4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/ring_buffer.h>
7 #include <linux/trace_clock.h>
8 #include <linux/spinlock.h>
9 #include <linux/debugfs.h>
10 #include <linux/uaccess.h>
11 #include <linux/hardirq.h>
12 #include <linux/kmemcheck.h>
13 #include <linux/module.h>
14 #include <linux/percpu.h>
15 #include <linux/mutex.h>
16 #include <linux/slab.h>
17 #include <linux/init.h>
18 #include <linux/hash.h>
19 #include <linux/list.h>
20 #include <linux/cpu.h>
23 #include <asm/local.h>
27 * The ring buffer header is special. We must manually up keep it.
29 int ring_buffer_print_entry_header(struct trace_seq
*s
)
33 ret
= trace_seq_printf(s
, "# compressed entry header\n");
34 ret
= trace_seq_printf(s
, "\ttype_len : 5 bits\n");
35 ret
= trace_seq_printf(s
, "\ttime_delta : 27 bits\n");
36 ret
= trace_seq_printf(s
, "\tarray : 32 bits\n");
37 ret
= trace_seq_printf(s
, "\n");
38 ret
= trace_seq_printf(s
, "\tpadding : type == %d\n",
39 RINGBUF_TYPE_PADDING
);
40 ret
= trace_seq_printf(s
, "\ttime_extend : type == %d\n",
41 RINGBUF_TYPE_TIME_EXTEND
);
42 ret
= trace_seq_printf(s
, "\tdata max type_len == %d\n",
43 RINGBUF_TYPE_DATA_TYPE_LEN_MAX
);
49 * The ring buffer is made up of a list of pages. A separate list of pages is
50 * allocated for each CPU. A writer may only write to a buffer that is
51 * associated with the CPU it is currently executing on. A reader may read
52 * from any per cpu buffer.
54 * The reader is special. For each per cpu buffer, the reader has its own
55 * reader page. When a reader has read the entire reader page, this reader
56 * page is swapped with another page in the ring buffer.
58 * Now, as long as the writer is off the reader page, the reader can do what
59 * ever it wants with that page. The writer will never write to that page
60 * again (as long as it is out of the ring buffer).
62 * Here's some silly ASCII art.
65 * |reader| RING BUFFER
67 * +------+ +---+ +---+ +---+
76 * |reader| RING BUFFER
77 * |page |------------------v
78 * +------+ +---+ +---+ +---+
87 * |reader| RING BUFFER
88 * |page |------------------v
89 * +------+ +---+ +---+ +---+
94 * +------------------------------+
98 * |buffer| RING BUFFER
99 * |page |------------------v
100 * +------+ +---+ +---+ +---+
102 * | New +---+ +---+ +---+
105 * +------------------------------+
108 * After we make this swap, the reader can hand this page off to the splice
109 * code and be done with it. It can even allocate a new page if it needs to
110 * and swap that into the ring buffer.
112 * We will be using cmpxchg soon to make all this lockless.
117 * A fast way to enable or disable all ring buffers is to
118 * call tracing_on or tracing_off. Turning off the ring buffers
119 * prevents all ring buffers from being recorded to.
120 * Turning this switch on, makes it OK to write to the
121 * ring buffer, if the ring buffer is enabled itself.
123 * There's three layers that must be on in order to write
124 * to the ring buffer.
126 * 1) This global flag must be set.
127 * 2) The ring buffer must be enabled for recording.
128 * 3) The per cpu buffer must be enabled for recording.
130 * In case of an anomaly, this global flag has a bit set that
131 * will permantly disable all ring buffers.
135 * Global flag to disable all recording to ring buffers
136 * This has two bits: ON, DISABLED
140 * 0 0 : ring buffers are off
141 * 1 0 : ring buffers are on
142 * X 1 : ring buffers are permanently disabled
146 RB_BUFFERS_ON_BIT
= 0,
147 RB_BUFFERS_DISABLED_BIT
= 1,
151 RB_BUFFERS_ON
= 1 << RB_BUFFERS_ON_BIT
,
152 RB_BUFFERS_DISABLED
= 1 << RB_BUFFERS_DISABLED_BIT
,
155 static unsigned long ring_buffer_flags __read_mostly
= RB_BUFFERS_ON
;
157 /* Used for individual buffers (after the counter) */
158 #define RB_BUFFER_OFF (1 << 20)
160 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
163 * tracing_off_permanent - permanently disable ring buffers
165 * This function, once called, will disable all ring buffers
168 void tracing_off_permanent(void)
170 set_bit(RB_BUFFERS_DISABLED_BIT
, &ring_buffer_flags
);
173 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
174 #define RB_ALIGNMENT 4U
175 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
176 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
178 #if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
179 # define RB_FORCE_8BYTE_ALIGNMENT 0
180 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
182 # define RB_FORCE_8BYTE_ALIGNMENT 1
183 # define RB_ARCH_ALIGNMENT 8U
186 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
187 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
190 RB_LEN_TIME_EXTEND
= 8,
191 RB_LEN_TIME_STAMP
= 16,
194 #define skip_time_extend(event) \
195 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
197 static inline int rb_null_event(struct ring_buffer_event
*event
)
199 return event
->type_len
== RINGBUF_TYPE_PADDING
&& !event
->time_delta
;
202 static void rb_event_set_padding(struct ring_buffer_event
*event
)
204 /* padding has a NULL time_delta */
205 event
->type_len
= RINGBUF_TYPE_PADDING
;
206 event
->time_delta
= 0;
210 rb_event_data_length(struct ring_buffer_event
*event
)
215 length
= event
->type_len
* RB_ALIGNMENT
;
217 length
= event
->array
[0];
218 return length
+ RB_EVNT_HDR_SIZE
;
222 * Return the length of the given event. Will return
223 * the length of the time extend if the event is a
226 static inline unsigned
227 rb_event_length(struct ring_buffer_event
*event
)
229 switch (event
->type_len
) {
230 case RINGBUF_TYPE_PADDING
:
231 if (rb_null_event(event
))
234 return event
->array
[0] + RB_EVNT_HDR_SIZE
;
236 case RINGBUF_TYPE_TIME_EXTEND
:
237 return RB_LEN_TIME_EXTEND
;
239 case RINGBUF_TYPE_TIME_STAMP
:
240 return RB_LEN_TIME_STAMP
;
242 case RINGBUF_TYPE_DATA
:
243 return rb_event_data_length(event
);
252 * Return total length of time extend and data,
253 * or just the event length for all other events.
255 static inline unsigned
256 rb_event_ts_length(struct ring_buffer_event
*event
)
260 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
261 /* time extends include the data event after it */
262 len
= RB_LEN_TIME_EXTEND
;
263 event
= skip_time_extend(event
);
265 return len
+ rb_event_length(event
);
269 * ring_buffer_event_length - return the length of the event
270 * @event: the event to get the length of
272 * Returns the size of the data load of a data event.
273 * If the event is something other than a data event, it
274 * returns the size of the event itself. With the exception
275 * of a TIME EXTEND, where it still returns the size of the
276 * data load of the data event after it.
278 unsigned ring_buffer_event_length(struct ring_buffer_event
*event
)
282 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
283 event
= skip_time_extend(event
);
285 length
= rb_event_length(event
);
286 if (event
->type_len
> RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
288 length
-= RB_EVNT_HDR_SIZE
;
289 if (length
> RB_MAX_SMALL_DATA
+ sizeof(event
->array
[0]))
290 length
-= sizeof(event
->array
[0]);
293 EXPORT_SYMBOL_GPL(ring_buffer_event_length
);
295 /* inline for ring buffer fast paths */
297 rb_event_data(struct ring_buffer_event
*event
)
299 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
300 event
= skip_time_extend(event
);
301 BUG_ON(event
->type_len
> RINGBUF_TYPE_DATA_TYPE_LEN_MAX
);
302 /* If length is in len field, then array[0] has the data */
304 return (void *)&event
->array
[0];
305 /* Otherwise length is in array[0] and array[1] has the data */
306 return (void *)&event
->array
[1];
310 * ring_buffer_event_data - return the data of the event
311 * @event: the event to get the data from
313 void *ring_buffer_event_data(struct ring_buffer_event
*event
)
315 return rb_event_data(event
);
317 EXPORT_SYMBOL_GPL(ring_buffer_event_data
);
319 #define for_each_buffer_cpu(buffer, cpu) \
320 for_each_cpu(cpu, buffer->cpumask)
323 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
324 #define TS_DELTA_TEST (~TS_MASK)
326 /* Flag when events were overwritten */
327 #define RB_MISSED_EVENTS (1 << 31)
328 /* Missed count stored at end */
329 #define RB_MISSED_STORED (1 << 30)
331 struct buffer_data_page
{
332 u64 time_stamp
; /* page time stamp */
333 local_t commit
; /* write committed index */
334 unsigned char data
[]; /* data of buffer page */
338 * Note, the buffer_page list must be first. The buffer pages
339 * are allocated in cache lines, which means that each buffer
340 * page will be at the beginning of a cache line, and thus
341 * the least significant bits will be zero. We use this to
342 * add flags in the list struct pointers, to make the ring buffer
346 struct list_head list
; /* list of buffer pages */
347 local_t write
; /* index for next write */
348 unsigned read
; /* index for next read */
349 local_t entries
; /* entries on this page */
350 unsigned long real_end
; /* real end of data */
351 struct buffer_data_page
*page
; /* Actual data page */
355 * The buffer page counters, write and entries, must be reset
356 * atomically when crossing page boundaries. To synchronize this
357 * update, two counters are inserted into the number. One is
358 * the actual counter for the write position or count on the page.
360 * The other is a counter of updaters. Before an update happens
361 * the update partition of the counter is incremented. This will
362 * allow the updater to update the counter atomically.
364 * The counter is 20 bits, and the state data is 12.
366 #define RB_WRITE_MASK 0xfffff
367 #define RB_WRITE_INTCNT (1 << 20)
369 static void rb_init_page(struct buffer_data_page
*bpage
)
371 local_set(&bpage
->commit
, 0);
375 * ring_buffer_page_len - the size of data on the page.
376 * @page: The page to read
378 * Returns the amount of data on the page, including buffer page header.
380 size_t ring_buffer_page_len(void *page
)
382 return local_read(&((struct buffer_data_page
*)page
)->commit
)
387 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
390 static void free_buffer_page(struct buffer_page
*bpage
)
392 free_page((unsigned long)bpage
->page
);
397 * We need to fit the time_stamp delta into 27 bits.
399 static inline int test_time_stamp(u64 delta
)
401 if (delta
& TS_DELTA_TEST
)
406 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
408 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
409 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
411 int ring_buffer_print_page_header(struct trace_seq
*s
)
413 struct buffer_data_page field
;
416 ret
= trace_seq_printf(s
, "\tfield: u64 timestamp;\t"
417 "offset:0;\tsize:%u;\tsigned:%u;\n",
418 (unsigned int)sizeof(field
.time_stamp
),
419 (unsigned int)is_signed_type(u64
));
421 ret
= trace_seq_printf(s
, "\tfield: local_t commit;\t"
422 "offset:%u;\tsize:%u;\tsigned:%u;\n",
423 (unsigned int)offsetof(typeof(field
), commit
),
424 (unsigned int)sizeof(field
.commit
),
425 (unsigned int)is_signed_type(long));
427 ret
= trace_seq_printf(s
, "\tfield: int overwrite;\t"
428 "offset:%u;\tsize:%u;\tsigned:%u;\n",
429 (unsigned int)offsetof(typeof(field
), commit
),
431 (unsigned int)is_signed_type(long));
433 ret
= trace_seq_printf(s
, "\tfield: char data;\t"
434 "offset:%u;\tsize:%u;\tsigned:%u;\n",
435 (unsigned int)offsetof(typeof(field
), data
),
436 (unsigned int)BUF_PAGE_SIZE
,
437 (unsigned int)is_signed_type(char));
443 * head_page == tail_page && head == tail then buffer is empty.
445 struct ring_buffer_per_cpu
{
447 atomic_t record_disabled
;
448 struct ring_buffer
*buffer
;
449 raw_spinlock_t reader_lock
; /* serialize readers */
450 arch_spinlock_t lock
;
451 struct lock_class_key lock_key
;
452 struct list_head
*pages
;
453 struct buffer_page
*head_page
; /* read from head */
454 struct buffer_page
*tail_page
; /* write to tail */
455 struct buffer_page
*commit_page
; /* committed pages */
456 struct buffer_page
*reader_page
;
457 unsigned long lost_events
;
458 unsigned long last_overrun
;
459 local_t entries_bytes
;
460 local_t commit_overrun
;
466 unsigned long read_bytes
;
475 atomic_t record_disabled
;
476 cpumask_var_t cpumask
;
478 struct lock_class_key
*reader_lock_key
;
482 struct ring_buffer_per_cpu
**buffers
;
484 #ifdef CONFIG_HOTPLUG_CPU
485 struct notifier_block cpu_notify
;
490 struct ring_buffer_iter
{
491 struct ring_buffer_per_cpu
*cpu_buffer
;
493 struct buffer_page
*head_page
;
494 struct buffer_page
*cache_reader_page
;
495 unsigned long cache_read
;
499 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
500 #define RB_WARN_ON(b, cond) \
502 int _____ret = unlikely(cond); \
504 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
505 struct ring_buffer_per_cpu *__b = \
507 atomic_inc(&__b->buffer->record_disabled); \
509 atomic_inc(&b->record_disabled); \
515 /* Up this if you want to test the TIME_EXTENTS and normalization */
516 #define DEBUG_SHIFT 0
518 static inline u64
rb_time_stamp(struct ring_buffer
*buffer
)
520 /* shift to debug/test normalization and TIME_EXTENTS */
521 return buffer
->clock() << DEBUG_SHIFT
;
524 u64
ring_buffer_time_stamp(struct ring_buffer
*buffer
, int cpu
)
528 preempt_disable_notrace();
529 time
= rb_time_stamp(buffer
);
530 preempt_enable_no_resched_notrace();
534 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp
);
536 void ring_buffer_normalize_time_stamp(struct ring_buffer
*buffer
,
539 /* Just stupid testing the normalize function and deltas */
542 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp
);
545 * Making the ring buffer lockless makes things tricky.
546 * Although writes only happen on the CPU that they are on,
547 * and they only need to worry about interrupts. Reads can
550 * The reader page is always off the ring buffer, but when the
551 * reader finishes with a page, it needs to swap its page with
552 * a new one from the buffer. The reader needs to take from
553 * the head (writes go to the tail). But if a writer is in overwrite
554 * mode and wraps, it must push the head page forward.
556 * Here lies the problem.
558 * The reader must be careful to replace only the head page, and
559 * not another one. As described at the top of the file in the
560 * ASCII art, the reader sets its old page to point to the next
561 * page after head. It then sets the page after head to point to
562 * the old reader page. But if the writer moves the head page
563 * during this operation, the reader could end up with the tail.
565 * We use cmpxchg to help prevent this race. We also do something
566 * special with the page before head. We set the LSB to 1.
568 * When the writer must push the page forward, it will clear the
569 * bit that points to the head page, move the head, and then set
570 * the bit that points to the new head page.
572 * We also don't want an interrupt coming in and moving the head
573 * page on another writer. Thus we use the second LSB to catch
576 * head->list->prev->next bit 1 bit 0
579 * Points to head page 0 1
582 * Note we can not trust the prev pointer of the head page, because:
584 * +----+ +-----+ +-----+
585 * | |------>| T |---X--->| N |
587 * +----+ +-----+ +-----+
590 * +----------| R |----------+ |
594 * Key: ---X--> HEAD flag set in pointer
599 * (see __rb_reserve_next() to see where this happens)
601 * What the above shows is that the reader just swapped out
602 * the reader page with a page in the buffer, but before it
603 * could make the new header point back to the new page added
604 * it was preempted by a writer. The writer moved forward onto
605 * the new page added by the reader and is about to move forward
608 * You can see, it is legitimate for the previous pointer of
609 * the head (or any page) not to point back to itself. But only
613 #define RB_PAGE_NORMAL 0UL
614 #define RB_PAGE_HEAD 1UL
615 #define RB_PAGE_UPDATE 2UL
618 #define RB_FLAG_MASK 3UL
620 /* PAGE_MOVED is not part of the mask */
621 #define RB_PAGE_MOVED 4UL
624 * rb_list_head - remove any bit
626 static struct list_head
*rb_list_head(struct list_head
*list
)
628 unsigned long val
= (unsigned long)list
;
630 return (struct list_head
*)(val
& ~RB_FLAG_MASK
);
634 * rb_is_head_page - test if the given page is the head page
636 * Because the reader may move the head_page pointer, we can
637 * not trust what the head page is (it may be pointing to
638 * the reader page). But if the next page is a header page,
639 * its flags will be non zero.
642 rb_is_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
643 struct buffer_page
*page
, struct list_head
*list
)
647 val
= (unsigned long)list
->next
;
649 if ((val
& ~RB_FLAG_MASK
) != (unsigned long)&page
->list
)
650 return RB_PAGE_MOVED
;
652 return val
& RB_FLAG_MASK
;
658 * The unique thing about the reader page, is that, if the
659 * writer is ever on it, the previous pointer never points
660 * back to the reader page.
662 static int rb_is_reader_page(struct buffer_page
*page
)
664 struct list_head
*list
= page
->list
.prev
;
666 return rb_list_head(list
->next
) != &page
->list
;
670 * rb_set_list_to_head - set a list_head to be pointing to head.
672 static void rb_set_list_to_head(struct ring_buffer_per_cpu
*cpu_buffer
,
673 struct list_head
*list
)
677 ptr
= (unsigned long *)&list
->next
;
678 *ptr
|= RB_PAGE_HEAD
;
679 *ptr
&= ~RB_PAGE_UPDATE
;
683 * rb_head_page_activate - sets up head page
685 static void rb_head_page_activate(struct ring_buffer_per_cpu
*cpu_buffer
)
687 struct buffer_page
*head
;
689 head
= cpu_buffer
->head_page
;
694 * Set the previous list pointer to have the HEAD flag.
696 rb_set_list_to_head(cpu_buffer
, head
->list
.prev
);
699 static void rb_list_head_clear(struct list_head
*list
)
701 unsigned long *ptr
= (unsigned long *)&list
->next
;
703 *ptr
&= ~RB_FLAG_MASK
;
707 * rb_head_page_dactivate - clears head page ptr (for free list)
710 rb_head_page_deactivate(struct ring_buffer_per_cpu
*cpu_buffer
)
712 struct list_head
*hd
;
714 /* Go through the whole list and clear any pointers found. */
715 rb_list_head_clear(cpu_buffer
->pages
);
717 list_for_each(hd
, cpu_buffer
->pages
)
718 rb_list_head_clear(hd
);
721 static int rb_head_page_set(struct ring_buffer_per_cpu
*cpu_buffer
,
722 struct buffer_page
*head
,
723 struct buffer_page
*prev
,
724 int old_flag
, int new_flag
)
726 struct list_head
*list
;
727 unsigned long val
= (unsigned long)&head
->list
;
732 val
&= ~RB_FLAG_MASK
;
734 ret
= cmpxchg((unsigned long *)&list
->next
,
735 val
| old_flag
, val
| new_flag
);
737 /* check if the reader took the page */
738 if ((ret
& ~RB_FLAG_MASK
) != val
)
739 return RB_PAGE_MOVED
;
741 return ret
& RB_FLAG_MASK
;
744 static int rb_head_page_set_update(struct ring_buffer_per_cpu
*cpu_buffer
,
745 struct buffer_page
*head
,
746 struct buffer_page
*prev
,
749 return rb_head_page_set(cpu_buffer
, head
, prev
,
750 old_flag
, RB_PAGE_UPDATE
);
753 static int rb_head_page_set_head(struct ring_buffer_per_cpu
*cpu_buffer
,
754 struct buffer_page
*head
,
755 struct buffer_page
*prev
,
758 return rb_head_page_set(cpu_buffer
, head
, prev
,
759 old_flag
, RB_PAGE_HEAD
);
762 static int rb_head_page_set_normal(struct ring_buffer_per_cpu
*cpu_buffer
,
763 struct buffer_page
*head
,
764 struct buffer_page
*prev
,
767 return rb_head_page_set(cpu_buffer
, head
, prev
,
768 old_flag
, RB_PAGE_NORMAL
);
771 static inline void rb_inc_page(struct ring_buffer_per_cpu
*cpu_buffer
,
772 struct buffer_page
**bpage
)
774 struct list_head
*p
= rb_list_head((*bpage
)->list
.next
);
776 *bpage
= list_entry(p
, struct buffer_page
, list
);
779 static struct buffer_page
*
780 rb_set_head_page(struct ring_buffer_per_cpu
*cpu_buffer
)
782 struct buffer_page
*head
;
783 struct buffer_page
*page
;
784 struct list_head
*list
;
787 if (RB_WARN_ON(cpu_buffer
, !cpu_buffer
->head_page
))
791 list
= cpu_buffer
->pages
;
792 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->prev
->next
) != list
))
795 page
= head
= cpu_buffer
->head_page
;
797 * It is possible that the writer moves the header behind
798 * where we started, and we miss in one loop.
799 * A second loop should grab the header, but we'll do
800 * three loops just because I'm paranoid.
802 for (i
= 0; i
< 3; i
++) {
804 if (rb_is_head_page(cpu_buffer
, page
, page
->list
.prev
)) {
805 cpu_buffer
->head_page
= page
;
808 rb_inc_page(cpu_buffer
, &page
);
809 } while (page
!= head
);
812 RB_WARN_ON(cpu_buffer
, 1);
817 static int rb_head_page_replace(struct buffer_page
*old
,
818 struct buffer_page
*new)
820 unsigned long *ptr
= (unsigned long *)&old
->list
.prev
->next
;
824 val
= *ptr
& ~RB_FLAG_MASK
;
827 ret
= cmpxchg(ptr
, val
, (unsigned long)&new->list
);
833 * rb_tail_page_update - move the tail page forward
835 * Returns 1 if moved tail page, 0 if someone else did.
837 static int rb_tail_page_update(struct ring_buffer_per_cpu
*cpu_buffer
,
838 struct buffer_page
*tail_page
,
839 struct buffer_page
*next_page
)
841 struct buffer_page
*old_tail
;
842 unsigned long old_entries
;
843 unsigned long old_write
;
847 * The tail page now needs to be moved forward.
849 * We need to reset the tail page, but without messing
850 * with possible erasing of data brought in by interrupts
851 * that have moved the tail page and are currently on it.
853 * We add a counter to the write field to denote this.
855 old_write
= local_add_return(RB_WRITE_INTCNT
, &next_page
->write
);
856 old_entries
= local_add_return(RB_WRITE_INTCNT
, &next_page
->entries
);
859 * Just make sure we have seen our old_write and synchronize
860 * with any interrupts that come in.
865 * If the tail page is still the same as what we think
866 * it is, then it is up to us to update the tail
869 if (tail_page
== cpu_buffer
->tail_page
) {
870 /* Zero the write counter */
871 unsigned long val
= old_write
& ~RB_WRITE_MASK
;
872 unsigned long eval
= old_entries
& ~RB_WRITE_MASK
;
875 * This will only succeed if an interrupt did
876 * not come in and change it. In which case, we
877 * do not want to modify it.
879 * We add (void) to let the compiler know that we do not care
880 * about the return value of these functions. We use the
881 * cmpxchg to only update if an interrupt did not already
882 * do it for us. If the cmpxchg fails, we don't care.
884 (void)local_cmpxchg(&next_page
->write
, old_write
, val
);
885 (void)local_cmpxchg(&next_page
->entries
, old_entries
, eval
);
888 * No need to worry about races with clearing out the commit.
889 * it only can increment when a commit takes place. But that
890 * only happens in the outer most nested commit.
892 local_set(&next_page
->page
->commit
, 0);
894 old_tail
= cmpxchg(&cpu_buffer
->tail_page
,
895 tail_page
, next_page
);
897 if (old_tail
== tail_page
)
904 static int rb_check_bpage(struct ring_buffer_per_cpu
*cpu_buffer
,
905 struct buffer_page
*bpage
)
907 unsigned long val
= (unsigned long)bpage
;
909 if (RB_WARN_ON(cpu_buffer
, val
& RB_FLAG_MASK
))
916 * rb_check_list - make sure a pointer to a list has the last bits zero
918 static int rb_check_list(struct ring_buffer_per_cpu
*cpu_buffer
,
919 struct list_head
*list
)
921 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->prev
) != list
->prev
))
923 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->next
) != list
->next
))
929 * check_pages - integrity check of buffer pages
930 * @cpu_buffer: CPU buffer with pages to test
932 * As a safety measure we check to make sure the data pages have not
935 static int rb_check_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
937 struct list_head
*head
= cpu_buffer
->pages
;
938 struct buffer_page
*bpage
, *tmp
;
940 rb_head_page_deactivate(cpu_buffer
);
942 if (RB_WARN_ON(cpu_buffer
, head
->next
->prev
!= head
))
944 if (RB_WARN_ON(cpu_buffer
, head
->prev
->next
!= head
))
947 if (rb_check_list(cpu_buffer
, head
))
950 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
951 if (RB_WARN_ON(cpu_buffer
,
952 bpage
->list
.next
->prev
!= &bpage
->list
))
954 if (RB_WARN_ON(cpu_buffer
,
955 bpage
->list
.prev
->next
!= &bpage
->list
))
957 if (rb_check_list(cpu_buffer
, &bpage
->list
))
961 rb_head_page_activate(cpu_buffer
);
966 static int rb_allocate_pages(struct ring_buffer_per_cpu
*cpu_buffer
,
969 struct buffer_page
*bpage
, *tmp
;
975 for (i
= 0; i
< nr_pages
; i
++) {
978 * __GFP_NORETRY flag makes sure that the allocation fails
979 * gracefully without invoking oom-killer and the system is
982 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
983 GFP_KERNEL
| __GFP_NORETRY
,
984 cpu_to_node(cpu_buffer
->cpu
));
988 rb_check_bpage(cpu_buffer
, bpage
);
990 list_add(&bpage
->list
, &pages
);
992 page
= alloc_pages_node(cpu_to_node(cpu_buffer
->cpu
),
993 GFP_KERNEL
| __GFP_NORETRY
, 0);
996 bpage
->page
= page_address(page
);
997 rb_init_page(bpage
->page
);
1001 * The ring buffer page list is a circular list that does not
1002 * start and end with a list head. All page list items point to
1005 cpu_buffer
->pages
= pages
.next
;
1008 rb_check_pages(cpu_buffer
);
1013 list_for_each_entry_safe(bpage
, tmp
, &pages
, list
) {
1014 list_del_init(&bpage
->list
);
1015 free_buffer_page(bpage
);
1020 static struct ring_buffer_per_cpu
*
1021 rb_allocate_cpu_buffer(struct ring_buffer
*buffer
, int cpu
)
1023 struct ring_buffer_per_cpu
*cpu_buffer
;
1024 struct buffer_page
*bpage
;
1028 cpu_buffer
= kzalloc_node(ALIGN(sizeof(*cpu_buffer
), cache_line_size()),
1029 GFP_KERNEL
, cpu_to_node(cpu
));
1033 cpu_buffer
->cpu
= cpu
;
1034 cpu_buffer
->buffer
= buffer
;
1035 raw_spin_lock_init(&cpu_buffer
->reader_lock
);
1036 lockdep_set_class(&cpu_buffer
->reader_lock
, buffer
->reader_lock_key
);
1037 cpu_buffer
->lock
= (arch_spinlock_t
)__ARCH_SPIN_LOCK_UNLOCKED
;
1039 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
1040 GFP_KERNEL
, cpu_to_node(cpu
));
1042 goto fail_free_buffer
;
1044 rb_check_bpage(cpu_buffer
, bpage
);
1046 cpu_buffer
->reader_page
= bpage
;
1047 page
= alloc_pages_node(cpu_to_node(cpu
), GFP_KERNEL
, 0);
1049 goto fail_free_reader
;
1050 bpage
->page
= page_address(page
);
1051 rb_init_page(bpage
->page
);
1053 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
1055 ret
= rb_allocate_pages(cpu_buffer
, buffer
->pages
);
1057 goto fail_free_reader
;
1059 cpu_buffer
->head_page
1060 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
1061 cpu_buffer
->tail_page
= cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
1063 rb_head_page_activate(cpu_buffer
);
1068 free_buffer_page(cpu_buffer
->reader_page
);
1075 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu
*cpu_buffer
)
1077 struct list_head
*head
= cpu_buffer
->pages
;
1078 struct buffer_page
*bpage
, *tmp
;
1080 free_buffer_page(cpu_buffer
->reader_page
);
1082 rb_head_page_deactivate(cpu_buffer
);
1085 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
1086 list_del_init(&bpage
->list
);
1087 free_buffer_page(bpage
);
1089 bpage
= list_entry(head
, struct buffer_page
, list
);
1090 free_buffer_page(bpage
);
1096 #ifdef CONFIG_HOTPLUG_CPU
1097 static int rb_cpu_notify(struct notifier_block
*self
,
1098 unsigned long action
, void *hcpu
);
1102 * ring_buffer_alloc - allocate a new ring_buffer
1103 * @size: the size in bytes per cpu that is needed.
1104 * @flags: attributes to set for the ring buffer.
1106 * Currently the only flag that is available is the RB_FL_OVERWRITE
1107 * flag. This flag means that the buffer will overwrite old data
1108 * when the buffer wraps. If this flag is not set, the buffer will
1109 * drop data when the tail hits the head.
1111 struct ring_buffer
*__ring_buffer_alloc(unsigned long size
, unsigned flags
,
1112 struct lock_class_key
*key
)
1114 struct ring_buffer
*buffer
;
1118 /* keep it in its own cache line */
1119 buffer
= kzalloc(ALIGN(sizeof(*buffer
), cache_line_size()),
1124 if (!alloc_cpumask_var(&buffer
->cpumask
, GFP_KERNEL
))
1125 goto fail_free_buffer
;
1127 buffer
->pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1128 buffer
->flags
= flags
;
1129 buffer
->clock
= trace_clock_local
;
1130 buffer
->reader_lock_key
= key
;
1132 /* need at least two pages */
1133 if (buffer
->pages
< 2)
1137 * In case of non-hotplug cpu, if the ring-buffer is allocated
1138 * in early initcall, it will not be notified of secondary cpus.
1139 * In that off case, we need to allocate for all possible cpus.
1141 #ifdef CONFIG_HOTPLUG_CPU
1143 cpumask_copy(buffer
->cpumask
, cpu_online_mask
);
1145 cpumask_copy(buffer
->cpumask
, cpu_possible_mask
);
1147 buffer
->cpus
= nr_cpu_ids
;
1149 bsize
= sizeof(void *) * nr_cpu_ids
;
1150 buffer
->buffers
= kzalloc(ALIGN(bsize
, cache_line_size()),
1152 if (!buffer
->buffers
)
1153 goto fail_free_cpumask
;
1155 for_each_buffer_cpu(buffer
, cpu
) {
1156 buffer
->buffers
[cpu
] =
1157 rb_allocate_cpu_buffer(buffer
, cpu
);
1158 if (!buffer
->buffers
[cpu
])
1159 goto fail_free_buffers
;
1162 #ifdef CONFIG_HOTPLUG_CPU
1163 buffer
->cpu_notify
.notifier_call
= rb_cpu_notify
;
1164 buffer
->cpu_notify
.priority
= 0;
1165 register_cpu_notifier(&buffer
->cpu_notify
);
1169 mutex_init(&buffer
->mutex
);
1174 for_each_buffer_cpu(buffer
, cpu
) {
1175 if (buffer
->buffers
[cpu
])
1176 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1178 kfree(buffer
->buffers
);
1181 free_cpumask_var(buffer
->cpumask
);
1188 EXPORT_SYMBOL_GPL(__ring_buffer_alloc
);
1191 * ring_buffer_free - free a ring buffer.
1192 * @buffer: the buffer to free.
1195 ring_buffer_free(struct ring_buffer
*buffer
)
1201 #ifdef CONFIG_HOTPLUG_CPU
1202 unregister_cpu_notifier(&buffer
->cpu_notify
);
1205 for_each_buffer_cpu(buffer
, cpu
)
1206 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1210 kfree(buffer
->buffers
);
1211 free_cpumask_var(buffer
->cpumask
);
1215 EXPORT_SYMBOL_GPL(ring_buffer_free
);
1217 void ring_buffer_set_clock(struct ring_buffer
*buffer
,
1220 buffer
->clock
= clock
;
1223 static void rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
);
1226 rb_remove_pages(struct ring_buffer_per_cpu
*cpu_buffer
, unsigned nr_pages
)
1228 struct buffer_page
*bpage
;
1229 struct list_head
*p
;
1232 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1233 rb_head_page_deactivate(cpu_buffer
);
1235 for (i
= 0; i
< nr_pages
; i
++) {
1236 if (RB_WARN_ON(cpu_buffer
, list_empty(cpu_buffer
->pages
)))
1238 p
= cpu_buffer
->pages
->next
;
1239 bpage
= list_entry(p
, struct buffer_page
, list
);
1240 list_del_init(&bpage
->list
);
1241 free_buffer_page(bpage
);
1243 if (RB_WARN_ON(cpu_buffer
, list_empty(cpu_buffer
->pages
)))
1246 rb_reset_cpu(cpu_buffer
);
1247 rb_check_pages(cpu_buffer
);
1250 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1254 rb_insert_pages(struct ring_buffer_per_cpu
*cpu_buffer
,
1255 struct list_head
*pages
, unsigned nr_pages
)
1257 struct buffer_page
*bpage
;
1258 struct list_head
*p
;
1261 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1262 rb_head_page_deactivate(cpu_buffer
);
1264 for (i
= 0; i
< nr_pages
; i
++) {
1265 if (RB_WARN_ON(cpu_buffer
, list_empty(pages
)))
1268 bpage
= list_entry(p
, struct buffer_page
, list
);
1269 list_del_init(&bpage
->list
);
1270 list_add_tail(&bpage
->list
, cpu_buffer
->pages
);
1272 rb_reset_cpu(cpu_buffer
);
1273 rb_check_pages(cpu_buffer
);
1276 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1280 * ring_buffer_resize - resize the ring buffer
1281 * @buffer: the buffer to resize.
1282 * @size: the new size.
1284 * Minimum size is 2 * BUF_PAGE_SIZE.
1286 * Returns -1 on failure.
1288 int ring_buffer_resize(struct ring_buffer
*buffer
, unsigned long size
)
1290 struct ring_buffer_per_cpu
*cpu_buffer
;
1291 unsigned nr_pages
, rm_pages
, new_pages
;
1292 struct buffer_page
*bpage
, *tmp
;
1293 unsigned long buffer_size
;
1298 * Always succeed at resizing a non-existent buffer:
1303 size
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1304 size
*= BUF_PAGE_SIZE
;
1305 buffer_size
= buffer
->pages
* BUF_PAGE_SIZE
;
1307 /* we need a minimum of two pages */
1308 if (size
< BUF_PAGE_SIZE
* 2)
1309 size
= BUF_PAGE_SIZE
* 2;
1311 if (size
== buffer_size
)
1314 atomic_inc(&buffer
->record_disabled
);
1316 /* Make sure all writers are done with this buffer. */
1317 synchronize_sched();
1319 mutex_lock(&buffer
->mutex
);
1322 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1324 if (size
< buffer_size
) {
1326 /* easy case, just free pages */
1327 if (RB_WARN_ON(buffer
, nr_pages
>= buffer
->pages
))
1330 rm_pages
= buffer
->pages
- nr_pages
;
1332 for_each_buffer_cpu(buffer
, cpu
) {
1333 cpu_buffer
= buffer
->buffers
[cpu
];
1334 rb_remove_pages(cpu_buffer
, rm_pages
);
1340 * This is a bit more difficult. We only want to add pages
1341 * when we can allocate enough for all CPUs. We do this
1342 * by allocating all the pages and storing them on a local
1343 * link list. If we succeed in our allocation, then we
1344 * add these pages to the cpu_buffers. Otherwise we just free
1345 * them all and return -ENOMEM;
1347 if (RB_WARN_ON(buffer
, nr_pages
<= buffer
->pages
))
1350 new_pages
= nr_pages
- buffer
->pages
;
1352 for_each_buffer_cpu(buffer
, cpu
) {
1353 for (i
= 0; i
< new_pages
; i
++) {
1356 * __GFP_NORETRY flag makes sure that the allocation
1357 * fails gracefully without invoking oom-killer and
1358 * the system is not destabilized.
1360 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
),
1362 GFP_KERNEL
| __GFP_NORETRY
,
1366 list_add(&bpage
->list
, &pages
);
1367 page
= alloc_pages_node(cpu_to_node(cpu
),
1368 GFP_KERNEL
| __GFP_NORETRY
, 0);
1371 bpage
->page
= page_address(page
);
1372 rb_init_page(bpage
->page
);
1376 for_each_buffer_cpu(buffer
, cpu
) {
1377 cpu_buffer
= buffer
->buffers
[cpu
];
1378 rb_insert_pages(cpu_buffer
, &pages
, new_pages
);
1381 if (RB_WARN_ON(buffer
, !list_empty(&pages
)))
1385 buffer
->pages
= nr_pages
;
1387 mutex_unlock(&buffer
->mutex
);
1389 atomic_dec(&buffer
->record_disabled
);
1394 list_for_each_entry_safe(bpage
, tmp
, &pages
, list
) {
1395 list_del_init(&bpage
->list
);
1396 free_buffer_page(bpage
);
1399 mutex_unlock(&buffer
->mutex
);
1400 atomic_dec(&buffer
->record_disabled
);
1404 * Something went totally wrong, and we are too paranoid
1405 * to even clean up the mess.
1409 mutex_unlock(&buffer
->mutex
);
1410 atomic_dec(&buffer
->record_disabled
);
1413 EXPORT_SYMBOL_GPL(ring_buffer_resize
);
1415 void ring_buffer_change_overwrite(struct ring_buffer
*buffer
, int val
)
1417 mutex_lock(&buffer
->mutex
);
1419 buffer
->flags
|= RB_FL_OVERWRITE
;
1421 buffer
->flags
&= ~RB_FL_OVERWRITE
;
1422 mutex_unlock(&buffer
->mutex
);
1424 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite
);
1426 static inline void *
1427 __rb_data_page_index(struct buffer_data_page
*bpage
, unsigned index
)
1429 return bpage
->data
+ index
;
1432 static inline void *__rb_page_index(struct buffer_page
*bpage
, unsigned index
)
1434 return bpage
->page
->data
+ index
;
1437 static inline struct ring_buffer_event
*
1438 rb_reader_event(struct ring_buffer_per_cpu
*cpu_buffer
)
1440 return __rb_page_index(cpu_buffer
->reader_page
,
1441 cpu_buffer
->reader_page
->read
);
1444 static inline struct ring_buffer_event
*
1445 rb_iter_head_event(struct ring_buffer_iter
*iter
)
1447 return __rb_page_index(iter
->head_page
, iter
->head
);
1450 static inline unsigned long rb_page_write(struct buffer_page
*bpage
)
1452 return local_read(&bpage
->write
) & RB_WRITE_MASK
;
1455 static inline unsigned rb_page_commit(struct buffer_page
*bpage
)
1457 return local_read(&bpage
->page
->commit
);
1460 static inline unsigned long rb_page_entries(struct buffer_page
*bpage
)
1462 return local_read(&bpage
->entries
) & RB_WRITE_MASK
;
1465 /* Size is determined by what has been committed */
1466 static inline unsigned rb_page_size(struct buffer_page
*bpage
)
1468 return rb_page_commit(bpage
);
1471 static inline unsigned
1472 rb_commit_index(struct ring_buffer_per_cpu
*cpu_buffer
)
1474 return rb_page_commit(cpu_buffer
->commit_page
);
1477 static inline unsigned
1478 rb_event_index(struct ring_buffer_event
*event
)
1480 unsigned long addr
= (unsigned long)event
;
1482 return (addr
& ~PAGE_MASK
) - BUF_PAGE_HDR_SIZE
;
1486 rb_event_is_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
1487 struct ring_buffer_event
*event
)
1489 unsigned long addr
= (unsigned long)event
;
1490 unsigned long index
;
1492 index
= rb_event_index(event
);
1495 return cpu_buffer
->commit_page
->page
== (void *)addr
&&
1496 rb_commit_index(cpu_buffer
) == index
;
1500 rb_set_commit_to_write(struct ring_buffer_per_cpu
*cpu_buffer
)
1502 unsigned long max_count
;
1505 * We only race with interrupts and NMIs on this CPU.
1506 * If we own the commit event, then we can commit
1507 * all others that interrupted us, since the interruptions
1508 * are in stack format (they finish before they come
1509 * back to us). This allows us to do a simple loop to
1510 * assign the commit to the tail.
1513 max_count
= cpu_buffer
->buffer
->pages
* 100;
1515 while (cpu_buffer
->commit_page
!= cpu_buffer
->tail_page
) {
1516 if (RB_WARN_ON(cpu_buffer
, !(--max_count
)))
1518 if (RB_WARN_ON(cpu_buffer
,
1519 rb_is_reader_page(cpu_buffer
->tail_page
)))
1521 local_set(&cpu_buffer
->commit_page
->page
->commit
,
1522 rb_page_write(cpu_buffer
->commit_page
));
1523 rb_inc_page(cpu_buffer
, &cpu_buffer
->commit_page
);
1524 cpu_buffer
->write_stamp
=
1525 cpu_buffer
->commit_page
->page
->time_stamp
;
1526 /* add barrier to keep gcc from optimizing too much */
1529 while (rb_commit_index(cpu_buffer
) !=
1530 rb_page_write(cpu_buffer
->commit_page
)) {
1532 local_set(&cpu_buffer
->commit_page
->page
->commit
,
1533 rb_page_write(cpu_buffer
->commit_page
));
1534 RB_WARN_ON(cpu_buffer
,
1535 local_read(&cpu_buffer
->commit_page
->page
->commit
) &
1540 /* again, keep gcc from optimizing */
1544 * If an interrupt came in just after the first while loop
1545 * and pushed the tail page forward, we will be left with
1546 * a dangling commit that will never go forward.
1548 if (unlikely(cpu_buffer
->commit_page
!= cpu_buffer
->tail_page
))
1552 static void rb_reset_reader_page(struct ring_buffer_per_cpu
*cpu_buffer
)
1554 cpu_buffer
->read_stamp
= cpu_buffer
->reader_page
->page
->time_stamp
;
1555 cpu_buffer
->reader_page
->read
= 0;
1558 static void rb_inc_iter(struct ring_buffer_iter
*iter
)
1560 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
1563 * The iterator could be on the reader page (it starts there).
1564 * But the head could have moved, since the reader was
1565 * found. Check for this case and assign the iterator
1566 * to the head page instead of next.
1568 if (iter
->head_page
== cpu_buffer
->reader_page
)
1569 iter
->head_page
= rb_set_head_page(cpu_buffer
);
1571 rb_inc_page(cpu_buffer
, &iter
->head_page
);
1573 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
1577 /* Slow path, do not inline */
1578 static noinline
struct ring_buffer_event
*
1579 rb_add_time_stamp(struct ring_buffer_event
*event
, u64 delta
)
1581 event
->type_len
= RINGBUF_TYPE_TIME_EXTEND
;
1583 /* Not the first event on the page? */
1584 if (rb_event_index(event
)) {
1585 event
->time_delta
= delta
& TS_MASK
;
1586 event
->array
[0] = delta
>> TS_SHIFT
;
1588 /* nope, just zero it */
1589 event
->time_delta
= 0;
1590 event
->array
[0] = 0;
1593 return skip_time_extend(event
);
1597 * ring_buffer_update_event - update event type and data
1598 * @event: the even to update
1599 * @type: the type of event
1600 * @length: the size of the event field in the ring buffer
1602 * Update the type and data fields of the event. The length
1603 * is the actual size that is written to the ring buffer,
1604 * and with this, we can determine what to place into the
1608 rb_update_event(struct ring_buffer_per_cpu
*cpu_buffer
,
1609 struct ring_buffer_event
*event
, unsigned length
,
1610 int add_timestamp
, u64 delta
)
1612 /* Only a commit updates the timestamp */
1613 if (unlikely(!rb_event_is_commit(cpu_buffer
, event
)))
1617 * If we need to add a timestamp, then we
1618 * add it to the start of the resevered space.
1620 if (unlikely(add_timestamp
)) {
1621 event
= rb_add_time_stamp(event
, delta
);
1622 length
-= RB_LEN_TIME_EXTEND
;
1626 event
->time_delta
= delta
;
1627 length
-= RB_EVNT_HDR_SIZE
;
1628 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
) {
1629 event
->type_len
= 0;
1630 event
->array
[0] = length
;
1632 event
->type_len
= DIV_ROUND_UP(length
, RB_ALIGNMENT
);
1636 * rb_handle_head_page - writer hit the head page
1638 * Returns: +1 to retry page
1643 rb_handle_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
1644 struct buffer_page
*tail_page
,
1645 struct buffer_page
*next_page
)
1647 struct buffer_page
*new_head
;
1652 entries
= rb_page_entries(next_page
);
1655 * The hard part is here. We need to move the head
1656 * forward, and protect against both readers on
1657 * other CPUs and writers coming in via interrupts.
1659 type
= rb_head_page_set_update(cpu_buffer
, next_page
, tail_page
,
1663 * type can be one of four:
1664 * NORMAL - an interrupt already moved it for us
1665 * HEAD - we are the first to get here.
1666 * UPDATE - we are the interrupt interrupting
1668 * MOVED - a reader on another CPU moved the next
1669 * pointer to its reader page. Give up
1676 * We changed the head to UPDATE, thus
1677 * it is our responsibility to update
1680 local_add(entries
, &cpu_buffer
->overrun
);
1681 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
1684 * The entries will be zeroed out when we move the
1688 /* still more to do */
1691 case RB_PAGE_UPDATE
:
1693 * This is an interrupt that interrupt the
1694 * previous update. Still more to do.
1697 case RB_PAGE_NORMAL
:
1699 * An interrupt came in before the update
1700 * and processed this for us.
1701 * Nothing left to do.
1706 * The reader is on another CPU and just did
1707 * a swap with our next_page.
1712 RB_WARN_ON(cpu_buffer
, 1); /* WTF??? */
1717 * Now that we are here, the old head pointer is
1718 * set to UPDATE. This will keep the reader from
1719 * swapping the head page with the reader page.
1720 * The reader (on another CPU) will spin till
1723 * We just need to protect against interrupts
1724 * doing the job. We will set the next pointer
1725 * to HEAD. After that, we set the old pointer
1726 * to NORMAL, but only if it was HEAD before.
1727 * otherwise we are an interrupt, and only
1728 * want the outer most commit to reset it.
1730 new_head
= next_page
;
1731 rb_inc_page(cpu_buffer
, &new_head
);
1733 ret
= rb_head_page_set_head(cpu_buffer
, new_head
, next_page
,
1737 * Valid returns are:
1738 * HEAD - an interrupt came in and already set it.
1739 * NORMAL - One of two things:
1740 * 1) We really set it.
1741 * 2) A bunch of interrupts came in and moved
1742 * the page forward again.
1746 case RB_PAGE_NORMAL
:
1750 RB_WARN_ON(cpu_buffer
, 1);
1755 * It is possible that an interrupt came in,
1756 * set the head up, then more interrupts came in
1757 * and moved it again. When we get back here,
1758 * the page would have been set to NORMAL but we
1759 * just set it back to HEAD.
1761 * How do you detect this? Well, if that happened
1762 * the tail page would have moved.
1764 if (ret
== RB_PAGE_NORMAL
) {
1766 * If the tail had moved passed next, then we need
1767 * to reset the pointer.
1769 if (cpu_buffer
->tail_page
!= tail_page
&&
1770 cpu_buffer
->tail_page
!= next_page
)
1771 rb_head_page_set_normal(cpu_buffer
, new_head
,
1777 * If this was the outer most commit (the one that
1778 * changed the original pointer from HEAD to UPDATE),
1779 * then it is up to us to reset it to NORMAL.
1781 if (type
== RB_PAGE_HEAD
) {
1782 ret
= rb_head_page_set_normal(cpu_buffer
, next_page
,
1785 if (RB_WARN_ON(cpu_buffer
,
1786 ret
!= RB_PAGE_UPDATE
))
1793 static unsigned rb_calculate_event_length(unsigned length
)
1795 struct ring_buffer_event event
; /* Used only for sizeof array */
1797 /* zero length can cause confusions */
1801 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
)
1802 length
+= sizeof(event
.array
[0]);
1804 length
+= RB_EVNT_HDR_SIZE
;
1805 length
= ALIGN(length
, RB_ARCH_ALIGNMENT
);
1811 rb_reset_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
1812 struct buffer_page
*tail_page
,
1813 unsigned long tail
, unsigned long length
)
1815 struct ring_buffer_event
*event
;
1818 * Only the event that crossed the page boundary
1819 * must fill the old tail_page with padding.
1821 if (tail
>= BUF_PAGE_SIZE
) {
1823 * If the page was filled, then we still need
1824 * to update the real_end. Reset it to zero
1825 * and the reader will ignore it.
1827 if (tail
== BUF_PAGE_SIZE
)
1828 tail_page
->real_end
= 0;
1830 local_sub(length
, &tail_page
->write
);
1834 event
= __rb_page_index(tail_page
, tail
);
1835 kmemcheck_annotate_bitfield(event
, bitfield
);
1837 /* account for padding bytes */
1838 local_add(BUF_PAGE_SIZE
- tail
, &cpu_buffer
->entries_bytes
);
1841 * Save the original length to the meta data.
1842 * This will be used by the reader to add lost event
1845 tail_page
->real_end
= tail
;
1848 * If this event is bigger than the minimum size, then
1849 * we need to be careful that we don't subtract the
1850 * write counter enough to allow another writer to slip
1852 * We put in a discarded commit instead, to make sure
1853 * that this space is not used again.
1855 * If we are less than the minimum size, we don't need to
1858 if (tail
> (BUF_PAGE_SIZE
- RB_EVNT_MIN_SIZE
)) {
1859 /* No room for any events */
1861 /* Mark the rest of the page with padding */
1862 rb_event_set_padding(event
);
1864 /* Set the write back to the previous setting */
1865 local_sub(length
, &tail_page
->write
);
1869 /* Put in a discarded event */
1870 event
->array
[0] = (BUF_PAGE_SIZE
- tail
) - RB_EVNT_HDR_SIZE
;
1871 event
->type_len
= RINGBUF_TYPE_PADDING
;
1872 /* time delta must be non zero */
1873 event
->time_delta
= 1;
1875 /* Set write to end of buffer */
1876 length
= (tail
+ length
) - BUF_PAGE_SIZE
;
1877 local_sub(length
, &tail_page
->write
);
1881 * This is the slow path, force gcc not to inline it.
1883 static noinline
struct ring_buffer_event
*
1884 rb_move_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
1885 unsigned long length
, unsigned long tail
,
1886 struct buffer_page
*tail_page
, u64 ts
)
1888 struct buffer_page
*commit_page
= cpu_buffer
->commit_page
;
1889 struct ring_buffer
*buffer
= cpu_buffer
->buffer
;
1890 struct buffer_page
*next_page
;
1893 next_page
= tail_page
;
1895 rb_inc_page(cpu_buffer
, &next_page
);
1898 * If for some reason, we had an interrupt storm that made
1899 * it all the way around the buffer, bail, and warn
1902 if (unlikely(next_page
== commit_page
)) {
1903 local_inc(&cpu_buffer
->commit_overrun
);
1908 * This is where the fun begins!
1910 * We are fighting against races between a reader that
1911 * could be on another CPU trying to swap its reader
1912 * page with the buffer head.
1914 * We are also fighting against interrupts coming in and
1915 * moving the head or tail on us as well.
1917 * If the next page is the head page then we have filled
1918 * the buffer, unless the commit page is still on the
1921 if (rb_is_head_page(cpu_buffer
, next_page
, &tail_page
->list
)) {
1924 * If the commit is not on the reader page, then
1925 * move the header page.
1927 if (!rb_is_reader_page(cpu_buffer
->commit_page
)) {
1929 * If we are not in overwrite mode,
1930 * this is easy, just stop here.
1932 if (!(buffer
->flags
& RB_FL_OVERWRITE
))
1935 ret
= rb_handle_head_page(cpu_buffer
,
1944 * We need to be careful here too. The
1945 * commit page could still be on the reader
1946 * page. We could have a small buffer, and
1947 * have filled up the buffer with events
1948 * from interrupts and such, and wrapped.
1950 * Note, if the tail page is also the on the
1951 * reader_page, we let it move out.
1953 if (unlikely((cpu_buffer
->commit_page
!=
1954 cpu_buffer
->tail_page
) &&
1955 (cpu_buffer
->commit_page
==
1956 cpu_buffer
->reader_page
))) {
1957 local_inc(&cpu_buffer
->commit_overrun
);
1963 ret
= rb_tail_page_update(cpu_buffer
, tail_page
, next_page
);
1966 * Nested commits always have zero deltas, so
1967 * just reread the time stamp
1969 ts
= rb_time_stamp(buffer
);
1970 next_page
->page
->time_stamp
= ts
;
1975 rb_reset_tail(cpu_buffer
, tail_page
, tail
, length
);
1977 /* fail and let the caller try again */
1978 return ERR_PTR(-EAGAIN
);
1982 rb_reset_tail(cpu_buffer
, tail_page
, tail
, length
);
1987 static struct ring_buffer_event
*
1988 __rb_reserve_next(struct ring_buffer_per_cpu
*cpu_buffer
,
1989 unsigned long length
, u64 ts
,
1990 u64 delta
, int add_timestamp
)
1992 struct buffer_page
*tail_page
;
1993 struct ring_buffer_event
*event
;
1994 unsigned long tail
, write
;
1997 * If the time delta since the last event is too big to
1998 * hold in the time field of the event, then we append a
1999 * TIME EXTEND event ahead of the data event.
2001 if (unlikely(add_timestamp
))
2002 length
+= RB_LEN_TIME_EXTEND
;
2004 tail_page
= cpu_buffer
->tail_page
;
2005 write
= local_add_return(length
, &tail_page
->write
);
2007 /* set write to only the index of the write */
2008 write
&= RB_WRITE_MASK
;
2009 tail
= write
- length
;
2011 /* See if we shot pass the end of this buffer page */
2012 if (unlikely(write
> BUF_PAGE_SIZE
))
2013 return rb_move_tail(cpu_buffer
, length
, tail
,
2016 /* We reserved something on the buffer */
2018 event
= __rb_page_index(tail_page
, tail
);
2019 kmemcheck_annotate_bitfield(event
, bitfield
);
2020 rb_update_event(cpu_buffer
, event
, length
, add_timestamp
, delta
);
2022 local_inc(&tail_page
->entries
);
2025 * If this is the first commit on the page, then update
2029 tail_page
->page
->time_stamp
= ts
;
2031 /* account for these added bytes */
2032 local_add(length
, &cpu_buffer
->entries_bytes
);
2038 rb_try_to_discard(struct ring_buffer_per_cpu
*cpu_buffer
,
2039 struct ring_buffer_event
*event
)
2041 unsigned long new_index
, old_index
;
2042 struct buffer_page
*bpage
;
2043 unsigned long index
;
2046 new_index
= rb_event_index(event
);
2047 old_index
= new_index
+ rb_event_ts_length(event
);
2048 addr
= (unsigned long)event
;
2051 bpage
= cpu_buffer
->tail_page
;
2053 if (bpage
->page
== (void *)addr
&& rb_page_write(bpage
) == old_index
) {
2054 unsigned long write_mask
=
2055 local_read(&bpage
->write
) & ~RB_WRITE_MASK
;
2056 unsigned long event_length
= rb_event_length(event
);
2058 * This is on the tail page. It is possible that
2059 * a write could come in and move the tail page
2060 * and write to the next page. That is fine
2061 * because we just shorten what is on this page.
2063 old_index
+= write_mask
;
2064 new_index
+= write_mask
;
2065 index
= local_cmpxchg(&bpage
->write
, old_index
, new_index
);
2066 if (index
== old_index
) {
2067 /* update counters */
2068 local_sub(event_length
, &cpu_buffer
->entries_bytes
);
2073 /* could not discard */
2077 static void rb_start_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2079 local_inc(&cpu_buffer
->committing
);
2080 local_inc(&cpu_buffer
->commits
);
2083 static inline void rb_end_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2085 unsigned long commits
;
2087 if (RB_WARN_ON(cpu_buffer
,
2088 !local_read(&cpu_buffer
->committing
)))
2092 commits
= local_read(&cpu_buffer
->commits
);
2093 /* synchronize with interrupts */
2095 if (local_read(&cpu_buffer
->committing
) == 1)
2096 rb_set_commit_to_write(cpu_buffer
);
2098 local_dec(&cpu_buffer
->committing
);
2100 /* synchronize with interrupts */
2104 * Need to account for interrupts coming in between the
2105 * updating of the commit page and the clearing of the
2106 * committing counter.
2108 if (unlikely(local_read(&cpu_buffer
->commits
) != commits
) &&
2109 !local_read(&cpu_buffer
->committing
)) {
2110 local_inc(&cpu_buffer
->committing
);
2115 static struct ring_buffer_event
*
2116 rb_reserve_next_event(struct ring_buffer
*buffer
,
2117 struct ring_buffer_per_cpu
*cpu_buffer
,
2118 unsigned long length
)
2120 struct ring_buffer_event
*event
;
2126 rb_start_commit(cpu_buffer
);
2128 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2130 * Due to the ability to swap a cpu buffer from a buffer
2131 * it is possible it was swapped before we committed.
2132 * (committing stops a swap). We check for it here and
2133 * if it happened, we have to fail the write.
2136 if (unlikely(ACCESS_ONCE(cpu_buffer
->buffer
) != buffer
)) {
2137 local_dec(&cpu_buffer
->committing
);
2138 local_dec(&cpu_buffer
->commits
);
2143 length
= rb_calculate_event_length(length
);
2149 * We allow for interrupts to reenter here and do a trace.
2150 * If one does, it will cause this original code to loop
2151 * back here. Even with heavy interrupts happening, this
2152 * should only happen a few times in a row. If this happens
2153 * 1000 times in a row, there must be either an interrupt
2154 * storm or we have something buggy.
2157 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 1000))
2160 ts
= rb_time_stamp(cpu_buffer
->buffer
);
2161 diff
= ts
- cpu_buffer
->write_stamp
;
2163 /* make sure this diff is calculated here */
2166 /* Did the write stamp get updated already? */
2167 if (likely(ts
>= cpu_buffer
->write_stamp
)) {
2169 if (unlikely(test_time_stamp(delta
))) {
2170 int local_clock_stable
= 1;
2171 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2172 local_clock_stable
= sched_clock_stable
;
2174 WARN_ONCE(delta
> (1ULL << 59),
2175 KERN_WARNING
"Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2176 (unsigned long long)delta
,
2177 (unsigned long long)ts
,
2178 (unsigned long long)cpu_buffer
->write_stamp
,
2179 local_clock_stable
? "" :
2180 "If you just came from a suspend/resume,\n"
2181 "please switch to the trace global clock:\n"
2182 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2187 event
= __rb_reserve_next(cpu_buffer
, length
, ts
,
2188 delta
, add_timestamp
);
2189 if (unlikely(PTR_ERR(event
) == -EAGAIN
))
2198 rb_end_commit(cpu_buffer
);
2202 #ifdef CONFIG_TRACING
2204 #define TRACE_RECURSIVE_DEPTH 16
2206 /* Keep this code out of the fast path cache */
2207 static noinline
void trace_recursive_fail(void)
2209 /* Disable all tracing before we do anything else */
2210 tracing_off_permanent();
2212 printk_once(KERN_WARNING
"Tracing recursion: depth[%ld]:"
2213 "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2214 trace_recursion_buffer(),
2215 hardirq_count() >> HARDIRQ_SHIFT
,
2216 softirq_count() >> SOFTIRQ_SHIFT
,
2222 static inline int trace_recursive_lock(void)
2224 trace_recursion_inc();
2226 if (likely(trace_recursion_buffer() < TRACE_RECURSIVE_DEPTH
))
2229 trace_recursive_fail();
2234 static inline void trace_recursive_unlock(void)
2236 WARN_ON_ONCE(!trace_recursion_buffer());
2238 trace_recursion_dec();
2243 #define trace_recursive_lock() (0)
2244 #define trace_recursive_unlock() do { } while (0)
2249 * ring_buffer_lock_reserve - reserve a part of the buffer
2250 * @buffer: the ring buffer to reserve from
2251 * @length: the length of the data to reserve (excluding event header)
2253 * Returns a reseverd event on the ring buffer to copy directly to.
2254 * The user of this interface will need to get the body to write into
2255 * and can use the ring_buffer_event_data() interface.
2257 * The length is the length of the data needed, not the event length
2258 * which also includes the event header.
2260 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2261 * If NULL is returned, then nothing has been allocated or locked.
2263 struct ring_buffer_event
*
2264 ring_buffer_lock_reserve(struct ring_buffer
*buffer
, unsigned long length
)
2266 struct ring_buffer_per_cpu
*cpu_buffer
;
2267 struct ring_buffer_event
*event
;
2270 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
2273 /* If we are tracing schedule, we don't want to recurse */
2274 preempt_disable_notrace();
2276 if (atomic_read(&buffer
->record_disabled
))
2279 if (trace_recursive_lock())
2282 cpu
= raw_smp_processor_id();
2284 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2287 cpu_buffer
= buffer
->buffers
[cpu
];
2289 if (atomic_read(&cpu_buffer
->record_disabled
))
2292 if (length
> BUF_MAX_DATA_SIZE
)
2295 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
2302 trace_recursive_unlock();
2305 preempt_enable_notrace();
2308 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve
);
2311 rb_update_write_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
2312 struct ring_buffer_event
*event
)
2317 * The event first in the commit queue updates the
2320 if (rb_event_is_commit(cpu_buffer
, event
)) {
2322 * A commit event that is first on a page
2323 * updates the write timestamp with the page stamp
2325 if (!rb_event_index(event
))
2326 cpu_buffer
->write_stamp
=
2327 cpu_buffer
->commit_page
->page
->time_stamp
;
2328 else if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
2329 delta
= event
->array
[0];
2331 delta
+= event
->time_delta
;
2332 cpu_buffer
->write_stamp
+= delta
;
2334 cpu_buffer
->write_stamp
+= event
->time_delta
;
2338 static void rb_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
2339 struct ring_buffer_event
*event
)
2341 local_inc(&cpu_buffer
->entries
);
2342 rb_update_write_stamp(cpu_buffer
, event
);
2343 rb_end_commit(cpu_buffer
);
2347 * ring_buffer_unlock_commit - commit a reserved
2348 * @buffer: The buffer to commit to
2349 * @event: The event pointer to commit.
2351 * This commits the data to the ring buffer, and releases any locks held.
2353 * Must be paired with ring_buffer_lock_reserve.
2355 int ring_buffer_unlock_commit(struct ring_buffer
*buffer
,
2356 struct ring_buffer_event
*event
)
2358 struct ring_buffer_per_cpu
*cpu_buffer
;
2359 int cpu
= raw_smp_processor_id();
2361 cpu_buffer
= buffer
->buffers
[cpu
];
2363 rb_commit(cpu_buffer
, event
);
2365 trace_recursive_unlock();
2367 preempt_enable_notrace();
2371 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit
);
2373 static inline void rb_event_discard(struct ring_buffer_event
*event
)
2375 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
2376 event
= skip_time_extend(event
);
2378 /* array[0] holds the actual length for the discarded event */
2379 event
->array
[0] = rb_event_data_length(event
) - RB_EVNT_HDR_SIZE
;
2380 event
->type_len
= RINGBUF_TYPE_PADDING
;
2381 /* time delta must be non zero */
2382 if (!event
->time_delta
)
2383 event
->time_delta
= 1;
2387 * Decrement the entries to the page that an event is on.
2388 * The event does not even need to exist, only the pointer
2389 * to the page it is on. This may only be called before the commit
2393 rb_decrement_entry(struct ring_buffer_per_cpu
*cpu_buffer
,
2394 struct ring_buffer_event
*event
)
2396 unsigned long addr
= (unsigned long)event
;
2397 struct buffer_page
*bpage
= cpu_buffer
->commit_page
;
2398 struct buffer_page
*start
;
2402 /* Do the likely case first */
2403 if (likely(bpage
->page
== (void *)addr
)) {
2404 local_dec(&bpage
->entries
);
2409 * Because the commit page may be on the reader page we
2410 * start with the next page and check the end loop there.
2412 rb_inc_page(cpu_buffer
, &bpage
);
2415 if (bpage
->page
== (void *)addr
) {
2416 local_dec(&bpage
->entries
);
2419 rb_inc_page(cpu_buffer
, &bpage
);
2420 } while (bpage
!= start
);
2422 /* commit not part of this buffer?? */
2423 RB_WARN_ON(cpu_buffer
, 1);
2427 * ring_buffer_commit_discard - discard an event that has not been committed
2428 * @buffer: the ring buffer
2429 * @event: non committed event to discard
2431 * Sometimes an event that is in the ring buffer needs to be ignored.
2432 * This function lets the user discard an event in the ring buffer
2433 * and then that event will not be read later.
2435 * This function only works if it is called before the the item has been
2436 * committed. It will try to free the event from the ring buffer
2437 * if another event has not been added behind it.
2439 * If another event has been added behind it, it will set the event
2440 * up as discarded, and perform the commit.
2442 * If this function is called, do not call ring_buffer_unlock_commit on
2445 void ring_buffer_discard_commit(struct ring_buffer
*buffer
,
2446 struct ring_buffer_event
*event
)
2448 struct ring_buffer_per_cpu
*cpu_buffer
;
2451 /* The event is discarded regardless */
2452 rb_event_discard(event
);
2454 cpu
= smp_processor_id();
2455 cpu_buffer
= buffer
->buffers
[cpu
];
2458 * This must only be called if the event has not been
2459 * committed yet. Thus we can assume that preemption
2460 * is still disabled.
2462 RB_WARN_ON(buffer
, !local_read(&cpu_buffer
->committing
));
2464 rb_decrement_entry(cpu_buffer
, event
);
2465 if (rb_try_to_discard(cpu_buffer
, event
))
2469 * The commit is still visible by the reader, so we
2470 * must still update the timestamp.
2472 rb_update_write_stamp(cpu_buffer
, event
);
2474 rb_end_commit(cpu_buffer
);
2476 trace_recursive_unlock();
2478 preempt_enable_notrace();
2481 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit
);
2484 * ring_buffer_write - write data to the buffer without reserving
2485 * @buffer: The ring buffer to write to.
2486 * @length: The length of the data being written (excluding the event header)
2487 * @data: The data to write to the buffer.
2489 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2490 * one function. If you already have the data to write to the buffer, it
2491 * may be easier to simply call this function.
2493 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2494 * and not the length of the event which would hold the header.
2496 int ring_buffer_write(struct ring_buffer
*buffer
,
2497 unsigned long length
,
2500 struct ring_buffer_per_cpu
*cpu_buffer
;
2501 struct ring_buffer_event
*event
;
2506 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
2509 preempt_disable_notrace();
2511 if (atomic_read(&buffer
->record_disabled
))
2514 cpu
= raw_smp_processor_id();
2516 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2519 cpu_buffer
= buffer
->buffers
[cpu
];
2521 if (atomic_read(&cpu_buffer
->record_disabled
))
2524 if (length
> BUF_MAX_DATA_SIZE
)
2527 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
2531 body
= rb_event_data(event
);
2533 memcpy(body
, data
, length
);
2535 rb_commit(cpu_buffer
, event
);
2539 preempt_enable_notrace();
2543 EXPORT_SYMBOL_GPL(ring_buffer_write
);
2545 static int rb_per_cpu_empty(struct ring_buffer_per_cpu
*cpu_buffer
)
2547 struct buffer_page
*reader
= cpu_buffer
->reader_page
;
2548 struct buffer_page
*head
= rb_set_head_page(cpu_buffer
);
2549 struct buffer_page
*commit
= cpu_buffer
->commit_page
;
2551 /* In case of error, head will be NULL */
2552 if (unlikely(!head
))
2555 return reader
->read
== rb_page_commit(reader
) &&
2556 (commit
== reader
||
2558 head
->read
== rb_page_commit(commit
)));
2562 * ring_buffer_record_disable - stop all writes into the buffer
2563 * @buffer: The ring buffer to stop writes to.
2565 * This prevents all writes to the buffer. Any attempt to write
2566 * to the buffer after this will fail and return NULL.
2568 * The caller should call synchronize_sched() after this.
2570 void ring_buffer_record_disable(struct ring_buffer
*buffer
)
2572 atomic_inc(&buffer
->record_disabled
);
2574 EXPORT_SYMBOL_GPL(ring_buffer_record_disable
);
2577 * ring_buffer_record_enable - enable writes to the buffer
2578 * @buffer: The ring buffer to enable writes
2580 * Note, multiple disables will need the same number of enables
2581 * to truly enable the writing (much like preempt_disable).
2583 void ring_buffer_record_enable(struct ring_buffer
*buffer
)
2585 atomic_dec(&buffer
->record_disabled
);
2587 EXPORT_SYMBOL_GPL(ring_buffer_record_enable
);
2590 * ring_buffer_record_off - stop all writes into the buffer
2591 * @buffer: The ring buffer to stop writes to.
2593 * This prevents all writes to the buffer. Any attempt to write
2594 * to the buffer after this will fail and return NULL.
2596 * This is different than ring_buffer_record_disable() as
2597 * it works like an on/off switch, where as the disable() verison
2598 * must be paired with a enable().
2600 void ring_buffer_record_off(struct ring_buffer
*buffer
)
2603 unsigned int new_rd
;
2606 rd
= atomic_read(&buffer
->record_disabled
);
2607 new_rd
= rd
| RB_BUFFER_OFF
;
2608 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
2610 EXPORT_SYMBOL_GPL(ring_buffer_record_off
);
2613 * ring_buffer_record_on - restart writes into the buffer
2614 * @buffer: The ring buffer to start writes to.
2616 * This enables all writes to the buffer that was disabled by
2617 * ring_buffer_record_off().
2619 * This is different than ring_buffer_record_enable() as
2620 * it works like an on/off switch, where as the enable() verison
2621 * must be paired with a disable().
2623 void ring_buffer_record_on(struct ring_buffer
*buffer
)
2626 unsigned int new_rd
;
2629 rd
= atomic_read(&buffer
->record_disabled
);
2630 new_rd
= rd
& ~RB_BUFFER_OFF
;
2631 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
2633 EXPORT_SYMBOL_GPL(ring_buffer_record_on
);
2636 * ring_buffer_record_is_on - return true if the ring buffer can write
2637 * @buffer: The ring buffer to see if write is enabled
2639 * Returns true if the ring buffer is in a state that it accepts writes.
2641 int ring_buffer_record_is_on(struct ring_buffer
*buffer
)
2643 return !atomic_read(&buffer
->record_disabled
);
2647 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2648 * @buffer: The ring buffer to stop writes to.
2649 * @cpu: The CPU buffer to stop
2651 * This prevents all writes to the buffer. Any attempt to write
2652 * to the buffer after this will fail and return NULL.
2654 * The caller should call synchronize_sched() after this.
2656 void ring_buffer_record_disable_cpu(struct ring_buffer
*buffer
, int cpu
)
2658 struct ring_buffer_per_cpu
*cpu_buffer
;
2660 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2663 cpu_buffer
= buffer
->buffers
[cpu
];
2664 atomic_inc(&cpu_buffer
->record_disabled
);
2666 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu
);
2669 * ring_buffer_record_enable_cpu - enable writes to the buffer
2670 * @buffer: The ring buffer to enable writes
2671 * @cpu: The CPU to enable.
2673 * Note, multiple disables will need the same number of enables
2674 * to truly enable the writing (much like preempt_disable).
2676 void ring_buffer_record_enable_cpu(struct ring_buffer
*buffer
, int cpu
)
2678 struct ring_buffer_per_cpu
*cpu_buffer
;
2680 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2683 cpu_buffer
= buffer
->buffers
[cpu
];
2684 atomic_dec(&cpu_buffer
->record_disabled
);
2686 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu
);
2689 * The total entries in the ring buffer is the running counter
2690 * of entries entered into the ring buffer, minus the sum of
2691 * the entries read from the ring buffer and the number of
2692 * entries that were overwritten.
2694 static inline unsigned long
2695 rb_num_of_entries(struct ring_buffer_per_cpu
*cpu_buffer
)
2697 return local_read(&cpu_buffer
->entries
) -
2698 (local_read(&cpu_buffer
->overrun
) + cpu_buffer
->read
);
2702 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
2703 * @buffer: The ring buffer
2704 * @cpu: The per CPU buffer to read from.
2706 unsigned long ring_buffer_oldest_event_ts(struct ring_buffer
*buffer
, int cpu
)
2708 unsigned long flags
;
2709 struct ring_buffer_per_cpu
*cpu_buffer
;
2710 struct buffer_page
*bpage
;
2713 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2716 cpu_buffer
= buffer
->buffers
[cpu
];
2717 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
2719 * if the tail is on reader_page, oldest time stamp is on the reader
2722 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
2723 bpage
= cpu_buffer
->reader_page
;
2725 bpage
= rb_set_head_page(cpu_buffer
);
2726 ret
= bpage
->page
->time_stamp
;
2727 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
2731 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts
);
2734 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
2735 * @buffer: The ring buffer
2736 * @cpu: The per CPU buffer to read from.
2738 unsigned long ring_buffer_bytes_cpu(struct ring_buffer
*buffer
, int cpu
)
2740 struct ring_buffer_per_cpu
*cpu_buffer
;
2743 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2746 cpu_buffer
= buffer
->buffers
[cpu
];
2747 ret
= local_read(&cpu_buffer
->entries_bytes
) - cpu_buffer
->read_bytes
;
2751 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu
);
2754 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2755 * @buffer: The ring buffer
2756 * @cpu: The per CPU buffer to get the entries from.
2758 unsigned long ring_buffer_entries_cpu(struct ring_buffer
*buffer
, int cpu
)
2760 struct ring_buffer_per_cpu
*cpu_buffer
;
2762 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2765 cpu_buffer
= buffer
->buffers
[cpu
];
2767 return rb_num_of_entries(cpu_buffer
);
2769 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu
);
2772 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
2773 * @buffer: The ring buffer
2774 * @cpu: The per CPU buffer to get the number of overruns from
2776 unsigned long ring_buffer_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
2778 struct ring_buffer_per_cpu
*cpu_buffer
;
2781 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2784 cpu_buffer
= buffer
->buffers
[cpu
];
2785 ret
= local_read(&cpu_buffer
->overrun
);
2789 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu
);
2792 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
2793 * @buffer: The ring buffer
2794 * @cpu: The per CPU buffer to get the number of overruns from
2797 ring_buffer_commit_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
2799 struct ring_buffer_per_cpu
*cpu_buffer
;
2802 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2805 cpu_buffer
= buffer
->buffers
[cpu
];
2806 ret
= local_read(&cpu_buffer
->commit_overrun
);
2810 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu
);
2813 * ring_buffer_entries - get the number of entries in a buffer
2814 * @buffer: The ring buffer
2816 * Returns the total number of entries in the ring buffer
2819 unsigned long ring_buffer_entries(struct ring_buffer
*buffer
)
2821 struct ring_buffer_per_cpu
*cpu_buffer
;
2822 unsigned long entries
= 0;
2825 /* if you care about this being correct, lock the buffer */
2826 for_each_buffer_cpu(buffer
, cpu
) {
2827 cpu_buffer
= buffer
->buffers
[cpu
];
2828 entries
+= rb_num_of_entries(cpu_buffer
);
2833 EXPORT_SYMBOL_GPL(ring_buffer_entries
);
2836 * ring_buffer_overruns - get the number of overruns in buffer
2837 * @buffer: The ring buffer
2839 * Returns the total number of overruns in the ring buffer
2842 unsigned long ring_buffer_overruns(struct ring_buffer
*buffer
)
2844 struct ring_buffer_per_cpu
*cpu_buffer
;
2845 unsigned long overruns
= 0;
2848 /* if you care about this being correct, lock the buffer */
2849 for_each_buffer_cpu(buffer
, cpu
) {
2850 cpu_buffer
= buffer
->buffers
[cpu
];
2851 overruns
+= local_read(&cpu_buffer
->overrun
);
2856 EXPORT_SYMBOL_GPL(ring_buffer_overruns
);
2858 static void rb_iter_reset(struct ring_buffer_iter
*iter
)
2860 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
2862 /* Iterator usage is expected to have record disabled */
2863 if (list_empty(&cpu_buffer
->reader_page
->list
)) {
2864 iter
->head_page
= rb_set_head_page(cpu_buffer
);
2865 if (unlikely(!iter
->head_page
))
2867 iter
->head
= iter
->head_page
->read
;
2869 iter
->head_page
= cpu_buffer
->reader_page
;
2870 iter
->head
= cpu_buffer
->reader_page
->read
;
2873 iter
->read_stamp
= cpu_buffer
->read_stamp
;
2875 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
2876 iter
->cache_reader_page
= cpu_buffer
->reader_page
;
2877 iter
->cache_read
= cpu_buffer
->read
;
2881 * ring_buffer_iter_reset - reset an iterator
2882 * @iter: The iterator to reset
2884 * Resets the iterator, so that it will start from the beginning
2887 void ring_buffer_iter_reset(struct ring_buffer_iter
*iter
)
2889 struct ring_buffer_per_cpu
*cpu_buffer
;
2890 unsigned long flags
;
2895 cpu_buffer
= iter
->cpu_buffer
;
2897 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
2898 rb_iter_reset(iter
);
2899 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
2901 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset
);
2904 * ring_buffer_iter_empty - check if an iterator has no more to read
2905 * @iter: The iterator to check
2907 int ring_buffer_iter_empty(struct ring_buffer_iter
*iter
)
2909 struct ring_buffer_per_cpu
*cpu_buffer
;
2911 cpu_buffer
= iter
->cpu_buffer
;
2913 return iter
->head_page
== cpu_buffer
->commit_page
&&
2914 iter
->head
== rb_commit_index(cpu_buffer
);
2916 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty
);
2919 rb_update_read_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
2920 struct ring_buffer_event
*event
)
2924 switch (event
->type_len
) {
2925 case RINGBUF_TYPE_PADDING
:
2928 case RINGBUF_TYPE_TIME_EXTEND
:
2929 delta
= event
->array
[0];
2931 delta
+= event
->time_delta
;
2932 cpu_buffer
->read_stamp
+= delta
;
2935 case RINGBUF_TYPE_TIME_STAMP
:
2936 /* FIXME: not implemented */
2939 case RINGBUF_TYPE_DATA
:
2940 cpu_buffer
->read_stamp
+= event
->time_delta
;
2950 rb_update_iter_read_stamp(struct ring_buffer_iter
*iter
,
2951 struct ring_buffer_event
*event
)
2955 switch (event
->type_len
) {
2956 case RINGBUF_TYPE_PADDING
:
2959 case RINGBUF_TYPE_TIME_EXTEND
:
2960 delta
= event
->array
[0];
2962 delta
+= event
->time_delta
;
2963 iter
->read_stamp
+= delta
;
2966 case RINGBUF_TYPE_TIME_STAMP
:
2967 /* FIXME: not implemented */
2970 case RINGBUF_TYPE_DATA
:
2971 iter
->read_stamp
+= event
->time_delta
;
2980 static struct buffer_page
*
2981 rb_get_reader_page(struct ring_buffer_per_cpu
*cpu_buffer
)
2983 struct buffer_page
*reader
= NULL
;
2984 unsigned long overwrite
;
2985 unsigned long flags
;
2989 local_irq_save(flags
);
2990 arch_spin_lock(&cpu_buffer
->lock
);
2994 * This should normally only loop twice. But because the
2995 * start of the reader inserts an empty page, it causes
2996 * a case where we will loop three times. There should be no
2997 * reason to loop four times (that I know of).
2999 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 3)) {
3004 reader
= cpu_buffer
->reader_page
;
3006 /* If there's more to read, return this page */
3007 if (cpu_buffer
->reader_page
->read
< rb_page_size(reader
))
3010 /* Never should we have an index greater than the size */
3011 if (RB_WARN_ON(cpu_buffer
,
3012 cpu_buffer
->reader_page
->read
> rb_page_size(reader
)))
3015 /* check if we caught up to the tail */
3017 if (cpu_buffer
->commit_page
== cpu_buffer
->reader_page
)
3021 * Reset the reader page to size zero.
3023 local_set(&cpu_buffer
->reader_page
->write
, 0);
3024 local_set(&cpu_buffer
->reader_page
->entries
, 0);
3025 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
3026 cpu_buffer
->reader_page
->real_end
= 0;
3030 * Splice the empty reader page into the list around the head.
3032 reader
= rb_set_head_page(cpu_buffer
);
3033 cpu_buffer
->reader_page
->list
.next
= rb_list_head(reader
->list
.next
);
3034 cpu_buffer
->reader_page
->list
.prev
= reader
->list
.prev
;
3037 * cpu_buffer->pages just needs to point to the buffer, it
3038 * has no specific buffer page to point to. Lets move it out
3039 * of our way so we don't accidentally swap it.
3041 cpu_buffer
->pages
= reader
->list
.prev
;
3043 /* The reader page will be pointing to the new head */
3044 rb_set_list_to_head(cpu_buffer
, &cpu_buffer
->reader_page
->list
);
3047 * We want to make sure we read the overruns after we set up our
3048 * pointers to the next object. The writer side does a
3049 * cmpxchg to cross pages which acts as the mb on the writer
3050 * side. Note, the reader will constantly fail the swap
3051 * while the writer is updating the pointers, so this
3052 * guarantees that the overwrite recorded here is the one we
3053 * want to compare with the last_overrun.
3056 overwrite
= local_read(&(cpu_buffer
->overrun
));
3059 * Here's the tricky part.
3061 * We need to move the pointer past the header page.
3062 * But we can only do that if a writer is not currently
3063 * moving it. The page before the header page has the
3064 * flag bit '1' set if it is pointing to the page we want.
3065 * but if the writer is in the process of moving it
3066 * than it will be '2' or already moved '0'.
3069 ret
= rb_head_page_replace(reader
, cpu_buffer
->reader_page
);
3072 * If we did not convert it, then we must try again.
3078 * Yeah! We succeeded in replacing the page.
3080 * Now make the new head point back to the reader page.
3082 rb_list_head(reader
->list
.next
)->prev
= &cpu_buffer
->reader_page
->list
;
3083 rb_inc_page(cpu_buffer
, &cpu_buffer
->head_page
);
3085 /* Finally update the reader page to the new head */
3086 cpu_buffer
->reader_page
= reader
;
3087 rb_reset_reader_page(cpu_buffer
);
3089 if (overwrite
!= cpu_buffer
->last_overrun
) {
3090 cpu_buffer
->lost_events
= overwrite
- cpu_buffer
->last_overrun
;
3091 cpu_buffer
->last_overrun
= overwrite
;
3097 arch_spin_unlock(&cpu_buffer
->lock
);
3098 local_irq_restore(flags
);
3103 static void rb_advance_reader(struct ring_buffer_per_cpu
*cpu_buffer
)
3105 struct ring_buffer_event
*event
;
3106 struct buffer_page
*reader
;
3109 reader
= rb_get_reader_page(cpu_buffer
);
3111 /* This function should not be called when buffer is empty */
3112 if (RB_WARN_ON(cpu_buffer
, !reader
))
3115 event
= rb_reader_event(cpu_buffer
);
3117 if (event
->type_len
<= RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
3120 rb_update_read_stamp(cpu_buffer
, event
);
3122 length
= rb_event_length(event
);
3123 cpu_buffer
->reader_page
->read
+= length
;
3126 static void rb_advance_iter(struct ring_buffer_iter
*iter
)
3128 struct ring_buffer_per_cpu
*cpu_buffer
;
3129 struct ring_buffer_event
*event
;
3132 cpu_buffer
= iter
->cpu_buffer
;
3135 * Check if we are at the end of the buffer.
3137 if (iter
->head
>= rb_page_size(iter
->head_page
)) {
3138 /* discarded commits can make the page empty */
3139 if (iter
->head_page
== cpu_buffer
->commit_page
)
3145 event
= rb_iter_head_event(iter
);
3147 length
= rb_event_length(event
);
3150 * This should not be called to advance the header if we are
3151 * at the tail of the buffer.
3153 if (RB_WARN_ON(cpu_buffer
,
3154 (iter
->head_page
== cpu_buffer
->commit_page
) &&
3155 (iter
->head
+ length
> rb_commit_index(cpu_buffer
))))
3158 rb_update_iter_read_stamp(iter
, event
);
3160 iter
->head
+= length
;
3162 /* check for end of page padding */
3163 if ((iter
->head
>= rb_page_size(iter
->head_page
)) &&
3164 (iter
->head_page
!= cpu_buffer
->commit_page
))
3165 rb_advance_iter(iter
);
3168 static int rb_lost_events(struct ring_buffer_per_cpu
*cpu_buffer
)
3170 return cpu_buffer
->lost_events
;
3173 static struct ring_buffer_event
*
3174 rb_buffer_peek(struct ring_buffer_per_cpu
*cpu_buffer
, u64
*ts
,
3175 unsigned long *lost_events
)
3177 struct ring_buffer_event
*event
;
3178 struct buffer_page
*reader
;
3183 * We repeat when a time extend is encountered.
3184 * Since the time extend is always attached to a data event,
3185 * we should never loop more than once.
3186 * (We never hit the following condition more than twice).
3188 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 2))
3191 reader
= rb_get_reader_page(cpu_buffer
);
3195 event
= rb_reader_event(cpu_buffer
);
3197 switch (event
->type_len
) {
3198 case RINGBUF_TYPE_PADDING
:
3199 if (rb_null_event(event
))
3200 RB_WARN_ON(cpu_buffer
, 1);
3202 * Because the writer could be discarding every
3203 * event it creates (which would probably be bad)
3204 * if we were to go back to "again" then we may never
3205 * catch up, and will trigger the warn on, or lock
3206 * the box. Return the padding, and we will release
3207 * the current locks, and try again.
3211 case RINGBUF_TYPE_TIME_EXTEND
:
3212 /* Internal data, OK to advance */
3213 rb_advance_reader(cpu_buffer
);
3216 case RINGBUF_TYPE_TIME_STAMP
:
3217 /* FIXME: not implemented */
3218 rb_advance_reader(cpu_buffer
);
3221 case RINGBUF_TYPE_DATA
:
3223 *ts
= cpu_buffer
->read_stamp
+ event
->time_delta
;
3224 ring_buffer_normalize_time_stamp(cpu_buffer
->buffer
,
3225 cpu_buffer
->cpu
, ts
);
3228 *lost_events
= rb_lost_events(cpu_buffer
);
3237 EXPORT_SYMBOL_GPL(ring_buffer_peek
);
3239 static struct ring_buffer_event
*
3240 rb_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3242 struct ring_buffer
*buffer
;
3243 struct ring_buffer_per_cpu
*cpu_buffer
;
3244 struct ring_buffer_event
*event
;
3247 cpu_buffer
= iter
->cpu_buffer
;
3248 buffer
= cpu_buffer
->buffer
;
3251 * Check if someone performed a consuming read to
3252 * the buffer. A consuming read invalidates the iterator
3253 * and we need to reset the iterator in this case.
3255 if (unlikely(iter
->cache_read
!= cpu_buffer
->read
||
3256 iter
->cache_reader_page
!= cpu_buffer
->reader_page
))
3257 rb_iter_reset(iter
);
3260 if (ring_buffer_iter_empty(iter
))
3264 * We repeat when a time extend is encountered.
3265 * Since the time extend is always attached to a data event,
3266 * we should never loop more than once.
3267 * (We never hit the following condition more than twice).
3269 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 2))
3272 if (rb_per_cpu_empty(cpu_buffer
))
3275 if (iter
->head
>= local_read(&iter
->head_page
->page
->commit
)) {
3280 event
= rb_iter_head_event(iter
);
3282 switch (event
->type_len
) {
3283 case RINGBUF_TYPE_PADDING
:
3284 if (rb_null_event(event
)) {
3288 rb_advance_iter(iter
);
3291 case RINGBUF_TYPE_TIME_EXTEND
:
3292 /* Internal data, OK to advance */
3293 rb_advance_iter(iter
);
3296 case RINGBUF_TYPE_TIME_STAMP
:
3297 /* FIXME: not implemented */
3298 rb_advance_iter(iter
);
3301 case RINGBUF_TYPE_DATA
:
3303 *ts
= iter
->read_stamp
+ event
->time_delta
;
3304 ring_buffer_normalize_time_stamp(buffer
,
3305 cpu_buffer
->cpu
, ts
);
3315 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek
);
3317 static inline int rb_ok_to_lock(void)
3320 * If an NMI die dumps out the content of the ring buffer
3321 * do not grab locks. We also permanently disable the ring
3322 * buffer too. A one time deal is all you get from reading
3323 * the ring buffer from an NMI.
3325 if (likely(!in_nmi()))
3328 tracing_off_permanent();
3333 * ring_buffer_peek - peek at the next event to be read
3334 * @buffer: The ring buffer to read
3335 * @cpu: The cpu to peak at
3336 * @ts: The timestamp counter of this event.
3337 * @lost_events: a variable to store if events were lost (may be NULL)
3339 * This will return the event that will be read next, but does
3340 * not consume the data.
3342 struct ring_buffer_event
*
3343 ring_buffer_peek(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3344 unsigned long *lost_events
)
3346 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
3347 struct ring_buffer_event
*event
;
3348 unsigned long flags
;
3351 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3354 dolock
= rb_ok_to_lock();
3356 local_irq_save(flags
);
3358 raw_spin_lock(&cpu_buffer
->reader_lock
);
3359 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3360 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3361 rb_advance_reader(cpu_buffer
);
3363 raw_spin_unlock(&cpu_buffer
->reader_lock
);
3364 local_irq_restore(flags
);
3366 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3373 * ring_buffer_iter_peek - peek at the next event to be read
3374 * @iter: The ring buffer iterator
3375 * @ts: The timestamp counter of this event.
3377 * This will return the event that will be read next, but does
3378 * not increment the iterator.
3380 struct ring_buffer_event
*
3381 ring_buffer_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3383 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3384 struct ring_buffer_event
*event
;
3385 unsigned long flags
;
3388 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3389 event
= rb_iter_peek(iter
, ts
);
3390 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3392 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3399 * ring_buffer_consume - return an event and consume it
3400 * @buffer: The ring buffer to get the next event from
3401 * @cpu: the cpu to read the buffer from
3402 * @ts: a variable to store the timestamp (may be NULL)
3403 * @lost_events: a variable to store if events were lost (may be NULL)
3405 * Returns the next event in the ring buffer, and that event is consumed.
3406 * Meaning, that sequential reads will keep returning a different event,
3407 * and eventually empty the ring buffer if the producer is slower.
3409 struct ring_buffer_event
*
3410 ring_buffer_consume(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3411 unsigned long *lost_events
)
3413 struct ring_buffer_per_cpu
*cpu_buffer
;
3414 struct ring_buffer_event
*event
= NULL
;
3415 unsigned long flags
;
3418 dolock
= rb_ok_to_lock();
3421 /* might be called in atomic */
3424 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3427 cpu_buffer
= buffer
->buffers
[cpu
];
3428 local_irq_save(flags
);
3430 raw_spin_lock(&cpu_buffer
->reader_lock
);
3432 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3434 cpu_buffer
->lost_events
= 0;
3435 rb_advance_reader(cpu_buffer
);
3439 raw_spin_unlock(&cpu_buffer
->reader_lock
);
3440 local_irq_restore(flags
);
3445 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3450 EXPORT_SYMBOL_GPL(ring_buffer_consume
);
3453 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3454 * @buffer: The ring buffer to read from
3455 * @cpu: The cpu buffer to iterate over
3457 * This performs the initial preparations necessary to iterate
3458 * through the buffer. Memory is allocated, buffer recording
3459 * is disabled, and the iterator pointer is returned to the caller.
3461 * Disabling buffer recordng prevents the reading from being
3462 * corrupted. This is not a consuming read, so a producer is not
3465 * After a sequence of ring_buffer_read_prepare calls, the user is
3466 * expected to make at least one call to ring_buffer_prepare_sync.
3467 * Afterwards, ring_buffer_read_start is invoked to get things going
3470 * This overall must be paired with ring_buffer_finish.
3472 struct ring_buffer_iter
*
3473 ring_buffer_read_prepare(struct ring_buffer
*buffer
, int cpu
)
3475 struct ring_buffer_per_cpu
*cpu_buffer
;
3476 struct ring_buffer_iter
*iter
;
3478 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3481 iter
= kmalloc(sizeof(*iter
), GFP_KERNEL
);
3485 cpu_buffer
= buffer
->buffers
[cpu
];
3487 iter
->cpu_buffer
= cpu_buffer
;
3489 atomic_inc(&cpu_buffer
->record_disabled
);
3493 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare
);
3496 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3498 * All previously invoked ring_buffer_read_prepare calls to prepare
3499 * iterators will be synchronized. Afterwards, read_buffer_read_start
3500 * calls on those iterators are allowed.
3503 ring_buffer_read_prepare_sync(void)
3505 synchronize_sched();
3507 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync
);
3510 * ring_buffer_read_start - start a non consuming read of the buffer
3511 * @iter: The iterator returned by ring_buffer_read_prepare
3513 * This finalizes the startup of an iteration through the buffer.
3514 * The iterator comes from a call to ring_buffer_read_prepare and
3515 * an intervening ring_buffer_read_prepare_sync must have been
3518 * Must be paired with ring_buffer_finish.
3521 ring_buffer_read_start(struct ring_buffer_iter
*iter
)
3523 struct ring_buffer_per_cpu
*cpu_buffer
;
3524 unsigned long flags
;
3529 cpu_buffer
= iter
->cpu_buffer
;
3531 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3532 arch_spin_lock(&cpu_buffer
->lock
);
3533 rb_iter_reset(iter
);
3534 arch_spin_unlock(&cpu_buffer
->lock
);
3535 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3537 EXPORT_SYMBOL_GPL(ring_buffer_read_start
);
3540 * ring_buffer_finish - finish reading the iterator of the buffer
3541 * @iter: The iterator retrieved by ring_buffer_start
3543 * This re-enables the recording to the buffer, and frees the
3547 ring_buffer_read_finish(struct ring_buffer_iter
*iter
)
3549 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3551 atomic_dec(&cpu_buffer
->record_disabled
);
3554 EXPORT_SYMBOL_GPL(ring_buffer_read_finish
);
3557 * ring_buffer_read - read the next item in the ring buffer by the iterator
3558 * @iter: The ring buffer iterator
3559 * @ts: The time stamp of the event read.
3561 * This reads the next event in the ring buffer and increments the iterator.
3563 struct ring_buffer_event
*
3564 ring_buffer_read(struct ring_buffer_iter
*iter
, u64
*ts
)
3566 struct ring_buffer_event
*event
;
3567 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3568 unsigned long flags
;
3570 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3572 event
= rb_iter_peek(iter
, ts
);
3576 if (event
->type_len
== RINGBUF_TYPE_PADDING
)
3579 rb_advance_iter(iter
);
3581 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3585 EXPORT_SYMBOL_GPL(ring_buffer_read
);
3588 * ring_buffer_size - return the size of the ring buffer (in bytes)
3589 * @buffer: The ring buffer.
3591 unsigned long ring_buffer_size(struct ring_buffer
*buffer
)
3593 return BUF_PAGE_SIZE
* buffer
->pages
;
3595 EXPORT_SYMBOL_GPL(ring_buffer_size
);
3598 rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
)
3600 rb_head_page_deactivate(cpu_buffer
);
3602 cpu_buffer
->head_page
3603 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
3604 local_set(&cpu_buffer
->head_page
->write
, 0);
3605 local_set(&cpu_buffer
->head_page
->entries
, 0);
3606 local_set(&cpu_buffer
->head_page
->page
->commit
, 0);
3608 cpu_buffer
->head_page
->read
= 0;
3610 cpu_buffer
->tail_page
= cpu_buffer
->head_page
;
3611 cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
3613 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
3614 local_set(&cpu_buffer
->reader_page
->write
, 0);
3615 local_set(&cpu_buffer
->reader_page
->entries
, 0);
3616 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
3617 cpu_buffer
->reader_page
->read
= 0;
3619 local_set(&cpu_buffer
->commit_overrun
, 0);
3620 local_set(&cpu_buffer
->entries_bytes
, 0);
3621 local_set(&cpu_buffer
->overrun
, 0);
3622 local_set(&cpu_buffer
->entries
, 0);
3623 local_set(&cpu_buffer
->committing
, 0);
3624 local_set(&cpu_buffer
->commits
, 0);
3625 cpu_buffer
->read
= 0;
3626 cpu_buffer
->read_bytes
= 0;
3628 cpu_buffer
->write_stamp
= 0;
3629 cpu_buffer
->read_stamp
= 0;
3631 cpu_buffer
->lost_events
= 0;
3632 cpu_buffer
->last_overrun
= 0;
3634 rb_head_page_activate(cpu_buffer
);
3638 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3639 * @buffer: The ring buffer to reset a per cpu buffer of
3640 * @cpu: The CPU buffer to be reset
3642 void ring_buffer_reset_cpu(struct ring_buffer
*buffer
, int cpu
)
3644 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
3645 unsigned long flags
;
3647 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3650 atomic_inc(&cpu_buffer
->record_disabled
);
3652 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3654 if (RB_WARN_ON(cpu_buffer
, local_read(&cpu_buffer
->committing
)))
3657 arch_spin_lock(&cpu_buffer
->lock
);
3659 rb_reset_cpu(cpu_buffer
);
3661 arch_spin_unlock(&cpu_buffer
->lock
);
3664 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3666 atomic_dec(&cpu_buffer
->record_disabled
);
3668 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu
);
3671 * ring_buffer_reset - reset a ring buffer
3672 * @buffer: The ring buffer to reset all cpu buffers
3674 void ring_buffer_reset(struct ring_buffer
*buffer
)
3678 for_each_buffer_cpu(buffer
, cpu
)
3679 ring_buffer_reset_cpu(buffer
, cpu
);
3681 EXPORT_SYMBOL_GPL(ring_buffer_reset
);
3684 * rind_buffer_empty - is the ring buffer empty?
3685 * @buffer: The ring buffer to test
3687 int ring_buffer_empty(struct ring_buffer
*buffer
)
3689 struct ring_buffer_per_cpu
*cpu_buffer
;
3690 unsigned long flags
;
3695 dolock
= rb_ok_to_lock();
3697 /* yes this is racy, but if you don't like the race, lock the buffer */
3698 for_each_buffer_cpu(buffer
, cpu
) {
3699 cpu_buffer
= buffer
->buffers
[cpu
];
3700 local_irq_save(flags
);
3702 raw_spin_lock(&cpu_buffer
->reader_lock
);
3703 ret
= rb_per_cpu_empty(cpu_buffer
);
3705 raw_spin_unlock(&cpu_buffer
->reader_lock
);
3706 local_irq_restore(flags
);
3714 EXPORT_SYMBOL_GPL(ring_buffer_empty
);
3717 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3718 * @buffer: The ring buffer
3719 * @cpu: The CPU buffer to test
3721 int ring_buffer_empty_cpu(struct ring_buffer
*buffer
, int cpu
)
3723 struct ring_buffer_per_cpu
*cpu_buffer
;
3724 unsigned long flags
;
3728 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3731 dolock
= rb_ok_to_lock();
3733 cpu_buffer
= buffer
->buffers
[cpu
];
3734 local_irq_save(flags
);
3736 raw_spin_lock(&cpu_buffer
->reader_lock
);
3737 ret
= rb_per_cpu_empty(cpu_buffer
);
3739 raw_spin_unlock(&cpu_buffer
->reader_lock
);
3740 local_irq_restore(flags
);
3744 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu
);
3746 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3748 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3749 * @buffer_a: One buffer to swap with
3750 * @buffer_b: The other buffer to swap with
3752 * This function is useful for tracers that want to take a "snapshot"
3753 * of a CPU buffer and has another back up buffer lying around.
3754 * it is expected that the tracer handles the cpu buffer not being
3755 * used at the moment.
3757 int ring_buffer_swap_cpu(struct ring_buffer
*buffer_a
,
3758 struct ring_buffer
*buffer_b
, int cpu
)
3760 struct ring_buffer_per_cpu
*cpu_buffer_a
;
3761 struct ring_buffer_per_cpu
*cpu_buffer_b
;
3764 if (!cpumask_test_cpu(cpu
, buffer_a
->cpumask
) ||
3765 !cpumask_test_cpu(cpu
, buffer_b
->cpumask
))
3768 /* At least make sure the two buffers are somewhat the same */
3769 if (buffer_a
->pages
!= buffer_b
->pages
)
3774 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
3777 if (atomic_read(&buffer_a
->record_disabled
))
3780 if (atomic_read(&buffer_b
->record_disabled
))
3783 cpu_buffer_a
= buffer_a
->buffers
[cpu
];
3784 cpu_buffer_b
= buffer_b
->buffers
[cpu
];
3786 if (atomic_read(&cpu_buffer_a
->record_disabled
))
3789 if (atomic_read(&cpu_buffer_b
->record_disabled
))
3793 * We can't do a synchronize_sched here because this
3794 * function can be called in atomic context.
3795 * Normally this will be called from the same CPU as cpu.
3796 * If not it's up to the caller to protect this.
3798 atomic_inc(&cpu_buffer_a
->record_disabled
);
3799 atomic_inc(&cpu_buffer_b
->record_disabled
);
3802 if (local_read(&cpu_buffer_a
->committing
))
3804 if (local_read(&cpu_buffer_b
->committing
))
3807 buffer_a
->buffers
[cpu
] = cpu_buffer_b
;
3808 buffer_b
->buffers
[cpu
] = cpu_buffer_a
;
3810 cpu_buffer_b
->buffer
= buffer_a
;
3811 cpu_buffer_a
->buffer
= buffer_b
;
3816 atomic_dec(&cpu_buffer_a
->record_disabled
);
3817 atomic_dec(&cpu_buffer_b
->record_disabled
);
3821 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu
);
3822 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
3825 * ring_buffer_alloc_read_page - allocate a page to read from buffer
3826 * @buffer: the buffer to allocate for.
3828 * This function is used in conjunction with ring_buffer_read_page.
3829 * When reading a full page from the ring buffer, these functions
3830 * can be used to speed up the process. The calling function should
3831 * allocate a few pages first with this function. Then when it
3832 * needs to get pages from the ring buffer, it passes the result
3833 * of this function into ring_buffer_read_page, which will swap
3834 * the page that was allocated, with the read page of the buffer.
3837 * The page allocated, or NULL on error.
3839 void *ring_buffer_alloc_read_page(struct ring_buffer
*buffer
, int cpu
)
3841 struct buffer_data_page
*bpage
;
3844 page
= alloc_pages_node(cpu_to_node(cpu
),
3845 GFP_KERNEL
| __GFP_NORETRY
, 0);
3849 bpage
= page_address(page
);
3851 rb_init_page(bpage
);
3855 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page
);
3858 * ring_buffer_free_read_page - free an allocated read page
3859 * @buffer: the buffer the page was allocate for
3860 * @data: the page to free
3862 * Free a page allocated from ring_buffer_alloc_read_page.
3864 void ring_buffer_free_read_page(struct ring_buffer
*buffer
, void *data
)
3866 free_page((unsigned long)data
);
3868 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page
);
3871 * ring_buffer_read_page - extract a page from the ring buffer
3872 * @buffer: buffer to extract from
3873 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
3874 * @len: amount to extract
3875 * @cpu: the cpu of the buffer to extract
3876 * @full: should the extraction only happen when the page is full.
3878 * This function will pull out a page from the ring buffer and consume it.
3879 * @data_page must be the address of the variable that was returned
3880 * from ring_buffer_alloc_read_page. This is because the page might be used
3881 * to swap with a page in the ring buffer.
3884 * rpage = ring_buffer_alloc_read_page(buffer);
3887 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
3889 * process_page(rpage, ret);
3891 * When @full is set, the function will not return true unless
3892 * the writer is off the reader page.
3894 * Note: it is up to the calling functions to handle sleeps and wakeups.
3895 * The ring buffer can be used anywhere in the kernel and can not
3896 * blindly call wake_up. The layer that uses the ring buffer must be
3897 * responsible for that.
3900 * >=0 if data has been transferred, returns the offset of consumed data.
3901 * <0 if no data has been transferred.
3903 int ring_buffer_read_page(struct ring_buffer
*buffer
,
3904 void **data_page
, size_t len
, int cpu
, int full
)
3906 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
3907 struct ring_buffer_event
*event
;
3908 struct buffer_data_page
*bpage
;
3909 struct buffer_page
*reader
;
3910 unsigned long missed_events
;
3911 unsigned long flags
;
3912 unsigned int commit
;
3917 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3921 * If len is not big enough to hold the page header, then
3922 * we can not copy anything.
3924 if (len
<= BUF_PAGE_HDR_SIZE
)
3927 len
-= BUF_PAGE_HDR_SIZE
;
3936 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3938 reader
= rb_get_reader_page(cpu_buffer
);
3942 event
= rb_reader_event(cpu_buffer
);
3944 read
= reader
->read
;
3945 commit
= rb_page_commit(reader
);
3947 /* Check if any events were dropped */
3948 missed_events
= cpu_buffer
->lost_events
;
3951 * If this page has been partially read or
3952 * if len is not big enough to read the rest of the page or
3953 * a writer is still on the page, then
3954 * we must copy the data from the page to the buffer.
3955 * Otherwise, we can simply swap the page with the one passed in.
3957 if (read
|| (len
< (commit
- read
)) ||
3958 cpu_buffer
->reader_page
== cpu_buffer
->commit_page
) {
3959 struct buffer_data_page
*rpage
= cpu_buffer
->reader_page
->page
;
3960 unsigned int rpos
= read
;
3961 unsigned int pos
= 0;
3967 if (len
> (commit
- read
))
3968 len
= (commit
- read
);
3970 /* Always keep the time extend and data together */
3971 size
= rb_event_ts_length(event
);
3976 /* save the current timestamp, since the user will need it */
3977 save_timestamp
= cpu_buffer
->read_stamp
;
3979 /* Need to copy one event at a time */
3981 /* We need the size of one event, because
3982 * rb_advance_reader only advances by one event,
3983 * whereas rb_event_ts_length may include the size of
3984 * one or two events.
3985 * We have already ensured there's enough space if this
3986 * is a time extend. */
3987 size
= rb_event_length(event
);
3988 memcpy(bpage
->data
+ pos
, rpage
->data
+ rpos
, size
);
3992 rb_advance_reader(cpu_buffer
);
3993 rpos
= reader
->read
;
3999 event
= rb_reader_event(cpu_buffer
);
4000 /* Always keep the time extend and data together */
4001 size
= rb_event_ts_length(event
);
4002 } while (len
>= size
);
4005 local_set(&bpage
->commit
, pos
);
4006 bpage
->time_stamp
= save_timestamp
;
4008 /* we copied everything to the beginning */
4011 /* update the entry counter */
4012 cpu_buffer
->read
+= rb_page_entries(reader
);
4013 cpu_buffer
->read_bytes
+= BUF_PAGE_SIZE
;
4015 /* swap the pages */
4016 rb_init_page(bpage
);
4017 bpage
= reader
->page
;
4018 reader
->page
= *data_page
;
4019 local_set(&reader
->write
, 0);
4020 local_set(&reader
->entries
, 0);
4025 * Use the real_end for the data size,
4026 * This gives us a chance to store the lost events
4029 if (reader
->real_end
)
4030 local_set(&bpage
->commit
, reader
->real_end
);
4034 cpu_buffer
->lost_events
= 0;
4036 commit
= local_read(&bpage
->commit
);
4038 * Set a flag in the commit field if we lost events
4040 if (missed_events
) {
4041 /* If there is room at the end of the page to save the
4042 * missed events, then record it there.
4044 if (BUF_PAGE_SIZE
- commit
>= sizeof(missed_events
)) {
4045 memcpy(&bpage
->data
[commit
], &missed_events
,
4046 sizeof(missed_events
));
4047 local_add(RB_MISSED_STORED
, &bpage
->commit
);
4048 commit
+= sizeof(missed_events
);
4050 local_add(RB_MISSED_EVENTS
, &bpage
->commit
);
4054 * This page may be off to user land. Zero it out here.
4056 if (commit
< BUF_PAGE_SIZE
)
4057 memset(&bpage
->data
[commit
], 0, BUF_PAGE_SIZE
- commit
);
4060 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4065 EXPORT_SYMBOL_GPL(ring_buffer_read_page
);
4067 #ifdef CONFIG_HOTPLUG_CPU
4068 static int rb_cpu_notify(struct notifier_block
*self
,
4069 unsigned long action
, void *hcpu
)
4071 struct ring_buffer
*buffer
=
4072 container_of(self
, struct ring_buffer
, cpu_notify
);
4073 long cpu
= (long)hcpu
;
4076 case CPU_UP_PREPARE
:
4077 case CPU_UP_PREPARE_FROZEN
:
4078 if (cpumask_test_cpu(cpu
, buffer
->cpumask
))
4081 buffer
->buffers
[cpu
] =
4082 rb_allocate_cpu_buffer(buffer
, cpu
);
4083 if (!buffer
->buffers
[cpu
]) {
4084 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4089 cpumask_set_cpu(cpu
, buffer
->cpumask
);
4091 case CPU_DOWN_PREPARE
:
4092 case CPU_DOWN_PREPARE_FROZEN
:
4095 * If we were to free the buffer, then the user would
4096 * lose any trace that was in the buffer.