cachepc-linux

Fork of AMDESE/linux with modifications for CachePC side-channel attack
git clone https://git.sinitax.com/sinitax/cachepc-linux
Log | Files | Refs | README | LICENSE | sfeed.txt

ring-buffer-design.rst (31321B)


      1.. SPDX-License-Identifier: GPL-2.0 OR GFDL-1.2-no-invariants-only
      2
      3===========================
      4Lockless Ring Buffer Design
      5===========================
      6
      7Copyright 2009 Red Hat Inc.
      8
      9:Author:   Steven Rostedt <srostedt@redhat.com>
     10:License:  The GNU Free Documentation License, Version 1.2
     11           (dual licensed under the GPL v2)
     12:Reviewers:  Mathieu Desnoyers, Huang Ying, Hidetoshi Seto,
     13	     and Frederic Weisbecker.
     14
     15
     16Written for: 2.6.31
     17
     18Terminology used in this Document
     19---------------------------------
     20
     21tail
     22	- where new writes happen in the ring buffer.
     23
     24head
     25	- where new reads happen in the ring buffer.
     26
     27producer
     28	- the task that writes into the ring buffer (same as writer)
     29
     30writer
     31	- same as producer
     32
     33consumer
     34	- the task that reads from the buffer (same as reader)
     35
     36reader
     37	- same as consumer.
     38
     39reader_page
     40	- A page outside the ring buffer used solely (for the most part)
     41	  by the reader.
     42
     43head_page
     44	- a pointer to the page that the reader will use next
     45
     46tail_page
     47	- a pointer to the page that will be written to next
     48
     49commit_page
     50	- a pointer to the page with the last finished non-nested write.
     51
     52cmpxchg
     53	- hardware-assisted atomic transaction that performs the following::
     54
     55	    A = B if previous A == C
     56
     57	    R = cmpxchg(A, C, B) is saying that we replace A with B if and only
     58		if current A is equal to C, and we put the old (current)
     59		A into R
     60
     61	    R gets the previous A regardless if A is updated with B or not.
     62
     63	  To see if the update was successful a compare of ``R == C``
     64	  may be used.
     65
     66The Generic Ring Buffer
     67-----------------------
     68
     69The ring buffer can be used in either an overwrite mode or in
     70producer/consumer mode.
     71
     72Producer/consumer mode is where if the producer were to fill up the
     73buffer before the consumer could free up anything, the producer
     74will stop writing to the buffer. This will lose most recent events.
     75
     76Overwrite mode is where if the producer were to fill up the buffer
     77before the consumer could free up anything, the producer will
     78overwrite the older data. This will lose the oldest events.
     79
     80No two writers can write at the same time (on the same per-cpu buffer),
     81but a writer may interrupt another writer, but it must finish writing
     82before the previous writer may continue. This is very important to the
     83algorithm. The writers act like a "stack". The way interrupts works
     84enforces this behavior::
     85
     86
     87  writer1 start
     88     <preempted> writer2 start
     89         <preempted> writer3 start
     90                     writer3 finishes
     91                 writer2 finishes
     92  writer1 finishes
     93
     94This is very much like a writer being preempted by an interrupt and
     95the interrupt doing a write as well.
     96
     97Readers can happen at any time. But no two readers may run at the
     98same time, nor can a reader preempt/interrupt another reader. A reader
     99cannot preempt/interrupt a writer, but it may read/consume from the
    100buffer at the same time as a writer is writing, but the reader must be
    101on another processor to do so. A reader may read on its own processor
    102and can be preempted by a writer.
    103
    104A writer can preempt a reader, but a reader cannot preempt a writer.
    105But a reader can read the buffer at the same time (on another processor)
    106as a writer.
    107
    108The ring buffer is made up of a list of pages held together by a linked list.
    109
    110At initialization a reader page is allocated for the reader that is not
    111part of the ring buffer.
    112
    113The head_page, tail_page and commit_page are all initialized to point
    114to the same page.
    115
    116The reader page is initialized to have its next pointer pointing to
    117the head page, and its previous pointer pointing to a page before
    118the head page.
    119
    120The reader has its own page to use. At start up time, this page is
    121allocated but is not attached to the list. When the reader wants
    122to read from the buffer, if its page is empty (like it is on start-up),
    123it will swap its page with the head_page. The old reader page will
    124become part of the ring buffer and the head_page will be removed.
    125The page after the inserted page (old reader_page) will become the
    126new head page.
    127
    128Once the new page is given to the reader, the reader could do what
    129it wants with it, as long as a writer has left that page.
    130
    131A sample of how the reader page is swapped: Note this does not
    132show the head page in the buffer, it is for demonstrating a swap
    133only.
    134
    135::
    136
    137  +------+
    138  |reader|          RING BUFFER
    139  |page  |
    140  +------+
    141                  +---+   +---+   +---+
    142                  |   |-->|   |-->|   |
    143                  |   |<--|   |<--|   |
    144                  +---+   +---+   +---+
    145                   ^ |             ^ |
    146                   | +-------------+ |
    147                   +-----------------+
    148
    149
    150  +------+
    151  |reader|          RING BUFFER
    152  |page  |-------------------+
    153  +------+                   v
    154    |             +---+   +---+   +---+
    155    |             |   |-->|   |-->|   |
    156    |             |   |<--|   |<--|   |<-+
    157    |             +---+   +---+   +---+  |
    158    |              ^ |             ^ |   |
    159    |              | +-------------+ |   |
    160    |              +-----------------+   |
    161    +------------------------------------+
    162
    163  +------+
    164  |reader|          RING BUFFER
    165  |page  |-------------------+
    166  +------+ <---------------+ v
    167    |  ^          +---+   +---+   +---+
    168    |  |          |   |-->|   |-->|   |
    169    |  |          |   |   |   |<--|   |<-+
    170    |  |          +---+   +---+   +---+  |
    171    |  |             |             ^ |   |
    172    |  |             +-------------+ |   |
    173    |  +-----------------------------+   |
    174    +------------------------------------+
    175
    176  +------+
    177  |buffer|          RING BUFFER
    178  |page  |-------------------+
    179  +------+ <---------------+ v
    180    |  ^          +---+   +---+   +---+
    181    |  |          |   |   |   |-->|   |
    182    |  |  New     |   |   |   |<--|   |<-+
    183    |  | Reader   +---+   +---+   +---+  |
    184    |  |  page ----^                 |   |
    185    |  |                             |   |
    186    |  +-----------------------------+   |
    187    +------------------------------------+
    188
    189
    190
    191It is possible that the page swapped is the commit page and the tail page,
    192if what is in the ring buffer is less than what is held in a buffer page.
    193
    194::
    195
    196            reader page    commit page   tail page
    197                |              |             |
    198                v              |             |
    199               +---+           |             |
    200               |   |<----------+             |
    201               |   |<------------------------+
    202               |   |------+
    203               +---+      |
    204                          |
    205                          v
    206      +---+    +---+    +---+    +---+
    207  <---|   |--->|   |--->|   |--->|   |--->
    208  --->|   |<---|   |<---|   |<---|   |<---
    209      +---+    +---+    +---+    +---+
    210
    211This case is still valid for this algorithm.
    212When the writer leaves the page, it simply goes into the ring buffer
    213since the reader page still points to the next location in the ring
    214buffer.
    215
    216
    217The main pointers:
    218
    219  reader page
    220	    - The page used solely by the reader and is not part
    221              of the ring buffer (may be swapped in)
    222
    223  head page
    224	    - the next page in the ring buffer that will be swapped
    225              with the reader page.
    226
    227  tail page
    228	    - the page where the next write will take place.
    229
    230  commit page
    231	    - the page that last finished a write.
    232
    233The commit page only is updated by the outermost writer in the
    234writer stack. A writer that preempts another writer will not move the
    235commit page.
    236
    237When data is written into the ring buffer, a position is reserved
    238in the ring buffer and passed back to the writer. When the writer
    239is finished writing data into that position, it commits the write.
    240
    241Another write (or a read) may take place at anytime during this
    242transaction. If another write happens it must finish before continuing
    243with the previous write.
    244
    245
    246   Write reserve::
    247
    248       Buffer page
    249      +---------+
    250      |written  |
    251      +---------+  <--- given back to writer (current commit)
    252      |reserved |
    253      +---------+ <--- tail pointer
    254      | empty   |
    255      +---------+
    256
    257   Write commit::
    258
    259       Buffer page
    260      +---------+
    261      |written  |
    262      +---------+
    263      |written  |
    264      +---------+  <--- next position for write (current commit)
    265      | empty   |
    266      +---------+
    267
    268
    269 If a write happens after the first reserve::
    270
    271       Buffer page
    272      +---------+
    273      |written  |
    274      +---------+  <-- current commit
    275      |reserved |
    276      +---------+  <--- given back to second writer
    277      |reserved |
    278      +---------+ <--- tail pointer
    279
    280  After second writer commits::
    281
    282
    283       Buffer page
    284      +---------+
    285      |written  |
    286      +---------+  <--(last full commit)
    287      |reserved |
    288      +---------+
    289      |pending  |
    290      |commit   |
    291      +---------+ <--- tail pointer
    292
    293  When the first writer commits::
    294
    295       Buffer page
    296      +---------+
    297      |written  |
    298      +---------+
    299      |written  |
    300      +---------+
    301      |written  |
    302      +---------+  <--(last full commit and tail pointer)
    303
    304
    305The commit pointer points to the last write location that was
    306committed without preempting another write. When a write that
    307preempted another write is committed, it only becomes a pending commit
    308and will not be a full commit until all writes have been committed.
    309
    310The commit page points to the page that has the last full commit.
    311The tail page points to the page with the last write (before
    312committing).
    313
    314The tail page is always equal to or after the commit page. It may
    315be several pages ahead. If the tail page catches up to the commit
    316page then no more writes may take place (regardless of the mode
    317of the ring buffer: overwrite and produce/consumer).
    318
    319The order of pages is::
    320
    321 head page
    322 commit page
    323 tail page
    324
    325Possible scenario::
    326
    327                               tail page
    328    head page         commit page  |
    329        |                 |        |
    330        v                 v        v
    331      +---+    +---+    +---+    +---+
    332  <---|   |--->|   |--->|   |--->|   |--->
    333  --->|   |<---|   |<---|   |<---|   |<---
    334      +---+    +---+    +---+    +---+
    335
    336There is a special case that the head page is after either the commit page
    337and possibly the tail page. That is when the commit (and tail) page has been
    338swapped with the reader page. This is because the head page is always
    339part of the ring buffer, but the reader page is not. Whenever there
    340has been less than a full page that has been committed inside the ring buffer,
    341and a reader swaps out a page, it will be swapping out the commit page.
    342
    343::
    344
    345            reader page    commit page   tail page
    346                |              |             |
    347                v              |             |
    348               +---+           |             |
    349               |   |<----------+             |
    350               |   |<------------------------+
    351               |   |------+
    352               +---+      |
    353                          |
    354                          v
    355      +---+    +---+    +---+    +---+
    356  <---|   |--->|   |--->|   |--->|   |--->
    357  --->|   |<---|   |<---|   |<---|   |<---
    358      +---+    +---+    +---+    +---+
    359                          ^
    360                          |
    361                      head page
    362
    363
    364In this case, the head page will not move when the tail and commit
    365move back into the ring buffer.
    366
    367The reader cannot swap a page into the ring buffer if the commit page
    368is still on that page. If the read meets the last commit (real commit
    369not pending or reserved), then there is nothing more to read.
    370The buffer is considered empty until another full commit finishes.
    371
    372When the tail meets the head page, if the buffer is in overwrite mode,
    373the head page will be pushed ahead one. If the buffer is in producer/consumer
    374mode, the write will fail.
    375
    376Overwrite mode::
    377
    378              tail page
    379                 |
    380                 v
    381      +---+    +---+    +---+    +---+
    382  <---|   |--->|   |--->|   |--->|   |--->
    383  --->|   |<---|   |<---|   |<---|   |<---
    384      +---+    +---+    +---+    +---+
    385                          ^
    386                          |
    387                      head page
    388
    389
    390              tail page
    391                 |
    392                 v
    393      +---+    +---+    +---+    +---+
    394  <---|   |--->|   |--->|   |--->|   |--->
    395  --->|   |<---|   |<---|   |<---|   |<---
    396      +---+    +---+    +---+    +---+
    397                                   ^
    398                                   |
    399                               head page
    400
    401
    402                      tail page
    403                          |
    404                          v
    405      +---+    +---+    +---+    +---+
    406  <---|   |--->|   |--->|   |--->|   |--->
    407  --->|   |<---|   |<---|   |<---|   |<---
    408      +---+    +---+    +---+    +---+
    409                                   ^
    410                                   |
    411                               head page
    412
    413Note, the reader page will still point to the previous head page.
    414But when a swap takes place, it will use the most recent head page.
    415
    416
    417Making the Ring Buffer Lockless:
    418--------------------------------
    419
    420The main idea behind the lockless algorithm is to combine the moving
    421of the head_page pointer with the swapping of pages with the reader.
    422State flags are placed inside the pointer to the page. To do this,
    423each page must be aligned in memory by 4 bytes. This will allow the 2
    424least significant bits of the address to be used as flags, since
    425they will always be zero for the address. To get the address,
    426simply mask out the flags::
    427
    428  MASK = ~3
    429
    430  address & MASK
    431
    432Two flags will be kept by these two bits:
    433
    434   HEADER
    435	- the page being pointed to is a head page
    436
    437   UPDATE
    438	- the page being pointed to is being updated by a writer
    439          and was or is about to be a head page.
    440
    441::
    442
    443	      reader page
    444		  |
    445		  v
    446		+---+
    447		|   |------+
    448		+---+      |
    449			    |
    450			    v
    451	+---+    +---+    +---+    +---+
    452    <---|   |--->|   |-H->|   |--->|   |--->
    453    --->|   |<---|   |<---|   |<---|   |<---
    454	+---+    +---+    +---+    +---+
    455
    456
    457The above pointer "-H->" would have the HEADER flag set. That is
    458the next page is the next page to be swapped out by the reader.
    459This pointer means the next page is the head page.
    460
    461When the tail page meets the head pointer, it will use cmpxchg to
    462change the pointer to the UPDATE state::
    463
    464
    465              tail page
    466                 |
    467                 v
    468      +---+    +---+    +---+    +---+
    469  <---|   |--->|   |-H->|   |--->|   |--->
    470  --->|   |<---|   |<---|   |<---|   |<---
    471      +---+    +---+    +---+    +---+
    472
    473              tail page
    474                 |
    475                 v
    476      +---+    +---+    +---+    +---+
    477  <---|   |--->|   |-U->|   |--->|   |--->
    478  --->|   |<---|   |<---|   |<---|   |<---
    479      +---+    +---+    +---+    +---+
    480
    481"-U->" represents a pointer in the UPDATE state.
    482
    483Any access to the reader will need to take some sort of lock to serialize
    484the readers. But the writers will never take a lock to write to the
    485ring buffer. This means we only need to worry about a single reader,
    486and writes only preempt in "stack" formation.
    487
    488When the reader tries to swap the page with the ring buffer, it
    489will also use cmpxchg. If the flag bit in the pointer to the
    490head page does not have the HEADER flag set, the compare will fail
    491and the reader will need to look for the new head page and try again.
    492Note, the flags UPDATE and HEADER are never set at the same time.
    493
    494The reader swaps the reader page as follows::
    495
    496  +------+
    497  |reader|          RING BUFFER
    498  |page  |
    499  +------+
    500                  +---+    +---+    +---+
    501                  |   |--->|   |--->|   |
    502                  |   |<---|   |<---|   |
    503                  +---+    +---+    +---+
    504                   ^ |               ^ |
    505                   | +---------------+ |
    506                   +-----H-------------+
    507
    508The reader sets the reader page next pointer as HEADER to the page after
    509the head page::
    510
    511
    512  +------+
    513  |reader|          RING BUFFER
    514  |page  |-------H-----------+
    515  +------+                   v
    516    |             +---+    +---+    +---+
    517    |             |   |--->|   |--->|   |
    518    |             |   |<---|   |<---|   |<-+
    519    |             +---+    +---+    +---+  |
    520    |              ^ |               ^ |   |
    521    |              | +---------------+ |   |
    522    |              +-----H-------------+   |
    523    +--------------------------------------+
    524
    525It does a cmpxchg with the pointer to the previous head page to make it
    526point to the reader page. Note that the new pointer does not have the HEADER
    527flag set.  This action atomically moves the head page forward::
    528
    529  +------+
    530  |reader|          RING BUFFER
    531  |page  |-------H-----------+
    532  +------+                   v
    533    |  ^          +---+   +---+   +---+
    534    |  |          |   |-->|   |-->|   |
    535    |  |          |   |<--|   |<--|   |<-+
    536    |  |          +---+   +---+   +---+  |
    537    |  |             |             ^ |   |
    538    |  |             +-------------+ |   |
    539    |  +-----------------------------+   |
    540    +------------------------------------+
    541
    542After the new head page is set, the previous pointer of the head page is
    543updated to the reader page::
    544
    545  +------+
    546  |reader|          RING BUFFER
    547  |page  |-------H-----------+
    548  +------+ <---------------+ v
    549    |  ^          +---+   +---+   +---+
    550    |  |          |   |-->|   |-->|   |
    551    |  |          |   |   |   |<--|   |<-+
    552    |  |          +---+   +---+   +---+  |
    553    |  |             |             ^ |   |
    554    |  |             +-------------+ |   |
    555    |  +-----------------------------+   |
    556    +------------------------------------+
    557
    558  +------+
    559  |buffer|          RING BUFFER
    560  |page  |-------H-----------+  <--- New head page
    561  +------+ <---------------+ v
    562    |  ^          +---+   +---+   +---+
    563    |  |          |   |   |   |-->|   |
    564    |  |  New     |   |   |   |<--|   |<-+
    565    |  | Reader   +---+   +---+   +---+  |
    566    |  |  page ----^                 |   |
    567    |  |                             |   |
    568    |  +-----------------------------+   |
    569    +------------------------------------+
    570
    571Another important point: The page that the reader page points back to
    572by its previous pointer (the one that now points to the new head page)
    573never points back to the reader page. That is because the reader page is
    574not part of the ring buffer. Traversing the ring buffer via the next pointers
    575will always stay in the ring buffer. Traversing the ring buffer via the
    576prev pointers may not.
    577
    578Note, the way to determine a reader page is simply by examining the previous
    579pointer of the page. If the next pointer of the previous page does not
    580point back to the original page, then the original page is a reader page::
    581
    582
    583             +--------+
    584             | reader |  next   +----+
    585             |  page  |-------->|    |<====== (buffer page)
    586             +--------+         +----+
    587                 |                | ^
    588                 |                v | next
    589            prev |              +----+
    590                 +------------->|    |
    591                                +----+
    592
    593The way the head page moves forward:
    594
    595When the tail page meets the head page and the buffer is in overwrite mode
    596and more writes take place, the head page must be moved forward before the
    597writer may move the tail page. The way this is done is that the writer
    598performs a cmpxchg to convert the pointer to the head page from the HEADER
    599flag to have the UPDATE flag set. Once this is done, the reader will
    600not be able to swap the head page from the buffer, nor will it be able to
    601move the head page, until the writer is finished with the move.
    602
    603This eliminates any races that the reader can have on the writer. The reader
    604must spin, and this is why the reader cannot preempt the writer::
    605
    606              tail page
    607                 |
    608                 v
    609      +---+    +---+    +---+    +---+
    610  <---|   |--->|   |-H->|   |--->|   |--->
    611  --->|   |<---|   |<---|   |<---|   |<---
    612      +---+    +---+    +---+    +---+
    613
    614              tail page
    615                 |
    616                 v
    617      +---+    +---+    +---+    +---+
    618  <---|   |--->|   |-U->|   |--->|   |--->
    619  --->|   |<---|   |<---|   |<---|   |<---
    620      +---+    +---+    +---+    +---+
    621
    622The following page will be made into the new head page::
    623
    624             tail page
    625                 |
    626                 v
    627      +---+    +---+    +---+    +---+
    628  <---|   |--->|   |-U->|   |-H->|   |--->
    629  --->|   |<---|   |<---|   |<---|   |<---
    630      +---+    +---+    +---+    +---+
    631
    632After the new head page has been set, we can set the old head page
    633pointer back to NORMAL::
    634
    635             tail page
    636                 |
    637                 v
    638      +---+    +---+    +---+    +---+
    639  <---|   |--->|   |--->|   |-H->|   |--->
    640  --->|   |<---|   |<---|   |<---|   |<---
    641      +---+    +---+    +---+    +---+
    642
    643After the head page has been moved, the tail page may now move forward::
    644
    645                      tail page
    646                          |
    647                          v
    648      +---+    +---+    +---+    +---+
    649  <---|   |--->|   |--->|   |-H->|   |--->
    650  --->|   |<---|   |<---|   |<---|   |<---
    651      +---+    +---+    +---+    +---+
    652
    653
    654The above are the trivial updates. Now for the more complex scenarios.
    655
    656
    657As stated before, if enough writes preempt the first write, the
    658tail page may make it all the way around the buffer and meet the commit
    659page. At this time, we must start dropping writes (usually with some kind
    660of warning to the user). But what happens if the commit was still on the
    661reader page? The commit page is not part of the ring buffer. The tail page
    662must account for this::
    663
    664
    665            reader page    commit page
    666                |              |
    667                v              |
    668               +---+           |
    669               |   |<----------+
    670               |   |
    671               |   |------+
    672               +---+      |
    673                          |
    674                          v
    675      +---+    +---+    +---+    +---+
    676  <---|   |--->|   |-H->|   |--->|   |--->
    677  --->|   |<---|   |<---|   |<---|   |<---
    678      +---+    +---+    +---+    +---+
    679                 ^
    680                 |
    681             tail page
    682
    683If the tail page were to simply push the head page forward, the commit when
    684leaving the reader page would not be pointing to the correct page.
    685
    686The solution to this is to test if the commit page is on the reader page
    687before pushing the head page. If it is, then it can be assumed that the
    688tail page wrapped the buffer, and we must drop new writes.
    689
    690This is not a race condition, because the commit page can only be moved
    691by the outermost writer (the writer that was preempted).
    692This means that the commit will not move while a writer is moving the
    693tail page. The reader cannot swap the reader page if it is also being
    694used as the commit page. The reader can simply check that the commit
    695is off the reader page. Once the commit page leaves the reader page
    696it will never go back on it unless a reader does another swap with the
    697buffer page that is also the commit page.
    698
    699
    700Nested writes
    701-------------
    702
    703In the pushing forward of the tail page we must first push forward
    704the head page if the head page is the next page. If the head page
    705is not the next page, the tail page is simply updated with a cmpxchg.
    706
    707Only writers move the tail page. This must be done atomically to protect
    708against nested writers::
    709
    710  temp_page = tail_page
    711  next_page = temp_page->next
    712  cmpxchg(tail_page, temp_page, next_page)
    713
    714The above will update the tail page if it is still pointing to the expected
    715page. If this fails, a nested write pushed it forward, the current write
    716does not need to push it::
    717
    718
    719             temp page
    720                 |
    721                 v
    722              tail page
    723                 |
    724                 v
    725      +---+    +---+    +---+    +---+
    726  <---|   |--->|   |--->|   |--->|   |--->
    727  --->|   |<---|   |<---|   |<---|   |<---
    728      +---+    +---+    +---+    +---+
    729
    730Nested write comes in and moves the tail page forward::
    731
    732                      tail page (moved by nested writer)
    733              temp page   |
    734                 |        |
    735                 v        v
    736      +---+    +---+    +---+    +---+
    737  <---|   |--->|   |--->|   |--->|   |--->
    738  --->|   |<---|   |<---|   |<---|   |<---
    739      +---+    +---+    +---+    +---+
    740
    741The above would fail the cmpxchg, but since the tail page has already
    742been moved forward, the writer will just try again to reserve storage
    743on the new tail page.
    744
    745But the moving of the head page is a bit more complex::
    746
    747              tail page
    748                 |
    749                 v
    750      +---+    +---+    +---+    +---+
    751  <---|   |--->|   |-H->|   |--->|   |--->
    752  --->|   |<---|   |<---|   |<---|   |<---
    753      +---+    +---+    +---+    +---+
    754
    755The write converts the head page pointer to UPDATE::
    756
    757              tail page
    758                 |
    759                 v
    760      +---+    +---+    +---+    +---+
    761  <---|   |--->|   |-U->|   |--->|   |--->
    762  --->|   |<---|   |<---|   |<---|   |<---
    763      +---+    +---+    +---+    +---+
    764
    765But if a nested writer preempts here, it will see that the next
    766page is a head page, but it is also nested. It will detect that
    767it is nested and will save that information. The detection is the
    768fact that it sees the UPDATE flag instead of a HEADER or NORMAL
    769pointer.
    770
    771The nested writer will set the new head page pointer::
    772
    773             tail page
    774                 |
    775                 v
    776      +---+    +---+    +---+    +---+
    777  <---|   |--->|   |-U->|   |-H->|   |--->
    778  --->|   |<---|   |<---|   |<---|   |<---
    779      +---+    +---+    +---+    +---+
    780
    781But it will not reset the update back to normal. Only the writer
    782that converted a pointer from HEAD to UPDATE will convert it back
    783to NORMAL::
    784
    785                      tail page
    786                          |
    787                          v
    788      +---+    +---+    +---+    +---+
    789  <---|   |--->|   |-U->|   |-H->|   |--->
    790  --->|   |<---|   |<---|   |<---|   |<---
    791      +---+    +---+    +---+    +---+
    792
    793After the nested writer finishes, the outermost writer will convert
    794the UPDATE pointer to NORMAL::
    795
    796
    797                      tail page
    798                          |
    799                          v
    800      +---+    +---+    +---+    +---+
    801  <---|   |--->|   |--->|   |-H->|   |--->
    802  --->|   |<---|   |<---|   |<---|   |<---
    803      +---+    +---+    +---+    +---+
    804
    805
    806It can be even more complex if several nested writes came in and moved
    807the tail page ahead several pages::
    808
    809
    810  (first writer)
    811
    812              tail page
    813                 |
    814                 v
    815      +---+    +---+    +---+    +---+
    816  <---|   |--->|   |-H->|   |--->|   |--->
    817  --->|   |<---|   |<---|   |<---|   |<---
    818      +---+    +---+    +---+    +---+
    819
    820The write converts the head page pointer to UPDATE::
    821
    822              tail page
    823                 |
    824                 v
    825      +---+    +---+    +---+    +---+
    826  <---|   |--->|   |-U->|   |--->|   |--->
    827  --->|   |<---|   |<---|   |<---|   |<---
    828      +---+    +---+    +---+    +---+
    829
    830Next writer comes in, and sees the update and sets up the new
    831head page::
    832
    833  (second writer)
    834
    835             tail page
    836                 |
    837                 v
    838      +---+    +---+    +---+    +---+
    839  <---|   |--->|   |-U->|   |-H->|   |--->
    840  --->|   |<---|   |<---|   |<---|   |<---
    841      +---+    +---+    +---+    +---+
    842
    843The nested writer moves the tail page forward. But does not set the old
    844update page to NORMAL because it is not the outermost writer::
    845
    846                      tail page
    847                          |
    848                          v
    849      +---+    +---+    +---+    +---+
    850  <---|   |--->|   |-U->|   |-H->|   |--->
    851  --->|   |<---|   |<---|   |<---|   |<---
    852      +---+    +---+    +---+    +---+
    853
    854Another writer preempts and sees the page after the tail page is a head page.
    855It changes it from HEAD to UPDATE::
    856
    857  (third writer)
    858
    859                      tail page
    860                          |
    861                          v
    862      +---+    +---+    +---+    +---+
    863  <---|   |--->|   |-U->|   |-U->|   |--->
    864  --->|   |<---|   |<---|   |<---|   |<---
    865      +---+    +---+    +---+    +---+
    866
    867The writer will move the head page forward::
    868
    869
    870  (third writer)
    871
    872                      tail page
    873                          |
    874                          v
    875      +---+    +---+    +---+    +---+
    876  <---|   |--->|   |-U->|   |-U->|   |-H->
    877  --->|   |<---|   |<---|   |<---|   |<---
    878      +---+    +---+    +---+    +---+
    879
    880But now that the third writer did change the HEAD flag to UPDATE it
    881will convert it to normal::
    882
    883
    884  (third writer)
    885
    886                      tail page
    887                          |
    888                          v
    889      +---+    +---+    +---+    +---+
    890  <---|   |--->|   |-U->|   |--->|   |-H->
    891  --->|   |<---|   |<---|   |<---|   |<---
    892      +---+    +---+    +---+    +---+
    893
    894
    895Then it will move the tail page, and return back to the second writer::
    896
    897
    898  (second writer)
    899
    900                               tail page
    901                                   |
    902                                   v
    903      +---+    +---+    +---+    +---+
    904  <---|   |--->|   |-U->|   |--->|   |-H->
    905  --->|   |<---|   |<---|   |<---|   |<---
    906      +---+    +---+    +---+    +---+
    907
    908
    909The second writer will fail to move the tail page because it was already
    910moved, so it will try again and add its data to the new tail page.
    911It will return to the first writer::
    912
    913
    914  (first writer)
    915
    916                               tail page
    917                                   |
    918                                   v
    919      +---+    +---+    +---+    +---+
    920  <---|   |--->|   |-U->|   |--->|   |-H->
    921  --->|   |<---|   |<---|   |<---|   |<---
    922      +---+    +---+    +---+    +---+
    923
    924The first writer cannot know atomically if the tail page moved
    925while it updates the HEAD page. It will then update the head page to
    926what it thinks is the new head page::
    927
    928
    929  (first writer)
    930
    931                               tail page
    932                                   |
    933                                   v
    934      +---+    +---+    +---+    +---+
    935  <---|   |--->|   |-U->|   |-H->|   |-H->
    936  --->|   |<---|   |<---|   |<---|   |<---
    937      +---+    +---+    +---+    +---+
    938
    939Since the cmpxchg returns the old value of the pointer the first writer
    940will see it succeeded in updating the pointer from NORMAL to HEAD.
    941But as we can see, this is not good enough. It must also check to see
    942if the tail page is either where it use to be or on the next page::
    943
    944
    945  (first writer)
    946
    947                 A        B    tail page
    948                 |        |        |
    949                 v        v        v
    950      +---+    +---+    +---+    +---+
    951  <---|   |--->|   |-U->|   |-H->|   |-H->
    952  --->|   |<---|   |<---|   |<---|   |<---
    953      +---+    +---+    +---+    +---+
    954
    955If tail page != A and tail page != B, then it must reset the pointer
    956back to NORMAL. The fact that it only needs to worry about nested
    957writers means that it only needs to check this after setting the HEAD page::
    958
    959
    960  (first writer)
    961
    962                 A        B    tail page
    963                 |        |        |
    964                 v        v        v
    965      +---+    +---+    +---+    +---+
    966  <---|   |--->|   |-U->|   |--->|   |-H->
    967  --->|   |<---|   |<---|   |<---|   |<---
    968      +---+    +---+    +---+    +---+
    969
    970Now the writer can update the head page. This is also why the head page must
    971remain in UPDATE and only reset by the outermost writer. This prevents
    972the reader from seeing the incorrect head page::
    973
    974
    975  (first writer)
    976
    977                 A        B    tail page
    978                 |        |        |
    979                 v        v        v
    980      +---+    +---+    +---+    +---+
    981  <---|   |--->|   |--->|   |--->|   |-H->
    982  --->|   |<---|   |<---|   |<---|   |<---
    983      +---+    +---+    +---+    +---+