4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/trace_events.h>
7 #include <linux/ring_buffer.h>
8 #include <linux/trace_clock.h>
9 #include <linux/sched/clock.h>
10 #include <linux/trace_seq.h>
11 #include <linux/spinlock.h>
12 #include <linux/irq_work.h>
13 #include <linux/uaccess.h>
14 #include <linux/hardirq.h>
15 #include <linux/kthread.h> /* for self test */
16 #include <linux/module.h>
17 #include <linux/percpu.h>
18 #include <linux/mutex.h>
19 #include <linux/delay.h>
20 #include <linux/slab.h>
21 #include <linux/init.h>
22 #include <linux/hash.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
26 #include <asm/local.h>
28 static void update_pages_handler(struct work_struct *work);
31 * The ring buffer header is special. We must manually up keep it.
33 int ring_buffer_print_entry_header(struct trace_seq *s)
35 trace_seq_puts(s, "# compressed entry header\n");
36 trace_seq_puts(s, "\ttype_len : 5 bits\n");
37 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
38 trace_seq_puts(s, "\tarray : 32 bits\n");
39 trace_seq_putc(s, '\n');
40 trace_seq_printf(s, "\tpadding : type == %d\n",
41 RINGBUF_TYPE_PADDING);
42 trace_seq_printf(s, "\ttime_extend : type == %d\n",
43 RINGBUF_TYPE_TIME_EXTEND);
44 trace_seq_printf(s, "\tdata max type_len == %d\n",
45 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
47 return !trace_seq_has_overflowed(s);
51 * The ring buffer is made up of a list of pages. A separate list of pages is
52 * allocated for each CPU. A writer may only write to a buffer that is
53 * associated with the CPU it is currently executing on. A reader may read
54 * from any per cpu buffer.
56 * The reader is special. For each per cpu buffer, the reader has its own
57 * reader page. When a reader has read the entire reader page, this reader
58 * page is swapped with another page in the ring buffer.
60 * Now, as long as the writer is off the reader page, the reader can do what
61 * ever it wants with that page. The writer will never write to that page
62 * again (as long as it is out of the ring buffer).
64 * Here's some silly ASCII art.
67 * |reader| RING BUFFER
69 * +------+ +---+ +---+ +---+
78 * |reader| RING BUFFER
79 * |page |------------------v
80 * +------+ +---+ +---+ +---+
89 * |reader| RING BUFFER
90 * |page |------------------v
91 * +------+ +---+ +---+ +---+
96 * +------------------------------+
100 * |buffer| RING BUFFER
101 * |page |------------------v
102 * +------+ +---+ +---+ +---+
104 * | New +---+ +---+ +---+
107 * +------------------------------+
110 * After we make this swap, the reader can hand this page off to the splice
111 * code and be done with it. It can even allocate a new page if it needs to
112 * and swap that into the ring buffer.
114 * We will be using cmpxchg soon to make all this lockless.
118 /* Used for individual buffers (after the counter) */
119 #define RB_BUFFER_OFF (1 << 20)
121 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
123 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
124 #define RB_ALIGNMENT 4U
125 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
126 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
128 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
129 # define RB_FORCE_8BYTE_ALIGNMENT 0
130 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
132 # define RB_FORCE_8BYTE_ALIGNMENT 1
133 # define RB_ARCH_ALIGNMENT 8U
136 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
138 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
139 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
142 RB_LEN_TIME_EXTEND = 8,
143 RB_LEN_TIME_STAMP = 16,
146 #define skip_time_extend(event) \
147 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
149 static inline int rb_null_event(struct ring_buffer_event *event)
151 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
154 static void rb_event_set_padding(struct ring_buffer_event *event)
156 /* padding has a NULL time_delta */
157 event->type_len = RINGBUF_TYPE_PADDING;
158 event->time_delta = 0;
162 rb_event_data_length(struct ring_buffer_event *event)
167 length = event->type_len * RB_ALIGNMENT;
169 length = event->array[0];
170 return length + RB_EVNT_HDR_SIZE;
174 * Return the length of the given event. Will return
175 * the length of the time extend if the event is a
178 static inline unsigned
179 rb_event_length(struct ring_buffer_event *event)
181 switch (event->type_len) {
182 case RINGBUF_TYPE_PADDING:
183 if (rb_null_event(event))
186 return event->array[0] + RB_EVNT_HDR_SIZE;
188 case RINGBUF_TYPE_TIME_EXTEND:
189 return RB_LEN_TIME_EXTEND;
191 case RINGBUF_TYPE_TIME_STAMP:
192 return RB_LEN_TIME_STAMP;
194 case RINGBUF_TYPE_DATA:
195 return rb_event_data_length(event);
204 * Return total length of time extend and data,
205 * or just the event length for all other events.
207 static inline unsigned
208 rb_event_ts_length(struct ring_buffer_event *event)
212 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
213 /* time extends include the data event after it */
214 len = RB_LEN_TIME_EXTEND;
215 event = skip_time_extend(event);
217 return len + rb_event_length(event);
221 * ring_buffer_event_length - return the length of the event
222 * @event: the event to get the length of
224 * Returns the size of the data load of a data event.
225 * If the event is something other than a data event, it
226 * returns the size of the event itself. With the exception
227 * of a TIME EXTEND, where it still returns the size of the
228 * data load of the data event after it.
230 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
234 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
235 event = skip_time_extend(event);
237 length = rb_event_length(event);
238 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
240 length -= RB_EVNT_HDR_SIZE;
241 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
242 length -= sizeof(event->array[0]);
245 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
247 /* inline for ring buffer fast paths */
248 static __always_inline void *
249 rb_event_data(struct ring_buffer_event *event)
251 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
252 event = skip_time_extend(event);
253 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
254 /* If length is in len field, then array[0] has the data */
256 return (void *)&event->array[0];
257 /* Otherwise length is in array[0] and array[1] has the data */
258 return (void *)&event->array[1];
262 * ring_buffer_event_data - return the data of the event
263 * @event: the event to get the data from
265 void *ring_buffer_event_data(struct ring_buffer_event *event)
267 return rb_event_data(event);
269 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
271 #define for_each_buffer_cpu(buffer, cpu) \
272 for_each_cpu(cpu, buffer->cpumask)
275 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
276 #define TS_DELTA_TEST (~TS_MASK)
278 /* Flag when events were overwritten */
279 #define RB_MISSED_EVENTS (1 << 31)
280 /* Missed count stored at end */
281 #define RB_MISSED_STORED (1 << 30)
283 #define RB_MISSED_FLAGS (RB_MISSED_EVENTS|RB_MISSED_STORED)
285 struct buffer_data_page {
286 u64 time_stamp; /* page time stamp */
287 local_t commit; /* write committed index */
288 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
292 * Note, the buffer_page list must be first. The buffer pages
293 * are allocated in cache lines, which means that each buffer
294 * page will be at the beginning of a cache line, and thus
295 * the least significant bits will be zero. We use this to
296 * add flags in the list struct pointers, to make the ring buffer
300 struct list_head list; /* list of buffer pages */
301 local_t write; /* index for next write */
302 unsigned read; /* index for next read */
303 local_t entries; /* entries on this page */
304 unsigned long real_end; /* real end of data */
305 struct buffer_data_page *page; /* Actual data page */
309 * The buffer page counters, write and entries, must be reset
310 * atomically when crossing page boundaries. To synchronize this
311 * update, two counters are inserted into the number. One is
312 * the actual counter for the write position or count on the page.
314 * The other is a counter of updaters. Before an update happens
315 * the update partition of the counter is incremented. This will
316 * allow the updater to update the counter atomically.
318 * The counter is 20 bits, and the state data is 12.
320 #define RB_WRITE_MASK 0xfffff
321 #define RB_WRITE_INTCNT (1 << 20)
323 static void rb_init_page(struct buffer_data_page *bpage)
325 local_set(&bpage->commit, 0);
329 * ring_buffer_page_len - the size of data on the page.
330 * @page: The page to read
332 * Returns the amount of data on the page, including buffer page header.
334 size_t ring_buffer_page_len(void *page)
336 struct buffer_data_page *bpage = page;
338 return (local_read(&bpage->commit) & ~RB_MISSED_FLAGS)
343 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
346 static void free_buffer_page(struct buffer_page *bpage)
348 free_page((unsigned long)bpage->page);
353 * We need to fit the time_stamp delta into 27 bits.
355 static inline int test_time_stamp(u64 delta)
357 if (delta & TS_DELTA_TEST)
362 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
364 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
365 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
367 int ring_buffer_print_page_header(struct trace_seq *s)
369 struct buffer_data_page field;
371 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
372 "offset:0;\tsize:%u;\tsigned:%u;\n",
373 (unsigned int)sizeof(field.time_stamp),
374 (unsigned int)is_signed_type(u64));
376 trace_seq_printf(s, "\tfield: local_t commit;\t"
377 "offset:%u;\tsize:%u;\tsigned:%u;\n",
378 (unsigned int)offsetof(typeof(field), commit),
379 (unsigned int)sizeof(field.commit),
380 (unsigned int)is_signed_type(long));
382 trace_seq_printf(s, "\tfield: int overwrite;\t"
383 "offset:%u;\tsize:%u;\tsigned:%u;\n",
384 (unsigned int)offsetof(typeof(field), commit),
386 (unsigned int)is_signed_type(long));
388 trace_seq_printf(s, "\tfield: char data;\t"
389 "offset:%u;\tsize:%u;\tsigned:%u;\n",
390 (unsigned int)offsetof(typeof(field), data),
391 (unsigned int)BUF_PAGE_SIZE,
392 (unsigned int)is_signed_type(char));
394 return !trace_seq_has_overflowed(s);
398 struct irq_work work;
399 wait_queue_head_t waiters;
400 wait_queue_head_t full_waiters;
401 bool waiters_pending;
402 bool full_waiters_pending;
407 * Structure to hold event state and handle nested events.
409 struct rb_event_info {
412 unsigned long length;
413 struct buffer_page *tail_page;
418 * Used for which event context the event is in.
425 * See trace_recursive_lock() comment below for more details.
437 * head_page == tail_page && head == tail then buffer is empty.
439 struct ring_buffer_per_cpu {
441 atomic_t record_disabled;
442 struct ring_buffer *buffer;
443 raw_spinlock_t reader_lock; /* serialize readers */
444 arch_spinlock_t lock;
445 struct lock_class_key lock_key;
446 struct buffer_data_page *free_page;
447 unsigned long nr_pages;
448 unsigned int current_context;
449 struct list_head *pages;
450 struct buffer_page *head_page; /* read from head */
451 struct buffer_page *tail_page; /* write to tail */
452 struct buffer_page *commit_page; /* committed pages */
453 struct buffer_page *reader_page;
454 unsigned long lost_events;
455 unsigned long last_overrun;
456 local_t entries_bytes;
459 local_t commit_overrun;
460 local_t dropped_events;
464 unsigned long read_bytes;
467 /* ring buffer pages to update, > 0 to add, < 0 to remove */
468 long nr_pages_to_update;
469 struct list_head new_pages; /* new pages to add */
470 struct work_struct update_pages_work;
471 struct completion update_done;
473 struct rb_irq_work irq_work;
479 atomic_t record_disabled;
480 atomic_t resize_disabled;
481 cpumask_var_t cpumask;
483 struct lock_class_key *reader_lock_key;
487 struct ring_buffer_per_cpu **buffers;
489 struct hlist_node node;
492 struct rb_irq_work irq_work;
495 struct ring_buffer_iter {
496 struct ring_buffer_per_cpu *cpu_buffer;
498 struct buffer_page *head_page;
499 struct buffer_page *cache_reader_page;
500 unsigned long cache_read;
505 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
507 * Schedules a delayed work to wake up any task that is blocked on the
508 * ring buffer waiters queue.
510 static void rb_wake_up_waiters(struct irq_work *work)
512 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
514 wake_up_all(&rbwork->waiters);
515 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
516 rbwork->wakeup_full = false;
517 rbwork->full_waiters_pending = false;
518 wake_up_all(&rbwork->full_waiters);
523 * ring_buffer_wait - wait for input to the ring buffer
524 * @buffer: buffer to wait on
525 * @cpu: the cpu buffer to wait on
526 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
528 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
529 * as data is added to any of the @buffer's cpu buffers. Otherwise
530 * it will wait for data to be added to a specific cpu buffer.
532 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
534 struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
536 struct rb_irq_work *work;
540 * Depending on what the caller is waiting for, either any
541 * data in any cpu buffer, or a specific buffer, put the
542 * caller on the appropriate wait queue.
544 if (cpu == RING_BUFFER_ALL_CPUS) {
545 work = &buffer->irq_work;
546 /* Full only makes sense on per cpu reads */
549 if (!cpumask_test_cpu(cpu, buffer->cpumask))
551 cpu_buffer = buffer->buffers[cpu];
552 work = &cpu_buffer->irq_work;
558 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
560 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
563 * The events can happen in critical sections where
564 * checking a work queue can cause deadlocks.
565 * After adding a task to the queue, this flag is set
566 * only to notify events to try to wake up the queue
569 * We don't clear it even if the buffer is no longer
570 * empty. The flag only causes the next event to run
571 * irq_work to do the work queue wake up. The worse
572 * that can happen if we race with !trace_empty() is that
573 * an event will cause an irq_work to try to wake up
576 * There's no reason to protect this flag either, as
577 * the work queue and irq_work logic will do the necessary
578 * synchronization for the wake ups. The only thing
579 * that is necessary is that the wake up happens after
580 * a task has been queued. It's OK for spurious wake ups.
583 work->full_waiters_pending = true;
585 work->waiters_pending = true;
587 if (signal_pending(current)) {
592 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
595 if (cpu != RING_BUFFER_ALL_CPUS &&
596 !ring_buffer_empty_cpu(buffer, cpu)) {
603 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
604 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
605 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
615 finish_wait(&work->full_waiters, &wait);
617 finish_wait(&work->waiters, &wait);
623 * ring_buffer_poll_wait - poll on buffer input
624 * @buffer: buffer to wait on
625 * @cpu: the cpu buffer to wait on
626 * @filp: the file descriptor
627 * @poll_table: The poll descriptor
629 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
630 * as data is added to any of the @buffer's cpu buffers. Otherwise
631 * it will wait for data to be added to a specific cpu buffer.
633 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
636 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
637 struct file *filp, poll_table *poll_table)
639 struct ring_buffer_per_cpu *cpu_buffer;
640 struct rb_irq_work *work;
642 if (cpu == RING_BUFFER_ALL_CPUS)
643 work = &buffer->irq_work;
645 if (!cpumask_test_cpu(cpu, buffer->cpumask))
648 cpu_buffer = buffer->buffers[cpu];
649 work = &cpu_buffer->irq_work;
652 poll_wait(filp, &work->waiters, poll_table);
653 work->waiters_pending = true;
655 * There's a tight race between setting the waiters_pending and
656 * checking if the ring buffer is empty. Once the waiters_pending bit
657 * is set, the next event will wake the task up, but we can get stuck
658 * if there's only a single event in.
