers[cpu];
4176 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4180 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4183 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4184 * @buffer: The ring buffer
4185 * @cpu: The per CPU buffer to get the entries from.
4187 unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4189 struct ring_buffer_per_cpu *cpu_buffer;
4191 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4194 cpu_buffer = buffer->buffers[cpu];
4196 return rb_num_of_entries(cpu_buffer);
4198 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4201 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4202 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4203 * @buffer: The ring buffer
4204 * @cpu: The per CPU buffer to get the number of overruns from
4206 unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4208 struct ring_buffer_per_cpu *cpu_buffer;
4211 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4214 cpu_buffer = buffer->buffers[cpu];
4215 ret = local_read(&cpu_buffer->overrun);
4219 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4222 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4223 * commits failing due to the buffer wrapping around while there are uncommitted
4224 * events, such as during an interrupt storm.
4225 * @buffer: The ring buffer
4226 * @cpu: The per CPU buffer to get the number of overruns from
4229 ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4231 struct ring_buffer_per_cpu *cpu_buffer;
4234 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4237 cpu_buffer = buffer->buffers[cpu];
4238 ret = local_read(&cpu_buffer->commit_overrun);
4242 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4245 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4246 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4247 * @buffer: The ring buffer
4248 * @cpu: The per CPU buffer to get the number of overruns from
4251 ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4253 struct ring_buffer_per_cpu *cpu_buffer;
4256 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4259 cpu_buffer = buffer->buffers[cpu];
4260 ret = local_read(&cpu_buffer->dropped_events);
4264 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4267 * ring_buffer_read_events_cpu - get the number of events successfully read
4268 * @buffer: The ring buffer
4269 * @cpu: The per CPU buffer to get the number of events read
4272 ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4274 struct ring_buffer_per_cpu *cpu_buffer;
4276 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4279 cpu_buffer = buffer->buffers[cpu];
4280 return cpu_buffer->read;
4282 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4285 * ring_buffer_entries - get the number of entries in a buffer
4286 * @buffer: The ring buffer
4288 * Returns the total number of entries in the ring buffer
4291 unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4293 struct ring_buffer_per_cpu *cpu_buffer;
4294 unsigned long entries = 0;
4297 /* if you care about this being correct, lock the buffer */
4298 for_each_buffer_cpu(buffer, cpu) {
4299 cpu_buffer = buffer->buffers[cpu];
4300 entries += rb_num_of_entries(cpu_buffer);
4305 EXPORT_SYMBOL_GPL(ring_buffer_entries);
4308 * ring_buffer_overruns - get the number of overruns in buffer
4309 * @buffer: The ring buffer
4311 * Returns the total number of overruns in the ring buffer
4314 unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4316 struct ring_buffer_per_cpu *cpu_buffer;
4317 unsigned long overruns = 0;
4320 /* if you care about this being correct, lock the buffer */
4321 for_each_buffer_cpu(buffer, cpu) {
4322 cpu_buffer = buffer->buffers[cpu];
4323 overruns += local_read(&cpu_buffer->overrun);
4328 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4330 static void rb_iter_reset(struct ring_buffer_iter *iter)
4332 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4334 /* Iterator usage is expected to have record disabled */
4335 iter->head_page = cpu_buffer->reader_page;
4336 iter->head = cpu_buffer->reader_page->read;
4337 iter->next_event = iter->head;
4339 iter->cache_reader_page = iter->head_page;
4340 iter->cache_read = cpu_buffer->read;
4341 iter->cache_pages_removed = cpu_buffer->pages_removed;
4344 iter->read_stamp = cpu_buffer->read_stamp;
4345 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4347 iter->read_stamp = iter->head_page->page->time_stamp;
4348 iter->page_stamp = iter->read_stamp;
4353 * ring_buffer_iter_reset - reset an iterator
4354 * @iter: The iterator to reset
4356 * Resets the iterator, so that it will start from the beginning
4359 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4361 struct ring_buffer_per_cpu *cpu_buffer;
4362 unsigned long flags;
4367 cpu_buffer = iter->cpu_buffer;
4369 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4370 rb_iter_reset(iter);
4371 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4373 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4376 * ring_buffer_iter_empty - check if an iterator has no more to read
4377 * @iter: The iterator to check
4379 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4381 struct ring_buffer_per_cpu *cpu_buffer;
4382 struct buffer_page *reader;
4383 struct buffer_page *head_page;
4384 struct buffer_page *commit_page;
4385 struct buffer_page *curr_commit_page;
4390 cpu_buffer = iter->cpu_buffer;
4391 reader = cpu_buffer->reader_page;
4392 head_page = cpu_buffer->head_page;
4393 commit_page = READ_ONCE(cpu_buffer->commit_page);
4394 commit_ts = commit_page->page->time_stamp;
4397 * When the writer goes across pages, it issues a cmpxchg which
4398 * is a mb(), which will synchronize with the rmb here.
4399 * (see rb_tail_page_update())
4402 commit = rb_page_commit(commit_page);
4403 /* We want to make sure that the commit page doesn't change */
4406 /* Make sure commit page didn't change */
4407 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4408 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4410 /* If the commit page changed, then there's more data */
4411 if (curr_commit_page != commit_page ||
4412 curr_commit_ts != commit_ts)
4415 /* Still racy, as it may return a false positive, but that's OK */
4416 return ((iter->head_page == commit_page && iter->head >= commit) ||
4417 (iter->head_page == reader && commit_page == head_page &&
4418 head_page->read == commit &&
4419 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4421 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4424 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4425 struct ring_buffer_event *event)
4429 switch (event->type_len) {
4430 case RINGBUF_TYPE_PADDING:
4433 case RINGBUF_TYPE_TIME_EXTEND:
4434 delta = rb_event_time_stamp(event);
4435 cpu_buffer->read_stamp += delta;
4438 case RINGBUF_TYPE_TIME_STAMP:
4439 delta = rb_event_time_stamp(event);
4440 delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
4441 cpu_buffer->read_stamp = delta;
4444 case RINGBUF_TYPE_DATA:
4445 cpu_buffer->read_stamp += event->time_delta;
4449 RB_WARN_ON(cpu_buffer, 1);
4454 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4455 struct ring_buffer_event *event)
4459 switch (event->type_len) {
4460 case RINGBUF_TYPE_PADDING:
4463 case RINGBUF_TYPE_TIME_EXTEND:
4464 delta = rb_event_time_stamp(event);
4465 iter->read_stamp += delta;
4468 case RINGBUF_TYPE_TIME_STAMP:
4469 delta = rb_event_time_stamp(event);
4470 delta = rb_fix_abs_ts(delta, iter->read_stamp);
4471 iter->read_stamp = delta;
4474 case RINGBUF_TYPE_DATA:
4475 iter->read_stamp += event->time_delta;
4479 RB_WARN_ON(iter->cpu_buffer, 1);
4483 static struct buffer_page *
4484 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4486 struct buffer_page *reader = NULL;
4487 unsigned long bsize = READ_ONCE(cpu_buffer->buffer->subbuf_size);
4488 unsigned long overwrite;
4489 unsigned long flags;
4493 local_irq_save(flags);
4494 arch_spin_lock(&cpu_buffer->lock);
4498 * This should normally only loop twice. But because the
4499 * start of the reader inserts an empty page, it causes
4500 * a case where we will loop three times. There should be no
4501 * reason to loop four times (that I know of).
