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 *************************************************************************
12 ** This file implements an external (disk-based) database using BTrees.
13 ** See the header comment on "btreeInt.h" for additional information.
14 ** Including a description of file format and an overview of operation.
19 ** The header string that appears at the beginning of every
22 static const char zMagicHeader
[] = SQLITE_FILE_HEADER
;
25 ** Set this global variable to 1 to enable tracing using the TRACE
29 int sqlite3BtreeTrace
=1; /* True to enable tracing */
30 # define TRACE(X) if(sqlite3BtreeTrace){printf X;fflush(stdout);}
36 ** Extract a 2-byte big-endian integer from an array of unsigned bytes.
37 ** But if the value is zero, make it 65536.
39 ** This routine is used to extract the "offset to cell content area" value
40 ** from the header of a btree page. If the page size is 65536 and the page
41 ** is empty, the offset should be 65536, but the 2-byte value stores zero.
42 ** This routine makes the necessary adjustment to 65536.
44 #define get2byteNotZero(X) (((((int)get2byte(X))-1)&0xffff)+1)
47 ** Values passed as the 5th argument to allocateBtreePage()
49 #define BTALLOC_ANY 0 /* Allocate any page */
50 #define BTALLOC_EXACT 1 /* Allocate exact page if possible */
51 #define BTALLOC_LE 2 /* Allocate any page <= the parameter */
54 ** Macro IfNotOmitAV(x) returns (x) if SQLITE_OMIT_AUTOVACUUM is not
55 ** defined, or 0 if it is. For example:
57 ** bIncrVacuum = IfNotOmitAV(pBtShared->incrVacuum);
59 #ifndef SQLITE_OMIT_AUTOVACUUM
60 #define IfNotOmitAV(expr) (expr)
62 #define IfNotOmitAV(expr) 0
65 #ifndef SQLITE_OMIT_SHARED_CACHE
67 ** A list of BtShared objects that are eligible for participation
68 ** in shared cache. This variable has file scope during normal builds,
69 ** but the test harness needs to access it so we make it global for
72 ** Access to this variable is protected by SQLITE_MUTEX_STATIC_MASTER.
75 BtShared
*SQLITE_WSD sqlite3SharedCacheList
= 0;
77 static BtShared
*SQLITE_WSD sqlite3SharedCacheList
= 0;
79 #endif /* SQLITE_OMIT_SHARED_CACHE */
81 #ifndef SQLITE_OMIT_SHARED_CACHE
83 ** Enable or disable the shared pager and schema features.
85 ** This routine has no effect on existing database connections.
86 ** The shared cache setting effects only future calls to
87 ** sqlite3_open(), sqlite3_open16(), or sqlite3_open_v2().
89 int sqlite3_enable_shared_cache(int enable
){
90 sqlite3GlobalConfig
.sharedCacheEnabled
= enable
;
97 #ifdef SQLITE_OMIT_SHARED_CACHE
99 ** The functions querySharedCacheTableLock(), setSharedCacheTableLock(),
100 ** and clearAllSharedCacheTableLocks()
101 ** manipulate entries in the BtShared.pLock linked list used to store
102 ** shared-cache table level locks. If the library is compiled with the
103 ** shared-cache feature disabled, then there is only ever one user
104 ** of each BtShared structure and so this locking is not necessary.
105 ** So define the lock related functions as no-ops.
107 #define querySharedCacheTableLock(a,b,c) SQLITE_OK
108 #define setSharedCacheTableLock(a,b,c) SQLITE_OK
109 #define clearAllSharedCacheTableLocks(a)
110 #define downgradeAllSharedCacheTableLocks(a)
111 #define hasSharedCacheTableLock(a,b,c,d) 1
112 #define hasReadConflicts(a, b) 0
116 ** Implementation of the SQLITE_CORRUPT_PAGE() macro. Takes a single
117 ** (MemPage*) as an argument. The (MemPage*) must not be NULL.
119 ** If SQLITE_DEBUG is not defined, then this macro is equivalent to
120 ** SQLITE_CORRUPT_BKPT. Or, if SQLITE_DEBUG is set, then the log message
121 ** normally produced as a side-effect of SQLITE_CORRUPT_BKPT is augmented
122 ** with the page number and filename associated with the (MemPage*).
125 int corruptPageError(int lineno
, MemPage
*p
){
127 sqlite3BeginBenignMalloc();
128 zMsg
= sqlite3_mprintf("database corruption page %d of %s",
129 (int)p
->pgno
, sqlite3PagerFilename(p
->pBt
->pPager
, 0)
131 sqlite3EndBenignMalloc();
133 sqlite3ReportError(SQLITE_CORRUPT
, lineno
, zMsg
);
136 return SQLITE_CORRUPT_BKPT
;
138 # define SQLITE_CORRUPT_PAGE(pMemPage) corruptPageError(__LINE__, pMemPage)
140 # define SQLITE_CORRUPT_PAGE(pMemPage) SQLITE_CORRUPT_PGNO(pMemPage->pgno)
143 #ifndef SQLITE_OMIT_SHARED_CACHE
147 **** This function is only used as part of an assert() statement. ***
149 ** Check to see if pBtree holds the required locks to read or write to the
150 ** table with root page iRoot. Return 1 if it does and 0 if not.
152 ** For example, when writing to a table with root-page iRoot via
153 ** Btree connection pBtree:
155 ** assert( hasSharedCacheTableLock(pBtree, iRoot, 0, WRITE_LOCK) );
157 ** When writing to an index that resides in a sharable database, the
158 ** caller should have first obtained a lock specifying the root page of
159 ** the corresponding table. This makes things a bit more complicated,
160 ** as this module treats each table as a separate structure. To determine
161 ** the table corresponding to the index being written, this
162 ** function has to search through the database schema.
164 ** Instead of a lock on the table/index rooted at page iRoot, the caller may
165 ** hold a write-lock on the schema table (root page 1). This is also
168 static int hasSharedCacheTableLock(
169 Btree
*pBtree
, /* Handle that must hold lock */
170 Pgno iRoot
, /* Root page of b-tree */
171 int isIndex
, /* True if iRoot is the root of an index b-tree */
172 int eLockType
/* Required lock type (READ_LOCK or WRITE_LOCK) */
174 Schema
*pSchema
= (Schema
*)pBtree
->pBt
->pSchema
;
178 /* If this database is not shareable, or if the client is reading
179 ** and has the read-uncommitted flag set, then no lock is required.
180 ** Return true immediately.
182 if( (pBtree
->sharable
==0)
183 || (eLockType
==READ_LOCK
&& (pBtree
->db
->flags
& SQLITE_ReadUncommit
))
188 /* If the client is reading or writing an index and the schema is
189 ** not loaded, then it is too difficult to actually check to see if
190 ** the correct locks are held. So do not bother - just return true.
191 ** This case does not come up very often anyhow.
193 if( isIndex
&& (!pSchema
|| (pSchema
->schemaFlags
&DB_SchemaLoaded
)==0) ){
197 /* Figure out the root-page that the lock should be held on. For table
198 ** b-trees, this is just the root page of the b-tree being read or
199 ** written. For index b-trees, it is the root page of the associated
203 for(p
=sqliteHashFirst(&pSchema
->idxHash
); p
; p
=sqliteHashNext(p
)){
204 Index
*pIdx
= (Index
*)sqliteHashData(p
);
205 if( pIdx
->tnum
==(int)iRoot
){
207 /* Two or more indexes share the same root page. There must
208 ** be imposter tables. So just return true. The assert is not
209 ** useful in that case. */
212 iTab
= pIdx
->pTable
->tnum
;
219 /* Search for the required lock. Either a write-lock on root-page iTab, a
220 ** write-lock on the schema table, or (if the client is reading) a
221 ** read-lock on iTab will suffice. Return 1 if any of these are found. */
222 for(pLock
=pBtree
->pBt
->pLock
; pLock
; pLock
=pLock
->pNext
){
223 if( pLock
->pBtree
==pBtree
224 && (pLock
->iTable
==iTab
|| (pLock
->eLock
==WRITE_LOCK
&& pLock
->iTable
==1))
225 && pLock
->eLock
>=eLockType
231 /* Failed to find the required lock. */
234 #endif /* SQLITE_DEBUG */
238 **** This function may be used as part of assert() statements only. ****
240 ** Return true if it would be illegal for pBtree to write into the
241 ** table or index rooted at iRoot because other shared connections are
242 ** simultaneously reading that same table or index.
244 ** It is illegal for pBtree to write if some other Btree object that
245 ** shares the same BtShared object is currently reading or writing
246 ** the iRoot table. Except, if the other Btree object has the
247 ** read-uncommitted flag set, then it is OK for the other object to
248 ** have a read cursor.
250 ** For example, before writing to any part of the table or index
251 ** rooted at page iRoot, one should call:
253 ** assert( !hasReadConflicts(pBtree, iRoot) );
255 static int hasReadConflicts(Btree
*pBtree
, Pgno iRoot
){
257 for(p
=pBtree
->pBt
->pCursor
; p
; p
=p
->pNext
){
258 if( p
->pgnoRoot
==iRoot
260 && 0==(p
->pBtree
->db
->flags
& SQLITE_ReadUncommit
)
267 #endif /* #ifdef SQLITE_DEBUG */
270 ** Query to see if Btree handle p may obtain a lock of type eLock
271 ** (READ_LOCK or WRITE_LOCK) on the table with root-page iTab. Return
272 ** SQLITE_OK if the lock may be obtained (by calling
273 ** setSharedCacheTableLock()), or SQLITE_LOCKED if not.
275 static int querySharedCacheTableLock(Btree
*p
, Pgno iTab
, u8 eLock
){
276 BtShared
*pBt
= p
->pBt
;
279 assert( sqlite3BtreeHoldsMutex(p
) );
280 assert( eLock
==READ_LOCK
|| eLock
==WRITE_LOCK
);
282 assert( !(p
->db
->flags
&SQLITE_ReadUncommit
)||eLock
==WRITE_LOCK
||iTab
==1 );
284 /* If requesting a write-lock, then the Btree must have an open write
285 ** transaction on this file. And, obviously, for this to be so there
286 ** must be an open write transaction on the file itself.
288 assert( eLock
==READ_LOCK
|| (p
==pBt
->pWriter
&& p
->inTrans
==TRANS_WRITE
) );
289 assert( eLock
==READ_LOCK
|| pBt
->inTransaction
==TRANS_WRITE
);
291 /* This routine is a no-op if the shared-cache is not enabled */
296 /* If some other connection is holding an exclusive lock, the
297 ** requested lock may not be obtained.
299 if( pBt
->pWriter
!=p
&& (pBt
->btsFlags
& BTS_EXCLUSIVE
)!=0 ){
300 sqlite3ConnectionBlocked(p
->db
, pBt
->pWriter
->db
);
301 return SQLITE_LOCKED_SHAREDCACHE
;
304 for(pIter
=pBt
->pLock
; pIter
; pIter
=pIter
->pNext
){
305 /* The condition (pIter->eLock!=eLock) in the following if(...)
306 ** statement is a simplification of:
308 ** (eLock==WRITE_LOCK || pIter->eLock==WRITE_LOCK)
310 ** since we know that if eLock==WRITE_LOCK, then no other connection
311 ** may hold a WRITE_LOCK on any table in this file (since there can
312 ** only be a single writer).
314 assert( pIter
->eLock
==READ_LOCK
|| pIter
->eLock
==WRITE_LOCK
);
315 assert( eLock
==READ_LOCK
|| pIter
->pBtree
==p
|| pIter
->eLock
==READ_LOCK
);
316 if( pIter
->pBtree
!=p
&& pIter
->iTable
==iTab
&& pIter
->eLock
!=eLock
){
317 sqlite3ConnectionBlocked(p
->db
, pIter
->pBtree
->db
);
318 if( eLock
==WRITE_LOCK
){
319 assert( p
==pBt
->pWriter
);
320 pBt
->btsFlags
|= BTS_PENDING
;
322 return SQLITE_LOCKED_SHAREDCACHE
;
327 #endif /* !SQLITE_OMIT_SHARED_CACHE */
329 #ifndef SQLITE_OMIT_SHARED_CACHE
331 ** Add a lock on the table with root-page iTable to the shared-btree used
332 ** by Btree handle p. Parameter eLock must be either READ_LOCK or
335 ** This function assumes the following:
337 ** (a) The specified Btree object p is connected to a sharable
338 ** database (one with the BtShared.sharable flag set), and
340 ** (b) No other Btree objects hold a lock that conflicts
341 ** with the requested lock (i.e. querySharedCacheTableLock() has
342 ** already been called and returned SQLITE_OK).
344 ** SQLITE_OK is returned if the lock is added successfully. SQLITE_NOMEM
345 ** is returned if a malloc attempt fails.
347 static int setSharedCacheTableLock(Btree
*p
, Pgno iTable
, u8 eLock
){
348 BtShared
*pBt
= p
->pBt
;
352 assert( sqlite3BtreeHoldsMutex(p
) );
353 assert( eLock
==READ_LOCK
|| eLock
==WRITE_LOCK
);
356 /* A connection with the read-uncommitted flag set will never try to
357 ** obtain a read-lock using this function. The only read-lock obtained
358 ** by a connection in read-uncommitted mode is on the sqlite_master
359 ** table, and that lock is obtained in BtreeBeginTrans(). */
360 assert( 0==(p
->db
->flags
&SQLITE_ReadUncommit
) || eLock
==WRITE_LOCK
);
362 /* This function should only be called on a sharable b-tree after it
363 ** has been determined that no other b-tree holds a conflicting lock. */
364 assert( p
->sharable
);
365 assert( SQLITE_OK
==querySharedCacheTableLock(p
, iTable
, eLock
) );
367 /* First search the list for an existing lock on this table. */
368 for(pIter
=pBt
->pLock
; pIter
; pIter
=pIter
->pNext
){
369 if( pIter
->iTable
==iTable
&& pIter
->pBtree
==p
){
375 /* If the above search did not find a BtLock struct associating Btree p
376 ** with table iTable, allocate one and link it into the list.
379 pLock
= (BtLock
*)sqlite3MallocZero(sizeof(BtLock
));
381 return SQLITE_NOMEM_BKPT
;
383 pLock
->iTable
= iTable
;
385 pLock
->pNext
= pBt
->pLock
;
389 /* Set the BtLock.eLock variable to the maximum of the current lock
390 ** and the requested lock. This means if a write-lock was already held
391 ** and a read-lock requested, we don't incorrectly downgrade the lock.
393 assert( WRITE_LOCK
>READ_LOCK
);
394 if( eLock
>pLock
->eLock
){
395 pLock
->eLock
= eLock
;
400 #endif /* !SQLITE_OMIT_SHARED_CACHE */
402 #ifndef SQLITE_OMIT_SHARED_CACHE
404 ** Release all the table locks (locks obtained via calls to
405 ** the setSharedCacheTableLock() procedure) held by Btree object p.
407 ** This function assumes that Btree p has an open read or write
408 ** transaction. If it does not, then the BTS_PENDING flag
409 ** may be incorrectly cleared.
411 static void clearAllSharedCacheTableLocks(Btree
*p
){
412 BtShared
*pBt
= p
->pBt
;
413 BtLock
**ppIter
= &pBt
->pLock
;
415 assert( sqlite3BtreeHoldsMutex(p
) );
416 assert( p
->sharable
|| 0==*ppIter
);
417 assert( p
->inTrans
>0 );
420 BtLock
*pLock
= *ppIter
;
421 assert( (pBt
->btsFlags
& BTS_EXCLUSIVE
)==0 || pBt
->pWriter
==pLock
->pBtree
);
422 assert( pLock
->pBtree
->inTrans
>=pLock
->eLock
);
423 if( pLock
->pBtree
==p
){
424 *ppIter
= pLock
->pNext
;
425 assert( pLock
->iTable
!=1 || pLock
==&p
->lock
);
426 if( pLock
->iTable
!=1 ){
430 ppIter
= &pLock
->pNext
;
434 assert( (pBt
->btsFlags
& BTS_PENDING
)==0 || pBt
->pWriter
);
435 if( pBt
->pWriter
==p
){
437 pBt
->btsFlags
&= ~(BTS_EXCLUSIVE
|BTS_PENDING
);
438 }else if( pBt
->nTransaction
==2 ){
439 /* This function is called when Btree p is concluding its
440 ** transaction. If there currently exists a writer, and p is not
441 ** that writer, then the number of locks held by connections other
442 ** than the writer must be about to drop to zero. In this case
443 ** set the BTS_PENDING flag to 0.
445 ** If there is not currently a writer, then BTS_PENDING must
446 ** be zero already. So this next line is harmless in that case.
448 pBt
->btsFlags
&= ~BTS_PENDING
;
453 ** This function changes all write-locks held by Btree p into read-locks.
455 static void downgradeAllSharedCacheTableLocks(Btree
*p
){
456 BtShared
*pBt
= p
->pBt
;
457 if( pBt
->pWriter
==p
){
460 pBt
->btsFlags
&= ~(BTS_EXCLUSIVE
|BTS_PENDING
);
461 for(pLock
=pBt
->pLock
; pLock
; pLock
=pLock
->pNext
){
462 assert( pLock
->eLock
==READ_LOCK
|| pLock
->pBtree
==p
);
463 pLock
->eLock
= READ_LOCK
;
468 #endif /* SQLITE_OMIT_SHARED_CACHE */
470 static void releasePage(MemPage
*pPage
); /* Forward reference */
471 static void releasePageOne(MemPage
*pPage
); /* Forward reference */
472 static void releasePageNotNull(MemPage
*pPage
); /* Forward reference */
475 ***** This routine is used inside of assert() only ****
477 ** Verify that the cursor holds the mutex on its BtShared
480 static int cursorHoldsMutex(BtCursor
*p
){
481 return sqlite3_mutex_held(p
->pBt
->mutex
);
484 /* Verify that the cursor and the BtShared agree about what is the current
485 ** database connetion. This is important in shared-cache mode. If the database
486 ** connection pointers get out-of-sync, it is possible for routines like
487 ** btreeInitPage() to reference an stale connection pointer that references a
488 ** a connection that has already closed. This routine is used inside assert()
489 ** statements only and for the purpose of double-checking that the btree code
490 ** does keep the database connection pointers up-to-date.
492 static int cursorOwnsBtShared(BtCursor
*p
){
493 assert( cursorHoldsMutex(p
) );
494 return (p
->pBtree
->db
==p
->pBt
->db
);
499 ** Invalidate the overflow cache of the cursor passed as the first argument.
500 ** on the shared btree structure pBt.
502 #define invalidateOverflowCache(pCur) (pCur->curFlags &= ~BTCF_ValidOvfl)
505 ** Invalidate the overflow page-list cache for all cursors opened
506 ** on the shared btree structure pBt.
508 static void invalidateAllOverflowCache(BtShared
*pBt
){
510 assert( sqlite3_mutex_held(pBt
->mutex
) );
511 for(p
=pBt
->pCursor
; p
; p
=p
->pNext
){
512 invalidateOverflowCache(p
);
516 #ifndef SQLITE_OMIT_INCRBLOB
518 ** This function is called before modifying the contents of a table
519 ** to invalidate any incrblob cursors that are open on the
520 ** row or one of the rows being modified.
522 ** If argument isClearTable is true, then the entire contents of the
523 ** table is about to be deleted. In this case invalidate all incrblob
524 ** cursors open on any row within the table with root-page pgnoRoot.
526 ** Otherwise, if argument isClearTable is false, then the row with
527 ** rowid iRow is being replaced or deleted. In this case invalidate
528 ** only those incrblob cursors open on that specific row.
530 static void invalidateIncrblobCursors(
531 Btree
*pBtree
, /* The database file to check */
532 Pgno pgnoRoot
, /* The table that might be changing */
533 i64 iRow
, /* The rowid that might be changing */
534 int isClearTable
/* True if all rows are being deleted */
537 if( pBtree
->hasIncrblobCur
==0 ) return;
538 assert( sqlite3BtreeHoldsMutex(pBtree
) );
539 pBtree
->hasIncrblobCur
= 0;
540 for(p
=pBtree
->pBt
->pCursor
; p
; p
=p
->pNext
){
541 if( (p
->curFlags
& BTCF_Incrblob
)!=0 ){
542 pBtree
->hasIncrblobCur
= 1;
543 if( p
->pgnoRoot
==pgnoRoot
&& (isClearTable
|| p
->info
.nKey
==iRow
) ){
544 p
->eState
= CURSOR_INVALID
;
551 /* Stub function when INCRBLOB is omitted */
552 #define invalidateIncrblobCursors(w,x,y,z)
553 #endif /* SQLITE_OMIT_INCRBLOB */
556 ** Set bit pgno of the BtShared.pHasContent bitvec. This is called
557 ** when a page that previously contained data becomes a free-list leaf
560 ** The BtShared.pHasContent bitvec exists to work around an obscure
561 ** bug caused by the interaction of two useful IO optimizations surrounding
562 ** free-list leaf pages:
564 ** 1) When all data is deleted from a page and the page becomes
565 ** a free-list leaf page, the page is not written to the database
566 ** (as free-list leaf pages contain no meaningful data). Sometimes
567 ** such a page is not even journalled (as it will not be modified,
568 ** why bother journalling it?).
570 ** 2) When a free-list leaf page is reused, its content is not read
571 ** from the database or written to the journal file (why should it
572 ** be, if it is not at all meaningful?).
574 ** By themselves, these optimizations work fine and provide a handy
575 ** performance boost to bulk delete or insert operations. However, if
576 ** a page is moved to the free-list and then reused within the same
577 ** transaction, a problem comes up. If the page is not journalled when
578 ** it is moved to the free-list and it is also not journalled when it
579 ** is extracted from the free-list and reused, then the original data
580 ** may be lost. In the event of a rollback, it may not be possible
581 ** to restore the database to its original configuration.
583 ** The solution is the BtShared.pHasContent bitvec. Whenever a page is
584 ** moved to become a free-list leaf page, the corresponding bit is
585 ** set in the bitvec. Whenever a leaf page is extracted from the free-list,
586 ** optimization 2 above is omitted if the corresponding bit is already
587 ** set in BtShared.pHasContent. The contents of the bitvec are cleared
588 ** at the end of every transaction.
590 static int btreeSetHasContent(BtShared
*pBt
, Pgno pgno
){
592 if( !pBt
->pHasContent
){
593 assert( pgno
<=pBt
->nPage
);
594 pBt
->pHasContent
= sqlite3BitvecCreate(pBt
->nPage
);
595 if( !pBt
->pHasContent
){
596 rc
= SQLITE_NOMEM_BKPT
;
599 if( rc
==SQLITE_OK
&& pgno
<=sqlite3BitvecSize(pBt
->pHasContent
) ){
600 rc
= sqlite3BitvecSet(pBt
->pHasContent
, pgno
);
606 ** Query the BtShared.pHasContent vector.
608 ** This function is called when a free-list leaf page is removed from the
609 ** free-list for reuse. It returns false if it is safe to retrieve the
610 ** page from the pager layer with the 'no-content' flag set. True otherwise.
612 static int btreeGetHasContent(BtShared
*pBt
, Pgno pgno
){
613 Bitvec
*p
= pBt
->pHasContent
;
614 return (p
&& (pgno
>sqlite3BitvecSize(p
) || sqlite3BitvecTest(p
, pgno
)));
618 ** Clear (destroy) the BtShared.pHasContent bitvec. This should be
619 ** invoked at the conclusion of each write-transaction.
621 static void btreeClearHasContent(BtShared
*pBt
){
622 sqlite3BitvecDestroy(pBt
->pHasContent
);
623 pBt
->pHasContent
= 0;
627 ** Release all of the apPage[] pages for a cursor.
629 static void btreeReleaseAllCursorPages(BtCursor
*pCur
){
631 if( pCur
->iPage
>=0 ){
632 for(i
=0; i
<pCur
->iPage
; i
++){
633 releasePageNotNull(pCur
->apPage
[i
]);
635 releasePageNotNull(pCur
->pPage
);
641 ** The cursor passed as the only argument must point to a valid entry
642 ** when this function is called (i.e. have eState==CURSOR_VALID). This
643 ** function saves the current cursor key in variables pCur->nKey and
644 ** pCur->pKey. SQLITE_OK is returned if successful or an SQLite error
647 ** If the cursor is open on an intkey table, then the integer key
648 ** (the rowid) is stored in pCur->nKey and pCur->pKey is left set to
649 ** NULL. If the cursor is open on a non-intkey table, then pCur->pKey is
650 ** set to point to a malloced buffer pCur->nKey bytes in size containing
653 static int saveCursorKey(BtCursor
*pCur
){
655 assert( CURSOR_VALID
==pCur
->eState
);
656 assert( 0==pCur
->pKey
);
657 assert( cursorHoldsMutex(pCur
) );
659 if( pCur
->curIntKey
){
660 /* Only the rowid is required for a table btree */
661 pCur
->nKey
= sqlite3BtreeIntegerKey(pCur
);
663 /* For an index btree, save the complete key content */
665 pCur
->nKey
= sqlite3BtreePayloadSize(pCur
);
666 pKey
= sqlite3Malloc( pCur
->nKey
);
668 rc
= sqlite3BtreePayload(pCur
, 0, (int)pCur
->nKey
, pKey
);
675 rc
= SQLITE_NOMEM_BKPT
;
678 assert( !pCur
->curIntKey
|| !pCur
->pKey
);
683 ** Save the current cursor position in the variables BtCursor.nKey
684 ** and BtCursor.pKey. The cursor's state is set to CURSOR_REQUIRESEEK.
686 ** The caller must ensure that the cursor is valid (has eState==CURSOR_VALID)
687 ** prior to calling this routine.
689 static int saveCursorPosition(BtCursor
*pCur
){
692 assert( CURSOR_VALID
==pCur
->eState
|| CURSOR_SKIPNEXT
==pCur
->eState
);
693 assert( 0==pCur
->pKey
);
694 assert( cursorHoldsMutex(pCur
) );
696 if( pCur
->eState
==CURSOR_SKIPNEXT
){
697 pCur
->eState
= CURSOR_VALID
;
702 rc
= saveCursorKey(pCur
);
704 btreeReleaseAllCursorPages(pCur
);
705 pCur
->eState
= CURSOR_REQUIRESEEK
;
708 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
|BTCF_AtLast
);
712 /* Forward reference */
713 static int SQLITE_NOINLINE
saveCursorsOnList(BtCursor
*,Pgno
,BtCursor
*);
716 ** Save the positions of all cursors (except pExcept) that are open on
717 ** the table with root-page iRoot. "Saving the cursor position" means that
718 ** the location in the btree is remembered in such a way that it can be
719 ** moved back to the same spot after the btree has been modified. This
720 ** routine is called just before cursor pExcept is used to modify the
721 ** table, for example in BtreeDelete() or BtreeInsert().
723 ** If there are two or more cursors on the same btree, then all such
724 ** cursors should have their BTCF_Multiple flag set. The btreeCursor()
725 ** routine enforces that rule. This routine only needs to be called in
726 ** the uncommon case when pExpect has the BTCF_Multiple flag set.
728 ** If pExpect!=NULL and if no other cursors are found on the same root-page,
729 ** then the BTCF_Multiple flag on pExpect is cleared, to avoid another
730 ** pointless call to this routine.
732 ** Implementation note: This routine merely checks to see if any cursors
733 ** need to be saved. It calls out to saveCursorsOnList() in the (unusual)
734 ** event that cursors are in need to being saved.
736 static int saveAllCursors(BtShared
*pBt
, Pgno iRoot
, BtCursor
*pExcept
){
738 assert( sqlite3_mutex_held(pBt
->mutex
) );
739 assert( pExcept
==0 || pExcept
->pBt
==pBt
);
740 for(p
=pBt
->pCursor
; p
; p
=p
->pNext
){
741 if( p
!=pExcept
&& (0==iRoot
|| p
->pgnoRoot
==iRoot
) ) break;
743 if( p
) return saveCursorsOnList(p
, iRoot
, pExcept
);
744 if( pExcept
) pExcept
->curFlags
&= ~BTCF_Multiple
;
748 /* This helper routine to saveAllCursors does the actual work of saving
749 ** the cursors if and when a cursor is found that actually requires saving.
750 ** The common case is that no cursors need to be saved, so this routine is
751 ** broken out from its caller to avoid unnecessary stack pointer movement.
753 static int SQLITE_NOINLINE
saveCursorsOnList(
754 BtCursor
*p
, /* The first cursor that needs saving */
755 Pgno iRoot
, /* Only save cursor with this iRoot. Save all if zero */
756 BtCursor
*pExcept
/* Do not save this cursor */
759 if( p
!=pExcept
&& (0==iRoot
|| p
->pgnoRoot
==iRoot
) ){
760 if( p
->eState
==CURSOR_VALID
|| p
->eState
==CURSOR_SKIPNEXT
){
761 int rc
= saveCursorPosition(p
);
766 testcase( p
->iPage
>=0 );
767 btreeReleaseAllCursorPages(p
);
776 ** Clear the current cursor position.
778 void sqlite3BtreeClearCursor(BtCursor
*pCur
){
779 assert( cursorHoldsMutex(pCur
) );
780 sqlite3_free(pCur
->pKey
);
782 pCur
->eState
= CURSOR_INVALID
;
786 ** In this version of BtreeMoveto, pKey is a packed index record
787 ** such as is generated by the OP_MakeRecord opcode. Unpack the
788 ** record and then call BtreeMovetoUnpacked() to do the work.
790 static int btreeMoveto(
791 BtCursor
*pCur
, /* Cursor open on the btree to be searched */
792 const void *pKey
, /* Packed key if the btree is an index */
793 i64 nKey
, /* Integer key for tables. Size of pKey for indices */
794 int bias
, /* Bias search to the high end */
795 int *pRes
/* Write search results here */
797 int rc
; /* Status code */
798 UnpackedRecord
*pIdxKey
; /* Unpacked index key */
801 assert( nKey
==(i64
)(int)nKey
);
802 pIdxKey
= sqlite3VdbeAllocUnpackedRecord(pCur
->pKeyInfo
);
803 if( pIdxKey
==0 ) return SQLITE_NOMEM_BKPT
;
804 sqlite3VdbeRecordUnpack(pCur
->pKeyInfo
, (int)nKey
, pKey
, pIdxKey
);
805 if( pIdxKey
->nField
==0 ){
806 rc
= SQLITE_CORRUPT_BKPT
;
812 rc
= sqlite3BtreeMovetoUnpacked(pCur
, pIdxKey
, nKey
, bias
, pRes
);
815 sqlite3DbFree(pCur
->pKeyInfo
->db
, pIdxKey
);
821 ** Restore the cursor to the position it was in (or as close to as possible)
822 ** when saveCursorPosition() was called. Note that this call deletes the
823 ** saved position info stored by saveCursorPosition(), so there can be
824 ** at most one effective restoreCursorPosition() call after each
825 ** saveCursorPosition().
827 static int btreeRestoreCursorPosition(BtCursor
*pCur
){
830 assert( cursorOwnsBtShared(pCur
) );
831 assert( pCur
->eState
>=CURSOR_REQUIRESEEK
);
832 if( pCur
->eState
==CURSOR_FAULT
){
833 return pCur
->skipNext
;
835 pCur
->eState
= CURSOR_INVALID
;
836 rc
= btreeMoveto(pCur
, pCur
->pKey
, pCur
->nKey
, 0, &skipNext
);
838 sqlite3_free(pCur
->pKey
);
840 assert( pCur
->eState
==CURSOR_VALID
|| pCur
->eState
==CURSOR_INVALID
);
841 pCur
->skipNext
|= skipNext
;
842 if( pCur
->skipNext
&& pCur
->eState
==CURSOR_VALID
){
843 pCur
->eState
= CURSOR_SKIPNEXT
;
849 #define restoreCursorPosition(p) \
850 (p->eState>=CURSOR_REQUIRESEEK ? \
851 btreeRestoreCursorPosition(p) : \
855 ** Determine whether or not a cursor has moved from the position where
856 ** it was last placed, or has been invalidated for any other reason.
857 ** Cursors can move when the row they are pointing at is deleted out
858 ** from under them, for example. Cursor might also move if a btree
861 ** Calling this routine with a NULL cursor pointer returns false.
863 ** Use the separate sqlite3BtreeCursorRestore() routine to restore a cursor
864 ** back to where it ought to be if this routine returns true.
866 int sqlite3BtreeCursorHasMoved(BtCursor
*pCur
){
867 return pCur
->eState
!=CURSOR_VALID
;
871 ** Return a pointer to a fake BtCursor object that will always answer
872 ** false to the sqlite3BtreeCursorHasMoved() routine above. The fake
873 ** cursor returned must not be used with any other Btree interface.
875 BtCursor
*sqlite3BtreeFakeValidCursor(void){
876 static u8 fakeCursor
= CURSOR_VALID
;
877 assert( offsetof(BtCursor
, eState
)==0 );
878 return (BtCursor
*)&fakeCursor
;
882 ** This routine restores a cursor back to its original position after it
883 ** has been moved by some outside activity (such as a btree rebalance or
884 ** a row having been deleted out from under the cursor).
886 ** On success, the *pDifferentRow parameter is false if the cursor is left
887 ** pointing at exactly the same row. *pDifferntRow is the row the cursor
888 ** was pointing to has been deleted, forcing the cursor to point to some
891 ** This routine should only be called for a cursor that just returned
892 ** TRUE from sqlite3BtreeCursorHasMoved().
894 int sqlite3BtreeCursorRestore(BtCursor
*pCur
, int *pDifferentRow
){
898 assert( pCur
->eState
!=CURSOR_VALID
);
899 rc
= restoreCursorPosition(pCur
);
904 if( pCur
->eState
!=CURSOR_VALID
){
907 assert( pCur
->skipNext
==0 );
913 #ifdef SQLITE_ENABLE_CURSOR_HINTS
915 ** Provide hints to the cursor. The particular hint given (and the type
916 ** and number of the varargs parameters) is determined by the eHintType
917 ** parameter. See the definitions of the BTREE_HINT_* macros for details.
919 void sqlite3BtreeCursorHint(BtCursor
*pCur
, int eHintType
, ...){
920 /* Used only by system that substitute their own storage engine */
925 ** Provide flag hints to the cursor.
927 void sqlite3BtreeCursorHintFlags(BtCursor
*pCur
, unsigned x
){
928 assert( x
==BTREE_SEEK_EQ
|| x
==BTREE_BULKLOAD
|| x
==0 );
933 #ifndef SQLITE_OMIT_AUTOVACUUM
935 ** Given a page number of a regular database page, return the page
936 ** number for the pointer-map page that contains the entry for the
937 ** input page number.
939 ** Return 0 (not a valid page) for pgno==1 since there is
940 ** no pointer map associated with page 1. The integrity_check logic
941 ** requires that ptrmapPageno(*,1)!=1.
943 static Pgno
ptrmapPageno(BtShared
*pBt
, Pgno pgno
){
944 int nPagesPerMapPage
;
946 assert( sqlite3_mutex_held(pBt
->mutex
) );
947 if( pgno
<2 ) return 0;
948 nPagesPerMapPage
= (pBt
->usableSize
/5)+1;
949 iPtrMap
= (pgno
-2)/nPagesPerMapPage
;
950 ret
= (iPtrMap
*nPagesPerMapPage
) + 2;
951 if( ret
==PENDING_BYTE_PAGE(pBt
) ){
958 ** Write an entry into the pointer map.
960 ** This routine updates the pointer map entry for page number 'key'
961 ** so that it maps to type 'eType' and parent page number 'pgno'.
963 ** If *pRC is initially non-zero (non-SQLITE_OK) then this routine is
964 ** a no-op. If an error occurs, the appropriate error code is written
967 static void ptrmapPut(BtShared
*pBt
, Pgno key
, u8 eType
, Pgno parent
, int *pRC
){
968 DbPage
*pDbPage
; /* The pointer map page */
969 u8
*pPtrmap
; /* The pointer map data */
970 Pgno iPtrmap
; /* The pointer map page number */
971 int offset
; /* Offset in pointer map page */
972 int rc
; /* Return code from subfunctions */
976 assert( sqlite3_mutex_held(pBt
->mutex
) );
977 /* The master-journal page number must never be used as a pointer map page */
978 assert( 0==PTRMAP_ISPAGE(pBt
, PENDING_BYTE_PAGE(pBt
)) );
980 assert( pBt
->autoVacuum
);
982 *pRC
= SQLITE_CORRUPT_BKPT
;
985 iPtrmap
= PTRMAP_PAGENO(pBt
, key
);
986 rc
= sqlite3PagerGet(pBt
->pPager
, iPtrmap
, &pDbPage
, 0);
991 offset
= PTRMAP_PTROFFSET(iPtrmap
, key
);
993 *pRC
= SQLITE_CORRUPT_BKPT
;
996 assert( offset
<= (int)pBt
->usableSize
-5 );
997 pPtrmap
= (u8
*)sqlite3PagerGetData(pDbPage
);
999 if( eType
!=pPtrmap
[offset
] || get4byte(&pPtrmap
[offset
+1])!=parent
){
1000 TRACE(("PTRMAP_UPDATE: %d->(%d,%d)\n", key
, eType
, parent
));
1001 *pRC
= rc
= sqlite3PagerWrite(pDbPage
);
1002 if( rc
==SQLITE_OK
){
1003 pPtrmap
[offset
] = eType
;
1004 put4byte(&pPtrmap
[offset
+1], parent
);
1009 sqlite3PagerUnref(pDbPage
);
1013 ** Read an entry from the pointer map.
1015 ** This routine retrieves the pointer map entry for page 'key', writing
1016 ** the type and parent page number to *pEType and *pPgno respectively.
1017 ** An error code is returned if something goes wrong, otherwise SQLITE_OK.
1019 static int ptrmapGet(BtShared
*pBt
, Pgno key
, u8
*pEType
, Pgno
*pPgno
){
1020 DbPage
*pDbPage
; /* The pointer map page */
1021 int iPtrmap
; /* Pointer map page index */
1022 u8
*pPtrmap
; /* Pointer map page data */
1023 int offset
; /* Offset of entry in pointer map */
1026 assert( sqlite3_mutex_held(pBt
->mutex
) );
1028 iPtrmap
= PTRMAP_PAGENO(pBt
, key
);
1029 rc
= sqlite3PagerGet(pBt
->pPager
, iPtrmap
, &pDbPage
, 0);
1033 pPtrmap
= (u8
*)sqlite3PagerGetData(pDbPage
);
1035 offset
= PTRMAP_PTROFFSET(iPtrmap
, key
);
1037 sqlite3PagerUnref(pDbPage
);
1038 return SQLITE_CORRUPT_BKPT
;
1040 assert( offset
<= (int)pBt
->usableSize
-5 );
1041 assert( pEType
!=0 );
1042 *pEType
= pPtrmap
[offset
];
1043 if( pPgno
) *pPgno
= get4byte(&pPtrmap
[offset
+1]);
1045 sqlite3PagerUnref(pDbPage
);
1046 if( *pEType
<1 || *pEType
>5 ) return SQLITE_CORRUPT_PGNO(iPtrmap
);
1050 #else /* if defined SQLITE_OMIT_AUTOVACUUM */
1051 #define ptrmapPut(w,x,y,z,rc)
1052 #define ptrmapGet(w,x,y,z) SQLITE_OK
1053 #define ptrmapPutOvflPtr(x, y, rc)
1057 ** Given a btree page and a cell index (0 means the first cell on
1058 ** the page, 1 means the second cell, and so forth) return a pointer
1059 ** to the cell content.
1061 ** findCellPastPtr() does the same except it skips past the initial
1062 ** 4-byte child pointer found on interior pages, if there is one.
1064 ** This routine works only for pages that do not contain overflow cells.
1066 #define findCell(P,I) \
1067 ((P)->aData + ((P)->maskPage & get2byteAligned(&(P)->aCellIdx[2*(I)])))
1068 #define findCellPastPtr(P,I) \
1069 ((P)->aDataOfst + ((P)->maskPage & get2byteAligned(&(P)->aCellIdx[2*(I)])))
1073 ** This is common tail processing for btreeParseCellPtr() and
1074 ** btreeParseCellPtrIndex() for the case when the cell does not fit entirely
1075 ** on a single B-tree page. Make necessary adjustments to the CellInfo
1078 static SQLITE_NOINLINE
void btreeParseCellAdjustSizeForOverflow(
1079 MemPage
*pPage
, /* Page containing the cell */
1080 u8
*pCell
, /* Pointer to the cell text. */
1081 CellInfo
*pInfo
/* Fill in this structure */
1083 /* If the payload will not fit completely on the local page, we have
1084 ** to decide how much to store locally and how much to spill onto
1085 ** overflow pages. The strategy is to minimize the amount of unused
1086 ** space on overflow pages while keeping the amount of local storage
1087 ** in between minLocal and maxLocal.
1089 ** Warning: changing the way overflow payload is distributed in any
1090 ** way will result in an incompatible file format.
1092 int minLocal
; /* Minimum amount of payload held locally */
1093 int maxLocal
; /* Maximum amount of payload held locally */
1094 int surplus
; /* Overflow payload available for local storage */
1096 minLocal
= pPage
->minLocal
;
1097 maxLocal
= pPage
->maxLocal
;
1098 surplus
= minLocal
+ (pInfo
->nPayload
- minLocal
)%(pPage
->pBt
->usableSize
-4);
1099 testcase( surplus
==maxLocal
);
1100 testcase( surplus
==maxLocal
+1 );
1101 if( surplus
<= maxLocal
){
1102 pInfo
->nLocal
= (u16
)surplus
;
1104 pInfo
->nLocal
= (u16
)minLocal
;
1106 pInfo
->nSize
= (u16
)(&pInfo
->pPayload
[pInfo
->nLocal
] - pCell
) + 4;
1110 ** The following routines are implementations of the MemPage.xParseCell()
1113 ** Parse a cell content block and fill in the CellInfo structure.
1115 ** btreeParseCellPtr() => table btree leaf nodes
1116 ** btreeParseCellNoPayload() => table btree internal nodes
1117 ** btreeParseCellPtrIndex() => index btree nodes
1119 ** There is also a wrapper function btreeParseCell() that works for
1120 ** all MemPage types and that references the cell by index rather than
1123 static void btreeParseCellPtrNoPayload(
1124 MemPage
*pPage
, /* Page containing the cell */
1125 u8
*pCell
, /* Pointer to the cell text. */
1126 CellInfo
*pInfo
/* Fill in this structure */
1128 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1129 assert( pPage
->leaf
==0 );
1130 assert( pPage
->childPtrSize
==4 );
1131 #ifndef SQLITE_DEBUG
1132 UNUSED_PARAMETER(pPage
);
1134 pInfo
->nSize
= 4 + getVarint(&pCell
[4], (u64
*)&pInfo
->nKey
);
1135 pInfo
->nPayload
= 0;
1137 pInfo
->pPayload
= 0;
1140 static void btreeParseCellPtr(
1141 MemPage
*pPage
, /* Page containing the cell */
1142 u8
*pCell
, /* Pointer to the cell text. */
1143 CellInfo
*pInfo
/* Fill in this structure */
1145 u8
*pIter
; /* For scanning through pCell */
1146 u32 nPayload
; /* Number of bytes of cell payload */
1147 u64 iKey
; /* Extracted Key value */
1149 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1150 assert( pPage
->leaf
==0 || pPage
->leaf
==1 );
1151 assert( pPage
->intKeyLeaf
);
1152 assert( pPage
->childPtrSize
==0 );
1155 /* The next block of code is equivalent to:
1157 ** pIter += getVarint32(pIter, nPayload);
1159 ** The code is inlined to avoid a function call.
1162 if( nPayload
>=0x80 ){
1163 u8
*pEnd
= &pIter
[8];
1166 nPayload
= (nPayload
<<7) | (*++pIter
& 0x7f);
1167 }while( (*pIter
)>=0x80 && pIter
<pEnd
);
1171 /* The next block of code is equivalent to:
1173 ** pIter += getVarint(pIter, (u64*)&pInfo->nKey);
1175 ** The code is inlined to avoid a function call.
1179 u8
*pEnd
= &pIter
[7];
1182 iKey
= (iKey
<<7) | (*++pIter
& 0x7f);
1183 if( (*pIter
)<0x80 ) break;
1185 iKey
= (iKey
<<8) | *++pIter
;
1192 pInfo
->nKey
= *(i64
*)&iKey
;
1193 pInfo
->nPayload
= nPayload
;
1194 pInfo
->pPayload
= pIter
;
1195 testcase( nPayload
==pPage
->maxLocal
);
1196 testcase( nPayload
==pPage
->maxLocal
+1 );
1197 if( nPayload
<=pPage
->maxLocal
){
1198 /* This is the (easy) common case where the entire payload fits
1199 ** on the local page. No overflow is required.
1201 pInfo
->nSize
= nPayload
+ (u16
)(pIter
- pCell
);
1202 if( pInfo
->nSize
<4 ) pInfo
->nSize
= 4;
1203 pInfo
->nLocal
= (u16
)nPayload
;
1205 btreeParseCellAdjustSizeForOverflow(pPage
, pCell
, pInfo
);
1208 static void btreeParseCellPtrIndex(
1209 MemPage
*pPage
, /* Page containing the cell */
1210 u8
*pCell
, /* Pointer to the cell text. */
1211 CellInfo
*pInfo
/* Fill in this structure */
1213 u8
*pIter
; /* For scanning through pCell */
1214 u32 nPayload
; /* Number of bytes of cell payload */
1216 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1217 assert( pPage
->leaf
==0 || pPage
->leaf
==1 );
1218 assert( pPage
->intKeyLeaf
==0 );
1219 pIter
= pCell
+ pPage
->childPtrSize
;
1221 if( nPayload
>=0x80 ){
1222 u8
*pEnd
= &pIter
[8];
1225 nPayload
= (nPayload
<<7) | (*++pIter
& 0x7f);
1226 }while( *(pIter
)>=0x80 && pIter
<pEnd
);
1229 pInfo
->nKey
= nPayload
;
1230 pInfo
->nPayload
= nPayload
;
1231 pInfo
->pPayload
= pIter
;
1232 testcase( nPayload
==pPage
->maxLocal
);
1233 testcase( nPayload
==pPage
->maxLocal
+1 );
1234 if( nPayload
<=pPage
->maxLocal
){
1235 /* This is the (easy) common case where the entire payload fits
1236 ** on the local page. No overflow is required.
1238 pInfo
->nSize
= nPayload
+ (u16
)(pIter
- pCell
);
1239 if( pInfo
->nSize
<4 ) pInfo
->nSize
= 4;
1240 pInfo
->nLocal
= (u16
)nPayload
;
1242 btreeParseCellAdjustSizeForOverflow(pPage
, pCell
, pInfo
);
1245 static void btreeParseCell(
1246 MemPage
*pPage
, /* Page containing the cell */
1247 int iCell
, /* The cell index. First cell is 0 */
1248 CellInfo
*pInfo
/* Fill in this structure */
1250 pPage
->xParseCell(pPage
, findCell(pPage
, iCell
), pInfo
);
1254 ** The following routines are implementations of the MemPage.xCellSize
1257 ** Compute the total number of bytes that a Cell needs in the cell
1258 ** data area of the btree-page. The return number includes the cell
1259 ** data header and the local payload, but not any overflow page or
1260 ** the space used by the cell pointer.
1262 ** cellSizePtrNoPayload() => table internal nodes
1263 ** cellSizePtr() => all index nodes & table leaf nodes
1265 static u16
cellSizePtr(MemPage
*pPage
, u8
*pCell
){
1266 u8
*pIter
= pCell
+ pPage
->childPtrSize
; /* For looping over bytes of pCell */
1267 u8
*pEnd
; /* End mark for a varint */
1268 u32 nSize
; /* Size value to return */
1271 /* The value returned by this function should always be the same as
1272 ** the (CellInfo.nSize) value found by doing a full parse of the
1273 ** cell. If SQLITE_DEBUG is defined, an assert() at the bottom of
1274 ** this function verifies that this invariant is not violated. */
1276 pPage
->xParseCell(pPage
, pCell
, &debuginfo
);
1284 nSize
= (nSize
<<7) | (*++pIter
& 0x7f);
1285 }while( *(pIter
)>=0x80 && pIter
<pEnd
);
1288 if( pPage
->intKey
){
1289 /* pIter now points at the 64-bit integer key value, a variable length
1290 ** integer. The following block moves pIter to point at the first byte
1291 ** past the end of the key value. */
1293 while( (*pIter
++)&0x80 && pIter
<pEnd
);
1295 testcase( nSize
==pPage
->maxLocal
);
1296 testcase( nSize
==pPage
->maxLocal
+1 );
1297 if( nSize
<=pPage
->maxLocal
){
1298 nSize
+= (u32
)(pIter
- pCell
);
1299 if( nSize
<4 ) nSize
= 4;
1301 int minLocal
= pPage
->minLocal
;
1302 nSize
= minLocal
+ (nSize
- minLocal
) % (pPage
->pBt
->usableSize
- 4);
1303 testcase( nSize
==pPage
->maxLocal
);
1304 testcase( nSize
==pPage
->maxLocal
+1 );
1305 if( nSize
>pPage
->maxLocal
){
1308 nSize
+= 4 + (u16
)(pIter
- pCell
);
1310 assert( nSize
==debuginfo
.nSize
|| CORRUPT_DB
);
1313 static u16
cellSizePtrNoPayload(MemPage
*pPage
, u8
*pCell
){
1314 u8
*pIter
= pCell
+ 4; /* For looping over bytes of pCell */
1315 u8
*pEnd
; /* End mark for a varint */
1318 /* The value returned by this function should always be the same as
1319 ** the (CellInfo.nSize) value found by doing a full parse of the
1320 ** cell. If SQLITE_DEBUG is defined, an assert() at the bottom of
1321 ** this function verifies that this invariant is not violated. */
1323 pPage
->xParseCell(pPage
, pCell
, &debuginfo
);
1325 UNUSED_PARAMETER(pPage
);
1328 assert( pPage
->childPtrSize
==4 );
1330 while( (*pIter
++)&0x80 && pIter
<pEnd
);
1331 assert( debuginfo
.nSize
==(u16
)(pIter
- pCell
) || CORRUPT_DB
);
1332 return (u16
)(pIter
- pCell
);
1337 /* This variation on cellSizePtr() is used inside of assert() statements
1339 static u16
cellSize(MemPage
*pPage
, int iCell
){
1340 return pPage
->xCellSize(pPage
, findCell(pPage
, iCell
));
1344 #ifndef SQLITE_OMIT_AUTOVACUUM
1346 ** If the cell pCell, part of page pPage contains a pointer
1347 ** to an overflow page, insert an entry into the pointer-map
1348 ** for the overflow page.
