1 Lockless Ring Buffer Design
2 ===========================
4 Copyright 2009 Red Hat Inc.
5 Author: Steven Rostedt <srostedt@redhat.com>
6 License: The GNU Free Documentation License, Version 1.2
7 (dual licensed under the GPL v2)
8 Reviewers: Mathieu Desnoyers, Huang Ying, Hidetoshi Seto,
9 and Frederic Weisbecker.
14 Terminology used in this Document
15 ---------------------------------
17 tail - where new writes happen in the ring buffer.
19 head - where new reads happen in the ring buffer.
21 producer - the task that writes into the ring buffer (same as writer)
23 writer - same as producer
25 consumer - the task that reads from the buffer (same as reader)
27 reader - same as consumer.
29 reader_page - A page outside the ring buffer used solely (for the most part)
32 head_page - a pointer to the page that the reader will use next
34 tail_page - a pointer to the page that will be written to next
36 commit_page - a pointer to the page with the last finished non-nested write.
38 cmpxchg - hardware-assisted atomic transaction that performs the following:
40 A = B iff previous A == C
42 R = cmpxchg(A, C, B) is saying that we replace A with B if and only if
43 current A is equal to C, and we put the old (current) A into R
45 R gets the previous A regardless if A is updated with B or not.
47 To see if the update was successful a compare of R == C may be used.
49 The Generic Ring Buffer
50 -----------------------
52 The ring buffer can be used in either an overwrite mode or in
53 producer/consumer mode.
55 Producer/consumer mode is where if the producer were to fill up the
56 buffer before the consumer could free up anything, the producer
57 will stop writing to the buffer. This will lose most recent events.
59 Overwrite mode is where if the producer were to fill up the buffer
60 before the consumer could free up anything, the producer will
61 overwrite the older data. This will lose the oldest events.
63 No two writers can write at the same time (on the same per-cpu buffer),
64 but a writer may interrupt another writer, but it must finish writing
65 before the previous writer may continue. This is very important to the
66 algorithm. The writers act like a "stack". The way interrupts works
67 enforces this behavior.
71 <preempted> writer2 start
72 <preempted> writer3 start
77 This is very much like a writer being preempted by an interrupt and
78 the interrupt doing a write as well.
80 Readers can happen at any time. But no two readers may run at the
81 same time, nor can a reader preempt/interrupt another reader. A reader
82 cannot preempt/interrupt a writer, but it may read/consume from the
83 buffer at the same time as a writer is writing, but the reader must be
84 on another processor to do so. A reader may read on its own processor
85 and can be preempted by a writer.
87 A writer can preempt a reader, but a reader cannot preempt a writer.
88 But a reader can read the buffer at the same time (on another processor)
91 The ring buffer is made up of a list of pages held together by a linked list.
93 At initialization a reader page is allocated for the reader that is not
94 part of the ring buffer.
96 The head_page, tail_page and commit_page are all initialized to point
99 The reader page is initialized to have its next pointer pointing to
100 the head page, and its previous pointer pointing to a page before
103 The reader has its own page to use. At start up time, this page is
104 allocated but is not attached to the list. When the reader wants
105 to read from the buffer, if its page is empty (like it is on start-up),
106 it will swap its page with the head_page. The old reader page will
107 become part of the ring buffer and the head_page will be removed.
108 The page after the inserted page (old reader_page) will become the
111 Once the new page is given to the reader, the reader could do what
112 it wants with it, as long as a writer has left that page.
114 A sample of how the reader page is swapped: Note this does not
115 show the head page in the buffer, it is for demonstrating a swap
133 |page |-------------------+
138 | +---+ +---+ +---+ |
140 | | +-------------+ | |
141 | +-----------------+ |
142 +------------------------------------+
146 |page |-------------------+
147 +------+ <---------------+ v
148 | ^ +---+ +---+ +---+
151 | | +---+ +---+ +---+ |
153 | | +-------------+ | |
154 | +-----------------------------+ |
155 +------------------------------------+
159 |page |-------------------+
160 +------+ <---------------+ v
161 | ^ +---+ +---+ +---+
163 | | New | | | |<--| |<-+
164 | | Reader +---+ +---+ +---+ |
167 | +-----------------------------+ |
168 +------------------------------------+
172 It is possible that the page swapped is the commit page and the tail page,
173 if what is in the ring buffer is less than what is held in a buffer page.
176 reader page commit page tail page
181 | |<------------------------+
186 +---+ +---+ +---+ +---+
187 <---| |--->| |--->| |--->| |--->
188 --->| |<---| |<---| |<---| |<---
189 +---+ +---+ +---+ +---+
191 This case is still valid for this algorithm.
192 When the writer leaves the page, it simply goes into the ring buffer
193 since the reader page still points to the next location in the ring
199 reader page - The page used solely by the reader and is not part
200 of the ring buffer (may be swapped in)
202 head page - the next page in the ring buffer that will be swapped
203 with the reader page.
205 tail page - the page where the next write will take place.
207 commit page - the page that last finished a write.
209 The commit page only is updated by the outermost writer in the
210 writer stack. A writer that preempts another writer will not move the
213 When data is written into the ring buffer, a position is reserved
214 in the ring buffer and passed back to the writer. When the writer
215 is finished writing data into that position, it commits the write.
217 Another write (or a read) may take place at anytime during this
218 transaction. If another write happens it must finish before continuing
219 with the previous write.
227 +---------+ <--- given back to writer (current commit)
229 +---------+ <--- tail pointer
240 +---------+ <--- next position for write (current commit)
245 If a write happens after the first reserve:
250 +---------+ <-- current commit
252 +---------+ <--- given back to second writer
254 +---------+ <--- tail pointer
256 After second writer commits:
262 +---------+ <--(last full commit)
267 +---------+ <--- tail pointer
269 When the first writer commits:
278 +---------+ <--(last full commit and tail pointer)
281 The commit pointer points to the last write location that was
282 committed without preempting another write. When a write that
283 preempted another write is committed, it only becomes a pending commit
284 and will not be a full commit until all writes have been committed.
286 The commit page points to the page that has the last full commit.
287 The tail page points to the page with the last write (before
290 The tail page is always equal to or after the commit page. It may
291 be several pages ahead. If the tail page catches up to the commit
292 page then no more writes may take place (regardless of the mode
293 of the ring buffer: overwrite and produce/consumer).
295 The order of pages is:
303 head page commit page |
306 +---+ +---+ +---+ +---+
307 <---| |--->| |--->| |--->| |--->
308 --->| |<---| |<---| |<---| |<---
309 +---+ +---+ +---+ +---+
311 There is a special case that the head page is after either the commit page
312 and possibly the tail page. That is when the commit (and tail) page has been
313 swapped with the reader page. This is because the head page is always
314 part of the ring buffer, but the reader page is not. Whenever there
315 has been less than a full page that has been committed inside the ring buffer,
316 and a reader swaps out a page, it will be swapping out the commit page.
319 reader page commit page tail page
324 | |<------------------------+
329 +---+ +---+ +---+ +---+
330 <---| |--->| |--->| |--->| |--->
331 --->| |<---| |<---| |<---| |<---
332 +---+ +---+ +---+ +---+
338 In this case, the head page will not move when the tail and commit
339 move back into the ring buffer.
341 The reader cannot swap a page into the ring buffer if the commit page
342 is still on that page. If the read meets the last commit (real commit
343 not pending or reserved), then there is nothing more to read.
344 The buffer is considered empty until another full commit finishes.
346 When the tail meets the head page, if the buffer is in overwrite mode,
347 the head page will be pushed ahead one. If the buffer is in producer/consumer
348 mode, the write will fail.
355 +---+ +---+ +---+ +---+
356 <---| |--->| |--->| |--->| |--->
357 --->| |<---| |<---| |<---| |<---
358 +---+ +---+ +---+ +---+
367 +---+ +---+ +---+ +---+
368 <---| |--->| |--->| |--->| |--->
369 --->| |<---| |<---| |<---| |<---
370 +---+ +---+ +---+ +---+
379 +---+ +---+ +---+ +---+
380 <---| |--->| |--->| |--->| |--->
381 --->| |<---| |<---| |<---| |<---
382 +---+ +---+ +---+ +---+
387 Note, the reader page will still point to the previous head page.
388 But when a swap takes place, it will use the most recent head page.
391 Making the Ring Buffer Lockless:
392 --------------------------------
394 The main idea behind the lockless algorithm is to combine the moving
395 of the head_page pointer with the swapping of pages with the reader.
396 State flags are placed inside the pointer to the page. To do this,
397 each page must be aligned in memory by 4 bytes. This will allow the 2
398 least significant bits of the address to be used as flags, since
399 they will always be zero for the address. To get the address,
400 simply mask out the flags.
406 Two flags will be kept by these two bits:
408 HEADER - the page being pointed to is a head page
410 UPDATE - the page being pointed to is being updated by a writer
411 and was or is about to be a head page.
422 +---+ +---+ +---+ +---+
423 <---| |--->| |-H->| |--->| |--->
424 --->| |<---| |<---| |<---| |<---
425 +---+ +---+ +---+ +---+
428 The above pointer "-H->" would have the HEADER flag set. That is
429 the next page is the next page to be swapped out by the reader.
430 This pointer means the next page is the head page.
432 When the tail page meets the head pointer, it will use cmpxchg to
433 change the pointer to the UPDATE state:
439 +---+ +---+ +---+ +---+
440 <---| |--->| |-H->| |--->| |--->
441 --->| |<---| |<---| |<---| |<---
442 +---+ +---+ +---+ +---+
447 +---+ +---+ +---+ +---+
448 <---| |--->| |-U->| |--->| |--->
449 --->| |<---| |<---| |<---| |<---
450 +---+ +---+ +---+ +---+
452 "-U->" represents a pointer in the UPDATE state.
454 Any access to the reader will need to take some sort of lock to serialize
455 the readers. But the writers will never take a lock to write to the
456 ring buffer. This means we only need to worry about a single reader,
457 and writes only preempt in "stack" formation.
459 When the reader tries to swap the page with the ring buffer, it
460 will also use cmpxchg. If the flag bit in the pointer to the
461 head page does not have the HEADER flag set, the compare will fail
462 and the reader will need to look for the new head page and try again.
463 Note, the flags UPDATE and HEADER are never set at the same time.
465 The reader swaps the reader page as follows:
476 | +---------------+ |
477 +-----H-------------+
479 The reader sets the reader page next pointer as HEADER to the page after
485 |page |-------H-----------+
489 | | |<---| |<---| |<-+
490 | +---+ +---+ +---+ |
492 | | +---------------+ | |
493 | +-----H-------------+ |
494 +--------------------------------------+
496 It does a cmpxchg with the pointer to the previous head page to make it
497 point to the reader page. Note that the new pointer does not have the HEADER
498 flag set. This action atomically moves the head page forward.
502 |page |-------H-----------+
504 | ^ +---+ +---+ +---+
506 | | | |<--| |<--| |<-+
507 | | +---+ +---+ +---+ |
509 | | +-------------+ | |
510 | +-----------------------------+ |
511 +------------------------------------+
513 After the new head page is set, the previous pointer of the head page is
514 updated to the reader page.
518 |page |-------H-----------+
519 +------+ <---------------+ v
520 | ^ +---+ +---+ +---+
523 | | +---+ +---+ +---+ |
525 | | +-------------+ | |
526 | +-----------------------------+ |
527 +------------------------------------+
531 |page |-------H-----------+ <--- New head page
532 +------+ <---------------+ v
533 | ^ +---+ +---+ +---+
535 | | New | | | |<--| |<-+
536 | | Reader +---+ +---+ +---+ |
539 | +-----------------------------+ |
540 +------------------------------------+
542 Another important point: The page that the reader page points back to
543 by its previous pointer (the one that now points to the new head page)
544 never points back to the reader page. That is because the reader page is
545 not part of the ring buffer. Traversing the ring buffer via the next pointers
546 will always stay in the ring buffer. Traversing the ring buffer via the
547 prev pointers may not.
549 Note, the way to determine a reader page is simply by examining the previous
550 pointer of the page. If the next pointer of the previous page does not
551 point back to the original page, then the original page is a reader page:
555 | reader | next +----+
556 | page |-------->| |<====== (buffer page)
564 The way the head page moves forward:
566 When the tail page meets the head page and the buffer is in overwrite mode
567 and more writes take place, the head page must be moved forward before the
568 writer may move the tail page. The way this is done is that the writer
569 performs a cmpxchg to convert the pointer to the head page from the HEADER
570 flag to have the UPDATE flag set. Once this is done, the reader will
571 not be able to swap the head page from the buffer, nor will it be able to
572 move the head page, until the writer is finished with the move.
574 This eliminates any races that the reader can have on the writer. The reader
575 must spin, and this is why the reader cannot preempt the writer.
580 +---+ +---+ +---+ +---+
581 <---| |--->| |-H->| |--->| |--->
582 --->| |<---| |<---| |<---| |<---
583 +---+ +---+ +---+ +---+
588 +---+ +---+ +---+ +---+
589 <---| |--->| |-U->| |--->| |--->
590 --->| |<---| |<---| |<---| |<---
591 +---+ +---+ +---+ +---+
593 The following page will be made into the new head page.
598 +---+ +---+ +---+ +---+
599 <---| |--->| |-U->| |-H->| |--->
600 --->| |<---| |<---| |<---| |<---
601 +---+ +---+ +---+ +---+
603 After the new head page has been set, we can set the old head page
604 pointer back to NORMAL.
609 +---+ +---+ +---+ +---+
610 <---| |--->| |--->| |-H->| |--->
611 --->| |<---| |<---| |<---| |<---
612 +---+ +---+ +---+ +---+
614 After the head page has been moved, the tail page may now move forward.
619 +---+ +---+ +---+ +---+
620 <---| |--->| |--->| |-H->| |--->
621 --->| |<---| |<---| |<---| |<---
622 +---+ +---+ +---+ +---+
625 The above are the trivial updates. Now for the more complex scenarios.
628 As stated before, if enough writes preempt the first write, the
629 tail page may make it all the way around the buffer and meet the commit
630 page. At this time, we must start dropping writes (usually with some kind
631 of warning to the user). But what happens if the commit was still on the
632 reader page? The commit page is not part of the ring buffer. The tail page
633 must account for this.
636 reader page commit page
646 +---+ +---+ +---+ +---+
647 <---| |--->| |-H->| |--->| |--->
648 --->| |<---| |<---| |<---| |<---
649 +---+ +---+ +---+ +---+
654 If the tail page were to simply push the head page forward, the commit when
655 leaving the reader page would not be pointing to the correct page.
657 The solution to this is to test if the commit page is on the reader page
658 before pushing the head page. If it is, then it can be assumed that the
659 tail page wrapped the buffer, and we must drop new writes.
661 This is not a race condition, because the commit page can only be moved
662 by the outermost writer (the writer that was preempted).
663 This means that the commit will not move while a writer is moving the
664 tail page. The reader cannot swap the reader page if it is also being
665 used as the commit page. The reader can simply check that the commit
666 is off the reader page. Once the commit page leaves the reader page
667 it will never go back on it unless a reader does another swap with the
668 buffer page that is also the commit page.
674 In the pushing forward of the tail page we must first push forward
675 the head page if the head page is the next page. If the head page
676 is not the next page, the tail page is simply updated with a cmpxchg.
678 Only writers move the tail page. This must be done atomically to protect
679 against nested writers.
681 temp_page = tail_page
682 next_page = temp_page->next
683 cmpxchg(tail_page, temp_page, next_page)
685 The above will update the tail page if it is still pointing to the expected
686 page. If this fails, a nested write pushed it forward, the the current write
687 does not need to push it.
696 +---+ +---+ +---+ +---+
697 <---| |--->| |--->| |--->| |--->
698 --->| |<---| |<---| |<---| |<---
699 +---+ +---+ +---+ +---+
701 Nested write comes in and moves the tail page forward:
703 tail page (moved by nested writer)
707 +---+ +---+ +---+ +---+
708 <---| |--->| |--->| |--->| |--->
709 --->| |<---| |<---| |<---| |<---
710 +---+ +---+ +---+ +---+
712 The above would fail the cmpxchg, but since the tail page has already
713 been moved forward, the writer will just try again to reserve storage
714 on the new tail page.
716 But the moving of the head page is a bit more complex.
721 +---+ +---+ +---+ +---+
722 <---| |--->| |-H->| |--->| |--->
723 --->| |<---| |<---| |<---| |<---
724 +---+ +---+ +---+ +---+
726 The write converts the head page pointer to UPDATE.
731 +---+ +---+ +---+ +---+
732 <---| |--->| |-U->| |--->| |--->
733 --->| |<---| |<---| |<---| |<---
734 +---+ +---+ +---+ +---+
736 But if a nested writer preempts here, it will see that the next
737 page is a head page, but it is also nested. It will detect that
738 it is nested and will save that information. The detection is the
739 fact that it sees the UPDATE flag instead of a HEADER or NORMAL
742 The nested writer will set the new head page pointer.
747 +---+ +---+ +---+ +---+
748 <---| |--->| |-U->| |-H->| |--->
749 --->| |<---| |<---| |<---| |<---
750 +---+ +---+ +---+ +---+
752 But it will not reset the update back to normal. Only the writer
753 that converted a pointer from HEAD to UPDATE will convert it back
759 +---+ +---+ +---+ +---+
760 <---| |--->| |-U->| |-H->| |--->
761 --->| |<---| |<---| |<---| |<---
762 +---+ +---+ +---+ +---+
764 After the nested writer finishes, the outermost writer will convert
765 the UPDATE pointer to NORMAL.
771 +---+ +---+ +---+ +---+
772 <---| |--->| |--->| |-H->| |--->
773 --->| |<---| |<---| |<---| |<---
774 +---+ +---+ +---+ +---+
777 It can be even more complex if several nested writes came in and moved
778 the tail page ahead several pages:
786 +---+ +---+ +---+ +---+
787 <---| |--->| |-H->| |--->| |--->
788 --->| |<---| |<---| |<---| |<---
789 +---+ +---+ +---+ +---+
791 The write converts the head page pointer to UPDATE.
796 +---+ +---+ +---+ +---+
797 <---| |--->| |-U->| |--->| |--->
798 --->| |<---| |<---| |<---| |<---
799 +---+ +---+ +---+ +---+
801 Next writer comes in, and sees the update and sets up the new
809 +---+ +---+ +---+ +---+
810 <---| |--->| |-U->| |-H->| |--->
811 --->| |<---| |<---| |<---| |<---
812 +---+ +---+ +---+ +---+
814 The nested writer moves the tail page forward. But does not set the old
815 update page to NORMAL because it is not the outermost writer.
820 +---+ +---+ +---+ +---+
821 <---| |--->| |-U->| |-H->| |--->
822 --->| |<---| |<---| |<---| |<---
823 +---+ +---+ +---+ +---+
825 Another writer preempts and sees the page after the tail page is a head page.
826 It changes it from HEAD to UPDATE.
833 +---+ +---+ +---+ +---+
834 <---| |--->| |-U->| |-U->| |--->
835 --->| |<---| |<---| |<---| |<---
836 +---+ +---+ +---+ +---+
838 The writer will move the head page forward:
846 +---+ +---+ +---+ +---+
847 <---| |--->| |-U->| |-U->| |-H->
848 --->| |<---| |<---| |<---| |<---
849 +---+ +---+ +---+ +---+
851 But now that the third writer did change the HEAD flag to UPDATE it
852 will convert it to normal:
860 +---+ +---+ +---+ +---+
861 <---| |--->| |-U->| |--->| |-H->
862 --->| |<---| |<---| |<---| |<---
863 +---+ +---+ +---+ +---+
866 Then it will move the tail page, and return back to the second writer.
874 +---+ +---+ +---+ +---+
875 <---| |--->| |-U->| |--->| |-H->
876 --->| |<---| |<---| |<---| |<---
877 +---+ +---+ +---+ +---+
880 The second writer will fail to move the tail page because it was already
881 moved, so it will try again and add its data to the new tail page.
882 It will return to the first writer.
890 +---+ +---+ +---+ +---+
891 <---| |--->| |-U->| |--->| |-H->
892 --->| |<---| |<---| |<---| |<---
893 +---+ +---+ +---+ +---+
895 The first writer cannot know atomically if the tail page moved
896 while it updates the HEAD page. It will then update the head page to
897 what it thinks is the new head page.
905 +---+ +---+ +---+ +---+
906 <---| |--->| |-U->| |-H->| |-H->
907 --->| |<---| |<---| |<---| |<---
908 +---+ +---+ +---+ +---+
910 Since the cmpxchg returns the old value of the pointer the first writer
911 will see it succeeded in updating the pointer from NORMAL to HEAD.
912 But as we can see, this is not good enough. It must also check to see
913 if the tail page is either where it use to be or on the next page:
921 +---+ +---+ +---+ +---+
922 <---| |--->| |-U->| |-H->| |-H->
923 --->| |<---| |<---| |<---| |<---
924 +---+ +---+ +---+ +---+
926 If tail page != A and tail page != B, then it must reset the pointer
927 back to NORMAL. The fact that it only needs to worry about nested
928 writers means that it only needs to check this after setting the HEAD page.
936 +---+ +---+ +---+ +---+
937 <---| |--->| |-U->| |--->| |-H->
938 --->| |<---| |<---| |<---| |<---
939 +---+ +---+ +---+ +---+
941 Now the writer can update the head page. This is also why the head page must
942 remain in UPDATE and only reset by the outermost writer. This prevents
943 the reader from seeing the incorrect head page.
951 +---+ +---+ +---+ +---+
952 <---| |--->| |--->| |--->| |-H->
953 --->| |<---| |<---| |<---| |<---
954 +---+ +---+ +---+ +---+