4503 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4508 reader = cpu_buffer->reader_page;
4510 /* If there's more to read, return this page */
4511 if (cpu_buffer->reader_page->read < rb_page_size(reader))
4514 /* Never should we have an index greater than the size */
4515 if (RB_WARN_ON(cpu_buffer,
4516 cpu_buffer->reader_page->read > rb_page_size(reader)))
4519 /* check if we caught up to the tail */
4521 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4524 /* Don't bother swapping if the ring buffer is empty */
4525 if (rb_num_of_entries(cpu_buffer) == 0)
4529 * Reset the reader page to size zero.
4531 local_set(&cpu_buffer->reader_page->write, 0);
4532 local_set(&cpu_buffer->reader_page->entries, 0);
4533 local_set(&cpu_buffer->reader_page->page->commit, 0);
4534 cpu_buffer->reader_page->real_end = 0;
4538 * Splice the empty reader page into the list around the head.
4540 reader = rb_set_head_page(cpu_buffer);
4543 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4544 cpu_buffer->reader_page->list.prev = reader->list.prev;
4547 * cpu_buffer->pages just needs to point to the buffer, it
4548 * has no specific buffer page to point to. Lets move it out
4549 * of our way so we don't accidentally swap it.
4551 cpu_buffer->pages = reader->list.prev;
4553 /* The reader page will be pointing to the new head */
4554 rb_set_list_to_head(&cpu_buffer->reader_page->list);
4557 * We want to make sure we read the overruns after we set up our
4558 * pointers to the next object. The writer side does a
4559 * cmpxchg to cross pages which acts as the mb on the writer
4560 * side. Note, the reader will constantly fail the swap
4561 * while the writer is updating the pointers, so this
4562 * guarantees that the overwrite recorded here is the one we
4563 * want to compare with the last_overrun.
4566 overwrite = local_read(&(cpu_buffer->overrun));
4569 * Here's the tricky part.
4571 * We need to move the pointer past the header page.
4572 * But we can only do that if a writer is not currently
4573 * moving it. The page before the header page has the
4574 * flag bit '1' set if it is pointing to the page we want.
4575 * but if the writer is in the process of moving it
4576 * than it will be '2' or already moved '0'.
4579 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4582 * If we did not convert it, then we must try again.
4588 * Yay! We succeeded in replacing the page.
4590 * Now make the new head point back to the reader page.
4592 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4593 rb_inc_page(&cpu_buffer->head_page);
4595 local_inc(&cpu_buffer->pages_read);
4597 /* Finally update the reader page to the new head */
4598 cpu_buffer->reader_page = reader;
4599 cpu_buffer->reader_page->read = 0;
4601 if (overwrite != cpu_buffer->last_overrun) {
4602 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4603 cpu_buffer->last_overrun = overwrite;
4609 /* Update the read_stamp on the first event */
4610 if (reader && reader->read == 0)
4611 cpu_buffer->read_stamp = reader->page->time_stamp;
4613 arch_spin_unlock(&cpu_buffer->lock);
4614 local_irq_restore(flags);
4617 * The writer has preempt disable, wait for it. But not forever
4618 * Although, 1 second is pretty much "forever"
4620 #define USECS_WAIT 1000000
4621 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
4622 /* If the write is past the end of page, a writer is still updating it */
4623 if (likely(!reader || rb_page_write(reader) <= bsize))
4628 /* Get the latest version of the reader write value */
4632 /* The writer is not moving forward? Something is wrong */
4633 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
4637 * Make sure we see any padding after the write update
4638 * (see rb_reset_tail()).
4640 * In addition, a writer may be writing on the reader page
4641 * if the page has not been fully filled, so the read barrier
4642 * is also needed to make sure we see the content of what is
4643 * committed by the writer (see rb_set_commit_to_write()).
4651 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4653 struct ring_buffer_event *event;
4654 struct buffer_page *reader;
4657 reader = rb_get_reader_page(cpu_buffer);
4659 /* This function should not be called when buffer is empty */
4660 if (RB_WARN_ON(cpu_buffer, !reader))
4663 event = rb_reader_event(cpu_buffer);
4665 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4668 rb_update_read_stamp(cpu_buffer, event);
4670 length = rb_event_length(event);
4671 cpu_buffer->reader_page->read += length;
4672 cpu_buffer->read_bytes += length;
4675 static void rb_advance_iter(struct ring_buffer_iter *iter)
4677 struct ring_buffer_per_cpu *cpu_buffer;
4679 cpu_buffer = iter->cpu_buffer;
4681 /* If head == next_event then we need to jump to the next event */
4682 if (iter->head == iter->next_event) {
4683 /* If the event gets overwritten again, there's nothing to do */
4684 if (rb_iter_head_event(iter) == NULL)
4688 iter->head = iter->next_event;
4691 * Check if we are at the end of the buffer.
4693 if (iter->next_event >= rb_page_size(iter->head_page)) {
4694 /* discarded commits can make the page empty */
4695 if (iter->head_page == cpu_buffer->commit_page)
4701 rb_update_iter_read_stamp(iter, iter->event);
4704 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4706 return cpu_buffer->lost_events;
4709 static struct ring_buffer_event *
4710 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4711 unsigned long *lost_events)
4713 struct ring_buffer_event *event;
4714 struct buffer_page *reader;
4721 * We repeat when a time extend is encountered.
4722 * Since the time extend is always attached to a data event,
4723 * we should never loop more than once.
4724 * (We never hit the following condition more than twice).
4726 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4729 reader = rb_get_reader_page(cpu_buffer);
4733 event = rb_reader_event(cpu_buffer);
4735 switch (event->type_len) {
4736 case RINGBUF_TYPE_PADDING:
4737 if (rb_null_event(event))
4738 RB_WARN_ON(cpu_buffer, 1);
4740 * Because the writer could be discarding every
4741 * event it creates (which would probably be bad)
4742 * if we were to go back to "again" then we may never
4743 * catch up, and will trigger the warn on, or lock
4744 * the box. Return the padding, and we will release
4745 * the current locks, and try again.
4749 case RINGBUF_TYPE_TIME_EXTEND:
4750 /* Internal data, OK to advance */
4751 rb_advance_reader(cpu_buffer);
4754 case RINGBUF_TYPE_TIME_STAMP:
4756 *ts = rb_event_time_stamp(event);
4757 *ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
4758 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4759 cpu_buffer->cpu, ts);
4761 /* Internal data, OK to advance */
4762 rb_advance_reader(cpu_buffer);
4765 case RINGBUF_TYPE_DATA:
4767 *ts = cpu_buffer->read_stamp + event->time_delta;
4768 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4769 cpu_buffer->cpu, ts);
4772 *lost_events = rb_lost_events(cpu_buffer);
4776 RB_WARN_ON(cpu_buffer, 1);
4781 EXPORT_SYMBOL_GPL(ring_buffer_peek);
4783 static struct ring_buffer_event *
4784 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4786 struct trace_buffer *buffer;
4787 struct ring_buffer_per_cpu *cpu_buffer;
4788 struct ring_buffer_event *event;
4794 cpu_buffer = iter->cpu_buffer;
4795 buffer = cpu_buffer->buffer;
4798 * Check if someone performed a consuming read to the buffer
4799 * or removed some pages from the buffer. In these cases,
4800 * iterator was invalidated and we need to reset it.
4802 if (unlikely(iter->cache_read != cpu_buffer->read ||
4803 iter->cache_reader_page != cpu_buffer->reader_page ||
4804 iter->cache_pages_removed != cpu_buffer->pages_removed))
4805 rb_iter_reset(iter);
4808 if (ring_buffer_iter_empty(iter))
4812 * As the writer can mess with what the iterator is trying
4813 * to read, just give up if we fail to get an event after
4814 * three tries. The iterator is not as reliable when reading
4815 * the ring buffer with an active write as the consumer is.
4816 * Do not warn if the three failures is reached.
4821 if (rb_per_cpu_empty(cpu_buffer))
4824 if (iter->head >= rb_page_size(iter->head_page)) {
4829 event = rb_iter_head_event(iter);
4833 switch (event->type_len) {
4834 case RINGBUF_TYPE_PADDING:
4835 if (rb_null_event(event)) {
4839 rb_advance_iter(iter);
4842 case RINGBUF_TYPE_TIME_EXTEND:
4843 /* Internal data, OK to advance */
4844 rb_advance_iter(iter);
4847 case RINGBUF_TYPE_TIME_STAMP:
4849 *ts = rb_event_time_stamp(event);
4850 *ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
4851 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4852 cpu_buffer->cpu, ts);
4854 /* Internal data, OK to advance */
4855 rb_advance_iter(iter);
4858 case RINGBUF_TYPE_DATA:
4860 *ts = iter->read_stamp + event->time_delta;
4861 ring_buffer_normalize_time_stamp(buffer,
4862 cpu_buffer->cpu, ts);
4867 RB_WARN_ON(cpu_buffer, 1);
4872 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4874 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4876 if (likely(!in_nmi())) {
4877 raw_spin_lock(&cpu_buffer->reader_lock);
4882 * If an NMI die dumps out the content of the ring buffer
4883 * trylock must be used to prevent a deadlock if the NMI
4884 * preempted a task that holds the ring buffer locks. If
4885 * we get the lock then all is fine, if not, then continue
4886 * to do the read, but this can corrupt the ring buffer,
4887 * so it must be permanently disabled from future writes.
4888 * Reading from NMI is a oneshot deal.
4890 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4893 /* Continue without locking, but disable the ring buffer */
4894 atomic_inc(&cpu_buffer->record_disabled);
4899 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4902 raw_spin_unlock(&cpu_buffer->reader_lock);
4906 * ring_buffer_peek - peek at the next event to be read
4907 * @buffer: The ring buffer to read
4908 * @cpu: The cpu to peak at
4909 * @ts: The timestamp counter of this event.
4910 * @lost_events: a variable to store if events were lost (may be NULL)
4912 * This will return the event that will be read next, but does
4913 * not consume the data.
4915 struct ring_buffer_event *
4916 ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4917 unsigned long *lost_events)
4919 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4920 struct ring_buffer_event *event;
4921 unsigned long flags;
4924 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4928 local_irq_save(flags);
4929 dolock = rb_reader_lock(cpu_buffer);
4930 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4931 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4932 rb_advance_reader(cpu_buffer);
4933 rb_reader_unlock(cpu_buffer, dolock);
4934 local_irq_restore(flags);
4936 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4942 /** ring_buffer_iter_dropped - report if there are dropped events
4943 * @iter: The ring buffer iterator
4945 * Returns true if there was dropped events since the last peek.
4947 bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4949 bool ret = iter->missed_events != 0;
4951 iter->missed_events = 0;
4954 EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
4957 * ring_buffer_iter_peek - peek at the next event to be read
4958 * @iter: The ring buffer iterator
4959 * @ts: The timestamp counter of this event.
4961 * This will return the event that will be read next, but does
4962 * not increment the iterator.
4964 struct ring_buffer_event *
4965 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4967 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4968 struct ring_buffer_event *event;
4969 unsigned long flags;
4972 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4973 event = rb_iter_peek(iter, ts);
4974 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4976 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4983 * ring_buffer_consume - return an event and consume it
4984 * @buffer: The ring buffer to get the next event from
4985 * @cpu: the cpu to read the buffer from
4986 * @ts: a variable to store the timestamp (may be NULL)
4987 * @lost_events: a variable to store if events were lost (may be NULL)
4989 * Returns the next event in the ring buffer, and that event is consumed.
4990 * Meaning, that sequential reads will keep returning a different event,
4991 * and eventually empty the ring buffer if the producer is slower.
4993 struct ring_buffer_event *
4994 ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
4995 unsigned long *lost_events)
4997 struct ring_buffer_per_cpu *cpu_buffer;
4998 struct ring_buffer_event *event = NULL;
4999 unsigned long flags;
5003 /* might be called in atomic */
5006 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5009 cpu_buffer = buffer->buffers[cpu];
5010 local_irq_save(flags);
5011 dolock = rb_reader_lock(cpu_buffer);
5013 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
5015 cpu_buffer->lost_events = 0;
5016 rb_advance_reader(cpu_buffer);
5019 rb_reader_unlock(cpu_buffer, dolock);
5020 local_irq_restore(flags);
5025 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5030 EXPORT_SYMBOL_GPL(ring_buffer_consume);
5033 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
5034 * @buffer: The ring buffer to read from
5035 * @cpu: The cpu buffer to iterate over
5036 * @flags: gfp flags to use for memory allocation
5038 * This performs the initial preparations necessary to iterate
5039 * through the buffer. Memory is allocated, buffer recording
5040 * is disabled, and the iterator pointer is returned to the caller.
5042 * Disabling buffer recording prevents the reading from being
5043 * corrupted. This is not a consuming read, so a producer is not
5046 * After a sequence of ring_buffer_read_prepare calls, the user is
5047 * expected to make at least one call to ring_buffer_read_prepare_sync.
5048 * Afterwards, ring_buffer_read_start is invoked to get things going
5051 * This overall must be paired with ring_buffer_read_finish.
5053 struct ring_buffer_iter *
5054 ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
5056 struct ring_buffer_per_cpu *cpu_buffer;
5057 struct ring_buffer_iter *iter;
5059 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5062 iter = kzalloc(sizeof(*iter), flags);
5066 /* Holds the entire event: data and meta data */
5067 iter->event_size = buffer->subbuf_size;
5068 iter->event = kmalloc(iter->event_size, flags);
5074 cpu_buffer = buffer->buffers[cpu];
5076 iter->cpu_buffer = cpu_buffer;
5078 atomic_inc(&cpu_buffer->resize_disabled);
5082 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
5085 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
5087 * All previously invoked ring_buffer_read_prepare calls to prepare
5088 * iterators will be synchronized. Afterwards, read_buffer_read_start
5089 * calls on those iterators are allowed.