660 * FIXME: Ideally, we need a memory barrier on the writer side as well,
661 * but adding a memory barrier to all events will cause too much of a
662 * performance hit in the fast path. We only need a memory barrier when
663 * the buffer goes from empty to having content. But as this race is
664 * extremely small, and it's not a problem if another event comes in, we
669 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
670 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
671 return POLLIN | POLLRDNORM;
675 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
676 #define RB_WARN_ON(b, cond) \
678 int _____ret = unlikely(cond); \
680 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
681 struct ring_buffer_per_cpu *__b = \
683 atomic_inc(&__b->buffer->record_disabled); \
685 atomic_inc(&b->record_disabled); \
691 /* Up this if you want to test the TIME_EXTENTS and normalization */
692 #define DEBUG_SHIFT 0
694 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
696 /* shift to debug/test normalization and TIME_EXTENTS */
697 return buffer->clock() << DEBUG_SHIFT;
700 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
704 preempt_disable_notrace();
705 time = rb_time_stamp(buffer);
706 preempt_enable_notrace();
710 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
712 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
715 /* Just stupid testing the normalize function and deltas */
718 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
721 * Making the ring buffer lockless makes things tricky.
722 * Although writes only happen on the CPU that they are on,
723 * and they only need to worry about interrupts. Reads can
726 * The reader page is always off the ring buffer, but when the
727 * reader finishes with a page, it needs to swap its page with
728 * a new one from the buffer. The reader needs to take from
729 * the head (writes go to the tail). But if a writer is in overwrite
730 * mode and wraps, it must push the head page forward.
732 * Here lies the problem.
734 * The reader must be careful to replace only the head page, and
735 * not another one. As described at the top of the file in the
736 * ASCII art, the reader sets its old page to point to the next
737 * page after head. It then sets the page after head to point to
738 * the old reader page. But if the writer moves the head page
739 * during this operation, the reader could end up with the tail.
741 * We use cmpxchg to help prevent this race. We also do something
742 * special with the page before head. We set the LSB to 1.
744 * When the writer must push the page forward, it will clear the
745 * bit that points to the head page, move the head, and then set
746 * the bit that points to the new head page.
748 * We also don't want an interrupt coming in and moving the head
749 * page on another writer. Thus we use the second LSB to catch
752 * head->list->prev->next bit 1 bit 0
755 * Points to head page 0 1
758 * Note we can not trust the prev pointer of the head page, because:
760 * +----+ +-----+ +-----+
761 * | |------>| T |---X--->| N |
763 * +----+ +-----+ +-----+
766 * +----------| R |----------+ |
770 * Key: ---X--> HEAD flag set in pointer
775 * (see __rb_reserve_next() to see where this happens)
777 * What the above shows is that the reader just swapped out
778 * the reader page with a page in the buffer, but before it
779 * could make the new header point back to the new page added
780 * it was preempted by a writer. The writer moved forward onto
781 * the new page added by the reader and is about to move forward
784 * You can see, it is legitimate for the previous pointer of
785 * the head (or any page) not to point back to itself. But only
789 #define RB_PAGE_NORMAL 0UL
790 #define RB_PAGE_HEAD 1UL
791 #define RB_PAGE_UPDATE 2UL
794 #define RB_FLAG_MASK 3UL
796 /* PAGE_MOVED is not part of the mask */
797 #define RB_PAGE_MOVED 4UL
800 * rb_list_head - remove any bit
802 static struct list_head *rb_list_head(struct list_head *list)
804 unsigned long val = (unsigned long)list;
806 return (struct list_head *)(val & ~RB_FLAG_MASK);
810 * rb_is_head_page - test if the given page is the head page
812 * Because the reader may move the head_page pointer, we can
813 * not trust what the head page is (it may be pointing to
814 * the reader page). But if the next page is a header page,
815 * its flags will be non zero.
818 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
819 struct buffer_page *page, struct list_head *list)
823 val = (unsigned long)list->next;
825 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
826 return RB_PAGE_MOVED;
828 return val & RB_FLAG_MASK;
834 * The unique thing about the reader page, is that, if the
835 * writer is ever on it, the previous pointer never points
836 * back to the reader page.
838 static bool rb_is_reader_page(struct buffer_page *page)
840 struct list_head *list = page->list.prev;
842 return rb_list_head(list->next) != &page->list;
846 * rb_set_list_to_head - set a list_head to be pointing to head.
848 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
849 struct list_head *list)
853 ptr = (unsigned long *)&list->next;
854 *ptr |= RB_PAGE_HEAD;
855 *ptr &= ~RB_PAGE_UPDATE;
859 * rb_head_page_activate - sets up head page
861 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
863 struct buffer_page *head;
865 head = cpu_buffer->head_page;
870 * Set the previous list pointer to have the HEAD flag.
872 rb_set_list_to_head(cpu_buffer, head->list.prev);
875 static void rb_list_head_clear(struct list_head *list)
877 unsigned long *ptr = (unsigned long *)&list->next;
879 *ptr &= ~RB_FLAG_MASK;
883 * rb_head_page_dactivate - clears head page ptr (for free list)
886 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
888 struct list_head *hd;
890 /* Go through the whole list and clear any pointers found. */
891 rb_list_head_clear(cpu_buffer->pages);
893 list_for_each(hd, cpu_buffer->pages)
894 rb_list_head_clear(hd);
897 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
898 struct buffer_page *head,
899 struct buffer_page *prev,
900 int old_flag, int new_flag)
902 struct list_head *list;
903 unsigned long val = (unsigned long)&head->list;
908 val &= ~RB_FLAG_MASK;
910 ret = cmpxchg((unsigned long *)&list->next,
911 val | old_flag, val | new_flag);
913 /* check if the reader took the page */
914 if ((ret & ~RB_FLAG_MASK) != val)
915 return RB_PAGE_MOVED;
917 return ret & RB_FLAG_MASK;
920 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
921 struct buffer_page *head,
922 struct buffer_page *prev,
925 return rb_head_page_set(cpu_buffer, head, prev,
926 old_flag, RB_PAGE_UPDATE);
929 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
930 struct buffer_page *head,
931 struct buffer_page *prev,
934 return rb_head_page_set(cpu_buffer, head, prev,
935 old_flag, RB_PAGE_HEAD);
938 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
939 struct buffer_page *head,
940 struct buffer_page *prev,
943 return rb_head_page_set(cpu_buffer, head, prev,
944 old_flag, RB_PAGE_NORMAL);
947 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
948 struct buffer_page **bpage)
950 struct list_head *p = rb_list_head((*bpage)->list.next);
952 *bpage = list_entry(p, struct buffer_page, list);
955 static struct buffer_page *
956 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
958 struct buffer_page *head;
959 struct buffer_page *page;
960 struct list_head *list;
963 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
967 list = cpu_buffer->pages;
968 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
971 page = head = cpu_buffer->head_page;
973 * It is possible that the writer moves the header behind
974 * where we started, and we miss in one loop.
975 * A second loop should grab the header, but we'll do
976 * three loops just because I'm paranoid.
978 for (i = 0; i < 3; i++) {
980 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
981 cpu_buffer->head_page = page;
984 rb_inc_page(cpu_buffer, &page);
985 } while (page != head);
988 RB_WARN_ON(cpu_buffer, 1);
993 static int rb_head_page_replace(struct buffer_page *old,
994 struct buffer_page *new)
996 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1000 val = *ptr & ~RB_FLAG_MASK;
1001 val |= RB_PAGE_HEAD;
1003 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1009 * rb_tail_page_update - move the tail page forward
1011 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1012 struct buffer_page *tail_page,
1013 struct buffer_page *next_page)
1015 unsigned long old_entries;
1016 unsigned long old_write;
1019 * The tail page now needs to be moved forward.
1021 * We need to reset the tail page, but without messing
1022 * with possible erasing of data brought in by interrupts
1023 * that have moved the tail page and are currently on it.
1025 * We add a counter to the write field to denote this.
1027 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1028 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1031 * Just make sure we have seen our old_write and synchronize
1032 * with any interrupts that come in.
1037 * If the tail page is still the same as what we think
1038 * it is, then it is up to us to update the tail
1041 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1042 /* Zero the write counter */
1043 unsigned long val = old_write & ~RB_WRITE_MASK;
1044 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1047 * This will only succeed if an interrupt did
1048 * not come in and change it. In which case, we
1049 * do not want to modify it.
1051 * We add (void) to let the compiler know that we do not care
1052 * about the return value of these functions. We use the
1053 * cmpxchg to only update if an interrupt did not already
1054 * do it for us. If the cmpxchg fails, we don't care.
1056 (void)local_cmpxchg(&next_page->write, old_write, val);
1057 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1060 * No need to worry about races with clearing out the commit.
1061 * it only can increment when a commit takes place. But that
1062 * only happens in the outer most nested commit.
1064 local_set(&next_page->page->commit, 0);
1066 /* Again, either we update tail_page or an interrupt does */
1067 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1071 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1072 struct buffer_page *bpage)
1074 unsigned long val = (unsigned long)bpage;
1076 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1083 * rb_check_list - make sure a pointer to a list has the last bits zero
1085 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1086 struct list_head *list)
1088 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1090 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1096 * rb_check_pages - integrity check of buffer pages
1097 * @cpu_buffer: CPU buffer with pages to test
1099 * As a safety measure we check to make sure the data pages have not
1102 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1104 struct list_head *head = cpu_buffer->pages;
1105 struct buffer_page *bpage, *tmp;
1107 /* Reset the head page if it exists */
1108 if (cpu_buffer->head_page)
1109 rb_set_head_page(cpu_buffer);
1111 rb_head_page_deactivate(cpu_buffer);
1113 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1115 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1118 if (rb_check_list(cpu_buffer, head))
1121 list_for_each_entry_safe(bpage, tmp, head, list) {
1122 if (RB_WARN_ON(cpu_buffer,
1123 bpage->list.next->prev != &bpage->list))
1125 if (RB_WARN_ON(cpu_buffer,
1126 bpage->list.prev->next != &bpage->list))
1128 if (rb_check_list(cpu_buffer, &bpage->list))
1132 rb_head_page_activate(cpu_buffer);
1137 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1139 struct buffer_page *bpage, *tmp;
1142 /* Check if the available memory is there first */
1143 i = si_mem_available();
1147 for (i = 0; i < nr_pages; i++) {
1150 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1151 * gracefully without invoking oom-killer and the system is not
1154 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1155 GFP_KERNEL | __GFP_RETRY_MAYFAIL,
1160 list_add(&bpage->list, pages);
1162 page = alloc_pages_node(cpu_to_node(cpu),
1163 GFP_KERNEL | __GFP_RETRY_MAYFAIL, 0);
1166 bpage->page = page_address(page);
1167 rb_init_page(bpage->page);
1173 list_for_each_entry_safe(bpage, tmp, pages, list) {
1174 list_del_init(&bpage->list);
1175 free_buffer_page(bpage);
1181 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1182 unsigned long nr_pages)
1188 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1192 * The ring buffer page list is a circular list that does not
1193 * start and end with a list head. All page list items point to
1196 cpu_buffer->pages = pages.next;
1199 cpu_buffer->nr_pages = nr_pages;
1201 rb_check_pages(cpu_buffer);
1206 static struct ring_buffer_per_cpu *
1207 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1209 struct ring_buffer_per_cpu *cpu_buffer;
1210 struct buffer_page *bpage;
1214 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1215 GFP_KERNEL, cpu_to_node(cpu));
1219 cpu_buffer->cpu = cpu;
1220 cpu_buffer->buffer = buffer;
1221 raw_spin_lock_init(&cpu_buffer->reader_lock);
1222 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1223 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1224 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1225 init_completion(&cpu_buffer->update_done);
1226 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1227 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1228 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1230 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1231 GFP_KERNEL, cpu_to_node(cpu));
1233 goto fail_free_buffer;
1235 rb_check_bpage(cpu_buffer, bpage);
1237 cpu_buffer->reader_page = bpage;
1238 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1240 goto fail_free_reader;
1241 bpage->page = page_address(page);
1242 rb_init_page(bpage->page);
1244 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1245 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1247 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1249 goto fail_free_reader;
1251 cpu_buffer->head_page
1252 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1253 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1255 rb_head_page_activate(cpu_buffer);
1260 free_buffer_page(cpu_buffer->reader_page);
1267 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1269 struct list_head *head = cpu_buffer->pages;
1270 struct buffer_page *bpage, *tmp;
1272 free_buffer_page(cpu_buffer->reader_page);
1275 rb_head_page_deactivate(cpu_buffer);
1277 list_for_each_entry_safe(bpage, tmp, head, list) {
1278 list_del_init(&bpage->list);
1279 free_buffer_page(bpage);
1281 bpage = list_entry(head, struct buffer_page, list);
1282 free_buffer_page(bpage);
1289 * __ring_buffer_alloc - allocate a new ring_buffer
1290 * @size: the size in bytes per cpu that is needed.
1291 * @flags: attributes to set for the ring buffer.
1293 * Currently the only flag that is available is the RB_FL_OVERWRITE
1294 * flag. This flag means that the buffer will overwrite old data
1295 * when the buffer wraps. If this flag is not set, the buffer will
1296 * drop data when the tail hits the head.
1298 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1299 struct lock_class_key *key)
1301 struct ring_buffer *buffer;
1307 /* keep it in its own cache line */
1308 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1313 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1314 goto fail_free_buffer;
1316 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1317 buffer->flags = flags;
1318 buffer->clock = trace_clock_local;
1319 buffer->reader_lock_key = key;
1321 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1322 init_waitqueue_head(&buffer->irq_work.waiters);
1324 /* need at least two pages */
1328 buffer->cpus = nr_cpu_ids;
1330 bsize = sizeof(void *) * nr_cpu_ids;
1331 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1333 if (!buffer->buffers)
1334 goto fail_free_cpumask;
1336 cpu = raw_smp_processor_id();
1337 cpumask_set_cpu(cpu, buffer->cpumask);
1338 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1339 if (!buffer->buffers[cpu])
1340 goto fail_free_buffers;
1342 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1344 goto fail_free_buffers;
1346 mutex_init(&buffer->mutex);
1351 for_each_buffer_cpu(buffer, cpu) {
1352 if (buffer->buffers[cpu])
1353 rb_free_cpu_buffer(buffer->buffers[cpu]);
1355 kfree(buffer->buffers);
1358 free_cpumask_var(buffer->cpumask);
1364 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1367 * ring_buffer_free - free a ring buffer.
