4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/ftrace_event.h>
7 #include <linux/ring_buffer.h>
8 #include <linux/trace_clock.h>
9 #include <linux/trace_seq.h>
10 #include <linux/spinlock.h>
11 #include <linux/debugfs.h>
12 #include <linux/uaccess.h>
13 #include <linux/hardirq.h>
14 #include <linux/kmemcheck.h>
15 #include <linux/module.h>
16 #include <linux/percpu.h>
17 #include <linux/mutex.h>
18 #include <linux/slab.h>
19 #include <linux/init.h>
20 #include <linux/hash.h>
21 #include <linux/list.h>
22 #include <linux/cpu.h>
25 #include <asm/local.h>
27 static void update_pages_handler(struct work_struct
*work
);
30 * The ring buffer header is special. We must manually up keep it.
32 int ring_buffer_print_entry_header(struct trace_seq
*s
)
36 ret
= trace_seq_printf(s
, "# compressed entry header\n");
37 ret
= trace_seq_printf(s
, "\ttype_len : 5 bits\n");
38 ret
= trace_seq_printf(s
, "\ttime_delta : 27 bits\n");
39 ret
= trace_seq_printf(s
, "\tarray : 32 bits\n");
40 ret
= trace_seq_printf(s
, "\n");
41 ret
= trace_seq_printf(s
, "\tpadding : type == %d\n",
42 RINGBUF_TYPE_PADDING
);
43 ret
= trace_seq_printf(s
, "\ttime_extend : type == %d\n",
44 RINGBUF_TYPE_TIME_EXTEND
);
45 ret
= trace_seq_printf(s
, "\tdata max type_len == %d\n",
46 RINGBUF_TYPE_DATA_TYPE_LEN_MAX
);
52 * The ring buffer is made up of a list of pages. A separate list of pages is
53 * allocated for each CPU. A writer may only write to a buffer that is
54 * associated with the CPU it is currently executing on. A reader may read
55 * from any per cpu buffer.
57 * The reader is special. For each per cpu buffer, the reader has its own
58 * reader page. When a reader has read the entire reader page, this reader
59 * page is swapped with another page in the ring buffer.
61 * Now, as long as the writer is off the reader page, the reader can do what
62 * ever it wants with that page. The writer will never write to that page
63 * again (as long as it is out of the ring buffer).
65 * Here's some silly ASCII art.
68 * |reader| RING BUFFER
70 * +------+ +---+ +---+ +---+
79 * |reader| RING BUFFER
80 * |page |------------------v
81 * +------+ +---+ +---+ +---+
90 * |reader| RING BUFFER
91 * |page |------------------v
92 * +------+ +---+ +---+ +---+
97 * +------------------------------+
101 * |buffer| RING BUFFER
102 * |page |------------------v
103 * +------+ +---+ +---+ +---+
105 * | New +---+ +---+ +---+
108 * +------------------------------+
111 * After we make this swap, the reader can hand this page off to the splice
112 * code and be done with it. It can even allocate a new page if it needs to
113 * and swap that into the ring buffer.
115 * We will be using cmpxchg soon to make all this lockless.
120 * A fast way to enable or disable all ring buffers is to
121 * call tracing_on or tracing_off. Turning off the ring buffers
122 * prevents all ring buffers from being recorded to.
123 * Turning this switch on, makes it OK to write to the
124 * ring buffer, if the ring buffer is enabled itself.
126 * There's three layers that must be on in order to write
127 * to the ring buffer.
129 * 1) This global flag must be set.
130 * 2) The ring buffer must be enabled for recording.
131 * 3) The per cpu buffer must be enabled for recording.
133 * In case of an anomaly, this global flag has a bit set that
134 * will permantly disable all ring buffers.
138 * Global flag to disable all recording to ring buffers
139 * This has two bits: ON, DISABLED
143 * 0 0 : ring buffers are off
144 * 1 0 : ring buffers are on
145 * X 1 : ring buffers are permanently disabled
149 RB_BUFFERS_ON_BIT
= 0,
150 RB_BUFFERS_DISABLED_BIT
= 1,
154 RB_BUFFERS_ON
= 1 << RB_BUFFERS_ON_BIT
,
155 RB_BUFFERS_DISABLED
= 1 << RB_BUFFERS_DISABLED_BIT
,
158 static unsigned long ring_buffer_flags __read_mostly
= RB_BUFFERS_ON
;
160 /* Used for individual buffers (after the counter) */
161 #define RB_BUFFER_OFF (1 << 20)
163 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
166 * tracing_off_permanent - permanently disable ring buffers
168 * This function, once called, will disable all ring buffers
171 void tracing_off_permanent(void)
173 set_bit(RB_BUFFERS_DISABLED_BIT
, &ring_buffer_flags
);
176 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
177 #define RB_ALIGNMENT 4U
178 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
179 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
181 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
182 # define RB_FORCE_8BYTE_ALIGNMENT 0
183 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
185 # define RB_FORCE_8BYTE_ALIGNMENT 1
186 # define RB_ARCH_ALIGNMENT 8U
189 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
191 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
192 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
195 RB_LEN_TIME_EXTEND
= 8,
196 RB_LEN_TIME_STAMP
= 16,
199 #define skip_time_extend(event) \
200 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
202 static inline int rb_null_event(struct ring_buffer_event
*event
)
204 return event
->type_len
== RINGBUF_TYPE_PADDING
&& !event
->time_delta
;
207 static void rb_event_set_padding(struct ring_buffer_event
*event
)
209 /* padding has a NULL time_delta */
210 event
->type_len
= RINGBUF_TYPE_PADDING
;
211 event
->time_delta
= 0;
215 rb_event_data_length(struct ring_buffer_event
*event
)
220 length
= event
->type_len
* RB_ALIGNMENT
;
222 length
= event
->array
[0];
223 return length
+ RB_EVNT_HDR_SIZE
;
227 * Return the length of the given event. Will return
228 * the length of the time extend if the event is a
231 static inline unsigned
232 rb_event_length(struct ring_buffer_event
*event
)
234 switch (event
->type_len
) {
235 case RINGBUF_TYPE_PADDING
:
236 if (rb_null_event(event
))
239 return event
->array
[0] + RB_EVNT_HDR_SIZE
;
241 case RINGBUF_TYPE_TIME_EXTEND
:
242 return RB_LEN_TIME_EXTEND
;
244 case RINGBUF_TYPE_TIME_STAMP
:
245 return RB_LEN_TIME_STAMP
;
247 case RINGBUF_TYPE_DATA
:
248 return rb_event_data_length(event
);
257 * Return total length of time extend and data,
258 * or just the event length for all other events.
260 static inline unsigned
261 rb_event_ts_length(struct ring_buffer_event
*event
)
265 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
266 /* time extends include the data event after it */
267 len
= RB_LEN_TIME_EXTEND
;
268 event
= skip_time_extend(event
);
270 return len
+ rb_event_length(event
);
274 * ring_buffer_event_length - return the length of the event
275 * @event: the event to get the length of
277 * Returns the size of the data load of a data event.
278 * If the event is something other than a data event, it
279 * returns the size of the event itself. With the exception
280 * of a TIME EXTEND, where it still returns the size of the
281 * data load of the data event after it.
283 unsigned ring_buffer_event_length(struct ring_buffer_event
*event
)
287 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
288 event
= skip_time_extend(event
);
290 length
= rb_event_length(event
);
291 if (event
->type_len
> RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
293 length
-= RB_EVNT_HDR_SIZE
;
294 if (length
> RB_MAX_SMALL_DATA
+ sizeof(event
->array
[0]))
295 length
-= sizeof(event
->array
[0]);
298 EXPORT_SYMBOL_GPL(ring_buffer_event_length
);
300 /* inline for ring buffer fast paths */
302 rb_event_data(struct ring_buffer_event
*event
)
304 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
305 event
= skip_time_extend(event
);
306 BUG_ON(event
->type_len
> RINGBUF_TYPE_DATA_TYPE_LEN_MAX
);
307 /* If length is in len field, then array[0] has the data */
309 return (void *)&event
->array
[0];
310 /* Otherwise length is in array[0] and array[1] has the data */
311 return (void *)&event
->array
[1];
315 * ring_buffer_event_data - return the data of the event
316 * @event: the event to get the data from
318 void *ring_buffer_event_data(struct ring_buffer_event
*event
)
320 return rb_event_data(event
);
322 EXPORT_SYMBOL_GPL(ring_buffer_event_data
);
324 #define for_each_buffer_cpu(buffer, cpu) \
325 for_each_cpu(cpu, buffer->cpumask)
328 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
329 #define TS_DELTA_TEST (~TS_MASK)
331 /* Flag when events were overwritten */
332 #define RB_MISSED_EVENTS (1 << 31)
333 /* Missed count stored at end */
334 #define RB_MISSED_STORED (1 << 30)
336 struct buffer_data_page
{
337 u64 time_stamp
; /* page time stamp */
338 local_t commit
; /* write committed index */
339 unsigned char data
[] RB_ALIGN_DATA
; /* data of buffer page */
343 * Note, the buffer_page list must be first. The buffer pages
344 * are allocated in cache lines, which means that each buffer
345 * page will be at the beginning of a cache line, and thus
346 * the least significant bits will be zero. We use this to
347 * add flags in the list struct pointers, to make the ring buffer
351 struct list_head list
; /* list of buffer pages */
352 local_t write
; /* index for next write */
353 unsigned read
; /* index for next read */
354 local_t entries
; /* entries on this page */
355 unsigned long real_end
; /* real end of data */
356 struct buffer_data_page
*page
; /* Actual data page */
360 * The buffer page counters, write and entries, must be reset
361 * atomically when crossing page boundaries. To synchronize this
362 * update, two counters are inserted into the number. One is
363 * the actual counter for the write position or count on the page.
365 * The other is a counter of updaters. Before an update happens
366 * the update partition of the counter is incremented. This will
367 * allow the updater to update the counter atomically.
369 * The counter is 20 bits, and the state data is 12.
371 #define RB_WRITE_MASK 0xfffff
372 #define RB_WRITE_INTCNT (1 << 20)
374 static void rb_init_page(struct buffer_data_page
*bpage
)
376 local_set(&bpage
->commit
, 0);
380 * ring_buffer_page_len - the size of data on the page.
381 * @page: The page to read
383 * Returns the amount of data on the page, including buffer page header.
385 size_t ring_buffer_page_len(void *page
)
387 return local_read(&((struct buffer_data_page
*)page
)->commit
)
392 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
395 static void free_buffer_page(struct buffer_page
*bpage
)
397 free_page((unsigned long)bpage
->page
);
402 * We need to fit the time_stamp delta into 27 bits.
404 static inline int test_time_stamp(u64 delta
)
406 if (delta
& TS_DELTA_TEST
)
411 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
413 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
414 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
416 int ring_buffer_print_page_header(struct trace_seq
*s
)
418 struct buffer_data_page field
;
421 ret
= trace_seq_printf(s
, "\tfield: u64 timestamp;\t"
422 "offset:0;\tsize:%u;\tsigned:%u;\n",
423 (unsigned int)sizeof(field
.time_stamp
),
424 (unsigned int)is_signed_type(u64
));
426 ret
= trace_seq_printf(s
, "\tfield: local_t commit;\t"
427 "offset:%u;\tsize:%u;\tsigned:%u;\n",
428 (unsigned int)offsetof(typeof(field
), commit
),
429 (unsigned int)sizeof(field
.commit
),
430 (unsigned int)is_signed_type(long));
432 ret
= trace_seq_printf(s
, "\tfield: int overwrite;\t"
433 "offset:%u;\tsize:%u;\tsigned:%u;\n",
434 (unsigned int)offsetof(typeof(field
), commit
),
436 (unsigned int)is_signed_type(long));
438 ret
= trace_seq_printf(s
, "\tfield: char data;\t"
439 "offset:%u;\tsize:%u;\tsigned:%u;\n",
440 (unsigned int)offsetof(typeof(field
), data
),
441 (unsigned int)BUF_PAGE_SIZE
,
442 (unsigned int)is_signed_type(char));
448 * head_page == tail_page && head == tail then buffer is empty.
450 struct ring_buffer_per_cpu
{
452 atomic_t record_disabled
;
453 struct ring_buffer
*buffer
;
454 raw_spinlock_t reader_lock
; /* serialize readers */
455 arch_spinlock_t lock
;
456 struct lock_class_key lock_key
;
457 unsigned int nr_pages
;
458 struct list_head
*pages
;
459 struct buffer_page
*head_page
; /* read from head */
460 struct buffer_page
*tail_page
; /* write to tail */
461 struct buffer_page
*commit_page
; /* committed pages */
462 struct buffer_page
*reader_page
;
463 unsigned long lost_events
;
464 unsigned long last_overrun
;
465 local_t entries_bytes
;
468 local_t commit_overrun
;
469 local_t dropped_events
;
473 unsigned long read_bytes
;
476 /* ring buffer pages to update, > 0 to add, < 0 to remove */
477 int nr_pages_to_update
;
478 struct list_head new_pages
; /* new pages to add */
479 struct work_struct update_pages_work
;
480 struct completion update_done
;
486 atomic_t record_disabled
;
487 atomic_t resize_disabled
;
488 cpumask_var_t cpumask
;
490 struct lock_class_key
*reader_lock_key
;
494 struct ring_buffer_per_cpu
**buffers
;
496 #ifdef CONFIG_HOTPLUG_CPU
497 struct notifier_block cpu_notify
;
502 struct ring_buffer_iter
{
503 struct ring_buffer_per_cpu
*cpu_buffer
;
505 struct buffer_page
*head_page
;
506 struct buffer_page
*cache_reader_page
;
507 unsigned long cache_read
;
511 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
512 #define RB_WARN_ON(b, cond) \
514 int _____ret = unlikely(cond); \
516 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
517 struct ring_buffer_per_cpu *__b = \
519 atomic_inc(&__b->buffer->record_disabled); \
521 atomic_inc(&b->record_disabled); \
527 /* Up this if you want to test the TIME_EXTENTS and normalization */
528 #define DEBUG_SHIFT 0
530 static inline u64
rb_time_stamp(struct ring_buffer
*buffer
)
532 /* shift to debug/test normalization and TIME_EXTENTS */
533 return buffer
->clock() << DEBUG_SHIFT
;
536 u64
ring_buffer_time_stamp(struct ring_buffer
*buffer
, int cpu
)
540 preempt_disable_notrace();
541 time
= rb_time_stamp(buffer
);
542 preempt_enable_no_resched_notrace();
546 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp
);
548 void ring_buffer_normalize_time_stamp(struct ring_buffer
*buffer
,
551 /* Just stupid testing the normalize function and deltas */
554 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp
);
557 * Making the ring buffer lockless makes things tricky.
558 * Although writes only happen on the CPU that they are on,
559 * and they only need to worry about interrupts. Reads can
562 * The reader page is always off the ring buffer, but when the
563 * reader finishes with a page, it needs to swap its page with
564 * a new one from the buffer. The reader needs to take from
565 * the head (writes go to the tail). But if a writer is in overwrite
566 * mode and wraps, it must push the head page forward.
568 * Here lies the problem.
570 * The reader must be careful to replace only the head page, and
571 * not another one. As described at the top of the file in the
572 * ASCII art, the reader sets its old page to point to the next
573 * page after head. It then sets the page after head to point to
574 * the old reader page. But if the writer moves the head page
575 * during this operation, the reader could end up with the tail.
577 * We use cmpxchg to help prevent this race. We also do something
578 * special with the page before head. We set the LSB to 1.
580 * When the writer must push the page forward, it will clear the
581 * bit that points to the head page, move the head, and then set
582 * the bit that points to the new head page.
584 * We also don't want an interrupt coming in and moving the head
585 * page on another writer. Thus we use the second LSB to catch
588 * head->list->prev->next bit 1 bit 0
591 * Points to head page 0 1
594 * Note we can not trust the prev pointer of the head page, because:
596 * +----+ +-----+ +-----+
597 * | |------>| T |---X--->| N |
599 * +----+ +-----+ +-----+
602 * +----------| R |----------+ |
606 * Key: ---X--> HEAD flag set in pointer
611 * (see __rb_reserve_next() to see where this happens)
613 * What the above shows is that the reader just swapped out
614 * the reader page with a page in the buffer, but before it
615 * could make the new header point back to the new page added
616 * it was preempted by a writer. The writer moved forward onto
617 * the new page added by the reader and is about to move forward
620 * You can see, it is legitimate for the previous pointer of
621 * the head (or any page) not to point back to itself. But only
625 #define RB_PAGE_NORMAL 0UL
626 #define RB_PAGE_HEAD 1UL
627 #define RB_PAGE_UPDATE 2UL
630 #define RB_FLAG_MASK 3UL
632 /* PAGE_MOVED is not part of the mask */
633 #define RB_PAGE_MOVED 4UL
636 * rb_list_head - remove any bit
638 static struct list_head
*rb_list_head(struct list_head
*list
)
640 unsigned long val
= (unsigned long)list
;
642 return (struct list_head
*)(val
& ~RB_FLAG_MASK
);
646 * rb_is_head_page - test if the given page is the head page
648 * Because the reader may move the head_page pointer, we can
649 * not trust what the head page is (it may be pointing to
650 * the reader page). But if the next page is a header page,
651 * its flags will be non zero.
654 rb_is_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
655 struct buffer_page
*page
, struct list_head
*list
)
659 val
= (unsigned long)list
->next
;
661 if ((val
& ~RB_FLAG_MASK
) != (unsigned long)&page
->list
)
662 return RB_PAGE_MOVED
;
664 return val
& RB_FLAG_MASK
;
670 * The unique thing about the reader page, is that, if the
671 * writer is ever on it, the previous pointer never points
672 * back to the reader page.
674 static int rb_is_reader_page(struct buffer_page
*page
)
676 struct list_head
*list
= page
->list
.prev
;
678 return rb_list_head(list
->next
) != &page
->list
;
682 * rb_set_list_to_head - set a list_head to be pointing to head.
684 static void rb_set_list_to_head(struct ring_buffer_per_cpu
*cpu_buffer
,
685 struct list_head
*list
)
689 ptr
= (unsigned long *)&list
->next
;
690 *ptr
|= RB_PAGE_HEAD
;
691 *ptr
&= ~RB_PAGE_UPDATE
;
695 * rb_head_page_activate - sets up head page
697 static void rb_head_page_activate(struct ring_buffer_per_cpu
*cpu_buffer
)
699 struct buffer_page
*head
;
701 head
= cpu_buffer
->head_page
;
706 * Set the previous list pointer to have the HEAD flag.
708 rb_set_list_to_head(cpu_buffer
, head
->list
.prev
);
711 static void rb_list_head_clear(struct list_head
*list
)
713 unsigned long *ptr
= (unsigned long *)&list
->next
;
715 *ptr
&= ~RB_FLAG_MASK
;
719 * rb_head_page_dactivate - clears head page ptr (for free list)
722 rb_head_page_deactivate(struct ring_buffer_per_cpu
*cpu_buffer
)
724 struct list_head
*hd
;
726 /* Go through the whole list and clear any pointers found. */
727 rb_list_head_clear(cpu_buffer
->pages
);
729 list_for_each(hd
, cpu_buffer
->pages
)
730 rb_list_head_clear(hd
);
733 static int rb_head_page_set(struct ring_buffer_per_cpu
*cpu_buffer
,
734 struct buffer_page
*head
,
735 struct buffer_page
*prev
,
736 int old_flag
, int new_flag
)
738 struct list_head
*list
;
739 unsigned long val
= (unsigned long)&head
->list
;
744 val
&= ~RB_FLAG_MASK
;
746 ret
= cmpxchg((unsigned long *)&list
->next
,
747 val
| old_flag
, val
| new_flag
);
749 /* check if the reader took the page */
750 if ((ret
& ~RB_FLAG_MASK
) != val
)
751 return RB_PAGE_MOVED
;
753 return ret
& RB_FLAG_MASK
;
756 static int rb_head_page_set_update(struct ring_buffer_per_cpu
*cpu_buffer
,
757 struct buffer_page
*head
,
758 struct buffer_page
*prev
,
761 return rb_head_page_set(cpu_buffer
, head
, prev
,
762 old_flag
, RB_PAGE_UPDATE
);
765 static int rb_head_page_set_head(struct ring_buffer_per_cpu
*cpu_buffer
,
766 struct buffer_page
*head
,
767 struct buffer_page
*prev
,
770 return rb_head_page_set(cpu_buffer
, head
, prev
,
771 old_flag
, RB_PAGE_HEAD
);
774 static int rb_head_page_set_normal(struct ring_buffer_per_cpu
*cpu_buffer
,
775 struct buffer_page
*head
,
776 struct buffer_page
*prev
,
779 return rb_head_page_set(cpu_buffer
, head
, prev
,
780 old_flag
, RB_PAGE_NORMAL
);
783 static inline void rb_inc_page(struct ring_buffer_per_cpu
*cpu_buffer
,
784 struct buffer_page
**bpage
)
786 struct list_head
*p
= rb_list_head((*bpage
)->list
.next
);
788 *bpage
= list_entry(p
, struct buffer_page
, list
);
791 static struct buffer_page
*
792 rb_set_head_page(struct ring_buffer_per_cpu
*cpu_buffer
)
794 struct buffer_page
*head
;
795 struct buffer_page
*page
;
796 struct list_head
*list
;
799 if (RB_WARN_ON(cpu_buffer
, !cpu_buffer
->head_page
))
803 list
= cpu_buffer
->pages
;
804 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->prev
->next
) != list
))
807 page
= head
= cpu_buffer
->head_page
;
809 * It is possible that the writer moves the header behind
810 * where we started, and we miss in one loop.
811 * A second loop should grab the header, but we'll do
812 * three loops just because I'm paranoid.
814 for (i
= 0; i
< 3; i
++) {
816 if (rb_is_head_page(cpu_buffer
, page
, page
->list
.prev
)) {
817 cpu_buffer
->head_page
= page
;
820 rb_inc_page(cpu_buffer
, &page
);
821 } while (page
!= head
);
824 RB_WARN_ON(cpu_buffer
, 1);
829 static int rb_head_page_replace(struct buffer_page
*old
,
830 struct buffer_page
*new)
832 unsigned long *ptr
= (unsigned long *)&old
->list
.prev
->next
;
836 val
= *ptr
& ~RB_FLAG_MASK
;
839 ret
= cmpxchg(ptr
, val
, (unsigned long)&new->list
);
845 * rb_tail_page_update - move the tail page forward
847 * Returns 1 if moved tail page, 0 if someone else did.
849 static int rb_tail_page_update(struct ring_buffer_per_cpu
*cpu_buffer
,
850 struct buffer_page
*tail_page
,
851 struct buffer_page
*next_page
)
853 struct buffer_page
*old_tail
;
854 unsigned long old_entries
;
855 unsigned long old_write
;
859 * The tail page now needs to be moved forward.
861 * We need to reset the tail page, but without messing
862 * with possible erasing of data brought in by interrupts
863 * that have moved the tail page and are currently on it.
865 * We add a counter to the write field to denote this.
867 old_write
= local_add_return(RB_WRITE_INTCNT
, &next_page
->write
);
868 old_entries
= local_add_return(RB_WRITE_INTCNT
, &next_page
->entries
);
871 * Just make sure we have seen our old_write and synchronize
872 * with any interrupts that come in.
877 * If the tail page is still the same as what we think
878 * it is, then it is up to us to update the tail
881 if (tail_page
== cpu_buffer
->tail_page
) {
882 /* Zero the write counter */
883 unsigned long val
= old_write
& ~RB_WRITE_MASK
;
884 unsigned long eval
= old_entries
& ~RB_WRITE_MASK
;
887 * This will only succeed if an interrupt did
888 * not come in and change it. In which case, we
889 * do not want to modify it.
891 * We add (void) to let the compiler know that we do not care
892 * about the return value of these functions. We use the
893 * cmpxchg to only update if an interrupt did not already
894 * do it for us. If the cmpxchg fails, we don't care.
896 (void)local_cmpxchg(&next_page
->write
, old_write
, val
);
897 (void)local_cmpxchg(&next_page
->entries
, old_entries
, eval
);
900 * No need to worry about races with clearing out the commit.
901 * it only can increment when a commit takes place. But that
902 * only happens in the outer most nested commit.
904 local_set(&next_page
->page
->commit
, 0);
906 old_tail
= cmpxchg(&cpu_buffer
->tail_page
,
907 tail_page
, next_page
);
909 if (old_tail
== tail_page
)
916 static int rb_check_bpage(struct ring_buffer_per_cpu
*cpu_buffer
,
917 struct buffer_page
*bpage
)
919 unsigned long val
= (unsigned long)bpage
;
921 if (RB_WARN_ON(cpu_buffer
, val
& RB_FLAG_MASK
))
928 * rb_check_list - make sure a pointer to a list has the last bits zero
930 static int rb_check_list(struct ring_buffer_per_cpu
*cpu_buffer
,
931 struct list_head
*list
)
933 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->prev
) != list
->prev
))
935 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->next
) != list
->next
))
941 * check_pages - integrity check of buffer pages
942 * @cpu_buffer: CPU buffer with pages to test
944 * As a safety measure we check to make sure the data pages have not
947 static int rb_check_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
949 struct list_head
*head
= cpu_buffer
->pages
;
950 struct buffer_page
*bpage
, *tmp
;
952 /* Reset the head page if it exists */
953 if (cpu_buffer
->head_page
)
954 rb_set_head_page(cpu_buffer
);
956 rb_head_page_deactivate(cpu_buffer
);
958 if (RB_WARN_ON(cpu_buffer
, head
->next
->prev
!= head
))
960 if (RB_WARN_ON(cpu_buffer
, head
->prev
->next
!= head
))
963 if (rb_check_list(cpu_buffer
, head
))
966 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
967 if (RB_WARN_ON(cpu_buffer
,
968 bpage
->list
.next
->prev
!= &bpage
->list
))
970 if (RB_WARN_ON(cpu_buffer
,
971 bpage
->list
.prev
->next
!= &bpage
->list
))
973 if (rb_check_list(cpu_buffer
, &bpage
->list
))
977 rb_head_page_activate(cpu_buffer
);
982 static int __rb_allocate_pages(int nr_pages
, struct list_head
*pages
, int cpu
)
985 struct buffer_page
*bpage
, *tmp
;
987 for (i
= 0; i
< nr_pages
; i
++) {
990 * __GFP_NORETRY flag makes sure that the allocation fails
991 * gracefully without invoking oom-killer and the system is
994 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
995 GFP_KERNEL
| __GFP_NORETRY
,
1000 list_add(&bpage
->list
, pages
);
1002 page
= alloc_pages_node(cpu_to_node(cpu
),
1003 GFP_KERNEL
| __GFP_NORETRY
, 0);
1006 bpage
->page
= page_address(page
);
1007 rb_init_page(bpage
->page
);
1013 list_for_each_entry_safe(bpage
, tmp
, pages
, list
) {
1014 list_del_init(&bpage
->list
);
1015 free_buffer_page(bpage
);
1021 static int rb_allocate_pages(struct ring_buffer_per_cpu
*cpu_buffer
,
1028 if (__rb_allocate_pages(nr_pages
, &pages
, cpu_buffer
->cpu
))
1032 * The ring buffer page list is a circular list that does not
1033 * start and end with a list head. All page list items point to
1036 cpu_buffer
->pages
= pages
.next
;
1039 cpu_buffer
->nr_pages
= nr_pages
;
1041 rb_check_pages(cpu_buffer
);
1046 static struct ring_buffer_per_cpu
*
1047 rb_allocate_cpu_buffer(struct ring_buffer
*buffer
, int nr_pages
, int cpu
)
1049 struct ring_buffer_per_cpu
*cpu_buffer
;
1050 struct buffer_page
*bpage
;
1054 cpu_buffer
= kzalloc_node(ALIGN(sizeof(*cpu_buffer
), cache_line_size()),
1055 GFP_KERNEL
, cpu_to_node(cpu
));
1059 cpu_buffer
->cpu
= cpu
;
1060 cpu_buffer
->buffer
= buffer
;
1061 raw_spin_lock_init(&cpu_buffer
->reader_lock
);
1062 lockdep_set_class(&cpu_buffer
->reader_lock
, buffer
->reader_lock_key
);
1063 cpu_buffer
->lock
= (arch_spinlock_t
)__ARCH_SPIN_LOCK_UNLOCKED
;
1064 INIT_WORK(&cpu_buffer
->update_pages_work
, update_pages_handler
);
1065 init_completion(&cpu_buffer
->update_done
);
1067 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
1068 GFP_KERNEL
, cpu_to_node(cpu
));
1070 goto fail_free_buffer
;
1072 rb_check_bpage(cpu_buffer
, bpage
);
1074 cpu_buffer
->reader_page
= bpage
;
1075 page
= alloc_pages_node(cpu_to_node(cpu
), GFP_KERNEL
, 0);
1077 goto fail_free_reader
;
1078 bpage
->page
= page_address(page
);
1079 rb_init_page(bpage
->page
);
1081 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
1082 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1084 ret
= rb_allocate_pages(cpu_buffer
, nr_pages
);
1086 goto fail_free_reader
;
1088 cpu_buffer
->head_page
1089 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
1090 cpu_buffer
->tail_page
= cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
1092 rb_head_page_activate(cpu_buffer
);
1097 free_buffer_page(cpu_buffer
->reader_page
);
1104 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu
*cpu_buffer
)
1106 struct list_head
*head
= cpu_buffer
->pages
;
1107 struct buffer_page
*bpage
, *tmp
;
1109 free_buffer_page(cpu_buffer
->reader_page
);
1111 rb_head_page_deactivate(cpu_buffer
);
1114 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
1115 list_del_init(&bpage
->list
);
1116 free_buffer_page(bpage
);
1118 bpage
= list_entry(head
, struct buffer_page
, list
);
1119 free_buffer_page(bpage
);
1125 #ifdef CONFIG_HOTPLUG_CPU
1126 static int rb_cpu_notify(struct notifier_block
*self
,
1127 unsigned long action
, void *hcpu
);
1131 * ring_buffer_alloc - allocate a new ring_buffer
1132 * @size: the size in bytes per cpu that is needed.
1133 * @flags: attributes to set for the ring buffer.
1135 * Currently the only flag that is available is the RB_FL_OVERWRITE
1136 * flag. This flag means that the buffer will overwrite old data
1137 * when the buffer wraps. If this flag is not set, the buffer will
1138 * drop data when the tail hits the head.
1140 struct ring_buffer
*__ring_buffer_alloc(unsigned long size
, unsigned flags
,
1141 struct lock_class_key
*key
)
1143 struct ring_buffer
*buffer
;
1147 /* keep it in its own cache line */
1148 buffer
= kzalloc(ALIGN(sizeof(*buffer
), cache_line_size()),
1153 if (!alloc_cpumask_var(&buffer
->cpumask
, GFP_KERNEL
))
1154 goto fail_free_buffer
;
1156 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1157 buffer
->flags
= flags
;
1158 buffer
->clock
= trace_clock_local
;
1159 buffer
->reader_lock_key
= key
;
1161 /* need at least two pages */
1166 * In case of non-hotplug cpu, if the ring-buffer is allocated
1167 * in early initcall, it will not be notified of secondary cpus.
1168 * In that off case, we need to allocate for all possible cpus.
1170 #ifdef CONFIG_HOTPLUG_CPU
1172 cpumask_copy(buffer
->cpumask
, cpu_online_mask
);
1174 cpumask_copy(buffer
->cpumask
, cpu_possible_mask
);
1176 buffer
->cpus
= nr_cpu_ids
;
1178 bsize
= sizeof(void *) * nr_cpu_ids
;
1179 buffer
->buffers
= kzalloc(ALIGN(bsize
, cache_line_size()),
1181 if (!buffer
->buffers
)
1182 goto fail_free_cpumask
;
1184 for_each_buffer_cpu(buffer
, cpu
) {
1185 buffer
->buffers
[cpu
] =
1186 rb_allocate_cpu_buffer(buffer
, nr_pages
, cpu
);
1187 if (!buffer
->buffers
[cpu
])
1188 goto fail_free_buffers
;
1191 #ifdef CONFIG_HOTPLUG_CPU
1192 buffer
->cpu_notify
.notifier_call
= rb_cpu_notify
;
1193 buffer
->cpu_notify
.priority
= 0;
1194 register_cpu_notifier(&buffer
->cpu_notify
);
1198 mutex_init(&buffer
->mutex
);
1203 for_each_buffer_cpu(buffer
, cpu
) {
1204 if (buffer
->buffers
[cpu
])
1205 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1207 kfree(buffer
->buffers
);
1210 free_cpumask_var(buffer
->cpumask
);
1217 EXPORT_SYMBOL_GPL(__ring_buffer_alloc
);
1220 * ring_buffer_free - free a ring buffer.
1221 * @buffer: the buffer to free.
1224 ring_buffer_free(struct ring_buffer
*buffer
)
1230 #ifdef CONFIG_HOTPLUG_CPU
1231 unregister_cpu_notifier(&buffer
->cpu_notify
);
1234 for_each_buffer_cpu(buffer
, cpu
)
1235 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1239 kfree(buffer
->buffers
);
1240 free_cpumask_var(buffer
->cpumask
);
1244 EXPORT_SYMBOL_GPL(ring_buffer_free
);
1246 void ring_buffer_set_clock(struct ring_buffer
*buffer
,
1249 buffer
->clock
= clock
;
1252 static void rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
);
1254 static inline unsigned long rb_page_entries(struct buffer_page
*bpage
)
1256 return local_read(&bpage
->entries
) & RB_WRITE_MASK
;
1259 static inline unsigned long rb_page_write(struct buffer_page
*bpage
)
1261 return local_read(&bpage
->write
) & RB_WRITE_MASK
;
1265 rb_remove_pages(struct ring_buffer_per_cpu
*cpu_buffer
, unsigned int nr_pages
)
1267 struct list_head
*tail_page
, *to_remove
, *next_page
;
1268 struct buffer_page
*to_remove_page
, *tmp_iter_page
;
1269 struct buffer_page
*last_page
, *first_page
;
1270 unsigned int nr_removed
;
1271 unsigned long head_bit
;
1276 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1277 atomic_inc(&cpu_buffer
->record_disabled
);
1279 * We don't race with the readers since we have acquired the reader
1280 * lock. We also don't race with writers after disabling recording.
1281 * This makes it easy to figure out the first and the last page to be
1282 * removed from the list. We unlink all the pages in between including
1283 * the first and last pages. This is done in a busy loop so that we
1284 * lose the least number of traces.
1285 * The pages are freed after we restart recording and unlock readers.
1287 tail_page
= &cpu_buffer
->tail_page
->list
;
1290 * tail page might be on reader page, we remove the next page
1291 * from the ring buffer
1293 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
1294 tail_page
= rb_list_head(tail_page
->next
);
1295 to_remove
= tail_page
;
1297 /* start of pages to remove */
1298 first_page
= list_entry(rb_list_head(to_remove
->next
),
1299 struct buffer_page
, list
);
1301 for (nr_removed
= 0; nr_removed
< nr_pages
; nr_removed
++) {
1302 to_remove
= rb_list_head(to_remove
)->next
;
1303 head_bit
|= (unsigned long)to_remove
& RB_PAGE_HEAD
;
1306 next_page
= rb_list_head(to_remove
)->next
;
1309 * Now we remove all pages between tail_page and next_page.
1310 * Make sure that we have head_bit value preserved for the
1313 tail_page
->next
= (struct list_head
*)((unsigned long)next_page
|
1315 next_page
= rb_list_head(next_page
);
1316 next_page
->prev
= tail_page
;
1318 /* make sure pages points to a valid page in the ring buffer */
1319 cpu_buffer
->pages
= next_page
;
1321 /* update head page */
1323 cpu_buffer
->head_page
= list_entry(next_page
,
1324 struct buffer_page
, list
);
1327 * change read pointer to make sure any read iterators reset
1330 cpu_buffer
->read
= 0;
1332 /* pages are removed, resume tracing and then free the pages */
1333 atomic_dec(&cpu_buffer
->record_disabled
);
1334 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1336 RB_WARN_ON(cpu_buffer
, list_empty(cpu_buffer
->pages
));
1338 /* last buffer page to remove */
1339 last_page
= list_entry(rb_list_head(to_remove
), struct buffer_page
,
1341 tmp_iter_page
= first_page
;
1344 to_remove_page
= tmp_iter_page
;
1345 rb_inc_page(cpu_buffer
, &tmp_iter_page
);
1347 /* update the counters */
1348 page_entries
= rb_page_entries(to_remove_page
);
1351 * If something was added to this page, it was full
1352 * since it is not the tail page. So we deduct the
1353 * bytes consumed in ring buffer from here.
1354 * Increment overrun to account for the lost events.
1356 local_add(page_entries
, &cpu_buffer
->overrun
);
1357 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
1361 * We have already removed references to this list item, just
1362 * free up the buffer_page and its page
1364 free_buffer_page(to_remove_page
);
1367 } while (to_remove_page
!= last_page
);
1369 RB_WARN_ON(cpu_buffer
, nr_removed
);
1371 return nr_removed
== 0;
1375 rb_insert_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1377 struct list_head
*pages
= &cpu_buffer
->new_pages
;
1378 int retries
, success
;
1380 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1382 * We are holding the reader lock, so the reader page won't be swapped
1383 * in the ring buffer. Now we are racing with the writer trying to
1384 * move head page and the tail page.
1385 * We are going to adapt the reader page update process where:
1386 * 1. We first splice the start and end of list of new pages between
1387 * the head page and its previous page.
1388 * 2. We cmpxchg the prev_page->next to point from head page to the
1389 * start of new pages list.
1390 * 3. Finally, we update the head->prev to the end of new list.
1392 * We will try this process 10 times, to make sure that we don't keep
1398 struct list_head
*head_page
, *prev_page
, *r
;
1399 struct list_head
*last_page
, *first_page
;
1400 struct list_head
*head_page_with_bit
;
1402 head_page
= &rb_set_head_page(cpu_buffer
)->list
;
1405 prev_page
= head_page
->prev
;
1407 first_page
= pages
->next
;
1408 last_page
= pages
->prev
;
1410 head_page_with_bit
= (struct list_head
*)
1411 ((unsigned long)head_page
| RB_PAGE_HEAD
);
1413 last_page
->next
= head_page_with_bit
;
1414 first_page
->prev
= prev_page
;
1416 r
= cmpxchg(&prev_page
->next
, head_page_with_bit
, first_page
);
1418 if (r
== head_page_with_bit
) {
1420 * yay, we replaced the page pointer to our new list,
1421 * now, we just have to update to head page's prev
1422 * pointer to point to end of list
1424 head_page
->prev
= last_page
;
1431 INIT_LIST_HEAD(pages
);
1433 * If we weren't successful in adding in new pages, warn and stop
1436 RB_WARN_ON(cpu_buffer
, !success
);
1437 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1439 /* free pages if they weren't inserted */
1441 struct buffer_page
*bpage
, *tmp
;
1442 list_for_each_entry_safe(bpage
, tmp
, &cpu_buffer
->new_pages
,
1444 list_del_init(&bpage
->list
);
1445 free_buffer_page(bpage
);
1451 static void rb_update_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1455 if (cpu_buffer
->nr_pages_to_update
> 0)
1456 success
= rb_insert_pages(cpu_buffer
);
1458 success
= rb_remove_pages(cpu_buffer
,
1459 -cpu_buffer
->nr_pages_to_update
);
1462 cpu_buffer
->nr_pages
+= cpu_buffer
->nr_pages_to_update
;
1465 static void update_pages_handler(struct work_struct
*work
)
1467 struct ring_buffer_per_cpu
*cpu_buffer
= container_of(work
,
1468 struct ring_buffer_per_cpu
, update_pages_work
);
1469 rb_update_pages(cpu_buffer
);
1470 complete(&cpu_buffer
->update_done
);
1474 * ring_buffer_resize - resize the ring buffer
1475 * @buffer: the buffer to resize.
1476 * @size: the new size.
1478 * Minimum size is 2 * BUF_PAGE_SIZE.
1480 * Returns 0 on success and < 0 on failure.
1482 int ring_buffer_resize(struct ring_buffer
*buffer
, unsigned long size
,
1485 struct ring_buffer_per_cpu
*cpu_buffer
;
1490 * Always succeed at resizing a non-existent buffer:
1495 /* Make sure the requested buffer exists */
1496 if (cpu_id
!= RING_BUFFER_ALL_CPUS
&&
1497 !cpumask_test_cpu(cpu_id
, buffer
->cpumask
))
1500 size
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1501 size
*= BUF_PAGE_SIZE
;
1503 /* we need a minimum of two pages */
1504 if (size
< BUF_PAGE_SIZE
* 2)
1505 size
= BUF_PAGE_SIZE
* 2;
1507 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1510 * Don't succeed if resizing is disabled, as a reader might be
1511 * manipulating the ring buffer and is expecting a sane state while
1514 if (atomic_read(&buffer
->resize_disabled
))
1517 /* prevent another thread from changing buffer sizes */
1518 mutex_lock(&buffer
->mutex
);
1520 if (cpu_id
== RING_BUFFER_ALL_CPUS
) {
1521 /* calculate the pages to update */
1522 for_each_buffer_cpu(buffer
, cpu
) {
1523 cpu_buffer
= buffer
->buffers
[cpu
];
1525 cpu_buffer
->nr_pages_to_update
= nr_pages
-
1526 cpu_buffer
->nr_pages
;
1528 * nothing more to do for removing pages or no update
1530 if (cpu_buffer
->nr_pages_to_update
<= 0)
1533 * to add pages, make sure all new pages can be
1534 * allocated without receiving ENOMEM
1536 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1537 if (__rb_allocate_pages(cpu_buffer
->nr_pages_to_update
,
1538 &cpu_buffer
->new_pages
, cpu
)) {
1539 /* not enough memory for new pages */
1547 * Fire off all the required work handlers
1548 * We can't schedule on offline CPUs, but it's not necessary
1549 * since we can change their buffer sizes without any race.
1551 for_each_buffer_cpu(buffer
, cpu
) {
1552 cpu_buffer
= buffer
->buffers
[cpu
];
1553 if (!cpu_buffer
->nr_pages_to_update
)
1556 if (cpu_online(cpu
))
1557 schedule_work_on(cpu
,
1558 &cpu_buffer
->update_pages_work
);
1560 rb_update_pages(cpu_buffer
);
1563 /* wait for all the updates to complete */
1564 for_each_buffer_cpu(buffer
, cpu
) {
1565 cpu_buffer
= buffer
->buffers
[cpu
];
1566 if (!cpu_buffer
->nr_pages_to_update
)
1569 if (cpu_online(cpu
))
1570 wait_for_completion(&cpu_buffer
->update_done
);
1571 cpu_buffer
->nr_pages_to_update
= 0;
1576 /* Make sure this CPU has been intitialized */
1577 if (!cpumask_test_cpu(cpu_id
, buffer
->cpumask
))
1580 cpu_buffer
= buffer
->buffers
[cpu_id
];
1582 if (nr_pages
== cpu_buffer
->nr_pages
)
1585 cpu_buffer
->nr_pages_to_update
= nr_pages
-
1586 cpu_buffer
->nr_pages
;
1588 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1589 if (cpu_buffer
->nr_pages_to_update
> 0 &&
1590 __rb_allocate_pages(cpu_buffer
->nr_pages_to_update
,
1591 &cpu_buffer
->new_pages
, cpu_id
)) {
1598 if (cpu_online(cpu_id
)) {
1599 schedule_work_on(cpu_id
,
1600 &cpu_buffer
->update_pages_work
);
1601 wait_for_completion(&cpu_buffer
->update_done
);
1603 rb_update_pages(cpu_buffer
);
1605 cpu_buffer
->nr_pages_to_update
= 0;
1611 * The ring buffer resize can happen with the ring buffer
1612 * enabled, so that the update disturbs the tracing as little
1613 * as possible. But if the buffer is disabled, we do not need
1614 * to worry about that, and we can take the time to verify
1615 * that the buffer is not corrupt.
1617 if (atomic_read(&buffer
->record_disabled
)) {
1618 atomic_inc(&buffer
->record_disabled
);
1620 * Even though the buffer was disabled, we must make sure
1621 * that it is truly disabled before calling rb_check_pages.
1622 * There could have been a race between checking
1623 * record_disable and incrementing it.
1625 synchronize_sched();
1626 for_each_buffer_cpu(buffer
, cpu
) {
1627 cpu_buffer
= buffer
->buffers
[cpu
];
1628 rb_check_pages(cpu_buffer
);
1630 atomic_dec(&buffer
->record_disabled
);
1633 mutex_unlock(&buffer
->mutex
);
1637 for_each_buffer_cpu(buffer
, cpu
) {
1638 struct buffer_page
*bpage
, *tmp
;
1640 cpu_buffer
= buffer
->buffers
[cpu
];
1641 cpu_buffer
->nr_pages_to_update
= 0;
1643 if (list_empty(&cpu_buffer
->new_pages
))
1646 list_for_each_entry_safe(bpage
, tmp
, &cpu_buffer
->new_pages
,
1648 list_del_init(&bpage
->list
);
1649 free_buffer_page(bpage
);
1652 mutex_unlock(&buffer
->mutex
);
1655 EXPORT_SYMBOL_GPL(ring_buffer_resize
);
1657 void ring_buffer_change_overwrite(struct ring_buffer
*buffer
, int val
)
1659 mutex_lock(&buffer
->mutex
);
1661 buffer
->flags
|= RB_FL_OVERWRITE
;
1663 buffer
->flags
&= ~RB_FL_OVERWRITE
;
1664 mutex_unlock(&buffer
->mutex
);
1666 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite
);
1668 static inline void *
1669 __rb_data_page_index(struct buffer_data_page
*bpage
, unsigned index
)
1671 return bpage
->data
+ index
;
1674 static inline void *__rb_page_index(struct buffer_page
*bpage
, unsigned index
)
1676 return bpage
->page
->data
+ index
;
1679 static inline struct ring_buffer_event
*
1680 rb_reader_event(struct ring_buffer_per_cpu
*cpu_buffer
)
1682 return __rb_page_index(cpu_buffer
->reader_page
,
1683 cpu_buffer
->reader_page
->read
);
1686 static inline struct ring_buffer_event
*
1687 rb_iter_head_event(struct ring_buffer_iter
*iter
)
1689 return __rb_page_index(iter
->head_page
, iter
->head
);
1692 static inline unsigned rb_page_commit(struct buffer_page
*bpage
)
1694 return local_read(&bpage
->page
->commit
);
1697 /* Size is determined by what has been committed */
1698 static inline unsigned rb_page_size(struct buffer_page
*bpage
)
1700 return rb_page_commit(bpage
);
1703 static inline unsigned
1704 rb_commit_index(struct ring_buffer_per_cpu
*cpu_buffer
)
1706 return rb_page_commit(cpu_buffer
->commit_page
);
1709 static inline unsigned
1710 rb_event_index(struct ring_buffer_event
*event
)
1712 unsigned long addr
= (unsigned long)event
;
1714 return (addr
& ~PAGE_MASK
) - BUF_PAGE_HDR_SIZE
;
1718 rb_event_is_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
1719 struct ring_buffer_event
*event
)
1721 unsigned long addr
= (unsigned long)event
;
1722 unsigned long index
;
1724 index
= rb_event_index(event
);
1727 return cpu_buffer
->commit_page
->page
== (void *)addr
&&
1728 rb_commit_index(cpu_buffer
) == index
;
1732 rb_set_commit_to_write(struct ring_buffer_per_cpu
*cpu_buffer
)
1734 unsigned long max_count
;
1737 * We only race with interrupts and NMIs on this CPU.
1738 * If we own the commit event, then we can commit
1739 * all others that interrupted us, since the interruptions
1740 * are in stack format (they finish before they come
1741 * back to us). This allows us to do a simple loop to
1742 * assign the commit to the tail.
1745 max_count
= cpu_buffer
->nr_pages
* 100;
1747 while (cpu_buffer
->commit_page
!= cpu_buffer
->tail_page
) {
1748 if (RB_WARN_ON(cpu_buffer
, !(--max_count
)))
1750 if (RB_WARN_ON(cpu_buffer
,
1751 rb_is_reader_page(cpu_buffer
->tail_page
)))
1753 local_set(&cpu_buffer
->commit_page
->page
->commit
,
1754 rb_page_write(cpu_buffer
->commit_page
));
1755 rb_inc_page(cpu_buffer
, &cpu_buffer
->commit_page
);
1756 cpu_buffer
->write_stamp
=
1757 cpu_buffer
->commit_page
->page
->time_stamp
;
1758 /* add barrier to keep gcc from optimizing too much */
1761 while (rb_commit_index(cpu_buffer
) !=
1762 rb_page_write(cpu_buffer
->commit_page
)) {
1764 local_set(&cpu_buffer
->commit_page
->page
->commit
,
1765 rb_page_write(cpu_buffer
->commit_page
));
1766 RB_WARN_ON(cpu_buffer
,
1767 local_read(&cpu_buffer
->commit_page
->page
->commit
) &
1772 /* again, keep gcc from optimizing */
1776 * If an interrupt came in just after the first while loop
1777 * and pushed the tail page forward, we will be left with
1778 * a dangling commit that will never go forward.
1780 if (unlikely(cpu_buffer
->commit_page
!= cpu_buffer
->tail_page
))
1784 static void rb_reset_reader_page(struct ring_buffer_per_cpu
*cpu_buffer
)
1786 cpu_buffer
->read_stamp
= cpu_buffer
->reader_page
->page
->time_stamp
;
1787 cpu_buffer
->reader_page
->read
= 0;
1790 static void rb_inc_iter(struct ring_buffer_iter
*iter
)
1792 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
1795 * The iterator could be on the reader page (it starts there).
1796 * But the head could have moved, since the reader was
1797 * found. Check for this case and assign the iterator
1798 * to the head page instead of next.
1800 if (iter
->head_page
== cpu_buffer
->reader_page
)
1801 iter
->head_page
= rb_set_head_page(cpu_buffer
);
1803 rb_inc_page(cpu_buffer
, &iter
->head_page
);
1805 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
1809 /* Slow path, do not inline */
1810 static noinline
struct ring_buffer_event
*
1811 rb_add_time_stamp(struct ring_buffer_event
*event
, u64 delta
)
1813 event
->type_len
= RINGBUF_TYPE_TIME_EXTEND
;
1815 /* Not the first event on the page? */
1816 if (rb_event_index(event
)) {
1817 event
->time_delta
= delta
& TS_MASK
;
1818 event
->array
[0] = delta
>> TS_SHIFT
;
1820 /* nope, just zero it */
1821 event
->time_delta
= 0;
1822 event
->array
[0] = 0;
1825 return skip_time_extend(event
);
1829 * rb_update_event - update event type and data
1830 * @event: the even to update
1831 * @type: the type of event
1832 * @length: the size of the event field in the ring buffer
1834 * Update the type and data fields of the event. The length
1835 * is the actual size that is written to the ring buffer,
1836 * and with this, we can determine what to place into the
1840 rb_update_event(struct ring_buffer_per_cpu
*cpu_buffer
,
1841 struct ring_buffer_event
*event
, unsigned length
,
1842 int add_timestamp
, u64 delta
)
1844 /* Only a commit updates the timestamp */
1845 if (unlikely(!rb_event_is_commit(cpu_buffer
, event
)))
1849 * If we need to add a timestamp, then we
1850 * add it to the start of the resevered space.
1852 if (unlikely(add_timestamp
)) {
1853 event
= rb_add_time_stamp(event
, delta
);
1854 length
-= RB_LEN_TIME_EXTEND
;
1858 event
->time_delta
= delta
;
1859 length
-= RB_EVNT_HDR_SIZE
;
1860 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
) {
1861 event
->type_len
= 0;
1862 event
->array
[0] = length
;
1864 event
->type_len
= DIV_ROUND_UP(length
, RB_ALIGNMENT
);
1868 * rb_handle_head_page - writer hit the head page
1870 * Returns: +1 to retry page
1875 rb_handle_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
1876 struct buffer_page
*tail_page
,
1877 struct buffer_page
*next_page
)
1879 struct buffer_page
*new_head
;
1884 entries
= rb_page_entries(next_page
);
1887 * The hard part is here. We need to move the head
1888 * forward, and protect against both readers on
1889 * other CPUs and writers coming in via interrupts.
1891 type
= rb_head_page_set_update(cpu_buffer
, next_page
, tail_page
,
1895 * type can be one of four:
1896 * NORMAL - an interrupt already moved it for us
1897 * HEAD - we are the first to get here.
1898 * UPDATE - we are the interrupt interrupting
1900 * MOVED - a reader on another CPU moved the next
1901 * pointer to its reader page. Give up
1908 * We changed the head to UPDATE, thus
1909 * it is our responsibility to update
1912 local_add(entries
, &cpu_buffer
->overrun
);
1913 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
1916 * The entries will be zeroed out when we move the
1920 /* still more to do */
1923 case RB_PAGE_UPDATE
:
1925 * This is an interrupt that interrupt the
1926 * previous update. Still more to do.
1929 case RB_PAGE_NORMAL
:
1931 * An interrupt came in before the update
1932 * and processed this for us.
1933 * Nothing left to do.
1938 * The reader is on another CPU and just did
1939 * a swap with our next_page.
1944 RB_WARN_ON(cpu_buffer
, 1); /* WTF??? */
1949 * Now that we are here, the old head pointer is
1950 * set to UPDATE. This will keep the reader from
1951 * swapping the head page with the reader page.
1952 * The reader (on another CPU) will spin till
1955 * We just need to protect against interrupts
1956 * doing the job. We will set the next pointer
1957 * to HEAD. After that, we set the old pointer
1958 * to NORMAL, but only if it was HEAD before.
1959 * otherwise we are an interrupt, and only
1960 * want the outer most commit to reset it.
1962 new_head
= next_page
;
1963 rb_inc_page(cpu_buffer
, &new_head
);
1965 ret
= rb_head_page_set_head(cpu_buffer
, new_head
, next_page
,
1969 * Valid returns are:
1970 * HEAD - an interrupt came in and already set it.
1971 * NORMAL - One of two things:
1972 * 1) We really set it.
1973 * 2) A bunch of interrupts came in and moved
1974 * the page forward again.
1978 case RB_PAGE_NORMAL
:
1982 RB_WARN_ON(cpu_buffer
, 1);
1987 * It is possible that an interrupt came in,
1988 * set the head up, then more interrupts came in
1989 * and moved it again. When we get back here,
1990 * the page would have been set to NORMAL but we
1991 * just set it back to HEAD.
1993 * How do you detect this? Well, if that happened
1994 * the tail page would have moved.
1996 if (ret
== RB_PAGE_NORMAL
) {
1998 * If the tail had moved passed next, then we need
1999 * to reset the pointer.
2001 if (cpu_buffer
->tail_page
!= tail_page
&&
2002 cpu_buffer
->tail_page
!= next_page
)
2003 rb_head_page_set_normal(cpu_buffer
, new_head
,
2009 * If this was the outer most commit (the one that
2010 * changed the original pointer from HEAD to UPDATE),
2011 * then it is up to us to reset it to NORMAL.
2013 if (type
== RB_PAGE_HEAD
) {
2014 ret
= rb_head_page_set_normal(cpu_buffer
, next_page
,
2017 if (RB_WARN_ON(cpu_buffer
,
2018 ret
!= RB_PAGE_UPDATE
))
2025 static unsigned rb_calculate_event_length(unsigned length
)
2027 struct ring_buffer_event event
; /* Used only for sizeof array */
2029 /* zero length can cause confusions */
2033 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
)
2034 length
+= sizeof(event
.array
[0]);
2036 length
+= RB_EVNT_HDR_SIZE
;
2037 length
= ALIGN(length
, RB_ARCH_ALIGNMENT
);
2043 rb_reset_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
2044 struct buffer_page
*tail_page
,
2045 unsigned long tail
, unsigned long length
)
2047 struct ring_buffer_event
*event
;
2050 * Only the event that crossed the page boundary
2051 * must fill the old tail_page with padding.
2053 if (tail
>= BUF_PAGE_SIZE
) {
2055 * If the page was filled, then we still need
2056 * to update the real_end. Reset it to zero
2057 * and the reader will ignore it.
2059 if (tail
== BUF_PAGE_SIZE
)
2060 tail_page
->real_end
= 0;
2062 local_sub(length
, &tail_page
->write
);
2066 event
= __rb_page_index(tail_page
, tail
);
2067 kmemcheck_annotate_bitfield(event
, bitfield
);
2069 /* account for padding bytes */
2070 local_add(BUF_PAGE_SIZE
- tail
, &cpu_buffer
->entries_bytes
);
2073 * Save the original length to the meta data.
2074 * This will be used by the reader to add lost event
2077 tail_page
->real_end
= tail
;
2080 * If this event is bigger than the minimum size, then
2081 * we need to be careful that we don't subtract the
2082 * write counter enough to allow another writer to slip
2084 * We put in a discarded commit instead, to make sure
2085 * that this space is not used again.
2087 * If we are less than the minimum size, we don't need to
2090 if (tail
> (BUF_PAGE_SIZE
- RB_EVNT_MIN_SIZE
)) {
2091 /* No room for any events */
2093 /* Mark the rest of the page with padding */
2094 rb_event_set_padding(event
);
2096 /* Set the write back to the previous setting */
2097 local_sub(length
, &tail_page
->write
);
2101 /* Put in a discarded event */
2102 event
->array
[0] = (BUF_PAGE_SIZE
- tail
) - RB_EVNT_HDR_SIZE
;
2103 event
->type_len
= RINGBUF_TYPE_PADDING
;
2104 /* time delta must be non zero */
2105 event
->time_delta
= 1;
2107 /* Set write to end of buffer */
2108 length
= (tail
+ length
) - BUF_PAGE_SIZE
;
2109 local_sub(length
, &tail_page
->write
);
2113 * This is the slow path, force gcc not to inline it.
2115 static noinline
struct ring_buffer_event
*
2116 rb_move_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
2117 unsigned long length
, unsigned long tail
,
2118 struct buffer_page
*tail_page
, u64 ts
)
2120 struct buffer_page
*commit_page
= cpu_buffer
->commit_page
;
2121 struct ring_buffer
*buffer
= cpu_buffer
->buffer
;
2122 struct buffer_page
*next_page
;
2125 next_page
= tail_page
;
2127 rb_inc_page(cpu_buffer
, &next_page
);
2130 * If for some reason, we had an interrupt storm that made
2131 * it all the way around the buffer, bail, and warn
2134 if (unlikely(next_page
== commit_page
)) {
2135 local_inc(&cpu_buffer
->commit_overrun
);
2140 * This is where the fun begins!
2142 * We are fighting against races between a reader that
2143 * could be on another CPU trying to swap its reader
2144 * page with the buffer head.
2146 * We are also fighting against interrupts coming in and
2147 * moving the head or tail on us as well.
2149 * If the next page is the head page then we have filled
2150 * the buffer, unless the commit page is still on the
2153 if (rb_is_head_page(cpu_buffer
, next_page
, &tail_page
->list
)) {
2156 * If the commit is not on the reader page, then
2157 * move the header page.
2159 if (!rb_is_reader_page(cpu_buffer
->commit_page
)) {
2161 * If we are not in overwrite mode,
2162 * this is easy, just stop here.
2164 if (!(buffer
->flags
& RB_FL_OVERWRITE
)) {
2165 local_inc(&cpu_buffer
->dropped_events
);
2169 ret
= rb_handle_head_page(cpu_buffer
,
2178 * We need to be careful here too. The
2179 * commit page could still be on the reader
2180 * page. We could have a small buffer, and
2181 * have filled up the buffer with events
2182 * from interrupts and such, and wrapped.
2184 * Note, if the tail page is also the on the
2185 * reader_page, we let it move out.
2187 if (unlikely((cpu_buffer
->commit_page
!=
2188 cpu_buffer
->tail_page
) &&
2189 (cpu_buffer
->commit_page
==
2190 cpu_buffer
->reader_page
))) {
2191 local_inc(&cpu_buffer
->commit_overrun
);
2197 ret
= rb_tail_page_update(cpu_buffer
, tail_page
, next_page
);
2200 * Nested commits always have zero deltas, so
2201 * just reread the time stamp
2203 ts
= rb_time_stamp(buffer
);
2204 next_page
->page
->time_stamp
= ts
;
2209 rb_reset_tail(cpu_buffer
, tail_page
, tail
, length
);
2211 /* fail and let the caller try again */
2212 return ERR_PTR(-EAGAIN
);
2216 rb_reset_tail(cpu_buffer
, tail_page
, tail
, length
);
2221 static struct ring_buffer_event
*
2222 __rb_reserve_next(struct ring_buffer_per_cpu
*cpu_buffer
,
2223 unsigned long length
, u64 ts
,
2224 u64 delta
, int add_timestamp
)
2226 struct buffer_page
*tail_page
;
2227 struct ring_buffer_event
*event
;
2228 unsigned long tail
, write
;
2231 * If the time delta since the last event is too big to
2232 * hold in the time field of the event, then we append a
2233 * TIME EXTEND event ahead of the data event.
2235 if (unlikely(add_timestamp
))
2236 length
+= RB_LEN_TIME_EXTEND
;
2238 tail_page
= cpu_buffer
->tail_page
;
2239 write
= local_add_return(length
, &tail_page
->write
);
2241 /* set write to only the index of the write */
2242 write
&= RB_WRITE_MASK
;
2243 tail
= write
- length
;
2245 /* See if we shot pass the end of this buffer page */
2246 if (unlikely(write
> BUF_PAGE_SIZE
))
2247 return rb_move_tail(cpu_buffer
, length
, tail
,
2250 /* We reserved something on the buffer */
2252 event
= __rb_page_index(tail_page
, tail
);
2253 kmemcheck_annotate_bitfield(event
, bitfield
);
2254 rb_update_event(cpu_buffer
, event
, length
, add_timestamp
, delta
);
2256 local_inc(&tail_page
->entries
);
2259 * If this is the first commit on the page, then update
2263 tail_page
->page
->time_stamp
= ts
;
2265 /* account for these added bytes */
2266 local_add(length
, &cpu_buffer
->entries_bytes
);
2272 rb_try_to_discard(struct ring_buffer_per_cpu
*cpu_buffer
,
2273 struct ring_buffer_event
*event
)
2275 unsigned long new_index
, old_index
;
2276 struct buffer_page
*bpage
;
2277 unsigned long index
;
2280 new_index
= rb_event_index(event
);
2281 old_index
= new_index
+ rb_event_ts_length(event
);
2282 addr
= (unsigned long)event
;
2285 bpage
= cpu_buffer
->tail_page
;
2287 if (bpage
->page
== (void *)addr
&& rb_page_write(bpage
) == old_index
) {
2288 unsigned long write_mask
=
2289 local_read(&bpage
->write
) & ~RB_WRITE_MASK
;
2290 unsigned long event_length
= rb_event_length(event
);
2292 * This is on the tail page. It is possible that
2293 * a write could come in and move the tail page
2294 * and write to the next page. That is fine
2295 * because we just shorten what is on this page.
2297 old_index
+= write_mask
;
2298 new_index
+= write_mask
;
2299 index
= local_cmpxchg(&bpage
->write
, old_index
, new_index
);
2300 if (index
== old_index
) {
2301 /* update counters */
2302 local_sub(event_length
, &cpu_buffer
->entries_bytes
);
2307 /* could not discard */
2311 static void rb_start_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2313 local_inc(&cpu_buffer
->committing
);
2314 local_inc(&cpu_buffer
->commits
);
2317 static inline void rb_end_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2319 unsigned long commits
;
2321 if (RB_WARN_ON(cpu_buffer
,
2322 !local_read(&cpu_buffer
->committing
)))
2326 commits
= local_read(&cpu_buffer
->commits
);
2327 /* synchronize with interrupts */
2329 if (local_read(&cpu_buffer
->committing
) == 1)
2330 rb_set_commit_to_write(cpu_buffer
);
2332 local_dec(&cpu_buffer
->committing
);
2334 /* synchronize with interrupts */
2338 * Need to account for interrupts coming in between the
2339 * updating of the commit page and the clearing of the
2340 * committing counter.
2342 if (unlikely(local_read(&cpu_buffer
->commits
) != commits
) &&
2343 !local_read(&cpu_buffer
->committing
)) {
2344 local_inc(&cpu_buffer
->committing
);
2349 static struct ring_buffer_event
*
2350 rb_reserve_next_event(struct ring_buffer
*buffer
,
2351 struct ring_buffer_per_cpu
*cpu_buffer
,
2352 unsigned long length
)
2354 struct ring_buffer_event
*event
;
2360 rb_start_commit(cpu_buffer
);
2362 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2364 * Due to the ability to swap a cpu buffer from a buffer
2365 * it is possible it was swapped before we committed.
2366 * (committing stops a swap). We check for it here and
2367 * if it happened, we have to fail the write.
2370 if (unlikely(ACCESS_ONCE(cpu_buffer
->buffer
) != buffer
)) {
2371 local_dec(&cpu_buffer
->committing
);
2372 local_dec(&cpu_buffer
->commits
);
2377 length
= rb_calculate_event_length(length
);
2383 * We allow for interrupts to reenter here and do a trace.
2384 * If one does, it will cause this original code to loop
2385 * back here. Even with heavy interrupts happening, this
2386 * should only happen a few times in a row. If this happens
2387 * 1000 times in a row, there must be either an interrupt
2388 * storm or we have something buggy.
2391 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 1000))
2394 ts
= rb_time_stamp(cpu_buffer
->buffer
);
2395 diff
= ts
- cpu_buffer
->write_stamp
;
2397 /* make sure this diff is calculated here */
2400 /* Did the write stamp get updated already? */
2401 if (likely(ts
>= cpu_buffer
->write_stamp
)) {
2403 if (unlikely(test_time_stamp(delta
))) {
2404 int local_clock_stable
= 1;
2405 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2406 local_clock_stable
= sched_clock_stable
;
2408 WARN_ONCE(delta
> (1ULL << 59),
2409 KERN_WARNING
"Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2410 (unsigned long long)delta
,
2411 (unsigned long long)ts
,
2412 (unsigned long long)cpu_buffer
->write_stamp
,
2413 local_clock_stable
? "" :
2414 "If you just came from a suspend/resume,\n"
2415 "please switch to the trace global clock:\n"
2416 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2421 event
= __rb_reserve_next(cpu_buffer
, length
, ts
,
2422 delta
, add_timestamp
);
2423 if (unlikely(PTR_ERR(event
) == -EAGAIN
))
2432 rb_end_commit(cpu_buffer
);
2436 #ifdef CONFIG_TRACING
2439 * The lock and unlock are done within a preempt disable section.
2440 * The current_context per_cpu variable can only be modified
2441 * by the current task between lock and unlock. But it can
2442 * be modified more than once via an interrupt. To pass this
2443 * information from the lock to the unlock without having to
2444 * access the 'in_interrupt()' functions again (which do show
2445 * a bit of overhead in something as critical as function tracing,
2446 * we use a bitmask trick.
2448 * bit 0 = NMI context
2449 * bit 1 = IRQ context
2450 * bit 2 = SoftIRQ context
2451 * bit 3 = normal context.
2453 * This works because this is the order of contexts that can
2454 * preempt other contexts. A SoftIRQ never preempts an IRQ
2457 * When the context is determined, the corresponding bit is
2458 * checked and set (if it was set, then a recursion of that context
2461 * On unlock, we need to clear this bit. To do so, just subtract
2462 * 1 from the current_context and AND it to itself.
2466 * 101 & 100 = 100 (clearing bit zero)
2469 * 1010 & 1001 = 1000 (clearing bit 1)
2471 * The least significant bit can be cleared this way, and it
2472 * just so happens that it is the same bit corresponding to
2473 * the current context.
2475 static DEFINE_PER_CPU(unsigned int, current_context
);
2477 static __always_inline
int trace_recursive_lock(void)
2479 unsigned int val
= this_cpu_read(current_context
);
2482 if (in_interrupt()) {
2492 if (unlikely(val
& (1 << bit
)))
2496 this_cpu_write(current_context
, val
);
2501 static __always_inline
void trace_recursive_unlock(void)
2503 unsigned int val
= this_cpu_read(current_context
);
2506 val
&= this_cpu_read(current_context
);
2507 this_cpu_write(current_context
, val
);
2512 #define trace_recursive_lock() (0)
2513 #define trace_recursive_unlock() do { } while (0)
2518 * ring_buffer_lock_reserve - reserve a part of the buffer
2519 * @buffer: the ring buffer to reserve from
2520 * @length: the length of the data to reserve (excluding event header)
2522 * Returns a reseverd event on the ring buffer to copy directly to.
2523 * The user of this interface will need to get the body to write into
2524 * and can use the ring_buffer_event_data() interface.
2526 * The length is the length of the data needed, not the event length
2527 * which also includes the event header.
2529 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2530 * If NULL is returned, then nothing has been allocated or locked.
2532 struct ring_buffer_event
*
2533 ring_buffer_lock_reserve(struct ring_buffer
*buffer
, unsigned long length
)
2535 struct ring_buffer_per_cpu
*cpu_buffer
;
2536 struct ring_buffer_event
*event
;
2539 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
2542 /* If we are tracing schedule, we don't want to recurse */
2543 preempt_disable_notrace();
2545 if (atomic_read(&buffer
->record_disabled
))
2548 if (trace_recursive_lock())
2551 cpu
= raw_smp_processor_id();
2553 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2556 cpu_buffer
= buffer
->buffers
[cpu
];
2558 if (atomic_read(&cpu_buffer
->record_disabled
))
2561 if (length
> BUF_MAX_DATA_SIZE
)
2564 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
2571 trace_recursive_unlock();
2574 preempt_enable_notrace();
2577 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve
);
2580 rb_update_write_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
2581 struct ring_buffer_event
*event
)
2586 * The event first in the commit queue updates the
2589 if (rb_event_is_commit(cpu_buffer
, event
)) {
2591 * A commit event that is first on a page
2592 * updates the write timestamp with the page stamp
2594 if (!rb_event_index(event
))
2595 cpu_buffer
->write_stamp
=
2596 cpu_buffer
->commit_page
->page
->time_stamp
;
2597 else if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
2598 delta
= event
->array
[0];
2600 delta
+= event
->time_delta
;
2601 cpu_buffer
->write_stamp
+= delta
;
2603 cpu_buffer
->write_stamp
+= event
->time_delta
;
2607 static void rb_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
2608 struct ring_buffer_event
*event
)
2610 local_inc(&cpu_buffer
->entries
);
2611 rb_update_write_stamp(cpu_buffer
, event
);
2612 rb_end_commit(cpu_buffer
);
2616 * ring_buffer_unlock_commit - commit a reserved
2617 * @buffer: The buffer to commit to
2618 * @event: The event pointer to commit.
2620 * This commits the data to the ring buffer, and releases any locks held.
2622 * Must be paired with ring_buffer_lock_reserve.
2624 int ring_buffer_unlock_commit(struct ring_buffer
*buffer
,
2625 struct ring_buffer_event
*event
)
2627 struct ring_buffer_per_cpu
*cpu_buffer
;
2628 int cpu
= raw_smp_processor_id();
2630 cpu_buffer
= buffer
->buffers
[cpu
];
2632 rb_commit(cpu_buffer
, event
);
2634 trace_recursive_unlock();
2636 preempt_enable_notrace();
2640 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit
);
2642 static inline void rb_event_discard(struct ring_buffer_event
*event
)
2644 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
2645 event
= skip_time_extend(event
);
2647 /* array[0] holds the actual length for the discarded event */
2648 event
->array
[0] = rb_event_data_length(event
) - RB_EVNT_HDR_SIZE
;
2649 event
->type_len
= RINGBUF_TYPE_PADDING
;
2650 /* time delta must be non zero */
2651 if (!event
->time_delta
)
2652 event
->time_delta
= 1;
2656 * Decrement the entries to the page that an event is on.
2657 * The event does not even need to exist, only the pointer
2658 * to the page it is on. This may only be called before the commit
2662 rb_decrement_entry(struct ring_buffer_per_cpu
*cpu_buffer
,
2663 struct ring_buffer_event
*event
)
2665 unsigned long addr
= (unsigned long)event
;
2666 struct buffer_page
*bpage
= cpu_buffer
->commit_page
;
2667 struct buffer_page
*start
;
2671 /* Do the likely case first */
2672 if (likely(bpage
->page
== (void *)addr
)) {
2673 local_dec(&bpage
->entries
);
2678 * Because the commit page may be on the reader page we
2679 * start with the next page and check the end loop there.
2681 rb_inc_page(cpu_buffer
, &bpage
);
2684 if (bpage
->page
== (void *)addr
) {
2685 local_dec(&bpage
->entries
);
2688 rb_inc_page(cpu_buffer
, &bpage
);
2689 } while (bpage
!= start
);
2691 /* commit not part of this buffer?? */
2692 RB_WARN_ON(cpu_buffer
, 1);
2696 * ring_buffer_commit_discard - discard an event that has not been committed
2697 * @buffer: the ring buffer
2698 * @event: non committed event to discard
2700 * Sometimes an event that is in the ring buffer needs to be ignored.
2701 * This function lets the user discard an event in the ring buffer
2702 * and then that event will not be read later.
2704 * This function only works if it is called before the the item has been
2705 * committed. It will try to free the event from the ring buffer
2706 * if another event has not been added behind it.
2708 * If another event has been added behind it, it will set the event
2709 * up as discarded, and perform the commit.
2711 * If this function is called, do not call ring_buffer_unlock_commit on
2714 void ring_buffer_discard_commit(struct ring_buffer
*buffer
,
2715 struct ring_buffer_event
*event
)
2717 struct ring_buffer_per_cpu
*cpu_buffer
;
2720 /* The event is discarded regardless */
2721 rb_event_discard(event
);
2723 cpu
= smp_processor_id();
2724 cpu_buffer
= buffer
->buffers
[cpu
];
2727 * This must only be called if the event has not been
2728 * committed yet. Thus we can assume that preemption
2729 * is still disabled.
2731 RB_WARN_ON(buffer
, !local_read(&cpu_buffer
->committing
));
2733 rb_decrement_entry(cpu_buffer
, event
);
2734 if (rb_try_to_discard(cpu_buffer
, event
))
2738 * The commit is still visible by the reader, so we
2739 * must still update the timestamp.
2741 rb_update_write_stamp(cpu_buffer
, event
);
2743 rb_end_commit(cpu_buffer
);
2745 trace_recursive_unlock();
2747 preempt_enable_notrace();
2750 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit
);
2753 * ring_buffer_write - write data to the buffer without reserving
2754 * @buffer: The ring buffer to write to.
2755 * @length: The length of the data being written (excluding the event header)
2756 * @data: The data to write to the buffer.
2758 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2759 * one function. If you already have the data to write to the buffer, it
2760 * may be easier to simply call this function.
2762 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2763 * and not the length of the event which would hold the header.
2765 int ring_buffer_write(struct ring_buffer
*buffer
,
2766 unsigned long length
,
2769 struct ring_buffer_per_cpu
*cpu_buffer
;
2770 struct ring_buffer_event
*event
;
2775 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
2778 preempt_disable_notrace();
2780 if (atomic_read(&buffer
->record_disabled
))
2783 cpu
= raw_smp_processor_id();
2785 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2788 cpu_buffer
= buffer
->buffers
[cpu
];
2790 if (atomic_read(&cpu_buffer
->record_disabled
))
2793 if (length
> BUF_MAX_DATA_SIZE
)
2796 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
2800 body
= rb_event_data(event
);
2802 memcpy(body
, data
, length
);
2804 rb_commit(cpu_buffer
, event
);
2808 preempt_enable_notrace();
2812 EXPORT_SYMBOL_GPL(ring_buffer_write
);
2814 static int rb_per_cpu_empty(struct ring_buffer_per_cpu
*cpu_buffer
)
2816 struct buffer_page
*reader
= cpu_buffer
->reader_page
;
2817 struct buffer_page
*head
= rb_set_head_page(cpu_buffer
);
2818 struct buffer_page
*commit
= cpu_buffer
->commit_page
;
2820 /* In case of error, head will be NULL */
2821 if (unlikely(!head
))
2824 return reader
->read
== rb_page_commit(reader
) &&
2825 (commit
== reader
||
2827 head
->read
== rb_page_commit(commit
)));
2831 * ring_buffer_record_disable - stop all writes into the buffer
2832 * @buffer: The ring buffer to stop writes to.
2834 * This prevents all writes to the buffer. Any attempt to write
2835 * to the buffer after this will fail and return NULL.
2837 * The caller should call synchronize_sched() after this.
2839 void ring_buffer_record_disable(struct ring_buffer
*buffer
)
2841 atomic_inc(&buffer
->record_disabled
);
2843 EXPORT_SYMBOL_GPL(ring_buffer_record_disable
);
2846 * ring_buffer_record_enable - enable writes to the buffer
2847 * @buffer: The ring buffer to enable writes
2849 * Note, multiple disables will need the same number of enables
2850 * to truly enable the writing (much like preempt_disable).
2852 void ring_buffer_record_enable(struct ring_buffer
*buffer
)
2854 atomic_dec(&buffer
->record_disabled
);
2856 EXPORT_SYMBOL_GPL(ring_buffer_record_enable
);
2859 * ring_buffer_record_off - stop all writes into the buffer
2860 * @buffer: The ring buffer to stop writes to.
2862 * This prevents all writes to the buffer. Any attempt to write
2863 * to the buffer after this will fail and return NULL.
2865 * This is different than ring_buffer_record_disable() as
2866 * it works like an on/off switch, where as the disable() version
2867 * must be paired with a enable().
2869 void ring_buffer_record_off(struct ring_buffer
*buffer
)
2872 unsigned int new_rd
;
2875 rd
= atomic_read(&buffer
->record_disabled
);
2876 new_rd
= rd
| RB_BUFFER_OFF
;
2877 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
2879 EXPORT_SYMBOL_GPL(ring_buffer_record_off
);
2882 * ring_buffer_record_on - restart writes into the buffer
2883 * @buffer: The ring buffer to start writes to.
2885 * This enables all writes to the buffer that was disabled by
2886 * ring_buffer_record_off().
2888 * This is different than ring_buffer_record_enable() as
2889 * it works like an on/off switch, where as the enable() version
2890 * must be paired with a disable().
2892 void ring_buffer_record_on(struct ring_buffer
*buffer
)
2895 unsigned int new_rd
;
2898 rd
= atomic_read(&buffer
->record_disabled
);
2899 new_rd
= rd
& ~RB_BUFFER_OFF
;
2900 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
2902 EXPORT_SYMBOL_GPL(ring_buffer_record_on
);
2905 * ring_buffer_record_is_on - return true if the ring buffer can write
2906 * @buffer: The ring buffer to see if write is enabled
2908 * Returns true if the ring buffer is in a state that it accepts writes.
2910 int ring_buffer_record_is_on(struct ring_buffer
*buffer
)
2912 return !atomic_read(&buffer
->record_disabled
);
2916 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2917 * @buffer: The ring buffer to stop writes to.
2918 * @cpu: The CPU buffer to stop
2920 * This prevents all writes to the buffer. Any attempt to write
2921 * to the buffer after this will fail and return NULL.
2923 * The caller should call synchronize_sched() after this.
2925 void ring_buffer_record_disable_cpu(struct ring_buffer
*buffer
, int cpu
)
2927 struct ring_buffer_per_cpu
*cpu_buffer
;
2929 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2932 cpu_buffer
= buffer
->buffers
[cpu
];
2933 atomic_inc(&cpu_buffer
->record_disabled
);
2935 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu
);
2938 * ring_buffer_record_enable_cpu - enable writes to the buffer
2939 * @buffer: The ring buffer to enable writes
2940 * @cpu: The CPU to enable.
2942 * Note, multiple disables will need the same number of enables
2943 * to truly enable the writing (much like preempt_disable).
2945 void ring_buffer_record_enable_cpu(struct ring_buffer
*buffer
, int cpu
)
2947 struct ring_buffer_per_cpu
*cpu_buffer
;
2949 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2952 cpu_buffer
= buffer
->buffers
[cpu
];
2953 atomic_dec(&cpu_buffer
->record_disabled
);
2955 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu
);
2958 * The total entries in the ring buffer is the running counter
2959 * of entries entered into the ring buffer, minus the sum of
2960 * the entries read from the ring buffer and the number of
2961 * entries that were overwritten.
2963 static inline unsigned long
2964 rb_num_of_entries(struct ring_buffer_per_cpu
*cpu_buffer
)
2966 return local_read(&cpu_buffer
->entries
) -
2967 (local_read(&cpu_buffer
->overrun
) + cpu_buffer
->read
);
2971 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
2972 * @buffer: The ring buffer
2973 * @cpu: The per CPU buffer to read from.
2975 u64
ring_buffer_oldest_event_ts(struct ring_buffer
*buffer
, int cpu
)
2977 unsigned long flags
;
2978 struct ring_buffer_per_cpu
*cpu_buffer
;
2979 struct buffer_page
*bpage
;
2982 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2985 cpu_buffer
= buffer
->buffers
[cpu
];
2986 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
2988 * if the tail is on reader_page, oldest time stamp is on the reader
2991 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
2992 bpage
= cpu_buffer
->reader_page
;
2994 bpage
= rb_set_head_page(cpu_buffer
);
2996 ret
= bpage
->page
->time_stamp
;
2997 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3001 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts
);
3004 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3005 * @buffer: The ring buffer
3006 * @cpu: The per CPU buffer to read from.
3008 unsigned long ring_buffer_bytes_cpu(struct ring_buffer
*buffer
, int cpu
)
3010 struct ring_buffer_per_cpu
*cpu_buffer
;
3013 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3016 cpu_buffer
= buffer
->buffers
[cpu
];
3017 ret
= local_read(&cpu_buffer
->entries_bytes
) - cpu_buffer
->read_bytes
;
3021 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu
);
3024 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3025 * @buffer: The ring buffer
3026 * @cpu: The per CPU buffer to get the entries from.
3028 unsigned long ring_buffer_entries_cpu(struct ring_buffer
*buffer
, int cpu
)
3030 struct ring_buffer_per_cpu
*cpu_buffer
;
3032 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3035 cpu_buffer
= buffer
->buffers
[cpu
];
3037 return rb_num_of_entries(cpu_buffer
);
3039 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu
);
3042 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3043 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3044 * @buffer: The ring buffer
3045 * @cpu: The per CPU buffer to get the number of overruns from
3047 unsigned long ring_buffer_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
3049 struct ring_buffer_per_cpu
*cpu_buffer
;
3052 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3055 cpu_buffer
= buffer
->buffers
[cpu
];
3056 ret
= local_read(&cpu_buffer
->overrun
);
3060 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu
);
3063 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3064 * commits failing due to the buffer wrapping around while there are uncommitted
3065 * events, such as during an interrupt storm.
3066 * @buffer: The ring buffer
3067 * @cpu: The per CPU buffer to get the number of overruns from
3070 ring_buffer_commit_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
3072 struct ring_buffer_per_cpu
*cpu_buffer
;
3075 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3078 cpu_buffer
= buffer
->buffers
[cpu
];
3079 ret
= local_read(&cpu_buffer
->commit_overrun
);
3083 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu
);
3086 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3087 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3088 * @buffer: The ring buffer
3089 * @cpu: The per CPU buffer to get the number of overruns from
3092 ring_buffer_dropped_events_cpu(struct ring_buffer
*buffer
, int cpu
)
3094 struct ring_buffer_per_cpu
*cpu_buffer
;
3097 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3100 cpu_buffer
= buffer
->buffers
[cpu
];
3101 ret
= local_read(&cpu_buffer
->dropped_events
);
3105 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu
);
3108 * ring_buffer_read_events_cpu - get the number of events successfully read
3109 * @buffer: The ring buffer
3110 * @cpu: The per CPU buffer to get the number of events read
3113 ring_buffer_read_events_cpu(struct ring_buffer
*buffer
, int cpu
)
3115 struct ring_buffer_per_cpu
*cpu_buffer
;
3117 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3120 cpu_buffer
= buffer
->buffers
[cpu
];
3121 return cpu_buffer
->read
;
3123 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu
);
3126 * ring_buffer_entries - get the number of entries in a buffer
3127 * @buffer: The ring buffer
3129 * Returns the total number of entries in the ring buffer
3132 unsigned long ring_buffer_entries(struct ring_buffer
*buffer
)
3134 struct ring_buffer_per_cpu
*cpu_buffer
;
3135 unsigned long entries
= 0;
3138 /* if you care about this being correct, lock the buffer */
3139 for_each_buffer_cpu(buffer
, cpu
) {
3140 cpu_buffer
= buffer
->buffers
[cpu
];
3141 entries
+= rb_num_of_entries(cpu_buffer
);
3146 EXPORT_SYMBOL_GPL(ring_buffer_entries
);
3149 * ring_buffer_overruns - get the number of overruns in buffer
3150 * @buffer: The ring buffer
3152 * Returns the total number of overruns in the ring buffer
3155 unsigned long ring_buffer_overruns(struct ring_buffer
*buffer
)
3157 struct ring_buffer_per_cpu
*cpu_buffer
;
3158 unsigned long overruns
= 0;
3161 /* if you care about this being correct, lock the buffer */
3162 for_each_buffer_cpu(buffer
, cpu
) {
3163 cpu_buffer
= buffer
->buffers
[cpu
];
3164 overruns
+= local_read(&cpu_buffer
->overrun
);
3169 EXPORT_SYMBOL_GPL(ring_buffer_overruns
);
3171 static void rb_iter_reset(struct ring_buffer_iter
*iter
)
3173 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3175 /* Iterator usage is expected to have record disabled */
3176 if (list_empty(&cpu_buffer
->reader_page
->list
)) {
3177 iter
->head_page
= rb_set_head_page(cpu_buffer
);
3178 if (unlikely(!iter
->head_page
))
3180 iter
->head
= iter
->head_page
->read
;
3182 iter
->head_page
= cpu_buffer
->reader_page
;
3183 iter
->head
= cpu_buffer
->reader_page
->read
;
3186 iter
->read_stamp
= cpu_buffer
->read_stamp
;
3188 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
3189 iter
->cache_reader_page
= cpu_buffer
->reader_page
;
3190 iter
->cache_read
= cpu_buffer
->read
;
3194 * ring_buffer_iter_reset - reset an iterator
3195 * @iter: The iterator to reset
3197 * Resets the iterator, so that it will start from the beginning
3200 void ring_buffer_iter_reset(struct ring_buffer_iter
*iter
)
3202 struct ring_buffer_per_cpu
*cpu_buffer
;
3203 unsigned long flags
;
3208 cpu_buffer
= iter
->cpu_buffer
;
3210 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3211 rb_iter_reset(iter
);
3212 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3214 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset
);
3217 * ring_buffer_iter_empty - check if an iterator has no more to read
3218 * @iter: The iterator to check
3220 int ring_buffer_iter_empty(struct ring_buffer_iter
*iter
)
3222 struct ring_buffer_per_cpu
*cpu_buffer
;
3224 cpu_buffer
= iter
->cpu_buffer
;
3226 return iter
->head_page
== cpu_buffer
->commit_page
&&
3227 iter
->head
== rb_commit_index(cpu_buffer
);
3229 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty
);
3232 rb_update_read_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
3233 struct ring_buffer_event
*event
)
3237 switch (event
->type_len
) {
3238 case RINGBUF_TYPE_PADDING
:
3241 case RINGBUF_TYPE_TIME_EXTEND
:
3242 delta
= event
->array
[0];
3244 delta
+= event
->time_delta
;
3245 cpu_buffer
->read_stamp
+= delta
;
3248 case RINGBUF_TYPE_TIME_STAMP
:
3249 /* FIXME: not implemented */
3252 case RINGBUF_TYPE_DATA
:
3253 cpu_buffer
->read_stamp
+= event
->time_delta
;
3263 rb_update_iter_read_stamp(struct ring_buffer_iter
*iter
,
3264 struct ring_buffer_event
*event
)
3268 switch (event
->type_len
) {
3269 case RINGBUF_TYPE_PADDING
:
3272 case RINGBUF_TYPE_TIME_EXTEND
:
3273 delta
= event
->array
[0];
3275 delta
+= event
->time_delta
;
3276 iter
->read_stamp
+= delta
;
3279 case RINGBUF_TYPE_TIME_STAMP
:
3280 /* FIXME: not implemented */
3283 case RINGBUF_TYPE_DATA
:
3284 iter
->read_stamp
+= event
->time_delta
;
3293 static struct buffer_page
*
3294 rb_get_reader_page(struct ring_buffer_per_cpu
*cpu_buffer
)
3296 struct buffer_page
*reader
= NULL
;
3297 unsigned long overwrite
;
3298 unsigned long flags
;
3302 local_irq_save(flags
);
3303 arch_spin_lock(&cpu_buffer
->lock
);
3307 * This should normally only loop twice. But because the
3308 * start of the reader inserts an empty page, it causes
3309 * a case where we will loop three times. There should be no
3310 * reason to loop four times (that I know of).
3312 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 3)) {
3317 reader
= cpu_buffer
->reader_page
;
3319 /* If there's more to read, return this page */
3320 if (cpu_buffer
->reader_page
->read
< rb_page_size(reader
))
3323 /* Never should we have an index greater than the size */
3324 if (RB_WARN_ON(cpu_buffer
,
3325 cpu_buffer
->reader_page
->read
> rb_page_size(reader
)))
3328 /* check if we caught up to the tail */
3330 if (cpu_buffer
->commit_page
== cpu_buffer
->reader_page
)
3333 /* Don't bother swapping if the ring buffer is empty */
3334 if (rb_num_of_entries(cpu_buffer
) == 0)
3338 * Reset the reader page to size zero.
3340 local_set(&cpu_buffer
->reader_page
->write
, 0);
3341 local_set(&cpu_buffer
->reader_page
->entries
, 0);
3342 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
3343 cpu_buffer
->reader_page
->real_end
= 0;
3347 * Splice the empty reader page into the list around the head.
3349 reader
= rb_set_head_page(cpu_buffer
);
3352 cpu_buffer
->reader_page
->list
.next
= rb_list_head(reader
->list
.next
);
3353 cpu_buffer
->reader_page
->list
.prev
= reader
->list
.prev
;
3356 * cpu_buffer->pages just needs to point to the buffer, it
3357 * has no specific buffer page to point to. Lets move it out
3358 * of our way so we don't accidentally swap it.
3360 cpu_buffer
->pages
= reader
->list
.prev
;
3362 /* The reader page will be pointing to the new head */
3363 rb_set_list_to_head(cpu_buffer
, &cpu_buffer
->reader_page
->list
);
3366 * We want to make sure we read the overruns after we set up our
3367 * pointers to the next object. The writer side does a
3368 * cmpxchg to cross pages which acts as the mb on the writer
3369 * side. Note, the reader will constantly fail the swap
3370 * while the writer is updating the pointers, so this
3371 * guarantees that the overwrite recorded here is the one we
3372 * want to compare with the last_overrun.
3375 overwrite
= local_read(&(cpu_buffer
->overrun
));
3378 * Here's the tricky part.
3380 * We need to move the pointer past the header page.
3381 * But we can only do that if a writer is not currently
3382 * moving it. The page before the header page has the
3383 * flag bit '1' set if it is pointing to the page we want.
3384 * but if the writer is in the process of moving it
3385 * than it will be '2' or already moved '0'.
3388 ret
= rb_head_page_replace(reader
, cpu_buffer
->reader_page
);
3391 * If we did not convert it, then we must try again.
3397 * Yeah! We succeeded in replacing the page.
3399 * Now make the new head point back to the reader page.
3401 rb_list_head(reader
->list
.next
)->prev
= &cpu_buffer
->reader_page
->list
;
3402 rb_inc_page(cpu_buffer
, &cpu_buffer
->head_page
);
3404 /* Finally update the reader page to the new head */
3405 cpu_buffer
->reader_page
= reader
;
3406 rb_reset_reader_page(cpu_buffer
);
3408 if (overwrite
!= cpu_buffer
->last_overrun
) {
3409 cpu_buffer
->lost_events
= overwrite
- cpu_buffer
->last_overrun
;
3410 cpu_buffer
->last_overrun
= overwrite
;
3416 arch_spin_unlock(&cpu_buffer
->lock
);
3417 local_irq_restore(flags
);
3422 static void rb_advance_reader(struct ring_buffer_per_cpu
*cpu_buffer
)
3424 struct ring_buffer_event
*event
;
3425 struct buffer_page
*reader
;
3428 reader
= rb_get_reader_page(cpu_buffer
);
3430 /* This function should not be called when buffer is empty */
3431 if (RB_WARN_ON(cpu_buffer
, !reader
))
3434 event
= rb_reader_event(cpu_buffer
);
3436 if (event
->type_len
<= RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
3439 rb_update_read_stamp(cpu_buffer
, event
);
3441 length
= rb_event_length(event
);
3442 cpu_buffer
->reader_page
->read
+= length
;
3445 static void rb_advance_iter(struct ring_buffer_iter
*iter
)
3447 struct ring_buffer_per_cpu
*cpu_buffer
;
3448 struct ring_buffer_event
*event
;
3451 cpu_buffer
= iter
->cpu_buffer
;
3454 * Check if we are at the end of the buffer.
3456 if (iter
->head
>= rb_page_size(iter
->head_page
)) {
3457 /* discarded commits can make the page empty */
3458 if (iter
->head_page
== cpu_buffer
->commit_page
)
3464 event
= rb_iter_head_event(iter
);
3466 length
= rb_event_length(event
);
3469 * This should not be called to advance the header if we are
3470 * at the tail of the buffer.
3472 if (RB_WARN_ON(cpu_buffer
,
3473 (iter
->head_page
== cpu_buffer
->commit_page
) &&
3474 (iter
->head
+ length
> rb_commit_index(cpu_buffer
))))
3477 rb_update_iter_read_stamp(iter
, event
);
3479 iter
->head
+= length
;
3481 /* check for end of page padding */
3482 if ((iter
->head
>= rb_page_size(iter
->head_page
)) &&
3483 (iter
->head_page
!= cpu_buffer
->commit_page
))
3487 static int rb_lost_events(struct ring_buffer_per_cpu
*cpu_buffer
)
3489 return cpu_buffer
->lost_events
;
3492 static struct ring_buffer_event
*
3493 rb_buffer_peek(struct ring_buffer_per_cpu
*cpu_buffer
, u64
*ts
,
3494 unsigned long *lost_events
)
3496 struct ring_buffer_event
*event
;
3497 struct buffer_page
*reader
;
3502 * We repeat when a time extend is encountered.
3503 * Since the time extend is always attached to a data event,
3504 * we should never loop more than once.
3505 * (We never hit the following condition more than twice).
3507 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 2))
3510 reader
= rb_get_reader_page(cpu_buffer
);
3514 event
= rb_reader_event(cpu_buffer
);
3516 switch (event
->type_len
) {
3517 case RINGBUF_TYPE_PADDING
:
3518 if (rb_null_event(event
))
3519 RB_WARN_ON(cpu_buffer
, 1);
3521 * Because the writer could be discarding every
3522 * event it creates (which would probably be bad)
3523 * if we were to go back to "again" then we may never
3524 * catch up, and will trigger the warn on, or lock
3525 * the box. Return the padding, and we will release
3526 * the current locks, and try again.
3530 case RINGBUF_TYPE_TIME_EXTEND
:
3531 /* Internal data, OK to advance */
3532 rb_advance_reader(cpu_buffer
);
3535 case RINGBUF_TYPE_TIME_STAMP
:
3536 /* FIXME: not implemented */
3537 rb_advance_reader(cpu_buffer
);
3540 case RINGBUF_TYPE_DATA
:
3542 *ts
= cpu_buffer
->read_stamp
+ event
->time_delta
;
3543 ring_buffer_normalize_time_stamp(cpu_buffer
->buffer
,
3544 cpu_buffer
->cpu
, ts
);
3547 *lost_events
= rb_lost_events(cpu_buffer
);
3556 EXPORT_SYMBOL_GPL(ring_buffer_peek
);
3558 static struct ring_buffer_event
*
3559 rb_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3561 struct ring_buffer
*buffer
;
3562 struct ring_buffer_per_cpu
*cpu_buffer
;
3563 struct ring_buffer_event
*event
;
3566 cpu_buffer
= iter
->cpu_buffer
;
3567 buffer
= cpu_buffer
->buffer
;
3570 * Check if someone performed a consuming read to
3571 * the buffer. A consuming read invalidates the iterator
3572 * and we need to reset the iterator in this case.
3574 if (unlikely(iter
->cache_read
!= cpu_buffer
->read
||
3575 iter
->cache_reader_page
!= cpu_buffer
->reader_page
))
3576 rb_iter_reset(iter
);
3579 if (ring_buffer_iter_empty(iter
))
3583 * We repeat when a time extend is encountered.
3584 * Since the time extend is always attached to a data event,
3585 * we should never loop more than once.
3586 * (We never hit the following condition more than twice).
3588 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 2))
3591 if (rb_per_cpu_empty(cpu_buffer
))
3594 if (iter
->head
>= local_read(&iter
->head_page
->page
->commit
)) {
3599 event
= rb_iter_head_event(iter
);
3601 switch (event
->type_len
) {
3602 case RINGBUF_TYPE_PADDING
:
3603 if (rb_null_event(event
)) {
3607 rb_advance_iter(iter
);
3610 case RINGBUF_TYPE_TIME_EXTEND
:
3611 /* Internal data, OK to advance */
3612 rb_advance_iter(iter
);
3615 case RINGBUF_TYPE_TIME_STAMP
:
3616 /* FIXME: not implemented */
3617 rb_advance_iter(iter
);
3620 case RINGBUF_TYPE_DATA
:
3622 *ts
= iter
->read_stamp
+ event
->time_delta
;
3623 ring_buffer_normalize_time_stamp(buffer
,
3624 cpu_buffer
->cpu
, ts
);
3634 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek
);
3636 static inline int rb_ok_to_lock(void)
3639 * If an NMI die dumps out the content of the ring buffer
3640 * do not grab locks. We also permanently disable the ring
3641 * buffer too. A one time deal is all you get from reading
3642 * the ring buffer from an NMI.
3644 if (likely(!in_nmi()))
3647 tracing_off_permanent();
3652 * ring_buffer_peek - peek at the next event to be read
3653 * @buffer: The ring buffer to read
3654 * @cpu: The cpu to peak at
3655 * @ts: The timestamp counter of this event.
3656 * @lost_events: a variable to store if events were lost (may be NULL)
3658 * This will return the event that will be read next, but does
3659 * not consume the data.
3661 struct ring_buffer_event
*
3662 ring_buffer_peek(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3663 unsigned long *lost_events
)
3665 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
3666 struct ring_buffer_event
*event
;
3667 unsigned long flags
;
3670 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3673 dolock
= rb_ok_to_lock();
3675 local_irq_save(flags
);
3677 raw_spin_lock(&cpu_buffer
->reader_lock
);
3678 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3679 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3680 rb_advance_reader(cpu_buffer
);
3682 raw_spin_unlock(&cpu_buffer
->reader_lock
);
3683 local_irq_restore(flags
);
3685 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3692 * ring_buffer_iter_peek - peek at the next event to be read
3693 * @iter: The ring buffer iterator
3694 * @ts: The timestamp counter of this event.
3696 * This will return the event that will be read next, but does
3697 * not increment the iterator.
3699 struct ring_buffer_event
*
3700 ring_buffer_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3702 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3703 struct ring_buffer_event
*event
;
3704 unsigned long flags
;
3707 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3708 event
= rb_iter_peek(iter
, ts
);
3709 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3711 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3718 * ring_buffer_consume - return an event and consume it
3719 * @buffer: The ring buffer to get the next event from
3720 * @cpu: the cpu to read the buffer from
3721 * @ts: a variable to store the timestamp (may be NULL)
3722 * @lost_events: a variable to store if events were lost (may be NULL)
3724 * Returns the next event in the ring buffer, and that event is consumed.
3725 * Meaning, that sequential reads will keep returning a different event,
3726 * and eventually empty the ring buffer if the producer is slower.
3728 struct ring_buffer_event
*
3729 ring_buffer_consume(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3730 unsigned long *lost_events
)
3732 struct ring_buffer_per_cpu
*cpu_buffer
;
3733 struct ring_buffer_event
*event
= NULL
;
3734 unsigned long flags
;
3737 dolock
= rb_ok_to_lock();
3740 /* might be called in atomic */
3743 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3746 cpu_buffer
= buffer
->buffers
[cpu
];
3747 local_irq_save(flags
);
3749 raw_spin_lock(&cpu_buffer
->reader_lock
);
3751 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3753 cpu_buffer
->lost_events
= 0;
3754 rb_advance_reader(cpu_buffer
);
3758 raw_spin_unlock(&cpu_buffer
->reader_lock
);
3759 local_irq_restore(flags
);
3764 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3769 EXPORT_SYMBOL_GPL(ring_buffer_consume
);
3772 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3773 * @buffer: The ring buffer to read from
3774 * @cpu: The cpu buffer to iterate over
3776 * This performs the initial preparations necessary to iterate
3777 * through the buffer. Memory is allocated, buffer recording
3778 * is disabled, and the iterator pointer is returned to the caller.
3780 * Disabling buffer recordng prevents the reading from being
3781 * corrupted. This is not a consuming read, so a producer is not
3784 * After a sequence of ring_buffer_read_prepare calls, the user is
3785 * expected to make at least one call to ring_buffer_prepare_sync.
3786 * Afterwards, ring_buffer_read_start is invoked to get things going
3789 * This overall must be paired with ring_buffer_finish.
3791 struct ring_buffer_iter
*
3792 ring_buffer_read_prepare(struct ring_buffer
*buffer
, int cpu
)
3794 struct ring_buffer_per_cpu
*cpu_buffer
;
3795 struct ring_buffer_iter
*iter
;
3797 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3800 iter
= kmalloc(sizeof(*iter
), GFP_KERNEL
);
3804 cpu_buffer
= buffer
->buffers
[cpu
];
3806 iter
->cpu_buffer
= cpu_buffer
;
3808 atomic_inc(&buffer
->resize_disabled
);
3809 atomic_inc(&cpu_buffer
->record_disabled
);
3813 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare
);
3816 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3818 * All previously invoked ring_buffer_read_prepare calls to prepare
3819 * iterators will be synchronized. Afterwards, read_buffer_read_start
3820 * calls on those iterators are allowed.
3823 ring_buffer_read_prepare_sync(void)
3825 synchronize_sched();
3827 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync
);
3830 * ring_buffer_read_start - start a non consuming read of the buffer
3831 * @iter: The iterator returned by ring_buffer_read_prepare
3833 * This finalizes the startup of an iteration through the buffer.
3834 * The iterator comes from a call to ring_buffer_read_prepare and
3835 * an intervening ring_buffer_read_prepare_sync must have been
3838 * Must be paired with ring_buffer_finish.
3841 ring_buffer_read_start(struct ring_buffer_iter
*iter
)
3843 struct ring_buffer_per_cpu
*cpu_buffer
;
3844 unsigned long flags
;
3849 cpu_buffer
= iter
->cpu_buffer
;
3851 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3852 arch_spin_lock(&cpu_buffer
->lock
);
3853 rb_iter_reset(iter
);
3854 arch_spin_unlock(&cpu_buffer
->lock
);
3855 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3857 EXPORT_SYMBOL_GPL(ring_buffer_read_start
);
3860 * ring_buffer_finish - finish reading the iterator of the buffer
3861 * @iter: The iterator retrieved by ring_buffer_start
3863 * This re-enables the recording to the buffer, and frees the
3867 ring_buffer_read_finish(struct ring_buffer_iter
*iter
)
3869 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3870 unsigned long flags
;
3873 * Ring buffer is disabled from recording, here's a good place
3874 * to check the integrity of the ring buffer.
3875 * Must prevent readers from trying to read, as the check
3876 * clears the HEAD page and readers require it.
3878 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3879 rb_check_pages(cpu_buffer
);
3880 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3882 atomic_dec(&cpu_buffer
->record_disabled
);
3883 atomic_dec(&cpu_buffer
->buffer
->resize_disabled
);
3886 EXPORT_SYMBOL_GPL(ring_buffer_read_finish
);
3889 * ring_buffer_read - read the next item in the ring buffer by the iterator
3890 * @iter: The ring buffer iterator
3891 * @ts: The time stamp of the event read.
3893 * This reads the next event in the ring buffer and increments the iterator.
3895 struct ring_buffer_event
*
3896 ring_buffer_read(struct ring_buffer_iter
*iter
, u64
*ts
)
3898 struct ring_buffer_event
*event
;
3899 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3900 unsigned long flags
;
3902 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3904 event
= rb_iter_peek(iter
, ts
);
3908 if (event
->type_len
== RINGBUF_TYPE_PADDING
)
3911 rb_advance_iter(iter
);
3913 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3917 EXPORT_SYMBOL_GPL(ring_buffer_read
);
3920 * ring_buffer_size - return the size of the ring buffer (in bytes)
3921 * @buffer: The ring buffer.
3923 unsigned long ring_buffer_size(struct ring_buffer
*buffer
, int cpu
)
3926 * Earlier, this method returned
3927 * BUF_PAGE_SIZE * buffer->nr_pages
3928 * Since the nr_pages field is now removed, we have converted this to
3929 * return the per cpu buffer value.
3931 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3934 return BUF_PAGE_SIZE
* buffer
->buffers
[cpu
]->nr_pages
;
3936 EXPORT_SYMBOL_GPL(ring_buffer_size
);
3939 rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
)
3941 rb_head_page_deactivate(cpu_buffer
);
3943 cpu_buffer
->head_page
3944 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
3945 local_set(&cpu_buffer
->head_page
->write
, 0);
3946 local_set(&cpu_buffer
->head_page
->entries
, 0);
3947 local_set(&cpu_buffer
->head_page
->page
->commit
, 0);
3949 cpu_buffer
->head_page
->read
= 0;
3951 cpu_buffer
->tail_page
= cpu_buffer
->head_page
;
3952 cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
3954 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
3955 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
3956 local_set(&cpu_buffer
->reader_page
->write
, 0);
3957 local_set(&cpu_buffer
->reader_page
->entries
, 0);
3958 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
3959 cpu_buffer
->reader_page
->read
= 0;
3961 local_set(&cpu_buffer
->entries_bytes
, 0);
3962 local_set(&cpu_buffer
->overrun
, 0);
3963 local_set(&cpu_buffer
->commit_overrun
, 0);
3964 local_set(&cpu_buffer
->dropped_events
, 0);
3965 local_set(&cpu_buffer
->entries
, 0);
3966 local_set(&cpu_buffer
->committing
, 0);
3967 local_set(&cpu_buffer
->commits
, 0);
3968 cpu_buffer
->read
= 0;
3969 cpu_buffer
->read_bytes
= 0;
3971 cpu_buffer
->write_stamp
= 0;
3972 cpu_buffer
->read_stamp
= 0;
3974 cpu_buffer
->lost_events
= 0;
3975 cpu_buffer
->last_overrun
= 0;
3977 rb_head_page_activate(cpu_buffer
);
3981 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3982 * @buffer: The ring buffer to reset a per cpu buffer of
3983 * @cpu: The CPU buffer to be reset
3985 void ring_buffer_reset_cpu(struct ring_buffer
*buffer
, int cpu
)
3987 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
3988 unsigned long flags
;
3990 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3993 atomic_inc(&buffer
->resize_disabled
);
3994 atomic_inc(&cpu_buffer
->record_disabled
);
3996 /* Make sure all commits have finished */
3997 synchronize_sched();
3999 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4001 if (RB_WARN_ON(cpu_buffer
, local_read(&cpu_buffer
->committing
)))
4004 arch_spin_lock(&cpu_buffer
->lock
);
4006 rb_reset_cpu(cpu_buffer
);
4008 arch_spin_unlock(&cpu_buffer
->lock
);
4011 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4013 atomic_dec(&cpu_buffer
->record_disabled
);
4014 atomic_dec(&buffer
->resize_disabled
);
4016 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu
);
4019 * ring_buffer_reset - reset a ring buffer
4020 * @buffer: The ring buffer to reset all cpu buffers
4022 void ring_buffer_reset(struct ring_buffer
*buffer
)
4026 for_each_buffer_cpu(buffer
, cpu
)
4027 ring_buffer_reset_cpu(buffer
, cpu
);
4029 EXPORT_SYMBOL_GPL(ring_buffer_reset
);
4032 * rind_buffer_empty - is the ring buffer empty?
4033 * @buffer: The ring buffer to test
4035 int ring_buffer_empty(struct ring_buffer
*buffer
)
4037 struct ring_buffer_per_cpu
*cpu_buffer
;
4038 unsigned long flags
;
4043 dolock
= rb_ok_to_lock();
4045 /* yes this is racy, but if you don't like the race, lock the buffer */
4046 for_each_buffer_cpu(buffer
, cpu
) {
4047 cpu_buffer
= buffer
->buffers
[cpu
];
4048 local_irq_save(flags
);
4050 raw_spin_lock(&cpu_buffer
->reader_lock
);
4051 ret
= rb_per_cpu_empty(cpu_buffer
);
4053 raw_spin_unlock(&cpu_buffer
->reader_lock
);
4054 local_irq_restore(flags
);
4062 EXPORT_SYMBOL_GPL(ring_buffer_empty
);
4065 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4066 * @buffer: The ring buffer
4067 * @cpu: The CPU buffer to test
4069 int ring_buffer_empty_cpu(struct ring_buffer
*buffer
, int cpu
)
4071 struct ring_buffer_per_cpu
*cpu_buffer
;
4072 unsigned long flags
;
4076 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4079 dolock
= rb_ok_to_lock();
4081 cpu_buffer
= buffer
->buffers
[cpu
];
4082 local_irq_save(flags
);
4084 raw_spin_lock(&cpu_buffer
->reader_lock
);
4085 ret
= rb_per_cpu_empty(cpu_buffer
);
4087 raw_spin_unlock(&cpu_buffer
->reader_lock
);
4088 local_irq_restore(flags
);
4092 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu
);
4094 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4096 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4097 * @buffer_a: One buffer to swap with
4098 * @buffer_b: The other buffer to swap with
4100 * This function is useful for tracers that want to take a "snapshot"
4101 * of a CPU buffer and has another back up buffer lying around.
4102 * it is expected that the tracer handles the cpu buffer not being
4103 * used at the moment.
4105 int ring_buffer_swap_cpu(struct ring_buffer
*buffer_a
,
4106 struct ring_buffer
*buffer_b
, int cpu
)
4108 struct ring_buffer_per_cpu
*cpu_buffer_a
;
4109 struct ring_buffer_per_cpu
*cpu_buffer_b
;
4112 if (!cpumask_test_cpu(cpu
, buffer_a
->cpumask
) ||
4113 !cpumask_test_cpu(cpu
, buffer_b
->cpumask
))
4116 cpu_buffer_a
= buffer_a
->buffers
[cpu
];
4117 cpu_buffer_b
= buffer_b
->buffers
[cpu
];
4119 /* At least make sure the two buffers are somewhat the same */
4120 if (cpu_buffer_a
->nr_pages
!= cpu_buffer_b
->nr_pages
)
4125 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
4128 if (atomic_read(&buffer_a
->record_disabled
))
4131 if (atomic_read(&buffer_b
->record_disabled
))
4134 if (atomic_read(&cpu_buffer_a
->record_disabled
))
4137 if (atomic_read(&cpu_buffer_b
->record_disabled
))
4141 * We can't do a synchronize_sched here because this
4142 * function can be called in atomic context.
4143 * Normally this will be called from the same CPU as cpu.
4144 * If not it's up to the caller to protect this.
4146 atomic_inc(&cpu_buffer_a
->record_disabled
);
4147 atomic_inc(&cpu_buffer_b
->record_disabled
);
4150 if (local_read(&cpu_buffer_a
->committing
))
4152 if (local_read(&cpu_buffer_b
->committing
))
4155 buffer_a
->buffers
[cpu
] = cpu_buffer_b
;
4156 buffer_b
->buffers
[cpu
] = cpu_buffer_a
;
4158 cpu_buffer_b
->buffer
= buffer_a
;
4159 cpu_buffer_a
->buffer
= buffer_b
;
4164 atomic_dec(&cpu_buffer_a
->record_disabled
);
4165 atomic_dec(&cpu_buffer_b
->record_disabled
);
4169 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu
);
4170 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4173 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4174 * @buffer: the buffer to allocate for.
4176 * This function is used in conjunction with ring_buffer_read_page.
4177 * When reading a full page from the ring buffer, these functions
4178 * can be used to speed up the process. The calling function should
4179 * allocate a few pages first with this function. Then when it
4180 * needs to get pages from the ring buffer, it passes the result
4181 * of this function into ring_buffer_read_page, which will swap
4182 * the page that was allocated, with the read page of the buffer.
4185 * The page allocated, or NULL on error.
4187 void *ring_buffer_alloc_read_page(struct ring_buffer
*buffer
, int cpu
)
4189 struct buffer_data_page
*bpage
;
4192 page
= alloc_pages_node(cpu_to_node(cpu
),
4193 GFP_KERNEL
| __GFP_NORETRY
, 0);
4197 bpage
= page_address(page
);
4199 rb_init_page(bpage
);
4203 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page
);
4206 * ring_buffer_free_read_page - free an allocated read page
4207 * @buffer: the buffer the page was allocate for
4208 * @data: the page to free
4210 * Free a page allocated from ring_buffer_alloc_read_page.
4212 void ring_buffer_free_read_page(struct ring_buffer
*buffer
, void *data
)
4214 free_page((unsigned long)data
);
4216 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page
);
4219 * ring_buffer_read_page - extract a page from the ring buffer
4220 * @buffer: buffer to extract from
4221 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4222 * @len: amount to extract
4223 * @cpu: the cpu of the buffer to extract
4224 * @full: should the extraction only happen when the page is full.
4226 * This function will pull out a page from the ring buffer and consume it.
4227 * @data_page must be the address of the variable that was returned
4228 * from ring_buffer_alloc_read_page. This is because the page might be used
4229 * to swap with a page in the ring buffer.
4232 * rpage = ring_buffer_alloc_read_page(buffer);
4235 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4237 * process_page(rpage, ret);
4239 * When @full is set, the function will not return true unless
4240 * the writer is off the reader page.
4242 * Note: it is up to the calling functions to handle sleeps and wakeups.
4243 * The ring buffer can be used anywhere in the kernel and can not
4244 * blindly call wake_up. The layer that uses the ring buffer must be
4245 * responsible for that.
4248 * >=0 if data has been transferred, returns the offset of consumed data.
4249 * <0 if no data has been transferred.
4251 int ring_buffer_read_page(struct ring_buffer
*buffer
,
4252 void **data_page
, size_t len
, int cpu
, int full
)
4254 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4255 struct ring_buffer_event
*event
;
4256 struct buffer_data_page
*bpage
;
4257 struct buffer_page
*reader
;
4258 unsigned long missed_events
;
4259 unsigned long flags
;
4260 unsigned int commit
;
4265 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4269 * If len is not big enough to hold the page header, then
4270 * we can not copy anything.
4272 if (len
<= BUF_PAGE_HDR_SIZE
)
4275 len
-= BUF_PAGE_HDR_SIZE
;
4284 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4286 reader
= rb_get_reader_page(cpu_buffer
);
4290 event
= rb_reader_event(cpu_buffer
);
4292 read
= reader
->read
;
4293 commit
= rb_page_commit(reader
);
4295 /* Check if any events were dropped */
4296 missed_events
= cpu_buffer
->lost_events
;
4299 * If this page has been partially read or
4300 * if len is not big enough to read the rest of the page or
4301 * a writer is still on the page, then
4302 * we must copy the data from the page to the buffer.
4303 * Otherwise, we can simply swap the page with the one passed in.
4305 if (read
|| (len
< (commit
- read
)) ||
4306 cpu_buffer
->reader_page
== cpu_buffer
->commit_page
) {
4307 struct buffer_data_page
*rpage
= cpu_buffer
->reader_page
->page
;
4308 unsigned int rpos
= read
;
4309 unsigned int pos
= 0;
4315 if (len
> (commit
- read
))
4316 len
= (commit
- read
);
4318 /* Always keep the time extend and data together */
4319 size
= rb_event_ts_length(event
);
4324 /* save the current timestamp, since the user will need it */
4325 save_timestamp
= cpu_buffer
->read_stamp
;
4327 /* Need to copy one event at a time */
4329 /* We need the size of one event, because
4330 * rb_advance_reader only advances by one event,
4331 * whereas rb_event_ts_length may include the size of
4332 * one or two events.
4333 * We have already ensured there's enough space if this
4334 * is a time extend. */
4335 size
= rb_event_length(event
);
4336 memcpy(bpage
->data
+ pos
, rpage
->data
+ rpos
, size
);
4340 rb_advance_reader(cpu_buffer
);
4341 rpos
= reader
->read
;
4347 event
= rb_reader_event(cpu_buffer
);
4348 /* Always keep the time extend and data together */
4349 size
= rb_event_ts_length(event
);
4350 } while (len
>= size
);
4353 local_set(&bpage
->commit
, pos
);
4354 bpage
->time_stamp
= save_timestamp
;
4356 /* we copied everything to the beginning */
4359 /* update the entry counter */
4360 cpu_buffer
->read
+= rb_page_entries(reader
);
4361 cpu_buffer
->read_bytes
+= BUF_PAGE_SIZE
;
4363 /* swap the pages */
4364 rb_init_page(bpage
);
4365 bpage
= reader
->page
;
4366 reader
->page
= *data_page
;
4367 local_set(&reader
->write
, 0);
4368 local_set(&reader
->entries
, 0);
4373 * Use the real_end for the data size,
4374 * This gives us a chance to store the lost events
4377 if (reader
->real_end
)
4378 local_set(&bpage
->commit
, reader
->real_end
);
4382 cpu_buffer
->lost_events
= 0;
4384 commit
= local_read(&bpage
->commit
);
4386 * Set a flag in the commit field if we lost events
4388 if (missed_events
) {
4389 /* If there is room at the end of the page to save the
4390 * missed events, then record it there.
4392 if (BUF_PAGE_SIZE
- commit
>= sizeof(missed_events
)) {
4393 memcpy(&bpage
->data
[commit
], &missed_events
,
4394 sizeof(missed_events
));
4395 local_add(RB_MISSED_STORED
, &bpage
->commit
);
4396 commit
+= sizeof(missed_events
);
4398 local_add(RB_MISSED_EVENTS
, &bpage
->commit
);
4402 * This page may be off to user land. Zero it out here.
4404 if (commit
< BUF_PAGE_SIZE
)
4405 memset(&bpage
->data
[commit
], 0, BUF_PAGE_SIZE
- commit
);
4408 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4413 EXPORT_SYMBOL_GPL(ring_buffer_read_page
);
4415 #ifdef CONFIG_HOTPLUG_CPU
4416 static int rb_cpu_notify(struct notifier_block
*self
,
4417 unsigned long action
, void *hcpu
)
4419 struct ring_buffer
*buffer
=
4420 container_of(self
, struct ring_buffer
, cpu_notify
);
4421 long cpu
= (long)hcpu
;
4422 int cpu_i
, nr_pages_same
;
4423 unsigned int nr_pages
;
4426 case CPU_UP_PREPARE
:
4427 case CPU_UP_PREPARE_FROZEN
:
4428 if (cpumask_test_cpu(cpu
, buffer
->cpumask
))
4433 /* check if all cpu sizes are same */
4434 for_each_buffer_cpu(buffer
, cpu_i
) {
4435 /* fill in the size from first enabled cpu */
4437 nr_pages
= buffer
->buffers
[cpu_i
]->nr_pages
;
4438 if (nr_pages
!= buffer
->buffers
[cpu_i
]->nr_pages
) {
4443 /* allocate minimum pages, user can later expand it */
4446 buffer
->buffers
[cpu
] =
4447 rb_allocate_cpu_buffer(buffer
, nr_pages
, cpu
);
4448 if (!buffer
->buffers
[cpu
]) {
4449 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4454 cpumask_set_cpu(cpu
, buffer
->cpumask
);
4456 case CPU_DOWN_PREPARE
:
4457 case CPU_DOWN_PREPARE_FROZEN
:
4460 * If we were to free the buffer, then the user would
4461 * lose any trace that was in the buffer.