1350 static void ptrmapPutOvflPtr(MemPage
*pPage
, u8
*pCell
, int *pRC
){
1354 pPage
->xParseCell(pPage
, pCell
, &info
);
1355 if( info
.nLocal
<info
.nPayload
){
1356 Pgno ovfl
= get4byte(&pCell
[info
.nSize
-4]);
1357 ptrmapPut(pPage
->pBt
, ovfl
, PTRMAP_OVERFLOW1
, pPage
->pgno
, pRC
);
1364 ** Defragment the page given. This routine reorganizes cells within the
1365 ** page so that there are no free-blocks on the free-block list.
1367 ** Parameter nMaxFrag is the maximum amount of fragmented space that may be
1368 ** present in the page after this routine returns.
1370 ** EVIDENCE-OF: R-44582-60138 SQLite may from time to time reorganize a
1371 ** b-tree page so that there are no freeblocks or fragment bytes, all
1372 ** unused bytes are contained in the unallocated space region, and all
1373 ** cells are packed tightly at the end of the page.
1375 static int defragmentPage(MemPage
*pPage
, int nMaxFrag
){
1376 int i
; /* Loop counter */
1377 int pc
; /* Address of the i-th cell */
1378 int hdr
; /* Offset to the page header */
1379 int size
; /* Size of a cell */
1380 int usableSize
; /* Number of usable bytes on a page */
1381 int cellOffset
; /* Offset to the cell pointer array */
1382 int cbrk
; /* Offset to the cell content area */
1383 int nCell
; /* Number of cells on the page */
1384 unsigned char *data
; /* The page data */
1385 unsigned char *temp
; /* Temp area for cell content */
1386 unsigned char *src
; /* Source of content */
1387 int iCellFirst
; /* First allowable cell index */
1388 int iCellLast
; /* Last possible cell index */
1390 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1391 assert( pPage
->pBt
!=0 );
1392 assert( pPage
->pBt
->usableSize
<= SQLITE_MAX_PAGE_SIZE
);
1393 assert( pPage
->nOverflow
==0 );
1394 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1396 src
= data
= pPage
->aData
;
1397 hdr
= pPage
->hdrOffset
;
1398 cellOffset
= pPage
->cellOffset
;
1399 nCell
= pPage
->nCell
;
1400 assert( nCell
==get2byte(&data
[hdr
+3]) );
1401 iCellFirst
= cellOffset
+ 2*nCell
;
1402 usableSize
= pPage
->pBt
->usableSize
;
1404 /* This block handles pages with two or fewer free blocks and nMaxFrag
1405 ** or fewer fragmented bytes. In this case it is faster to move the
1406 ** two (or one) blocks of cells using memmove() and add the required
1407 ** offsets to each pointer in the cell-pointer array than it is to
1408 ** reconstruct the entire page. */
1409 if( (int)data
[hdr
+7]<=nMaxFrag
){
1410 int iFree
= get2byte(&data
[hdr
+1]);
1412 int iFree2
= get2byte(&data
[iFree
]);
1414 /* pageFindSlot() has already verified that free blocks are sorted
1415 ** in order of offset within the page, and that no block extends
1416 ** past the end of the page. Provided the two free slots do not
1417 ** overlap, this guarantees that the memmove() calls below will not
1418 ** overwrite the usableSize byte buffer, even if the database page
1420 assert( iFree2
==0 || iFree2
>iFree
);
1421 assert( iFree
+get2byte(&data
[iFree
+2]) <= usableSize
);
1422 assert( iFree2
==0 || iFree2
+get2byte(&data
[iFree2
+2]) <= usableSize
);
1424 if( 0==iFree2
|| (data
[iFree2
]==0 && data
[iFree2
+1]==0) ){
1425 u8
*pEnd
= &data
[cellOffset
+ nCell
*2];
1428 int sz
= get2byte(&data
[iFree
+2]);
1429 int top
= get2byte(&data
[hdr
+5]);
1431 return SQLITE_CORRUPT_PAGE(pPage
);
1434 assert( iFree
+sz
<=iFree2
); /* Verified by pageFindSlot() */
1435 sz2
= get2byte(&data
[iFree2
+2]);
1436 assert( iFree
+sz
+sz2
+iFree2
-(iFree
+sz
) <= usableSize
);
1437 memmove(&data
[iFree
+sz
+sz2
], &data
[iFree
+sz
], iFree2
-(iFree
+sz
));
1441 assert( cbrk
+(iFree
-top
) <= usableSize
);
1442 memmove(&data
[cbrk
], &data
[top
], iFree
-top
);
1443 for(pAddr
=&data
[cellOffset
]; pAddr
<pEnd
; pAddr
+=2){
1444 pc
= get2byte(pAddr
);
1445 if( pc
<iFree
){ put2byte(pAddr
, pc
+sz
); }
1446 else if( pc
<iFree2
){ put2byte(pAddr
, pc
+sz2
); }
1448 goto defragment_out
;
1454 iCellLast
= usableSize
- 4;
1455 for(i
=0; i
<nCell
; i
++){
1456 u8
*pAddr
; /* The i-th cell pointer */
1457 pAddr
= &data
[cellOffset
+ i
*2];
1458 pc
= get2byte(pAddr
);
1459 testcase( pc
==iCellFirst
);
1460 testcase( pc
==iCellLast
);
1461 /* These conditions have already been verified in btreeInitPage()
1462 ** if PRAGMA cell_size_check=ON.
1464 if( pc
<iCellFirst
|| pc
>iCellLast
){
1465 return SQLITE_CORRUPT_PAGE(pPage
);
1467 assert( pc
>=iCellFirst
&& pc
<=iCellLast
);
1468 size
= pPage
->xCellSize(pPage
, &src
[pc
]);
1470 if( cbrk
<iCellFirst
|| pc
+size
>usableSize
){
1471 return SQLITE_CORRUPT_PAGE(pPage
);
1473 assert( cbrk
+size
<=usableSize
&& cbrk
>=iCellFirst
);
1474 testcase( cbrk
+size
==usableSize
);
1475 testcase( pc
+size
==usableSize
);
1476 put2byte(pAddr
, cbrk
);
1479 if( cbrk
==pc
) continue;
1480 temp
= sqlite3PagerTempSpace(pPage
->pBt
->pPager
);
1481 x
= get2byte(&data
[hdr
+5]);
1482 memcpy(&temp
[x
], &data
[x
], (cbrk
+size
) - x
);
1485 memcpy(&data
[cbrk
], &src
[pc
], size
);
1490 if( data
[hdr
+7]+cbrk
-iCellFirst
!=pPage
->nFree
){
1491 return SQLITE_CORRUPT_PAGE(pPage
);
1493 assert( cbrk
>=iCellFirst
);
1494 put2byte(&data
[hdr
+5], cbrk
);
1497 memset(&data
[iCellFirst
], 0, cbrk
-iCellFirst
);
1498 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1503 ** Search the free-list on page pPg for space to store a cell nByte bytes in
1504 ** size. If one can be found, return a pointer to the space and remove it
1505 ** from the free-list.
1507 ** If no suitable space can be found on the free-list, return NULL.
1509 ** This function may detect corruption within pPg. If corruption is
1510 ** detected then *pRc is set to SQLITE_CORRUPT and NULL is returned.
1512 ** Slots on the free list that are between 1 and 3 bytes larger than nByte
1513 ** will be ignored if adding the extra space to the fragmentation count
1514 ** causes the fragmentation count to exceed 60.
1516 static u8
*pageFindSlot(MemPage
*pPg
, int nByte
, int *pRc
){
1517 const int hdr
= pPg
->hdrOffset
;
1518 u8
* const aData
= pPg
->aData
;
1519 int iAddr
= hdr
+ 1;
1520 int pc
= get2byte(&aData
[iAddr
]);
1522 int usableSize
= pPg
->pBt
->usableSize
;
1523 int size
; /* Size of the free slot */
1526 while( pc
<=usableSize
-4 ){
1527 /* EVIDENCE-OF: R-22710-53328 The third and fourth bytes of each
1528 ** freeblock form a big-endian integer which is the size of the freeblock
1529 ** in bytes, including the 4-byte header. */
1530 size
= get2byte(&aData
[pc
+2]);
1531 if( (x
= size
- nByte
)>=0 ){
1534 if( size
+pc
> usableSize
){
1535 *pRc
= SQLITE_CORRUPT_PAGE(pPg
);
1538 /* EVIDENCE-OF: R-11498-58022 In a well-formed b-tree page, the total
1539 ** number of bytes in fragments may not exceed 60. */
1540 if( aData
[hdr
+7]>57 ) return 0;
1542 /* Remove the slot from the free-list. Update the number of
1543 ** fragmented bytes within the page. */
1544 memcpy(&aData
[iAddr
], &aData
[pc
], 2);
1545 aData
[hdr
+7] += (u8
)x
;
1547 /* The slot remains on the free-list. Reduce its size to account
1548 ** for the portion used by the new allocation. */
1549 put2byte(&aData
[pc
+2], x
);
1551 return &aData
[pc
+ x
];
1554 pc
= get2byte(&aData
[pc
]);
1555 if( pc
<iAddr
+size
) break;
1558 *pRc
= SQLITE_CORRUPT_PAGE(pPg
);
1565 ** Allocate nByte bytes of space from within the B-Tree page passed
1566 ** as the first argument. Write into *pIdx the index into pPage->aData[]
1567 ** of the first byte of allocated space. Return either SQLITE_OK or
1568 ** an error code (usually SQLITE_CORRUPT).
1570 ** The caller guarantees that there is sufficient space to make the
1571 ** allocation. This routine might need to defragment in order to bring
1572 ** all the space together, however. This routine will avoid using
1573 ** the first two bytes past the cell pointer area since presumably this
1574 ** allocation is being made in order to insert a new cell, so we will
1575 ** also end up needing a new cell pointer.
1577 static int allocateSpace(MemPage
*pPage
, int nByte
, int *pIdx
){
1578 const int hdr
= pPage
->hdrOffset
; /* Local cache of pPage->hdrOffset */
1579 u8
* const data
= pPage
->aData
; /* Local cache of pPage->aData */
1580 int top
; /* First byte of cell content area */
1581 int rc
= SQLITE_OK
; /* Integer return code */
1582 int gap
; /* First byte of gap between cell pointers and cell content */
1584 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1585 assert( pPage
->pBt
);
1586 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1587 assert( nByte
>=0 ); /* Minimum cell size is 4 */
1588 assert( pPage
->nFree
>=nByte
);
1589 assert( pPage
->nOverflow
==0 );
1590 assert( nByte
< (int)(pPage
->pBt
->usableSize
-8) );
1592 assert( pPage
->cellOffset
== hdr
+ 12 - 4*pPage
->leaf
);
1593 gap
= pPage
->cellOffset
+ 2*pPage
->nCell
;
1594 assert( gap
<=65536 );
1595 /* EVIDENCE-OF: R-29356-02391 If the database uses a 65536-byte page size
1596 ** and the reserved space is zero (the usual value for reserved space)
1597 ** then the cell content offset of an empty page wants to be 65536.
1598 ** However, that integer is too large to be stored in a 2-byte unsigned
1599 ** integer, so a value of 0 is used in its place. */
1600 top
= get2byte(&data
[hdr
+5]);
1601 assert( top
<=(int)pPage
->pBt
->usableSize
); /* Prevent by getAndInitPage() */
1603 if( top
==0 && pPage
->pBt
->usableSize
==65536 ){
1606 return SQLITE_CORRUPT_PAGE(pPage
);
1610 /* If there is enough space between gap and top for one more cell pointer
1611 ** array entry offset, and if the freelist is not empty, then search the
1612 ** freelist looking for a free slot big enough to satisfy the request.
1614 testcase( gap
+2==top
);
1615 testcase( gap
+1==top
);
1616 testcase( gap
==top
);
1617 if( (data
[hdr
+2] || data
[hdr
+1]) && gap
+2<=top
){
1618 u8
*pSpace
= pageFindSlot(pPage
, nByte
, &rc
);
1620 assert( pSpace
>=data
&& (pSpace
- data
)<65536 );
1621 *pIdx
= (int)(pSpace
- data
);
1628 /* The request could not be fulfilled using a freelist slot. Check
1629 ** to see if defragmentation is necessary.
1631 testcase( gap
+2+nByte
==top
);
1632 if( gap
+2+nByte
>top
){
1633 assert( pPage
->nCell
>0 || CORRUPT_DB
);
1634 rc
= defragmentPage(pPage
, MIN(4, pPage
->nFree
- (2+nByte
)));
1636 top
= get2byteNotZero(&data
[hdr
+5]);
1637 assert( gap
+2+nByte
<=top
);
1641 /* Allocate memory from the gap in between the cell pointer array
1642 ** and the cell content area. The btreeInitPage() call has already
1643 ** validated the freelist. Given that the freelist is valid, there
1644 ** is no way that the allocation can extend off the end of the page.
1645 ** The assert() below verifies the previous sentence.
1648 put2byte(&data
[hdr
+5], top
);
1649 assert( top
+nByte
<= (int)pPage
->pBt
->usableSize
);
1655 ** Return a section of the pPage->aData to the freelist.
1656 ** The first byte of the new free block is pPage->aData[iStart]
1657 ** and the size of the block is iSize bytes.
1659 ** Adjacent freeblocks are coalesced.
1661 ** Note that even though the freeblock list was checked by btreeInitPage(),
1662 ** that routine will not detect overlap between cells or freeblocks. Nor
1663 ** does it detect cells or freeblocks that encrouch into the reserved bytes
1664 ** at the end of the page. So do additional corruption checks inside this
1665 ** routine and return SQLITE_CORRUPT if any problems are found.
1667 static int freeSpace(MemPage
*pPage
, u16 iStart
, u16 iSize
){
1668 u16 iPtr
; /* Address of ptr to next freeblock */
1669 u16 iFreeBlk
; /* Address of the next freeblock */
1670 u8 hdr
; /* Page header size. 0 or 100 */
1671 u8 nFrag
= 0; /* Reduction in fragmentation */
1672 u16 iOrigSize
= iSize
; /* Original value of iSize */
1673 u16 x
; /* Offset to cell content area */
1674 u32 iEnd
= iStart
+ iSize
; /* First byte past the iStart buffer */
1675 unsigned char *data
= pPage
->aData
; /* Page content */
1677 assert( pPage
->pBt
!=0 );
1678 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1679 assert( CORRUPT_DB
|| iStart
>=pPage
->hdrOffset
+6+pPage
->childPtrSize
);
1680 assert( CORRUPT_DB
|| iEnd
<= pPage
->pBt
->usableSize
);
1681 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1682 assert( iSize
>=4 ); /* Minimum cell size is 4 */
1683 assert( iStart
<=pPage
->pBt
->usableSize
-4 );
1685 /* The list of freeblocks must be in ascending order. Find the
1686 ** spot on the list where iStart should be inserted.
1688 hdr
= pPage
->hdrOffset
;
1690 if( data
[iPtr
+1]==0 && data
[iPtr
]==0 ){
1691 iFreeBlk
= 0; /* Shortcut for the case when the freelist is empty */
1693 while( (iFreeBlk
= get2byte(&data
[iPtr
]))<iStart
){
1694 if( iFreeBlk
<iPtr
+4 ){
1695 if( iFreeBlk
==0 ) break;
1696 return SQLITE_CORRUPT_PAGE(pPage
);
1700 if( iFreeBlk
>pPage
->pBt
->usableSize
-4 ){
1701 return SQLITE_CORRUPT_PAGE(pPage
);
1703 assert( iFreeBlk
>iPtr
|| iFreeBlk
==0 );
1706 ** iFreeBlk: First freeblock after iStart, or zero if none
1707 ** iPtr: The address of a pointer to iFreeBlk
1709 ** Check to see if iFreeBlk should be coalesced onto the end of iStart.
1711 if( iFreeBlk
&& iEnd
+3>=iFreeBlk
){
1712 nFrag
= iFreeBlk
- iEnd
;
1713 if( iEnd
>iFreeBlk
) return SQLITE_CORRUPT_PAGE(pPage
);
1714 iEnd
= iFreeBlk
+ get2byte(&data
[iFreeBlk
+2]);
1715 if( iEnd
> pPage
->pBt
->usableSize
){
1716 return SQLITE_CORRUPT_PAGE(pPage
);
1718 iSize
= iEnd
- iStart
;
1719 iFreeBlk
= get2byte(&data
[iFreeBlk
]);
1722 /* If iPtr is another freeblock (that is, if iPtr is not the freelist
1723 ** pointer in the page header) then check to see if iStart should be
1724 ** coalesced onto the end of iPtr.
1727 int iPtrEnd
= iPtr
+ get2byte(&data
[iPtr
+2]);
1728 if( iPtrEnd
+3>=iStart
){
1729 if( iPtrEnd
>iStart
) return SQLITE_CORRUPT_PAGE(pPage
);
1730 nFrag
+= iStart
- iPtrEnd
;
1731 iSize
= iEnd
- iPtr
;
1735 if( nFrag
>data
[hdr
+7] ) return SQLITE_CORRUPT_PAGE(pPage
);
1736 data
[hdr
+7] -= nFrag
;
1738 x
= get2byte(&data
[hdr
+5]);
1740 /* The new freeblock is at the beginning of the cell content area,
1741 ** so just extend the cell content area rather than create another
1742 ** freelist entry */
1743 if( iStart
<x
|| iPtr
!=hdr
+1 ) return SQLITE_CORRUPT_PAGE(pPage
);
1744 put2byte(&data
[hdr
+1], iFreeBlk
);
1745 put2byte(&data
[hdr
+5], iEnd
);
1747 /* Insert the new freeblock into the freelist */
1748 put2byte(&data
[iPtr
], iStart
);
1750 if( pPage
->pBt
->btsFlags
& BTS_FAST_SECURE
){
1751 /* Overwrite deleted information with zeros when the secure_delete
1752 ** option is enabled */
1753 memset(&data
[iStart
], 0, iSize
);
1755 put2byte(&data
[iStart
], iFreeBlk
);
1756 put2byte(&data
[iStart
+2], iSize
);
1757 pPage
->nFree
+= iOrigSize
;
1762 ** Decode the flags byte (the first byte of the header) for a page
1763 ** and initialize fields of the MemPage structure accordingly.
1765 ** Only the following combinations are supported. Anything different
1766 ** indicates a corrupt database files:
1769 ** PTF_ZERODATA | PTF_LEAF
1770 ** PTF_LEAFDATA | PTF_INTKEY
1771 ** PTF_LEAFDATA | PTF_INTKEY | PTF_LEAF
1773 static int decodeFlags(MemPage
*pPage
, int flagByte
){
1774 BtShared
*pBt
; /* A copy of pPage->pBt */
1776 assert( pPage
->hdrOffset
==(pPage
->pgno
==1 ? 100 : 0) );
1777 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1778 pPage
->leaf
= (u8
)(flagByte
>>3); assert( PTF_LEAF
== 1<<3 );
1779 flagByte
&= ~PTF_LEAF
;
1780 pPage
->childPtrSize
= 4-4*pPage
->leaf
;
1781 pPage
->xCellSize
= cellSizePtr
;
1783 if( flagByte
==(PTF_LEAFDATA
| PTF_INTKEY
) ){
1784 /* EVIDENCE-OF: R-07291-35328 A value of 5 (0x05) means the page is an
1785 ** interior table b-tree page. */
1786 assert( (PTF_LEAFDATA
|PTF_INTKEY
)==5 );
1787 /* EVIDENCE-OF: R-26900-09176 A value of 13 (0x0d) means the page is a
1788 ** leaf table b-tree page. */
1789 assert( (PTF_LEAFDATA
|PTF_INTKEY
|PTF_LEAF
)==13 );
1792 pPage
->intKeyLeaf
= 1;
1793 pPage
->xParseCell
= btreeParseCellPtr
;
1795 pPage
->intKeyLeaf
= 0;
1796 pPage
->xCellSize
= cellSizePtrNoPayload
;
1797 pPage
->xParseCell
= btreeParseCellPtrNoPayload
;
1799 pPage
->maxLocal
= pBt
->maxLeaf
;
1800 pPage
->minLocal
= pBt
->minLeaf
;
1801 }else if( flagByte
==PTF_ZERODATA
){
1802 /* EVIDENCE-OF: R-43316-37308 A value of 2 (0x02) means the page is an
1803 ** interior index b-tree page. */
1804 assert( (PTF_ZERODATA
)==2 );
1805 /* EVIDENCE-OF: R-59615-42828 A value of 10 (0x0a) means the page is a
1806 ** leaf index b-tree page. */
1807 assert( (PTF_ZERODATA
|PTF_LEAF
)==10 );
1809 pPage
->intKeyLeaf
= 0;
1810 pPage
->xParseCell
= btreeParseCellPtrIndex
;
1811 pPage
->maxLocal
= pBt
->maxLocal
;
1812 pPage
->minLocal
= pBt
->minLocal
;
1814 /* EVIDENCE-OF: R-47608-56469 Any other value for the b-tree page type is
1816 return SQLITE_CORRUPT_PAGE(pPage
);
1818 pPage
->max1bytePayload
= pBt
->max1bytePayload
;
1823 ** Initialize the auxiliary information for a disk block.
1825 ** Return SQLITE_OK on success. If we see that the page does
1826 ** not contain a well-formed database page, then return
1827 ** SQLITE_CORRUPT. Note that a return of SQLITE_OK does not
1828 ** guarantee that the page is well-formed. It only shows that
1829 ** we failed to detect any corruption.
1831 static int btreeInitPage(MemPage
*pPage
){
1832 int pc
; /* Address of a freeblock within pPage->aData[] */
1833 u8 hdr
; /* Offset to beginning of page header */
1834 u8
*data
; /* Equal to pPage->aData */
1835 BtShared
*pBt
; /* The main btree structure */
1836 int usableSize
; /* Amount of usable space on each page */
1837 u16 cellOffset
; /* Offset from start of page to first cell pointer */
1838 int nFree
; /* Number of unused bytes on the page */
1839 int top
; /* First byte of the cell content area */
1840 int iCellFirst
; /* First allowable cell or freeblock offset */
1841 int iCellLast
; /* Last possible cell or freeblock offset */
1843 assert( pPage
->pBt
!=0 );
1844 assert( pPage
->pBt
->db
!=0 );
1845 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1846 assert( pPage
->pgno
==sqlite3PagerPagenumber(pPage
->pDbPage
) );
1847 assert( pPage
== sqlite3PagerGetExtra(pPage
->pDbPage
) );
1848 assert( pPage
->aData
== sqlite3PagerGetData(pPage
->pDbPage
) );
1849 assert( pPage
->isInit
==0 );
1852 hdr
= pPage
->hdrOffset
;
1853 data
= pPage
->aData
;
1854 /* EVIDENCE-OF: R-28594-02890 The one-byte flag at offset 0 indicating
1855 ** the b-tree page type. */
1856 if( decodeFlags(pPage
, data
[hdr
]) ){
1857 return SQLITE_CORRUPT_PAGE(pPage
);
1859 assert( pBt
->pageSize
>=512 && pBt
->pageSize
<=65536 );
1860 pPage
->maskPage
= (u16
)(pBt
->pageSize
- 1);
1861 pPage
->nOverflow
= 0;
1862 usableSize
= pBt
->usableSize
;
1863 pPage
->cellOffset
= cellOffset
= hdr
+ 8 + pPage
->childPtrSize
;
1864 pPage
->aDataEnd
= &data
[usableSize
];
1865 pPage
->aCellIdx
= &data
[cellOffset
];
1866 pPage
->aDataOfst
= &data
[pPage
->childPtrSize
];
1867 /* EVIDENCE-OF: R-58015-48175 The two-byte integer at offset 5 designates
1868 ** the start of the cell content area. A zero value for this integer is
1869 ** interpreted as 65536. */
1870 top
= get2byteNotZero(&data
[hdr
+5]);
1871 /* EVIDENCE-OF: R-37002-32774 The two-byte integer at offset 3 gives the
1872 ** number of cells on the page. */
1873 pPage
->nCell
= get2byte(&data
[hdr
+3]);
1874 if( pPage
->nCell
>MX_CELL(pBt
) ){
1875 /* To many cells for a single page. The page must be corrupt */
1876 return SQLITE_CORRUPT_PAGE(pPage
);
1878 testcase( pPage
->nCell
==MX_CELL(pBt
) );
1879 /* EVIDENCE-OF: R-24089-57979 If a page contains no cells (which is only
1880 ** possible for a root page of a table that contains no rows) then the
1881 ** offset to the cell content area will equal the page size minus the
1882 ** bytes of reserved space. */
1883 assert( pPage
->nCell
>0 || top
==usableSize
|| CORRUPT_DB
);
1885 /* A malformed database page might cause us to read past the end
1886 ** of page when parsing a cell.
1888 ** The following block of code checks early to see if a cell extends
1889 ** past the end of a page boundary and causes SQLITE_CORRUPT to be
1890 ** returned if it does.
1892 iCellFirst
= cellOffset
+ 2*pPage
->nCell
;
1893 iCellLast
= usableSize
- 4;
1894 if( pBt
->db
->flags
& SQLITE_CellSizeCk
){
1895 int i
; /* Index into the cell pointer array */
1896 int sz
; /* Size of a cell */
1898 if( !pPage
->leaf
) iCellLast
--;
1899 for(i
=0; i
<pPage
->nCell
; i
++){
1900 pc
= get2byteAligned(&data
[cellOffset
+i
*2]);
1901 testcase( pc
==iCellFirst
);
1902 testcase( pc
==iCellLast
);
1903 if( pc
<iCellFirst
|| pc
>iCellLast
){
1904 return SQLITE_CORRUPT_PAGE(pPage
);
1906 sz
= pPage
->xCellSize(pPage
, &data
[pc
]);
1907 testcase( pc
+sz
==usableSize
);
1908 if( pc
+sz
>usableSize
){
1909 return SQLITE_CORRUPT_PAGE(pPage
);
1912 if( !pPage
->leaf
) iCellLast
++;
1915 /* Compute the total free space on the page
1916 ** EVIDENCE-OF: R-23588-34450 The two-byte integer at offset 1 gives the
1917 ** start of the first freeblock on the page, or is zero if there are no
1919 pc
= get2byte(&data
[hdr
+1]);
1920 nFree
= data
[hdr
+7] + top
; /* Init nFree to non-freeblock free space */
1923 if( pc
<iCellFirst
){
1924 /* EVIDENCE-OF: R-55530-52930 In a well-formed b-tree page, there will
1925 ** always be at least one cell before the first freeblock.
1927 return SQLITE_CORRUPT_PAGE(pPage
);
1931 /* Freeblock off the end of the page */
1932 return SQLITE_CORRUPT_PAGE(pPage
);
1934 next
= get2byte(&data
[pc
]);
1935 size
= get2byte(&data
[pc
+2]);
1936 nFree
= nFree
+ size
;
1937 if( next
<=pc
+size
+3 ) break;
1941 /* Freeblock not in ascending order */
1942 return SQLITE_CORRUPT_PAGE(pPage
);
1944 if( pc
+size
>(unsigned int)usableSize
){
1945 /* Last freeblock extends past page end */
1946 return SQLITE_CORRUPT_PAGE(pPage
);
1950 /* At this point, nFree contains the sum of the offset to the start
1951 ** of the cell-content area plus the number of free bytes within
1952 ** the cell-content area. If this is greater than the usable-size
1953 ** of the page, then the page must be corrupted. This check also
1954 ** serves to verify that the offset to the start of the cell-content
1955 ** area, according to the page header, lies within the page.
1957 if( nFree
>usableSize
){
1958 return SQLITE_CORRUPT_PAGE(pPage
);
1960 pPage
->nFree
= (u16
)(nFree
- iCellFirst
);
1966 ** Set up a raw page so that it looks like a database page holding
1969 static void zeroPage(MemPage
*pPage
, int flags
){
1970 unsigned char *data
= pPage
->aData
;
1971 BtShared
*pBt
= pPage
->pBt
;
1972 u8 hdr
= pPage
->hdrOffset
;
1975 assert( sqlite3PagerPagenumber(pPage
->pDbPage
)==pPage
->pgno
);
1976 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
1977 assert( sqlite3PagerGetData(pPage
->pDbPage
) == data
);
1978 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1979 assert( sqlite3_mutex_held(pBt
->mutex
) );
1980 if( pBt
->btsFlags
& BTS_FAST_SECURE
){
1981 memset(&data
[hdr
], 0, pBt
->usableSize
- hdr
);
1983 data
[hdr
] = (char)flags
;
1984 first
= hdr
+ ((flags
&PTF_LEAF
)==0 ? 12 : 8);
1985 memset(&data
[hdr
+1], 0, 4);
1987 put2byte(&data
[hdr
+5], pBt
->usableSize
);
1988 pPage
->nFree
= (u16
)(pBt
->usableSize
- first
);
1989 decodeFlags(pPage
, flags
);
1990 pPage
->cellOffset
= first
;
1991 pPage
->aDataEnd
= &data
[pBt
->usableSize
];
1992 pPage
->aCellIdx
= &data
[first
];
1993 pPage
->aDataOfst
= &data
[pPage
->childPtrSize
];
1994 pPage
->nOverflow
= 0;
1995 assert( pBt
->pageSize
>=512 && pBt
->pageSize
<=65536 );
1996 pPage
->maskPage
= (u16
)(pBt
->pageSize
- 1);
2003 ** Convert a DbPage obtained from the pager into a MemPage used by
2006 static MemPage
*btreePageFromDbPage(DbPage
*pDbPage
, Pgno pgno
, BtShared
*pBt
){
2007 MemPage
*pPage
= (MemPage
*)sqlite3PagerGetExtra(pDbPage
);
2008 if( pgno
!=pPage
->pgno
){
2009 pPage
->aData
= sqlite3PagerGetData(pDbPage
);
2010 pPage
->pDbPage
= pDbPage
;
2013 pPage
->hdrOffset
= pgno
==1 ? 100 : 0;
2015 assert( pPage
->aData
==sqlite3PagerGetData(pDbPage
) );
2020 ** Get a page from the pager. Initialize the MemPage.pBt and
2021 ** MemPage.aData elements if needed. See also: btreeGetUnusedPage().
2023 ** If the PAGER_GET_NOCONTENT flag is set, it means that we do not care
2024 ** about the content of the page at this time. So do not go to the disk
2025 ** to fetch the content. Just fill in the content with zeros for now.
2026 ** If in the future we call sqlite3PagerWrite() on this page, that
2027 ** means we have started to be concerned about content and the disk
2028 ** read should occur at that point.
2030 static int btreeGetPage(
2031 BtShared
*pBt
, /* The btree */
2032 Pgno pgno
, /* Number of the page to fetch */
2033 MemPage
**ppPage
, /* Return the page in this parameter */
2034 int flags
/* PAGER_GET_NOCONTENT or PAGER_GET_READONLY */
2039 assert( flags
==0 || flags
==PAGER_GET_NOCONTENT
|| flags
==PAGER_GET_READONLY
);
2040 assert( sqlite3_mutex_held(pBt
->mutex
) );
2041 rc
= sqlite3PagerGet(pBt
->pPager
, pgno
, (DbPage
**)&pDbPage
, flags
);
2043 *ppPage
= btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2048 ** Retrieve a page from the pager cache. If the requested page is not
2049 ** already in the pager cache return NULL. Initialize the MemPage.pBt and
2050 ** MemPage.aData elements if needed.
2052 static MemPage
*btreePageLookup(BtShared
*pBt
, Pgno pgno
){
2054 assert( sqlite3_mutex_held(pBt
->mutex
) );
2055 pDbPage
= sqlite3PagerLookup(pBt
->pPager
, pgno
);
2057 return btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2063 ** Return the size of the database file in pages. If there is any kind of
2064 ** error, return ((unsigned int)-1).
2066 static Pgno
btreePagecount(BtShared
*pBt
){
2069 u32
sqlite3BtreeLastPage(Btree
*p
){
2070 assert( sqlite3BtreeHoldsMutex(p
) );
2071 assert( ((p
->pBt
->nPage
)&0x80000000)==0 );
2072 return btreePagecount(p
->pBt
);
2076 ** Get a page from the pager and initialize it.
2078 ** If pCur!=0 then the page is being fetched as part of a moveToChild()
2079 ** call. Do additional sanity checking on the page in this case.
2080 ** And if the fetch fails, this routine must decrement pCur->iPage.
2082 ** The page is fetched as read-write unless pCur is not NULL and is
2083 ** a read-only cursor.
2085 ** If an error occurs, then *ppPage is undefined. It
2086 ** may remain unchanged, or it may be set to an invalid value.
2088 static int getAndInitPage(
2089 BtShared
*pBt
, /* The database file */
2090 Pgno pgno
, /* Number of the page to get */
2091 MemPage
**ppPage
, /* Write the page pointer here */
2092 BtCursor
*pCur
, /* Cursor to receive the page, or NULL */
2093 int bReadOnly
/* True for a read-only page */
2097 assert( sqlite3_mutex_held(pBt
->mutex
) );
2098 assert( pCur
==0 || ppPage
==&pCur
->pPage
);
2099 assert( pCur
==0 || bReadOnly
==pCur
->curPagerFlags
);
2100 assert( pCur
==0 || pCur
->iPage
>0 );
2102 if( pgno
>btreePagecount(pBt
) ){
2103 rc
= SQLITE_CORRUPT_BKPT
;
2104 goto getAndInitPage_error
;
2106 rc
= sqlite3PagerGet(pBt
->pPager
, pgno
, (DbPage
**)&pDbPage
, bReadOnly
);
2108 goto getAndInitPage_error
;
2110 *ppPage
= (MemPage
*)sqlite3PagerGetExtra(pDbPage
);
2111 if( (*ppPage
)->isInit
==0 ){
2112 btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2113 rc
= btreeInitPage(*ppPage
);
2114 if( rc
!=SQLITE_OK
){
2115 releasePage(*ppPage
);
2116 goto getAndInitPage_error
;
2119 assert( (*ppPage
)->pgno
==pgno
);
2120 assert( (*ppPage
)->aData
==sqlite3PagerGetData(pDbPage
) );
2122 /* If obtaining a child page for a cursor, we must verify that the page is
2123 ** compatible with the root page. */
2124 if( pCur
&& ((*ppPage
)->nCell
<1 || (*ppPage
)->intKey
!=pCur
->curIntKey
) ){
2125 rc
= SQLITE_CORRUPT_PGNO(pgno
);
2126 releasePage(*ppPage
);
2127 goto getAndInitPage_error
;
2131 getAndInitPage_error
:
2134 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
2136 testcase( pgno
==0 );
2137 assert( pgno
!=0 || rc
==SQLITE_CORRUPT
);
2142 ** Release a MemPage. This should be called once for each prior
2143 ** call to btreeGetPage.
2145 ** Page1 is a special case and must be released using releasePageOne().
2147 static void releasePageNotNull(MemPage
*pPage
){
2148 assert( pPage
->aData
);
2149 assert( pPage
->pBt
);
2150 assert( pPage
->pDbPage
!=0 );
2151 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
2152 assert( sqlite3PagerGetData(pPage
->pDbPage
)==pPage
->aData
);
2153 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2154 sqlite3PagerUnrefNotNull(pPage
->pDbPage
);
2156 static void releasePage(MemPage
*pPage
){
2157 if( pPage
) releasePageNotNull(pPage
);
2159 static void releasePageOne(MemPage
*pPage
){
2161 assert( pPage
->aData
);
2162 assert( pPage
->pBt
);
2163 assert( pPage
->pDbPage
!=0 );
2164 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
2165 assert( sqlite3PagerGetData(pPage
->pDbPage
)==pPage
->aData
);
2166 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2167 sqlite3PagerUnrefPageOne(pPage
->pDbPage
);
2171 ** Get an unused page.
2173 ** This works just like btreeGetPage() with the addition:
2175 ** * If the page is already in use for some other purpose, immediately
2176 ** release it and return an SQLITE_CURRUPT error.
2177 ** * Make sure the isInit flag is clear
2179 static int btreeGetUnusedPage(
2180 BtShared
*pBt
, /* The btree */
2181 Pgno pgno
, /* Number of the page to fetch */
2182 MemPage
**ppPage
, /* Return the page in this parameter */
2183 int flags
/* PAGER_GET_NOCONTENT or PAGER_GET_READONLY */
2185 int rc
= btreeGetPage(pBt
, pgno
, ppPage
, flags
);
2186 if( rc
==SQLITE_OK
){
2187 if( sqlite3PagerPageRefcount((*ppPage
)->pDbPage
)>1 ){
2188 releasePage(*ppPage
);
2190 return SQLITE_CORRUPT_BKPT
;
2192 (*ppPage
)->isInit
= 0;
2201 ** During a rollback, when the pager reloads information into the cache
2202 ** so that the cache is restored to its original state at the start of
2203 ** the transaction, for each page restored this routine is called.
2205 ** This routine needs to reset the extra data section at the end of the
2206 ** page to agree with the restored data.
2208 static void pageReinit(DbPage
*pData
){
2210 pPage
= (MemPage
*)sqlite3PagerGetExtra(pData
);
2211 assert( sqlite3PagerPageRefcount(pData
)>0 );
2212 if( pPage
->isInit
){
2213 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2215 if( sqlite3PagerPageRefcount(pData
)>1 ){
2216 /* pPage might not be a btree page; it might be an overflow page
2217 ** or ptrmap page or a free page. In those cases, the following
2218 ** call to btreeInitPage() will likely return SQLITE_CORRUPT.
2219 ** But no harm is done by this. And it is very important that
2220 ** btreeInitPage() be called on every btree page so we make
2221 ** the call for every page that comes in for re-initing. */
2222 btreeInitPage(pPage
);
2228 ** Invoke the busy handler for a btree.
2230 static int btreeInvokeBusyHandler(void *pArg
){
2231 BtShared
*pBt
= (BtShared
*)pArg
;
2233 assert( sqlite3_mutex_held(pBt
->db
->mutex
) );
2234 return sqlite3InvokeBusyHandler(&pBt
->db
->busyHandler
,
2235 sqlite3PagerFile(pBt
->pPager
));
2239 ** Open a database file.
2241 ** zFilename is the name of the database file. If zFilename is NULL
2242 ** then an ephemeral database is created. The ephemeral database might
2243 ** be exclusively in memory, or it might use a disk-based memory cache.
2244 ** Either way, the ephemeral database will be automatically deleted
2245 ** when sqlite3BtreeClose() is called.
2247 ** If zFilename is ":memory:" then an in-memory database is created
2248 ** that is automatically destroyed when it is closed.
2250 ** The "flags" parameter is a bitmask that might contain bits like
2251 ** BTREE_OMIT_JOURNAL and/or BTREE_MEMORY.
2253 ** If the database is already opened in the same database connection
2254 ** and we are in shared cache mode, then the open will fail with an
2255 ** SQLITE_CONSTRAINT error. We cannot allow two or more BtShared
2256 ** objects in the same database connection since doing so will lead
2257 ** to problems with locking.
2259 int sqlite3BtreeOpen(
2260 sqlite3_vfs
*pVfs
, /* VFS to use for this b-tree */
2261 const char *zFilename
, /* Name of the file containing the BTree database */
2262 sqlite3
*db
, /* Associated database handle */
2263 Btree
**ppBtree
, /* Pointer to new Btree object written here */
2264 int flags
, /* Options */
2265 int vfsFlags
/* Flags passed through to sqlite3_vfs.xOpen() */
2267 BtShared
*pBt
= 0; /* Shared part of btree structure */
2268 Btree
*p
; /* Handle to return */
2269 sqlite3_mutex
*mutexOpen
= 0; /* Prevents a race condition. Ticket #3537 */
2270 int rc
= SQLITE_OK
; /* Result code from this function */
2271 u8 nReserve
; /* Byte of unused space on each page */
2272 unsigned char zDbHeader
[100]; /* Database header content */
2274 /* True if opening an ephemeral, temporary database */
2275 const int isTempDb
= zFilename
==0 || zFilename
[0]==0;
2277 /* Set the variable isMemdb to true for an in-memory database, or
2278 ** false for a file-based database.
2280 #ifdef SQLITE_OMIT_MEMORYDB
2281 const int isMemdb
= 0;
2283 const int isMemdb
= (zFilename
&& strcmp(zFilename
, ":memory:")==0)
2284 || (isTempDb
&& sqlite3TempInMemory(db
))
2285 || (vfsFlags
& SQLITE_OPEN_MEMORY
)!=0;
2290 assert( sqlite3_mutex_held(db
->mutex
) );
2291 assert( (flags
&0xff)==flags
); /* flags fit in 8 bits */
2293 /* Only a BTREE_SINGLE database can be BTREE_UNORDERED */
2294 assert( (flags
& BTREE_UNORDERED
)==0 || (flags
& BTREE_SINGLE
)!=0 );
2296 /* A BTREE_SINGLE database is always a temporary and/or ephemeral */
2297 assert( (flags
& BTREE_SINGLE
)==0 || isTempDb
);
2300 flags
|= BTREE_MEMORY
;
2302 if( (vfsFlags
& SQLITE_OPEN_MAIN_DB
)!=0 && (isMemdb
|| isTempDb
) ){
2303 vfsFlags
= (vfsFlags
& ~SQLITE_OPEN_MAIN_DB
) | SQLITE_OPEN_TEMP_DB
;
2305 p
= sqlite3MallocZero(sizeof(Btree
));
2307 return SQLITE_NOMEM_BKPT
;
2309 p
->inTrans
= TRANS_NONE
;
2311 #ifndef SQLITE_OMIT_SHARED_CACHE
2316 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2318 ** If this Btree is a candidate for shared cache, try to find an
2319 ** existing BtShared object that we can share with
2321 if( isTempDb
==0 && (isMemdb
==0 || (vfsFlags
&SQLITE_OPEN_URI
)!=0) ){
2322 if( vfsFlags
& SQLITE_OPEN_SHAREDCACHE
){
2323 int nFilename
= sqlite3Strlen30(zFilename
)+1;
2324 int nFullPathname
= pVfs
->mxPathname
+1;
2325 char *zFullPathname
= sqlite3Malloc(MAX(nFullPathname
,nFilename
));
2326 MUTEX_LOGIC( sqlite3_mutex
*mutexShared
; )
2329 if( !zFullPathname
){
2331 return SQLITE_NOMEM_BKPT
;
2334 memcpy(zFullPathname
, zFilename
, nFilename
);
2336 rc
= sqlite3OsFullPathname(pVfs
, zFilename
,
2337 nFullPathname
, zFullPathname
);
2339 sqlite3_free(zFullPathname
);
2344 #if SQLITE_THREADSAFE
2345 mutexOpen
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_OPEN
);
2346 sqlite3_mutex_enter(mutexOpen
);
2347 mutexShared
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER
);
2348 sqlite3_mutex_enter(mutexShared
);
2350 for(pBt
=GLOBAL(BtShared
*,sqlite3SharedCacheList
); pBt
; pBt
=pBt
->pNext
){
2351 assert( pBt
->nRef
>0 );
2352 if( 0==strcmp(zFullPathname
, sqlite3PagerFilename(pBt
->pPager
, 0))
2353 && sqlite3PagerVfs(pBt
->pPager
)==pVfs
){
2355 for(iDb
=db
->nDb
-1; iDb
>=0; iDb
--){
2356 Btree
*pExisting
= db
->aDb
[iDb
].pBt
;
2357 if( pExisting
&& pExisting
->pBt
==pBt
){
2358 sqlite3_mutex_leave(mutexShared
);
2359 sqlite3_mutex_leave(mutexOpen
);
2360 sqlite3_free(zFullPathname
);
2362 return SQLITE_CONSTRAINT
;
2370 sqlite3_mutex_leave(mutexShared
);
2371 sqlite3_free(zFullPathname
);
2375 /* In debug mode, we mark all persistent databases as sharable
2376 ** even when they are not. This exercises the locking code and
2377 ** gives more opportunity for asserts(sqlite3_mutex_held())
2378 ** statements to find locking problems.
2387 ** The following asserts make sure that structures used by the btree are
2388 ** the right size. This is to guard against size changes that result
2389 ** when compiling on a different architecture.
2391 assert( sizeof(i64
)==8 );
2392 assert( sizeof(u64
)==8 );
2393 assert( sizeof(u32
)==4 );
2394 assert( sizeof(u16
)==2 );
2395 assert( sizeof(Pgno
)==4 );
2397 pBt
= sqlite3MallocZero( sizeof(*pBt
) );
2399 rc
= SQLITE_NOMEM_BKPT
;
2400 goto btree_open_out
;
2402 rc
= sqlite3PagerOpen(pVfs
, &pBt
->pPager
, zFilename
,
2403 sizeof(MemPage
), flags
, vfsFlags
, pageReinit
);
2404 if( rc
==SQLITE_OK
){
2405 sqlite3PagerSetMmapLimit(pBt
->pPager
, db
->szMmap
);
2406 rc
= sqlite3PagerReadFileheader(pBt
->pPager
,sizeof(zDbHeader
),zDbHeader
);
2408 if( rc
!=SQLITE_OK
){
2409 goto btree_open_out
;
2411 pBt
->openFlags
= (u8
)flags
;
2413 sqlite3PagerSetBusyHandler(pBt
->pPager
, btreeInvokeBusyHandler
, pBt
);
2418 if( sqlite3PagerIsreadonly(pBt
->pPager
) ) pBt
->btsFlags
|= BTS_READ_ONLY
;
2419 #if defined(SQLITE_SECURE_DELETE)
2420 pBt
->btsFlags
|= BTS_SECURE_DELETE
;
2421 #elif defined(SQLITE_FAST_SECURE_DELETE)
2422 pBt
->btsFlags
|= BTS_OVERWRITE
;
2424 /* EVIDENCE-OF: R-51873-39618 The page size for a database file is
2425 ** determined by the 2-byte integer located at an offset of 16 bytes from
2426 ** the beginning of the database file. */
2427 pBt
->pageSize
= (zDbHeader
[16]<<8) | (zDbHeader
[17]<<16);
2428 if( pBt
->pageSize
<512 || pBt
->pageSize
>SQLITE_MAX_PAGE_SIZE
2429 || ((pBt
->pageSize
-1)&pBt
->pageSize
)!=0 ){
2431 #ifndef SQLITE_OMIT_AUTOVACUUM
2432 /* If the magic name ":memory:" will create an in-memory database, then
2433 ** leave the autoVacuum mode at 0 (do not auto-vacuum), even if
2434 ** SQLITE_DEFAULT_AUTOVACUUM is true. On the other hand, if
2435 ** SQLITE_OMIT_MEMORYDB has been defined, then ":memory:" is just a
2436 ** regular file-name. In this case the auto-vacuum applies as per normal.
2438 if( zFilename
&& !isMemdb
){
2439 pBt
->autoVacuum
= (SQLITE_DEFAULT_AUTOVACUUM
? 1 : 0);
2440 pBt
->incrVacuum
= (SQLITE_DEFAULT_AUTOVACUUM
==2 ? 1 : 0);
2445 /* EVIDENCE-OF: R-37497-42412 The size of the reserved region is
2446 ** determined by the one-byte unsigned integer found at an offset of 20
2447 ** into the database file header. */
2448 nReserve
= zDbHeader
[20];
2449 pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
2450 #ifndef SQLITE_OMIT_AUTOVACUUM
2451 pBt
->autoVacuum
= (get4byte(&zDbHeader
[36 + 4*4])?1:0);
2452 pBt
->incrVacuum
= (get4byte(&zDbHeader
[36 + 7*4])?1:0);
2455 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
, nReserve
);
2456 if( rc
) goto btree_open_out
;
2457 pBt
->usableSize
= pBt
->pageSize
- nReserve
;
2458 assert( (pBt
->pageSize
& 7)==0 ); /* 8-byte alignment of pageSize */
2460 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2461 /* Add the new BtShared object to the linked list sharable BtShareds.
2465 MUTEX_LOGIC( sqlite3_mutex
*mutexShared
; )
2466 MUTEX_LOGIC( mutexShared
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER
);)
2467 if( SQLITE_THREADSAFE
&& sqlite3GlobalConfig
.bCoreMutex
){
2468 pBt
->mutex
= sqlite3MutexAlloc(SQLITE_MUTEX_FAST
);
2469 if( pBt
->mutex
==0 ){
2470 rc
= SQLITE_NOMEM_BKPT
;
2471 goto btree_open_out
;
2474 sqlite3_mutex_enter(mutexShared
);
2475 pBt
->pNext
= GLOBAL(BtShared
*,sqlite3SharedCacheList
);
2476 GLOBAL(BtShared
*,sqlite3SharedCacheList
) = pBt
;
2477 sqlite3_mutex_leave(mutexShared
);
2482 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2483 /* If the new Btree uses a sharable pBtShared, then link the new
2484 ** Btree into the list of all sharable Btrees for the same connection.
2485 ** The list is kept in ascending order by pBt address.
2490 for(i
=0; i
<db
->nDb
; i
++){
2491 if( (pSib
= db
->aDb
[i
].pBt
)!=0 && pSib
->sharable
){
2492 while( pSib
->pPrev
){ pSib
= pSib
->pPrev
; }
2493 if( (uptr
)p
->pBt
<(uptr
)pSib
->pBt
){
2498 while( pSib
->pNext
&& (uptr
)pSib
->pNext
->pBt
<(uptr
)p
->pBt
){
2501 p
->pNext
= pSib
->pNext
;
2504 p
->pNext
->pPrev
= p
;
2516 if( rc
!=SQLITE_OK
){
2517 if( pBt
&& pBt
->pPager
){
2518 sqlite3PagerClose(pBt
->pPager
, 0);
2524 sqlite3_file
*pFile
;
2526 /* If the B-Tree was successfully opened, set the pager-cache size to the
2527 ** default value. Except, when opening on an existing shared pager-cache,
2528 ** do not change the pager-cache size.
2530 if( sqlite3BtreeSchema(p
, 0, 0)==0 ){
2531 sqlite3PagerSetCachesize(p
->pBt
->pPager
, SQLITE_DEFAULT_CACHE_SIZE
);
2534 pFile
= sqlite3PagerFile(pBt
->pPager
);
2535 if( pFile
->pMethods
){
2536 sqlite3OsFileControlHint(pFile
, SQLITE_FCNTL_PDB
, (void*)&pBt
->db
);
2540 assert( sqlite3_mutex_held(mutexOpen
) );
2541 sqlite3_mutex_leave(mutexOpen
);
2543 assert( rc
!=SQLITE_OK
|| sqlite3BtreeConnectionCount(*ppBtree
)>0 );
2548 ** Decrement the BtShared.nRef counter. When it reaches zero,
2549 ** remove the BtShared structure from the sharing list. Return
2550 ** true if the BtShared.nRef counter reaches zero and return
2551 ** false if it is still positive.
2553 static int removeFromSharingList(BtShared
*pBt
){
2554 #ifndef SQLITE_OMIT_SHARED_CACHE
2555 MUTEX_LOGIC( sqlite3_mutex
*pMaster
; )
2559 assert( sqlite3_mutex_notheld(pBt
->mutex
) );
2560 MUTEX_LOGIC( pMaster
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER
); )
2561 sqlite3_mutex_enter(pMaster
);
2564 if( GLOBAL(BtShared
*,sqlite3SharedCacheList
)==pBt
){
2565 GLOBAL(BtShared
*,sqlite3SharedCacheList
) = pBt
->pNext
;
2567 pList
= GLOBAL(BtShared
*,sqlite3SharedCacheList
);
2568 while( ALWAYS(pList
) && pList
->pNext
!=pBt
){
2571 if( ALWAYS(pList
) ){
2572 pList
->pNext
= pBt
->pNext
;
2575 if( SQLITE_THREADSAFE
){
2576 sqlite3_mutex_free(pBt
->mutex
);
2580 sqlite3_mutex_leave(pMaster
);
2588 ** Make sure pBt->pTmpSpace points to an allocation of
2589 ** MX_CELL_SIZE(pBt) bytes with a 4-byte prefix for a left-child
2592 static void allocateTempSpace(BtShared
*pBt
){
2593 if( !pBt
->pTmpSpace
){
2594 pBt
->pTmpSpace
= sqlite3PageMalloc( pBt
->pageSize
);
2596 /* One of the uses of pBt->pTmpSpace is to format cells before
2597 ** inserting them into a leaf page (function fillInCell()). If
2598 ** a cell is less than 4 bytes in size, it is rounded up to 4 bytes
2599 ** by the various routines that manipulate binary cells. Which
2600 ** can mean that fillInCell() only initializes the first 2 or 3
2601 ** bytes of pTmpSpace, but that the first 4 bytes are copied from
2602 ** it into a database page. This is not actually a problem, but it
2603 ** does cause a valgrind error when the 1 or 2 bytes of unitialized
2604 ** data is passed to system call write(). So to avoid this error,
2605 ** zero the first 4 bytes of temp space here.
2607 ** Also: Provide four bytes of initialized space before the
2608 ** beginning of pTmpSpace as an area available to prepend the
2609 ** left-child pointer to the beginning of a cell.
2611 if( pBt
->pTmpSpace
){
2612 memset(pBt
->pTmpSpace
, 0, 8);
2613 pBt
->pTmpSpace
+= 4;
2619 ** Free the pBt->pTmpSpace allocation
2621 static void freeTempSpace(BtShared
*pBt
){
2622 if( pBt
->pTmpSpace
){
2623 pBt
->pTmpSpace
-= 4;
2624 sqlite3PageFree(pBt
->pTmpSpace
);
2630 ** Close an open database and invalidate all cursors.
2632 int sqlite3BtreeClose(Btree
*p
){
2633 BtShared
*pBt
= p
->pBt
;
2636 /* Close all cursors opened via this handle. */
2637 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2638 sqlite3BtreeEnter(p
);
2639 pCur
= pBt
->pCursor
;
2641 BtCursor
*pTmp
= pCur
;
2643 if( pTmp
->pBtree
==p
){
2644 sqlite3BtreeCloseCursor(pTmp
);
2648 /* Rollback any active transaction and free the handle structure.
2649 ** The call to sqlite3BtreeRollback() drops any table-locks held by
2652 sqlite3BtreeRollback(p
, SQLITE_OK
, 0);
2653 sqlite3BtreeLeave(p
);
2655 /* If there are still other outstanding references to the shared-btree
2656 ** structure, return now. The remainder of this procedure cleans
2657 ** up the shared-btree.
2659 assert( p
->wantToLock
==0 && p
->locked
==0 );
2660 if( !p
->sharable
|| removeFromSharingList(pBt
) ){
2661 /* The pBt is no longer on the sharing list, so we can access
2662 ** it without having to hold the mutex.
2664 ** Clean out and delete the BtShared object.
2666 assert( !pBt
->pCursor
);
2667 sqlite3PagerClose(pBt
->pPager
, p
->db
);
2668 if( pBt
->xFreeSchema
&& pBt
->pSchema
){
2669 pBt
->xFreeSchema(pBt
->pSchema
);
2671 sqlite3DbFree(0, pBt
->pSchema
);
2676 #ifndef SQLITE_OMIT_SHARED_CACHE
2677 assert( p
->wantToLock
==0 );
2678 assert( p
->locked
==0 );
2679 if( p
->pPrev
) p
->pPrev
->pNext
= p
->pNext
;
2680 if( p
->pNext
) p
->pNext
->pPrev
= p
->pPrev
;
2688 ** Change the "soft" limit on the number of pages in the cache.
2689 ** Unused and unmodified pages will be recycled when the number of
2690 ** pages in the cache exceeds this soft limit. But the size of the
2691 ** cache is allowed to grow larger than this limit if it contains
2692 ** dirty pages or pages still in active use.
2694 int sqlite3BtreeSetCacheSize(Btree
*p
, int mxPage
){
2695 BtShared
*pBt
= p
->pBt
;
2696 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2697 sqlite3BtreeEnter(p
);
2698 sqlite3PagerSetCachesize(pBt
->pPager
, mxPage
);
2699 sqlite3BtreeLeave(p
);
2704 ** Change the "spill" limit on the number of pages in the cache.
2705 ** If the number of pages exceeds this limit during a write transaction,
2706 ** the pager might attempt to "spill" pages to the journal early in
2707 ** order to free up memory.
2709 ** The value returned is the current spill size. If zero is passed
2710 ** as an argument, no changes are made to the spill size setting, so
2711 ** using mxPage of 0 is a way to query the current spill size.
2713 int sqlite3BtreeSetSpillSize(Btree
*p
, int mxPage
){
2714 BtShared
*pBt
= p
->pBt
;
2716 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2717 sqlite3BtreeEnter(p
);
2718 res
= sqlite3PagerSetSpillsize(pBt
->pPager
, mxPage
);
2719 sqlite3BtreeLeave(p
);
2723 #if SQLITE_MAX_MMAP_SIZE>0
2725 ** Change the limit on the amount of the database file that may be
2728 int sqlite3BtreeSetMmapLimit(Btree
*p
, sqlite3_int64 szMmap
){
2729 BtShared
*pBt
= p
->pBt
;
2730 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2731 sqlite3BtreeEnter(p
);
2732 sqlite3PagerSetMmapLimit(pBt
->pPager
, szMmap
);
2733 sqlite3BtreeLeave(p
);
2736 #endif /* SQLITE_MAX_MMAP_SIZE>0 */
2739 ** Change the way data is synced to disk in order to increase or decrease
2740 ** how well the database resists damage due to OS crashes and power
2741 ** failures. Level 1 is the same as asynchronous (no syncs() occur and
2742 ** there is a high probability of damage) Level 2 is the default. There
2743 ** is a very low but non-zero probability of damage. Level 3 reduces the
2744 ** probability of damage to near zero but with a write performance reduction.
2746 #ifndef SQLITE_OMIT_PAGER_PRAGMAS
2747 int sqlite3BtreeSetPagerFlags(
2748 Btree
*p
, /* The btree to set the safety level on */
2749 unsigned pgFlags
/* Various PAGER_* flags */
2751 BtShared
*pBt
= p
->pBt
;
2752 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2753 sqlite3BtreeEnter(p
);
2754 sqlite3PagerSetFlags(pBt
->pPager
, pgFlags
);
2755 sqlite3BtreeLeave(p
);
2761 ** Change the default pages size and the number of reserved bytes per page.
2762 ** Or, if the page size has already been fixed, return SQLITE_READONLY
2763 ** without changing anything.
2765 ** The page size must be a power of 2 between 512 and 65536. If the page
2766 ** size supplied does not meet this constraint then the page size is not
2769 ** Page sizes are constrained to be a power of two so that the region
2770 ** of the database file used for locking (beginning at PENDING_BYTE,
2771 ** the first byte past the 1GB boundary, 0x40000000) needs to occur
2772 ** at the beginning of a page.
2774 ** If parameter nReserve is less than zero, then the number of reserved
2775 ** bytes per page is left unchanged.
2777 ** If the iFix!=0 then the BTS_PAGESIZE_FIXED flag is set so that the page size
2778 ** and autovacuum mode can no longer be changed.
2780 int sqlite3BtreeSetPageSize(Btree
*p
, int pageSize
, int nReserve
, int iFix
){
2782 BtShared
*pBt
= p
->pBt
;
2783 assert( nReserve
>=-1 && nReserve
<=255 );
2784 sqlite3BtreeEnter(p
);
2785 #if SQLITE_HAS_CODEC
2786 if( nReserve
>pBt
->optimalReserve
) pBt
->optimalReserve
= (u8
)nReserve
;
2788 if( pBt
->btsFlags
& BTS_PAGESIZE_FIXED
){
2789 sqlite3BtreeLeave(p
);
2790 return SQLITE_READONLY
;
2793 nReserve
= pBt
->pageSize
- pBt
->usableSize
;
2795 assert( nReserve
>=0 && nReserve
<=255 );
2796 if( pageSize
>=512 && pageSize
<=SQLITE_MAX_PAGE_SIZE
&&
2797 ((pageSize
-1)&pageSize
)==0 ){
2798 assert( (pageSize
& 7)==0 );
2799 assert( !pBt
->pCursor
);
2800 pBt
->pageSize
= (u32
)pageSize
;
2803 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
, nReserve
);
2804 pBt
->usableSize
= pBt
->pageSize
- (u16
)nReserve
;
2805 if( iFix
) pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
2806 sqlite3BtreeLeave(p
);
2811 ** Return the currently defined page size
2813 int sqlite3BtreeGetPageSize(Btree
*p
){
2814 return p
->pBt
->pageSize
;
2818 ** This function is similar to sqlite3BtreeGetReserve(), except that it
2819 ** may only be called if it is guaranteed that the b-tree mutex is already
2822 ** This is useful in one special case in the backup API code where it is
2823 ** known that the shared b-tree mutex is held, but the mutex on the
2824 ** database handle that owns *p is not. In this case if sqlite3BtreeEnter()
2825 ** were to be called, it might collide with some other operation on the
2826 ** database handle that owns *p, causing undefined behavior.
2828 int sqlite3BtreeGetReserveNoMutex(Btree
*p
){
2830 assert( sqlite3_mutex_held(p
->pBt
->mutex
) );
2831 n
= p
->pBt
->pageSize
- p
->pBt
->usableSize
;
2836 ** Return the number of bytes of space at the end of every page that
2837 ** are intentually left unused. This is the "reserved" space that is
2838 ** sometimes used by extensions.
2840 ** If SQLITE_HAS_MUTEX is defined then the number returned is the
2841 ** greater of the current reserved space and the maximum requested
2844 int sqlite3BtreeGetOptimalReserve(Btree
*p
){
2846 sqlite3BtreeEnter(p
);
2847 n
= sqlite3BtreeGetReserveNoMutex(p
);
2848 #ifdef SQLITE_HAS_CODEC
2849 if( n
<p
->pBt
->optimalReserve
) n
= p
->pBt
->optimalReserve
;
2851 sqlite3BtreeLeave(p
);
2857 ** Set the maximum page count for a database if mxPage is positive.
2858 ** No changes are made if mxPage is 0 or negative.
2859 ** Regardless of the value of mxPage, return the maximum page count.
2861 int sqlite3BtreeMaxPageCount(Btree
*p
, int mxPage
){
2863 sqlite3BtreeEnter(p
);
2864 n
= sqlite3PagerMaxPageCount(p
->pBt
->pPager
, mxPage
);
2865 sqlite3BtreeLeave(p
);
2870 ** Change the values for the BTS_SECURE_DELETE and BTS_OVERWRITE flags:
2872 ** newFlag==0 Both BTS_SECURE_DELETE and BTS_OVERWRITE are cleared
2873 ** newFlag==1 BTS_SECURE_DELETE set and BTS_OVERWRITE is cleared
2874 ** newFlag==2 BTS_SECURE_DELETE cleared and BTS_OVERWRITE is set
2875 ** newFlag==(-1) No changes
2877 ** This routine acts as a query if newFlag is less than zero
2879 ** With BTS_OVERWRITE set, deleted content is overwritten by zeros, but
2880 ** freelist leaf pages are not written back to the database. Thus in-page
2881 ** deleted content is cleared, but freelist deleted content is not.
2883 ** With BTS_SECURE_DELETE, operation is like BTS_OVERWRITE with the addition
2884 ** that freelist leaf pages are written back into the database, increasing
2885 ** the amount of disk I/O.
2887 int sqlite3BtreeSecureDelete(Btree
*p
, int newFlag
){
2889 if( p
==0 ) return 0;
2890 sqlite3BtreeEnter(p
);
2891 assert( BTS_OVERWRITE
==BTS_SECURE_DELETE
*2 );
2892 assert( BTS_FAST_SECURE
==(BTS_OVERWRITE
|BTS_SECURE_DELETE
) );
2894 p
->pBt
->btsFlags
&= ~BTS_FAST_SECURE
;
2895 p
->pBt
->btsFlags
|= BTS_SECURE_DELETE
*newFlag
;
2897 b
= (p
->pBt
->btsFlags
& BTS_FAST_SECURE
)/BTS_SECURE_DELETE
;
2898 sqlite3BtreeLeave(p
);
2903 ** Change the 'auto-vacuum' property of the database. If the 'autoVacuum'
2904 ** parameter is non-zero, then auto-vacuum mode is enabled. If zero, it
2905 ** is disabled. The default value for the auto-vacuum property is
2906 ** determined by the SQLITE_DEFAULT_AUTOVACUUM macro.
2908 int sqlite3BtreeSetAutoVacuum(Btree
*p
, int autoVacuum
){
2909 #ifdef SQLITE_OMIT_AUTOVACUUM
2910 return SQLITE_READONLY
;
2912 BtShared
*pBt
= p
->pBt
;
2914 u8 av
= (u8
)autoVacuum
;
2916 sqlite3BtreeEnter(p
);
2917 if( (pBt
->btsFlags
& BTS_PAGESIZE_FIXED
)!=0 && (av
?1:0)!=pBt
->autoVacuum
){
2918 rc
= SQLITE_READONLY
;
2920 pBt
->autoVacuum
= av
?1:0;
2921 pBt
->incrVacuum
= av
==2 ?1:0;
2923 sqlite3BtreeLeave(p
);
2929 ** Return the value of the 'auto-vacuum' property. If auto-vacuum is
2930 ** enabled 1 is returned. Otherwise 0.
2932 int sqlite3BtreeGetAutoVacuum(Btree
*p
){
2933 #ifdef SQLITE_OMIT_AUTOVACUUM
2934 return BTREE_AUTOVACUUM_NONE
;
2937 sqlite3BtreeEnter(p
);
2939 (!p
->pBt
->autoVacuum
)?BTREE_AUTOVACUUM_NONE
:
2940 (!p
->pBt
->incrVacuum
)?BTREE_AUTOVACUUM_FULL
:
2941 BTREE_AUTOVACUUM_INCR
2943 sqlite3BtreeLeave(p
);
2949 ** If the user has not set the safety-level for this database connection
2950 ** using "PRAGMA synchronous", and if the safety-level is not already
2951 ** set to the value passed to this function as the second parameter,
2954 #if SQLITE_DEFAULT_SYNCHRONOUS!=SQLITE_DEFAULT_WAL_SYNCHRONOUS \
2955 && !defined(SQLITE_OMIT_WAL)
2956 static void setDefaultSyncFlag(BtShared
*pBt
, u8 safety_level
){
2959 if( (db
=pBt
->db
)!=0 && (pDb
=db
->aDb
)!=0 ){
2960 while( pDb
->pBt
==0 || pDb
->pBt
->pBt
!=pBt
){ pDb
++; }
2961 if( pDb
->bSyncSet
==0
2962 && pDb
->safety_level
!=safety_level
2965 pDb
->safety_level
= safety_level
;
2966 sqlite3PagerSetFlags(pBt
->pPager
,
2967 pDb
->safety_level
| (db
->flags
& PAGER_FLAGS_MASK
));
2972 # define setDefaultSyncFlag(pBt,safety_level)
2975 /* Forward declaration */
2976 static int newDatabase(BtShared
*);
2980 ** Get a reference to pPage1 of the database file. This will
2981 ** also acquire a readlock on that file.
2983 ** SQLITE_OK is returned on success. If the file is not a
2984 ** well-formed database file, then SQLITE_CORRUPT is returned.
2985 ** SQLITE_BUSY is returned if the database is locked. SQLITE_NOMEM
2986 ** is returned if we run out of memory.
2988 static int lockBtree(BtShared
*pBt
){
2989 int rc
; /* Result code from subfunctions */
2990 MemPage
*pPage1
; /* Page 1 of the database file */
2991 int nPage
; /* Number of pages in the database */
2992 int nPageFile
= 0; /* Number of pages in the database file */
2993 int nPageHeader
; /* Number of pages in the database according to hdr */
2995 assert( sqlite3_mutex_held(pBt
->mutex
) );
2996 assert( pBt
->pPage1
==0 );
2997 rc
= sqlite3PagerSharedLock(pBt
->pPager
);
2998 if( rc
!=SQLITE_OK
) return rc
;
2999 rc
= btreeGetPage(pBt
, 1, &pPage1
, 0);
3000 if( rc
!=SQLITE_OK
) return rc
;
3002 /* Do some checking to help insure the file we opened really is
3003 ** a valid database file.
3005 nPage
= nPageHeader
= get4byte(28+(u8
*)pPage1
->aData
);
3006 sqlite3PagerPagecount(pBt
->pPager
, &nPageFile
);
3007 if( nPage
==0 || memcmp(24+(u8
*)pPage1
->aData
, 92+(u8
*)pPage1
->aData
,4)!=0 ){
3010 if( (pBt
->db
->flags
& SQLITE_ResetDatabase
)!=0 ){
3016 u8
*page1
= pPage1
->aData
;
3018 /* EVIDENCE-OF: R-43737-39999 Every valid SQLite database file begins
3019 ** with the following 16 bytes (in hex): 53 51 4c 69 74 65 20 66 6f 72 6d
3020 ** 61 74 20 33 00. */
3021 if( memcmp(page1
, zMagicHeader
, 16)!=0 ){
3022 goto page1_init_failed
;
3025 #ifdef SQLITE_OMIT_WAL
3027 pBt
->btsFlags
|= BTS_READ_ONLY
;
3030 goto page1_init_failed
;
3034 pBt
->btsFlags
|= BTS_READ_ONLY
;
3037 goto page1_init_failed
;
3040 /* If the write version is set to 2, this database should be accessed
3041 ** in WAL mode. If the log is not already open, open it now. Then
3042 ** return SQLITE_OK and return without populating BtShared.pPage1.
3043 ** The caller detects this and calls this function again. This is
3044 ** required as the version of page 1 currently in the page1 buffer
3045 ** may not be the latest version - there may be a newer one in the log
3048 if( page1
[19]==2 && (pBt
->btsFlags
& BTS_NO_WAL
)==0 ){
3050 rc
= sqlite3PagerOpenWal(pBt
->pPager
, &isOpen
);
3051 if( rc
!=SQLITE_OK
){
3052 goto page1_init_failed
;
3054 setDefaultSyncFlag(pBt
, SQLITE_DEFAULT_WAL_SYNCHRONOUS
+1);
3056 releasePageOne(pPage1
);
3062 setDefaultSyncFlag(pBt
, SQLITE_DEFAULT_SYNCHRONOUS
+1);
3066 /* EVIDENCE-OF: R-15465-20813 The maximum and minimum embedded payload
3067 ** fractions and the leaf payload fraction values must be 64, 32, and 32.
3069 ** The original design allowed these amounts to vary, but as of
3070 ** version 3.6.0, we require them to be fixed.
3072 if( memcmp(&page1
[21], "\100\040\040",3)!=0 ){
3073 goto page1_init_failed
;
3075 /* EVIDENCE-OF: R-51873-39618 The page size for a database file is
3076 ** determined by the 2-byte integer located at an offset of 16 bytes from
3077 ** the beginning of the database file. */
3078 pageSize
= (page1
[16]<<8) | (page1
[17]<<16);
3079 /* EVIDENCE-OF: R-25008-21688 The size of a page is a power of two
3080 ** between 512 and 65536 inclusive. */
3081 if( ((pageSize
-1)&pageSize
)!=0
3082 || pageSize
>SQLITE_MAX_PAGE_SIZE
3085 goto page1_init_failed
;
3087 assert( (pageSize
& 7)==0 );
3088 /* EVIDENCE-OF: R-59310-51205 The "reserved space" size in the 1-byte
3089 ** integer at offset 20 is the number of bytes of space at the end of
3090 ** each page to reserve for extensions.
3092 ** EVIDENCE-OF: R-37497-42412 The size of the reserved region is
3093 ** determined by the one-byte unsigned integer found at an offset of 20
3094 ** into the database file header. */
3095 usableSize
= pageSize
- page1
[20];
3096 if( (u32
)pageSize
!=pBt
->pageSize
){
3097 /* After reading the first page of the database assuming a page size
3098 ** of BtShared.pageSize, we have discovered that the page-size is
3099 ** actually pageSize. Unlock the database, leave pBt->pPage1 at
3100 ** zero and return SQLITE_OK. The caller will call this function
3101 ** again with the correct page-size.
3103 releasePageOne(pPage1
);
3104 pBt
->usableSize
= usableSize
;
3105 pBt
->pageSize
= pageSize
;
3107 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
,
3108 pageSize
-usableSize
);
3111 if( (pBt
->db
->flags
& SQLITE_WriteSchema
)==0 && nPage
>nPageFile
){
3112 rc
= SQLITE_CORRUPT_BKPT
;
3113 goto page1_init_failed
;
3115 /* EVIDENCE-OF: R-28312-64704 However, the usable size is not allowed to
3116 ** be less than 480. In other words, if the page size is 512, then the
3117 ** reserved space size cannot exceed 32. */
3118 if( usableSize
<480 ){
3119 goto page1_init_failed
;
3121 pBt
->pageSize
= pageSize
;
3122 pBt
->usableSize
= usableSize
;
3123 #ifndef SQLITE_OMIT_AUTOVACUUM
3124 pBt
->autoVacuum
= (get4byte(&page1
[36 + 4*4])?1:0);
3125 pBt
->incrVacuum
= (get4byte(&page1
[36 + 7*4])?1:0);
3129 /* maxLocal is the maximum amount of payload to store locally for
3130 ** a cell. Make sure it is small enough so that at least minFanout
3131 ** cells can will fit on one page. We assume a 10-byte page header.
3132 ** Besides the payload, the cell must store:
3133 ** 2-byte pointer to the cell
3134 ** 4-byte child pointer
3135 ** 9-byte nKey value
3136 ** 4-byte nData value
3137 ** 4-byte overflow page pointer
3138 ** So a cell consists of a 2-byte pointer, a header which is as much as
3139 ** 17 bytes long, 0 to N bytes of payload, and an optional 4 byte overflow
3142 pBt
->maxLocal
= (u16
)((pBt
->usableSize
-12)*64/255 - 23);
3143 pBt
->minLocal
= (u16
)((pBt
->usableSize
-12)*32/255 - 23);
3144 pBt
->maxLeaf
= (u16
)(pBt
->usableSize
- 35);
3145 pBt
->minLeaf
= (u16
)((pBt
->usableSize
-12)*32/255 - 23);
3146 if( pBt
->maxLocal
>127 ){
3147 pBt
->max1bytePayload
= 127;
3149 pBt
->max1bytePayload
= (u8
)pBt
->maxLocal
;
3151 assert( pBt
->maxLeaf
+ 23 <= MX_CELL_SIZE(pBt
) );
3152 pBt
->pPage1
= pPage1
;
3157 releasePageOne(pPage1
);
3164 ** Return the number of cursors open on pBt. This is for use
3165 ** in assert() expressions, so it is only compiled if NDEBUG is not
3168 ** Only write cursors are counted if wrOnly is true. If wrOnly is
3169 ** false then all cursors are counted.
3171 ** For the purposes of this routine, a cursor is any cursor that
3172 ** is capable of reading or writing to the database. Cursors that
3173 ** have been tripped into the CURSOR_FAULT state are not counted.
3175 static int countValidCursors(BtShared
*pBt
, int wrOnly
){
3178 for(pCur
=pBt
->pCursor
; pCur
; pCur
=pCur
->pNext
){
3179 if( (wrOnly
==0 || (pCur
->curFlags
& BTCF_WriteFlag
)!=0)
3180 && pCur
->eState
!=CURSOR_FAULT
) r
++;
3187 ** If there are no outstanding cursors and we are not in the middle
3188 ** of a transaction but there is a read lock on the database, then
3189 ** this routine unrefs the first page of the database file which
3190 ** has the effect of releasing the read lock.
3192 ** If there is a transaction in progress, this routine is a no-op.
3194 static void unlockBtreeIfUnused(BtShared
*pBt
){
3195 assert( sqlite3_mutex_held(pBt
->mutex
) );
3196 assert( countValidCursors(pBt
,0)==0 || pBt
->inTransaction
>TRANS_NONE
);
3197 if( pBt
->inTransaction
==TRANS_NONE
&& pBt
->pPage1
!=0 ){
3198 MemPage
*pPage1
= pBt
->pPage1
;
3199 assert( pPage1
->aData
);
3200 assert( sqlite3PagerRefcount(pBt
->pPager
)==1 );
3202 releasePageOne(pPage1
);
3207 ** If pBt points to an empty file then convert that empty file
3208 ** into a new empty database by initializing the first page of
3211 static int newDatabase(BtShared
*pBt
){
3213 unsigned char *data
;
3216 assert( sqlite3_mutex_held(pBt
->mutex
) );
3223 rc
= sqlite3PagerWrite(pP1
->pDbPage
);
3225 memcpy(data
, zMagicHeader
, sizeof(zMagicHeader
));
3226 assert( sizeof(zMagicHeader
)==16 );
3227 data
[16] = (u8
)((pBt
->pageSize
>>8)&0xff);
3228 data
[17] = (u8
)((pBt
->pageSize
>>16)&0xff);
3231 assert( pBt
->usableSize
<=pBt
->pageSize
&& pBt
->usableSize
+255>=pBt
->pageSize
);
3232 data
[20] = (u8
)(pBt
->pageSize
- pBt
->usableSize
);
3236 memset(&data
[24], 0, 100-24);
3237 zeroPage(pP1
, PTF_INTKEY
|PTF_LEAF
|PTF_LEAFDATA
);
3238 pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
3239 #ifndef SQLITE_OMIT_AUTOVACUUM
3240 assert( pBt
->autoVacuum
==1 || pBt
->autoVacuum
==0 );
3241 assert( pBt
->incrVacuum
==1 || pBt
->incrVacuum
==0 );
3242 put4byte(&data
[36 + 4*4], pBt
->autoVacuum
);
3243 put4byte(&data
[36 + 7*4], pBt
->incrVacuum
);
3251 ** Initialize the first page of the database file (creating a database
3252 ** consisting of a single page and no schema objects). Return SQLITE_OK
3253 ** if successful, or an SQLite error code otherwise.
3255 int sqlite3BtreeNewDb(Btree
*p
){
3257 sqlite3BtreeEnter(p
);
3259 rc
= newDatabase(p
->pBt
);
3260 sqlite3BtreeLeave(p
);
3265 ** Attempt to start a new transaction. A write-transaction
3266 ** is started if the second argument is nonzero, otherwise a read-
3267 ** transaction. If the second argument is 2 or more and exclusive
3268 ** transaction is started, meaning that no other process is allowed
3269 ** to access the database. A preexisting transaction may not be
3270 ** upgraded to exclusive by calling this routine a second time - the
3271 ** exclusivity flag only works for a new transaction.
3273 ** A write-transaction must be started before attempting any
3274 ** changes to the database. None of the following routines
3275 ** will work unless a transaction is started first:
3277 ** sqlite3BtreeCreateTable()
3278 ** sqlite3BtreeCreateIndex()
3279 ** sqlite3BtreeClearTable()
3280 ** sqlite3BtreeDropTable()
3281 ** sqlite3BtreeInsert()
3282 ** sqlite3BtreeDelete()
3283 ** sqlite3BtreeUpdateMeta()
3285 ** If an initial attempt to acquire the lock fails because of lock contention
3286 ** and the database was previously unlocked, then invoke the busy handler
3287 ** if there is one. But if there was previously a read-lock, do not
3288 ** invoke the busy handler - just return SQLITE_BUSY. SQLITE_BUSY is
3289 ** returned when there is already a read-lock in order to avoid a deadlock.
3291 ** Suppose there are two processes A and B. A has a read lock and B has
3292 ** a reserved lock. B tries to promote to exclusive but is blocked because
3293 ** of A's read lock. A tries to promote to reserved but is blocked by B.
3294 ** One or the other of the two processes must give way or there can be
3295 ** no progress. By returning SQLITE_BUSY and not invoking the busy callback
3296 ** when A already has a read lock, we encourage A to give up and let B
3299 int sqlite3BtreeBeginTrans(Btree
*p
, int wrflag
){
3300 BtShared
*pBt
= p
->pBt
;
3303 sqlite3BtreeEnter(p
);
3306 /* If the btree is already in a write-transaction, or it
3307 ** is already in a read-transaction and a read-transaction
3308 ** is requested, this is a no-op.
3310 if( p
->inTrans
==TRANS_WRITE
|| (p
->inTrans
==TRANS_READ
&& !wrflag
) ){
3313 assert( pBt
->inTransaction
==TRANS_WRITE
|| IfNotOmitAV(pBt
->bDoTruncate
)==0 );
3315 /* Write transactions are not possible on a read-only database */
3316 if( (pBt
->btsFlags
& BTS_READ_ONLY
)!=0 && wrflag
){
3317 rc
= SQLITE_READONLY
;
3321 #ifndef SQLITE_OMIT_SHARED_CACHE
3323 sqlite3
*pBlock
= 0;
3324 /* If another database handle has already opened a write transaction
3325 ** on this shared-btree structure and a second write transaction is
3326 ** requested, return SQLITE_LOCKED.
3328 if( (wrflag
&& pBt
->inTransaction
==TRANS_WRITE
)
3329 || (pBt
->btsFlags
& BTS_PENDING
)!=0
3331 pBlock
= pBt
->pWriter
->db
;
3332 }else if( wrflag
>1 ){
3334 for(pIter
=pBt
->pLock
; pIter
; pIter
=pIter
->pNext
){
3335 if( pIter
->pBtree
!=p
){
3336 pBlock
= pIter
->pBtree
->db
;
3342 sqlite3ConnectionBlocked(p
->db
, pBlock
);
3343 rc
= SQLITE_LOCKED_SHAREDCACHE
;
3349 /* Any read-only or read-write transaction implies a read-lock on
3350 ** page 1. So if some other shared-cache client already has a write-lock
3351 ** on page 1, the transaction cannot be opened. */
3352 rc
= querySharedCacheTableLock(p
, MASTER_ROOT
, READ_LOCK
);
3353 if( SQLITE_OK
!=rc
) goto trans_begun
;
3355 pBt
->btsFlags
&= ~BTS_INITIALLY_EMPTY
;
3356 if( pBt
->nPage
==0 ) pBt
->btsFlags
|= BTS_INITIALLY_EMPTY
;
3358 /* Call lockBtree() until either pBt->pPage1 is populated or
3359 ** lockBtree() returns something other than SQLITE_OK. lockBtree()
3360 ** may return SQLITE_OK but leave pBt->pPage1 set to 0 if after
3361 ** reading page 1 it discovers that the page-size of the database
3362 ** file is not pBt->pageSize. In this case lockBtree() will update
3363 ** pBt->pageSize to the page-size of the file on disk.
3365 while( pBt
->pPage1
==0 && SQLITE_OK
==(rc
= lockBtree(pBt
)) );
3367 if( rc
==SQLITE_OK
&& wrflag
){
3368 if( (pBt
->btsFlags
& BTS_READ_ONLY
)!=0 ){
3369 rc
= SQLITE_READONLY
;
3371 rc
= sqlite3PagerBegin(pBt
->pPager
,wrflag
>1,sqlite3TempInMemory(p
->db
));
3372 if( rc
==SQLITE_OK
){
3373 rc
= newDatabase(pBt
);
3378 if( rc
!=SQLITE_OK
){
3379 unlockBtreeIfUnused(pBt
);
3381 }while( (rc
&0xFF)==SQLITE_BUSY
&& pBt
->inTransaction
==TRANS_NONE
&&
3382 btreeInvokeBusyHandler(pBt
) );
3383 sqlite3PagerResetLockTimeout(pBt
->pPager
);
3385 if( rc
==SQLITE_OK
){
3386 if( p
->inTrans
==TRANS_NONE
){
3387 pBt
->nTransaction
++;
3388 #ifndef SQLITE_OMIT_SHARED_CACHE
3390 assert( p
->lock
.pBtree
==p
&& p
->lock
.iTable
==1 );
3391 p
->lock
.eLock
= READ_LOCK
;
3392 p
->lock
.pNext
= pBt
->pLock
;
3393 pBt
->pLock
= &p
->lock
;
3397 p
->inTrans
= (wrflag
?TRANS_WRITE
:TRANS_READ
);
3398 if( p
->inTrans
>pBt
->inTransaction
){
3399 pBt
->inTransaction
= p
->inTrans
;
3402 MemPage
*pPage1
= pBt
->pPage1
;
3403 #ifndef SQLITE_OMIT_SHARED_CACHE
3404 assert( !pBt
->pWriter
);
3406 pBt
->btsFlags
&= ~BTS_EXCLUSIVE
;
3407 if( wrflag
>1 ) pBt
->btsFlags
|= BTS_EXCLUSIVE
;
3410 /* If the db-size header field is incorrect (as it may be if an old
3411 ** client has been writing the database file), update it now. Doing
3412 ** this sooner rather than later means the database size can safely
3413 ** re-read the database size from page 1 if a savepoint or transaction
3414 ** rollback occurs within the transaction.
3416 if( pBt
->nPage
!=get4byte(&pPage1
->aData
[28]) ){
3417 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
3418 if( rc
==SQLITE_OK
){
3419 put4byte(&pPage1
->aData
[28], pBt
->nPage
);
3427 if( rc
==SQLITE_OK
&& wrflag
){
3428 /* This call makes sure that the pager has the correct number of
3429 ** open savepoints. If the second parameter is greater than 0 and
3430 ** the sub-journal is not already open, then it will be opened here.
3432 rc
= sqlite3PagerOpenSavepoint(pBt
->pPager
, p
->db
->nSavepoint
);
3436 sqlite3BtreeLeave(p
);
3440 #ifndef SQLITE_OMIT_AUTOVACUUM
3443 ** Set the pointer-map entries for all children of page pPage. Also, if
3444 ** pPage contains cells that point to overflow pages, set the pointer
3445 ** map entries for the overflow pages as well.
3447 static int setChildPtrmaps(MemPage
*pPage
){
3448 int i
; /* Counter variable */
3449 int nCell
; /* Number of cells in page pPage */
3450 int rc
; /* Return code */
3451 BtShared
*pBt
= pPage
->pBt
;
3452 Pgno pgno
= pPage
->pgno
;
3454 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
3455 rc
= pPage
->isInit
? SQLITE_OK
: btreeInitPage(pPage
);
3456 if( rc
!=SQLITE_OK
) return rc
;
3457 nCell
= pPage
->nCell
;
3459 for(i
=0; i
<nCell
; i
++){
3460 u8
*pCell
= findCell(pPage
, i
);
3462 ptrmapPutOvflPtr(pPage
, pCell
, &rc
);
3465 Pgno childPgno
= get4byte(pCell
);
3466 ptrmapPut(pBt
, childPgno
, PTRMAP_BTREE
, pgno
, &rc
);
3471 Pgno childPgno
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
3472 ptrmapPut(pBt
, childPgno
, PTRMAP_BTREE
, pgno
, &rc
);
3479 ** Somewhere on pPage is a pointer to page iFrom. Modify this pointer so
3480 ** that it points to iTo. Parameter eType describes the type of pointer to
3481 ** be modified, as follows:
3483 ** PTRMAP_BTREE: pPage is a btree-page. The pointer points at a child
3486 ** PTRMAP_OVERFLOW1: pPage is a btree-page. The pointer points at an overflow
3487 ** page pointed to by one of the cells on pPage.
3489 ** PTRMAP_OVERFLOW2: pPage is an overflow-page. The pointer points at the next
3490 ** overflow page in the list.
3492 static int modifyPagePointer(MemPage
*pPage
, Pgno iFrom
, Pgno iTo
, u8 eType
){
3493 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
3494 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
3495 if( eType
==PTRMAP_OVERFLOW2
){
3496 /* The pointer is always the first 4 bytes of the page in this case. */
3497 if( get4byte(pPage
->aData
)!=iFrom
){
3498 return SQLITE_CORRUPT_PAGE(pPage
);
3500 put4byte(pPage
->aData
, iTo
);
3506 rc
= pPage
->isInit
? SQLITE_OK
: btreeInitPage(pPage
);
3508 nCell
= pPage
->nCell
;
3510 for(i
=0; i
<nCell
; i
++){
3511 u8
*pCell
= findCell(pPage
, i
);
3512 if( eType
==PTRMAP_OVERFLOW1
){
3514 pPage
->xParseCell(pPage
, pCell
, &info
);
3515 if( info
.nLocal
<info
.nPayload
){
3516 if( pCell
+info
.nSize
> pPage
->aData
+pPage
->pBt
->usableSize
){
3517 return SQLITE_CORRUPT_PAGE(pPage
);
3519 if( iFrom
==get4byte(pCell
+info
.nSize
-4) ){
3520 put4byte(pCell
+info
.nSize
-4, iTo
);
3525 if( get4byte(pCell
)==iFrom
){
3526 put4byte(pCell
, iTo
);
3533 if( eType
!=PTRMAP_BTREE
||
3534 get4byte(&pPage
->aData
[pPage
->hdrOffset
+8])!=iFrom
){
3535 return SQLITE_CORRUPT_PAGE(pPage
);
3537 put4byte(&pPage
->aData
[pPage
->hdrOffset
+8], iTo
);
3545 ** Move the open database page pDbPage to location iFreePage in the
3546 ** database. The pDbPage reference remains valid.
3548 ** The isCommit flag indicates that there is no need to remember that
3549 ** the journal needs to be sync()ed before database page pDbPage->pgno
3550 ** can be written to. The caller has already promised not to write to that
3553 static int relocatePage(
3554 BtShared
*pBt
, /* Btree */
3555 MemPage
*pDbPage
, /* Open page to move */
3556 u8 eType
, /* Pointer map 'type' entry for pDbPage */
3557 Pgno iPtrPage
, /* Pointer map 'page-no' entry for pDbPage */
3558 Pgno iFreePage
, /* The location to move pDbPage to */
3559 int isCommit
/* isCommit flag passed to sqlite3PagerMovepage */
3561 MemPage
*pPtrPage
; /* The page that contains a pointer to pDbPage */
3562 Pgno iDbPage
= pDbPage
->pgno
;
3563 Pager
*pPager
= pBt
->pPager
;
3566 assert( eType
==PTRMAP_OVERFLOW2
|| eType
==PTRMAP_OVERFLOW1
||
3567 eType
==PTRMAP_BTREE
|| eType
==PTRMAP_ROOTPAGE
);
3568 assert( sqlite3_mutex_held(pBt
->mutex
) );
3569 assert( pDbPage
->pBt
==pBt
);
3571 /* Move page iDbPage from its current location to page number iFreePage */
3572 TRACE(("AUTOVACUUM: Moving %d to free page %d (ptr page %d type %d)\n",
3573 iDbPage
, iFreePage
, iPtrPage
, eType
));
3574 rc
= sqlite3PagerMovepage(pPager
, pDbPage
->pDbPage
, iFreePage
, isCommit
);
3575 if( rc
!=SQLITE_OK
){
3578 pDbPage
->pgno
= iFreePage
;
3580 /* If pDbPage was a btree-page, then it may have child pages and/or cells
3581 ** that point to overflow pages. The pointer map entries for all these
3582 ** pages need to be changed.
3584 ** If pDbPage is an overflow page, then the first 4 bytes may store a
3585 ** pointer to a subsequent overflow page. If this is the case, then
3586 ** the pointer map needs to be updated for the subsequent overflow page.
3588 if( eType
==PTRMAP_BTREE
|| eType
==PTRMAP_ROOTPAGE
){
3589 rc
= setChildPtrmaps(pDbPage
);
3590 if( rc
!=SQLITE_OK
){
3594 Pgno nextOvfl
= get4byte(pDbPage
->aData
);
3596 ptrmapPut(pBt
, nextOvfl
, PTRMAP_OVERFLOW2
, iFreePage
, &rc
);
3597 if( rc
!=SQLITE_OK
){
3603 /* Fix the database pointer on page iPtrPage that pointed at iDbPage so
3604 ** that it points at iFreePage. Also fix the pointer map entry for
3607 if( eType
!=PTRMAP_ROOTPAGE
){
3608 rc
= btreeGetPage(pBt
, iPtrPage
, &pPtrPage
, 0);
3609 if( rc
!=SQLITE_OK
){
3612 rc
= sqlite3PagerWrite(pPtrPage
->pDbPage
);
3613 if( rc
!=SQLITE_OK
){
3614 releasePage(pPtrPage
);
3617 rc
= modifyPagePointer(pPtrPage
, iDbPage
, iFreePage
, eType
);
3618 releasePage(pPtrPage
);
3619 if( rc
==SQLITE_OK
){
3620 ptrmapPut(pBt
, iFreePage
, eType
, iPtrPage
, &rc
);
3626 /* Forward declaration required by incrVacuumStep(). */
3627 static int allocateBtreePage(BtShared
*, MemPage
**, Pgno
*, Pgno
, u8
);
3630 ** Perform a single step of an incremental-vacuum. If successful, return
3631 ** SQLITE_OK. If there is no work to do (and therefore no point in
3632 ** calling this function again), return SQLITE_DONE. Or, if an error
3633 ** occurs, return some other error code.
3635 ** More specifically, this function attempts to re-organize the database so
3636 ** that the last page of the file currently in use is no longer in use.
3638 ** Parameter nFin is the number of pages that this database would contain
3639 ** were this function called until it returns SQLITE_DONE.
3641 ** If the bCommit parameter is non-zero, this function assumes that the
3642 ** caller will keep calling incrVacuumStep() until it returns SQLITE_DONE
3643 ** or an error. bCommit is passed true for an auto-vacuum-on-commit
3644 ** operation, or false for an incremental vacuum.
3646 static int incrVacuumStep(BtShared
*pBt
, Pgno nFin
, Pgno iLastPg
, int bCommit
){
3647 Pgno nFreeList
; /* Number of pages still on the free-list */
3650 assert( sqlite3_mutex_held(pBt
->mutex
) );
3651 assert( iLastPg
>nFin
);
3653 if( !PTRMAP_ISPAGE(pBt
, iLastPg
) && iLastPg
!=PENDING_BYTE_PAGE(pBt
) ){
3657 nFreeList
= get4byte(&pBt
->pPage1
->aData
[36]);
3662 rc
= ptrmapGet(pBt
, iLastPg
, &eType
, &iPtrPage
);
3663 if( rc
!=SQLITE_OK
){
3666 if( eType
==PTRMAP_ROOTPAGE
){
3667 return SQLITE_CORRUPT_BKPT
;
3670 if( eType
==PTRMAP_FREEPAGE
){
3672 /* Remove the page from the files free-list. This is not required
3673 ** if bCommit is non-zero. In that case, the free-list will be
3674 ** truncated to zero after this function returns, so it doesn't
3675 ** matter if it still contains some garbage entries.
3679 rc
= allocateBtreePage(pBt
, &pFreePg
, &iFreePg
, iLastPg
, BTALLOC_EXACT
);
3680 if( rc
!=SQLITE_OK
){
3683 assert( iFreePg
==iLastPg
);
3684 releasePage(pFreePg
);
3687 Pgno iFreePg
; /* Index of free page to move pLastPg to */
3689 u8 eMode
= BTALLOC_ANY
; /* Mode parameter for allocateBtreePage() */
3690 Pgno iNear
= 0; /* nearby parameter for allocateBtreePage() */
3692 rc
= btreeGetPage(pBt
, iLastPg
, &pLastPg
, 0);
3693 if( rc
!=SQLITE_OK
){
3697 /* If bCommit is zero, this loop runs exactly once and page pLastPg
3698 ** is swapped with the first free page pulled off the free list.
3700 ** On the other hand, if bCommit is greater than zero, then keep
3701 ** looping until a free-page located within the first nFin pages
3702 ** of the file is found.
3710 rc
= allocateBtreePage(pBt
, &pFreePg
, &iFreePg
, iNear
, eMode
);
3711 if( rc
!=SQLITE_OK
){
3712 releasePage(pLastPg
);
3715 releasePage(pFreePg
);
3716 }while( bCommit
&& iFreePg
>nFin
);
3717 assert( iFreePg
<iLastPg
);
3719 rc
= relocatePage(pBt
, pLastPg
, eType
, iPtrPage
, iFreePg
, bCommit
);
3720 releasePage(pLastPg
);
3721 if( rc
!=SQLITE_OK
){
3730 }while( iLastPg
==PENDING_BYTE_PAGE(pBt
) || PTRMAP_ISPAGE(pBt
, iLastPg
) );
3731 pBt
->bDoTruncate
= 1;
3732 pBt
->nPage
= iLastPg
;
3738 ** The database opened by the first argument is an auto-vacuum database
3739 ** nOrig pages in size containing nFree free pages. Return the expected
3740 ** size of the database in pages following an auto-vacuum operation.
3742 static Pgno
finalDbSize(BtShared
*pBt
, Pgno nOrig
, Pgno nFree
){
3743 int nEntry
; /* Number of entries on one ptrmap page */
3744 Pgno nPtrmap
; /* Number of PtrMap pages to be freed */
3745 Pgno nFin
; /* Return value */
3747 nEntry
= pBt
->usableSize
/5;
3748 nPtrmap
= (nFree
-nOrig
+PTRMAP_PAGENO(pBt
, nOrig
)+nEntry
)/nEntry
;
3749 nFin
= nOrig
- nFree
- nPtrmap
;
3750 if( nOrig
>PENDING_BYTE_PAGE(pBt
) && nFin
<PENDING_BYTE_PAGE(pBt
) ){
3753 while( PTRMAP_ISPAGE(pBt
, nFin
) || nFin
==PENDING_BYTE_PAGE(pBt
) ){
3761 ** A write-transaction must be opened before calling this function.
3762 ** It performs a single unit of work towards an incremental vacuum.
3764 ** If the incremental vacuum is finished after this function has run,
3765 ** SQLITE_DONE is returned. If it is not finished, but no error occurred,
3766 ** SQLITE_OK is returned. Otherwise an SQLite error code.
3768 int sqlite3BtreeIncrVacuum(Btree
*p
){
3770 BtShared
*pBt
= p
->pBt
;
3772 sqlite3BtreeEnter(p
);
3773 assert( pBt
->inTransaction
==TRANS_WRITE
&& p
->inTrans
==TRANS_WRITE
);
3774 if( !pBt
->autoVacuum
){
3777 Pgno nOrig
= btreePagecount(pBt
);
3778 Pgno nFree
= get4byte(&pBt
->pPage1
->aData
[36]);
3779 Pgno nFin
= finalDbSize(pBt
, nOrig
, nFree
);
3782 rc
= SQLITE_CORRUPT_BKPT
;
3783 }else if( nFree
>0 ){
3784 rc
= saveAllCursors(pBt
, 0, 0);
3785 if( rc
==SQLITE_OK
){
3786 invalidateAllOverflowCache(pBt
);
3787 rc
= incrVacuumStep(pBt
, nFin
, nOrig
, 0);
3789 if( rc
==SQLITE_OK
){
3790 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
3791 put4byte(&pBt
->pPage1
->aData
[28], pBt
->nPage
);
3797 sqlite3BtreeLeave(p
);
3802 ** This routine is called prior to sqlite3PagerCommit when a transaction
3803 ** is committed for an auto-vacuum database.
3805 ** If SQLITE_OK is returned, then *pnTrunc is set to the number of pages
3806 ** the database file should be truncated to during the commit process.
3807 ** i.e. the database has been reorganized so that only the first *pnTrunc
3808 ** pages are in use.
3810 static int autoVacuumCommit(BtShared
*pBt
){
3812 Pager
*pPager
= pBt
->pPager
;
3813 VVA_ONLY( int nRef
= sqlite3PagerRefcount(pPager
); )
3815 assert( sqlite3_mutex_held(pBt
->mutex
) );
3816 invalidateAllOverflowCache(pBt
);
3817 assert(pBt
->autoVacuum
);
3818 if( !pBt
->incrVacuum
){
3819 Pgno nFin
; /* Number of pages in database after autovacuuming */
3820 Pgno nFree
; /* Number of pages on the freelist initially */
3821 Pgno iFree
; /* The next page to be freed */
3822 Pgno nOrig
; /* Database size before freeing */
3824 nOrig
= btreePagecount(pBt
);
3825 if( PTRMAP_ISPAGE(pBt
, nOrig
) || nOrig
==PENDING_BYTE_PAGE(pBt
) ){
3826 /* It is not possible to create a database for which the final page
3827 ** is either a pointer-map page or the pending-byte page. If one
3828 ** is encountered, this indicates corruption.
3830 return SQLITE_CORRUPT_BKPT
;
3833 nFree
= get4byte(&pBt
->pPage1
->aData
[36]);
3834 nFin
= finalDbSize(pBt
, nOrig
, nFree
);
3835 if( nFin
>nOrig
) return SQLITE_CORRUPT_BKPT
;
3837 rc
= saveAllCursors(pBt
, 0, 0);
3839 for(iFree
=nOrig
; iFree
>nFin
&& rc
==SQLITE_OK
; iFree
--){
3840 rc
= incrVacuumStep(pBt
, nFin
, iFree
, 1);
3842 if( (rc
==SQLITE_DONE
|| rc
==SQLITE_OK
) && nFree
>0 ){
3843 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
3844 put4byte(&pBt
->pPage1
->aData
[32], 0);
3845 put4byte(&pBt
->pPage1
->aData
[36], 0);
3846 put4byte(&pBt
->pPage1
->aData
[28], nFin
);
3847 pBt
->bDoTruncate
= 1;
3850 if( rc
!=SQLITE_OK
){
3851 sqlite3PagerRollback(pPager
);
3855 assert( nRef
>=sqlite3PagerRefcount(pPager
) );
3859 #else /* ifndef SQLITE_OMIT_AUTOVACUUM */
3860 # define setChildPtrmaps(x) SQLITE_OK
3864 ** This routine does the first phase of a two-phase commit. This routine
3865 ** causes a rollback journal to be created (if it does not already exist)
3866 ** and populated with enough information so that if a power loss occurs
3867 ** the database can be restored to its original state by playing back
3868 ** the journal. Then the contents of the journal are flushed out to
3869 ** the disk. After the journal is safely on oxide, the changes to the
3870 ** database are written into the database file and flushed to oxide.
3871 ** At the end of this call, the rollback journal still exists on the
3872 ** disk and we are still holding all locks, so the transaction has not
3873 ** committed. See sqlite3BtreeCommitPhaseTwo() for the second phase of the
3876 ** This call is a no-op if no write-transaction is currently active on pBt.
3878 ** Otherwise, sync the database file for the btree pBt. zMaster points to
3879 ** the name of a master journal file that should be written into the
3880 ** individual journal file, or is NULL, indicating no master journal file
3881 ** (single database transaction).
3883 ** When this is called, the master journal should already have been
3884 ** created, populated with this journal pointer and synced to disk.
3886 ** Once this is routine has returned, the only thing required to commit
3887 ** the write-transaction for this database file is to delete the journal.
3889 int sqlite3BtreeCommitPhaseOne(Btree
*p
, const char *zMaster
){
3891 if( p
->inTrans
==TRANS_WRITE
){
3892 BtShared
*pBt
= p
->pBt
;
3893 sqlite3BtreeEnter(p
);
3894 #ifndef SQLITE_OMIT_AUTOVACUUM
3895 if( pBt
->autoVacuum
){
3896 rc
= autoVacuumCommit(pBt
);
3897 if( rc
!=SQLITE_OK
){
3898 sqlite3BtreeLeave(p
);
3902 if( pBt
->bDoTruncate
){
3903 sqlite3PagerTruncateImage(pBt
->pPager
, pBt
->nPage
);
3906 rc
= sqlite3PagerCommitPhaseOne(pBt
->pPager
, zMaster
, 0);
3907 sqlite3BtreeLeave(p
);
3913 ** This function is called from both BtreeCommitPhaseTwo() and BtreeRollback()
3914 ** at the conclusion of a transaction.
3916 static void btreeEndTransaction(Btree
*p
){
3917 BtShared
*pBt
= p
->pBt
;
3918 sqlite3
*db
= p
->db
;
3919 assert( sqlite3BtreeHoldsMutex(p
) );
3921 #ifndef SQLITE_OMIT_AUTOVACUUM
3922 pBt
->bDoTruncate
= 0;
3924 if( p
->inTrans
>TRANS_NONE
&& db
->nVdbeRead
>1 ){
3925 /* If there are other active statements that belong to this database
3926 ** handle, downgrade to a read-only transaction. The other statements
3927 ** may still be reading from the database. */
3928 downgradeAllSharedCacheTableLocks(p
);
3929 p
->inTrans
= TRANS_READ
;
3931 /* If the handle had any kind of transaction open, decrement the
3932 ** transaction count of the shared btree. If the transaction count
3933 ** reaches 0, set the shared state to TRANS_NONE. The unlockBtreeIfUnused()
3934 ** call below will unlock the pager. */
3935 if( p
->inTrans
!=TRANS_NONE
){
3936 clearAllSharedCacheTableLocks(p
);
3937 pBt
->nTransaction
--;
3938 if( 0==pBt
->nTransaction
){
3939 pBt
->inTransaction
= TRANS_NONE
;
3943 /* Set the current transaction state to TRANS_NONE and unlock the
3944 ** pager if this call closed the only read or write transaction. */
3945 p
->inTrans
= TRANS_NONE
;
3946 unlockBtreeIfUnused(pBt
);
3953 ** Commit the transaction currently in progress.
3955 ** This routine implements the second phase of a 2-phase commit. The
3956 ** sqlite3BtreeCommitPhaseOne() routine does the first phase and should
3957 ** be invoked prior to calling this routine. The sqlite3BtreeCommitPhaseOne()
3958 ** routine did all the work of writing information out to disk and flushing the
3959 ** contents so that they are written onto the disk platter. All this
3960 ** routine has to do is delete or truncate or zero the header in the
3961 ** the rollback journal (which causes the transaction to commit) and
3964 ** Normally, if an error occurs while the pager layer is attempting to
3965 ** finalize the underlying journal file, this function returns an error and
3966 ** the upper layer will attempt a rollback. However, if the second argument
3967 ** is non-zero then this b-tree transaction is part of a multi-file
3968 ** transaction. In this case, the transaction has already been committed
3969 ** (by deleting a master journal file) and the caller will ignore this
3970 ** functions return code. So, even if an error occurs in the pager layer,
3971 ** reset the b-tree objects internal state to indicate that the write
3972 ** transaction has been closed. This is quite safe, as the pager will have
3973 ** transitioned to the error state.
3975 ** This will release the write lock on the database file. If there
3976 ** are no active cursors, it also releases the read lock.
3978 int sqlite3BtreeCommitPhaseTwo(Btree
*p
, int bCleanup
){
3980 if( p
->inTrans
==TRANS_NONE
) return SQLITE_OK
;
3981 sqlite3BtreeEnter(p
);
3984 /* If the handle has a write-transaction open, commit the shared-btrees
3985 ** transaction and set the shared state to TRANS_READ.
3987 if( p
->inTrans
==TRANS_WRITE
){
3989 BtShared
*pBt
= p
->pBt
;
3990 assert( pBt
->inTransaction
==TRANS_WRITE
);
3991 assert( pBt
->nTransaction
>0 );
3992 rc
= sqlite3PagerCommitPhaseTwo(pBt
->pPager
);
3993 if( rc
!=SQLITE_OK
&& bCleanup
==0 ){
3994 sqlite3BtreeLeave(p
);
3997 p
->iDataVersion
--; /* Compensate for pPager->iDataVersion++; */
3998 pBt
->inTransaction
= TRANS_READ
;
3999 btreeClearHasContent(pBt
);
4002 btreeEndTransaction(p
);
4003 sqlite3BtreeLeave(p
);
4008 ** Do both phases of a commit.
4010 int sqlite3BtreeCommit(Btree
*p
){
4012 sqlite3BtreeEnter(p
);
4013 rc
= sqlite3BtreeCommitPhaseOne(p
, 0);
4014 if( rc
==SQLITE_OK
){
4015 rc
= sqlite3BtreeCommitPhaseTwo(p
, 0);
4017 sqlite3BtreeLeave(p
);
4022 ** This routine sets the state to CURSOR_FAULT and the error
4023 ** code to errCode for every cursor on any BtShared that pBtree
4024 ** references. Or if the writeOnly flag is set to 1, then only
4025 ** trip write cursors and leave read cursors unchanged.
4027 ** Every cursor is a candidate to be tripped, including cursors
4028 ** that belong to other database connections that happen to be
4029 ** sharing the cache with pBtree.
4031 ** This routine gets called when a rollback occurs. If the writeOnly
4032 ** flag is true, then only write-cursors need be tripped - read-only
4033 ** cursors save their current positions so that they may continue
4034 ** following the rollback. Or, if writeOnly is false, all cursors are
4035 ** tripped. In general, writeOnly is false if the transaction being
4036 ** rolled back modified the database schema. In this case b-tree root
4037 ** pages may be moved or deleted from the database altogether, making
4038 ** it unsafe for read cursors to continue.
4040 ** If the writeOnly flag is true and an error is encountered while
4041 ** saving the current position of a read-only cursor, all cursors,
4042 ** including all read-cursors are tripped.
4044 ** SQLITE_OK is returned if successful, or if an error occurs while
4045 ** saving a cursor position, an SQLite error code.
4047 int sqlite3BtreeTripAllCursors(Btree
*pBtree
, int errCode
, int writeOnly
){
4051 assert( (writeOnly
==0 || writeOnly
==1) && BTCF_WriteFlag
==1 );
4053 sqlite3BtreeEnter(pBtree
);
4054 for(p
=pBtree
->pBt
->pCursor
; p
; p
=p
->pNext
){
4055 if( writeOnly
&& (p
->curFlags
& BTCF_WriteFlag
)==0 ){
4056 if( p
->eState
==CURSOR_VALID
|| p
->eState
==CURSOR_SKIPNEXT
){
4057 rc
= saveCursorPosition(p
);
4058 if( rc
!=SQLITE_OK
){
4059 (void)sqlite3BtreeTripAllCursors(pBtree
, rc
, 0);
4064 sqlite3BtreeClearCursor(p
);
4065 p
->eState
= CURSOR_FAULT
;
4066 p
->skipNext
= errCode
;
4068 btreeReleaseAllCursorPages(p
);
4070 sqlite3BtreeLeave(pBtree
);
4076 ** Rollback the transaction in progress.
4078 ** If tripCode is not SQLITE_OK then cursors will be invalidated (tripped).
4079 ** Only write cursors are tripped if writeOnly is true but all cursors are
4080 ** tripped if writeOnly is false. Any attempt to use
4081 ** a tripped cursor will result in an error.
4083 ** This will release the write lock on the database file. If there
4084 ** are no active cursors, it also releases the read lock.
4086 int sqlite3BtreeRollback(Btree
*p
, int tripCode
, int writeOnly
){
4088 BtShared
*pBt
= p
->pBt
;
4091 assert( writeOnly
==1 || writeOnly
==0 );
4092 assert( tripCode
==SQLITE_ABORT_ROLLBACK
|| tripCode
==SQLITE_OK
);
4093 sqlite3BtreeEnter(p
);
4094 if( tripCode
==SQLITE_OK
){
4095 rc
= tripCode
= saveAllCursors(pBt
, 0, 0);
4096 if( rc
) writeOnly
= 0;
4101 int rc2
= sqlite3BtreeTripAllCursors(p
, tripCode
, writeOnly
);
4102 assert( rc
==SQLITE_OK
|| (writeOnly
==0 && rc2
==SQLITE_OK
) );
4103 if( rc2
!=SQLITE_OK
) rc
= rc2
;
4107 if( p
->inTrans
==TRANS_WRITE
){
4110 assert( TRANS_WRITE
==pBt
->inTransaction
);
4111 rc2
= sqlite3PagerRollback(pBt
->pPager
);
4112 if( rc2
!=SQLITE_OK
){
4116 /* The rollback may have destroyed the pPage1->aData value. So
4117 ** call btreeGetPage() on page 1 again to make
4118 ** sure pPage1->aData is set correctly. */
4119 if( btreeGetPage(pBt
, 1, &pPage1
, 0)==SQLITE_OK
){
4120 int nPage
= get4byte(28+(u8
*)pPage1
->aData
);
4121 testcase( nPage
==0 );
4122 if( nPage
==0 ) sqlite3PagerPagecount(pBt
->pPager
, &nPage
);
4123 testcase( pBt
->nPage
!=nPage
);
4125 releasePageOne(pPage1
);
4127 assert( countValidCursors(pBt
, 1)==0 );
4128 pBt
->inTransaction
= TRANS_READ
;
4129 btreeClearHasContent(pBt
);
4132 btreeEndTransaction(p
);
4133 sqlite3BtreeLeave(p
);
4138 ** Start a statement subtransaction. The subtransaction can be rolled
4139 ** back independently of the main transaction. You must start a transaction
4140 ** before starting a subtransaction. The subtransaction is ended automatically
4141 ** if the main transaction commits or rolls back.
4143 ** Statement subtransactions are used around individual SQL statements
4144 ** that are contained within a BEGIN...COMMIT block. If a constraint
4145 ** error occurs within the statement, the effect of that one statement
4146 ** can be rolled back without having to rollback the entire transaction.
4148 ** A statement sub-transaction is implemented as an anonymous savepoint. The
4149 ** value passed as the second parameter is the total number of savepoints,
4150 ** including the new anonymous savepoint, open on the B-Tree. i.e. if there
4151 ** are no active savepoints and no other statement-transactions open,
4152 ** iStatement is 1. This anonymous savepoint can be released or rolled back
4153 ** using the sqlite3BtreeSavepoint() function.
4155 int sqlite3BtreeBeginStmt(Btree
*p
, int iStatement
){
4157 BtShared
*pBt
= p
->pBt
;
4158 sqlite3BtreeEnter(p
);
4159 assert( p
->inTrans
==TRANS_WRITE
);
4160 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
4161 assert( iStatement
>0 );
4162 assert( iStatement
>p
->db
->nSavepoint
);
4163 assert( pBt
->inTransaction
==TRANS_WRITE
);
4164 /* At the pager level, a statement transaction is a savepoint with
4165 ** an index greater than all savepoints created explicitly using
4166 ** SQL statements. It is illegal to open, release or rollback any
4167 ** such savepoints while the statement transaction savepoint is active.
4169 rc
= sqlite3PagerOpenSavepoint(pBt
->pPager
, iStatement
);
4170 sqlite3BtreeLeave(p
);
4175 ** The second argument to this function, op, is always SAVEPOINT_ROLLBACK
4176 ** or SAVEPOINT_RELEASE. This function either releases or rolls back the
4177 ** savepoint identified by parameter iSavepoint, depending on the value
4180 ** Normally, iSavepoint is greater than or equal to zero. However, if op is
4181 ** SAVEPOINT_ROLLBACK, then iSavepoint may also be -1. In this case the
4182 ** contents of the entire transaction are rolled back. This is different
4183 ** from a normal transaction rollback, as no locks are released and the
4184 ** transaction remains open.
4186 int sqlite3BtreeSavepoint(Btree
*p
, int op
, int iSavepoint
){
4188 if( p
&& p
->inTrans
==TRANS_WRITE
){
4189 BtShared
*pBt
= p
->pBt
;
4190 assert( op
==SAVEPOINT_RELEASE
|| op
==SAVEPOINT_ROLLBACK
);
4191 assert( iSavepoint
>=0 || (iSavepoint
==-1 && op
==SAVEPOINT_ROLLBACK
) );
4192 sqlite3BtreeEnter(p
);
4193 if( op
==SAVEPOINT_ROLLBACK
){
4194 rc
= saveAllCursors(pBt
, 0, 0);
4196 if( rc
==SQLITE_OK
){
4197 rc
= sqlite3PagerSavepoint(pBt
->pPager
, op
, iSavepoint
);
4199 if( rc
==SQLITE_OK
){
4200 if( iSavepoint
<0 && (pBt
->btsFlags
& BTS_INITIALLY_EMPTY
)!=0 ){
4203 rc
= newDatabase(pBt
);
4204 pBt
->nPage
= get4byte(28 + pBt
->pPage1
->aData
);
4206 /* The database size was written into the offset 28 of the header
4207 ** when the transaction started, so we know that the value at offset
4208 ** 28 is nonzero. */
4209 assert( pBt
->nPage
>0 );
4211 sqlite3BtreeLeave(p
);
4217 ** Create a new cursor for the BTree whose root is on the page
4218 ** iTable. If a read-only cursor is requested, it is assumed that
4219 ** the caller already has at least a read-only transaction open
4220 ** on the database already. If a write-cursor is requested, then
4221 ** the caller is assumed to have an open write transaction.
4223 ** If the BTREE_WRCSR bit of wrFlag is clear, then the cursor can only
4224 ** be used for reading. If the BTREE_WRCSR bit is set, then the cursor
4225 ** can be used for reading or for writing if other conditions for writing
4226 ** are also met. These are the conditions that must be met in order
4227 ** for writing to be allowed:
4229 ** 1: The cursor must have been opened with wrFlag containing BTREE_WRCSR
4231 ** 2: Other database connections that share the same pager cache
4232 ** but which are not in the READ_UNCOMMITTED state may not have
4233 ** cursors open with wrFlag==0 on the same table. Otherwise
4234 ** the changes made by this write cursor would be visible to
4235 ** the read cursors in the other database connection.
4237 ** 3: The database must be writable (not on read-only media)
4239 ** 4: There must be an active transaction.
4241 ** The BTREE_FORDELETE bit of wrFlag may optionally be set if BTREE_WRCSR
4242 ** is set. If FORDELETE is set, that is a hint to the implementation that
4243 ** this cursor will only be used to seek to and delete entries of an index
4244 ** as part of a larger DELETE statement. The FORDELETE hint is not used by
4245 ** this implementation. But in a hypothetical alternative storage engine
4246 ** in which index entries are automatically deleted when corresponding table
4247 ** rows are deleted, the FORDELETE flag is a hint that all SEEK and DELETE
4248 ** operations on this cursor can be no-ops and all READ operations can
4249 ** return a null row (2-bytes: 0x01 0x00).
4251 ** No checking is done to make sure that page iTable really is the
4252 ** root page of a b-tree. If it is not, then the cursor acquired
4253 ** will not work correctly.
4255 ** It is assumed that the sqlite3BtreeCursorZero() has been called
4256 ** on pCur to initialize the memory space prior to invoking this routine.
4258 static int btreeCursor(
4259 Btree
*p
, /* The btree */
4260 int iTable
, /* Root page of table to open */
4261 int wrFlag
, /* 1 to write. 0 read-only */
4262 struct KeyInfo
*pKeyInfo
, /* First arg to comparison function */
4263 BtCursor
*pCur
/* Space for new cursor */
4265 BtShared
*pBt
= p
->pBt
; /* Shared b-tree handle */
4266 BtCursor
*pX
; /* Looping over other all cursors */
4268 assert( sqlite3BtreeHoldsMutex(p
) );
4270 || wrFlag
==BTREE_WRCSR
4271 || wrFlag
==(BTREE_WRCSR
|BTREE_FORDELETE
)
4274 /* The following assert statements verify that if this is a sharable
4275 ** b-tree database, the connection is holding the required table locks,
4276 ** and that no other connection has any open cursor that conflicts with
4278 assert( hasSharedCacheTableLock(p
, iTable
, pKeyInfo
!=0, (wrFlag
?2:1)) );
4279 assert( wrFlag
==0 || !hasReadConflicts(p
, iTable
) );
4281 /* Assert that the caller has opened the required transaction. */
4282 assert( p
->inTrans
>TRANS_NONE
);
4283 assert( wrFlag
==0 || p
->inTrans
==TRANS_WRITE
);
4284 assert( pBt
->pPage1
&& pBt
->pPage1
->aData
);
4285 assert( wrFlag
==0 || (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
4288 allocateTempSpace(pBt
);
4289 if( pBt
->pTmpSpace
==0 ) return SQLITE_NOMEM_BKPT
;
4291 if( iTable
==1 && btreePagecount(pBt
)==0 ){
4292 assert( wrFlag
==0 );
4296 /* Now that no other errors can occur, finish filling in the BtCursor
4297 ** variables and link the cursor into the BtShared list. */
4298 pCur
->pgnoRoot
= (Pgno
)iTable
;
4300 pCur
->pKeyInfo
= pKeyInfo
;
4303 pCur
->curFlags
= wrFlag
? BTCF_WriteFlag
: 0;
4304 pCur
->curPagerFlags
= wrFlag
? 0 : PAGER_GET_READONLY
;
4305 /* If there are two or more cursors on the same btree, then all such
4306 ** cursors *must* have the BTCF_Multiple flag set. */
4307 for(pX
=pBt
->pCursor
; pX
; pX
=pX
->pNext
){
4308 if( pX
->pgnoRoot
==(Pgno
)iTable
){
4309 pX
->curFlags
|= BTCF_Multiple
;
4310 pCur
->curFlags
|= BTCF_Multiple
;
4313 pCur
->pNext
= pBt
->pCursor
;
4314 pBt
->pCursor
= pCur
;
4315 pCur
->eState
= CURSOR_INVALID
;
4318 int sqlite3BtreeCursor(
4319 Btree
*p
, /* The btree */
4320 int iTable
, /* Root page of table to open */
4321 int wrFlag
, /* 1 to write. 0 read-only */
4322 struct KeyInfo
*pKeyInfo
, /* First arg to xCompare() */
4323 BtCursor
*pCur
/* Write new cursor here */
4327 rc
= SQLITE_CORRUPT_BKPT
;
4329 sqlite3BtreeEnter(p
);
4330 rc
= btreeCursor(p
, iTable
, wrFlag
, pKeyInfo
, pCur
);
4331 sqlite3BtreeLeave(p
);
4337 ** Return the size of a BtCursor object in bytes.
4339 ** This interfaces is needed so that users of cursors can preallocate
4340 ** sufficient storage to hold a cursor. The BtCursor object is opaque
4341 ** to users so they cannot do the sizeof() themselves - they must call
4344 int sqlite3BtreeCursorSize(void){
4345 return ROUND8(sizeof(BtCursor
));
4349 ** Initialize memory that will be converted into a BtCursor object.
4351 ** The simple approach here would be to memset() the entire object
4352 ** to zero. But it turns out that the apPage[] and aiIdx[] arrays
4353 ** do not need to be zeroed and they are large, so we can save a lot
4354 ** of run-time by skipping the initialization of those elements.
4356 void sqlite3BtreeCursorZero(BtCursor
*p
){
4357 memset(p
, 0, offsetof(BtCursor
, BTCURSOR_FIRST_UNINIT
));
4361 ** Close a cursor. The read lock on the database file is released
4362 ** when the last cursor is closed.
4364 int sqlite3BtreeCloseCursor(BtCursor
*pCur
){
4365 Btree
*pBtree
= pCur
->pBtree
;
4367 BtShared
*pBt
= pCur
->pBt
;
4368 sqlite3BtreeEnter(pBtree
);
4369 assert( pBt
->pCursor
!=0 );
4370 if( pBt
->pCursor
==pCur
){
4371 pBt
->pCursor
= pCur
->pNext
;
4373 BtCursor
*pPrev
= pBt
->pCursor
;
4375 if( pPrev
->pNext
==pCur
){
4376 pPrev
->pNext
= pCur
->pNext
;
4379 pPrev
= pPrev
->pNext
;
4380 }while( ALWAYS(pPrev
) );
4382 btreeReleaseAllCursorPages(pCur
);
4383 unlockBtreeIfUnused(pBt
);
4384 sqlite3_free(pCur
->aOverflow
);
4385 sqlite3_free(pCur
->pKey
);
4386 sqlite3BtreeLeave(pBtree
);
4392 ** Make sure the BtCursor* given in the argument has a valid
4393 ** BtCursor.info structure. If it is not already valid, call
4394 ** btreeParseCell() to fill it in.
4396 ** BtCursor.info is a cache of the information in the current cell.
4397 ** Using this cache reduces the number of calls to btreeParseCell().
4400 static int cellInfoEqual(CellInfo
*a
, CellInfo
*b
){
4401 if( a
->nKey
!=b
->nKey
) return 0;
4402 if( a
->pPayload
!=b
->pPayload
) return 0;
4403 if( a
->nPayload
!=b
->nPayload
) return 0;
4404 if( a
->nLocal
!=b
->nLocal
) return 0;
4405 if( a
->nSize
!=b
->nSize
) return 0;
4408 static void assertCellInfo(BtCursor
*pCur
){
4410 memset(&info
, 0, sizeof(info
));
4411 btreeParseCell(pCur
->pPage
, pCur
->ix
, &info
);
4412 assert( CORRUPT_DB
|| cellInfoEqual(&info
, &pCur
->info
) );
4415 #define assertCellInfo(x)
4417 static SQLITE_NOINLINE
void getCellInfo(BtCursor
*pCur
){
4418 if( pCur
->info
.nSize
==0 ){
4419 pCur
->curFlags
|= BTCF_ValidNKey
;
4420 btreeParseCell(pCur
->pPage
,pCur
->ix
,&pCur
->info
);
4422 assertCellInfo(pCur
);
4426 #ifndef NDEBUG /* The next routine used only within assert() statements */
4428 ** Return true if the given BtCursor is valid. A valid cursor is one
4429 ** that is currently pointing to a row in a (non-empty) table.
4430 ** This is a verification routine is used only within assert() statements.
4432 int sqlite3BtreeCursorIsValid(BtCursor
*pCur
){
4433 return pCur
&& pCur
->eState
==CURSOR_VALID
;
4436 int sqlite3BtreeCursorIsValidNN(BtCursor
*pCur
){
4438 return pCur
->eState
==CURSOR_VALID
;
4442 ** Return the value of the integer key or "rowid" for a table btree.
4443 ** This routine is only valid for a cursor that is pointing into a
4444 ** ordinary table btree. If the cursor points to an index btree or
4445 ** is invalid, the result of this routine is undefined.
4447 i64
sqlite3BtreeIntegerKey(BtCursor
*pCur
){
4448 assert( cursorHoldsMutex(pCur
) );
4449 assert( pCur
->eState
==CURSOR_VALID
);
4450 assert( pCur
->curIntKey
);
4452 return pCur
->info
.nKey
;
4455 #ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC
4457 ** Return the offset into the database file for the start of the
4458 ** payload to which the cursor is pointing.
4460 i64
sqlite3BtreeOffset(BtCursor
*pCur
){
4461 assert( cursorHoldsMutex(pCur
) );
4462 assert( pCur
->eState
==CURSOR_VALID
);
4464 return (i64
)pCur
->pBt
->pageSize
*((i64
)pCur
->pPage
->pgno
- 1) +
4465 (i64
)(pCur
->info
.pPayload
- pCur
->pPage
->aData
);
4467 #endif /* SQLITE_ENABLE_OFFSET_SQL_FUNC */
4470 ** Return the number of bytes of payload for the entry that pCur is
4471 ** currently pointing to. For table btrees, this will be the amount
4472 ** of data. For index btrees, this will be the size of the key.
4474 ** The caller must guarantee that the cursor is pointing to a non-NULL
4475 ** valid entry. In other words, the calling procedure must guarantee
4476 ** that the cursor has Cursor.eState==CURSOR_VALID.
4478 u32
sqlite3BtreePayloadSize(BtCursor
*pCur
){
4479 assert( cursorHoldsMutex(pCur
) );
4480 assert( pCur
->eState
==CURSOR_VALID
);
4482 return pCur
->info
.nPayload
;
4486 ** Given the page number of an overflow page in the database (parameter
4487 ** ovfl), this function finds the page number of the next page in the
4488 ** linked list of overflow pages. If possible, it uses the auto-vacuum
4489 ** pointer-map data instead of reading the content of page ovfl to do so.
4491 ** If an error occurs an SQLite error code is returned. Otherwise:
4493 ** The page number of the next overflow page in the linked list is
4494 ** written to *pPgnoNext. If page ovfl is the last page in its linked
4495 ** list, *pPgnoNext is set to zero.
4497 ** If ppPage is not NULL, and a reference to the MemPage object corresponding
4498 ** to page number pOvfl was obtained, then *ppPage is set to point to that
4499 ** reference. It is the responsibility of the caller to call releasePage()
4500 ** on *ppPage to free the reference. In no reference was obtained (because
4501 ** the pointer-map was used to obtain the value for *pPgnoNext), then
4502 ** *ppPage is set to zero.
4504 static int getOverflowPage(
4505 BtShared
*pBt
, /* The database file */
4506 Pgno ovfl
, /* Current overflow page number */
4507 MemPage
**ppPage
, /* OUT: MemPage handle (may be NULL) */
4508 Pgno
*pPgnoNext
/* OUT: Next overflow page number */
4514 assert( sqlite3_mutex_held(pBt
->mutex
) );
4517 #ifndef SQLITE_OMIT_AUTOVACUUM
4518 /* Try to find the next page in the overflow list using the
4519 ** autovacuum pointer-map pages. Guess that the next page in
4520 ** the overflow list is page number (ovfl+1). If that guess turns
4521 ** out to be wrong, fall back to loading the data of page
4522 ** number ovfl to determine the next page number.
4524 if( pBt
->autoVacuum
){
4526 Pgno iGuess
= ovfl
+1;
4529 while( PTRMAP_ISPAGE(pBt
, iGuess
) || iGuess
==PENDING_BYTE_PAGE(pBt
) ){
4533 if( iGuess
<=btreePagecount(pBt
) ){
4534 rc
= ptrmapGet(pBt
, iGuess
, &eType
, &pgno
);
4535 if( rc
==SQLITE_OK
&& eType
==PTRMAP_OVERFLOW2
&& pgno
==ovfl
){
4543 assert( next
==0 || rc
==SQLITE_DONE
);
4544 if( rc
==SQLITE_OK
){
4545 rc
= btreeGetPage(pBt
, ovfl
, &pPage
, (ppPage
==0) ? PAGER_GET_READONLY
: 0);
4546 assert( rc
==SQLITE_OK
|| pPage
==0 );
4547 if( rc
==SQLITE_OK
){
4548 next
= get4byte(pPage
->aData
);
4558 return (rc
==SQLITE_DONE
? SQLITE_OK
: rc
);
4562 ** Copy data from a buffer to a page, or from a page to a buffer.
4564 ** pPayload is a pointer to data stored on database page pDbPage.
4565 ** If argument eOp is false, then nByte bytes of data are copied
4566 ** from pPayload to the buffer pointed at by pBuf. If eOp is true,
4567 ** then sqlite3PagerWrite() is called on pDbPage and nByte bytes
4568 ** of data are copied from the buffer pBuf to pPayload.
4570 ** SQLITE_OK is returned on success, otherwise an error code.
4572 static int copyPayload(
4573 void *pPayload
, /* Pointer to page data */
4574 void *pBuf
, /* Pointer to buffer */
4575 int nByte
, /* Number of bytes to copy */
4576 int eOp
, /* 0 -> copy from page, 1 -> copy to page */
4577 DbPage
*pDbPage
/* Page containing pPayload */
4580 /* Copy data from buffer to page (a write operation) */
4581 int rc
= sqlite3PagerWrite(pDbPage
);
4582 if( rc
!=SQLITE_OK
){
4585 memcpy(pPayload
, pBuf
, nByte
);
4587 /* Copy data from page to buffer (a read operation) */
4588 memcpy(pBuf
, pPayload
, nByte
);
4594 ** This function is used to read or overwrite payload information
4595 ** for the entry that the pCur cursor is pointing to. The eOp
4596 ** argument is interpreted as follows:
4598 ** 0: The operation is a read. Populate the overflow cache.
4599 ** 1: The operation is a write. Populate the overflow cache.
4601 ** A total of "amt" bytes are read or written beginning at "offset".
4602 ** Data is read to or from the buffer pBuf.
4604 ** The content being read or written might appear on the main page
4605 ** or be scattered out on multiple overflow pages.
4607 ** If the current cursor entry uses one or more overflow pages
4608 ** this function may allocate space for and lazily populate
4609 ** the overflow page-list cache array (BtCursor.aOverflow).
4610 ** Subsequent calls use this cache to make seeking to the supplied offset
4613 ** Once an overflow page-list cache has been allocated, it must be
4614 ** invalidated if some other cursor writes to the same table, or if
4615 ** the cursor is moved to a different row. Additionally, in auto-vacuum
4616 ** mode, the following events may invalidate an overflow page-list cache.
4618 ** * An incremental vacuum,
4619 ** * A commit in auto_vacuum="full" mode,
4620 ** * Creating a table (may require moving an overflow page).
4622 static int accessPayload(
4623 BtCursor
*pCur
, /* Cursor pointing to entry to read from */
4624 u32 offset
, /* Begin reading this far into payload */
4625 u32 amt
, /* Read this many bytes */
4626 unsigned char *pBuf
, /* Write the bytes into this buffer */
4627 int eOp
/* zero to read. non-zero to write. */
4629 unsigned char *aPayload
;
4632 MemPage
*pPage
= pCur
->pPage
; /* Btree page of current entry */
4633 BtShared
*pBt
= pCur
->pBt
; /* Btree this cursor belongs to */
4634 #ifdef SQLITE_DIRECT_OVERFLOW_READ
4635 unsigned char * const pBufStart
= pBuf
; /* Start of original out buffer */
4639 assert( eOp
==0 || eOp
==1 );
4640 assert( pCur
->eState
==CURSOR_VALID
);
4641 assert( pCur
->ix
<pPage
->nCell
);
4642 assert( cursorHoldsMutex(pCur
) );
4645 aPayload
= pCur
->info
.pPayload
;
4646 assert( offset
+amt
<= pCur
->info
.nPayload
);
4648 assert( aPayload
> pPage
->aData
);
4649 if( (uptr
)(aPayload
- pPage
->aData
) > (pBt
->usableSize
- pCur
->info
.nLocal
) ){
4650 /* Trying to read or write past the end of the data is an error. The
4651 ** conditional above is really:
4652 ** &aPayload[pCur->info.nLocal] > &pPage->aData[pBt->usableSize]
4653 ** but is recast into its current form to avoid integer overflow problems
4655 return SQLITE_CORRUPT_PAGE(pPage
);
4658 /* Check if data must be read/written to/from the btree page itself. */
4659 if( offset
<pCur
->info
.nLocal
){
4661 if( a
+offset
>pCur
->info
.nLocal
){
4662 a
= pCur
->info
.nLocal
- offset
;
4664 rc
= copyPayload(&aPayload
[offset
], pBuf
, a
, eOp
, pPage
->pDbPage
);
4669 offset
-= pCur
->info
.nLocal
;
4673 if( rc
==SQLITE_OK
&& amt
>0 ){
4674 const u32 ovflSize
= pBt
->usableSize
- 4; /* Bytes content per ovfl page */
4677 nextPage
= get4byte(&aPayload
[pCur
->info
.nLocal
]);
4679 /* If the BtCursor.aOverflow[] has not been allocated, allocate it now.
4681 ** The aOverflow[] array is sized at one entry for each overflow page
4682 ** in the overflow chain. The page number of the first overflow page is
4683 ** stored in aOverflow[0], etc. A value of 0 in the aOverflow[] array
4684 ** means "not yet known" (the cache is lazily populated).
4686 if( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 ){
4687 int nOvfl
= (pCur
->info
.nPayload
-pCur
->info
.nLocal
+ovflSize
-1)/ovflSize
;
4688 if( pCur
->aOverflow
==0
4689 || nOvfl
*(int)sizeof(Pgno
) > sqlite3MallocSize(pCur
->aOverflow
)
4691 Pgno
*aNew
= (Pgno
*)sqlite3Realloc(
4692 pCur
->aOverflow
, nOvfl
*2*sizeof(Pgno
)
4695 return SQLITE_NOMEM_BKPT
;
4697 pCur
->aOverflow
= aNew
;
4700 memset(pCur
->aOverflow
, 0, nOvfl
*sizeof(Pgno
));
4701 pCur
->curFlags
|= BTCF_ValidOvfl
;
4703 /* If the overflow page-list cache has been allocated and the
4704 ** entry for the first required overflow page is valid, skip
4707 if( pCur
->aOverflow
[offset
/ovflSize
] ){
4708 iIdx
= (offset
/ovflSize
);
4709 nextPage
= pCur
->aOverflow
[iIdx
];
4710 offset
= (offset
%ovflSize
);
4714 assert( rc
==SQLITE_OK
&& amt
>0 );
4716 /* If required, populate the overflow page-list cache. */
4717 assert( pCur
->aOverflow
[iIdx
]==0
4718 || pCur
->aOverflow
[iIdx
]==nextPage
4720 pCur
->aOverflow
[iIdx
] = nextPage
;
4722 if( offset
>=ovflSize
){
4723 /* The only reason to read this page is to obtain the page
4724 ** number for the next page in the overflow chain. The page
4725 ** data is not required. So first try to lookup the overflow
4726 ** page-list cache, if any, then fall back to the getOverflowPage()
4729 assert( pCur
->curFlags
& BTCF_ValidOvfl
);
4730 assert( pCur
->pBtree
->db
==pBt
->db
);
4731 if( pCur
->aOverflow
[iIdx
+1] ){
4732 nextPage
= pCur
->aOverflow
[iIdx
+1];
4734 rc
= getOverflowPage(pBt
, nextPage
, 0, &nextPage
);
4738 /* Need to read this page properly. It contains some of the
4739 ** range of data that is being read (eOp==0) or written (eOp!=0).
4741 #ifdef SQLITE_DIRECT_OVERFLOW_READ
4742 sqlite3_file
*fd
; /* File from which to do direct overflow read */
4745 if( a
+ offset
> ovflSize
){
4746 a
= ovflSize
- offset
;
4749 #ifdef SQLITE_DIRECT_OVERFLOW_READ
4750 /* If all the following are true:
4752 ** 1) this is a read operation, and
4753 ** 2) data is required from the start of this overflow page, and
4754 ** 3) there is no open write-transaction, and
4755 ** 4) the database is file-backed, and
4756 ** 5) the page is not in the WAL file
4757 ** 6) at least 4 bytes have already been read into the output buffer
4759 ** then data can be read directly from the database file into the
4760 ** output buffer, bypassing the page-cache altogether. This speeds
4761 ** up loading large records that span many overflow pages.
4763 if( eOp
==0 /* (1) */
4764 && offset
==0 /* (2) */
4765 && pBt
->inTransaction
==TRANS_READ
/* (3) */
4766 && (fd
= sqlite3PagerFile(pBt
->pPager
))->pMethods
/* (4) */
4767 && 0==sqlite3PagerUseWal(pBt
->pPager
, nextPage
) /* (5) */
4768 && &pBuf
[-4]>=pBufStart
/* (6) */
4771 u8
*aWrite
= &pBuf
[-4];
4772 assert( aWrite
>=pBufStart
); /* due to (6) */
4773 memcpy(aSave
, aWrite
, 4);
4774 rc
= sqlite3OsRead(fd
, aWrite
, a
+4, (i64
)pBt
->pageSize
*(nextPage
-1));
4775 nextPage
= get4byte(aWrite
);
4776 memcpy(aWrite
, aSave
, 4);
4782 rc
= sqlite3PagerGet(pBt
->pPager
, nextPage
, &pDbPage
,
4783 (eOp
==0 ? PAGER_GET_READONLY
: 0)
4785 if( rc
==SQLITE_OK
){
4786 aPayload
= sqlite3PagerGetData(pDbPage
);
4787 nextPage
= get4byte(aPayload
);
4788 rc
= copyPayload(&aPayload
[offset
+4], pBuf
, a
, eOp
, pDbPage
);
4789 sqlite3PagerUnref(pDbPage
);
4794 if( amt
==0 ) return rc
;
4802 if( rc
==SQLITE_OK
&& amt
>0 ){
4803 /* Overflow chain ends prematurely */
4804 return SQLITE_CORRUPT_PAGE(pPage
);
4810 ** Read part of the payload for the row at which that cursor pCur is currently
4811 ** pointing. "amt" bytes will be transferred into pBuf[]. The transfer
4812 ** begins at "offset".
4814 ** pCur can be pointing to either a table or an index b-tree.
4815 ** If pointing to a table btree, then the content section is read. If
4816 ** pCur is pointing to an index b-tree then the key section is read.
4818 ** For sqlite3BtreePayload(), the caller must ensure that pCur is pointing
4819 ** to a valid row in the table. For sqlite3BtreePayloadChecked(), the
4820 ** cursor might be invalid or might need to be restored before being read.
4822 ** Return SQLITE_OK on success or an error code if anything goes
4823 ** wrong. An error is returned if "offset+amt" is larger than
4824 ** the available payload.
4826 int sqlite3BtreePayload(BtCursor
*pCur
, u32 offset
, u32 amt
, void *pBuf
){
4827 assert( cursorHoldsMutex(pCur
) );
4828 assert( pCur
->eState
==CURSOR_VALID
);
4829 assert( pCur
->iPage
>=0 && pCur
->pPage
);
4830 assert( pCur
->ix
<pCur
->pPage
->nCell
);
4831 return accessPayload(pCur
, offset
, amt
, (unsigned char*)pBuf
, 0);
4835 ** This variant of sqlite3BtreePayload() works even if the cursor has not
4836 ** in the CURSOR_VALID state. It is only used by the sqlite3_blob_read()
4839 #ifndef SQLITE_OMIT_INCRBLOB
4840 static SQLITE_NOINLINE
int accessPayloadChecked(
4847 if ( pCur
->eState
==CURSOR_INVALID
){
4848 return SQLITE_ABORT
;
4850 assert( cursorOwnsBtShared(pCur
) );
4851 rc
= btreeRestoreCursorPosition(pCur
);
4852 return rc
? rc
: accessPayload(pCur
, offset
, amt
, pBuf
, 0);
4854 int sqlite3BtreePayloadChecked(BtCursor
*pCur
, u32 offset
, u32 amt
, void *pBuf
){
4855 if( pCur
->eState
==CURSOR_VALID
){
4856 assert( cursorOwnsBtShared(pCur
) );
4857 return accessPayload(pCur
, offset
, amt
, pBuf
, 0);
4859 return accessPayloadChecked(pCur
, offset
, amt
, pBuf
);
4862 #endif /* SQLITE_OMIT_INCRBLOB */
4865 ** Return a pointer to payload information from the entry that the
4866 ** pCur cursor is pointing to. The pointer is to the beginning of
4867 ** the key if index btrees (pPage->intKey==0) and is the data for
4868 ** table btrees (pPage->intKey==1). The number of bytes of available
4869 ** key/data is written into *pAmt. If *pAmt==0, then the value
4870 ** returned will not be a valid pointer.
4872 ** This routine is an optimization. It is common for the entire key
4873 ** and data to fit on the local page and for there to be no overflow
4874 ** pages. When that is so, this routine can be used to access the
4875 ** key and data without making a copy. If the key and/or data spills
4876 ** onto overflow pages, then accessPayload() must be used to reassemble
4877 ** the key/data and copy it into a preallocated buffer.
4879 ** The pointer returned by this routine looks directly into the cached
4880 ** page of the database. The data might change or move the next time
4881 ** any btree routine is called.
4883 static const void *fetchPayload(
4884 BtCursor
*pCur
, /* Cursor pointing to entry to read from */
4885 u32
*pAmt
/* Write the number of available bytes here */
4888 assert( pCur
!=0 && pCur
->iPage
>=0 && pCur
->pPage
);
4889 assert( pCur
->eState
==CURSOR_VALID
);
4890 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
4891 assert( cursorOwnsBtShared(pCur
) );
4892 assert( pCur
->ix
<pCur
->pPage
->nCell
);
4893 assert( pCur
->info
.nSize
>0 );
4894 assert( pCur
->info
.pPayload
>pCur
->pPage
->aData
|| CORRUPT_DB
);
4895 assert( pCur
->info
.pPayload
<pCur
->pPage
->aDataEnd
||CORRUPT_DB
);
4896 amt
= pCur
->info
.nLocal
;
4897 if( amt
>(int)(pCur
->pPage
->aDataEnd
- pCur
->info
.pPayload
) ){
4898 /* There is too little space on the page for the expected amount
4899 ** of local content. Database must be corrupt. */
4900 assert( CORRUPT_DB
);
4901 amt
= MAX(0, (int)(pCur
->pPage
->aDataEnd
- pCur
->info
.pPayload
));
4904 return (void*)pCur
->info
.pPayload
;
4909 ** For the entry that cursor pCur is point to, return as
4910 ** many bytes of the key or data as are available on the local
4911 ** b-tree page. Write the number of available bytes into *pAmt.
4913 ** The pointer returned is ephemeral. The key/data may move
4914 ** or be destroyed on the next call to any Btree routine,
4915 ** including calls from other threads against the same cache.
4916 ** Hence, a mutex on the BtShared should be held prior to calling
4919 ** These routines is used to get quick access to key and data
4920 ** in the common case where no overflow pages are used.
4922 const void *sqlite3BtreePayloadFetch(BtCursor
*pCur
, u32
*pAmt
){
4923 return fetchPayload(pCur
, pAmt
);
4928 ** Move the cursor down to a new child page. The newPgno argument is the
4929 ** page number of the child page to move to.
4931 ** This function returns SQLITE_CORRUPT if the page-header flags field of
4932 ** the new child page does not match the flags field of the parent (i.e.
4933 ** if an intkey page appears to be the parent of a non-intkey page, or
4936 static int moveToChild(BtCursor
*pCur
, u32 newPgno
){
4937 BtShared
*pBt
= pCur
->pBt
;
4939 assert( cursorOwnsBtShared(pCur
) );
4940 assert( pCur
->eState
==CURSOR_VALID
);
4941 assert( pCur
->iPage
<BTCURSOR_MAX_DEPTH
);
4942 assert( pCur
->iPage
>=0 );
4943 if( pCur
->iPage
>=(BTCURSOR_MAX_DEPTH
-1) ){
4944 return SQLITE_CORRUPT_BKPT
;
4946 pCur
->info
.nSize
= 0;
4947 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
4948 pCur
->aiIdx
[pCur
->iPage
] = pCur
->ix
;
4949 pCur
->apPage
[pCur
->iPage
] = pCur
->pPage
;
4952 return getAndInitPage(pBt
, newPgno
, &pCur
->pPage
, pCur
, pCur
->curPagerFlags
);
4957 ** Page pParent is an internal (non-leaf) tree page. This function
4958 ** asserts that page number iChild is the left-child if the iIdx'th
4959 ** cell in page pParent. Or, if iIdx is equal to the total number of
4960 ** cells in pParent, that page number iChild is the right-child of
4963 static void assertParentIndex(MemPage
*pParent
, int iIdx
, Pgno iChild
){
4964 if( CORRUPT_DB
) return; /* The conditions tested below might not be true
4965 ** in a corrupt database */
4966 assert( iIdx
<=pParent
->nCell
);
4967 if( iIdx
==pParent
->nCell
){
4968 assert( get4byte(&pParent
->aData
[pParent
->hdrOffset
+8])==iChild
);
4970 assert( get4byte(findCell(pParent
, iIdx
))==iChild
);
4974 # define assertParentIndex(x,y,z)
4978 ** Move the cursor up to the parent page.
4980 ** pCur->idx is set to the cell index that contains the pointer
4981 ** to the page we are coming from. If we are coming from the
4982 ** right-most child page then pCur->idx is set to one more than
4983 ** the largest cell index.
4985 static void moveToParent(BtCursor
*pCur
){
4987 assert( cursorOwnsBtShared(pCur
) );
4988 assert( pCur
->eState
==CURSOR_VALID
);
4989 assert( pCur
->iPage
>0 );
4990 assert( pCur
->pPage
);
4992 pCur
->apPage
[pCur
->iPage
-1],
4993 pCur
->aiIdx
[pCur
->iPage
-1],
4996 testcase( pCur
->aiIdx
[pCur
->iPage
-1] > pCur
->apPage
[pCur
->iPage
-1]->nCell
);
4997 pCur
->info
.nSize
= 0;
4998 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
4999 pCur
->ix
= pCur
->aiIdx
[pCur
->iPage
-1];
5000 pLeaf
= pCur
->pPage
;
5001 pCur
->pPage
= pCur
->apPage
[--pCur
->iPage
];
5002 releasePageNotNull(pLeaf
);
5006 ** Move the cursor to point to the root page of its b-tree structure.
5008 ** If the table has a virtual root page, then the cursor is moved to point
5009 ** to the virtual root page instead of the actual root page. A table has a
5010 ** virtual root page when the actual root page contains no cells and a
5011 ** single child page. This can only happen with the table rooted at page 1.
5013 ** If the b-tree structure is empty, the cursor state is set to
5014 ** CURSOR_INVALID and this routine returns SQLITE_EMPTY. Otherwise,
5015 ** the cursor is set to point to the first cell located on the root
5016 ** (or virtual root) page and the cursor state is set to CURSOR_VALID.
5018 ** If this function returns successfully, it may be assumed that the
5019 ** page-header flags indicate that the [virtual] root-page is the expected
5020 ** kind of b-tree page (i.e. if when opening the cursor the caller did not
5021 ** specify a KeyInfo structure the flags byte is set to 0x05 or 0x0D,
5022 ** indicating a table b-tree, or if the caller did specify a KeyInfo
5023 ** structure the flags byte is set to 0x02 or 0x0A, indicating an index
5026 static int moveToRoot(BtCursor
*pCur
){
5030 assert( cursorOwnsBtShared(pCur
) );
5031 assert( CURSOR_INVALID
< CURSOR_REQUIRESEEK
);
5032 assert( CURSOR_VALID
< CURSOR_REQUIRESEEK
);
5033 assert( CURSOR_FAULT
> CURSOR_REQUIRESEEK
);
5034 assert( pCur
->eState
< CURSOR_REQUIRESEEK
|| pCur
->iPage
<0 );
5035 assert( pCur
->pgnoRoot
>0 || pCur
->iPage
<0 );
5037 if( pCur
->iPage
>=0 ){
5039 releasePageNotNull(pCur
->pPage
);
5040 while( --pCur
->iPage
){
5041 releasePageNotNull(pCur
->apPage
[pCur
->iPage
]);
5043 pCur
->pPage
= pCur
->apPage
[0];
5046 }else if( pCur
->pgnoRoot
==0 ){
5047 pCur
->eState
= CURSOR_INVALID
;
5048 return SQLITE_EMPTY
;
5050 assert( pCur
->iPage
==(-1) );
5051 if( pCur
->eState
>=CURSOR_REQUIRESEEK
){
5052 if( pCur
->eState
==CURSOR_FAULT
){
5053 assert( pCur
->skipNext
!=SQLITE_OK
);
5054 return pCur
->skipNext
;
5056 sqlite3BtreeClearCursor(pCur
);
5058 rc
= getAndInitPage(pCur
->pBtree
->pBt
, pCur
->pgnoRoot
, &pCur
->pPage
,
5059 0, pCur
->curPagerFlags
);
5060 if( rc
!=SQLITE_OK
){
5061 pCur
->eState
= CURSOR_INVALID
;
5065 pCur
->curIntKey
= pCur
->pPage
->intKey
;
5067 pRoot
= pCur
->pPage
;
5068 assert( pRoot
->pgno
==pCur
->pgnoRoot
);
5070 /* If pCur->pKeyInfo is not NULL, then the caller that opened this cursor
5071 ** expected to open it on an index b-tree. Otherwise, if pKeyInfo is
5072 ** NULL, the caller expects a table b-tree. If this is not the case,
5073 ** return an SQLITE_CORRUPT error.
5075 ** Earlier versions of SQLite assumed that this test could not fail
5076 ** if the root page was already loaded when this function was called (i.e.
5077 ** if pCur->iPage>=0). But this is not so if the database is corrupted
5078 ** in such a way that page pRoot is linked into a second b-tree table
5079 ** (or the freelist). */
5080 assert( pRoot
->intKey
==1 || pRoot
->intKey
==0 );
5081 if( pRoot
->isInit
==0 || (pCur
->pKeyInfo
==0)!=pRoot
->intKey
){
5082 return SQLITE_CORRUPT_PAGE(pCur
->pPage
);
5087 pCur
->info
.nSize
= 0;
5088 pCur
->curFlags
&= ~(BTCF_AtLast
|BTCF_ValidNKey
|BTCF_ValidOvfl
);
5090 pRoot
= pCur
->pPage
;
5091 if( pRoot
->nCell
>0 ){
5092 pCur
->eState
= CURSOR_VALID
;
5093 }else if( !pRoot
->leaf
){
5095 if( pRoot
->pgno
!=1 ) return SQLITE_CORRUPT_BKPT
;
5096 subpage
= get4byte(&pRoot
->aData
[pRoot
->hdrOffset
+8]);
5097 pCur
->eState
= CURSOR_VALID
;
5098 rc
= moveToChild(pCur
, subpage
);
5100 pCur
->eState
= CURSOR_INVALID
;
5107 ** Move the cursor down to the left-most leaf entry beneath the
5108 ** entry to which it is currently pointing.
5110 ** The left-most leaf is the one with the smallest key - the first
5111 ** in ascending order.
5113 static int moveToLeftmost(BtCursor
*pCur
){
5118 assert( cursorOwnsBtShared(pCur
) );
5119 assert( pCur
->eState
==CURSOR_VALID
);
5120 while( rc
==SQLITE_OK
&& !(pPage
= pCur
->pPage
)->leaf
){
5121 assert( pCur
->ix
<pPage
->nCell
);
5122 pgno
= get4byte(findCell(pPage
, pCur
->ix
));
5123 rc
= moveToChild(pCur
, pgno
);
5129 ** Move the cursor down to the right-most leaf entry beneath the
5130 ** page to which it is currently pointing. Notice the difference
5131 ** between moveToLeftmost() and moveToRightmost(). moveToLeftmost()
5132 ** finds the left-most entry beneath the *entry* whereas moveToRightmost()
5133 ** finds the right-most entry beneath the *page*.
5135 ** The right-most entry is the one with the largest key - the last
5136 ** key in ascending order.
5138 static int moveToRightmost(BtCursor
*pCur
){
5143 assert( cursorOwnsBtShared(pCur
) );
5144 assert( pCur
->eState
==CURSOR_VALID
);
5145 while( !(pPage
= pCur
->pPage
)->leaf
){
5146 pgno
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
5147 pCur
->ix
= pPage
->nCell
;
5148 rc
= moveToChild(pCur
, pgno
);
5151 pCur
->ix
= pPage
->nCell
-1;
5152 assert( pCur
->info
.nSize
==0 );
5153 assert( (pCur
->curFlags
& BTCF_ValidNKey
)==0 );
5157 /* Move the cursor to the first entry in the table. Return SQLITE_OK
5158 ** on success. Set *pRes to 0 if the cursor actually points to something
5159 ** or set *pRes to 1 if the table is empty.
5161 int sqlite3BtreeFirst(BtCursor
*pCur
, int *pRes
){
5164 assert( cursorOwnsBtShared(pCur
) );
5165 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5166 rc
= moveToRoot(pCur
);
5167 if( rc
==SQLITE_OK
){
5168 assert( pCur
->pPage
->nCell
>0 );
5170 rc
= moveToLeftmost(pCur
);
5171 }else if( rc
==SQLITE_EMPTY
){
5172 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5179 /* Move the cursor to the last entry in the table. Return SQLITE_OK
5180 ** on success. Set *pRes to 0 if the cursor actually points to something
5181 ** or set *pRes to 1 if the table is empty.
5183 int sqlite3BtreeLast(BtCursor
*pCur
, int *pRes
){
5186 assert( cursorOwnsBtShared(pCur
) );
5187 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5189 /* If the cursor already points to the last entry, this is a no-op. */
5190 if( CURSOR_VALID
==pCur
->eState
&& (pCur
->curFlags
& BTCF_AtLast
)!=0 ){
5192 /* This block serves to assert() that the cursor really does point
5193 ** to the last entry in the b-tree. */
5195 for(ii
=0; ii
<pCur
->iPage
; ii
++){
5196 assert( pCur
->aiIdx
[ii
]==pCur
->apPage
[ii
]->nCell
);
5198 assert( pCur
->ix
==pCur
->pPage
->nCell
-1 );
5199 assert( pCur
->pPage
->leaf
);
5204 rc
= moveToRoot(pCur
);
5205 if( rc
==SQLITE_OK
){
5206 assert( pCur
->eState
==CURSOR_VALID
);
5208 rc
= moveToRightmost(pCur
);
5209 if( rc
==SQLITE_OK
){
5210 pCur
->curFlags
|= BTCF_AtLast
;
5212 pCur
->curFlags
&= ~BTCF_AtLast
;
5214 }else if( rc
==SQLITE_EMPTY
){
5215 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5222 /* Move the cursor so that it points to an entry near the key
5223 ** specified by pIdxKey or intKey. Return a success code.
5225 ** For INTKEY tables, the intKey parameter is used. pIdxKey
5226 ** must be NULL. For index tables, pIdxKey is used and intKey
5229 ** If an exact match is not found, then the cursor is always
5230 ** left pointing at a leaf page which would hold the entry if it
5231 ** were present. The cursor might point to an entry that comes
5232 ** before or after the key.
5234 ** An integer is written into *pRes which is the result of
5235 ** comparing the key with the entry to which the cursor is
5236 ** pointing. The meaning of the integer written into
5237 ** *pRes is as follows:
5239 ** *pRes<0 The cursor is left pointing at an entry that
5240 ** is smaller than intKey/pIdxKey or if the table is empty
5241 ** and the cursor is therefore left point to nothing.
5243 ** *pRes==0 The cursor is left pointing at an entry that
5244 ** exactly matches intKey/pIdxKey.
5246 ** *pRes>0 The cursor is left pointing at an entry that
5247 ** is larger than intKey/pIdxKey.
5249 ** For index tables, the pIdxKey->eqSeen field is set to 1 if there
5250 ** exists an entry in the table that exactly matches pIdxKey.
5252 int sqlite3BtreeMovetoUnpacked(
5253 BtCursor
*pCur
, /* The cursor to be moved */
5254 UnpackedRecord
*pIdxKey
, /* Unpacked index key */
5255 i64 intKey
, /* The table key */
5256 int biasRight
, /* If true, bias the search to the high end */
5257 int *pRes
/* Write search results here */
5260 RecordCompare xRecordCompare
;
5262 assert( cursorOwnsBtShared(pCur
) );
5263 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5265 assert( (pIdxKey
==0)==(pCur
->pKeyInfo
==0) );
5266 assert( pCur
->eState
!=CURSOR_VALID
|| (pIdxKey
==0)==(pCur
->curIntKey
!=0) );
5268 /* If the cursor is already positioned at the point we are trying
5269 ** to move to, then just return without doing any work */
5271 && pCur
->eState
==CURSOR_VALID
&& (pCur
->curFlags
& BTCF_ValidNKey
)!=0
5273 if( pCur
->info
.nKey
==intKey
){
5277 if( pCur
->info
.nKey
<intKey
){
5278 if( (pCur
->curFlags
& BTCF_AtLast
)!=0 ){
5282 /* If the requested key is one more than the previous key, then
5283 ** try to get there using sqlite3BtreeNext() rather than a full
5284 ** binary search. This is an optimization only. The correct answer
5285 ** is still obtained without this case, only a little more slowely */
5286 if( pCur
->info
.nKey
+1==intKey
&& !pCur
->skipNext
){
5288 rc
= sqlite3BtreeNext(pCur
, 0);
5289 if( rc
==SQLITE_OK
){
5291 if( pCur
->info
.nKey
==intKey
){
5294 }else if( rc
==SQLITE_DONE
){
5304 xRecordCompare
= sqlite3VdbeFindCompare(pIdxKey
);
5305 pIdxKey
->errCode
= 0;
5306 assert( pIdxKey
->default_rc
==1
5307 || pIdxKey
->default_rc
==0
5308 || pIdxKey
->default_rc
==-1
5311 xRecordCompare
= 0; /* All keys are integers */
5314 rc
= moveToRoot(pCur
);
5316 if( rc
==SQLITE_EMPTY
){
5317 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5323 assert( pCur
->pPage
);
5324 assert( pCur
->pPage
->isInit
);
5325 assert( pCur
->eState
==CURSOR_VALID
);
5326 assert( pCur
->pPage
->nCell
> 0 );
5327 assert( pCur
->iPage
==0 || pCur
->apPage
[0]->intKey
==pCur
->curIntKey
);
5328 assert( pCur
->curIntKey
|| pIdxKey
);
5330 int lwr
, upr
, idx
, c
;
5332 MemPage
*pPage
= pCur
->pPage
;
5333 u8
*pCell
; /* Pointer to current cell in pPage */
5335 /* pPage->nCell must be greater than zero. If this is the root-page
5336 ** the cursor would have been INVALID above and this for(;;) loop
5337 ** not run. If this is not the root-page, then the moveToChild() routine
5338 ** would have already detected db corruption. Similarly, pPage must
5339 ** be the right kind (index or table) of b-tree page. Otherwise
5340 ** a moveToChild() or moveToRoot() call would have detected corruption. */
5341 assert( pPage
->nCell
>0 );
5342 assert( pPage
->intKey
==(pIdxKey
==0) );
5344 upr
= pPage
->nCell
-1;
5345 assert( biasRight
==0 || biasRight
==1 );
5346 idx
= upr
>>(1-biasRight
); /* idx = biasRight ? upr : (lwr+upr)/2; */
5347 pCur
->ix
= (u16
)idx
;
5348 if( xRecordCompare
==0 ){
5351 pCell
= findCellPastPtr(pPage
, idx
);
5352 if( pPage
->intKeyLeaf
){
5353 while( 0x80 <= *(pCell
++) ){
5354 if( pCell
>=pPage
->aDataEnd
){
5355 return SQLITE_CORRUPT_PAGE(pPage
);
5359 getVarint(pCell
, (u64
*)&nCellKey
);
5360 if( nCellKey
<intKey
){
5362 if( lwr
>upr
){ c
= -1; break; }
5363 }else if( nCellKey
>intKey
){
5365 if( lwr
>upr
){ c
= +1; break; }
5367 assert( nCellKey
==intKey
);
5368 pCur
->ix
= (u16
)idx
;
5371 goto moveto_next_layer
;
5373 pCur
->curFlags
|= BTCF_ValidNKey
;
5374 pCur
->info
.nKey
= nCellKey
;
5375 pCur
->info
.nSize
= 0;
5380 assert( lwr
+upr
>=0 );
5381 idx
= (lwr
+upr
)>>1; /* idx = (lwr+upr)/2; */
5385 int nCell
; /* Size of the pCell cell in bytes */
5386 pCell
= findCellPastPtr(pPage
, idx
);
5388 /* The maximum supported page-size is 65536 bytes. This means that
5389 ** the maximum number of record bytes stored on an index B-Tree
5390 ** page is less than 16384 bytes and may be stored as a 2-byte
5391 ** varint. This information is used to attempt to avoid parsing
5392 ** the entire cell by checking for the cases where the record is
5393 ** stored entirely within the b-tree page by inspecting the first
5394 ** 2 bytes of the cell.
5397 if( nCell
<=pPage
->max1bytePayload
){
5398 /* This branch runs if the record-size field of the cell is a
5399 ** single byte varint and the record fits entirely on the main
5401 testcase( pCell
+nCell
+1==pPage
->aDataEnd
);
5402 c
= xRecordCompare(nCell
, (void*)&pCell
[1], pIdxKey
);
5403 }else if( !(pCell
[1] & 0x80)
5404 && (nCell
= ((nCell
&0x7f)<<7) + pCell
[1])<=pPage
->maxLocal
5406 /* The record-size field is a 2 byte varint and the record
5407 ** fits entirely on the main b-tree page. */
5408 testcase( pCell
+nCell
+2==pPage
->aDataEnd
);
5409 c
= xRecordCompare(nCell
, (void*)&pCell
[2], pIdxKey
);
5411 /* The record flows over onto one or more overflow pages. In
5412 ** this case the whole cell needs to be parsed, a buffer allocated
5413 ** and accessPayload() used to retrieve the record into the
5414 ** buffer before VdbeRecordCompare() can be called.
5416 ** If the record is corrupt, the xRecordCompare routine may read
5417 ** up to two varints past the end of the buffer. An extra 18
5418 ** bytes of padding is allocated at the end of the buffer in
5419 ** case this happens. */
5421 u8
* const pCellBody
= pCell
- pPage
->childPtrSize
;
5422 pPage
->xParseCell(pPage
, pCellBody
, &pCur
->info
);
5423 nCell
= (int)pCur
->info
.nKey
;
5424 testcase( nCell
<0 ); /* True if key size is 2^32 or more */
5425 testcase( nCell
==0 ); /* Invalid key size: 0x80 0x80 0x00 */
5426 testcase( nCell
==1 ); /* Invalid key size: 0x80 0x80 0x01 */
5427 testcase( nCell
==2 ); /* Minimum legal index key size */
5429 rc
= SQLITE_CORRUPT_PAGE(pPage
);
5432 pCellKey
= sqlite3Malloc( nCell
+18 );
5434 rc
= SQLITE_NOMEM_BKPT
;
5437 pCur
->ix
= (u16
)idx
;
5438 rc
= accessPayload(pCur
, 0, nCell
, (unsigned char*)pCellKey
, 0);
5439 pCur
->curFlags
&= ~BTCF_ValidOvfl
;
5441 sqlite3_free(pCellKey
);
5444 c
= xRecordCompare(nCell
, pCellKey
, pIdxKey
);
5445 sqlite3_free(pCellKey
);
5448 (pIdxKey
->errCode
!=SQLITE_CORRUPT
|| c
==0)
5449 && (pIdxKey
->errCode
!=SQLITE_NOMEM
|| pCur
->pBtree
->db
->mallocFailed
)
5459 pCur
->ix
= (u16
)idx
;
5460 if( pIdxKey
->errCode
) rc
= SQLITE_CORRUPT_BKPT
;
5463 if( lwr
>upr
) break;
5464 assert( lwr
+upr
>=0 );
5465 idx
= (lwr
+upr
)>>1; /* idx = (lwr+upr)/2 */
5468 assert( lwr
==upr
+1 || (pPage
->intKey
&& !pPage
->leaf
) );
5469 assert( pPage
->isInit
);
5471 assert( pCur
->ix
<pCur
->pPage
->nCell
);
5472 pCur
->ix
= (u16
)idx
;
5478 if( lwr
>=pPage
->nCell
){
5479 chldPg
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
5481 chldPg
= get4byte(findCell(pPage
, lwr
));
5483 pCur
->ix
= (u16
)lwr
;
5484 rc
= moveToChild(pCur
, chldPg
);
5488 pCur
->info
.nSize
= 0;
5489 assert( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 );
5495 ** Return TRUE if the cursor is not pointing at an entry of the table.
5497 ** TRUE will be returned after a call to sqlite3BtreeNext() moves
5498 ** past the last entry in the table or sqlite3BtreePrev() moves past
5499 ** the first entry. TRUE is also returned if the table is empty.
5501 int sqlite3BtreeEof(BtCursor
*pCur
){
5502 /* TODO: What if the cursor is in CURSOR_REQUIRESEEK but all table entries
5503 ** have been deleted? This API will need to change to return an error code
5504 ** as well as the boolean result value.
5506 return (CURSOR_VALID
!=pCur
->eState
);
5510 ** Return an estimate for the number of rows in the table that pCur is
5511 ** pointing to. Return a negative number if no estimate is currently
5514 i64
sqlite3BtreeRowCountEst(BtCursor
*pCur
){
5518 assert( cursorOwnsBtShared(pCur
) );
5519 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5521 /* Currently this interface is only called by the OP_IfSmaller
5522 ** opcode, and it that case the cursor will always be valid and
5523 ** will always point to a leaf node. */
5524 if( NEVER(pCur
->eState
!=CURSOR_VALID
) ) return -1;
5525 if( NEVER(pCur
->pPage
->leaf
==0) ) return -1;
5527 n
= pCur
->pPage
->nCell
;
5528 for(i
=0; i
<pCur
->iPage
; i
++){
5529 n
*= pCur
->apPage
[i
]->nCell
;
5535 ** Advance the cursor to the next entry in the database.
5538 ** SQLITE_OK success
5539 ** SQLITE_DONE cursor is already pointing at the last element
5540 ** otherwise some kind of error occurred
5542 ** The main entry point is sqlite3BtreeNext(). That routine is optimized
5543 ** for the common case of merely incrementing the cell counter BtCursor.aiIdx
5544 ** to the next cell on the current page. The (slower) btreeNext() helper
5545 ** routine is called when it is necessary to move to a different page or
5546 ** to restore the cursor.
5548 ** If bit 0x01 of the F argument in sqlite3BtreeNext(C,F) is 1, then the
5549 ** cursor corresponds to an SQL index and this routine could have been
5550 ** skipped if the SQL index had been a unique index. The F argument
5551 ** is a hint to the implement. SQLite btree implementation does not use
5552 ** this hint, but COMDB2 does.
5554 static SQLITE_NOINLINE
int btreeNext(BtCursor
*pCur
){
5559 assert( cursorOwnsBtShared(pCur
) );
5560 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5561 if( pCur
->eState
!=CURSOR_VALID
){
5562 assert( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 );
5563 rc
= restoreCursorPosition(pCur
);
5564 if( rc
!=SQLITE_OK
){
5567 if( CURSOR_INVALID
==pCur
->eState
){
5570 if( pCur
->skipNext
){
5571 assert( pCur
->eState
==CURSOR_VALID
|| pCur
->eState
==CURSOR_SKIPNEXT
);
5572 pCur
->eState
= CURSOR_VALID
;
5573 if( pCur
->skipNext
>0 ){
5581 pPage
= pCur
->pPage
;
5583 assert( pPage
->isInit
);
5585 /* If the database file is corrupt, it is possible for the value of idx
5586 ** to be invalid here. This can only occur if a second cursor modifies
5587 ** the page while cursor pCur is holding a reference to it. Which can
5588 ** only happen if the database is corrupt in such a way as to link the
5589 ** page into more than one b-tree structure. */
5590 testcase( idx
>pPage
->nCell
);
5592 if( idx
>=pPage
->nCell
){
5594 rc
= moveToChild(pCur
, get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]));
5596 return moveToLeftmost(pCur
);
5599 if( pCur
->iPage
==0 ){
5600 pCur
->eState
= CURSOR_INVALID
;
5604 pPage
= pCur
->pPage
;
5605 }while( pCur
->ix
>=pPage
->nCell
);
5606 if( pPage
->intKey
){
5607 return sqlite3BtreeNext(pCur
, 0);
5615 return moveToLeftmost(pCur
);
5618 int sqlite3BtreeNext(BtCursor
*pCur
, int flags
){
5620 UNUSED_PARAMETER( flags
); /* Used in COMDB2 but not native SQLite */
5621 assert( cursorOwnsBtShared(pCur
) );
5622 assert( flags
==0 || flags
==1 );
5623 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5624 pCur
->info
.nSize
= 0;
5625 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
5626 if( pCur
->eState
!=CURSOR_VALID
) return btreeNext(pCur
);
5627 pPage
= pCur
->pPage
;
5628 if( (++pCur
->ix
)>=pPage
->nCell
){
5630 return btreeNext(pCur
);
5635 return moveToLeftmost(pCur
);
5640 ** Step the cursor to the back to the previous entry in the database.
5643 ** SQLITE_OK success
5644 ** SQLITE_DONE the cursor is already on the first element of the table
5645 ** otherwise some kind of error occurred
5647 ** The main entry point is sqlite3BtreePrevious(). That routine is optimized
5648 ** for the common case of merely decrementing the cell counter BtCursor.aiIdx
5649 ** to the previous cell on the current page. The (slower) btreePrevious()
5650 ** helper routine is called when it is necessary to move to a different page
5651 ** or to restore the cursor.
5653 ** If bit 0x01 of the F argument to sqlite3BtreePrevious(C,F) is 1, then
5654 ** the cursor corresponds to an SQL index and this routine could have been
5655 ** skipped if the SQL index had been a unique index. The F argument is a
5656 ** hint to the implement. The native SQLite btree implementation does not
5657 ** use this hint, but COMDB2 does.
5659 static SQLITE_NOINLINE
int btreePrevious(BtCursor
*pCur
){
5663 assert( cursorOwnsBtShared(pCur
) );
5664 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5665 assert( (pCur
->curFlags
& (BTCF_AtLast
|BTCF_ValidOvfl
|BTCF_ValidNKey
))==0 );
5666 assert( pCur
->info
.nSize
==0 );
5667 if( pCur
->eState
!=CURSOR_VALID
){
5668 rc
= restoreCursorPosition(pCur
);
5669 if( rc
!=SQLITE_OK
){
5672 if( CURSOR_INVALID
==pCur
->eState
){
5675 if( pCur
->skipNext
){
5676 assert( pCur
->eState
==CURSOR_VALID
|| pCur
->eState
==CURSOR_SKIPNEXT
);
5677 pCur
->eState
= CURSOR_VALID
;
5678 if( pCur
->skipNext
<0 ){
5686 pPage
= pCur
->pPage
;
5687 assert( pPage
->isInit
);
5690 rc
= moveToChild(pCur
, get4byte(findCell(pPage
, idx
)));
5692 rc
= moveToRightmost(pCur
);
5694 while( pCur
->ix
==0 ){
5695 if( pCur
->iPage
==0 ){
5696 pCur
->eState
= CURSOR_INVALID
;
5701 assert( pCur
->info
.nSize
==0 );
5702 assert( (pCur
->curFlags
& (BTCF_ValidOvfl
))==0 );
5705 pPage
= pCur
->pPage
;
5706 if( pPage
->intKey
&& !pPage
->leaf
){
5707 rc
= sqlite3BtreePrevious(pCur
, 0);
5714 int sqlite3BtreePrevious(BtCursor
*pCur
, int flags
){
5715 assert( cursorOwnsBtShared(pCur
) );
5716 assert( flags
==0 || flags
==1 );
5717 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5718 UNUSED_PARAMETER( flags
); /* Used in COMDB2 but not native SQLite */
5719 pCur
->curFlags
&= ~(BTCF_AtLast
|BTCF_ValidOvfl
|BTCF_ValidNKey
);
5720 pCur
->info
.nSize
= 0;
5721 if( pCur
->eState
!=CURSOR_VALID
5723 || pCur
->pPage
->leaf
==0
5725 return btreePrevious(pCur
);
5732 ** Allocate a new page from the database file.
5734 ** The new page is marked as dirty. (In other words, sqlite3PagerWrite()
5735 ** has already been called on the new page.) The new page has also
5736 ** been referenced and the calling routine is responsible for calling
5737 ** sqlite3PagerUnref() on the new page when it is done.
5739 ** SQLITE_OK is returned on success. Any other return value indicates
5740 ** an error. *ppPage is set to NULL in the event of an error.
5742 ** If the "nearby" parameter is not 0, then an effort is made to
5743 ** locate a page close to the page number "nearby". This can be used in an
5744 ** attempt to keep related pages close to each other in the database file,
5745 ** which in turn can make database access faster.
5747 ** If the eMode parameter is BTALLOC_EXACT and the nearby page exists
5748 ** anywhere on the free-list, then it is guaranteed to be returned. If
5749 ** eMode is BTALLOC_LT then the page returned will be less than or equal
5750 ** to nearby if any such page exists. If eMode is BTALLOC_ANY then there
5751 ** are no restrictions on which page is returned.
5753 static int allocateBtreePage(
5754 BtShared
*pBt
, /* The btree */
5755 MemPage
**ppPage
, /* Store pointer to the allocated page here */
5756 Pgno
*pPgno
, /* Store the page number here */
5757 Pgno nearby
, /* Search for a page near this one */
5758 u8 eMode
/* BTALLOC_EXACT, BTALLOC_LT, or BTALLOC_ANY */
5762 u32 n
; /* Number of pages on the freelist */
5763 u32 k
; /* Number of leaves on the trunk of the freelist */
5764 MemPage
*pTrunk
= 0;
5765 MemPage
*pPrevTrunk
= 0;
5766 Pgno mxPage
; /* Total size of the database file */
5768 assert( sqlite3_mutex_held(pBt
->mutex
) );
5769 assert( eMode
==BTALLOC_ANY
|| (nearby
>0 && IfNotOmitAV(pBt
->autoVacuum
)) );
5770 pPage1
= pBt
->pPage1
;
5771 mxPage
= btreePagecount(pBt
);
5772 /* EVIDENCE-OF: R-05119-02637 The 4-byte big-endian integer at offset 36
5773 ** stores stores the total number of pages on the freelist. */
5774 n
= get4byte(&pPage1
->aData
[36]);
5775 testcase( n
==mxPage
-1 );
5777 return SQLITE_CORRUPT_BKPT
;
5780 /* There are pages on the freelist. Reuse one of those pages. */
5782 u8 searchList
= 0; /* If the free-list must be searched for 'nearby' */
5783 u32 nSearch
= 0; /* Count of the number of search attempts */
5785 /* If eMode==BTALLOC_EXACT and a query of the pointer-map
5786 ** shows that the page 'nearby' is somewhere on the free-list, then
5787 ** the entire-list will be searched for that page.
5789 #ifndef SQLITE_OMIT_AUTOVACUUM
5790 if( eMode
==BTALLOC_EXACT
){
5791 if( nearby
<=mxPage
){
5794 assert( pBt
->autoVacuum
);
5795 rc
= ptrmapGet(pBt
, nearby
, &eType
, 0);
5797 if( eType
==PTRMAP_FREEPAGE
){
5801 }else if( eMode
==BTALLOC_LE
){
5806 /* Decrement the free-list count by 1. Set iTrunk to the index of the
5807 ** first free-list trunk page. iPrevTrunk is initially 1.
5809 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
5811 put4byte(&pPage1
->aData
[36], n
-1);
5813 /* The code within this loop is run only once if the 'searchList' variable
5814 ** is not true. Otherwise, it runs once for each trunk-page on the
5815 ** free-list until the page 'nearby' is located (eMode==BTALLOC_EXACT)
5816 ** or until a page less than 'nearby' is located (eMode==BTALLOC_LT)
5819 pPrevTrunk
= pTrunk
;
5821 /* EVIDENCE-OF: R-01506-11053 The first integer on a freelist trunk page
5822 ** is the page number of the next freelist trunk page in the list or
5823 ** zero if this is the last freelist trunk page. */
5824 iTrunk
= get4byte(&pPrevTrunk
->aData
[0]);
5826 /* EVIDENCE-OF: R-59841-13798 The 4-byte big-endian integer at offset 32
5827 ** stores the page number of the first page of the freelist, or zero if
5828 ** the freelist is empty. */
5829 iTrunk
= get4byte(&pPage1
->aData
[32]);
5831 testcase( iTrunk
==mxPage
);
5832 if( iTrunk
>mxPage
|| nSearch
++ > n
){
5833 rc
= SQLITE_CORRUPT_PGNO(pPrevTrunk
? pPrevTrunk
->pgno
: 1);
5835 rc
= btreeGetUnusedPage(pBt
, iTrunk
, &pTrunk
, 0);
5839 goto end_allocate_page
;
5841 assert( pTrunk
!=0 );
5842 assert( pTrunk
->aData
!=0 );
5843 /* EVIDENCE-OF: R-13523-04394 The second integer on a freelist trunk page
5844 ** is the number of leaf page pointers to follow. */
5845 k
= get4byte(&pTrunk
->aData
[4]);
5846 if( k
==0 && !searchList
){
5847 /* The trunk has no leaves and the list is not being searched.
5848 ** So extract the trunk page itself and use it as the newly
5849 ** allocated page */
5850 assert( pPrevTrunk
==0 );
5851 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
5853 goto end_allocate_page
;
5856 memcpy(&pPage1
->aData
[32], &pTrunk
->aData
[0], 4);
5859 TRACE(("ALLOCATE: %d trunk - %d free pages left\n", *pPgno
, n
-1));
5860 }else if( k
>(u32
)(pBt
->usableSize
/4 - 2) ){
5861 /* Value of k is out of range. Database corruption */
5862 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
5863 goto end_allocate_page
;
5864 #ifndef SQLITE_OMIT_AUTOVACUUM
5865 }else if( searchList
5866 && (nearby
==iTrunk
|| (iTrunk
<nearby
&& eMode
==BTALLOC_LE
))
5868 /* The list is being searched and this trunk page is the page
5869 ** to allocate, regardless of whether it has leaves.
5874 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
5876 goto end_allocate_page
;
5880 memcpy(&pPage1
->aData
[32], &pTrunk
->aData
[0], 4);
5882 rc
= sqlite3PagerWrite(pPrevTrunk
->pDbPage
);
5883 if( rc
!=SQLITE_OK
){
5884 goto end_allocate_page
;
5886 memcpy(&pPrevTrunk
->aData
[0], &pTrunk
->aData
[0], 4);
5889 /* The trunk page is required by the caller but it contains
5890 ** pointers to free-list leaves. The first leaf becomes a trunk
5891 ** page in this case.
5894 Pgno iNewTrunk
= get4byte(&pTrunk
->aData
[8]);
5895 if( iNewTrunk
>mxPage
){
5896 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
5897 goto end_allocate_page
;
5899 testcase( iNewTrunk
==mxPage
);
5900 rc
= btreeGetUnusedPage(pBt
, iNewTrunk
, &pNewTrunk
, 0);
5901 if( rc
!=SQLITE_OK
){
5902 goto end_allocate_page
;
5904 rc
= sqlite3PagerWrite(pNewTrunk
->pDbPage
);
5905 if( rc
!=SQLITE_OK
){
5906 releasePage(pNewTrunk
);
5907 goto end_allocate_page
;
5909 memcpy(&pNewTrunk
->aData
[0], &pTrunk
->aData
[0], 4);
5910 put4byte(&pNewTrunk
->aData
[4], k
-1);
5911 memcpy(&pNewTrunk
->aData
[8], &pTrunk
->aData
[12], (k
-1)*4);
5912 releasePage(pNewTrunk
);
5914 assert( sqlite3PagerIswriteable(pPage1
->pDbPage
) );
5915 put4byte(&pPage1
->aData
[32], iNewTrunk
);
5917 rc
= sqlite3PagerWrite(pPrevTrunk
->pDbPage
);
5919 goto end_allocate_page
;
5921 put4byte(&pPrevTrunk
->aData
[0], iNewTrunk
);
5925 TRACE(("ALLOCATE: %d trunk - %d free pages left\n", *pPgno
, n
-1));
5928 /* Extract a leaf from the trunk */
5931 unsigned char *aData
= pTrunk
->aData
;
5935 if( eMode
==BTALLOC_LE
){
5937 iPage
= get4byte(&aData
[8+i
*4]);
5938 if( iPage
<=nearby
){
5945 dist
= sqlite3AbsInt32(get4byte(&aData
[8]) - nearby
);
5947 int d2
= sqlite3AbsInt32(get4byte(&aData
[8+i
*4]) - nearby
);
5958 iPage
= get4byte(&aData
[8+closest
*4]);
5959 testcase( iPage
==mxPage
);
5961 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
5962 goto end_allocate_page
;
5964 testcase( iPage
==mxPage
);
5966 || (iPage
==nearby
|| (iPage
<nearby
&& eMode
==BTALLOC_LE
))
5970 TRACE(("ALLOCATE: %d was leaf %d of %d on trunk %d"
5971 ": %d more free pages\n",
5972 *pPgno
, closest
+1, k
, pTrunk
->pgno
, n
-1));
5973 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
5974 if( rc
) goto end_allocate_page
;
5976 memcpy(&aData
[8+closest
*4], &aData
[4+k
*4], 4);
5978 put4byte(&aData
[4], k
-1);
5979 noContent
= !btreeGetHasContent(pBt
, *pPgno
)? PAGER_GET_NOCONTENT
: 0;
5980 rc
= btreeGetUnusedPage(pBt
, *pPgno
, ppPage
, noContent
);
5981 if( rc
==SQLITE_OK
){
5982 rc
= sqlite3PagerWrite((*ppPage
)->pDbPage
);
5983 if( rc
!=SQLITE_OK
){
5984 releasePage(*ppPage
);
5991 releasePage(pPrevTrunk
);
5993 }while( searchList
);
5995 /* There are no pages on the freelist, so append a new page to the
5998 ** Normally, new pages allocated by this block can be requested from the
5999 ** pager layer with the 'no-content' flag set. This prevents the pager
6000 ** from trying to read the pages content from disk. However, if the
6001 ** current transaction has already run one or more incremental-vacuum
6002 ** steps, then the page we are about to allocate may contain content
6003 ** that is required in the event of a rollback. In this case, do
6004 ** not set the no-content flag. This causes the pager to load and journal
6005 ** the current page content before overwriting it.
6007 ** Note that the pager will not actually attempt to load or journal
6008 ** content for any page that really does lie past the end of the database
6009 ** file on disk. So the effects of disabling the no-content optimization
6010 ** here are confined to those pages that lie between the end of the
6011 ** database image and the end of the database file.
6013 int bNoContent
= (0==IfNotOmitAV(pBt
->bDoTruncate
))? PAGER_GET_NOCONTENT
:0;
6015 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
6018 if( pBt
->nPage
==PENDING_BYTE_PAGE(pBt
) ) pBt
->nPage
++;
6020 #ifndef SQLITE_OMIT_AUTOVACUUM
6021 if( pBt
->autoVacuum
&& PTRMAP_ISPAGE(pBt
, pBt
->nPage
) ){
6022 /* If *pPgno refers to a pointer-map page, allocate two new pages
6023 ** at the end of the file instead of one. The first allocated page
6024 ** becomes a new pointer-map page, the second is used by the caller.
6027 TRACE(("ALLOCATE: %d from end of file (pointer-map page)\n", pBt
->nPage
));
6028 assert( pBt
->nPage
!=PENDING_BYTE_PAGE(pBt
) );
6029 rc
= btreeGetUnusedPage(pBt
, pBt
->nPage
, &pPg
, bNoContent
);
6030 if( rc
==SQLITE_OK
){
6031 rc
= sqlite3PagerWrite(pPg
->pDbPage
);
6036 if( pBt
->nPage
==PENDING_BYTE_PAGE(pBt
) ){ pBt
->nPage
++; }
6039 put4byte(28 + (u8
*)pBt
->pPage1
->aData
, pBt
->nPage
);
6040 *pPgno
= pBt
->nPage
;
6042 assert( *pPgno
!=PENDING_BYTE_PAGE(pBt
) );
6043 rc
= btreeGetUnusedPage(pBt
, *pPgno
, ppPage
, bNoContent
);
6045 rc
= sqlite3PagerWrite((*ppPage
)->pDbPage
);
6046 if( rc
!=SQLITE_OK
){
6047 releasePage(*ppPage
);
6050 TRACE(("ALLOCATE: %d from end of file\n", *pPgno
));
6053 assert( *pPgno
!=PENDING_BYTE_PAGE(pBt
) );
6056 releasePage(pTrunk
);
6057 releasePage(pPrevTrunk
);
6058 assert( rc
!=SQLITE_OK
|| sqlite3PagerPageRefcount((*ppPage
)->pDbPage
)<=1 );
6059 assert( rc
!=SQLITE_OK
|| (*ppPage
)->isInit
==0 );
6064 ** This function is used to add page iPage to the database file free-list.
6065 ** It is assumed that the page is not already a part of the free-list.
6067 ** The value passed as the second argument to this function is optional.
6068 ** If the caller happens to have a pointer to the MemPage object
6069 ** corresponding to page iPage handy, it may pass it as the second value.
6070 ** Otherwise, it may pass NULL.
6072 ** If a pointer to a MemPage object is passed as the second argument,
6073 ** its reference count is not altered by this function.
6075 static int freePage2(BtShared
*pBt
, MemPage
*pMemPage
, Pgno iPage
){
6076 MemPage
*pTrunk
= 0; /* Free-list trunk page */
6077 Pgno iTrunk
= 0; /* Page number of free-list trunk page */
6078 MemPage
*pPage1
= pBt
->pPage1
; /* Local reference to page 1 */
6079 MemPage
*pPage
; /* Page being freed. May be NULL. */
6080 int rc
; /* Return Code */
6081 int nFree
; /* Initial number of pages on free-list */
6083 assert( sqlite3_mutex_held(pBt
->mutex
) );
6084 assert( CORRUPT_DB
|| iPage
>1 );
6085 assert( !pMemPage
|| pMemPage
->pgno
==iPage
);
6087 if( iPage
<2 ) return SQLITE_CORRUPT_BKPT
;
6090 sqlite3PagerRef(pPage
->pDbPage
);
6092 pPage
= btreePageLookup(pBt
, iPage
);
6095 /* Increment the free page count on pPage1 */
6096 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
6097 if( rc
) goto freepage_out
;
6098 nFree
= get4byte(&pPage1
->aData
[36]);
6099 put4byte(&pPage1
->aData
[36], nFree
+1);
6101 if( pBt
->btsFlags
& BTS_SECURE_DELETE
){
6102 /* If the secure_delete option is enabled, then
6103 ** always fully overwrite deleted information with zeros.
6105 if( (!pPage
&& ((rc
= btreeGetPage(pBt
, iPage
, &pPage
, 0))!=0) )
6106 || ((rc
= sqlite3PagerWrite(pPage
->pDbPage
))!=0)
6110 memset(pPage
->aData
, 0, pPage
->pBt
->pageSize
);
6113 /* If the database supports auto-vacuum, write an entry in the pointer-map
6114 ** to indicate that the page is free.
6117 ptrmapPut(pBt
, iPage
, PTRMAP_FREEPAGE
, 0, &rc
);
6118 if( rc
) goto freepage_out
;
6121 /* Now manipulate the actual database free-list structure. There are two
6122 ** possibilities. If the free-list is currently empty, or if the first
6123 ** trunk page in the free-list is full, then this page will become a
6124 ** new free-list trunk page. Otherwise, it will become a leaf of the
6125 ** first trunk page in the current free-list. This block tests if it
6126 ** is possible to add the page as a new free-list leaf.
6129 u32 nLeaf
; /* Initial number of leaf cells on trunk page */
6131 iTrunk
= get4byte(&pPage1
->aData
[32]);
6132 rc
= btreeGetPage(pBt
, iTrunk
, &pTrunk
, 0);
6133 if( rc
!=SQLITE_OK
){
6137 nLeaf
= get4byte(&pTrunk
->aData
[4]);
6138 assert( pBt
->usableSize
>32 );
6139 if( nLeaf
> (u32
)pBt
->usableSize
/4 - 2 ){
6140 rc
= SQLITE_CORRUPT_BKPT
;
6143 if( nLeaf
< (u32
)pBt
->usableSize
/4 - 8 ){
6144 /* In this case there is room on the trunk page to insert the page
6145 ** being freed as a new leaf.
6147 ** Note that the trunk page is not really full until it contains
6148 ** usableSize/4 - 2 entries, not usableSize/4 - 8 entries as we have
6149 ** coded. But due to a coding error in versions of SQLite prior to
6150 ** 3.6.0, databases with freelist trunk pages holding more than
6151 ** usableSize/4 - 8 entries will be reported as corrupt. In order
6152 ** to maintain backwards compatibility with older versions of SQLite,
6153 ** we will continue to restrict the number of entries to usableSize/4 - 8
6154 ** for now. At some point in the future (once everyone has upgraded
6155 ** to 3.6.0 or later) we should consider fixing the conditional above
6156 ** to read "usableSize/4-2" instead of "usableSize/4-8".
6158 ** EVIDENCE-OF: R-19920-11576 However, newer versions of SQLite still
6159 ** avoid using the last six entries in the freelist trunk page array in
6160 ** order that database files created by newer versions of SQLite can be
6161 ** read by older versions of SQLite.
6163 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
6164 if( rc
==SQLITE_OK
){
6165 put4byte(&pTrunk
->aData
[4], nLeaf
+1);
6166 put4byte(&pTrunk
->aData
[8+nLeaf
*4], iPage
);
6167 if( pPage
&& (pBt
->btsFlags
& BTS_SECURE_DELETE
)==0 ){
6168 sqlite3PagerDontWrite(pPage
->pDbPage
);
6170 rc
= btreeSetHasContent(pBt
, iPage
);
6172 TRACE(("FREE-PAGE: %d leaf on trunk page %d\n",pPage
->pgno
,pTrunk
->pgno
));
6177 /* If control flows to this point, then it was not possible to add the
6178 ** the page being freed as a leaf page of the first trunk in the free-list.
6179 ** Possibly because the free-list is empty, or possibly because the
6180 ** first trunk in the free-list is full. Either way, the page being freed
6181 ** will become the new first trunk page in the free-list.
6183 if( pPage
==0 && SQLITE_OK
!=(rc
= btreeGetPage(pBt
, iPage
, &pPage
, 0)) ){
6186 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
6187 if( rc
!=SQLITE_OK
){
6190 put4byte(pPage
->aData
, iTrunk
);
6191 put4byte(&pPage
->aData
[4], 0);
6192 put4byte(&pPage1
->aData
[32], iPage
);
6193 TRACE(("FREE-PAGE: %d new trunk page replacing %d\n", pPage
->pgno
, iTrunk
));
6200 releasePage(pTrunk
);
6203 static void freePage(MemPage
*pPage
, int *pRC
){
6204 if( (*pRC
)==SQLITE_OK
){
6205 *pRC
= freePage2(pPage
->pBt
, pPage
, pPage
->pgno
);
6210 ** Free any overflow pages associated with the given Cell. Store
6211 ** size information about the cell in pInfo.
6213 static int clearCell(
6214 MemPage
*pPage
, /* The page that contains the Cell */
6215 unsigned char *pCell
, /* First byte of the Cell */
6216 CellInfo
*pInfo
/* Size information about the cell */
6224 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6225 pPage
->xParseCell(pPage
, pCell
, pInfo
);
6226 if( pInfo
->nLocal
==pInfo
->nPayload
){
6227 return SQLITE_OK
; /* No overflow pages. Return without doing anything */
6229 if( pCell
+pInfo
->nSize
-1 > pPage
->aData
+pPage
->maskPage
){
6230 /* Cell extends past end of page */
6231 return SQLITE_CORRUPT_PAGE(pPage
);
6233 ovflPgno
= get4byte(pCell
+ pInfo
->nSize
- 4);
6235 assert( pBt
->usableSize
> 4 );
6236 ovflPageSize
= pBt
->usableSize
- 4;
6237 nOvfl
= (pInfo
->nPayload
- pInfo
->nLocal
+ ovflPageSize
- 1)/ovflPageSize
;
6239 (CORRUPT_DB
&& (pInfo
->nPayload
+ ovflPageSize
)<ovflPageSize
)
6244 if( ovflPgno
<2 || ovflPgno
>btreePagecount(pBt
) ){
6245 /* 0 is not a legal page number and page 1 cannot be an
6246 ** overflow page. Therefore if ovflPgno<2 or past the end of the
6247 ** file the database must be corrupt. */
6248 return SQLITE_CORRUPT_BKPT
;
6251 rc
= getOverflowPage(pBt
, ovflPgno
, &pOvfl
, &iNext
);
6255 if( ( pOvfl
|| ((pOvfl
= btreePageLookup(pBt
, ovflPgno
))!=0) )
6256 && sqlite3PagerPageRefcount(pOvfl
->pDbPage
)!=1
6258 /* There is no reason any cursor should have an outstanding reference
6259 ** to an overflow page belonging to a cell that is being deleted/updated.
6260 ** So if there exists more than one reference to this page, then it
6261 ** must not really be an overflow page and the database must be corrupt.
6262 ** It is helpful to detect this before calling freePage2(), as
6263 ** freePage2() may zero the page contents if secure-delete mode is
6264 ** enabled. If this 'overflow' page happens to be a page that the
6265 ** caller is iterating through or using in some other way, this
6266 ** can be problematic.
6268 rc
= SQLITE_CORRUPT_BKPT
;
6270 rc
= freePage2(pBt
, pOvfl
, ovflPgno
);
6274 sqlite3PagerUnref(pOvfl
->pDbPage
);
6283 ** Create the byte sequence used to represent a cell on page pPage
6284 ** and write that byte sequence into pCell[]. Overflow pages are
6285 ** allocated and filled in as necessary. The calling procedure
6286 ** is responsible for making sure sufficient space has been allocated
6289 ** Note that pCell does not necessary need to point to the pPage->aData
6290 ** area. pCell might point to some temporary storage. The cell will
6291 ** be constructed in this temporary area then copied into pPage->aData
6294 static int fillInCell(
6295 MemPage
*pPage
, /* The page that contains the cell */
6296 unsigned char *pCell
, /* Complete text of the cell */
6297 const BtreePayload
*pX
, /* Payload with which to construct the cell */
6298 int *pnSize
/* Write cell size here */
6302 int nSrc
, n
, rc
, mn
;
6304 MemPage
*pToRelease
;
6305 unsigned char *pPrior
;
6306 unsigned char *pPayload
;
6311 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6313 /* pPage is not necessarily writeable since pCell might be auxiliary
6314 ** buffer space that is separate from the pPage buffer area */
6315 assert( pCell
<pPage
->aData
|| pCell
>=&pPage
->aData
[pPage
->pBt
->pageSize
]
6316 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6318 /* Fill in the header. */
6319 nHeader
= pPage
->childPtrSize
;
6320 if( pPage
->intKey
){
6321 nPayload
= pX
->nData
+ pX
->nZero
;
6324 assert( pPage
->intKeyLeaf
); /* fillInCell() only called for leaves */
6325 nHeader
+= putVarint32(&pCell
[nHeader
], nPayload
);
6326 nHeader
+= putVarint(&pCell
[nHeader
], *(u64
*)&pX
->nKey
);
6328 assert( pX
->nKey
<=0x7fffffff && pX
->pKey
!=0 );
6329 nSrc
= nPayload
= (int)pX
->nKey
;
6331 nHeader
+= putVarint32(&pCell
[nHeader
], nPayload
);
6334 /* Fill in the payload */
6335 pPayload
= &pCell
[nHeader
];
6336 if( nPayload
<=pPage
->maxLocal
){
6337 /* This is the common case where everything fits on the btree page
6338 ** and no overflow pages are required. */
6339 n
= nHeader
+ nPayload
;
6344 assert( nSrc
<=nPayload
);
6345 testcase( nSrc
<nPayload
);
6346 memcpy(pPayload
, pSrc
, nSrc
);
6347 memset(pPayload
+nSrc
, 0, nPayload
-nSrc
);
6351 /* If we reach this point, it means that some of the content will need
6352 ** to spill onto overflow pages.
6354 mn
= pPage
->minLocal
;
6355 n
= mn
+ (nPayload
- mn
) % (pPage
->pBt
->usableSize
- 4);
6356 testcase( n
==pPage
->maxLocal
);
6357 testcase( n
==pPage
->maxLocal
+1 );
6358 if( n
> pPage
->maxLocal
) n
= mn
;
6360 *pnSize
= n
+ nHeader
+ 4;
6361 pPrior
= &pCell
[nHeader
+n
];
6366 /* At this point variables should be set as follows:
6368 ** nPayload Total payload size in bytes
6369 ** pPayload Begin writing payload here
6370 ** spaceLeft Space available at pPayload. If nPayload>spaceLeft,
6371 ** that means content must spill into overflow pages.
6372 ** *pnSize Size of the local cell (not counting overflow pages)
6373 ** pPrior Where to write the pgno of the first overflow page
6375 ** Use a call to btreeParseCellPtr() to verify that the values above
6376 ** were computed correctly.
6381 pPage
->xParseCell(pPage
, pCell
, &info
);
6382 assert( nHeader
==(int)(info
.pPayload
- pCell
) );
6383 assert( info
.nKey
==pX
->nKey
);
6384 assert( *pnSize
== info
.nSize
);
6385 assert( spaceLeft
== info
.nLocal
);
6389 /* Write the payload into the local Cell and any extra into overflow pages */
6392 if( n
>spaceLeft
) n
= spaceLeft
;
6394 /* If pToRelease is not zero than pPayload points into the data area
6395 ** of pToRelease. Make sure pToRelease is still writeable. */
6396 assert( pToRelease
==0 || sqlite3PagerIswriteable(pToRelease
->pDbPage
) );
6398 /* If pPayload is part of the data area of pPage, then make sure pPage
6399 ** is still writeable */
6400 assert( pPayload
<pPage
->aData
|| pPayload
>=&pPage
->aData
[pBt
->pageSize
]
6401 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6404 memcpy(pPayload
, pSrc
, n
);
6407 memcpy(pPayload
, pSrc
, n
);
6409 memset(pPayload
, 0, n
);
6412 if( nPayload
<=0 ) break;
6419 #ifndef SQLITE_OMIT_AUTOVACUUM
6420 Pgno pgnoPtrmap
= pgnoOvfl
; /* Overflow page pointer-map entry page */
6421 if( pBt
->autoVacuum
){
6425 PTRMAP_ISPAGE(pBt
, pgnoOvfl
) || pgnoOvfl
==PENDING_BYTE_PAGE(pBt
)
6429 rc
= allocateBtreePage(pBt
, &pOvfl
, &pgnoOvfl
, pgnoOvfl
, 0);
6430 #ifndef SQLITE_OMIT_AUTOVACUUM
6431 /* If the database supports auto-vacuum, and the second or subsequent
6432 ** overflow page is being allocated, add an entry to the pointer-map
6433 ** for that page now.
6435 ** If this is the first overflow page, then write a partial entry
6436 ** to the pointer-map. If we write nothing to this pointer-map slot,
6437 ** then the optimistic overflow chain processing in clearCell()
6438 ** may misinterpret the uninitialized values and delete the
6439 ** wrong pages from the database.
6441 if( pBt
->autoVacuum
&& rc
==SQLITE_OK
){
6442 u8 eType
= (pgnoPtrmap
?PTRMAP_OVERFLOW2
:PTRMAP_OVERFLOW1
);
6443 ptrmapPut(pBt
, pgnoOvfl
, eType
, pgnoPtrmap
, &rc
);
6450 releasePage(pToRelease
);
6454 /* If pToRelease is not zero than pPrior points into the data area
6455 ** of pToRelease. Make sure pToRelease is still writeable. */
6456 assert( pToRelease
==0 || sqlite3PagerIswriteable(pToRelease
->pDbPage
) );
6458 /* If pPrior is part of the data area of pPage, then make sure pPage
6459 ** is still writeable */
6460 assert( pPrior
<pPage
->aData
|| pPrior
>=&pPage
->aData
[pBt
->pageSize
]
6461 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6463 put4byte(pPrior
, pgnoOvfl
);
6464 releasePage(pToRelease
);
6466 pPrior
= pOvfl
->aData
;
6467 put4byte(pPrior
, 0);
6468 pPayload
= &pOvfl
->aData
[4];
6469 spaceLeft
= pBt
->usableSize
- 4;
6472 releasePage(pToRelease
);
6477 ** Remove the i-th cell from pPage. This routine effects pPage only.
6478 ** The cell content is not freed or deallocated. It is assumed that
6479 ** the cell content has been copied someplace else. This routine just
6480 ** removes the reference to the cell from pPage.
6482 ** "sz" must be the number of bytes in the cell.
6484 static void dropCell(MemPage
*pPage
, int idx
, int sz
, int *pRC
){
6485 u32 pc
; /* Offset to cell content of cell being deleted */
6486 u8
*data
; /* pPage->aData */
6487 u8
*ptr
; /* Used to move bytes around within data[] */
6488 int rc
; /* The return code */
6489 int hdr
; /* Beginning of the header. 0 most pages. 100 page 1 */
6492 assert( idx
>=0 && idx
<pPage
->nCell
);
6493 assert( CORRUPT_DB
|| sz
==cellSize(pPage
, idx
) );
6494 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
6495 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6496 data
= pPage
->aData
;
6497 ptr
= &pPage
->aCellIdx
[2*idx
];
6499 hdr
= pPage
->hdrOffset
;
6500 testcase( pc
==get2byte(&data
[hdr
+5]) );
6501 testcase( pc
+sz
==pPage
->pBt
->usableSize
);
6502 if( pc
+sz
> pPage
->pBt
->usableSize
){
6503 *pRC
= SQLITE_CORRUPT_BKPT
;
6506 rc
= freeSpace(pPage
, pc
, sz
);
6512 if( pPage
->nCell
==0 ){
6513 memset(&data
[hdr
+1], 0, 4);
6515 put2byte(&data
[hdr
+5], pPage
->pBt
->usableSize
);
6516 pPage
->nFree
= pPage
->pBt
->usableSize
- pPage
->hdrOffset
6517 - pPage
->childPtrSize
- 8;
6519 memmove(ptr
, ptr
+2, 2*(pPage
->nCell
- idx
));
6520 put2byte(&data
[hdr
+3], pPage
->nCell
);
6526 ** Insert a new cell on pPage at cell index "i". pCell points to the
6527 ** content of the cell.
6529 ** If the cell content will fit on the page, then put it there. If it
6530 ** will not fit, then make a copy of the cell content into pTemp if
6531 ** pTemp is not null. Regardless of pTemp, allocate a new entry
6532 ** in pPage->apOvfl[] and make it point to the cell content (either
6533 ** in pTemp or the original pCell) and also record its index.
6534 ** Allocating a new entry in pPage->aCell[] implies that
6535 ** pPage->nOverflow is incremented.
6537 ** *pRC must be SQLITE_OK when this routine is called.
6539 static void insertCell(
6540 MemPage
*pPage
, /* Page into which we are copying */
6541 int i
, /* New cell becomes the i-th cell of the page */
6542 u8
*pCell
, /* Content of the new cell */
6543 int sz
, /* Bytes of content in pCell */
6544 u8
*pTemp
, /* Temp storage space for pCell, if needed */
6545 Pgno iChild
, /* If non-zero, replace first 4 bytes with this value */
6546 int *pRC
/* Read and write return code from here */
6548 int idx
= 0; /* Where to write new cell content in data[] */
6549 int j
; /* Loop counter */
6550 u8
*data
; /* The content of the whole page */
6551 u8
*pIns
; /* The point in pPage->aCellIdx[] where no cell inserted */
6553 assert( *pRC
==SQLITE_OK
);
6554 assert( i
>=0 && i
<=pPage
->nCell
+pPage
->nOverflow
);
6555 assert( MX_CELL(pPage
->pBt
)<=10921 );
6556 assert( pPage
->nCell
<=MX_CELL(pPage
->pBt
) || CORRUPT_DB
);
6557 assert( pPage
->nOverflow
<=ArraySize(pPage
->apOvfl
) );
6558 assert( ArraySize(pPage
->apOvfl
)==ArraySize(pPage
->aiOvfl
) );
6559 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6560 /* The cell should normally be sized correctly. However, when moving a
6561 ** malformed cell from a leaf page to an interior page, if the cell size
6562 ** wanted to be less than 4 but got rounded up to 4 on the leaf, then size
6563 ** might be less than 8 (leaf-size + pointer) on the interior node. Hence
6564 ** the term after the || in the following assert(). */
6565 assert( sz
==pPage
->xCellSize(pPage
, pCell
) || (sz
==8 && iChild
>0) );
6566 if( pPage
->nOverflow
|| sz
+2>pPage
->nFree
){
6568 memcpy(pTemp
, pCell
, sz
);
6572 put4byte(pCell
, iChild
);
6574 j
= pPage
->nOverflow
++;
6575 /* Comparison against ArraySize-1 since we hold back one extra slot
6576 ** as a contingency. In other words, never need more than 3 overflow
6577 ** slots but 4 are allocated, just to be safe. */
6578 assert( j
< ArraySize(pPage
->apOvfl
)-1 );
6579 pPage
->apOvfl
[j
] = pCell
;
6580 pPage
->aiOvfl
[j
] = (u16
)i
;
6582 /* When multiple overflows occur, they are always sequential and in
6583 ** sorted order. This invariants arise because multiple overflows can
6584 ** only occur when inserting divider cells into the parent page during
6585 ** balancing, and the dividers are adjacent and sorted.
6587 assert( j
==0 || pPage
->aiOvfl
[j
-1]<(u16
)i
); /* Overflows in sorted order */
6588 assert( j
==0 || i
==pPage
->aiOvfl
[j
-1]+1 ); /* Overflows are sequential */
6590 int rc
= sqlite3PagerWrite(pPage
->pDbPage
);
6591 if( rc
!=SQLITE_OK
){
6595 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
6596 data
= pPage
->aData
;
6597 assert( &data
[pPage
->cellOffset
]==pPage
->aCellIdx
);
6598 rc
= allocateSpace(pPage
, sz
, &idx
);
6599 if( rc
){ *pRC
= rc
; return; }
6600 /* The allocateSpace() routine guarantees the following properties
6601 ** if it returns successfully */
6603 assert( idx
>= pPage
->cellOffset
+2*pPage
->nCell
+2 || CORRUPT_DB
);
6604 assert( idx
+sz
<= (int)pPage
->pBt
->usableSize
);
6605 pPage
->nFree
-= (u16
)(2 + sz
);
6606 memcpy(&data
[idx
], pCell
, sz
);
6608 put4byte(&data
[idx
], iChild
);
6610 pIns
= pPage
->aCellIdx
+ i
*2;
6611 memmove(pIns
+2, pIns
, 2*(pPage
->nCell
- i
));
6612 put2byte(pIns
, idx
);
6614 /* increment the cell count */
6615 if( (++data
[pPage
->hdrOffset
+4])==0 ) data
[pPage
->hdrOffset
+3]++;
6616 assert( get2byte(&data
[pPage
->hdrOffset
+3])==pPage
->nCell
);
6617 #ifndef SQLITE_OMIT_AUTOVACUUM
6618 if( pPage
->pBt
->autoVacuum
){
6619 /* The cell may contain a pointer to an overflow page. If so, write
6620 ** the entry for the overflow page into the pointer map.
6622 ptrmapPutOvflPtr(pPage
, pCell
, pRC
);
6629 ** A CellArray object contains a cache of pointers and sizes for a
6630 ** consecutive sequence of cells that might be held on multiple pages.
6632 typedef struct CellArray CellArray
;
6634 int nCell
; /* Number of cells in apCell[] */
6635 MemPage
*pRef
; /* Reference page */
6636 u8
**apCell
; /* All cells begin balanced */
6637 u16
*szCell
; /* Local size of all cells in apCell[] */
6641 ** Make sure the cell sizes at idx, idx+1, ..., idx+N-1 have been
6644 static void populateCellCache(CellArray
*p
, int idx
, int N
){
6645 assert( idx
>=0 && idx
+N
<=p
->nCell
);
6647 assert( p
->apCell
[idx
]!=0 );
6648 if( p
->szCell
[idx
]==0 ){
6649 p
->szCell
[idx
] = p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[idx
]);
6651 assert( CORRUPT_DB
||
6652 p
->szCell
[idx
]==p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[idx
]) );
6660 ** Return the size of the Nth element of the cell array
6662 static SQLITE_NOINLINE u16
computeCellSize(CellArray
*p
, int N
){
6663 assert( N
>=0 && N
<p
->nCell
);
6664 assert( p
->szCell
[N
]==0 );
6665 p
->szCell
[N
] = p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[N
]);
6666 return p
->szCell
[N
];
6668 static u16
cachedCellSize(CellArray
*p
, int N
){
6669 assert( N
>=0 && N
<p
->nCell
);
6670 if( p
->szCell
[N
] ) return p
->szCell
[N
];
6671 return computeCellSize(p
, N
);
6675 ** Array apCell[] contains pointers to nCell b-tree page cells. The
6676 ** szCell[] array contains the size in bytes of each cell. This function
6677 ** replaces the current contents of page pPg with the contents of the cell
6680 ** Some of the cells in apCell[] may currently be stored in pPg. This
6681 ** function works around problems caused by this by making a copy of any
6682 ** such cells before overwriting the page data.
6684 ** The MemPage.nFree field is invalidated by this function. It is the
6685 ** responsibility of the caller to set it correctly.
6687 static int rebuildPage(
6688 MemPage
*pPg
, /* Edit this page */
6689 int nCell
, /* Final number of cells on page */
6690 u8
**apCell
, /* Array of cells */
6691 u16
*szCell
/* Array of cell sizes */
6693 const int hdr
= pPg
->hdrOffset
; /* Offset of header on pPg */
6694 u8
* const aData
= pPg
->aData
; /* Pointer to data for pPg */
6695 const int usableSize
= pPg
->pBt
->usableSize
;
6696 u8
* const pEnd
= &aData
[usableSize
];
6698 u8
*pCellptr
= pPg
->aCellIdx
;
6699 u8
*pTmp
= sqlite3PagerTempSpace(pPg
->pBt
->pPager
);
6702 i
= get2byte(&aData
[hdr
+5]);
6703 memcpy(&pTmp
[i
], &aData
[i
], usableSize
- i
);
6706 for(i
=0; i
<nCell
; i
++){
6707 u8
*pCell
= apCell
[i
];
6708 if( SQLITE_WITHIN(pCell
,aData
,pEnd
) ){
6709 pCell
= &pTmp
[pCell
- aData
];
6712 put2byte(pCellptr
, (pData
- aData
));
6714 if( pData
< pCellptr
) return SQLITE_CORRUPT_BKPT
;
6715 memcpy(pData
, pCell
, szCell
[i
]);
6716 assert( szCell
[i
]==pPg
->xCellSize(pPg
, pCell
) || CORRUPT_DB
);
6717 testcase( szCell
[i
]!=pPg
->xCellSize(pPg
,pCell
) );
6720 /* The pPg->nFree field is now set incorrectly. The caller will fix it. */
6724 put2byte(&aData
[hdr
+1], 0);
6725 put2byte(&aData
[hdr
+3], pPg
->nCell
);
6726 put2byte(&aData
[hdr
+5], pData
- aData
);
6727 aData
[hdr
+7] = 0x00;
6732 ** Array apCell[] contains nCell pointers to b-tree cells. Array szCell
6733 ** contains the size in bytes of each such cell. This function attempts to
6734 ** add the cells stored in the array to page pPg. If it cannot (because
6735 ** the page needs to be defragmented before the cells will fit), non-zero
6736 ** is returned. Otherwise, if the cells are added successfully, zero is
6739 ** Argument pCellptr points to the first entry in the cell-pointer array
6740 ** (part of page pPg) to populate. After cell apCell[0] is written to the
6741 ** page body, a 16-bit offset is written to pCellptr. And so on, for each
6742 ** cell in the array. It is the responsibility of the caller to ensure
6743 ** that it is safe to overwrite this part of the cell-pointer array.
6745 ** When this function is called, *ppData points to the start of the
6746 ** content area on page pPg. If the size of the content area is extended,
6747 ** *ppData is updated to point to the new start of the content area
6748 ** before returning.
6750 ** Finally, argument pBegin points to the byte immediately following the
6751 ** end of the space required by this page for the cell-pointer area (for
6752 ** all cells - not just those inserted by the current call). If the content
6753 ** area must be extended to before this point in order to accomodate all
6754 ** cells in apCell[], then the cells do not fit and non-zero is returned.
6756 static int pageInsertArray(
6757 MemPage
*pPg
, /* Page to add cells to */
6758 u8
*pBegin
, /* End of cell-pointer array */
6759 u8
**ppData
, /* IN/OUT: Page content -area pointer */
6760 u8
*pCellptr
, /* Pointer to cell-pointer area */
6761 int iFirst
, /* Index of first cell to add */
6762 int nCell
, /* Number of cells to add to pPg */
6763 CellArray
*pCArray
/* Array of cells */
6766 u8
*aData
= pPg
->aData
;
6767 u8
*pData
= *ppData
;
6768 int iEnd
= iFirst
+ nCell
;
6769 assert( CORRUPT_DB
|| pPg
->hdrOffset
==0 ); /* Never called on page 1 */
6770 for(i
=iFirst
; i
<iEnd
; i
++){
6773 sz
= cachedCellSize(pCArray
, i
);
6774 if( (aData
[1]==0 && aData
[2]==0) || (pSlot
= pageFindSlot(pPg
,sz
,&rc
))==0 ){
6775 if( (pData
- pBegin
)<sz
) return 1;
6779 /* pSlot and pCArray->apCell[i] will never overlap on a well-formed
6780 ** database. But they might for a corrupt database. Hence use memmove()
6781 ** since memcpy() sends SIGABORT with overlapping buffers on OpenBSD */
6782 assert( (pSlot
+sz
)<=pCArray
->apCell
[i
]
6783 || pSlot
>=(pCArray
->apCell
[i
]+sz
)
6785 memmove(pSlot
, pCArray
->apCell
[i
], sz
);
6786 put2byte(pCellptr
, (pSlot
- aData
));
6794 ** Array apCell[] contains nCell pointers to b-tree cells. Array szCell
6795 ** contains the size in bytes of each such cell. This function adds the
6796 ** space associated with each cell in the array that is currently stored
6797 ** within the body of pPg to the pPg free-list. The cell-pointers and other
6798 ** fields of the page are not updated.
6800 ** This function returns the total number of cells added to the free-list.
6802 static int pageFreeArray(
6803 MemPage
*pPg
, /* Page to edit */
6804 int iFirst
, /* First cell to delete */
6805 int nCell
, /* Cells to delete */
6806 CellArray
*pCArray
/* Array of cells */
6808 u8
* const aData
= pPg
->aData
;
6809 u8
* const pEnd
= &aData
[pPg
->pBt
->usableSize
];
6810 u8
* const pStart
= &aData
[pPg
->hdrOffset
+ 8 + pPg
->childPtrSize
];
6813 int iEnd
= iFirst
+ nCell
;
6817 for(i
=iFirst
; i
<iEnd
; i
++){
6818 u8
*pCell
= pCArray
->apCell
[i
];
6819 if( SQLITE_WITHIN(pCell
, pStart
, pEnd
) ){
6821 /* No need to use cachedCellSize() here. The sizes of all cells that
6822 ** are to be freed have already been computing while deciding which
6823 ** cells need freeing */
6824 sz
= pCArray
->szCell
[i
]; assert( sz
>0 );
6825 if( pFree
!=(pCell
+ sz
) ){
6827 assert( pFree
>aData
&& (pFree
- aData
)<65536 );
6828 freeSpace(pPg
, (u16
)(pFree
- aData
), szFree
);
6832 if( pFree
+sz
>pEnd
) return 0;
6841 assert( pFree
>aData
&& (pFree
- aData
)<65536 );
6842 freeSpace(pPg
, (u16
)(pFree
- aData
), szFree
);
6848 ** apCell[] and szCell[] contains pointers to and sizes of all cells in the
6849 ** pages being balanced. The current page, pPg, has pPg->nCell cells starting
6850 ** with apCell[iOld]. After balancing, this page should hold nNew cells
6851 ** starting at apCell[iNew].
6853 ** This routine makes the necessary adjustments to pPg so that it contains
6854 ** the correct cells after being balanced.
6856 ** The pPg->nFree field is invalid when this function returns. It is the
6857 ** responsibility of the caller to set it correctly.
6859 static int editPage(
6860 MemPage
*pPg
, /* Edit this page */
6861 int iOld
, /* Index of first cell currently on page */
6862 int iNew
, /* Index of new first cell on page */
6863 int nNew
, /* Final number of cells on page */
6864 CellArray
*pCArray
/* Array of cells and sizes */
6866 u8
* const aData
= pPg
->aData
;
6867 const int hdr
= pPg
->hdrOffset
;
6868 u8
*pBegin
= &pPg
->aCellIdx
[nNew
* 2];
6869 int nCell
= pPg
->nCell
; /* Cells stored on pPg */
6873 int iOldEnd
= iOld
+ pPg
->nCell
+ pPg
->nOverflow
;
6874 int iNewEnd
= iNew
+ nNew
;
6877 u8
*pTmp
= sqlite3PagerTempSpace(pPg
->pBt
->pPager
);
6878 memcpy(pTmp
, aData
, pPg
->pBt
->usableSize
);
6881 /* Remove cells from the start and end of the page */
6883 int nShift
= pageFreeArray(pPg
, iOld
, iNew
-iOld
, pCArray
);
6884 memmove(pPg
->aCellIdx
, &pPg
->aCellIdx
[nShift
*2], nCell
*2);
6887 if( iNewEnd
< iOldEnd
){
6888 nCell
-= pageFreeArray(pPg
, iNewEnd
, iOldEnd
- iNewEnd
, pCArray
);
6891 pData
= &aData
[get2byteNotZero(&aData
[hdr
+5])];
6892 if( pData
<pBegin
) goto editpage_fail
;
6894 /* Add cells to the start of the page */
6896 int nAdd
= MIN(nNew
,iOld
-iNew
);
6897 assert( (iOld
-iNew
)<nNew
|| nCell
==0 || CORRUPT_DB
);
6898 pCellptr
= pPg
->aCellIdx
;
6899 memmove(&pCellptr
[nAdd
*2], pCellptr
, nCell
*2);
6900 if( pageInsertArray(
6901 pPg
, pBegin
, &pData
, pCellptr
,
6903 ) ) goto editpage_fail
;
6907 /* Add any overflow cells */
6908 for(i
=0; i
<pPg
->nOverflow
; i
++){
6909 int iCell
= (iOld
+ pPg
->aiOvfl
[i
]) - iNew
;
6910 if( iCell
>=0 && iCell
<nNew
){
6911 pCellptr
= &pPg
->aCellIdx
[iCell
* 2];
6912 memmove(&pCellptr
[2], pCellptr
, (nCell
- iCell
) * 2);
6914 if( pageInsertArray(
6915 pPg
, pBegin
, &pData
, pCellptr
,
6916 iCell
+iNew
, 1, pCArray
6917 ) ) goto editpage_fail
;
6921 /* Append cells to the end of the page */
6922 pCellptr
= &pPg
->aCellIdx
[nCell
*2];
6923 if( pageInsertArray(
6924 pPg
, pBegin
, &pData
, pCellptr
,
6925 iNew
+nCell
, nNew
-nCell
, pCArray
6926 ) ) goto editpage_fail
;
6931 put2byte(&aData
[hdr
+3], pPg
->nCell
);
6932 put2byte(&aData
[hdr
+5], pData
- aData
);
6935 for(i
=0; i
<nNew
&& !CORRUPT_DB
; i
++){
6936 u8
*pCell
= pCArray
->apCell
[i
+iNew
];
6937 int iOff
= get2byteAligned(&pPg
->aCellIdx
[i
*2]);
6938 if( SQLITE_WITHIN(pCell
, aData
, &aData
[pPg
->pBt
->usableSize
]) ){
6939 pCell
= &pTmp
[pCell
- aData
];
6941 assert( 0==memcmp(pCell
, &aData
[iOff
],
6942 pCArray
->pRef
->xCellSize(pCArray
->pRef
, pCArray
->apCell
[i
+iNew
])) );
6948 /* Unable to edit this page. Rebuild it from scratch instead. */
6949 populateCellCache(pCArray
, iNew
, nNew
);
6950 return rebuildPage(pPg
, nNew
, &pCArray
->apCell
[iNew
], &pCArray
->szCell
[iNew
]);
6954 ** The following parameters determine how many adjacent pages get involved
6955 ** in a balancing operation. NN is the number of neighbors on either side
6956 ** of the page that participate in the balancing operation. NB is the
6957 ** total number of pages that participate, including the target page and
6958 ** NN neighbors on either side.
6960 ** The minimum value of NN is 1 (of course). Increasing NN above 1
6961 ** (to 2 or 3) gives a modest improvement in SELECT and DELETE performance
6962 ** in exchange for a larger degradation in INSERT and UPDATE performance.
6963 ** The value of NN appears to give the best results overall.
6965 #define NN 1 /* Number of neighbors on either side of pPage */
6966 #define NB (NN*2+1) /* Total pages involved in the balance */
6969 #ifndef SQLITE_OMIT_QUICKBALANCE
6971 ** This version of balance() handles the common special case where
6972 ** a new entry is being inserted on the extreme right-end of the
6973 ** tree, in other words, when the new entry will become the largest
6974 ** entry in the tree.
6976 ** Instead of trying to balance the 3 right-most leaf pages, just add
6977 ** a new page to the right-hand side and put the one new entry in
6978 ** that page. This leaves the right side of the tree somewhat
6979 ** unbalanced. But odds are that we will be inserting new entries
6980 ** at the end soon afterwards so the nearly empty page will quickly
6981 ** fill up. On average.
6983 ** pPage is the leaf page which is the right-most page in the tree.
6984 ** pParent is its parent. pPage must have a single overflow entry
6985 ** which is also the right-most entry on the page.
6987 ** The pSpace buffer is used to store a temporary copy of the divider
6988 ** cell that will be inserted into pParent. Such a cell consists of a 4
6989 ** byte page number followed by a variable length integer. In other
6990 ** words, at most 13 bytes. Hence the pSpace buffer must be at
6991 ** least 13 bytes in size.
6993 static int balance_quick(MemPage
*pParent
, MemPage
*pPage
, u8
*pSpace
){
6994 BtShared
*const pBt
= pPage
->pBt
; /* B-Tree Database */
6995 MemPage
*pNew
; /* Newly allocated page */
6996 int rc
; /* Return Code */
6997 Pgno pgnoNew
; /* Page number of pNew */
6999 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
7000 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7001 assert( pPage
->nOverflow
==1 );
7003 /* This error condition is now caught prior to reaching this function */
7004 if( NEVER(pPage
->nCell
==0) ) return SQLITE_CORRUPT_BKPT
;
7006 /* Allocate a new page. This page will become the right-sibling of
7007 ** pPage. Make the parent page writable, so that the new divider cell
7008 ** may be inserted. If both these operations are successful, proceed.
7010 rc
= allocateBtreePage(pBt
, &pNew
, &pgnoNew
, 0, 0);
7012 if( rc
==SQLITE_OK
){
7014 u8
*pOut
= &pSpace
[4];
7015 u8
*pCell
= pPage
->apOvfl
[0];
7016 u16 szCell
= pPage
->xCellSize(pPage
, pCell
);
7019 assert( sqlite3PagerIswriteable(pNew
->pDbPage
) );
7020 assert( pPage
->aData
[0]==(PTF_INTKEY
|PTF_LEAFDATA
|PTF_LEAF
) );
7021 zeroPage(pNew
, PTF_INTKEY
|PTF_LEAFDATA
|PTF_LEAF
);
7022 rc
= rebuildPage(pNew
, 1, &pCell
, &szCell
);
7023 if( NEVER(rc
) ) return rc
;
7024 pNew
->nFree
= pBt
->usableSize
- pNew
->cellOffset
- 2 - szCell
;
7026 /* If this is an auto-vacuum database, update the pointer map
7027 ** with entries for the new page, and any pointer from the
7028 ** cell on the page to an overflow page. If either of these
7029 ** operations fails, the return code is set, but the contents
7030 ** of the parent page are still manipulated by thh code below.
7031 ** That is Ok, at this point the parent page is guaranteed to
7032 ** be marked as dirty. Returning an error code will cause a
7033 ** rollback, undoing any changes made to the parent page.
7036 ptrmapPut(pBt
, pgnoNew
, PTRMAP_BTREE
, pParent
->pgno
, &rc
);
7037 if( szCell
>pNew
->minLocal
){
7038 ptrmapPutOvflPtr(pNew
, pCell
, &rc
);
7042 /* Create a divider cell to insert into pParent. The divider cell
7043 ** consists of a 4-byte page number (the page number of pPage) and
7044 ** a variable length key value (which must be the same value as the
7045 ** largest key on pPage).
7047 ** To find the largest key value on pPage, first find the right-most
7048 ** cell on pPage. The first two fields of this cell are the
7049 ** record-length (a variable length integer at most 32-bits in size)
7050 ** and the key value (a variable length integer, may have any value).
7051 ** The first of the while(...) loops below skips over the record-length
7052 ** field. The second while(...) loop copies the key value from the
7053 ** cell on pPage into the pSpace buffer.
7055 pCell
= findCell(pPage
, pPage
->nCell
-1);
7057 while( (*(pCell
++)&0x80) && pCell
<pStop
);
7059 while( ((*(pOut
++) = *(pCell
++))&0x80) && pCell
<pStop
);
7061 /* Insert the new divider cell into pParent. */
7062 if( rc
==SQLITE_OK
){
7063 insertCell(pParent
, pParent
->nCell
, pSpace
, (int)(pOut
-pSpace
),
7064 0, pPage
->pgno
, &rc
);
7067 /* Set the right-child pointer of pParent to point to the new page. */
7068 put4byte(&pParent
->aData
[pParent
->hdrOffset
+8], pgnoNew
);
7070 /* Release the reference to the new page. */
7076 #endif /* SQLITE_OMIT_QUICKBALANCE */
7080 ** This function does not contribute anything to the operation of SQLite.
7081 ** it is sometimes activated temporarily while debugging code responsible
7082 ** for setting pointer-map entries.
7084 static int ptrmapCheckPages(MemPage
**apPage
, int nPage
){
7086 for(i
=0; i
<nPage
; i
++){
7089 MemPage
*pPage
= apPage
[i
];
7090 BtShared
*pBt
= pPage
->pBt
;
7091 assert( pPage
->isInit
);
7093 for(j
=0; j
<pPage
->nCell
; j
++){
7097 z
= findCell(pPage
, j
);
7098 pPage
->xParseCell(pPage
, z
, &info
);
7099 if( info
.nLocal
<info
.nPayload
){
7100 Pgno ovfl
= get4byte(&z
[info
.nSize
-4]);
7101 ptrmapGet(pBt
, ovfl
, &e
, &n
);
7102 assert( n
==pPage
->pgno
&& e
==PTRMAP_OVERFLOW1
);
7105 Pgno child
= get4byte(z
);
7106 ptrmapGet(pBt
, child
, &e
, &n
);
7107 assert( n
==pPage
->pgno
&& e
==PTRMAP_BTREE
);
7111 Pgno child
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
7112 ptrmapGet(pBt
, child
, &e
, &n
);
7113 assert( n
==pPage
->pgno
&& e
==PTRMAP_BTREE
);
7121 ** This function is used to copy the contents of the b-tree node stored
7122 ** on page pFrom to page pTo. If page pFrom was not a leaf page, then
7123 ** the pointer-map entries for each child page are updated so that the
7124 ** parent page stored in the pointer map is page pTo. If pFrom contained
7125 ** any cells with overflow page pointers, then the corresponding pointer
7126 ** map entries are also updated so that the parent page is page pTo.
7128 ** If pFrom is currently carrying any overflow cells (entries in the
7129 ** MemPage.apOvfl[] array), they are not copied to pTo.
7131 ** Before returning, page pTo is reinitialized using btreeInitPage().
7133 ** The performance of this function is not critical. It is only used by
7134 ** the balance_shallower() and balance_deeper() procedures, neither of
7135 ** which are called often under normal circumstances.
7137 static void copyNodeContent(MemPage
*pFrom
, MemPage
*pTo
, int *pRC
){
7138 if( (*pRC
)==SQLITE_OK
){
7139 BtShared
* const pBt
= pFrom
->pBt
;
7140 u8
* const aFrom
= pFrom
->aData
;
7141 u8
* const aTo
= pTo
->aData
;
7142 int const iFromHdr
= pFrom
->hdrOffset
;
7143 int const iToHdr
= ((pTo
->pgno
==1) ? 100 : 0);
7148 assert( pFrom
->isInit
);
7149 assert( pFrom
->nFree
>=iToHdr
);
7150 assert( get2byte(&aFrom
[iFromHdr
+5]) <= (int)pBt
->usableSize
);
7152 /* Copy the b-tree node content from page pFrom to page pTo. */
7153 iData
= get2byte(&aFrom
[iFromHdr
+5]);
7154 memcpy(&aTo
[iData
], &aFrom
[iData
], pBt
->usableSize
-iData
);
7155 memcpy(&aTo
[iToHdr
], &aFrom
[iFromHdr
], pFrom
->cellOffset
+ 2*pFrom
->nCell
);
7157 /* Reinitialize page pTo so that the contents of the MemPage structure
7158 ** match the new data. The initialization of pTo can actually fail under
7159 ** fairly obscure circumstances, even though it is a copy of initialized
7163 rc
= btreeInitPage(pTo
);
7164 if( rc
!=SQLITE_OK
){
7169 /* If this is an auto-vacuum database, update the pointer-map entries
7170 ** for any b-tree or overflow pages that pTo now contains the pointers to.
7173 *pRC
= setChildPtrmaps(pTo
);
7179 ** This routine redistributes cells on the iParentIdx'th child of pParent
7180 ** (hereafter "the page") and up to 2 siblings so that all pages have about the
7181 ** same amount of free space. Usually a single sibling on either side of the
7182 ** page are used in the balancing, though both siblings might come from one
7183 ** side if the page is the first or last child of its parent. If the page
7184 ** has fewer than 2 siblings (something which can only happen if the page
7185 ** is a root page or a child of a root page) then all available siblings
7186 ** participate in the balancing.
7188 ** The number of siblings of the page might be increased or decreased by
7189 ** one or two in an effort to keep pages nearly full but not over full.
7191 ** Note that when this routine is called, some of the cells on the page
7192 ** might not actually be stored in MemPage.aData[]. This can happen
7193 ** if the page is overfull. This routine ensures that all cells allocated
7194 ** to the page and its siblings fit into MemPage.aData[] before returning.
7196 ** In the course of balancing the page and its siblings, cells may be
7197 ** inserted into or removed from the parent page (pParent). Doing so
7198 ** may cause the parent page to become overfull or underfull. If this
7199 ** happens, it is the responsibility of the caller to invoke the correct
7200 ** balancing routine to fix this problem (see the balance() routine).
7202 ** If this routine fails for any reason, it might leave the database
7203 ** in a corrupted state. So if this routine fails, the database should
7206 ** The third argument to this function, aOvflSpace, is a pointer to a
7207 ** buffer big enough to hold one page. If while inserting cells into the parent
7208 ** page (pParent) the parent page becomes overfull, this buffer is
7209 ** used to store the parent's overflow cells. Because this function inserts
7210 ** a maximum of four divider cells into the parent page, and the maximum
7211 ** size of a cell stored within an internal node is always less than 1/4
7212 ** of the page-size, the aOvflSpace[] buffer is guaranteed to be large
7213 ** enough for all overflow cells.
7215 ** If aOvflSpace is set to a null pointer, this function returns
7218 static int balance_nonroot(
7219 MemPage
*pParent
, /* Parent page of siblings being balanced */
7220 int iParentIdx
, /* Index of "the page" in pParent */
7221 u8
*aOvflSpace
, /* page-size bytes of space for parent ovfl */
7222 int isRoot
, /* True if pParent is a root-page */
7223 int bBulk
/* True if this call is part of a bulk load */
7225 BtShared
*pBt
; /* The whole database */
7226 int nMaxCells
= 0; /* Allocated size of apCell, szCell, aFrom. */
7227 int nNew
= 0; /* Number of pages in apNew[] */
7228 int nOld
; /* Number of pages in apOld[] */
7229 int i
, j
, k
; /* Loop counters */
7230 int nxDiv
; /* Next divider slot in pParent->aCell[] */
7231 int rc
= SQLITE_OK
; /* The return code */
7232 u16 leafCorrection
; /* 4 if pPage is a leaf. 0 if not */
7233 int leafData
; /* True if pPage is a leaf of a LEAFDATA tree */
7234 int usableSpace
; /* Bytes in pPage beyond the header */
7235 int pageFlags
; /* Value of pPage->aData[0] */
7236 int iSpace1
= 0; /* First unused byte of aSpace1[] */
7237 int iOvflSpace
= 0; /* First unused byte of aOvflSpace[] */
7238 int szScratch
; /* Size of scratch memory requested */
7239 MemPage
*apOld
[NB
]; /* pPage and up to two siblings */
7240 MemPage
*apNew
[NB
+2]; /* pPage and up to NB siblings after balancing */
7241 u8
*pRight
; /* Location in parent of right-sibling pointer */
7242 u8
*apDiv
[NB
-1]; /* Divider cells in pParent */
7243 int cntNew
[NB
+2]; /* Index in b.paCell[] of cell after i-th page */
7244 int cntOld
[NB
+2]; /* Old index in b.apCell[] */
7245 int szNew
[NB
+2]; /* Combined size of cells placed on i-th page */
7246 u8
*aSpace1
; /* Space for copies of dividers cells */
7247 Pgno pgno
; /* Temp var to store a page number in */
7248 u8 abDone
[NB
+2]; /* True after i'th new page is populated */
7249 Pgno aPgno
[NB
+2]; /* Page numbers of new pages before shuffling */
7250 Pgno aPgOrder
[NB
+2]; /* Copy of aPgno[] used for sorting pages */
7251 u16 aPgFlags
[NB
+2]; /* flags field of new pages before shuffling */
7252 CellArray b
; /* Parsed information on cells being balanced */
7254 memset(abDone
, 0, sizeof(abDone
));
7258 assert( sqlite3_mutex_held(pBt
->mutex
) );
7259 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7262 TRACE(("BALANCE: begin page %d child of %d\n", pPage
->pgno
, pParent
->pgno
));
7265 /* At this point pParent may have at most one overflow cell. And if
7266 ** this overflow cell is present, it must be the cell with
7267 ** index iParentIdx. This scenario comes about when this function
7268 ** is called (indirectly) from sqlite3BtreeDelete().
7270 assert( pParent
->nOverflow
==0 || pParent
->nOverflow
==1 );
7271 assert( pParent
->nOverflow
==0 || pParent
->aiOvfl
[0]==iParentIdx
);
7274 return SQLITE_NOMEM_BKPT
;
7277 /* Find the sibling pages to balance. Also locate the cells in pParent
7278 ** that divide the siblings. An attempt is made to find NN siblings on
7279 ** either side of pPage. More siblings are taken from one side, however,
7280 ** if there are fewer than NN siblings on the other side. If pParent
7281 ** has NB or fewer children then all children of pParent are taken.
7283 ** This loop also drops the divider cells from the parent page. This
7284 ** way, the remainder of the function does not have to deal with any
7285 ** overflow cells in the parent page, since if any existed they will
7286 ** have already been removed.
7288 i
= pParent
->nOverflow
+ pParent
->nCell
;
7292 assert( bBulk
==0 || bBulk
==1 );
7293 if( iParentIdx
==0 ){
7295 }else if( iParentIdx
==i
){
7298 nxDiv
= iParentIdx
-1;
7303 if( (i
+nxDiv
-pParent
->nOverflow
)==pParent
->nCell
){
7304 pRight
= &pParent
->aData
[pParent
->hdrOffset
+8];
7306 pRight
= findCell(pParent
, i
+nxDiv
-pParent
->nOverflow
);
7308 pgno
= get4byte(pRight
);
7310 rc
= getAndInitPage(pBt
, pgno
, &apOld
[i
], 0, 0);
7312 memset(apOld
, 0, (i
+1)*sizeof(MemPage
*));
7313 goto balance_cleanup
;
7315 nMaxCells
+= 1+apOld
[i
]->nCell
+apOld
[i
]->nOverflow
;
7316 if( (i
--)==0 ) break;
7318 if( pParent
->nOverflow
&& i
+nxDiv
==pParent
->aiOvfl
[0] ){
7319 apDiv
[i
] = pParent
->apOvfl
[0];
7320 pgno
= get4byte(apDiv
[i
]);
7321 szNew
[i
] = pParent
->xCellSize(pParent
, apDiv
[i
]);
7322 pParent
->nOverflow
= 0;
7324 apDiv
[i
] = findCell(pParent
, i
+nxDiv
-pParent
->nOverflow
);
7325 pgno
= get4byte(apDiv
[i
]);
7326 szNew
[i
] = pParent
->xCellSize(pParent
, apDiv
[i
]);
7328 /* Drop the cell from the parent page. apDiv[i] still points to
7329 ** the cell within the parent, even though it has been dropped.
7330 ** This is safe because dropping a cell only overwrites the first
7331 ** four bytes of it, and this function does not need the first
7332 ** four bytes of the divider cell. So the pointer is safe to use
7335 ** But not if we are in secure-delete mode. In secure-delete mode,
7336 ** the dropCell() routine will overwrite the entire cell with zeroes.
7337 ** In this case, temporarily copy the cell into the aOvflSpace[]
7338 ** buffer. It will be copied out again as soon as the aSpace[] buffer
7340 if( pBt
->btsFlags
& BTS_FAST_SECURE
){
7343 iOff
= SQLITE_PTR_TO_INT(apDiv
[i
]) - SQLITE_PTR_TO_INT(pParent
->aData
);
7344 if( (iOff
+szNew
[i
])>(int)pBt
->usableSize
){
7345 rc
= SQLITE_CORRUPT_BKPT
;
7346 memset(apOld
, 0, (i
+1)*sizeof(MemPage
*));
7347 goto balance_cleanup
;
7349 memcpy(&aOvflSpace
[iOff
], apDiv
[i
], szNew
[i
]);
7350 apDiv
[i
] = &aOvflSpace
[apDiv
[i
]-pParent
->aData
];
7353 dropCell(pParent
, i
+nxDiv
-pParent
->nOverflow
, szNew
[i
], &rc
);
7357 /* Make nMaxCells a multiple of 4 in order to preserve 8-byte
7359 nMaxCells
= (nMaxCells
+ 3)&~3;
7362 ** Allocate space for memory structures
7365 nMaxCells
*sizeof(u8
*) /* b.apCell */
7366 + nMaxCells
*sizeof(u16
) /* b.szCell */
7367 + pBt
->pageSize
; /* aSpace1 */
7369 assert( szScratch
<=6*(int)pBt
->pageSize
);
7370 b
.apCell
= sqlite3StackAllocRaw(0, szScratch
);
7372 rc
= SQLITE_NOMEM_BKPT
;
7373 goto balance_cleanup
;
7375 b
.szCell
= (u16
*)&b
.apCell
[nMaxCells
];
7376 aSpace1
= (u8
*)&b
.szCell
[nMaxCells
];
7377 assert( EIGHT_BYTE_ALIGNMENT(aSpace1
) );
7380 ** Load pointers to all cells on sibling pages and the divider cells
7381 ** into the local b.apCell[] array. Make copies of the divider cells
7382 ** into space obtained from aSpace1[]. The divider cells have already
7383 ** been removed from pParent.
7385 ** If the siblings are on leaf pages, then the child pointers of the
7386 ** divider cells are stripped from the cells before they are copied
7387 ** into aSpace1[]. In this way, all cells in b.apCell[] are without
7388 ** child pointers. If siblings are not leaves, then all cell in
7389 ** b.apCell[] include child pointers. Either way, all cells in b.apCell[]
7392 ** leafCorrection: 4 if pPage is a leaf. 0 if pPage is not a leaf.
7393 ** leafData: 1 if pPage holds key+data and pParent holds only keys.
7396 leafCorrection
= b
.pRef
->leaf
*4;
7397 leafData
= b
.pRef
->intKeyLeaf
;
7398 for(i
=0; i
<nOld
; i
++){
7399 MemPage
*pOld
= apOld
[i
];
7400 int limit
= pOld
->nCell
;
7401 u8
*aData
= pOld
->aData
;
7402 u16 maskPage
= pOld
->maskPage
;
7403 u8
*piCell
= aData
+ pOld
->cellOffset
;
7406 /* Verify that all sibling pages are of the same "type" (table-leaf,
7407 ** table-interior, index-leaf, or index-interior).
7409 if( pOld
->aData
[0]!=apOld
[0]->aData
[0] ){
7410 rc
= SQLITE_CORRUPT_BKPT
;
7411 goto balance_cleanup
;
7414 /* Load b.apCell[] with pointers to all cells in pOld. If pOld
7415 ** contains overflow cells, include them in the b.apCell[] array
7416 ** in the correct spot.
7418 ** Note that when there are multiple overflow cells, it is always the
7419 ** case that they are sequential and adjacent. This invariant arises
7420 ** because multiple overflows can only occurs when inserting divider
7421 ** cells into a parent on a prior balance, and divider cells are always
7422 ** adjacent and are inserted in order. There is an assert() tagged
7423 ** with "NOTE 1" in the overflow cell insertion loop to prove this
7426 ** This must be done in advance. Once the balance starts, the cell
7427 ** offset section of the btree page will be overwritten and we will no
7428 ** long be able to find the cells if a pointer to each cell is not saved
7431 memset(&b
.szCell
[b
.nCell
], 0, sizeof(b
.szCell
[0])*(limit
+pOld
->nOverflow
));
7432 if( pOld
->nOverflow
>0 ){
7433 limit
= pOld
->aiOvfl
[0];
7434 for(j
=0; j
<limit
; j
++){
7435 b
.apCell
[b
.nCell
] = aData
+ (maskPage
& get2byteAligned(piCell
));
7439 for(k
=0; k
<pOld
->nOverflow
; k
++){
7440 assert( k
==0 || pOld
->aiOvfl
[k
-1]+1==pOld
->aiOvfl
[k
] );/* NOTE 1 */
7441 b
.apCell
[b
.nCell
] = pOld
->apOvfl
[k
];
7445 piEnd
= aData
+ pOld
->cellOffset
+ 2*pOld
->nCell
;
7446 while( piCell
<piEnd
){
7447 assert( b
.nCell
<nMaxCells
);
7448 b
.apCell
[b
.nCell
] = aData
+ (maskPage
& get2byteAligned(piCell
));
7453 cntOld
[i
] = b
.nCell
;
7454 if( i
<nOld
-1 && !leafData
){
7455 u16 sz
= (u16
)szNew
[i
];
7457 assert( b
.nCell
<nMaxCells
);
7458 b
.szCell
[b
.nCell
] = sz
;
7459 pTemp
= &aSpace1
[iSpace1
];
7461 assert( sz
<=pBt
->maxLocal
+23 );
7462 assert( iSpace1
<= (int)pBt
->pageSize
);
7463 memcpy(pTemp
, apDiv
[i
], sz
);
7464 b
.apCell
[b
.nCell
] = pTemp
+leafCorrection
;
7465 assert( leafCorrection
==0 || leafCorrection
==4 );
7466 b
.szCell
[b
.nCell
] = b
.szCell
[b
.nCell
] - leafCorrection
;
7468 assert( leafCorrection
==0 );
7469 assert( pOld
->hdrOffset
==0 );
7470 /* The right pointer of the child page pOld becomes the left
7471 ** pointer of the divider cell */
7472 memcpy(b
.apCell
[b
.nCell
], &pOld
->aData
[8], 4);
7474 assert( leafCorrection
==4 );
7475 while( b
.szCell
[b
.nCell
]<4 ){
7476 /* Do not allow any cells smaller than 4 bytes. If a smaller cell
7477 ** does exist, pad it with 0x00 bytes. */
7478 assert( b
.szCell
[b
.nCell
]==3 || CORRUPT_DB
);
7479 assert( b
.apCell
[b
.nCell
]==&aSpace1
[iSpace1
-3] || CORRUPT_DB
);
7480 aSpace1
[iSpace1
++] = 0x00;
7481 b
.szCell
[b
.nCell
]++;
7489 ** Figure out the number of pages needed to hold all b.nCell cells.
7490 ** Store this number in "k". Also compute szNew[] which is the total
7491 ** size of all cells on the i-th page and cntNew[] which is the index
7492 ** in b.apCell[] of the cell that divides page i from page i+1.
7493 ** cntNew[k] should equal b.nCell.
7495 ** Values computed by this block:
7497 ** k: The total number of sibling pages
7498 ** szNew[i]: Spaced used on the i-th sibling page.
7499 ** cntNew[i]: Index in b.apCell[] and b.szCell[] for the first cell to
7500 ** the right of the i-th sibling page.
7501 ** usableSpace: Number of bytes of space available on each sibling.
7504 usableSpace
= pBt
->usableSize
- 12 + leafCorrection
;
7505 for(i
=0; i
<nOld
; i
++){
7506 MemPage
*p
= apOld
[i
];
7507 szNew
[i
] = usableSpace
- p
->nFree
;
7508 for(j
=0; j
<p
->nOverflow
; j
++){
7509 szNew
[i
] += 2 + p
->xCellSize(p
, p
->apOvfl
[j
]);
7511 cntNew
[i
] = cntOld
[i
];
7516 while( szNew
[i
]>usableSpace
){
7519 if( k
>NB
+2 ){ rc
= SQLITE_CORRUPT_BKPT
; goto balance_cleanup
; }
7521 cntNew
[k
-1] = b
.nCell
;
7523 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]-1);
7526 if( cntNew
[i
]<b
.nCell
){
7527 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7535 while( cntNew
[i
]<b
.nCell
){
7536 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7537 if( szNew
[i
]+sz
>usableSpace
) break;
7541 if( cntNew
[i
]<b
.nCell
){
7542 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7549 if( cntNew
[i
]>=b
.nCell
){
7551 }else if( cntNew
[i
] <= (i
>0 ? cntNew
[i
-1] : 0) ){
7552 rc
= SQLITE_CORRUPT_BKPT
;
7553 goto balance_cleanup
;
7558 ** The packing computed by the previous block is biased toward the siblings
7559 ** on the left side (siblings with smaller keys). The left siblings are
7560 ** always nearly full, while the right-most sibling might be nearly empty.
7561 ** The next block of code attempts to adjust the packing of siblings to
7562 ** get a better balance.
7564 ** This adjustment is more than an optimization. The packing above might
7565 ** be so out of balance as to be illegal. For example, the right-most
7566 ** sibling might be completely empty. This adjustment is not optional.
7568 for(i
=k
-1; i
>0; i
--){
7569 int szRight
= szNew
[i
]; /* Size of sibling on the right */
7570 int szLeft
= szNew
[i
-1]; /* Size of sibling on the left */
7571 int r
; /* Index of right-most cell in left sibling */
7572 int d
; /* Index of first cell to the left of right sibling */
7574 r
= cntNew
[i
-1] - 1;
7575 d
= r
+ 1 - leafData
;
7576 (void)cachedCellSize(&b
, d
);
7578 assert( d
<nMaxCells
);
7579 assert( r
<nMaxCells
);
7580 (void)cachedCellSize(&b
, r
);
7582 && (bBulk
|| szRight
+b
.szCell
[d
]+2 > szLeft
-(b
.szCell
[r
]+(i
==k
-1?0:2)))){
7585 szRight
+= b
.szCell
[d
] + 2;
7586 szLeft
-= b
.szCell
[r
] + 2;
7592 szNew
[i
-1] = szLeft
;
7593 if( cntNew
[i
-1] <= (i
>1 ? cntNew
[i
-2] : 0) ){
7594 rc
= SQLITE_CORRUPT_BKPT
;
7595 goto balance_cleanup
;
7599 /* Sanity check: For a non-corrupt database file one of the follwing
7601 ** (1) We found one or more cells (cntNew[0])>0), or
7602 ** (2) pPage is a virtual root page. A virtual root page is when
7603 ** the real root page is page 1 and we are the only child of
7606 assert( cntNew
[0]>0 || (pParent
->pgno
==1 && pParent
->nCell
==0) || CORRUPT_DB
);
7607 TRACE(("BALANCE: old: %d(nc=%d) %d(nc=%d) %d(nc=%d)\n",
7608 apOld
[0]->pgno
, apOld
[0]->nCell
,
7609 nOld
>=2 ? apOld
[1]->pgno
: 0, nOld
>=2 ? apOld
[1]->nCell
: 0,
7610 nOld
>=3 ? apOld
[2]->pgno
: 0, nOld
>=3 ? apOld
[2]->nCell
: 0
7614 ** Allocate k new pages. Reuse old pages where possible.
7616 pageFlags
= apOld
[0]->aData
[0];
7620 pNew
= apNew
[i
] = apOld
[i
];
7622 rc
= sqlite3PagerWrite(pNew
->pDbPage
);
7624 if( rc
) goto balance_cleanup
;
7627 rc
= allocateBtreePage(pBt
, &pNew
, &pgno
, (bBulk
? 1 : pgno
), 0);
7628 if( rc
) goto balance_cleanup
;
7629 zeroPage(pNew
, pageFlags
);
7632 cntOld
[i
] = b
.nCell
;
7634 /* Set the pointer-map entry for the new sibling page. */
7636 ptrmapPut(pBt
, pNew
->pgno
, PTRMAP_BTREE
, pParent
->pgno
, &rc
);
7637 if( rc
!=SQLITE_OK
){
7638 goto balance_cleanup
;
7645 ** Reassign page numbers so that the new pages are in ascending order.
7646 ** This helps to keep entries in the disk file in order so that a scan
7647 ** of the table is closer to a linear scan through the file. That in turn
7648 ** helps the operating system to deliver pages from the disk more rapidly.
7650 ** An O(n^2) insertion sort algorithm is used, but since n is never more
7651 ** than (NB+2) (a small constant), that should not be a problem.
7653 ** When NB==3, this one optimization makes the database about 25% faster
7654 ** for large insertions and deletions.
7656 for(i
=0; i
<nNew
; i
++){
7657 aPgOrder
[i
] = aPgno
[i
] = apNew
[i
]->pgno
;
7658 aPgFlags
[i
] = apNew
[i
]->pDbPage
->flags
;
7660 if( aPgno
[j
]==aPgno
[i
] ){
7661 /* This branch is taken if the set of sibling pages somehow contains
7662 ** duplicate entries. This can happen if the database is corrupt.
7663 ** It would be simpler to detect this as part of the loop below, but
7664 ** we do the detection here in order to avoid populating the pager
7665 ** cache with two separate objects associated with the same
7667 assert( CORRUPT_DB
);
7668 rc
= SQLITE_CORRUPT_BKPT
;
7669 goto balance_cleanup
;
7673 for(i
=0; i
<nNew
; i
++){
7674 int iBest
= 0; /* aPgno[] index of page number to use */
7675 for(j
=1; j
<nNew
; j
++){
7676 if( aPgOrder
[j
]<aPgOrder
[iBest
] ) iBest
= j
;
7678 pgno
= aPgOrder
[iBest
];
7679 aPgOrder
[iBest
] = 0xffffffff;
7682 sqlite3PagerRekey(apNew
[iBest
]->pDbPage
, pBt
->nPage
+iBest
+1, 0);
7684 sqlite3PagerRekey(apNew
[i
]->pDbPage
, pgno
, aPgFlags
[iBest
]);
7685 apNew
[i
]->pgno
= pgno
;
7689 TRACE(("BALANCE: new: %d(%d nc=%d) %d(%d nc=%d) %d(%d nc=%d) "
7690 "%d(%d nc=%d) %d(%d nc=%d)\n",
7691 apNew
[0]->pgno
, szNew
[0], cntNew
[0],
7692 nNew
>=2 ? apNew
[1]->pgno
: 0, nNew
>=2 ? szNew
[1] : 0,
7693 nNew
>=2 ? cntNew
[1] - cntNew
[0] - !leafData
: 0,
7694 nNew
>=3 ? apNew
[2]->pgno
: 0, nNew
>=3 ? szNew
[2] : 0,
7695 nNew
>=3 ? cntNew
[2] - cntNew
[1] - !leafData
: 0,
7696 nNew
>=4 ? apNew
[3]->pgno
: 0, nNew
>=4 ? szNew
[3] : 0,
7697 nNew
>=4 ? cntNew
[3] - cntNew
[2] - !leafData
: 0,
7698 nNew
>=5 ? apNew
[4]->pgno
: 0, nNew
>=5 ? szNew
[4] : 0,
7699 nNew
>=5 ? cntNew
[4] - cntNew
[3] - !leafData
: 0
7702 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7703 put4byte(pRight
, apNew
[nNew
-1]->pgno
);
7705 /* If the sibling pages are not leaves, ensure that the right-child pointer
7706 ** of the right-most new sibling page is set to the value that was
7707 ** originally in the same field of the right-most old sibling page. */
7708 if( (pageFlags
& PTF_LEAF
)==0 && nOld
!=nNew
){
7709 MemPage
*pOld
= (nNew
>nOld
? apNew
: apOld
)[nOld
-1];
7710 memcpy(&apNew
[nNew
-1]->aData
[8], &pOld
->aData
[8], 4);
7713 /* Make any required updates to pointer map entries associated with
7714 ** cells stored on sibling pages following the balance operation. Pointer
7715 ** map entries associated with divider cells are set by the insertCell()
7716 ** routine. The associated pointer map entries are:
7718 ** a) if the cell contains a reference to an overflow chain, the
7719 ** entry associated with the first page in the overflow chain, and
7721 ** b) if the sibling pages are not leaves, the child page associated
7724 ** If the sibling pages are not leaves, then the pointer map entry
7725 ** associated with the right-child of each sibling may also need to be
7726 ** updated. This happens below, after the sibling pages have been
7727 ** populated, not here.
7730 MemPage
*pNew
= apNew
[0];
7731 u8
*aOld
= pNew
->aData
;
7732 int cntOldNext
= pNew
->nCell
+ pNew
->nOverflow
;
7733 int usableSize
= pBt
->usableSize
;
7737 for(i
=0; i
<b
.nCell
; i
++){
7738 u8
*pCell
= b
.apCell
[i
];
7739 if( i
==cntOldNext
){
7740 MemPage
*pOld
= (++iOld
)<nNew
? apNew
[iOld
] : apOld
[iOld
];
7741 cntOldNext
+= pOld
->nCell
+ pOld
->nOverflow
+ !leafData
;
7744 if( i
==cntNew
[iNew
] ){
7745 pNew
= apNew
[++iNew
];
7746 if( !leafData
) continue;
7749 /* Cell pCell is destined for new sibling page pNew. Originally, it
7750 ** was either part of sibling page iOld (possibly an overflow cell),
7751 ** or else the divider cell to the left of sibling page iOld. So,
7752 ** if sibling page iOld had the same page number as pNew, and if
7753 ** pCell really was a part of sibling page iOld (not a divider or
7754 ** overflow cell), we can skip updating the pointer map entries. */
7756 || pNew
->pgno
!=aPgno
[iOld
]
7757 || !SQLITE_WITHIN(pCell
,aOld
,&aOld
[usableSize
])
7759 if( !leafCorrection
){
7760 ptrmapPut(pBt
, get4byte(pCell
), PTRMAP_BTREE
, pNew
->pgno
, &rc
);
7762 if( cachedCellSize(&b
,i
)>pNew
->minLocal
){
7763 ptrmapPutOvflPtr(pNew
, pCell
, &rc
);
7765 if( rc
) goto balance_cleanup
;
7770 /* Insert new divider cells into pParent. */
7771 for(i
=0; i
<nNew
-1; i
++){
7775 MemPage
*pNew
= apNew
[i
];
7778 assert( j
<nMaxCells
);
7779 assert( b
.apCell
[j
]!=0 );
7780 pCell
= b
.apCell
[j
];
7781 sz
= b
.szCell
[j
] + leafCorrection
;
7782 pTemp
= &aOvflSpace
[iOvflSpace
];
7784 memcpy(&pNew
->aData
[8], pCell
, 4);
7785 }else if( leafData
){
7786 /* If the tree is a leaf-data tree, and the siblings are leaves,
7787 ** then there is no divider cell in b.apCell[]. Instead, the divider
7788 ** cell consists of the integer key for the right-most cell of
7789 ** the sibling-page assembled above only.
7793 pNew
->xParseCell(pNew
, b
.apCell
[j
], &info
);
7795 sz
= 4 + putVarint(&pCell
[4], info
.nKey
);
7799 /* Obscure case for non-leaf-data trees: If the cell at pCell was
7800 ** previously stored on a leaf node, and its reported size was 4
7801 ** bytes, then it may actually be smaller than this
7802 ** (see btreeParseCellPtr(), 4 bytes is the minimum size of
7803 ** any cell). But it is important to pass the correct size to
7804 ** insertCell(), so reparse the cell now.
7806 ** This can only happen for b-trees used to evaluate "IN (SELECT ...)"
7807 ** and WITHOUT ROWID tables with exactly one column which is the
7810 if( b
.szCell
[j
]==4 ){
7811 assert(leafCorrection
==4);
7812 sz
= pParent
->xCellSize(pParent
, pCell
);
7816 assert( sz
<=pBt
->maxLocal
+23 );
7817 assert( iOvflSpace
<= (int)pBt
->pageSize
);
7818 insertCell(pParent
, nxDiv
+i
, pCell
, sz
, pTemp
, pNew
->pgno
, &rc
);
7819 if( rc
!=SQLITE_OK
) goto balance_cleanup
;
7820 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7823 /* Now update the actual sibling pages. The order in which they are updated
7824 ** is important, as this code needs to avoid disrupting any page from which
7825 ** cells may still to be read. In practice, this means:
7827 ** (1) If cells are moving left (from apNew[iPg] to apNew[iPg-1])
7828 ** then it is not safe to update page apNew[iPg] until after
7829 ** the left-hand sibling apNew[iPg-1] has been updated.
7831 ** (2) If cells are moving right (from apNew[iPg] to apNew[iPg+1])
7832 ** then it is not safe to update page apNew[iPg] until after
7833 ** the right-hand sibling apNew[iPg+1] has been updated.
7835 ** If neither of the above apply, the page is safe to update.
7837 ** The iPg value in the following loop starts at nNew-1 goes down
7838 ** to 0, then back up to nNew-1 again, thus making two passes over
7839 ** the pages. On the initial downward pass, only condition (1) above
7840 ** needs to be tested because (2) will always be true from the previous
7841 ** step. On the upward pass, both conditions are always true, so the
7842 ** upwards pass simply processes pages that were missed on the downward
7845 for(i
=1-nNew
; i
<nNew
; i
++){
7846 int iPg
= i
<0 ? -i
: i
;
7847 assert( iPg
>=0 && iPg
<nNew
);
7848 if( abDone
[iPg
] ) continue; /* Skip pages already processed */
7849 if( i
>=0 /* On the upwards pass, or... */
7850 || cntOld
[iPg
-1]>=cntNew
[iPg
-1] /* Condition (1) is true */
7856 /* Verify condition (1): If cells are moving left, update iPg
7857 ** only after iPg-1 has already been updated. */
7858 assert( iPg
==0 || cntOld
[iPg
-1]>=cntNew
[iPg
-1] || abDone
[iPg
-1] );
7860 /* Verify condition (2): If cells are moving right, update iPg
7861 ** only after iPg+1 has already been updated. */
7862 assert( cntNew
[iPg
]>=cntOld
[iPg
] || abDone
[iPg
+1] );
7866 nNewCell
= cntNew
[0];
7868 iOld
= iPg
<nOld
? (cntOld
[iPg
-1] + !leafData
) : b
.nCell
;
7869 iNew
= cntNew
[iPg
-1] + !leafData
;
7870 nNewCell
= cntNew
[iPg
] - iNew
;
7873 rc
= editPage(apNew
[iPg
], iOld
, iNew
, nNewCell
, &b
);
7874 if( rc
) goto balance_cleanup
;
7876 apNew
[iPg
]->nFree
= usableSpace
-szNew
[iPg
];
7877 assert( apNew
[iPg
]->nOverflow
==0 );
7878 assert( apNew
[iPg
]->nCell
==nNewCell
);
7882 /* All pages have been processed exactly once */
7883 assert( memcmp(abDone
, "\01\01\01\01\01", nNew
)==0 );
7888 if( isRoot
&& pParent
->nCell
==0 && pParent
->hdrOffset
<=apNew
[0]->nFree
){
7889 /* The root page of the b-tree now contains no cells. The only sibling
7890 ** page is the right-child of the parent. Copy the contents of the
7891 ** child page into the parent, decreasing the overall height of the
7892 ** b-tree structure by one. This is described as the "balance-shallower"
7893 ** sub-algorithm in some documentation.
7895 ** If this is an auto-vacuum database, the call to copyNodeContent()
7896 ** sets all pointer-map entries corresponding to database image pages
7897 ** for which the pointer is stored within the content being copied.
7899 ** It is critical that the child page be defragmented before being
7900 ** copied into the parent, because if the parent is page 1 then it will
7901 ** by smaller than the child due to the database header, and so all the
7902 ** free space needs to be up front.
7904 assert( nNew
==1 || CORRUPT_DB
);
7905 rc
= defragmentPage(apNew
[0], -1);
7906 testcase( rc
!=SQLITE_OK
);
7907 assert( apNew
[0]->nFree
==
7908 (get2byte(&apNew
[0]->aData
[5])-apNew
[0]->cellOffset
-apNew
[0]->nCell
*2)
7911 copyNodeContent(apNew
[0], pParent
, &rc
);
7912 freePage(apNew
[0], &rc
);
7913 }else if( ISAUTOVACUUM
&& !leafCorrection
){
7914 /* Fix the pointer map entries associated with the right-child of each
7915 ** sibling page. All other pointer map entries have already been taken
7917 for(i
=0; i
<nNew
; i
++){
7918 u32 key
= get4byte(&apNew
[i
]->aData
[8]);
7919 ptrmapPut(pBt
, key
, PTRMAP_BTREE
, apNew
[i
]->pgno
, &rc
);
7923 assert( pParent
->isInit
);
7924 TRACE(("BALANCE: finished: old=%d new=%d cells=%d\n",
7925 nOld
, nNew
, b
.nCell
));
7927 /* Free any old pages that were not reused as new pages.
7929 for(i
=nNew
; i
<nOld
; i
++){
7930 freePage(apOld
[i
], &rc
);
7934 if( ISAUTOVACUUM
&& rc
==SQLITE_OK
&& apNew
[0]->isInit
){
7935 /* The ptrmapCheckPages() contains assert() statements that verify that
7936 ** all pointer map pages are set correctly. This is helpful while
7937 ** debugging. This is usually disabled because a corrupt database may
7938 ** cause an assert() statement to fail. */
7939 ptrmapCheckPages(apNew
, nNew
);
7940 ptrmapCheckPages(&pParent
, 1);
7945 ** Cleanup before returning.
7948 sqlite3StackFree(0, b
.apCell
);
7949 for(i
=0; i
<nOld
; i
++){
7950 releasePage(apOld
[i
]);
7952 for(i
=0; i
<nNew
; i
++){
7953 releasePage(apNew
[i
]);
7961 ** This function is called when the root page of a b-tree structure is
7962 ** overfull (has one or more overflow pages).
7964 ** A new child page is allocated and the contents of the current root
7965 ** page, including overflow cells, are copied into the child. The root
7966 ** page is then overwritten to make it an empty page with the right-child
7967 ** pointer pointing to the new page.
7969 ** Before returning, all pointer-map entries corresponding to pages
7970 ** that the new child-page now contains pointers to are updated. The
7971 ** entry corresponding to the new right-child pointer of the root
7972 ** page is also updated.
7974 ** If successful, *ppChild is set to contain a reference to the child
7975 ** page and SQLITE_OK is returned. In this case the caller is required
7976 ** to call releasePage() on *ppChild exactly once. If an error occurs,
7977 ** an error code is returned and *ppChild is set to 0.
7979 static int balance_deeper(MemPage
*pRoot
, MemPage
**ppChild
){
7980 int rc
; /* Return value from subprocedures */
7981 MemPage
*pChild
= 0; /* Pointer to a new child page */
7982 Pgno pgnoChild
= 0; /* Page number of the new child page */
7983 BtShared
*pBt
= pRoot
->pBt
; /* The BTree */
7985 assert( pRoot
->nOverflow
>0 );
7986 assert( sqlite3_mutex_held(pBt
->mutex
) );
7988 /* Make pRoot, the root page of the b-tree, writable. Allocate a new
7989 ** page that will become the new right-child of pPage. Copy the contents
7990 ** of the node stored on pRoot into the new child page.
7992 rc
= sqlite3PagerWrite(pRoot
->pDbPage
);
7993 if( rc
==SQLITE_OK
){
7994 rc
= allocateBtreePage(pBt
,&pChild
,&pgnoChild
,pRoot
->pgno
,0);
7995 copyNodeContent(pRoot
, pChild
, &rc
);
7997 ptrmapPut(pBt
, pgnoChild
, PTRMAP_BTREE
, pRoot
->pgno
, &rc
);
8002 releasePage(pChild
);
8005 assert( sqlite3PagerIswriteable(pChild
->pDbPage
) );
8006 assert( sqlite3PagerIswriteable(pRoot
->pDbPage
) );
8007 assert( pChild
->nCell
==pRoot
->nCell
);
8009 TRACE(("BALANCE: copy root %d into %d\n", pRoot
->pgno
, pChild
->pgno
));
8011 /* Copy the overflow cells from pRoot to pChild */
8012 memcpy(pChild
->aiOvfl
, pRoot
->aiOvfl
,
8013 pRoot
->nOverflow
*sizeof(pRoot
->aiOvfl
[0]));
8014 memcpy(pChild
->apOvfl
, pRoot
->apOvfl
,
8015 pRoot
->nOverflow
*sizeof(pRoot
->apOvfl
[0]));
8016 pChild
->nOverflow
= pRoot
->nOverflow
;
8018 /* Zero the contents of pRoot. Then install pChild as the right-child. */
8019 zeroPage(pRoot
, pChild
->aData
[0] & ~PTF_LEAF
);
8020 put4byte(&pRoot
->aData
[pRoot
->hdrOffset
+8], pgnoChild
);
8027 ** The page that pCur currently points to has just been modified in
8028 ** some way. This function figures out if this modification means the
8029 ** tree needs to be balanced, and if so calls the appropriate balancing
8030 ** routine. Balancing routines are:
8034 ** balance_nonroot()
8036 static int balance(BtCursor
*pCur
){
8038 const int nMin
= pCur
->pBt
->usableSize
* 2 / 3;
8039 u8 aBalanceQuickSpace
[13];
8042 VVA_ONLY( int balance_quick_called
= 0 );
8043 VVA_ONLY( int balance_deeper_called
= 0 );
8046 int iPage
= pCur
->iPage
;
8047 MemPage
*pPage
= pCur
->pPage
;
8050 if( pPage
->nOverflow
){
8051 /* The root page of the b-tree is overfull. In this case call the
8052 ** balance_deeper() function to create a new child for the root-page
8053 ** and copy the current contents of the root-page to it. The
8054 ** next iteration of the do-loop will balance the child page.
8056 assert( balance_deeper_called
==0 );
8057 VVA_ONLY( balance_deeper_called
++ );
8058 rc
= balance_deeper(pPage
, &pCur
->apPage
[1]);
8059 if( rc
==SQLITE_OK
){
8063 pCur
->apPage
[0] = pPage
;
8064 pCur
->pPage
= pCur
->apPage
[1];
8065 assert( pCur
->pPage
->nOverflow
);
8070 }else if( pPage
->nOverflow
==0 && pPage
->nFree
<=nMin
){
8073 MemPage
* const pParent
= pCur
->apPage
[iPage
-1];
8074 int const iIdx
= pCur
->aiIdx
[iPage
-1];
8076 rc
= sqlite3PagerWrite(pParent
->pDbPage
);
8077 if( rc
==SQLITE_OK
){
8078 #ifndef SQLITE_OMIT_QUICKBALANCE
8079 if( pPage
->intKeyLeaf
8080 && pPage
->nOverflow
==1
8081 && pPage
->aiOvfl
[0]==pPage
->nCell
8083 && pParent
->nCell
==iIdx
8085 /* Call balance_quick() to create a new sibling of pPage on which
8086 ** to store the overflow cell. balance_quick() inserts a new cell
8087 ** into pParent, which may cause pParent overflow. If this
8088 ** happens, the next iteration of the do-loop will balance pParent
8089 ** use either balance_nonroot() or balance_deeper(). Until this
8090 ** happens, the overflow cell is stored in the aBalanceQuickSpace[]
8093 ** The purpose of the following assert() is to check that only a
8094 ** single call to balance_quick() is made for each call to this
8095 ** function. If this were not verified, a subtle bug involving reuse
8096 ** of the aBalanceQuickSpace[] might sneak in.
8098 assert( balance_quick_called
==0 );
8099 VVA_ONLY( balance_quick_called
++ );
8100 rc
= balance_quick(pParent
, pPage
, aBalanceQuickSpace
);
8104 /* In this case, call balance_nonroot() to redistribute cells
8105 ** between pPage and up to 2 of its sibling pages. This involves
8106 ** modifying the contents of pParent, which may cause pParent to
8107 ** become overfull or underfull. The next iteration of the do-loop
8108 ** will balance the parent page to correct this.
8110 ** If the parent page becomes overfull, the overflow cell or cells
8111 ** are stored in the pSpace buffer allocated immediately below.
8112 ** A subsequent iteration of the do-loop will deal with this by
8113 ** calling balance_nonroot() (balance_deeper() may be called first,
8114 ** but it doesn't deal with overflow cells - just moves them to a
8115 ** different page). Once this subsequent call to balance_nonroot()
8116 ** has completed, it is safe to release the pSpace buffer used by
8117 ** the previous call, as the overflow cell data will have been
8118 ** copied either into the body of a database page or into the new
8119 ** pSpace buffer passed to the latter call to balance_nonroot().
8121 u8
*pSpace
= sqlite3PageMalloc(pCur
->pBt
->pageSize
);
8122 rc
= balance_nonroot(pParent
, iIdx
, pSpace
, iPage
==1,
8123 pCur
->hints
&BTREE_BULKLOAD
);
8125 /* If pFree is not NULL, it points to the pSpace buffer used
8126 ** by a previous call to balance_nonroot(). Its contents are
8127 ** now stored either on real database pages or within the
8128 ** new pSpace buffer, so it may be safely freed here. */
8129 sqlite3PageFree(pFree
);
8132 /* The pSpace buffer will be freed after the next call to
8133 ** balance_nonroot(), or just before this function returns, whichever
8139 pPage
->nOverflow
= 0;
8141 /* The next iteration of the do-loop balances the parent page. */
8144 assert( pCur
->iPage
>=0 );
8145 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
8147 }while( rc
==SQLITE_OK
);
8150 sqlite3PageFree(pFree
);
8155 /* Overwrite content from pX into pDest. Only do the write if the
8156 ** content is different from what is already there.
8158 static int btreeOverwriteContent(
8159 MemPage
*pPage
, /* MemPage on which writing will occur */
8160 u8
*pDest
, /* Pointer to the place to start writing */
8161 const BtreePayload
*pX
, /* Source of data to write */
8162 int iOffset
, /* Offset of first byte to write */
8163 int iAmt
/* Number of bytes to be written */
8165 int nData
= pX
->nData
- iOffset
;
8167 /* Overwritting with zeros */
8169 for(i
=0; i
<iAmt
&& pDest
[i
]==0; i
++){}
8171 int rc
= sqlite3PagerWrite(pPage
->pDbPage
);
8173 memset(pDest
+ i
, 0, iAmt
- i
);
8177 /* Mixed read data and zeros at the end. Make a recursive call
8178 ** to write the zeros then fall through to write the real data */
8179 int rc
= btreeOverwriteContent(pPage
, pDest
+nData
, pX
, iOffset
+nData
,
8184 if( memcmp(pDest
, ((u8
*)pX
->pData
) + iOffset
, iAmt
)!=0 ){
8185 int rc
= sqlite3PagerWrite(pPage
->pDbPage
);
8187 memcpy(pDest
, ((u8
*)pX
->pData
) + iOffset
, iAmt
);
8194 ** Overwrite the cell that cursor pCur is pointing to with fresh content
8197 static int btreeOverwriteCell(BtCursor
*pCur
, const BtreePayload
*pX
){
8198 int iOffset
; /* Next byte of pX->pData to write */
8199 int nTotal
= pX
->nData
+ pX
->nZero
; /* Total bytes of to write */
8200 int rc
; /* Return code */
8201 MemPage
*pPage
= pCur
->pPage
; /* Page being written */
8202 BtShared
*pBt
; /* Btree */
8203 Pgno ovflPgno
; /* Next overflow page to write */
8204 u32 ovflPageSize
; /* Size to write on overflow page */
8206 if( pCur
->info
.pPayload
+ pCur
->info
.nLocal
> pPage
->aDataEnd
){
8207 return SQLITE_CORRUPT_BKPT
;
8209 /* Overwrite the local portion first */
8210 rc
= btreeOverwriteContent(pPage
, pCur
->info
.pPayload
, pX
,
8211 0, pCur
->info
.nLocal
);
8213 if( pCur
->info
.nLocal
==nTotal
) return SQLITE_OK
;
8215 /* Now overwrite the overflow pages */
8216 iOffset
= pCur
->info
.nLocal
;
8217 ovflPgno
= get4byte(pCur
->info
.pPayload
+ iOffset
);
8219 ovflPageSize
= pBt
->usableSize
- 4;
8221 rc
= btreeGetPage(pBt
, ovflPgno
, &pPage
, 0);
8223 if( sqlite3PagerPageRefcount(pPage
->pDbPage
)!=1 ){
8224 rc
= SQLITE_CORRUPT_BKPT
;
8226 if( iOffset
+ovflPageSize
<nTotal
){
8227 ovflPgno
= get4byte(pPage
->aData
);
8229 ovflPageSize
= nTotal
- iOffset
;
8231 rc
= btreeOverwriteContent(pPage
, pPage
->aData
+4, pX
,
8232 iOffset
, ovflPageSize
);
8234 sqlite3PagerUnref(pPage
->pDbPage
);
8236 iOffset
+= ovflPageSize
;
8237 }while( iOffset
<nTotal
);
8243 ** Insert a new record into the BTree. The content of the new record
8244 ** is described by the pX object. The pCur cursor is used only to
8245 ** define what table the record should be inserted into, and is left
8246 ** pointing at a random location.
8248 ** For a table btree (used for rowid tables), only the pX.nKey value of
8249 ** the key is used. The pX.pKey value must be NULL. The pX.nKey is the
8250 ** rowid or INTEGER PRIMARY KEY of the row. The pX.nData,pData,nZero fields
8251 ** hold the content of the row.
8253 ** For an index btree (used for indexes and WITHOUT ROWID tables), the
8254 ** key is an arbitrary byte sequence stored in pX.pKey,nKey. The
8255 ** pX.pData,nData,nZero fields must be zero.
8257 ** If the seekResult parameter is non-zero, then a successful call to
8258 ** MovetoUnpacked() to seek cursor pCur to (pKey,nKey) has already
8259 ** been performed. In other words, if seekResult!=0 then the cursor
8260 ** is currently pointing to a cell that will be adjacent to the cell
8261 ** to be inserted. If seekResult<0 then pCur points to a cell that is
8262 ** smaller then (pKey,nKey). If seekResult>0 then pCur points to a cell
8263 ** that is larger than (pKey,nKey).
8265 ** If seekResult==0, that means pCur is pointing at some unknown location.
8266 ** In that case, this routine must seek the cursor to the correct insertion
8267 ** point for (pKey,nKey) before doing the insertion. For index btrees,
8268 ** if pX->nMem is non-zero, then pX->aMem contains pointers to the unpacked
8269 ** key values and pX->aMem can be used instead of pX->pKey to avoid having
8270 ** to decode the key.
8272 int sqlite3BtreeInsert(
8273 BtCursor
*pCur
, /* Insert data into the table of this cursor */
8274 const BtreePayload
*pX
, /* Content of the row to be inserted */
8275 int flags
, /* True if this is likely an append */
8276 int seekResult
/* Result of prior MovetoUnpacked() call */
8279 int loc
= seekResult
; /* -1: before desired location +1: after */
8283 Btree
*p
= pCur
->pBtree
;
8284 BtShared
*pBt
= p
->pBt
;
8285 unsigned char *oldCell
;
8286 unsigned char *newCell
= 0;
8288 assert( (flags
& (BTREE_SAVEPOSITION
|BTREE_APPEND
))==flags
);
8290 if( pCur
->eState
==CURSOR_FAULT
){
8291 assert( pCur
->skipNext
!=SQLITE_OK
);
8292 return pCur
->skipNext
;
8295 assert( cursorOwnsBtShared(pCur
) );
8296 assert( (pCur
->curFlags
& BTCF_WriteFlag
)!=0
8297 && pBt
->inTransaction
==TRANS_WRITE
8298 && (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
8299 assert( hasSharedCacheTableLock(p
, pCur
->pgnoRoot
, pCur
->pKeyInfo
!=0, 2) );
8301 /* Assert that the caller has been consistent. If this cursor was opened
8302 ** expecting an index b-tree, then the caller should be inserting blob
8303 ** keys with no associated data. If the cursor was opened expecting an
8304 ** intkey table, the caller should be inserting integer keys with a
8305 ** blob of associated data. */
8306 assert( (pX
->pKey
==0)==(pCur
->pKeyInfo
==0) );
8308 /* Save the positions of any other cursors open on this table.
8310 ** In some cases, the call to btreeMoveto() below is a no-op. For
8311 ** example, when inserting data into a table with auto-generated integer
8312 ** keys, the VDBE layer invokes sqlite3BtreeLast() to figure out the
8313 ** integer key to use. It then calls this function to actually insert the
8314 ** data into the intkey B-Tree. In this case btreeMoveto() recognizes
8315 ** that the cursor is already where it needs to be and returns without
8316 ** doing any work. To avoid thwarting these optimizations, it is important
8317 ** not to clear the cursor here.
8319 if( pCur
->curFlags
& BTCF_Multiple
){
8320 rc
= saveAllCursors(pBt
, pCur
->pgnoRoot
, pCur
);
8324 if( pCur
->pKeyInfo
==0 ){
8325 assert( pX
->pKey
==0 );
8326 /* If this is an insert into a table b-tree, invalidate any incrblob
8327 ** cursors open on the row being replaced */
8328 invalidateIncrblobCursors(p
, pCur
->pgnoRoot
, pX
->nKey
, 0);
8330 /* If BTREE_SAVEPOSITION is set, the cursor must already be pointing
8331 ** to a row with the same key as the new entry being inserted. */
8332 assert( (flags
& BTREE_SAVEPOSITION
)==0 ||
8333 ((pCur
->curFlags
&BTCF_ValidNKey
)!=0 && pX
->nKey
==pCur
->info
.nKey
) );
8335 /* If the cursor is currently on the last row and we are appending a
8336 ** new row onto the end, set the "loc" to avoid an unnecessary
8337 ** btreeMoveto() call */
8338 if( (pCur
->curFlags
&BTCF_ValidNKey
)!=0 && pX
->nKey
==pCur
->info
.nKey
){
8339 /* The current is currently pointing to the entry that is to be
8341 if( pCur
->info
.nSize
!=0 && pCur
->info
.nPayload
==pX
->nData
+pX
->nZero
){
8342 return btreeOverwriteCell(pCur
, pX
);
8346 rc
= sqlite3BtreeMovetoUnpacked(pCur
, 0, pX
->nKey
, flags
!=0, &loc
);
8349 }else if( loc
==0 && (flags
& BTREE_SAVEPOSITION
)==0 ){
8352 r
.pKeyInfo
= pCur
->pKeyInfo
;
8354 r
.nField
= pX
->nMem
;
8360 rc
= sqlite3BtreeMovetoUnpacked(pCur
, &r
, 0, flags
!=0, &loc
);
8362 rc
= btreeMoveto(pCur
, pX
->pKey
, pX
->nKey
, flags
!=0, &loc
);
8366 assert( pCur
->eState
==CURSOR_VALID
|| (pCur
->eState
==CURSOR_INVALID
&& loc
) );
8368 pPage
= pCur
->pPage
;
8369 assert( pPage
->intKey
|| pX
->nKey
>=0 );
8370 assert( pPage
->leaf
|| !pPage
->intKey
);
8372 TRACE(("INSERT: table=%d nkey=%lld ndata=%d page=%d %s\n",
8373 pCur
->pgnoRoot
, pX
->nKey
, pX
->nData
, pPage
->pgno
,
8374 loc
==0 ? "overwrite" : "new entry"));
8375 assert( pPage
->isInit
);
8376 newCell
= pBt
->pTmpSpace
;
8377 assert( newCell
!=0 );
8378 rc
= fillInCell(pPage
, newCell
, pX
, &szNew
);
8379 if( rc
) goto end_insert
;
8380 assert( szNew
==pPage
->xCellSize(pPage
, newCell
) );
8381 assert( szNew
<= MX_CELL_SIZE(pBt
) );
8385 assert( idx
<pPage
->nCell
);
8386 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
8390 oldCell
= findCell(pPage
, idx
);
8392 memcpy(newCell
, oldCell
, 4);
8394 rc
= clearCell(pPage
, oldCell
, &info
);
8395 if( info
.nSize
==szNew
&& info
.nLocal
==info
.nPayload
8396 && (!ISAUTOVACUUM
|| szNew
<pPage
->minLocal
)
8398 /* Overwrite the old cell with the new if they are the same size.
8399 ** We could also try to do this if the old cell is smaller, then add
8400 ** the leftover space to the free list. But experiments show that
8401 ** doing that is no faster then skipping this optimization and just
8402 ** calling dropCell() and insertCell().
8404 ** This optimization cannot be used on an autovacuum database if the
8405 ** new entry uses overflow pages, as the insertCell() call below is
8406 ** necessary to add the PTRMAP_OVERFLOW1 pointer-map entry. */
8407 assert( rc
==SQLITE_OK
); /* clearCell never fails when nLocal==nPayload */
8408 if( oldCell
+szNew
> pPage
->aDataEnd
) return SQLITE_CORRUPT_BKPT
;
8409 memcpy(oldCell
, newCell
, szNew
);
8412 dropCell(pPage
, idx
, info
.nSize
, &rc
);
8413 if( rc
) goto end_insert
;
8414 }else if( loc
<0 && pPage
->nCell
>0 ){
8415 assert( pPage
->leaf
);
8417 pCur
->curFlags
&= ~BTCF_ValidNKey
;
8419 assert( pPage
->leaf
);
8421 insertCell(pPage
, idx
, newCell
, szNew
, 0, 0, &rc
);
8422 assert( pPage
->nOverflow
==0 || rc
==SQLITE_OK
);
8423 assert( rc
!=SQLITE_OK
|| pPage
->nCell
>0 || pPage
->nOverflow
>0 );
8425 /* If no error has occurred and pPage has an overflow cell, call balance()
8426 ** to redistribute the cells within the tree. Since balance() may move
8427 ** the cursor, zero the BtCursor.info.nSize and BTCF_ValidNKey
8430 ** Previous versions of SQLite called moveToRoot() to move the cursor
8431 ** back to the root page as balance() used to invalidate the contents
8432 ** of BtCursor.apPage[] and BtCursor.aiIdx[]. Instead of doing that,
8433 ** set the cursor state to "invalid". This makes common insert operations
8436 ** There is a subtle but important optimization here too. When inserting
8437 ** multiple records into an intkey b-tree using a single cursor (as can
8438 ** happen while processing an "INSERT INTO ... SELECT" statement), it
8439 ** is advantageous to leave the cursor pointing to the last entry in
8440 ** the b-tree if possible. If the cursor is left pointing to the last
8441 ** entry in the table, and the next row inserted has an integer key
8442 ** larger than the largest existing key, it is possible to insert the
8443 ** row without seeking the cursor. This can be a big performance boost.
8445 pCur
->info
.nSize
= 0;
8446 if( pPage
->nOverflow
){
8447 assert( rc
==SQLITE_OK
);
8448 pCur
->curFlags
&= ~(BTCF_ValidNKey
);
8451 /* Must make sure nOverflow is reset to zero even if the balance()
8452 ** fails. Internal data structure corruption will result otherwise.
8453 ** Also, set the cursor state to invalid. This stops saveCursorPosition()
8454 ** from trying to save the current position of the cursor. */
8455 pCur
->pPage
->nOverflow
= 0;
8456 pCur
->eState
= CURSOR_INVALID
;
8457 if( (flags
& BTREE_SAVEPOSITION
) && rc
==SQLITE_OK
){
8458 btreeReleaseAllCursorPages(pCur
);
8459 if( pCur
->pKeyInfo
){
8460 assert( pCur
->pKey
==0 );
8461 pCur
->pKey
= sqlite3Malloc( pX
->nKey
);
8462 if( pCur
->pKey
==0 ){
8465 memcpy(pCur
->pKey
, pX
->pKey
, pX
->nKey
);
8468 pCur
->eState
= CURSOR_REQUIRESEEK
;
8469 pCur
->nKey
= pX
->nKey
;
8472 assert( pCur
->iPage
<0 || pCur
->pPage
->nOverflow
==0 );
8479 ** Delete the entry that the cursor is pointing to.
8481 ** If the BTREE_SAVEPOSITION bit of the flags parameter is zero, then
8482 ** the cursor is left pointing at an arbitrary location after the delete.
8483 ** But if that bit is set, then the cursor is left in a state such that
8484 ** the next call to BtreeNext() or BtreePrev() moves it to the same row
8485 ** as it would have been on if the call to BtreeDelete() had been omitted.
8487 ** The BTREE_AUXDELETE bit of flags indicates that is one of several deletes
8488 ** associated with a single table entry and its indexes. Only one of those
8489 ** deletes is considered the "primary" delete. The primary delete occurs
8490 ** on a cursor that is not a BTREE_FORDELETE cursor. All but one delete
8491 ** operation on non-FORDELETE cursors is tagged with the AUXDELETE flag.
8492 ** The BTREE_AUXDELETE bit is a hint that is not used by this implementation,
8493 ** but which might be used by alternative storage engines.
8495 int sqlite3BtreeDelete(BtCursor
*pCur
, u8 flags
){
8496 Btree
*p
= pCur
->pBtree
;
8497 BtShared
*pBt
= p
->pBt
;
8498 int rc
; /* Return code */
8499 MemPage
*pPage
; /* Page to delete cell from */
8500 unsigned char *pCell
; /* Pointer to cell to delete */
8501 int iCellIdx
; /* Index of cell to delete */
8502 int iCellDepth
; /* Depth of node containing pCell */
8503 CellInfo info
; /* Size of the cell being deleted */
8504 int bSkipnext
= 0; /* Leaf cursor in SKIPNEXT state */
8505 u8 bPreserve
= flags
& BTREE_SAVEPOSITION
; /* Keep cursor valid */
8507 assert( cursorOwnsBtShared(pCur
) );
8508 assert( pBt
->inTransaction
==TRANS_WRITE
);
8509 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
8510 assert( pCur
->curFlags
& BTCF_WriteFlag
);
8511 assert( hasSharedCacheTableLock(p
, pCur
->pgnoRoot
, pCur
->pKeyInfo
!=0, 2) );
8512 assert( !hasReadConflicts(p
, pCur
->pgnoRoot
) );
8513 assert( pCur
->ix
<pCur
->pPage
->nCell
);
8514 assert( pCur
->eState
==CURSOR_VALID
);
8515 assert( (flags
& ~(BTREE_SAVEPOSITION
| BTREE_AUXDELETE
))==0 );
8517 iCellDepth
= pCur
->iPage
;
8518 iCellIdx
= pCur
->ix
;
8519 pPage
= pCur
->pPage
;
8520 pCell
= findCell(pPage
, iCellIdx
);
8522 /* If the bPreserve flag is set to true, then the cursor position must
8523 ** be preserved following this delete operation. If the current delete
8524 ** will cause a b-tree rebalance, then this is done by saving the cursor
8525 ** key and leaving the cursor in CURSOR_REQUIRESEEK state before
8528 ** Or, if the current delete will not cause a rebalance, then the cursor
8529 ** will be left in CURSOR_SKIPNEXT state pointing to the entry immediately
8530 ** before or after the deleted entry. In this case set bSkipnext to true. */
8533 || (pPage
->nFree
+cellSizePtr(pPage
,pCell
)+2)>(int)(pBt
->usableSize
*2/3)
8535 /* A b-tree rebalance will be required after deleting this entry.
8536 ** Save the cursor key. */
8537 rc
= saveCursorKey(pCur
);
8544 /* If the page containing the entry to delete is not a leaf page, move
8545 ** the cursor to the largest entry in the tree that is smaller than
8546 ** the entry being deleted. This cell will replace the cell being deleted
8547 ** from the internal node. The 'previous' entry is used for this instead
8548 ** of the 'next' entry, as the previous entry is always a part of the
8549 ** sub-tree headed by the child page of the cell being deleted. This makes
8550 ** balancing the tree following the delete operation easier. */
8552 rc
= sqlite3BtreePrevious(pCur
, 0);
8553 assert( rc
!=SQLITE_DONE
);
8557 /* Save the positions of any other cursors open on this table before
8558 ** making any modifications. */
8559 if( pCur
->curFlags
& BTCF_Multiple
){
8560 rc
= saveAllCursors(pBt
, pCur
->pgnoRoot
, pCur
);
8564 /* If this is a delete operation to remove a row from a table b-tree,
8565 ** invalidate any incrblob cursors open on the row being deleted. */
8566 if( pCur
->pKeyInfo
==0 ){
8567 invalidateIncrblobCursors(p
, pCur
->pgnoRoot
, pCur
->info
.nKey
, 0);
8570 /* Make the page containing the entry to be deleted writable. Then free any
8571 ** overflow pages associated with the entry and finally remove the cell
8572 ** itself from within the page. */
8573 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
8575 rc
= clearCell(pPage
, pCell
, &info
);
8576 dropCell(pPage
, iCellIdx
, info
.nSize
, &rc
);
8579 /* If the cell deleted was not located on a leaf page, then the cursor
8580 ** is currently pointing to the largest entry in the sub-tree headed
8581 ** by the child-page of the cell that was just deleted from an internal
8582 ** node. The cell from the leaf node needs to be moved to the internal
8583 ** node to replace the deleted cell. */
8585 MemPage
*pLeaf
= pCur
->pPage
;
8588 unsigned char *pTmp
;
8590 if( iCellDepth
<pCur
->iPage
-1 ){
8591 n
= pCur
->apPage
[iCellDepth
+1]->pgno
;
8593 n
= pCur
->pPage
->pgno
;
8595 pCell
= findCell(pLeaf
, pLeaf
->nCell
-1);
8596 if( pCell
<&pLeaf
->aData
[4] ) return SQLITE_CORRUPT_BKPT
;
8597 nCell
= pLeaf
->xCellSize(pLeaf
, pCell
);
8598 assert( MX_CELL_SIZE(pBt
) >= nCell
);
8599 pTmp
= pBt
->pTmpSpace
;
8601 rc
= sqlite3PagerWrite(pLeaf
->pDbPage
);
8602 if( rc
==SQLITE_OK
){
8603 insertCell(pPage
, iCellIdx
, pCell
-4, nCell
+4, pTmp
, n
, &rc
);
8605 dropCell(pLeaf
, pLeaf
->nCell
-1, nCell
, &rc
);
8609 /* Balance the tree. If the entry deleted was located on a leaf page,
8610 ** then the cursor still points to that page. In this case the first
8611 ** call to balance() repairs the tree, and the if(...) condition is
8614 ** Otherwise, if the entry deleted was on an internal node page, then
8615 ** pCur is pointing to the leaf page from which a cell was removed to
8616 ** replace the cell deleted from the internal node. This is slightly
8617 ** tricky as the leaf node may be underfull, and the internal node may
8618 ** be either under or overfull. In this case run the balancing algorithm
8619 ** on the leaf node first. If the balance proceeds far enough up the
8620 ** tree that we can be sure that any problem in the internal node has
8621 ** been corrected, so be it. Otherwise, after balancing the leaf node,
8622 ** walk the cursor up the tree to the internal node and balance it as
8625 if( rc
==SQLITE_OK
&& pCur
->iPage
>iCellDepth
){
8626 releasePageNotNull(pCur
->pPage
);
8628 while( pCur
->iPage
>iCellDepth
){
8629 releasePage(pCur
->apPage
[pCur
->iPage
--]);
8631 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
8635 if( rc
==SQLITE_OK
){
8637 assert( bPreserve
&& (pCur
->iPage
==iCellDepth
|| CORRUPT_DB
) );
8638 assert( pPage
==pCur
->pPage
|| CORRUPT_DB
);
8639 assert( (pPage
->nCell
>0 || CORRUPT_DB
) && iCellIdx
<=pPage
->nCell
);
8640 pCur
->eState
= CURSOR_SKIPNEXT
;
8641 if( iCellIdx
>=pPage
->nCell
){
8642 pCur
->skipNext
= -1;
8643 pCur
->ix
= pPage
->nCell
-1;
8648 rc
= moveToRoot(pCur
);
8650 btreeReleaseAllCursorPages(pCur
);
8651 pCur
->eState
= CURSOR_REQUIRESEEK
;
8653 if( rc
==SQLITE_EMPTY
) rc
= SQLITE_OK
;
8660 ** Create a new BTree table. Write into *piTable the page
8661 ** number for the root page of the new table.
8663 ** The type of type is determined by the flags parameter. Only the
8664 ** following values of flags are currently in use. Other values for
8665 ** flags might not work:
8667 ** BTREE_INTKEY|BTREE_LEAFDATA Used for SQL tables with rowid keys
8668 ** BTREE_ZERODATA Used for SQL indices
8670 static int btreeCreateTable(Btree
*p
, int *piTable
, int createTabFlags
){
8671 BtShared
*pBt
= p
->pBt
;
8675 int ptfFlags
; /* Page-type flage for the root page of new table */
8677 assert( sqlite3BtreeHoldsMutex(p
) );
8678 assert( pBt
->inTransaction
==TRANS_WRITE
);
8679 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
8681 #ifdef SQLITE_OMIT_AUTOVACUUM
8682 rc
= allocateBtreePage(pBt
, &pRoot
, &pgnoRoot
, 1, 0);
8687 if( pBt
->autoVacuum
){
8688 Pgno pgnoMove
; /* Move a page here to make room for the root-page */
8689 MemPage
*pPageMove
; /* The page to move to. */
8691 /* Creating a new table may probably require moving an existing database
8692 ** to make room for the new tables root page. In case this page turns
8693 ** out to be an overflow page, delete all overflow page-map caches
8694 ** held by open cursors.
8696 invalidateAllOverflowCache(pBt
);
8698 /* Read the value of meta[3] from the database to determine where the
8699 ** root page of the new table should go. meta[3] is the largest root-page
8700 ** created so far, so the new root-page is (meta[3]+1).
8702 sqlite3BtreeGetMeta(p
, BTREE_LARGEST_ROOT_PAGE
, &pgnoRoot
);
8705 /* The new root-page may not be allocated on a pointer-map page, or the
8706 ** PENDING_BYTE page.
8708 while( pgnoRoot
==PTRMAP_PAGENO(pBt
, pgnoRoot
) ||
8709 pgnoRoot
==PENDING_BYTE_PAGE(pBt
) ){
8712 assert( pgnoRoot
>=3 || CORRUPT_DB
);
8713 testcase( pgnoRoot
<3 );
8715 /* Allocate a page. The page that currently resides at pgnoRoot will
8716 ** be moved to the allocated page (unless the allocated page happens
8717 ** to reside at pgnoRoot).
8719 rc
= allocateBtreePage(pBt
, &pPageMove
, &pgnoMove
, pgnoRoot
, BTALLOC_EXACT
);
8720 if( rc
!=SQLITE_OK
){
8724 if( pgnoMove
!=pgnoRoot
){
8725 /* pgnoRoot is the page that will be used for the root-page of
8726 ** the new table (assuming an error did not occur). But we were
8727 ** allocated pgnoMove. If required (i.e. if it was not allocated
8728 ** by extending the file), the current page at position pgnoMove
8729 ** is already journaled.
8734 /* Save the positions of any open cursors. This is required in
8735 ** case they are holding a reference to an xFetch reference
8736 ** corresponding to page pgnoRoot. */
8737 rc
= saveAllCursors(pBt
, 0, 0);
8738 releasePage(pPageMove
);
8739 if( rc
!=SQLITE_OK
){
8743 /* Move the page currently at pgnoRoot to pgnoMove. */
8744 rc
= btreeGetPage(pBt
, pgnoRoot
, &pRoot
, 0);
8745 if( rc
!=SQLITE_OK
){
8748 rc
= ptrmapGet(pBt
, pgnoRoot
, &eType
, &iPtrPage
);
8749 if( eType
==PTRMAP_ROOTPAGE
|| eType
==PTRMAP_FREEPAGE
){
8750 rc
= SQLITE_CORRUPT_BKPT
;
8752 if( rc
!=SQLITE_OK
){
8756 assert( eType
!=PTRMAP_ROOTPAGE
);
8757 assert( eType
!=PTRMAP_FREEPAGE
);
8758 rc
= relocatePage(pBt
, pRoot
, eType
, iPtrPage
, pgnoMove
, 0);
8761 /* Obtain the page at pgnoRoot */
8762 if( rc
!=SQLITE_OK
){
8765 rc
= btreeGetPage(pBt
, pgnoRoot
, &pRoot
, 0);
8766 if( rc
!=SQLITE_OK
){
8769 rc
= sqlite3PagerWrite(pRoot
->pDbPage
);
8770 if( rc
!=SQLITE_OK
){
8778 /* Update the pointer-map and meta-data with the new root-page number. */
8779 ptrmapPut(pBt
, pgnoRoot
, PTRMAP_ROOTPAGE
, 0, &rc
);
8785 /* When the new root page was allocated, page 1 was made writable in
8786 ** order either to increase the database filesize, or to decrement the
8787 ** freelist count. Hence, the sqlite3BtreeUpdateMeta() call cannot fail.
8789 assert( sqlite3PagerIswriteable(pBt
->pPage1
->pDbPage
) );
8790 rc
= sqlite3BtreeUpdateMeta(p
, 4, pgnoRoot
);
8797 rc
= allocateBtreePage(pBt
, &pRoot
, &pgnoRoot
, 1, 0);
8801 assert( sqlite3PagerIswriteable(pRoot
->pDbPage
) );
8802 if( createTabFlags
& BTREE_INTKEY
){
8803 ptfFlags
= PTF_INTKEY
| PTF_LEAFDATA
| PTF_LEAF
;
8805 ptfFlags
= PTF_ZERODATA
| PTF_LEAF
;
8807 zeroPage(pRoot
, ptfFlags
);
8808 sqlite3PagerUnref(pRoot
->pDbPage
);
8809 assert( (pBt
->openFlags
& BTREE_SINGLE
)==0 || pgnoRoot
==2 );
8810 *piTable
= (int)pgnoRoot
;
8813 int sqlite3BtreeCreateTable(Btree
*p
, int *piTable
, int flags
){
8815 sqlite3BtreeEnter(p
);
8816 rc
= btreeCreateTable(p
, piTable
, flags
);
8817 sqlite3BtreeLeave(p
);
8822 ** Erase the given database page and all its children. Return
8823 ** the page to the freelist.
8825 static int clearDatabasePage(
8826 BtShared
*pBt
, /* The BTree that contains the table */
8827 Pgno pgno
, /* Page number to clear */
8828 int freePageFlag
, /* Deallocate page if true */
8829 int *pnChange
/* Add number of Cells freed to this counter */
8833 unsigned char *pCell
;
8838 assert( sqlite3_mutex_held(pBt
->mutex
) );
8839 if( pgno
>btreePagecount(pBt
) ){
8840 return SQLITE_CORRUPT_BKPT
;
8842 rc
= getAndInitPage(pBt
, pgno
, &pPage
, 0, 0);
8845 rc
= SQLITE_CORRUPT_BKPT
;
8846 goto cleardatabasepage_out
;
8849 hdr
= pPage
->hdrOffset
;
8850 for(i
=0; i
<pPage
->nCell
; i
++){
8851 pCell
= findCell(pPage
, i
);
8853 rc
= clearDatabasePage(pBt
, get4byte(pCell
), 1, pnChange
);
8854 if( rc
) goto cleardatabasepage_out
;
8856 rc
= clearCell(pPage
, pCell
, &info
);
8857 if( rc
) goto cleardatabasepage_out
;
8860 rc
= clearDatabasePage(pBt
, get4byte(&pPage
->aData
[hdr
+8]), 1, pnChange
);
8861 if( rc
) goto cleardatabasepage_out
;
8862 }else if( pnChange
){
8863 assert( pPage
->intKey
|| CORRUPT_DB
);
8864 testcase( !pPage
->intKey
);
8865 *pnChange
+= pPage
->nCell
;
8868 freePage(pPage
, &rc
);
8869 }else if( (rc
= sqlite3PagerWrite(pPage
->pDbPage
))==0 ){
8870 zeroPage(pPage
, pPage
->aData
[hdr
] | PTF_LEAF
);
8873 cleardatabasepage_out
:
8880 ** Delete all information from a single table in the database. iTable is
8881 ** the page number of the root of the table. After this routine returns,
8882 ** the root page is empty, but still exists.
8884 ** This routine will fail with SQLITE_LOCKED if there are any open
8885 ** read cursors on the table. Open write cursors are moved to the
8886 ** root of the table.
8888 ** If pnChange is not NULL, then table iTable must be an intkey table. The
8889 ** integer value pointed to by pnChange is incremented by the number of
8890 ** entries in the table.
8892 int sqlite3BtreeClearTable(Btree
*p
, int iTable
, int *pnChange
){
8894 BtShared
*pBt
= p
->pBt
;
8895 sqlite3BtreeEnter(p
);
8896 assert( p
->inTrans
==TRANS_WRITE
);
8898 rc
= saveAllCursors(pBt
, (Pgno
)iTable
, 0);
8900 if( SQLITE_OK
==rc
){
8901 /* Invalidate all incrblob cursors open on table iTable (assuming iTable
8902 ** is the root of a table b-tree - if it is not, the following call is
8904 invalidateIncrblobCursors(p
, (Pgno
)iTable
, 0, 1);
8905 rc
= clearDatabasePage(pBt
, (Pgno
)iTable
, 0, pnChange
);
8907 sqlite3BtreeLeave(p
);
8912 ** Delete all information from the single table that pCur is open on.
8914 ** This routine only work for pCur on an ephemeral table.
8916 int sqlite3BtreeClearTableOfCursor(BtCursor
*pCur
){
8917 return sqlite3BtreeClearTable(pCur
->pBtree
, pCur
->pgnoRoot
, 0);
8921 ** Erase all information in a table and add the root of the table to
8922 ** the freelist. Except, the root of the principle table (the one on
8923 ** page 1) is never added to the freelist.
8925 ** This routine will fail with SQLITE_LOCKED if there are any open
8926 ** cursors on the table.
8928 ** If AUTOVACUUM is enabled and the page at iTable is not the last
8929 ** root page in the database file, then the last root page
8930 ** in the database file is moved into the slot formerly occupied by
8931 ** iTable and that last slot formerly occupied by the last root page
8932 ** is added to the freelist instead of iTable. In this say, all
8933 ** root pages are kept at the beginning of the database file, which
8934 ** is necessary for AUTOVACUUM to work right. *piMoved is set to the
8935 ** page number that used to be the last root page in the file before
8936 ** the move. If no page gets moved, *piMoved is set to 0.
8937 ** The last root page is recorded in meta[3] and the value of
8938 ** meta[3] is updated by this procedure.
8940 static int btreeDropTable(Btree
*p
, Pgno iTable
, int *piMoved
){
8943 BtShared
*pBt
= p
->pBt
;
8945 assert( sqlite3BtreeHoldsMutex(p
) );
8946 assert( p
->inTrans
==TRANS_WRITE
);
8947 assert( iTable
>=2 );
8949 rc
= btreeGetPage(pBt
, (Pgno
)iTable
, &pPage
, 0);
8951 rc
= sqlite3BtreeClearTable(p
, iTable
, 0);
8959 #ifdef SQLITE_OMIT_AUTOVACUUM
8960 freePage(pPage
, &rc
);
8963 if( pBt
->autoVacuum
){
8965 sqlite3BtreeGetMeta(p
, BTREE_LARGEST_ROOT_PAGE
, &maxRootPgno
);
8967 if( iTable
==maxRootPgno
){
8968 /* If the table being dropped is the table with the largest root-page
8969 ** number in the database, put the root page on the free list.
8971 freePage(pPage
, &rc
);
8973 if( rc
!=SQLITE_OK
){
8977 /* The table being dropped does not have the largest root-page
8978 ** number in the database. So move the page that does into the
8979 ** gap left by the deleted root-page.
8983 rc
= btreeGetPage(pBt
, maxRootPgno
, &pMove
, 0);
8984 if( rc
!=SQLITE_OK
){
8987 rc
= relocatePage(pBt
, pMove
, PTRMAP_ROOTPAGE
, 0, iTable
, 0);
8989 if( rc
!=SQLITE_OK
){
8993 rc
= btreeGetPage(pBt
, maxRootPgno
, &pMove
, 0);
8994 freePage(pMove
, &rc
);
8996 if( rc
!=SQLITE_OK
){
8999 *piMoved
= maxRootPgno
;
9002 /* Set the new 'max-root-page' value in the database header. This
9003 ** is the old value less one, less one more if that happens to
9004 ** be a root-page number, less one again if that is the
9005 ** PENDING_BYTE_PAGE.
9008 while( maxRootPgno
==PENDING_BYTE_PAGE(pBt
)
9009 || PTRMAP_ISPAGE(pBt
, maxRootPgno
) ){
9012 assert( maxRootPgno
!=PENDING_BYTE_PAGE(pBt
) );
9014 rc
= sqlite3BtreeUpdateMeta(p
, 4, maxRootPgno
);
9016 freePage(pPage
, &rc
);
9022 int sqlite3BtreeDropTable(Btree
*p
, int iTable
, int *piMoved
){
9024 sqlite3BtreeEnter(p
);
9025 rc
= btreeDropTable(p
, iTable
, piMoved
);
9026 sqlite3BtreeLeave(p
);
9032 ** This function may only be called if the b-tree connection already
9033 ** has a read or write transaction open on the database.
9035 ** Read the meta-information out of a database file. Meta[0]
9036 ** is the number of free pages currently in the database. Meta[1]
9037 ** through meta[15] are available for use by higher layers. Meta[0]
9038 ** is read-only, the others are read/write.
9040 ** The schema layer numbers meta values differently. At the schema
9041 ** layer (and the SetCookie and ReadCookie opcodes) the number of
9042 ** free pages is not visible. So Cookie[0] is the same as Meta[1].
9044 ** This routine treats Meta[BTREE_DATA_VERSION] as a special case. Instead
9045 ** of reading the value out of the header, it instead loads the "DataVersion"
9046 ** from the pager. The BTREE_DATA_VERSION value is not actually stored in the
9047 ** database file. It is a number computed by the pager. But its access
9048 ** pattern is the same as header meta values, and so it is convenient to
9049 ** read it from this routine.
9051 void sqlite3BtreeGetMeta(Btree
*p
, int idx
, u32
*pMeta
){
9052 BtShared
*pBt
= p
->pBt
;
9054 sqlite3BtreeEnter(p
);
9055 assert( p
->inTrans
>TRANS_NONE
);
9056 assert( SQLITE_OK
==querySharedCacheTableLock(p
, MASTER_ROOT
, READ_LOCK
) );
9057 assert( pBt
->pPage1
);
9058 assert( idx
>=0 && idx
<=15 );
9060 if( idx
==BTREE_DATA_VERSION
){
9061 *pMeta
= sqlite3PagerDataVersion(pBt
->pPager
) + p
->iDataVersion
;
9063 *pMeta
= get4byte(&pBt
->pPage1
->aData
[36 + idx
*4]);
9066 /* If auto-vacuum is disabled in this build and this is an auto-vacuum
9067 ** database, mark the database as read-only. */
9068 #ifdef SQLITE_OMIT_AUTOVACUUM
9069 if( idx
==BTREE_LARGEST_ROOT_PAGE
&& *pMeta
>0 ){
9070 pBt
->btsFlags
|= BTS_READ_ONLY
;
9074 sqlite3BtreeLeave(p
);
9078 ** Write meta-information back into the database. Meta[0] is
9079 ** read-only and may not be written.
9081 int sqlite3BtreeUpdateMeta(Btree
*p
, int idx
, u32 iMeta
){
9082 BtShared
*pBt
= p
->pBt
;
9085 assert( idx
>=1 && idx
<=15 );
9086 sqlite3BtreeEnter(p
);
9087 assert( p
->inTrans
==TRANS_WRITE
);
9088 assert( pBt
->pPage1
!=0 );
9089 pP1
= pBt
->pPage1
->aData
;
9090 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
9091 if( rc
==SQLITE_OK
){
9092 put4byte(&pP1
[36 + idx
*4], iMeta
);
9093 #ifndef SQLITE_OMIT_AUTOVACUUM
9094 if( idx
==BTREE_INCR_VACUUM
){
9095 assert( pBt
->autoVacuum
|| iMeta
==0 );
9096 assert( iMeta
==0 || iMeta
==1 );
9097 pBt
->incrVacuum
= (u8
)iMeta
;
9101 sqlite3BtreeLeave(p
);
9105 #ifndef SQLITE_OMIT_BTREECOUNT
9107 ** The first argument, pCur, is a cursor opened on some b-tree. Count the
9108 ** number of entries in the b-tree and write the result to *pnEntry.
9110 ** SQLITE_OK is returned if the operation is successfully executed.
9111 ** Otherwise, if an error is encountered (i.e. an IO error or database
9112 ** corruption) an SQLite error code is returned.
9114 int sqlite3BtreeCount(BtCursor
*pCur
, i64
*pnEntry
){
9115 i64 nEntry
= 0; /* Value to return in *pnEntry */
9116 int rc
; /* Return code */
9118 rc
= moveToRoot(pCur
);
9119 if( rc
==SQLITE_EMPTY
){
9124 /* Unless an error occurs, the following loop runs one iteration for each
9125 ** page in the B-Tree structure (not including overflow pages).
9127 while( rc
==SQLITE_OK
){
9128 int iIdx
; /* Index of child node in parent */
9129 MemPage
*pPage
; /* Current page of the b-tree */
9131 /* If this is a leaf page or the tree is not an int-key tree, then
9132 ** this page contains countable entries. Increment the entry counter
9135 pPage
= pCur
->pPage
;
9136 if( pPage
->leaf
|| !pPage
->intKey
){
9137 nEntry
+= pPage
->nCell
;
9140 /* pPage is a leaf node. This loop navigates the cursor so that it
9141 ** points to the first interior cell that it points to the parent of
9142 ** the next page in the tree that has not yet been visited. The
9143 ** pCur->aiIdx[pCur->iPage] value is set to the index of the parent cell
9144 ** of the page, or to the number of cells in the page if the next page
9145 ** to visit is the right-child of its parent.
9147 ** If all pages in the tree have been visited, return SQLITE_OK to the
9152 if( pCur
->iPage
==0 ){
9153 /* All pages of the b-tree have been visited. Return successfully. */
9155 return moveToRoot(pCur
);
9158 }while ( pCur
->ix
>=pCur
->pPage
->nCell
);
9161 pPage
= pCur
->pPage
;
9164 /* Descend to the child node of the cell that the cursor currently
9165 ** points at. This is the right-child if (iIdx==pPage->nCell).
9168 if( iIdx
==pPage
->nCell
){
9169 rc
= moveToChild(pCur
, get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]));
9171 rc
= moveToChild(pCur
, get4byte(findCell(pPage
, iIdx
)));
9175 /* An error has occurred. Return an error code. */
9181 ** Return the pager associated with a BTree. This routine is used for
9182 ** testing and debugging only.
9184 Pager
*sqlite3BtreePager(Btree
*p
){
9185 return p
->pBt
->pPager
;
9188 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9190 ** Append a message to the error message string.
9192 static void checkAppendMsg(
9193 IntegrityCk
*pCheck
,
9194 const char *zFormat
,
9198 if( !pCheck
->mxErr
) return;
9201 va_start(ap
, zFormat
);
9202 if( pCheck
->errMsg
.nChar
){
9203 sqlite3StrAccumAppend(&pCheck
->errMsg
, "\n", 1);
9206 sqlite3XPrintf(&pCheck
->errMsg
, pCheck
->zPfx
, pCheck
->v1
, pCheck
->v2
);
9208 sqlite3VXPrintf(&pCheck
->errMsg
, zFormat
, ap
);
9210 if( pCheck
->errMsg
.accError
==STRACCUM_NOMEM
){
9211 pCheck
->mallocFailed
= 1;
9214 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9216 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9219 ** Return non-zero if the bit in the IntegrityCk.aPgRef[] array that
9220 ** corresponds to page iPg is already set.
9222 static int getPageReferenced(IntegrityCk
*pCheck
, Pgno iPg
){
9223 assert( iPg
<=pCheck
->nPage
&& sizeof(pCheck
->aPgRef
[0])==1 );
9224 return (pCheck
->aPgRef
[iPg
/8] & (1 << (iPg
& 0x07)));
9228 ** Set the bit in the IntegrityCk.aPgRef[] array that corresponds to page iPg.
9230 static void setPageReferenced(IntegrityCk
*pCheck
, Pgno iPg
){
9231 assert( iPg
<=pCheck
->nPage
&& sizeof(pCheck
->aPgRef
[0])==1 );
9232 pCheck
->aPgRef
[iPg
/8] |= (1 << (iPg
& 0x07));
9237 ** Add 1 to the reference count for page iPage. If this is the second
9238 ** reference to the page, add an error message to pCheck->zErrMsg.
9239 ** Return 1 if there are 2 or more references to the page and 0 if
9240 ** if this is the first reference to the page.
9242 ** Also check that the page number is in bounds.
9244 static int checkRef(IntegrityCk
*pCheck
, Pgno iPage
){
9245 if( iPage
==0 ) return 1;
9246 if( iPage
>pCheck
->nPage
){
9247 checkAppendMsg(pCheck
, "invalid page number %d", iPage
);
9250 if( getPageReferenced(pCheck
, iPage
) ){
9251 checkAppendMsg(pCheck
, "2nd reference to page %d", iPage
);
9254 setPageReferenced(pCheck
, iPage
);
9258 #ifndef SQLITE_OMIT_AUTOVACUUM
9260 ** Check that the entry in the pointer-map for page iChild maps to
9261 ** page iParent, pointer type ptrType. If not, append an error message
9264 static void checkPtrmap(
9265 IntegrityCk
*pCheck
, /* Integrity check context */
9266 Pgno iChild
, /* Child page number */
9267 u8 eType
, /* Expected pointer map type */
9268 Pgno iParent
/* Expected pointer map parent page number */
9274 rc
= ptrmapGet(pCheck
->pBt
, iChild
, &ePtrmapType
, &iPtrmapParent
);
9275 if( rc
!=SQLITE_OK
){
9276 if( rc
==SQLITE_NOMEM
|| rc
==SQLITE_IOERR_NOMEM
) pCheck
->mallocFailed
= 1;
9277 checkAppendMsg(pCheck
, "Failed to read ptrmap key=%d", iChild
);
9281 if( ePtrmapType
!=eType
|| iPtrmapParent
!=iParent
){
9282 checkAppendMsg(pCheck
,
9283 "Bad ptr map entry key=%d expected=(%d,%d) got=(%d,%d)",
9284 iChild
, eType
, iParent
, ePtrmapType
, iPtrmapParent
);
9290 ** Check the integrity of the freelist or of an overflow page list.
9291 ** Verify that the number of pages on the list is N.
9293 static void checkList(
9294 IntegrityCk
*pCheck
, /* Integrity checking context */
9295 int isFreeList
, /* True for a freelist. False for overflow page list */
9296 int iPage
, /* Page number for first page in the list */
9297 int N
/* Expected number of pages in the list */
9302 while( N
-- > 0 && pCheck
->mxErr
){
9304 unsigned char *pOvflData
;
9306 checkAppendMsg(pCheck
,
9307 "%d of %d pages missing from overflow list starting at %d",
9308 N
+1, expected
, iFirst
);
9311 if( checkRef(pCheck
, iPage
) ) break;
9312 if( sqlite3PagerGet(pCheck
->pPager
, (Pgno
)iPage
, &pOvflPage
, 0) ){
9313 checkAppendMsg(pCheck
, "failed to get page %d", iPage
);
9316 pOvflData
= (unsigned char *)sqlite3PagerGetData(pOvflPage
);
9318 int n
= get4byte(&pOvflData
[4]);
9319 #ifndef SQLITE_OMIT_AUTOVACUUM
9320 if( pCheck
->pBt
->autoVacuum
){
9321 checkPtrmap(pCheck
, iPage
, PTRMAP_FREEPAGE
, 0);
9324 if( n
>(int)pCheck
->pBt
->usableSize
/4-2 ){
9325 checkAppendMsg(pCheck
,
9326 "freelist leaf count too big on page %d", iPage
);
9330 Pgno iFreePage
= get4byte(&pOvflData
[8+i
*4]);
9331 #ifndef SQLITE_OMIT_AUTOVACUUM
9332 if( pCheck
->pBt
->autoVacuum
){
9333 checkPtrmap(pCheck
, iFreePage
, PTRMAP_FREEPAGE
, 0);
9336 checkRef(pCheck
, iFreePage
);
9341 #ifndef SQLITE_OMIT_AUTOVACUUM
9343 /* If this database supports auto-vacuum and iPage is not the last
9344 ** page in this overflow list, check that the pointer-map entry for
9345 ** the following page matches iPage.
9347 if( pCheck
->pBt
->autoVacuum
&& N
>0 ){
9348 i
= get4byte(pOvflData
);
9349 checkPtrmap(pCheck
, i
, PTRMAP_OVERFLOW2
, iPage
);
9353 iPage
= get4byte(pOvflData
);
9354 sqlite3PagerUnref(pOvflPage
);
9356 if( isFreeList
&& N
<(iPage
!=0) ){
9357 checkAppendMsg(pCheck
, "free-page count in header is too small");
9361 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9364 ** An implementation of a min-heap.
9366 ** aHeap[0] is the number of elements on the heap. aHeap[1] is the
9367 ** root element. The daughter nodes of aHeap[N] are aHeap[N*2]
9368 ** and aHeap[N*2+1].
9370 ** The heap property is this: Every node is less than or equal to both
9371 ** of its daughter nodes. A consequence of the heap property is that the
9372 ** root node aHeap[1] is always the minimum value currently in the heap.
9374 ** The btreeHeapInsert() routine inserts an unsigned 32-bit number onto
9375 ** the heap, preserving the heap property. The btreeHeapPull() routine
9376 ** removes the root element from the heap (the minimum value in the heap)
9377 ** and then moves other nodes around as necessary to preserve the heap
9380 ** This heap is used for cell overlap and coverage testing. Each u32
9381 ** entry represents the span of a cell or freeblock on a btree page.
9382 ** The upper 16 bits are the index of the first byte of a range and the
9383 ** lower 16 bits are the index of the last byte of that range.
9385 static void btreeHeapInsert(u32
*aHeap
, u32 x
){
9386 u32 j
, i
= ++aHeap
[0];
9388 while( (j
= i
/2)>0 && aHeap
[j
]>aHeap
[i
] ){
9390 aHeap
[j
] = aHeap
[i
];
9395 static int btreeHeapPull(u32
*aHeap
, u32
*pOut
){
9397 if( (x
= aHeap
[0])==0 ) return 0;
9399 aHeap
[1] = aHeap
[x
];
9400 aHeap
[x
] = 0xffffffff;
9403 while( (j
= i
*2)<=aHeap
[0] ){
9404 if( aHeap
[j
]>aHeap
[j
+1] ) j
++;
9405 if( aHeap
[i
]<aHeap
[j
] ) break;
9407 aHeap
[i
] = aHeap
[j
];
9414 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9416 ** Do various sanity checks on a single page of a tree. Return
9417 ** the tree depth. Root pages return 0. Parents of root pages
9418 ** return 1, and so forth.
9420 ** These checks are done:
9422 ** 1. Make sure that cells and freeblocks do not overlap
9423 ** but combine to completely cover the page.
9424 ** 2. Make sure integer cell keys are in order.
9425 ** 3. Check the integrity of overflow pages.
9426 ** 4. Recursively call checkTreePage on all children.
9427 ** 5. Verify that the depth of all children is the same.
9429 static int checkTreePage(
9430 IntegrityCk
*pCheck
, /* Context for the sanity check */
9431 int iPage
, /* Page number of the page to check */
9432 i64
*piMinKey
, /* Write minimum integer primary key here */
9433 i64 maxKey
/* Error if integer primary key greater than this */
9435 MemPage
*pPage
= 0; /* The page being analyzed */
9436 int i
; /* Loop counter */
9437 int rc
; /* Result code from subroutine call */
9438 int depth
= -1, d2
; /* Depth of a subtree */
9439 int pgno
; /* Page number */
9440 int nFrag
; /* Number of fragmented bytes on the page */
9441 int hdr
; /* Offset to the page header */
9442 int cellStart
; /* Offset to the start of the cell pointer array */
9443 int nCell
; /* Number of cells */
9444 int doCoverageCheck
= 1; /* True if cell coverage checking should be done */
9445 int keyCanBeEqual
= 1; /* True if IPK can be equal to maxKey
9446 ** False if IPK must be strictly less than maxKey */
9447 u8
*data
; /* Page content */
9448 u8
*pCell
; /* Cell content */
9449 u8
*pCellIdx
; /* Next element of the cell pointer array */
9450 BtShared
*pBt
; /* The BtShared object that owns pPage */
9451 u32 pc
; /* Address of a cell */
9452 u32 usableSize
; /* Usable size of the page */
9453 u32 contentOffset
; /* Offset to the start of the cell content area */
9454 u32
*heap
= 0; /* Min-heap used for checking cell coverage */
9455 u32 x
, prev
= 0; /* Next and previous entry on the min-heap */
9456 const char *saved_zPfx
= pCheck
->zPfx
;
9457 int saved_v1
= pCheck
->v1
;
9458 int saved_v2
= pCheck
->v2
;
9461 /* Check that the page exists
9464 usableSize
= pBt
->usableSize
;
9465 if( iPage
==0 ) return 0;
9466 if( checkRef(pCheck
, iPage
) ) return 0;
9467 pCheck
->zPfx
= "Page %d: ";
9469 if( (rc
= btreeGetPage(pBt
, (Pgno
)iPage
, &pPage
, 0))!=0 ){
9470 checkAppendMsg(pCheck
,
9471 "unable to get the page. error code=%d", rc
);
9475 /* Clear MemPage.isInit to make sure the corruption detection code in
9476 ** btreeInitPage() is executed. */
9477 savedIsInit
= pPage
->isInit
;
9479 if( (rc
= btreeInitPage(pPage
))!=0 ){
9480 assert( rc
==SQLITE_CORRUPT
); /* The only possible error from InitPage */
9481 checkAppendMsg(pCheck
,
9482 "btreeInitPage() returns error code %d", rc
);
9485 data
= pPage
->aData
;
9486 hdr
= pPage
->hdrOffset
;
9488 /* Set up for cell analysis */
9489 pCheck
->zPfx
= "On tree page %d cell %d: ";
9490 contentOffset
= get2byteNotZero(&data
[hdr
+5]);
9491 assert( contentOffset
<=usableSize
); /* Enforced by btreeInitPage() */
9493 /* EVIDENCE-OF: R-37002-32774 The two-byte integer at offset 3 gives the
9494 ** number of cells on the page. */
9495 nCell
= get2byte(&data
[hdr
+3]);
9496 assert( pPage
->nCell
==nCell
);
9498 /* EVIDENCE-OF: R-23882-45353 The cell pointer array of a b-tree page
9499 ** immediately follows the b-tree page header. */
9500 cellStart
= hdr
+ 12 - 4*pPage
->leaf
;
9501 assert( pPage
->aCellIdx
==&data
[cellStart
] );
9502 pCellIdx
= &data
[cellStart
+ 2*(nCell
-1)];
9505 /* Analyze the right-child page of internal pages */
9506 pgno
= get4byte(&data
[hdr
+8]);
9507 #ifndef SQLITE_OMIT_AUTOVACUUM
9508 if( pBt
->autoVacuum
){
9509 pCheck
->zPfx
= "On page %d at right child: ";
9510 checkPtrmap(pCheck
, pgno
, PTRMAP_BTREE
, iPage
);
9513 depth
= checkTreePage(pCheck
, pgno
, &maxKey
, maxKey
);
9516 /* For leaf pages, the coverage check will occur in the same loop
9517 ** as the other cell checks, so initialize the heap. */
9518 heap
= pCheck
->heap
;
9522 /* EVIDENCE-OF: R-02776-14802 The cell pointer array consists of K 2-byte
9523 ** integer offsets to the cell contents. */
9524 for(i
=nCell
-1; i
>=0 && pCheck
->mxErr
; i
--){
9527 /* Check cell size */
9529 assert( pCellIdx
==&data
[cellStart
+ i
*2] );
9530 pc
= get2byteAligned(pCellIdx
);
9532 if( pc
<contentOffset
|| pc
>usableSize
-4 ){
9533 checkAppendMsg(pCheck
, "Offset %d out of range %d..%d",
9534 pc
, contentOffset
, usableSize
-4);
9535 doCoverageCheck
= 0;
9539 pPage
->xParseCell(pPage
, pCell
, &info
);
9540 if( pc
+info
.nSize
>usableSize
){
9541 checkAppendMsg(pCheck
, "Extends off end of page");
9542 doCoverageCheck
= 0;
9546 /* Check for integer primary key out of range */
9547 if( pPage
->intKey
){
9548 if( keyCanBeEqual
? (info
.nKey
> maxKey
) : (info
.nKey
>= maxKey
) ){
9549 checkAppendMsg(pCheck
, "Rowid %lld out of order", info
.nKey
);
9552 keyCanBeEqual
= 0; /* Only the first key on the page may ==maxKey */
9555 /* Check the content overflow list */
9556 if( info
.nPayload
>info
.nLocal
){
9557 int nPage
; /* Number of pages on the overflow chain */
9558 Pgno pgnoOvfl
; /* First page of the overflow chain */
9559 assert( pc
+ info
.nSize
- 4 <= usableSize
);
9560 nPage
= (info
.nPayload
- info
.nLocal
+ usableSize
- 5)/(usableSize
- 4);
9561 pgnoOvfl
= get4byte(&pCell
[info
.nSize
- 4]);
9562 #ifndef SQLITE_OMIT_AUTOVACUUM
9563 if( pBt
->autoVacuum
){
9564 checkPtrmap(pCheck
, pgnoOvfl
, PTRMAP_OVERFLOW1
, iPage
);
9567 checkList(pCheck
, 0, pgnoOvfl
, nPage
);
9571 /* Check sanity of left child page for internal pages */
9572 pgno
= get4byte(pCell
);
9573 #ifndef SQLITE_OMIT_AUTOVACUUM
9574 if( pBt
->autoVacuum
){
9575 checkPtrmap(pCheck
, pgno
, PTRMAP_BTREE
, iPage
);
9578 d2
= checkTreePage(pCheck
, pgno
, &maxKey
, maxKey
);
9581 checkAppendMsg(pCheck
, "Child page depth differs");
9585 /* Populate the coverage-checking heap for leaf pages */
9586 btreeHeapInsert(heap
, (pc
<<16)|(pc
+info
.nSize
-1));
9591 /* Check for complete coverage of the page
9594 if( doCoverageCheck
&& pCheck
->mxErr
>0 ){
9595 /* For leaf pages, the min-heap has already been initialized and the
9596 ** cells have already been inserted. But for internal pages, that has
9597 ** not yet been done, so do it now */
9599 heap
= pCheck
->heap
;
9601 for(i
=nCell
-1; i
>=0; i
--){
9603 pc
= get2byteAligned(&data
[cellStart
+i
*2]);
9604 size
= pPage
->xCellSize(pPage
, &data
[pc
]);
9605 btreeHeapInsert(heap
, (pc
<<16)|(pc
+size
-1));
9608 /* Add the freeblocks to the min-heap
9610 ** EVIDENCE-OF: R-20690-50594 The second field of the b-tree page header
9611 ** is the offset of the first freeblock, or zero if there are no
9612 ** freeblocks on the page.
9614 i
= get2byte(&data
[hdr
+1]);
9617 assert( (u32
)i
<=usableSize
-4 ); /* Enforced by btreeInitPage() */
9618 size
= get2byte(&data
[i
+2]);
9619 assert( (u32
)(i
+size
)<=usableSize
); /* Enforced by btreeInitPage() */
9620 btreeHeapInsert(heap
, (((u32
)i
)<<16)|(i
+size
-1));
9621 /* EVIDENCE-OF: R-58208-19414 The first 2 bytes of a freeblock are a
9622 ** big-endian integer which is the offset in the b-tree page of the next
9623 ** freeblock in the chain, or zero if the freeblock is the last on the
9625 j
= get2byte(&data
[i
]);
9626 /* EVIDENCE-OF: R-06866-39125 Freeblocks are always connected in order of
9627 ** increasing offset. */
9628 assert( j
==0 || j
>i
+size
); /* Enforced by btreeInitPage() */
9629 assert( (u32
)j
<=usableSize
-4 ); /* Enforced by btreeInitPage() */
9632 /* Analyze the min-heap looking for overlap between cells and/or
9633 ** freeblocks, and counting the number of untracked bytes in nFrag.
9635 ** Each min-heap entry is of the form: (start_address<<16)|end_address.
9636 ** There is an implied first entry the covers the page header, the cell
9637 ** pointer index, and the gap between the cell pointer index and the start
9640 ** The loop below pulls entries from the min-heap in order and compares
9641 ** the start_address against the previous end_address. If there is an
9642 ** overlap, that means bytes are used multiple times. If there is a gap,
9643 ** that gap is added to the fragmentation count.
9646 prev
= contentOffset
- 1; /* Implied first min-heap entry */
9647 while( btreeHeapPull(heap
,&x
) ){
9648 if( (prev
&0xffff)>=(x
>>16) ){
9649 checkAppendMsg(pCheck
,
9650 "Multiple uses for byte %u of page %d", x
>>16, iPage
);
9653 nFrag
+= (x
>>16) - (prev
&0xffff) - 1;
9657 nFrag
+= usableSize
- (prev
&0xffff) - 1;
9658 /* EVIDENCE-OF: R-43263-13491 The total number of bytes in all fragments
9659 ** is stored in the fifth field of the b-tree page header.
9660 ** EVIDENCE-OF: R-07161-27322 The one-byte integer at offset 7 gives the
9661 ** number of fragmented free bytes within the cell content area.
9663 if( heap
[0]==0 && nFrag
!=data
[hdr
+7] ){
9664 checkAppendMsg(pCheck
,
9665 "Fragmentation of %d bytes reported as %d on page %d",
9666 nFrag
, data
[hdr
+7], iPage
);
9671 if( !doCoverageCheck
) pPage
->isInit
= savedIsInit
;
9673 pCheck
->zPfx
= saved_zPfx
;
9674 pCheck
->v1
= saved_v1
;
9675 pCheck
->v2
= saved_v2
;
9678 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9680 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9682 ** This routine does a complete check of the given BTree file. aRoot[] is
9683 ** an array of pages numbers were each page number is the root page of
9684 ** a table. nRoot is the number of entries in aRoot.
9686 ** A read-only or read-write transaction must be opened before calling
9689 ** Write the number of error seen in *pnErr. Except for some memory
9690 ** allocation errors, an error message held in memory obtained from
9691 ** malloc is returned if *pnErr is non-zero. If *pnErr==0 then NULL is
9692 ** returned. If a memory allocation error occurs, NULL is returned.
9694 char *sqlite3BtreeIntegrityCheck(
9695 Btree
*p
, /* The btree to be checked */
9696 int *aRoot
, /* An array of root pages numbers for individual trees */
9697 int nRoot
, /* Number of entries in aRoot[] */
9698 int mxErr
, /* Stop reporting errors after this many */
9699 int *pnErr
/* Write number of errors seen to this variable */
9703 BtShared
*pBt
= p
->pBt
;
9704 int savedDbFlags
= pBt
->db
->flags
;
9706 VVA_ONLY( int nRef
);
9708 sqlite3BtreeEnter(p
);
9709 assert( p
->inTrans
>TRANS_NONE
&& pBt
->inTransaction
>TRANS_NONE
);
9710 VVA_ONLY( nRef
= sqlite3PagerRefcount(pBt
->pPager
) );
9713 sCheck
.pPager
= pBt
->pPager
;
9714 sCheck
.nPage
= btreePagecount(sCheck
.pBt
);
9715 sCheck
.mxErr
= mxErr
;
9717 sCheck
.mallocFailed
= 0;
9723 sqlite3StrAccumInit(&sCheck
.errMsg
, 0, zErr
, sizeof(zErr
), SQLITE_MAX_LENGTH
);
9724 sCheck
.errMsg
.printfFlags
= SQLITE_PRINTF_INTERNAL
;
9725 if( sCheck
.nPage
==0 ){
9726 goto integrity_ck_cleanup
;
9729 sCheck
.aPgRef
= sqlite3MallocZero((sCheck
.nPage
/ 8)+ 1);
9730 if( !sCheck
.aPgRef
){
9731 sCheck
.mallocFailed
= 1;
9732 goto integrity_ck_cleanup
;
9734 sCheck
.heap
= (u32
*)sqlite3PageMalloc( pBt
->pageSize
);
9735 if( sCheck
.heap
==0 ){
9736 sCheck
.mallocFailed
= 1;
9737 goto integrity_ck_cleanup
;
9740 i
= PENDING_BYTE_PAGE(pBt
);
9741 if( i
<=sCheck
.nPage
) setPageReferenced(&sCheck
, i
);
9743 /* Check the integrity of the freelist
9745 sCheck
.zPfx
= "Main freelist: ";
9746 checkList(&sCheck
, 1, get4byte(&pBt
->pPage1
->aData
[32]),
9747 get4byte(&pBt
->pPage1
->aData
[36]));
9750 /* Check all the tables.
9752 testcase( pBt
->db
->flags
& SQLITE_CellSizeCk
);
9753 pBt
->db
->flags
&= ~SQLITE_CellSizeCk
;
9754 for(i
=0; (int)i
<nRoot
&& sCheck
.mxErr
; i
++){
9756 if( aRoot
[i
]==0 ) continue;
9757 #ifndef SQLITE_OMIT_AUTOVACUUM
9758 if( pBt
->autoVacuum
&& aRoot
[i
]>1 ){
9759 checkPtrmap(&sCheck
, aRoot
[i
], PTRMAP_ROOTPAGE
, 0);
9762 checkTreePage(&sCheck
, aRoot
[i
], ¬Used
, LARGEST_INT64
);
9764 pBt
->db
->flags
= savedDbFlags
;
9766 /* Make sure every page in the file is referenced
9768 for(i
=1; i
<=sCheck
.nPage
&& sCheck
.mxErr
; i
++){
9769 #ifdef SQLITE_OMIT_AUTOVACUUM
9770 if( getPageReferenced(&sCheck
, i
)==0 ){
9771 checkAppendMsg(&sCheck
, "Page %d is never used", i
);
9774 /* If the database supports auto-vacuum, make sure no tables contain
9775 ** references to pointer-map pages.
9777 if( getPageReferenced(&sCheck
, i
)==0 &&
9778 (PTRMAP_PAGENO(pBt
, i
)!=i
|| !pBt
->autoVacuum
) ){
9779 checkAppendMsg(&sCheck
, "Page %d is never used", i
);
9781 if( getPageReferenced(&sCheck
, i
)!=0 &&
9782 (PTRMAP_PAGENO(pBt
, i
)==i
&& pBt
->autoVacuum
) ){
9783 checkAppendMsg(&sCheck
, "Pointer map page %d is referenced", i
);
9788 /* Clean up and report errors.
9790 integrity_ck_cleanup
:
9791 sqlite3PageFree(sCheck
.heap
);
9792 sqlite3_free(sCheck
.aPgRef
);
9793 if( sCheck
.mallocFailed
){
9794 sqlite3StrAccumReset(&sCheck
.errMsg
);
9797 *pnErr
= sCheck
.nErr
;
9798 if( sCheck
.nErr
==0 ) sqlite3StrAccumReset(&sCheck
.errMsg
);
9799 /* Make sure this analysis did not leave any unref() pages. */
9800 assert( nRef
==sqlite3PagerRefcount(pBt
->pPager
) );
9801 sqlite3BtreeLeave(p
);
9802 return sqlite3StrAccumFinish(&sCheck
.errMsg
);
9804 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9807 ** Return the full pathname of the underlying database file. Return
9808 ** an empty string if the database is in-memory or a TEMP database.
9810 ** The pager filename is invariant as long as the pager is
9811 ** open so it is safe to access without the BtShared mutex.
9813 const char *sqlite3BtreeGetFilename(Btree
*p
){
9814 assert( p
->pBt
->pPager
!=0 );
9815 return sqlite3PagerFilename(p
->pBt
->pPager
, 1);
9819 ** Return the pathname of the journal file for this database. The return
9820 ** value of this routine is the same regardless of whether the journal file
9821 ** has been created or not.
9823 ** The pager journal filename is invariant as long as the pager is
9824 ** open so it is safe to access without the BtShared mutex.
9826 const char *sqlite3BtreeGetJournalname(Btree
*p
){
9827 assert( p
->pBt
->pPager
!=0 );
9828 return sqlite3PagerJournalname(p
->pBt
->pPager
);
9832 ** Return non-zero if a transaction is active.
9834 int sqlite3BtreeIsInTrans(Btree
*p
){
9835 assert( p
==0 || sqlite3_mutex_held(p
->db
->mutex
) );
9836 return (p
&& (p
->inTrans
==TRANS_WRITE
));
9839 #ifndef SQLITE_OMIT_WAL
9841 ** Run a checkpoint on the Btree passed as the first argument.
9843 ** Return SQLITE_LOCKED if this or any other connection has an open
9844 ** transaction on the shared-cache the argument Btree is connected to.
9846 ** Parameter eMode is one of SQLITE_CHECKPOINT_PASSIVE, FULL or RESTART.
9848 int sqlite3BtreeCheckpoint(Btree
*p
, int eMode
, int *pnLog
, int *pnCkpt
){
9851 BtShared
*pBt
= p
->pBt
;
9852 sqlite3BtreeEnter(p
);
9853 if( pBt
->inTransaction
!=TRANS_NONE
){
9856 rc
= sqlite3PagerCheckpoint(pBt
->pPager
, p
->db
, eMode
, pnLog
, pnCkpt
);
9858 sqlite3BtreeLeave(p
);
9865 ** Return non-zero if a read (or write) transaction is active.
9867 int sqlite3BtreeIsInReadTrans(Btree
*p
){
9869 assert( sqlite3_mutex_held(p
->db
->mutex
) );
9870 return p
->inTrans
!=TRANS_NONE
;
9873 int sqlite3BtreeIsInBackup(Btree
*p
){
9875 assert( sqlite3_mutex_held(p
->db
->mutex
) );
9876 return p
->nBackup
!=0;
9880 ** This function returns a pointer to a blob of memory associated with
9881 ** a single shared-btree. The memory is used by client code for its own
9882 ** purposes (for example, to store a high-level schema associated with
9883 ** the shared-btree). The btree layer manages reference counting issues.
9885 ** The first time this is called on a shared-btree, nBytes bytes of memory
9886 ** are allocated, zeroed, and returned to the caller. For each subsequent
9887 ** call the nBytes parameter is ignored and a pointer to the same blob
9888 ** of memory returned.
9890 ** If the nBytes parameter is 0 and the blob of memory has not yet been
9891 ** allocated, a null pointer is returned. If the blob has already been
9892 ** allocated, it is returned as normal.
9894 ** Just before the shared-btree is closed, the function passed as the
9895 ** xFree argument when the memory allocation was made is invoked on the
9896 ** blob of allocated memory. The xFree function should not call sqlite3_free()
9897 ** on the memory, the btree layer does that.
9899 void *sqlite3BtreeSchema(Btree
*p
, int nBytes
, void(*xFree
)(void *)){
9900 BtShared
*pBt
= p
->pBt
;
9901 sqlite3BtreeEnter(p
);
9902 if( !pBt
->pSchema
&& nBytes
){
9903 pBt
->pSchema
= sqlite3DbMallocZero(0, nBytes
);
9904 pBt
->xFreeSchema
= xFree
;
9906 sqlite3BtreeLeave(p
);
9907 return pBt
->pSchema
;
9911 ** Return SQLITE_LOCKED_SHAREDCACHE if another user of the same shared
9912 ** btree as the argument handle holds an exclusive lock on the
9913 ** sqlite_master table. Otherwise SQLITE_OK.
9915 int sqlite3BtreeSchemaLocked(Btree
*p
){
9917 assert( sqlite3_mutex_held(p
->db
->mutex
) );
9918 sqlite3BtreeEnter(p
);
9919 rc
= querySharedCacheTableLock(p
, MASTER_ROOT
, READ_LOCK
);
9920 assert( rc
==SQLITE_OK
|| rc
==SQLITE_LOCKED_SHAREDCACHE
);
9921 sqlite3BtreeLeave(p
);
9926 #ifndef SQLITE_OMIT_SHARED_CACHE
9928 ** Obtain a lock on the table whose root page is iTab. The
9929 ** lock is a write lock if isWritelock is true or a read lock
9932 int sqlite3BtreeLockTable(Btree
*p
, int iTab
, u8 isWriteLock
){
9934 assert( p
->inTrans
!=TRANS_NONE
);
9936 u8 lockType
= READ_LOCK
+ isWriteLock
;
9937 assert( READ_LOCK
+1==WRITE_LOCK
);
9938 assert( isWriteLock
==0 || isWriteLock
==1 );
9940 sqlite3BtreeEnter(p
);
9941 rc
= querySharedCacheTableLock(p
, iTab
, lockType
);
9942 if( rc
==SQLITE_OK
){
9943 rc
= setSharedCacheTableLock(p
, iTab
, lockType
);
9945 sqlite3BtreeLeave(p
);
9951 #ifndef SQLITE_OMIT_INCRBLOB
9953 ** Argument pCsr must be a cursor opened for writing on an
9954 ** INTKEY table currently pointing at a valid table entry.
9955 ** This function modifies the data stored as part of that entry.
9957 ** Only the data content may only be modified, it is not possible to
9958 ** change the length of the data stored. If this function is called with
9959 ** parameters that attempt to write past the end of the existing data,
9960 ** no modifications are made and SQLITE_CORRUPT is returned.
9962 int sqlite3BtreePutData(BtCursor
*pCsr
, u32 offset
, u32 amt
, void *z
){
9964 assert( cursorOwnsBtShared(pCsr
) );
9965 assert( sqlite3_mutex_held(pCsr
->pBtree
->db
->mutex
) );
9966 assert( pCsr
->curFlags
& BTCF_Incrblob
);
9968 rc
= restoreCursorPosition(pCsr
);
9969 if( rc
!=SQLITE_OK
){
9972 assert( pCsr
->eState
!=CURSOR_REQUIRESEEK
);
9973 if( pCsr
->eState
!=CURSOR_VALID
){
9974 return SQLITE_ABORT
;
9977 /* Save the positions of all other cursors open on this table. This is
9978 ** required in case any of them are holding references to an xFetch
9979 ** version of the b-tree page modified by the accessPayload call below.
9981 ** Note that pCsr must be open on a INTKEY table and saveCursorPosition()
9982 ** and hence saveAllCursors() cannot fail on a BTREE_INTKEY table, hence
9983 ** saveAllCursors can only return SQLITE_OK.
9985 VVA_ONLY(rc
=) saveAllCursors(pCsr
->pBt
, pCsr
->pgnoRoot
, pCsr
);
9986 assert( rc
==SQLITE_OK
);
9988 /* Check some assumptions:
9989 ** (a) the cursor is open for writing,
9990 ** (b) there is a read/write transaction open,
9991 ** (c) the connection holds a write-lock on the table (if required),
9992 ** (d) there are no conflicting read-locks, and
9993 ** (e) the cursor points at a valid row of an intKey table.
9995 if( (pCsr
->curFlags
& BTCF_WriteFlag
)==0 ){
9996 return SQLITE_READONLY
;
9998 assert( (pCsr
->pBt
->btsFlags
& BTS_READ_ONLY
)==0
9999 && pCsr
->pBt
->inTransaction
==TRANS_WRITE
);
10000 assert( hasSharedCacheTableLock(pCsr
->pBtree
, pCsr
->pgnoRoot
, 0, 2) );
10001 assert( !hasReadConflicts(pCsr
->pBtree
, pCsr
->pgnoRoot
) );
10002 assert( pCsr
->pPage
->intKey
);
10004 return accessPayload(pCsr
, offset
, amt
, (unsigned char *)z
, 1);
10008 ** Mark this cursor as an incremental blob cursor.
10010 void sqlite3BtreeIncrblobCursor(BtCursor
*pCur
){
10011 pCur
->curFlags
|= BTCF_Incrblob
;
10012 pCur
->pBtree
->hasIncrblobCur
= 1;
10017 ** Set both the "read version" (single byte at byte offset 18) and
10018 ** "write version" (single byte at byte offset 19) fields in the database
10019 ** header to iVersion.
10021 int sqlite3BtreeSetVersion(Btree
*pBtree
, int iVersion
){
10022 BtShared
*pBt
= pBtree
->pBt
;
10023 int rc
; /* Return code */
10025 assert( iVersion
==1 || iVersion
==2 );
10027 /* If setting the version fields to 1, do not automatically open the
10028 ** WAL connection, even if the version fields are currently set to 2.
10030 pBt
->btsFlags
&= ~BTS_NO_WAL
;
10031 if( iVersion
==1 ) pBt
->btsFlags
|= BTS_NO_WAL
;
10033 rc
= sqlite3BtreeBeginTrans(pBtree
, 0);
10034 if( rc
==SQLITE_OK
){
10035 u8
*aData
= pBt
->pPage1
->aData
;
10036 if( aData
[18]!=(u8
)iVersion
|| aData
[19]!=(u8
)iVersion
){
10037 rc
= sqlite3BtreeBeginTrans(pBtree
, 2);
10038 if( rc
==SQLITE_OK
){
10039 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
10040 if( rc
==SQLITE_OK
){
10041 aData
[18] = (u8
)iVersion
;
10042 aData
[19] = (u8
)iVersion
;
10048 pBt
->btsFlags
&= ~BTS_NO_WAL
;
10053 ** Return true if the cursor has a hint specified. This routine is
10054 ** only used from within assert() statements
10056 int sqlite3BtreeCursorHasHint(BtCursor
*pCsr
, unsigned int mask
){
10057 return (pCsr
->hints
& mask
)!=0;
10061 ** Return true if the given Btree is read-only.
10063 int sqlite3BtreeIsReadonly(Btree
*p
){
10064 return (p
->pBt
->btsFlags
& BTS_READ_ONLY
)!=0;
10068 ** Return the size of the header added to each page by this module.
10070 int sqlite3HeaderSizeBtree(void){ return ROUND8(sizeof(MemPage
)); }
10072 #if !defined(SQLITE_OMIT_SHARED_CACHE)
10074 ** Return true if the Btree passed as the only argument is sharable.
10076 int sqlite3BtreeSharable(Btree
*p
){
10077 return p
->sharable
;
10081 ** Return the number of connections to the BtShared object accessed by
10082 ** the Btree handle passed as the only argument. For private caches
10083 ** this is always 1. For shared caches it may be 1 or greater.
10085 int sqlite3BtreeConnectionCount(Btree
*p
){
10086 testcase( p
->sharable
);
10087 return p
->pBt
->nRef
;