5092 ring_buffer_read_prepare_sync(void)
5096 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
5099 * ring_buffer_read_start - start a non consuming read of the buffer
5100 * @iter: The iterator returned by ring_buffer_read_prepare
5102 * This finalizes the startup of an iteration through the buffer.
5103 * The iterator comes from a call to ring_buffer_read_prepare and
5104 * an intervening ring_buffer_read_prepare_sync must have been
5107 * Must be paired with ring_buffer_read_finish.
5110 ring_buffer_read_start(struct ring_buffer_iter *iter)
5112 struct ring_buffer_per_cpu *cpu_buffer;
5113 unsigned long flags;
5118 cpu_buffer = iter->cpu_buffer;
5120 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5121 arch_spin_lock(&cpu_buffer->lock);
5122 rb_iter_reset(iter);
5123 arch_spin_unlock(&cpu_buffer->lock);
5124 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5126 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5129 * ring_buffer_read_finish - finish reading the iterator of the buffer
5130 * @iter: The iterator retrieved by ring_buffer_start
5132 * This re-enables the recording to the buffer, and frees the
5136 ring_buffer_read_finish(struct ring_buffer_iter *iter)
5138 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5139 unsigned long flags;
5142 * Ring buffer is disabled from recording, here's a good place
5143 * to check the integrity of the ring buffer.
5144 * Must prevent readers from trying to read, as the check
5145 * clears the HEAD page and readers require it.
5147 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5148 rb_check_pages(cpu_buffer);
5149 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5151 atomic_dec(&cpu_buffer->resize_disabled);
5155 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5158 * ring_buffer_iter_advance - advance the iterator to the next location
5159 * @iter: The ring buffer iterator
5161 * Move the location of the iterator such that the next read will
5162 * be the next location of the iterator.
5164 void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5166 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5167 unsigned long flags;
5169 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5171 rb_advance_iter(iter);
5173 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5175 EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5178 * ring_buffer_size - return the size of the ring buffer (in bytes)
5179 * @buffer: The ring buffer.
5180 * @cpu: The CPU to get ring buffer size from.
5182 unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5184 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5187 return buffer->subbuf_size * buffer->buffers[cpu]->nr_pages;
5189 EXPORT_SYMBOL_GPL(ring_buffer_size);
5192 * ring_buffer_max_event_size - return the max data size of an event
5193 * @buffer: The ring buffer.
5195 * Returns the maximum size an event can be.
5197 unsigned long ring_buffer_max_event_size(struct trace_buffer *buffer)
5199 /* If abs timestamp is requested, events have a timestamp too */
5200 if (ring_buffer_time_stamp_abs(buffer))
5201 return buffer->max_data_size - RB_LEN_TIME_EXTEND;
5202 return buffer->max_data_size;
5204 EXPORT_SYMBOL_GPL(ring_buffer_max_event_size);
5206 static void rb_clear_buffer_page(struct buffer_page *page)
5208 local_set(&page->write, 0);
5209 local_set(&page->entries, 0);
5210 rb_init_page(page->page);
5215 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5217 struct buffer_page *page;
5219 rb_head_page_deactivate(cpu_buffer);
5221 cpu_buffer->head_page
5222 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5223 rb_clear_buffer_page(cpu_buffer->head_page);
5224 list_for_each_entry(page, cpu_buffer->pages, list) {
5225 rb_clear_buffer_page(page);
5228 cpu_buffer->tail_page = cpu_buffer->head_page;
5229 cpu_buffer->commit_page = cpu_buffer->head_page;
5231 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5232 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5233 rb_clear_buffer_page(cpu_buffer->reader_page);
5235 local_set(&cpu_buffer->entries_bytes, 0);
5236 local_set(&cpu_buffer->overrun, 0);
5237 local_set(&cpu_buffer->commit_overrun, 0);
5238 local_set(&cpu_buffer->dropped_events, 0);
5239 local_set(&cpu_buffer->entries, 0);
5240 local_set(&cpu_buffer->committing, 0);
5241 local_set(&cpu_buffer->commits, 0);
5242 local_set(&cpu_buffer->pages_touched, 0);
5243 local_set(&cpu_buffer->pages_lost, 0);
5244 local_set(&cpu_buffer->pages_read, 0);
5245 cpu_buffer->last_pages_touch = 0;
5246 cpu_buffer->shortest_full = 0;
5247 cpu_buffer->read = 0;
5248 cpu_buffer->read_bytes = 0;
5250 rb_time_set(&cpu_buffer->write_stamp, 0);
5251 rb_time_set(&cpu_buffer->before_stamp, 0);
5253 memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5255 cpu_buffer->lost_events = 0;
5256 cpu_buffer->last_overrun = 0;
5258 rb_head_page_activate(cpu_buffer);
5259 cpu_buffer->pages_removed = 0;
5262 /* Must have disabled the cpu buffer then done a synchronize_rcu */
5263 static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5265 unsigned long flags;
5267 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5269 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5272 arch_spin_lock(&cpu_buffer->lock);
5274 rb_reset_cpu(cpu_buffer);
5276 arch_spin_unlock(&cpu_buffer->lock);
5279 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5283 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5284 * @buffer: The ring buffer to reset a per cpu buffer of
5285 * @cpu: The CPU buffer to be reset
5287 void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5289 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5291 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5294 /* prevent another thread from changing buffer sizes */
5295 mutex_lock(&buffer->mutex);
5297 atomic_inc(&cpu_buffer->resize_disabled);
5298 atomic_inc(&cpu_buffer->record_disabled);
5300 /* Make sure all commits have finished */
5303 reset_disabled_cpu_buffer(cpu_buffer);
5305 atomic_dec(&cpu_buffer->record_disabled);
5306 atomic_dec(&cpu_buffer->resize_disabled);
5308 mutex_unlock(&buffer->mutex);
5310 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5312 /* Flag to ensure proper resetting of atomic variables */
5313 #define RESET_BIT (1 << 30)
5316 * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
5317 * @buffer: The ring buffer to reset a per cpu buffer of
5319 void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5321 struct ring_buffer_per_cpu *cpu_buffer;
5324 /* prevent another thread from changing buffer sizes */
5325 mutex_lock(&buffer->mutex);
5327 for_each_online_buffer_cpu(buffer, cpu) {
5328 cpu_buffer = buffer->buffers[cpu];
5330 atomic_add(RESET_BIT, &cpu_buffer->resize_disabled);
5331 atomic_inc(&cpu_buffer->record_disabled);
5334 /* Make sure all commits have finished */
5337 for_each_buffer_cpu(buffer, cpu) {
5338 cpu_buffer = buffer->buffers[cpu];
5341 * If a CPU came online during the synchronize_rcu(), then
5344 if (!(atomic_read(&cpu_buffer->resize_disabled) & RESET_BIT))
5347 reset_disabled_cpu_buffer(cpu_buffer);
5349 atomic_dec(&cpu_buffer->record_disabled);
5350 atomic_sub(RESET_BIT, &cpu_buffer->resize_disabled);
5353 mutex_unlock(&buffer->mutex);
5357 * ring_buffer_reset - reset a ring buffer
5358 * @buffer: The ring buffer to reset all cpu buffers
5360 void ring_buffer_reset(struct trace_buffer *buffer)
5362 struct ring_buffer_per_cpu *cpu_buffer;
5365 /* prevent another thread from changing buffer sizes */
5366 mutex_lock(&buffer->mutex);
5368 for_each_buffer_cpu(buffer, cpu) {
5369 cpu_buffer = buffer->buffers[cpu];
5371 atomic_inc(&cpu_buffer->resize_disabled);
5372 atomic_inc(&cpu_buffer->record_disabled);
5375 /* Make sure all commits have finished */
5378 for_each_buffer_cpu(buffer, cpu) {
5379 cpu_buffer = buffer->buffers[cpu];
5381 reset_disabled_cpu_buffer(cpu_buffer);
5383 atomic_dec(&cpu_buffer->record_disabled);
5384 atomic_dec(&cpu_buffer->resize_disabled);
5387 mutex_unlock(&buffer->mutex);
5389 EXPORT_SYMBOL_GPL(ring_buffer_reset);
5392 * ring_buffer_empty - is the ring buffer empty?
