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/trace_seq.h>
10 #include <linux/spinlock.h>
11 #include <linux/irq_work.h>
12 #include <linux/uaccess.h>
13 #include <linux/hardirq.h>
14 #include <linux/kthread.h> /* for self test */
15 #include <linux/kmemcheck.h>
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 */
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 unsigned long nr_pages;
447 unsigned int current_context;
448 struct list_head *pages;
449 struct buffer_page *head_page; /* read from head */
450 struct buffer_page *tail_page; /* write to tail */
451 struct buffer_page *commit_page; /* committed pages */
452 struct buffer_page *reader_page;
453 unsigned long lost_events;
454 unsigned long last_overrun;
455 local_t entries_bytes;
458 local_t commit_overrun;
459 local_t dropped_events;
463 unsigned long read_bytes;
466 /* ring buffer pages to update, > 0 to add, < 0 to remove */
467 long nr_pages_to_update;
468 struct list_head new_pages; /* new pages to add */
469 struct work_struct update_pages_work;
470 struct completion update_done;
472 struct rb_irq_work irq_work;
478 atomic_t record_disabled;
479 atomic_t resize_disabled;
480 cpumask_var_t cpumask;
482 struct lock_class_key *reader_lock_key;
486 struct ring_buffer_per_cpu **buffers;
488 #ifdef CONFIG_HOTPLUG_CPU
489 struct notifier_block cpu_notify;
493 struct rb_irq_work irq_work;
496 struct ring_buffer_iter {
497 struct ring_buffer_per_cpu *cpu_buffer;
499 struct buffer_page *head_page;
500 struct buffer_page *cache_reader_page;
501 unsigned long cache_read;
506 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
508 * Schedules a delayed work to wake up any task that is blocked on the
509 * ring buffer waiters queue.
511 static void rb_wake_up_waiters(struct irq_work *work)
513 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
515 wake_up_all(&rbwork->waiters);
516 if (rbwork->wakeup_full) {
517 rbwork->wakeup_full = 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 for (i = 0; i < nr_pages; i++) {
1145 * __GFP_NORETRY flag makes sure that the allocation fails
1146 * gracefully without invoking oom-killer and the system is
1149 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1150 GFP_KERNEL | __GFP_NORETRY,
1155 list_add(&bpage->list, pages);
1157 page = alloc_pages_node(cpu_to_node(cpu),
1158 GFP_KERNEL | __GFP_NORETRY, 0);
1161 bpage->page = page_address(page);
1162 rb_init_page(bpage->page);
1168 list_for_each_entry_safe(bpage, tmp, pages, list) {
1169 list_del_init(&bpage->list);
1170 free_buffer_page(bpage);
1176 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1177 unsigned long nr_pages)
1183 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1187 * The ring buffer page list is a circular list that does not
1188 * start and end with a list head. All page list items point to
1191 cpu_buffer->pages = pages.next;
1194 cpu_buffer->nr_pages = nr_pages;
1196 rb_check_pages(cpu_buffer);
1201 static struct ring_buffer_per_cpu *
1202 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1204 struct ring_buffer_per_cpu *cpu_buffer;
1205 struct buffer_page *bpage;
1209 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1210 GFP_KERNEL, cpu_to_node(cpu));
1214 cpu_buffer->cpu = cpu;
1215 cpu_buffer->buffer = buffer;
1216 raw_spin_lock_init(&cpu_buffer->reader_lock);
1217 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1218 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1219 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1220 init_completion(&cpu_buffer->update_done);
1221 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1222 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1223 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1225 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1226 GFP_KERNEL, cpu_to_node(cpu));
1228 goto fail_free_buffer;
1230 rb_check_bpage(cpu_buffer, bpage);
1232 cpu_buffer->reader_page = bpage;
1233 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1235 goto fail_free_reader;
1236 bpage->page = page_address(page);
1237 rb_init_page(bpage->page);
1239 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1240 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1242 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1244 goto fail_free_reader;
1246 cpu_buffer->head_page
1247 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1248 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1250 rb_head_page_activate(cpu_buffer);
1255 free_buffer_page(cpu_buffer->reader_page);
1262 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1264 struct list_head *head = cpu_buffer->pages;
1265 struct buffer_page *bpage, *tmp;
1267 free_buffer_page(cpu_buffer->reader_page);
1269 rb_head_page_deactivate(cpu_buffer);
1272 list_for_each_entry_safe(bpage, tmp, head, list) {
1273 list_del_init(&bpage->list);
1274 free_buffer_page(bpage);
1276 bpage = list_entry(head, struct buffer_page, list);
1277 free_buffer_page(bpage);
1283 #ifdef CONFIG_HOTPLUG_CPU
1284 static int rb_cpu_notify(struct notifier_block *self,
1285 unsigned long action, void *hcpu);
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;
1306 /* keep it in its own cache line */
1307 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1312 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1313 goto fail_free_buffer;
1315 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1316 buffer->flags = flags;
1317 buffer->clock = trace_clock_local;
1318 buffer->reader_lock_key = key;
1320 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1321 init_waitqueue_head(&buffer->irq_work.waiters);
1323 /* need at least two pages */
1328 * In case of non-hotplug cpu, if the ring-buffer is allocated
1329 * in early initcall, it will not be notified of secondary cpus.
1330 * In that off case, we need to allocate for all possible cpus.
1332 #ifdef CONFIG_HOTPLUG_CPU
1333 cpu_notifier_register_begin();
1334 cpumask_copy(buffer->cpumask, cpu_online_mask);
1336 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1338 buffer->cpus = nr_cpu_ids;
1340 bsize = sizeof(void *) * nr_cpu_ids;
1341 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1343 if (!buffer->buffers)
1344 goto fail_free_cpumask;
1346 for_each_buffer_cpu(buffer, cpu) {
1347 buffer->buffers[cpu] =
1348 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1349 if (!buffer->buffers[cpu])
1350 goto fail_free_buffers;
1353 #ifdef CONFIG_HOTPLUG_CPU
1354 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1355 buffer->cpu_notify.priority = 0;
1356 __register_cpu_notifier(&buffer->cpu_notify);
1357 cpu_notifier_register_done();
1360 mutex_init(&buffer->mutex);
1365 for_each_buffer_cpu(buffer, cpu) {
1366 if (buffer->buffers[cpu])
1367 rb_free_cpu_buffer(buffer->buffers[cpu]);
1369 kfree(buffer->buffers);
1372 free_cpumask_var(buffer->cpumask);
1373 #ifdef CONFIG_HOTPLUG_CPU
1374 cpu_notifier_register_done();
1381 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1384 * ring_buffer_free - free a ring buffer.
1385 * @buffer: the buffer to free.
1388 ring_buffer_free(struct ring_buffer *buffer)
1392 #ifdef CONFIG_HOTPLUG_CPU
1393 cpu_notifier_register_begin();
1394 __unregister_cpu_notifier(&buffer->cpu_notify);
1397 for_each_buffer_cpu(buffer, cpu)
1398 rb_free_cpu_buffer(buffer->buffers[cpu]);
1400 #ifdef CONFIG_HOTPLUG_CPU
1401 cpu_notifier_register_done();
1404 kfree(buffer->buffers);
1405 free_cpumask_var(buffer->cpumask);
1409 EXPORT_SYMBOL_GPL(ring_buffer_free);
1411 void ring_buffer_set_clock(struct ring_buffer *buffer,
1414 buffer->clock = clock;
1417 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1419 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1421 return local_read(&bpage->entries) & RB_WRITE_MASK;
1424 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1426 return local_read(&bpage->write) & RB_WRITE_MASK;
1430 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1432 struct list_head *tail_page, *to_remove, *next_page;
1433 struct buffer_page *to_remove_page, *tmp_iter_page;
1434 struct buffer_page *last_page, *first_page;
1435 unsigned long nr_removed;
1436 unsigned long head_bit;
1441 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1442 atomic_inc(&cpu_buffer->record_disabled);
1444 * We don't race with the readers since we have acquired the reader
1445 * lock. We also don't race with writers after disabling recording.
1446 * This makes it easy to figure out the first and the last page to be
1447 * removed from the list. We unlink all the pages in between including
1448 * the first and last pages. This is done in a busy loop so that we
1449 * lose the least number of traces.
1450 * The pages are freed after we restart recording and unlock readers.
1452 tail_page = &cpu_buffer->tail_page->list;
1455 * tail page might be on reader page, we remove the next page
1456 * from the ring buffer
1458 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1459 tail_page = rb_list_head(tail_page->next);
1460 to_remove = tail_page;
1462 /* start of pages to remove */
1463 first_page = list_entry(rb_list_head(to_remove->next),
1464 struct buffer_page, list);
1466 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1467 to_remove = rb_list_head(to_remove)->next;
1468 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1471 next_page = rb_list_head(to_remove)->next;
1474 * Now we remove all pages between tail_page and next_page.
1475 * Make sure that we have head_bit value preserved for the
1478 tail_page->next = (struct list_head *)((unsigned long)next_page |
1480 next_page = rb_list_head(next_page);
1481 next_page->prev = tail_page;
1483 /* make sure pages points to a valid page in the ring buffer */
1484 cpu_buffer->pages = next_page;
1486 /* update head page */
1488 cpu_buffer->head_page = list_entry(next_page,
1489 struct buffer_page, list);
1492 * change read pointer to make sure any read iterators reset
1495 cpu_buffer->read = 0;
1497 /* pages are removed, resume tracing and then free the pages */
1498 atomic_dec(&cpu_buffer->record_disabled);
1499 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1501 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1503 /* last buffer page to remove */
1504 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1506 tmp_iter_page = first_page;
1511 to_remove_page = tmp_iter_page;
1512 rb_inc_page(cpu_buffer, &tmp_iter_page);
1514 /* update the counters */
1515 page_entries = rb_page_entries(to_remove_page);
1518 * If something was added to this page, it was full
1519 * since it is not the tail page. So we deduct the
1520 * bytes consumed in ring buffer from here.
1521 * Increment overrun to account for the lost events.
1523 local_add(page_entries, &cpu_buffer->overrun);
1524 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1528 * We have already removed references to this list item, just
1529 * free up the buffer_page and its page
1531 free_buffer_page(to_remove_page);
1534 } while (to_remove_page != last_page);
1536 RB_WARN_ON(cpu_buffer, nr_removed);
1538 return nr_removed == 0;
1542 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1544 struct list_head *pages = &cpu_buffer->new_pages;
1545 int retries, success;
1547 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1549 * We are holding the reader lock, so the reader page won't be swapped
1550 * in the ring buffer. Now we are racing with the writer trying to
1551 * move head page and the tail page.
1552 * We are going to adapt the reader page update process where:
1553 * 1. We first splice the start and end of list of new pages between
1554 * the head page and its previous page.
1555 * 2. We cmpxchg the prev_page->next to point from head page to the
1556 * start of new pages list.
1557 * 3. Finally, we update the head->prev to the end of new list.
1559 * We will try this process 10 times, to make sure that we don't keep
1565 struct list_head *head_page, *prev_page, *r;
1566 struct list_head *last_page, *first_page;
1567 struct list_head *head_page_with_bit;
1569 head_page = &rb_set_head_page(cpu_buffer)->list;
1572 prev_page = head_page->prev;
1574 first_page = pages->next;
1575 last_page = pages->prev;
1577 head_page_with_bit = (struct list_head *)
1578 ((unsigned long)head_page | RB_PAGE_HEAD);
1580 last_page->next = head_page_with_bit;
1581 first_page->prev = prev_page;
1583 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1585 if (r == head_page_with_bit) {
1587 * yay, we replaced the page pointer to our new list,
1588 * now, we just have to update to head page's prev
1589 * pointer to point to end of list
1591 head_page->prev = last_page;
1598 INIT_LIST_HEAD(pages);
1600 * If we weren't successful in adding in new pages, warn and stop
1603 RB_WARN_ON(cpu_buffer, !success);
1604 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1606 /* free pages if they weren't inserted */
1608 struct buffer_page *bpage, *tmp;
1609 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1611 list_del_init(&bpage->list);
1612 free_buffer_page(bpage);
1618 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1622 if (cpu_buffer->nr_pages_to_update > 0)
1623 success = rb_insert_pages(cpu_buffer);
1625 success = rb_remove_pages(cpu_buffer,
1626 -cpu_buffer->nr_pages_to_update);
1629 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1632 static void update_pages_handler(struct work_struct *work)
1634 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1635 struct ring_buffer_per_cpu, update_pages_work);
1636 rb_update_pages(cpu_buffer);
1637 complete(&cpu_buffer->update_done);
1641 * ring_buffer_resize - resize the ring buffer
1642 * @buffer: the buffer to resize.
1643 * @size: the new size.
1644 * @cpu_id: the cpu buffer to resize
1646 * Minimum size is 2 * BUF_PAGE_SIZE.
1648 * Returns 0 on success and < 0 on failure.
1650 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1653 struct ring_buffer_per_cpu *cpu_buffer;
1654 unsigned long nr_pages;
1658 * Always succeed at resizing a non-existent buffer:
1663 /* Make sure the requested buffer exists */
1664 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1665 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1668 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1670 /* we need a minimum of two pages */
1674 size = nr_pages * BUF_PAGE_SIZE;
1677 * Don't succeed if resizing is disabled, as a reader might be
1678 * manipulating the ring buffer and is expecting a sane state while
1681 if (atomic_read(&buffer->resize_disabled))
1684 /* prevent another thread from changing buffer sizes */
1685 mutex_lock(&buffer->mutex);
1687 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1688 /* calculate the pages to update */
1689 for_each_buffer_cpu(buffer, cpu) {
1690 cpu_buffer = buffer->buffers[cpu];
1692 cpu_buffer->nr_pages_to_update = nr_pages -
1693 cpu_buffer->nr_pages;
1695 * nothing more to do for removing pages or no update
1697 if (cpu_buffer->nr_pages_to_update <= 0)
1700 * to add pages, make sure all new pages can be
1701 * allocated without receiving ENOMEM
1703 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1704 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1705 &cpu_buffer->new_pages, cpu)) {
1706 /* not enough memory for new pages */
1714 * Fire off all the required work handlers
1715 * We can't schedule on offline CPUs, but it's not necessary
1716 * since we can change their buffer sizes without any race.
1718 for_each_buffer_cpu(buffer, cpu) {
1719 cpu_buffer = buffer->buffers[cpu];
1720 if (!cpu_buffer->nr_pages_to_update)
1723 /* Can't run something on an offline CPU. */
1724 if (!cpu_online(cpu)) {
1725 rb_update_pages(cpu_buffer);
1726 cpu_buffer->nr_pages_to_update = 0;
1728 schedule_work_on(cpu,
1729 &cpu_buffer->update_pages_work);
1733 /* wait for all the updates to complete */
1734 for_each_buffer_cpu(buffer, cpu) {
1735 cpu_buffer = buffer->buffers[cpu];
1736 if (!cpu_buffer->nr_pages_to_update)
1739 if (cpu_online(cpu))
1740 wait_for_completion(&cpu_buffer->update_done);
1741 cpu_buffer->nr_pages_to_update = 0;
1746 /* Make sure this CPU has been intitialized */
1747 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1750 cpu_buffer = buffer->buffers[cpu_id];
1752 if (nr_pages == cpu_buffer->nr_pages)
1755 cpu_buffer->nr_pages_to_update = nr_pages -
1756 cpu_buffer->nr_pages;
1758 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1759 if (cpu_buffer->nr_pages_to_update > 0 &&
1760 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1761 &cpu_buffer->new_pages, cpu_id)) {
1768 /* Can't run something on an offline CPU. */
1769 if (!cpu_online(cpu_id))
1770 rb_update_pages(cpu_buffer);
1772 schedule_work_on(cpu_id,
1773 &cpu_buffer->update_pages_work);
1774 wait_for_completion(&cpu_buffer->update_done);
1777 cpu_buffer->nr_pages_to_update = 0;
1783 * The ring buffer resize can happen with the ring buffer
1784 * enabled, so that the update disturbs the tracing as little
1785 * as possible. But if the buffer is disabled, we do not need
1786 * to worry about that, and we can take the time to verify
1787 * that the buffer is not corrupt.
1789 if (atomic_read(&buffer->record_disabled)) {
1790 atomic_inc(&buffer->record_disabled);
1792 * Even though the buffer was disabled, we must make sure
1793 * that it is truly disabled before calling rb_check_pages.
1794 * There could have been a race between checking
1795 * record_disable and incrementing it.
1797 synchronize_sched();
1798 for_each_buffer_cpu(buffer, cpu) {
1799 cpu_buffer = buffer->buffers[cpu];
1800 rb_check_pages(cpu_buffer);
1802 atomic_dec(&buffer->record_disabled);
1805 mutex_unlock(&buffer->mutex);
1809 for_each_buffer_cpu(buffer, cpu) {
1810 struct buffer_page *bpage, *tmp;
1812 cpu_buffer = buffer->buffers[cpu];
1813 cpu_buffer->nr_pages_to_update = 0;
1815 if (list_empty(&cpu_buffer->new_pages))
1818 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1820 list_del_init(&bpage->list);
1821 free_buffer_page(bpage);
1824 mutex_unlock(&buffer->mutex);
1827 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1829 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1831 mutex_lock(&buffer->mutex);
1833 buffer->flags |= RB_FL_OVERWRITE;
1835 buffer->flags &= ~RB_FL_OVERWRITE;
1836 mutex_unlock(&buffer->mutex);
1838 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1840 static inline void *
1841 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1843 return bpage->data + index;
1846 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1848 return bpage->page->data + index;
1851 static inline struct ring_buffer_event *
1852 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1854 return __rb_page_index(cpu_buffer->reader_page,
1855 cpu_buffer->reader_page->read);
1858 static inline struct ring_buffer_event *
1859 rb_iter_head_event(struct ring_buffer_iter *iter)
1861 return __rb_page_index(iter->head_page, iter->head);
1864 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1866 return local_read(&bpage->page->commit);
1869 /* Size is determined by what has been committed */
1870 static inline unsigned rb_page_size(struct buffer_page *bpage)
1872 return rb_page_commit(bpage);
1875 static inline unsigned
1876 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1878 return rb_page_commit(cpu_buffer->commit_page);
1881 static inline unsigned
1882 rb_event_index(struct ring_buffer_event *event)
1884 unsigned long addr = (unsigned long)event;
1886 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1889 static void rb_inc_iter(struct ring_buffer_iter *iter)
1891 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1894 * The iterator could be on the reader page (it starts there).
1895 * But the head could have moved, since the reader was
1896 * found. Check for this case and assign the iterator
1897 * to the head page instead of next.
1899 if (iter->head_page == cpu_buffer->reader_page)
1900 iter->head_page = rb_set_head_page(cpu_buffer);
1902 rb_inc_page(cpu_buffer, &iter->head_page);
1904 iter->read_stamp = iter->head_page->page->time_stamp;
1909 * rb_handle_head_page - writer hit the head page
1911 * Returns: +1 to retry page
1916 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1917 struct buffer_page *tail_page,
1918 struct buffer_page *next_page)
1920 struct buffer_page *new_head;
1925 entries = rb_page_entries(next_page);
1928 * The hard part is here. We need to move the head
1929 * forward, and protect against both readers on
1930 * other CPUs and writers coming in via interrupts.
1932 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1936 * type can be one of four:
1937 * NORMAL - an interrupt already moved it for us
1938 * HEAD - we are the first to get here.
1939 * UPDATE - we are the interrupt interrupting
1941 * MOVED - a reader on another CPU moved the next
1942 * pointer to its reader page. Give up
1949 * We changed the head to UPDATE, thus
1950 * it is our responsibility to update
1953 local_add(entries, &cpu_buffer->overrun);
1954 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1957 * The entries will be zeroed out when we move the
1961 /* still more to do */
1964 case RB_PAGE_UPDATE:
1966 * This is an interrupt that interrupt the
1967 * previous update. Still more to do.
1970 case RB_PAGE_NORMAL:
1972 * An interrupt came in before the update
1973 * and processed this for us.
1974 * Nothing left to do.
1979 * The reader is on another CPU and just did
1980 * a swap with our next_page.
1985 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1990 * Now that we are here, the old head pointer is
1991 * set to UPDATE. This will keep the reader from
1992 * swapping the head page with the reader page.
1993 * The reader (on another CPU) will spin till
1996 * We just need to protect against interrupts
1997 * doing the job. We will set the next pointer
1998 * to HEAD. After that, we set the old pointer
1999 * to NORMAL, but only if it was HEAD before.
2000 * otherwise we are an interrupt, and only
2001 * want the outer most commit to reset it.
2003 new_head = next_page;
2004 rb_inc_page(cpu_buffer, &new_head);
2006 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2010 * Valid returns are:
2011 * HEAD - an interrupt came in and already set it.
2012 * NORMAL - One of two things:
2013 * 1) We really set it.
2014 * 2) A bunch of interrupts came in and moved
2015 * the page forward again.
2019 case RB_PAGE_NORMAL:
2023 RB_WARN_ON(cpu_buffer, 1);
2028 * It is possible that an interrupt came in,
2029 * set the head up, then more interrupts came in
2030 * and moved it again. When we get back here,
2031 * the page would have been set to NORMAL but we
2032 * just set it back to HEAD.
2034 * How do you detect this? Well, if that happened
2035 * the tail page would have moved.
2037 if (ret == RB_PAGE_NORMAL) {
2038 struct buffer_page *buffer_tail_page;
2040 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2042 * If the tail had moved passed next, then we need
2043 * to reset the pointer.
2045 if (buffer_tail_page != tail_page &&
2046 buffer_tail_page != next_page)
2047 rb_head_page_set_normal(cpu_buffer, new_head,
2053 * If this was the outer most commit (the one that
2054 * changed the original pointer from HEAD to UPDATE),
2055 * then it is up to us to reset it to NORMAL.
2057 if (type == RB_PAGE_HEAD) {
2058 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2061 if (RB_WARN_ON(cpu_buffer,
2062 ret != RB_PAGE_UPDATE))
2070 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2071 unsigned long tail, struct rb_event_info *info)
2073 struct buffer_page *tail_page = info->tail_page;
2074 struct ring_buffer_event *event;
2075 unsigned long length = info->length;
2078 * Only the event that crossed the page boundary
2079 * must fill the old tail_page with padding.
2081 if (tail >= BUF_PAGE_SIZE) {
2083 * If the page was filled, then we still need
2084 * to update the real_end. Reset it to zero
2085 * and the reader will ignore it.
2087 if (tail == BUF_PAGE_SIZE)
2088 tail_page->real_end = 0;
2090 local_sub(length, &tail_page->write);
2094 event = __rb_page_index(tail_page, tail);
2095 kmemcheck_annotate_bitfield(event, bitfield);
2097 /* account for padding bytes */
2098 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2101 * Save the original length to the meta data.
2102 * This will be used by the reader to add lost event
2105 tail_page->real_end = tail;
2108 * If this event is bigger than the minimum size, then
2109 * we need to be careful that we don't subtract the
2110 * write counter enough to allow another writer to slip
2112 * We put in a discarded commit instead, to make sure
2113 * that this space is not used again.
2115 * If we are less than the minimum size, we don't need to
2118 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2119 /* No room for any events */
2121 /* Mark the rest of the page with padding */
2122 rb_event_set_padding(event);
2124 /* Set the write back to the previous setting */
2125 local_sub(length, &tail_page->write);
2129 /* Put in a discarded event */
2130 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2131 event->type_len = RINGBUF_TYPE_PADDING;
2132 /* time delta must be non zero */
2133 event->time_delta = 1;
2135 /* Set write to end of buffer */
2136 length = (tail + length) - BUF_PAGE_SIZE;
2137 local_sub(length, &tail_page->write);
2140 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2143 * This is the slow path, force gcc not to inline it.
2145 static noinline struct ring_buffer_event *
2146 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2147 unsigned long tail, struct rb_event_info *info)
2149 struct buffer_page *tail_page = info->tail_page;
2150 struct buffer_page *commit_page = cpu_buffer->commit_page;
2151 struct ring_buffer *buffer = cpu_buffer->buffer;
2152 struct buffer_page *next_page;
2155 next_page = tail_page;
2157 rb_inc_page(cpu_buffer, &next_page);
2160 * If for some reason, we had an interrupt storm that made
2161 * it all the way around the buffer, bail, and warn
2164 if (unlikely(next_page == commit_page)) {
2165 local_inc(&cpu_buffer->commit_overrun);
2170 * This is where the fun begins!
2172 * We are fighting against races between a reader that
2173 * could be on another CPU trying to swap its reader
2174 * page with the buffer head.
2176 * We are also fighting against interrupts coming in and
2177 * moving the head or tail on us as well.
2179 * If the next page is the head page then we have filled
2180 * the buffer, unless the commit page is still on the
2183 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2186 * If the commit is not on the reader page, then
2187 * move the header page.
2189 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2191 * If we are not in overwrite mode,
2192 * this is easy, just stop here.
2194 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2195 local_inc(&cpu_buffer->dropped_events);
2199 ret = rb_handle_head_page(cpu_buffer,
2208 * We need to be careful here too. The
2209 * commit page could still be on the reader
2210 * page. We could have a small buffer, and
2211 * have filled up the buffer with events
2212 * from interrupts and such, and wrapped.
2214 * Note, if the tail page is also the on the
2215 * reader_page, we let it move out.
2217 if (unlikely((cpu_buffer->commit_page !=
2218 cpu_buffer->tail_page) &&
2219 (cpu_buffer->commit_page ==
2220 cpu_buffer->reader_page))) {
2221 local_inc(&cpu_buffer->commit_overrun);
2227 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2231 rb_reset_tail(cpu_buffer, tail, info);
2233 /* Commit what we have for now. */
2234 rb_end_commit(cpu_buffer);
2235 /* rb_end_commit() decs committing */
2236 local_inc(&cpu_buffer->committing);
2238 /* fail and let the caller try again */
2239 return ERR_PTR(-EAGAIN);
2243 rb_reset_tail(cpu_buffer, tail, info);
2248 /* Slow path, do not inline */
2249 static noinline struct ring_buffer_event *
2250 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2252 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2254 /* Not the first event on the page? */
2255 if (rb_event_index(event)) {
2256 event->time_delta = delta & TS_MASK;
2257 event->array[0] = delta >> TS_SHIFT;
2259 /* nope, just zero it */
2260 event->time_delta = 0;
2261 event->array[0] = 0;
2264 return skip_time_extend(event);
2267 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2268 struct ring_buffer_event *event);
2271 * rb_update_event - update event type and data
2272 * @event: the event to update
2273 * @type: the type of event
2274 * @length: the size of the event field in the ring buffer
2276 * Update the type and data fields of the event. The length
2277 * is the actual size that is written to the ring buffer,
2278 * and with this, we can determine what to place into the
2282 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2283 struct ring_buffer_event *event,
2284 struct rb_event_info *info)
2286 unsigned length = info->length;
2287 u64 delta = info->delta;
2289 /* Only a commit updates the timestamp */
2290 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2294 * If we need to add a timestamp, then we
2295 * add it to the start of the resevered space.
2297 if (unlikely(info->add_timestamp)) {
2298 event = rb_add_time_stamp(event, delta);
2299 length -= RB_LEN_TIME_EXTEND;
2303 event->time_delta = delta;
2304 length -= RB_EVNT_HDR_SIZE;
2305 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2306 event->type_len = 0;
2307 event->array[0] = length;
2309 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2312 static unsigned rb_calculate_event_length(unsigned length)
2314 struct ring_buffer_event event; /* Used only for sizeof array */
2316 /* zero length can cause confusions */
2320 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2321 length += sizeof(event.array[0]);
2323 length += RB_EVNT_HDR_SIZE;
2324 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2327 * In case the time delta is larger than the 27 bits for it
2328 * in the header, we need to add a timestamp. If another
2329 * event comes in when trying to discard this one to increase
2330 * the length, then the timestamp will be added in the allocated
2331 * space of this event. If length is bigger than the size needed
2332 * for the TIME_EXTEND, then padding has to be used. The events
2333 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2334 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2335 * As length is a multiple of 4, we only need to worry if it
2336 * is 12 (RB_LEN_TIME_EXTEND + 4).
2338 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2339 length += RB_ALIGNMENT;
2344 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2345 static inline bool sched_clock_stable(void)
2352 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2353 struct ring_buffer_event *event)
2355 unsigned long new_index, old_index;
2356 struct buffer_page *bpage;
2357 unsigned long index;
2360 new_index = rb_event_index(event);
2361 old_index = new_index + rb_event_ts_length(event);
2362 addr = (unsigned long)event;
2365 bpage = READ_ONCE(cpu_buffer->tail_page);
2367 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2368 unsigned long write_mask =
2369 local_read(&bpage->write) & ~RB_WRITE_MASK;
2370 unsigned long event_length = rb_event_length(event);
2372 * This is on the tail page. It is possible that
2373 * a write could come in and move the tail page
2374 * and write to the next page. That is fine
2375 * because we just shorten what is on this page.
2377 old_index += write_mask;
2378 new_index += write_mask;
2379 index = local_cmpxchg(&bpage->write, old_index, new_index);
2380 if (index == old_index) {
2381 /* update counters */
2382 local_sub(event_length, &cpu_buffer->entries_bytes);
2387 /* could not discard */
2391 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2393 local_inc(&cpu_buffer->committing);
2394 local_inc(&cpu_buffer->commits);
2398 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2400 unsigned long max_count;
2403 * We only race with interrupts and NMIs on this CPU.
2404 * If we own the commit event, then we can commit
2405 * all others that interrupted us, since the interruptions
2406 * are in stack format (they finish before they come
2407 * back to us). This allows us to do a simple loop to
2408 * assign the commit to the tail.
2411 max_count = cpu_buffer->nr_pages * 100;
2413 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2414 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2416 if (RB_WARN_ON(cpu_buffer,
2417 rb_is_reader_page(cpu_buffer->tail_page)))
2419 local_set(&cpu_buffer->commit_page->page->commit,
2420 rb_page_write(cpu_buffer->commit_page));
2421 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2422 /* Only update the write stamp if the page has an event */
2423 if (rb_page_write(cpu_buffer->commit_page))
2424 cpu_buffer->write_stamp =
2425 cpu_buffer->commit_page->page->time_stamp;
2426 /* add barrier to keep gcc from optimizing too much */
2429 while (rb_commit_index(cpu_buffer) !=
2430 rb_page_write(cpu_buffer->commit_page)) {
2432 local_set(&cpu_buffer->commit_page->page->commit,
2433 rb_page_write(cpu_buffer->commit_page));
2434 RB_WARN_ON(cpu_buffer,
2435 local_read(&cpu_buffer->commit_page->page->commit) &
2440 /* again, keep gcc from optimizing */
2444 * If an interrupt came in just after the first while loop
2445 * and pushed the tail page forward, we will be left with
2446 * a dangling commit that will never go forward.
2448 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2452 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2454 unsigned long commits;
2456 if (RB_WARN_ON(cpu_buffer,
2457 !local_read(&cpu_buffer->committing)))
2461 commits = local_read(&cpu_buffer->commits);
2462 /* synchronize with interrupts */
2464 if (local_read(&cpu_buffer->committing) == 1)
2465 rb_set_commit_to_write(cpu_buffer);
2467 local_dec(&cpu_buffer->committing);
2469 /* synchronize with interrupts */
2473 * Need to account for interrupts coming in between the
2474 * updating of the commit page and the clearing of the
2475 * committing counter.
2477 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2478 !local_read(&cpu_buffer->committing)) {
2479 local_inc(&cpu_buffer->committing);
2484 static inline void rb_event_discard(struct ring_buffer_event *event)
2486 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2487 event = skip_time_extend(event);
2489 /* array[0] holds the actual length for the discarded event */
2490 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2491 event->type_len = RINGBUF_TYPE_PADDING;
2492 /* time delta must be non zero */
2493 if (!event->time_delta)
2494 event->time_delta = 1;
2498 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2499 struct ring_buffer_event *event)
2501 unsigned long addr = (unsigned long)event;
2502 unsigned long index;
2504 index = rb_event_index(event);
2507 return cpu_buffer->commit_page->page == (void *)addr &&
2508 rb_commit_index(cpu_buffer) == index;
2512 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2513 struct ring_buffer_event *event)
2518 * The event first in the commit queue updates the
2521 if (rb_event_is_commit(cpu_buffer, event)) {
2523 * A commit event that is first on a page
2524 * updates the write timestamp with the page stamp
2526 if (!rb_event_index(event))
2527 cpu_buffer->write_stamp =
2528 cpu_buffer->commit_page->page->time_stamp;
2529 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2530 delta = event->array[0];
2532 delta += event->time_delta;
2533 cpu_buffer->write_stamp += delta;
2535 cpu_buffer->write_stamp += event->time_delta;
2539 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2540 struct ring_buffer_event *event)
2542 local_inc(&cpu_buffer->entries);
2543 rb_update_write_stamp(cpu_buffer, event);
2544 rb_end_commit(cpu_buffer);
2547 static __always_inline void
2548 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2552 if (buffer->irq_work.waiters_pending) {
2553 buffer->irq_work.waiters_pending = false;
2554 /* irq_work_queue() supplies it's own memory barriers */
2555 irq_work_queue(&buffer->irq_work.work);
2558 if (cpu_buffer->irq_work.waiters_pending) {
2559 cpu_buffer->irq_work.waiters_pending = false;
2560 /* irq_work_queue() supplies it's own memory barriers */
2561 irq_work_queue(&cpu_buffer->irq_work.work);
2564 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2566 if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2567 cpu_buffer->irq_work.wakeup_full = true;
2568 cpu_buffer->irq_work.full_waiters_pending = false;
2569 /* irq_work_queue() supplies it's own memory barriers */
2570 irq_work_queue(&cpu_buffer->irq_work.work);
2575 * The lock and unlock are done within a preempt disable section.
2576 * The current_context per_cpu variable can only be modified
2577 * by the current task between lock and unlock. But it can
2578 * be modified more than once via an interrupt. To pass this
2579 * information from the lock to the unlock without having to
2580 * access the 'in_interrupt()' functions again (which do show
2581 * a bit of overhead in something as critical as function tracing,
2582 * we use a bitmask trick.
2584 * bit 1 = NMI context
2585 * bit 2 = IRQ context
2586 * bit 3 = SoftIRQ context
2587 * bit 4 = normal context.
2589 * This works because this is the order of contexts that can
2590 * preempt other contexts. A SoftIRQ never preempts an IRQ
2593 * When the context is determined, the corresponding bit is
2594 * checked and set (if it was set, then a recursion of that context
2597 * On unlock, we need to clear this bit. To do so, just subtract
2598 * 1 from the current_context and AND it to itself.
2602 * 101 & 100 = 100 (clearing bit zero)
2605 * 1010 & 1001 = 1000 (clearing bit 1)
2607 * The least significant bit can be cleared this way, and it
2608 * just so happens that it is the same bit corresponding to
2609 * the current context.
2611 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
2612 * is set when a recursion is detected at the current context, and if
2613 * the TRANSITION bit is already set, it will fail the recursion.
2614 * This is needed because there's a lag between the changing of
2615 * interrupt context and updating the preempt count. In this case,
2616 * a false positive will be found. To handle this, one extra recursion
2617 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
2618 * bit is already set, then it is considered a recursion and the function
2619 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
2621 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
2622 * to be cleared. Even if it wasn't the context that set it. That is,
2623 * if an interrupt comes in while NORMAL bit is set and the ring buffer
2624 * is called before preempt_count() is updated, since the check will
2625 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
2626 * NMI then comes in, it will set the NMI bit, but when the NMI code
2627 * does the trace_recursive_unlock() it will clear the TRANSTION bit
2628 * and leave the NMI bit set. But this is fine, because the interrupt
2629 * code that set the TRANSITION bit will then clear the NMI bit when it
2630 * calls trace_recursive_unlock(). If another NMI comes in, it will
2631 * set the TRANSITION bit and continue.
2633 * Note: The TRANSITION bit only handles a single transition between context.
2636 static __always_inline int
2637 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2639 unsigned int val = cpu_buffer->current_context;
2642 if (in_interrupt()) {
2648 bit = RB_CTX_SOFTIRQ;
2650 bit = RB_CTX_NORMAL;
2652 if (unlikely(val & (1 << bit))) {
2654 * It is possible that this was called by transitioning
2655 * between interrupt context, and preempt_count() has not
2656 * been updated yet. In this case, use the TRANSITION bit.
2658 bit = RB_CTX_TRANSITION;
2659 if (val & (1 << bit))
2664 cpu_buffer->current_context = val;
2669 static __always_inline void
2670 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2672 cpu_buffer->current_context &= cpu_buffer->current_context - 1;
2676 * ring_buffer_unlock_commit - commit a reserved
2677 * @buffer: The buffer to commit to
2678 * @event: The event pointer to commit.
2680 * This commits the data to the ring buffer, and releases any locks held.
2682 * Must be paired with ring_buffer_lock_reserve.
2684 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2685 struct ring_buffer_event *event)
2687 struct ring_buffer_per_cpu *cpu_buffer;
2688 int cpu = raw_smp_processor_id();
2690 cpu_buffer = buffer->buffers[cpu];
2692 rb_commit(cpu_buffer, event);
2694 rb_wakeups(buffer, cpu_buffer);
2696 trace_recursive_unlock(cpu_buffer);
2698 preempt_enable_notrace();
2702 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2704 static noinline void
2705 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2706 struct rb_event_info *info)
2708 WARN_ONCE(info->delta > (1ULL << 59),
2709 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2710 (unsigned long long)info->delta,
2711 (unsigned long long)info->ts,
2712 (unsigned long long)cpu_buffer->write_stamp,
2713 sched_clock_stable() ? "" :
2714 "If you just came from a suspend/resume,\n"
2715 "please switch to the trace global clock:\n"
2716 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2717 info->add_timestamp = 1;
2720 static struct ring_buffer_event *
2721 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2722 struct rb_event_info *info)
2724 struct ring_buffer_event *event;
2725 struct buffer_page *tail_page;
2726 unsigned long tail, write;
2729 * If the time delta since the last event is too big to
2730 * hold in the time field of the event, then we append a
2731 * TIME EXTEND event ahead of the data event.
2733 if (unlikely(info->add_timestamp))
2734 info->length += RB_LEN_TIME_EXTEND;
2736 /* Don't let the compiler play games with cpu_buffer->tail_page */
2737 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2738 write = local_add_return(info->length, &tail_page->write);
2740 /* set write to only the index of the write */
2741 write &= RB_WRITE_MASK;
2742 tail = write - info->length;
2745 * If this is the first commit on the page, then it has the same
2746 * timestamp as the page itself.
2751 /* See if we shot pass the end of this buffer page */
2752 if (unlikely(write > BUF_PAGE_SIZE))
2753 return rb_move_tail(cpu_buffer, tail, info);
2755 /* We reserved something on the buffer */
2757 event = __rb_page_index(tail_page, tail);
2758 kmemcheck_annotate_bitfield(event, bitfield);
2759 rb_update_event(cpu_buffer, event, info);
2761 local_inc(&tail_page->entries);
2764 * If this is the first commit on the page, then update
2768 tail_page->page->time_stamp = info->ts;
2770 /* account for these added bytes */
2771 local_add(info->length, &cpu_buffer->entries_bytes);
2776 static struct ring_buffer_event *
2777 rb_reserve_next_event(struct ring_buffer *buffer,
2778 struct ring_buffer_per_cpu *cpu_buffer,
2779 unsigned long length)
2781 struct ring_buffer_event *event;
2782 struct rb_event_info info;
2786 rb_start_commit(cpu_buffer);
2788 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2790 * Due to the ability to swap a cpu buffer from a buffer
2791 * it is possible it was swapped before we committed.
2792 * (committing stops a swap). We check for it here and
2793 * if it happened, we have to fail the write.
2796 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2797 local_dec(&cpu_buffer->committing);
2798 local_dec(&cpu_buffer->commits);
2803 info.length = rb_calculate_event_length(length);
2805 info.add_timestamp = 0;
2809 * We allow for interrupts to reenter here and do a trace.
2810 * If one does, it will cause this original code to loop
2811 * back here. Even with heavy interrupts happening, this
2812 * should only happen a few times in a row. If this happens
2813 * 1000 times in a row, there must be either an interrupt
2814 * storm or we have something buggy.
2817 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2820 info.ts = rb_time_stamp(cpu_buffer->buffer);
2821 diff = info.ts - cpu_buffer->write_stamp;
2823 /* make sure this diff is calculated here */
2826 /* Did the write stamp get updated already? */
2827 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2829 if (unlikely(test_time_stamp(info.delta)))
2830 rb_handle_timestamp(cpu_buffer, &info);
2833 event = __rb_reserve_next(cpu_buffer, &info);
2835 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2836 if (info.add_timestamp)
2837 info.length -= RB_LEN_TIME_EXTEND;
2847 rb_end_commit(cpu_buffer);
2852 * ring_buffer_lock_reserve - reserve a part of the buffer
2853 * @buffer: the ring buffer to reserve from
2854 * @length: the length of the data to reserve (excluding event header)
2856 * Returns a reseverd event on the ring buffer to copy directly to.
2857 * The user of this interface will need to get the body to write into
2858 * and can use the ring_buffer_event_data() interface.
2860 * The length is the length of the data needed, not the event length
2861 * which also includes the event header.
2863 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2864 * If NULL is returned, then nothing has been allocated or locked.
2866 struct ring_buffer_event *
2867 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2869 struct ring_buffer_per_cpu *cpu_buffer;
2870 struct ring_buffer_event *event;
2873 /* If we are tracing schedule, we don't want to recurse */
2874 preempt_disable_notrace();
2876 if (unlikely(atomic_read(&buffer->record_disabled)))
2879 cpu = raw_smp_processor_id();
2881 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2884 cpu_buffer = buffer->buffers[cpu];
2886 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2889 if (unlikely(length > BUF_MAX_DATA_SIZE))
2892 if (unlikely(trace_recursive_lock(cpu_buffer)))
2895 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2902 trace_recursive_unlock(cpu_buffer);
2904 preempt_enable_notrace();
2907 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2910 * Decrement the entries to the page that an event is on.
2911 * The event does not even need to exist, only the pointer
2912 * to the page it is on. This may only be called before the commit
2916 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2917 struct ring_buffer_event *event)
2919 unsigned long addr = (unsigned long)event;
2920 struct buffer_page *bpage = cpu_buffer->commit_page;
2921 struct buffer_page *start;
2925 /* Do the likely case first */
2926 if (likely(bpage->page == (void *)addr)) {
2927 local_dec(&bpage->entries);
2932 * Because the commit page may be on the reader page we
2933 * start with the next page and check the end loop there.
2935 rb_inc_page(cpu_buffer, &bpage);
2938 if (bpage->page == (void *)addr) {
2939 local_dec(&bpage->entries);
2942 rb_inc_page(cpu_buffer, &bpage);
2943 } while (bpage != start);
2945 /* commit not part of this buffer?? */
2946 RB_WARN_ON(cpu_buffer, 1);
2950 * ring_buffer_commit_discard - discard an event that has not been committed
2951 * @buffer: the ring buffer
2952 * @event: non committed event to discard
2954 * Sometimes an event that is in the ring buffer needs to be ignored.
2955 * This function lets the user discard an event in the ring buffer
2956 * and then that event will not be read later.
2958 * This function only works if it is called before the the item has been
2959 * committed. It will try to free the event from the ring buffer
2960 * if another event has not been added behind it.
2962 * If another event has been added behind it, it will set the event
2963 * up as discarded, and perform the commit.
2965 * If this function is called, do not call ring_buffer_unlock_commit on
2968 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2969 struct ring_buffer_event *event)
2971 struct ring_buffer_per_cpu *cpu_buffer;
2974 /* The event is discarded regardless */
2975 rb_event_discard(event);
2977 cpu = smp_processor_id();
2978 cpu_buffer = buffer->buffers[cpu];
2981 * This must only be called if the event has not been
2982 * committed yet. Thus we can assume that preemption
2983 * is still disabled.
2985 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2987 rb_decrement_entry(cpu_buffer, event);
2988 if (rb_try_to_discard(cpu_buffer, event))
2992 * The commit is still visible by the reader, so we
2993 * must still update the timestamp.
2995 rb_update_write_stamp(cpu_buffer, event);
2997 rb_end_commit(cpu_buffer);
2999 trace_recursive_unlock(cpu_buffer);
3001 preempt_enable_notrace();
3004 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3007 * ring_buffer_write - write data to the buffer without reserving
3008 * @buffer: The ring buffer to write to.
3009 * @length: The length of the data being written (excluding the event header)
3010 * @data: The data to write to the buffer.
3012 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3013 * one function. If you already have the data to write to the buffer, it
3014 * may be easier to simply call this function.
3016 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3017 * and not the length of the event which would hold the header.
3019 int ring_buffer_write(struct ring_buffer *buffer,
3020 unsigned long length,
3023 struct ring_buffer_per_cpu *cpu_buffer;
3024 struct ring_buffer_event *event;
3029 preempt_disable_notrace();
3031 if (atomic_read(&buffer->record_disabled))
3034 cpu = raw_smp_processor_id();
3036 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3039 cpu_buffer = buffer->buffers[cpu];
3041 if (atomic_read(&cpu_buffer->record_disabled))
3044 if (length > BUF_MAX_DATA_SIZE)
3047 if (unlikely(trace_recursive_lock(cpu_buffer)))
3050 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3054 body = rb_event_data(event);
3056 memcpy(body, data, length);
3058 rb_commit(cpu_buffer, event);
3060 rb_wakeups(buffer, cpu_buffer);
3065 trace_recursive_unlock(cpu_buffer);
3068 preempt_enable_notrace();
3072 EXPORT_SYMBOL_GPL(ring_buffer_write);
3074 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3076 struct buffer_page *reader = cpu_buffer->reader_page;
3077 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3078 struct buffer_page *commit = cpu_buffer->commit_page;
3080 /* In case of error, head will be NULL */
3081 if (unlikely(!head))
3084 /* Reader should exhaust content in reader page */
3085 if (reader->read != rb_page_commit(reader))
3089 * If writers are committing on the reader page, knowing all
3090 * committed content has been read, the ring buffer is empty.
3092 if (commit == reader)
3096 * If writers are committing on a page other than reader page
3097 * and head page, there should always be content to read.
3103 * Writers are committing on the head page, we just need
3104 * to care about there're committed data, and the reader will
3105 * swap reader page with head page when it is to read data.
3107 return rb_page_commit(commit) == 0;
3111 * ring_buffer_record_disable - stop all writes into the buffer
3112 * @buffer: The ring buffer to stop writes to.
3114 * This prevents all writes to the buffer. Any attempt to write
3115 * to the buffer after this will fail and return NULL.
3117 * The caller should call synchronize_sched() after this.
3119 void ring_buffer_record_disable(struct ring_buffer *buffer)
3121 atomic_inc(&buffer->record_disabled);
3123 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3126 * ring_buffer_record_enable - enable writes to the buffer
3127 * @buffer: The ring buffer to enable writes
3129 * Note, multiple disables will need the same number of enables
3130 * to truly enable the writing (much like preempt_disable).
3132 void ring_buffer_record_enable(struct ring_buffer *buffer)
3134 atomic_dec(&buffer->record_disabled);
3136 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3139 * ring_buffer_record_off - stop all writes into the buffer
3140 * @buffer: The ring buffer to stop writes to.
3142 * This prevents all writes to the buffer. Any attempt to write
3143 * to the buffer after this will fail and return NULL.
3145 * This is different than ring_buffer_record_disable() as
3146 * it works like an on/off switch, where as the disable() version
3147 * must be paired with a enable().
3149 void ring_buffer_record_off(struct ring_buffer *buffer)
3152 unsigned int new_rd;
3155 rd = atomic_read(&buffer->record_disabled);
3156 new_rd = rd | RB_BUFFER_OFF;
3157 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3159 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3162 * ring_buffer_record_on - restart writes into the buffer
3163 * @buffer: The ring buffer to start writes to.
3165 * This enables all writes to the buffer that was disabled by
3166 * ring_buffer_record_off().
3168 * This is different than ring_buffer_record_enable() as
3169 * it works like an on/off switch, where as the enable() version
3170 * must be paired with a disable().
3172 void ring_buffer_record_on(struct ring_buffer *buffer)
3175 unsigned int new_rd;
3178 rd = atomic_read(&buffer->record_disabled);
3179 new_rd = rd & ~RB_BUFFER_OFF;
3180 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3182 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3185 * ring_buffer_record_is_on - return true if the ring buffer can write
3186 * @buffer: The ring buffer to see if write is enabled
3188 * Returns true if the ring buffer is in a state that it accepts writes.
3190 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3192 return !atomic_read(&buffer->record_disabled);
3196 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3197 * @buffer: The ring buffer to see if write is set enabled
3199 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3200 * Note that this does NOT mean it is in a writable state.
3202 * It may return true when the ring buffer has been disabled by
3203 * ring_buffer_record_disable(), as that is a temporary disabling of
3206 int ring_buffer_record_is_set_on(struct ring_buffer *buffer)
3208 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3212 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3213 * @buffer: The ring buffer to stop writes to.
3214 * @cpu: The CPU buffer to stop
3216 * This prevents all writes to the buffer. Any attempt to write
3217 * to the buffer after this will fail and return NULL.
3219 * The caller should call synchronize_sched() after this.
3221 void ring_buffer_record_disable_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_inc(&cpu_buffer->record_disabled);
3231 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3234 * ring_buffer_record_enable_cpu - enable writes to the buffer
3235 * @buffer: The ring buffer to enable writes
3236 * @cpu: The CPU to enable.
3238 * Note, multiple disables will need the same number of enables
3239 * to truly enable the writing (much like preempt_disable).
3241 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3243 struct ring_buffer_per_cpu *cpu_buffer;
3245 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3248 cpu_buffer = buffer->buffers[cpu];
3249 atomic_dec(&cpu_buffer->record_disabled);
3251 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3254 * The total entries in the ring buffer is the running counter
3255 * of entries entered into the ring buffer, minus the sum of
3256 * the entries read from the ring buffer and the number of
3257 * entries that were overwritten.
3259 static inline unsigned long
3260 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3262 return local_read(&cpu_buffer->entries) -
3263 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3267 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3268 * @buffer: The ring buffer
3269 * @cpu: The per CPU buffer to read from.
3271 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3273 unsigned long flags;
3274 struct ring_buffer_per_cpu *cpu_buffer;
3275 struct buffer_page *bpage;
3278 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3281 cpu_buffer = buffer->buffers[cpu];
3282 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3284 * if the tail is on reader_page, oldest time stamp is on the reader
3287 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3288 bpage = cpu_buffer->reader_page;
3290 bpage = rb_set_head_page(cpu_buffer);
3292 ret = bpage->page->time_stamp;
3293 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3297 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3300 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3301 * @buffer: The ring buffer
3302 * @cpu: The per CPU buffer to read from.
3304 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3306 struct ring_buffer_per_cpu *cpu_buffer;
3309 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3312 cpu_buffer = buffer->buffers[cpu];
3313 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3317 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3320 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3321 * @buffer: The ring buffer
3322 * @cpu: The per CPU buffer to get the entries from.
3324 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3326 struct ring_buffer_per_cpu *cpu_buffer;
3328 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3331 cpu_buffer = buffer->buffers[cpu];
3333 return rb_num_of_entries(cpu_buffer);
3335 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3338 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3339 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3340 * @buffer: The ring buffer
3341 * @cpu: The per CPU buffer to get the number of overruns from
3343 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3345 struct ring_buffer_per_cpu *cpu_buffer;
3348 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3351 cpu_buffer = buffer->buffers[cpu];
3352 ret = local_read(&cpu_buffer->overrun);
3356 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3359 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3360 * commits failing due to the buffer wrapping around while there are uncommitted
3361 * events, such as during an interrupt storm.
3362 * @buffer: The ring buffer
3363 * @cpu: The per CPU buffer to get the number of overruns from
3366 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3368 struct ring_buffer_per_cpu *cpu_buffer;
3371 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3374 cpu_buffer = buffer->buffers[cpu];
3375 ret = local_read(&cpu_buffer->commit_overrun);
3379 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3382 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3383 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3384 * @buffer: The ring buffer
3385 * @cpu: The per CPU buffer to get the number of overruns from
3388 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3390 struct ring_buffer_per_cpu *cpu_buffer;
3393 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3396 cpu_buffer = buffer->buffers[cpu];
3397 ret = local_read(&cpu_buffer->dropped_events);
3401 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3404 * ring_buffer_read_events_cpu - get the number of events successfully read
3405 * @buffer: The ring buffer
3406 * @cpu: The per CPU buffer to get the number of events read
3409 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3411 struct ring_buffer_per_cpu *cpu_buffer;
3413 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3416 cpu_buffer = buffer->buffers[cpu];
3417 return cpu_buffer->read;
3419 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3422 * ring_buffer_entries - get the number of entries in a buffer
3423 * @buffer: The ring buffer
3425 * Returns the total number of entries in the ring buffer
3428 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3430 struct ring_buffer_per_cpu *cpu_buffer;
3431 unsigned long entries = 0;
3434 /* if you care about this being correct, lock the buffer */
3435 for_each_buffer_cpu(buffer, cpu) {
3436 cpu_buffer = buffer->buffers[cpu];
3437 entries += rb_num_of_entries(cpu_buffer);
3442 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3445 * ring_buffer_overruns - get the number of overruns in buffer
3446 * @buffer: The ring buffer
3448 * Returns the total number of overruns in the ring buffer
3451 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3453 struct ring_buffer_per_cpu *cpu_buffer;
3454 unsigned long overruns = 0;
3457 /* if you care about this being correct, lock the buffer */
3458 for_each_buffer_cpu(buffer, cpu) {
3459 cpu_buffer = buffer->buffers[cpu];
3460 overruns += local_read(&cpu_buffer->overrun);
3465 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3467 static void rb_iter_reset(struct ring_buffer_iter *iter)
3469 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3471 /* Iterator usage is expected to have record disabled */
3472 iter->head_page = cpu_buffer->reader_page;
3473 iter->head = cpu_buffer->reader_page->read;
3475 iter->cache_reader_page = iter->head_page;
3476 iter->cache_read = cpu_buffer->read;
3479 iter->read_stamp = cpu_buffer->read_stamp;
3481 iter->read_stamp = iter->head_page->page->time_stamp;
3485 * ring_buffer_iter_reset - reset an iterator
3486 * @iter: The iterator to reset
3488 * Resets the iterator, so that it will start from the beginning
3491 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3493 struct ring_buffer_per_cpu *cpu_buffer;
3494 unsigned long flags;
3499 cpu_buffer = iter->cpu_buffer;
3501 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3502 rb_iter_reset(iter);
3503 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3505 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3508 * ring_buffer_iter_empty - check if an iterator has no more to read
3509 * @iter: The iterator to check
3511 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3513 struct ring_buffer_per_cpu *cpu_buffer;
3514 struct buffer_page *reader;
3515 struct buffer_page *head_page;
3516 struct buffer_page *commit_page;
3519 cpu_buffer = iter->cpu_buffer;
3521 /* Remember, trace recording is off when iterator is in use */
3522 reader = cpu_buffer->reader_page;
3523 head_page = cpu_buffer->head_page;
3524 commit_page = cpu_buffer->commit_page;
3525 commit = rb_page_commit(commit_page);
3527 return ((iter->head_page == commit_page && iter->head == commit) ||
3528 (iter->head_page == reader && commit_page == head_page &&
3529 head_page->read == commit &&
3530 iter->head == rb_page_commit(cpu_buffer->reader_page)));
3532 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3535 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3536 struct ring_buffer_event *event)
3540 switch (event->type_len) {
3541 case RINGBUF_TYPE_PADDING:
3544 case RINGBUF_TYPE_TIME_EXTEND:
3545 delta = event->array[0];
3547 delta += event->time_delta;
3548 cpu_buffer->read_stamp += delta;
3551 case RINGBUF_TYPE_TIME_STAMP:
3552 /* FIXME: not implemented */
3555 case RINGBUF_TYPE_DATA:
3556 cpu_buffer->read_stamp += event->time_delta;
3566 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3567 struct ring_buffer_event *event)
3571 switch (event->type_len) {
3572 case RINGBUF_TYPE_PADDING:
3575 case RINGBUF_TYPE_TIME_EXTEND:
3576 delta = event->array[0];
3578 delta += event->time_delta;
3579 iter->read_stamp += delta;
3582 case RINGBUF_TYPE_TIME_STAMP:
3583 /* FIXME: not implemented */
3586 case RINGBUF_TYPE_DATA:
3587 iter->read_stamp += event->time_delta;
3596 static struct buffer_page *
3597 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3599 struct buffer_page *reader = NULL;
3600 unsigned long overwrite;
3601 unsigned long flags;
3605 local_irq_save(flags);
3606 arch_spin_lock(&cpu_buffer->lock);
3610 * This should normally only loop twice. But because the
3611 * start of the reader inserts an empty page, it causes
3612 * a case where we will loop three times. There should be no
3613 * reason to loop four times (that I know of).
3615 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3620 reader = cpu_buffer->reader_page;
3622 /* If there's more to read, return this page */
3623 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3626 /* Never should we have an index greater than the size */
3627 if (RB_WARN_ON(cpu_buffer,
3628 cpu_buffer->reader_page->read > rb_page_size(reader)))
3631 /* check if we caught up to the tail */
3633 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3636 /* Don't bother swapping if the ring buffer is empty */
3637 if (rb_num_of_entries(cpu_buffer) == 0)
3641 * Reset the reader page to size zero.
3643 local_set(&cpu_buffer->reader_page->write, 0);
3644 local_set(&cpu_buffer->reader_page->entries, 0);
3645 local_set(&cpu_buffer->reader_page->page->commit, 0);
3646 cpu_buffer->reader_page->real_end = 0;
3650 * Splice the empty reader page into the list around the head.
3652 reader = rb_set_head_page(cpu_buffer);
3655 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3656 cpu_buffer->reader_page->list.prev = reader->list.prev;
3659 * cpu_buffer->pages just needs to point to the buffer, it
3660 * has no specific buffer page to point to. Lets move it out
3661 * of our way so we don't accidentally swap it.
3663 cpu_buffer->pages = reader->list.prev;
3665 /* The reader page will be pointing to the new head */
3666 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3669 * We want to make sure we read the overruns after we set up our
3670 * pointers to the next object. The writer side does a
3671 * cmpxchg to cross pages which acts as the mb on the writer
3672 * side. Note, the reader will constantly fail the swap
3673 * while the writer is updating the pointers, so this
3674 * guarantees that the overwrite recorded here is the one we
3675 * want to compare with the last_overrun.
3678 overwrite = local_read(&(cpu_buffer->overrun));
3681 * Here's the tricky part.
3683 * We need to move the pointer past the header page.
3684 * But we can only do that if a writer is not currently
3685 * moving it. The page before the header page has the
3686 * flag bit '1' set if it is pointing to the page we want.
3687 * but if the writer is in the process of moving it
3688 * than it will be '2' or already moved '0'.
3691 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3694 * If we did not convert it, then we must try again.
3700 * Yeah! We succeeded in replacing the page.
3702 * Now make the new head point back to the reader page.
3704 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3705 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3707 /* Finally update the reader page to the new head */
3708 cpu_buffer->reader_page = reader;
3709 cpu_buffer->reader_page->read = 0;
3711 if (overwrite != cpu_buffer->last_overrun) {
3712 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3713 cpu_buffer->last_overrun = overwrite;
3719 /* Update the read_stamp on the first event */
3720 if (reader && reader->read == 0)
3721 cpu_buffer->read_stamp = reader->page->time_stamp;
3723 arch_spin_unlock(&cpu_buffer->lock);
3724 local_irq_restore(flags);
3729 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3731 struct ring_buffer_event *event;
3732 struct buffer_page *reader;
3735 reader = rb_get_reader_page(cpu_buffer);
3737 /* This function should not be called when buffer is empty */
3738 if (RB_WARN_ON(cpu_buffer, !reader))
3741 event = rb_reader_event(cpu_buffer);
3743 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3746 rb_update_read_stamp(cpu_buffer, event);
3748 length = rb_event_length(event);
3749 cpu_buffer->reader_page->read += length;
3752 static void rb_advance_iter(struct ring_buffer_iter *iter)
3754 struct ring_buffer_per_cpu *cpu_buffer;
3755 struct ring_buffer_event *event;
3758 cpu_buffer = iter->cpu_buffer;
3761 * Check if we are at the end of the buffer.
3763 if (iter->head >= rb_page_size(iter->head_page)) {
3764 /* discarded commits can make the page empty */
3765 if (iter->head_page == cpu_buffer->commit_page)
3771 event = rb_iter_head_event(iter);
3773 length = rb_event_length(event);
3776 * This should not be called to advance the header if we are
3777 * at the tail of the buffer.
3779 if (RB_WARN_ON(cpu_buffer,
3780 (iter->head_page == cpu_buffer->commit_page) &&
3781 (iter->head + length > rb_commit_index(cpu_buffer))))
3784 rb_update_iter_read_stamp(iter, event);
3786 iter->head += length;
3788 /* check for end of page padding */
3789 if ((iter->head >= rb_page_size(iter->head_page)) &&
3790 (iter->head_page != cpu_buffer->commit_page))
3794 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3796 return cpu_buffer->lost_events;
3799 static struct ring_buffer_event *
3800 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3801 unsigned long *lost_events)
3803 struct ring_buffer_event *event;
3804 struct buffer_page *reader;
3809 * We repeat when a time extend is encountered.
3810 * Since the time extend is always attached to a data event,
3811 * we should never loop more than once.
3812 * (We never hit the following condition more than twice).
3814 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3817 reader = rb_get_reader_page(cpu_buffer);
3821 event = rb_reader_event(cpu_buffer);
3823 switch (event->type_len) {
3824 case RINGBUF_TYPE_PADDING:
3825 if (rb_null_event(event))
3826 RB_WARN_ON(cpu_buffer, 1);
3828 * Because the writer could be discarding every
3829 * event it creates (which would probably be bad)
3830 * if we were to go back to "again" then we may never
3831 * catch up, and will trigger the warn on, or lock
3832 * the box. Return the padding, and we will release
3833 * the current locks, and try again.
3837 case RINGBUF_TYPE_TIME_EXTEND:
3838 /* Internal data, OK to advance */
3839 rb_advance_reader(cpu_buffer);
3842 case RINGBUF_TYPE_TIME_STAMP:
3843 /* FIXME: not implemented */
3844 rb_advance_reader(cpu_buffer);
3847 case RINGBUF_TYPE_DATA:
3849 *ts = cpu_buffer->read_stamp + event->time_delta;
3850 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3851 cpu_buffer->cpu, ts);
3854 *lost_events = rb_lost_events(cpu_buffer);
3863 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3865 static struct ring_buffer_event *
3866 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3868 struct ring_buffer *buffer;
3869 struct ring_buffer_per_cpu *cpu_buffer;
3870 struct ring_buffer_event *event;
3873 cpu_buffer = iter->cpu_buffer;
3874 buffer = cpu_buffer->buffer;
3877 * Check if someone performed a consuming read to
3878 * the buffer. A consuming read invalidates the iterator
3879 * and we need to reset the iterator in this case.
3881 if (unlikely(iter->cache_read != cpu_buffer->read ||
3882 iter->cache_reader_page != cpu_buffer->reader_page))
3883 rb_iter_reset(iter);
3886 if (ring_buffer_iter_empty(iter))
3890 * We repeat when a time extend is encountered or we hit
3891 * the end of the page. Since the time extend is always attached
3892 * to a data event, we should never loop more than three times.
3893 * Once for going to next page, once on time extend, and
3894 * finally once to get the event.
3895 * (We never hit the following condition more than thrice).
3897 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3900 if (rb_per_cpu_empty(cpu_buffer))
3903 if (iter->head >= rb_page_size(iter->head_page)) {
3908 event = rb_iter_head_event(iter);
3910 switch (event->type_len) {
3911 case RINGBUF_TYPE_PADDING:
3912 if (rb_null_event(event)) {
3916 rb_advance_iter(iter);
3919 case RINGBUF_TYPE_TIME_EXTEND:
3920 /* Internal data, OK to advance */
3921 rb_advance_iter(iter);
3924 case RINGBUF_TYPE_TIME_STAMP:
3925 /* FIXME: not implemented */
3926 rb_advance_iter(iter);
3929 case RINGBUF_TYPE_DATA:
3931 *ts = iter->read_stamp + event->time_delta;
3932 ring_buffer_normalize_time_stamp(buffer,
3933 cpu_buffer->cpu, ts);
3943 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3945 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3947 if (likely(!in_nmi())) {
3948 raw_spin_lock(&cpu_buffer->reader_lock);
3953 * If an NMI die dumps out the content of the ring buffer
3954 * trylock must be used to prevent a deadlock if the NMI
3955 * preempted a task that holds the ring buffer locks. If
3956 * we get the lock then all is fine, if not, then continue
3957 * to do the read, but this can corrupt the ring buffer,
3958 * so it must be permanently disabled from future writes.
3959 * Reading from NMI is a oneshot deal.
3961 if (raw_spin_trylock(&cpu_buffer->reader_lock))
3964 /* Continue without locking, but disable the ring buffer */
3965 atomic_inc(&cpu_buffer->record_disabled);
3970 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3973 raw_spin_unlock(&cpu_buffer->reader_lock);
3978 * ring_buffer_peek - peek at the next event to be read
3979 * @buffer: The ring buffer to read
3980 * @cpu: The cpu to peak at
3981 * @ts: The timestamp counter of this event.
3982 * @lost_events: a variable to store if events were lost (may be NULL)
3984 * This will return the event that will be read next, but does
3985 * not consume the data.
3987 struct ring_buffer_event *
3988 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3989 unsigned long *lost_events)
3991 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3992 struct ring_buffer_event *event;
3993 unsigned long flags;
3996 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4000 local_irq_save(flags);
4001 dolock = rb_reader_lock(cpu_buffer);
4002 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4003 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4004 rb_advance_reader(cpu_buffer);
4005 rb_reader_unlock(cpu_buffer, dolock);
4006 local_irq_restore(flags);
4008 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4015 * ring_buffer_iter_peek - peek at the next event to be read
4016 * @iter: The ring buffer iterator
4017 * @ts: The timestamp counter of this event.
4019 * This will return the event that will be read next, but does
4020 * not increment the iterator.
4022 struct ring_buffer_event *
4023 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4025 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4026 struct ring_buffer_event *event;
4027 unsigned long flags;
4030 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4031 event = rb_iter_peek(iter, ts);
4032 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4034 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4041 * ring_buffer_consume - return an event and consume it
4042 * @buffer: The ring buffer to get the next event from
4043 * @cpu: the cpu to read the buffer from
4044 * @ts: a variable to store the timestamp (may be NULL)
4045 * @lost_events: a variable to store if events were lost (may be NULL)
4047 * Returns the next event in the ring buffer, and that event is consumed.
4048 * Meaning, that sequential reads will keep returning a different event,
4049 * and eventually empty the ring buffer if the producer is slower.
4051 struct ring_buffer_event *
4052 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
4053 unsigned long *lost_events)
4055 struct ring_buffer_per_cpu *cpu_buffer;
4056 struct ring_buffer_event *event = NULL;
4057 unsigned long flags;
4061 /* might be called in atomic */
4064 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4067 cpu_buffer = buffer->buffers[cpu];
4068 local_irq_save(flags);
4069 dolock = rb_reader_lock(cpu_buffer);
4071 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4073 cpu_buffer->lost_events = 0;
4074 rb_advance_reader(cpu_buffer);
4077 rb_reader_unlock(cpu_buffer, dolock);
4078 local_irq_restore(flags);
4083 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4088 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4091 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4092 * @buffer: The ring buffer to read from
4093 * @cpu: The cpu buffer to iterate over
4094 * @flags: gfp flags to use for memory allocation
4096 * This performs the initial preparations necessary to iterate
4097 * through the buffer. Memory is allocated, buffer recording
4098 * is disabled, and the iterator pointer is returned to the caller.
4100 * Disabling buffer recordng prevents the reading from being
4101 * corrupted. This is not a consuming read, so a producer is not
4104 * After a sequence of ring_buffer_read_prepare calls, the user is
4105 * expected to make at least one call to ring_buffer_read_prepare_sync.
4106 * Afterwards, ring_buffer_read_start is invoked to get things going
4109 * This overall must be paired with ring_buffer_read_finish.
4111 struct ring_buffer_iter *
4112 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu, gfp_t flags)
4114 struct ring_buffer_per_cpu *cpu_buffer;
4115 struct ring_buffer_iter *iter;
4117 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4120 iter = kmalloc(sizeof(*iter), flags);
4124 cpu_buffer = buffer->buffers[cpu];
4126 iter->cpu_buffer = cpu_buffer;
4128 atomic_inc(&buffer->resize_disabled);
4129 atomic_inc(&cpu_buffer->record_disabled);
4133 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4136 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4138 * All previously invoked ring_buffer_read_prepare calls to prepare
4139 * iterators will be synchronized. Afterwards, read_buffer_read_start
4140 * calls on those iterators are allowed.
4143 ring_buffer_read_prepare_sync(void)
4145 synchronize_sched();
4147 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4150 * ring_buffer_read_start - start a non consuming read of the buffer
4151 * @iter: The iterator returned by ring_buffer_read_prepare
4153 * This finalizes the startup of an iteration through the buffer.
4154 * The iterator comes from a call to ring_buffer_read_prepare and
4155 * an intervening ring_buffer_read_prepare_sync must have been
4158 * Must be paired with ring_buffer_read_finish.
4161 ring_buffer_read_start(struct ring_buffer_iter *iter)
4163 struct ring_buffer_per_cpu *cpu_buffer;
4164 unsigned long flags;
4169 cpu_buffer = iter->cpu_buffer;
4171 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4172 arch_spin_lock(&cpu_buffer->lock);
4173 rb_iter_reset(iter);
4174 arch_spin_unlock(&cpu_buffer->lock);
4175 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4177 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4180 * ring_buffer_read_finish - finish reading the iterator of the buffer
4181 * @iter: The iterator retrieved by ring_buffer_start
4183 * This re-enables the recording to the buffer, and frees the
4187 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4189 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4190 unsigned long flags;
4193 * Ring buffer is disabled from recording, here's a good place
4194 * to check the integrity of the ring buffer.
4195 * Must prevent readers from trying to read, as the check
4196 * clears the HEAD page and readers require it.
4198 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4199 rb_check_pages(cpu_buffer);
4200 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4202 atomic_dec(&cpu_buffer->record_disabled);
4203 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4206 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4209 * ring_buffer_read - read the next item in the ring buffer by the iterator
4210 * @iter: The ring buffer iterator
4211 * @ts: The time stamp of the event read.
4213 * This reads the next event in the ring buffer and increments the iterator.
4215 struct ring_buffer_event *
4216 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4218 struct ring_buffer_event *event;
4219 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4220 unsigned long flags;
4222 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4224 event = rb_iter_peek(iter, ts);
4228 if (event->type_len == RINGBUF_TYPE_PADDING)
4231 rb_advance_iter(iter);
4233 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4237 EXPORT_SYMBOL_GPL(ring_buffer_read);
4240 * ring_buffer_size - return the size of the ring buffer (in bytes)
4241 * @buffer: The ring buffer.
4243 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4246 * Earlier, this method returned
4247 * BUF_PAGE_SIZE * buffer->nr_pages
4248 * Since the nr_pages field is now removed, we have converted this to
4249 * return the per cpu buffer value.
4251 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4254 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4256 EXPORT_SYMBOL_GPL(ring_buffer_size);
4259 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4261 rb_head_page_deactivate(cpu_buffer);
4263 cpu_buffer->head_page
4264 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4265 local_set(&cpu_buffer->head_page->write, 0);
4266 local_set(&cpu_buffer->head_page->entries, 0);
4267 local_set(&cpu_buffer->head_page->page->commit, 0);
4269 cpu_buffer->head_page->read = 0;
4271 cpu_buffer->tail_page = cpu_buffer->head_page;
4272 cpu_buffer->commit_page = cpu_buffer->head_page;
4274 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4275 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4276 local_set(&cpu_buffer->reader_page->write, 0);
4277 local_set(&cpu_buffer->reader_page->entries, 0);
4278 local_set(&cpu_buffer->reader_page->page->commit, 0);
4279 cpu_buffer->reader_page->read = 0;
4281 local_set(&cpu_buffer->entries_bytes, 0);
4282 local_set(&cpu_buffer->overrun, 0);
4283 local_set(&cpu_buffer->commit_overrun, 0);
4284 local_set(&cpu_buffer->dropped_events, 0);
4285 local_set(&cpu_buffer->entries, 0);
4286 local_set(&cpu_buffer->committing, 0);
4287 local_set(&cpu_buffer->commits, 0);
4288 cpu_buffer->read = 0;
4289 cpu_buffer->read_bytes = 0;
4291 cpu_buffer->write_stamp = 0;
4292 cpu_buffer->read_stamp = 0;
4294 cpu_buffer->lost_events = 0;
4295 cpu_buffer->last_overrun = 0;
4297 rb_head_page_activate(cpu_buffer);
4301 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4302 * @buffer: The ring buffer to reset a per cpu buffer of
4303 * @cpu: The CPU buffer to be reset
4305 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4307 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4308 unsigned long flags;
4310 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4312 /* prevent another thread from changing buffer sizes */
4313 mutex_lock(&buffer->mutex);
4315 atomic_inc(&buffer->resize_disabled);
4316 atomic_inc(&cpu_buffer->record_disabled);
4318 /* Make sure all commits have finished */
4319 synchronize_sched();
4321 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4323 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4326 arch_spin_lock(&cpu_buffer->lock);
4328 rb_reset_cpu(cpu_buffer);
4330 arch_spin_unlock(&cpu_buffer->lock);
4333 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4335 atomic_dec(&cpu_buffer->record_disabled);
4336 atomic_dec(&buffer->resize_disabled);
4338 mutex_unlock(&buffer->mutex);
4340 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4343 * ring_buffer_reset - reset a ring buffer
4344 * @buffer: The ring buffer to reset all cpu buffers
4346 void ring_buffer_reset(struct ring_buffer *buffer)
4350 for_each_buffer_cpu(buffer, cpu)
4351 ring_buffer_reset_cpu(buffer, cpu);
4353 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4356 * rind_buffer_empty - is the ring buffer empty?
4357 * @buffer: The ring buffer to test
4359 bool ring_buffer_empty(struct ring_buffer *buffer)
4361 struct ring_buffer_per_cpu *cpu_buffer;
4362 unsigned long flags;
4367 /* yes this is racy, but if you don't like the race, lock the buffer */
4368 for_each_buffer_cpu(buffer, cpu) {
4369 cpu_buffer = buffer->buffers[cpu];
4370 local_irq_save(flags);
4371 dolock = rb_reader_lock(cpu_buffer);
4372 ret = rb_per_cpu_empty(cpu_buffer);
4373 rb_reader_unlock(cpu_buffer, dolock);
4374 local_irq_restore(flags);
4382 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4385 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4386 * @buffer: The ring buffer
4387 * @cpu: The CPU buffer to test
4389 bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4391 struct ring_buffer_per_cpu *cpu_buffer;
4392 unsigned long flags;
4396 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4399 cpu_buffer = buffer->buffers[cpu];
4400 local_irq_save(flags);
4401 dolock = rb_reader_lock(cpu_buffer);
4402 ret = rb_per_cpu_empty(cpu_buffer);
4403 rb_reader_unlock(cpu_buffer, dolock);
4404 local_irq_restore(flags);
4408 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4410 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4412 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4413 * @buffer_a: One buffer to swap with
4414 * @buffer_b: The other buffer to swap with
4416 * This function is useful for tracers that want to take a "snapshot"
4417 * of a CPU buffer and has another back up buffer lying around.
4418 * it is expected that the tracer handles the cpu buffer not being
4419 * used at the moment.
4421 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4422 struct ring_buffer *buffer_b, int cpu)
4424 struct ring_buffer_per_cpu *cpu_buffer_a;
4425 struct ring_buffer_per_cpu *cpu_buffer_b;
4428 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4429 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4432 cpu_buffer_a = buffer_a->buffers[cpu];
4433 cpu_buffer_b = buffer_b->buffers[cpu];
4435 /* At least make sure the two buffers are somewhat the same */
4436 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4441 if (atomic_read(&buffer_a->record_disabled))
4444 if (atomic_read(&buffer_b->record_disabled))
4447 if (atomic_read(&cpu_buffer_a->record_disabled))
4450 if (atomic_read(&cpu_buffer_b->record_disabled))
4454 * We can't do a synchronize_sched here because this
4455 * function can be called in atomic context.
4456 * Normally this will be called from the same CPU as cpu.
4457 * If not it's up to the caller to protect this.
4459 atomic_inc(&cpu_buffer_a->record_disabled);
4460 atomic_inc(&cpu_buffer_b->record_disabled);
4463 if (local_read(&cpu_buffer_a->committing))
4465 if (local_read(&cpu_buffer_b->committing))
4468 buffer_a->buffers[cpu] = cpu_buffer_b;
4469 buffer_b->buffers[cpu] = cpu_buffer_a;
4471 cpu_buffer_b->buffer = buffer_a;
4472 cpu_buffer_a->buffer = buffer_b;
4477 atomic_dec(&cpu_buffer_a->record_disabled);
4478 atomic_dec(&cpu_buffer_b->record_disabled);
4482 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4483 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4486 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4487 * @buffer: the buffer to allocate for.
4488 * @cpu: the cpu buffer to allocate.
4490 * This function is used in conjunction with ring_buffer_read_page.
4491 * When reading a full page from the ring buffer, these functions
4492 * can be used to speed up the process. The calling function should
4493 * allocate a few pages first with this function. Then when it
4494 * needs to get pages from the ring buffer, it passes the result
4495 * of this function into ring_buffer_read_page, which will swap
4496 * the page that was allocated, with the read page of the buffer.
4499 * The page allocated, or NULL on error.
4501 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4503 struct buffer_data_page *bpage;
4506 page = alloc_pages_node(cpu_to_node(cpu),
4507 GFP_KERNEL | __GFP_NORETRY, 0);
4511 bpage = page_address(page);
4513 rb_init_page(bpage);
4517 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4520 * ring_buffer_free_read_page - free an allocated read page
4521 * @buffer: the buffer the page was allocate for
4522 * @data: the page to free
4524 * Free a page allocated from ring_buffer_alloc_read_page.
4526 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4528 free_page((unsigned long)data);
4530 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4533 * ring_buffer_read_page - extract a page from the ring buffer
4534 * @buffer: buffer to extract from
4535 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4536 * @len: amount to extract
4537 * @cpu: the cpu of the buffer to extract
4538 * @full: should the extraction only happen when the page is full.
4540 * This function will pull out a page from the ring buffer and consume it.
4541 * @data_page must be the address of the variable that was returned
4542 * from ring_buffer_alloc_read_page. This is because the page might be used
4543 * to swap with a page in the ring buffer.
4546 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4549 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4551 * process_page(rpage, ret);
4553 * When @full is set, the function will not return true unless
4554 * the writer is off the reader page.
4556 * Note: it is up to the calling functions to handle sleeps and wakeups.
4557 * The ring buffer can be used anywhere in the kernel and can not
4558 * blindly call wake_up. The layer that uses the ring buffer must be
4559 * responsible for that.
4562 * >=0 if data has been transferred, returns the offset of consumed data.
4563 * <0 if no data has been transferred.
4565 int ring_buffer_read_page(struct ring_buffer *buffer,
4566 void **data_page, size_t len, int cpu, int full)
4568 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4569 struct ring_buffer_event *event;
4570 struct buffer_data_page *bpage;
4571 struct buffer_page *reader;
4572 unsigned long missed_events;
4573 unsigned long flags;
4574 unsigned int commit;
4579 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4583 * If len is not big enough to hold the page header, then
4584 * we can not copy anything.
4586 if (len <= BUF_PAGE_HDR_SIZE)
4589 len -= BUF_PAGE_HDR_SIZE;
4598 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4600 reader = rb_get_reader_page(cpu_buffer);
4604 event = rb_reader_event(cpu_buffer);
4606 read = reader->read;
4607 commit = rb_page_commit(reader);
4609 /* Check if any events were dropped */
4610 missed_events = cpu_buffer->lost_events;
4613 * If this page has been partially read or
4614 * if len is not big enough to read the rest of the page or
4615 * a writer is still on the page, then
4616 * we must copy the data from the page to the buffer.
4617 * Otherwise, we can simply swap the page with the one passed in.
4619 if (read || (len < (commit - read)) ||
4620 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4621 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4622 unsigned int rpos = read;
4623 unsigned int pos = 0;
4629 if (len > (commit - read))
4630 len = (commit - read);
4632 /* Always keep the time extend and data together */
4633 size = rb_event_ts_length(event);
4638 /* save the current timestamp, since the user will need it */
4639 save_timestamp = cpu_buffer->read_stamp;
4641 /* Need to copy one event at a time */
4643 /* We need the size of one event, because
4644 * rb_advance_reader only advances by one event,
4645 * whereas rb_event_ts_length may include the size of
4646 * one or two events.
4647 * We have already ensured there's enough space if this
4648 * is a time extend. */
4649 size = rb_event_length(event);
4650 memcpy(bpage->data + pos, rpage->data + rpos, size);
4654 rb_advance_reader(cpu_buffer);
4655 rpos = reader->read;
4661 event = rb_reader_event(cpu_buffer);
4662 /* Always keep the time extend and data together */
4663 size = rb_event_ts_length(event);
4664 } while (len >= size);
4667 local_set(&bpage->commit, pos);
4668 bpage->time_stamp = save_timestamp;
4670 /* we copied everything to the beginning */
4673 /* update the entry counter */
4674 cpu_buffer->read += rb_page_entries(reader);
4675 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4677 /* swap the pages */
4678 rb_init_page(bpage);
4679 bpage = reader->page;
4680 reader->page = *data_page;
4681 local_set(&reader->write, 0);
4682 local_set(&reader->entries, 0);
4687 * Use the real_end for the data size,
4688 * This gives us a chance to store the lost events
4691 if (reader->real_end)
4692 local_set(&bpage->commit, reader->real_end);
4696 cpu_buffer->lost_events = 0;
4698 commit = local_read(&bpage->commit);
4700 * Set a flag in the commit field if we lost events
4702 if (missed_events) {
4703 /* If there is room at the end of the page to save the
4704 * missed events, then record it there.
4706 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4707 memcpy(&bpage->data[commit], &missed_events,
4708 sizeof(missed_events));
4709 local_add(RB_MISSED_STORED, &bpage->commit);
4710 commit += sizeof(missed_events);
4712 local_add(RB_MISSED_EVENTS, &bpage->commit);
4716 * This page may be off to user land. Zero it out here.
4718 if (commit < BUF_PAGE_SIZE)
4719 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4722 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4727 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4729 #ifdef CONFIG_HOTPLUG_CPU
4730 static int rb_cpu_notify(struct notifier_block *self,
4731 unsigned long action, void *hcpu)
4733 struct ring_buffer *buffer =
4734 container_of(self, struct ring_buffer, cpu_notify);
4735 long cpu = (long)hcpu;
4738 unsigned long nr_pages;
4741 case CPU_UP_PREPARE:
4742 case CPU_UP_PREPARE_FROZEN:
4743 if (cpumask_test_cpu(cpu, buffer->cpumask))
4748 /* check if all cpu sizes are same */
4749 for_each_buffer_cpu(buffer, cpu_i) {
4750 /* fill in the size from first enabled cpu */
4752 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4753 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4758 /* allocate minimum pages, user can later expand it */
4761 buffer->buffers[cpu] =
4762 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4763 if (!buffer->buffers[cpu]) {
4764 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4769 cpumask_set_cpu(cpu, buffer->cpumask);
4771 case CPU_DOWN_PREPARE:
4772 case CPU_DOWN_PREPARE_FROZEN:
4775 * If we were to free the buffer, then the user would
4776 * lose any trace that was in the buffer.
4786 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4788 * This is a basic integrity check of the ring buffer.
4789 * Late in the boot cycle this test will run when configured in.
4790 * It will kick off a thread per CPU that will go into a loop
4791 * writing to the per cpu ring buffer various sizes of data.
4792 * Some of the data will be large items, some small.
4794 * Another thread is created that goes into a spin, sending out
4795 * IPIs to the other CPUs to also write into the ring buffer.
4796 * this is to test the nesting ability of the buffer.
4798 * Basic stats are recorded and reported. If something in the
4799 * ring buffer should happen that's not expected, a big warning
4800 * is displayed and all ring buffers are disabled.
4802 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4804 struct rb_test_data {
4805 struct ring_buffer *buffer;
4806 unsigned long events;
4807 unsigned long bytes_written;
4808 unsigned long bytes_alloc;
4809 unsigned long bytes_dropped;
4810 unsigned long events_nested;
4811 unsigned long bytes_written_nested;
4812 unsigned long bytes_alloc_nested;
4813 unsigned long bytes_dropped_nested;
4814 int min_size_nested;
4815 int max_size_nested;
4822 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4825 #define RB_TEST_BUFFER_SIZE 1048576
4827 static char rb_string[] __initdata =
4828 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4829 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4830 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4832 static bool rb_test_started __initdata;
4839 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4841 struct ring_buffer_event *event;
4842 struct rb_item *item;
4849 /* Have nested writes different that what is written */
4850 cnt = data->cnt + (nested ? 27 : 0);
4852 /* Multiply cnt by ~e, to make some unique increment */
4853 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4855 len = size + sizeof(struct rb_item);
4857 started = rb_test_started;
4858 /* read rb_test_started before checking buffer enabled */
4861 event = ring_buffer_lock_reserve(data->buffer, len);
4863 /* Ignore dropped events before test starts. */
4866 data->bytes_dropped += len;
4868 data->bytes_dropped_nested += len;
4873 event_len = ring_buffer_event_length(event);
4875 if (RB_WARN_ON(data->buffer, event_len < len))
4878 item = ring_buffer_event_data(event);
4880 memcpy(item->str, rb_string, size);
4883 data->bytes_alloc_nested += event_len;
4884 data->bytes_written_nested += len;
4885 data->events_nested++;
4886 if (!data->min_size_nested || len < data->min_size_nested)
4887 data->min_size_nested = len;
4888 if (len > data->max_size_nested)
4889 data->max_size_nested = len;
4891 data->bytes_alloc += event_len;
4892 data->bytes_written += len;
4894 if (!data->min_size || len < data->min_size)
4895 data->max_size = len;
4896 if (len > data->max_size)
4897 data->max_size = len;
4901 ring_buffer_unlock_commit(data->buffer, event);
4906 static __init int rb_test(void *arg)
4908 struct rb_test_data *data = arg;
4910 while (!kthread_should_stop()) {
4911 rb_write_something(data, false);
4914 set_current_state(TASK_INTERRUPTIBLE);
4915 /* Now sleep between a min of 100-300us and a max of 1ms */
4916 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4922 static __init void rb_ipi(void *ignore)
4924 struct rb_test_data *data;
4925 int cpu = smp_processor_id();
4927 data = &rb_data[cpu];
4928 rb_write_something(data, true);
4931 static __init int rb_hammer_test(void *arg)
4933 while (!kthread_should_stop()) {
4935 /* Send an IPI to all cpus to write data! */
4936 smp_call_function(rb_ipi, NULL, 1);
4937 /* No sleep, but for non preempt, let others run */
4944 static __init int test_ringbuffer(void)
4946 struct task_struct *rb_hammer;
4947 struct ring_buffer *buffer;
4951 pr_info("Running ring buffer tests...\n");
4953 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4954 if (WARN_ON(!buffer))
4957 /* Disable buffer so that threads can't write to it yet */
4958 ring_buffer_record_off(buffer);
4960 for_each_online_cpu(cpu) {
4961 rb_data[cpu].buffer = buffer;
4962 rb_data[cpu].cpu = cpu;
4963 rb_data[cpu].cnt = cpu;
4964 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4965 "rbtester/%d", cpu);
4966 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
4967 pr_cont("FAILED\n");
4968 ret = PTR_ERR(rb_threads[cpu]);
4972 kthread_bind(rb_threads[cpu], cpu);
4973 wake_up_process(rb_threads[cpu]);
4976 /* Now create the rb hammer! */
4977 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4978 if (WARN_ON(IS_ERR(rb_hammer))) {
4979 pr_cont("FAILED\n");
4980 ret = PTR_ERR(rb_hammer);
4984 ring_buffer_record_on(buffer);
4986 * Show buffer is enabled before setting rb_test_started.
4987 * Yes there's a small race window where events could be
4988 * dropped and the thread wont catch it. But when a ring
4989 * buffer gets enabled, there will always be some kind of
4990 * delay before other CPUs see it. Thus, we don't care about
4991 * those dropped events. We care about events dropped after
4992 * the threads see that the buffer is active.
4995 rb_test_started = true;
4997 set_current_state(TASK_INTERRUPTIBLE);
4998 /* Just run for 10 seconds */;
4999 schedule_timeout(10 * HZ);
5001 kthread_stop(rb_hammer);
5004 for_each_online_cpu(cpu) {
5005 if (!rb_threads[cpu])
5007 kthread_stop(rb_threads[cpu]);
5010 ring_buffer_free(buffer);
5015 pr_info("finished\n");
5016 for_each_online_cpu(cpu) {
5017 struct ring_buffer_event *event;
5018 struct rb_test_data *data = &rb_data[cpu];
5019 struct rb_item *item;
5020 unsigned long total_events;
5021 unsigned long total_dropped;
5022 unsigned long total_written;
5023 unsigned long total_alloc;
5024 unsigned long total_read = 0;
5025 unsigned long total_size = 0;
5026 unsigned long total_len = 0;
5027 unsigned long total_lost = 0;
5030 int small_event_size;
5034 total_events = data->events + data->events_nested;
5035 total_written = data->bytes_written + data->bytes_written_nested;
5036 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5037 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5039 big_event_size = data->max_size + data->max_size_nested;
5040 small_event_size = data->min_size + data->min_size_nested;
5042 pr_info("CPU %d:\n", cpu);
5043 pr_info(" events: %ld\n", total_events);
5044 pr_info(" dropped bytes: %ld\n", total_dropped);
5045 pr_info(" alloced bytes: %ld\n", total_alloc);
5046 pr_info(" written bytes: %ld\n", total_written);
5047 pr_info(" biggest event: %d\n", big_event_size);
5048 pr_info(" smallest event: %d\n", small_event_size);
5050 if (RB_WARN_ON(buffer, total_dropped))
5055 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5057 item = ring_buffer_event_data(event);
5058 total_len += ring_buffer_event_length(event);
5059 total_size += item->size + sizeof(struct rb_item);
5060 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5061 pr_info("FAILED!\n");
5062 pr_info("buffer had: %.*s\n", item->size, item->str);
5063 pr_info("expected: %.*s\n", item->size, rb_string);
5064 RB_WARN_ON(buffer, 1);
5075 pr_info(" read events: %ld\n", total_read);
5076 pr_info(" lost events: %ld\n", total_lost);
5077 pr_info(" total events: %ld\n", total_lost + total_read);
5078 pr_info(" recorded len bytes: %ld\n", total_len);
5079 pr_info(" recorded size bytes: %ld\n", total_size);
5081 pr_info(" With dropped events, record len and size may not match\n"
5082 " alloced and written from above\n");
5084 if (RB_WARN_ON(buffer, total_len != total_alloc ||
5085 total_size != total_written))
5088 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5094 pr_info("Ring buffer PASSED!\n");
5096 ring_buffer_free(buffer);
5100 late_initcall(test_ringbuffer);
5101 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */