4 ** The author disclaims copyright to this source code. In place of
5 ** a legal notice, here is a blessing:
7 ** May you do good and not evil.
8 ** May you find forgiveness for yourself and forgive others.
9 ** May you share freely, never taking more than you give.
11 *************************************************************************
13 ** This file contains the implementation of a write-ahead log (WAL) used in
14 ** "journal_mode=WAL" mode.
16 ** WRITE-AHEAD LOG (WAL) FILE FORMAT
18 ** A WAL file consists of a header followed by zero or more "frames".
19 ** Each frame records the revised content of a single page from the
20 ** database file. All changes to the database are recorded by writing
21 ** frames into the WAL. Transactions commit when a frame is written that
22 ** contains a commit marker. A single WAL can and usually does record
23 ** multiple transactions. Periodically, the content of the WAL is
24 ** transferred back into the database file in an operation called a
27 ** A single WAL file can be used multiple times. In other words, the
28 ** WAL can fill up with frames and then be checkpointed and then new
29 ** frames can overwrite the old ones. A WAL always grows from beginning
30 ** toward the end. Checksums and counters attached to each frame are
31 ** used to determine which frames within the WAL are valid and which
32 ** are leftovers from prior checkpoints.
34 ** The WAL header is 32 bytes in size and consists of the following eight
35 ** big-endian 32-bit unsigned integer values:
37 ** 0: Magic number. 0x377f0682 or 0x377f0683
38 ** 4: File format version. Currently 3007000
39 ** 8: Database page size. Example: 1024
40 ** 12: Checkpoint sequence number
41 ** 16: Salt-1, random integer incremented with each checkpoint
42 ** 20: Salt-2, a different random integer changing with each ckpt
43 ** 24: Checksum-1 (first part of checksum for first 24 bytes of header).
44 ** 28: Checksum-2 (second part of checksum for first 24 bytes of header).
46 ** Immediately following the wal-header are zero or more frames. Each
47 ** frame consists of a 24-byte frame-header followed by a <page-size> bytes
48 ** of page data. The frame-header is six big-endian 32-bit unsigned
49 ** integer values, as follows:
52 ** 4: For commit records, the size of the database image in pages
53 ** after the commit. For all other records, zero.
54 ** 8: Salt-1 (copied from the header)
55 ** 12: Salt-2 (copied from the header)
59 ** A frame is considered valid if and only if the following conditions are
62 ** (1) The salt-1 and salt-2 values in the frame-header match
63 ** salt values in the wal-header
65 ** (2) The checksum values in the final 8 bytes of the frame-header
66 ** exactly match the checksum computed consecutively on the
67 ** WAL header and the first 8 bytes and the content of all frames
68 ** up to and including the current frame.
70 ** The checksum is computed using 32-bit big-endian integers if the
71 ** magic number in the first 4 bytes of the WAL is 0x377f0683 and it
72 ** is computed using little-endian if the magic number is 0x377f0682.
73 ** The checksum values are always stored in the frame header in a
74 ** big-endian format regardless of which byte order is used to compute
75 ** the checksum. The checksum is computed by interpreting the input as
76 ** an even number of unsigned 32-bit integers: x[0] through x[N]. The
77 ** algorithm used for the checksum is as follows:
79 ** for i from 0 to n-1 step 2:
84 ** Note that s0 and s1 are both weighted checksums using fibonacci weights
85 ** in reverse order (the largest fibonacci weight occurs on the first element
86 ** of the sequence being summed.) The s1 value spans all 32-bit
87 ** terms of the sequence whereas s0 omits the final term.
89 ** On a checkpoint, the WAL is first VFS.xSync-ed, then valid content of the
90 ** WAL is transferred into the database, then the database is VFS.xSync-ed.
91 ** The VFS.xSync operations serve as write barriers - all writes launched
92 ** before the xSync must complete before any write that launches after the
95 ** After each checkpoint, the salt-1 value is incremented and the salt-2
96 ** value is randomized. This prevents old and new frames in the WAL from
97 ** being considered valid at the same time and being checkpointing together
102 ** To read a page from the database (call it page number P), a reader
103 ** first checks the WAL to see if it contains page P. If so, then the
104 ** last valid instance of page P that is a followed by a commit frame
105 ** or is a commit frame itself becomes the value read. If the WAL
106 ** contains no copies of page P that are valid and which are a commit
107 ** frame or are followed by a commit frame, then page P is read from
108 ** the database file.
110 ** To start a read transaction, the reader records the index of the last
111 ** valid frame in the WAL. The reader uses this recorded "mxFrame" value
112 ** for all subsequent read operations. New transactions can be appended
113 ** to the WAL, but as long as the reader uses its original mxFrame value
114 ** and ignores the newly appended content, it will see a consistent snapshot
115 ** of the database from a single point in time. This technique allows
116 ** multiple concurrent readers to view different versions of the database
117 ** content simultaneously.
119 ** The reader algorithm in the previous paragraphs works correctly, but
120 ** because frames for page P can appear anywhere within the WAL, the
121 ** reader has to scan the entire WAL looking for page P frames. If the
122 ** WAL is large (multiple megabytes is typical) that scan can be slow,
123 ** and read performance suffers. To overcome this problem, a separate
124 ** data structure called the wal-index is maintained to expedite the
125 ** search for frames of a particular page.
129 ** Conceptually, the wal-index is shared memory, though VFS implementations
130 ** might choose to implement the wal-index using a mmapped file. Because
131 ** the wal-index is shared memory, SQLite does not support journal_mode=WAL
132 ** on a network filesystem. All users of the database must be able to
135 ** In the default unix and windows implementation, the wal-index is a mmapped
136 ** file whose name is the database name with a "-shm" suffix added. For that
137 ** reason, the wal-index is sometimes called the "shm" file.
139 ** The wal-index is transient. After a crash, the wal-index can (and should
140 ** be) reconstructed from the original WAL file. In fact, the VFS is required
141 ** to either truncate or zero the header of the wal-index when the last
142 ** connection to it closes. Because the wal-index is transient, it can
143 ** use an architecture-specific format; it does not have to be cross-platform.
144 ** Hence, unlike the database and WAL file formats which store all values
145 ** as big endian, the wal-index can store multi-byte values in the native
146 ** byte order of the host computer.
148 ** The purpose of the wal-index is to answer this question quickly: Given
149 ** a page number P and a maximum frame index M, return the index of the
150 ** last frame in the wal before frame M for page P in the WAL, or return
151 ** NULL if there are no frames for page P in the WAL prior to M.
153 ** The wal-index consists of a header region, followed by an one or
154 ** more index blocks.
156 ** The wal-index header contains the total number of frames within the WAL
157 ** in the mxFrame field.
159 ** Each index block except for the first contains information on
160 ** HASHTABLE_NPAGE frames. The first index block contains information on
161 ** HASHTABLE_NPAGE_ONE frames. The values of HASHTABLE_NPAGE_ONE and
162 ** HASHTABLE_NPAGE are selected so that together the wal-index header and
163 ** first index block are the same size as all other index blocks in the
166 ** Each index block contains two sections, a page-mapping that contains the
167 ** database page number associated with each wal frame, and a hash-table
168 ** that allows readers to query an index block for a specific page number.
169 ** The page-mapping is an array of HASHTABLE_NPAGE (or HASHTABLE_NPAGE_ONE
170 ** for the first index block) 32-bit page numbers. The first entry in the
171 ** first index-block contains the database page number corresponding to the
172 ** first frame in the WAL file. The first entry in the second index block
173 ** in the WAL file corresponds to the (HASHTABLE_NPAGE_ONE+1)th frame in
174 ** the log, and so on.
176 ** The last index block in a wal-index usually contains less than the full
177 ** complement of HASHTABLE_NPAGE (or HASHTABLE_NPAGE_ONE) page-numbers,
178 ** depending on the contents of the WAL file. This does not change the
179 ** allocated size of the page-mapping array - the page-mapping array merely
180 ** contains unused entries.
182 ** Even without using the hash table, the last frame for page P
183 ** can be found by scanning the page-mapping sections of each index block
184 ** starting with the last index block and moving toward the first, and
185 ** within each index block, starting at the end and moving toward the
186 ** beginning. The first entry that equals P corresponds to the frame
187 ** holding the content for that page.
189 ** The hash table consists of HASHTABLE_NSLOT 16-bit unsigned integers.
190 ** HASHTABLE_NSLOT = 2*HASHTABLE_NPAGE, and there is one entry in the
191 ** hash table for each page number in the mapping section, so the hash
192 ** table is never more than half full. The expected number of collisions
193 ** prior to finding a match is 1. Each entry of the hash table is an
194 ** 1-based index of an entry in the mapping section of the same
195 ** index block. Let K be the 1-based index of the largest entry in
196 ** the mapping section. (For index blocks other than the last, K will
197 ** always be exactly HASHTABLE_NPAGE (4096) and for the last index block
198 ** K will be (mxFrame%HASHTABLE_NPAGE).) Unused slots of the hash table
199 ** contain a value of 0.
201 ** To look for page P in the hash table, first compute a hash iKey on
204 ** iKey = (P * 383) % HASHTABLE_NSLOT
206 ** Then start scanning entries of the hash table, starting with iKey
207 ** (wrapping around to the beginning when the end of the hash table is
208 ** reached) until an unused hash slot is found. Let the first unused slot
209 ** be at index iUnused. (iUnused might be less than iKey if there was
210 ** wrap-around.) Because the hash table is never more than half full,
211 ** the search is guaranteed to eventually hit an unused entry. Let
212 ** iMax be the value between iKey and iUnused, closest to iUnused,
213 ** where aHash[iMax]==P. If there is no iMax entry (if there exists
214 ** no hash slot such that aHash[i]==p) then page P is not in the
215 ** current index block. Otherwise the iMax-th mapping entry of the
216 ** current index block corresponds to the last entry that references
219 ** A hash search begins with the last index block and moves toward the
220 ** first index block, looking for entries corresponding to page P. On
221 ** average, only two or three slots in each index block need to be
222 ** examined in order to either find the last entry for page P, or to
223 ** establish that no such entry exists in the block. Each index block
224 ** holds over 4000 entries. So two or three index blocks are sufficient
225 ** to cover a typical 10 megabyte WAL file, assuming 1K pages. 8 or 10
226 ** comparisons (on average) suffice to either locate a frame in the
227 ** WAL or to establish that the frame does not exist in the WAL. This
228 ** is much faster than scanning the entire 10MB WAL.
230 ** Note that entries are added in order of increasing K. Hence, one
231 ** reader might be using some value K0 and a second reader that started
232 ** at a later time (after additional transactions were added to the WAL
233 ** and to the wal-index) might be using a different value K1, where K1>K0.
234 ** Both readers can use the same hash table and mapping section to get
235 ** the correct result. There may be entries in the hash table with
236 ** K>K0 but to the first reader, those entries will appear to be unused
237 ** slots in the hash table and so the first reader will get an answer as
238 ** if no values greater than K0 had ever been inserted into the hash table
239 ** in the first place - which is what reader one wants. Meanwhile, the
240 ** second reader using K1 will see additional values that were inserted
241 ** later, which is exactly what reader two wants.
243 ** When a rollback occurs, the value of K is decreased. Hash table entries
244 ** that correspond to frames greater than the new K value are removed
245 ** from the hash table at this point.
247 #ifndef SQLITE_OMIT_WAL
252 ** Trace output macros
254 #if defined(SQLITE_TEST) && defined(SQLITE_DEBUG)
255 int sqlite3WalTrace
= 0;
256 # define WALTRACE(X) if(sqlite3WalTrace) sqlite3DebugPrintf X
262 ** WAL mode depends on atomic aligned 32-bit loads and stores in a few
263 ** places. The following macros try to make this explicit.
265 #if GCC_VESRION>=5004000
266 # define AtomicLoad(PTR) __atomic_load_n((PTR),__ATOMIC_RELAXED)
267 # define AtomicStore(PTR,VAL) __atomic_store_n((PTR),(VAL),__ATOMIC_RELAXED)
269 # define AtomicLoad(PTR) (*(PTR))
270 # define AtomicStore(PTR,VAL) (*(PTR) = (VAL))
274 ** The maximum (and only) versions of the wal and wal-index formats
275 ** that may be interpreted by this version of SQLite.
277 ** If a client begins recovering a WAL file and finds that (a) the checksum
278 ** values in the wal-header are correct and (b) the version field is not
279 ** WAL_MAX_VERSION, recovery fails and SQLite returns SQLITE_CANTOPEN.
281 ** Similarly, if a client successfully reads a wal-index header (i.e. the
282 ** checksum test is successful) and finds that the version field is not
283 ** WALINDEX_MAX_VERSION, then no read-transaction is opened and SQLite
284 ** returns SQLITE_CANTOPEN.
286 #define WAL_MAX_VERSION 3007000
287 #define WALINDEX_MAX_VERSION 3007000
290 ** Index numbers for various locking bytes. WAL_NREADER is the number
291 ** of available reader locks and should be at least 3. The default
292 ** is SQLITE_SHM_NLOCK==8 and WAL_NREADER==5.
294 ** Technically, the various VFSes are free to implement these locks however
295 ** they see fit. However, compatibility is encouraged so that VFSes can
296 ** interoperate. The standard implemention used on both unix and windows
297 ** is for the index number to indicate a byte offset into the
298 ** WalCkptInfo.aLock[] array in the wal-index header. In other words, all
299 ** locks are on the shm file. The WALINDEX_LOCK_OFFSET constant (which
300 ** should be 120) is the location in the shm file for the first locking
303 #define WAL_WRITE_LOCK 0
304 #define WAL_ALL_BUT_WRITE 1
305 #define WAL_CKPT_LOCK 1
306 #define WAL_RECOVER_LOCK 2
307 #define WAL_READ_LOCK(I) (3+(I))
308 #define WAL_NREADER (SQLITE_SHM_NLOCK-3)
311 /* Object declarations */
312 typedef struct WalIndexHdr WalIndexHdr
;
313 typedef struct WalIterator WalIterator
;
314 typedef struct WalCkptInfo WalCkptInfo
;
318 ** The following object holds a copy of the wal-index header content.
320 ** The actual header in the wal-index consists of two copies of this
321 ** object followed by one instance of the WalCkptInfo object.
322 ** For all versions of SQLite through 3.10.0 and probably beyond,
323 ** the locking bytes (WalCkptInfo.aLock) start at offset 120 and
324 ** the total header size is 136 bytes.
326 ** The szPage value can be any power of 2 between 512 and 32768, inclusive.
327 ** Or it can be 1 to represent a 65536-byte page. The latter case was
328 ** added in 3.7.1 when support for 64K pages was added.
331 u32 iVersion
; /* Wal-index version */
332 u32 unused
; /* Unused (padding) field */
333 u32 iChange
; /* Counter incremented each transaction */
334 u8 isInit
; /* 1 when initialized */
335 u8 bigEndCksum
; /* True if checksums in WAL are big-endian */
336 u16 szPage
; /* Database page size in bytes. 1==64K */
337 u32 mxFrame
; /* Index of last valid frame in the WAL */
338 u32 nPage
; /* Size of database in pages */
339 u32 aFrameCksum
[2]; /* Checksum of last frame in log */
340 u32 aSalt
[2]; /* Two salt values copied from WAL header */
341 u32 aCksum
[2]; /* Checksum over all prior fields */
345 ** A copy of the following object occurs in the wal-index immediately
346 ** following the second copy of the WalIndexHdr. This object stores
347 ** information used by checkpoint.
349 ** nBackfill is the number of frames in the WAL that have been written
350 ** back into the database. (We call the act of moving content from WAL to
351 ** database "backfilling".) The nBackfill number is never greater than
352 ** WalIndexHdr.mxFrame. nBackfill can only be increased by threads
353 ** holding the WAL_CKPT_LOCK lock (which includes a recovery thread).
354 ** However, a WAL_WRITE_LOCK thread can move the value of nBackfill from
355 ** mxFrame back to zero when the WAL is reset.
357 ** nBackfillAttempted is the largest value of nBackfill that a checkpoint
358 ** has attempted to achieve. Normally nBackfill==nBackfillAtempted, however
359 ** the nBackfillAttempted is set before any backfilling is done and the
360 ** nBackfill is only set after all backfilling completes. So if a checkpoint
361 ** crashes, nBackfillAttempted might be larger than nBackfill. The
362 ** WalIndexHdr.mxFrame must never be less than nBackfillAttempted.
364 ** The aLock[] field is a set of bytes used for locking. These bytes should
365 ** never be read or written.
367 ** There is one entry in aReadMark[] for each reader lock. If a reader
368 ** holds read-lock K, then the value in aReadMark[K] is no greater than
369 ** the mxFrame for that reader. The value READMARK_NOT_USED (0xffffffff)
370 ** for any aReadMark[] means that entry is unused. aReadMark[0] is
371 ** a special case; its value is never used and it exists as a place-holder
372 ** to avoid having to offset aReadMark[] indexs by one. Readers holding
373 ** WAL_READ_LOCK(0) always ignore the entire WAL and read all content
374 ** directly from the database.
376 ** The value of aReadMark[K] may only be changed by a thread that
377 ** is holding an exclusive lock on WAL_READ_LOCK(K). Thus, the value of
378 ** aReadMark[K] cannot changed while there is a reader is using that mark
379 ** since the reader will be holding a shared lock on WAL_READ_LOCK(K).
381 ** The checkpointer may only transfer frames from WAL to database where
382 ** the frame numbers are less than or equal to every aReadMark[] that is
383 ** in use (that is, every aReadMark[j] for which there is a corresponding
384 ** WAL_READ_LOCK(j)). New readers (usually) pick the aReadMark[] with the
385 ** largest value and will increase an unused aReadMark[] to mxFrame if there
386 ** is not already an aReadMark[] equal to mxFrame. The exception to the
387 ** previous sentence is when nBackfill equals mxFrame (meaning that everything
388 ** in the WAL has been backfilled into the database) then new readers
389 ** will choose aReadMark[0] which has value 0 and hence such reader will
390 ** get all their all content directly from the database file and ignore
393 ** Writers normally append new frames to the end of the WAL. However,
394 ** if nBackfill equals mxFrame (meaning that all WAL content has been
395 ** written back into the database) and if no readers are using the WAL
396 ** (in other words, if there are no WAL_READ_LOCK(i) where i>0) then
397 ** the writer will first "reset" the WAL back to the beginning and start
398 ** writing new content beginning at frame 1.
400 ** We assume that 32-bit loads are atomic and so no locks are needed in
401 ** order to read from any aReadMark[] entries.
404 u32 nBackfill
; /* Number of WAL frames backfilled into DB */
405 u32 aReadMark
[WAL_NREADER
]; /* Reader marks */
406 u8 aLock
[SQLITE_SHM_NLOCK
]; /* Reserved space for locks */
407 u32 nBackfillAttempted
; /* WAL frames perhaps written, or maybe not */
408 u32 notUsed0
; /* Available for future enhancements */
410 #define READMARK_NOT_USED 0xffffffff
413 /* A block of WALINDEX_LOCK_RESERVED bytes beginning at
414 ** WALINDEX_LOCK_OFFSET is reserved for locks. Since some systems
415 ** only support mandatory file-locks, we do not read or write data
416 ** from the region of the file on which locks are applied.
418 #define WALINDEX_LOCK_OFFSET (sizeof(WalIndexHdr)*2+offsetof(WalCkptInfo,aLock))
419 #define WALINDEX_HDR_SIZE (sizeof(WalIndexHdr)*2+sizeof(WalCkptInfo))
421 /* Size of header before each frame in wal */
422 #define WAL_FRAME_HDRSIZE 24
424 /* Size of write ahead log header, including checksum. */
425 #define WAL_HDRSIZE 32
427 /* WAL magic value. Either this value, or the same value with the least
428 ** significant bit also set (WAL_MAGIC | 0x00000001) is stored in 32-bit
429 ** big-endian format in the first 4 bytes of a WAL file.
431 ** If the LSB is set, then the checksums for each frame within the WAL
432 ** file are calculated by treating all data as an array of 32-bit
433 ** big-endian words. Otherwise, they are calculated by interpreting
434 ** all data as 32-bit little-endian words.
436 #define WAL_MAGIC 0x377f0682
439 ** Return the offset of frame iFrame in the write-ahead log file,
440 ** assuming a database page size of szPage bytes. The offset returned
441 ** is to the start of the write-ahead log frame-header.