1368 * @buffer: the buffer to free.
1371 ring_buffer_free(struct ring_buffer *buffer)
1375 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1377 for_each_buffer_cpu(buffer, cpu)
1378 rb_free_cpu_buffer(buffer->buffers[cpu]);
1380 kfree(buffer->buffers);
1381 free_cpumask_var(buffer->cpumask);
1385 EXPORT_SYMBOL_GPL(ring_buffer_free);
1387 void ring_buffer_set_clock(struct ring_buffer *buffer,
1390 buffer->clock = clock;
1393 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1395 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1397 return local_read(&bpage->entries) & RB_WRITE_MASK;
1400 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1402 return local_read(&bpage->write) & RB_WRITE_MASK;
1406 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1408 struct list_head *tail_page, *to_remove, *next_page;
1409 struct buffer_page *to_remove_page, *tmp_iter_page;
1410 struct buffer_page *last_page, *first_page;
1411 unsigned long nr_removed;
1412 unsigned long head_bit;
1417 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1418 atomic_inc(&cpu_buffer->record_disabled);
1420 * We don't race with the readers since we have acquired the reader
1421 * lock. We also don't race with writers after disabling recording.
1422 * This makes it easy to figure out the first and the last page to be
1423 * removed from the list. We unlink all the pages in between including
1424 * the first and last pages. This is done in a busy loop so that we
1425 * lose the least number of traces.
1426 * The pages are freed after we restart recording and unlock readers.
1428 tail_page = &cpu_buffer->tail_page->list;
1431 * tail page might be on reader page, we remove the next page
1432 * from the ring buffer
1434 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1435 tail_page = rb_list_head(tail_page->next);
1436 to_remove = tail_page;
1438 /* start of pages to remove */
1439 first_page = list_entry(rb_list_head(to_remove->next),
1440 struct buffer_page, list);
1442 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1443 to_remove = rb_list_head(to_remove)->next;
1444 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1447 next_page = rb_list_head(to_remove)->next;
1450 * Now we remove all pages between tail_page and next_page.
1451 * Make sure that we have head_bit value preserved for the
1454 tail_page->next = (struct list_head *)((unsigned long)next_page |
1456 next_page = rb_list_head(next_page);
1457 next_page->prev = tail_page;
1459 /* make sure pages points to a valid page in the ring buffer */
1460 cpu_buffer->pages = next_page;
1462 /* update head page */
1464 cpu_buffer->head_page = list_entry(next_page,
1465 struct buffer_page, list);
1468 * change read pointer to make sure any read iterators reset
1471 cpu_buffer->read = 0;
1473 /* pages are removed, resume tracing and then free the pages */
1474 atomic_dec(&cpu_buffer->record_disabled);
1475 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1477 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1479 /* last buffer page to remove */
1480 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1482 tmp_iter_page = first_page;
1487 to_remove_page = tmp_iter_page;
1488 rb_inc_page(cpu_buffer, &tmp_iter_page);
1490 /* update the counters */
1491 page_entries = rb_page_entries(to_remove_page);
1494 * If something was added to this page, it was full
1495 * since it is not the tail page. So we deduct the
1496 * bytes consumed in ring buffer from here.
1497 * Increment overrun to account for the lost events.
1499 local_add(page_entries, &cpu_buffer->overrun);
1500 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1504 * We have already removed references to this list item, just
1505 * free up the buffer_page and its page
1507 free_buffer_page(to_remove_page);
1510 } while (to_remove_page != last_page);
1512 RB_WARN_ON(cpu_buffer, nr_removed);
1514 return nr_removed == 0;
1518 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1520 struct list_head *pages = &cpu_buffer->new_pages;
1521 int retries, success;
1523 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1525 * We are holding the reader lock, so the reader page won't be swapped
1526 * in the ring buffer. Now we are racing with the writer trying to
1527 * move head page and the tail page.
1528 * We are going to adapt the reader page update process where:
1529 * 1. We first splice the start and end of list of new pages between
1530 * the head page and its previous page.
1531 * 2. We cmpxchg the prev_page->next to point from head page to the
1532 * start of new pages list.
1533 * 3. Finally, we update the head->prev to the end of new list.
1535 * We will try this process 10 times, to make sure that we don't keep
1541 struct list_head *head_page, *prev_page, *r;
1542 struct list_head *last_page, *first_page;
1543 struct list_head *head_page_with_bit;
1545 head_page = &rb_set_head_page(cpu_buffer)->list;
1548 prev_page = head_page->prev;
1550 first_page = pages->next;
1551 last_page = pages->prev;
1553 head_page_with_bit = (struct list_head *)
1554 ((unsigned long)head_page | RB_PAGE_HEAD);
1556 last_page->next = head_page_with_bit;
1557 first_page->prev = prev_page;
1559 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1561 if (r == head_page_with_bit) {
1563 * yay, we replaced the page pointer to our new list,
1564 * now, we just have to update to head page's prev
1565 * pointer to point to end of list
1567 head_page->prev = last_page;
1574 INIT_LIST_HEAD(pages);
1576 * If we weren't successful in adding in new pages, warn and stop
1579 RB_WARN_ON(cpu_buffer, !success);
1580 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1582 /* free pages if they weren't inserted */
1584 struct buffer_page *bpage, *tmp;
1585 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1587 list_del_init(&bpage->list);
1588 free_buffer_page(bpage);
1594 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1598 if (cpu_buffer->nr_pages_to_update > 0)
1599 success = rb_insert_pages(cpu_buffer);
1601 success = rb_remove_pages(cpu_buffer,
1602 -cpu_buffer->nr_pages_to_update);
1605 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1608 static void update_pages_handler(struct work_struct *work)
1610 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1611 struct ring_buffer_per_cpu, update_pages_work);
1612 rb_update_pages(cpu_buffer);
1613 complete(&cpu_buffer->update_done);
1617 * ring_buffer_resize - resize the ring buffer
1618 * @buffer: the buffer to resize.
1619 * @size: the new size.
1620 * @cpu_id: the cpu buffer to resize
1622 * Minimum size is 2 * BUF_PAGE_SIZE.
1624 * Returns 0 on success and < 0 on failure.
1626 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1629 struct ring_buffer_per_cpu *cpu_buffer;
1630 unsigned long nr_pages;
1634 * Always succeed at resizing a non-existent buffer:
1639 /* Make sure the requested buffer exists */
1640 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1641 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1644 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1646 /* we need a minimum of two pages */
1650 size = nr_pages * BUF_PAGE_SIZE;
1653 * Don't succeed if resizing is disabled, as a reader might be
1654 * manipulating the ring buffer and is expecting a sane state while
1657 if (atomic_read(&buffer->resize_disabled))
1660 /* prevent another thread from changing buffer sizes */
1661 mutex_lock(&buffer->mutex);
1663 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1664 /* calculate the pages to update */
1665 for_each_buffer_cpu(buffer, cpu) {
1666 cpu_buffer = buffer->buffers[cpu];
1668 cpu_buffer->nr_pages_to_update = nr_pages -
1669 cpu_buffer->nr_pages;
1671 * nothing more to do for removing pages or no update
1673 if (cpu_buffer->nr_pages_to_update <= 0)
1676 * to add pages, make sure all new pages can be
1677 * allocated without receiving ENOMEM
1679 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1680 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1681 &cpu_buffer->new_pages, cpu)) {
1682 /* not enough memory for new pages */
1690 * Fire off all the required work handlers
1691 * We can't schedule on offline CPUs, but it's not necessary
1692 * since we can change their buffer sizes without any race.
1694 for_each_buffer_cpu(buffer, cpu) {
1695 cpu_buffer = buffer->buffers[cpu];
1696 if (!cpu_buffer->nr_pages_to_update)
1699 /* Can't run something on an offline CPU. */
1700 if (!cpu_online(cpu)) {
1701 rb_update_pages(cpu_buffer);
1702 cpu_buffer->nr_pages_to_update = 0;
1704 schedule_work_on(cpu,
1705 &cpu_buffer->update_pages_work);
1709 /* wait for all the updates to complete */
1710 for_each_buffer_cpu(buffer, cpu) {
1711 cpu_buffer = buffer->buffers[cpu];
1712 if (!cpu_buffer->nr_pages_to_update)
1715 if (cpu_online(cpu))
1716 wait_for_completion(&cpu_buffer->update_done);
1717 cpu_buffer->nr_pages_to_update = 0;
1722 /* Make sure this CPU has been intitialized */
1723 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1726 cpu_buffer = buffer->buffers[cpu_id];
1728 if (nr_pages == cpu_buffer->nr_pages)
1731 cpu_buffer->nr_pages_to_update = nr_pages -
1732 cpu_buffer->nr_pages;
1734 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1735 if (cpu_buffer->nr_pages_to_update > 0 &&
1736 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1737 &cpu_buffer->new_pages, cpu_id)) {
1744 /* Can't run something on an offline CPU. */
1745 if (!cpu_online(cpu_id))
1746 rb_update_pages(cpu_buffer);
1748 schedule_work_on(cpu_id,
1749 &cpu_buffer->update_pages_work);
1750 wait_for_completion(&cpu_buffer->update_done);
1753 cpu_buffer->nr_pages_to_update = 0;
1759 * The ring buffer resize can happen with the ring buffer
1760 * enabled, so that the update disturbs the tracing as little
1761 * as possible. But if the buffer is disabled, we do not need
1762 * to worry about that, and we can take the time to verify
1763 * that the buffer is not corrupt.
1765 if (atomic_read(&buffer->record_disabled)) {
1766 atomic_inc(&buffer->record_disabled);
1768 * Even though the buffer was disabled, we must make sure
1769 * that it is truly disabled before calling rb_check_pages.
1770 * There could have been a race between checking
1771 * record_disable and incrementing it.
1773 synchronize_sched();
1774 for_each_buffer_cpu(buffer, cpu) {
1775 cpu_buffer = buffer->buffers[cpu];
1776 rb_check_pages(cpu_buffer);
1778 atomic_dec(&buffer->record_disabled);
1781 mutex_unlock(&buffer->mutex);
1785 for_each_buffer_cpu(buffer, cpu) {
1786 struct buffer_page *bpage, *tmp;
1788 cpu_buffer = buffer->buffers[cpu];
1789 cpu_buffer->nr_pages_to_update = 0;
1791 if (list_empty(&cpu_buffer->new_pages))
1794 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1796 list_del_init(&bpage->list);
1797 free_buffer_page(bpage);
1800 mutex_unlock(&buffer->mutex);
1803 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1805 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1807 mutex_lock(&buffer->mutex);
1809 buffer->flags |= RB_FL_OVERWRITE;
1811 buffer->flags &= ~RB_FL_OVERWRITE;
1812 mutex_unlock(&buffer->mutex);
1814 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1816 static __always_inline void *
1817 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1819 return bpage->data + index;
1822 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1824 return bpage->page->data + index;
1827 static __always_inline struct ring_buffer_event *
1828 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1830 return __rb_page_index(cpu_buffer->reader_page,
1831 cpu_buffer->reader_page->read);
1834 static __always_inline struct ring_buffer_event *
1835 rb_iter_head_event(struct ring_buffer_iter *iter)
1837 return __rb_page_index(iter->head_page, iter->head);
1840 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1842 return local_read(&bpage->page->commit);
1845 /* Size is determined by what has been committed */
1846 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1848 return rb_page_commit(bpage);
1851 static __always_inline unsigned
1852 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1854 return rb_page_commit(cpu_buffer->commit_page);
1857 static __always_inline unsigned
1858 rb_event_index(struct ring_buffer_event *event)
1860 unsigned long addr = (unsigned long)event;
1862 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1865 static void rb_inc_iter(struct ring_buffer_iter *iter)
1867 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1870 * The iterator could be on the reader page (it starts there).
1871 * But the head could have moved, since the reader was
1872 * found. Check for this case and assign the iterator
1873 * to the head page instead of next.
1875 if (iter->head_page == cpu_buffer->reader_page)
1876 iter->head_page = rb_set_head_page(cpu_buffer);
1878 rb_inc_page(cpu_buffer, &iter->head_page);
1880 iter->read_stamp = iter->head_page->page->time_stamp;
1885 * rb_handle_head_page - writer hit the head page
1887 * Returns: +1 to retry page
1892 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1893 struct buffer_page *tail_page,
1894 struct buffer_page *next_page)
1896 struct buffer_page *new_head;
1901 entries = rb_page_entries(next_page);
1904 * The hard part is here. We need to move the head
1905 * forward, and protect against both readers on
1906 * other CPUs and writers coming in via interrupts.
1908 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1912 * type can be one of four:
1913 * NORMAL - an interrupt already moved it for us
1914 * HEAD - we are the first to get here.
1915 * UPDATE - we are the interrupt interrupting
1917 * MOVED - a reader on another CPU moved the next
1918 * pointer to its reader page. Give up
1925 * We changed the head to UPDATE, thus
1926 * it is our responsibility to update
1929 local_add(entries, &cpu_buffer->overrun);
1930 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1933 * The entries will be zeroed out when we move the
1937 /* still more to do */
1940 case RB_PAGE_UPDATE:
1942 * This is an interrupt that interrupt the
1943 * previous update. Still more to do.
1946 case RB_PAGE_NORMAL:
1948 * An interrupt came in before the update
1949 * and processed this for us.
1950 * Nothing left to do.
1955 * The reader is on another CPU and just did
1956 * a swap with our next_page.
1961 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1966 * Now that we are here, the old head pointer is
1967 * set to UPDATE. This will keep the reader from
1968 * swapping the head page with the reader page.
1969 * The reader (on another CPU) will spin till
1972 * We just need to protect against interrupts
1973 * doing the job. We will set the next pointer
1974 * to HEAD. After that, we set the old pointer
1975 * to NORMAL, but only if it was HEAD before.
1976 * otherwise we are an interrupt, and only
1977 * want the outer most commit to reset it.
1979 new_head = next_page;
1980 rb_inc_page(cpu_buffer, &new_head);
1982 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1986 * Valid returns are:
1987 * HEAD - an interrupt came in and already set it.
1988 * NORMAL - One of two things:
1989 * 1) We really set it.
1990 * 2) A bunch of interrupts came in and moved
1991 * the page forward again.
1995 case RB_PAGE_NORMAL:
1999 RB_WARN_ON(cpu_buffer, 1);
2004 * It is possible that an interrupt came in,
2005 * set the head up, then more interrupts came in
2006 * and moved it again. When we get back here,
2007 * the page would have been set to NORMAL but we
2008 * just set it back to HEAD.
2010 * How do you detect this? Well, if that happened
2011 * the tail page would have moved.
2013 if (ret == RB_PAGE_NORMAL) {
2014 struct buffer_page *buffer_tail_page;
2016 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2018 * If the tail had moved passed next, then we need
2019 * to reset the pointer.
2021 if (buffer_tail_page != tail_page &&
2022 buffer_tail_page != next_page)
2023 rb_head_page_set_normal(cpu_buffer, new_head,
2029 * If this was the outer most commit (the one that
2030 * changed the original pointer from HEAD to UPDATE),
2031 * then it is up to us to reset it to NORMAL.
2033 if (type == RB_PAGE_HEAD) {
2034 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2037 if (RB_WARN_ON(cpu_buffer,
2038 ret != RB_PAGE_UPDATE))
2046 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2047 unsigned long tail, struct rb_event_info *info)
2049 struct buffer_page *tail_page = info->tail_page;
2050 struct ring_buffer_event *event;
2051 unsigned long length = info->length;
2054 * Only the event that crossed the page boundary
2055 * must fill the old tail_page with padding.
2057 if (tail >= BUF_PAGE_SIZE) {
2059 * If the page was filled, then we still need
2060 * to update the real_end. Reset it to zero
2061 * and the reader will ignore it.
2063 if (tail == BUF_PAGE_SIZE)
2064 tail_page->real_end = 0;
2066 local_sub(length, &tail_page->write);
2070 event = __rb_page_index(tail_page, tail);
2072 /* account for padding bytes */
2073 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2076 * Save the original length to the meta data.
2077 * This will be used by the reader to add lost event
2080 tail_page->real_end = tail;
2083 * If this event is bigger than the minimum size, then
2084 * we need to be careful that we don't subtract the
2085 * write counter enough to allow another writer to slip
2087 * We put in a discarded commit instead, to make sure
2088 * that this space is not used again.
2090 * If we are less than the minimum size, we don't need to
2093 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2094 /* No room for any events */
2096 /* Mark the rest of the page with padding */
2097 rb_event_set_padding(event);
2099 /* Make sure the padding is visible before the write update */
2102 /* Set the write back to the previous setting */
2103 local_sub(length, &tail_page->write);
2107 /* Put in a discarded event */
2108 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2109 event->type_len = RINGBUF_TYPE_PADDING;
2110 /* time delta must be non zero */
2111 event->time_delta = 1;
2113 /* Make sure the padding is visible before the tail_page->write update */
2116 /* Set write to end of buffer */
2117 length = (tail + length) - BUF_PAGE_SIZE;
2118 local_sub(length, &tail_page->write);
2121 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2124 * This is the slow path, force gcc not to inline it.
2126 static noinline struct ring_buffer_event *
2127 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2128 unsigned long tail, struct rb_event_info *info)
2130 struct buffer_page *tail_page = info->tail_page;
2131 struct buffer_page *commit_page = cpu_buffer->commit_page;
2132 struct ring_buffer *buffer = cpu_buffer->buffer;
2133 struct buffer_page *next_page;
2136 next_page = tail_page;
2138 rb_inc_page(cpu_buffer, &next_page);
2141 * If for some reason, we had an interrupt storm that made
2142 * it all the way around the buffer, bail, and warn
2145 if (unlikely(next_page == commit_page)) {
2146 local_inc(&cpu_buffer->commit_overrun);
2151 * This is where the fun begins!
2153 * We are fighting against races between a reader that
2154 * could be on another CPU trying to swap its reader
2155 * page with the buffer head.
2157 * We are also fighting against interrupts coming in and
2158 * moving the head or tail on us as well.
2160 * If the next page is the head page then we have filled
2161 * the buffer, unless the commit page is still on the
2164 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2167 * If the commit is not on the reader page, then
2168 * move the header page.
2170 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2172 * If we are not in overwrite mode,
2173 * this is easy, just stop here.
2175 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2176 local_inc(&cpu_buffer->dropped_events);
2180 ret = rb_handle_head_page(cpu_buffer,
2189 * We need to be careful here too. The
2190 * commit page could still be on the reader
2191 * page. We could have a small buffer, and
2192 * have filled up the buffer with events
2193 * from interrupts and such, and wrapped.
2195 * Note, if the tail page is also the on the
2196 * reader_page, we let it move out.
2198 if (unlikely((cpu_buffer->commit_page !=
2199 cpu_buffer->tail_page) &&
2200 (cpu_buffer->commit_page ==
2201 cpu_buffer->reader_page))) {
2202 local_inc(&cpu_buffer->commit_overrun);
2208 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2212 rb_reset_tail(cpu_buffer, tail, info);
2214 /* Commit what we have for now. */
2215 rb_end_commit(cpu_buffer);
2216 /* rb_end_commit() decs committing */
2217 local_inc(&cpu_buffer->committing);
2219 /* fail and let the caller try again */
2220 return ERR_PTR(-EAGAIN);
2224 rb_reset_tail(cpu_buffer, tail, info);
2229 /* Slow path, do not inline */
2230 static noinline struct ring_buffer_event *
2231 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2233 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2235 /* Not the first event on the page? */
2236 if (rb_event_index(event)) {
2237 event->time_delta = delta & TS_MASK;
2238 event->array[0] = delta >> TS_SHIFT;
2240 /* nope, just zero it */
2241 event->time_delta = 0;
2242 event->array[0] = 0;
2245 return skip_time_extend(event);
2248 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2249 struct ring_buffer_event *event);
2252 * rb_update_event - update event type and data
2253 * @event: the event to update
2254 * @type: the type of event
2255 * @length: the size of the event field in the ring buffer
2257 * Update the type and data fields of the event. The length
2258 * is the actual size that is written to the ring buffer,
2259 * and with this, we can determine what to place into the
2263 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2264 struct ring_buffer_event *event,
2265 struct rb_event_info *info)
2267 unsigned length = info->length;
2268 u64 delta = info->delta;
2270 /* Only a commit updates the timestamp */
2271 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2275 * If we need to add a timestamp, then we
2276 * add it to the start of the resevered space.
2278 if (unlikely(info->add_timestamp)) {
2279 event = rb_add_time_stamp(event, delta);
2280 length -= RB_LEN_TIME_EXTEND;
2284 event->time_delta = delta;
2285 length -= RB_EVNT_HDR_SIZE;
2286 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2287 event->type_len = 0;
2288 event->array[0] = length;
2290 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2293 static unsigned rb_calculate_event_length(unsigned length)
2295 struct ring_buffer_event event; /* Used only for sizeof array */
2297 /* zero length can cause confusions */
2301 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2302 length += sizeof(event.array[0]);
2304 length += RB_EVNT_HDR_SIZE;
2305 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2308 * In case the time delta is larger than the 27 bits for it
2309 * in the header, we need to add a timestamp. If another
2310 * event comes in when trying to discard this one to increase
2311 * the length, then the timestamp will be added in the allocated
2312 * space of this event. If length is bigger than the size needed
2313 * for the TIME_EXTEND, then padding has to be used. The events
2314 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2315 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2316 * As length is a multiple of 4, we only need to worry if it
2317 * is 12 (RB_LEN_TIME_EXTEND + 4).
2319 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2320 length += RB_ALIGNMENT;
2325 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2326 static inline bool sched_clock_stable(void)
2333 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2334 struct ring_buffer_event *event)
2336 unsigned long new_index, old_index;
2337 struct buffer_page *bpage;
2338 unsigned long index;
2341 new_index = rb_event_index(event);
2342 old_index = new_index + rb_event_ts_length(event);
2343 addr = (unsigned long)event;
2346 bpage = READ_ONCE(cpu_buffer->tail_page);
2348 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2349 unsigned long write_mask =
2350 local_read(&bpage->write) & ~RB_WRITE_MASK;
2351 unsigned long event_length = rb_event_length(event);
2353 * This is on the tail page. It is possible that
2354 * a write could come in and move the tail page
2355 * and write to the next page. That is fine
2356 * because we just shorten what is on this page.
2358 old_index += write_mask;
2359 new_index += write_mask;
2360 index = local_cmpxchg(&bpage->write, old_index, new_index);
2361 if (index == old_index) {
2362 /* update counters */
2363 local_sub(event_length, &cpu_buffer->entries_bytes);
2368 /* could not discard */
2372 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2374 local_inc(&cpu_buffer->committing);
2375 local_inc(&cpu_buffer->commits);
2378 static __always_inline void
2379 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2381 unsigned long max_count;
2384 * We only race with interrupts and NMIs on this CPU.
2385 * If we own the commit event, then we can commit
2386 * all others that interrupted us, since the interruptions
2387 * are in stack format (they finish before they come
2388 * back to us). This allows us to do a simple loop to
2389 * assign the commit to the tail.
2392 max_count = cpu_buffer->nr_pages * 100;
2394 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2395 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2397 if (RB_WARN_ON(cpu_buffer,
2398 rb_is_reader_page(cpu_buffer->tail_page)))
2400 local_set(&cpu_buffer->commit_page->page->commit,
2401 rb_page_write(cpu_buffer->commit_page));
2402 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2403 /* Only update the write stamp if the page has an event */
2404 if (rb_page_write(cpu_buffer->commit_page))
2405 cpu_buffer->write_stamp =
2406 cpu_buffer->commit_page->page->time_stamp;
2407 /* add barrier to keep gcc from optimizing too much */
2410 while (rb_commit_index(cpu_buffer) !=
2411 rb_page_write(cpu_buffer->commit_page)) {
2413 local_set(&cpu_buffer->commit_page->page->commit,
2414 rb_page_write(cpu_buffer->commit_page));
2415 RB_WARN_ON(cpu_buffer,
2416 local_read(&cpu_buffer->commit_page->page->commit) &
2421 /* again, keep gcc from optimizing */
2425 * If an interrupt came in just after the first while loop
2426 * and pushed the tail page forward, we will be left with
2427 * a dangling commit that will never go forward.
2429 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2433 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2435 unsigned long commits;
2437 if (RB_WARN_ON(cpu_buffer,
2438 !local_read(&cpu_buffer->committing)))
2442 commits = local_read(&cpu_buffer->commits);
2443 /* synchronize with interrupts */
2445 if (local_read(&cpu_buffer->committing) == 1)
2446 rb_set_commit_to_write(cpu_buffer);
2448 local_dec(&cpu_buffer->committing);
2450 /* synchronize with interrupts */
2454 * Need to account for interrupts coming in between the
2455 * updating of the commit page and the clearing of the
2456 * committing counter.
2458 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2459 !local_read(&cpu_buffer->committing)) {
2460 local_inc(&cpu_buffer->committing);
2465 static inline void rb_event_discard(struct ring_buffer_event *event)
2467 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2468 event = skip_time_extend(event);
2470 /* array[0] holds the actual length for the discarded event */
2471 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2472 event->type_len = RINGBUF_TYPE_PADDING;
2473 /* time delta must be non zero */
2474 if (!event->time_delta)
2475 event->time_delta = 1;
2478 static __always_inline bool
2479 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2480 struct ring_buffer_event *event)
2482 unsigned long addr = (unsigned long)event;
2483 unsigned long index;
2485 index = rb_event_index(event);
2488 return cpu_buffer->commit_page->page == (void *)addr &&
2489 rb_commit_index(cpu_buffer) == index;
2492 static __always_inline void
2493 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2494 struct ring_buffer_event *event)
2499 * The event first in the commit queue updates the
2502 if (rb_event_is_commit(cpu_buffer, event)) {
2504 * A commit event that is first on a page
2505 * updates the write timestamp with the page stamp
2507 if (!rb_event_index(event))
2508 cpu_buffer->write_stamp =
2509 cpu_buffer->commit_page->page->time_stamp;
2510 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2511 delta = event->array[0];
2513 delta += event->time_delta;
2514 cpu_buffer->write_stamp += delta;
2516 cpu_buffer->write_stamp += event->time_delta;
2520 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2521 struct ring_buffer_event *event)
2523 local_inc(&cpu_buffer->entries);
2524 rb_update_write_stamp(cpu_buffer, event);
2525 rb_end_commit(cpu_buffer);
2528 static __always_inline void
2529 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2533 if (buffer->irq_work.waiters_pending) {
2534 buffer->irq_work.waiters_pending = false;
2535 /* irq_work_queue() supplies it's own memory barriers */
2536 irq_work_queue(&buffer->irq_work.work);
2539 if (cpu_buffer->irq_work.waiters_pending) {
2540 cpu_buffer->irq_work.waiters_pending = false;
2541 /* irq_work_queue() supplies it's own memory barriers */
2542 irq_work_queue(&cpu_buffer->irq_work.work);
2545 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2547 if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2548 cpu_buffer->irq_work.wakeup_full = true;
2549 cpu_buffer->irq_work.full_waiters_pending = false;
2550 /* irq_work_queue() supplies it's own memory barriers */
2551 irq_work_queue(&cpu_buffer->irq_work.work);
2556 * The lock and unlock are done within a preempt disable section.
2557 * The current_context per_cpu variable can only be modified
2558 * by the current task between lock and unlock. But it can
2559 * be modified more than once via an interrupt. To pass this
2560 * information from the lock to the unlock without having to
2561 * access the 'in_interrupt()' functions again (which do show
2562 * a bit of overhead in something as critical as function tracing,
2563 * we use a bitmask trick.
2565 * bit 1 = NMI context
2566 * bit 2 = IRQ context
2567 * bit 3 = SoftIRQ context
2568 * bit 4 = normal context.
2570 * This works because this is the order of contexts that can
2571 * preempt other contexts. A SoftIRQ never preempts an IRQ
2574 * When the context is determined, the corresponding bit is
2575 * checked and set (if it was set, then a recursion of that context
2578 * On unlock, we need to clear this bit. To do so, just subtract
2579 * 1 from the current_context and AND it to itself.
2583 * 101 & 100 = 100 (clearing bit zero)
2586 * 1010 & 1001 = 1000 (clearing bit 1)
2588 * The least significant bit can be cleared this way, and it
2589 * just so happens that it is the same bit corresponding to
2590 * the current context.
2592 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
2593 * is set when a recursion is detected at the current context, and if
2594 * the TRANSITION bit is already set, it will fail the recursion.
2595 * This is needed because there's a lag between the changing of
2596 * interrupt context and updating the preempt count. In this case,
2597 * a false positive will be found. To handle this, one extra recursion
2598 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
2599 * bit is already set, then it is considered a recursion and the function
2600 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
2602 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
2603 * to be cleared. Even if it wasn't the context that set it. That is,
2604 * if an interrupt comes in while NORMAL bit is set and the ring buffer
2605 * is called before preempt_count() is updated, since the check will
2606 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
2607 * NMI then comes in, it will set the NMI bit, but when the NMI code
2608 * does the trace_recursive_unlock() it will clear the TRANSTION bit
2609 * and leave the NMI bit set. But this is fine, because the interrupt
2610 * code that set the TRANSITION bit will then clear the NMI bit when it
2611 * calls trace_recursive_unlock(). If another NMI comes in, it will
2612 * set the TRANSITION bit and continue.
2614 * Note: The TRANSITION bit only handles a single transition between context.
2617 static __always_inline int
2618 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2620 unsigned int val = cpu_buffer->current_context;
2623 if (in_interrupt()) {
2629 bit = RB_CTX_SOFTIRQ;
2631 bit = RB_CTX_NORMAL;
2633 if (unlikely(val & (1 << bit))) {
2635 * It is possible that this was called by transitioning
2636 * between interrupt context, and preempt_count() has not
2637 * been updated yet. In this case, use the TRANSITION bit.
2639 bit = RB_CTX_TRANSITION;
2640 if (val & (1 << bit))
2645 cpu_buffer->current_context = val;
2650 static __always_inline void
2651 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2653 cpu_buffer->current_context &= cpu_buffer->current_context - 1;
2657 * ring_buffer_unlock_commit - commit a reserved
2658 * @buffer: The buffer to commit to
2659 * @event: The event pointer to commit.
2661 * This commits the data to the ring buffer, and releases any locks held.
2663 * Must be paired with ring_buffer_lock_reserve.
2665 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2666 struct ring_buffer_event *event)
2668 struct ring_buffer_per_cpu *cpu_buffer;
2669 int cpu = raw_smp_processor_id();
2671 cpu_buffer = buffer->buffers[cpu];
2673 rb_commit(cpu_buffer, event);
2675 rb_wakeups(buffer, cpu_buffer);
2677 trace_recursive_unlock(cpu_buffer);
2679 preempt_enable_notrace();
2683 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2685 static noinline void
2686 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2687 struct rb_event_info *info)
2689 WARN_ONCE(info->delta > (1ULL << 59),
2690 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2691 (unsigned long long)info->delta,
2692 (unsigned long long)info->ts,
2693 (unsigned long long)cpu_buffer->write_stamp,
2694 sched_clock_stable() ? "" :
2695 "If you just came from a suspend/resume,\n"
2696 "please switch to the trace global clock:\n"
2697 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2698 info->add_timestamp = 1;
2701 static struct ring_buffer_event *
2702 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2703 struct rb_event_info *info)
2705 struct ring_buffer_event *event;
2706 struct buffer_page *tail_page;
2707 unsigned long tail, write;
2710 * If the time delta since the last event is too big to
2711 * hold in the time field of the event, then we append a
2712 * TIME EXTEND event ahead of the data event.
2714 if (unlikely(info->add_timestamp))
2715 info->length += RB_LEN_TIME_EXTEND;
2717 /* Don't let the compiler play games with cpu_buffer->tail_page */
2718 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2719 write = local_add_return(info->length, &tail_page->write);
2721 /* set write to only the index of the write */
2722 write &= RB_WRITE_MASK;
2723 tail = write - info->length;
2726 * If this is the first commit on the page, then it has the same
2727 * timestamp as the page itself.
2732 /* See if we shot pass the end of this buffer page */
2733 if (unlikely(write > BUF_PAGE_SIZE))
2734 return rb_move_tail(cpu_buffer, tail, info);
2736 /* We reserved something on the buffer */
2738 event = __rb_page_index(tail_page, tail);
2739 rb_update_event(cpu_buffer, event, info);
2741 local_inc(&tail_page->entries);
2744 * If this is the first commit on the page, then update
2748 tail_page->page->time_stamp = info->ts;
2750 /* account for these added bytes */
2751 local_add(info->length, &cpu_buffer->entries_bytes);
2756 static __always_inline struct ring_buffer_event *
2757 rb_reserve_next_event(struct ring_buffer *buffer,
2758 struct ring_buffer_per_cpu *cpu_buffer,
2759 unsigned long length)
2761 struct ring_buffer_event *event;
2762 struct rb_event_info info;
2766 rb_start_commit(cpu_buffer);
2768 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2770 * Due to the ability to swap a cpu buffer from a buffer
2771 * it is possible it was swapped before we committed.
2772 * (committing stops a swap). We check for it here and
2773 * if it happened, we have to fail the write.
2776 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2777 local_dec(&cpu_buffer->committing);
2778 local_dec(&cpu_buffer->commits);
2783 info.length = rb_calculate_event_length(length);
2785 info.add_timestamp = 0;
2789 * We allow for interrupts to reenter here and do a trace.
2790 * If one does, it will cause this original code to loop
2791 * back here. Even with heavy interrupts happening, this
2792 * should only happen a few times in a row. If this happens
2793 * 1000 times in a row, there must be either an interrupt
2794 * storm or we have something buggy.
2797 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2800 info.ts = rb_time_stamp(cpu_buffer->buffer);
2801 diff = info.ts - cpu_buffer->write_stamp;
2803 /* make sure this diff is calculated here */
2806 /* Did the write stamp get updated already? */
2807 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2809 if (unlikely(test_time_stamp(info.delta)))
2810 rb_handle_timestamp(cpu_buffer, &info);
2813 event = __rb_reserve_next(cpu_buffer, &info);
2815 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2816 if (info.add_timestamp)
2817 info.length -= RB_LEN_TIME_EXTEND;
2827 rb_end_commit(cpu_buffer);
2832 * ring_buffer_lock_reserve - reserve a part of the buffer
2833 * @buffer: the ring buffer to reserve from
2834 * @length: the length of the data to reserve (excluding event header)
2836 * Returns a reseverd event on the ring buffer to copy directly to.
2837 * The user of this interface will need to get the body to write into
2838 * and can use the ring_buffer_event_data() interface.
2840 * The length is the length of the data needed, not the event length
2841 * which also includes the event header.
2843 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2844 * If NULL is returned, then nothing has been allocated or locked.
2846 struct ring_buffer_event *
2847 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2849 struct ring_buffer_per_cpu *cpu_buffer;
2850 struct ring_buffer_event *event;
2853 /* If we are tracing schedule, we don't want to recurse */
2854 preempt_disable_notrace();
2856 if (unlikely(atomic_read(&buffer->record_disabled)))
2859 cpu = raw_smp_processor_id();
2861 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2864 cpu_buffer = buffer->buffers[cpu];
2866 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2869 if (unlikely(length > BUF_MAX_DATA_SIZE))
2872 if (unlikely(trace_recursive_lock(cpu_buffer)))
2875 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2882 trace_recursive_unlock(cpu_buffer);
2884 preempt_enable_notrace();
2887 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2890 * Decrement the entries to the page that an event is on.
2891 * The event does not even need to exist, only the pointer
2892 * to the page it is on. This may only be called before the commit
2896 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2897 struct ring_buffer_event *event)
2899 unsigned long addr = (unsigned long)event;
2900 struct buffer_page *bpage = cpu_buffer->commit_page;
2901 struct buffer_page *start;
2905 /* Do the likely case first */
2906 if (likely(bpage->page == (void *)addr)) {
2907 local_dec(&bpage->entries);
2912 * Because the commit page may be on the reader page we
2913 * start with the next page and check the end loop there.
2915 rb_inc_page(cpu_buffer, &bpage);
2918 if (bpage->page == (void *)addr) {
2919 local_dec(&bpage->entries);
2922 rb_inc_page(cpu_buffer, &bpage);
2923 } while (bpage != start);
2925 /* commit not part of this buffer?? */
2926 RB_WARN_ON(cpu_buffer, 1);
2930 * ring_buffer_commit_discard - discard an event that has not been committed
2931 * @buffer: the ring buffer
2932 * @event: non committed event to discard
2934 * Sometimes an event that is in the ring buffer needs to be ignored.
2935 * This function lets the user discard an event in the ring buffer
2936 * and then that event will not be read later.
2938 * This function only works if it is called before the the item has been
2939 * committed. It will try to free the event from the ring buffer
2940 * if another event has not been added behind it.
2942 * If another event has been added behind it, it will set the event
2943 * up as discarded, and perform the commit.
2945 * If this function is called, do not call ring_buffer_unlock_commit on
2948 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2949 struct ring_buffer_event *event)
2951 struct ring_buffer_per_cpu *cpu_buffer;
2954 /* The event is discarded regardless */
2955 rb_event_discard(event);
2957 cpu = smp_processor_id();
2958 cpu_buffer = buffer->buffers[cpu];
2961 * This must only be called if the event has not been
2962 * committed yet. Thus we can assume that preemption
2963 * is still disabled.
2965 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2967 rb_decrement_entry(cpu_buffer, event);
2968 if (rb_try_to_discard(cpu_buffer, event))
2972 * The commit is still visible by the reader, so we
2973 * must still update the timestamp.
2975 rb_update_write_stamp(cpu_buffer, event);
2977 rb_end_commit(cpu_buffer);
2979 trace_recursive_unlock(cpu_buffer);
2981 preempt_enable_notrace();
2984 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2987 * ring_buffer_write - write data to the buffer without reserving
2988 * @buffer: The ring buffer to write to.
2989 * @length: The length of the data being written (excluding the event header)
2990 * @data: The data to write to the buffer.
2992 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2993 * one function. If you already have the data to write to the buffer, it
2994 * may be easier to simply call this function.
2996 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2997 * and not the length of the event which would hold the header.
2999 int ring_buffer_write(struct ring_buffer *buffer,
3000 unsigned long length,
3003 struct ring_buffer_per_cpu *cpu_buffer;
3004 struct ring_buffer_event *event;
3009 preempt_disable_notrace();
3011 if (atomic_read(&buffer->record_disabled))
3014 cpu = raw_smp_processor_id();
3016 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3019 cpu_buffer = buffer->buffers[cpu];
3021 if (atomic_read(&cpu_buffer->record_disabled))
3024 if (length > BUF_MAX_DATA_SIZE)
3027 if (unlikely(trace_recursive_lock(cpu_buffer)))
3030 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3034 body = rb_event_data(event);
3036 memcpy(body, data, length);
3038 rb_commit(cpu_buffer, event);
3040 rb_wakeups(buffer, cpu_buffer);
3045 trace_recursive_unlock(cpu_buffer);
3048 preempt_enable_notrace();
3052 EXPORT_SYMBOL_GPL(ring_buffer_write);
3054 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3056 struct buffer_page *reader = cpu_buffer->reader_page;
3057 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3058 struct buffer_page *commit = cpu_buffer->commit_page;
3060 /* In case of error, head will be NULL */
3061 if (unlikely(!head))
3064 /* Reader should exhaust content in reader page */
3065 if (reader->read != rb_page_commit(reader))
3069 * If writers are committing on the reader page, knowing all
3070 * committed content has been read, the ring buffer is empty.
3072 if (commit == reader)
3076 * If writers are committing on a page other than reader page
3077 * and head page, there should always be content to read.
3083 * Writers are committing on the head page, we just need
3084 * to care about there're committed data, and the reader will
3085 * swap reader page with head page when it is to read data.
3087 return rb_page_commit(commit) == 0;
3091 * ring_buffer_record_disable - stop all writes into the buffer
3092 * @buffer: The ring buffer to stop writes to.
3094 * This prevents all writes to the buffer. Any attempt to write
3095 * to the buffer after this will fail and return NULL.
3097 * The caller should call synchronize_sched() after this.
3099 void ring_buffer_record_disable(struct ring_buffer *buffer)
3101 atomic_inc(&buffer->record_disabled);
3103 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3106 * ring_buffer_record_enable - enable writes to the buffer
3107 * @buffer: The ring buffer to enable writes
3109 * Note, multiple disables will need the same number of enables
3110 * to truly enable the writing (much like preempt_disable).
3112 void ring_buffer_record_enable(struct ring_buffer *buffer)
3114 atomic_dec(&buffer->record_disabled);
3116 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3119 * ring_buffer_record_off - stop all writes into the buffer
3120 * @buffer: The ring buffer to stop writes to.
3122 * This prevents all writes to the buffer. Any attempt to write
3123 * to the buffer after this will fail and return NULL.
3125 * This is different than ring_buffer_record_disable() as
3126 * it works like an on/off switch, where as the disable() version
3127 * must be paired with a enable().
3129 void ring_buffer_record_off(struct ring_buffer *buffer)
3132 unsigned int new_rd;
3135 rd = atomic_read(&buffer->record_disabled);
3136 new_rd = rd | RB_BUFFER_OFF;
3137 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3139 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3142 * ring_buffer_record_on - restart writes into the buffer
3143 * @buffer: The ring buffer to start writes to.
3145 * This enables all writes to the buffer that was disabled by
3146 * ring_buffer_record_off().
3148 * This is different than ring_buffer_record_enable() as
3149 * it works like an on/off switch, where as the enable() version
3150 * must be paired with a disable().
3152 void ring_buffer_record_on(struct ring_buffer *buffer)
3155 unsigned int new_rd;
3158 rd = atomic_read(&buffer->record_disabled);
3159 new_rd = rd & ~RB_BUFFER_OFF;
3160 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3162 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3165 * ring_buffer_record_is_on - return true if the ring buffer can write
3166 * @buffer: The ring buffer to see if write is enabled
3168 * Returns true if the ring buffer is in a state that it accepts writes.
3170 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3172 return !atomic_read(&buffer->record_disabled);
3176 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3177 * @buffer: The ring buffer to see if write is set enabled
3179 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3180 * Note that this does NOT mean it is in a writable state.
3182 * It may return true when the ring buffer has been disabled by
3183 * ring_buffer_record_disable(), as that is a temporary disabling of
3186 int ring_buffer_record_is_set_on(struct ring_buffer *buffer)
3188 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3192 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3193 * @buffer: The ring buffer to stop writes to.
3194 * @cpu: The CPU buffer to stop
3196 * This prevents all writes to the buffer. Any attempt to write
3197 * to the buffer after this will fail and return NULL.
3199 * The caller should call synchronize_sched() after this.
3201 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3203 struct ring_buffer_per_cpu *cpu_buffer;
3205 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3208 cpu_buffer = buffer->buffers[cpu];
3209 atomic_inc(&cpu_buffer->record_disabled);
3211 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3214 * ring_buffer_record_enable_cpu - enable writes to the buffer
3215 * @buffer: The ring buffer to enable writes
3216 * @cpu: The CPU to enable.
3218 * Note, multiple disables will need the same number of enables
3219 * to truly enable the writing (much like preempt_disable).
3221 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3223 struct ring_buffer_per_cpu *cpu_buffer;
3225 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3228 cpu_buffer = buffer->buffers[cpu];
3229 atomic_dec(&cpu_buffer->record_disabled);
3231 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3234 * The total entries in the ring buffer is the running counter
3235 * of entries entered into the ring buffer, minus the sum of
3236 * the entries read from the ring buffer and the number of
3237 * entries that were overwritten.
3239 static inline unsigned long
3240 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3242 return local_read(&cpu_buffer->entries) -
3243 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3247 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3248 * @buffer: The ring buffer
3249 * @cpu: The per CPU buffer to read from.
3251 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3253 unsigned long flags;
3254 struct ring_buffer_per_cpu *cpu_buffer;
3255 struct buffer_page *bpage;
3258 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3261 cpu_buffer = buffer->buffers[cpu];
3262 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3264 * if the tail is on reader_page, oldest time stamp is on the reader
3267 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3268 bpage = cpu_buffer->reader_page;
3270 bpage = rb_set_head_page(cpu_buffer);
3272 ret = bpage->page->time_stamp;
3273 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3277 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3280 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3281 * @buffer: The ring buffer
3282 * @cpu: The per CPU buffer to read from.
3284 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3286 struct ring_buffer_per_cpu *cpu_buffer;
3289 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3292 cpu_buffer = buffer->buffers[cpu];
3293 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3297 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3300 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3301 * @buffer: The ring buffer
3302 * @cpu: The per CPU buffer to get the entries from.
3304 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3306 struct ring_buffer_per_cpu *cpu_buffer;
3308 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3311 cpu_buffer = buffer->buffers[cpu];
3313 return rb_num_of_entries(cpu_buffer);
3315 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3318 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3319 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3320 * @buffer: The ring buffer
3321 * @cpu: The per CPU buffer to get the number of overruns from
3323 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3325 struct ring_buffer_per_cpu *cpu_buffer;
3328 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3331 cpu_buffer = buffer->buffers[cpu];
3332 ret = local_read(&cpu_buffer->overrun);
3336 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3339 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3340 * commits failing due to the buffer wrapping around while there are uncommitted
3341 * events, such as during an interrupt storm.
3342 * @buffer: The ring buffer
3343 * @cpu: The per CPU buffer to get the number of overruns from
3346 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3348 struct ring_buffer_per_cpu *cpu_buffer;
3351 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3354 cpu_buffer = buffer->buffers[cpu];
3355 ret = local_read(&cpu_buffer->commit_overrun);
3359 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3362 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3363 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3364 * @buffer: The ring buffer
3365 * @cpu: The per CPU buffer to get the number of overruns from
3368 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3370 struct ring_buffer_per_cpu *cpu_buffer;
3373 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3376 cpu_buffer = buffer->buffers[cpu];
3377 ret = local_read(&cpu_buffer->dropped_events);
3381 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3384 * ring_buffer_read_events_cpu - get the number of events successfully read
3385 * @buffer: The ring buffer
3386 * @cpu: The per CPU buffer to get the number of events read
3389 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3391 struct ring_buffer_per_cpu *cpu_buffer;
3393 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3396 cpu_buffer = buffer->buffers[cpu];
3397 return cpu_buffer->read;
3399 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3402 * ring_buffer_entries - get the number of entries in a buffer
3403 * @buffer: The ring buffer
3405 * Returns the total number of entries in the ring buffer
3408 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3410 struct ring_buffer_per_cpu *cpu_buffer;
3411 unsigned long entries = 0;
3414 /* if you care about this being correct, lock the buffer */
3415 for_each_buffer_cpu(buffer, cpu) {
3416 cpu_buffer = buffer->buffers[cpu];
3417 entries += rb_num_of_entries(cpu_buffer);
3422 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3425 * ring_buffer_overruns - get the number of overruns in buffer
3426 * @buffer: The ring buffer
3428 * Returns the total number of overruns in the ring buffer
3431 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3433 struct ring_buffer_per_cpu *cpu_buffer;
3434 unsigned long overruns = 0;
3437 /* if you care about this being correct, lock the buffer */
3438 for_each_buffer_cpu(buffer, cpu) {
3439 cpu_buffer = buffer->buffers[cpu];
3440 overruns += local_read(&cpu_buffer->overrun);
3445 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3447 static void rb_iter_reset(struct ring_buffer_iter *iter)
3449 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3451 /* Iterator usage is expected to have record disabled */
3452 iter->head_page = cpu_buffer->reader_page;
3453 iter->head = cpu_buffer->reader_page->read;
3455 iter->cache_reader_page = iter->head_page;
3456 iter->cache_read = cpu_buffer->read;
3459 iter->read_stamp = cpu_buffer->read_stamp;
3461 iter->read_stamp = iter->head_page->page->time_stamp;
3465 * ring_buffer_iter_reset - reset an iterator
3466 * @iter: The iterator to reset
3468 * Resets the iterator, so that it will start from the beginning
3471 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3473 struct ring_buffer_per_cpu *cpu_buffer;
3474 unsigned long flags;
3479 cpu_buffer = iter->cpu_buffer;
3481 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3482 rb_iter_reset(iter);
3483 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3485 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3488 * ring_buffer_iter_empty - check if an iterator has no more to read
3489 * @iter: The iterator to check
3491 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3493 struct ring_buffer_per_cpu *cpu_buffer;
3494 struct buffer_page *reader;
3495 struct buffer_page *head_page;
3496 struct buffer_page *commit_page;
3499 cpu_buffer = iter->cpu_buffer;
3501 /* Remember, trace recording is off when iterator is in use */
3502 reader = cpu_buffer->reader_page;
3503 head_page = cpu_buffer->head_page;
3504 commit_page = cpu_buffer->commit_page;
3505 commit = rb_page_commit(commit_page);
3507 return ((iter->head_page == commit_page && iter->head == commit) ||
3508 (iter->head_page == reader && commit_page == head_page &&
3509 head_page->read == commit &&
3510 iter->head == rb_page_commit(cpu_buffer->reader_page)));
3512 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3515 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3516 struct ring_buffer_event *event)
3520 switch (event->type_len) {
3521 case RINGBUF_TYPE_PADDING:
3524 case RINGBUF_TYPE_TIME_EXTEND:
3525 delta = event->array[0];
3527 delta += event->time_delta;
3528 cpu_buffer->read_stamp += delta;
3531 case RINGBUF_TYPE_TIME_STAMP:
3532 /* FIXME: not implemented */
3535 case RINGBUF_TYPE_DATA:
3536 cpu_buffer->read_stamp += event->time_delta;
3546 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3547 struct ring_buffer_event *event)
3551 switch (event->type_len) {
3552 case RINGBUF_TYPE_PADDING:
3555 case RINGBUF_TYPE_TIME_EXTEND:
3556 delta = event->array[0];
3558 delta += event->time_delta;
3559 iter->read_stamp += delta;
3562 case RINGBUF_TYPE_TIME_STAMP:
3563 /* FIXME: not implemented */
3566 case RINGBUF_TYPE_DATA:
3567 iter->read_stamp += event->time_delta;
3576 static struct buffer_page *
3577 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3579 struct buffer_page *reader = NULL;
3580 unsigned long overwrite;
3581 unsigned long flags;
3585 local_irq_save(flags);
3586 arch_spin_lock(&cpu_buffer->lock);
3590 * This should normally only loop twice. But because the
3591 * start of the reader inserts an empty page, it causes
3592 * a case where we will loop three times. There should be no
3593 * reason to loop four times (that I know of).
3595 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3600 reader = cpu_buffer->reader_page;
3602 /* If there's more to read, return this page */
3603 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3606 /* Never should we have an index greater than the size */
3607 if (RB_WARN_ON(cpu_buffer,
3608 cpu_buffer->reader_page->read > rb_page_size(reader)))
3611 /* check if we caught up to the tail */
3613 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3616 /* Don't bother swapping if the ring buffer is empty */
3617 if (rb_num_of_entries(cpu_buffer) == 0)
3621 * Reset the reader page to size zero.
3623 local_set(&cpu_buffer->reader_page->write, 0);
3624 local_set(&cpu_buffer->reader_page->entries, 0);
3625 local_set(&cpu_buffer->reader_page->page->commit, 0);
3626 cpu_buffer->reader_page->real_end = 0;
3630 * Splice the empty reader page into the list around the head.
3632 reader = rb_set_head_page(cpu_buffer);
3635 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3636 cpu_buffer->reader_page->list.prev = reader->list.prev;
3639 * cpu_buffer->pages just needs to point to the buffer, it
3640 * has no specific buffer page to point to. Lets move it out
3641 * of our way so we don't accidentally swap it.
3643 cpu_buffer->pages = reader->list.prev;
3645 /* The reader page will be pointing to the new head */
3646 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3649 * We want to make sure we read the overruns after we set up our
3650 * pointers to the next object. The writer side does a
3651 * cmpxchg to cross pages which acts as the mb on the writer
3652 * side. Note, the reader will constantly fail the swap
3653 * while the writer is updating the pointers, so this
3654 * guarantees that the overwrite recorded here is the one we
3655 * want to compare with the last_overrun.
3658 overwrite = local_read(&(cpu_buffer->overrun));
3661 * Here's the tricky part.
3663 * We need to move the pointer past the header page.
3664 * But we can only do that if a writer is not currently
3665 * moving it. The page before the header page has the
3666 * flag bit '1' set if it is pointing to the page we want.
3667 * but if the writer is in the process of moving it
3668 * than it will be '2' or already moved '0'.
3671 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3674 * If we did not convert it, then we must try again.
3680 * Yeah! We succeeded in replacing the page.
3682 * Now make the new head point back to the reader page.
3684 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3685 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3687 /* Finally update the reader page to the new head */
3688 cpu_buffer->reader_page = reader;
3689 cpu_buffer->reader_page->read = 0;
3691 if (overwrite != cpu_buffer->last_overrun) {
3692 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3693 cpu_buffer->last_overrun = overwrite;
3699 /* Update the read_stamp on the first event */
3700 if (reader && reader->read == 0)
3701 cpu_buffer->read_stamp = reader->page->time_stamp;
3703 arch_spin_unlock(&cpu_buffer->lock);
3704 local_irq_restore(flags);
3707 * The writer has preempt disable, wait for it. But not forever
3708 * Although, 1 second is pretty much "forever"
3710 #define USECS_WAIT 1000000
3711 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
3712 /* If the write is past the end of page, a writer is still updating it */
3713 if (likely(!reader || rb_page_write(reader) <= BUF_PAGE_SIZE))
3718 /* Get the latest version of the reader write value */
3722 /* The writer is not moving forward? Something is wrong */
3723 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
3727 * Make sure we see any padding after the write update
3728 * (see rb_reset_tail())
3736 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3738 struct ring_buffer_event *event;
3739 struct buffer_page *reader;
3742 reader = rb_get_reader_page(cpu_buffer);
3744 /* This function should not be called when buffer is empty */
3745 if (RB_WARN_ON(cpu_buffer, !reader))
3748 event = rb_reader_event(cpu_buffer);
3750 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3753 rb_update_read_stamp(cpu_buffer, event);
3755 length = rb_event_length(event);
3756 cpu_buffer->reader_page->read += length;
3759 static void rb_advance_iter(struct ring_buffer_iter *iter)
3761 struct ring_buffer_per_cpu *cpu_buffer;
3762 struct ring_buffer_event *event;
3765 cpu_buffer = iter->cpu_buffer;
3768 * Check if we are at the end of the buffer.
3770 if (iter->head >= rb_page_size(iter->head_page)) {
3771 /* discarded commits can make the page empty */
3772 if (iter->head_page == cpu_buffer->commit_page)
3778 event = rb_iter_head_event(iter);
3780 length = rb_event_length(event);
3783 * This should not be called to advance the header if we are
3784 * at the tail of the buffer.
3786 if (RB_WARN_ON(cpu_buffer,
3787 (iter->head_page == cpu_buffer->commit_page) &&
3788 (iter->head + length > rb_commit_index(cpu_buffer))))
3791 rb_update_iter_read_stamp(iter, event);
3793 iter->head += length;
3795 /* check for end of page padding */
3796 if ((iter->head >= rb_page_size(iter->head_page)) &&
3797 (iter->head_page != cpu_buffer->commit_page))
3801 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3803 return cpu_buffer->lost_events;
3806 static struct ring_buffer_event *
3807 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3808 unsigned long *lost_events)
3810 struct ring_buffer_event *event;
3811 struct buffer_page *reader;
3816 * We repeat when a time extend is encountered.
3817 * Since the time extend is always attached to a data event,
3818 * we should never loop more than once.
3819 * (We never hit the following condition more than twice).
3821 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3824 reader = rb_get_reader_page(cpu_buffer);
3828 event = rb_reader_event(cpu_buffer);
3830 switch (event->type_len) {
3831 case RINGBUF_TYPE_PADDING:
3832 if (rb_null_event(event))
3833 RB_WARN_ON(cpu_buffer, 1);
3835 * Because the writer could be discarding every
3836 * event it creates (which would probably be bad)
3837 * if we were to go back to "again" then we may never
3838 * catch up, and will trigger the warn on, or lock
3839 * the box. Return the padding, and we will release
3840 * the current locks, and try again.
3844 case RINGBUF_TYPE_TIME_EXTEND:
3845 /* Internal data, OK to advance */
3846 rb_advance_reader(cpu_buffer);
3849 case RINGBUF_TYPE_TIME_STAMP:
3850 /* FIXME: not implemented */
3851 rb_advance_reader(cpu_buffer);
3854 case RINGBUF_TYPE_DATA:
3856 *ts = cpu_buffer->read_stamp + event->time_delta;
3857 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3858 cpu_buffer->cpu, ts);
3861 *lost_events = rb_lost_events(cpu_buffer);
3870 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3872 static struct ring_buffer_event *
3873 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3875 struct ring_buffer *buffer;
3876 struct ring_buffer_per_cpu *cpu_buffer;
3877 struct ring_buffer_event *event;
3880 cpu_buffer = iter->cpu_buffer;
3881 buffer = cpu_buffer->buffer;
3884 * Check if someone performed a consuming read to
3885 * the buffer. A consuming read invalidates the iterator
3886 * and we need to reset the iterator in this case.
3888 if (unlikely(iter->cache_read != cpu_buffer->read ||
3889 iter->cache_reader_page != cpu_buffer->reader_page))
3890 rb_iter_reset(iter);
3893 if (ring_buffer_iter_empty(iter))
3897 * We repeat when a time extend is encountered or we hit
3898 * the end of the page. Since the time extend is always attached
3899 * to a data event, we should never loop more than three times.
3900 * Once for going to next page, once on time extend, and
3901 * finally once to get the event.
3902 * (We never hit the following condition more than thrice).
3904 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3907 if (rb_per_cpu_empty(cpu_buffer))
3910 if (iter->head >= rb_page_size(iter->head_page)) {
3915 event = rb_iter_head_event(iter);
3917 switch (event->type_len) {
3918 case RINGBUF_TYPE_PADDING:
3919 if (rb_null_event(event)) {
3923 rb_advance_iter(iter);
3926 case RINGBUF_TYPE_TIME_EXTEND:
3927 /* Internal data, OK to advance */
3928 rb_advance_iter(iter);
3931 case RINGBUF_TYPE_TIME_STAMP:
3932 /* FIXME: not implemented */
3933 rb_advance_iter(iter);
3936 case RINGBUF_TYPE_DATA:
3938 *ts = iter->read_stamp + event->time_delta;
3939 ring_buffer_normalize_time_stamp(buffer,
3940 cpu_buffer->cpu, ts);
3950 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3952 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3954 if (likely(!in_nmi())) {
3955 raw_spin_lock(&cpu_buffer->reader_lock);
3960 * If an NMI die dumps out the content of the ring buffer
3961 * trylock must be used to prevent a deadlock if the NMI
3962 * preempted a task that holds the ring buffer locks. If
3963 * we get the lock then all is fine, if not, then continue
3964 * to do the read, but this can corrupt the ring buffer,
3965 * so it must be permanently disabled from future writes.
3966 * Reading from NMI is a oneshot deal.
3968 if (raw_spin_trylock(&cpu_buffer->reader_lock))
3971 /* Continue without locking, but disable the ring buffer */
3972 atomic_inc(&cpu_buffer->record_disabled);
3977 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3980 raw_spin_unlock(&cpu_buffer->reader_lock);
3985 * ring_buffer_peek - peek at the next event to be read
3986 * @buffer: The ring buffer to read
3987 * @cpu: The cpu to peak at
3988 * @ts: The timestamp counter of this event.
3989 * @lost_events: a variable to store if events were lost (may be NULL)
3991 * This will return the event that will be read next, but does
3992 * not consume the data.
3994 struct ring_buffer_event *
3995 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3996 unsigned long *lost_events)
3998 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3999 struct ring_buffer_event *event;
4000 unsigned long flags;
4003 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4007 local_irq_save(flags);
4008 dolock = rb_reader_lock(cpu_buffer);
4009 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4010 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4011 rb_advance_reader(cpu_buffer);
4012 rb_reader_unlock(cpu_buffer, dolock);
4013 local_irq_restore(flags);
4015 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4022 * ring_buffer_iter_peek - peek at the next event to be read
4023 * @iter: The ring buffer iterator
4024 * @ts: The timestamp counter of this event.
4026 * This will return the event that will be read next, but does
4027 * not increment the iterator.
4029 struct ring_buffer_event *
4030 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4032 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4033 struct ring_buffer_event *event;
4034 unsigned long flags;
4037 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4038 event = rb_iter_peek(iter, ts);
4039 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4041 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4048 * ring_buffer_consume - return an event and consume it
4049 * @buffer: The ring buffer to get the next event from
4050 * @cpu: the cpu to read the buffer from
4051 * @ts: a variable to store the timestamp (may be NULL)
4052 * @lost_events: a variable to store if events were lost (may be NULL)
4054 * Returns the next event in the ring buffer, and that event is consumed.
4055 * Meaning, that sequential reads will keep returning a different event,
4056 * and eventually empty the ring buffer if the producer is slower.
4058 struct ring_buffer_event *
4059 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
4060 unsigned long *lost_events)
4062 struct ring_buffer_per_cpu *cpu_buffer;
4063 struct ring_buffer_event *event = NULL;
4064 unsigned long flags;
4068 /* might be called in atomic */
4071 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4074 cpu_buffer = buffer->buffers[cpu];
4075 local_irq_save(flags);
4076 dolock = rb_reader_lock(cpu_buffer);
4078 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4080 cpu_buffer->lost_events = 0;
4081 rb_advance_reader(cpu_buffer);
4084 rb_reader_unlock(cpu_buffer, dolock);
4085 local_irq_restore(flags);
4090 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4095 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4098 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4099 * @buffer: The ring buffer to read from
4100 * @cpu: The cpu buffer to iterate over
4101 * @flags: gfp flags to use for memory allocation
4103 * This performs the initial preparations necessary to iterate
4104 * through the buffer. Memory is allocated, buffer recording
4105 * is disabled, and the iterator pointer is returned to the caller.
4107 * Disabling buffer recordng prevents the reading from being
4108 * corrupted. This is not a consuming read, so a producer is not
4111 * After a sequence of ring_buffer_read_prepare calls, the user is
4112 * expected to make at least one call to ring_buffer_read_prepare_sync.
4113 * Afterwards, ring_buffer_read_start is invoked to get things going
4116 * This overall must be paired with ring_buffer_read_finish.
4118 struct ring_buffer_iter *
4119 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu, gfp_t flags)
4121 struct ring_buffer_per_cpu *cpu_buffer;
4122 struct ring_buffer_iter *iter;
4124 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4127 iter = kmalloc(sizeof(*iter), flags);
4131 cpu_buffer = buffer->buffers[cpu];
4133 iter->cpu_buffer = cpu_buffer;
4135 atomic_inc(&buffer->resize_disabled);
4136 atomic_inc(&cpu_buffer->record_disabled);
4140 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4143 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4145 * All previously invoked ring_buffer_read_prepare calls to prepare
4146 * iterators will be synchronized. Afterwards, read_buffer_read_start
4147 * calls on those iterators are allowed.
4150 ring_buffer_read_prepare_sync(void)
4152 synchronize_sched();
4154 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4157 * ring_buffer_read_start - start a non consuming read of the buffer
4158 * @iter: The iterator returned by ring_buffer_read_prepare
4160 * This finalizes the startup of an iteration through the buffer.
4161 * The iterator comes from a call to ring_buffer_read_prepare and
4162 * an intervening ring_buffer_read_prepare_sync must have been
4165 * Must be paired with ring_buffer_read_finish.
4168 ring_buffer_read_start(struct ring_buffer_iter *iter)
4170 struct ring_buffer_per_cpu *cpu_buffer;
4171 unsigned long flags;
4176 cpu_buffer = iter->cpu_buffer;
4178 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4179 arch_spin_lock(&cpu_buffer->lock);
4180 rb_iter_reset(iter);
4181 arch_spin_unlock(&cpu_buffer->lock);
4182 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4184 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4187 * ring_buffer_read_finish - finish reading the iterator of the buffer
4188 * @iter: The iterator retrieved by ring_buffer_start
4190 * This re-enables the recording to the buffer, and frees the
4194 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4196 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4197 unsigned long flags;
4200 * Ring buffer is disabled from recording, here's a good place
4201 * to check the integrity of the ring buffer.
4202 * Must prevent readers from trying to read, as the check
4203 * clears the HEAD page and readers require it.
4205 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4206 rb_check_pages(cpu_buffer);
4207 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4209 atomic_dec(&cpu_buffer->record_disabled);
4210 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4213 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4216 * ring_buffer_read - read the next item in the ring buffer by the iterator
4217 * @iter: The ring buffer iterator
4218 * @ts: The time stamp of the event read.
4220 * This reads the next event in the ring buffer and increments the iterator.
4222 struct ring_buffer_event *
4223 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4225 struct ring_buffer_event *event;
4226 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4227 unsigned long flags;
4229 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4231 event = rb_iter_peek(iter, ts);
4235 if (event->type_len == RINGBUF_TYPE_PADDING)
4238 rb_advance_iter(iter);
4240 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4244 EXPORT_SYMBOL_GPL(ring_buffer_read);
4247 * ring_buffer_size - return the size of the ring buffer (in bytes)
4248 * @buffer: The ring buffer.
4250 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4253 * Earlier, this method returned
4254 * BUF_PAGE_SIZE * buffer->nr_pages
4255 * Since the nr_pages field is now removed, we have converted this to
4256 * return the per cpu buffer value.
4258 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4261 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4263 EXPORT_SYMBOL_GPL(ring_buffer_size);
4266 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4268 rb_head_page_deactivate(cpu_buffer);
4270 cpu_buffer->head_page
4271 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4272 local_set(&cpu_buffer->head_page->write, 0);
4273 local_set(&cpu_buffer->head_page->entries, 0);
4274 local_set(&cpu_buffer->head_page->page->commit, 0);
4276 cpu_buffer->head_page->read = 0;
4278 cpu_buffer->tail_page = cpu_buffer->head_page;
4279 cpu_buffer->commit_page = cpu_buffer->head_page;
4281 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4282 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4283 local_set(&cpu_buffer->reader_page->write, 0);
4284 local_set(&cpu_buffer->reader_page->entries, 0);
4285 local_set(&cpu_buffer->reader_page->page->commit, 0);
4286 cpu_buffer->reader_page->read = 0;
4288 local_set(&cpu_buffer->entries_bytes, 0);
4289 local_set(&cpu_buffer->overrun, 0);
4290 local_set(&cpu_buffer->commit_overrun, 0);
4291 local_set(&cpu_buffer->dropped_events, 0);
4292 local_set(&cpu_buffer->entries, 0);
4293 local_set(&cpu_buffer->committing, 0);
4294 local_set(&cpu_buffer->commits, 0);
4295 cpu_buffer->read = 0;
4296 cpu_buffer->read_bytes = 0;
4298 cpu_buffer->write_stamp = 0;
4299 cpu_buffer->read_stamp = 0;
4301 cpu_buffer->lost_events = 0;
4302 cpu_buffer->last_overrun = 0;
4304 rb_head_page_activate(cpu_buffer);
4308 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4309 * @buffer: The ring buffer to reset a per cpu buffer of
4310 * @cpu: The CPU buffer to be reset
4312 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4314 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4315 unsigned long flags;
4317 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4319 /* prevent another thread from changing buffer sizes */
4320 mutex_lock(&buffer->mutex);
4322 atomic_inc(&buffer->resize_disabled);
4323 atomic_inc(&cpu_buffer->record_disabled);
4325 /* Make sure all commits have finished */
4326 synchronize_sched();
4328 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4330 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4333 arch_spin_lock(&cpu_buffer->lock);
4335 rb_reset_cpu(cpu_buffer);
4337 arch_spin_unlock(&cpu_buffer->lock);
4340 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4342 atomic_dec(&cpu_buffer->record_disabled);
4343 atomic_dec(&buffer->resize_disabled);
4345 mutex_unlock(&buffer->mutex);
4347 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4350 * ring_buffer_reset - reset a ring buffer
4351 * @buffer: The ring buffer to reset all cpu buffers
4353 void ring_buffer_reset(struct ring_buffer *buffer)
4357 for_each_buffer_cpu(buffer, cpu)
4358 ring_buffer_reset_cpu(buffer, cpu);
4360 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4363 * rind_buffer_empty - is the ring buffer empty?
4364 * @buffer: The ring buffer to test
4366 bool ring_buffer_empty(struct ring_buffer *buffer)
4368 struct ring_buffer_per_cpu *cpu_buffer;
4369 unsigned long flags;
4374 /* yes this is racy, but if you don't like the race, lock the buffer */
4375 for_each_buffer_cpu(buffer, cpu) {
4376 cpu_buffer = buffer->buffers[cpu];
4377 local_irq_save(flags);
4378 dolock = rb_reader_lock(cpu_buffer);
4379 ret = rb_per_cpu_empty(cpu_buffer);
4380 rb_reader_unlock(cpu_buffer, dolock);
4381 local_irq_restore(flags);
4389 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4392 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4393 * @buffer: The ring buffer
4394 * @cpu: The CPU buffer to test
4396 bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4398 struct ring_buffer_per_cpu *cpu_buffer;
4399 unsigned long flags;
4403 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4406 cpu_buffer = buffer->buffers[cpu];
4407 local_irq_save(flags);
4408 dolock = rb_reader_lock(cpu_buffer);
4409 ret = rb_per_cpu_empty(cpu_buffer);
4410 rb_reader_unlock(cpu_buffer, dolock);
4411 local_irq_restore(flags);
4415 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4417 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4419 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4420 * @buffer_a: One buffer to swap with
4421 * @buffer_b: The other buffer to swap with
4423 * This function is useful for tracers that want to take a "snapshot"
4424 * of a CPU buffer and has another back up buffer lying around.
4425 * it is expected that the tracer handles the cpu buffer not being
4426 * used at the moment.
4428 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4429 struct ring_buffer *buffer_b, int cpu)
4431 struct ring_buffer_per_cpu *cpu_buffer_a;
4432 struct ring_buffer_per_cpu *cpu_buffer_b;
4435 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4436 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4439 cpu_buffer_a = buffer_a->buffers[cpu];
4440 cpu_buffer_b = buffer_b->buffers[cpu];
4442 /* At least make sure the two buffers are somewhat the same */
4443 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4448 if (atomic_read(&buffer_a->record_disabled))
4451 if (atomic_read(&buffer_b->record_disabled))
4454 if (atomic_read(&cpu_buffer_a->record_disabled))
4457 if (atomic_read(&cpu_buffer_b->record_disabled))
4461 * We can't do a synchronize_sched here because this
4462 * function can be called in atomic context.
4463 * Normally this will be called from the same CPU as cpu.
4464 * If not it's up to the caller to protect this.
4466 atomic_inc(&cpu_buffer_a->record_disabled);
4467 atomic_inc(&cpu_buffer_b->record_disabled);
4470 if (local_read(&cpu_buffer_a->committing))
4472 if (local_read(&cpu_buffer_b->committing))
4475 buffer_a->buffers[cpu] = cpu_buffer_b;
4476 buffer_b->buffers[cpu] = cpu_buffer_a;
4478 cpu_buffer_b->buffer = buffer_a;
4479 cpu_buffer_a->buffer = buffer_b;
4484 atomic_dec(&cpu_buffer_a->record_disabled);
4485 atomic_dec(&cpu_buffer_b->record_disabled);
4489 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4490 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4493 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4494 * @buffer: the buffer to allocate for.
4495 * @cpu: the cpu buffer to allocate.
4497 * This function is used in conjunction with ring_buffer_read_page.
4498 * When reading a full page from the ring buffer, these functions
4499 * can be used to speed up the process. The calling function should
4500 * allocate a few pages first with this function. Then when it
4501 * needs to get pages from the ring buffer, it passes the result
4502 * of this function into ring_buffer_read_page, which will swap
4503 * the page that was allocated, with the read page of the buffer.
4506 * The page allocated, or ERR_PTR
4508 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4510 struct ring_buffer_per_cpu *cpu_buffer;
4511 struct buffer_data_page *bpage = NULL;
4512 unsigned long flags;
4515 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4516 return ERR_PTR(-ENODEV);
4518 cpu_buffer = buffer->buffers[cpu];
4519 local_irq_save(flags);
4520 arch_spin_lock(&cpu_buffer->lock);
4522 if (cpu_buffer->free_page) {
4523 bpage = cpu_buffer->free_page;
4524 cpu_buffer->free_page = NULL;
4527 arch_spin_unlock(&cpu_buffer->lock);
4528 local_irq_restore(flags);
4533 page = alloc_pages_node(cpu_to_node(cpu),
4534 GFP_KERNEL | __GFP_NORETRY, 0);
4536 return ERR_PTR(-ENOMEM);
4538 bpage = page_address(page);
4541 rb_init_page(bpage);
4545 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4548 * ring_buffer_free_read_page - free an allocated read page
4549 * @buffer: the buffer the page was allocate for
4550 * @cpu: the cpu buffer the page came from
4551 * @data: the page to free
4553 * Free a page allocated from ring_buffer_alloc_read_page.
4555 void ring_buffer_free_read_page(struct ring_buffer *buffer, int cpu, void *data)
4557 struct ring_buffer_per_cpu *cpu_buffer;
4558 struct buffer_data_page *bpage = data;
4559 struct page *page = virt_to_page(bpage);
4560 unsigned long flags;
4562 if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
4565 cpu_buffer = buffer->buffers[cpu];
4567 /* If the page is still in use someplace else, we can't reuse it */
4568 if (page_ref_count(page) > 1)
4571 local_irq_save(flags);
4572 arch_spin_lock(&cpu_buffer->lock);
4574 if (!cpu_buffer->free_page) {
4575 cpu_buffer->free_page = bpage;
4579 arch_spin_unlock(&cpu_buffer->lock);
4580 local_irq_restore(flags);
4583 free_page((unsigned long)bpage);
4585 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4588 * ring_buffer_read_page - extract a page from the ring buffer
4589 * @buffer: buffer to extract from
4590 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4591 * @len: amount to extract
4592 * @cpu: the cpu of the buffer to extract
4593 * @full: should the extraction only happen when the page is full.
4595 * This function will pull out a page from the ring buffer and consume it.
4596 * @data_page must be the address of the variable that was returned
4597 * from ring_buffer_alloc_read_page. This is because the page might be used
4598 * to swap with a page in the ring buffer.
4601 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4602 * if (IS_ERR(rpage))
4603 * return PTR_ERR(rpage);
4604 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4606 * process_page(rpage, ret);
4608 * When @full is set, the function will not return true unless
4609 * the writer is off the reader page.
4611 * Note: it is up to the calling functions to handle sleeps and wakeups.
4612 * The ring buffer can be used anywhere in the kernel and can not
4613 * blindly call wake_up. The layer that uses the ring buffer must be
4614 * responsible for that.
4617 * >=0 if data has been transferred, returns the offset of consumed data.
4618 * <0 if no data has been transferred.
4620 int ring_buffer_read_page(struct ring_buffer *buffer,
4621 void **data_page, size_t len, int cpu, int full)
4623 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4624 struct ring_buffer_event *event;
4625 struct buffer_data_page *bpage;
4626 struct buffer_page *reader;
4627 unsigned long missed_events;
4628 unsigned long flags;
4629 unsigned int commit;
4634 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4638 * If len is not big enough to hold the page header, then
4639 * we can not copy anything.
4641 if (len <= BUF_PAGE_HDR_SIZE)
4644 len -= BUF_PAGE_HDR_SIZE;
4653 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4655 reader = rb_get_reader_page(cpu_buffer);
4659 event = rb_reader_event(cpu_buffer);
4661 read = reader->read;
4662 commit = rb_page_commit(reader);
4664 /* Check if any events were dropped */
4665 missed_events = cpu_buffer->lost_events;
4668 * If this page has been partially read or
4669 * if len is not big enough to read the rest of the page or
4670 * a writer is still on the page, then
4671 * we must copy the data from the page to the buffer.
4672 * Otherwise, we can simply swap the page with the one passed in.
4674 if (read || (len < (commit - read)) ||
4675 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4676 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4677 unsigned int rpos = read;
4678 unsigned int pos = 0;
4682 * If a full page is expected, this can still be returned
4683 * if there's been a previous partial read and the
4684 * rest of the page can be read and the commit page is off
4688 (!read || (len < (commit - read)) ||
4689 cpu_buffer->reader_page == cpu_buffer->commit_page))
4692 if (len > (commit - read))
4693 len = (commit - read);
4695 /* Always keep the time extend and data together */
4696 size = rb_event_ts_length(event);
4701 /* save the current timestamp, since the user will need it */
4702 save_timestamp = cpu_buffer->read_stamp;
4704 /* Need to copy one event at a time */
4706 /* We need the size of one event, because
4707 * rb_advance_reader only advances by one event,
4708 * whereas rb_event_ts_length may include the size of
4709 * one or two events.
4710 * We have already ensured there's enough space if this
4711 * is a time extend. */
4712 size = rb_event_length(event);
4713 memcpy(bpage->data + pos, rpage->data + rpos, size);
4717 rb_advance_reader(cpu_buffer);
4718 rpos = reader->read;
4724 event = rb_reader_event(cpu_buffer);
4725 /* Always keep the time extend and data together */
4726 size = rb_event_ts_length(event);
4727 } while (len >= size);
4730 local_set(&bpage->commit, pos);
4731 bpage->time_stamp = save_timestamp;
4733 /* we copied everything to the beginning */
4736 /* update the entry counter */
4737 cpu_buffer->read += rb_page_entries(reader);
4738 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4740 /* swap the pages */
4741 rb_init_page(bpage);
4742 bpage = reader->page;
4743 reader->page = *data_page;
4744 local_set(&reader->write, 0);
4745 local_set(&reader->entries, 0);
4750 * Use the real_end for the data size,
4751 * This gives us a chance to store the lost events
4754 if (reader->real_end)
4755 local_set(&bpage->commit, reader->real_end);
4759 cpu_buffer->lost_events = 0;
4761 commit = local_read(&bpage->commit);
4763 * Set a flag in the commit field if we lost events
4765 if (missed_events) {
4766 /* If there is room at the end of the page to save the
4767 * missed events, then record it there.
4769 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4770 memcpy(&bpage->data[commit], &missed_events,
4771 sizeof(missed_events));
4772 local_add(RB_MISSED_STORED, &bpage->commit);
4773 commit += sizeof(missed_events);
4775 local_add(RB_MISSED_EVENTS, &bpage->commit);
4779 * This page may be off to user land. Zero it out here.
4781 if (commit < BUF_PAGE_SIZE)
4782 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4785 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4790 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4793 * We only allocate new buffers, never free them if the CPU goes down.
4794 * If we were to free the buffer, then the user would lose any trace that was in
4797 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4799 struct ring_buffer *buffer;
4802 unsigned long nr_pages;
4804 buffer = container_of(node, struct ring_buffer, node);
4805 if (cpumask_test_cpu(cpu, buffer->cpumask))
4810 /* check if all cpu sizes are same */
4811 for_each_buffer_cpu(buffer, cpu_i) {
4812 /* fill in the size from first enabled cpu */
4814 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4815 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4820 /* allocate minimum pages, user can later expand it */
4823 buffer->buffers[cpu] =
4824 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4825 if (!buffer->buffers[cpu]) {
4826 WARN(1, "failed to allocate ring buffer on CPU %u\n",
4831 cpumask_set_cpu(cpu, buffer->cpumask);
4835 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4837 * This is a basic integrity check of the ring buffer.
4838 * Late in the boot cycle this test will run when configured in.
4839 * It will kick off a thread per CPU that will go into a loop
4840 * writing to the per cpu ring buffer various sizes of data.
4841 * Some of the data will be large items, some small.
4843 * Another thread is created that goes into a spin, sending out
4844 * IPIs to the other CPUs to also write into the ring buffer.
4845 * this is to test the nesting ability of the buffer.
4847 * Basic stats are recorded and reported. If something in the
4848 * ring buffer should happen that's not expected, a big warning
4849 * is displayed and all ring buffers are disabled.
4851 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4853 struct rb_test_data {
4854 struct ring_buffer *buffer;
4855 unsigned long events;
4856 unsigned long bytes_written;
4857 unsigned long bytes_alloc;
4858 unsigned long bytes_dropped;
4859 unsigned long events_nested;
4860 unsigned long bytes_written_nested;
4861 unsigned long bytes_alloc_nested;
4862 unsigned long bytes_dropped_nested;
4863 int min_size_nested;
4864 int max_size_nested;
4871 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4874 #define RB_TEST_BUFFER_SIZE 1048576
4876 static char rb_string[] __initdata =
4877 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4878 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4879 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4881 static bool rb_test_started __initdata;
4888 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4890 struct ring_buffer_event *event;
4891 struct rb_item *item;
4898 /* Have nested writes different that what is written */
4899 cnt = data->cnt + (nested ? 27 : 0);
4901 /* Multiply cnt by ~e, to make some unique increment */
4902 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4904 len = size + sizeof(struct rb_item);
4906 started = rb_test_started;
4907 /* read rb_test_started before checking buffer enabled */
4910 event = ring_buffer_lock_reserve(data->buffer, len);
4912 /* Ignore dropped events before test starts. */
4915 data->bytes_dropped += len;
4917 data->bytes_dropped_nested += len;
4922 event_len = ring_buffer_event_length(event);
4924 if (RB_WARN_ON(data->buffer, event_len < len))
4927 item = ring_buffer_event_data(event);
4929 memcpy(item->str, rb_string, size);
4932 data->bytes_alloc_nested += event_len;
4933 data->bytes_written_nested += len;
4934 data->events_nested++;
4935 if (!data->min_size_nested || len < data->min_size_nested)
4936 data->min_size_nested = len;
4937 if (len > data->max_size_nested)
4938 data->max_size_nested = len;
4940 data->bytes_alloc += event_len;
4941 data->bytes_written += len;
4943 if (!data->min_size || len < data->min_size)
4944 data->max_size = len;
4945 if (len > data->max_size)
4946 data->max_size = len;
4950 ring_buffer_unlock_commit(data->buffer, event);
4955 static __init int rb_test(void *arg)
4957 struct rb_test_data *data = arg;
4959 while (!kthread_should_stop()) {
4960 rb_write_something(data, false);
4963 set_current_state(TASK_INTERRUPTIBLE);
4964 /* Now sleep between a min of 100-300us and a max of 1ms */
4965 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4971 static __init void rb_ipi(void *ignore)
4973 struct rb_test_data *data;
4974 int cpu = smp_processor_id();
4976 data = &rb_data[cpu];
4977 rb_write_something(data, true);
4980 static __init int rb_hammer_test(void *arg)
4982 while (!kthread_should_stop()) {
4984 /* Send an IPI to all cpus to write data! */
4985 smp_call_function(rb_ipi, NULL, 1);
4986 /* No sleep, but for non preempt, let others run */
4993 static __init int test_ringbuffer(void)
4995 struct task_struct *rb_hammer;
4996 struct ring_buffer *buffer;
5000 pr_info("Running ring buffer tests...\n");
5002 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5003 if (WARN_ON(!buffer))
5006 /* Disable buffer so that threads can't write to it yet */
5007 ring_buffer_record_off(buffer);
5009 for_each_online_cpu(cpu) {
5010 rb_data[cpu].buffer = buffer;
5011 rb_data[cpu].cpu = cpu;
5012 rb_data[cpu].cnt = cpu;
5013 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5014 "rbtester/%d", cpu);
5015 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5016 pr_cont("FAILED\n");
5017 ret = PTR_ERR(rb_threads[cpu]);
5021 kthread_bind(rb_threads[cpu], cpu);
5022 wake_up_process(rb_threads[cpu]);
5025 /* Now create the rb hammer! */
5026 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5027 if (WARN_ON(IS_ERR(rb_hammer))) {
5028 pr_cont("FAILED\n");
5029 ret = PTR_ERR(rb_hammer);
5033 ring_buffer_record_on(buffer);
5035 * Show buffer is enabled before setting rb_test_started.
5036 * Yes there's a small race window where events could be
5037 * dropped and the thread wont catch it. But when a ring
5038 * buffer gets enabled, there will always be some kind of
5039 * delay before other CPUs see it. Thus, we don't care about
5040 * those dropped events. We care about events dropped after
5041 * the threads see that the buffer is active.
5044 rb_test_started = true;
5046 set_current_state(TASK_INTERRUPTIBLE);
5047 /* Just run for 10 seconds */;
5048 schedule_timeout(10 * HZ);
5050 kthread_stop(rb_hammer);
5053 for_each_online_cpu(cpu) {
5054 if (!rb_threads[cpu])
5056 kthread_stop(rb_threads[cpu]);
5059 ring_buffer_free(buffer);
5064 pr_info("finished\n");
5065 for_each_online_cpu(cpu) {
5066 struct ring_buffer_event *event;
5067 struct rb_test_data *data = &rb_data[cpu];
5068 struct rb_item *item;
5069 unsigned long total_events;
5070 unsigned long total_dropped;
5071 unsigned long total_written;
5072 unsigned long total_alloc;
5073 unsigned long total_read = 0;
5074 unsigned long total_size = 0;
5075 unsigned long total_len = 0;
5076 unsigned long total_lost = 0;
5079 int small_event_size;
5083 total_events = data->events + data->events_nested;
5084 total_written = data->bytes_written + data->bytes_written_nested;
5085 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5086 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5088 big_event_size = data->max_size + data->max_size_nested;
5089 small_event_size = data->min_size + data->min_size_nested;
5091 pr_info("CPU %d:\n", cpu);
5092 pr_info(" events: %ld\n", total_events);
5093 pr_info(" dropped bytes: %ld\n", total_dropped);
5094 pr_info(" alloced bytes: %ld\n", total_alloc);
5095 pr_info(" written bytes: %ld\n", total_written);
5096 pr_info(" biggest event: %d\n", big_event_size);
5097 pr_info(" smallest event: %d\n", small_event_size);
5099 if (RB_WARN_ON(buffer, total_dropped))
5104 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5106 item = ring_buffer_event_data(event);
5107 total_len += ring_buffer_event_length(event);
5108 total_size += item->size + sizeof(struct rb_item);
5109 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5110 pr_info("FAILED!\n");
5111 pr_info("buffer had: %.*s\n", item->size, item->str);
5112 pr_info("expected: %.*s\n", item->size, rb_string);
5113 RB_WARN_ON(buffer, 1);
5124 pr_info(" read events: %ld\n", total_read);
5125 pr_info(" lost events: %ld\n", total_lost);
5126 pr_info(" total events: %ld\n", total_lost + total_read);
5127 pr_info(" recorded len bytes: %ld\n", total_len);
5128 pr_info(" recorded size bytes: %ld\n", total_size);
5130 pr_info(" With dropped events, record len and size may not match\n"
5131 " alloced and written from above\n");
5133 if (RB_WARN_ON(buffer, total_len != total_alloc ||
5134 total_size != total_written))
5137 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5143 pr_info("Ring buffer PASSED!\n");
5145 ring_buffer_free(buffer);
5149 late_initcall(test_ringbuffer);
5150 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */