0
0
mirror of https://github.com/sqlite/sqlite.git synced 2024-12-01 09:12:16 +01:00
sqlite/doc/vfs-shm.txt
drh d9e5c4f6ed Refactoring the VFS-SHM methods used by WAL. This version compiles and
runs non-WAL test cases but crashes and burns on wal.test.

FossilOrigin-Name: 2b00152c1ac0b3735aa6cfab61259ff04d81c701
2010-05-12 18:01:39 +00:00

131 lines
5.4 KiB
Plaintext

The 5 states of an historical rollback lock as implemented by the
xLock, xUnlock, and xCheckReservedLock methods of the sqlite3_io_methods
objec are:
UNLOCKED
SHARED
RESERVED
PENDING
EXCLUSIVE
The wal-index file has a similar locking hierarchy implemented using
the xShmLock method of the sqlite3_vfs object, but with 7
states. Each connection to a wal-index file must be in one of
the following 7 states:
UNLOCKED
READ
READ_FULL
WRITE
PENDING
CHECKPOINT
RECOVER
These roughly correspond to the 5 states of a rollback lock except
that SHARED is split out into 2 states: READ and READ_FULL and
there is an extra RECOVER state used for wal-index reconstruction.
The meanings of the various wal-index locking states is as follows:
UNLOCKED - The wal-index is not in use.
READ - Some prefix of the wal-index is being read. Additional
wal-index information can be appended at any time. The
newly appended content will be ignored by the holder of
the READ lock.
READ_FULL - The entire wal-index is being read. No new information
can be added to the wal-index. The holder of a READ_FULL
lock promises never to read pages from the database file
that are available anywhere in the wal-index.
WRITE - It is OK to append to the wal-index file and to adjust
the header to indicate the new "last valid frame".
PENDING - Waiting on all READ locks to clear so that a
CHECKPOINT lock can be acquired.
CHECKPOINT - It is OK to write any WAL data into the database file
and zero the last valid frame field of the wal-index
header. The wal-index file itself may not be changed
other than to zero the last valid frame field in the
header.
RECOVER - Held during wal-index recovery. Used to prevent a
race if multiple clients try to recover a wal-index at
the same time.
A particular lock manager implementation may coalesce one or more of
the wal-index locking states, though with a reduction in concurrency.
For example, an implemention might implement only exclusive locking,
in which case all states would be equivalent to CHECKPOINT, meaning that
only one reader or one writer or one checkpointer could be active at a
time. Or, an implementation might combine READ and READ_FULL into
a single state equivalent to READ, meaning that a writer could
coexist with a reader, but no reader or writers could coexist with a
checkpointer.
The lock manager must obey the following rules:
(1) A READ cannot coexist with CHECKPOINT.
(2) A READ_FULL cannot coexist with WRITE.
(3) None of WRITE, PENDING, CHECKPOINT, or RECOVER can coexist.
The SQLite core will obey the next set of rules. These rules are
assertions on the behavior of the SQLite core which might be verified
during testing using an instrumented lock manager.
(5) No part of the wal-index will be read without holding either some
kind of SHM lock or an EXCLUSIVE lock on the original database.
The original database is the file named in the 2nd parameter to
the xShmOpen method.
(6) A holder of a READ_FULL will never read any page of the database
file that is contained anywhere in the wal-index.
(7) No part of the wal-index other than the header will be written nor
will the size of the wal-index grow without holding a WRITE or
an EXCLUSIVE on the original database file.
(8) The wal-index header will not be written without holding one of
WRITE, CHECKPOINT, or RECOVER on the wal-index or an EXCLUSIVE on
the original database files.
(9) A CHECKPOINT or RECOVER must be held on the wal-index, or an
EXCLUSIVE on the original database file, in order to reset the
last valid frame counter in the header of the wal-index back to zero.
(10) A WRITE can only increase the last valid frame pointer in the header.
The SQLite core will only ever send requests for UNLOCK, READ, WRITE,
CHECKPOINT, or RECOVER to the lock manager. The SQLite core will never
request a READ_FULL or PENDING lock though the lock manager may deliver
those locking states in response to READ and CHECKPOINT requests,
respectively, if and only if the requested READ or CHECKPOINT cannot
be delivered.
The following are the allowed lock transitions:
Original-State Request New-State
-------------- ---------- ----------
(11a) UNLOCK READ READ
(11b) UNLOCK READ READ_FULL
(11c) UNLOCK CHECKPOINT PENDING
(11d) UNLOCK CHECKPOINT CHECKPOINT
(11e) READ UNLOCK UNLOCK
(11f) READ WRITE WRITE
(11g) READ RECOVER RECOVER
(11h) READ_FULL UNLOCK UNLOCK
(11i) READ_FULL WRITE WRITE
(11j) READ_FULL RECOVER RECOVER
(11k) WRITE READ READ
(11l) PENDING UNLOCK UNLOCK
(11m) PENDING CHECKPOINT CHECKPOINT
(11n) CHECKPOINT UNLOCK UNLOCK
(11o) RECOVER READ READ
These 15 transitions are all that needs to be supported. The lock
manager implementation can assert that fact. The other 27 possible
transitions among the 7 locking states will never occur.