443 #define walFrameOffset(iFrame, szPage) ( \
444 WAL_HDRSIZE + ((iFrame)-1)*(i64)((szPage)+WAL_FRAME_HDRSIZE) \
448 ** An open write-ahead log file is represented by an instance of the
452 sqlite3_vfs
*pVfs
; /* The VFS used to create pDbFd */
453 sqlite3_file
*pDbFd
; /* File handle for the database file */
454 sqlite3_file
*pWalFd
; /* File handle for WAL file */
455 u32 iCallback
; /* Value to pass to log callback (or 0) */
456 i64 mxWalSize
; /* Truncate WAL to this size upon reset */
457 int nWiData
; /* Size of array apWiData */
458 int szFirstBlock
; /* Size of first block written to WAL file */
459 volatile u32
**apWiData
; /* Pointer to wal-index content in memory */
460 u32 szPage
; /* Database page size */
461 i16 readLock
; /* Which read lock is being held. -1 for none */
462 u8 syncFlags
; /* Flags to use to sync header writes */
463 u8 exclusiveMode
; /* Non-zero if connection is in exclusive mode */
464 u8 writeLock
; /* True if in a write transaction */
465 u8 ckptLock
; /* True if holding a checkpoint lock */
466 u8 readOnly
; /* WAL_RDWR, WAL_RDONLY, or WAL_SHM_RDONLY */
467 u8 truncateOnCommit
; /* True to truncate WAL file on commit */
468 u8 syncHeader
; /* Fsync the WAL header if true */
469 u8 padToSectorBoundary
; /* Pad transactions out to the next sector */
470 u8 bShmUnreliable
; /* SHM content is read-only and unreliable */
471 WalIndexHdr hdr
; /* Wal-index header for current transaction */
472 u32 minFrame
; /* Ignore wal frames before this one */
473 u32 iReCksum
; /* On commit, recalculate checksums from here */
474 const char *zWalName
; /* Name of WAL file */
475 u32 nCkpt
; /* Checkpoint sequence counter in the wal-header */
477 u8 lockError
; /* True if a locking error has occurred */
479 #ifdef SQLITE_ENABLE_SNAPSHOT
480 WalIndexHdr
*pSnapshot
; /* Start transaction here if not NULL */
485 ** Candidate values for Wal.exclusiveMode.
487 #define WAL_NORMAL_MODE 0
488 #define WAL_EXCLUSIVE_MODE 1
489 #define WAL_HEAPMEMORY_MODE 2
492 ** Possible values for WAL.readOnly
494 #define WAL_RDWR 0 /* Normal read/write connection */
495 #define WAL_RDONLY 1 /* The WAL file is readonly */
496 #define WAL_SHM_RDONLY 2 /* The SHM file is readonly */
499 ** Each page of the wal-index mapping contains a hash-table made up of
500 ** an array of HASHTABLE_NSLOT elements of the following type.
505 ** This structure is used to implement an iterator that loops through
506 ** all frames in the WAL in database page order. Where two or more frames
507 ** correspond to the same database page, the iterator visits only the
508 ** frame most recently written to the WAL (in other words, the frame with
509 ** the largest index).
511 ** The internals of this structure are only accessed by:
513 ** walIteratorInit() - Create a new iterator,
514 ** walIteratorNext() - Step an iterator,
515 ** walIteratorFree() - Free an iterator.
517 ** This functionality is used by the checkpoint code (see walCheckpoint()).
520 int iPrior
; /* Last result returned from the iterator */
521 int nSegment
; /* Number of entries in aSegment[] */
523 int iNext
; /* Next slot in aIndex[] not yet returned */
524 ht_slot
*aIndex
; /* i0, i1, i2... such that aPgno[iN] ascend */
525 u32
*aPgno
; /* Array of page numbers. */
526 int nEntry
; /* Nr. of entries in aPgno[] and aIndex[] */
527 int iZero
; /* Frame number associated with aPgno[0] */
528 } aSegment
[1]; /* One for every 32KB page in the wal-index */
532 ** Define the parameters of the hash tables in the wal-index file. There
533 ** is a hash-table following every HASHTABLE_NPAGE page numbers in the
536 ** Changing any of these constants will alter the wal-index format and
537 ** create incompatibilities.
539 #define HASHTABLE_NPAGE 4096 /* Must be power of 2 */
540 #define HASHTABLE_HASH_1 383 /* Should be prime */
541 #define HASHTABLE_NSLOT (HASHTABLE_NPAGE*2) /* Must be a power of 2 */
544 ** The block of page numbers associated with the first hash-table in a
545 ** wal-index is smaller than usual. This is so that there is a complete
546 ** hash-table on each aligned 32KB page of the wal-index.
548 #define HASHTABLE_NPAGE_ONE (HASHTABLE_NPAGE - (WALINDEX_HDR_SIZE/sizeof(u32)))
550 /* The wal-index is divided into pages of WALINDEX_PGSZ bytes each. */
551 #define WALINDEX_PGSZ ( \
552 sizeof(ht_slot)*HASHTABLE_NSLOT + HASHTABLE_NPAGE*sizeof(u32) \
556 ** Obtain a pointer to the iPage'th page of the wal-index. The wal-index
557 ** is broken into pages of WALINDEX_PGSZ bytes. Wal-index pages are
558 ** numbered from zero.
560 ** If the wal-index is currently smaller the iPage pages then the size
561 ** of the wal-index might be increased, but only if it is safe to do
562 ** so. It is safe to enlarge the wal-index if pWal->writeLock is true
563 ** or pWal->exclusiveMode==WAL_HEAPMEMORY_MODE.
565 ** If this call is successful, *ppPage is set to point to the wal-index
566 ** page and SQLITE_OK is returned. If an error (an OOM or VFS error) occurs,
567 ** then an SQLite error code is returned and *ppPage is set to 0.
569 static SQLITE_NOINLINE
int walIndexPageRealloc(
570 Wal
*pWal
, /* The WAL context */
571 int iPage
, /* The page we seek */
572 volatile u32
**ppPage
/* Write the page pointer here */
576 /* Enlarge the pWal->apWiData[] array if required */
577 if( pWal
->nWiData
<=iPage
){
578 sqlite3_int64 nByte
= sizeof(u32
*)*(iPage
+1);
579 volatile u32
**apNew
;
580 apNew
= (volatile u32
**)sqlite3_realloc64((void *)pWal
->apWiData
, nByte
);
583 return SQLITE_NOMEM_BKPT
;
585 memset((void*)&apNew
[pWal
->nWiData
], 0,
586 sizeof(u32
*)*(iPage
+1-pWal
->nWiData
));
587 pWal
->apWiData
= apNew
;
588 pWal
->nWiData
= iPage
+1;
591 /* Request a pointer to the required page from the VFS */
592 assert( pWal
->apWiData
[iPage
]==0 );
593 if( pWal
->exclusiveMode
==WAL_HEAPMEMORY_MODE
){
594 pWal
->apWiData
[iPage
] = (u32
volatile *)sqlite3MallocZero(WALINDEX_PGSZ
);
595 if( !pWal
->apWiData
[iPage
] ) rc
= SQLITE_NOMEM_BKPT
;
597 rc
= sqlite3OsShmMap(pWal
->pDbFd
, iPage
, WALINDEX_PGSZ
,
598 pWal
->writeLock
, (void volatile **)&pWal
->apWiData
[iPage
]
600 assert( pWal
->apWiData
[iPage
]!=0 || rc
!=SQLITE_OK
|| pWal
->writeLock
==0 );
601 testcase( pWal
->apWiData
[iPage
]==0 && rc
==SQLITE_OK
);
602 if( (rc
&0xff)==SQLITE_READONLY
){
603 pWal
->readOnly
|= WAL_SHM_RDONLY
;
604 if( rc
==SQLITE_READONLY
){
610 *ppPage
= pWal
->apWiData
[iPage
];
611 assert( iPage
==0 || *ppPage
|| rc
!=SQLITE_OK
);
614 static int walIndexPage(
615 Wal
*pWal
, /* The WAL context */
616 int iPage
, /* The page we seek */
617 volatile u32
**ppPage
/* Write the page pointer here */
619 if( pWal
->nWiData
<=iPage
|| (*ppPage
= pWal
->apWiData
[iPage
])==0 ){
620 return walIndexPageRealloc(pWal
, iPage
, ppPage
);
626 ** Return a pointer to the WalCkptInfo structure in the wal-index.
628 static volatile WalCkptInfo
*walCkptInfo(Wal
*pWal
){
629 assert( pWal
->nWiData
>0 && pWal
->apWiData
[0] );
630 return (volatile WalCkptInfo
*)&(pWal
->apWiData
[0][sizeof(WalIndexHdr
)/2]);
634 ** Return a pointer to the WalIndexHdr structure in the wal-index.
636 static volatile WalIndexHdr
*walIndexHdr(Wal
*pWal
){
637 assert( pWal
->nWiData
>0 && pWal
->apWiData
[0] );
638 return (volatile WalIndexHdr
*)pWal
->apWiData
[0];
642 ** The argument to this macro must be of type u32. On a little-endian
643 ** architecture, it returns the u32 value that results from interpreting
644 ** the 4 bytes as a big-endian value. On a big-endian architecture, it
645 ** returns the value that would be produced by interpreting the 4 bytes
646 ** of the input value as a little-endian integer.
648 #define BYTESWAP32(x) ( \
649 (((x)&0x000000FF)<<24) + (((x)&0x0000FF00)<<8) \
650 + (((x)&0x00FF0000)>>8) + (((x)&0xFF000000)>>24) \
654 ** Generate or extend an 8 byte checksum based on the data in
655 ** array aByte[] and the initial values of aIn[0] and aIn[1] (or
656 ** initial values of 0 and 0 if aIn==NULL).
658 ** The checksum is written back into aOut[] before returning.
660 ** nByte must be a positive multiple of 8.
662 static void walChecksumBytes(
663 int nativeCksum
, /* True for native byte-order, false for non-native */
664 u8
*a
, /* Content to be checksummed */
665 int nByte
, /* Bytes of content in a[]. Must be a multiple of 8. */
666 const u32
*aIn
, /* Initial checksum value input */
667 u32
*aOut
/* OUT: Final checksum value output */
670 u32
*aData
= (u32
*)a
;
671 u32
*aEnd
= (u32
*)&a
[nByte
];
681 assert( (nByte
&0x00000007)==0 );
682 assert( nByte
<=65536 );
688 }while( aData
<aEnd
);
691 s1
+= BYTESWAP32(aData
[0]) + s2
;
692 s2
+= BYTESWAP32(aData
[1]) + s1
;
694 }while( aData
<aEnd
);
701 static void walShmBarrier(Wal
*pWal
){
702 if( pWal
->exclusiveMode
!=WAL_HEAPMEMORY_MODE
){
703 sqlite3OsShmBarrier(pWal
->pDbFd
);
708 ** Write the header information in pWal->hdr into the wal-index.
710 ** The checksum on pWal->hdr is updated before it is written.
712 static void walIndexWriteHdr(Wal
*pWal
){
713 volatile WalIndexHdr
*aHdr
= walIndexHdr(pWal
);
714 const int nCksum
= offsetof(WalIndexHdr
, aCksum
);
716 assert( pWal
->writeLock
);
717 pWal
->hdr
.isInit
= 1;
718 pWal
->hdr
.iVersion
= WALINDEX_MAX_VERSION
;
719 walChecksumBytes(1, (u8
*)&pWal
->hdr
, nCksum
, 0, pWal
->hdr
.aCksum
);
720 memcpy((void*)&aHdr
[1], (const void*)&pWal
->hdr
, sizeof(WalIndexHdr
));
722 memcpy((void*)&aHdr
[0], (const void*)&pWal
->hdr
, sizeof(WalIndexHdr
));
726 ** This function encodes a single frame header and writes it to a buffer
727 ** supplied by the caller. A frame-header is made up of a series of
728 ** 4-byte big-endian integers, as follows:
731 ** 4: For commit records, the size of the database image in pages
732 ** after the commit. For all other records, zero.
733 ** 8: Salt-1 (copied from the wal-header)
734 ** 12: Salt-2 (copied from the wal-header)
738 static void walEncodeFrame(
739 Wal
*pWal
, /* The write-ahead log */
740 u32 iPage
, /* Database page number for frame */
741 u32 nTruncate
, /* New db size (or 0 for non-commit frames) */
742 u8
*aData
, /* Pointer to page data */
743 u8
*aFrame
/* OUT: Write encoded frame here */
745 int nativeCksum
; /* True for native byte-order checksums */
746 u32
*aCksum
= pWal
->hdr
.aFrameCksum
;
747 assert( WAL_FRAME_HDRSIZE
==24 );
748 sqlite3Put4byte(&aFrame
[0], iPage
);
749 sqlite3Put4byte(&aFrame
[4], nTruncate
);
750 if( pWal
->iReCksum
==0 ){
751 memcpy(&aFrame
[8], pWal
->hdr
.aSalt
, 8);
753 nativeCksum
= (pWal
->hdr
.bigEndCksum
==SQLITE_BIGENDIAN
);
754 walChecksumBytes(nativeCksum
, aFrame
, 8, aCksum
, aCksum
);
755 walChecksumBytes(nativeCksum
, aData
, pWal
->szPage
, aCksum
, aCksum
);
757 sqlite3Put4byte(&aFrame
[16], aCksum
[0]);
758 sqlite3Put4byte(&aFrame
[20], aCksum
[1]);
760 memset(&aFrame
[8], 0, 16);
765 ** Check to see if the frame with header in aFrame[] and content
766 ** in aData[] is valid. If it is a valid frame, fill *piPage and
767 ** *pnTruncate and return true. Return if the frame is not valid.
769 static int walDecodeFrame(
770 Wal
*pWal
, /* The write-ahead log */
771 u32
*piPage
, /* OUT: Database page number for frame */
772 u32
*pnTruncate
, /* OUT: New db size (or 0 if not commit) */
773 u8
*aData
, /* Pointer to page data (for checksum) */
774 u8
*aFrame
/* Frame data */
776 int nativeCksum
; /* True for native byte-order checksums */
777 u32
*aCksum
= pWal
->hdr
.aFrameCksum
;
778 u32 pgno
; /* Page number of the frame */
779 assert( WAL_FRAME_HDRSIZE
==24 );
781 /* A frame is only valid if the salt values in the frame-header
782 ** match the salt values in the wal-header.
784 if( memcmp(&pWal
->hdr
.aSalt
, &aFrame
[8], 8)!=0 ){
788 /* A frame is only valid if the page number is creater than zero.
790 pgno
= sqlite3Get4byte(&aFrame
[0]);
795 /* A frame is only valid if a checksum of the WAL header,
796 ** all prior frams, the first 16 bytes of this frame-header,
797 ** and the frame-data matches the checksum in the last 8
798 ** bytes of this frame-header.
800 nativeCksum
= (pWal
->hdr
.bigEndCksum
==SQLITE_BIGENDIAN
);
801 walChecksumBytes(nativeCksum
, aFrame
, 8, aCksum
, aCksum
);
802 walChecksumBytes(nativeCksum
, aData
, pWal
->szPage
, aCksum
, aCksum
);
803 if( aCksum
[0]!=sqlite3Get4byte(&aFrame
[16])
804 || aCksum
[1]!=sqlite3Get4byte(&aFrame
[20])
806 /* Checksum failed. */
810 /* If we reach this point, the frame is valid. Return the page number
811 ** and the new database size.
814 *pnTruncate
= sqlite3Get4byte(&aFrame
[4]);
819 #if defined(SQLITE_TEST) && defined(SQLITE_DEBUG)
821 ** Names of locks. This routine is used to provide debugging output and is not
822 ** a part of an ordinary build.
824 static const char *walLockName(int lockIdx
){
825 if( lockIdx
==WAL_WRITE_LOCK
){
827 }else if( lockIdx
==WAL_CKPT_LOCK
){
829 }else if( lockIdx
==WAL_RECOVER_LOCK
){
830 return "RECOVER-LOCK";
832 static char zName
[15];
833 sqlite3_snprintf(sizeof(zName
), zName
, "READ-LOCK[%d]",
834 lockIdx
-WAL_READ_LOCK(0));
838 #endif /*defined(SQLITE_TEST) || defined(SQLITE_DEBUG) */
842 ** Set or release locks on the WAL. Locks are either shared or exclusive.
843 ** A lock cannot be moved directly between shared and exclusive - it must go
844 ** through the unlocked state first.
846 ** In locking_mode=EXCLUSIVE, all of these routines become no-ops.
848 static int walLockShared(Wal
*pWal
, int lockIdx
){
850 if( pWal
->exclusiveMode
) return SQLITE_OK
;
851 rc
= sqlite3OsShmLock(pWal
->pDbFd
, lockIdx
, 1,
852 SQLITE_SHM_LOCK
| SQLITE_SHM_SHARED
);
853 WALTRACE(("WAL%p: acquire SHARED-%s %s\n", pWal
,
854 walLockName(lockIdx
), rc
? "failed" : "ok"));
855 VVA_ONLY( pWal
->lockError
= (u8
)(rc
!=SQLITE_OK
&& rc
!=SQLITE_BUSY
); )
858 static void walUnlockShared(Wal
*pWal
, int lockIdx
){
859 if( pWal
->exclusiveMode
) return;
860 (void)sqlite3OsShmLock(pWal
->pDbFd
, lockIdx
, 1,
861 SQLITE_SHM_UNLOCK
| SQLITE_SHM_SHARED
);
862 WALTRACE(("WAL%p: release SHARED-%s\n", pWal
, walLockName(lockIdx
)));
864 static int walLockExclusive(Wal
*pWal
, int lockIdx
, int n
){
866 if( pWal
->exclusiveMode
) return SQLITE_OK
;
867 rc
= sqlite3OsShmLock(pWal
->pDbFd
, lockIdx
, n
,
868 SQLITE_SHM_LOCK
| SQLITE_SHM_EXCLUSIVE
);
869 WALTRACE(("WAL%p: acquire EXCLUSIVE-%s cnt=%d %s\n", pWal
,
870 walLockName(lockIdx
), n
, rc
? "failed" : "ok"));
871 VVA_ONLY( pWal
->lockError
= (u8
)(rc
!=SQLITE_OK
&& rc
!=SQLITE_BUSY
); )
874 static void walUnlockExclusive(Wal
*pWal
, int lockIdx
, int n
){
875 if( pWal
->exclusiveMode
) return;
876 (void)sqlite3OsShmLock(pWal
->pDbFd
, lockIdx
, n
,
877 SQLITE_SHM_UNLOCK
| SQLITE_SHM_EXCLUSIVE
);
878 WALTRACE(("WAL%p: release EXCLUSIVE-%s cnt=%d\n", pWal
,
879 walLockName(lockIdx
), n
));
883 ** Compute a hash on a page number. The resulting hash value must land
884 ** between 0 and (HASHTABLE_NSLOT-1). The walHashNext() function advances
885 ** the hash to the next value in the event of a collision.
887 static int walHash(u32 iPage
){
889 assert( (HASHTABLE_NSLOT
& (HASHTABLE_NSLOT
-1))==0 );
890 return (iPage
*HASHTABLE_HASH_1
) & (HASHTABLE_NSLOT
-1);
892 static int walNextHash(int iPriorHash
){
893 return (iPriorHash
+1)&(HASHTABLE_NSLOT
-1);
897 ** An instance of the WalHashLoc object is used to describe the location
898 ** of a page hash table in the wal-index. This becomes the return value
899 ** from walHashGet().
901 typedef struct WalHashLoc WalHashLoc
;
903 volatile ht_slot
*aHash
; /* Start of the wal-index hash table */
904 volatile u32
*aPgno
; /* aPgno[1] is the page of first frame indexed */
905 u32 iZero
; /* One less than the frame number of first indexed*/
909 ** Return pointers to the hash table and page number array stored on
910 ** page iHash of the wal-index. The wal-index is broken into 32KB pages
911 ** numbered starting from 0.
913 ** Set output variable pLoc->aHash to point to the start of the hash table
914 ** in the wal-index file. Set pLoc->iZero to one less than the frame
915 ** number of the first frame indexed by this hash table. If a
916 ** slot in the hash table is set to N, it refers to frame number
917 ** (pLoc->iZero+N) in the log.
919 ** Finally, set pLoc->aPgno so that pLoc->aPgno[1] is the page number of the
920 ** first frame indexed by the hash table, frame (pLoc->iZero+1).
922 static int walHashGet(
923 Wal
*pWal
, /* WAL handle */
924 int iHash
, /* Find the iHash'th table */
925 WalHashLoc
*pLoc
/* OUT: Hash table location */
927 int rc
; /* Return code */
929 rc
= walIndexPage(pWal
, iHash
, &pLoc
->aPgno
);
930 assert( rc
==SQLITE_OK
|| iHash
>0 );
933 pLoc
->aHash
= (volatile ht_slot
*)&pLoc
->aPgno
[HASHTABLE_NPAGE
];
935 pLoc
->aPgno
= &pLoc
->aPgno
[WALINDEX_HDR_SIZE
/sizeof(u32
)];
938 pLoc
->iZero
= HASHTABLE_NPAGE_ONE
+ (iHash
-1)*HASHTABLE_NPAGE
;
940 pLoc
->aPgno
= &pLoc
->aPgno
[-1];
946 ** Return the number of the wal-index page that contains the hash-table
947 ** and page-number array that contain entries corresponding to WAL frame
948 ** iFrame. The wal-index is broken up into 32KB pages. Wal-index pages
949 ** are numbered starting from 0.
951 static int walFramePage(u32 iFrame
){
952 int iHash
= (iFrame
+HASHTABLE_NPAGE
-HASHTABLE_NPAGE_ONE
-1) / HASHTABLE_NPAGE
;
953 assert( (iHash
==0 || iFrame
>HASHTABLE_NPAGE_ONE
)
954 && (iHash
>=1 || iFrame
<=HASHTABLE_NPAGE_ONE
)
955 && (iHash
<=1 || iFrame
>(HASHTABLE_NPAGE_ONE
+HASHTABLE_NPAGE
))
956 && (iHash
>=2 || iFrame
<=HASHTABLE_NPAGE_ONE
+HASHTABLE_NPAGE
)
957 && (iHash
<=2 || iFrame
>(HASHTABLE_NPAGE_ONE
+2*HASHTABLE_NPAGE
))
963 ** Return the page number associated with frame iFrame in this WAL.
965 static u32
walFramePgno(Wal
*pWal
, u32 iFrame
){
966 int iHash
= walFramePage(iFrame
);
968 return pWal
->apWiData
[0][WALINDEX_HDR_SIZE
/sizeof(u32
) + iFrame
- 1];
970 return pWal
->apWiData
[iHash
][(iFrame
-1-HASHTABLE_NPAGE_ONE
)%HASHTABLE_NPAGE
];
974 ** Remove entries from the hash table that point to WAL slots greater
975 ** than pWal->hdr.mxFrame.
977 ** This function is called whenever pWal->hdr.mxFrame is decreased due
978 ** to a rollback or savepoint.
980 ** At most only the hash table containing pWal->hdr.mxFrame needs to be
981 ** updated. Any later hash tables will be automatically cleared when
982 ** pWal->hdr.mxFrame advances to the point where those hash tables are
985 static void walCleanupHash(Wal
*pWal
){
986 WalHashLoc sLoc
; /* Hash table location */
987 int iLimit
= 0; /* Zero values greater than this */
988 int nByte
; /* Number of bytes to zero in aPgno[] */
989 int i
; /* Used to iterate through aHash[] */
990 int rc
; /* Return code form walHashGet() */
992 assert( pWal
->writeLock
);
993 testcase( pWal
->hdr
.mxFrame
==HASHTABLE_NPAGE_ONE
-1 );
994 testcase( pWal
->hdr
.mxFrame
==HASHTABLE_NPAGE_ONE
);
995 testcase( pWal
->hdr
.mxFrame
==HASHTABLE_NPAGE_ONE
+1 );
997 if( pWal
->hdr
.mxFrame
==0 ) return;
999 /* Obtain pointers to the hash-table and page-number array containing
1000 ** the entry that corresponds to frame pWal->hdr.mxFrame. It is guaranteed
1001 ** that the page said hash-table and array reside on is already mapped.(1)
1003 assert( pWal
->nWiData
>walFramePage(pWal
->hdr
.mxFrame
) );
1004 assert( pWal
->apWiData
[walFramePage(pWal
->hdr
.mxFrame
)] );
1005 rc
= walHashGet(pWal
, walFramePage(pWal
->hdr
.mxFrame
), &sLoc
);
1006 if( NEVER(rc
) ) return; /* Defense-in-depth, in case (1) above is wrong */
1008 /* Zero all hash-table entries that correspond to frame numbers greater
1009 ** than pWal->hdr.mxFrame.
1011 iLimit
= pWal
->hdr
.mxFrame
- sLoc
.iZero
;
1013 for(i
=0; i
<HASHTABLE_NSLOT
; i
++){
1014 if( sLoc
.aHash
[i
]>iLimit
){
1019 /* Zero the entries in the aPgno array that correspond to frames with
1020 ** frame numbers greater than pWal->hdr.mxFrame.
1022 nByte
= (int)((char *)sLoc
.aHash
- (char *)&sLoc
.aPgno
[iLimit
+1]);
1023 memset((void *)&sLoc
.aPgno
[iLimit
+1], 0, nByte
);
1025 #ifdef SQLITE_ENABLE_EXPENSIVE_ASSERT
1026 /* Verify that the every entry in the mapping region is still reachable
1027 ** via the hash table even after the cleanup.
1030 int j
; /* Loop counter */
1031 int iKey
; /* Hash key */
1032 for(j
=1; j
<=iLimit
; j
++){
1033 for(iKey
=walHash(sLoc
.aPgno
[j
]);sLoc
.aHash
[iKey
];iKey
=walNextHash(iKey
)){
1034 if( sLoc
.aHash
[iKey
]==j
) break;
1036 assert( sLoc
.aHash
[iKey
]==j
);
1039 #endif /* SQLITE_ENABLE_EXPENSIVE_ASSERT */
1044 ** Set an entry in the wal-index that will map database page number
1045 ** pPage into WAL frame iFrame.
1047 static int walIndexAppend(Wal
*pWal
, u32 iFrame
, u32 iPage
){
1048 int rc
; /* Return code */
1049 WalHashLoc sLoc
; /* Wal-index hash table location */
1051 rc
= walHashGet(pWal
, walFramePage(iFrame
), &sLoc
);
1053 /* Assuming the wal-index file was successfully mapped, populate the
1054 ** page number array and hash table entry.
1056 if( rc
==SQLITE_OK
){
1057 int iKey
; /* Hash table key */
1058 int idx
; /* Value to write to hash-table slot */
1059 int nCollide
; /* Number of hash collisions */
1061 idx
= iFrame
- sLoc
.iZero
;
1062 assert( idx
<= HASHTABLE_NSLOT
/2 + 1 );
1064 /* If this is the first entry to be added to this hash-table, zero the
1065 ** entire hash table and aPgno[] array before proceeding.
1068 int nByte
= (int)((u8
*)&sLoc
.aHash
[HASHTABLE_NSLOT
]
1069 - (u8
*)&sLoc
.aPgno
[1]);
1070 memset((void*)&sLoc
.aPgno
[1], 0, nByte
);
1073 /* If the entry in aPgno[] is already set, then the previous writer
1074 ** must have exited unexpectedly in the middle of a transaction (after
1075 ** writing one or more dirty pages to the WAL to free up memory).
1076 ** Remove the remnants of that writers uncommitted transaction from
1077 ** the hash-table before writing any new entries.
1079 if( sLoc
.aPgno
[idx
] ){
1080 walCleanupHash(pWal
);
1081 assert( !sLoc
.aPgno
[idx
] );
1084 /* Write the aPgno[] array entry and the hash-table slot. */
1086 for(iKey
=walHash(iPage
); sLoc
.aHash
[iKey
]; iKey
=walNextHash(iKey
)){
1087 if( (nCollide
--)==0 ) return SQLITE_CORRUPT_BKPT
;
1089 sLoc
.aPgno
[idx
] = iPage
;
1090 sLoc
.aHash
[iKey
] = (ht_slot
)idx
;
1092 #ifdef SQLITE_ENABLE_EXPENSIVE_ASSERT
1093 /* Verify that the number of entries in the hash table exactly equals
1094 ** the number of entries in the mapping region.
1097 int i
; /* Loop counter */
1098 int nEntry
= 0; /* Number of entries in the hash table */
1099 for(i
=0; i
<HASHTABLE_NSLOT
; i
++){ if( sLoc
.aHash
[i
] ) nEntry
++; }
1100 assert( nEntry
==idx
);
1103 /* Verify that the every entry in the mapping region is reachable
1104 ** via the hash table. This turns out to be a really, really expensive
1105 ** thing to check, so only do this occasionally - not on every
1108 if( (idx
&0x3ff)==0 ){
1109 int i
; /* Loop counter */
1110 for(i
=1; i
<=idx
; i
++){
1111 for(iKey
=walHash(sLoc
.aPgno
[i
]);
1113 iKey
=walNextHash(iKey
)){
1114 if( sLoc
.aHash
[iKey
]==i
) break;
1116 assert( sLoc
.aHash
[iKey
]==i
);
1119 #endif /* SQLITE_ENABLE_EXPENSIVE_ASSERT */
1128 ** Recover the wal-index by reading the write-ahead log file.
1130 ** This routine first tries to establish an exclusive lock on the
1131 ** wal-index to prevent other threads/processes from doing anything
1132 ** with the WAL or wal-index while recovery is running. The
1133 ** WAL_RECOVER_LOCK is also held so that other threads will know
1134 ** that this thread is running recovery. If unable to establish
1135 ** the necessary locks, this routine returns SQLITE_BUSY.
1137 static int walIndexRecover(Wal
*pWal
){
1138 int rc
; /* Return Code */
1139 i64 nSize
; /* Size of log file */
1140 u32 aFrameCksum
[2] = {0, 0};
1141 int iLock
; /* Lock offset to lock for checkpoint */
1143 /* Obtain an exclusive lock on all byte in the locking range not already
1144 ** locked by the caller. The caller is guaranteed to have locked the
1145 ** WAL_WRITE_LOCK byte, and may have also locked the WAL_CKPT_LOCK byte.
1146 ** If successful, the same bytes that are locked here are unlocked before
1147 ** this function returns.
1149 assert( pWal
->ckptLock
==1 || pWal
->ckptLock
==0 );
1150 assert( WAL_ALL_BUT_WRITE
==WAL_WRITE_LOCK
+1 );
1151 assert( WAL_CKPT_LOCK
==WAL_ALL_BUT_WRITE
);
1152 assert( pWal
->writeLock
);
1153 iLock
= WAL_ALL_BUT_WRITE
+ pWal
->ckptLock
;
1154 rc
= walLockExclusive(pWal
, iLock
, WAL_READ_LOCK(0)-iLock
);
1155 if( rc
==SQLITE_OK
){
1156 rc
= walLockExclusive(pWal
, WAL_READ_LOCK(1), WAL_NREADER
-1);
1157 if( rc
!=SQLITE_OK
){
1158 walUnlockExclusive(pWal
, iLock
, WAL_READ_LOCK(0)-iLock
);
1165 WALTRACE(("WAL%p: recovery begin...\n", pWal
));
1167 memset(&pWal
->hdr
, 0, sizeof(WalIndexHdr
));
1169 rc
= sqlite3OsFileSize(pWal
->pWalFd
, &nSize
);
1170 if( rc
!=SQLITE_OK
){
1171 goto recovery_error
;
1174 if( nSize
>WAL_HDRSIZE
){
1175 u8 aBuf
[WAL_HDRSIZE
]; /* Buffer to load WAL header into */
1176 u8
*aFrame
= 0; /* Malloc'd buffer to load entire frame */
1177 int szFrame
; /* Number of bytes in buffer aFrame[] */
1178 u8
*aData
; /* Pointer to data part of aFrame buffer */
1179 int iFrame
; /* Index of last frame read */
1180 i64 iOffset
; /* Next offset to read from log file */
1181 int szPage
; /* Page size according to the log */
1182 u32 magic
; /* Magic value read from WAL header */
1183 u32 version
; /* Magic value read from WAL header */
1184 int isValid
; /* True if this frame is valid */
1186 /* Read in the WAL header. */
1187 rc
= sqlite3OsRead(pWal
->pWalFd
, aBuf
, WAL_HDRSIZE
, 0);
1188 if( rc
!=SQLITE_OK
){
1189 goto recovery_error
;
1192 /* If the database page size is not a power of two, or is greater than
1193 ** SQLITE_MAX_PAGE_SIZE, conclude that the WAL file contains no valid
1194 ** data. Similarly, if the 'magic' value is invalid, ignore the whole
1197 magic
= sqlite3Get4byte(&aBuf
[0]);
1198 szPage
= sqlite3Get4byte(&aBuf
[8]);
1199 if( (magic
&0xFFFFFFFE)!=WAL_MAGIC
1200 || szPage
&(szPage
-1)
1201 || szPage
>SQLITE_MAX_PAGE_SIZE
1206 pWal
->hdr
.bigEndCksum
= (u8
)(magic
&0x00000001);
1207 pWal
->szPage
= szPage
;
1208 pWal
->nCkpt
= sqlite3Get4byte(&aBuf
[12]);
1209 memcpy(&pWal
->hdr
.aSalt
, &aBuf
[16], 8);
1211 /* Verify that the WAL header checksum is correct */
1212 walChecksumBytes(pWal
->hdr
.bigEndCksum
==SQLITE_BIGENDIAN
,
1213 aBuf
, WAL_HDRSIZE
-2*4, 0, pWal
->hdr
.aFrameCksum
1215 if( pWal
->hdr
.aFrameCksum
[0]!=sqlite3Get4byte(&aBuf
[24])
1216 || pWal
->hdr
.aFrameCksum
[1]!=sqlite3Get4byte(&aBuf
[28])
1221 /* Verify that the version number on the WAL format is one that
1222 ** are able to understand */
1223 version
= sqlite3Get4byte(&aBuf
[4]);
1224 if( version
!=WAL_MAX_VERSION
){
1225 rc
= SQLITE_CANTOPEN_BKPT
;
1229 /* Malloc a buffer to read frames into. */
1230 szFrame
= szPage
+ WAL_FRAME_HDRSIZE
;
1231 aFrame
= (u8
*)sqlite3_malloc64(szFrame
);
1233 rc
= SQLITE_NOMEM_BKPT
;
1234 goto recovery_error
;
1236 aData
= &aFrame
[WAL_FRAME_HDRSIZE
];
1238 /* Read all frames from the log file. */
1240 for(iOffset
=WAL_HDRSIZE
; (iOffset
+szFrame
)<=nSize
; iOffset
+=szFrame
){
1241 u32 pgno
; /* Database page number for frame */
1242 u32 nTruncate
; /* dbsize field from frame header */
1244 /* Read and decode the next log frame. */
1246 rc
= sqlite3OsRead(pWal
->pWalFd
, aFrame
, szFrame
, iOffset
);
1247 if( rc
!=SQLITE_OK
) break;
1248 isValid
= walDecodeFrame(pWal
, &pgno
, &nTruncate
, aData
, aFrame
);
1249 if( !isValid
) break;
1250 rc
= walIndexAppend(pWal
, iFrame
, pgno
);
1251 if( rc
!=SQLITE_OK
) break;
1253 /* If nTruncate is non-zero, this is a commit record. */
1255 pWal
->hdr
.mxFrame
= iFrame
;
1256 pWal
->hdr
.nPage
= nTruncate
;
1257 pWal
->hdr
.szPage
= (u16
)((szPage
&0xff00) | (szPage
>>16));
1258 testcase( szPage
<=32768 );
1259 testcase( szPage
>=65536 );
1260 aFrameCksum
[0] = pWal
->hdr
.aFrameCksum
[0];
1261 aFrameCksum
[1] = pWal
->hdr
.aFrameCksum
[1];
1265 sqlite3_free(aFrame
);
1269 if( rc
==SQLITE_OK
){
1270 volatile WalCkptInfo
*pInfo
;
1272 pWal
->hdr
.aFrameCksum
[0] = aFrameCksum
[0];
1273 pWal
->hdr
.aFrameCksum
[1] = aFrameCksum
[1];
1274 walIndexWriteHdr(pWal
);
1276 /* Reset the checkpoint-header. This is safe because this thread is
1277 ** currently holding locks that exclude all other readers, writers and
1280 pInfo
= walCkptInfo(pWal
);
1281 pInfo
->nBackfill
= 0;
1282 pInfo
->nBackfillAttempted
= pWal
->hdr
.mxFrame
;
1283 pInfo
->aReadMark
[0] = 0;
1284 for(i
=1; i
<WAL_NREADER
; i
++) pInfo
->aReadMark
[i
] = READMARK_NOT_USED
;
1285 if( pWal
->hdr
.mxFrame
) pInfo
->aReadMark
[1] = pWal
->hdr
.mxFrame
;
1287 /* If more than one frame was recovered from the log file, report an
1288 ** event via sqlite3_log(). This is to help with identifying performance
1289 ** problems caused by applications routinely shutting down without
1290 ** checkpointing the log file.
1292 if( pWal
->hdr
.nPage
){
1293 sqlite3_log(SQLITE_NOTICE_RECOVER_WAL
,
1294 "recovered %d frames from WAL file %s",
1295 pWal
->hdr
.mxFrame
, pWal
->zWalName
1301 WALTRACE(("WAL%p: recovery %s\n", pWal
, rc
? "failed" : "ok"));
1302 walUnlockExclusive(pWal
, iLock
, WAL_READ_LOCK(0)-iLock
);
1303 walUnlockExclusive(pWal
, WAL_READ_LOCK(1), WAL_NREADER
-1);
1308 ** Close an open wal-index.
1310 static void walIndexClose(Wal
*pWal
, int isDelete
){
1311 if( pWal
->exclusiveMode
==WAL_HEAPMEMORY_MODE
|| pWal
->bShmUnreliable
){
1313 for(i
=0; i
<pWal
->nWiData
; i
++){
1314 sqlite3_free((void *)pWal
->apWiData
[i
]);
1315 pWal
->apWiData
[i
] = 0;
1318 if( pWal
->exclusiveMode
!=WAL_HEAPMEMORY_MODE
){
1319 sqlite3OsShmUnmap(pWal
->pDbFd
, isDelete
);
1324 ** Open a connection to the WAL file zWalName. The database file must
1325 ** already be opened on connection pDbFd. The buffer that zWalName points
1326 ** to must remain valid for the lifetime of the returned Wal* handle.
1328 ** A SHARED lock should be held on the database file when this function
1329 ** is called. The purpose of this SHARED lock is to prevent any other
1330 ** client from unlinking the WAL or wal-index file. If another process
1331 ** were to do this just after this client opened one of these files, the
1332 ** system would be badly broken.
1334 ** If the log file is successfully opened, SQLITE_OK is returned and
1335 ** *ppWal is set to point to a new WAL handle. If an error occurs,
1336 ** an SQLite error code is returned and *ppWal is left unmodified.
1339 sqlite3_vfs
*pVfs
, /* vfs module to open wal and wal-index */
1340 sqlite3_file
*pDbFd
, /* The open database file */
1341 const char *zWalName
, /* Name of the WAL file */
1342 int bNoShm
, /* True to run in heap-memory mode */
1343 i64 mxWalSize
, /* Truncate WAL to this size on reset */
1344 Wal
**ppWal
/* OUT: Allocated Wal handle */
1346 int rc
; /* Return Code */
1347 Wal
*pRet
; /* Object to allocate and return */
1348 int flags
; /* Flags passed to OsOpen() */
1350 assert( zWalName
&& zWalName
[0] );
1353 /* In the amalgamation, the os_unix.c and os_win.c source files come before
1354 ** this source file. Verify that the #defines of the locking byte offsets
1355 ** in os_unix.c and os_win.c agree with the WALINDEX_LOCK_OFFSET value.
1356 ** For that matter, if the lock offset ever changes from its initial design
1357 ** value of 120, we need to know that so there is an assert() to check it.
1359 assert( 120==WALINDEX_LOCK_OFFSET
);
1360 assert( 136==WALINDEX_HDR_SIZE
);
1362 assert( WIN_SHM_BASE
==WALINDEX_LOCK_OFFSET
);
1364 #ifdef UNIX_SHM_BASE
1365 assert( UNIX_SHM_BASE
==WALINDEX_LOCK_OFFSET
);
1369 /* Allocate an instance of struct Wal to return. */
1371 pRet
= (Wal
*)sqlite3MallocZero(sizeof(Wal
) + pVfs
->szOsFile
);
1373 return SQLITE_NOMEM_BKPT
;
1377 pRet
->pWalFd
= (sqlite3_file
*)&pRet
[1];
1378 pRet
->pDbFd
= pDbFd
;
1379 pRet
->readLock
= -1;
1380 pRet
->mxWalSize
= mxWalSize
;
1381 pRet
->zWalName
= zWalName
;
1382 pRet
->syncHeader
= 1;
1383 pRet
->padToSectorBoundary
= 1;
1384 pRet
->exclusiveMode
= (bNoShm
? WAL_HEAPMEMORY_MODE
: WAL_NORMAL_MODE
);
1386 /* Open file handle on the write-ahead log file. */
1387 flags
= (SQLITE_OPEN_READWRITE
|SQLITE_OPEN_CREATE
|SQLITE_OPEN_WAL
);
1388 rc
= sqlite3OsOpen(pVfs
, zWalName
, pRet
->pWalFd
, flags
, &flags
);
1389 if( rc
==SQLITE_OK
&& flags
&SQLITE_OPEN_READONLY
){
1390 pRet
->readOnly
= WAL_RDONLY
;
1393 if( rc
!=SQLITE_OK
){
1394 walIndexClose(pRet
, 0);
1395 sqlite3OsClose(pRet
->pWalFd
);
1398 int iDC
= sqlite3OsDeviceCharacteristics(pDbFd
);
1399 if( iDC
& SQLITE_IOCAP_SEQUENTIAL
){ pRet
->syncHeader
= 0; }
1400 if( iDC
& SQLITE_IOCAP_POWERSAFE_OVERWRITE
){
1401 pRet
->padToSectorBoundary
= 0;
1404 WALTRACE(("WAL%d: opened\n", pRet
));
1410 ** Change the size to which the WAL file is trucated on each reset.
1412 void sqlite3WalLimit(Wal
*pWal
, i64 iLimit
){
1413 if( pWal
) pWal
->mxWalSize
= iLimit
;
1417 ** Find the smallest page number out of all pages held in the WAL that
1418 ** has not been returned by any prior invocation of this method on the
1419 ** same WalIterator object. Write into *piFrame the frame index where
1420 ** that page was last written into the WAL. Write into *piPage the page
1423 ** Return 0 on success. If there are no pages in the WAL with a page
1424 ** number larger than *piPage, then return 1.
1426 static int walIteratorNext(
1427 WalIterator
*p
, /* Iterator */
1428 u32
*piPage
, /* OUT: The page number of the next page */
1429 u32
*piFrame
/* OUT: Wal frame index of next page */
1431 u32 iMin
; /* Result pgno must be greater than iMin */
1432 u32 iRet
= 0xFFFFFFFF; /* 0xffffffff is never a valid page number */
1433 int i
; /* For looping through segments */
1436 assert( iMin
<0xffffffff );
1437 for(i
=p
->nSegment
-1; i
>=0; i
--){
1438 struct WalSegment
*pSegment
= &p
->aSegment
[i
];
1439 while( pSegment
->iNext
<pSegment
->nEntry
){
1440 u32 iPg
= pSegment
->aPgno
[pSegment
->aIndex
[pSegment
->iNext
]];
1444 *piFrame
= pSegment
->iZero
+ pSegment
->aIndex
[pSegment
->iNext
];
1452 *piPage
= p
->iPrior
= iRet
;
1453 return (iRet
==0xFFFFFFFF);
1457 ** This function merges two sorted lists into a single sorted list.
1459 ** aLeft[] and aRight[] are arrays of indices. The sort key is
1460 ** aContent[aLeft[]] and aContent[aRight[]]. Upon entry, the following
1461 ** is guaranteed for all J<K:
1463 ** aContent[aLeft[J]] < aContent[aLeft[K]]
1464 ** aContent[aRight[J]] < aContent[aRight[K]]
1466 ** This routine overwrites aRight[] with a new (probably longer) sequence
1467 ** of indices such that the aRight[] contains every index that appears in
1468 ** either aLeft[] or the old aRight[] and such that the second condition
1469 ** above is still met.
1471 ** The aContent[aLeft[X]] values will be unique for all X. And the
1472 ** aContent[aRight[X]] values will be unique too. But there might be
1473 ** one or more combinations of X and Y such that
1475 ** aLeft[X]!=aRight[Y] && aContent[aLeft[X]] == aContent[aRight[Y]]
1477 ** When that happens, omit the aLeft[X] and use the aRight[Y] index.
1479 static void walMerge(
1480 const u32
*aContent
, /* Pages in wal - keys for the sort */
1481 ht_slot
*aLeft
, /* IN: Left hand input list */
1482 int nLeft
, /* IN: Elements in array *paLeft */
1483 ht_slot
**paRight
, /* IN/OUT: Right hand input list */
1484 int *pnRight
, /* IN/OUT: Elements in *paRight */
1485 ht_slot
*aTmp
/* Temporary buffer */
1487 int iLeft
= 0; /* Current index in aLeft */
1488 int iRight
= 0; /* Current index in aRight */
1489 int iOut
= 0; /* Current index in output buffer */
1490 int nRight
= *pnRight
;
1491 ht_slot
*aRight
= *paRight
;
1493 assert( nLeft
>0 && nRight
>0 );
1494 while( iRight
<nRight
|| iLeft
<nLeft
){
1499 && (iRight
>=nRight
|| aContent
[aLeft
[iLeft
]]<aContent
[aRight
[iRight
]])
1501 logpage
= aLeft
[iLeft
++];
1503 logpage
= aRight
[iRight
++];
1505 dbpage
= aContent
[logpage
];
1507 aTmp
[iOut
++] = logpage
;
1508 if( iLeft
<nLeft
&& aContent
[aLeft
[iLeft
]]==dbpage
) iLeft
++;
1510 assert( iLeft
>=nLeft
|| aContent
[aLeft
[iLeft
]]>dbpage
);
1511 assert( iRight
>=nRight
|| aContent
[aRight
[iRight
]]>dbpage
);
1516 memcpy(aLeft
, aTmp
, sizeof(aTmp
[0])*iOut
);
1520 ** Sort the elements in list aList using aContent[] as the sort key.
1521 ** Remove elements with duplicate keys, preferring to keep the
1522 ** larger aList[] values.
1524 ** The aList[] entries are indices into aContent[]. The values in
1525 ** aList[] are to be sorted so that for all J<K:
1527 ** aContent[aList[J]] < aContent[aList[K]]
1529 ** For any X and Y such that
1531 ** aContent[aList[X]] == aContent[aList[Y]]
1533 ** Keep the larger of the two values aList[X] and aList[Y] and discard
1536 static void walMergesort(
1537 const u32
*aContent
, /* Pages in wal */
1538 ht_slot
*aBuffer
, /* Buffer of at least *pnList items to use */
1539 ht_slot
*aList
, /* IN/OUT: List to sort */
1540 int *pnList
/* IN/OUT: Number of elements in aList[] */
1543 int nList
; /* Number of elements in aList */
1544 ht_slot
*aList
; /* Pointer to sub-list content */
1547 const int nList
= *pnList
; /* Size of input list */
1548 int nMerge
= 0; /* Number of elements in list aMerge */
1549 ht_slot
*aMerge
= 0; /* List to be merged */
1550 int iList
; /* Index into input list */
1551 u32 iSub
= 0; /* Index into aSub array */
1552 struct Sublist aSub
[13]; /* Array of sub-lists */
1554 memset(aSub
, 0, sizeof(aSub
));
1555 assert( nList
<=HASHTABLE_NPAGE
&& nList
>0 );
1556 assert( HASHTABLE_NPAGE
==(1<<(ArraySize(aSub
)-1)) );
1558 for(iList
=0; iList
<nList
; iList
++){
1560 aMerge
= &aList
[iList
];
1561 for(iSub
=0; iList
& (1<<iSub
); iSub
++){
1563 assert( iSub
<ArraySize(aSub
) );
1565 assert( p
->aList
&& p
->nList
<=(1<<iSub
) );
1566 assert( p
->aList
==&aList
[iList
&~((2<<iSub
)-1)] );
1567 walMerge(aContent
, p
->aList
, p
->nList
, &aMerge
, &nMerge
, aBuffer
);
1569 aSub
[iSub
].aList
= aMerge
;
1570 aSub
[iSub
].nList
= nMerge
;
1573 for(iSub
++; iSub
<ArraySize(aSub
); iSub
++){
1574 if( nList
& (1<<iSub
) ){
1576 assert( iSub
<ArraySize(aSub
) );
1578 assert( p
->nList
<=(1<<iSub
) );
1579 assert( p
->aList
==&aList
[nList
&~((2<<iSub
)-1)] );
1580 walMerge(aContent
, p
->aList
, p
->nList
, &aMerge
, &nMerge
, aBuffer
);
1583 assert( aMerge
==aList
);
1589 for(i
=1; i
<*pnList
; i
++){
1590 assert( aContent
[aList
[i
]] > aContent
[aList
[i
-1]] );
1597 ** Free an iterator allocated by walIteratorInit().
1599 static void walIteratorFree(WalIterator
*p
){
1604 ** Construct a WalInterator object that can be used to loop over all
1605 ** pages in the WAL following frame nBackfill in ascending order. Frames
1606 ** nBackfill or earlier may be included - excluding them is an optimization
1607 ** only. The caller must hold the checkpoint lock.
1609 ** On success, make *pp point to the newly allocated WalInterator object
1610 ** return SQLITE_OK. Otherwise, return an error code. If this routine
1611 ** returns an error, the value of *pp is undefined.
1613 ** The calling routine should invoke walIteratorFree() to destroy the
1614 ** WalIterator object when it has finished with it.
1616 static int walIteratorInit(Wal
*pWal
, u32 nBackfill
, WalIterator
**pp
){
1617 WalIterator
*p
; /* Return value */
1618 int nSegment
; /* Number of segments to merge */
1619 u32 iLast
; /* Last frame in log */
1620 sqlite3_int64 nByte
; /* Number of bytes to allocate */
1621 int i
; /* Iterator variable */
1622 ht_slot
*aTmp
; /* Temp space used by merge-sort */
1623 int rc
= SQLITE_OK
; /* Return Code */
1625 /* This routine only runs while holding the checkpoint lock. And
1626 ** it only runs if there is actually content in the log (mxFrame>0).
1628 assert( pWal
->ckptLock
&& pWal
->hdr
.mxFrame
>0 );
1629 iLast
= pWal
->hdr
.mxFrame
;
1631 /* Allocate space for the WalIterator object. */
1632 nSegment
= walFramePage(iLast
) + 1;
1633 nByte
= sizeof(WalIterator
)
1634 + (nSegment
-1)*sizeof(struct WalSegment
)
1635 + iLast
*sizeof(ht_slot
);
1636 p
= (WalIterator
*)sqlite3_malloc64(nByte
);
1638 return SQLITE_NOMEM_BKPT
;
1640 memset(p
, 0, nByte
);
1641 p
->nSegment
= nSegment
;
1643 /* Allocate temporary space used by the merge-sort routine. This block
1644 ** of memory will be freed before this function returns.
1646 aTmp
= (ht_slot
*)sqlite3_malloc64(
1647 sizeof(ht_slot
) * (iLast
>HASHTABLE_NPAGE
?HASHTABLE_NPAGE
:iLast
)
1650 rc
= SQLITE_NOMEM_BKPT
;
1653 for(i
=walFramePage(nBackfill
+1); rc
==SQLITE_OK
&& i
<nSegment
; i
++){
1656 rc
= walHashGet(pWal
, i
, &sLoc
);
1657 if( rc
==SQLITE_OK
){
1658 int j
; /* Counter variable */
1659 int nEntry
; /* Number of entries in this segment */
1660 ht_slot
*aIndex
; /* Sorted index for this segment */
1663 if( (i
+1)==nSegment
){
1664 nEntry
= (int)(iLast
- sLoc
.iZero
);
1666 nEntry
= (int)((u32
*)sLoc
.aHash
- (u32
*)sLoc
.aPgno
);
1668 aIndex
= &((ht_slot
*)&p
->aSegment
[p
->nSegment
])[sLoc
.iZero
];
1671 for(j
=0; j
<nEntry
; j
++){
1672 aIndex
[j
] = (ht_slot
)j
;
1674 walMergesort((u32
*)sLoc
.aPgno
, aTmp
, aIndex
, &nEntry
);
1675 p
->aSegment
[i
].iZero
= sLoc
.iZero
;
1676 p
->aSegment
[i
].nEntry
= nEntry
;
1677 p
->aSegment
[i
].aIndex
= aIndex
;
1678 p
->aSegment
[i
].aPgno
= (u32
*)sLoc
.aPgno
;
1683 if( rc
!=SQLITE_OK
){
1692 ** Attempt to obtain the exclusive WAL lock defined by parameters lockIdx and
1693 ** n. If the attempt fails and parameter xBusy is not NULL, then it is a
1694 ** busy-handler function. Invoke it and retry the lock until either the
1695 ** lock is successfully obtained or the busy-handler returns 0.
1697 static int walBusyLock(
1698 Wal
*pWal
, /* WAL connection */
1699 int (*xBusy
)(void*), /* Function to call when busy */
1700 void *pBusyArg
, /* Context argument for xBusyHandler */
1701 int lockIdx
, /* Offset of first byte to lock */
1702 int n
/* Number of bytes to lock */
1706 rc
= walLockExclusive(pWal
, lockIdx
, n
);
1707 }while( xBusy
&& rc
==SQLITE_BUSY
&& xBusy(pBusyArg
) );
1712 ** The cache of the wal-index header must be valid to call this function.
1713 ** Return the page-size in bytes used by the database.
1715 static int walPagesize(Wal
*pWal
){
1716 return (pWal
->hdr
.szPage
&0xfe00) + ((pWal
->hdr
.szPage
&0x0001)<<16);
1720 ** The following is guaranteed when this function is called:
1722 ** a) the WRITER lock is held,
1723 ** b) the entire log file has been checkpointed, and
1724 ** c) any existing readers are reading exclusively from the database
1725 ** file - there are no readers that may attempt to read a frame from
1728 ** This function updates the shared-memory structures so that the next
1729 ** client to write to the database (which may be this one) does so by
1730 ** writing frames into the start of the log file.
1732 ** The value of parameter salt1 is used as the aSalt[1] value in the
1733 ** new wal-index header. It should be passed a pseudo-random value (i.e.
1734 ** one obtained from sqlite3_randomness()).
1736 static void walRestartHdr(Wal
*pWal
, u32 salt1
){
1737 volatile WalCkptInfo
*pInfo
= walCkptInfo(pWal
);
1738 int i
; /* Loop counter */
1739 u32
*aSalt
= pWal
->hdr
.aSalt
; /* Big-endian salt values */
1741 pWal
->hdr
.mxFrame
= 0;
1742 sqlite3Put4byte((u8
*)&aSalt
[0], 1 + sqlite3Get4byte((u8
*)&aSalt
[0]));
1743 memcpy(&pWal
->hdr
.aSalt
[1], &salt1
, 4);
1744 walIndexWriteHdr(pWal
);
1745 pInfo
->nBackfill
= 0;
1746 pInfo
->nBackfillAttempted
= 0;
1747 pInfo
->aReadMark
[1] = 0;
1748 for(i
=2; i
<WAL_NREADER
; i
++) pInfo
->aReadMark
[i
] = READMARK_NOT_USED
;
1749 assert( pInfo
->aReadMark
[0]==0 );
1753 ** Copy as much content as we can from the WAL back into the database file
1754 ** in response to an sqlite3_wal_checkpoint() request or the equivalent.
1756 ** The amount of information copies from WAL to database might be limited
1757 ** by active readers. This routine will never overwrite a database page
1758 ** that a concurrent reader might be using.
1760 ** All I/O barrier operations (a.k.a fsyncs) occur in this routine when
1761 ** SQLite is in WAL-mode in synchronous=NORMAL. That means that if
1762 ** checkpoints are always run by a background thread or background
1763 ** process, foreground threads will never block on a lengthy fsync call.
1765 ** Fsync is called on the WAL before writing content out of the WAL and
1766 ** into the database. This ensures that if the new content is persistent
1767 ** in the WAL and can be recovered following a power-loss or hard reset.
1769 ** Fsync is also called on the database file if (and only if) the entire
1770 ** WAL content is copied into the database file. This second fsync makes
1771 ** it safe to delete the WAL since the new content will persist in the
1774 ** This routine uses and updates the nBackfill field of the wal-index header.
1775 ** This is the only routine that will increase the value of nBackfill.
1776 ** (A WAL reset or recovery will revert nBackfill to zero, but not increase
1779 ** The caller must be holding sufficient locks to ensure that no other
1780 ** checkpoint is running (in any other thread or process) at the same
1783 static int walCheckpoint(
1784 Wal
*pWal
, /* Wal connection */
1785 sqlite3
*db
, /* Check for interrupts on this handle */
1786 int eMode
, /* One of PASSIVE, FULL or RESTART */
1787 int (*xBusy
)(void*), /* Function to call when busy */
1788 void *pBusyArg
, /* Context argument for xBusyHandler */
1789 int sync_flags
, /* Flags for OsSync() (or 0) */
1790 u8
*zBuf
/* Temporary buffer to use */
1792 int rc
= SQLITE_OK
; /* Return code */
1793 int szPage
; /* Database page-size */
1794 WalIterator
*pIter
= 0; /* Wal iterator context */
1795 u32 iDbpage
= 0; /* Next database page to write */
1796 u32 iFrame
= 0; /* Wal frame containing data for iDbpage */
1797 u32 mxSafeFrame
; /* Max frame that can be backfilled */
1798 u32 mxPage
; /* Max database page to write */
1799 int i
; /* Loop counter */
1800 volatile WalCkptInfo
*pInfo
; /* The checkpoint status information */
1802 szPage
= walPagesize(pWal
);
1803 testcase( szPage
<=32768 );
1804 testcase( szPage
>=65536 );
1805 pInfo
= walCkptInfo(pWal
);
1806 if( pInfo
->nBackfill
<pWal
->hdr
.mxFrame
){
1808 /* EVIDENCE-OF: R-62920-47450 The busy-handler callback is never invoked
1809 ** in the SQLITE_CHECKPOINT_PASSIVE mode. */
1810 assert( eMode
!=SQLITE_CHECKPOINT_PASSIVE
|| xBusy
==0 );
1812 /* Compute in mxSafeFrame the index of the last frame of the WAL that is
1813 ** safe to write into the database. Frames beyond mxSafeFrame might
1814 ** overwrite database pages that are in use by active readers and thus
1815 ** cannot be backfilled from the WAL.
1817 mxSafeFrame
= pWal
->hdr
.mxFrame
;
1818 mxPage
= pWal
->hdr
.nPage
;
1819 for(i
=1; i
<WAL_NREADER
; i
++){
1820 /* Thread-sanitizer reports that the following is an unsafe read,
1821 ** as some other thread may be in the process of updating the value
1822 ** of the aReadMark[] slot. The assumption here is that if that is
1823 ** happening, the other client may only be increasing the value,
1824 ** not decreasing it. So assuming either that either the "old" or
1825 ** "new" version of the value is read, and not some arbitrary value
1826 ** that would never be written by a real client, things are still
1829 ** Astute readers have pointed out that the assumption stated in the
1830 ** last sentence of the previous paragraph is not guaranteed to be
1831 ** true for all conforming systems. However, the assumption is true
1832 ** for all compilers and architectures in common use today (circa
1833 ** 2019-11-27) and the alternatives are both slow and complex, and
1834 ** so we will continue to go with the current design for now. If this
1835 ** bothers you, or if you really are running on a system where aligned
1836 ** 32-bit reads and writes are not atomic, then you can simply avoid
1837 ** the use of WAL mode, or only use WAL mode together with
1838 ** PRAGMA locking_mode=EXCLUSIVE and all will be well.
1840 u32 y
= pInfo
->aReadMark
[i
];
1841 if( mxSafeFrame
>y
){
1842 assert( y
<=pWal
->hdr
.mxFrame
);
1843 rc
= walBusyLock(pWal
, xBusy
, pBusyArg
, WAL_READ_LOCK(i
), 1);
1844 if( rc
==SQLITE_OK
){
1845 pInfo
->aReadMark
[i
] = (i
==1 ? mxSafeFrame
: READMARK_NOT_USED
);
1846 walUnlockExclusive(pWal
, WAL_READ_LOCK(i
), 1);
1847 }else if( rc
==SQLITE_BUSY
){
1851 goto walcheckpoint_out
;
1856 /* Allocate the iterator */
1857 if( pInfo
->nBackfill
<mxSafeFrame
){
1858 rc
= walIteratorInit(pWal
, pInfo
->nBackfill
, &pIter
);
1859 assert( rc
==SQLITE_OK
|| pIter
==0 );
1863 && (rc
= walBusyLock(pWal
, xBusy
, pBusyArg
, WAL_READ_LOCK(0),1))==SQLITE_OK
1865 u32 nBackfill
= pInfo
->nBackfill
;
1867 pInfo
->nBackfillAttempted
= mxSafeFrame
;
1869 /* Sync the WAL to disk */
1870 rc
= sqlite3OsSync(pWal
->pWalFd
, CKPT_SYNC_FLAGS(sync_flags
));
1872 /* If the database may grow as a result of this checkpoint, hint
1873 ** about the eventual size of the db file to the VFS layer.
1875 if( rc
==SQLITE_OK
){
1876 i64 nReq
= ((i64
)mxPage
* szPage
);
1877 i64 nSize
; /* Current size of database file */
1878 rc
= sqlite3OsFileSize(pWal
->pDbFd
, &nSize
);
1879 if( rc
==SQLITE_OK
&& nSize
<nReq
){
1880 sqlite3OsFileControlHint(pWal
->pDbFd
, SQLITE_FCNTL_SIZE_HINT
, &nReq
);
1885 /* Iterate through the contents of the WAL, copying data to the db file */
1886 while( rc
==SQLITE_OK
&& 0==walIteratorNext(pIter
, &iDbpage
, &iFrame
) ){
1888 assert( walFramePgno(pWal
, iFrame
)==iDbpage
);
1889 if( db
->u1
.isInterrupted
){
1890 rc
= db
->mallocFailed
? SQLITE_NOMEM_BKPT
: SQLITE_INTERRUPT
;
1893 if( iFrame
<=nBackfill
|| iFrame
>mxSafeFrame
|| iDbpage
>mxPage
){
1896 iOffset
= walFrameOffset(iFrame
, szPage
) + WAL_FRAME_HDRSIZE
;
1897 /* testcase( IS_BIG_INT(iOffset) ); // requires a 4GiB WAL file */
1898 rc
= sqlite3OsRead(pWal
->pWalFd
, zBuf
, szPage
, iOffset
);
1899 if( rc
!=SQLITE_OK
) break;
1900 iOffset
= (iDbpage
-1)*(i64
)szPage
;
1901 testcase( IS_BIG_INT(iOffset
) );
1902 rc
= sqlite3OsWrite(pWal
->pDbFd
, zBuf
, szPage
, iOffset
);
1903 if( rc
!=SQLITE_OK
) break;
1906 /* If work was actually accomplished... */
1907 if( rc
==SQLITE_OK
){
1908 if( mxSafeFrame
==walIndexHdr(pWal
)->mxFrame
){
1909 i64 szDb
= pWal
->hdr
.nPage
*(i64
)szPage
;
1910 testcase( IS_BIG_INT(szDb
) );
1911 rc
= sqlite3OsTruncate(pWal
->pDbFd
, szDb
);
1912 if( rc
==SQLITE_OK
){
1913 rc
= sqlite3OsSync(pWal
->pDbFd
, CKPT_SYNC_FLAGS(sync_flags
));
1916 if( rc
==SQLITE_OK
){
1917 rc
= sqlite3OsFileControl(pWal
->pDbFd
, SQLITE_FCNTL_CKPT_DONE
, 0);
1918 if( rc
==SQLITE_NOTFOUND
) rc
= SQLITE_OK
;
1920 if( rc
==SQLITE_OK
){
1921 pInfo
->nBackfill
= mxSafeFrame
;
1925 /* Release the reader lock held while backfilling */
1926 walUnlockExclusive(pWal
, WAL_READ_LOCK(0), 1);
1929 if( rc
==SQLITE_BUSY
){
1930 /* Reset the return code so as not to report a checkpoint failure
1931 ** just because there are active readers. */
1936 /* If this is an SQLITE_CHECKPOINT_RESTART or TRUNCATE operation, and the
1937 ** entire wal file has been copied into the database file, then block
1938 ** until all readers have finished using the wal file. This ensures that
1939 ** the next process to write to the database restarts the wal file.
1941 if( rc
==SQLITE_OK
&& eMode
!=SQLITE_CHECKPOINT_PASSIVE
){
1942 assert( pWal
->writeLock
);
1943 if( pInfo
->nBackfill
<pWal
->hdr
.mxFrame
){
1945 }else if( eMode
>=SQLITE_CHECKPOINT_RESTART
){
1947 sqlite3_randomness(4, &salt1
);
1948 assert( pInfo
->nBackfill
==pWal
->hdr
.mxFrame
);
1949 rc
= walBusyLock(pWal
, xBusy
, pBusyArg
, WAL_READ_LOCK(1), WAL_NREADER
-1);
1950 if( rc
==SQLITE_OK
){
1951 if( eMode
==SQLITE_CHECKPOINT_TRUNCATE
){
1952 /* IMPLEMENTATION-OF: R-44699-57140 This mode works the same way as
1953 ** SQLITE_CHECKPOINT_RESTART with the addition that it also
1954 ** truncates the log file to zero bytes just prior to a
1955 ** successful return.
1957 ** In theory, it might be safe to do this without updating the
1958 ** wal-index header in shared memory, as all subsequent reader or
1959 ** writer clients should see that the entire log file has been
1960 ** checkpointed and behave accordingly. This seems unsafe though,
1961 ** as it would leave the system in a state where the contents of
1962 ** the wal-index header do not match the contents of the
1963 ** file-system. To avoid this, update the wal-index header to
1964 ** indicate that the log file contains zero valid frames. */
1965 walRestartHdr(pWal
, salt1
);
1966 rc
= sqlite3OsTruncate(pWal
->pWalFd
, 0);
1968 walUnlockExclusive(pWal
, WAL_READ_LOCK(1), WAL_NREADER
-1);
1974 walIteratorFree(pIter
);
1979 ** If the WAL file is currently larger than nMax bytes in size, truncate
1980 ** it to exactly nMax bytes. If an error occurs while doing so, ignore it.
1982 static void walLimitSize(Wal
*pWal
, i64 nMax
){
1985 sqlite3BeginBenignMalloc();
1986 rx
= sqlite3OsFileSize(pWal
->pWalFd
, &sz
);
1987 if( rx
==SQLITE_OK
&& (sz
> nMax
) ){
1988 rx
= sqlite3OsTruncate(pWal
->pWalFd
, nMax
);
1990 sqlite3EndBenignMalloc();
1992 sqlite3_log(rx
, "cannot limit WAL size: %s", pWal
->zWalName
);
1997 ** Close a connection to a log file.
1999 int sqlite3WalClose(
2000 Wal
*pWal
, /* Wal to close */
2001 sqlite3
*db
, /* For interrupt flag */
2002 int sync_flags
, /* Flags to pass to OsSync() (or 0) */
2004 u8
*zBuf
/* Buffer of at least nBuf bytes */
2008 int isDelete
= 0; /* True to unlink wal and wal-index files */
2010 /* If an EXCLUSIVE lock can be obtained on the database file (using the
2011 ** ordinary, rollback-mode locking methods, this guarantees that the
2012 ** connection associated with this log file is the only connection to
2013 ** the database. In this case checkpoint the database and unlink both
2014 ** the wal and wal-index files.
2016 ** The EXCLUSIVE lock is not released before returning.
2019 && SQLITE_OK
==(rc
= sqlite3OsLock(pWal
->pDbFd
, SQLITE_LOCK_EXCLUSIVE
))
2021 if( pWal
->exclusiveMode
==WAL_NORMAL_MODE
){
2022 pWal
->exclusiveMode
= WAL_EXCLUSIVE_MODE
;
2024 rc
= sqlite3WalCheckpoint(pWal
, db
,
2025 SQLITE_CHECKPOINT_PASSIVE
, 0, 0, sync_flags
, nBuf
, zBuf
, 0, 0
2027 if( rc
==SQLITE_OK
){
2029 sqlite3OsFileControlHint(
2030 pWal
->pDbFd
, SQLITE_FCNTL_PERSIST_WAL
, &bPersist
2033 /* Try to delete the WAL file if the checkpoint completed and
2034 ** fsyned (rc==SQLITE_OK) and if we are not in persistent-wal
2035 ** mode (!bPersist) */
2037 }else if( pWal
->mxWalSize
>=0 ){
2038 /* Try to truncate the WAL file to zero bytes if the checkpoint
2039 ** completed and fsynced (rc==SQLITE_OK) and we are in persistent
2040 ** WAL mode (bPersist) and if the PRAGMA journal_size_limit is a
2041 ** non-negative value (pWal->mxWalSize>=0). Note that we truncate
2042 ** to zero bytes as truncating to the journal_size_limit might
2043 ** leave a corrupt WAL file on disk. */
2044 walLimitSize(pWal
, 0);
2049 walIndexClose(pWal
, isDelete
);
2050 sqlite3OsClose(pWal
->pWalFd
);
2052 sqlite3BeginBenignMalloc();
2053 sqlite3OsDelete(pWal
->pVfs
, pWal
->zWalName
, 0);
2054 sqlite3EndBenignMalloc();
2056 WALTRACE(("WAL%p: closed\n", pWal
));
2057 sqlite3_free((void *)pWal
->apWiData
);
2064 ** Try to read the wal-index header. Return 0 on success and 1 if
2065 ** there is a problem.
2067 ** The wal-index is in shared memory. Another thread or process might
2068 ** be writing the header at the same time this procedure is trying to
2069 ** read it, which might result in inconsistency. A dirty read is detected
2070 ** by verifying that both copies of the header are the same and also by
2071 ** a checksum on the header.
2073 ** If and only if the read is consistent and the header is different from
2074 ** pWal->hdr, then pWal->hdr is updated to the content of the new header
2075 ** and *pChanged is set to 1.
2077 ** If the checksum cannot be verified return non-zero. If the header
2078 ** is read successfully and the checksum verified, return zero.
2080 static int walIndexTryHdr(Wal
*pWal
, int *pChanged
){
2081 u32 aCksum
[2]; /* Checksum on the header content */
2082 WalIndexHdr h1
, h2
; /* Two copies of the header content */
2083 WalIndexHdr
volatile *aHdr
; /* Header in shared memory */
2085 /* The first page of the wal-index must be mapped at this point. */
2086 assert( pWal
->nWiData
>0 && pWal
->apWiData
[0] );
2088 /* Read the header. This might happen concurrently with a write to the
2089 ** same area of shared memory on a different CPU in a SMP,
2090 ** meaning it is possible that an inconsistent snapshot is read
2091 ** from the file. If this happens, return non-zero.
2093 ** There are two copies of the header at the beginning of the wal-index.
2094 ** When reading, read [0] first then [1]. Writes are in the reverse order.
2095 ** Memory barriers are used to prevent the compiler or the hardware from
2096 ** reordering the reads and writes.
2098 aHdr
= walIndexHdr(pWal
);
2099 memcpy(&h1
, (void *)&aHdr
[0], sizeof(h1
));
2100 walShmBarrier(pWal
);
2101 memcpy(&h2
, (void *)&aHdr
[1], sizeof(h2
));
2103 if( memcmp(&h1
, &h2
, sizeof(h1
))!=0 ){
2104 return 1; /* Dirty read */
2107 return 1; /* Malformed header - probably all zeros */
2109 walChecksumBytes(1, (u8
*)&h1
, sizeof(h1
)-sizeof(h1
.aCksum
), 0, aCksum
);
2110 if( aCksum
[0]!=h1
.aCksum
[0] || aCksum
[1]!=h1
.aCksum
[1] ){
2111 return 1; /* Checksum does not match */
2114 if( memcmp(&pWal
->hdr
, &h1
, sizeof(WalIndexHdr
)) ){
2116 memcpy(&pWal
->hdr
, &h1
, sizeof(WalIndexHdr
));
2117 pWal
->szPage
= (pWal
->hdr
.szPage
&0xfe00) + ((pWal
->hdr
.szPage
&0x0001)<<16);
2118 testcase( pWal
->szPage
<=32768 );
2119 testcase( pWal
->szPage
>=65536 );
2122 /* The header was successfully read. Return zero. */
2127 ** This is the value that walTryBeginRead returns when it needs to
2130 #define WAL_RETRY (-1)
2133 ** Read the wal-index header from the wal-index and into pWal->hdr.
2134 ** If the wal-header appears to be corrupt, try to reconstruct the
2135 ** wal-index from the WAL before returning.
2137 ** Set *pChanged to 1 if the wal-index header value in pWal->hdr is
2138 ** changed by this operation. If pWal->hdr is unchanged, set *pChanged
2141 ** If the wal-index header is successfully read, return SQLITE_OK.
2142 ** Otherwise an SQLite error code.
2144 static int walIndexReadHdr(Wal
*pWal
, int *pChanged
){
2145 int rc
; /* Return code */
2146 int badHdr
; /* True if a header read failed */
2147 volatile u32
*page0
; /* Chunk of wal-index containing header */
2149 /* Ensure that page 0 of the wal-index (the page that contains the
2150 ** wal-index header) is mapped. Return early if an error occurs here.
2153 rc
= walIndexPage(pWal
, 0, &page0
);
2154 if( rc
!=SQLITE_OK
){
2155 assert( rc
!=SQLITE_READONLY
); /* READONLY changed to OK in walIndexPage */
2156 if( rc
==SQLITE_READONLY_CANTINIT
){
2157 /* The SQLITE_READONLY_CANTINIT return means that the shared-memory
2158 ** was openable but is not writable, and this thread is unable to
2159 ** confirm that another write-capable connection has the shared-memory
2160 ** open, and hence the content of the shared-memory is unreliable,
2161 ** since the shared-memory might be inconsistent with the WAL file
2162 ** and there is no writer on hand to fix it. */
2164 assert( pWal
->writeLock
==0 );
2165 assert( pWal
->readOnly
& WAL_SHM_RDONLY
);
2166 pWal
->bShmUnreliable
= 1;
2167 pWal
->exclusiveMode
= WAL_HEAPMEMORY_MODE
;
2170 return rc
; /* Any other non-OK return is just an error */
2173 /* page0 can be NULL if the SHM is zero bytes in size and pWal->writeLock
2174 ** is zero, which prevents the SHM from growing */
2175 testcase( page0
!=0 );
2177 assert( page0
!=0 || pWal
->writeLock
==0 );
2179 /* If the first page of the wal-index has been mapped, try to read the
2180 ** wal-index header immediately, without holding any lock. This usually
2181 ** works, but may fail if the wal-index header is corrupt or currently
2182 ** being modified by another thread or process.
2184 badHdr
= (page0
? walIndexTryHdr(pWal
, pChanged
) : 1);
2186 /* If the first attempt failed, it might have been due to a race
2187 ** with a writer. So get a WRITE lock and try again.
2189 assert( badHdr
==0 || pWal
->writeLock
==0 );
2191 if( pWal
->bShmUnreliable
==0 && (pWal
->readOnly
& WAL_SHM_RDONLY
) ){
2192 if( SQLITE_OK
==(rc
= walLockShared(pWal
, WAL_WRITE_LOCK
)) ){
2193 walUnlockShared(pWal
, WAL_WRITE_LOCK
);
2194 rc
= SQLITE_READONLY_RECOVERY
;
2196 }else if( SQLITE_OK
==(rc
= walLockExclusive(pWal
, WAL_WRITE_LOCK
, 1)) ){
2197 pWal
->writeLock
= 1;
2198 if( SQLITE_OK
==(rc
= walIndexPage(pWal
, 0, &page0
)) ){
2199 badHdr
= walIndexTryHdr(pWal
, pChanged
);
2201 /* If the wal-index header is still malformed even while holding
2202 ** a WRITE lock, it can only mean that the header is corrupted and
2203 ** needs to be reconstructed. So run recovery to do exactly that.
2205 rc
= walIndexRecover(pWal
);
2209 pWal
->writeLock
= 0;
2210 walUnlockExclusive(pWal
, WAL_WRITE_LOCK
, 1);
2214 /* If the header is read successfully, check the version number to make
2215 ** sure the wal-index was not constructed with some future format that
2216 ** this version of SQLite cannot understand.
2218 if( badHdr
==0 && pWal
->hdr
.iVersion
!=WALINDEX_MAX_VERSION
){
2219 rc
= SQLITE_CANTOPEN_BKPT
;
2221 if( pWal
->bShmUnreliable
){
2222 if( rc
!=SQLITE_OK
){
2223 walIndexClose(pWal
, 0);
2224 pWal
->bShmUnreliable
= 0;
2225 assert( pWal
->nWiData
>0 && pWal
->apWiData
[0]==0 );
2226 /* walIndexRecover() might have returned SHORT_READ if a concurrent
2227 ** writer truncated the WAL out from under it. If that happens, it
2228 ** indicates that a writer has fixed the SHM file for us, so retry */
2229 if( rc
==SQLITE_IOERR_SHORT_READ
) rc
= WAL_RETRY
;
2231 pWal
->exclusiveMode
= WAL_NORMAL_MODE
;
2238 ** Open a transaction in a connection where the shared-memory is read-only
2239 ** and where we cannot verify that there is a separate write-capable connection
2240 ** on hand to keep the shared-memory up-to-date with the WAL file.
2242 ** This can happen, for example, when the shared-memory is implemented by
2243 ** memory-mapping a *-shm file, where a prior writer has shut down and
2244 ** left the *-shm file on disk, and now the present connection is trying
2245 ** to use that database but lacks write permission on the *-shm file.
2246 ** Other scenarios are also possible, depending on the VFS implementation.
2250 ** The *-wal file has been read and an appropriate wal-index has been
2251 ** constructed in pWal->apWiData[] using heap memory instead of shared
2254 ** If this function returns SQLITE_OK, then the read transaction has
2255 ** been successfully opened. In this case output variable (*pChanged)
2256 ** is set to true before returning if the caller should discard the
2257 ** contents of the page cache before proceeding. Or, if it returns
2258 ** WAL_RETRY, then the heap memory wal-index has been discarded and
2259 ** the caller should retry opening the read transaction from the
2260 ** beginning (including attempting to map the *-shm file).
2262 ** If an error occurs, an SQLite error code is returned.
2264 static int walBeginShmUnreliable(Wal
*pWal
, int *pChanged
){
2265 i64 szWal
; /* Size of wal file on disk in bytes */
2266 i64 iOffset
; /* Current offset when reading wal file */
2267 u8 aBuf
[WAL_HDRSIZE
]; /* Buffer to load WAL header into */
2268 u8
*aFrame
= 0; /* Malloc'd buffer to load entire frame */
2269 int szFrame
; /* Number of bytes in buffer aFrame[] */
2270 u8
*aData
; /* Pointer to data part of aFrame buffer */
2271 volatile void *pDummy
; /* Dummy argument for xShmMap */
2272 int rc
; /* Return code */
2273 u32 aSaveCksum
[2]; /* Saved copy of pWal->hdr.aFrameCksum */
2275 assert( pWal
->bShmUnreliable
);
2276 assert( pWal
->readOnly
& WAL_SHM_RDONLY
);
2277 assert( pWal
->nWiData
>0 && pWal
->apWiData
[0] );
2279 /* Take WAL_READ_LOCK(0). This has the effect of preventing any
2280 ** writers from running a checkpoint, but does not stop them
2281 ** from running recovery. */
2282 rc
= walLockShared(pWal
, WAL_READ_LOCK(0));
2283 if( rc
!=SQLITE_OK
){
2284 if( rc
==SQLITE_BUSY
) rc
= WAL_RETRY
;
2285 goto begin_unreliable_shm_out
;
2289 /* Check to see if a separate writer has attached to the shared-memory area,
2290 ** thus making the shared-memory "reliable" again. Do this by invoking
2291 ** the xShmMap() routine of the VFS and looking to see if the return
2292 ** is SQLITE_READONLY instead of SQLITE_READONLY_CANTINIT.
2294 ** If the shared-memory is now "reliable" return WAL_RETRY, which will
2295 ** cause the heap-memory WAL-index to be discarded and the actual
2296 ** shared memory to be used in its place.
2298 ** This step is important because, even though this connection is holding
2299 ** the WAL_READ_LOCK(0) which prevents a checkpoint, a writer might
2300 ** have already checkpointed the WAL file and, while the current
2301 ** is active, wrap the WAL and start overwriting frames that this
2302 ** process wants to use.
2304 ** Once sqlite3OsShmMap() has been called for an sqlite3_file and has
2305 ** returned any SQLITE_READONLY value, it must return only SQLITE_READONLY
2306 ** or SQLITE_READONLY_CANTINIT or some error for all subsequent invocations,
2307 ** even if some external agent does a "chmod" to make the shared-memory
2308 ** writable by us, until sqlite3OsShmUnmap() has been called.
2309 ** This is a requirement on the VFS implementation.
2311 rc
= sqlite3OsShmMap(pWal
->pDbFd
, 0, WALINDEX_PGSZ
, 0, &pDummy
);
2312 assert( rc
!=SQLITE_OK
); /* SQLITE_OK not possible for read-only connection */
2313 if( rc
!=SQLITE_READONLY_CANTINIT
){
2314 rc
= (rc
==SQLITE_READONLY
? WAL_RETRY
: rc
);
2315 goto begin_unreliable_shm_out
;
2318 /* We reach this point only if the real shared-memory is still unreliable.
2319 ** Assume the in-memory WAL-index substitute is correct and load it
2322 memcpy(&pWal
->hdr
, (void*)walIndexHdr(pWal
), sizeof(WalIndexHdr
));
2324 /* Make sure some writer hasn't come in and changed the WAL file out
2325 ** from under us, then disconnected, while we were not looking.
2327 rc
= sqlite3OsFileSize(pWal
->pWalFd
, &szWal
);
2328 if( rc
!=SQLITE_OK
){
2329 goto begin_unreliable_shm_out
;
2331 if( szWal
<WAL_HDRSIZE
){
2332 /* If the wal file is too small to contain a wal-header and the
2333 ** wal-index header has mxFrame==0, then it must be safe to proceed
2334 ** reading the database file only. However, the page cache cannot
2335 ** be trusted, as a read/write connection may have connected, written
2336 ** the db, run a checkpoint, truncated the wal file and disconnected
2337 ** since this client's last read transaction. */
2339 rc
= (pWal
->hdr
.mxFrame
==0 ? SQLITE_OK
: WAL_RETRY
);
2340 goto begin_unreliable_shm_out
;
2343 /* Check the salt keys at the start of the wal file still match. */
2344 rc
= sqlite3OsRead(pWal
->pWalFd
, aBuf
, WAL_HDRSIZE
, 0);
2345 if( rc
!=SQLITE_OK
){
2346 goto begin_unreliable_shm_out
;
2348 if( memcmp(&pWal
->hdr
.aSalt
, &aBuf
[16], 8) ){
2349 /* Some writer has wrapped the WAL file while we were not looking.
2350 ** Return WAL_RETRY which will cause the in-memory WAL-index to be
2353 goto begin_unreliable_shm_out
;
2356 /* Allocate a buffer to read frames into */
2357 szFrame
= pWal
->hdr
.szPage
+ WAL_FRAME_HDRSIZE
;
2358 aFrame
= (u8
*)sqlite3_malloc64(szFrame
);
2360 rc
= SQLITE_NOMEM_BKPT
;
2361 goto begin_unreliable_shm_out
;
2363 aData
= &aFrame
[WAL_FRAME_HDRSIZE
];
2365 /* Check to see if a complete transaction has been appended to the
2366 ** wal file since the heap-memory wal-index was created. If so, the
2367 ** heap-memory wal-index is discarded and WAL_RETRY returned to
2369 aSaveCksum
[0] = pWal
->hdr
.aFrameCksum
[0];
2370 aSaveCksum
[1] = pWal
->hdr
.aFrameCksum
[1];
2371 for(iOffset
=walFrameOffset(pWal
->hdr
.mxFrame
+1, pWal
->hdr
.szPage
);
2372 iOffset
+szFrame
<=szWal
;
2375 u32 pgno
; /* Database page number for frame */
2376 u32 nTruncate
; /* dbsize field from frame header */
2378 /* Read and decode the next log frame. */
2379 rc
= sqlite3OsRead(pWal
->pWalFd
, aFrame
, szFrame
, iOffset
);
2380 if( rc
!=SQLITE_OK
) break;
2381 if( !walDecodeFrame(pWal
, &pgno
, &nTruncate
, aData
, aFrame
) ) break;
2383 /* If nTruncate is non-zero, then a complete transaction has been
2384 ** appended to this wal file. Set rc to WAL_RETRY and break out of
2391 pWal
->hdr
.aFrameCksum
[0] = aSaveCksum
[0];
2392 pWal
->hdr
.aFrameCksum
[1] = aSaveCksum
[1];
2394 begin_unreliable_shm_out
:
2395 sqlite3_free(aFrame
);
2396 if( rc
!=SQLITE_OK
){
2398 for(i
=0; i
<pWal
->nWiData
; i
++){
2399 sqlite3_free((void*)pWal
->apWiData
[i
]);
2400 pWal
->apWiData
[i
] = 0;
2402 pWal
->bShmUnreliable
= 0;
2403 sqlite3WalEndReadTransaction(pWal
);
2410 ** Attempt to start a read transaction. This might fail due to a race or
2411 ** other transient condition. When that happens, it returns WAL_RETRY to
2412 ** indicate to the caller that it is safe to retry immediately.
2414 ** On success return SQLITE_OK. On a permanent failure (such an
2415 ** I/O error or an SQLITE_BUSY because another process is running
2416 ** recovery) return a positive error code.
2418 ** The useWal parameter is true to force the use of the WAL and disable
2419 ** the case where the WAL is bypassed because it has been completely
2420 ** checkpointed. If useWal==0 then this routine calls walIndexReadHdr()
2421 ** to make a copy of the wal-index header into pWal->hdr. If the
2422 ** wal-index header has changed, *pChanged is set to 1 (as an indication
2423 ** to the caller that the local page cache is obsolete and needs to be
2424 ** flushed.) When useWal==1, the wal-index header is assumed to already
2425 ** be loaded and the pChanged parameter is unused.
2427 ** The caller must set the cnt parameter to the number of prior calls to
2428 ** this routine during the current read attempt that returned WAL_RETRY.
2429 ** This routine will start taking more aggressive measures to clear the
2430 ** race conditions after multiple WAL_RETRY returns, and after an excessive
2431 ** number of errors will ultimately return SQLITE_PROTOCOL. The
2432 ** SQLITE_PROTOCOL return indicates that some other process has gone rogue
2433 ** and is not honoring the locking protocol. There is a vanishingly small
2434 ** chance that SQLITE_PROTOCOL could be returned because of a run of really
2435 ** bad luck when there is lots of contention for the wal-index, but that
2436 ** possibility is so small that it can be safely neglected, we believe.
2438 ** On success, this routine obtains a read lock on
2439 ** WAL_READ_LOCK(pWal->readLock). The pWal->readLock integer is
2440 ** in the range 0 <= pWal->readLock < WAL_NREADER. If pWal->readLock==(-1)
2441 ** that means the Wal does not hold any read lock. The reader must not
2442 ** access any database page that is modified by a WAL frame up to and
2443 ** including frame number aReadMark[pWal->readLock]. The reader will
2444 ** use WAL frames up to and including pWal->hdr.mxFrame if pWal->readLock>0
2445 ** Or if pWal->readLock==0, then the reader will ignore the WAL
2446 ** completely and get all content directly from the database file.
2447 ** If the useWal parameter is 1 then the WAL will never be ignored and
2448 ** this routine will always set pWal->readLock>0 on success.
2449 ** When the read transaction is completed, the caller must release the
2450 ** lock on WAL_READ_LOCK(pWal->readLock) and set pWal->readLock to -1.
2452 ** This routine uses the nBackfill and aReadMark[] fields of the header
2453 ** to select a particular WAL_READ_LOCK() that strives to let the
2454 ** checkpoint process do as much work as possible. This routine might
2455 ** update values of the aReadMark[] array in the header, but if it does
2456 ** so it takes care to hold an exclusive lock on the corresponding
2457 ** WAL_READ_LOCK() while changing values.
2459 static int walTryBeginRead(Wal
*pWal
, int *pChanged
, int useWal
, int cnt
){
2460 volatile WalCkptInfo
*pInfo
; /* Checkpoint information in wal-index */
2461 u32 mxReadMark
; /* Largest aReadMark[] value */
2462 int mxI
; /* Index of largest aReadMark[] value */
2463 int i
; /* Loop counter */
2464 int rc
= SQLITE_OK
; /* Return code */
2465 u32 mxFrame
; /* Wal frame to lock to */
2467 assert( pWal
->readLock
<0 ); /* Not currently locked */
2469 /* useWal may only be set for read/write connections */
2470 assert( (pWal
->readOnly
& WAL_SHM_RDONLY
)==0 || useWal
==0 );
2472 /* Take steps to avoid spinning forever if there is a protocol error.
2474 ** Circumstances that cause a RETRY should only last for the briefest
2475 ** instances of time. No I/O or other system calls are done while the
2476 ** locks are held, so the locks should not be held for very long. But
2477 ** if we are unlucky, another process that is holding a lock might get
2478 ** paged out or take a page-fault that is time-consuming to resolve,
2479 ** during the few nanoseconds that it is holding the lock. In that case,
2480 ** it might take longer than normal for the lock to free.
2482 ** After 5 RETRYs, we begin calling sqlite3OsSleep(). The first few
2483 ** calls to sqlite3OsSleep() have a delay of 1 microsecond. Really this
2484 ** is more of a scheduler yield than an actual delay. But on the 10th
2485 ** an subsequent retries, the delays start becoming longer and longer,
2486 ** so that on the 100th (and last) RETRY we delay for 323 milliseconds.
2487 ** The total delay time before giving up is less than 10 seconds.
2490 int nDelay
= 1; /* Pause time in microseconds */
2492 VVA_ONLY( pWal
->lockError
= 1; )
2493 return SQLITE_PROTOCOL
;
2495 if( cnt
>=10 ) nDelay
= (cnt
-9)*(cnt
-9)*39;
2496 sqlite3OsSleep(pWal
->pVfs
, nDelay
);
2500 assert( rc
==SQLITE_OK
);
2501 if( pWal
->bShmUnreliable
==0 ){
2502 rc
= walIndexReadHdr(pWal
, pChanged
);
2504 if( rc
==SQLITE_BUSY
){
2505 /* If there is not a recovery running in another thread or process
2506 ** then convert BUSY errors to WAL_RETRY. If recovery is known to
2507 ** be running, convert BUSY to BUSY_RECOVERY. There is a race here
2508 ** which might cause WAL_RETRY to be returned even if BUSY_RECOVERY
2509 ** would be technically correct. But the race is benign since with
2510 ** WAL_RETRY this routine will be called again and will probably be
2511 ** right on the second iteration.
2513 if( pWal
->apWiData
[0]==0 ){
2514 /* This branch is taken when the xShmMap() method returns SQLITE_BUSY.
2515 ** We assume this is a transient condition, so return WAL_RETRY. The
2516 ** xShmMap() implementation used by the default unix and win32 VFS
2517 ** modules may return SQLITE_BUSY due to a race condition in the
2518 ** code that determines whether or not the shared-memory region
2519 ** must be zeroed before the requested page is returned.
2522 }else if( SQLITE_OK
==(rc
= walLockShared(pWal
, WAL_RECOVER_LOCK
)) ){
2523 walUnlockShared(pWal
, WAL_RECOVER_LOCK
);
2525 }else if( rc
==SQLITE_BUSY
){
2526 rc
= SQLITE_BUSY_RECOVERY
;
2529 if( rc
!=SQLITE_OK
){
2532 else if( pWal
->bShmUnreliable
){
2533 return walBeginShmUnreliable(pWal
, pChanged
);
2537 assert( pWal
->nWiData
>0 );
2538 assert( pWal
->apWiData
[0]!=0 );
2539 pInfo
= walCkptInfo(pWal
);
2540 if( !useWal
&& pInfo
->nBackfill
==pWal
->hdr
.mxFrame
2541 #ifdef SQLITE_ENABLE_SNAPSHOT
2542 && (pWal
->pSnapshot
==0 || pWal
->hdr
.mxFrame
==0)
2545 /* The WAL has been completely backfilled (or it is empty).
2546 ** and can be safely ignored.
2548 rc
= walLockShared(pWal
, WAL_READ_LOCK(0));
2549 walShmBarrier(pWal
);
2550 if( rc
==SQLITE_OK
){
2551 if( memcmp((void *)walIndexHdr(pWal
), &pWal
->hdr
, sizeof(WalIndexHdr
)) ){
2552 /* It is not safe to allow the reader to continue here if frames
2553 ** may have been appended to the log before READ_LOCK(0) was obtained.
2554 ** When holding READ_LOCK(0), the reader ignores the entire log file,
2555 ** which implies that the database file contains a trustworthy
2556 ** snapshot. Since holding READ_LOCK(0) prevents a checkpoint from
2557 ** happening, this is usually correct.
2559 ** However, if frames have been appended to the log (or if the log
2560 ** is wrapped and written for that matter) before the READ_LOCK(0)
2561 ** is obtained, that is not necessarily true. A checkpointer may
2562 ** have started to backfill the appended frames but crashed before
2563 ** it finished. Leaving a corrupt image in the database file.
2565 walUnlockShared(pWal
, WAL_READ_LOCK(0));
2570 }else if( rc
!=SQLITE_BUSY
){
2575 /* If we get this far, it means that the reader will want to use
2576 ** the WAL to get at content from recent commits. The job now is
2577 ** to select one of the aReadMark[] entries that is closest to
2578 ** but not exceeding pWal->hdr.mxFrame and lock that entry.
2582 mxFrame
= pWal
->hdr
.mxFrame
;
2583 #ifdef SQLITE_ENABLE_SNAPSHOT
2584 if( pWal
->pSnapshot
&& pWal
->pSnapshot
->mxFrame
<mxFrame
){
2585 mxFrame
= pWal
->pSnapshot
->mxFrame
;
2588 for(i
=1; i
<WAL_NREADER
; i
++){
2589 u32 thisMark
= AtomicLoad(pInfo
->aReadMark
+i
);
2590 if( mxReadMark
<=thisMark
&& thisMark
<=mxFrame
){
2591 assert( thisMark
!=READMARK_NOT_USED
);
2592 mxReadMark
= thisMark
;
2596 if( (pWal
->readOnly
& WAL_SHM_RDONLY
)==0
2597 && (mxReadMark
<mxFrame
|| mxI
==0)
2599 for(i
=1; i
<WAL_NREADER
; i
++){
2600 rc
= walLockExclusive(pWal
, WAL_READ_LOCK(i
), 1);
2601 if( rc
==SQLITE_OK
){
2602 mxReadMark
= AtomicStore(pInfo
->aReadMark
+i
,mxFrame
);
2604 walUnlockExclusive(pWal
, WAL_READ_LOCK(i
), 1);
2606 }else if( rc
!=SQLITE_BUSY
){
2612 assert( rc
==SQLITE_BUSY
|| (pWal
->readOnly
& WAL_SHM_RDONLY
)!=0 );
2613 return rc
==SQLITE_BUSY
? WAL_RETRY
: SQLITE_READONLY_CANTINIT
;
2616 rc
= walLockShared(pWal
, WAL_READ_LOCK(mxI
));
2618 return rc
==SQLITE_BUSY
? WAL_RETRY
: rc
;
2620 /* Now that the read-lock has been obtained, check that neither the
2621 ** value in the aReadMark[] array or the contents of the wal-index
2622 ** header have changed.
2624 ** It is necessary to check that the wal-index header did not change
2625 ** between the time it was read and when the shared-lock was obtained
2626 ** on WAL_READ_LOCK(mxI) was obtained to account for the possibility
2627 ** that the log file may have been wrapped by a writer, or that frames
2628 ** that occur later in the log than pWal->hdr.mxFrame may have been
2629 ** copied into the database by a checkpointer. If either of these things
2630 ** happened, then reading the database with the current value of
2631 ** pWal->hdr.mxFrame risks reading a corrupted snapshot. So, retry
2634 ** Before checking that the live wal-index header has not changed
2635 ** since it was read, set Wal.minFrame to the first frame in the wal
2636 ** file that has not yet been checkpointed. This client will not need
2637 ** to read any frames earlier than minFrame from the wal file - they
2638 ** can be safely read directly from the database file.
2640 ** Because a ShmBarrier() call is made between taking the copy of
2641 ** nBackfill and checking that the wal-header in shared-memory still
2642 ** matches the one cached in pWal->hdr, it is guaranteed that the
2643 ** checkpointer that set nBackfill was not working with a wal-index
2644 ** header newer than that cached in pWal->hdr. If it were, that could
2645 ** cause a problem. The checkpointer could omit to checkpoint
2646 ** a version of page X that lies before pWal->minFrame (call that version
2647 ** A) on the basis that there is a newer version (version B) of the same
2648 ** page later in the wal file. But if version B happens to like past
2649 ** frame pWal->hdr.mxFrame - then the client would incorrectly assume
2650 ** that it can read version A from the database file. However, since
2651 ** we can guarantee that the checkpointer that set nBackfill could not
2652 ** see any pages past pWal->hdr.mxFrame, this problem does not come up.
2654 pWal
->minFrame
= AtomicLoad(&pInfo
->nBackfill
)+1;
2655 walShmBarrier(pWal
);
2656 if( AtomicLoad(pInfo
->aReadMark
+mxI
)!=mxReadMark
2657 || memcmp((void *)walIndexHdr(pWal
), &pWal
->hdr
, sizeof(WalIndexHdr
))
2659 walUnlockShared(pWal
, WAL_READ_LOCK(mxI
));
2662 assert( mxReadMark
<=pWal
->hdr
.mxFrame
);
2663 pWal
->readLock
= (i16
)mxI
;
2668 #ifdef SQLITE_ENABLE_SNAPSHOT
2670 ** Attempt to reduce the value of the WalCkptInfo.nBackfillAttempted
2671 ** variable so that older snapshots can be accessed. To do this, loop
2672 ** through all wal frames from nBackfillAttempted to (nBackfill+1),
2673 ** comparing their content to the corresponding page with the database
2674 ** file, if any. Set nBackfillAttempted to the frame number of the
2675 ** first frame for which the wal file content matches the db file.
2677 ** This is only really safe if the file-system is such that any page
2678 ** writes made by earlier checkpointers were atomic operations, which
2679 ** is not always true. It is also possible that nBackfillAttempted
2680 ** may be left set to a value larger than expected, if a wal frame
2681 ** contains content that duplicate of an earlier version of the same
2684 ** SQLITE_OK is returned if successful, or an SQLite error code if an
2685 ** error occurs. It is not an error if nBackfillAttempted cannot be
2686 ** decreased at all.
2688 int sqlite3WalSnapshotRecover(Wal
*pWal
){
2691 assert( pWal
->readLock
>=0 );
2692 rc
= walLockExclusive(pWal
, WAL_CKPT_LOCK
, 1);
2693 if( rc
==SQLITE_OK
){
2694 volatile WalCkptInfo
*pInfo
= walCkptInfo(pWal
);
2695 int szPage
= (int)pWal
->szPage
;
2696 i64 szDb
; /* Size of db file in bytes */
2698 rc
= sqlite3OsFileSize(pWal
->pDbFd
, &szDb
);
2699 if( rc
==SQLITE_OK
){
2700 void *pBuf1
= sqlite3_malloc(szPage
);
2701 void *pBuf2
= sqlite3_malloc(szPage
);
2702 if( pBuf1
==0 || pBuf2
==0 ){
2705 u32 i
= pInfo
->nBackfillAttempted
;
2706 for(i
=pInfo
->nBackfillAttempted
; i
>pInfo
->nBackfill
; i
--){
2707 WalHashLoc sLoc
; /* Hash table location */
2708 u32 pgno
; /* Page number in db file */
2709 i64 iDbOff
; /* Offset of db file entry */
2710 i64 iWalOff
; /* Offset of wal file entry */
2712 rc
= walHashGet(pWal
, walFramePage(i
), &sLoc
);
2713 if( rc
!=SQLITE_OK
) break;
2714 pgno
= sLoc
.aPgno
[i
-sLoc
.iZero
];
2715 iDbOff
= (i64
)(pgno
-1) * szPage
;
2717 if( iDbOff
+szPage
<=szDb
){
2718 iWalOff
= walFrameOffset(i
, szPage
) + WAL_FRAME_HDRSIZE
;
2719 rc
= sqlite3OsRead(pWal
->pWalFd
, pBuf1
, szPage
, iWalOff
);
2721 if( rc
==SQLITE_OK
){
2722 rc
= sqlite3OsRead(pWal
->pDbFd
, pBuf2
, szPage
, iDbOff
);
2725 if( rc
!=SQLITE_OK
|| 0==memcmp(pBuf1
, pBuf2
, szPage
) ){
2730 pInfo
->nBackfillAttempted
= i
-1;
2734 sqlite3_free(pBuf1
);
2735 sqlite3_free(pBuf2
);
2737 walUnlockExclusive(pWal
, WAL_CKPT_LOCK
, 1);
2742 #endif /* SQLITE_ENABLE_SNAPSHOT */
2745 ** Begin a read transaction on the database.
2747 ** This routine used to be called sqlite3OpenSnapshot() and with good reason:
2748 ** it takes a snapshot of the state of the WAL and wal-index for the current
2749 ** instant in time. The current thread will continue to use this snapshot.
2750 ** Other threads might append new content to the WAL and wal-index but
2751 ** that extra content is ignored by the current thread.
2753 ** If the database contents have changes since the previous read
2754 ** transaction, then *pChanged is set to 1 before returning. The
2755 ** Pager layer will use this to know that its cache is stale and
2756 ** needs to be flushed.
2758 int sqlite3WalBeginReadTransaction(Wal
*pWal
, int *pChanged
){
2759 int rc
; /* Return code */
2760 int cnt
= 0; /* Number of TryBeginRead attempts */
2762 #ifdef SQLITE_ENABLE_SNAPSHOT
2764 WalIndexHdr
*pSnapshot
= pWal
->pSnapshot
;
2765 if( pSnapshot
&& memcmp(pSnapshot
, &pWal
->hdr
, sizeof(WalIndexHdr
))!=0 ){
2771 rc
= walTryBeginRead(pWal
, pChanged
, 0, ++cnt
);
2772 }while( rc
==WAL_RETRY
);
2773 testcase( (rc
&0xff)==SQLITE_BUSY
);
2774 testcase( (rc
&0xff)==SQLITE_IOERR
);
2775 testcase( rc
==SQLITE_PROTOCOL
);
2776 testcase( rc
==SQLITE_OK
);
2778 #ifdef SQLITE_ENABLE_SNAPSHOT
2779 if( rc
==SQLITE_OK
){
2780 if( pSnapshot
&& memcmp(pSnapshot
, &pWal
->hdr
, sizeof(WalIndexHdr
))!=0 ){
2781 /* At this point the client has a lock on an aReadMark[] slot holding
2782 ** a value equal to or smaller than pSnapshot->mxFrame, but pWal->hdr
2783 ** is populated with the wal-index header corresponding to the head
2784 ** of the wal file. Verify that pSnapshot is still valid before
2785 ** continuing. Reasons why pSnapshot might no longer be valid:
2787 ** (1) The WAL file has been reset since the snapshot was taken.
2788 ** In this case, the salt will have changed.
2790 ** (2) A checkpoint as been attempted that wrote frames past
2791 ** pSnapshot->mxFrame into the database file. Note that the
2792 ** checkpoint need not have completed for this to cause problems.
2794 volatile WalCkptInfo
*pInfo
= walCkptInfo(pWal
);
2796 assert( pWal
->readLock
>0 || pWal
->hdr
.mxFrame
==0 );
2797 assert( pInfo
->aReadMark
[pWal
->readLock
]<=pSnapshot
->mxFrame
);
2799 /* It is possible that there is a checkpointer thread running
2800 ** concurrent with this code. If this is the case, it may be that the
2801 ** checkpointer has already determined that it will checkpoint
2802 ** snapshot X, where X is later in the wal file than pSnapshot, but
2803 ** has not yet set the pInfo->nBackfillAttempted variable to indicate
2804 ** its intent. To avoid the race condition this leads to, ensure that
2805 ** there is no checkpointer process by taking a shared CKPT lock
2806 ** before checking pInfo->nBackfillAttempted.
2808 ** TODO: Does the aReadMark[] lock prevent a checkpointer from doing
2811 rc
= walLockShared(pWal
, WAL_CKPT_LOCK
);
2813 if( rc
==SQLITE_OK
){
2814 /* Check that the wal file has not been wrapped. Assuming that it has
2815 ** not, also check that no checkpointer has attempted to checkpoint any
2816 ** frames beyond pSnapshot->mxFrame. If either of these conditions are
2817 ** true, return SQLITE_ERROR_SNAPSHOT. Otherwise, overwrite pWal->hdr
2818 ** with *pSnapshot and set *pChanged as appropriate for opening the
2820 if( !memcmp(pSnapshot
->aSalt
, pWal
->hdr
.aSalt
, sizeof(pWal
->hdr
.aSalt
))
2821 && pSnapshot
->mxFrame
>=pInfo
->nBackfillAttempted
2823 assert( pWal
->readLock
>0 );
2824 memcpy(&pWal
->hdr
, pSnapshot
, sizeof(WalIndexHdr
));
2825 *pChanged
= bChanged
;
2827 rc
= SQLITE_ERROR_SNAPSHOT
;
2830 /* Release the shared CKPT lock obtained above. */
2831 walUnlockShared(pWal
, WAL_CKPT_LOCK
);
2836 if( rc
!=SQLITE_OK
){
2837 sqlite3WalEndReadTransaction(pWal
);
2846 ** Finish with a read transaction. All this does is release the
2849 void sqlite3WalEndReadTransaction(Wal
*pWal
){
2850 sqlite3WalEndWriteTransaction(pWal
);
2851 if( pWal
->readLock
>=0 ){
2852 walUnlockShared(pWal
, WAL_READ_LOCK(pWal
->readLock
));
2853 pWal
->readLock
= -1;
2858 ** Search the wal file for page pgno. If found, set *piRead to the frame that
2859 ** contains the page. Otherwise, if pgno is not in the wal file, set *piRead
2862 ** Return SQLITE_OK if successful, or an error code if an error occurs. If an
2863 ** error does occur, the final value of *piRead is undefined.
2865 int sqlite3WalFindFrame(
2866 Wal
*pWal
, /* WAL handle */
2867 Pgno pgno
, /* Database page number to read data for */
2868 u32
*piRead
/* OUT: Frame number (or zero) */
2870 u32 iRead
= 0; /* If !=0, WAL frame to return data from */
2871 u32 iLast
= pWal
->hdr
.mxFrame
; /* Last page in WAL for this reader */
2872 int iHash
; /* Used to loop through N hash tables */
2875 /* This routine is only be called from within a read transaction. */
2876 assert( pWal
->readLock
>=0 || pWal
->lockError
);
2878 /* If the "last page" field of the wal-index header snapshot is 0, then
2879 ** no data will be read from the wal under any circumstances. Return early
2880 ** in this case as an optimization. Likewise, if pWal->readLock==0,
2881 ** then the WAL is ignored by the reader so return early, as if the
2884 if( iLast
==0 || (pWal
->readLock
==0 && pWal
->bShmUnreliable
==0) ){
2889 /* Search the hash table or tables for an entry matching page number
2890 ** pgno. Each iteration of the following for() loop searches one
2891 ** hash table (each hash table indexes up to HASHTABLE_NPAGE frames).
2893 ** This code might run concurrently to the code in walIndexAppend()
2894 ** that adds entries to the wal-index (and possibly to this hash
2895 ** table). This means the value just read from the hash
2896 ** slot (aHash[iKey]) may have been added before or after the
2897 ** current read transaction was opened. Values added after the
2898 ** read transaction was opened may have been written incorrectly -
2899 ** i.e. these slots may contain garbage data. However, we assume
2900 ** that any slots written before the current read transaction was
2901 ** opened remain unmodified.
2903 ** For the reasons above, the if(...) condition featured in the inner
2904 ** loop of the following block is more stringent that would be required
2905 ** if we had exclusive access to the hash-table:
2907 ** (aPgno[iFrame]==pgno):
2908 ** This condition filters out normal hash-table collisions.
2911 ** This condition filters out entries that were added to the hash
2912 ** table after the current read-transaction had started.
2914 iMinHash
= walFramePage(pWal
->minFrame
);
2915 for(iHash
=walFramePage(iLast
); iHash
>=iMinHash
; iHash
--){
2916 WalHashLoc sLoc
; /* Hash table location */
2917 int iKey
; /* Hash slot index */
2918 int nCollide
; /* Number of hash collisions remaining */
2919 int rc
; /* Error code */
2921 rc
= walHashGet(pWal
, iHash
, &sLoc
);
2922 if( rc
!=SQLITE_OK
){
2925 nCollide
= HASHTABLE_NSLOT
;
2926 for(iKey
=walHash(pgno
); sLoc
.aHash
[iKey
]; iKey
=walNextHash(iKey
)){
2927 u32 iH
= sLoc
.aHash
[iKey
];
2928 u32 iFrame
= iH
+ sLoc
.iZero
;
2929 if( iFrame
<=iLast
&& iFrame
>=pWal
->minFrame
&& sLoc
.aPgno
[iH
]==pgno
){
2930 assert( iFrame
>iRead
|| CORRUPT_DB
);
2933 if( (nCollide
--)==0 ){
2934 return SQLITE_CORRUPT_BKPT
;
2940 #ifdef SQLITE_ENABLE_EXPENSIVE_ASSERT
2941 /* If expensive assert() statements are available, do a linear search
2942 ** of the wal-index file content. Make sure the results agree with the
2943 ** result obtained using the hash indexes above. */
2947 assert( pWal
->bShmUnreliable
|| pWal
->minFrame
>0 );
2948 for(iTest
=iLast
; iTest
>=pWal
->minFrame
&& iTest
>0; iTest
--){
2949 if( walFramePgno(pWal
, iTest
)==pgno
){
2954 assert( iRead
==iRead2
);
2963 ** Read the contents of frame iRead from the wal file into buffer pOut
2964 ** (which is nOut bytes in size). Return SQLITE_OK if successful, or an
2965 ** error code otherwise.
2967 int sqlite3WalReadFrame(
2968 Wal
*pWal
, /* WAL handle */
2969 u32 iRead
, /* Frame to read */
2970 int nOut
, /* Size of buffer pOut in bytes */
2971 u8
*pOut
/* Buffer to write page data to */
2975 sz
= pWal
->hdr
.szPage
;
2976 sz
= (sz
&0xfe00) + ((sz
&0x0001)<<16);
2977 testcase( sz
<=32768 );
2978 testcase( sz
>=65536 );
2979 iOffset
= walFrameOffset(iRead
, sz
) + WAL_FRAME_HDRSIZE
;
2980 /* testcase( IS_BIG_INT(iOffset) ); // requires a 4GiB WAL */
2981 return sqlite3OsRead(pWal
->pWalFd
, pOut
, (nOut
>sz
? sz
: nOut
), iOffset
);
2985 ** Return the size of the database in pages (or zero, if unknown).
2987 Pgno
sqlite3WalDbsize(Wal
*pWal
){
2988 if( pWal
&& ALWAYS(pWal
->readLock
>=0) ){
2989 return pWal
->hdr
.nPage
;
2996 ** This function starts a write transaction on the WAL.
2998 ** A read transaction must have already been started by a prior call
2999 ** to sqlite3WalBeginReadTransaction().
3001 ** If another thread or process has written into the database since
3002 ** the read transaction was started, then it is not possible for this
3003 ** thread to write as doing so would cause a fork. So this routine
3004 ** returns SQLITE_BUSY in that case and no write transaction is started.
3006 ** There can only be a single writer active at a time.
3008 int sqlite3WalBeginWriteTransaction(Wal
*pWal
){
3011 /* Cannot start a write transaction without first holding a read
3013 assert( pWal
->readLock
>=0 );
3014 assert( pWal
->writeLock
==0 && pWal
->iReCksum
==0 );
3016 if( pWal
->readOnly
){
3017 return SQLITE_READONLY
;
3020 /* Only one writer allowed at a time. Get the write lock. Return
3021 ** SQLITE_BUSY if unable.
3023 rc
= walLockExclusive(pWal
, WAL_WRITE_LOCK
, 1);
3027 pWal
->writeLock
= 1;
3029 /* If another connection has written to the database file since the
3030 ** time the read transaction on this connection was started, then
3031 ** the write is disallowed.
3033 if( memcmp(&pWal
->hdr
, (void *)walIndexHdr(pWal
), sizeof(WalIndexHdr
))!=0 ){
3034 walUnlockExclusive(pWal
, WAL_WRITE_LOCK
, 1);
3035 pWal
->writeLock
= 0;
3036 rc
= SQLITE_BUSY_SNAPSHOT
;
3043 ** End a write transaction. The commit has already been done. This
3044 ** routine merely releases the lock.
3046 int sqlite3WalEndWriteTransaction(Wal
*pWal
){
3047 if( pWal
->writeLock
){
3048 walUnlockExclusive(pWal
, WAL_WRITE_LOCK
, 1);
3049 pWal
->writeLock
= 0;
3051 pWal
->truncateOnCommit
= 0;
3057 ** If any data has been written (but not committed) to the log file, this
3058 ** function moves the write-pointer back to the start of the transaction.
3060 ** Additionally, the callback function is invoked for each frame written
3061 ** to the WAL since the start of the transaction. If the callback returns
3062 ** other than SQLITE_OK, it is not invoked again and the error code is
3063 ** returned to the caller.
3065 ** Otherwise, if the callback function does not return an error, this
3066 ** function returns SQLITE_OK.
3068 int sqlite3WalUndo(Wal
*pWal
, int (*xUndo
)(void *, Pgno
), void *pUndoCtx
){
3070 if( ALWAYS(pWal
->writeLock
) ){
3071 Pgno iMax
= pWal
->hdr
.mxFrame
;
3074 /* Restore the clients cache of the wal-index header to the state it
3075 ** was in before the client began writing to the database.
3077 memcpy(&pWal
->hdr
, (void *)walIndexHdr(pWal
), sizeof(WalIndexHdr
));
3079 for(iFrame
=pWal
->hdr
.mxFrame
+1;
3080 ALWAYS(rc
==SQLITE_OK
) && iFrame
<=iMax
;
3083 /* This call cannot fail. Unless the page for which the page number
3084 ** is passed as the second argument is (a) in the cache and
3085 ** (b) has an outstanding reference, then xUndo is either a no-op
3086 ** (if (a) is false) or simply expels the page from the cache (if (b)
3089 ** If the upper layer is doing a rollback, it is guaranteed that there
3090 ** are no outstanding references to any page other than page 1. And
3091 ** page 1 is never written to the log until the transaction is
3092 ** committed. As a result, the call to xUndo may not fail.
3094 assert( walFramePgno(pWal
, iFrame
)!=1 );
3095 rc
= xUndo(pUndoCtx
, walFramePgno(pWal
, iFrame
));
3097 if( iMax
!=pWal
->hdr
.mxFrame
) walCleanupHash(pWal
);
3103 ** Argument aWalData must point to an array of WAL_SAVEPOINT_NDATA u32
3104 ** values. This function populates the array with values required to
3105 ** "rollback" the write position of the WAL handle back to the current
3106 ** point in the event of a savepoint rollback (via WalSavepointUndo()).
3108 void sqlite3WalSavepoint(Wal
*pWal
, u32
*aWalData
){
3109 assert( pWal
->writeLock
);
3110 aWalData
[0] = pWal
->hdr
.mxFrame
;
3111 aWalData
[1] = pWal
->hdr
.aFrameCksum
[0];
3112 aWalData
[2] = pWal
->hdr
.aFrameCksum
[1];
3113 aWalData
[3] = pWal
->nCkpt
;
3117 ** Move the write position of the WAL back to the point identified by
3118 ** the values in the aWalData[] array. aWalData must point to an array
3119 ** of WAL_SAVEPOINT_NDATA u32 values that has been previously populated
3120 ** by a call to WalSavepoint().
3122 int sqlite3WalSavepointUndo(Wal
*pWal
, u32
*aWalData
){
3125 assert( pWal
->writeLock
);
3126 assert( aWalData
[3]!=pWal
->nCkpt
|| aWalData
[0]<=pWal
->hdr
.mxFrame
);
3128 if( aWalData
[3]!=pWal
->nCkpt
){
3129 /* This savepoint was opened immediately after the write-transaction
3130 ** was started. Right after that, the writer decided to wrap around
3131 ** to the start of the log. Update the savepoint values to match.
3134 aWalData
[3] = pWal
->nCkpt
;
3137 if( aWalData
[0]<pWal
->hdr
.mxFrame
){
3138 pWal
->hdr
.mxFrame
= aWalData
[0];
3139 pWal
->hdr
.aFrameCksum
[0] = aWalData
[1];
3140 pWal
->hdr
.aFrameCksum
[1] = aWalData
[2];
3141 walCleanupHash(pWal
);
3148 ** This function is called just before writing a set of frames to the log
3149 ** file (see sqlite3WalFrames()). It checks to see if, instead of appending
3150 ** to the current log file, it is possible to overwrite the start of the
3151 ** existing log file with the new frames (i.e. "reset" the log). If so,
3152 ** it sets pWal->hdr.mxFrame to 0. Otherwise, pWal->hdr.mxFrame is left
3155 ** SQLITE_OK is returned if no error is encountered (regardless of whether
3156 ** or not pWal->hdr.mxFrame is modified). An SQLite error code is returned
3157 ** if an error occurs.
3159 static int walRestartLog(Wal
*pWal
){
3163 if( pWal
->readLock
==0 ){
3164 volatile WalCkptInfo
*pInfo
= walCkptInfo(pWal
);
3165 assert( pInfo
->nBackfill
==pWal
->hdr
.mxFrame
);
3166 if( pInfo
->nBackfill
>0 ){
3168 sqlite3_randomness(4, &salt1
);
3169 rc
= walLockExclusive(pWal
, WAL_READ_LOCK(1), WAL_NREADER
-1);
3170 if( rc
==SQLITE_OK
){
3171 /* If all readers are using WAL_READ_LOCK(0) (in other words if no
3172 ** readers are currently using the WAL), then the transactions
3173 ** frames will overwrite the start of the existing log. Update the
3174 ** wal-index header to reflect this.
3176 ** In theory it would be Ok to update the cache of the header only
3177 ** at this point. But updating the actual wal-index header is also
3178 ** safe and means there is no special case for sqlite3WalUndo()
3179 ** to handle if this transaction is rolled back. */
3180 walRestartHdr(pWal
, salt1
);
3181 walUnlockExclusive(pWal
, WAL_READ_LOCK(1), WAL_NREADER
-1);
3182 }else if( rc
!=SQLITE_BUSY
){
3186 walUnlockShared(pWal
, WAL_READ_LOCK(0));
3187 pWal
->readLock
= -1;
3191 rc
= walTryBeginRead(pWal
, ¬Used
, 1, ++cnt
);
3192 }while( rc
==WAL_RETRY
);
3193 assert( (rc
&0xff)!=SQLITE_BUSY
); /* BUSY not possible when useWal==1 */
3194 testcase( (rc
&0xff)==SQLITE_IOERR
);
3195 testcase( rc
==SQLITE_PROTOCOL
);
3196 testcase( rc
==SQLITE_OK
);
3202 ** Information about the current state of the WAL file and where
3203 ** the next fsync should occur - passed from sqlite3WalFrames() into
3206 typedef struct WalWriter
{
3207 Wal
*pWal
; /* The complete WAL information */
3208 sqlite3_file
*pFd
; /* The WAL file to which we write */
3209 sqlite3_int64 iSyncPoint
; /* Fsync at this offset */
3210 int syncFlags
; /* Flags for the fsync */
3211 int szPage
; /* Size of one page */
3215 ** Write iAmt bytes of content into the WAL file beginning at iOffset.
3216 ** Do a sync when crossing the p->iSyncPoint boundary.
3218 ** In other words, if iSyncPoint is in between iOffset and iOffset+iAmt,
3219 ** first write the part before iSyncPoint, then sync, then write the
3222 static int walWriteToLog(
3223 WalWriter
*p
, /* WAL to write to */
3224 void *pContent
, /* Content to be written */
3225 int iAmt
, /* Number of bytes to write */
3226 sqlite3_int64 iOffset
/* Start writing at this offset */
3229 if( iOffset
<p
->iSyncPoint
&& iOffset
+iAmt
>=p
->iSyncPoint
){
3230 int iFirstAmt
= (int)(p
->iSyncPoint
- iOffset
);
3231 rc
= sqlite3OsWrite(p
->pFd
, pContent
, iFirstAmt
, iOffset
);
3233 iOffset
+= iFirstAmt
;
3235 pContent
= (void*)(iFirstAmt
+ (char*)pContent
);
3236 assert( WAL_SYNC_FLAGS(p
->syncFlags
)!=0 );
3237 rc
= sqlite3OsSync(p
->pFd
, WAL_SYNC_FLAGS(p
->syncFlags
));
3238 if( iAmt
==0 || rc
) return rc
;
3240 rc
= sqlite3OsWrite(p
->pFd
, pContent
, iAmt
, iOffset
);
3245 ** Write out a single frame of the WAL
3247 static int walWriteOneFrame(
3248 WalWriter
*p
, /* Where to write the frame */
3249 PgHdr
*pPage
, /* The page of the frame to be written */
3250 int nTruncate
, /* The commit flag. Usually 0. >0 for commit */
3251 sqlite3_int64 iOffset
/* Byte offset at which to write */
3253 int rc
; /* Result code from subfunctions */
3254 void *pData
; /* Data actually written */
3255 u8 aFrame
[WAL_FRAME_HDRSIZE
]; /* Buffer to assemble frame-header in */
3256 pData
= pPage
->pData
;
3257 walEncodeFrame(p
->pWal
, pPage
->pgno
, nTruncate
, pData
, aFrame
);
3258 rc
= walWriteToLog(p
, aFrame
, sizeof(aFrame
), iOffset
);
3260 /* Write the page data */
3261 rc
= walWriteToLog(p
, pData
, p
->szPage
, iOffset
+sizeof(aFrame
));
3266 ** This function is called as part of committing a transaction within which
3267 ** one or more frames have been overwritten. It updates the checksums for
3268 ** all frames written to the wal file by the current transaction starting
3269 ** with the earliest to have been overwritten.
3271 ** SQLITE_OK is returned if successful, or an SQLite error code otherwise.
3273 static int walRewriteChecksums(Wal
*pWal
, u32 iLast
){
3274 const int szPage
= pWal
->szPage
;/* Database page size */
3275 int rc
= SQLITE_OK
; /* Return code */
3276 u8
*aBuf
; /* Buffer to load data from wal file into */
3277 u8 aFrame
[WAL_FRAME_HDRSIZE
]; /* Buffer to assemble frame-headers in */
3278 u32 iRead
; /* Next frame to read from wal file */
3281 aBuf
= sqlite3_malloc(szPage
+ WAL_FRAME_HDRSIZE
);
3282 if( aBuf
==0 ) return SQLITE_NOMEM_BKPT
;
3284 /* Find the checksum values to use as input for the recalculating the
3285 ** first checksum. If the first frame is frame 1 (implying that the current
3286 ** transaction restarted the wal file), these values must be read from the
3287 ** wal-file header. Otherwise, read them from the frame header of the
3288 ** previous frame. */
3289 assert( pWal
->iReCksum
>0 );
3290 if( pWal
->iReCksum
==1 ){
3293 iCksumOff
= walFrameOffset(pWal
->iReCksum
-1, szPage
) + 16;
3295 rc
= sqlite3OsRead(pWal
->pWalFd
, aBuf
, sizeof(u32
)*2, iCksumOff
);
3296 pWal
->hdr
.aFrameCksum
[0] = sqlite3Get4byte(aBuf
);
3297 pWal
->hdr
.aFrameCksum
[1] = sqlite3Get4byte(&aBuf
[sizeof(u32
)]);
3299 iRead
= pWal
->iReCksum
;
3301 for(; rc
==SQLITE_OK
&& iRead
<=iLast
; iRead
++){
3302 i64 iOff
= walFrameOffset(iRead
, szPage
);
3303 rc
= sqlite3OsRead(pWal
->pWalFd
, aBuf
, szPage
+WAL_FRAME_HDRSIZE
, iOff
);
3304 if( rc
==SQLITE_OK
){
3306 iPgno
= sqlite3Get4byte(aBuf
);
3307 nDbSize
= sqlite3Get4byte(&aBuf
[4]);
3309 walEncodeFrame(pWal
, iPgno
, nDbSize
, &aBuf
[WAL_FRAME_HDRSIZE
], aFrame
);
3310 rc
= sqlite3OsWrite(pWal
->pWalFd
, aFrame
, sizeof(aFrame
), iOff
);
3319 ** Write a set of frames to the log. The caller must hold the write-lock
3320 ** on the log file (obtained using sqlite3WalBeginWriteTransaction()).
3322 int sqlite3WalFrames(
3323 Wal
*pWal
, /* Wal handle to write to */
3324 int szPage
, /* Database page-size in bytes */
3325 PgHdr
*pList
, /* List of dirty pages to write */
3326 Pgno nTruncate
, /* Database size after this commit */
3327 int isCommit
, /* True if this is a commit */
3328 int sync_flags
/* Flags to pass to OsSync() (or 0) */
3330 int rc
; /* Used to catch return codes */
3331 u32 iFrame
; /* Next frame address */
3332 PgHdr
*p
; /* Iterator to run through pList with. */
3333 PgHdr
*pLast
= 0; /* Last frame in list */
3334 int nExtra
= 0; /* Number of extra copies of last page */
3335 int szFrame
; /* The size of a single frame */
3336 i64 iOffset
; /* Next byte to write in WAL file */
3337 WalWriter w
; /* The writer */
3338 u32 iFirst
= 0; /* First frame that may be overwritten */
3339 WalIndexHdr
*pLive
; /* Pointer to shared header */
3342 assert( pWal
->writeLock
);
3344 /* If this frame set completes a transaction, then nTruncate>0. If
3345 ** nTruncate==0 then this frame set does not complete the transaction. */
3346 assert( (isCommit
!=0)==(nTruncate
!=0) );
3348 #if defined(SQLITE_TEST) && defined(SQLITE_DEBUG)
3349 { int cnt
; for(cnt
=0, p
=pList
; p
; p
=p
->pDirty
, cnt
++){}
3350 WALTRACE(("WAL%p: frame write begin. %d frames. mxFrame=%d. %s\n",
3351 pWal
, cnt
, pWal
->hdr
.mxFrame
, isCommit
? "Commit" : "Spill"));
3355 pLive
= (WalIndexHdr
*)walIndexHdr(pWal
);
3356 if( memcmp(&pWal
->hdr
, (void *)pLive
, sizeof(WalIndexHdr
))!=0 ){
3357 iFirst
= pLive
->mxFrame
+1;
3360 /* See if it is possible to write these frames into the start of the
3361 ** log file, instead of appending to it at pWal->hdr.mxFrame.
3363 if( SQLITE_OK
!=(rc
= walRestartLog(pWal
)) ){
3367 /* If this is the first frame written into the log, write the WAL
3368 ** header to the start of the WAL file. See comments at the top of
3369 ** this source file for a description of the WAL header format.
3371 iFrame
= pWal
->hdr
.mxFrame
;
3373 u8 aWalHdr
[WAL_HDRSIZE
]; /* Buffer to assemble wal-header in */
3374 u32 aCksum
[2]; /* Checksum for wal-header */
3376 sqlite3Put4byte(&aWalHdr
[0], (WAL_MAGIC
| SQLITE_BIGENDIAN
));
3377 sqlite3Put4byte(&aWalHdr
[4], WAL_MAX_VERSION
);
3378 sqlite3Put4byte(&aWalHdr
[8], szPage
);
3379 sqlite3Put4byte(&aWalHdr
[12], pWal
->nCkpt
);
3380 if( pWal
->nCkpt
==0 ) sqlite3_randomness(8, pWal
->hdr
.aSalt
);
3381 memcpy(&aWalHdr
[16], pWal
->hdr
.aSalt
, 8);
3382 walChecksumBytes(1, aWalHdr
, WAL_HDRSIZE
-2*4, 0, aCksum
);
3383 sqlite3Put4byte(&aWalHdr
[24], aCksum
[0]);
3384 sqlite3Put4byte(&aWalHdr
[28], aCksum
[1]);
3386 pWal
->szPage
= szPage
;
3387 pWal
->hdr
.bigEndCksum
= SQLITE_BIGENDIAN
;
3388 pWal
->hdr
.aFrameCksum
[0] = aCksum
[0];
3389 pWal
->hdr
.aFrameCksum
[1] = aCksum
[1];
3390 pWal
->truncateOnCommit
= 1;
3392 rc
= sqlite3OsWrite(pWal
->pWalFd
, aWalHdr
, sizeof(aWalHdr
), 0);
3393 WALTRACE(("WAL%p: wal-header write %s\n", pWal
, rc
? "failed" : "ok"));
3394 if( rc
!=SQLITE_OK
){
3398 /* Sync the header (unless SQLITE_IOCAP_SEQUENTIAL is true or unless
3399 ** all syncing is turned off by PRAGMA synchronous=OFF). Otherwise
3400 ** an out-of-order write following a WAL restart could result in
3401 ** database corruption. See the ticket:
3403 ** https://sqlite.org/src/info/ff5be73dee
3405 if( pWal
->syncHeader
){
3406 rc
= sqlite3OsSync(pWal
->pWalFd
, CKPT_SYNC_FLAGS(sync_flags
));
3410 assert( (int)pWal
->szPage
==szPage
);
3412 /* Setup information needed to write frames into the WAL */
3414 w
.pFd
= pWal
->pWalFd
;
3416 w
.syncFlags
= sync_flags
;
3418 iOffset
= walFrameOffset(iFrame
+1, szPage
);
3419 szFrame
= szPage
+ WAL_FRAME_HDRSIZE
;
3421 /* Write all frames into the log file exactly once */
3422 for(p
=pList
; p
; p
=p
->pDirty
){
3423 int nDbSize
; /* 0 normally. Positive == commit flag */
3425 /* Check if this page has already been written into the wal file by
3426 ** the current transaction. If so, overwrite the existing frame and
3427 ** set Wal.writeLock to WAL_WRITELOCK_RECKSUM - indicating that
3428 ** checksums must be recomputed when the transaction is committed. */
3429 if( iFirst
&& (p
->pDirty
|| isCommit
==0) ){
3431 VVA_ONLY(rc
=) sqlite3WalFindFrame(pWal
, p
->pgno
, &iWrite
);
3432 assert( rc
==SQLITE_OK
|| iWrite
==0 );
3433 if( iWrite
>=iFirst
){
3434 i64 iOff
= walFrameOffset(iWrite
, szPage
) + WAL_FRAME_HDRSIZE
;
3436 if( pWal
->iReCksum
==0 || iWrite
<pWal
->iReCksum
){
3437 pWal
->iReCksum
= iWrite
;
3440 rc
= sqlite3OsWrite(pWal
->pWalFd
, pData
, szPage
, iOff
);
3442 p
->flags
&= ~PGHDR_WAL_APPEND
;
3448 assert( iOffset
==walFrameOffset(iFrame
, szPage
) );
3449 nDbSize
= (isCommit
&& p
->pDirty
==0) ? nTruncate
: 0;
3450 rc
= walWriteOneFrame(&w
, p
, nDbSize
, iOffset
);
3454 p
->flags
|= PGHDR_WAL_APPEND
;
3457 /* Recalculate checksums within the wal file if required. */
3458 if( isCommit
&& pWal
->iReCksum
){
3459 rc
= walRewriteChecksums(pWal
, iFrame
);
3463 /* If this is the end of a transaction, then we might need to pad
3464 ** the transaction and/or sync the WAL file.
3466 ** Padding and syncing only occur if this set of frames complete a
3467 ** transaction and if PRAGMA synchronous=FULL. If synchronous==NORMAL
3468 ** or synchronous==OFF, then no padding or syncing are needed.
3470 ** If SQLITE_IOCAP_POWERSAFE_OVERWRITE is defined, then padding is not
3471 ** needed and only the sync is done. If padding is needed, then the
3472 ** final frame is repeated (with its commit mark) until the next sector
3473 ** boundary is crossed. Only the part of the WAL prior to the last
3474 ** sector boundary is synced; the part of the last frame that extends
3475 ** past the sector boundary is written after the sync.
3477 if( isCommit
&& WAL_SYNC_FLAGS(sync_flags
)!=0 ){
3479 if( pWal
->padToSectorBoundary
){
3480 int sectorSize
= sqlite3SectorSize(pWal
->pWalFd
);
3481 w
.iSyncPoint
= ((iOffset
+sectorSize
-1)/sectorSize
)*sectorSize
;
3482 bSync
= (w
.iSyncPoint
==iOffset
);
3484 while( iOffset
<w
.iSyncPoint
){
3485 rc
= walWriteOneFrame(&w
, pLast
, nTruncate
, iOffset
);
3493 assert( rc
==SQLITE_OK
);
3494 rc
= sqlite3OsSync(w
.pFd
, WAL_SYNC_FLAGS(sync_flags
));
3498 /* If this frame set completes the first transaction in the WAL and
3499 ** if PRAGMA journal_size_limit is set, then truncate the WAL to the
3500 ** journal size limit, if possible.
3502 if( isCommit
&& pWal
->truncateOnCommit
&& pWal
->mxWalSize
>=0 ){
3503 i64 sz
= pWal
->mxWalSize
;
3504 if( walFrameOffset(iFrame
+nExtra
+1, szPage
)>pWal
->mxWalSize
){
3505 sz
= walFrameOffset(iFrame
+nExtra
+1, szPage
);
3507 walLimitSize(pWal
, sz
);
3508 pWal
->truncateOnCommit
= 0;
3511 /* Append data to the wal-index. It is not necessary to lock the
3512 ** wal-index to do this as the SQLITE_SHM_WRITE lock held on the wal-index
3513 ** guarantees that there are no other writers, and no data that may
3514 ** be in use by existing readers is being overwritten.
3516 iFrame
= pWal
->hdr
.mxFrame
;
3517 for(p
=pList
; p
&& rc
==SQLITE_OK
; p
=p
->pDirty
){
3518 if( (p
->flags
& PGHDR_WAL_APPEND
)==0 ) continue;
3520 rc
= walIndexAppend(pWal
, iFrame
, p
->pgno
);
3522 assert( pLast
!=0 || nExtra
==0 );
3523 while( rc
==SQLITE_OK
&& nExtra
>0 ){
3526 rc
= walIndexAppend(pWal
, iFrame
, pLast
->pgno
);
3529 if( rc
==SQLITE_OK
){
3530 /* Update the private copy of the header. */
3531 pWal
->hdr
.szPage
= (u16
)((szPage
&0xff00) | (szPage
>>16));
3532 testcase( szPage
<=32768 );
3533 testcase( szPage
>=65536 );
3534 pWal
->hdr
.mxFrame
= iFrame
;
3536 pWal
->hdr
.iChange
++;
3537 pWal
->hdr
.nPage
= nTruncate
;
3539 /* If this is a commit, update the wal-index header too. */
3541 walIndexWriteHdr(pWal
);
3542 pWal
->iCallback
= iFrame
;
3546 WALTRACE(("WAL%p: frame write %s\n", pWal
, rc
? "failed" : "ok"));
3551 ** This routine is called to implement sqlite3_wal_checkpoint() and
3552 ** related interfaces.
3554 ** Obtain a CHECKPOINT lock and then backfill as much information as
3555 ** we can from WAL into the database.
3557 ** If parameter xBusy is not NULL, it is a pointer to a busy-handler
3558 ** callback. In this case this function runs a blocking checkpoint.
3560 int sqlite3WalCheckpoint(
3561 Wal
*pWal
, /* Wal connection */
3562 sqlite3
*db
, /* Check this handle's interrupt flag */
3563 int eMode
, /* PASSIVE, FULL, RESTART, or TRUNCATE */
3564 int (*xBusy
)(void*), /* Function to call when busy */
3565 void *pBusyArg
, /* Context argument for xBusyHandler */
3566 int sync_flags
, /* Flags to sync db file with (or 0) */
3567 int nBuf
, /* Size of temporary buffer */
3568 u8
*zBuf
, /* Temporary buffer to use */
3569 int *pnLog
, /* OUT: Number of frames in WAL */
3570 int *pnCkpt
/* OUT: Number of backfilled frames in WAL */
3572 int rc
; /* Return code */
3573 int isChanged
= 0; /* True if a new wal-index header is loaded */
3574 int eMode2
= eMode
; /* Mode to pass to walCheckpoint() */
3575 int (*xBusy2
)(void*) = xBusy
; /* Busy handler for eMode2 */
3577 assert( pWal
->ckptLock
==0 );
3578 assert( pWal
->writeLock
==0 );
3580 /* EVIDENCE-OF: R-62920-47450 The busy-handler callback is never invoked
3581 ** in the SQLITE_CHECKPOINT_PASSIVE mode. */
3582 assert( eMode
!=SQLITE_CHECKPOINT_PASSIVE
|| xBusy
==0 );
3584 if( pWal
->readOnly
) return SQLITE_READONLY
;
3585 WALTRACE(("WAL%p: checkpoint begins\n", pWal
));
3587 /* IMPLEMENTATION-OF: R-62028-47212 All calls obtain an exclusive
3588 ** "checkpoint" lock on the database file. */
3589 rc
= walLockExclusive(pWal
, WAL_CKPT_LOCK
, 1);
3591 /* EVIDENCE-OF: R-10421-19736 If any other process is running a
3592 ** checkpoint operation at the same time, the lock cannot be obtained and
3593 ** SQLITE_BUSY is returned.
3594 ** EVIDENCE-OF: R-53820-33897 Even if there is a busy-handler configured,
3595 ** it will not be invoked in this case.
3597 testcase( rc
==SQLITE_BUSY
);
3598 testcase( xBusy
!=0 );
3603 /* IMPLEMENTATION-OF: R-59782-36818 The SQLITE_CHECKPOINT_FULL, RESTART and
3604 ** TRUNCATE modes also obtain the exclusive "writer" lock on the database
3607 ** EVIDENCE-OF: R-60642-04082 If the writer lock cannot be obtained
3608 ** immediately, and a busy-handler is configured, it is invoked and the
3609 ** writer lock retried until either the busy-handler returns 0 or the
3610 ** lock is successfully obtained.
3612 if( eMode
!=SQLITE_CHECKPOINT_PASSIVE
){
3613 rc
= walBusyLock(pWal
, xBusy
, pBusyArg
, WAL_WRITE_LOCK
, 1);
3614 if( rc
==SQLITE_OK
){
3615 pWal
->writeLock
= 1;
3616 }else if( rc
==SQLITE_BUSY
){
3617 eMode2
= SQLITE_CHECKPOINT_PASSIVE
;
3623 /* Read the wal-index header. */
3624 if( rc
==SQLITE_OK
){
3625 rc
= walIndexReadHdr(pWal
, &isChanged
);
3626 if( isChanged
&& pWal
->pDbFd
->pMethods
->iVersion
>=3 ){
3627 sqlite3OsUnfetch(pWal
->pDbFd
, 0, 0);
3631 /* Copy data from the log to the database file. */
3632 if( rc
==SQLITE_OK
){
3634 if( pWal
->hdr
.mxFrame
&& walPagesize(pWal
)!=nBuf
){
3635 rc
= SQLITE_CORRUPT_BKPT
;
3637 rc
= walCheckpoint(pWal
, db
, eMode2
, xBusy2
, pBusyArg
, sync_flags
, zBuf
);
3640 /* If no error occurred, set the output variables. */
3641 if( rc
==SQLITE_OK
|| rc
==SQLITE_BUSY
){
3642 if( pnLog
) *pnLog
= (int)pWal
->hdr
.mxFrame
;
3643 if( pnCkpt
) *pnCkpt
= (int)(walCkptInfo(pWal
)->nBackfill
);
3648 /* If a new wal-index header was loaded before the checkpoint was
3649 ** performed, then the pager-cache associated with pWal is now
3650 ** out of date. So zero the cached wal-index header to ensure that
3651 ** next time the pager opens a snapshot on this database it knows that
3652 ** the cache needs to be reset.
3654 memset(&pWal
->hdr
, 0, sizeof(WalIndexHdr
));
3657 /* Release the locks. */
3658 sqlite3WalEndWriteTransaction(pWal
);
3659 walUnlockExclusive(pWal
, WAL_CKPT_LOCK
, 1);
3661 WALTRACE(("WAL%p: checkpoint %s\n", pWal
, rc
? "failed" : "ok"));
3662 return (rc
==SQLITE_OK
&& eMode
!=eMode2
? SQLITE_BUSY
: rc
);
3665 /* Return the value to pass to a sqlite3_wal_hook callback, the
3666 ** number of frames in the WAL at the point of the last commit since
3667 ** sqlite3WalCallback() was called. If no commits have occurred since
3668 ** the last call, then return 0.
3670 int sqlite3WalCallback(Wal
*pWal
){
3673 ret
= pWal
->iCallback
;
3674 pWal
->iCallback
= 0;
3680 ** This function is called to change the WAL subsystem into or out
3681 ** of locking_mode=EXCLUSIVE.
3683 ** If op is zero, then attempt to change from locking_mode=EXCLUSIVE
3684 ** into locking_mode=NORMAL. This means that we must acquire a lock
3685 ** on the pWal->readLock byte. If the WAL is already in locking_mode=NORMAL
3686 ** or if the acquisition of the lock fails, then return 0. If the
3687 ** transition out of exclusive-mode is successful, return 1. This
3688 ** operation must occur while the pager is still holding the exclusive
3689 ** lock on the main database file.
3691 ** If op is one, then change from locking_mode=NORMAL into
3692 ** locking_mode=EXCLUSIVE. This means that the pWal->readLock must
3693 ** be released. Return 1 if the transition is made and 0 if the
3694 ** WAL is already in exclusive-locking mode - meaning that this
3695 ** routine is a no-op. The pager must already hold the exclusive lock
3696 ** on the main database file before invoking this operation.
3698 ** If op is negative, then do a dry-run of the op==1 case but do
3699 ** not actually change anything. The pager uses this to see if it
3700 ** should acquire the database exclusive lock prior to invoking
3703 int sqlite3WalExclusiveMode(Wal
*pWal
, int op
){
3705 assert( pWal
->writeLock
==0 );
3706 assert( pWal
->exclusiveMode
!=WAL_HEAPMEMORY_MODE
|| op
==-1 );
3708 /* pWal->readLock is usually set, but might be -1 if there was a
3709 ** prior error while attempting to acquire are read-lock. This cannot
3710 ** happen if the connection is actually in exclusive mode (as no xShmLock
3711 ** locks are taken in this case). Nor should the pager attempt to
3712 ** upgrade to exclusive-mode following such an error.
3714 assert( pWal
->readLock
>=0 || pWal
->lockError
);
3715 assert( pWal
->readLock
>=0 || (op
<=0 && pWal
->exclusiveMode
==0) );
3718 if( pWal
->exclusiveMode
!=WAL_NORMAL_MODE
){
3719 pWal
->exclusiveMode
= WAL_NORMAL_MODE
;
3720 if( walLockShared(pWal
, WAL_READ_LOCK(pWal
->readLock
))!=SQLITE_OK
){
3721 pWal
->exclusiveMode
= WAL_EXCLUSIVE_MODE
;
3723 rc
= pWal
->exclusiveMode
==WAL_NORMAL_MODE
;
3725 /* Already in locking_mode=NORMAL */
3729 assert( pWal
->exclusiveMode
==WAL_NORMAL_MODE
);
3730 assert( pWal
->readLock
>=0 );
3731 walUnlockShared(pWal
, WAL_READ_LOCK(pWal
->readLock
));
3732 pWal
->exclusiveMode
= WAL_EXCLUSIVE_MODE
;
3735 rc
= pWal
->exclusiveMode
==WAL_NORMAL_MODE
;
3741 ** Return true if the argument is non-NULL and the WAL module is using
3742 ** heap-memory for the wal-index. Otherwise, if the argument is NULL or the
3743 ** WAL module is using shared-memory, return false.
3745 int sqlite3WalHeapMemory(Wal
*pWal
){
3746 return (pWal
&& pWal
->exclusiveMode
==WAL_HEAPMEMORY_MODE
);
3749 #ifdef SQLITE_ENABLE_SNAPSHOT
3750 /* Create a snapshot object. The content of a snapshot is opaque to
3751 ** every other subsystem, so the WAL module can put whatever it needs
3754 int sqlite3WalSnapshotGet(Wal
*pWal
, sqlite3_snapshot
**ppSnapshot
){
3757 static const u32 aZero
[4] = { 0, 0, 0, 0 };
3759 assert( pWal
->readLock
>=0 && pWal
->writeLock
==0 );
3761 if( memcmp(&pWal
->hdr
.aFrameCksum
[0],aZero
,16)==0 ){
3763 return SQLITE_ERROR
;
3765 pRet
= (WalIndexHdr
*)sqlite3_malloc(sizeof(WalIndexHdr
));
3767 rc
= SQLITE_NOMEM_BKPT
;
3769 memcpy(pRet
, &pWal
->hdr
, sizeof(WalIndexHdr
));
3770 *ppSnapshot
= (sqlite3_snapshot
*)pRet
;
3776 /* Try to open on pSnapshot when the next read-transaction starts
3778 void sqlite3WalSnapshotOpen(Wal
*pWal
, sqlite3_snapshot
*pSnapshot
){
3779 pWal
->pSnapshot
= (WalIndexHdr
*)pSnapshot
;
3783 ** Return a +ve value if snapshot p1 is newer than p2. A -ve value if
3784 ** p1 is older than p2 and zero if p1 and p2 are the same snapshot.
3786 int sqlite3_snapshot_cmp(sqlite3_snapshot
*p1
, sqlite3_snapshot
*p2
){
3787 WalIndexHdr
*pHdr1
= (WalIndexHdr
*)p1
;
3788 WalIndexHdr
*pHdr2
= (WalIndexHdr
*)p2
;
3790 /* aSalt[0] is a copy of the value stored in the wal file header. It
3791 ** is incremented each time the wal file is restarted. */
3792 if( pHdr1
->aSalt
[0]<pHdr2
->aSalt
[0] ) return -1;
3793 if( pHdr1
->aSalt
[0]>pHdr2
->aSalt
[0] ) return +1;
3794 if( pHdr1
->mxFrame
<pHdr2
->mxFrame
) return -1;
3795 if( pHdr1
->mxFrame
>pHdr2
->mxFrame
) return +1;
3800 ** The caller currently has a read transaction open on the database.
3801 ** This function takes a SHARED lock on the CHECKPOINTER slot and then
3802 ** checks if the snapshot passed as the second argument is still
3803 ** available. If so, SQLITE_OK is returned.
3805 ** If the snapshot is not available, SQLITE_ERROR is returned. Or, if
3806 ** the CHECKPOINTER lock cannot be obtained, SQLITE_BUSY. If any error
3807 ** occurs (any value other than SQLITE_OK is returned), the CHECKPOINTER
3808 ** lock is released before returning.
3810 int sqlite3WalSnapshotCheck(Wal
*pWal
, sqlite3_snapshot
*pSnapshot
){
3812 rc
= walLockShared(pWal
, WAL_CKPT_LOCK
);
3813 if( rc
==SQLITE_OK
){
3814 WalIndexHdr
*pNew
= (WalIndexHdr
*)pSnapshot
;
3815 if( memcmp(pNew
->aSalt
, pWal
->hdr
.aSalt
, sizeof(pWal
->hdr
.aSalt
))
3816 || pNew
->mxFrame
<walCkptInfo(pWal
)->nBackfillAttempted
3818 rc
= SQLITE_ERROR_SNAPSHOT
;
3819 walUnlockShared(pWal
, WAL_CKPT_LOCK
);
3826 ** Release a lock obtained by an earlier successful call to
3827 ** sqlite3WalSnapshotCheck().
3829 void sqlite3WalSnapshotUnlock(Wal
*pWal
){
3831 walUnlockShared(pWal
, WAL_CKPT_LOCK
);
3835 #endif /* SQLITE_ENABLE_SNAPSHOT */
3837 #ifdef SQLITE_ENABLE_ZIPVFS
3839 ** If the argument is not NULL, it points to a Wal object that holds a
3840 ** read-lock. This function returns the database page-size if it is known,
3841 ** or zero if it is not (or if pWal is NULL).
3843 int sqlite3WalFramesize(Wal
*pWal
){
3844 assert( pWal
==0 || pWal
->readLock
>=0 );
3845 return (pWal
? pWal
->szPage
: 0);
3849 /* Return the sqlite3_file object for the WAL file
3851 sqlite3_file
*sqlite3WalFile(Wal
*pWal
){
3852 return pWal
->pWalFd
;
3855 #endif /* #ifndef SQLITE_OMIT_WAL */