5393 * @buffer: The ring buffer to test
5395 bool ring_buffer_empty(struct trace_buffer *buffer)
5397 struct ring_buffer_per_cpu *cpu_buffer;
5398 unsigned long flags;
5403 /* yes this is racy, but if you don't like the race, lock the buffer */
5404 for_each_buffer_cpu(buffer, cpu) {
5405 cpu_buffer = buffer->buffers[cpu];
5406 local_irq_save(flags);
5407 dolock = rb_reader_lock(cpu_buffer);
5408 ret = rb_per_cpu_empty(cpu_buffer);
5409 rb_reader_unlock(cpu_buffer, dolock);
5410 local_irq_restore(flags);
5418 EXPORT_SYMBOL_GPL(ring_buffer_empty);
5421 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5422 * @buffer: The ring buffer
5423 * @cpu: The CPU buffer to test
5425 bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5427 struct ring_buffer_per_cpu *cpu_buffer;
5428 unsigned long flags;
5432 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5435 cpu_buffer = buffer->buffers[cpu];
5436 local_irq_save(flags);
5437 dolock = rb_reader_lock(cpu_buffer);
5438 ret = rb_per_cpu_empty(cpu_buffer);
5439 rb_reader_unlock(cpu_buffer, dolock);
5440 local_irq_restore(flags);
5444 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5446 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5448 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5449 * @buffer_a: One buffer to swap with
5450 * @buffer_b: The other buffer to swap with
5451 * @cpu: the CPU of the buffers to swap
5453 * This function is useful for tracers that want to take a "snapshot"
5454 * of a CPU buffer and has another back up buffer lying around.
5455 * it is expected that the tracer handles the cpu buffer not being
5456 * used at the moment.
5458 int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5459 struct trace_buffer *buffer_b, int cpu)
5461 struct ring_buffer_per_cpu *cpu_buffer_a;
5462 struct ring_buffer_per_cpu *cpu_buffer_b;
5465 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5466 !cpumask_test_cpu(cpu, buffer_b->cpumask))
5469 cpu_buffer_a = buffer_a->buffers[cpu];
5470 cpu_buffer_b = buffer_b->buffers[cpu];
5472 /* At least make sure the two buffers are somewhat the same */
5473 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5476 if (buffer_a->subbuf_order != buffer_b->subbuf_order)
5481 if (atomic_read(&buffer_a->record_disabled))
5484 if (atomic_read(&buffer_b->record_disabled))
5487 if (atomic_read(&cpu_buffer_a->record_disabled))
5490 if (atomic_read(&cpu_buffer_b->record_disabled))
5494 * We can't do a synchronize_rcu here because this
5495 * function can be called in atomic context.
5496 * Normally this will be called from the same CPU as cpu.
5497 * If not it's up to the caller to protect this.
5499 atomic_inc(&cpu_buffer_a->record_disabled);
5500 atomic_inc(&cpu_buffer_b->record_disabled);
5503 if (local_read(&cpu_buffer_a->committing))
5505 if (local_read(&cpu_buffer_b->committing))
5509 * When resize is in progress, we cannot swap it because
5510 * it will mess the state of the cpu buffer.
5512 if (atomic_read(&buffer_a->resizing))
5514 if (atomic_read(&buffer_b->resizing))
5517 buffer_a->buffers[cpu] = cpu_buffer_b;
5518 buffer_b->buffers[cpu] = cpu_buffer_a;
5520 cpu_buffer_b->buffer = buffer_a;
5521 cpu_buffer_a->buffer = buffer_b;
5526 atomic_dec(&cpu_buffer_a->record_disabled);
5527 atomic_dec(&cpu_buffer_b->record_disabled);
5531 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5532 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5535 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5536 * @buffer: the buffer to allocate for.
5537 * @cpu: the cpu buffer to allocate.
5539 * This function is used in conjunction with ring_buffer_read_page.
5540 * When reading a full page from the ring buffer, these functions
5541 * can be used to speed up the process. The calling function should
5542 * allocate a few pages first with this function. Then when it
5543 * needs to get pages from the ring buffer, it passes the result
5544 * of this function into ring_buffer_read_page, which will swap
5545 * the page that was allocated, with the read page of the buffer.
5548 * The page allocated, or ERR_PTR
5550 struct buffer_data_read_page *
5551 ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5553 struct ring_buffer_per_cpu *cpu_buffer;
5554 struct buffer_data_read_page *bpage = NULL;
5555 unsigned long flags;
5558 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5559 return ERR_PTR(-ENODEV);
5561 bpage = kzalloc(sizeof(*bpage), GFP_KERNEL);
5563 return ERR_PTR(-ENOMEM);
5565 bpage->order = buffer->subbuf_order;
5566 cpu_buffer = buffer->buffers[cpu];
5567 local_irq_save(flags);
5568 arch_spin_lock(&cpu_buffer->lock);
5570 if (cpu_buffer->free_page) {
5571 bpage->data = cpu_buffer->free_page;
5572 cpu_buffer->free_page = NULL;
5575 arch_spin_unlock(&cpu_buffer->lock);
5576 local_irq_restore(flags);
5581 page = alloc_pages_node(cpu_to_node(cpu),
5582 GFP_KERNEL | __GFP_NORETRY | __GFP_ZERO,
5583 cpu_buffer->buffer->subbuf_order);
5586 return ERR_PTR(-ENOMEM);
5589 bpage->data = page_address(page);
5592 rb_init_page(bpage->data);
5596 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5599 * ring_buffer_free_read_page - free an allocated read page
5600 * @buffer: the buffer the page was allocate for
5601 * @cpu: the cpu buffer the page came from
5602 * @data_page: the page to free
5604 * Free a page allocated from ring_buffer_alloc_read_page.
5606 void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu,
5607 struct buffer_data_read_page *data_page)
5609 struct ring_buffer_per_cpu *cpu_buffer;
5610 struct buffer_data_page *bpage = data_page->data;
5611 struct page *page = virt_to_page(bpage);
5612 unsigned long flags;
5614 if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
5617 cpu_buffer = buffer->buffers[cpu];
5620 * If the page is still in use someplace else, or order of the page
5621 * is different from the subbuffer order of the buffer -
5624 if (page_ref_count(page) > 1 || data_page->order != buffer->subbuf_order)
5627 local_irq_save(flags);
5628 arch_spin_lock(&cpu_buffer->lock);
5630 if (!cpu_buffer->free_page) {
5631 cpu_buffer->free_page = bpage;
5635 arch_spin_unlock(&cpu_buffer->lock);
5636 local_irq_restore(flags);
5639 free_pages((unsigned long)bpage, data_page->order);
5642 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5645 * ring_buffer_read_page - extract a page from the ring buffer
5646 * @buffer: buffer to extract from
5647 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5648 * @len: amount to extract
5649 * @cpu: the cpu of the buffer to extract
5650 * @full: should the extraction only happen when the page is full.
5652 * This function will pull out a page from the ring buffer and consume it.
5653 * @data_page must be the address of the variable that was returned
5654 * from ring_buffer_alloc_read_page. This is because the page might be used
5655 * to swap with a page in the ring buffer.
5658 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
5659 * if (IS_ERR(rpage))
5660 * return PTR_ERR(rpage);
5661 * ret = ring_buffer_read_page(buffer, rpage, len, cpu, 0);
5663 * process_page(ring_buffer_read_page_data(rpage), ret);
5664 * ring_buffer_free_read_page(buffer, cpu, rpage);
5666 * When @full is set, the function will not return true unless
5667 * the writer is off the reader page.
5669 * Note: it is up to the calling functions to handle sleeps and wakeups.
5670 * The ring buffer can be used anywhere in the kernel and can not
5671 * blindly call wake_up. The layer that uses the ring buffer must be
5672 * responsible for that.
5675 * >=0 if data has been transferred, returns the offset of consumed data.
5676 * <0 if no data has been transferred.
5678 int ring_buffer_read_page(struct trace_buffer *buffer,
5679 struct buffer_data_read_page *data_page,
5680 size_t len, int cpu, int full)
5682 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5683 struct ring_buffer_event *event;
5684 struct buffer_data_page *bpage;
5685 struct buffer_page *reader;
5686 unsigned long missed_events;
5687 unsigned long flags;
5688 unsigned int commit;
5693 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5697 * If len is not big enough to hold the page header, then
5698 * we can not copy anything.
5700 if (len <= BUF_PAGE_HDR_SIZE)
5703 len -= BUF_PAGE_HDR_SIZE;
5705 if (!data_page || !data_page->data)
5707 if (data_page->order != buffer->subbuf_order)
5710 bpage = data_page->data;
5714 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5716 reader = rb_get_reader_page(cpu_buffer);
5720 event = rb_reader_event(cpu_buffer);
5722 read = reader->read;
5723 commit = rb_page_commit(reader);
5725 /* Check if any events were dropped */
5726 missed_events = cpu_buffer->lost_events;
5729 * If this page has been partially read or
5730 * if len is not big enough to read the rest of the page or
5731 * a writer is still on the page, then
5732 * we must copy the data from the page to the buffer.
5733 * Otherwise, we can simply swap the page with the one passed in.
5735 if (read || (len < (commit - read)) ||
5736 cpu_buffer->reader_page == cpu_buffer->commit_page) {
5737 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5738 unsigned int rpos = read;
5739 unsigned int pos = 0;
5743 * If a full page is expected, this can still be returned
5744 * if there's been a previous partial read and the
5745 * rest of the page can be read and the commit page is off
5749 (!read || (len < (commit - read)) ||
5750 cpu_buffer->reader_page == cpu_buffer->commit_page))
5753 if (len > (commit - read))
5754 len = (commit - read);
5756 /* Always keep the time extend and data together */
5757 size = rb_event_ts_length(event);
5762 /* save the current timestamp, since the user will need it */
5763 save_timestamp = cpu_buffer->read_stamp;
5765 /* Need to copy one event at a time */
5767 /* We need the size of one event, because
5768 * rb_advance_reader only advances by one event,
5769 * whereas rb_event_ts_length may include the size of
5770 * one or two events.
5771 * We have already ensured there's enough space if this
5772 * is a time extend. */
5773 size = rb_event_length(event);
5774 memcpy(bpage->data + pos, rpage->data + rpos, size);
5778 rb_advance_reader(cpu_buffer);
5779 rpos = reader->read;
5785 event = rb_reader_event(cpu_buffer);
5786 /* Always keep the time extend and data together */
5787 size = rb_event_ts_length(event);
5788 } while (len >= size);
5791 local_set(&bpage->commit, pos);
5792 bpage->time_stamp = save_timestamp;
5794 /* we copied everything to the beginning */
5797 /* update the entry counter */
5798 cpu_buffer->read += rb_page_entries(reader);
5799 cpu_buffer->read_bytes += rb_page_commit(reader);
5801 /* swap the pages */
5802 rb_init_page(bpage);
5803 bpage = reader->page;
5804 reader->page = data_page->data;
5805 local_set(&reader->write, 0);
5806 local_set(&reader->entries, 0);
5808 data_page->data = bpage;
5811 * Use the real_end for the data size,
5812 * This gives us a chance to store the lost events
5815 if (reader->real_end)
5816 local_set(&bpage->commit, reader->real_end);
5820 cpu_buffer->lost_events = 0;
5822 commit = local_read(&bpage->commit);
5824 * Set a flag in the commit field if we lost events
5826 if (missed_events) {
5827 /* If there is room at the end of the page to save the
5828 * missed events, then record it there.
5830 if (buffer->subbuf_size - commit >= sizeof(missed_events)) {
5831 memcpy(&bpage->data[commit], &missed_events,
5832 sizeof(missed_events));
5833 local_add(RB_MISSED_STORED, &bpage->commit);
5834 commit += sizeof(missed_events);
5836 local_add(RB_MISSED_EVENTS, &bpage->commit);
5840 * This page may be off to user land. Zero it out here.
5842 if (commit < buffer->subbuf_size)
5843 memset(&bpage->data[commit], 0, buffer->subbuf_size - commit);
5846 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5851 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5854 * ring_buffer_read_page_data - get pointer to the data in the page.
5855 * @page: the page to get the data from
5857 * Returns pointer to the actual data in this page.
5859 void *ring_buffer_read_page_data(struct buffer_data_read_page *page)
5863 EXPORT_SYMBOL_GPL(ring_buffer_read_page_data);
5866 * ring_buffer_subbuf_size_get - get size of the sub buffer.
5867 * @buffer: the buffer to get the sub buffer size from
5869 * Returns size of the sub buffer, in bytes.
5871 int ring_buffer_subbuf_size_get(struct trace_buffer *buffer)
5873 return buffer->subbuf_size + BUF_PAGE_HDR_SIZE;
5875 EXPORT_SYMBOL_GPL(ring_buffer_subbuf_size_get);
5878 * ring_buffer_subbuf_order_get - get order of system sub pages in one buffer page.
5879 * @buffer: The ring_buffer to get the system sub page order from
5881 * By default, one ring buffer sub page equals to one system page. This parameter
5882 * is configurable, per ring buffer. The size of the ring buffer sub page can be
5883 * extended, but must be an order of system page size.
5885 * Returns the order of buffer sub page size, in system pages:
5886 * 0 means the sub buffer size is 1 system page and so forth.
5887 * In case of an error < 0 is returned.
5889 int ring_buffer_subbuf_order_get(struct trace_buffer *buffer)
5894 return buffer->subbuf_order;
5896 EXPORT_SYMBOL_GPL(ring_buffer_subbuf_order_get);
5899 * ring_buffer_subbuf_order_set - set the size of ring buffer sub page.
5900 * @buffer: The ring_buffer to set the new page size.
5901 * @order: Order of the system pages in one sub buffer page
5903 * By default, one ring buffer pages equals to one system page. This API can be
5904 * used to set new size of the ring buffer page. The size must be order of
5905 * system page size, that's why the input parameter @order is the order of
5906 * system pages that are allocated for one ring buffer page:
5908 * 1 - 2 system pages
5909 * 3 - 4 system pages
5912 * Returns 0 on success or < 0 in case of an error.
5914 int ring_buffer_subbuf_order_set(struct trace_buffer *buffer, int order)
5916 struct ring_buffer_per_cpu *cpu_buffer;
5917 struct buffer_page *bpage, *tmp;
5918 int old_order, old_size;
5924 if (!buffer || order < 0)
5927 if (buffer->subbuf_order == order)
5930 psize = (1 << order) * PAGE_SIZE;
5931 if (psize <= BUF_PAGE_HDR_SIZE)
5934 /* Size of a subbuf cannot be greater than the write counter */
5935 if (psize > RB_WRITE_MASK + 1)
5938 old_order = buffer->subbuf_order;
5939 old_size = buffer->subbuf_size;
5941 /* prevent another thread from changing buffer sizes */
5942 mutex_lock(&buffer->mutex);
5943 atomic_inc(&buffer->record_disabled);
5945 /* Make sure all commits have finished */
5948 buffer->subbuf_order = order;
5949 buffer->subbuf_size = psize - BUF_PAGE_HDR_SIZE;
5951 /* Make sure all new buffers are allocated, before deleting the old ones */
5952 for_each_buffer_cpu(buffer, cpu) {
5954 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5957 cpu_buffer = buffer->buffers[cpu];
5959 /* Update the number of pages to match the new size */
5960 nr_pages = old_size * buffer->buffers[cpu]->nr_pages;
5961 nr_pages = DIV_ROUND_UP(nr_pages, buffer->subbuf_size);
5963 /* we need a minimum of two pages */
5967 cpu_buffer->nr_pages_to_update = nr_pages;
5969 /* Include the reader page */
5972 /* Allocate the new size buffer */
5973 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5974 if (__rb_allocate_pages(cpu_buffer, nr_pages,
5975 &cpu_buffer->new_pages)) {
5976 /* not enough memory for new pages */
5982 for_each_buffer_cpu(buffer, cpu) {
5984 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5987 cpu_buffer = buffer->buffers[cpu];
5989 /* Clear the head bit to make the link list normal to read */
5990 rb_head_page_deactivate(cpu_buffer);
5992 /* Now walk the list and free all the old sub buffers */
5993 list_for_each_entry_safe(bpage, tmp, cpu_buffer->pages, list) {
5994 list_del_init(&bpage->list);
5995 free_buffer_page(bpage);
5997 /* The above loop stopped an the last page needing to be freed */
5998 bpage = list_entry(cpu_buffer->pages, struct buffer_page, list);
5999 free_buffer_page(bpage);
6001 /* Free the current reader page */
6002 free_buffer_page(cpu_buffer->reader_page);
6004 /* One page was allocated for the reader page */
6005 cpu_buffer->reader_page = list_entry(cpu_buffer->new_pages.next,
6006 struct buffer_page, list);
6007 list_del_init(&cpu_buffer->reader_page->list);
6009 /* The cpu_buffer pages are a link list with no head */
6010 cpu_buffer->pages = cpu_buffer->new_pages.next;
6011 cpu_buffer->new_pages.next->prev = cpu_buffer->new_pages.prev;
6012 cpu_buffer->new_pages.prev->next = cpu_buffer->new_pages.next;
6014 /* Clear the new_pages list */
6015 INIT_LIST_HEAD(&cpu_buffer->new_pages);
6017 cpu_buffer->head_page
6018 = list_entry(cpu_buffer->pages, struct buffer_page, list);
6019 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
6021 cpu_buffer->nr_pages = cpu_buffer->nr_pages_to_update;
6022 cpu_buffer->nr_pages_to_update = 0;
6024 free_pages((unsigned long)cpu_buffer->free_page, old_order);
6025 cpu_buffer->free_page = NULL;
6027 rb_head_page_activate(cpu_buffer);
6029 rb_check_pages(cpu_buffer);
6032 atomic_dec(&buffer->record_disabled);
6033 mutex_unlock(&buffer->mutex);
6038 buffer->subbuf_order = old_order;
6039 buffer->subbuf_size = old_size;
6041 atomic_dec(&buffer->record_disabled);
6042 mutex_unlock(&buffer->mutex);
6044 for_each_buffer_cpu(buffer, cpu) {
6045 cpu_buffer = buffer->buffers[cpu];
6047 if (!cpu_buffer->nr_pages_to_update)
6050 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages, list) {
6051 list_del_init(&bpage->list);
6052 free_buffer_page(bpage);
6058 EXPORT_SYMBOL_GPL(ring_buffer_subbuf_order_set);
6061 * We only allocate new buffers, never free them if the CPU goes down.
6062 * If we were to free the buffer, then the user would lose any trace that was in
6065 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
6067 struct trace_buffer *buffer;
6070 unsigned long nr_pages;
6072 buffer = container_of(node, struct trace_buffer, node);
6073 if (cpumask_test_cpu(cpu, buffer->cpumask))
6078 /* check if all cpu sizes are same */
6079 for_each_buffer_cpu(buffer, cpu_i) {
6080 /* fill in the size from first enabled cpu */
6082 nr_pages = buffer->buffers[cpu_i]->nr_pages;
6083 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
6088 /* allocate minimum pages, user can later expand it */
6091 buffer->buffers[cpu] =
6092 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
6093 if (!buffer->buffers[cpu]) {
6094 WARN(1, "failed to allocate ring buffer on CPU %u\n",
6099 cpumask_set_cpu(cpu, buffer->cpumask);
6103 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
6105 * This is a basic integrity check of the ring buffer.
6106 * Late in the boot cycle this test will run when configured in.
6107 * It will kick off a thread per CPU that will go into a loop
6108 * writing to the per cpu ring buffer various sizes of data.
6109 * Some of the data will be large items, some small.
6111 * Another thread is created that goes into a spin, sending out
6112 * IPIs to the other CPUs to also write into the ring buffer.
6113 * this is to test the nesting ability of the buffer.
6115 * Basic stats are recorded and reported. If something in the
6116 * ring buffer should happen that's not expected, a big warning
6117 * is displayed and all ring buffers are disabled.
6119 static struct task_struct *rb_threads[NR_CPUS] __initdata;
6121 struct rb_test_data {
6122 struct trace_buffer *buffer;
6123 unsigned long events;
6124 unsigned long bytes_written;
6125 unsigned long bytes_alloc;
6126 unsigned long bytes_dropped;
6127 unsigned long events_nested;
6128 unsigned long bytes_written_nested;
6129 unsigned long bytes_alloc_nested;
6130 unsigned long bytes_dropped_nested;
6131 int min_size_nested;
6132 int max_size_nested;
6139 static struct rb_test_data rb_data[NR_CPUS] __initdata;
6142 #define RB_TEST_BUFFER_SIZE 1048576
6144 static char rb_string[] __initdata =
6145 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
6146 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
6147 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
6149 static bool rb_test_started __initdata;
6156 static __init int rb_write_something(struct rb_test_data *data, bool nested)
6158 struct ring_buffer_event *event;
6159 struct rb_item *item;
6166 /* Have nested writes different that what is written */
6167 cnt = data->cnt + (nested ? 27 : 0);
6169 /* Multiply cnt by ~e, to make some unique increment */
6170 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
6172 len = size + sizeof(struct rb_item);
6174 started = rb_test_started;
6175 /* read rb_test_started before checking buffer enabled */
6178 event = ring_buffer_lock_reserve(data->buffer, len);
6180 /* Ignore dropped events before test starts. */
6183 data->bytes_dropped += len;
6185 data->bytes_dropped_nested += len;
6190 event_len = ring_buffer_event_length(event);
6192 if (RB_WARN_ON(data->buffer, event_len < len))
6195 item = ring_buffer_event_data(event);
6197 memcpy(item->str, rb_string, size);
6200 data->bytes_alloc_nested += event_len;
6201 data->bytes_written_nested += len;
6202 data->events_nested++;
6203 if (!data->min_size_nested || len < data->min_size_nested)
6204 data->min_size_nested = len;
6205 if (len > data->max_size_nested)
6206 data->max_size_nested = len;
6208 data->bytes_alloc += event_len;
6209 data->bytes_written += len;
6211 if (!data->min_size || len < data->min_size)
6212 data->max_size = len;
6213 if (len > data->max_size)
6214 data->max_size = len;
6218 ring_buffer_unlock_commit(data->buffer);
6223 static __init int rb_test(void *arg)
6225 struct rb_test_data *data = arg;
6227 while (!kthread_should_stop()) {
6228 rb_write_something(data, false);
6231 set_current_state(TASK_INTERRUPTIBLE);
6232 /* Now sleep between a min of 100-300us and a max of 1ms */
6233 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
6239 static __init void rb_ipi(void *ignore)
6241 struct rb_test_data *data;
6242 int cpu = smp_processor_id();
6244 data = &rb_data[cpu];
6245 rb_write_something(data, true);
6248 static __init int rb_hammer_test(void *arg)
6250 while (!kthread_should_stop()) {
6252 /* Send an IPI to all cpus to write data! */
6253 smp_call_function(rb_ipi, NULL, 1);
6254 /* No sleep, but for non preempt, let others run */
6261 static __init int test_ringbuffer(void)
6263 struct task_struct *rb_hammer;
6264 struct trace_buffer *buffer;
6268 if (security_locked_down(LOCKDOWN_TRACEFS)) {
6269 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
6273 pr_info("Running ring buffer tests...\n");
6275 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
6276 if (WARN_ON(!buffer))
6279 /* Disable buffer so that threads can't write to it yet */
6280 ring_buffer_record_off(buffer);
6282 for_each_online_cpu(cpu) {
6283 rb_data[cpu].buffer = buffer;
6284 rb_data[cpu].cpu = cpu;
6285 rb_data[cpu].cnt = cpu;
6286 rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
6287 cpu, "rbtester/%u");
6288 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
6289 pr_cont("FAILED\n");
6290 ret = PTR_ERR(rb_threads[cpu]);
6295 /* Now create the rb hammer! */
6296 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
6297 if (WARN_ON(IS_ERR(rb_hammer))) {
6298 pr_cont("FAILED\n");
6299 ret = PTR_ERR(rb_hammer);
6303 ring_buffer_record_on(buffer);
6305 * Show buffer is enabled before setting rb_test_started.
6306 * Yes there's a small race window where events could be
6307 * dropped and the thread wont catch it. But when a ring
6308 * buffer gets enabled, there will always be some kind of
6309 * delay before other CPUs see it. Thus, we don't care about
6310 * those dropped events. We care about events dropped after
6311 * the threads see that the buffer is active.
6314 rb_test_started = true;
6316 set_current_state(TASK_INTERRUPTIBLE);
6317 /* Just run for 10 seconds */;
6318 schedule_timeout(10 * HZ);
6320 kthread_stop(rb_hammer);
6323 for_each_online_cpu(cpu) {
6324 if (!rb_threads[cpu])
6326 kthread_stop(rb_threads[cpu]);
6329 ring_buffer_free(buffer);
6334 pr_info("finished\n");
6335 for_each_online_cpu(cpu) {
6336 struct ring_buffer_event *event;
6337 struct rb_test_data *data = &rb_data[cpu];
6338 struct rb_item *item;
6339 unsigned long total_events;
6340 unsigned long total_dropped;
6341 unsigned long total_written;
6342 unsigned long total_alloc;
6343 unsigned long total_read = 0;
6344 unsigned long total_size = 0;
6345 unsigned long total_len = 0;
6346 unsigned long total_lost = 0;
6349 int small_event_size;
6353 total_events = data->events + data->events_nested;
6354 total_written = data->bytes_written + data->bytes_written_nested;
6355 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
6356 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
6358 big_event_size = data->max_size + data->max_size_nested;
6359 small_event_size = data->min_size + data->min_size_nested;
6361 pr_info("CPU %d:\n", cpu);
6362 pr_info(" events: %ld\n", total_events);
6363 pr_info(" dropped bytes: %ld\n", total_dropped);
6364 pr_info(" alloced bytes: %ld\n", total_alloc);
6365 pr_info(" written bytes: %ld\n", total_written);
6366 pr_info(" biggest event: %d\n", big_event_size);
6367 pr_info(" smallest event: %d\n", small_event_size);
6369 if (RB_WARN_ON(buffer, total_dropped))
6374 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
6376 item = ring_buffer_event_data(event);
6377 total_len += ring_buffer_event_length(event);
6378 total_size += item->size + sizeof(struct rb_item);
6379 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
6380 pr_info("FAILED!\n");
6381 pr_info("buffer had: %.*s\n", item->size, item->str);
6382 pr_info("expected: %.*s\n", item->size, rb_string);
6383 RB_WARN_ON(buffer, 1);
6394 pr_info(" read events: %ld\n", total_read);
6395 pr_info(" lost events: %ld\n", total_lost);
6396 pr_info(" total events: %ld\n", total_lost + total_read);
6397 pr_info(" recorded len bytes: %ld\n", total_len);
6398 pr_info(" recorded size bytes: %ld\n", total_size);
6400 pr_info(" With dropped events, record len and size may not match\n"
6401 " alloced and written from above\n");
6403 if (RB_WARN_ON(buffer, total_len != total_alloc ||
6404 total_size != total_written))
6407 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
6413 pr_info("Ring buffer PASSED!\n");
6415 ring_buffer_free(buffer);
6419 late_initcall(test_ringbuffer);
6420 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */