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
115 #ifndef SQLITE_OMIT_SHARED_CACHE
119 **** This function is only used as part of an assert() statement. ***
121 ** Check to see if pBtree holds the required locks to read or write to the
122 ** table with root page iRoot. Return 1 if it does and 0 if not.
124 ** For example, when writing to a table with root-page iRoot via
125 ** Btree connection pBtree:
127 ** assert( hasSharedCacheTableLock(pBtree, iRoot, 0, WRITE_LOCK) );
129 ** When writing to an index that resides in a sharable database, the
130 ** caller should have first obtained a lock specifying the root page of
131 ** the corresponding table. This makes things a bit more complicated,
132 ** as this module treats each table as a separate structure. To determine
133 ** the table corresponding to the index being written, this
134 ** function has to search through the database schema.
136 ** Instead of a lock on the table/index rooted at page iRoot, the caller may
137 ** hold a write-lock on the schema table (root page 1). This is also
140 static int hasSharedCacheTableLock(
141 Btree
*pBtree
, /* Handle that must hold lock */
142 Pgno iRoot
, /* Root page of b-tree */
143 int isIndex
, /* True if iRoot is the root of an index b-tree */
144 int eLockType
/* Required lock type (READ_LOCK or WRITE_LOCK) */
146 Schema
*pSchema
= (Schema
*)pBtree
->pBt
->pSchema
;
150 /* If this database is not shareable, or if the client is reading
151 ** and has the read-uncommitted flag set, then no lock is required.
152 ** Return true immediately.
154 if( (pBtree
->sharable
==0)
155 || (eLockType
==READ_LOCK
&& (pBtree
->db
->flags
& SQLITE_ReadUncommit
))
160 /* If the client is reading or writing an index and the schema is
161 ** not loaded, then it is too difficult to actually check to see if
162 ** the correct locks are held. So do not bother - just return true.
163 ** This case does not come up very often anyhow.
165 if( isIndex
&& (!pSchema
|| (pSchema
->schemaFlags
&DB_SchemaLoaded
)==0) ){
169 /* Figure out the root-page that the lock should be held on. For table
170 ** b-trees, this is just the root page of the b-tree being read or
171 ** written. For index b-trees, it is the root page of the associated
175 for(p
=sqliteHashFirst(&pSchema
->idxHash
); p
; p
=sqliteHashNext(p
)){
176 Index
*pIdx
= (Index
*)sqliteHashData(p
);
177 if( pIdx
->tnum
==(int)iRoot
){
179 /* Two or more indexes share the same root page. There must
180 ** be imposter tables. So just return true. The assert is not
181 ** useful in that case. */
184 iTab
= pIdx
->pTable
->tnum
;
191 /* Search for the required lock. Either a write-lock on root-page iTab, a
192 ** write-lock on the schema table, or (if the client is reading) a
193 ** read-lock on iTab will suffice. Return 1 if any of these are found. */
194 for(pLock
=pBtree
->pBt
->pLock
; pLock
; pLock
=pLock
->pNext
){
195 if( pLock
->pBtree
==pBtree
196 && (pLock
->iTable
==iTab
|| (pLock
->eLock
==WRITE_LOCK
&& pLock
->iTable
==1))
197 && pLock
->eLock
>=eLockType
203 /* Failed to find the required lock. */
206 #endif /* SQLITE_DEBUG */
210 **** This function may be used as part of assert() statements only. ****
212 ** Return true if it would be illegal for pBtree to write into the
213 ** table or index rooted at iRoot because other shared connections are
214 ** simultaneously reading that same table or index.
216 ** It is illegal for pBtree to write if some other Btree object that
217 ** shares the same BtShared object is currently reading or writing
218 ** the iRoot table. Except, if the other Btree object has the
219 ** read-uncommitted flag set, then it is OK for the other object to
220 ** have a read cursor.
222 ** For example, before writing to any part of the table or index
223 ** rooted at page iRoot, one should call:
225 ** assert( !hasReadConflicts(pBtree, iRoot) );
227 static int hasReadConflicts(Btree
*pBtree
, Pgno iRoot
){
229 for(p
=pBtree
->pBt
->pCursor
; p
; p
=p
->pNext
){
230 if( p
->pgnoRoot
==iRoot
232 && 0==(p
->pBtree
->db
->flags
& SQLITE_ReadUncommit
)
239 #endif /* #ifdef SQLITE_DEBUG */
242 ** Query to see if Btree handle p may obtain a lock of type eLock
243 ** (READ_LOCK or WRITE_LOCK) on the table with root-page iTab. Return
244 ** SQLITE_OK if the lock may be obtained (by calling
245 ** setSharedCacheTableLock()), or SQLITE_LOCKED if not.
247 static int querySharedCacheTableLock(Btree
*p
, Pgno iTab
, u8 eLock
){
248 BtShared
*pBt
= p
->pBt
;
251 assert( sqlite3BtreeHoldsMutex(p
) );
252 assert( eLock
==READ_LOCK
|| eLock
==WRITE_LOCK
);
254 assert( !(p
->db
->flags
&SQLITE_ReadUncommit
)||eLock
==WRITE_LOCK
||iTab
==1 );
256 /* If requesting a write-lock, then the Btree must have an open write
257 ** transaction on this file. And, obviously, for this to be so there
258 ** must be an open write transaction on the file itself.
260 assert( eLock
==READ_LOCK
|| (p
==pBt
->pWriter
&& p
->inTrans
==TRANS_WRITE
) );
261 assert( eLock
==READ_LOCK
|| pBt
->inTransaction
==TRANS_WRITE
);
263 /* This routine is a no-op if the shared-cache is not enabled */
268 /* If some other connection is holding an exclusive lock, the
269 ** requested lock may not be obtained.
271 if( pBt
->pWriter
!=p
&& (pBt
->btsFlags
& BTS_EXCLUSIVE
)!=0 ){
272 sqlite3ConnectionBlocked(p
->db
, pBt
->pWriter
->db
);
273 return SQLITE_LOCKED_SHAREDCACHE
;
276 for(pIter
=pBt
->pLock
; pIter
; pIter
=pIter
->pNext
){
277 /* The condition (pIter->eLock!=eLock) in the following if(...)
278 ** statement is a simplification of:
280 ** (eLock==WRITE_LOCK || pIter->eLock==WRITE_LOCK)
282 ** since we know that if eLock==WRITE_LOCK, then no other connection
283 ** may hold a WRITE_LOCK on any table in this file (since there can
284 ** only be a single writer).
286 assert( pIter
->eLock
==READ_LOCK
|| pIter
->eLock
==WRITE_LOCK
);
287 assert( eLock
==READ_LOCK
|| pIter
->pBtree
==p
|| pIter
->eLock
==READ_LOCK
);
288 if( pIter
->pBtree
!=p
&& pIter
->iTable
==iTab
&& pIter
->eLock
!=eLock
){
289 sqlite3ConnectionBlocked(p
->db
, pIter
->pBtree
->db
);
290 if( eLock
==WRITE_LOCK
){
291 assert( p
==pBt
->pWriter
);
292 pBt
->btsFlags
|= BTS_PENDING
;
294 return SQLITE_LOCKED_SHAREDCACHE
;
299 #endif /* !SQLITE_OMIT_SHARED_CACHE */
301 #ifndef SQLITE_OMIT_SHARED_CACHE
303 ** Add a lock on the table with root-page iTable to the shared-btree used
304 ** by Btree handle p. Parameter eLock must be either READ_LOCK or
307 ** This function assumes the following:
309 ** (a) The specified Btree object p is connected to a sharable
310 ** database (one with the BtShared.sharable flag set), and
312 ** (b) No other Btree objects hold a lock that conflicts
313 ** with the requested lock (i.e. querySharedCacheTableLock() has
314 ** already been called and returned SQLITE_OK).
316 ** SQLITE_OK is returned if the lock is added successfully. SQLITE_NOMEM
317 ** is returned if a malloc attempt fails.
319 static int setSharedCacheTableLock(Btree
*p
, Pgno iTable
, u8 eLock
){
320 BtShared
*pBt
= p
->pBt
;
324 assert( sqlite3BtreeHoldsMutex(p
) );
325 assert( eLock
==READ_LOCK
|| eLock
==WRITE_LOCK
);
328 /* A connection with the read-uncommitted flag set will never try to
329 ** obtain a read-lock using this function. The only read-lock obtained
330 ** by a connection in read-uncommitted mode is on the sqlite_master
331 ** table, and that lock is obtained in BtreeBeginTrans(). */
332 assert( 0==(p
->db
->flags
&SQLITE_ReadUncommit
) || eLock
==WRITE_LOCK
);
334 /* This function should only be called on a sharable b-tree after it
335 ** has been determined that no other b-tree holds a conflicting lock. */
336 assert( p
->sharable
);
337 assert( SQLITE_OK
==querySharedCacheTableLock(p
, iTable
, eLock
) );
339 /* First search the list for an existing lock on this table. */
340 for(pIter
=pBt
->pLock
; pIter
; pIter
=pIter
->pNext
){
341 if( pIter
->iTable
==iTable
&& pIter
->pBtree
==p
){
347 /* If the above search did not find a BtLock struct associating Btree p
348 ** with table iTable, allocate one and link it into the list.
351 pLock
= (BtLock
*)sqlite3MallocZero(sizeof(BtLock
));
353 return SQLITE_NOMEM_BKPT
;
355 pLock
->iTable
= iTable
;
357 pLock
->pNext
= pBt
->pLock
;
361 /* Set the BtLock.eLock variable to the maximum of the current lock
362 ** and the requested lock. This means if a write-lock was already held
363 ** and a read-lock requested, we don't incorrectly downgrade the lock.
365 assert( WRITE_LOCK
>READ_LOCK
);
366 if( eLock
>pLock
->eLock
){
367 pLock
->eLock
= eLock
;
372 #endif /* !SQLITE_OMIT_SHARED_CACHE */
374 #ifndef SQLITE_OMIT_SHARED_CACHE
376 ** Release all the table locks (locks obtained via calls to
377 ** the setSharedCacheTableLock() procedure) held by Btree object p.
379 ** This function assumes that Btree p has an open read or write
380 ** transaction. If it does not, then the BTS_PENDING flag
381 ** may be incorrectly cleared.
383 static void clearAllSharedCacheTableLocks(Btree
*p
){
384 BtShared
*pBt
= p
->pBt
;
385 BtLock
**ppIter
= &pBt
->pLock
;
387 assert( sqlite3BtreeHoldsMutex(p
) );
388 assert( p
->sharable
|| 0==*ppIter
);
389 assert( p
->inTrans
>0 );
392 BtLock
*pLock
= *ppIter
;
393 assert( (pBt
->btsFlags
& BTS_EXCLUSIVE
)==0 || pBt
->pWriter
==pLock
->pBtree
);
394 assert( pLock
->pBtree
->inTrans
>=pLock
->eLock
);
395 if( pLock
->pBtree
==p
){
396 *ppIter
= pLock
->pNext
;
397 assert( pLock
->iTable
!=1 || pLock
==&p
->lock
);
398 if( pLock
->iTable
!=1 ){
402 ppIter
= &pLock
->pNext
;
406 assert( (pBt
->btsFlags
& BTS_PENDING
)==0 || pBt
->pWriter
);
407 if( pBt
->pWriter
==p
){
409 pBt
->btsFlags
&= ~(BTS_EXCLUSIVE
|BTS_PENDING
);
410 }else if( pBt
->nTransaction
==2 ){
411 /* This function is called when Btree p is concluding its
412 ** transaction. If there currently exists a writer, and p is not
413 ** that writer, then the number of locks held by connections other
414 ** than the writer must be about to drop to zero. In this case
415 ** set the BTS_PENDING flag to 0.
417 ** If there is not currently a writer, then BTS_PENDING must
418 ** be zero already. So this next line is harmless in that case.
420 pBt
->btsFlags
&= ~BTS_PENDING
;
425 ** This function changes all write-locks held by Btree p into read-locks.
427 static void downgradeAllSharedCacheTableLocks(Btree
*p
){
428 BtShared
*pBt
= p
->pBt
;
429 if( pBt
->pWriter
==p
){
432 pBt
->btsFlags
&= ~(BTS_EXCLUSIVE
|BTS_PENDING
);
433 for(pLock
=pBt
->pLock
; pLock
; pLock
=pLock
->pNext
){
434 assert( pLock
->eLock
==READ_LOCK
|| pLock
->pBtree
==p
);
435 pLock
->eLock
= READ_LOCK
;
440 #endif /* SQLITE_OMIT_SHARED_CACHE */
442 static void releasePage(MemPage
*pPage
); /* Forward reference */
443 static void releasePageOne(MemPage
*pPage
); /* Forward reference */
444 static void releasePageNotNull(MemPage
*pPage
); /* Forward reference */
447 ***** This routine is used inside of assert() only ****
449 ** Verify that the cursor holds the mutex on its BtShared
452 static int cursorHoldsMutex(BtCursor
*p
){
453 return sqlite3_mutex_held(p
->pBt
->mutex
);
456 /* Verify that the cursor and the BtShared agree about what is the current
457 ** database connetion. This is important in shared-cache mode. If the database
458 ** connection pointers get out-of-sync, it is possible for routines like
459 ** btreeInitPage() to reference an stale connection pointer that references a
460 ** a connection that has already closed. This routine is used inside assert()
461 ** statements only and for the purpose of double-checking that the btree code
462 ** does keep the database connection pointers up-to-date.
464 static int cursorOwnsBtShared(BtCursor
*p
){
465 assert( cursorHoldsMutex(p
) );
466 return (p
->pBtree
->db
==p
->pBt
->db
);
471 ** Invalidate the overflow cache of the cursor passed as the first argument.
472 ** on the shared btree structure pBt.
474 #define invalidateOverflowCache(pCur) (pCur->curFlags &= ~BTCF_ValidOvfl)
477 ** Invalidate the overflow page-list cache for all cursors opened
478 ** on the shared btree structure pBt.
480 static void invalidateAllOverflowCache(BtShared
*pBt
){
482 assert( sqlite3_mutex_held(pBt
->mutex
) );
483 for(p
=pBt
->pCursor
; p
; p
=p
->pNext
){
484 invalidateOverflowCache(p
);
488 #ifndef SQLITE_OMIT_INCRBLOB
490 ** This function is called before modifying the contents of a table
491 ** to invalidate any incrblob cursors that are open on the
492 ** row or one of the rows being modified.
494 ** If argument isClearTable is true, then the entire contents of the
495 ** table is about to be deleted. In this case invalidate all incrblob
496 ** cursors open on any row within the table with root-page pgnoRoot.
498 ** Otherwise, if argument isClearTable is false, then the row with
499 ** rowid iRow is being replaced or deleted. In this case invalidate
500 ** only those incrblob cursors open on that specific row.
502 static void invalidateIncrblobCursors(
503 Btree
*pBtree
, /* The database file to check */
504 Pgno pgnoRoot
, /* The table that might be changing */
505 i64 iRow
, /* The rowid that might be changing */
506 int isClearTable
/* True if all rows are being deleted */
509 if( pBtree
->hasIncrblobCur
==0 ) return;
510 assert( sqlite3BtreeHoldsMutex(pBtree
) );
511 pBtree
->hasIncrblobCur
= 0;
512 for(p
=pBtree
->pBt
->pCursor
; p
; p
=p
->pNext
){
513 if( (p
->curFlags
& BTCF_Incrblob
)!=0 ){
514 pBtree
->hasIncrblobCur
= 1;
515 if( p
->pgnoRoot
==pgnoRoot
&& (isClearTable
|| p
->info
.nKey
==iRow
) ){
516 p
->eState
= CURSOR_INVALID
;
523 /* Stub function when INCRBLOB is omitted */
524 #define invalidateIncrblobCursors(w,x,y,z)
525 #endif /* SQLITE_OMIT_INCRBLOB */
528 ** Set bit pgno of the BtShared.pHasContent bitvec. This is called
529 ** when a page that previously contained data becomes a free-list leaf
532 ** The BtShared.pHasContent bitvec exists to work around an obscure
533 ** bug caused by the interaction of two useful IO optimizations surrounding
534 ** free-list leaf pages:
536 ** 1) When all data is deleted from a page and the page becomes
537 ** a free-list leaf page, the page is not written to the database
538 ** (as free-list leaf pages contain no meaningful data). Sometimes
539 ** such a page is not even journalled (as it will not be modified,
540 ** why bother journalling it?).
542 ** 2) When a free-list leaf page is reused, its content is not read
543 ** from the database or written to the journal file (why should it
544 ** be, if it is not at all meaningful?).
546 ** By themselves, these optimizations work fine and provide a handy
547 ** performance boost to bulk delete or insert operations. However, if
548 ** a page is moved to the free-list and then reused within the same
549 ** transaction, a problem comes up. If the page is not journalled when
550 ** it is moved to the free-list and it is also not journalled when it
551 ** is extracted from the free-list and reused, then the original data
552 ** may be lost. In the event of a rollback, it may not be possible
553 ** to restore the database to its original configuration.
555 ** The solution is the BtShared.pHasContent bitvec. Whenever a page is
556 ** moved to become a free-list leaf page, the corresponding bit is
557 ** set in the bitvec. Whenever a leaf page is extracted from the free-list,
558 ** optimization 2 above is omitted if the corresponding bit is already
559 ** set in BtShared.pHasContent. The contents of the bitvec are cleared
560 ** at the end of every transaction.
562 static int btreeSetHasContent(BtShared
*pBt
, Pgno pgno
){
564 if( !pBt
->pHasContent
){
565 assert( pgno
<=pBt
->nPage
);
566 pBt
->pHasContent
= sqlite3BitvecCreate(pBt
->nPage
);
567 if( !pBt
->pHasContent
){
568 rc
= SQLITE_NOMEM_BKPT
;
571 if( rc
==SQLITE_OK
&& pgno
<=sqlite3BitvecSize(pBt
->pHasContent
) ){
572 rc
= sqlite3BitvecSet(pBt
->pHasContent
, pgno
);
578 ** Query the BtShared.pHasContent vector.
580 ** This function is called when a free-list leaf page is removed from the
581 ** free-list for reuse. It returns false if it is safe to retrieve the
582 ** page from the pager layer with the 'no-content' flag set. True otherwise.
584 static int btreeGetHasContent(BtShared
*pBt
, Pgno pgno
){
585 Bitvec
*p
= pBt
->pHasContent
;
586 return (p
&& (pgno
>sqlite3BitvecSize(p
) || sqlite3BitvecTest(p
, pgno
)));
590 ** Clear (destroy) the BtShared.pHasContent bitvec. This should be
591 ** invoked at the conclusion of each write-transaction.
593 static void btreeClearHasContent(BtShared
*pBt
){
594 sqlite3BitvecDestroy(pBt
->pHasContent
);
595 pBt
->pHasContent
= 0;
599 ** Release all of the apPage[] pages for a cursor.
601 static void btreeReleaseAllCursorPages(BtCursor
*pCur
){
603 if( pCur
->iPage
>=0 ){
604 for(i
=0; i
<pCur
->iPage
; i
++){
605 releasePageNotNull(pCur
->apPage
[i
]);
607 releasePageNotNull(pCur
->pPage
);
613 ** The cursor passed as the only argument must point to a valid entry
614 ** when this function is called (i.e. have eState==CURSOR_VALID). This
615 ** function saves the current cursor key in variables pCur->nKey and
616 ** pCur->pKey. SQLITE_OK is returned if successful or an SQLite error
619 ** If the cursor is open on an intkey table, then the integer key
620 ** (the rowid) is stored in pCur->nKey and pCur->pKey is left set to
621 ** NULL. If the cursor is open on a non-intkey table, then pCur->pKey is
622 ** set to point to a malloced buffer pCur->nKey bytes in size containing
625 static int saveCursorKey(BtCursor
*pCur
){
627 assert( CURSOR_VALID
==pCur
->eState
);
628 assert( 0==pCur
->pKey
);
629 assert( cursorHoldsMutex(pCur
) );
631 if( pCur
->curIntKey
){
632 /* Only the rowid is required for a table btree */
633 pCur
->nKey
= sqlite3BtreeIntegerKey(pCur
);
635 /* For an index btree, save the complete key content */
637 pCur
->nKey
= sqlite3BtreePayloadSize(pCur
);
638 pKey
= sqlite3Malloc( pCur
->nKey
);
640 rc
= sqlite3BtreePayload(pCur
, 0, (int)pCur
->nKey
, pKey
);
647 rc
= SQLITE_NOMEM_BKPT
;
650 assert( !pCur
->curIntKey
|| !pCur
->pKey
);
655 ** Save the current cursor position in the variables BtCursor.nKey
656 ** and BtCursor.pKey. The cursor's state is set to CURSOR_REQUIRESEEK.
658 ** The caller must ensure that the cursor is valid (has eState==CURSOR_VALID)
659 ** prior to calling this routine.
661 static int saveCursorPosition(BtCursor
*pCur
){
664 assert( CURSOR_VALID
==pCur
->eState
|| CURSOR_SKIPNEXT
==pCur
->eState
);
665 assert( 0==pCur
->pKey
);
666 assert( cursorHoldsMutex(pCur
) );
668 if( pCur
->eState
==CURSOR_SKIPNEXT
){
669 pCur
->eState
= CURSOR_VALID
;
674 rc
= saveCursorKey(pCur
);
676 btreeReleaseAllCursorPages(pCur
);
677 pCur
->eState
= CURSOR_REQUIRESEEK
;
680 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
|BTCF_AtLast
);
684 /* Forward reference */
685 static int SQLITE_NOINLINE
saveCursorsOnList(BtCursor
*,Pgno
,BtCursor
*);
688 ** Save the positions of all cursors (except pExcept) that are open on
689 ** the table with root-page iRoot. "Saving the cursor position" means that
690 ** the location in the btree is remembered in such a way that it can be
691 ** moved back to the same spot after the btree has been modified. This
692 ** routine is called just before cursor pExcept is used to modify the
693 ** table, for example in BtreeDelete() or BtreeInsert().
695 ** If there are two or more cursors on the same btree, then all such
696 ** cursors should have their BTCF_Multiple flag set. The btreeCursor()
697 ** routine enforces that rule. This routine only needs to be called in
698 ** the uncommon case when pExpect has the BTCF_Multiple flag set.
700 ** If pExpect!=NULL and if no other cursors are found on the same root-page,
701 ** then the BTCF_Multiple flag on pExpect is cleared, to avoid another
702 ** pointless call to this routine.
704 ** Implementation note: This routine merely checks to see if any cursors
705 ** need to be saved. It calls out to saveCursorsOnList() in the (unusual)
706 ** event that cursors are in need to being saved.
708 static int saveAllCursors(BtShared
*pBt
, Pgno iRoot
, BtCursor
*pExcept
){
710 assert( sqlite3_mutex_held(pBt
->mutex
) );
711 assert( pExcept
==0 || pExcept
->pBt
==pBt
);
712 for(p
=pBt
->pCursor
; p
; p
=p
->pNext
){
713 if( p
!=pExcept
&& (0==iRoot
|| p
->pgnoRoot
==iRoot
) ) break;
715 if( p
) return saveCursorsOnList(p
, iRoot
, pExcept
);
716 if( pExcept
) pExcept
->curFlags
&= ~BTCF_Multiple
;
720 /* This helper routine to saveAllCursors does the actual work of saving
721 ** the cursors if and when a cursor is found that actually requires saving.
722 ** The common case is that no cursors need to be saved, so this routine is
723 ** broken out from its caller to avoid unnecessary stack pointer movement.
725 static int SQLITE_NOINLINE
saveCursorsOnList(
726 BtCursor
*p
, /* The first cursor that needs saving */
727 Pgno iRoot
, /* Only save cursor with this iRoot. Save all if zero */
728 BtCursor
*pExcept
/* Do not save this cursor */
731 if( p
!=pExcept
&& (0==iRoot
|| p
->pgnoRoot
==iRoot
) ){
732 if( p
->eState
==CURSOR_VALID
|| p
->eState
==CURSOR_SKIPNEXT
){
733 int rc
= saveCursorPosition(p
);
738 testcase( p
->iPage
>=0 );
739 btreeReleaseAllCursorPages(p
);
748 ** Clear the current cursor position.
750 void sqlite3BtreeClearCursor(BtCursor
*pCur
){
751 assert( cursorHoldsMutex(pCur
) );
752 sqlite3_free(pCur
->pKey
);
754 pCur
->eState
= CURSOR_INVALID
;
758 ** In this version of BtreeMoveto, pKey is a packed index record
759 ** such as is generated by the OP_MakeRecord opcode. Unpack the
760 ** record and then call BtreeMovetoUnpacked() to do the work.
762 static int btreeMoveto(
763 BtCursor
*pCur
, /* Cursor open on the btree to be searched */
764 const void *pKey
, /* Packed key if the btree is an index */
765 i64 nKey
, /* Integer key for tables. Size of pKey for indices */
766 int bias
, /* Bias search to the high end */
767 int *pRes
/* Write search results here */
769 int rc
; /* Status code */
770 UnpackedRecord
*pIdxKey
; /* Unpacked index key */
773 assert( nKey
==(i64
)(int)nKey
);
774 pIdxKey
= sqlite3VdbeAllocUnpackedRecord(pCur
->pKeyInfo
);
775 if( pIdxKey
==0 ) return SQLITE_NOMEM_BKPT
;
776 sqlite3VdbeRecordUnpack(pCur
->pKeyInfo
, (int)nKey
, pKey
, pIdxKey
);
777 if( pIdxKey
->nField
==0 ){
778 rc
= SQLITE_CORRUPT_BKPT
;
784 rc
= sqlite3BtreeMovetoUnpacked(pCur
, pIdxKey
, nKey
, bias
, pRes
);
787 sqlite3DbFree(pCur
->pKeyInfo
->db
, pIdxKey
);
793 ** Restore the cursor to the position it was in (or as close to as possible)
794 ** when saveCursorPosition() was called. Note that this call deletes the
795 ** saved position info stored by saveCursorPosition(), so there can be
796 ** at most one effective restoreCursorPosition() call after each
797 ** saveCursorPosition().
799 static int btreeRestoreCursorPosition(BtCursor
*pCur
){
802 assert( cursorOwnsBtShared(pCur
) );
803 assert( pCur
->eState
>=CURSOR_REQUIRESEEK
);
804 if( pCur
->eState
==CURSOR_FAULT
){
805 return pCur
->skipNext
;
807 pCur
->eState
= CURSOR_INVALID
;
808 rc
= btreeMoveto(pCur
, pCur
->pKey
, pCur
->nKey
, 0, &skipNext
);
810 sqlite3_free(pCur
->pKey
);
812 assert( pCur
->eState
==CURSOR_VALID
|| pCur
->eState
==CURSOR_INVALID
);
813 pCur
->skipNext
|= skipNext
;
814 if( pCur
->skipNext
&& pCur
->eState
==CURSOR_VALID
){
815 pCur
->eState
= CURSOR_SKIPNEXT
;
821 #define restoreCursorPosition(p) \
822 (p->eState>=CURSOR_REQUIRESEEK ? \
823 btreeRestoreCursorPosition(p) : \
827 ** Determine whether or not a cursor has moved from the position where
828 ** it was last placed, or has been invalidated for any other reason.
829 ** Cursors can move when the row they are pointing at is deleted out
830 ** from under them, for example. Cursor might also move if a btree
833 ** Calling this routine with a NULL cursor pointer returns false.
835 ** Use the separate sqlite3BtreeCursorRestore() routine to restore a cursor
836 ** back to where it ought to be if this routine returns true.
838 int sqlite3BtreeCursorHasMoved(BtCursor
*pCur
){
839 return pCur
->eState
!=CURSOR_VALID
;
843 ** Return a pointer to a fake BtCursor object that will always answer
844 ** false to the sqlite3BtreeCursorHasMoved() routine above. The fake
845 ** cursor returned must not be used with any other Btree interface.
847 BtCursor
*sqlite3BtreeFakeValidCursor(void){
848 static u8 fakeCursor
= CURSOR_VALID
;
849 assert( offsetof(BtCursor
, eState
)==0 );
850 return (BtCursor
*)&fakeCursor
;
854 ** This routine restores a cursor back to its original position after it
855 ** has been moved by some outside activity (such as a btree rebalance or
856 ** a row having been deleted out from under the cursor).
858 ** On success, the *pDifferentRow parameter is false if the cursor is left
859 ** pointing at exactly the same row. *pDifferntRow is the row the cursor
860 ** was pointing to has been deleted, forcing the cursor to point to some
863 ** This routine should only be called for a cursor that just returned
864 ** TRUE from sqlite3BtreeCursorHasMoved().
866 int sqlite3BtreeCursorRestore(BtCursor
*pCur
, int *pDifferentRow
){
870 assert( pCur
->eState
!=CURSOR_VALID
);
871 rc
= restoreCursorPosition(pCur
);
876 if( pCur
->eState
!=CURSOR_VALID
){
879 assert( pCur
->skipNext
==0 );
885 #ifdef SQLITE_ENABLE_CURSOR_HINTS
887 ** Provide hints to the cursor. The particular hint given (and the type
888 ** and number of the varargs parameters) is determined by the eHintType
889 ** parameter. See the definitions of the BTREE_HINT_* macros for details.
891 void sqlite3BtreeCursorHint(BtCursor
*pCur
, int eHintType
, ...){
892 /* Used only by system that substitute their own storage engine */
897 ** Provide flag hints to the cursor.
899 void sqlite3BtreeCursorHintFlags(BtCursor
*pCur
, unsigned x
){
900 assert( x
==BTREE_SEEK_EQ
|| x
==BTREE_BULKLOAD
|| x
==0 );
905 #ifndef SQLITE_OMIT_AUTOVACUUM
907 ** Given a page number of a regular database page, return the page
908 ** number for the pointer-map page that contains the entry for the
909 ** input page number.
911 ** Return 0 (not a valid page) for pgno==1 since there is
912 ** no pointer map associated with page 1. The integrity_check logic
913 ** requires that ptrmapPageno(*,1)!=1.
915 static Pgno
ptrmapPageno(BtShared
*pBt
, Pgno pgno
){
916 int nPagesPerMapPage
;
918 assert( sqlite3_mutex_held(pBt
->mutex
) );
919 if( pgno
<2 ) return 0;
920 nPagesPerMapPage
= (pBt
->usableSize
/5)+1;
921 iPtrMap
= (pgno
-2)/nPagesPerMapPage
;
922 ret
= (iPtrMap
*nPagesPerMapPage
) + 2;
923 if( ret
==PENDING_BYTE_PAGE(pBt
) ){
930 ** Write an entry into the pointer map.
932 ** This routine updates the pointer map entry for page number 'key'
933 ** so that it maps to type 'eType' and parent page number 'pgno'.
935 ** If *pRC is initially non-zero (non-SQLITE_OK) then this routine is
936 ** a no-op. If an error occurs, the appropriate error code is written
939 static void ptrmapPut(BtShared
*pBt
, Pgno key
, u8 eType
, Pgno parent
, int *pRC
){
940 DbPage
*pDbPage
; /* The pointer map page */
941 u8
*pPtrmap
; /* The pointer map data */
942 Pgno iPtrmap
; /* The pointer map page number */
943 int offset
; /* Offset in pointer map page */
944 int rc
; /* Return code from subfunctions */
948 assert( sqlite3_mutex_held(pBt
->mutex
) );
949 /* The master-journal page number must never be used as a pointer map page */
950 assert( 0==PTRMAP_ISPAGE(pBt
, PENDING_BYTE_PAGE(pBt
)) );
952 assert( pBt
->autoVacuum
);
954 *pRC
= SQLITE_CORRUPT_BKPT
;
957 iPtrmap
= PTRMAP_PAGENO(pBt
, key
);
958 rc
= sqlite3PagerGet(pBt
->pPager
, iPtrmap
, &pDbPage
, 0);
963 offset
= PTRMAP_PTROFFSET(iPtrmap
, key
);
965 *pRC
= SQLITE_CORRUPT_BKPT
;
968 assert( offset
<= (int)pBt
->usableSize
-5 );
969 pPtrmap
= (u8
*)sqlite3PagerGetData(pDbPage
);
971 if( eType
!=pPtrmap
[offset
] || get4byte(&pPtrmap
[offset
+1])!=parent
){
972 TRACE(("PTRMAP_UPDATE: %d->(%d,%d)\n", key
, eType
, parent
));
973 *pRC
= rc
= sqlite3PagerWrite(pDbPage
);
975 pPtrmap
[offset
] = eType
;
976 put4byte(&pPtrmap
[offset
+1], parent
);
981 sqlite3PagerUnref(pDbPage
);
985 ** Read an entry from the pointer map.
987 ** This routine retrieves the pointer map entry for page 'key', writing
988 ** the type and parent page number to *pEType and *pPgno respectively.
989 ** An error code is returned if something goes wrong, otherwise SQLITE_OK.
991 static int ptrmapGet(BtShared
*pBt
, Pgno key
, u8
*pEType
, Pgno
*pPgno
){
992 DbPage
*pDbPage
; /* The pointer map page */
993 int iPtrmap
; /* Pointer map page index */
994 u8
*pPtrmap
; /* Pointer map page data */
995 int offset
; /* Offset of entry in pointer map */
998 assert( sqlite3_mutex_held(pBt
->mutex
) );
1000 iPtrmap
= PTRMAP_PAGENO(pBt
, key
);
1001 rc
= sqlite3PagerGet(pBt
->pPager
, iPtrmap
, &pDbPage
, 0);
1005 pPtrmap
= (u8
*)sqlite3PagerGetData(pDbPage
);
1007 offset
= PTRMAP_PTROFFSET(iPtrmap
, key
);
1009 sqlite3PagerUnref(pDbPage
);
1010 return SQLITE_CORRUPT_BKPT
;
1012 assert( offset
<= (int)pBt
->usableSize
-5 );
1013 assert( pEType
!=0 );
1014 *pEType
= pPtrmap
[offset
];
1015 if( pPgno
) *pPgno
= get4byte(&pPtrmap
[offset
+1]);
1017 sqlite3PagerUnref(pDbPage
);
1018 if( *pEType
<1 || *pEType
>5 ) return SQLITE_CORRUPT_PGNO(iPtrmap
);
1022 #else /* if defined SQLITE_OMIT_AUTOVACUUM */
1023 #define ptrmapPut(w,x,y,z,rc)
1024 #define ptrmapGet(w,x,y,z) SQLITE_OK
1025 #define ptrmapPutOvflPtr(x, y, rc)
1029 ** Given a btree page and a cell index (0 means the first cell on
1030 ** the page, 1 means the second cell, and so forth) return a pointer
1031 ** to the cell content.
1033 ** findCellPastPtr() does the same except it skips past the initial
1034 ** 4-byte child pointer found on interior pages, if there is one.
1036 ** This routine works only for pages that do not contain overflow cells.
1038 #define findCell(P,I) \
1039 ((P)->aData + ((P)->maskPage & get2byteAligned(&(P)->aCellIdx[2*(I)])))
1040 #define findCellPastPtr(P,I) \
1041 ((P)->aDataOfst + ((P)->maskPage & get2byteAligned(&(P)->aCellIdx[2*(I)])))
1045 ** This is common tail processing for btreeParseCellPtr() and
1046 ** btreeParseCellPtrIndex() for the case when the cell does not fit entirely
1047 ** on a single B-tree page. Make necessary adjustments to the CellInfo
1050 static SQLITE_NOINLINE
void btreeParseCellAdjustSizeForOverflow(
1051 MemPage
*pPage
, /* Page containing the cell */
1052 u8
*pCell
, /* Pointer to the cell text. */
1053 CellInfo
*pInfo
/* Fill in this structure */
1055 /* If the payload will not fit completely on the local page, we have
1056 ** to decide how much to store locally and how much to spill onto
1057 ** overflow pages. The strategy is to minimize the amount of unused
1058 ** space on overflow pages while keeping the amount of local storage
1059 ** in between minLocal and maxLocal.
1061 ** Warning: changing the way overflow payload is distributed in any
1062 ** way will result in an incompatible file format.
1064 int minLocal
; /* Minimum amount of payload held locally */
1065 int maxLocal
; /* Maximum amount of payload held locally */
1066 int surplus
; /* Overflow payload available for local storage */
1068 minLocal
= pPage
->minLocal
;
1069 maxLocal
= pPage
->maxLocal
;
1070 surplus
= minLocal
+ (pInfo
->nPayload
- minLocal
)%(pPage
->pBt
->usableSize
-4);
1071 testcase( surplus
==maxLocal
);
1072 testcase( surplus
==maxLocal
+1 );
1073 if( surplus
<= maxLocal
){
1074 pInfo
->nLocal
= (u16
)surplus
;
1076 pInfo
->nLocal
= (u16
)minLocal
;
1078 pInfo
->nSize
= (u16
)(&pInfo
->pPayload
[pInfo
->nLocal
] - pCell
) + 4;
1082 ** The following routines are implementations of the MemPage.xParseCell()
1085 ** Parse a cell content block and fill in the CellInfo structure.
1087 ** btreeParseCellPtr() => table btree leaf nodes
1088 ** btreeParseCellNoPayload() => table btree internal nodes
1089 ** btreeParseCellPtrIndex() => index btree nodes
1091 ** There is also a wrapper function btreeParseCell() that works for
1092 ** all MemPage types and that references the cell by index rather than
1095 static void btreeParseCellPtrNoPayload(
1096 MemPage
*pPage
, /* Page containing the cell */
1097 u8
*pCell
, /* Pointer to the cell text. */
1098 CellInfo
*pInfo
/* Fill in this structure */
1100 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1101 assert( pPage
->leaf
==0 );
1102 assert( pPage
->childPtrSize
==4 );
1103 #ifndef SQLITE_DEBUG
1104 UNUSED_PARAMETER(pPage
);
1106 pInfo
->nSize
= 4 + getVarint(&pCell
[4], (u64
*)&pInfo
->nKey
);
1107 pInfo
->nPayload
= 0;
1109 pInfo
->pPayload
= 0;
1112 static void btreeParseCellPtr(
1113 MemPage
*pPage
, /* Page containing the cell */
1114 u8
*pCell
, /* Pointer to the cell text. */
1115 CellInfo
*pInfo
/* Fill in this structure */
1117 u8
*pIter
; /* For scanning through pCell */
1118 u32 nPayload
; /* Number of bytes of cell payload */
1119 u64 iKey
; /* Extracted Key value */
1121 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1122 assert( pPage
->leaf
==0 || pPage
->leaf
==1 );
1123 assert( pPage
->intKeyLeaf
);
1124 assert( pPage
->childPtrSize
==0 );
1127 /* The next block of code is equivalent to:
1129 ** pIter += getVarint32(pIter, nPayload);
1131 ** The code is inlined to avoid a function call.
1134 if( nPayload
>=0x80 ){
1135 u8
*pEnd
= &pIter
[8];
1138 nPayload
= (nPayload
<<7) | (*++pIter
& 0x7f);
1139 }while( (*pIter
)>=0x80 && pIter
<pEnd
);
1143 /* The next block of code is equivalent to:
1145 ** pIter += getVarint(pIter, (u64*)&pInfo->nKey);
1147 ** The code is inlined to avoid a function call.
1151 u8
*pEnd
= &pIter
[7];
1154 iKey
= (iKey
<<7) | (*++pIter
& 0x7f);
1155 if( (*pIter
)<0x80 ) break;
1157 iKey
= (iKey
<<8) | *++pIter
;
1164 pInfo
->nKey
= *(i64
*)&iKey
;
1165 pInfo
->nPayload
= nPayload
;
1166 pInfo
->pPayload
= pIter
;
1167 testcase( nPayload
==pPage
->maxLocal
);
1168 testcase( nPayload
==pPage
->maxLocal
+1 );
1169 if( nPayload
<=pPage
->maxLocal
){
1170 /* This is the (easy) common case where the entire payload fits
1171 ** on the local page. No overflow is required.
1173 pInfo
->nSize
= nPayload
+ (u16
)(pIter
- pCell
);
1174 if( pInfo
->nSize
<4 ) pInfo
->nSize
= 4;
1175 pInfo
->nLocal
= (u16
)nPayload
;
1177 btreeParseCellAdjustSizeForOverflow(pPage
, pCell
, pInfo
);
1180 static void btreeParseCellPtrIndex(
1181 MemPage
*pPage
, /* Page containing the cell */
1182 u8
*pCell
, /* Pointer to the cell text. */
1183 CellInfo
*pInfo
/* Fill in this structure */
1185 u8
*pIter
; /* For scanning through pCell */
1186 u32 nPayload
; /* Number of bytes of cell payload */
1188 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1189 assert( pPage
->leaf
==0 || pPage
->leaf
==1 );
1190 assert( pPage
->intKeyLeaf
==0 );
1191 pIter
= pCell
+ pPage
->childPtrSize
;
1193 if( nPayload
>=0x80 ){
1194 u8
*pEnd
= &pIter
[8];
1197 nPayload
= (nPayload
<<7) | (*++pIter
& 0x7f);
1198 }while( *(pIter
)>=0x80 && pIter
<pEnd
);
1201 pInfo
->nKey
= nPayload
;
1202 pInfo
->nPayload
= nPayload
;
1203 pInfo
->pPayload
= pIter
;
1204 testcase( nPayload
==pPage
->maxLocal
);
1205 testcase( nPayload
==pPage
->maxLocal
+1 );
1206 if( nPayload
<=pPage
->maxLocal
){
1207 /* This is the (easy) common case where the entire payload fits
1208 ** on the local page. No overflow is required.
1210 pInfo
->nSize
= nPayload
+ (u16
)(pIter
- pCell
);
1211 if( pInfo
->nSize
<4 ) pInfo
->nSize
= 4;
1212 pInfo
->nLocal
= (u16
)nPayload
;
1214 btreeParseCellAdjustSizeForOverflow(pPage
, pCell
, pInfo
);
1217 static void btreeParseCell(
1218 MemPage
*pPage
, /* Page containing the cell */
1219 int iCell
, /* The cell index. First cell is 0 */
1220 CellInfo
*pInfo
/* Fill in this structure */
1222 pPage
->xParseCell(pPage
, findCell(pPage
, iCell
), pInfo
);
1226 ** The following routines are implementations of the MemPage.xCellSize
1229 ** Compute the total number of bytes that a Cell needs in the cell
1230 ** data area of the btree-page. The return number includes the cell
1231 ** data header and the local payload, but not any overflow page or
1232 ** the space used by the cell pointer.
1234 ** cellSizePtrNoPayload() => table internal nodes
1235 ** cellSizePtr() => all index nodes & table leaf nodes
1237 static u16
cellSizePtr(MemPage
*pPage
, u8
*pCell
){
1238 u8
*pIter
= pCell
+ pPage
->childPtrSize
; /* For looping over bytes of pCell */
1239 u8
*pEnd
; /* End mark for a varint */
1240 u32 nSize
; /* Size value to return */
1243 /* The value returned by this function should always be the same as
1244 ** the (CellInfo.nSize) value found by doing a full parse of the
1245 ** cell. If SQLITE_DEBUG is defined, an assert() at the bottom of
1246 ** this function verifies that this invariant is not violated. */
1248 pPage
->xParseCell(pPage
, pCell
, &debuginfo
);
1256 nSize
= (nSize
<<7) | (*++pIter
& 0x7f);
1257 }while( *(pIter
)>=0x80 && pIter
<pEnd
);
1260 if( pPage
->intKey
){
1261 /* pIter now points at the 64-bit integer key value, a variable length
1262 ** integer. The following block moves pIter to point at the first byte
1263 ** past the end of the key value. */
1265 while( (*pIter
++)&0x80 && pIter
<pEnd
);
1267 testcase( nSize
==pPage
->maxLocal
);
1268 testcase( nSize
==pPage
->maxLocal
+1 );
1269 if( nSize
<=pPage
->maxLocal
){
1270 nSize
+= (u32
)(pIter
- pCell
);
1271 if( nSize
<4 ) nSize
= 4;
1273 int minLocal
= pPage
->minLocal
;
1274 nSize
= minLocal
+ (nSize
- minLocal
) % (pPage
->pBt
->usableSize
- 4);
1275 testcase( nSize
==pPage
->maxLocal
);
1276 testcase( nSize
==pPage
->maxLocal
+1 );
1277 if( nSize
>pPage
->maxLocal
){
1280 nSize
+= 4 + (u16
)(pIter
- pCell
);
1282 assert( nSize
==debuginfo
.nSize
|| CORRUPT_DB
);
1285 static u16
cellSizePtrNoPayload(MemPage
*pPage
, u8
*pCell
){
1286 u8
*pIter
= pCell
+ 4; /* For looping over bytes of pCell */
1287 u8
*pEnd
; /* End mark for a varint */
1290 /* The value returned by this function should always be the same as
1291 ** the (CellInfo.nSize) value found by doing a full parse of the
1292 ** cell. If SQLITE_DEBUG is defined, an assert() at the bottom of
1293 ** this function verifies that this invariant is not violated. */
1295 pPage
->xParseCell(pPage
, pCell
, &debuginfo
);
1297 UNUSED_PARAMETER(pPage
);
1300 assert( pPage
->childPtrSize
==4 );
1302 while( (*pIter
++)&0x80 && pIter
<pEnd
);
1303 assert( debuginfo
.nSize
==(u16
)(pIter
- pCell
) || CORRUPT_DB
);
1304 return (u16
)(pIter
- pCell
);
1309 /* This variation on cellSizePtr() is used inside of assert() statements
1311 static u16
cellSize(MemPage
*pPage
, int iCell
){
1312 return pPage
->xCellSize(pPage
, findCell(pPage
, iCell
));
1316 #ifndef SQLITE_OMIT_AUTOVACUUM
1318 ** If the cell pCell, part of page pPage contains a pointer
1319 ** to an overflow page, insert an entry into the pointer-map
1320 ** for the overflow page.
1322 static void ptrmapPutOvflPtr(MemPage
*pPage
, u8
*pCell
, int *pRC
){
1326 pPage
->xParseCell(pPage
, pCell
, &info
);
1327 if( info
.nLocal
<info
.nPayload
){
1328 Pgno ovfl
= get4byte(&pCell
[info
.nSize
-4]);
1329 ptrmapPut(pPage
->pBt
, ovfl
, PTRMAP_OVERFLOW1
, pPage
->pgno
, pRC
);
1336 ** Defragment the page given. This routine reorganizes cells within the
1337 ** page so that there are no free-blocks on the free-block list.
1339 ** Parameter nMaxFrag is the maximum amount of fragmented space that may be
1340 ** present in the page after this routine returns.
1342 ** EVIDENCE-OF: R-44582-60138 SQLite may from time to time reorganize a
1343 ** b-tree page so that there are no freeblocks or fragment bytes, all
1344 ** unused bytes are contained in the unallocated space region, and all
1345 ** cells are packed tightly at the end of the page.
1347 static int defragmentPage(MemPage
*pPage
, int nMaxFrag
){
1348 int i
; /* Loop counter */
1349 int pc
; /* Address of the i-th cell */
1350 int hdr
; /* Offset to the page header */
1351 int size
; /* Size of a cell */
1352 int usableSize
; /* Number of usable bytes on a page */
1353 int cellOffset
; /* Offset to the cell pointer array */
1354 int cbrk
; /* Offset to the cell content area */
1355 int nCell
; /* Number of cells on the page */
1356 unsigned char *data
; /* The page data */
1357 unsigned char *temp
; /* Temp area for cell content */
1358 unsigned char *src
; /* Source of content */
1359 int iCellFirst
; /* First allowable cell index */
1360 int iCellLast
; /* Last possible cell index */
1362 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1363 assert( pPage
->pBt
!=0 );
1364 assert( pPage
->pBt
->usableSize
<= SQLITE_MAX_PAGE_SIZE
);
1365 assert( pPage
->nOverflow
==0 );
1366 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1368 src
= data
= pPage
->aData
;
1369 hdr
= pPage
->hdrOffset
;
1370 cellOffset
= pPage
->cellOffset
;
1371 nCell
= pPage
->nCell
;
1372 assert( nCell
==get2byte(&data
[hdr
+3]) );
1373 iCellFirst
= cellOffset
+ 2*nCell
;
1374 usableSize
= pPage
->pBt
->usableSize
;
1376 /* This block handles pages with two or fewer free blocks and nMaxFrag
1377 ** or fewer fragmented bytes. In this case it is faster to move the
1378 ** two (or one) blocks of cells using memmove() and add the required
1379 ** offsets to each pointer in the cell-pointer array than it is to
1380 ** reconstruct the entire page. */
1381 if( (int)data
[hdr
+7]<=nMaxFrag
){
1382 int iFree
= get2byte(&data
[hdr
+1]);
1384 int iFree2
= get2byte(&data
[iFree
]);
1386 /* pageFindSlot() has already verified that free blocks are sorted
1387 ** in order of offset within the page, and that no block extends
1388 ** past the end of the page. Provided the two free slots do not
1389 ** overlap, this guarantees that the memmove() calls below will not
1390 ** overwrite the usableSize byte buffer, even if the database page
1392 assert( iFree2
==0 || iFree2
>iFree
);
1393 assert( iFree
+get2byte(&data
[iFree
+2]) <= usableSize
);
1394 assert( iFree2
==0 || iFree2
+get2byte(&data
[iFree2
+2]) <= usableSize
);
1396 if( 0==iFree2
|| (data
[iFree2
]==0 && data
[iFree2
+1]==0) ){
1397 u8
*pEnd
= &data
[cellOffset
+ nCell
*2];
1400 int sz
= get2byte(&data
[iFree
+2]);
1401 int top
= get2byte(&data
[hdr
+5]);
1403 assert( iFree
+sz
<=iFree2
); /* Verified by pageFindSlot() */
1404 sz2
= get2byte(&data
[iFree2
+2]);
1405 assert( iFree
+sz
+sz2
+iFree2
-(iFree
+sz
) <= usableSize
);
1406 memmove(&data
[iFree
+sz
+sz2
], &data
[iFree
+sz
], iFree2
-(iFree
+sz
));
1410 assert( cbrk
+(iFree
-top
) <= usableSize
);
1411 memmove(&data
[cbrk
], &data
[top
], iFree
-top
);
1412 for(pAddr
=&data
[cellOffset
]; pAddr
<pEnd
; pAddr
+=2){
1413 pc
= get2byte(pAddr
);
1414 if( pc
<iFree
){ put2byte(pAddr
, pc
+sz
); }
1415 else if( pc
<iFree2
){ put2byte(pAddr
, pc
+sz2
); }
1417 goto defragment_out
;
1423 iCellLast
= usableSize
- 4;
1424 for(i
=0; i
<nCell
; i
++){
1425 u8
*pAddr
; /* The i-th cell pointer */
1426 pAddr
= &data
[cellOffset
+ i
*2];
1427 pc
= get2byte(pAddr
);
1428 testcase( pc
==iCellFirst
);
1429 testcase( pc
==iCellLast
);
1430 /* These conditions have already been verified in btreeInitPage()
1431 ** if PRAGMA cell_size_check=ON.
1433 if( pc
<iCellFirst
|| pc
>iCellLast
){
1434 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1436 assert( pc
>=iCellFirst
&& pc
<=iCellLast
);
1437 size
= pPage
->xCellSize(pPage
, &src
[pc
]);
1439 if( cbrk
<iCellFirst
|| pc
+size
>usableSize
){
1440 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1442 assert( cbrk
+size
<=usableSize
&& cbrk
>=iCellFirst
);
1443 testcase( cbrk
+size
==usableSize
);
1444 testcase( pc
+size
==usableSize
);
1445 put2byte(pAddr
, cbrk
);
1448 if( cbrk
==pc
) continue;
1449 temp
= sqlite3PagerTempSpace(pPage
->pBt
->pPager
);
1450 x
= get2byte(&data
[hdr
+5]);
1451 memcpy(&temp
[x
], &data
[x
], (cbrk
+size
) - x
);
1454 memcpy(&data
[cbrk
], &src
[pc
], size
);
1459 if( data
[hdr
+7]+cbrk
-iCellFirst
!=pPage
->nFree
){
1460 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1462 assert( cbrk
>=iCellFirst
);
1463 put2byte(&data
[hdr
+5], cbrk
);
1466 memset(&data
[iCellFirst
], 0, cbrk
-iCellFirst
);
1467 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1472 ** Search the free-list on page pPg for space to store a cell nByte bytes in
1473 ** size. If one can be found, return a pointer to the space and remove it
1474 ** from the free-list.
1476 ** If no suitable space can be found on the free-list, return NULL.
1478 ** This function may detect corruption within pPg. If corruption is
1479 ** detected then *pRc is set to SQLITE_CORRUPT and NULL is returned.
1481 ** Slots on the free list that are between 1 and 3 bytes larger than nByte
1482 ** will be ignored if adding the extra space to the fragmentation count
1483 ** causes the fragmentation count to exceed 60.
1485 static u8
*pageFindSlot(MemPage
*pPg
, int nByte
, int *pRc
){
1486 const int hdr
= pPg
->hdrOffset
;
1487 u8
* const aData
= pPg
->aData
;
1488 int iAddr
= hdr
+ 1;
1489 int pc
= get2byte(&aData
[iAddr
]);
1491 int usableSize
= pPg
->pBt
->usableSize
;
1492 int size
; /* Size of the free slot */
1495 while( pc
<=usableSize
-4 ){
1496 /* EVIDENCE-OF: R-22710-53328 The third and fourth bytes of each
1497 ** freeblock form a big-endian integer which is the size of the freeblock
1498 ** in bytes, including the 4-byte header. */
1499 size
= get2byte(&aData
[pc
+2]);
1500 if( (x
= size
- nByte
)>=0 ){
1503 if( size
+pc
> usableSize
){
1504 *pRc
= SQLITE_CORRUPT_PGNO(pPg
->pgno
);
1507 /* EVIDENCE-OF: R-11498-58022 In a well-formed b-tree page, the total
1508 ** number of bytes in fragments may not exceed 60. */
1509 if( aData
[hdr
+7]>57 ) return 0;
1511 /* Remove the slot from the free-list. Update the number of
1512 ** fragmented bytes within the page. */
1513 memcpy(&aData
[iAddr
], &aData
[pc
], 2);
1514 aData
[hdr
+7] += (u8
)x
;
1516 /* The slot remains on the free-list. Reduce its size to account
1517 ** for the portion used by the new allocation. */
1518 put2byte(&aData
[pc
+2], x
);
1520 return &aData
[pc
+ x
];
1523 pc
= get2byte(&aData
[pc
]);
1524 if( pc
<iAddr
+size
) break;
1527 *pRc
= SQLITE_CORRUPT_PGNO(pPg
->pgno
);
1534 ** Allocate nByte bytes of space from within the B-Tree page passed
1535 ** as the first argument. Write into *pIdx the index into pPage->aData[]
1536 ** of the first byte of allocated space. Return either SQLITE_OK or
1537 ** an error code (usually SQLITE_CORRUPT).
1539 ** The caller guarantees that there is sufficient space to make the
1540 ** allocation. This routine might need to defragment in order to bring
1541 ** all the space together, however. This routine will avoid using
1542 ** the first two bytes past the cell pointer area since presumably this
1543 ** allocation is being made in order to insert a new cell, so we will
1544 ** also end up needing a new cell pointer.
1546 static int allocateSpace(MemPage
*pPage
, int nByte
, int *pIdx
){
1547 const int hdr
= pPage
->hdrOffset
; /* Local cache of pPage->hdrOffset */
1548 u8
* const data
= pPage
->aData
; /* Local cache of pPage->aData */
1549 int top
; /* First byte of cell content area */
1550 int rc
= SQLITE_OK
; /* Integer return code */
1551 int gap
; /* First byte of gap between cell pointers and cell content */
1553 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1554 assert( pPage
->pBt
);
1555 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1556 assert( nByte
>=0 ); /* Minimum cell size is 4 */
1557 assert( pPage
->nFree
>=nByte
);
1558 assert( pPage
->nOverflow
==0 );
1559 assert( nByte
< (int)(pPage
->pBt
->usableSize
-8) );
1561 assert( pPage
->cellOffset
== hdr
+ 12 - 4*pPage
->leaf
);
1562 gap
= pPage
->cellOffset
+ 2*pPage
->nCell
;
1563 assert( gap
<=65536 );
1564 /* EVIDENCE-OF: R-29356-02391 If the database uses a 65536-byte page size
1565 ** and the reserved space is zero (the usual value for reserved space)
1566 ** then the cell content offset of an empty page wants to be 65536.
1567 ** However, that integer is too large to be stored in a 2-byte unsigned
1568 ** integer, so a value of 0 is used in its place. */
1569 top
= get2byte(&data
[hdr
+5]);
1570 assert( top
<=(int)pPage
->pBt
->usableSize
); /* Prevent by getAndInitPage() */
1572 if( top
==0 && pPage
->pBt
->usableSize
==65536 ){
1575 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1579 /* If there is enough space between gap and top for one more cell pointer
1580 ** array entry offset, and if the freelist is not empty, then search the
1581 ** freelist looking for a free slot big enough to satisfy the request.
1583 testcase( gap
+2==top
);
1584 testcase( gap
+1==top
);
1585 testcase( gap
==top
);
1586 if( (data
[hdr
+2] || data
[hdr
+1]) && gap
+2<=top
){
1587 u8
*pSpace
= pageFindSlot(pPage
, nByte
, &rc
);
1589 assert( pSpace
>=data
&& (pSpace
- data
)<65536 );
1590 *pIdx
= (int)(pSpace
- data
);
1597 /* The request could not be fulfilled using a freelist slot. Check
1598 ** to see if defragmentation is necessary.
1600 testcase( gap
+2+nByte
==top
);
1601 if( gap
+2+nByte
>top
){
1602 assert( pPage
->nCell
>0 || CORRUPT_DB
);
1603 rc
= defragmentPage(pPage
, MIN(4, pPage
->nFree
- (2+nByte
)));
1605 top
= get2byteNotZero(&data
[hdr
+5]);
1606 assert( gap
+2+nByte
<=top
);
1610 /* Allocate memory from the gap in between the cell pointer array
1611 ** and the cell content area. The btreeInitPage() call has already
1612 ** validated the freelist. Given that the freelist is valid, there
1613 ** is no way that the allocation can extend off the end of the page.
1614 ** The assert() below verifies the previous sentence.
1617 put2byte(&data
[hdr
+5], top
);
1618 assert( top
+nByte
<= (int)pPage
->pBt
->usableSize
);
1624 ** Return a section of the pPage->aData to the freelist.
1625 ** The first byte of the new free block is pPage->aData[iStart]
1626 ** and the size of the block is iSize bytes.
1628 ** Adjacent freeblocks are coalesced.
1630 ** Note that even though the freeblock list was checked by btreeInitPage(),
1631 ** that routine will not detect overlap between cells or freeblocks. Nor
1632 ** does it detect cells or freeblocks that encrouch into the reserved bytes
1633 ** at the end of the page. So do additional corruption checks inside this
1634 ** routine and return SQLITE_CORRUPT if any problems are found.
1636 static int freeSpace(MemPage
*pPage
, u16 iStart
, u16 iSize
){
1637 u16 iPtr
; /* Address of ptr to next freeblock */
1638 u16 iFreeBlk
; /* Address of the next freeblock */
1639 u8 hdr
; /* Page header size. 0 or 100 */
1640 u8 nFrag
= 0; /* Reduction in fragmentation */
1641 u16 iOrigSize
= iSize
; /* Original value of iSize */
1642 u16 x
; /* Offset to cell content area */
1643 u32 iEnd
= iStart
+ iSize
; /* First byte past the iStart buffer */
1644 unsigned char *data
= pPage
->aData
; /* Page content */
1646 assert( pPage
->pBt
!=0 );
1647 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1648 assert( CORRUPT_DB
|| iStart
>=pPage
->hdrOffset
+6+pPage
->childPtrSize
);
1649 assert( CORRUPT_DB
|| iEnd
<= pPage
->pBt
->usableSize
);
1650 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1651 assert( iSize
>=4 ); /* Minimum cell size is 4 */
1652 assert( iStart
<=pPage
->pBt
->usableSize
-4 );
1654 /* The list of freeblocks must be in ascending order. Find the
1655 ** spot on the list where iStart should be inserted.
1657 hdr
= pPage
->hdrOffset
;
1659 if( data
[iPtr
+1]==0 && data
[iPtr
]==0 ){
1660 iFreeBlk
= 0; /* Shortcut for the case when the freelist is empty */
1662 while( (iFreeBlk
= get2byte(&data
[iPtr
]))<iStart
){
1663 if( iFreeBlk
<iPtr
+4 ){
1664 if( iFreeBlk
==0 ) break;
1665 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1669 if( iFreeBlk
>pPage
->pBt
->usableSize
-4 ){
1670 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1672 assert( iFreeBlk
>iPtr
|| iFreeBlk
==0 );
1675 ** iFreeBlk: First freeblock after iStart, or zero if none
1676 ** iPtr: The address of a pointer to iFreeBlk
1678 ** Check to see if iFreeBlk should be coalesced onto the end of iStart.
1680 if( iFreeBlk
&& iEnd
+3>=iFreeBlk
){
1681 nFrag
= iFreeBlk
- iEnd
;
1682 if( iEnd
>iFreeBlk
) return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1683 iEnd
= iFreeBlk
+ get2byte(&data
[iFreeBlk
+2]);
1684 if( iEnd
> pPage
->pBt
->usableSize
){
1685 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1687 iSize
= iEnd
- iStart
;
1688 iFreeBlk
= get2byte(&data
[iFreeBlk
]);
1691 /* If iPtr is another freeblock (that is, if iPtr is not the freelist
1692 ** pointer in the page header) then check to see if iStart should be
1693 ** coalesced onto the end of iPtr.
1696 int iPtrEnd
= iPtr
+ get2byte(&data
[iPtr
+2]);
1697 if( iPtrEnd
+3>=iStart
){
1698 if( iPtrEnd
>iStart
) return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1699 nFrag
+= iStart
- iPtrEnd
;
1700 iSize
= iEnd
- iPtr
;
1704 if( nFrag
>data
[hdr
+7] ) return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1705 data
[hdr
+7] -= nFrag
;
1707 x
= get2byte(&data
[hdr
+5]);
1709 /* The new freeblock is at the beginning of the cell content area,
1710 ** so just extend the cell content area rather than create another
1711 ** freelist entry */
1712 if( iStart
<x
|| iPtr
!=hdr
+1 ) return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1713 put2byte(&data
[hdr
+1], iFreeBlk
);
1714 put2byte(&data
[hdr
+5], iEnd
);
1716 /* Insert the new freeblock into the freelist */
1717 put2byte(&data
[iPtr
], iStart
);
1719 if( pPage
->pBt
->btsFlags
& BTS_FAST_SECURE
){
1720 /* Overwrite deleted information with zeros when the secure_delete
1721 ** option is enabled */
1722 memset(&data
[iStart
], 0, iSize
);
1724 put2byte(&data
[iStart
], iFreeBlk
);
1725 put2byte(&data
[iStart
+2], iSize
);
1726 pPage
->nFree
+= iOrigSize
;
1731 ** Decode the flags byte (the first byte of the header) for a page
1732 ** and initialize fields of the MemPage structure accordingly.
1734 ** Only the following combinations are supported. Anything different
1735 ** indicates a corrupt database files:
1738 ** PTF_ZERODATA | PTF_LEAF
1739 ** PTF_LEAFDATA | PTF_INTKEY
1740 ** PTF_LEAFDATA | PTF_INTKEY | PTF_LEAF
1742 static int decodeFlags(MemPage
*pPage
, int flagByte
){
1743 BtShared
*pBt
; /* A copy of pPage->pBt */
1745 assert( pPage
->hdrOffset
==(pPage
->pgno
==1 ? 100 : 0) );
1746 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1747 pPage
->leaf
= (u8
)(flagByte
>>3); assert( PTF_LEAF
== 1<<3 );
1748 flagByte
&= ~PTF_LEAF
;
1749 pPage
->childPtrSize
= 4-4*pPage
->leaf
;
1750 pPage
->xCellSize
= cellSizePtr
;
1752 if( flagByte
==(PTF_LEAFDATA
| PTF_INTKEY
) ){
1753 /* EVIDENCE-OF: R-07291-35328 A value of 5 (0x05) means the page is an
1754 ** interior table b-tree page. */
1755 assert( (PTF_LEAFDATA
|PTF_INTKEY
)==5 );
1756 /* EVIDENCE-OF: R-26900-09176 A value of 13 (0x0d) means the page is a
1757 ** leaf table b-tree page. */
1758 assert( (PTF_LEAFDATA
|PTF_INTKEY
|PTF_LEAF
)==13 );
1761 pPage
->intKeyLeaf
= 1;
1762 pPage
->xParseCell
= btreeParseCellPtr
;
1764 pPage
->intKeyLeaf
= 0;
1765 pPage
->xCellSize
= cellSizePtrNoPayload
;
1766 pPage
->xParseCell
= btreeParseCellPtrNoPayload
;
1768 pPage
->maxLocal
= pBt
->maxLeaf
;
1769 pPage
->minLocal
= pBt
->minLeaf
;
1770 }else if( flagByte
==PTF_ZERODATA
){
1771 /* EVIDENCE-OF: R-43316-37308 A value of 2 (0x02) means the page is an
1772 ** interior index b-tree page. */
1773 assert( (PTF_ZERODATA
)==2 );
1774 /* EVIDENCE-OF: R-59615-42828 A value of 10 (0x0a) means the page is a
1775 ** leaf index b-tree page. */
1776 assert( (PTF_ZERODATA
|PTF_LEAF
)==10 );
1778 pPage
->intKeyLeaf
= 0;
1779 pPage
->xParseCell
= btreeParseCellPtrIndex
;
1780 pPage
->maxLocal
= pBt
->maxLocal
;
1781 pPage
->minLocal
= pBt
->minLocal
;
1783 /* EVIDENCE-OF: R-47608-56469 Any other value for the b-tree page type is
1785 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1787 pPage
->max1bytePayload
= pBt
->max1bytePayload
;
1792 ** Initialize the auxiliary information for a disk block.
1794 ** Return SQLITE_OK on success. If we see that the page does
1795 ** not contain a well-formed database page, then return
1796 ** SQLITE_CORRUPT. Note that a return of SQLITE_OK does not
1797 ** guarantee that the page is well-formed. It only shows that
1798 ** we failed to detect any corruption.
1800 static int btreeInitPage(MemPage
*pPage
){
1801 int pc
; /* Address of a freeblock within pPage->aData[] */
1802 u8 hdr
; /* Offset to beginning of page header */
1803 u8
*data
; /* Equal to pPage->aData */
1804 BtShared
*pBt
; /* The main btree structure */
1805 int usableSize
; /* Amount of usable space on each page */
1806 u16 cellOffset
; /* Offset from start of page to first cell pointer */
1807 int nFree
; /* Number of unused bytes on the page */
1808 int top
; /* First byte of the cell content area */
1809 int iCellFirst
; /* First allowable cell or freeblock offset */
1810 int iCellLast
; /* Last possible cell or freeblock offset */
1812 assert( pPage
->pBt
!=0 );
1813 assert( pPage
->pBt
->db
!=0 );
1814 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1815 assert( pPage
->pgno
==sqlite3PagerPagenumber(pPage
->pDbPage
) );
1816 assert( pPage
== sqlite3PagerGetExtra(pPage
->pDbPage
) );
1817 assert( pPage
->aData
== sqlite3PagerGetData(pPage
->pDbPage
) );
1818 assert( pPage
->isInit
==0 );
1821 hdr
= pPage
->hdrOffset
;
1822 data
= pPage
->aData
;
1823 /* EVIDENCE-OF: R-28594-02890 The one-byte flag at offset 0 indicating
1824 ** the b-tree page type. */
1825 if( decodeFlags(pPage
, data
[hdr
]) ){
1826 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1828 assert( pBt
->pageSize
>=512 && pBt
->pageSize
<=65536 );
1829 pPage
->maskPage
= (u16
)(pBt
->pageSize
- 1);
1830 pPage
->nOverflow
= 0;
1831 usableSize
= pBt
->usableSize
;
1832 pPage
->cellOffset
= cellOffset
= hdr
+ 8 + pPage
->childPtrSize
;
1833 pPage
->aDataEnd
= &data
[usableSize
];
1834 pPage
->aCellIdx
= &data
[cellOffset
];
1835 pPage
->aDataOfst
= &data
[pPage
->childPtrSize
];
1836 /* EVIDENCE-OF: R-58015-48175 The two-byte integer at offset 5 designates
1837 ** the start of the cell content area. A zero value for this integer is
1838 ** interpreted as 65536. */
1839 top
= get2byteNotZero(&data
[hdr
+5]);
1840 /* EVIDENCE-OF: R-37002-32774 The two-byte integer at offset 3 gives the
1841 ** number of cells on the page. */
1842 pPage
->nCell
= get2byte(&data
[hdr
+3]);
1843 if( pPage
->nCell
>MX_CELL(pBt
) ){
1844 /* To many cells for a single page. The page must be corrupt */
1845 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1847 testcase( pPage
->nCell
==MX_CELL(pBt
) );
1848 /* EVIDENCE-OF: R-24089-57979 If a page contains no cells (which is only
1849 ** possible for a root page of a table that contains no rows) then the
1850 ** offset to the cell content area will equal the page size minus the
1851 ** bytes of reserved space. */
1852 assert( pPage
->nCell
>0 || top
==usableSize
|| CORRUPT_DB
);
1854 /* A malformed database page might cause us to read past the end
1855 ** of page when parsing a cell.
1857 ** The following block of code checks early to see if a cell extends
1858 ** past the end of a page boundary and causes SQLITE_CORRUPT to be
1859 ** returned if it does.
1861 iCellFirst
= cellOffset
+ 2*pPage
->nCell
;
1862 iCellLast
= usableSize
- 4;
1863 if( pBt
->db
->flags
& SQLITE_CellSizeCk
){
1864 int i
; /* Index into the cell pointer array */
1865 int sz
; /* Size of a cell */
1867 if( !pPage
->leaf
) iCellLast
--;
1868 for(i
=0; i
<pPage
->nCell
; i
++){
1869 pc
= get2byteAligned(&data
[cellOffset
+i
*2]);
1870 testcase( pc
==iCellFirst
);
1871 testcase( pc
==iCellLast
);
1872 if( pc
<iCellFirst
|| pc
>iCellLast
){
1873 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1875 sz
= pPage
->xCellSize(pPage
, &data
[pc
]);
1876 testcase( pc
+sz
==usableSize
);
1877 if( pc
+sz
>usableSize
){
1878 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1881 if( !pPage
->leaf
) iCellLast
++;
1884 /* Compute the total free space on the page
1885 ** EVIDENCE-OF: R-23588-34450 The two-byte integer at offset 1 gives the
1886 ** start of the first freeblock on the page, or is zero if there are no
1888 pc
= get2byte(&data
[hdr
+1]);
1889 nFree
= data
[hdr
+7] + top
; /* Init nFree to non-freeblock free space */
1892 if( pc
<iCellFirst
){
1893 /* EVIDENCE-OF: R-55530-52930 In a well-formed b-tree page, there will
1894 ** always be at least one cell before the first freeblock.
1896 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1900 /* Freeblock off the end of the page */
1901 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1903 next
= get2byte(&data
[pc
]);
1904 size
= get2byte(&data
[pc
+2]);
1905 nFree
= nFree
+ size
;
1906 if( next
<=pc
+size
+3 ) break;
1910 /* Freeblock not in ascending order */
1911 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1913 if( pc
+size
>(unsigned int)usableSize
){
1914 /* Last freeblock extends past page end */
1915 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1919 /* At this point, nFree contains the sum of the offset to the start
1920 ** of the cell-content area plus the number of free bytes within
1921 ** the cell-content area. If this is greater than the usable-size
1922 ** of the page, then the page must be corrupted. This check also
1923 ** serves to verify that the offset to the start of the cell-content
1924 ** area, according to the page header, lies within the page.
1926 if( nFree
>usableSize
){
1927 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1929 pPage
->nFree
= (u16
)(nFree
- iCellFirst
);
1935 ** Set up a raw page so that it looks like a database page holding
1938 static void zeroPage(MemPage
*pPage
, int flags
){
1939 unsigned char *data
= pPage
->aData
;
1940 BtShared
*pBt
= pPage
->pBt
;
1941 u8 hdr
= pPage
->hdrOffset
;
1944 assert( sqlite3PagerPagenumber(pPage
->pDbPage
)==pPage
->pgno
);
1945 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
1946 assert( sqlite3PagerGetData(pPage
->pDbPage
) == data
);
1947 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1948 assert( sqlite3_mutex_held(pBt
->mutex
) );
1949 if( pBt
->btsFlags
& BTS_FAST_SECURE
){
1950 memset(&data
[hdr
], 0, pBt
->usableSize
- hdr
);
1952 data
[hdr
] = (char)flags
;
1953 first
= hdr
+ ((flags
&PTF_LEAF
)==0 ? 12 : 8);
1954 memset(&data
[hdr
+1], 0, 4);
1956 put2byte(&data
[hdr
+5], pBt
->usableSize
);
1957 pPage
->nFree
= (u16
)(pBt
->usableSize
- first
);
1958 decodeFlags(pPage
, flags
);
1959 pPage
->cellOffset
= first
;
1960 pPage
->aDataEnd
= &data
[pBt
->usableSize
];
1961 pPage
->aCellIdx
= &data
[first
];
1962 pPage
->aDataOfst
= &data
[pPage
->childPtrSize
];
1963 pPage
->nOverflow
= 0;
1964 assert( pBt
->pageSize
>=512 && pBt
->pageSize
<=65536 );
1965 pPage
->maskPage
= (u16
)(pBt
->pageSize
- 1);
1972 ** Convert a DbPage obtained from the pager into a MemPage used by
1975 static MemPage
*btreePageFromDbPage(DbPage
*pDbPage
, Pgno pgno
, BtShared
*pBt
){
1976 MemPage
*pPage
= (MemPage
*)sqlite3PagerGetExtra(pDbPage
);
1977 if( pgno
!=pPage
->pgno
){
1978 pPage
->aData
= sqlite3PagerGetData(pDbPage
);
1979 pPage
->pDbPage
= pDbPage
;
1982 pPage
->hdrOffset
= pgno
==1 ? 100 : 0;
1984 assert( pPage
->aData
==sqlite3PagerGetData(pDbPage
) );
1989 ** Get a page from the pager. Initialize the MemPage.pBt and
1990 ** MemPage.aData elements if needed. See also: btreeGetUnusedPage().
1992 ** If the PAGER_GET_NOCONTENT flag is set, it means that we do not care
1993 ** about the content of the page at this time. So do not go to the disk
1994 ** to fetch the content. Just fill in the content with zeros for now.
1995 ** If in the future we call sqlite3PagerWrite() on this page, that
1996 ** means we have started to be concerned about content and the disk
1997 ** read should occur at that point.
1999 static int btreeGetPage(
2000 BtShared
*pBt
, /* The btree */
2001 Pgno pgno
, /* Number of the page to fetch */
2002 MemPage
**ppPage
, /* Return the page in this parameter */
2003 int flags
/* PAGER_GET_NOCONTENT or PAGER_GET_READONLY */
2008 assert( flags
==0 || flags
==PAGER_GET_NOCONTENT
|| flags
==PAGER_GET_READONLY
);
2009 assert( sqlite3_mutex_held(pBt
->mutex
) );
2010 rc
= sqlite3PagerGet(pBt
->pPager
, pgno
, (DbPage
**)&pDbPage
, flags
);
2012 *ppPage
= btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2017 ** Retrieve a page from the pager cache. If the requested page is not
2018 ** already in the pager cache return NULL. Initialize the MemPage.pBt and
2019 ** MemPage.aData elements if needed.
2021 static MemPage
*btreePageLookup(BtShared
*pBt
, Pgno pgno
){
2023 assert( sqlite3_mutex_held(pBt
->mutex
) );
2024 pDbPage
= sqlite3PagerLookup(pBt
->pPager
, pgno
);
2026 return btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2032 ** Return the size of the database file in pages. If there is any kind of
2033 ** error, return ((unsigned int)-1).
2035 static Pgno
btreePagecount(BtShared
*pBt
){
2038 u32
sqlite3BtreeLastPage(Btree
*p
){
2039 assert( sqlite3BtreeHoldsMutex(p
) );
2040 assert( ((p
->pBt
->nPage
)&0x8000000)==0 );
2041 return btreePagecount(p
->pBt
);
2045 ** Get a page from the pager and initialize it.
2047 ** If pCur!=0 then the page is being fetched as part of a moveToChild()
2048 ** call. Do additional sanity checking on the page in this case.
2049 ** And if the fetch fails, this routine must decrement pCur->iPage.
2051 ** The page is fetched as read-write unless pCur is not NULL and is
2052 ** a read-only cursor.
2054 ** If an error occurs, then *ppPage is undefined. It
2055 ** may remain unchanged, or it may be set to an invalid value.
2057 static int getAndInitPage(
2058 BtShared
*pBt
, /* The database file */
2059 Pgno pgno
, /* Number of the page to get */
2060 MemPage
**ppPage
, /* Write the page pointer here */
2061 BtCursor
*pCur
, /* Cursor to receive the page, or NULL */
2062 int bReadOnly
/* True for a read-only page */
2066 assert( sqlite3_mutex_held(pBt
->mutex
) );
2067 assert( pCur
==0 || ppPage
==&pCur
->pPage
);
2068 assert( pCur
==0 || bReadOnly
==pCur
->curPagerFlags
);
2069 assert( pCur
==0 || pCur
->iPage
>0 );
2071 if( pgno
>btreePagecount(pBt
) ){
2072 rc
= SQLITE_CORRUPT_BKPT
;
2073 goto getAndInitPage_error
;
2075 rc
= sqlite3PagerGet(pBt
->pPager
, pgno
, (DbPage
**)&pDbPage
, bReadOnly
);
2077 goto getAndInitPage_error
;
2079 *ppPage
= (MemPage
*)sqlite3PagerGetExtra(pDbPage
);
2080 if( (*ppPage
)->isInit
==0 ){
2081 btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2082 rc
= btreeInitPage(*ppPage
);
2083 if( rc
!=SQLITE_OK
){
2084 releasePage(*ppPage
);
2085 goto getAndInitPage_error
;
2088 assert( (*ppPage
)->pgno
==pgno
);
2089 assert( (*ppPage
)->aData
==sqlite3PagerGetData(pDbPage
) );
2091 /* If obtaining a child page for a cursor, we must verify that the page is
2092 ** compatible with the root page. */
2093 if( pCur
&& ((*ppPage
)->nCell
<1 || (*ppPage
)->intKey
!=pCur
->curIntKey
) ){
2094 rc
= SQLITE_CORRUPT_PGNO(pgno
);
2095 releasePage(*ppPage
);
2096 goto getAndInitPage_error
;
2100 getAndInitPage_error
:
2103 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
2105 testcase( pgno
==0 );
2106 assert( pgno
!=0 || rc
==SQLITE_CORRUPT
);
2111 ** Release a MemPage. This should be called once for each prior
2112 ** call to btreeGetPage.
2114 ** Page1 is a special case and must be released using releasePageOne().
2116 static void releasePageNotNull(MemPage
*pPage
){
2117 assert( pPage
->aData
);
2118 assert( pPage
->pBt
);
2119 assert( pPage
->pDbPage
!=0 );
2120 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
2121 assert( sqlite3PagerGetData(pPage
->pDbPage
)==pPage
->aData
);
2122 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2123 sqlite3PagerUnrefNotNull(pPage
->pDbPage
);
2125 static void releasePage(MemPage
*pPage
){
2126 if( pPage
) releasePageNotNull(pPage
);
2128 static void releasePageOne(MemPage
*pPage
){
2130 assert( pPage
->aData
);
2131 assert( pPage
->pBt
);
2132 assert( pPage
->pDbPage
!=0 );
2133 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
2134 assert( sqlite3PagerGetData(pPage
->pDbPage
)==pPage
->aData
);
2135 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2136 sqlite3PagerUnrefPageOne(pPage
->pDbPage
);
2140 ** Get an unused page.
2142 ** This works just like btreeGetPage() with the addition:
2144 ** * If the page is already in use for some other purpose, immediately
2145 ** release it and return an SQLITE_CURRUPT error.
2146 ** * Make sure the isInit flag is clear
2148 static int btreeGetUnusedPage(
2149 BtShared
*pBt
, /* The btree */
2150 Pgno pgno
, /* Number of the page to fetch */
2151 MemPage
**ppPage
, /* Return the page in this parameter */
2152 int flags
/* PAGER_GET_NOCONTENT or PAGER_GET_READONLY */
2154 int rc
= btreeGetPage(pBt
, pgno
, ppPage
, flags
);
2155 if( rc
==SQLITE_OK
){
2156 if( sqlite3PagerPageRefcount((*ppPage
)->pDbPage
)>1 ){
2157 releasePage(*ppPage
);
2159 return SQLITE_CORRUPT_BKPT
;
2161 (*ppPage
)->isInit
= 0;
2170 ** During a rollback, when the pager reloads information into the cache
2171 ** so that the cache is restored to its original state at the start of
2172 ** the transaction, for each page restored this routine is called.
2174 ** This routine needs to reset the extra data section at the end of the
2175 ** page to agree with the restored data.
2177 static void pageReinit(DbPage
*pData
){
2179 pPage
= (MemPage
*)sqlite3PagerGetExtra(pData
);
2180 assert( sqlite3PagerPageRefcount(pData
)>0 );
2181 if( pPage
->isInit
){
2182 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2184 if( sqlite3PagerPageRefcount(pData
)>1 ){
2185 /* pPage might not be a btree page; it might be an overflow page
2186 ** or ptrmap page or a free page. In those cases, the following
2187 ** call to btreeInitPage() will likely return SQLITE_CORRUPT.
2188 ** But no harm is done by this. And it is very important that
2189 ** btreeInitPage() be called on every btree page so we make
2190 ** the call for every page that comes in for re-initing. */
2191 btreeInitPage(pPage
);
2197 ** Invoke the busy handler for a btree.
2199 static int btreeInvokeBusyHandler(void *pArg
){
2200 BtShared
*pBt
= (BtShared
*)pArg
;
2202 assert( sqlite3_mutex_held(pBt
->db
->mutex
) );
2203 return sqlite3InvokeBusyHandler(&pBt
->db
->busyHandler
);
2207 ** Open a database file.
2209 ** zFilename is the name of the database file. If zFilename is NULL
2210 ** then an ephemeral database is created. The ephemeral database might
2211 ** be exclusively in memory, or it might use a disk-based memory cache.
2212 ** Either way, the ephemeral database will be automatically deleted
2213 ** when sqlite3BtreeClose() is called.
2215 ** If zFilename is ":memory:" then an in-memory database is created
2216 ** that is automatically destroyed when it is closed.
2218 ** The "flags" parameter is a bitmask that might contain bits like
2219 ** BTREE_OMIT_JOURNAL and/or BTREE_MEMORY.
2221 ** If the database is already opened in the same database connection
2222 ** and we are in shared cache mode, then the open will fail with an
2223 ** SQLITE_CONSTRAINT error. We cannot allow two or more BtShared
2224 ** objects in the same database connection since doing so will lead
2225 ** to problems with locking.
2227 int sqlite3BtreeOpen(
2228 sqlite3_vfs
*pVfs
, /* VFS to use for this b-tree */
2229 const char *zFilename
, /* Name of the file containing the BTree database */
2230 sqlite3
*db
, /* Associated database handle */
2231 Btree
**ppBtree
, /* Pointer to new Btree object written here */
2232 int flags
, /* Options */
2233 int vfsFlags
/* Flags passed through to sqlite3_vfs.xOpen() */
2235 BtShared
*pBt
= 0; /* Shared part of btree structure */
2236 Btree
*p
; /* Handle to return */
2237 sqlite3_mutex
*mutexOpen
= 0; /* Prevents a race condition. Ticket #3537 */
2238 int rc
= SQLITE_OK
; /* Result code from this function */
2239 u8 nReserve
; /* Byte of unused space on each page */
2240 unsigned char zDbHeader
[100]; /* Database header content */
2242 /* True if opening an ephemeral, temporary database */
2243 const int isTempDb
= zFilename
==0 || zFilename
[0]==0;
2245 /* Set the variable isMemdb to true for an in-memory database, or
2246 ** false for a file-based database.
2248 #ifdef SQLITE_OMIT_MEMORYDB
2249 const int isMemdb
= 0;
2251 const int isMemdb
= (zFilename
&& strcmp(zFilename
, ":memory:")==0)
2252 || (isTempDb
&& sqlite3TempInMemory(db
))
2253 || (vfsFlags
& SQLITE_OPEN_MEMORY
)!=0;
2258 assert( sqlite3_mutex_held(db
->mutex
) );
2259 assert( (flags
&0xff)==flags
); /* flags fit in 8 bits */
2261 /* Only a BTREE_SINGLE database can be BTREE_UNORDERED */
2262 assert( (flags
& BTREE_UNORDERED
)==0 || (flags
& BTREE_SINGLE
)!=0 );
2264 /* A BTREE_SINGLE database is always a temporary and/or ephemeral */
2265 assert( (flags
& BTREE_SINGLE
)==0 || isTempDb
);
2268 flags
|= BTREE_MEMORY
;
2270 if( (vfsFlags
& SQLITE_OPEN_MAIN_DB
)!=0 && (isMemdb
|| isTempDb
) ){
2271 vfsFlags
= (vfsFlags
& ~SQLITE_OPEN_MAIN_DB
) | SQLITE_OPEN_TEMP_DB
;
2273 p
= sqlite3MallocZero(sizeof(Btree
));
2275 return SQLITE_NOMEM_BKPT
;
2277 p
->inTrans
= TRANS_NONE
;
2279 #ifndef SQLITE_OMIT_SHARED_CACHE
2284 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2286 ** If this Btree is a candidate for shared cache, try to find an
2287 ** existing BtShared object that we can share with
2289 if( isTempDb
==0 && (isMemdb
==0 || (vfsFlags
&SQLITE_OPEN_URI
)!=0) ){
2290 if( vfsFlags
& SQLITE_OPEN_SHAREDCACHE
){
2291 int nFilename
= sqlite3Strlen30(zFilename
)+1;
2292 int nFullPathname
= pVfs
->mxPathname
+1;
2293 char *zFullPathname
= sqlite3Malloc(MAX(nFullPathname
,nFilename
));
2294 MUTEX_LOGIC( sqlite3_mutex
*mutexShared
; )
2297 if( !zFullPathname
){
2299 return SQLITE_NOMEM_BKPT
;
2302 memcpy(zFullPathname
, zFilename
, nFilename
);
2304 rc
= sqlite3OsFullPathname(pVfs
, zFilename
,
2305 nFullPathname
, zFullPathname
);
2307 sqlite3_free(zFullPathname
);
2312 #if SQLITE_THREADSAFE
2313 mutexOpen
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_OPEN
);
2314 sqlite3_mutex_enter(mutexOpen
);
2315 mutexShared
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER
);
2316 sqlite3_mutex_enter(mutexShared
);
2318 for(pBt
=GLOBAL(BtShared
*,sqlite3SharedCacheList
); pBt
; pBt
=pBt
->pNext
){
2319 assert( pBt
->nRef
>0 );
2320 if( 0==strcmp(zFullPathname
, sqlite3PagerFilename(pBt
->pPager
, 0))
2321 && sqlite3PagerVfs(pBt
->pPager
)==pVfs
){
2323 for(iDb
=db
->nDb
-1; iDb
>=0; iDb
--){
2324 Btree
*pExisting
= db
->aDb
[iDb
].pBt
;
2325 if( pExisting
&& pExisting
->pBt
==pBt
){
2326 sqlite3_mutex_leave(mutexShared
);
2327 sqlite3_mutex_leave(mutexOpen
);
2328 sqlite3_free(zFullPathname
);
2330 return SQLITE_CONSTRAINT
;
2338 sqlite3_mutex_leave(mutexShared
);
2339 sqlite3_free(zFullPathname
);
2343 /* In debug mode, we mark all persistent databases as sharable
2344 ** even when they are not. This exercises the locking code and
2345 ** gives more opportunity for asserts(sqlite3_mutex_held())
2346 ** statements to find locking problems.
2355 ** The following asserts make sure that structures used by the btree are
2356 ** the right size. This is to guard against size changes that result
2357 ** when compiling on a different architecture.
2359 assert( sizeof(i64
)==8 );
2360 assert( sizeof(u64
)==8 );
2361 assert( sizeof(u32
)==4 );
2362 assert( sizeof(u16
)==2 );
2363 assert( sizeof(Pgno
)==4 );
2365 pBt
= sqlite3MallocZero( sizeof(*pBt
) );
2367 rc
= SQLITE_NOMEM_BKPT
;
2368 goto btree_open_out
;
2370 rc
= sqlite3PagerOpen(pVfs
, &pBt
->pPager
, zFilename
,
2371 sizeof(MemPage
), flags
, vfsFlags
, pageReinit
);
2372 if( rc
==SQLITE_OK
){
2373 sqlite3PagerSetMmapLimit(pBt
->pPager
, db
->szMmap
);
2374 rc
= sqlite3PagerReadFileheader(pBt
->pPager
,sizeof(zDbHeader
),zDbHeader
);
2376 if( rc
!=SQLITE_OK
){
2377 goto btree_open_out
;
2379 pBt
->openFlags
= (u8
)flags
;
2381 sqlite3PagerSetBusyhandler(pBt
->pPager
, btreeInvokeBusyHandler
, pBt
);
2386 if( sqlite3PagerIsreadonly(pBt
->pPager
) ) pBt
->btsFlags
|= BTS_READ_ONLY
;
2387 #if defined(SQLITE_SECURE_DELETE)
2388 pBt
->btsFlags
|= BTS_SECURE_DELETE
;
2389 #elif defined(SQLITE_FAST_SECURE_DELETE)
2390 pBt
->btsFlags
|= BTS_OVERWRITE
;
2392 /* EVIDENCE-OF: R-51873-39618 The page size for a database file is
2393 ** determined by the 2-byte integer located at an offset of 16 bytes from
2394 ** the beginning of the database file. */
2395 pBt
->pageSize
= (zDbHeader
[16]<<8) | (zDbHeader
[17]<<16);
2396 if( pBt
->pageSize
<512 || pBt
->pageSize
>SQLITE_MAX_PAGE_SIZE
2397 || ((pBt
->pageSize
-1)&pBt
->pageSize
)!=0 ){
2399 #ifndef SQLITE_OMIT_AUTOVACUUM
2400 /* If the magic name ":memory:" will create an in-memory database, then
2401 ** leave the autoVacuum mode at 0 (do not auto-vacuum), even if
2402 ** SQLITE_DEFAULT_AUTOVACUUM is true. On the other hand, if
2403 ** SQLITE_OMIT_MEMORYDB has been defined, then ":memory:" is just a
2404 ** regular file-name. In this case the auto-vacuum applies as per normal.
2406 if( zFilename
&& !isMemdb
){
2407 pBt
->autoVacuum
= (SQLITE_DEFAULT_AUTOVACUUM
? 1 : 0);
2408 pBt
->incrVacuum
= (SQLITE_DEFAULT_AUTOVACUUM
==2 ? 1 : 0);
2413 /* EVIDENCE-OF: R-37497-42412 The size of the reserved region is
2414 ** determined by the one-byte unsigned integer found at an offset of 20
2415 ** into the database file header. */
2416 nReserve
= zDbHeader
[20];
2417 pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
2418 #ifndef SQLITE_OMIT_AUTOVACUUM
2419 pBt
->autoVacuum
= (get4byte(&zDbHeader
[36 + 4*4])?1:0);
2420 pBt
->incrVacuum
= (get4byte(&zDbHeader
[36 + 7*4])?1:0);
2423 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
, nReserve
);
2424 if( rc
) goto btree_open_out
;
2425 pBt
->usableSize
= pBt
->pageSize
- nReserve
;
2426 assert( (pBt
->pageSize
& 7)==0 ); /* 8-byte alignment of pageSize */
2428 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2429 /* Add the new BtShared object to the linked list sharable BtShareds.
2433 MUTEX_LOGIC( sqlite3_mutex
*mutexShared
; )
2434 MUTEX_LOGIC( mutexShared
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER
);)
2435 if( SQLITE_THREADSAFE
&& sqlite3GlobalConfig
.bCoreMutex
){
2436 pBt
->mutex
= sqlite3MutexAlloc(SQLITE_MUTEX_FAST
);
2437 if( pBt
->mutex
==0 ){
2438 rc
= SQLITE_NOMEM_BKPT
;
2439 goto btree_open_out
;
2442 sqlite3_mutex_enter(mutexShared
);
2443 pBt
->pNext
= GLOBAL(BtShared
*,sqlite3SharedCacheList
);
2444 GLOBAL(BtShared
*,sqlite3SharedCacheList
) = pBt
;
2445 sqlite3_mutex_leave(mutexShared
);
2450 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2451 /* If the new Btree uses a sharable pBtShared, then link the new
2452 ** Btree into the list of all sharable Btrees for the same connection.
2453 ** The list is kept in ascending order by pBt address.
2458 for(i
=0; i
<db
->nDb
; i
++){
2459 if( (pSib
= db
->aDb
[i
].pBt
)!=0 && pSib
->sharable
){
2460 while( pSib
->pPrev
){ pSib
= pSib
->pPrev
; }
2461 if( (uptr
)p
->pBt
<(uptr
)pSib
->pBt
){
2466 while( pSib
->pNext
&& (uptr
)pSib
->pNext
->pBt
<(uptr
)p
->pBt
){
2469 p
->pNext
= pSib
->pNext
;
2472 p
->pNext
->pPrev
= p
;
2484 if( rc
!=SQLITE_OK
){
2485 if( pBt
&& pBt
->pPager
){
2486 sqlite3PagerClose(pBt
->pPager
, 0);
2492 sqlite3_file
*pFile
;
2494 /* If the B-Tree was successfully opened, set the pager-cache size to the
2495 ** default value. Except, when opening on an existing shared pager-cache,
2496 ** do not change the pager-cache size.
2498 if( sqlite3BtreeSchema(p
, 0, 0)==0 ){
2499 sqlite3PagerSetCachesize(p
->pBt
->pPager
, SQLITE_DEFAULT_CACHE_SIZE
);
2502 pFile
= sqlite3PagerFile(pBt
->pPager
);
2503 if( pFile
->pMethods
){
2504 sqlite3OsFileControlHint(pFile
, SQLITE_FCNTL_PDB
, (void*)&pBt
->db
);
2508 assert( sqlite3_mutex_held(mutexOpen
) );
2509 sqlite3_mutex_leave(mutexOpen
);
2511 assert( rc
!=SQLITE_OK
|| sqlite3BtreeConnectionCount(*ppBtree
)>0 );
2516 ** Decrement the BtShared.nRef counter. When it reaches zero,
2517 ** remove the BtShared structure from the sharing list. Return
2518 ** true if the BtShared.nRef counter reaches zero and return
2519 ** false if it is still positive.
2521 static int removeFromSharingList(BtShared
*pBt
){
2522 #ifndef SQLITE_OMIT_SHARED_CACHE
2523 MUTEX_LOGIC( sqlite3_mutex
*pMaster
; )
2527 assert( sqlite3_mutex_notheld(pBt
->mutex
) );
2528 MUTEX_LOGIC( pMaster
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER
); )
2529 sqlite3_mutex_enter(pMaster
);
2532 if( GLOBAL(BtShared
*,sqlite3SharedCacheList
)==pBt
){
2533 GLOBAL(BtShared
*,sqlite3SharedCacheList
) = pBt
->pNext
;
2535 pList
= GLOBAL(BtShared
*,sqlite3SharedCacheList
);
2536 while( ALWAYS(pList
) && pList
->pNext
!=pBt
){
2539 if( ALWAYS(pList
) ){
2540 pList
->pNext
= pBt
->pNext
;
2543 if( SQLITE_THREADSAFE
){
2544 sqlite3_mutex_free(pBt
->mutex
);
2548 sqlite3_mutex_leave(pMaster
);
2556 ** Make sure pBt->pTmpSpace points to an allocation of
2557 ** MX_CELL_SIZE(pBt) bytes with a 4-byte prefix for a left-child
2560 static void allocateTempSpace(BtShared
*pBt
){
2561 if( !pBt
->pTmpSpace
){
2562 pBt
->pTmpSpace
= sqlite3PageMalloc( pBt
->pageSize
);
2564 /* One of the uses of pBt->pTmpSpace is to format cells before
2565 ** inserting them into a leaf page (function fillInCell()). If
2566 ** a cell is less than 4 bytes in size, it is rounded up to 4 bytes
2567 ** by the various routines that manipulate binary cells. Which
2568 ** can mean that fillInCell() only initializes the first 2 or 3
2569 ** bytes of pTmpSpace, but that the first 4 bytes are copied from
2570 ** it into a database page. This is not actually a problem, but it
2571 ** does cause a valgrind error when the 1 or 2 bytes of unitialized
2572 ** data is passed to system call write(). So to avoid this error,
2573 ** zero the first 4 bytes of temp space here.
2575 ** Also: Provide four bytes of initialized space before the
2576 ** beginning of pTmpSpace as an area available to prepend the
2577 ** left-child pointer to the beginning of a cell.
2579 if( pBt
->pTmpSpace
){
2580 memset(pBt
->pTmpSpace
, 0, 8);
2581 pBt
->pTmpSpace
+= 4;
2587 ** Free the pBt->pTmpSpace allocation
2589 static void freeTempSpace(BtShared
*pBt
){
2590 if( pBt
->pTmpSpace
){
2591 pBt
->pTmpSpace
-= 4;
2592 sqlite3PageFree(pBt
->pTmpSpace
);
2598 ** Close an open database and invalidate all cursors.
2600 int sqlite3BtreeClose(Btree
*p
){
2601 BtShared
*pBt
= p
->pBt
;
2604 /* Close all cursors opened via this handle. */
2605 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2606 sqlite3BtreeEnter(p
);
2607 pCur
= pBt
->pCursor
;
2609 BtCursor
*pTmp
= pCur
;
2611 if( pTmp
->pBtree
==p
){
2612 sqlite3BtreeCloseCursor(pTmp
);
2616 /* Rollback any active transaction and free the handle structure.
2617 ** The call to sqlite3BtreeRollback() drops any table-locks held by
2620 sqlite3BtreeRollback(p
, SQLITE_OK
, 0);
2621 sqlite3BtreeLeave(p
);
2623 /* If there are still other outstanding references to the shared-btree
2624 ** structure, return now. The remainder of this procedure cleans
2625 ** up the shared-btree.
2627 assert( p
->wantToLock
==0 && p
->locked
==0 );
2628 if( !p
->sharable
|| removeFromSharingList(pBt
) ){
2629 /* The pBt is no longer on the sharing list, so we can access
2630 ** it without having to hold the mutex.
2632 ** Clean out and delete the BtShared object.
2634 assert( !pBt
->pCursor
);
2635 sqlite3PagerClose(pBt
->pPager
, p
->db
);
2636 if( pBt
->xFreeSchema
&& pBt
->pSchema
){
2637 pBt
->xFreeSchema(pBt
->pSchema
);
2639 sqlite3DbFree(0, pBt
->pSchema
);
2644 #ifndef SQLITE_OMIT_SHARED_CACHE
2645 assert( p
->wantToLock
==0 );
2646 assert( p
->locked
==0 );
2647 if( p
->pPrev
) p
->pPrev
->pNext
= p
->pNext
;
2648 if( p
->pNext
) p
->pNext
->pPrev
= p
->pPrev
;
2656 ** Change the "soft" limit on the number of pages in the cache.
2657 ** Unused and unmodified pages will be recycled when the number of
2658 ** pages in the cache exceeds this soft limit. But the size of the
2659 ** cache is allowed to grow larger than this limit if it contains
2660 ** dirty pages or pages still in active use.
2662 int sqlite3BtreeSetCacheSize(Btree
*p
, int mxPage
){
2663 BtShared
*pBt
= p
->pBt
;
2664 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2665 sqlite3BtreeEnter(p
);
2666 sqlite3PagerSetCachesize(pBt
->pPager
, mxPage
);
2667 sqlite3BtreeLeave(p
);
2672 ** Change the "spill" limit on the number of pages in the cache.
2673 ** If the number of pages exceeds this limit during a write transaction,
2674 ** the pager might attempt to "spill" pages to the journal early in
2675 ** order to free up memory.
2677 ** The value returned is the current spill size. If zero is passed
2678 ** as an argument, no changes are made to the spill size setting, so
2679 ** using mxPage of 0 is a way to query the current spill size.
2681 int sqlite3BtreeSetSpillSize(Btree
*p
, int mxPage
){
2682 BtShared
*pBt
= p
->pBt
;
2684 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2685 sqlite3BtreeEnter(p
);
2686 res
= sqlite3PagerSetSpillsize(pBt
->pPager
, mxPage
);
2687 sqlite3BtreeLeave(p
);
2691 #if SQLITE_MAX_MMAP_SIZE>0
2693 ** Change the limit on the amount of the database file that may be
2696 int sqlite3BtreeSetMmapLimit(Btree
*p
, sqlite3_int64 szMmap
){
2697 BtShared
*pBt
= p
->pBt
;
2698 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2699 sqlite3BtreeEnter(p
);
2700 sqlite3PagerSetMmapLimit(pBt
->pPager
, szMmap
);
2701 sqlite3BtreeLeave(p
);
2704 #endif /* SQLITE_MAX_MMAP_SIZE>0 */
2707 ** Change the way data is synced to disk in order to increase or decrease
2708 ** how well the database resists damage due to OS crashes and power
2709 ** failures. Level 1 is the same as asynchronous (no syncs() occur and
2710 ** there is a high probability of damage) Level 2 is the default. There
2711 ** is a very low but non-zero probability of damage. Level 3 reduces the
2712 ** probability of damage to near zero but with a write performance reduction.
2714 #ifndef SQLITE_OMIT_PAGER_PRAGMAS
2715 int sqlite3BtreeSetPagerFlags(
2716 Btree
*p
, /* The btree to set the safety level on */
2717 unsigned pgFlags
/* Various PAGER_* flags */
2719 BtShared
*pBt
= p
->pBt
;
2720 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2721 sqlite3BtreeEnter(p
);
2722 sqlite3PagerSetFlags(pBt
->pPager
, pgFlags
);
2723 sqlite3BtreeLeave(p
);
2729 ** Change the default pages size and the number of reserved bytes per page.
2730 ** Or, if the page size has already been fixed, return SQLITE_READONLY
2731 ** without changing anything.
2733 ** The page size must be a power of 2 between 512 and 65536. If the page
2734 ** size supplied does not meet this constraint then the page size is not
2737 ** Page sizes are constrained to be a power of two so that the region
2738 ** of the database file used for locking (beginning at PENDING_BYTE,
2739 ** the first byte past the 1GB boundary, 0x40000000) needs to occur
2740 ** at the beginning of a page.
2742 ** If parameter nReserve is less than zero, then the number of reserved
2743 ** bytes per page is left unchanged.
2745 ** If the iFix!=0 then the BTS_PAGESIZE_FIXED flag is set so that the page size
2746 ** and autovacuum mode can no longer be changed.
2748 int sqlite3BtreeSetPageSize(Btree
*p
, int pageSize
, int nReserve
, int iFix
){
2750 BtShared
*pBt
= p
->pBt
;
2751 assert( nReserve
>=-1 && nReserve
<=255 );
2752 sqlite3BtreeEnter(p
);
2753 #if SQLITE_HAS_CODEC
2754 if( nReserve
>pBt
->optimalReserve
) pBt
->optimalReserve
= (u8
)nReserve
;
2756 if( pBt
->btsFlags
& BTS_PAGESIZE_FIXED
){
2757 sqlite3BtreeLeave(p
);
2758 return SQLITE_READONLY
;
2761 nReserve
= pBt
->pageSize
- pBt
->usableSize
;
2763 assert( nReserve
>=0 && nReserve
<=255 );
2764 if( pageSize
>=512 && pageSize
<=SQLITE_MAX_PAGE_SIZE
&&
2765 ((pageSize
-1)&pageSize
)==0 ){
2766 assert( (pageSize
& 7)==0 );
2767 assert( !pBt
->pCursor
);
2768 pBt
->pageSize
= (u32
)pageSize
;
2771 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
, nReserve
);
2772 pBt
->usableSize
= pBt
->pageSize
- (u16
)nReserve
;
2773 if( iFix
) pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
2774 sqlite3BtreeLeave(p
);
2779 ** Return the currently defined page size
2781 int sqlite3BtreeGetPageSize(Btree
*p
){
2782 return p
->pBt
->pageSize
;
2786 ** This function is similar to sqlite3BtreeGetReserve(), except that it
2787 ** may only be called if it is guaranteed that the b-tree mutex is already
2790 ** This is useful in one special case in the backup API code where it is
2791 ** known that the shared b-tree mutex is held, but the mutex on the
2792 ** database handle that owns *p is not. In this case if sqlite3BtreeEnter()
2793 ** were to be called, it might collide with some other operation on the
2794 ** database handle that owns *p, causing undefined behavior.
2796 int sqlite3BtreeGetReserveNoMutex(Btree
*p
){
2798 assert( sqlite3_mutex_held(p
->pBt
->mutex
) );
2799 n
= p
->pBt
->pageSize
- p
->pBt
->usableSize
;
2804 ** Return the number of bytes of space at the end of every page that
2805 ** are intentually left unused. This is the "reserved" space that is
2806 ** sometimes used by extensions.
2808 ** If SQLITE_HAS_MUTEX is defined then the number returned is the
2809 ** greater of the current reserved space and the maximum requested
2812 int sqlite3BtreeGetOptimalReserve(Btree
*p
){
2814 sqlite3BtreeEnter(p
);
2815 n
= sqlite3BtreeGetReserveNoMutex(p
);
2816 #ifdef SQLITE_HAS_CODEC
2817 if( n
<p
->pBt
->optimalReserve
) n
= p
->pBt
->optimalReserve
;
2819 sqlite3BtreeLeave(p
);
2825 ** Set the maximum page count for a database if mxPage is positive.
2826 ** No changes are made if mxPage is 0 or negative.
2827 ** Regardless of the value of mxPage, return the maximum page count.
2829 int sqlite3BtreeMaxPageCount(Btree
*p
, int mxPage
){
2831 sqlite3BtreeEnter(p
);
2832 n
= sqlite3PagerMaxPageCount(p
->pBt
->pPager
, mxPage
);
2833 sqlite3BtreeLeave(p
);
2838 ** Change the values for the BTS_SECURE_DELETE and BTS_OVERWRITE flags:
2840 ** newFlag==0 Both BTS_SECURE_DELETE and BTS_OVERWRITE are cleared
2841 ** newFlag==1 BTS_SECURE_DELETE set and BTS_OVERWRITE is cleared
2842 ** newFlag==2 BTS_SECURE_DELETE cleared and BTS_OVERWRITE is set
2843 ** newFlag==(-1) No changes
2845 ** This routine acts as a query if newFlag is less than zero
2847 ** With BTS_OVERWRITE set, deleted content is overwritten by zeros, but
2848 ** freelist leaf pages are not written back to the database. Thus in-page
2849 ** deleted content is cleared, but freelist deleted content is not.
2851 ** With BTS_SECURE_DELETE, operation is like BTS_OVERWRITE with the addition
2852 ** that freelist leaf pages are written back into the database, increasing
2853 ** the amount of disk I/O.
2855 int sqlite3BtreeSecureDelete(Btree
*p
, int newFlag
){
2857 if( p
==0 ) return 0;
2858 sqlite3BtreeEnter(p
);
2859 assert( BTS_OVERWRITE
==BTS_SECURE_DELETE
*2 );
2860 assert( BTS_FAST_SECURE
==(BTS_OVERWRITE
|BTS_SECURE_DELETE
) );
2862 p
->pBt
->btsFlags
&= ~BTS_FAST_SECURE
;
2863 p
->pBt
->btsFlags
|= BTS_SECURE_DELETE
*newFlag
;
2865 b
= (p
->pBt
->btsFlags
& BTS_FAST_SECURE
)/BTS_SECURE_DELETE
;
2866 sqlite3BtreeLeave(p
);
2871 ** Change the 'auto-vacuum' property of the database. If the 'autoVacuum'
2872 ** parameter is non-zero, then auto-vacuum mode is enabled. If zero, it
2873 ** is disabled. The default value for the auto-vacuum property is
2874 ** determined by the SQLITE_DEFAULT_AUTOVACUUM macro.
2876 int sqlite3BtreeSetAutoVacuum(Btree
*p
, int autoVacuum
){
2877 #ifdef SQLITE_OMIT_AUTOVACUUM
2878 return SQLITE_READONLY
;
2880 BtShared
*pBt
= p
->pBt
;
2882 u8 av
= (u8
)autoVacuum
;
2884 sqlite3BtreeEnter(p
);
2885 if( (pBt
->btsFlags
& BTS_PAGESIZE_FIXED
)!=0 && (av
?1:0)!=pBt
->autoVacuum
){
2886 rc
= SQLITE_READONLY
;
2888 pBt
->autoVacuum
= av
?1:0;
2889 pBt
->incrVacuum
= av
==2 ?1:0;
2891 sqlite3BtreeLeave(p
);
2897 ** Return the value of the 'auto-vacuum' property. If auto-vacuum is
2898 ** enabled 1 is returned. Otherwise 0.
2900 int sqlite3BtreeGetAutoVacuum(Btree
*p
){
2901 #ifdef SQLITE_OMIT_AUTOVACUUM
2902 return BTREE_AUTOVACUUM_NONE
;
2905 sqlite3BtreeEnter(p
);
2907 (!p
->pBt
->autoVacuum
)?BTREE_AUTOVACUUM_NONE
:
2908 (!p
->pBt
->incrVacuum
)?BTREE_AUTOVACUUM_FULL
:
2909 BTREE_AUTOVACUUM_INCR
2911 sqlite3BtreeLeave(p
);
2917 ** If the user has not set the safety-level for this database connection
2918 ** using "PRAGMA synchronous", and if the safety-level is not already
2919 ** set to the value passed to this function as the second parameter,
2922 #if SQLITE_DEFAULT_SYNCHRONOUS!=SQLITE_DEFAULT_WAL_SYNCHRONOUS
2923 static void setDefaultSyncFlag(BtShared
*pBt
, u8 safety_level
){
2926 if( (db
=pBt
->db
)!=0 && (pDb
=db
->aDb
)!=0 ){
2927 while( pDb
->pBt
==0 || pDb
->pBt
->pBt
!=pBt
){ pDb
++; }
2928 if( pDb
->bSyncSet
==0
2929 && pDb
->safety_level
!=safety_level
2932 pDb
->safety_level
= safety_level
;
2933 sqlite3PagerSetFlags(pBt
->pPager
,
2934 pDb
->safety_level
| (db
->flags
& PAGER_FLAGS_MASK
));
2939 # define setDefaultSyncFlag(pBt,safety_level)
2943 ** Get a reference to pPage1 of the database file. This will
2944 ** also acquire a readlock on that file.
2946 ** SQLITE_OK is returned on success. If the file is not a
2947 ** well-formed database file, then SQLITE_CORRUPT is returned.
2948 ** SQLITE_BUSY is returned if the database is locked. SQLITE_NOMEM
2949 ** is returned if we run out of memory.
2951 static int lockBtree(BtShared
*pBt
){
2952 int rc
; /* Result code from subfunctions */
2953 MemPage
*pPage1
; /* Page 1 of the database file */
2954 int nPage
; /* Number of pages in the database */
2955 int nPageFile
= 0; /* Number of pages in the database file */
2956 int nPageHeader
; /* Number of pages in the database according to hdr */
2958 assert( sqlite3_mutex_held(pBt
->mutex
) );
2959 assert( pBt
->pPage1
==0 );
2960 rc
= sqlite3PagerSharedLock(pBt
->pPager
);
2961 if( rc
!=SQLITE_OK
) return rc
;
2962 rc
= btreeGetPage(pBt
, 1, &pPage1
, 0);
2963 if( rc
!=SQLITE_OK
) return rc
;
2965 /* Do some checking to help insure the file we opened really is
2966 ** a valid database file.
2968 nPage
= nPageHeader
= get4byte(28+(u8
*)pPage1
->aData
);
2969 sqlite3PagerPagecount(pBt
->pPager
, &nPageFile
);
2970 if( nPage
==0 || memcmp(24+(u8
*)pPage1
->aData
, 92+(u8
*)pPage1
->aData
,4)!=0 ){
2976 u8
*page1
= pPage1
->aData
;
2978 /* EVIDENCE-OF: R-43737-39999 Every valid SQLite database file begins
2979 ** with the following 16 bytes (in hex): 53 51 4c 69 74 65 20 66 6f 72 6d
2980 ** 61 74 20 33 00. */
2981 if( memcmp(page1
, zMagicHeader
, 16)!=0 ){
2982 goto page1_init_failed
;
2985 #ifdef SQLITE_OMIT_WAL
2987 pBt
->btsFlags
|= BTS_READ_ONLY
;
2990 goto page1_init_failed
;
2994 pBt
->btsFlags
|= BTS_READ_ONLY
;
2997 goto page1_init_failed
;
3000 /* If the write version is set to 2, this database should be accessed
3001 ** in WAL mode. If the log is not already open, open it now. Then
3002 ** return SQLITE_OK and return without populating BtShared.pPage1.
3003 ** The caller detects this and calls this function again. This is
3004 ** required as the version of page 1 currently in the page1 buffer
3005 ** may not be the latest version - there may be a newer one in the log
3008 if( page1
[19]==2 && (pBt
->btsFlags
& BTS_NO_WAL
)==0 ){
3010 rc
= sqlite3PagerOpenWal(pBt
->pPager
, &isOpen
);
3011 if( rc
!=SQLITE_OK
){
3012 goto page1_init_failed
;
3014 setDefaultSyncFlag(pBt
, SQLITE_DEFAULT_WAL_SYNCHRONOUS
+1);
3016 releasePageOne(pPage1
);
3022 setDefaultSyncFlag(pBt
, SQLITE_DEFAULT_SYNCHRONOUS
+1);
3026 /* EVIDENCE-OF: R-15465-20813 The maximum and minimum embedded payload
3027 ** fractions and the leaf payload fraction values must be 64, 32, and 32.
3029 ** The original design allowed these amounts to vary, but as of
3030 ** version 3.6.0, we require them to be fixed.
3032 if( memcmp(&page1
[21], "\100\040\040",3)!=0 ){
3033 goto page1_init_failed
;
3035 /* EVIDENCE-OF: R-51873-39618 The page size for a database file is
3036 ** determined by the 2-byte integer located at an offset of 16 bytes from
3037 ** the beginning of the database file. */
3038 pageSize
= (page1
[16]<<8) | (page1
[17]<<16);
3039 /* EVIDENCE-OF: R-25008-21688 The size of a page is a power of two
3040 ** between 512 and 65536 inclusive. */
3041 if( ((pageSize
-1)&pageSize
)!=0
3042 || pageSize
>SQLITE_MAX_PAGE_SIZE
3045 goto page1_init_failed
;
3047 assert( (pageSize
& 7)==0 );
3048 /* EVIDENCE-OF: R-59310-51205 The "reserved space" size in the 1-byte
3049 ** integer at offset 20 is the number of bytes of space at the end of
3050 ** each page to reserve for extensions.
3052 ** EVIDENCE-OF: R-37497-42412 The size of the reserved region is
3053 ** determined by the one-byte unsigned integer found at an offset of 20
3054 ** into the database file header. */
3055 usableSize
= pageSize
- page1
[20];
3056 if( (u32
)pageSize
!=pBt
->pageSize
){
3057 /* After reading the first page of the database assuming a page size
3058 ** of BtShared.pageSize, we have discovered that the page-size is
3059 ** actually pageSize. Unlock the database, leave pBt->pPage1 at
3060 ** zero and return SQLITE_OK. The caller will call this function
3061 ** again with the correct page-size.
3063 releasePageOne(pPage1
);
3064 pBt
->usableSize
= usableSize
;
3065 pBt
->pageSize
= pageSize
;
3067 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
,
3068 pageSize
-usableSize
);
3071 if( (pBt
->db
->flags
& SQLITE_WriteSchema
)==0 && nPage
>nPageFile
){
3072 rc
= SQLITE_CORRUPT_BKPT
;
3073 goto page1_init_failed
;
3075 /* EVIDENCE-OF: R-28312-64704 However, the usable size is not allowed to
3076 ** be less than 480. In other words, if the page size is 512, then the
3077 ** reserved space size cannot exceed 32. */
3078 if( usableSize
<480 ){
3079 goto page1_init_failed
;
3081 pBt
->pageSize
= pageSize
;
3082 pBt
->usableSize
= usableSize
;
3083 #ifndef SQLITE_OMIT_AUTOVACUUM
3084 pBt
->autoVacuum
= (get4byte(&page1
[36 + 4*4])?1:0);
3085 pBt
->incrVacuum
= (get4byte(&page1
[36 + 7*4])?1:0);
3089 /* maxLocal is the maximum amount of payload to store locally for
3090 ** a cell. Make sure it is small enough so that at least minFanout
3091 ** cells can will fit on one page. We assume a 10-byte page header.
3092 ** Besides the payload, the cell must store:
3093 ** 2-byte pointer to the cell
3094 ** 4-byte child pointer
3095 ** 9-byte nKey value
3096 ** 4-byte nData value
3097 ** 4-byte overflow page pointer
3098 ** So a cell consists of a 2-byte pointer, a header which is as much as
3099 ** 17 bytes long, 0 to N bytes of payload, and an optional 4 byte overflow
3102 pBt
->maxLocal
= (u16
)((pBt
->usableSize
-12)*64/255 - 23);
3103 pBt
->minLocal
= (u16
)((pBt
->usableSize
-12)*32/255 - 23);
3104 pBt
->maxLeaf
= (u16
)(pBt
->usableSize
- 35);
3105 pBt
->minLeaf
= (u16
)((pBt
->usableSize
-12)*32/255 - 23);
3106 if( pBt
->maxLocal
>127 ){
3107 pBt
->max1bytePayload
= 127;
3109 pBt
->max1bytePayload
= (u8
)pBt
->maxLocal
;
3111 assert( pBt
->maxLeaf
+ 23 <= MX_CELL_SIZE(pBt
) );
3112 pBt
->pPage1
= pPage1
;
3117 releasePageOne(pPage1
);
3124 ** Return the number of cursors open on pBt. This is for use
3125 ** in assert() expressions, so it is only compiled if NDEBUG is not
3128 ** Only write cursors are counted if wrOnly is true. If wrOnly is
3129 ** false then all cursors are counted.
3131 ** For the purposes of this routine, a cursor is any cursor that
3132 ** is capable of reading or writing to the database. Cursors that
3133 ** have been tripped into the CURSOR_FAULT state are not counted.
3135 static int countValidCursors(BtShared
*pBt
, int wrOnly
){
3138 for(pCur
=pBt
->pCursor
; pCur
; pCur
=pCur
->pNext
){
3139 if( (wrOnly
==0 || (pCur
->curFlags
& BTCF_WriteFlag
)!=0)
3140 && pCur
->eState
!=CURSOR_FAULT
) r
++;
3147 ** If there are no outstanding cursors and we are not in the middle
3148 ** of a transaction but there is a read lock on the database, then
3149 ** this routine unrefs the first page of the database file which
3150 ** has the effect of releasing the read lock.
3152 ** If there is a transaction in progress, this routine is a no-op.
3154 static void unlockBtreeIfUnused(BtShared
*pBt
){
3155 assert( sqlite3_mutex_held(pBt
->mutex
) );
3156 assert( countValidCursors(pBt
,0)==0 || pBt
->inTransaction
>TRANS_NONE
);
3157 if( pBt
->inTransaction
==TRANS_NONE
&& pBt
->pPage1
!=0 ){
3158 MemPage
*pPage1
= pBt
->pPage1
;
3159 assert( pPage1
->aData
);
3160 assert( sqlite3PagerRefcount(pBt
->pPager
)==1 );
3162 releasePageOne(pPage1
);
3167 ** If pBt points to an empty file then convert that empty file
3168 ** into a new empty database by initializing the first page of
3171 static int newDatabase(BtShared
*pBt
){
3173 unsigned char *data
;
3176 assert( sqlite3_mutex_held(pBt
->mutex
) );
3183 rc
= sqlite3PagerWrite(pP1
->pDbPage
);
3185 memcpy(data
, zMagicHeader
, sizeof(zMagicHeader
));
3186 assert( sizeof(zMagicHeader
)==16 );
3187 data
[16] = (u8
)((pBt
->pageSize
>>8)&0xff);
3188 data
[17] = (u8
)((pBt
->pageSize
>>16)&0xff);
3191 assert( pBt
->usableSize
<=pBt
->pageSize
&& pBt
->usableSize
+255>=pBt
->pageSize
);
3192 data
[20] = (u8
)(pBt
->pageSize
- pBt
->usableSize
);
3196 memset(&data
[24], 0, 100-24);
3197 zeroPage(pP1
, PTF_INTKEY
|PTF_LEAF
|PTF_LEAFDATA
);
3198 pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
3199 #ifndef SQLITE_OMIT_AUTOVACUUM
3200 assert( pBt
->autoVacuum
==1 || pBt
->autoVacuum
==0 );
3201 assert( pBt
->incrVacuum
==1 || pBt
->incrVacuum
==0 );
3202 put4byte(&data
[36 + 4*4], pBt
->autoVacuum
);
3203 put4byte(&data
[36 + 7*4], pBt
->incrVacuum
);
3211 ** Initialize the first page of the database file (creating a database
3212 ** consisting of a single page and no schema objects). Return SQLITE_OK
3213 ** if successful, or an SQLite error code otherwise.
3215 int sqlite3BtreeNewDb(Btree
*p
){
3217 sqlite3BtreeEnter(p
);
3219 rc
= newDatabase(p
->pBt
);
3220 sqlite3BtreeLeave(p
);
3225 ** Attempt to start a new transaction. A write-transaction
3226 ** is started if the second argument is nonzero, otherwise a read-
3227 ** transaction. If the second argument is 2 or more and exclusive
3228 ** transaction is started, meaning that no other process is allowed
3229 ** to access the database. A preexisting transaction may not be
3230 ** upgraded to exclusive by calling this routine a second time - the
3231 ** exclusivity flag only works for a new transaction.
3233 ** A write-transaction must be started before attempting any
3234 ** changes to the database. None of the following routines
3235 ** will work unless a transaction is started first:
3237 ** sqlite3BtreeCreateTable()
3238 ** sqlite3BtreeCreateIndex()
3239 ** sqlite3BtreeClearTable()
3240 ** sqlite3BtreeDropTable()
3241 ** sqlite3BtreeInsert()
3242 ** sqlite3BtreeDelete()
3243 ** sqlite3BtreeUpdateMeta()
3245 ** If an initial attempt to acquire the lock fails because of lock contention
3246 ** and the database was previously unlocked, then invoke the busy handler
3247 ** if there is one. But if there was previously a read-lock, do not
3248 ** invoke the busy handler - just return SQLITE_BUSY. SQLITE_BUSY is
3249 ** returned when there is already a read-lock in order to avoid a deadlock.
3251 ** Suppose there are two processes A and B. A has a read lock and B has
3252 ** a reserved lock. B tries to promote to exclusive but is blocked because
3253 ** of A's read lock. A tries to promote to reserved but is blocked by B.
3254 ** One or the other of the two processes must give way or there can be
3255 ** no progress. By returning SQLITE_BUSY and not invoking the busy callback
3256 ** when A already has a read lock, we encourage A to give up and let B
3259 int sqlite3BtreeBeginTrans(Btree
*p
, int wrflag
){
3260 BtShared
*pBt
= p
->pBt
;
3263 sqlite3BtreeEnter(p
);
3266 /* If the btree is already in a write-transaction, or it
3267 ** is already in a read-transaction and a read-transaction
3268 ** is requested, this is a no-op.
3270 if( p
->inTrans
==TRANS_WRITE
|| (p
->inTrans
==TRANS_READ
&& !wrflag
) ){
3273 assert( pBt
->inTransaction
==TRANS_WRITE
|| IfNotOmitAV(pBt
->bDoTruncate
)==0 );
3275 /* Write transactions are not possible on a read-only database */
3276 if( (pBt
->btsFlags
& BTS_READ_ONLY
)!=0 && wrflag
){
3277 rc
= SQLITE_READONLY
;
3281 #ifndef SQLITE_OMIT_SHARED_CACHE
3283 sqlite3
*pBlock
= 0;
3284 /* If another database handle has already opened a write transaction
3285 ** on this shared-btree structure and a second write transaction is
3286 ** requested, return SQLITE_LOCKED.
3288 if( (wrflag
&& pBt
->inTransaction
==TRANS_WRITE
)
3289 || (pBt
->btsFlags
& BTS_PENDING
)!=0
3291 pBlock
= pBt
->pWriter
->db
;
3292 }else if( wrflag
>1 ){
3294 for(pIter
=pBt
->pLock
; pIter
; pIter
=pIter
->pNext
){
3295 if( pIter
->pBtree
!=p
){
3296 pBlock
= pIter
->pBtree
->db
;
3302 sqlite3ConnectionBlocked(p
->db
, pBlock
);
3303 rc
= SQLITE_LOCKED_SHAREDCACHE
;
3309 /* Any read-only or read-write transaction implies a read-lock on
3310 ** page 1. So if some other shared-cache client already has a write-lock
3311 ** on page 1, the transaction cannot be opened. */
3312 rc
= querySharedCacheTableLock(p
, MASTER_ROOT
, READ_LOCK
);
3313 if( SQLITE_OK
!=rc
) goto trans_begun
;
3315 pBt
->btsFlags
&= ~BTS_INITIALLY_EMPTY
;
3316 if( pBt
->nPage
==0 ) pBt
->btsFlags
|= BTS_INITIALLY_EMPTY
;
3318 /* Call lockBtree() until either pBt->pPage1 is populated or
3319 ** lockBtree() returns something other than SQLITE_OK. lockBtree()
3320 ** may return SQLITE_OK but leave pBt->pPage1 set to 0 if after
3321 ** reading page 1 it discovers that the page-size of the database
3322 ** file is not pBt->pageSize. In this case lockBtree() will update
3323 ** pBt->pageSize to the page-size of the file on disk.
3325 while( pBt
->pPage1
==0 && SQLITE_OK
==(rc
= lockBtree(pBt
)) );
3327 if( rc
==SQLITE_OK
&& wrflag
){
3328 if( (pBt
->btsFlags
& BTS_READ_ONLY
)!=0 ){
3329 rc
= SQLITE_READONLY
;
3331 rc
= sqlite3PagerBegin(pBt
->pPager
,wrflag
>1,sqlite3TempInMemory(p
->db
));
3332 if( rc
==SQLITE_OK
){
3333 rc
= newDatabase(pBt
);
3338 if( rc
!=SQLITE_OK
){
3339 unlockBtreeIfUnused(pBt
);
3341 }while( (rc
&0xFF)==SQLITE_BUSY
&& pBt
->inTransaction
==TRANS_NONE
&&
3342 btreeInvokeBusyHandler(pBt
) );
3344 if( rc
==SQLITE_OK
){
3345 if( p
->inTrans
==TRANS_NONE
){
3346 pBt
->nTransaction
++;
3347 #ifndef SQLITE_OMIT_SHARED_CACHE
3349 assert( p
->lock
.pBtree
==p
&& p
->lock
.iTable
==1 );
3350 p
->lock
.eLock
= READ_LOCK
;
3351 p
->lock
.pNext
= pBt
->pLock
;
3352 pBt
->pLock
= &p
->lock
;
3356 p
->inTrans
= (wrflag
?TRANS_WRITE
:TRANS_READ
);
3357 if( p
->inTrans
>pBt
->inTransaction
){
3358 pBt
->inTransaction
= p
->inTrans
;
3361 MemPage
*pPage1
= pBt
->pPage1
;
3362 #ifndef SQLITE_OMIT_SHARED_CACHE
3363 assert( !pBt
->pWriter
);
3365 pBt
->btsFlags
&= ~BTS_EXCLUSIVE
;
3366 if( wrflag
>1 ) pBt
->btsFlags
|= BTS_EXCLUSIVE
;
3369 /* If the db-size header field is incorrect (as it may be if an old
3370 ** client has been writing the database file), update it now. Doing
3371 ** this sooner rather than later means the database size can safely
3372 ** re-read the database size from page 1 if a savepoint or transaction
3373 ** rollback occurs within the transaction.
3375 if( pBt
->nPage
!=get4byte(&pPage1
->aData
[28]) ){
3376 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
3377 if( rc
==SQLITE_OK
){
3378 put4byte(&pPage1
->aData
[28], pBt
->nPage
);
3386 if( rc
==SQLITE_OK
&& wrflag
){
3387 /* This call makes sure that the pager has the correct number of
3388 ** open savepoints. If the second parameter is greater than 0 and
3389 ** the sub-journal is not already open, then it will be opened here.
3391 rc
= sqlite3PagerOpenSavepoint(pBt
->pPager
, p
->db
->nSavepoint
);
3395 sqlite3BtreeLeave(p
);
3399 #ifndef SQLITE_OMIT_AUTOVACUUM
3402 ** Set the pointer-map entries for all children of page pPage. Also, if
3403 ** pPage contains cells that point to overflow pages, set the pointer
3404 ** map entries for the overflow pages as well.
3406 static int setChildPtrmaps(MemPage
*pPage
){
3407 int i
; /* Counter variable */
3408 int nCell
; /* Number of cells in page pPage */
3409 int rc
; /* Return code */
3410 BtShared
*pBt
= pPage
->pBt
;
3411 Pgno pgno
= pPage
->pgno
;
3413 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
3414 rc
= pPage
->isInit
? SQLITE_OK
: btreeInitPage(pPage
);
3415 if( rc
!=SQLITE_OK
) return rc
;
3416 nCell
= pPage
->nCell
;
3418 for(i
=0; i
<nCell
; i
++){
3419 u8
*pCell
= findCell(pPage
, i
);
3421 ptrmapPutOvflPtr(pPage
, pCell
, &rc
);
3424 Pgno childPgno
= get4byte(pCell
);
3425 ptrmapPut(pBt
, childPgno
, PTRMAP_BTREE
, pgno
, &rc
);
3430 Pgno childPgno
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
3431 ptrmapPut(pBt
, childPgno
, PTRMAP_BTREE
, pgno
, &rc
);
3438 ** Somewhere on pPage is a pointer to page iFrom. Modify this pointer so
3439 ** that it points to iTo. Parameter eType describes the type of pointer to
3440 ** be modified, as follows:
3442 ** PTRMAP_BTREE: pPage is a btree-page. The pointer points at a child
3445 ** PTRMAP_OVERFLOW1: pPage is a btree-page. The pointer points at an overflow
3446 ** page pointed to by one of the cells on pPage.
3448 ** PTRMAP_OVERFLOW2: pPage is an overflow-page. The pointer points at the next
3449 ** overflow page in the list.
3451 static int modifyPagePointer(MemPage
*pPage
, Pgno iFrom
, Pgno iTo
, u8 eType
){
3452 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
3453 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
3454 if( eType
==PTRMAP_OVERFLOW2
){
3455 /* The pointer is always the first 4 bytes of the page in this case. */
3456 if( get4byte(pPage
->aData
)!=iFrom
){
3457 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
3459 put4byte(pPage
->aData
, iTo
);
3465 rc
= pPage
->isInit
? SQLITE_OK
: btreeInitPage(pPage
);
3467 nCell
= pPage
->nCell
;
3469 for(i
=0; i
<nCell
; i
++){
3470 u8
*pCell
= findCell(pPage
, i
);
3471 if( eType
==PTRMAP_OVERFLOW1
){
3473 pPage
->xParseCell(pPage
, pCell
, &info
);
3474 if( info
.nLocal
<info
.nPayload
){
3475 if( pCell
+info
.nSize
> pPage
->aData
+pPage
->pBt
->usableSize
){
3476 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
3478 if( iFrom
==get4byte(pCell
+info
.nSize
-4) ){
3479 put4byte(pCell
+info
.nSize
-4, iTo
);
3484 if( get4byte(pCell
)==iFrom
){
3485 put4byte(pCell
, iTo
);
3492 if( eType
!=PTRMAP_BTREE
||
3493 get4byte(&pPage
->aData
[pPage
->hdrOffset
+8])!=iFrom
){
3494 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
3496 put4byte(&pPage
->aData
[pPage
->hdrOffset
+8], iTo
);
3504 ** Move the open database page pDbPage to location iFreePage in the
3505 ** database. The pDbPage reference remains valid.
3507 ** The isCommit flag indicates that there is no need to remember that
3508 ** the journal needs to be sync()ed before database page pDbPage->pgno
3509 ** can be written to. The caller has already promised not to write to that
3512 static int relocatePage(
3513 BtShared
*pBt
, /* Btree */
3514 MemPage
*pDbPage
, /* Open page to move */
3515 u8 eType
, /* Pointer map 'type' entry for pDbPage */
3516 Pgno iPtrPage
, /* Pointer map 'page-no' entry for pDbPage */
3517 Pgno iFreePage
, /* The location to move pDbPage to */
3518 int isCommit
/* isCommit flag passed to sqlite3PagerMovepage */
3520 MemPage
*pPtrPage
; /* The page that contains a pointer to pDbPage */
3521 Pgno iDbPage
= pDbPage
->pgno
;
3522 Pager
*pPager
= pBt
->pPager
;
3525 assert( eType
==PTRMAP_OVERFLOW2
|| eType
==PTRMAP_OVERFLOW1
||
3526 eType
==PTRMAP_BTREE
|| eType
==PTRMAP_ROOTPAGE
);
3527 assert( sqlite3_mutex_held(pBt
->mutex
) );
3528 assert( pDbPage
->pBt
==pBt
);
3530 /* Move page iDbPage from its current location to page number iFreePage */
3531 TRACE(("AUTOVACUUM: Moving %d to free page %d (ptr page %d type %d)\n",
3532 iDbPage
, iFreePage
, iPtrPage
, eType
));
3533 rc
= sqlite3PagerMovepage(pPager
, pDbPage
->pDbPage
, iFreePage
, isCommit
);
3534 if( rc
!=SQLITE_OK
){
3537 pDbPage
->pgno
= iFreePage
;
3539 /* If pDbPage was a btree-page, then it may have child pages and/or cells
3540 ** that point to overflow pages. The pointer map entries for all these
3541 ** pages need to be changed.
3543 ** If pDbPage is an overflow page, then the first 4 bytes may store a
3544 ** pointer to a subsequent overflow page. If this is the case, then
3545 ** the pointer map needs to be updated for the subsequent overflow page.
3547 if( eType
==PTRMAP_BTREE
|| eType
==PTRMAP_ROOTPAGE
){
3548 rc
= setChildPtrmaps(pDbPage
);
3549 if( rc
!=SQLITE_OK
){
3553 Pgno nextOvfl
= get4byte(pDbPage
->aData
);
3555 ptrmapPut(pBt
, nextOvfl
, PTRMAP_OVERFLOW2
, iFreePage
, &rc
);
3556 if( rc
!=SQLITE_OK
){
3562 /* Fix the database pointer on page iPtrPage that pointed at iDbPage so
3563 ** that it points at iFreePage. Also fix the pointer map entry for
3566 if( eType
!=PTRMAP_ROOTPAGE
){
3567 rc
= btreeGetPage(pBt
, iPtrPage
, &pPtrPage
, 0);
3568 if( rc
!=SQLITE_OK
){
3571 rc
= sqlite3PagerWrite(pPtrPage
->pDbPage
);
3572 if( rc
!=SQLITE_OK
){
3573 releasePage(pPtrPage
);
3576 rc
= modifyPagePointer(pPtrPage
, iDbPage
, iFreePage
, eType
);
3577 releasePage(pPtrPage
);
3578 if( rc
==SQLITE_OK
){
3579 ptrmapPut(pBt
, iFreePage
, eType
, iPtrPage
, &rc
);
3585 /* Forward declaration required by incrVacuumStep(). */
3586 static int allocateBtreePage(BtShared
*, MemPage
**, Pgno
*, Pgno
, u8
);
3589 ** Perform a single step of an incremental-vacuum. If successful, return
3590 ** SQLITE_OK. If there is no work to do (and therefore no point in
3591 ** calling this function again), return SQLITE_DONE. Or, if an error
3592 ** occurs, return some other error code.
3594 ** More specifically, this function attempts to re-organize the database so
3595 ** that the last page of the file currently in use is no longer in use.
3597 ** Parameter nFin is the number of pages that this database would contain
3598 ** were this function called until it returns SQLITE_DONE.
3600 ** If the bCommit parameter is non-zero, this function assumes that the
3601 ** caller will keep calling incrVacuumStep() until it returns SQLITE_DONE
3602 ** or an error. bCommit is passed true for an auto-vacuum-on-commit
3603 ** operation, or false for an incremental vacuum.
3605 static int incrVacuumStep(BtShared
*pBt
, Pgno nFin
, Pgno iLastPg
, int bCommit
){
3606 Pgno nFreeList
; /* Number of pages still on the free-list */
3609 assert( sqlite3_mutex_held(pBt
->mutex
) );
3610 assert( iLastPg
>nFin
);
3612 if( !PTRMAP_ISPAGE(pBt
, iLastPg
) && iLastPg
!=PENDING_BYTE_PAGE(pBt
) ){
3616 nFreeList
= get4byte(&pBt
->pPage1
->aData
[36]);
3621 rc
= ptrmapGet(pBt
, iLastPg
, &eType
, &iPtrPage
);
3622 if( rc
!=SQLITE_OK
){
3625 if( eType
==PTRMAP_ROOTPAGE
){
3626 return SQLITE_CORRUPT_BKPT
;
3629 if( eType
==PTRMAP_FREEPAGE
){
3631 /* Remove the page from the files free-list. This is not required
3632 ** if bCommit is non-zero. In that case, the free-list will be
3633 ** truncated to zero after this function returns, so it doesn't
3634 ** matter if it still contains some garbage entries.
3638 rc
= allocateBtreePage(pBt
, &pFreePg
, &iFreePg
, iLastPg
, BTALLOC_EXACT
);
3639 if( rc
!=SQLITE_OK
){
3642 assert( iFreePg
==iLastPg
);
3643 releasePage(pFreePg
);
3646 Pgno iFreePg
; /* Index of free page to move pLastPg to */
3648 u8 eMode
= BTALLOC_ANY
; /* Mode parameter for allocateBtreePage() */
3649 Pgno iNear
= 0; /* nearby parameter for allocateBtreePage() */
3651 rc
= btreeGetPage(pBt
, iLastPg
, &pLastPg
, 0);
3652 if( rc
!=SQLITE_OK
){
3656 /* If bCommit is zero, this loop runs exactly once and page pLastPg
3657 ** is swapped with the first free page pulled off the free list.
3659 ** On the other hand, if bCommit is greater than zero, then keep
3660 ** looping until a free-page located within the first nFin pages
3661 ** of the file is found.
3669 rc
= allocateBtreePage(pBt
, &pFreePg
, &iFreePg
, iNear
, eMode
);
3670 if( rc
!=SQLITE_OK
){
3671 releasePage(pLastPg
);
3674 releasePage(pFreePg
);
3675 }while( bCommit
&& iFreePg
>nFin
);
3676 assert( iFreePg
<iLastPg
);
3678 rc
= relocatePage(pBt
, pLastPg
, eType
, iPtrPage
, iFreePg
, bCommit
);
3679 releasePage(pLastPg
);
3680 if( rc
!=SQLITE_OK
){
3689 }while( iLastPg
==PENDING_BYTE_PAGE(pBt
) || PTRMAP_ISPAGE(pBt
, iLastPg
) );
3690 pBt
->bDoTruncate
= 1;
3691 pBt
->nPage
= iLastPg
;
3697 ** The database opened by the first argument is an auto-vacuum database
3698 ** nOrig pages in size containing nFree free pages. Return the expected
3699 ** size of the database in pages following an auto-vacuum operation.
3701 static Pgno
finalDbSize(BtShared
*pBt
, Pgno nOrig
, Pgno nFree
){
3702 int nEntry
; /* Number of entries on one ptrmap page */
3703 Pgno nPtrmap
; /* Number of PtrMap pages to be freed */
3704 Pgno nFin
; /* Return value */
3706 nEntry
= pBt
->usableSize
/5;
3707 nPtrmap
= (nFree
-nOrig
+PTRMAP_PAGENO(pBt
, nOrig
)+nEntry
)/nEntry
;
3708 nFin
= nOrig
- nFree
- nPtrmap
;
3709 if( nOrig
>PENDING_BYTE_PAGE(pBt
) && nFin
<PENDING_BYTE_PAGE(pBt
) ){
3712 while( PTRMAP_ISPAGE(pBt
, nFin
) || nFin
==PENDING_BYTE_PAGE(pBt
) ){
3720 ** A write-transaction must be opened before calling this function.
3721 ** It performs a single unit of work towards an incremental vacuum.
3723 ** If the incremental vacuum is finished after this function has run,
3724 ** SQLITE_DONE is returned. If it is not finished, but no error occurred,
3725 ** SQLITE_OK is returned. Otherwise an SQLite error code.
3727 int sqlite3BtreeIncrVacuum(Btree
*p
){
3729 BtShared
*pBt
= p
->pBt
;
3731 sqlite3BtreeEnter(p
);
3732 assert( pBt
->inTransaction
==TRANS_WRITE
&& p
->inTrans
==TRANS_WRITE
);
3733 if( !pBt
->autoVacuum
){
3736 Pgno nOrig
= btreePagecount(pBt
);
3737 Pgno nFree
= get4byte(&pBt
->pPage1
->aData
[36]);
3738 Pgno nFin
= finalDbSize(pBt
, nOrig
, nFree
);
3741 rc
= SQLITE_CORRUPT_BKPT
;
3742 }else if( nFree
>0 ){
3743 rc
= saveAllCursors(pBt
, 0, 0);
3744 if( rc
==SQLITE_OK
){
3745 invalidateAllOverflowCache(pBt
);
3746 rc
= incrVacuumStep(pBt
, nFin
, nOrig
, 0);
3748 if( rc
==SQLITE_OK
){
3749 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
3750 put4byte(&pBt
->pPage1
->aData
[28], pBt
->nPage
);
3756 sqlite3BtreeLeave(p
);
3761 ** This routine is called prior to sqlite3PagerCommit when a transaction
3762 ** is committed for an auto-vacuum database.
3764 ** If SQLITE_OK is returned, then *pnTrunc is set to the number of pages
3765 ** the database file should be truncated to during the commit process.
3766 ** i.e. the database has been reorganized so that only the first *pnTrunc
3767 ** pages are in use.
3769 static int autoVacuumCommit(BtShared
*pBt
){
3771 Pager
*pPager
= pBt
->pPager
;
3772 VVA_ONLY( int nRef
= sqlite3PagerRefcount(pPager
); )
3774 assert( sqlite3_mutex_held(pBt
->mutex
) );
3775 invalidateAllOverflowCache(pBt
);
3776 assert(pBt
->autoVacuum
);
3777 if( !pBt
->incrVacuum
){
3778 Pgno nFin
; /* Number of pages in database after autovacuuming */
3779 Pgno nFree
; /* Number of pages on the freelist initially */
3780 Pgno iFree
; /* The next page to be freed */
3781 Pgno nOrig
; /* Database size before freeing */
3783 nOrig
= btreePagecount(pBt
);
3784 if( PTRMAP_ISPAGE(pBt
, nOrig
) || nOrig
==PENDING_BYTE_PAGE(pBt
) ){
3785 /* It is not possible to create a database for which the final page
3786 ** is either a pointer-map page or the pending-byte page. If one
3787 ** is encountered, this indicates corruption.
3789 return SQLITE_CORRUPT_BKPT
;
3792 nFree
= get4byte(&pBt
->pPage1
->aData
[36]);
3793 nFin
= finalDbSize(pBt
, nOrig
, nFree
);
3794 if( nFin
>nOrig
) return SQLITE_CORRUPT_BKPT
;
3796 rc
= saveAllCursors(pBt
, 0, 0);
3798 for(iFree
=nOrig
; iFree
>nFin
&& rc
==SQLITE_OK
; iFree
--){
3799 rc
= incrVacuumStep(pBt
, nFin
, iFree
, 1);
3801 if( (rc
==SQLITE_DONE
|| rc
==SQLITE_OK
) && nFree
>0 ){
3802 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
3803 put4byte(&pBt
->pPage1
->aData
[32], 0);
3804 put4byte(&pBt
->pPage1
->aData
[36], 0);
3805 put4byte(&pBt
->pPage1
->aData
[28], nFin
);
3806 pBt
->bDoTruncate
= 1;
3809 if( rc
!=SQLITE_OK
){
3810 sqlite3PagerRollback(pPager
);
3814 assert( nRef
>=sqlite3PagerRefcount(pPager
) );
3818 #else /* ifndef SQLITE_OMIT_AUTOVACUUM */
3819 # define setChildPtrmaps(x) SQLITE_OK
3823 ** This routine does the first phase of a two-phase commit. This routine
3824 ** causes a rollback journal to be created (if it does not already exist)
3825 ** and populated with enough information so that if a power loss occurs
3826 ** the database can be restored to its original state by playing back
3827 ** the journal. Then the contents of the journal are flushed out to
3828 ** the disk. After the journal is safely on oxide, the changes to the
3829 ** database are written into the database file and flushed to oxide.
3830 ** At the end of this call, the rollback journal still exists on the
3831 ** disk and we are still holding all locks, so the transaction has not
3832 ** committed. See sqlite3BtreeCommitPhaseTwo() for the second phase of the
3835 ** This call is a no-op if no write-transaction is currently active on pBt.
3837 ** Otherwise, sync the database file for the btree pBt. zMaster points to
3838 ** the name of a master journal file that should be written into the
3839 ** individual journal file, or is NULL, indicating no master journal file
3840 ** (single database transaction).
3842 ** When this is called, the master journal should already have been
3843 ** created, populated with this journal pointer and synced to disk.
3845 ** Once this is routine has returned, the only thing required to commit
3846 ** the write-transaction for this database file is to delete the journal.
3848 int sqlite3BtreeCommitPhaseOne(Btree
*p
, const char *zMaster
){
3850 if( p
->inTrans
==TRANS_WRITE
){
3851 BtShared
*pBt
= p
->pBt
;
3852 sqlite3BtreeEnter(p
);
3853 #ifndef SQLITE_OMIT_AUTOVACUUM
3854 if( pBt
->autoVacuum
){
3855 rc
= autoVacuumCommit(pBt
);
3856 if( rc
!=SQLITE_OK
){
3857 sqlite3BtreeLeave(p
);
3861 if( pBt
->bDoTruncate
){
3862 sqlite3PagerTruncateImage(pBt
->pPager
, pBt
->nPage
);
3865 rc
= sqlite3PagerCommitPhaseOne(pBt
->pPager
, zMaster
, 0);
3866 sqlite3BtreeLeave(p
);
3872 ** This function is called from both BtreeCommitPhaseTwo() and BtreeRollback()
3873 ** at the conclusion of a transaction.
3875 static void btreeEndTransaction(Btree
*p
){
3876 BtShared
*pBt
= p
->pBt
;
3877 sqlite3
*db
= p
->db
;
3878 assert( sqlite3BtreeHoldsMutex(p
) );
3880 #ifndef SQLITE_OMIT_AUTOVACUUM
3881 pBt
->bDoTruncate
= 0;
3883 if( p
->inTrans
>TRANS_NONE
&& db
->nVdbeRead
>1 ){
3884 /* If there are other active statements that belong to this database
3885 ** handle, downgrade to a read-only transaction. The other statements
3886 ** may still be reading from the database. */
3887 downgradeAllSharedCacheTableLocks(p
);
3888 p
->inTrans
= TRANS_READ
;
3890 /* If the handle had any kind of transaction open, decrement the
3891 ** transaction count of the shared btree. If the transaction count
3892 ** reaches 0, set the shared state to TRANS_NONE. The unlockBtreeIfUnused()
3893 ** call below will unlock the pager. */
3894 if( p
->inTrans
!=TRANS_NONE
){
3895 clearAllSharedCacheTableLocks(p
);
3896 pBt
->nTransaction
--;
3897 if( 0==pBt
->nTransaction
){
3898 pBt
->inTransaction
= TRANS_NONE
;
3902 /* Set the current transaction state to TRANS_NONE and unlock the
3903 ** pager if this call closed the only read or write transaction. */
3904 p
->inTrans
= TRANS_NONE
;
3905 unlockBtreeIfUnused(pBt
);
3912 ** Commit the transaction currently in progress.
3914 ** This routine implements the second phase of a 2-phase commit. The
3915 ** sqlite3BtreeCommitPhaseOne() routine does the first phase and should
3916 ** be invoked prior to calling this routine. The sqlite3BtreeCommitPhaseOne()
3917 ** routine did all the work of writing information out to disk and flushing the
3918 ** contents so that they are written onto the disk platter. All this
3919 ** routine has to do is delete or truncate or zero the header in the
3920 ** the rollback journal (which causes the transaction to commit) and
3923 ** Normally, if an error occurs while the pager layer is attempting to
3924 ** finalize the underlying journal file, this function returns an error and
3925 ** the upper layer will attempt a rollback. However, if the second argument
3926 ** is non-zero then this b-tree transaction is part of a multi-file
3927 ** transaction. In this case, the transaction has already been committed
3928 ** (by deleting a master journal file) and the caller will ignore this
3929 ** functions return code. So, even if an error occurs in the pager layer,
3930 ** reset the b-tree objects internal state to indicate that the write
3931 ** transaction has been closed. This is quite safe, as the pager will have
3932 ** transitioned to the error state.
3934 ** This will release the write lock on the database file. If there
3935 ** are no active cursors, it also releases the read lock.
3937 int sqlite3BtreeCommitPhaseTwo(Btree
*p
, int bCleanup
){
3939 if( p
->inTrans
==TRANS_NONE
) return SQLITE_OK
;
3940 sqlite3BtreeEnter(p
);
3943 /* If the handle has a write-transaction open, commit the shared-btrees
3944 ** transaction and set the shared state to TRANS_READ.
3946 if( p
->inTrans
==TRANS_WRITE
){
3948 BtShared
*pBt
= p
->pBt
;
3949 assert( pBt
->inTransaction
==TRANS_WRITE
);
3950 assert( pBt
->nTransaction
>0 );
3951 rc
= sqlite3PagerCommitPhaseTwo(pBt
->pPager
);
3952 if( rc
!=SQLITE_OK
&& bCleanup
==0 ){
3953 sqlite3BtreeLeave(p
);
3956 p
->iDataVersion
--; /* Compensate for pPager->iDataVersion++; */
3957 pBt
->inTransaction
= TRANS_READ
;
3958 btreeClearHasContent(pBt
);
3961 btreeEndTransaction(p
);
3962 sqlite3BtreeLeave(p
);
3967 ** Do both phases of a commit.
3969 int sqlite3BtreeCommit(Btree
*p
){
3971 sqlite3BtreeEnter(p
);
3972 rc
= sqlite3BtreeCommitPhaseOne(p
, 0);
3973 if( rc
==SQLITE_OK
){
3974 rc
= sqlite3BtreeCommitPhaseTwo(p
, 0);
3976 sqlite3BtreeLeave(p
);
3981 ** This routine sets the state to CURSOR_FAULT and the error
3982 ** code to errCode for every cursor on any BtShared that pBtree
3983 ** references. Or if the writeOnly flag is set to 1, then only
3984 ** trip write cursors and leave read cursors unchanged.
3986 ** Every cursor is a candidate to be tripped, including cursors
3987 ** that belong to other database connections that happen to be
3988 ** sharing the cache with pBtree.
3990 ** This routine gets called when a rollback occurs. If the writeOnly
3991 ** flag is true, then only write-cursors need be tripped - read-only
3992 ** cursors save their current positions so that they may continue
3993 ** following the rollback. Or, if writeOnly is false, all cursors are
3994 ** tripped. In general, writeOnly is false if the transaction being
3995 ** rolled back modified the database schema. In this case b-tree root
3996 ** pages may be moved or deleted from the database altogether, making
3997 ** it unsafe for read cursors to continue.
3999 ** If the writeOnly flag is true and an error is encountered while
4000 ** saving the current position of a read-only cursor, all cursors,
4001 ** including all read-cursors are tripped.
4003 ** SQLITE_OK is returned if successful, or if an error occurs while
4004 ** saving a cursor position, an SQLite error code.
4006 int sqlite3BtreeTripAllCursors(Btree
*pBtree
, int errCode
, int writeOnly
){
4010 assert( (writeOnly
==0 || writeOnly
==1) && BTCF_WriteFlag
==1 );
4012 sqlite3BtreeEnter(pBtree
);
4013 for(p
=pBtree
->pBt
->pCursor
; p
; p
=p
->pNext
){
4014 if( writeOnly
&& (p
->curFlags
& BTCF_WriteFlag
)==0 ){
4015 if( p
->eState
==CURSOR_VALID
|| p
->eState
==CURSOR_SKIPNEXT
){
4016 rc
= saveCursorPosition(p
);
4017 if( rc
!=SQLITE_OK
){
4018 (void)sqlite3BtreeTripAllCursors(pBtree
, rc
, 0);
4023 sqlite3BtreeClearCursor(p
);
4024 p
->eState
= CURSOR_FAULT
;
4025 p
->skipNext
= errCode
;
4027 btreeReleaseAllCursorPages(p
);
4029 sqlite3BtreeLeave(pBtree
);
4035 ** Rollback the transaction in progress.
4037 ** If tripCode is not SQLITE_OK then cursors will be invalidated (tripped).
4038 ** Only write cursors are tripped if writeOnly is true but all cursors are
4039 ** tripped if writeOnly is false. Any attempt to use
4040 ** a tripped cursor will result in an error.
4042 ** This will release the write lock on the database file. If there
4043 ** are no active cursors, it also releases the read lock.
4045 int sqlite3BtreeRollback(Btree
*p
, int tripCode
, int writeOnly
){
4047 BtShared
*pBt
= p
->pBt
;
4050 assert( writeOnly
==1 || writeOnly
==0 );
4051 assert( tripCode
==SQLITE_ABORT_ROLLBACK
|| tripCode
==SQLITE_OK
);
4052 sqlite3BtreeEnter(p
);
4053 if( tripCode
==SQLITE_OK
){
4054 rc
= tripCode
= saveAllCursors(pBt
, 0, 0);
4055 if( rc
) writeOnly
= 0;
4060 int rc2
= sqlite3BtreeTripAllCursors(p
, tripCode
, writeOnly
);
4061 assert( rc
==SQLITE_OK
|| (writeOnly
==0 && rc2
==SQLITE_OK
) );
4062 if( rc2
!=SQLITE_OK
) rc
= rc2
;
4066 if( p
->inTrans
==TRANS_WRITE
){
4069 assert( TRANS_WRITE
==pBt
->inTransaction
);
4070 rc2
= sqlite3PagerRollback(pBt
->pPager
);
4071 if( rc2
!=SQLITE_OK
){
4075 /* The rollback may have destroyed the pPage1->aData value. So
4076 ** call btreeGetPage() on page 1 again to make
4077 ** sure pPage1->aData is set correctly. */
4078 if( btreeGetPage(pBt
, 1, &pPage1
, 0)==SQLITE_OK
){
4079 int nPage
= get4byte(28+(u8
*)pPage1
->aData
);
4080 testcase( nPage
==0 );
4081 if( nPage
==0 ) sqlite3PagerPagecount(pBt
->pPager
, &nPage
);
4082 testcase( pBt
->nPage
!=nPage
);
4084 releasePageOne(pPage1
);
4086 assert( countValidCursors(pBt
, 1)==0 );
4087 pBt
->inTransaction
= TRANS_READ
;
4088 btreeClearHasContent(pBt
);
4091 btreeEndTransaction(p
);
4092 sqlite3BtreeLeave(p
);
4097 ** Start a statement subtransaction. The subtransaction can be rolled
4098 ** back independently of the main transaction. You must start a transaction
4099 ** before starting a subtransaction. The subtransaction is ended automatically
4100 ** if the main transaction commits or rolls back.
4102 ** Statement subtransactions are used around individual SQL statements
4103 ** that are contained within a BEGIN...COMMIT block. If a constraint
4104 ** error occurs within the statement, the effect of that one statement
4105 ** can be rolled back without having to rollback the entire transaction.
4107 ** A statement sub-transaction is implemented as an anonymous savepoint. The
4108 ** value passed as the second parameter is the total number of savepoints,
4109 ** including the new anonymous savepoint, open on the B-Tree. i.e. if there
4110 ** are no active savepoints and no other statement-transactions open,
4111 ** iStatement is 1. This anonymous savepoint can be released or rolled back
4112 ** using the sqlite3BtreeSavepoint() function.
4114 int sqlite3BtreeBeginStmt(Btree
*p
, int iStatement
){
4116 BtShared
*pBt
= p
->pBt
;
4117 sqlite3BtreeEnter(p
);
4118 assert( p
->inTrans
==TRANS_WRITE
);
4119 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
4120 assert( iStatement
>0 );
4121 assert( iStatement
>p
->db
->nSavepoint
);
4122 assert( pBt
->inTransaction
==TRANS_WRITE
);
4123 /* At the pager level, a statement transaction is a savepoint with
4124 ** an index greater than all savepoints created explicitly using
4125 ** SQL statements. It is illegal to open, release or rollback any
4126 ** such savepoints while the statement transaction savepoint is active.
4128 rc
= sqlite3PagerOpenSavepoint(pBt
->pPager
, iStatement
);
4129 sqlite3BtreeLeave(p
);
4134 ** The second argument to this function, op, is always SAVEPOINT_ROLLBACK
4135 ** or SAVEPOINT_RELEASE. This function either releases or rolls back the
4136 ** savepoint identified by parameter iSavepoint, depending on the value
4139 ** Normally, iSavepoint is greater than or equal to zero. However, if op is
4140 ** SAVEPOINT_ROLLBACK, then iSavepoint may also be -1. In this case the
4141 ** contents of the entire transaction are rolled back. This is different
4142 ** from a normal transaction rollback, as no locks are released and the
4143 ** transaction remains open.
4145 int sqlite3BtreeSavepoint(Btree
*p
, int op
, int iSavepoint
){
4147 if( p
&& p
->inTrans
==TRANS_WRITE
){
4148 BtShared
*pBt
= p
->pBt
;
4149 assert( op
==SAVEPOINT_RELEASE
|| op
==SAVEPOINT_ROLLBACK
);
4150 assert( iSavepoint
>=0 || (iSavepoint
==-1 && op
==SAVEPOINT_ROLLBACK
) );
4151 sqlite3BtreeEnter(p
);
4152 if( op
==SAVEPOINT_ROLLBACK
){
4153 rc
= saveAllCursors(pBt
, 0, 0);
4155 if( rc
==SQLITE_OK
){
4156 rc
= sqlite3PagerSavepoint(pBt
->pPager
, op
, iSavepoint
);
4158 if( rc
==SQLITE_OK
){
4159 if( iSavepoint
<0 && (pBt
->btsFlags
& BTS_INITIALLY_EMPTY
)!=0 ){
4162 rc
= newDatabase(pBt
);
4163 pBt
->nPage
= get4byte(28 + pBt
->pPage1
->aData
);
4165 /* The database size was written into the offset 28 of the header
4166 ** when the transaction started, so we know that the value at offset
4167 ** 28 is nonzero. */
4168 assert( pBt
->nPage
>0 );
4170 sqlite3BtreeLeave(p
);
4176 ** Create a new cursor for the BTree whose root is on the page
4177 ** iTable. If a read-only cursor is requested, it is assumed that
4178 ** the caller already has at least a read-only transaction open
4179 ** on the database already. If a write-cursor is requested, then
4180 ** the caller is assumed to have an open write transaction.
4182 ** If the BTREE_WRCSR bit of wrFlag is clear, then the cursor can only
4183 ** be used for reading. If the BTREE_WRCSR bit is set, then the cursor
4184 ** can be used for reading or for writing if other conditions for writing
4185 ** are also met. These are the conditions that must be met in order
4186 ** for writing to be allowed:
4188 ** 1: The cursor must have been opened with wrFlag containing BTREE_WRCSR
4190 ** 2: Other database connections that share the same pager cache
4191 ** but which are not in the READ_UNCOMMITTED state may not have
4192 ** cursors open with wrFlag==0 on the same table. Otherwise
4193 ** the changes made by this write cursor would be visible to
4194 ** the read cursors in the other database connection.
4196 ** 3: The database must be writable (not on read-only media)
4198 ** 4: There must be an active transaction.
4200 ** The BTREE_FORDELETE bit of wrFlag may optionally be set if BTREE_WRCSR
4201 ** is set. If FORDELETE is set, that is a hint to the implementation that
4202 ** this cursor will only be used to seek to and delete entries of an index
4203 ** as part of a larger DELETE statement. The FORDELETE hint is not used by
4204 ** this implementation. But in a hypothetical alternative storage engine
4205 ** in which index entries are automatically deleted when corresponding table
4206 ** rows are deleted, the FORDELETE flag is a hint that all SEEK and DELETE
4207 ** operations on this cursor can be no-ops and all READ operations can
4208 ** return a null row (2-bytes: 0x01 0x00).
4210 ** No checking is done to make sure that page iTable really is the
4211 ** root page of a b-tree. If it is not, then the cursor acquired
4212 ** will not work correctly.
4214 ** It is assumed that the sqlite3BtreeCursorZero() has been called
4215 ** on pCur to initialize the memory space prior to invoking this routine.
4217 static int btreeCursor(
4218 Btree
*p
, /* The btree */
4219 int iTable
, /* Root page of table to open */
4220 int wrFlag
, /* 1 to write. 0 read-only */
4221 struct KeyInfo
*pKeyInfo
, /* First arg to comparison function */
4222 BtCursor
*pCur
/* Space for new cursor */
4224 BtShared
*pBt
= p
->pBt
; /* Shared b-tree handle */
4225 BtCursor
*pX
; /* Looping over other all cursors */
4227 assert( sqlite3BtreeHoldsMutex(p
) );
4229 || wrFlag
==BTREE_WRCSR
4230 || wrFlag
==(BTREE_WRCSR
|BTREE_FORDELETE
)
4233 /* The following assert statements verify that if this is a sharable
4234 ** b-tree database, the connection is holding the required table locks,
4235 ** and that no other connection has any open cursor that conflicts with
4237 assert( hasSharedCacheTableLock(p
, iTable
, pKeyInfo
!=0, (wrFlag
?2:1)) );
4238 assert( wrFlag
==0 || !hasReadConflicts(p
, iTable
) );
4240 /* Assert that the caller has opened the required transaction. */
4241 assert( p
->inTrans
>TRANS_NONE
);
4242 assert( wrFlag
==0 || p
->inTrans
==TRANS_WRITE
);
4243 assert( pBt
->pPage1
&& pBt
->pPage1
->aData
);
4244 assert( wrFlag
==0 || (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
4247 allocateTempSpace(pBt
);
4248 if( pBt
->pTmpSpace
==0 ) return SQLITE_NOMEM_BKPT
;
4250 if( iTable
==1 && btreePagecount(pBt
)==0 ){
4251 assert( wrFlag
==0 );
4255 /* Now that no other errors can occur, finish filling in the BtCursor
4256 ** variables and link the cursor into the BtShared list. */
4257 pCur
->pgnoRoot
= (Pgno
)iTable
;
4259 pCur
->pKeyInfo
= pKeyInfo
;
4262 pCur
->curFlags
= wrFlag
? BTCF_WriteFlag
: 0;
4263 pCur
->curPagerFlags
= wrFlag
? 0 : PAGER_GET_READONLY
;
4264 /* If there are two or more cursors on the same btree, then all such
4265 ** cursors *must* have the BTCF_Multiple flag set. */
4266 for(pX
=pBt
->pCursor
; pX
; pX
=pX
->pNext
){
4267 if( pX
->pgnoRoot
==(Pgno
)iTable
){
4268 pX
->curFlags
|= BTCF_Multiple
;
4269 pCur
->curFlags
|= BTCF_Multiple
;
4272 pCur
->pNext
= pBt
->pCursor
;
4273 pBt
->pCursor
= pCur
;
4274 pCur
->eState
= CURSOR_INVALID
;
4277 int sqlite3BtreeCursor(
4278 Btree
*p
, /* The btree */
4279 int iTable
, /* Root page of table to open */
4280 int wrFlag
, /* 1 to write. 0 read-only */
4281 struct KeyInfo
*pKeyInfo
, /* First arg to xCompare() */
4282 BtCursor
*pCur
/* Write new cursor here */
4286 rc
= SQLITE_CORRUPT_BKPT
;
4288 sqlite3BtreeEnter(p
);
4289 rc
= btreeCursor(p
, iTable
, wrFlag
, pKeyInfo
, pCur
);
4290 sqlite3BtreeLeave(p
);
4296 ** Return the size of a BtCursor object in bytes.
4298 ** This interfaces is needed so that users of cursors can preallocate
4299 ** sufficient storage to hold a cursor. The BtCursor object is opaque
4300 ** to users so they cannot do the sizeof() themselves - they must call
4303 int sqlite3BtreeCursorSize(void){
4304 return ROUND8(sizeof(BtCursor
));
4308 ** Initialize memory that will be converted into a BtCursor object.
4310 ** The simple approach here would be to memset() the entire object
4311 ** to zero. But it turns out that the apPage[] and aiIdx[] arrays
4312 ** do not need to be zeroed and they are large, so we can save a lot
4313 ** of run-time by skipping the initialization of those elements.
4315 void sqlite3BtreeCursorZero(BtCursor
*p
){
4316 memset(p
, 0, offsetof(BtCursor
, iPage
));
4320 ** Close a cursor. The read lock on the database file is released
4321 ** when the last cursor is closed.
4323 int sqlite3BtreeCloseCursor(BtCursor
*pCur
){
4324 Btree
*pBtree
= pCur
->pBtree
;
4326 BtShared
*pBt
= pCur
->pBt
;
4327 sqlite3BtreeEnter(pBtree
);
4328 assert( pBt
->pCursor
!=0 );
4329 if( pBt
->pCursor
==pCur
){
4330 pBt
->pCursor
= pCur
->pNext
;
4332 BtCursor
*pPrev
= pBt
->pCursor
;
4334 if( pPrev
->pNext
==pCur
){
4335 pPrev
->pNext
= pCur
->pNext
;
4338 pPrev
= pPrev
->pNext
;
4339 }while( ALWAYS(pPrev
) );
4341 btreeReleaseAllCursorPages(pCur
);
4342 unlockBtreeIfUnused(pBt
);
4343 sqlite3_free(pCur
->aOverflow
);
4344 sqlite3_free(pCur
->pKey
);
4345 sqlite3BtreeLeave(pBtree
);
4351 ** Make sure the BtCursor* given in the argument has a valid
4352 ** BtCursor.info structure. If it is not already valid, call
4353 ** btreeParseCell() to fill it in.
4355 ** BtCursor.info is a cache of the information in the current cell.
4356 ** Using this cache reduces the number of calls to btreeParseCell().
4359 static void assertCellInfo(BtCursor
*pCur
){
4361 memset(&info
, 0, sizeof(info
));
4362 btreeParseCell(pCur
->pPage
, pCur
->ix
, &info
);
4363 assert( CORRUPT_DB
|| memcmp(&info
, &pCur
->info
, sizeof(info
))==0 );
4366 #define assertCellInfo(x)
4368 static SQLITE_NOINLINE
void getCellInfo(BtCursor
*pCur
){
4369 if( pCur
->info
.nSize
==0 ){
4370 pCur
->curFlags
|= BTCF_ValidNKey
;
4371 btreeParseCell(pCur
->pPage
,pCur
->ix
,&pCur
->info
);
4373 assertCellInfo(pCur
);
4377 #ifndef NDEBUG /* The next routine used only within assert() statements */
4379 ** Return true if the given BtCursor is valid. A valid cursor is one
4380 ** that is currently pointing to a row in a (non-empty) table.
4381 ** This is a verification routine is used only within assert() statements.
4383 int sqlite3BtreeCursorIsValid(BtCursor
*pCur
){
4384 return pCur
&& pCur
->eState
==CURSOR_VALID
;
4387 int sqlite3BtreeCursorIsValidNN(BtCursor
*pCur
){
4389 return pCur
->eState
==CURSOR_VALID
;
4393 ** Return the value of the integer key or "rowid" for a table btree.
4394 ** This routine is only valid for a cursor that is pointing into a
4395 ** ordinary table btree. If the cursor points to an index btree or
4396 ** is invalid, the result of this routine is undefined.
4398 i64
sqlite3BtreeIntegerKey(BtCursor
*pCur
){
4399 assert( cursorHoldsMutex(pCur
) );
4400 assert( pCur
->eState
==CURSOR_VALID
);
4401 assert( pCur
->curIntKey
);
4403 return pCur
->info
.nKey
;
4407 ** Return the number of bytes of payload for the entry that pCur is
4408 ** currently pointing to. For table btrees, this will be the amount
4409 ** of data. For index btrees, this will be the size of the key.
4411 ** The caller must guarantee that the cursor is pointing to a non-NULL
4412 ** valid entry. In other words, the calling procedure must guarantee
4413 ** that the cursor has Cursor.eState==CURSOR_VALID.
4415 u32
sqlite3BtreePayloadSize(BtCursor
*pCur
){
4416 assert( cursorHoldsMutex(pCur
) );
4417 assert( pCur
->eState
==CURSOR_VALID
);
4419 return pCur
->info
.nPayload
;
4423 ** Given the page number of an overflow page in the database (parameter
4424 ** ovfl), this function finds the page number of the next page in the
4425 ** linked list of overflow pages. If possible, it uses the auto-vacuum
4426 ** pointer-map data instead of reading the content of page ovfl to do so.
4428 ** If an error occurs an SQLite error code is returned. Otherwise:
4430 ** The page number of the next overflow page in the linked list is
4431 ** written to *pPgnoNext. If page ovfl is the last page in its linked
4432 ** list, *pPgnoNext is set to zero.
4434 ** If ppPage is not NULL, and a reference to the MemPage object corresponding
4435 ** to page number pOvfl was obtained, then *ppPage is set to point to that
4436 ** reference. It is the responsibility of the caller to call releasePage()
4437 ** on *ppPage to free the reference. In no reference was obtained (because
4438 ** the pointer-map was used to obtain the value for *pPgnoNext), then
4439 ** *ppPage is set to zero.
4441 static int getOverflowPage(
4442 BtShared
*pBt
, /* The database file */
4443 Pgno ovfl
, /* Current overflow page number */
4444 MemPage
**ppPage
, /* OUT: MemPage handle (may be NULL) */
4445 Pgno
*pPgnoNext
/* OUT: Next overflow page number */
4451 assert( sqlite3_mutex_held(pBt
->mutex
) );
4454 #ifndef SQLITE_OMIT_AUTOVACUUM
4455 /* Try to find the next page in the overflow list using the
4456 ** autovacuum pointer-map pages. Guess that the next page in
4457 ** the overflow list is page number (ovfl+1). If that guess turns
4458 ** out to be wrong, fall back to loading the data of page
4459 ** number ovfl to determine the next page number.
4461 if( pBt
->autoVacuum
){
4463 Pgno iGuess
= ovfl
+1;
4466 while( PTRMAP_ISPAGE(pBt
, iGuess
) || iGuess
==PENDING_BYTE_PAGE(pBt
) ){
4470 if( iGuess
<=btreePagecount(pBt
) ){
4471 rc
= ptrmapGet(pBt
, iGuess
, &eType
, &pgno
);
4472 if( rc
==SQLITE_OK
&& eType
==PTRMAP_OVERFLOW2
&& pgno
==ovfl
){
4480 assert( next
==0 || rc
==SQLITE_DONE
);
4481 if( rc
==SQLITE_OK
){
4482 rc
= btreeGetPage(pBt
, ovfl
, &pPage
, (ppPage
==0) ? PAGER_GET_READONLY
: 0);
4483 assert( rc
==SQLITE_OK
|| pPage
==0 );
4484 if( rc
==SQLITE_OK
){
4485 next
= get4byte(pPage
->aData
);
4495 return (rc
==SQLITE_DONE
? SQLITE_OK
: rc
);
4499 ** Copy data from a buffer to a page, or from a page to a buffer.
4501 ** pPayload is a pointer to data stored on database page pDbPage.
4502 ** If argument eOp is false, then nByte bytes of data are copied
4503 ** from pPayload to the buffer pointed at by pBuf. If eOp is true,
4504 ** then sqlite3PagerWrite() is called on pDbPage and nByte bytes
4505 ** of data are copied from the buffer pBuf to pPayload.
4507 ** SQLITE_OK is returned on success, otherwise an error code.
4509 static int copyPayload(
4510 void *pPayload
, /* Pointer to page data */
4511 void *pBuf
, /* Pointer to buffer */
4512 int nByte
, /* Number of bytes to copy */
4513 int eOp
, /* 0 -> copy from page, 1 -> copy to page */
4514 DbPage
*pDbPage
/* Page containing pPayload */
4517 /* Copy data from buffer to page (a write operation) */
4518 int rc
= sqlite3PagerWrite(pDbPage
);
4519 if( rc
!=SQLITE_OK
){
4522 memcpy(pPayload
, pBuf
, nByte
);
4524 /* Copy data from page to buffer (a read operation) */
4525 memcpy(pBuf
, pPayload
, nByte
);
4531 ** This function is used to read or overwrite payload information
4532 ** for the entry that the pCur cursor is pointing to. The eOp
4533 ** argument is interpreted as follows:
4535 ** 0: The operation is a read. Populate the overflow cache.
4536 ** 1: The operation is a write. Populate the overflow cache.
4538 ** A total of "amt" bytes are read or written beginning at "offset".
4539 ** Data is read to or from the buffer pBuf.
4541 ** The content being read or written might appear on the main page
4542 ** or be scattered out on multiple overflow pages.
4544 ** If the current cursor entry uses one or more overflow pages
4545 ** this function may allocate space for and lazily populate
4546 ** the overflow page-list cache array (BtCursor.aOverflow).
4547 ** Subsequent calls use this cache to make seeking to the supplied offset
4550 ** Once an overflow page-list cache has been allocated, it must be
4551 ** invalidated if some other cursor writes to the same table, or if
4552 ** the cursor is moved to a different row. Additionally, in auto-vacuum
4553 ** mode, the following events may invalidate an overflow page-list cache.
4555 ** * An incremental vacuum,
4556 ** * A commit in auto_vacuum="full" mode,
4557 ** * Creating a table (may require moving an overflow page).
4559 static int accessPayload(
4560 BtCursor
*pCur
, /* Cursor pointing to entry to read from */
4561 u32 offset
, /* Begin reading this far into payload */
4562 u32 amt
, /* Read this many bytes */
4563 unsigned char *pBuf
, /* Write the bytes into this buffer */
4564 int eOp
/* zero to read. non-zero to write. */
4566 unsigned char *aPayload
;
4569 MemPage
*pPage
= pCur
->pPage
; /* Btree page of current entry */
4570 BtShared
*pBt
= pCur
->pBt
; /* Btree this cursor belongs to */
4571 #ifdef SQLITE_DIRECT_OVERFLOW_READ
4572 unsigned char * const pBufStart
= pBuf
; /* Start of original out buffer */
4576 assert( eOp
==0 || eOp
==1 );
4577 assert( pCur
->eState
==CURSOR_VALID
);
4578 assert( pCur
->ix
<pPage
->nCell
);
4579 assert( cursorHoldsMutex(pCur
) );
4582 aPayload
= pCur
->info
.pPayload
;
4583 assert( offset
+amt
<= pCur
->info
.nPayload
);
4585 assert( aPayload
> pPage
->aData
);
4586 if( (uptr
)(aPayload
- pPage
->aData
) > (pBt
->usableSize
- pCur
->info
.nLocal
) ){
4587 /* Trying to read or write past the end of the data is an error. The
4588 ** conditional above is really:
4589 ** &aPayload[pCur->info.nLocal] > &pPage->aData[pBt->usableSize]
4590 ** but is recast into its current form to avoid integer overflow problems
4592 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
4595 /* Check if data must be read/written to/from the btree page itself. */
4596 if( offset
<pCur
->info
.nLocal
){
4598 if( a
+offset
>pCur
->info
.nLocal
){
4599 a
= pCur
->info
.nLocal
- offset
;
4601 rc
= copyPayload(&aPayload
[offset
], pBuf
, a
, eOp
, pPage
->pDbPage
);
4606 offset
-= pCur
->info
.nLocal
;
4610 if( rc
==SQLITE_OK
&& amt
>0 ){
4611 const u32 ovflSize
= pBt
->usableSize
- 4; /* Bytes content per ovfl page */
4614 nextPage
= get4byte(&aPayload
[pCur
->info
.nLocal
]);
4616 /* If the BtCursor.aOverflow[] has not been allocated, allocate it now.
4618 ** The aOverflow[] array is sized at one entry for each overflow page
4619 ** in the overflow chain. The page number of the first overflow page is
4620 ** stored in aOverflow[0], etc. A value of 0 in the aOverflow[] array
4621 ** means "not yet known" (the cache is lazily populated).
4623 if( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 ){
4624 int nOvfl
= (pCur
->info
.nPayload
-pCur
->info
.nLocal
+ovflSize
-1)/ovflSize
;
4625 if( nOvfl
>pCur
->nOvflAlloc
){
4626 Pgno
*aNew
= (Pgno
*)sqlite3Realloc(
4627 pCur
->aOverflow
, nOvfl
*2*sizeof(Pgno
)
4630 return SQLITE_NOMEM_BKPT
;
4632 pCur
->nOvflAlloc
= nOvfl
*2;
4633 pCur
->aOverflow
= aNew
;
4636 memset(pCur
->aOverflow
, 0, nOvfl
*sizeof(Pgno
));
4637 pCur
->curFlags
|= BTCF_ValidOvfl
;
4639 /* If the overflow page-list cache has been allocated and the
4640 ** entry for the first required overflow page is valid, skip
4643 if( pCur
->aOverflow
[offset
/ovflSize
] ){
4644 iIdx
= (offset
/ovflSize
);
4645 nextPage
= pCur
->aOverflow
[iIdx
];
4646 offset
= (offset
%ovflSize
);
4650 assert( rc
==SQLITE_OK
&& amt
>0 );
4652 /* If required, populate the overflow page-list cache. */
4653 assert( pCur
->aOverflow
[iIdx
]==0
4654 || pCur
->aOverflow
[iIdx
]==nextPage
4656 pCur
->aOverflow
[iIdx
] = nextPage
;
4658 if( offset
>=ovflSize
){
4659 /* The only reason to read this page is to obtain the page
4660 ** number for the next page in the overflow chain. The page
4661 ** data is not required. So first try to lookup the overflow
4662 ** page-list cache, if any, then fall back to the getOverflowPage()
4665 assert( pCur
->curFlags
& BTCF_ValidOvfl
);
4666 assert( pCur
->pBtree
->db
==pBt
->db
);
4667 if( pCur
->aOverflow
[iIdx
+1] ){
4668 nextPage
= pCur
->aOverflow
[iIdx
+1];
4670 rc
= getOverflowPage(pBt
, nextPage
, 0, &nextPage
);
4674 /* Need to read this page properly. It contains some of the
4675 ** range of data that is being read (eOp==0) or written (eOp!=0).
4677 #ifdef SQLITE_DIRECT_OVERFLOW_READ
4678 sqlite3_file
*fd
; /* File from which to do direct overflow read */
4681 if( a
+ offset
> ovflSize
){
4682 a
= ovflSize
- offset
;
4685 #ifdef SQLITE_DIRECT_OVERFLOW_READ
4686 /* If all the following are true:
4688 ** 1) this is a read operation, and
4689 ** 2) data is required from the start of this overflow page, and
4690 ** 3) there is no open write-transaction, and
4691 ** 4) the database is file-backed, and
4692 ** 5) the page is not in the WAL file
4693 ** 6) at least 4 bytes have already been read into the output buffer
4695 ** then data can be read directly from the database file into the
4696 ** output buffer, bypassing the page-cache altogether. This speeds
4697 ** up loading large records that span many overflow pages.
4699 if( eOp
==0 /* (1) */
4700 && offset
==0 /* (2) */
4701 && pBt
->inTransaction
==TRANS_READ
/* (3) */
4702 && (fd
= sqlite3PagerFile(pBt
->pPager
))->pMethods
/* (4) */
4703 && 0==sqlite3PagerUseWal(pBt
->pPager
, nextPage
) /* (5) */
4704 && &pBuf
[-4]>=pBufStart
/* (6) */
4707 u8
*aWrite
= &pBuf
[-4];
4708 assert( aWrite
>=pBufStart
); /* due to (6) */
4709 memcpy(aSave
, aWrite
, 4);
4710 rc
= sqlite3OsRead(fd
, aWrite
, a
+4, (i64
)pBt
->pageSize
*(nextPage
-1));
4711 nextPage
= get4byte(aWrite
);
4712 memcpy(aWrite
, aSave
, 4);
4718 rc
= sqlite3PagerGet(pBt
->pPager
, nextPage
, &pDbPage
,
4719 (eOp
==0 ? PAGER_GET_READONLY
: 0)
4721 if( rc
==SQLITE_OK
){
4722 aPayload
= sqlite3PagerGetData(pDbPage
);
4723 nextPage
= get4byte(aPayload
);
4724 rc
= copyPayload(&aPayload
[offset
+4], pBuf
, a
, eOp
, pDbPage
);
4725 sqlite3PagerUnref(pDbPage
);
4730 if( amt
==0 ) return rc
;
4738 if( rc
==SQLITE_OK
&& amt
>0 ){
4739 /* Overflow chain ends prematurely */
4740 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
4746 ** Read part of the payload for the row at which that cursor pCur is currently
4747 ** pointing. "amt" bytes will be transferred into pBuf[]. The transfer
4748 ** begins at "offset".
4750 ** pCur can be pointing to either a table or an index b-tree.
4751 ** If pointing to a table btree, then the content section is read. If
4752 ** pCur is pointing to an index b-tree then the key section is read.
4754 ** For sqlite3BtreePayload(), the caller must ensure that pCur is pointing
4755 ** to a valid row in the table. For sqlite3BtreePayloadChecked(), the
4756 ** cursor might be invalid or might need to be restored before being read.
4758 ** Return SQLITE_OK on success or an error code if anything goes
4759 ** wrong. An error is returned if "offset+amt" is larger than
4760 ** the available payload.
4762 int sqlite3BtreePayload(BtCursor
*pCur
, u32 offset
, u32 amt
, void *pBuf
){
4763 assert( cursorHoldsMutex(pCur
) );
4764 assert( pCur
->eState
==CURSOR_VALID
);
4765 assert( pCur
->iPage
>=0 && pCur
->pPage
);
4766 assert( pCur
->ix
<pCur
->pPage
->nCell
);
4767 return accessPayload(pCur
, offset
, amt
, (unsigned char*)pBuf
, 0);
4771 ** This variant of sqlite3BtreePayload() works even if the cursor has not
4772 ** in the CURSOR_VALID state. It is only used by the sqlite3_blob_read()
4775 #ifndef SQLITE_OMIT_INCRBLOB
4776 static SQLITE_NOINLINE
int accessPayloadChecked(
4783 if ( pCur
->eState
==CURSOR_INVALID
){
4784 return SQLITE_ABORT
;
4786 assert( cursorOwnsBtShared(pCur
) );
4787 rc
= btreeRestoreCursorPosition(pCur
);
4788 return rc
? rc
: accessPayload(pCur
, offset
, amt
, pBuf
, 0);
4790 int sqlite3BtreePayloadChecked(BtCursor
*pCur
, u32 offset
, u32 amt
, void *pBuf
){
4791 if( pCur
->eState
==CURSOR_VALID
){
4792 assert( cursorOwnsBtShared(pCur
) );
4793 return accessPayload(pCur
, offset
, amt
, pBuf
, 0);
4795 return accessPayloadChecked(pCur
, offset
, amt
, pBuf
);
4798 #endif /* SQLITE_OMIT_INCRBLOB */
4801 ** Return a pointer to payload information from the entry that the
4802 ** pCur cursor is pointing to. The pointer is to the beginning of
4803 ** the key if index btrees (pPage->intKey==0) and is the data for
4804 ** table btrees (pPage->intKey==1). The number of bytes of available
4805 ** key/data is written into *pAmt. If *pAmt==0, then the value
4806 ** returned will not be a valid pointer.
4808 ** This routine is an optimization. It is common for the entire key
4809 ** and data to fit on the local page and for there to be no overflow
4810 ** pages. When that is so, this routine can be used to access the
4811 ** key and data without making a copy. If the key and/or data spills
4812 ** onto overflow pages, then accessPayload() must be used to reassemble
4813 ** the key/data and copy it into a preallocated buffer.
4815 ** The pointer returned by this routine looks directly into the cached
4816 ** page of the database. The data might change or move the next time
4817 ** any btree routine is called.
4819 static const void *fetchPayload(
4820 BtCursor
*pCur
, /* Cursor pointing to entry to read from */
4821 u32
*pAmt
/* Write the number of available bytes here */
4824 assert( pCur
!=0 && pCur
->iPage
>=0 && pCur
->pPage
);
4825 assert( pCur
->eState
==CURSOR_VALID
);
4826 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
4827 assert( cursorOwnsBtShared(pCur
) );
4828 assert( pCur
->ix
<pCur
->pPage
->nCell
);
4829 assert( pCur
->info
.nSize
>0 );
4830 assert( pCur
->info
.pPayload
>pCur
->pPage
->aData
|| CORRUPT_DB
);
4831 assert( pCur
->info
.pPayload
<pCur
->pPage
->aDataEnd
||CORRUPT_DB
);
4832 amt
= (int)(pCur
->pPage
->aDataEnd
- pCur
->info
.pPayload
);
4833 if( pCur
->info
.nLocal
<amt
) amt
= pCur
->info
.nLocal
;
4835 return (void*)pCur
->info
.pPayload
;
4840 ** For the entry that cursor pCur is point to, return as
4841 ** many bytes of the key or data as are available on the local
4842 ** b-tree page. Write the number of available bytes into *pAmt.
4844 ** The pointer returned is ephemeral. The key/data may move
4845 ** or be destroyed on the next call to any Btree routine,
4846 ** including calls from other threads against the same cache.
4847 ** Hence, a mutex on the BtShared should be held prior to calling
4850 ** These routines is used to get quick access to key and data
4851 ** in the common case where no overflow pages are used.
4853 const void *sqlite3BtreePayloadFetch(BtCursor
*pCur
, u32
*pAmt
){
4854 return fetchPayload(pCur
, pAmt
);
4859 ** Move the cursor down to a new child page. The newPgno argument is the
4860 ** page number of the child page to move to.
4862 ** This function returns SQLITE_CORRUPT if the page-header flags field of
4863 ** the new child page does not match the flags field of the parent (i.e.
4864 ** if an intkey page appears to be the parent of a non-intkey page, or
4867 static int moveToChild(BtCursor
*pCur
, u32 newPgno
){
4868 BtShared
*pBt
= pCur
->pBt
;
4870 assert( cursorOwnsBtShared(pCur
) );
4871 assert( pCur
->eState
==CURSOR_VALID
);
4872 assert( pCur
->iPage
<BTCURSOR_MAX_DEPTH
);
4873 assert( pCur
->iPage
>=0 );
4874 if( pCur
->iPage
>=(BTCURSOR_MAX_DEPTH
-1) ){
4875 return SQLITE_CORRUPT_BKPT
;
4877 pCur
->info
.nSize
= 0;
4878 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
4879 pCur
->aiIdx
[pCur
->iPage
] = pCur
->ix
;
4880 pCur
->apPage
[pCur
->iPage
] = pCur
->pPage
;
4883 return getAndInitPage(pBt
, newPgno
, &pCur
->pPage
, pCur
, pCur
->curPagerFlags
);
4888 ** Page pParent is an internal (non-leaf) tree page. This function
4889 ** asserts that page number iChild is the left-child if the iIdx'th
4890 ** cell in page pParent. Or, if iIdx is equal to the total number of
4891 ** cells in pParent, that page number iChild is the right-child of
4894 static void assertParentIndex(MemPage
*pParent
, int iIdx
, Pgno iChild
){
4895 if( CORRUPT_DB
) return; /* The conditions tested below might not be true
4896 ** in a corrupt database */
4897 assert( iIdx
<=pParent
->nCell
);
4898 if( iIdx
==pParent
->nCell
){
4899 assert( get4byte(&pParent
->aData
[pParent
->hdrOffset
+8])==iChild
);
4901 assert( get4byte(findCell(pParent
, iIdx
))==iChild
);
4905 # define assertParentIndex(x,y,z)
4909 ** Move the cursor up to the parent page.
4911 ** pCur->idx is set to the cell index that contains the pointer
4912 ** to the page we are coming from. If we are coming from the
4913 ** right-most child page then pCur->idx is set to one more than
4914 ** the largest cell index.
4916 static void moveToParent(BtCursor
*pCur
){
4918 assert( cursorOwnsBtShared(pCur
) );
4919 assert( pCur
->eState
==CURSOR_VALID
);
4920 assert( pCur
->iPage
>0 );
4921 assert( pCur
->pPage
);
4923 pCur
->apPage
[pCur
->iPage
-1],
4924 pCur
->aiIdx
[pCur
->iPage
-1],
4927 testcase( pCur
->aiIdx
[pCur
->iPage
-1] > pCur
->apPage
[pCur
->iPage
-1]->nCell
);
4928 pCur
->info
.nSize
= 0;
4929 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
4930 pCur
->ix
= pCur
->aiIdx
[pCur
->iPage
-1];
4931 pLeaf
= pCur
->pPage
;
4932 pCur
->pPage
= pCur
->apPage
[--pCur
->iPage
];
4933 releasePageNotNull(pLeaf
);
4937 ** Move the cursor to point to the root page of its b-tree structure.
4939 ** If the table has a virtual root page, then the cursor is moved to point
4940 ** to the virtual root page instead of the actual root page. A table has a
4941 ** virtual root page when the actual root page contains no cells and a
4942 ** single child page. This can only happen with the table rooted at page 1.
4944 ** If the b-tree structure is empty, the cursor state is set to
4945 ** CURSOR_INVALID and this routine returns SQLITE_EMPTY. Otherwise,
4946 ** the cursor is set to point to the first cell located on the root
4947 ** (or virtual root) page and the cursor state is set to CURSOR_VALID.
4949 ** If this function returns successfully, it may be assumed that the
4950 ** page-header flags indicate that the [virtual] root-page is the expected
4951 ** kind of b-tree page (i.e. if when opening the cursor the caller did not
4952 ** specify a KeyInfo structure the flags byte is set to 0x05 or 0x0D,
4953 ** indicating a table b-tree, or if the caller did specify a KeyInfo
4954 ** structure the flags byte is set to 0x02 or 0x0A, indicating an index
4957 static int moveToRoot(BtCursor
*pCur
){
4961 assert( cursorOwnsBtShared(pCur
) );
4962 assert( CURSOR_INVALID
< CURSOR_REQUIRESEEK
);
4963 assert( CURSOR_VALID
< CURSOR_REQUIRESEEK
);
4964 assert( CURSOR_FAULT
> CURSOR_REQUIRESEEK
);
4965 assert( pCur
->eState
< CURSOR_REQUIRESEEK
|| pCur
->iPage
<0 );
4966 assert( pCur
->pgnoRoot
>0 || pCur
->iPage
<0 );
4968 if( pCur
->iPage
>=0 ){
4970 releasePageNotNull(pCur
->pPage
);
4971 while( --pCur
->iPage
){
4972 releasePageNotNull(pCur
->apPage
[pCur
->iPage
]);
4974 pCur
->pPage
= pCur
->apPage
[0];
4977 }else if( pCur
->pgnoRoot
==0 ){
4978 pCur
->eState
= CURSOR_INVALID
;
4979 return SQLITE_EMPTY
;
4981 assert( pCur
->iPage
==(-1) );
4982 if( pCur
->eState
>=CURSOR_REQUIRESEEK
){
4983 if( pCur
->eState
==CURSOR_FAULT
){
4984 assert( pCur
->skipNext
!=SQLITE_OK
);
4985 return pCur
->skipNext
;
4987 sqlite3BtreeClearCursor(pCur
);
4989 rc
= getAndInitPage(pCur
->pBtree
->pBt
, pCur
->pgnoRoot
, &pCur
->pPage
,
4990 0, pCur
->curPagerFlags
);
4991 if( rc
!=SQLITE_OK
){
4992 pCur
->eState
= CURSOR_INVALID
;
4996 pCur
->curIntKey
= pCur
->pPage
->intKey
;
4998 pRoot
= pCur
->pPage
;
4999 assert( pRoot
->pgno
==pCur
->pgnoRoot
);
5001 /* If pCur->pKeyInfo is not NULL, then the caller that opened this cursor
5002 ** expected to open it on an index b-tree. Otherwise, if pKeyInfo is
5003 ** NULL, the caller expects a table b-tree. If this is not the case,
5004 ** return an SQLITE_CORRUPT error.
5006 ** Earlier versions of SQLite assumed that this test could not fail
5007 ** if the root page was already loaded when this function was called (i.e.
5008 ** if pCur->iPage>=0). But this is not so if the database is corrupted
5009 ** in such a way that page pRoot is linked into a second b-tree table
5010 ** (or the freelist). */
5011 assert( pRoot
->intKey
==1 || pRoot
->intKey
==0 );
5012 if( pRoot
->isInit
==0 || (pCur
->pKeyInfo
==0)!=pRoot
->intKey
){
5013 return SQLITE_CORRUPT_PGNO(pCur
->pPage
->pgno
);
5018 pCur
->info
.nSize
= 0;
5019 pCur
->curFlags
&= ~(BTCF_AtLast
|BTCF_ValidNKey
|BTCF_ValidOvfl
);
5021 pRoot
= pCur
->pPage
;
5022 if( pRoot
->nCell
>0 ){
5023 pCur
->eState
= CURSOR_VALID
;
5024 }else if( !pRoot
->leaf
){
5026 if( pRoot
->pgno
!=1 ) return SQLITE_CORRUPT_BKPT
;
5027 subpage
= get4byte(&pRoot
->aData
[pRoot
->hdrOffset
+8]);
5028 pCur
->eState
= CURSOR_VALID
;
5029 rc
= moveToChild(pCur
, subpage
);
5031 pCur
->eState
= CURSOR_INVALID
;
5038 ** Move the cursor down to the left-most leaf entry beneath the
5039 ** entry to which it is currently pointing.
5041 ** The left-most leaf is the one with the smallest key - the first
5042 ** in ascending order.
5044 static int moveToLeftmost(BtCursor
*pCur
){
5049 assert( cursorOwnsBtShared(pCur
) );
5050 assert( pCur
->eState
==CURSOR_VALID
);
5051 while( rc
==SQLITE_OK
&& !(pPage
= pCur
->pPage
)->leaf
){
5052 assert( pCur
->ix
<pPage
->nCell
);
5053 pgno
= get4byte(findCell(pPage
, pCur
->ix
));
5054 rc
= moveToChild(pCur
, pgno
);
5060 ** Move the cursor down to the right-most leaf entry beneath the
5061 ** page to which it is currently pointing. Notice the difference
5062 ** between moveToLeftmost() and moveToRightmost(). moveToLeftmost()
5063 ** finds the left-most entry beneath the *entry* whereas moveToRightmost()
5064 ** finds the right-most entry beneath the *page*.
5066 ** The right-most entry is the one with the largest key - the last
5067 ** key in ascending order.
5069 static int moveToRightmost(BtCursor
*pCur
){
5074 assert( cursorOwnsBtShared(pCur
) );
5075 assert( pCur
->eState
==CURSOR_VALID
);
5076 while( !(pPage
= pCur
->pPage
)->leaf
){
5077 pgno
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
5078 pCur
->ix
= pPage
->nCell
;
5079 rc
= moveToChild(pCur
, pgno
);
5082 pCur
->ix
= pPage
->nCell
-1;
5083 assert( pCur
->info
.nSize
==0 );
5084 assert( (pCur
->curFlags
& BTCF_ValidNKey
)==0 );
5088 /* Move the cursor to the first entry in the table. Return SQLITE_OK
5089 ** on success. Set *pRes to 0 if the cursor actually points to something
5090 ** or set *pRes to 1 if the table is empty.
5092 int sqlite3BtreeFirst(BtCursor
*pCur
, int *pRes
){
5095 assert( cursorOwnsBtShared(pCur
) );
5096 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5097 rc
= moveToRoot(pCur
);
5098 if( rc
==SQLITE_OK
){
5099 assert( pCur
->pPage
->nCell
>0 );
5101 rc
= moveToLeftmost(pCur
);
5102 }else if( rc
==SQLITE_EMPTY
){
5103 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5110 /* Move the cursor to the last entry in the table. Return SQLITE_OK
5111 ** on success. Set *pRes to 0 if the cursor actually points to something
5112 ** or set *pRes to 1 if the table is empty.
5114 int sqlite3BtreeLast(BtCursor
*pCur
, int *pRes
){
5117 assert( cursorOwnsBtShared(pCur
) );
5118 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5120 /* If the cursor already points to the last entry, this is a no-op. */
5121 if( CURSOR_VALID
==pCur
->eState
&& (pCur
->curFlags
& BTCF_AtLast
)!=0 ){
5123 /* This block serves to assert() that the cursor really does point
5124 ** to the last entry in the b-tree. */
5126 for(ii
=0; ii
<pCur
->iPage
; ii
++){
5127 assert( pCur
->aiIdx
[ii
]==pCur
->apPage
[ii
]->nCell
);
5129 assert( pCur
->ix
==pCur
->pPage
->nCell
-1 );
5130 assert( pCur
->pPage
->leaf
);
5135 rc
= moveToRoot(pCur
);
5136 if( rc
==SQLITE_OK
){
5137 assert( pCur
->eState
==CURSOR_VALID
);
5139 rc
= moveToRightmost(pCur
);
5140 if( rc
==SQLITE_OK
){
5141 pCur
->curFlags
|= BTCF_AtLast
;
5143 pCur
->curFlags
&= ~BTCF_AtLast
;
5145 }else if( rc
==SQLITE_EMPTY
){
5146 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5153 /* Move the cursor so that it points to an entry near the key
5154 ** specified by pIdxKey or intKey. Return a success code.
5156 ** For INTKEY tables, the intKey parameter is used. pIdxKey
5157 ** must be NULL. For index tables, pIdxKey is used and intKey
5160 ** If an exact match is not found, then the cursor is always
5161 ** left pointing at a leaf page which would hold the entry if it
5162 ** were present. The cursor might point to an entry that comes
5163 ** before or after the key.
5165 ** An integer is written into *pRes which is the result of
5166 ** comparing the key with the entry to which the cursor is
5167 ** pointing. The meaning of the integer written into
5168 ** *pRes is as follows:
5170 ** *pRes<0 The cursor is left pointing at an entry that
5171 ** is smaller than intKey/pIdxKey or if the table is empty
5172 ** and the cursor is therefore left point to nothing.
5174 ** *pRes==0 The cursor is left pointing at an entry that
5175 ** exactly matches intKey/pIdxKey.
5177 ** *pRes>0 The cursor is left pointing at an entry that
5178 ** is larger than intKey/pIdxKey.
5180 ** For index tables, the pIdxKey->eqSeen field is set to 1 if there
5181 ** exists an entry in the table that exactly matches pIdxKey.
5183 int sqlite3BtreeMovetoUnpacked(
5184 BtCursor
*pCur
, /* The cursor to be moved */
5185 UnpackedRecord
*pIdxKey
, /* Unpacked index key */
5186 i64 intKey
, /* The table key */
5187 int biasRight
, /* If true, bias the search to the high end */
5188 int *pRes
/* Write search results here */
5191 RecordCompare xRecordCompare
;
5193 assert( cursorOwnsBtShared(pCur
) );
5194 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5196 assert( (pIdxKey
==0)==(pCur
->pKeyInfo
==0) );
5197 assert( pCur
->eState
!=CURSOR_VALID
|| (pIdxKey
==0)==(pCur
->curIntKey
!=0) );
5199 /* If the cursor is already positioned at the point we are trying
5200 ** to move to, then just return without doing any work */
5202 && pCur
->eState
==CURSOR_VALID
&& (pCur
->curFlags
& BTCF_ValidNKey
)!=0
5204 if( pCur
->info
.nKey
==intKey
){
5208 if( pCur
->info
.nKey
<intKey
){
5209 if( (pCur
->curFlags
& BTCF_AtLast
)!=0 ){
5213 /* If the requested key is one more than the previous key, then
5214 ** try to get there using sqlite3BtreeNext() rather than a full
5215 ** binary search. This is an optimization only. The correct answer
5216 ** is still obtained without this case, only a little more slowely */
5217 if( pCur
->info
.nKey
+1==intKey
&& !pCur
->skipNext
){
5219 rc
= sqlite3BtreeNext(pCur
, 0);
5220 if( rc
==SQLITE_OK
){
5222 if( pCur
->info
.nKey
==intKey
){
5225 }else if( rc
==SQLITE_DONE
){
5235 xRecordCompare
= sqlite3VdbeFindCompare(pIdxKey
);
5236 pIdxKey
->errCode
= 0;
5237 assert( pIdxKey
->default_rc
==1
5238 || pIdxKey
->default_rc
==0
5239 || pIdxKey
->default_rc
==-1
5242 xRecordCompare
= 0; /* All keys are integers */
5245 rc
= moveToRoot(pCur
);
5247 if( rc
==SQLITE_EMPTY
){
5248 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5254 assert( pCur
->pPage
);
5255 assert( pCur
->pPage
->isInit
);
5256 assert( pCur
->eState
==CURSOR_VALID
);
5257 assert( pCur
->pPage
->nCell
> 0 );
5258 assert( pCur
->iPage
==0 || pCur
->apPage
[0]->intKey
==pCur
->curIntKey
);
5259 assert( pCur
->curIntKey
|| pIdxKey
);
5261 int lwr
, upr
, idx
, c
;
5263 MemPage
*pPage
= pCur
->pPage
;
5264 u8
*pCell
; /* Pointer to current cell in pPage */
5266 /* pPage->nCell must be greater than zero. If this is the root-page
5267 ** the cursor would have been INVALID above and this for(;;) loop
5268 ** not run. If this is not the root-page, then the moveToChild() routine
5269 ** would have already detected db corruption. Similarly, pPage must
5270 ** be the right kind (index or table) of b-tree page. Otherwise
5271 ** a moveToChild() or moveToRoot() call would have detected corruption. */
5272 assert( pPage
->nCell
>0 );
5273 assert( pPage
->intKey
==(pIdxKey
==0) );
5275 upr
= pPage
->nCell
-1;
5276 assert( biasRight
==0 || biasRight
==1 );
5277 idx
= upr
>>(1-biasRight
); /* idx = biasRight ? upr : (lwr+upr)/2; */
5278 pCur
->ix
= (u16
)idx
;
5279 if( xRecordCompare
==0 ){
5282 pCell
= findCellPastPtr(pPage
, idx
);
5283 if( pPage
->intKeyLeaf
){
5284 while( 0x80 <= *(pCell
++) ){
5285 if( pCell
>=pPage
->aDataEnd
){
5286 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
5290 getVarint(pCell
, (u64
*)&nCellKey
);
5291 if( nCellKey
<intKey
){
5293 if( lwr
>upr
){ c
= -1; break; }
5294 }else if( nCellKey
>intKey
){
5296 if( lwr
>upr
){ c
= +1; break; }
5298 assert( nCellKey
==intKey
);
5299 pCur
->ix
= (u16
)idx
;
5302 goto moveto_next_layer
;
5304 pCur
->curFlags
|= BTCF_ValidNKey
;
5305 pCur
->info
.nKey
= nCellKey
;
5306 pCur
->info
.nSize
= 0;
5311 assert( lwr
+upr
>=0 );
5312 idx
= (lwr
+upr
)>>1; /* idx = (lwr+upr)/2; */
5316 int nCell
; /* Size of the pCell cell in bytes */
5317 pCell
= findCellPastPtr(pPage
, idx
);
5319 /* The maximum supported page-size is 65536 bytes. This means that
5320 ** the maximum number of record bytes stored on an index B-Tree
5321 ** page is less than 16384 bytes and may be stored as a 2-byte
5322 ** varint. This information is used to attempt to avoid parsing
5323 ** the entire cell by checking for the cases where the record is
5324 ** stored entirely within the b-tree page by inspecting the first
5325 ** 2 bytes of the cell.
5328 if( nCell
<=pPage
->max1bytePayload
){
5329 /* This branch runs if the record-size field of the cell is a
5330 ** single byte varint and the record fits entirely on the main
5332 testcase( pCell
+nCell
+1==pPage
->aDataEnd
);
5333 c
= xRecordCompare(nCell
, (void*)&pCell
[1], pIdxKey
);
5334 }else if( !(pCell
[1] & 0x80)
5335 && (nCell
= ((nCell
&0x7f)<<7) + pCell
[1])<=pPage
->maxLocal
5337 /* The record-size field is a 2 byte varint and the record
5338 ** fits entirely on the main b-tree page. */
5339 testcase( pCell
+nCell
+2==pPage
->aDataEnd
);
5340 c
= xRecordCompare(nCell
, (void*)&pCell
[2], pIdxKey
);
5342 /* The record flows over onto one or more overflow pages. In
5343 ** this case the whole cell needs to be parsed, a buffer allocated
5344 ** and accessPayload() used to retrieve the record into the
5345 ** buffer before VdbeRecordCompare() can be called.
5347 ** If the record is corrupt, the xRecordCompare routine may read
5348 ** up to two varints past the end of the buffer. An extra 18
5349 ** bytes of padding is allocated at the end of the buffer in
5350 ** case this happens. */
5352 u8
* const pCellBody
= pCell
- pPage
->childPtrSize
;
5353 pPage
->xParseCell(pPage
, pCellBody
, &pCur
->info
);
5354 nCell
= (int)pCur
->info
.nKey
;
5355 testcase( nCell
<0 ); /* True if key size is 2^32 or more */
5356 testcase( nCell
==0 ); /* Invalid key size: 0x80 0x80 0x00 */
5357 testcase( nCell
==1 ); /* Invalid key size: 0x80 0x80 0x01 */
5358 testcase( nCell
==2 ); /* Minimum legal index key size */
5360 rc
= SQLITE_CORRUPT_PGNO(pPage
->pgno
);
5363 pCellKey
= sqlite3Malloc( nCell
+18 );
5365 rc
= SQLITE_NOMEM_BKPT
;
5368 pCur
->ix
= (u16
)idx
;
5369 rc
= accessPayload(pCur
, 0, nCell
, (unsigned char*)pCellKey
, 0);
5370 pCur
->curFlags
&= ~BTCF_ValidOvfl
;
5372 sqlite3_free(pCellKey
);
5375 c
= xRecordCompare(nCell
, pCellKey
, pIdxKey
);
5376 sqlite3_free(pCellKey
);
5379 (pIdxKey
->errCode
!=SQLITE_CORRUPT
|| c
==0)
5380 && (pIdxKey
->errCode
!=SQLITE_NOMEM
|| pCur
->pBtree
->db
->mallocFailed
)
5390 pCur
->ix
= (u16
)idx
;
5391 if( pIdxKey
->errCode
) rc
= SQLITE_CORRUPT_BKPT
;
5394 if( lwr
>upr
) break;
5395 assert( lwr
+upr
>=0 );
5396 idx
= (lwr
+upr
)>>1; /* idx = (lwr+upr)/2 */
5399 assert( lwr
==upr
+1 || (pPage
->intKey
&& !pPage
->leaf
) );
5400 assert( pPage
->isInit
);
5402 assert( pCur
->ix
<pCur
->pPage
->nCell
);
5403 pCur
->ix
= (u16
)idx
;
5409 if( lwr
>=pPage
->nCell
){
5410 chldPg
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
5412 chldPg
= get4byte(findCell(pPage
, lwr
));
5414 pCur
->ix
= (u16
)lwr
;
5415 rc
= moveToChild(pCur
, chldPg
);
5419 pCur
->info
.nSize
= 0;
5420 assert( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 );
5426 ** Return TRUE if the cursor is not pointing at an entry of the table.
5428 ** TRUE will be returned after a call to sqlite3BtreeNext() moves
5429 ** past the last entry in the table or sqlite3BtreePrev() moves past
5430 ** the first entry. TRUE is also returned if the table is empty.
5432 int sqlite3BtreeEof(BtCursor
*pCur
){
5433 /* TODO: What if the cursor is in CURSOR_REQUIRESEEK but all table entries
5434 ** have been deleted? This API will need to change to return an error code
5435 ** as well as the boolean result value.
5437 return (CURSOR_VALID
!=pCur
->eState
);
5441 ** Return an estimate for the number of rows in the table that pCur is
5442 ** pointing to. Return a negative number if no estimate is currently
5445 i64
sqlite3BtreeRowCountEst(BtCursor
*pCur
){
5449 assert( cursorOwnsBtShared(pCur
) );
5450 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5452 /* Currently this interface is only called by the OP_IfSmaller
5453 ** opcode, and it that case the cursor will always be valid and
5454 ** will always point to a leaf node. */
5455 if( NEVER(pCur
->eState
!=CURSOR_VALID
) ) return -1;
5456 if( NEVER(pCur
->pPage
->leaf
==0) ) return -1;
5458 n
= pCur
->pPage
->nCell
;
5459 for(i
=0; i
<pCur
->iPage
; i
++){
5460 n
*= pCur
->apPage
[i
]->nCell
;
5466 ** Advance the cursor to the next entry in the database.
5469 ** SQLITE_OK success
5470 ** SQLITE_DONE cursor is already pointing at the last element
5471 ** otherwise some kind of error occurred
5473 ** The main entry point is sqlite3BtreeNext(). That routine is optimized
5474 ** for the common case of merely incrementing the cell counter BtCursor.aiIdx
5475 ** to the next cell on the current page. The (slower) btreeNext() helper
5476 ** routine is called when it is necessary to move to a different page or
5477 ** to restore the cursor.
5479 ** If bit 0x01 of the F argument in sqlite3BtreeNext(C,F) is 1, then the
5480 ** cursor corresponds to an SQL index and this routine could have been
5481 ** skipped if the SQL index had been a unique index. The F argument
5482 ** is a hint to the implement. SQLite btree implementation does not use
5483 ** this hint, but COMDB2 does.
5485 static SQLITE_NOINLINE
int btreeNext(BtCursor
*pCur
){
5490 assert( cursorOwnsBtShared(pCur
) );
5491 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5492 if( pCur
->eState
!=CURSOR_VALID
){
5493 assert( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 );
5494 rc
= restoreCursorPosition(pCur
);
5495 if( rc
!=SQLITE_OK
){
5498 if( CURSOR_INVALID
==pCur
->eState
){
5501 if( pCur
->skipNext
){
5502 assert( pCur
->eState
==CURSOR_VALID
|| pCur
->eState
==CURSOR_SKIPNEXT
);
5503 pCur
->eState
= CURSOR_VALID
;
5504 if( pCur
->skipNext
>0 ){
5512 pPage
= pCur
->pPage
;
5514 assert( pPage
->isInit
);
5516 /* If the database file is corrupt, it is possible for the value of idx
5517 ** to be invalid here. This can only occur if a second cursor modifies
5518 ** the page while cursor pCur is holding a reference to it. Which can
5519 ** only happen if the database is corrupt in such a way as to link the
5520 ** page into more than one b-tree structure. */
5521 testcase( idx
>pPage
->nCell
);
5523 if( idx
>=pPage
->nCell
){
5525 rc
= moveToChild(pCur
, get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]));
5527 return moveToLeftmost(pCur
);
5530 if( pCur
->iPage
==0 ){
5531 pCur
->eState
= CURSOR_INVALID
;
5535 pPage
= pCur
->pPage
;
5536 }while( pCur
->ix
>=pPage
->nCell
);
5537 if( pPage
->intKey
){
5538 return sqlite3BtreeNext(pCur
, 0);
5546 return moveToLeftmost(pCur
);
5549 int sqlite3BtreeNext(BtCursor
*pCur
, int flags
){
5551 UNUSED_PARAMETER( flags
); /* Used in COMDB2 but not native SQLite */
5552 assert( cursorOwnsBtShared(pCur
) );
5553 assert( flags
==0 || flags
==1 );
5554 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5555 pCur
->info
.nSize
= 0;
5556 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
5557 if( pCur
->eState
!=CURSOR_VALID
) return btreeNext(pCur
);
5558 pPage
= pCur
->pPage
;
5559 if( (++pCur
->ix
)>=pPage
->nCell
){
5561 return btreeNext(pCur
);
5566 return moveToLeftmost(pCur
);
5571 ** Step the cursor to the back to the previous entry in the database.
5574 ** SQLITE_OK success
5575 ** SQLITE_DONE the cursor is already on the first element of the table
5576 ** otherwise some kind of error occurred
5578 ** The main entry point is sqlite3BtreePrevious(). That routine is optimized
5579 ** for the common case of merely decrementing the cell counter BtCursor.aiIdx
5580 ** to the previous cell on the current page. The (slower) btreePrevious()
5581 ** helper routine is called when it is necessary to move to a different page
5582 ** or to restore the cursor.
5584 ** If bit 0x01 of the F argument to sqlite3BtreePrevious(C,F) is 1, then
5585 ** the cursor corresponds to an SQL index and this routine could have been
5586 ** skipped if the SQL index had been a unique index. The F argument is a
5587 ** hint to the implement. The native SQLite btree implementation does not
5588 ** use this hint, but COMDB2 does.
5590 static SQLITE_NOINLINE
int btreePrevious(BtCursor
*pCur
){
5594 assert( cursorOwnsBtShared(pCur
) );
5595 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5596 assert( (pCur
->curFlags
& (BTCF_AtLast
|BTCF_ValidOvfl
|BTCF_ValidNKey
))==0 );
5597 assert( pCur
->info
.nSize
==0 );
5598 if( pCur
->eState
!=CURSOR_VALID
){
5599 rc
= restoreCursorPosition(pCur
);
5600 if( rc
!=SQLITE_OK
){
5603 if( CURSOR_INVALID
==pCur
->eState
){
5606 if( pCur
->skipNext
){
5607 assert( pCur
->eState
==CURSOR_VALID
|| pCur
->eState
==CURSOR_SKIPNEXT
);
5608 pCur
->eState
= CURSOR_VALID
;
5609 if( pCur
->skipNext
<0 ){
5617 pPage
= pCur
->pPage
;
5618 assert( pPage
->isInit
);
5621 rc
= moveToChild(pCur
, get4byte(findCell(pPage
, idx
)));
5623 rc
= moveToRightmost(pCur
);
5625 while( pCur
->ix
==0 ){
5626 if( pCur
->iPage
==0 ){
5627 pCur
->eState
= CURSOR_INVALID
;
5632 assert( pCur
->info
.nSize
==0 );
5633 assert( (pCur
->curFlags
& (BTCF_ValidOvfl
))==0 );
5636 pPage
= pCur
->pPage
;
5637 if( pPage
->intKey
&& !pPage
->leaf
){
5638 rc
= sqlite3BtreePrevious(pCur
, 0);
5645 int sqlite3BtreePrevious(BtCursor
*pCur
, int flags
){
5646 assert( cursorOwnsBtShared(pCur
) );
5647 assert( flags
==0 || flags
==1 );
5648 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5649 UNUSED_PARAMETER( flags
); /* Used in COMDB2 but not native SQLite */
5650 pCur
->curFlags
&= ~(BTCF_AtLast
|BTCF_ValidOvfl
|BTCF_ValidNKey
);
5651 pCur
->info
.nSize
= 0;
5652 if( pCur
->eState
!=CURSOR_VALID
5654 || pCur
->pPage
->leaf
==0
5656 return btreePrevious(pCur
);
5663 ** Allocate a new page from the database file.
5665 ** The new page is marked as dirty. (In other words, sqlite3PagerWrite()
5666 ** has already been called on the new page.) The new page has also
5667 ** been referenced and the calling routine is responsible for calling
5668 ** sqlite3PagerUnref() on the new page when it is done.
5670 ** SQLITE_OK is returned on success. Any other return value indicates
5671 ** an error. *ppPage is set to NULL in the event of an error.
5673 ** If the "nearby" parameter is not 0, then an effort is made to
5674 ** locate a page close to the page number "nearby". This can be used in an
5675 ** attempt to keep related pages close to each other in the database file,
5676 ** which in turn can make database access faster.
5678 ** If the eMode parameter is BTALLOC_EXACT and the nearby page exists
5679 ** anywhere on the free-list, then it is guaranteed to be returned. If
5680 ** eMode is BTALLOC_LT then the page returned will be less than or equal
5681 ** to nearby if any such page exists. If eMode is BTALLOC_ANY then there
5682 ** are no restrictions on which page is returned.
5684 static int allocateBtreePage(
5685 BtShared
*pBt
, /* The btree */
5686 MemPage
**ppPage
, /* Store pointer to the allocated page here */
5687 Pgno
*pPgno
, /* Store the page number here */
5688 Pgno nearby
, /* Search for a page near this one */
5689 u8 eMode
/* BTALLOC_EXACT, BTALLOC_LT, or BTALLOC_ANY */
5693 u32 n
; /* Number of pages on the freelist */
5694 u32 k
; /* Number of leaves on the trunk of the freelist */
5695 MemPage
*pTrunk
= 0;
5696 MemPage
*pPrevTrunk
= 0;
5697 Pgno mxPage
; /* Total size of the database file */
5699 assert( sqlite3_mutex_held(pBt
->mutex
) );
5700 assert( eMode
==BTALLOC_ANY
|| (nearby
>0 && IfNotOmitAV(pBt
->autoVacuum
)) );
5701 pPage1
= pBt
->pPage1
;
5702 mxPage
= btreePagecount(pBt
);
5703 /* EVIDENCE-OF: R-05119-02637 The 4-byte big-endian integer at offset 36
5704 ** stores stores the total number of pages on the freelist. */
5705 n
= get4byte(&pPage1
->aData
[36]);
5706 testcase( n
==mxPage
-1 );
5708 return SQLITE_CORRUPT_BKPT
;
5711 /* There are pages on the freelist. Reuse one of those pages. */
5713 u8 searchList
= 0; /* If the free-list must be searched for 'nearby' */
5714 u32 nSearch
= 0; /* Count of the number of search attempts */
5716 /* If eMode==BTALLOC_EXACT and a query of the pointer-map
5717 ** shows that the page 'nearby' is somewhere on the free-list, then
5718 ** the entire-list will be searched for that page.
5720 #ifndef SQLITE_OMIT_AUTOVACUUM
5721 if( eMode
==BTALLOC_EXACT
){
5722 if( nearby
<=mxPage
){
5725 assert( pBt
->autoVacuum
);
5726 rc
= ptrmapGet(pBt
, nearby
, &eType
, 0);
5728 if( eType
==PTRMAP_FREEPAGE
){
5732 }else if( eMode
==BTALLOC_LE
){
5737 /* Decrement the free-list count by 1. Set iTrunk to the index of the
5738 ** first free-list trunk page. iPrevTrunk is initially 1.
5740 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
5742 put4byte(&pPage1
->aData
[36], n
-1);
5744 /* The code within this loop is run only once if the 'searchList' variable
5745 ** is not true. Otherwise, it runs once for each trunk-page on the
5746 ** free-list until the page 'nearby' is located (eMode==BTALLOC_EXACT)
5747 ** or until a page less than 'nearby' is located (eMode==BTALLOC_LT)
5750 pPrevTrunk
= pTrunk
;
5752 /* EVIDENCE-OF: R-01506-11053 The first integer on a freelist trunk page
5753 ** is the page number of the next freelist trunk page in the list or
5754 ** zero if this is the last freelist trunk page. */
5755 iTrunk
= get4byte(&pPrevTrunk
->aData
[0]);
5757 /* EVIDENCE-OF: R-59841-13798 The 4-byte big-endian integer at offset 32
5758 ** stores the page number of the first page of the freelist, or zero if
5759 ** the freelist is empty. */
5760 iTrunk
= get4byte(&pPage1
->aData
[32]);
5762 testcase( iTrunk
==mxPage
);
5763 if( iTrunk
>mxPage
|| nSearch
++ > n
){
5764 rc
= SQLITE_CORRUPT_PGNO(pPrevTrunk
? pPrevTrunk
->pgno
: 1);
5766 rc
= btreeGetUnusedPage(pBt
, iTrunk
, &pTrunk
, 0);
5770 goto end_allocate_page
;
5772 assert( pTrunk
!=0 );
5773 assert( pTrunk
->aData
!=0 );
5774 /* EVIDENCE-OF: R-13523-04394 The second integer on a freelist trunk page
5775 ** is the number of leaf page pointers to follow. */
5776 k
= get4byte(&pTrunk
->aData
[4]);
5777 if( k
==0 && !searchList
){
5778 /* The trunk has no leaves and the list is not being searched.
5779 ** So extract the trunk page itself and use it as the newly
5780 ** allocated page */
5781 assert( pPrevTrunk
==0 );
5782 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
5784 goto end_allocate_page
;
5787 memcpy(&pPage1
->aData
[32], &pTrunk
->aData
[0], 4);
5790 TRACE(("ALLOCATE: %d trunk - %d free pages left\n", *pPgno
, n
-1));
5791 }else if( k
>(u32
)(pBt
->usableSize
/4 - 2) ){
5792 /* Value of k is out of range. Database corruption */
5793 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
5794 goto end_allocate_page
;
5795 #ifndef SQLITE_OMIT_AUTOVACUUM
5796 }else if( searchList
5797 && (nearby
==iTrunk
|| (iTrunk
<nearby
&& eMode
==BTALLOC_LE
))
5799 /* The list is being searched and this trunk page is the page
5800 ** to allocate, regardless of whether it has leaves.
5805 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
5807 goto end_allocate_page
;
5811 memcpy(&pPage1
->aData
[32], &pTrunk
->aData
[0], 4);
5813 rc
= sqlite3PagerWrite(pPrevTrunk
->pDbPage
);
5814 if( rc
!=SQLITE_OK
){
5815 goto end_allocate_page
;
5817 memcpy(&pPrevTrunk
->aData
[0], &pTrunk
->aData
[0], 4);
5820 /* The trunk page is required by the caller but it contains
5821 ** pointers to free-list leaves. The first leaf becomes a trunk
5822 ** page in this case.
5825 Pgno iNewTrunk
= get4byte(&pTrunk
->aData
[8]);
5826 if( iNewTrunk
>mxPage
){
5827 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
5828 goto end_allocate_page
;
5830 testcase( iNewTrunk
==mxPage
);
5831 rc
= btreeGetUnusedPage(pBt
, iNewTrunk
, &pNewTrunk
, 0);
5832 if( rc
!=SQLITE_OK
){
5833 goto end_allocate_page
;
5835 rc
= sqlite3PagerWrite(pNewTrunk
->pDbPage
);
5836 if( rc
!=SQLITE_OK
){
5837 releasePage(pNewTrunk
);
5838 goto end_allocate_page
;
5840 memcpy(&pNewTrunk
->aData
[0], &pTrunk
->aData
[0], 4);
5841 put4byte(&pNewTrunk
->aData
[4], k
-1);
5842 memcpy(&pNewTrunk
->aData
[8], &pTrunk
->aData
[12], (k
-1)*4);
5843 releasePage(pNewTrunk
);
5845 assert( sqlite3PagerIswriteable(pPage1
->pDbPage
) );
5846 put4byte(&pPage1
->aData
[32], iNewTrunk
);
5848 rc
= sqlite3PagerWrite(pPrevTrunk
->pDbPage
);
5850 goto end_allocate_page
;
5852 put4byte(&pPrevTrunk
->aData
[0], iNewTrunk
);
5856 TRACE(("ALLOCATE: %d trunk - %d free pages left\n", *pPgno
, n
-1));
5859 /* Extract a leaf from the trunk */
5862 unsigned char *aData
= pTrunk
->aData
;
5866 if( eMode
==BTALLOC_LE
){
5868 iPage
= get4byte(&aData
[8+i
*4]);
5869 if( iPage
<=nearby
){
5876 dist
= sqlite3AbsInt32(get4byte(&aData
[8]) - nearby
);
5878 int d2
= sqlite3AbsInt32(get4byte(&aData
[8+i
*4]) - nearby
);
5889 iPage
= get4byte(&aData
[8+closest
*4]);
5890 testcase( iPage
==mxPage
);
5892 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
5893 goto end_allocate_page
;
5895 testcase( iPage
==mxPage
);
5897 || (iPage
==nearby
|| (iPage
<nearby
&& eMode
==BTALLOC_LE
))
5901 TRACE(("ALLOCATE: %d was leaf %d of %d on trunk %d"
5902 ": %d more free pages\n",
5903 *pPgno
, closest
+1, k
, pTrunk
->pgno
, n
-1));
5904 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
5905 if( rc
) goto end_allocate_page
;
5907 memcpy(&aData
[8+closest
*4], &aData
[4+k
*4], 4);
5909 put4byte(&aData
[4], k
-1);
5910 noContent
= !btreeGetHasContent(pBt
, *pPgno
)? PAGER_GET_NOCONTENT
: 0;
5911 rc
= btreeGetUnusedPage(pBt
, *pPgno
, ppPage
, noContent
);
5912 if( rc
==SQLITE_OK
){
5913 rc
= sqlite3PagerWrite((*ppPage
)->pDbPage
);
5914 if( rc
!=SQLITE_OK
){
5915 releasePage(*ppPage
);
5922 releasePage(pPrevTrunk
);
5924 }while( searchList
);
5926 /* There are no pages on the freelist, so append a new page to the
5929 ** Normally, new pages allocated by this block can be requested from the
5930 ** pager layer with the 'no-content' flag set. This prevents the pager
5931 ** from trying to read the pages content from disk. However, if the
5932 ** current transaction has already run one or more incremental-vacuum
5933 ** steps, then the page we are about to allocate may contain content
5934 ** that is required in the event of a rollback. In this case, do
5935 ** not set the no-content flag. This causes the pager to load and journal
5936 ** the current page content before overwriting it.
5938 ** Note that the pager will not actually attempt to load or journal
5939 ** content for any page that really does lie past the end of the database
5940 ** file on disk. So the effects of disabling the no-content optimization
5941 ** here are confined to those pages that lie between the end of the
5942 ** database image and the end of the database file.
5944 int bNoContent
= (0==IfNotOmitAV(pBt
->bDoTruncate
))? PAGER_GET_NOCONTENT
:0;
5946 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
5949 if( pBt
->nPage
==PENDING_BYTE_PAGE(pBt
) ) pBt
->nPage
++;
5951 #ifndef SQLITE_OMIT_AUTOVACUUM
5952 if( pBt
->autoVacuum
&& PTRMAP_ISPAGE(pBt
, pBt
->nPage
) ){
5953 /* If *pPgno refers to a pointer-map page, allocate two new pages
5954 ** at the end of the file instead of one. The first allocated page
5955 ** becomes a new pointer-map page, the second is used by the caller.
5958 TRACE(("ALLOCATE: %d from end of file (pointer-map page)\n", pBt
->nPage
));
5959 assert( pBt
->nPage
!=PENDING_BYTE_PAGE(pBt
) );
5960 rc
= btreeGetUnusedPage(pBt
, pBt
->nPage
, &pPg
, bNoContent
);
5961 if( rc
==SQLITE_OK
){
5962 rc
= sqlite3PagerWrite(pPg
->pDbPage
);
5967 if( pBt
->nPage
==PENDING_BYTE_PAGE(pBt
) ){ pBt
->nPage
++; }
5970 put4byte(28 + (u8
*)pBt
->pPage1
->aData
, pBt
->nPage
);
5971 *pPgno
= pBt
->nPage
;
5973 assert( *pPgno
!=PENDING_BYTE_PAGE(pBt
) );
5974 rc
= btreeGetUnusedPage(pBt
, *pPgno
, ppPage
, bNoContent
);
5976 rc
= sqlite3PagerWrite((*ppPage
)->pDbPage
);
5977 if( rc
!=SQLITE_OK
){
5978 releasePage(*ppPage
);
5981 TRACE(("ALLOCATE: %d from end of file\n", *pPgno
));
5984 assert( *pPgno
!=PENDING_BYTE_PAGE(pBt
) );
5987 releasePage(pTrunk
);
5988 releasePage(pPrevTrunk
);
5989 assert( rc
!=SQLITE_OK
|| sqlite3PagerPageRefcount((*ppPage
)->pDbPage
)<=1 );
5990 assert( rc
!=SQLITE_OK
|| (*ppPage
)->isInit
==0 );
5995 ** This function is used to add page iPage to the database file free-list.
5996 ** It is assumed that the page is not already a part of the free-list.
5998 ** The value passed as the second argument to this function is optional.
5999 ** If the caller happens to have a pointer to the MemPage object
6000 ** corresponding to page iPage handy, it may pass it as the second value.
6001 ** Otherwise, it may pass NULL.
6003 ** If a pointer to a MemPage object is passed as the second argument,
6004 ** its reference count is not altered by this function.
6006 static int freePage2(BtShared
*pBt
, MemPage
*pMemPage
, Pgno iPage
){
6007 MemPage
*pTrunk
= 0; /* Free-list trunk page */
6008 Pgno iTrunk
= 0; /* Page number of free-list trunk page */
6009 MemPage
*pPage1
= pBt
->pPage1
; /* Local reference to page 1 */
6010 MemPage
*pPage
; /* Page being freed. May be NULL. */
6011 int rc
; /* Return Code */
6012 int nFree
; /* Initial number of pages on free-list */
6014 assert( sqlite3_mutex_held(pBt
->mutex
) );
6015 assert( CORRUPT_DB
|| iPage
>1 );
6016 assert( !pMemPage
|| pMemPage
->pgno
==iPage
);
6018 if( iPage
<2 ) return SQLITE_CORRUPT_BKPT
;
6021 sqlite3PagerRef(pPage
->pDbPage
);
6023 pPage
= btreePageLookup(pBt
, iPage
);
6026 /* Increment the free page count on pPage1 */
6027 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
6028 if( rc
) goto freepage_out
;
6029 nFree
= get4byte(&pPage1
->aData
[36]);
6030 put4byte(&pPage1
->aData
[36], nFree
+1);
6032 if( pBt
->btsFlags
& BTS_SECURE_DELETE
){
6033 /* If the secure_delete option is enabled, then
6034 ** always fully overwrite deleted information with zeros.
6036 if( (!pPage
&& ((rc
= btreeGetPage(pBt
, iPage
, &pPage
, 0))!=0) )
6037 || ((rc
= sqlite3PagerWrite(pPage
->pDbPage
))!=0)
6041 memset(pPage
->aData
, 0, pPage
->pBt
->pageSize
);
6044 /* If the database supports auto-vacuum, write an entry in the pointer-map
6045 ** to indicate that the page is free.
6048 ptrmapPut(pBt
, iPage
, PTRMAP_FREEPAGE
, 0, &rc
);
6049 if( rc
) goto freepage_out
;
6052 /* Now manipulate the actual database free-list structure. There are two
6053 ** possibilities. If the free-list is currently empty, or if the first
6054 ** trunk page in the free-list is full, then this page will become a
6055 ** new free-list trunk page. Otherwise, it will become a leaf of the
6056 ** first trunk page in the current free-list. This block tests if it
6057 ** is possible to add the page as a new free-list leaf.
6060 u32 nLeaf
; /* Initial number of leaf cells on trunk page */
6062 iTrunk
= get4byte(&pPage1
->aData
[32]);
6063 rc
= btreeGetPage(pBt
, iTrunk
, &pTrunk
, 0);
6064 if( rc
!=SQLITE_OK
){
6068 nLeaf
= get4byte(&pTrunk
->aData
[4]);
6069 assert( pBt
->usableSize
>32 );
6070 if( nLeaf
> (u32
)pBt
->usableSize
/4 - 2 ){
6071 rc
= SQLITE_CORRUPT_BKPT
;
6074 if( nLeaf
< (u32
)pBt
->usableSize
/4 - 8 ){
6075 /* In this case there is room on the trunk page to insert the page
6076 ** being freed as a new leaf.
6078 ** Note that the trunk page is not really full until it contains
6079 ** usableSize/4 - 2 entries, not usableSize/4 - 8 entries as we have
6080 ** coded. But due to a coding error in versions of SQLite prior to
6081 ** 3.6.0, databases with freelist trunk pages holding more than
6082 ** usableSize/4 - 8 entries will be reported as corrupt. In order
6083 ** to maintain backwards compatibility with older versions of SQLite,
6084 ** we will continue to restrict the number of entries to usableSize/4 - 8
6085 ** for now. At some point in the future (once everyone has upgraded
6086 ** to 3.6.0 or later) we should consider fixing the conditional above
6087 ** to read "usableSize/4-2" instead of "usableSize/4-8".
6089 ** EVIDENCE-OF: R-19920-11576 However, newer versions of SQLite still
6090 ** avoid using the last six entries in the freelist trunk page array in
6091 ** order that database files created by newer versions of SQLite can be
6092 ** read by older versions of SQLite.
6094 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
6095 if( rc
==SQLITE_OK
){
6096 put4byte(&pTrunk
->aData
[4], nLeaf
+1);
6097 put4byte(&pTrunk
->aData
[8+nLeaf
*4], iPage
);
6098 if( pPage
&& (pBt
->btsFlags
& BTS_SECURE_DELETE
)==0 ){
6099 sqlite3PagerDontWrite(pPage
->pDbPage
);
6101 rc
= btreeSetHasContent(pBt
, iPage
);
6103 TRACE(("FREE-PAGE: %d leaf on trunk page %d\n",pPage
->pgno
,pTrunk
->pgno
));
6108 /* If control flows to this point, then it was not possible to add the
6109 ** the page being freed as a leaf page of the first trunk in the free-list.
6110 ** Possibly because the free-list is empty, or possibly because the
6111 ** first trunk in the free-list is full. Either way, the page being freed
6112 ** will become the new first trunk page in the free-list.
6114 if( pPage
==0 && SQLITE_OK
!=(rc
= btreeGetPage(pBt
, iPage
, &pPage
, 0)) ){
6117 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
6118 if( rc
!=SQLITE_OK
){
6121 put4byte(pPage
->aData
, iTrunk
);
6122 put4byte(&pPage
->aData
[4], 0);
6123 put4byte(&pPage1
->aData
[32], iPage
);
6124 TRACE(("FREE-PAGE: %d new trunk page replacing %d\n", pPage
->pgno
, iTrunk
));
6131 releasePage(pTrunk
);
6134 static void freePage(MemPage
*pPage
, int *pRC
){
6135 if( (*pRC
)==SQLITE_OK
){
6136 *pRC
= freePage2(pPage
->pBt
, pPage
, pPage
->pgno
);
6141 ** Free any overflow pages associated with the given Cell. Write the
6142 ** local Cell size (the number of bytes on the original page, omitting
6143 ** overflow) into *pnSize.
6145 static int clearCell(
6146 MemPage
*pPage
, /* The page that contains the Cell */
6147 unsigned char *pCell
, /* First byte of the Cell */
6148 CellInfo
*pInfo
/* Size information about the cell */
6156 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6157 pPage
->xParseCell(pPage
, pCell
, pInfo
);
6158 if( pInfo
->nLocal
==pInfo
->nPayload
){
6159 return SQLITE_OK
; /* No overflow pages. Return without doing anything */
6161 if( pCell
+pInfo
->nSize
-1 > pPage
->aData
+pPage
->maskPage
){
6162 /* Cell extends past end of page */
6163 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
6165 ovflPgno
= get4byte(pCell
+ pInfo
->nSize
- 4);
6167 assert( pBt
->usableSize
> 4 );
6168 ovflPageSize
= pBt
->usableSize
- 4;
6169 nOvfl
= (pInfo
->nPayload
- pInfo
->nLocal
+ ovflPageSize
- 1)/ovflPageSize
;
6171 (CORRUPT_DB
&& (pInfo
->nPayload
+ ovflPageSize
)<ovflPageSize
)
6176 if( ovflPgno
<2 || ovflPgno
>btreePagecount(pBt
) ){
6177 /* 0 is not a legal page number and page 1 cannot be an
6178 ** overflow page. Therefore if ovflPgno<2 or past the end of the
6179 ** file the database must be corrupt. */
6180 return SQLITE_CORRUPT_BKPT
;
6183 rc
= getOverflowPage(pBt
, ovflPgno
, &pOvfl
, &iNext
);
6187 if( ( pOvfl
|| ((pOvfl
= btreePageLookup(pBt
, ovflPgno
))!=0) )
6188 && sqlite3PagerPageRefcount(pOvfl
->pDbPage
)!=1
6190 /* There is no reason any cursor should have an outstanding reference
6191 ** to an overflow page belonging to a cell that is being deleted/updated.
6192 ** So if there exists more than one reference to this page, then it
6193 ** must not really be an overflow page and the database must be corrupt.
6194 ** It is helpful to detect this before calling freePage2(), as
6195 ** freePage2() may zero the page contents if secure-delete mode is
6196 ** enabled. If this 'overflow' page happens to be a page that the
6197 ** caller is iterating through or using in some other way, this
6198 ** can be problematic.
6200 rc
= SQLITE_CORRUPT_BKPT
;
6202 rc
= freePage2(pBt
, pOvfl
, ovflPgno
);
6206 sqlite3PagerUnref(pOvfl
->pDbPage
);
6215 ** Create the byte sequence used to represent a cell on page pPage
6216 ** and write that byte sequence into pCell[]. Overflow pages are
6217 ** allocated and filled in as necessary. The calling procedure
6218 ** is responsible for making sure sufficient space has been allocated
6221 ** Note that pCell does not necessary need to point to the pPage->aData
6222 ** area. pCell might point to some temporary storage. The cell will
6223 ** be constructed in this temporary area then copied into pPage->aData
6226 static int fillInCell(
6227 MemPage
*pPage
, /* The page that contains the cell */
6228 unsigned char *pCell
, /* Complete text of the cell */
6229 const BtreePayload
*pX
, /* Payload with which to construct the cell */
6230 int *pnSize
/* Write cell size here */
6234 int nSrc
, n
, rc
, mn
;
6236 MemPage
*pToRelease
;
6237 unsigned char *pPrior
;
6238 unsigned char *pPayload
;
6243 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6245 /* pPage is not necessarily writeable since pCell might be auxiliary
6246 ** buffer space that is separate from the pPage buffer area */
6247 assert( pCell
<pPage
->aData
|| pCell
>=&pPage
->aData
[pPage
->pBt
->pageSize
]
6248 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6250 /* Fill in the header. */
6251 nHeader
= pPage
->childPtrSize
;
6252 if( pPage
->intKey
){
6253 nPayload
= pX
->nData
+ pX
->nZero
;
6256 assert( pPage
->intKeyLeaf
); /* fillInCell() only called for leaves */
6257 nHeader
+= putVarint32(&pCell
[nHeader
], nPayload
);
6258 nHeader
+= putVarint(&pCell
[nHeader
], *(u64
*)&pX
->nKey
);
6260 assert( pX
->nKey
<=0x7fffffff && pX
->pKey
!=0 );
6261 nSrc
= nPayload
= (int)pX
->nKey
;
6263 nHeader
+= putVarint32(&pCell
[nHeader
], nPayload
);
6266 /* Fill in the payload */
6267 pPayload
= &pCell
[nHeader
];
6268 if( nPayload
<=pPage
->maxLocal
){
6269 /* This is the common case where everything fits on the btree page
6270 ** and no overflow pages are required. */
6271 n
= nHeader
+ nPayload
;
6276 assert( nSrc
<=nPayload
);
6277 testcase( nSrc
<nPayload
);
6278 memcpy(pPayload
, pSrc
, nSrc
);
6279 memset(pPayload
+nSrc
, 0, nPayload
-nSrc
);
6283 /* If we reach this point, it means that some of the content will need
6284 ** to spill onto overflow pages.
6286 mn
= pPage
->minLocal
;
6287 n
= mn
+ (nPayload
- mn
) % (pPage
->pBt
->usableSize
- 4);
6288 testcase( n
==pPage
->maxLocal
);
6289 testcase( n
==pPage
->maxLocal
+1 );
6290 if( n
> pPage
->maxLocal
) n
= mn
;
6292 *pnSize
= n
+ nHeader
+ 4;
6293 pPrior
= &pCell
[nHeader
+n
];
6298 /* At this point variables should be set as follows:
6300 ** nPayload Total payload size in bytes
6301 ** pPayload Begin writing payload here
6302 ** spaceLeft Space available at pPayload. If nPayload>spaceLeft,
6303 ** that means content must spill into overflow pages.
6304 ** *pnSize Size of the local cell (not counting overflow pages)
6305 ** pPrior Where to write the pgno of the first overflow page
6307 ** Use a call to btreeParseCellPtr() to verify that the values above
6308 ** were computed correctly.
6313 pPage
->xParseCell(pPage
, pCell
, &info
);
6314 assert( nHeader
==(int)(info
.pPayload
- pCell
) );
6315 assert( info
.nKey
==pX
->nKey
);
6316 assert( *pnSize
== info
.nSize
);
6317 assert( spaceLeft
== info
.nLocal
);
6321 /* Write the payload into the local Cell and any extra into overflow pages */
6324 if( n
>spaceLeft
) n
= spaceLeft
;
6326 /* If pToRelease is not zero than pPayload points into the data area
6327 ** of pToRelease. Make sure pToRelease is still writeable. */
6328 assert( pToRelease
==0 || sqlite3PagerIswriteable(pToRelease
->pDbPage
) );
6330 /* If pPayload is part of the data area of pPage, then make sure pPage
6331 ** is still writeable */
6332 assert( pPayload
<pPage
->aData
|| pPayload
>=&pPage
->aData
[pBt
->pageSize
]
6333 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6336 memcpy(pPayload
, pSrc
, n
);
6339 memcpy(pPayload
, pSrc
, n
);
6341 memset(pPayload
, 0, n
);
6344 if( nPayload
<=0 ) break;
6351 #ifndef SQLITE_OMIT_AUTOVACUUM
6352 Pgno pgnoPtrmap
= pgnoOvfl
; /* Overflow page pointer-map entry page */
6353 if( pBt
->autoVacuum
){
6357 PTRMAP_ISPAGE(pBt
, pgnoOvfl
) || pgnoOvfl
==PENDING_BYTE_PAGE(pBt
)
6361 rc
= allocateBtreePage(pBt
, &pOvfl
, &pgnoOvfl
, pgnoOvfl
, 0);
6362 #ifndef SQLITE_OMIT_AUTOVACUUM
6363 /* If the database supports auto-vacuum, and the second or subsequent
6364 ** overflow page is being allocated, add an entry to the pointer-map
6365 ** for that page now.
6367 ** If this is the first overflow page, then write a partial entry
6368 ** to the pointer-map. If we write nothing to this pointer-map slot,
6369 ** then the optimistic overflow chain processing in clearCell()
6370 ** may misinterpret the uninitialized values and delete the
6371 ** wrong pages from the database.
6373 if( pBt
->autoVacuum
&& rc
==SQLITE_OK
){
6374 u8 eType
= (pgnoPtrmap
?PTRMAP_OVERFLOW2
:PTRMAP_OVERFLOW1
);
6375 ptrmapPut(pBt
, pgnoOvfl
, eType
, pgnoPtrmap
, &rc
);
6382 releasePage(pToRelease
);
6386 /* If pToRelease is not zero than pPrior points into the data area
6387 ** of pToRelease. Make sure pToRelease is still writeable. */
6388 assert( pToRelease
==0 || sqlite3PagerIswriteable(pToRelease
->pDbPage
) );
6390 /* If pPrior is part of the data area of pPage, then make sure pPage
6391 ** is still writeable */
6392 assert( pPrior
<pPage
->aData
|| pPrior
>=&pPage
->aData
[pBt
->pageSize
]
6393 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6395 put4byte(pPrior
, pgnoOvfl
);
6396 releasePage(pToRelease
);
6398 pPrior
= pOvfl
->aData
;
6399 put4byte(pPrior
, 0);
6400 pPayload
= &pOvfl
->aData
[4];
6401 spaceLeft
= pBt
->usableSize
- 4;
6404 releasePage(pToRelease
);
6409 ** Remove the i-th cell from pPage. This routine effects pPage only.
6410 ** The cell content is not freed or deallocated. It is assumed that
6411 ** the cell content has been copied someplace else. This routine just
6412 ** removes the reference to the cell from pPage.
6414 ** "sz" must be the number of bytes in the cell.
6416 static void dropCell(MemPage
*pPage
, int idx
, int sz
, int *pRC
){
6417 u32 pc
; /* Offset to cell content of cell being deleted */
6418 u8
*data
; /* pPage->aData */
6419 u8
*ptr
; /* Used to move bytes around within data[] */
6420 int rc
; /* The return code */
6421 int hdr
; /* Beginning of the header. 0 most pages. 100 page 1 */
6424 assert( idx
>=0 && idx
<pPage
->nCell
);
6425 assert( CORRUPT_DB
|| sz
==cellSize(pPage
, idx
) );
6426 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
6427 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6428 data
= pPage
->aData
;
6429 ptr
= &pPage
->aCellIdx
[2*idx
];
6431 hdr
= pPage
->hdrOffset
;
6432 testcase( pc
==get2byte(&data
[hdr
+5]) );
6433 testcase( pc
+sz
==pPage
->pBt
->usableSize
);
6434 if( pc
+sz
> pPage
->pBt
->usableSize
){
6435 *pRC
= SQLITE_CORRUPT_BKPT
;
6438 rc
= freeSpace(pPage
, pc
, sz
);
6444 if( pPage
->nCell
==0 ){
6445 memset(&data
[hdr
+1], 0, 4);
6447 put2byte(&data
[hdr
+5], pPage
->pBt
->usableSize
);
6448 pPage
->nFree
= pPage
->pBt
->usableSize
- pPage
->hdrOffset
6449 - pPage
->childPtrSize
- 8;
6451 memmove(ptr
, ptr
+2, 2*(pPage
->nCell
- idx
));
6452 put2byte(&data
[hdr
+3], pPage
->nCell
);
6458 ** Insert a new cell on pPage at cell index "i". pCell points to the
6459 ** content of the cell.
6461 ** If the cell content will fit on the page, then put it there. If it
6462 ** will not fit, then make a copy of the cell content into pTemp if
6463 ** pTemp is not null. Regardless of pTemp, allocate a new entry
6464 ** in pPage->apOvfl[] and make it point to the cell content (either
6465 ** in pTemp or the original pCell) and also record its index.
6466 ** Allocating a new entry in pPage->aCell[] implies that
6467 ** pPage->nOverflow is incremented.
6469 ** *pRC must be SQLITE_OK when this routine is called.
6471 static void insertCell(
6472 MemPage
*pPage
, /* Page into which we are copying */
6473 int i
, /* New cell becomes the i-th cell of the page */
6474 u8
*pCell
, /* Content of the new cell */
6475 int sz
, /* Bytes of content in pCell */
6476 u8
*pTemp
, /* Temp storage space for pCell, if needed */
6477 Pgno iChild
, /* If non-zero, replace first 4 bytes with this value */
6478 int *pRC
/* Read and write return code from here */
6480 int idx
= 0; /* Where to write new cell content in data[] */
6481 int j
; /* Loop counter */
6482 u8
*data
; /* The content of the whole page */
6483 u8
*pIns
; /* The point in pPage->aCellIdx[] where no cell inserted */
6485 assert( *pRC
==SQLITE_OK
);
6486 assert( i
>=0 && i
<=pPage
->nCell
+pPage
->nOverflow
);
6487 assert( MX_CELL(pPage
->pBt
)<=10921 );
6488 assert( pPage
->nCell
<=MX_CELL(pPage
->pBt
) || CORRUPT_DB
);
6489 assert( pPage
->nOverflow
<=ArraySize(pPage
->apOvfl
) );
6490 assert( ArraySize(pPage
->apOvfl
)==ArraySize(pPage
->aiOvfl
) );
6491 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6492 /* The cell should normally be sized correctly. However, when moving a
6493 ** malformed cell from a leaf page to an interior page, if the cell size
6494 ** wanted to be less than 4 but got rounded up to 4 on the leaf, then size
6495 ** might be less than 8 (leaf-size + pointer) on the interior node. Hence
6496 ** the term after the || in the following assert(). */
6497 assert( sz
==pPage
->xCellSize(pPage
, pCell
) || (sz
==8 && iChild
>0) );
6498 if( pPage
->nOverflow
|| sz
+2>pPage
->nFree
){
6500 memcpy(pTemp
, pCell
, sz
);
6504 put4byte(pCell
, iChild
);
6506 j
= pPage
->nOverflow
++;
6507 /* Comparison against ArraySize-1 since we hold back one extra slot
6508 ** as a contingency. In other words, never need more than 3 overflow
6509 ** slots but 4 are allocated, just to be safe. */
6510 assert( j
< ArraySize(pPage
->apOvfl
)-1 );
6511 pPage
->apOvfl
[j
] = pCell
;
6512 pPage
->aiOvfl
[j
] = (u16
)i
;
6514 /* When multiple overflows occur, they are always sequential and in
6515 ** sorted order. This invariants arise because multiple overflows can
6516 ** only occur when inserting divider cells into the parent page during
6517 ** balancing, and the dividers are adjacent and sorted.
6519 assert( j
==0 || pPage
->aiOvfl
[j
-1]<(u16
)i
); /* Overflows in sorted order */
6520 assert( j
==0 || i
==pPage
->aiOvfl
[j
-1]+1 ); /* Overflows are sequential */
6522 int rc
= sqlite3PagerWrite(pPage
->pDbPage
);
6523 if( rc
!=SQLITE_OK
){
6527 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
6528 data
= pPage
->aData
;
6529 assert( &data
[pPage
->cellOffset
]==pPage
->aCellIdx
);
6530 rc
= allocateSpace(pPage
, sz
, &idx
);
6531 if( rc
){ *pRC
= rc
; return; }
6532 /* The allocateSpace() routine guarantees the following properties
6533 ** if it returns successfully */
6535 assert( idx
>= pPage
->cellOffset
+2*pPage
->nCell
+2 || CORRUPT_DB
);
6536 assert( idx
+sz
<= (int)pPage
->pBt
->usableSize
);
6537 pPage
->nFree
-= (u16
)(2 + sz
);
6538 memcpy(&data
[idx
], pCell
, sz
);
6540 put4byte(&data
[idx
], iChild
);
6542 pIns
= pPage
->aCellIdx
+ i
*2;
6543 memmove(pIns
+2, pIns
, 2*(pPage
->nCell
- i
));
6544 put2byte(pIns
, idx
);
6546 /* increment the cell count */
6547 if( (++data
[pPage
->hdrOffset
+4])==0 ) data
[pPage
->hdrOffset
+3]++;
6548 assert( get2byte(&data
[pPage
->hdrOffset
+3])==pPage
->nCell
);
6549 #ifndef SQLITE_OMIT_AUTOVACUUM
6550 if( pPage
->pBt
->autoVacuum
){
6551 /* The cell may contain a pointer to an overflow page. If so, write
6552 ** the entry for the overflow page into the pointer map.
6554 ptrmapPutOvflPtr(pPage
, pCell
, pRC
);
6561 ** A CellArray object contains a cache of pointers and sizes for a
6562 ** consecutive sequence of cells that might be held on multiple pages.
6564 typedef struct CellArray CellArray
;
6566 int nCell
; /* Number of cells in apCell[] */
6567 MemPage
*pRef
; /* Reference page */
6568 u8
**apCell
; /* All cells begin balanced */
6569 u16
*szCell
; /* Local size of all cells in apCell[] */
6573 ** Make sure the cell sizes at idx, idx+1, ..., idx+N-1 have been
6576 static void populateCellCache(CellArray
*p
, int idx
, int N
){
6577 assert( idx
>=0 && idx
+N
<=p
->nCell
);
6579 assert( p
->apCell
[idx
]!=0 );
6580 if( p
->szCell
[idx
]==0 ){
6581 p
->szCell
[idx
] = p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[idx
]);
6583 assert( CORRUPT_DB
||
6584 p
->szCell
[idx
]==p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[idx
]) );
6592 ** Return the size of the Nth element of the cell array
6594 static SQLITE_NOINLINE u16
computeCellSize(CellArray
*p
, int N
){
6595 assert( N
>=0 && N
<p
->nCell
);
6596 assert( p
->szCell
[N
]==0 );
6597 p
->szCell
[N
] = p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[N
]);
6598 return p
->szCell
[N
];
6600 static u16
cachedCellSize(CellArray
*p
, int N
){
6601 assert( N
>=0 && N
<p
->nCell
);
6602 if( p
->szCell
[N
] ) return p
->szCell
[N
];
6603 return computeCellSize(p
, N
);
6607 ** Array apCell[] contains pointers to nCell b-tree page cells. The
6608 ** szCell[] array contains the size in bytes of each cell. This function
6609 ** replaces the current contents of page pPg with the contents of the cell
6612 ** Some of the cells in apCell[] may currently be stored in pPg. This
6613 ** function works around problems caused by this by making a copy of any
6614 ** such cells before overwriting the page data.
6616 ** The MemPage.nFree field is invalidated by this function. It is the
6617 ** responsibility of the caller to set it correctly.
6619 static int rebuildPage(
6620 MemPage
*pPg
, /* Edit this page */
6621 int nCell
, /* Final number of cells on page */
6622 u8
**apCell
, /* Array of cells */
6623 u16
*szCell
/* Array of cell sizes */
6625 const int hdr
= pPg
->hdrOffset
; /* Offset of header on pPg */
6626 u8
* const aData
= pPg
->aData
; /* Pointer to data for pPg */
6627 const int usableSize
= pPg
->pBt
->usableSize
;
6628 u8
* const pEnd
= &aData
[usableSize
];
6630 u8
*pCellptr
= pPg
->aCellIdx
;
6631 u8
*pTmp
= sqlite3PagerTempSpace(pPg
->pBt
->pPager
);
6634 i
= get2byte(&aData
[hdr
+5]);
6635 memcpy(&pTmp
[i
], &aData
[i
], usableSize
- i
);
6638 for(i
=0; i
<nCell
; i
++){
6639 u8
*pCell
= apCell
[i
];
6640 if( SQLITE_WITHIN(pCell
,aData
,pEnd
) ){
6641 pCell
= &pTmp
[pCell
- aData
];
6644 put2byte(pCellptr
, (pData
- aData
));
6646 if( pData
< pCellptr
) return SQLITE_CORRUPT_BKPT
;
6647 memcpy(pData
, pCell
, szCell
[i
]);
6648 assert( szCell
[i
]==pPg
->xCellSize(pPg
, pCell
) || CORRUPT_DB
);
6649 testcase( szCell
[i
]!=pPg
->xCellSize(pPg
,pCell
) );
6652 /* The pPg->nFree field is now set incorrectly. The caller will fix it. */
6656 put2byte(&aData
[hdr
+1], 0);
6657 put2byte(&aData
[hdr
+3], pPg
->nCell
);
6658 put2byte(&aData
[hdr
+5], pData
- aData
);
6659 aData
[hdr
+7] = 0x00;
6664 ** Array apCell[] contains nCell pointers to b-tree cells. Array szCell
6665 ** contains the size in bytes of each such cell. This function attempts to
6666 ** add the cells stored in the array to page pPg. If it cannot (because
6667 ** the page needs to be defragmented before the cells will fit), non-zero
6668 ** is returned. Otherwise, if the cells are added successfully, zero is
6671 ** Argument pCellptr points to the first entry in the cell-pointer array
6672 ** (part of page pPg) to populate. After cell apCell[0] is written to the
6673 ** page body, a 16-bit offset is written to pCellptr. And so on, for each
6674 ** cell in the array. It is the responsibility of the caller to ensure
6675 ** that it is safe to overwrite this part of the cell-pointer array.
6677 ** When this function is called, *ppData points to the start of the
6678 ** content area on page pPg. If the size of the content area is extended,
6679 ** *ppData is updated to point to the new start of the content area
6680 ** before returning.
6682 ** Finally, argument pBegin points to the byte immediately following the
6683 ** end of the space required by this page for the cell-pointer area (for
6684 ** all cells - not just those inserted by the current call). If the content
6685 ** area must be extended to before this point in order to accomodate all
6686 ** cells in apCell[], then the cells do not fit and non-zero is returned.
6688 static int pageInsertArray(
6689 MemPage
*pPg
, /* Page to add cells to */
6690 u8
*pBegin
, /* End of cell-pointer array */
6691 u8
**ppData
, /* IN/OUT: Page content -area pointer */
6692 u8
*pCellptr
, /* Pointer to cell-pointer area */
6693 int iFirst
, /* Index of first cell to add */
6694 int nCell
, /* Number of cells to add to pPg */
6695 CellArray
*pCArray
/* Array of cells */
6698 u8
*aData
= pPg
->aData
;
6699 u8
*pData
= *ppData
;
6700 int iEnd
= iFirst
+ nCell
;
6701 assert( CORRUPT_DB
|| pPg
->hdrOffset
==0 ); /* Never called on page 1 */
6702 for(i
=iFirst
; i
<iEnd
; i
++){
6705 sz
= cachedCellSize(pCArray
, i
);
6706 if( (aData
[1]==0 && aData
[2]==0) || (pSlot
= pageFindSlot(pPg
,sz
,&rc
))==0 ){
6707 if( (pData
- pBegin
)<sz
) return 1;
6711 /* pSlot and pCArray->apCell[i] will never overlap on a well-formed
6712 ** database. But they might for a corrupt database. Hence use memmove()
6713 ** since memcpy() sends SIGABORT with overlapping buffers on OpenBSD */
6714 assert( (pSlot
+sz
)<=pCArray
->apCell
[i
]
6715 || pSlot
>=(pCArray
->apCell
[i
]+sz
)
6717 memmove(pSlot
, pCArray
->apCell
[i
], sz
);
6718 put2byte(pCellptr
, (pSlot
- aData
));
6726 ** Array apCell[] contains nCell pointers to b-tree cells. Array szCell
6727 ** contains the size in bytes of each such cell. This function adds the
6728 ** space associated with each cell in the array that is currently stored
6729 ** within the body of pPg to the pPg free-list. The cell-pointers and other
6730 ** fields of the page are not updated.
6732 ** This function returns the total number of cells added to the free-list.
6734 static int pageFreeArray(
6735 MemPage
*pPg
, /* Page to edit */
6736 int iFirst
, /* First cell to delete */
6737 int nCell
, /* Cells to delete */
6738 CellArray
*pCArray
/* Array of cells */
6740 u8
* const aData
= pPg
->aData
;
6741 u8
* const pEnd
= &aData
[pPg
->pBt
->usableSize
];
6742 u8
* const pStart
= &aData
[pPg
->hdrOffset
+ 8 + pPg
->childPtrSize
];
6745 int iEnd
= iFirst
+ nCell
;
6749 for(i
=iFirst
; i
<iEnd
; i
++){
6750 u8
*pCell
= pCArray
->apCell
[i
];
6751 if( SQLITE_WITHIN(pCell
, pStart
, pEnd
) ){
6753 /* No need to use cachedCellSize() here. The sizes of all cells that
6754 ** are to be freed have already been computing while deciding which
6755 ** cells need freeing */
6756 sz
= pCArray
->szCell
[i
]; assert( sz
>0 );
6757 if( pFree
!=(pCell
+ sz
) ){
6759 assert( pFree
>aData
&& (pFree
- aData
)<65536 );
6760 freeSpace(pPg
, (u16
)(pFree
- aData
), szFree
);
6764 if( pFree
+sz
>pEnd
) return 0;
6773 assert( pFree
>aData
&& (pFree
- aData
)<65536 );
6774 freeSpace(pPg
, (u16
)(pFree
- aData
), szFree
);
6780 ** apCell[] and szCell[] contains pointers to and sizes of all cells in the
6781 ** pages being balanced. The current page, pPg, has pPg->nCell cells starting
6782 ** with apCell[iOld]. After balancing, this page should hold nNew cells
6783 ** starting at apCell[iNew].
6785 ** This routine makes the necessary adjustments to pPg so that it contains
6786 ** the correct cells after being balanced.
6788 ** The pPg->nFree field is invalid when this function returns. It is the
6789 ** responsibility of the caller to set it correctly.
6791 static int editPage(
6792 MemPage
*pPg
, /* Edit this page */
6793 int iOld
, /* Index of first cell currently on page */
6794 int iNew
, /* Index of new first cell on page */
6795 int nNew
, /* Final number of cells on page */
6796 CellArray
*pCArray
/* Array of cells and sizes */
6798 u8
* const aData
= pPg
->aData
;
6799 const int hdr
= pPg
->hdrOffset
;
6800 u8
*pBegin
= &pPg
->aCellIdx
[nNew
* 2];
6801 int nCell
= pPg
->nCell
; /* Cells stored on pPg */
6805 int iOldEnd
= iOld
+ pPg
->nCell
+ pPg
->nOverflow
;
6806 int iNewEnd
= iNew
+ nNew
;
6809 u8
*pTmp
= sqlite3PagerTempSpace(pPg
->pBt
->pPager
);
6810 memcpy(pTmp
, aData
, pPg
->pBt
->usableSize
);
6813 /* Remove cells from the start and end of the page */
6815 int nShift
= pageFreeArray(pPg
, iOld
, iNew
-iOld
, pCArray
);
6816 memmove(pPg
->aCellIdx
, &pPg
->aCellIdx
[nShift
*2], nCell
*2);
6819 if( iNewEnd
< iOldEnd
){
6820 nCell
-= pageFreeArray(pPg
, iNewEnd
, iOldEnd
- iNewEnd
, pCArray
);
6823 pData
= &aData
[get2byteNotZero(&aData
[hdr
+5])];
6824 if( pData
<pBegin
) goto editpage_fail
;
6826 /* Add cells to the start of the page */
6828 int nAdd
= MIN(nNew
,iOld
-iNew
);
6829 assert( (iOld
-iNew
)<nNew
|| nCell
==0 || CORRUPT_DB
);
6830 pCellptr
= pPg
->aCellIdx
;
6831 memmove(&pCellptr
[nAdd
*2], pCellptr
, nCell
*2);
6832 if( pageInsertArray(
6833 pPg
, pBegin
, &pData
, pCellptr
,
6835 ) ) goto editpage_fail
;
6839 /* Add any overflow cells */
6840 for(i
=0; i
<pPg
->nOverflow
; i
++){
6841 int iCell
= (iOld
+ pPg
->aiOvfl
[i
]) - iNew
;
6842 if( iCell
>=0 && iCell
<nNew
){
6843 pCellptr
= &pPg
->aCellIdx
[iCell
* 2];
6844 memmove(&pCellptr
[2], pCellptr
, (nCell
- iCell
) * 2);
6846 if( pageInsertArray(
6847 pPg
, pBegin
, &pData
, pCellptr
,
6848 iCell
+iNew
, 1, pCArray
6849 ) ) goto editpage_fail
;
6853 /* Append cells to the end of the page */
6854 pCellptr
= &pPg
->aCellIdx
[nCell
*2];
6855 if( pageInsertArray(
6856 pPg
, pBegin
, &pData
, pCellptr
,
6857 iNew
+nCell
, nNew
-nCell
, pCArray
6858 ) ) goto editpage_fail
;
6863 put2byte(&aData
[hdr
+3], pPg
->nCell
);
6864 put2byte(&aData
[hdr
+5], pData
- aData
);
6867 for(i
=0; i
<nNew
&& !CORRUPT_DB
; i
++){
6868 u8
*pCell
= pCArray
->apCell
[i
+iNew
];
6869 int iOff
= get2byteAligned(&pPg
->aCellIdx
[i
*2]);
6870 if( SQLITE_WITHIN(pCell
, aData
, &aData
[pPg
->pBt
->usableSize
]) ){
6871 pCell
= &pTmp
[pCell
- aData
];
6873 assert( 0==memcmp(pCell
, &aData
[iOff
],
6874 pCArray
->pRef
->xCellSize(pCArray
->pRef
, pCArray
->apCell
[i
+iNew
])) );
6880 /* Unable to edit this page. Rebuild it from scratch instead. */
6881 populateCellCache(pCArray
, iNew
, nNew
);
6882 return rebuildPage(pPg
, nNew
, &pCArray
->apCell
[iNew
], &pCArray
->szCell
[iNew
]);
6886 ** The following parameters determine how many adjacent pages get involved
6887 ** in a balancing operation. NN is the number of neighbors on either side
6888 ** of the page that participate in the balancing operation. NB is the
6889 ** total number of pages that participate, including the target page and
6890 ** NN neighbors on either side.
6892 ** The minimum value of NN is 1 (of course). Increasing NN above 1
6893 ** (to 2 or 3) gives a modest improvement in SELECT and DELETE performance
6894 ** in exchange for a larger degradation in INSERT and UPDATE performance.
6895 ** The value of NN appears to give the best results overall.
6897 #define NN 1 /* Number of neighbors on either side of pPage */
6898 #define NB (NN*2+1) /* Total pages involved in the balance */
6901 #ifndef SQLITE_OMIT_QUICKBALANCE
6903 ** This version of balance() handles the common special case where
6904 ** a new entry is being inserted on the extreme right-end of the
6905 ** tree, in other words, when the new entry will become the largest
6906 ** entry in the tree.
6908 ** Instead of trying to balance the 3 right-most leaf pages, just add
6909 ** a new page to the right-hand side and put the one new entry in
6910 ** that page. This leaves the right side of the tree somewhat
6911 ** unbalanced. But odds are that we will be inserting new entries
6912 ** at the end soon afterwards so the nearly empty page will quickly
6913 ** fill up. On average.
6915 ** pPage is the leaf page which is the right-most page in the tree.
6916 ** pParent is its parent. pPage must have a single overflow entry
6917 ** which is also the right-most entry on the page.
6919 ** The pSpace buffer is used to store a temporary copy of the divider
6920 ** cell that will be inserted into pParent. Such a cell consists of a 4
6921 ** byte page number followed by a variable length integer. In other
6922 ** words, at most 13 bytes. Hence the pSpace buffer must be at
6923 ** least 13 bytes in size.
6925 static int balance_quick(MemPage
*pParent
, MemPage
*pPage
, u8
*pSpace
){
6926 BtShared
*const pBt
= pPage
->pBt
; /* B-Tree Database */
6927 MemPage
*pNew
; /* Newly allocated page */
6928 int rc
; /* Return Code */
6929 Pgno pgnoNew
; /* Page number of pNew */
6931 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6932 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
6933 assert( pPage
->nOverflow
==1 );
6935 /* This error condition is now caught prior to reaching this function */
6936 if( NEVER(pPage
->nCell
==0) ) return SQLITE_CORRUPT_BKPT
;
6938 /* Allocate a new page. This page will become the right-sibling of
6939 ** pPage. Make the parent page writable, so that the new divider cell
6940 ** may be inserted. If both these operations are successful, proceed.
6942 rc
= allocateBtreePage(pBt
, &pNew
, &pgnoNew
, 0, 0);
6944 if( rc
==SQLITE_OK
){
6946 u8
*pOut
= &pSpace
[4];
6947 u8
*pCell
= pPage
->apOvfl
[0];
6948 u16 szCell
= pPage
->xCellSize(pPage
, pCell
);
6951 assert( sqlite3PagerIswriteable(pNew
->pDbPage
) );
6952 assert( pPage
->aData
[0]==(PTF_INTKEY
|PTF_LEAFDATA
|PTF_LEAF
) );
6953 zeroPage(pNew
, PTF_INTKEY
|PTF_LEAFDATA
|PTF_LEAF
);
6954 rc
= rebuildPage(pNew
, 1, &pCell
, &szCell
);
6955 if( NEVER(rc
) ) return rc
;
6956 pNew
->nFree
= pBt
->usableSize
- pNew
->cellOffset
- 2 - szCell
;
6958 /* If this is an auto-vacuum database, update the pointer map
6959 ** with entries for the new page, and any pointer from the
6960 ** cell on the page to an overflow page. If either of these
6961 ** operations fails, the return code is set, but the contents
6962 ** of the parent page are still manipulated by thh code below.
6963 ** That is Ok, at this point the parent page is guaranteed to
6964 ** be marked as dirty. Returning an error code will cause a
6965 ** rollback, undoing any changes made to the parent page.
6968 ptrmapPut(pBt
, pgnoNew
, PTRMAP_BTREE
, pParent
->pgno
, &rc
);
6969 if( szCell
>pNew
->minLocal
){
6970 ptrmapPutOvflPtr(pNew
, pCell
, &rc
);
6974 /* Create a divider cell to insert into pParent. The divider cell
6975 ** consists of a 4-byte page number (the page number of pPage) and
6976 ** a variable length key value (which must be the same value as the
6977 ** largest key on pPage).
6979 ** To find the largest key value on pPage, first find the right-most
6980 ** cell on pPage. The first two fields of this cell are the
6981 ** record-length (a variable length integer at most 32-bits in size)
6982 ** and the key value (a variable length integer, may have any value).
6983 ** The first of the while(...) loops below skips over the record-length
6984 ** field. The second while(...) loop copies the key value from the
6985 ** cell on pPage into the pSpace buffer.
6987 pCell
= findCell(pPage
, pPage
->nCell
-1);
6989 while( (*(pCell
++)&0x80) && pCell
<pStop
);
6991 while( ((*(pOut
++) = *(pCell
++))&0x80) && pCell
<pStop
);
6993 /* Insert the new divider cell into pParent. */
6994 if( rc
==SQLITE_OK
){
6995 insertCell(pParent
, pParent
->nCell
, pSpace
, (int)(pOut
-pSpace
),
6996 0, pPage
->pgno
, &rc
);
6999 /* Set the right-child pointer of pParent to point to the new page. */
7000 put4byte(&pParent
->aData
[pParent
->hdrOffset
+8], pgnoNew
);
7002 /* Release the reference to the new page. */
7008 #endif /* SQLITE_OMIT_QUICKBALANCE */
7012 ** This function does not contribute anything to the operation of SQLite.
7013 ** it is sometimes activated temporarily while debugging code responsible
7014 ** for setting pointer-map entries.
7016 static int ptrmapCheckPages(MemPage
**apPage
, int nPage
){
7018 for(i
=0; i
<nPage
; i
++){
7021 MemPage
*pPage
= apPage
[i
];
7022 BtShared
*pBt
= pPage
->pBt
;
7023 assert( pPage
->isInit
);
7025 for(j
=0; j
<pPage
->nCell
; j
++){
7029 z
= findCell(pPage
, j
);
7030 pPage
->xParseCell(pPage
, z
, &info
);
7031 if( info
.nLocal
<info
.nPayload
){
7032 Pgno ovfl
= get4byte(&z
[info
.nSize
-4]);
7033 ptrmapGet(pBt
, ovfl
, &e
, &n
);
7034 assert( n
==pPage
->pgno
&& e
==PTRMAP_OVERFLOW1
);
7037 Pgno child
= get4byte(z
);
7038 ptrmapGet(pBt
, child
, &e
, &n
);
7039 assert( n
==pPage
->pgno
&& e
==PTRMAP_BTREE
);
7043 Pgno child
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
7044 ptrmapGet(pBt
, child
, &e
, &n
);
7045 assert( n
==pPage
->pgno
&& e
==PTRMAP_BTREE
);
7053 ** This function is used to copy the contents of the b-tree node stored
7054 ** on page pFrom to page pTo. If page pFrom was not a leaf page, then
7055 ** the pointer-map entries for each child page are updated so that the
7056 ** parent page stored in the pointer map is page pTo. If pFrom contained
7057 ** any cells with overflow page pointers, then the corresponding pointer
7058 ** map entries are also updated so that the parent page is page pTo.
7060 ** If pFrom is currently carrying any overflow cells (entries in the
7061 ** MemPage.apOvfl[] array), they are not copied to pTo.
7063 ** Before returning, page pTo is reinitialized using btreeInitPage().
7065 ** The performance of this function is not critical. It is only used by
7066 ** the balance_shallower() and balance_deeper() procedures, neither of
7067 ** which are called often under normal circumstances.
7069 static void copyNodeContent(MemPage
*pFrom
, MemPage
*pTo
, int *pRC
){
7070 if( (*pRC
)==SQLITE_OK
){
7071 BtShared
* const pBt
= pFrom
->pBt
;
7072 u8
* const aFrom
= pFrom
->aData
;
7073 u8
* const aTo
= pTo
->aData
;
7074 int const iFromHdr
= pFrom
->hdrOffset
;
7075 int const iToHdr
= ((pTo
->pgno
==1) ? 100 : 0);
7080 assert( pFrom
->isInit
);
7081 assert( pFrom
->nFree
>=iToHdr
);
7082 assert( get2byte(&aFrom
[iFromHdr
+5]) <= (int)pBt
->usableSize
);
7084 /* Copy the b-tree node content from page pFrom to page pTo. */
7085 iData
= get2byte(&aFrom
[iFromHdr
+5]);
7086 memcpy(&aTo
[iData
], &aFrom
[iData
], pBt
->usableSize
-iData
);
7087 memcpy(&aTo
[iToHdr
], &aFrom
[iFromHdr
], pFrom
->cellOffset
+ 2*pFrom
->nCell
);
7089 /* Reinitialize page pTo so that the contents of the MemPage structure
7090 ** match the new data. The initialization of pTo can actually fail under
7091 ** fairly obscure circumstances, even though it is a copy of initialized
7095 rc
= btreeInitPage(pTo
);
7096 if( rc
!=SQLITE_OK
){
7101 /* If this is an auto-vacuum database, update the pointer-map entries
7102 ** for any b-tree or overflow pages that pTo now contains the pointers to.
7105 *pRC
= setChildPtrmaps(pTo
);
7111 ** This routine redistributes cells on the iParentIdx'th child of pParent
7112 ** (hereafter "the page") and up to 2 siblings so that all pages have about the
7113 ** same amount of free space. Usually a single sibling on either side of the
7114 ** page are used in the balancing, though both siblings might come from one
7115 ** side if the page is the first or last child of its parent. If the page
7116 ** has fewer than 2 siblings (something which can only happen if the page
7117 ** is a root page or a child of a root page) then all available siblings
7118 ** participate in the balancing.
7120 ** The number of siblings of the page might be increased or decreased by
7121 ** one or two in an effort to keep pages nearly full but not over full.
7123 ** Note that when this routine is called, some of the cells on the page
7124 ** might not actually be stored in MemPage.aData[]. This can happen
7125 ** if the page is overfull. This routine ensures that all cells allocated
7126 ** to the page and its siblings fit into MemPage.aData[] before returning.
7128 ** In the course of balancing the page and its siblings, cells may be
7129 ** inserted into or removed from the parent page (pParent). Doing so
7130 ** may cause the parent page to become overfull or underfull. If this
7131 ** happens, it is the responsibility of the caller to invoke the correct
7132 ** balancing routine to fix this problem (see the balance() routine).
7134 ** If this routine fails for any reason, it might leave the database
7135 ** in a corrupted state. So if this routine fails, the database should
7138 ** The third argument to this function, aOvflSpace, is a pointer to a
7139 ** buffer big enough to hold one page. If while inserting cells into the parent
7140 ** page (pParent) the parent page becomes overfull, this buffer is
7141 ** used to store the parent's overflow cells. Because this function inserts
7142 ** a maximum of four divider cells into the parent page, and the maximum
7143 ** size of a cell stored within an internal node is always less than 1/4
7144 ** of the page-size, the aOvflSpace[] buffer is guaranteed to be large
7145 ** enough for all overflow cells.
7147 ** If aOvflSpace is set to a null pointer, this function returns
7150 static int balance_nonroot(
7151 MemPage
*pParent
, /* Parent page of siblings being balanced */
7152 int iParentIdx
, /* Index of "the page" in pParent */
7153 u8
*aOvflSpace
, /* page-size bytes of space for parent ovfl */
7154 int isRoot
, /* True if pParent is a root-page */
7155 int bBulk
/* True if this call is part of a bulk load */
7157 BtShared
*pBt
; /* The whole database */
7158 int nMaxCells
= 0; /* Allocated size of apCell, szCell, aFrom. */
7159 int nNew
= 0; /* Number of pages in apNew[] */
7160 int nOld
; /* Number of pages in apOld[] */
7161 int i
, j
, k
; /* Loop counters */
7162 int nxDiv
; /* Next divider slot in pParent->aCell[] */
7163 int rc
= SQLITE_OK
; /* The return code */
7164 u16 leafCorrection
; /* 4 if pPage is a leaf. 0 if not */
7165 int leafData
; /* True if pPage is a leaf of a LEAFDATA tree */
7166 int usableSpace
; /* Bytes in pPage beyond the header */
7167 int pageFlags
; /* Value of pPage->aData[0] */
7168 int iSpace1
= 0; /* First unused byte of aSpace1[] */
7169 int iOvflSpace
= 0; /* First unused byte of aOvflSpace[] */
7170 int szScratch
; /* Size of scratch memory requested */
7171 MemPage
*apOld
[NB
]; /* pPage and up to two siblings */
7172 MemPage
*apNew
[NB
+2]; /* pPage and up to NB siblings after balancing */
7173 u8
*pRight
; /* Location in parent of right-sibling pointer */
7174 u8
*apDiv
[NB
-1]; /* Divider cells in pParent */
7175 int cntNew
[NB
+2]; /* Index in b.paCell[] of cell after i-th page */
7176 int cntOld
[NB
+2]; /* Old index in b.apCell[] */
7177 int szNew
[NB
+2]; /* Combined size of cells placed on i-th page */
7178 u8
*aSpace1
; /* Space for copies of dividers cells */
7179 Pgno pgno
; /* Temp var to store a page number in */
7180 u8 abDone
[NB
+2]; /* True after i'th new page is populated */
7181 Pgno aPgno
[NB
+2]; /* Page numbers of new pages before shuffling */
7182 Pgno aPgOrder
[NB
+2]; /* Copy of aPgno[] used for sorting pages */
7183 u16 aPgFlags
[NB
+2]; /* flags field of new pages before shuffling */
7184 CellArray b
; /* Parsed information on cells being balanced */
7186 memset(abDone
, 0, sizeof(abDone
));
7190 assert( sqlite3_mutex_held(pBt
->mutex
) );
7191 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7194 TRACE(("BALANCE: begin page %d child of %d\n", pPage
->pgno
, pParent
->pgno
));
7197 /* At this point pParent may have at most one overflow cell. And if
7198 ** this overflow cell is present, it must be the cell with
7199 ** index iParentIdx. This scenario comes about when this function
7200 ** is called (indirectly) from sqlite3BtreeDelete().
7202 assert( pParent
->nOverflow
==0 || pParent
->nOverflow
==1 );
7203 assert( pParent
->nOverflow
==0 || pParent
->aiOvfl
[0]==iParentIdx
);
7206 return SQLITE_NOMEM_BKPT
;
7209 /* Find the sibling pages to balance. Also locate the cells in pParent
7210 ** that divide the siblings. An attempt is made to find NN siblings on
7211 ** either side of pPage. More siblings are taken from one side, however,
7212 ** if there are fewer than NN siblings on the other side. If pParent
7213 ** has NB or fewer children then all children of pParent are taken.
7215 ** This loop also drops the divider cells from the parent page. This
7216 ** way, the remainder of the function does not have to deal with any
7217 ** overflow cells in the parent page, since if any existed they will
7218 ** have already been removed.
7220 i
= pParent
->nOverflow
+ pParent
->nCell
;
7224 assert( bBulk
==0 || bBulk
==1 );
7225 if( iParentIdx
==0 ){
7227 }else if( iParentIdx
==i
){
7230 nxDiv
= iParentIdx
-1;
7235 if( (i
+nxDiv
-pParent
->nOverflow
)==pParent
->nCell
){
7236 pRight
= &pParent
->aData
[pParent
->hdrOffset
+8];
7238 pRight
= findCell(pParent
, i
+nxDiv
-pParent
->nOverflow
);
7240 pgno
= get4byte(pRight
);
7242 rc
= getAndInitPage(pBt
, pgno
, &apOld
[i
], 0, 0);
7244 memset(apOld
, 0, (i
+1)*sizeof(MemPage
*));
7245 goto balance_cleanup
;
7247 nMaxCells
+= 1+apOld
[i
]->nCell
+apOld
[i
]->nOverflow
;
7248 if( (i
--)==0 ) break;
7250 if( pParent
->nOverflow
&& i
+nxDiv
==pParent
->aiOvfl
[0] ){
7251 apDiv
[i
] = pParent
->apOvfl
[0];
7252 pgno
= get4byte(apDiv
[i
]);
7253 szNew
[i
] = pParent
->xCellSize(pParent
, apDiv
[i
]);
7254 pParent
->nOverflow
= 0;
7256 apDiv
[i
] = findCell(pParent
, i
+nxDiv
-pParent
->nOverflow
);
7257 pgno
= get4byte(apDiv
[i
]);
7258 szNew
[i
] = pParent
->xCellSize(pParent
, apDiv
[i
]);
7260 /* Drop the cell from the parent page. apDiv[i] still points to
7261 ** the cell within the parent, even though it has been dropped.
7262 ** This is safe because dropping a cell only overwrites the first
7263 ** four bytes of it, and this function does not need the first
7264 ** four bytes of the divider cell. So the pointer is safe to use
7267 ** But not if we are in secure-delete mode. In secure-delete mode,
7268 ** the dropCell() routine will overwrite the entire cell with zeroes.
7269 ** In this case, temporarily copy the cell into the aOvflSpace[]
7270 ** buffer. It will be copied out again as soon as the aSpace[] buffer
7272 if( pBt
->btsFlags
& BTS_FAST_SECURE
){
7275 iOff
= SQLITE_PTR_TO_INT(apDiv
[i
]) - SQLITE_PTR_TO_INT(pParent
->aData
);
7276 if( (iOff
+szNew
[i
])>(int)pBt
->usableSize
){
7277 rc
= SQLITE_CORRUPT_BKPT
;
7278 memset(apOld
, 0, (i
+1)*sizeof(MemPage
*));
7279 goto balance_cleanup
;
7281 memcpy(&aOvflSpace
[iOff
], apDiv
[i
], szNew
[i
]);
7282 apDiv
[i
] = &aOvflSpace
[apDiv
[i
]-pParent
->aData
];
7285 dropCell(pParent
, i
+nxDiv
-pParent
->nOverflow
, szNew
[i
], &rc
);
7289 /* Make nMaxCells a multiple of 4 in order to preserve 8-byte
7291 nMaxCells
= (nMaxCells
+ 3)&~3;
7294 ** Allocate space for memory structures
7297 nMaxCells
*sizeof(u8
*) /* b.apCell */
7298 + nMaxCells
*sizeof(u16
) /* b.szCell */
7299 + pBt
->pageSize
; /* aSpace1 */
7301 /* EVIDENCE-OF: R-28375-38319 SQLite will never request a scratch buffer
7302 ** that is more than 6 times the database page size. */
7303 assert( szScratch
<=6*(int)pBt
->pageSize
);
7304 b
.apCell
= sqlite3StackAllocRaw(0, szScratch
);
7306 rc
= SQLITE_NOMEM_BKPT
;
7307 goto balance_cleanup
;
7309 b
.szCell
= (u16
*)&b
.apCell
[nMaxCells
];
7310 aSpace1
= (u8
*)&b
.szCell
[nMaxCells
];
7311 assert( EIGHT_BYTE_ALIGNMENT(aSpace1
) );
7314 ** Load pointers to all cells on sibling pages and the divider cells
7315 ** into the local b.apCell[] array. Make copies of the divider cells
7316 ** into space obtained from aSpace1[]. The divider cells have already
7317 ** been removed from pParent.
7319 ** If the siblings are on leaf pages, then the child pointers of the
7320 ** divider cells are stripped from the cells before they are copied
7321 ** into aSpace1[]. In this way, all cells in b.apCell[] are without
7322 ** child pointers. If siblings are not leaves, then all cell in
7323 ** b.apCell[] include child pointers. Either way, all cells in b.apCell[]
7326 ** leafCorrection: 4 if pPage is a leaf. 0 if pPage is not a leaf.
7327 ** leafData: 1 if pPage holds key+data and pParent holds only keys.
7330 leafCorrection
= b
.pRef
->leaf
*4;
7331 leafData
= b
.pRef
->intKeyLeaf
;
7332 for(i
=0; i
<nOld
; i
++){
7333 MemPage
*pOld
= apOld
[i
];
7334 int limit
= pOld
->nCell
;
7335 u8
*aData
= pOld
->aData
;
7336 u16 maskPage
= pOld
->maskPage
;
7337 u8
*piCell
= aData
+ pOld
->cellOffset
;
7340 /* Verify that all sibling pages are of the same "type" (table-leaf,
7341 ** table-interior, index-leaf, or index-interior).
7343 if( pOld
->aData
[0]!=apOld
[0]->aData
[0] ){
7344 rc
= SQLITE_CORRUPT_BKPT
;
7345 goto balance_cleanup
;
7348 /* Load b.apCell[] with pointers to all cells in pOld. If pOld
7349 ** constains overflow cells, include them in the b.apCell[] array
7350 ** in the correct spot.
7352 ** Note that when there are multiple overflow cells, it is always the
7353 ** case that they are sequential and adjacent. This invariant arises
7354 ** because multiple overflows can only occurs when inserting divider
7355 ** cells into a parent on a prior balance, and divider cells are always
7356 ** adjacent and are inserted in order. There is an assert() tagged
7357 ** with "NOTE 1" in the overflow cell insertion loop to prove this
7360 ** This must be done in advance. Once the balance starts, the cell
7361 ** offset section of the btree page will be overwritten and we will no
7362 ** long be able to find the cells if a pointer to each cell is not saved
7365 memset(&b
.szCell
[b
.nCell
], 0, sizeof(b
.szCell
[0])*(limit
+pOld
->nOverflow
));
7366 if( pOld
->nOverflow
>0 ){
7367 limit
= pOld
->aiOvfl
[0];
7368 for(j
=0; j
<limit
; j
++){
7369 b
.apCell
[b
.nCell
] = aData
+ (maskPage
& get2byteAligned(piCell
));
7373 for(k
=0; k
<pOld
->nOverflow
; k
++){
7374 assert( k
==0 || pOld
->aiOvfl
[k
-1]+1==pOld
->aiOvfl
[k
] );/* NOTE 1 */
7375 b
.apCell
[b
.nCell
] = pOld
->apOvfl
[k
];
7379 piEnd
= aData
+ pOld
->cellOffset
+ 2*pOld
->nCell
;
7380 while( piCell
<piEnd
){
7381 assert( b
.nCell
<nMaxCells
);
7382 b
.apCell
[b
.nCell
] = aData
+ (maskPage
& get2byteAligned(piCell
));
7387 cntOld
[i
] = b
.nCell
;
7388 if( i
<nOld
-1 && !leafData
){
7389 u16 sz
= (u16
)szNew
[i
];
7391 assert( b
.nCell
<nMaxCells
);
7392 b
.szCell
[b
.nCell
] = sz
;
7393 pTemp
= &aSpace1
[iSpace1
];
7395 assert( sz
<=pBt
->maxLocal
+23 );
7396 assert( iSpace1
<= (int)pBt
->pageSize
);
7397 memcpy(pTemp
, apDiv
[i
], sz
);
7398 b
.apCell
[b
.nCell
] = pTemp
+leafCorrection
;
7399 assert( leafCorrection
==0 || leafCorrection
==4 );
7400 b
.szCell
[b
.nCell
] = b
.szCell
[b
.nCell
] - leafCorrection
;
7402 assert( leafCorrection
==0 );
7403 assert( pOld
->hdrOffset
==0 );
7404 /* The right pointer of the child page pOld becomes the left
7405 ** pointer of the divider cell */
7406 memcpy(b
.apCell
[b
.nCell
], &pOld
->aData
[8], 4);
7408 assert( leafCorrection
==4 );
7409 while( b
.szCell
[b
.nCell
]<4 ){
7410 /* Do not allow any cells smaller than 4 bytes. If a smaller cell
7411 ** does exist, pad it with 0x00 bytes. */
7412 assert( b
.szCell
[b
.nCell
]==3 || CORRUPT_DB
);
7413 assert( b
.apCell
[b
.nCell
]==&aSpace1
[iSpace1
-3] || CORRUPT_DB
);
7414 aSpace1
[iSpace1
++] = 0x00;
7415 b
.szCell
[b
.nCell
]++;
7423 ** Figure out the number of pages needed to hold all b.nCell cells.
7424 ** Store this number in "k". Also compute szNew[] which is the total
7425 ** size of all cells on the i-th page and cntNew[] which is the index
7426 ** in b.apCell[] of the cell that divides page i from page i+1.
7427 ** cntNew[k] should equal b.nCell.
7429 ** Values computed by this block:
7431 ** k: The total number of sibling pages
7432 ** szNew[i]: Spaced used on the i-th sibling page.
7433 ** cntNew[i]: Index in b.apCell[] and b.szCell[] for the first cell to
7434 ** the right of the i-th sibling page.
7435 ** usableSpace: Number of bytes of space available on each sibling.
7438 usableSpace
= pBt
->usableSize
- 12 + leafCorrection
;
7439 for(i
=0; i
<nOld
; i
++){
7440 MemPage
*p
= apOld
[i
];
7441 szNew
[i
] = usableSpace
- p
->nFree
;
7442 for(j
=0; j
<p
->nOverflow
; j
++){
7443 szNew
[i
] += 2 + p
->xCellSize(p
, p
->apOvfl
[j
]);
7445 cntNew
[i
] = cntOld
[i
];
7450 while( szNew
[i
]>usableSpace
){
7453 if( k
>NB
+2 ){ rc
= SQLITE_CORRUPT_BKPT
; goto balance_cleanup
; }
7455 cntNew
[k
-1] = b
.nCell
;
7457 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]-1);
7460 if( cntNew
[i
]<b
.nCell
){
7461 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7469 while( cntNew
[i
]<b
.nCell
){
7470 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7471 if( szNew
[i
]+sz
>usableSpace
) break;
7475 if( cntNew
[i
]<b
.nCell
){
7476 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7483 if( cntNew
[i
]>=b
.nCell
){
7485 }else if( cntNew
[i
] <= (i
>0 ? cntNew
[i
-1] : 0) ){
7486 rc
= SQLITE_CORRUPT_BKPT
;
7487 goto balance_cleanup
;
7492 ** The packing computed by the previous block is biased toward the siblings
7493 ** on the left side (siblings with smaller keys). The left siblings are
7494 ** always nearly full, while the right-most sibling might be nearly empty.
7495 ** The next block of code attempts to adjust the packing of siblings to
7496 ** get a better balance.
7498 ** This adjustment is more than an optimization. The packing above might
7499 ** be so out of balance as to be illegal. For example, the right-most
7500 ** sibling might be completely empty. This adjustment is not optional.
7502 for(i
=k
-1; i
>0; i
--){
7503 int szRight
= szNew
[i
]; /* Size of sibling on the right */
7504 int szLeft
= szNew
[i
-1]; /* Size of sibling on the left */
7505 int r
; /* Index of right-most cell in left sibling */
7506 int d
; /* Index of first cell to the left of right sibling */
7508 r
= cntNew
[i
-1] - 1;
7509 d
= r
+ 1 - leafData
;
7510 (void)cachedCellSize(&b
, d
);
7512 assert( d
<nMaxCells
);
7513 assert( r
<nMaxCells
);
7514 (void)cachedCellSize(&b
, r
);
7516 && (bBulk
|| szRight
+b
.szCell
[d
]+2 > szLeft
-(b
.szCell
[r
]+(i
==k
-1?0:2)))){
7519 szRight
+= b
.szCell
[d
] + 2;
7520 szLeft
-= b
.szCell
[r
] + 2;
7526 szNew
[i
-1] = szLeft
;
7527 if( cntNew
[i
-1] <= (i
>1 ? cntNew
[i
-2] : 0) ){
7528 rc
= SQLITE_CORRUPT_BKPT
;
7529 goto balance_cleanup
;
7533 /* Sanity check: For a non-corrupt database file one of the follwing
7535 ** (1) We found one or more cells (cntNew[0])>0), or
7536 ** (2) pPage is a virtual root page. A virtual root page is when
7537 ** the real root page is page 1 and we are the only child of
7540 assert( cntNew
[0]>0 || (pParent
->pgno
==1 && pParent
->nCell
==0) || CORRUPT_DB
);
7541 TRACE(("BALANCE: old: %d(nc=%d) %d(nc=%d) %d(nc=%d)\n",
7542 apOld
[0]->pgno
, apOld
[0]->nCell
,
7543 nOld
>=2 ? apOld
[1]->pgno
: 0, nOld
>=2 ? apOld
[1]->nCell
: 0,
7544 nOld
>=3 ? apOld
[2]->pgno
: 0, nOld
>=3 ? apOld
[2]->nCell
: 0
7548 ** Allocate k new pages. Reuse old pages where possible.
7550 pageFlags
= apOld
[0]->aData
[0];
7554 pNew
= apNew
[i
] = apOld
[i
];
7556 rc
= sqlite3PagerWrite(pNew
->pDbPage
);
7558 if( rc
) goto balance_cleanup
;
7561 rc
= allocateBtreePage(pBt
, &pNew
, &pgno
, (bBulk
? 1 : pgno
), 0);
7562 if( rc
) goto balance_cleanup
;
7563 zeroPage(pNew
, pageFlags
);
7566 cntOld
[i
] = b
.nCell
;
7568 /* Set the pointer-map entry for the new sibling page. */
7570 ptrmapPut(pBt
, pNew
->pgno
, PTRMAP_BTREE
, pParent
->pgno
, &rc
);
7571 if( rc
!=SQLITE_OK
){
7572 goto balance_cleanup
;
7579 ** Reassign page numbers so that the new pages are in ascending order.
7580 ** This helps to keep entries in the disk file in order so that a scan
7581 ** of the table is closer to a linear scan through the file. That in turn
7582 ** helps the operating system to deliver pages from the disk more rapidly.
7584 ** An O(n^2) insertion sort algorithm is used, but since n is never more
7585 ** than (NB+2) (a small constant), that should not be a problem.
7587 ** When NB==3, this one optimization makes the database about 25% faster
7588 ** for large insertions and deletions.
7590 for(i
=0; i
<nNew
; i
++){
7591 aPgOrder
[i
] = aPgno
[i
] = apNew
[i
]->pgno
;
7592 aPgFlags
[i
] = apNew
[i
]->pDbPage
->flags
;
7594 if( aPgno
[j
]==aPgno
[i
] ){
7595 /* This branch is taken if the set of sibling pages somehow contains
7596 ** duplicate entries. This can happen if the database is corrupt.
7597 ** It would be simpler to detect this as part of the loop below, but
7598 ** we do the detection here in order to avoid populating the pager
7599 ** cache with two separate objects associated with the same
7601 assert( CORRUPT_DB
);
7602 rc
= SQLITE_CORRUPT_BKPT
;
7603 goto balance_cleanup
;
7607 for(i
=0; i
<nNew
; i
++){
7608 int iBest
= 0; /* aPgno[] index of page number to use */
7609 for(j
=1; j
<nNew
; j
++){
7610 if( aPgOrder
[j
]<aPgOrder
[iBest
] ) iBest
= j
;
7612 pgno
= aPgOrder
[iBest
];
7613 aPgOrder
[iBest
] = 0xffffffff;
7616 sqlite3PagerRekey(apNew
[iBest
]->pDbPage
, pBt
->nPage
+iBest
+1, 0);
7618 sqlite3PagerRekey(apNew
[i
]->pDbPage
, pgno
, aPgFlags
[iBest
]);
7619 apNew
[i
]->pgno
= pgno
;
7623 TRACE(("BALANCE: new: %d(%d nc=%d) %d(%d nc=%d) %d(%d nc=%d) "
7624 "%d(%d nc=%d) %d(%d nc=%d)\n",
7625 apNew
[0]->pgno
, szNew
[0], cntNew
[0],
7626 nNew
>=2 ? apNew
[1]->pgno
: 0, nNew
>=2 ? szNew
[1] : 0,
7627 nNew
>=2 ? cntNew
[1] - cntNew
[0] - !leafData
: 0,
7628 nNew
>=3 ? apNew
[2]->pgno
: 0, nNew
>=3 ? szNew
[2] : 0,
7629 nNew
>=3 ? cntNew
[2] - cntNew
[1] - !leafData
: 0,
7630 nNew
>=4 ? apNew
[3]->pgno
: 0, nNew
>=4 ? szNew
[3] : 0,
7631 nNew
>=4 ? cntNew
[3] - cntNew
[2] - !leafData
: 0,
7632 nNew
>=5 ? apNew
[4]->pgno
: 0, nNew
>=5 ? szNew
[4] : 0,
7633 nNew
>=5 ? cntNew
[4] - cntNew
[3] - !leafData
: 0
7636 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7637 put4byte(pRight
, apNew
[nNew
-1]->pgno
);
7639 /* If the sibling pages are not leaves, ensure that the right-child pointer
7640 ** of the right-most new sibling page is set to the value that was
7641 ** originally in the same field of the right-most old sibling page. */
7642 if( (pageFlags
& PTF_LEAF
)==0 && nOld
!=nNew
){
7643 MemPage
*pOld
= (nNew
>nOld
? apNew
: apOld
)[nOld
-1];
7644 memcpy(&apNew
[nNew
-1]->aData
[8], &pOld
->aData
[8], 4);
7647 /* Make any required updates to pointer map entries associated with
7648 ** cells stored on sibling pages following the balance operation. Pointer
7649 ** map entries associated with divider cells are set by the insertCell()
7650 ** routine. The associated pointer map entries are:
7652 ** a) if the cell contains a reference to an overflow chain, the
7653 ** entry associated with the first page in the overflow chain, and
7655 ** b) if the sibling pages are not leaves, the child page associated
7658 ** If the sibling pages are not leaves, then the pointer map entry
7659 ** associated with the right-child of each sibling may also need to be
7660 ** updated. This happens below, after the sibling pages have been
7661 ** populated, not here.
7664 MemPage
*pNew
= apNew
[0];
7665 u8
*aOld
= pNew
->aData
;
7666 int cntOldNext
= pNew
->nCell
+ pNew
->nOverflow
;
7667 int usableSize
= pBt
->usableSize
;
7671 for(i
=0; i
<b
.nCell
; i
++){
7672 u8
*pCell
= b
.apCell
[i
];
7673 if( i
==cntOldNext
){
7674 MemPage
*pOld
= (++iOld
)<nNew
? apNew
[iOld
] : apOld
[iOld
];
7675 cntOldNext
+= pOld
->nCell
+ pOld
->nOverflow
+ !leafData
;
7678 if( i
==cntNew
[iNew
] ){
7679 pNew
= apNew
[++iNew
];
7680 if( !leafData
) continue;
7683 /* Cell pCell is destined for new sibling page pNew. Originally, it
7684 ** was either part of sibling page iOld (possibly an overflow cell),
7685 ** or else the divider cell to the left of sibling page iOld. So,
7686 ** if sibling page iOld had the same page number as pNew, and if
7687 ** pCell really was a part of sibling page iOld (not a divider or
7688 ** overflow cell), we can skip updating the pointer map entries. */
7690 || pNew
->pgno
!=aPgno
[iOld
]
7691 || !SQLITE_WITHIN(pCell
,aOld
,&aOld
[usableSize
])
7693 if( !leafCorrection
){
7694 ptrmapPut(pBt
, get4byte(pCell
), PTRMAP_BTREE
, pNew
->pgno
, &rc
);
7696 if( cachedCellSize(&b
,i
)>pNew
->minLocal
){
7697 ptrmapPutOvflPtr(pNew
, pCell
, &rc
);
7699 if( rc
) goto balance_cleanup
;
7704 /* Insert new divider cells into pParent. */
7705 for(i
=0; i
<nNew
-1; i
++){
7709 MemPage
*pNew
= apNew
[i
];
7712 assert( j
<nMaxCells
);
7713 assert( b
.apCell
[j
]!=0 );
7714 pCell
= b
.apCell
[j
];
7715 sz
= b
.szCell
[j
] + leafCorrection
;
7716 pTemp
= &aOvflSpace
[iOvflSpace
];
7718 memcpy(&pNew
->aData
[8], pCell
, 4);
7719 }else if( leafData
){
7720 /* If the tree is a leaf-data tree, and the siblings are leaves,
7721 ** then there is no divider cell in b.apCell[]. Instead, the divider
7722 ** cell consists of the integer key for the right-most cell of
7723 ** the sibling-page assembled above only.
7727 pNew
->xParseCell(pNew
, b
.apCell
[j
], &info
);
7729 sz
= 4 + putVarint(&pCell
[4], info
.nKey
);
7733 /* Obscure case for non-leaf-data trees: If the cell at pCell was
7734 ** previously stored on a leaf node, and its reported size was 4
7735 ** bytes, then it may actually be smaller than this
7736 ** (see btreeParseCellPtr(), 4 bytes is the minimum size of
7737 ** any cell). But it is important to pass the correct size to
7738 ** insertCell(), so reparse the cell now.
7740 ** This can only happen for b-trees used to evaluate "IN (SELECT ...)"
7741 ** and WITHOUT ROWID tables with exactly one column which is the
7744 if( b
.szCell
[j
]==4 ){
7745 assert(leafCorrection
==4);
7746 sz
= pParent
->xCellSize(pParent
, pCell
);
7750 assert( sz
<=pBt
->maxLocal
+23 );
7751 assert( iOvflSpace
<= (int)pBt
->pageSize
);
7752 insertCell(pParent
, nxDiv
+i
, pCell
, sz
, pTemp
, pNew
->pgno
, &rc
);
7753 if( rc
!=SQLITE_OK
) goto balance_cleanup
;
7754 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7757 /* Now update the actual sibling pages. The order in which they are updated
7758 ** is important, as this code needs to avoid disrupting any page from which
7759 ** cells may still to be read. In practice, this means:
7761 ** (1) If cells are moving left (from apNew[iPg] to apNew[iPg-1])
7762 ** then it is not safe to update page apNew[iPg] until after
7763 ** the left-hand sibling apNew[iPg-1] has been updated.
7765 ** (2) If cells are moving right (from apNew[iPg] to apNew[iPg+1])
7766 ** then it is not safe to update page apNew[iPg] until after
7767 ** the right-hand sibling apNew[iPg+1] has been updated.
7769 ** If neither of the above apply, the page is safe to update.
7771 ** The iPg value in the following loop starts at nNew-1 goes down
7772 ** to 0, then back up to nNew-1 again, thus making two passes over
7773 ** the pages. On the initial downward pass, only condition (1) above
7774 ** needs to be tested because (2) will always be true from the previous
7775 ** step. On the upward pass, both conditions are always true, so the
7776 ** upwards pass simply processes pages that were missed on the downward
7779 for(i
=1-nNew
; i
<nNew
; i
++){
7780 int iPg
= i
<0 ? -i
: i
;
7781 assert( iPg
>=0 && iPg
<nNew
);
7782 if( abDone
[iPg
] ) continue; /* Skip pages already processed */
7783 if( i
>=0 /* On the upwards pass, or... */
7784 || cntOld
[iPg
-1]>=cntNew
[iPg
-1] /* Condition (1) is true */
7790 /* Verify condition (1): If cells are moving left, update iPg
7791 ** only after iPg-1 has already been updated. */
7792 assert( iPg
==0 || cntOld
[iPg
-1]>=cntNew
[iPg
-1] || abDone
[iPg
-1] );
7794 /* Verify condition (2): If cells are moving right, update iPg
7795 ** only after iPg+1 has already been updated. */
7796 assert( cntNew
[iPg
]>=cntOld
[iPg
] || abDone
[iPg
+1] );
7800 nNewCell
= cntNew
[0];
7802 iOld
= iPg
<nOld
? (cntOld
[iPg
-1] + !leafData
) : b
.nCell
;
7803 iNew
= cntNew
[iPg
-1] + !leafData
;
7804 nNewCell
= cntNew
[iPg
] - iNew
;
7807 rc
= editPage(apNew
[iPg
], iOld
, iNew
, nNewCell
, &b
);
7808 if( rc
) goto balance_cleanup
;
7810 apNew
[iPg
]->nFree
= usableSpace
-szNew
[iPg
];
7811 assert( apNew
[iPg
]->nOverflow
==0 );
7812 assert( apNew
[iPg
]->nCell
==nNewCell
);
7816 /* All pages have been processed exactly once */
7817 assert( memcmp(abDone
, "\01\01\01\01\01", nNew
)==0 );
7822 if( isRoot
&& pParent
->nCell
==0 && pParent
->hdrOffset
<=apNew
[0]->nFree
){
7823 /* The root page of the b-tree now contains no cells. The only sibling
7824 ** page is the right-child of the parent. Copy the contents of the
7825 ** child page into the parent, decreasing the overall height of the
7826 ** b-tree structure by one. This is described as the "balance-shallower"
7827 ** sub-algorithm in some documentation.
7829 ** If this is an auto-vacuum database, the call to copyNodeContent()
7830 ** sets all pointer-map entries corresponding to database image pages
7831 ** for which the pointer is stored within the content being copied.
7833 ** It is critical that the child page be defragmented before being
7834 ** copied into the parent, because if the parent is page 1 then it will
7835 ** by smaller than the child due to the database header, and so all the
7836 ** free space needs to be up front.
7838 assert( nNew
==1 || CORRUPT_DB
);
7839 rc
= defragmentPage(apNew
[0], -1);
7840 testcase( rc
!=SQLITE_OK
);
7841 assert( apNew
[0]->nFree
==
7842 (get2byte(&apNew
[0]->aData
[5])-apNew
[0]->cellOffset
-apNew
[0]->nCell
*2)
7845 copyNodeContent(apNew
[0], pParent
, &rc
);
7846 freePage(apNew
[0], &rc
);
7847 }else if( ISAUTOVACUUM
&& !leafCorrection
){
7848 /* Fix the pointer map entries associated with the right-child of each
7849 ** sibling page. All other pointer map entries have already been taken
7851 for(i
=0; i
<nNew
; i
++){
7852 u32 key
= get4byte(&apNew
[i
]->aData
[8]);
7853 ptrmapPut(pBt
, key
, PTRMAP_BTREE
, apNew
[i
]->pgno
, &rc
);
7857 assert( pParent
->isInit
);
7858 TRACE(("BALANCE: finished: old=%d new=%d cells=%d\n",
7859 nOld
, nNew
, b
.nCell
));
7861 /* Free any old pages that were not reused as new pages.
7863 for(i
=nNew
; i
<nOld
; i
++){
7864 freePage(apOld
[i
], &rc
);
7868 if( ISAUTOVACUUM
&& rc
==SQLITE_OK
&& apNew
[0]->isInit
){
7869 /* The ptrmapCheckPages() contains assert() statements that verify that
7870 ** all pointer map pages are set correctly. This is helpful while
7871 ** debugging. This is usually disabled because a corrupt database may
7872 ** cause an assert() statement to fail. */
7873 ptrmapCheckPages(apNew
, nNew
);
7874 ptrmapCheckPages(&pParent
, 1);
7879 ** Cleanup before returning.
7882 sqlite3StackFree(0, b
.apCell
);
7883 for(i
=0; i
<nOld
; i
++){
7884 releasePage(apOld
[i
]);
7886 for(i
=0; i
<nNew
; i
++){
7887 releasePage(apNew
[i
]);
7895 ** This function is called when the root page of a b-tree structure is
7896 ** overfull (has one or more overflow pages).
7898 ** A new child page is allocated and the contents of the current root
7899 ** page, including overflow cells, are copied into the child. The root
7900 ** page is then overwritten to make it an empty page with the right-child
7901 ** pointer pointing to the new page.
7903 ** Before returning, all pointer-map entries corresponding to pages
7904 ** that the new child-page now contains pointers to are updated. The
7905 ** entry corresponding to the new right-child pointer of the root
7906 ** page is also updated.
7908 ** If successful, *ppChild is set to contain a reference to the child
7909 ** page and SQLITE_OK is returned. In this case the caller is required
7910 ** to call releasePage() on *ppChild exactly once. If an error occurs,
7911 ** an error code is returned and *ppChild is set to 0.
7913 static int balance_deeper(MemPage
*pRoot
, MemPage
**ppChild
){
7914 int rc
; /* Return value from subprocedures */
7915 MemPage
*pChild
= 0; /* Pointer to a new child page */
7916 Pgno pgnoChild
= 0; /* Page number of the new child page */
7917 BtShared
*pBt
= pRoot
->pBt
; /* The BTree */
7919 assert( pRoot
->nOverflow
>0 );
7920 assert( sqlite3_mutex_held(pBt
->mutex
) );
7922 /* Make pRoot, the root page of the b-tree, writable. Allocate a new
7923 ** page that will become the new right-child of pPage. Copy the contents
7924 ** of the node stored on pRoot into the new child page.
7926 rc
= sqlite3PagerWrite(pRoot
->pDbPage
);
7927 if( rc
==SQLITE_OK
){
7928 rc
= allocateBtreePage(pBt
,&pChild
,&pgnoChild
,pRoot
->pgno
,0);
7929 copyNodeContent(pRoot
, pChild
, &rc
);
7931 ptrmapPut(pBt
, pgnoChild
, PTRMAP_BTREE
, pRoot
->pgno
, &rc
);
7936 releasePage(pChild
);
7939 assert( sqlite3PagerIswriteable(pChild
->pDbPage
) );
7940 assert( sqlite3PagerIswriteable(pRoot
->pDbPage
) );
7941 assert( pChild
->nCell
==pRoot
->nCell
);
7943 TRACE(("BALANCE: copy root %d into %d\n", pRoot
->pgno
, pChild
->pgno
));
7945 /* Copy the overflow cells from pRoot to pChild */
7946 memcpy(pChild
->aiOvfl
, pRoot
->aiOvfl
,
7947 pRoot
->nOverflow
*sizeof(pRoot
->aiOvfl
[0]));
7948 memcpy(pChild
->apOvfl
, pRoot
->apOvfl
,
7949 pRoot
->nOverflow
*sizeof(pRoot
->apOvfl
[0]));
7950 pChild
->nOverflow
= pRoot
->nOverflow
;
7952 /* Zero the contents of pRoot. Then install pChild as the right-child. */
7953 zeroPage(pRoot
, pChild
->aData
[0] & ~PTF_LEAF
);
7954 put4byte(&pRoot
->aData
[pRoot
->hdrOffset
+8], pgnoChild
);
7961 ** The page that pCur currently points to has just been modified in
7962 ** some way. This function figures out if this modification means the
7963 ** tree needs to be balanced, and if so calls the appropriate balancing
7964 ** routine. Balancing routines are:
7968 ** balance_nonroot()
7970 static int balance(BtCursor
*pCur
){
7972 const int nMin
= pCur
->pBt
->usableSize
* 2 / 3;
7973 u8 aBalanceQuickSpace
[13];
7976 VVA_ONLY( int balance_quick_called
= 0 );
7977 VVA_ONLY( int balance_deeper_called
= 0 );
7980 int iPage
= pCur
->iPage
;
7981 MemPage
*pPage
= pCur
->pPage
;
7984 if( pPage
->nOverflow
){
7985 /* The root page of the b-tree is overfull. In this case call the
7986 ** balance_deeper() function to create a new child for the root-page
7987 ** and copy the current contents of the root-page to it. The
7988 ** next iteration of the do-loop will balance the child page.
7990 assert( balance_deeper_called
==0 );
7991 VVA_ONLY( balance_deeper_called
++ );
7992 rc
= balance_deeper(pPage
, &pCur
->apPage
[1]);
7993 if( rc
==SQLITE_OK
){
7997 pCur
->apPage
[0] = pPage
;
7998 pCur
->pPage
= pCur
->apPage
[1];
7999 assert( pCur
->pPage
->nOverflow
);
8004 }else if( pPage
->nOverflow
==0 && pPage
->nFree
<=nMin
){
8007 MemPage
* const pParent
= pCur
->apPage
[iPage
-1];
8008 int const iIdx
= pCur
->aiIdx
[iPage
-1];
8010 rc
= sqlite3PagerWrite(pParent
->pDbPage
);
8011 if( rc
==SQLITE_OK
){
8012 #ifndef SQLITE_OMIT_QUICKBALANCE
8013 if( pPage
->intKeyLeaf
8014 && pPage
->nOverflow
==1
8015 && pPage
->aiOvfl
[0]==pPage
->nCell
8017 && pParent
->nCell
==iIdx
8019 /* Call balance_quick() to create a new sibling of pPage on which
8020 ** to store the overflow cell. balance_quick() inserts a new cell
8021 ** into pParent, which may cause pParent overflow. If this
8022 ** happens, the next iteration of the do-loop will balance pParent
8023 ** use either balance_nonroot() or balance_deeper(). Until this
8024 ** happens, the overflow cell is stored in the aBalanceQuickSpace[]
8027 ** The purpose of the following assert() is to check that only a
8028 ** single call to balance_quick() is made for each call to this
8029 ** function. If this were not verified, a subtle bug involving reuse
8030 ** of the aBalanceQuickSpace[] might sneak in.
8032 assert( balance_quick_called
==0 );
8033 VVA_ONLY( balance_quick_called
++ );
8034 rc
= balance_quick(pParent
, pPage
, aBalanceQuickSpace
);
8038 /* In this case, call balance_nonroot() to redistribute cells
8039 ** between pPage and up to 2 of its sibling pages. This involves
8040 ** modifying the contents of pParent, which may cause pParent to
8041 ** become overfull or underfull. The next iteration of the do-loop
8042 ** will balance the parent page to correct this.
8044 ** If the parent page becomes overfull, the overflow cell or cells
8045 ** are stored in the pSpace buffer allocated immediately below.
8046 ** A subsequent iteration of the do-loop will deal with this by
8047 ** calling balance_nonroot() (balance_deeper() may be called first,
8048 ** but it doesn't deal with overflow cells - just moves them to a
8049 ** different page). Once this subsequent call to balance_nonroot()
8050 ** has completed, it is safe to release the pSpace buffer used by
8051 ** the previous call, as the overflow cell data will have been
8052 ** copied either into the body of a database page or into the new
8053 ** pSpace buffer passed to the latter call to balance_nonroot().
8055 u8
*pSpace
= sqlite3PageMalloc(pCur
->pBt
->pageSize
);
8056 rc
= balance_nonroot(pParent
, iIdx
, pSpace
, iPage
==1,
8057 pCur
->hints
&BTREE_BULKLOAD
);
8059 /* If pFree is not NULL, it points to the pSpace buffer used
8060 ** by a previous call to balance_nonroot(). Its contents are
8061 ** now stored either on real database pages or within the
8062 ** new pSpace buffer, so it may be safely freed here. */
8063 sqlite3PageFree(pFree
);
8066 /* The pSpace buffer will be freed after the next call to
8067 ** balance_nonroot(), or just before this function returns, whichever
8073 pPage
->nOverflow
= 0;
8075 /* The next iteration of the do-loop balances the parent page. */
8078 assert( pCur
->iPage
>=0 );
8079 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
8081 }while( rc
==SQLITE_OK
);
8084 sqlite3PageFree(pFree
);
8091 ** Insert a new record into the BTree. The content of the new record
8092 ** is described by the pX object. The pCur cursor is used only to
8093 ** define what table the record should be inserted into, and is left
8094 ** pointing at a random location.
8096 ** For a table btree (used for rowid tables), only the pX.nKey value of
8097 ** the key is used. The pX.pKey value must be NULL. The pX.nKey is the
8098 ** rowid or INTEGER PRIMARY KEY of the row. The pX.nData,pData,nZero fields
8099 ** hold the content of the row.
8101 ** For an index btree (used for indexes and WITHOUT ROWID tables), the
8102 ** key is an arbitrary byte sequence stored in pX.pKey,nKey. The
8103 ** pX.pData,nData,nZero fields must be zero.
8105 ** If the seekResult parameter is non-zero, then a successful call to
8106 ** MovetoUnpacked() to seek cursor pCur to (pKey,nKey) has already
8107 ** been performed. In other words, if seekResult!=0 then the cursor
8108 ** is currently pointing to a cell that will be adjacent to the cell
8109 ** to be inserted. If seekResult<0 then pCur points to a cell that is
8110 ** smaller then (pKey,nKey). If seekResult>0 then pCur points to a cell
8111 ** that is larger than (pKey,nKey).
8113 ** If seekResult==0, that means pCur is pointing at some unknown location.
8114 ** In that case, this routine must seek the cursor to the correct insertion
8115 ** point for (pKey,nKey) before doing the insertion. For index btrees,
8116 ** if pX->nMem is non-zero, then pX->aMem contains pointers to the unpacked
8117 ** key values and pX->aMem can be used instead of pX->pKey to avoid having
8118 ** to decode the key.
8120 int sqlite3BtreeInsert(
8121 BtCursor
*pCur
, /* Insert data into the table of this cursor */
8122 const BtreePayload
*pX
, /* Content of the row to be inserted */
8123 int flags
, /* True if this is likely an append */
8124 int seekResult
/* Result of prior MovetoUnpacked() call */
8127 int loc
= seekResult
; /* -1: before desired location +1: after */
8131 Btree
*p
= pCur
->pBtree
;
8132 BtShared
*pBt
= p
->pBt
;
8133 unsigned char *oldCell
;
8134 unsigned char *newCell
= 0;
8136 assert( (flags
& (BTREE_SAVEPOSITION
|BTREE_APPEND
))==flags
);
8138 if( pCur
->eState
==CURSOR_FAULT
){
8139 assert( pCur
->skipNext
!=SQLITE_OK
);
8140 return pCur
->skipNext
;
8143 assert( cursorOwnsBtShared(pCur
) );
8144 assert( (pCur
->curFlags
& BTCF_WriteFlag
)!=0
8145 && pBt
->inTransaction
==TRANS_WRITE
8146 && (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
8147 assert( hasSharedCacheTableLock(p
, pCur
->pgnoRoot
, pCur
->pKeyInfo
!=0, 2) );
8149 /* Assert that the caller has been consistent. If this cursor was opened
8150 ** expecting an index b-tree, then the caller should be inserting blob
8151 ** keys with no associated data. If the cursor was opened expecting an
8152 ** intkey table, the caller should be inserting integer keys with a
8153 ** blob of associated data. */
8154 assert( (pX
->pKey
==0)==(pCur
->pKeyInfo
==0) );
8156 /* Save the positions of any other cursors open on this table.
8158 ** In some cases, the call to btreeMoveto() below is a no-op. For
8159 ** example, when inserting data into a table with auto-generated integer
8160 ** keys, the VDBE layer invokes sqlite3BtreeLast() to figure out the
8161 ** integer key to use. It then calls this function to actually insert the
8162 ** data into the intkey B-Tree. In this case btreeMoveto() recognizes
8163 ** that the cursor is already where it needs to be and returns without
8164 ** doing any work. To avoid thwarting these optimizations, it is important
8165 ** not to clear the cursor here.
8167 if( pCur
->curFlags
& BTCF_Multiple
){
8168 rc
= saveAllCursors(pBt
, pCur
->pgnoRoot
, pCur
);
8172 if( pCur
->pKeyInfo
==0 ){
8173 assert( pX
->pKey
==0 );
8174 /* If this is an insert into a table b-tree, invalidate any incrblob
8175 ** cursors open on the row being replaced */
8176 invalidateIncrblobCursors(p
, pCur
->pgnoRoot
, pX
->nKey
, 0);
8178 /* If BTREE_SAVEPOSITION is set, the cursor must already be pointing
8179 ** to a row with the same key as the new entry being inserted. */
8180 assert( (flags
& BTREE_SAVEPOSITION
)==0 ||
8181 ((pCur
->curFlags
&BTCF_ValidNKey
)!=0 && pX
->nKey
==pCur
->info
.nKey
) );
8183 /* If the cursor is currently on the last row and we are appending a
8184 ** new row onto the end, set the "loc" to avoid an unnecessary
8185 ** btreeMoveto() call */
8186 if( (pCur
->curFlags
&BTCF_ValidNKey
)!=0 && pX
->nKey
==pCur
->info
.nKey
){
8189 rc
= sqlite3BtreeMovetoUnpacked(pCur
, 0, pX
->nKey
, flags
!=0, &loc
);
8192 }else if( loc
==0 && (flags
& BTREE_SAVEPOSITION
)==0 ){
8195 r
.pKeyInfo
= pCur
->pKeyInfo
;
8197 r
.nField
= pX
->nMem
;
8203 rc
= sqlite3BtreeMovetoUnpacked(pCur
, &r
, 0, flags
!=0, &loc
);
8205 rc
= btreeMoveto(pCur
, pX
->pKey
, pX
->nKey
, flags
!=0, &loc
);
8209 assert( pCur
->eState
==CURSOR_VALID
|| (pCur
->eState
==CURSOR_INVALID
&& loc
) );
8211 pPage
= pCur
->pPage
;
8212 assert( pPage
->intKey
|| pX
->nKey
>=0 );
8213 assert( pPage
->leaf
|| !pPage
->intKey
);
8215 TRACE(("INSERT: table=%d nkey=%lld ndata=%d page=%d %s\n",
8216 pCur
->pgnoRoot
, pX
->nKey
, pX
->nData
, pPage
->pgno
,
8217 loc
==0 ? "overwrite" : "new entry"));
8218 assert( pPage
->isInit
);
8219 newCell
= pBt
->pTmpSpace
;
8220 assert( newCell
!=0 );
8221 rc
= fillInCell(pPage
, newCell
, pX
, &szNew
);
8222 if( rc
) goto end_insert
;
8223 assert( szNew
==pPage
->xCellSize(pPage
, newCell
) );
8224 assert( szNew
<= MX_CELL_SIZE(pBt
) );
8228 assert( idx
<pPage
->nCell
);
8229 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
8233 oldCell
= findCell(pPage
, idx
);
8235 memcpy(newCell
, oldCell
, 4);
8237 rc
= clearCell(pPage
, oldCell
, &info
);
8238 if( info
.nSize
==szNew
&& info
.nLocal
==info
.nPayload
8239 && (!ISAUTOVACUUM
|| szNew
<pPage
->minLocal
)
8241 /* Overwrite the old cell with the new if they are the same size.
8242 ** We could also try to do this if the old cell is smaller, then add
8243 ** the leftover space to the free list. But experiments show that
8244 ** doing that is no faster then skipping this optimization and just
8245 ** calling dropCell() and insertCell().
8247 ** This optimization cannot be used on an autovacuum database if the
8248 ** new entry uses overflow pages, as the insertCell() call below is
8249 ** necessary to add the PTRMAP_OVERFLOW1 pointer-map entry. */
8250 assert( rc
==SQLITE_OK
); /* clearCell never fails when nLocal==nPayload */
8251 if( oldCell
+szNew
> pPage
->aDataEnd
) return SQLITE_CORRUPT_BKPT
;
8252 memcpy(oldCell
, newCell
, szNew
);
8255 dropCell(pPage
, idx
, info
.nSize
, &rc
);
8256 if( rc
) goto end_insert
;
8257 }else if( loc
<0 && pPage
->nCell
>0 ){
8258 assert( pPage
->leaf
);
8260 pCur
->curFlags
&= ~BTCF_ValidNKey
;
8262 assert( pPage
->leaf
);
8264 insertCell(pPage
, idx
, newCell
, szNew
, 0, 0, &rc
);
8265 assert( pPage
->nOverflow
==0 || rc
==SQLITE_OK
);
8266 assert( rc
!=SQLITE_OK
|| pPage
->nCell
>0 || pPage
->nOverflow
>0 );
8268 /* If no error has occurred and pPage has an overflow cell, call balance()
8269 ** to redistribute the cells within the tree. Since balance() may move
8270 ** the cursor, zero the BtCursor.info.nSize and BTCF_ValidNKey
8273 ** Previous versions of SQLite called moveToRoot() to move the cursor
8274 ** back to the root page as balance() used to invalidate the contents
8275 ** of BtCursor.apPage[] and BtCursor.aiIdx[]. Instead of doing that,
8276 ** set the cursor state to "invalid". This makes common insert operations
8279 ** There is a subtle but important optimization here too. When inserting
8280 ** multiple records into an intkey b-tree using a single cursor (as can
8281 ** happen while processing an "INSERT INTO ... SELECT" statement), it
8282 ** is advantageous to leave the cursor pointing to the last entry in
8283 ** the b-tree if possible. If the cursor is left pointing to the last
8284 ** entry in the table, and the next row inserted has an integer key
8285 ** larger than the largest existing key, it is possible to insert the
8286 ** row without seeking the cursor. This can be a big performance boost.
8288 pCur
->info
.nSize
= 0;
8289 if( pPage
->nOverflow
){
8290 assert( rc
==SQLITE_OK
);
8291 pCur
->curFlags
&= ~(BTCF_ValidNKey
);
8294 /* Must make sure nOverflow is reset to zero even if the balance()
8295 ** fails. Internal data structure corruption will result otherwise.
8296 ** Also, set the cursor state to invalid. This stops saveCursorPosition()
8297 ** from trying to save the current position of the cursor. */
8298 pCur
->pPage
->nOverflow
= 0;
8299 pCur
->eState
= CURSOR_INVALID
;
8300 if( (flags
& BTREE_SAVEPOSITION
) && rc
==SQLITE_OK
){
8301 btreeReleaseAllCursorPages(pCur
);
8302 if( pCur
->pKeyInfo
){
8303 assert( pCur
->pKey
==0 );
8304 pCur
->pKey
= sqlite3Malloc( pX
->nKey
);
8305 if( pCur
->pKey
==0 ){
8308 memcpy(pCur
->pKey
, pX
->pKey
, pX
->nKey
);
8311 pCur
->eState
= CURSOR_REQUIRESEEK
;
8312 pCur
->nKey
= pX
->nKey
;
8315 assert( pCur
->iPage
<0 || pCur
->pPage
->nOverflow
==0 );
8322 ** Delete the entry that the cursor is pointing to.
8324 ** If the BTREE_SAVEPOSITION bit of the flags parameter is zero, then
8325 ** the cursor is left pointing at an arbitrary location after the delete.
8326 ** But if that bit is set, then the cursor is left in a state such that
8327 ** the next call to BtreeNext() or BtreePrev() moves it to the same row
8328 ** as it would have been on if the call to BtreeDelete() had been omitted.
8330 ** The BTREE_AUXDELETE bit of flags indicates that is one of several deletes
8331 ** associated with a single table entry and its indexes. Only one of those
8332 ** deletes is considered the "primary" delete. The primary delete occurs
8333 ** on a cursor that is not a BTREE_FORDELETE cursor. All but one delete
8334 ** operation on non-FORDELETE cursors is tagged with the AUXDELETE flag.
8335 ** The BTREE_AUXDELETE bit is a hint that is not used by this implementation,
8336 ** but which might be used by alternative storage engines.
8338 int sqlite3BtreeDelete(BtCursor
*pCur
, u8 flags
){
8339 Btree
*p
= pCur
->pBtree
;
8340 BtShared
*pBt
= p
->pBt
;
8341 int rc
; /* Return code */
8342 MemPage
*pPage
; /* Page to delete cell from */
8343 unsigned char *pCell
; /* Pointer to cell to delete */
8344 int iCellIdx
; /* Index of cell to delete */
8345 int iCellDepth
; /* Depth of node containing pCell */
8346 CellInfo info
; /* Size of the cell being deleted */
8347 int bSkipnext
= 0; /* Leaf cursor in SKIPNEXT state */
8348 u8 bPreserve
= flags
& BTREE_SAVEPOSITION
; /* Keep cursor valid */
8350 assert( cursorOwnsBtShared(pCur
) );
8351 assert( pBt
->inTransaction
==TRANS_WRITE
);
8352 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
8353 assert( pCur
->curFlags
& BTCF_WriteFlag
);
8354 assert( hasSharedCacheTableLock(p
, pCur
->pgnoRoot
, pCur
->pKeyInfo
!=0, 2) );
8355 assert( !hasReadConflicts(p
, pCur
->pgnoRoot
) );
8356 assert( pCur
->ix
<pCur
->pPage
->nCell
);
8357 assert( pCur
->eState
==CURSOR_VALID
);
8358 assert( (flags
& ~(BTREE_SAVEPOSITION
| BTREE_AUXDELETE
))==0 );
8360 iCellDepth
= pCur
->iPage
;
8361 iCellIdx
= pCur
->ix
;
8362 pPage
= pCur
->pPage
;
8363 pCell
= findCell(pPage
, iCellIdx
);
8365 /* If the bPreserve flag is set to true, then the cursor position must
8366 ** be preserved following this delete operation. If the current delete
8367 ** will cause a b-tree rebalance, then this is done by saving the cursor
8368 ** key and leaving the cursor in CURSOR_REQUIRESEEK state before
8371 ** Or, if the current delete will not cause a rebalance, then the cursor
8372 ** will be left in CURSOR_SKIPNEXT state pointing to the entry immediately
8373 ** before or after the deleted entry. In this case set bSkipnext to true. */
8376 || (pPage
->nFree
+cellSizePtr(pPage
,pCell
)+2)>(int)(pBt
->usableSize
*2/3)
8378 /* A b-tree rebalance will be required after deleting this entry.
8379 ** Save the cursor key. */
8380 rc
= saveCursorKey(pCur
);
8387 /* If the page containing the entry to delete is not a leaf page, move
8388 ** the cursor to the largest entry in the tree that is smaller than
8389 ** the entry being deleted. This cell will replace the cell being deleted
8390 ** from the internal node. The 'previous' entry is used for this instead
8391 ** of the 'next' entry, as the previous entry is always a part of the
8392 ** sub-tree headed by the child page of the cell being deleted. This makes
8393 ** balancing the tree following the delete operation easier. */
8395 rc
= sqlite3BtreePrevious(pCur
, 0);
8396 assert( rc
!=SQLITE_DONE
);
8400 /* Save the positions of any other cursors open on this table before
8401 ** making any modifications. */
8402 if( pCur
->curFlags
& BTCF_Multiple
){
8403 rc
= saveAllCursors(pBt
, pCur
->pgnoRoot
, pCur
);
8407 /* If this is a delete operation to remove a row from a table b-tree,
8408 ** invalidate any incrblob cursors open on the row being deleted. */
8409 if( pCur
->pKeyInfo
==0 ){
8410 invalidateIncrblobCursors(p
, pCur
->pgnoRoot
, pCur
->info
.nKey
, 0);
8413 /* Make the page containing the entry to be deleted writable. Then free any
8414 ** overflow pages associated with the entry and finally remove the cell
8415 ** itself from within the page. */
8416 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
8418 rc
= clearCell(pPage
, pCell
, &info
);
8419 dropCell(pPage
, iCellIdx
, info
.nSize
, &rc
);
8422 /* If the cell deleted was not located on a leaf page, then the cursor
8423 ** is currently pointing to the largest entry in the sub-tree headed
8424 ** by the child-page of the cell that was just deleted from an internal
8425 ** node. The cell from the leaf node needs to be moved to the internal
8426 ** node to replace the deleted cell. */
8428 MemPage
*pLeaf
= pCur
->pPage
;
8431 unsigned char *pTmp
;
8433 if( iCellDepth
<pCur
->iPage
-1 ){
8434 n
= pCur
->apPage
[iCellDepth
+1]->pgno
;
8436 n
= pCur
->pPage
->pgno
;
8438 pCell
= findCell(pLeaf
, pLeaf
->nCell
-1);
8439 if( pCell
<&pLeaf
->aData
[4] ) return SQLITE_CORRUPT_BKPT
;
8440 nCell
= pLeaf
->xCellSize(pLeaf
, pCell
);
8441 assert( MX_CELL_SIZE(pBt
) >= nCell
);
8442 pTmp
= pBt
->pTmpSpace
;
8444 rc
= sqlite3PagerWrite(pLeaf
->pDbPage
);
8445 if( rc
==SQLITE_OK
){
8446 insertCell(pPage
, iCellIdx
, pCell
-4, nCell
+4, pTmp
, n
, &rc
);
8448 dropCell(pLeaf
, pLeaf
->nCell
-1, nCell
, &rc
);
8452 /* Balance the tree. If the entry deleted was located on a leaf page,
8453 ** then the cursor still points to that page. In this case the first
8454 ** call to balance() repairs the tree, and the if(...) condition is
8457 ** Otherwise, if the entry deleted was on an internal node page, then
8458 ** pCur is pointing to the leaf page from which a cell was removed to
8459 ** replace the cell deleted from the internal node. This is slightly
8460 ** tricky as the leaf node may be underfull, and the internal node may
8461 ** be either under or overfull. In this case run the balancing algorithm
8462 ** on the leaf node first. If the balance proceeds far enough up the
8463 ** tree that we can be sure that any problem in the internal node has
8464 ** been corrected, so be it. Otherwise, after balancing the leaf node,
8465 ** walk the cursor up the tree to the internal node and balance it as
8468 if( rc
==SQLITE_OK
&& pCur
->iPage
>iCellDepth
){
8469 releasePageNotNull(pCur
->pPage
);
8471 while( pCur
->iPage
>iCellDepth
){
8472 releasePage(pCur
->apPage
[pCur
->iPage
--]);
8474 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
8478 if( rc
==SQLITE_OK
){
8480 assert( bPreserve
&& (pCur
->iPage
==iCellDepth
|| CORRUPT_DB
) );
8481 assert( pPage
==pCur
->pPage
|| CORRUPT_DB
);
8482 assert( (pPage
->nCell
>0 || CORRUPT_DB
) && iCellIdx
<=pPage
->nCell
);
8483 pCur
->eState
= CURSOR_SKIPNEXT
;
8484 if( iCellIdx
>=pPage
->nCell
){
8485 pCur
->skipNext
= -1;
8486 pCur
->ix
= pPage
->nCell
-1;
8491 rc
= moveToRoot(pCur
);
8493 btreeReleaseAllCursorPages(pCur
);
8494 pCur
->eState
= CURSOR_REQUIRESEEK
;
8496 if( rc
==SQLITE_EMPTY
) rc
= SQLITE_OK
;
8503 ** Create a new BTree table. Write into *piTable the page
8504 ** number for the root page of the new table.
8506 ** The type of type is determined by the flags parameter. Only the
8507 ** following values of flags are currently in use. Other values for
8508 ** flags might not work:
8510 ** BTREE_INTKEY|BTREE_LEAFDATA Used for SQL tables with rowid keys
8511 ** BTREE_ZERODATA Used for SQL indices
8513 static int btreeCreateTable(Btree
*p
, int *piTable
, int createTabFlags
){
8514 BtShared
*pBt
= p
->pBt
;
8518 int ptfFlags
; /* Page-type flage for the root page of new table */
8520 assert( sqlite3BtreeHoldsMutex(p
) );
8521 assert( pBt
->inTransaction
==TRANS_WRITE
);
8522 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
8524 #ifdef SQLITE_OMIT_AUTOVACUUM
8525 rc
= allocateBtreePage(pBt
, &pRoot
, &pgnoRoot
, 1, 0);
8530 if( pBt
->autoVacuum
){
8531 Pgno pgnoMove
; /* Move a page here to make room for the root-page */
8532 MemPage
*pPageMove
; /* The page to move to. */
8534 /* Creating a new table may probably require moving an existing database
8535 ** to make room for the new tables root page. In case this page turns
8536 ** out to be an overflow page, delete all overflow page-map caches
8537 ** held by open cursors.
8539 invalidateAllOverflowCache(pBt
);
8541 /* Read the value of meta[3] from the database to determine where the
8542 ** root page of the new table should go. meta[3] is the largest root-page
8543 ** created so far, so the new root-page is (meta[3]+1).
8545 sqlite3BtreeGetMeta(p
, BTREE_LARGEST_ROOT_PAGE
, &pgnoRoot
);
8548 /* The new root-page may not be allocated on a pointer-map page, or the
8549 ** PENDING_BYTE page.
8551 while( pgnoRoot
==PTRMAP_PAGENO(pBt
, pgnoRoot
) ||
8552 pgnoRoot
==PENDING_BYTE_PAGE(pBt
) ){
8555 assert( pgnoRoot
>=3 || CORRUPT_DB
);
8556 testcase( pgnoRoot
<3 );
8558 /* Allocate a page. The page that currently resides at pgnoRoot will
8559 ** be moved to the allocated page (unless the allocated page happens
8560 ** to reside at pgnoRoot).
8562 rc
= allocateBtreePage(pBt
, &pPageMove
, &pgnoMove
, pgnoRoot
, BTALLOC_EXACT
);
8563 if( rc
!=SQLITE_OK
){
8567 if( pgnoMove
!=pgnoRoot
){
8568 /* pgnoRoot is the page that will be used for the root-page of
8569 ** the new table (assuming an error did not occur). But we were
8570 ** allocated pgnoMove. If required (i.e. if it was not allocated
8571 ** by extending the file), the current page at position pgnoMove
8572 ** is already journaled.
8577 /* Save the positions of any open cursors. This is required in
8578 ** case they are holding a reference to an xFetch reference
8579 ** corresponding to page pgnoRoot. */
8580 rc
= saveAllCursors(pBt
, 0, 0);
8581 releasePage(pPageMove
);
8582 if( rc
!=SQLITE_OK
){
8586 /* Move the page currently at pgnoRoot to pgnoMove. */
8587 rc
= btreeGetPage(pBt
, pgnoRoot
, &pRoot
, 0);
8588 if( rc
!=SQLITE_OK
){
8591 rc
= ptrmapGet(pBt
, pgnoRoot
, &eType
, &iPtrPage
);
8592 if( eType
==PTRMAP_ROOTPAGE
|| eType
==PTRMAP_FREEPAGE
){
8593 rc
= SQLITE_CORRUPT_BKPT
;
8595 if( rc
!=SQLITE_OK
){
8599 assert( eType
!=PTRMAP_ROOTPAGE
);
8600 assert( eType
!=PTRMAP_FREEPAGE
);
8601 rc
= relocatePage(pBt
, pRoot
, eType
, iPtrPage
, pgnoMove
, 0);
8604 /* Obtain the page at pgnoRoot */
8605 if( rc
!=SQLITE_OK
){
8608 rc
= btreeGetPage(pBt
, pgnoRoot
, &pRoot
, 0);
8609 if( rc
!=SQLITE_OK
){
8612 rc
= sqlite3PagerWrite(pRoot
->pDbPage
);
8613 if( rc
!=SQLITE_OK
){
8621 /* Update the pointer-map and meta-data with the new root-page number. */
8622 ptrmapPut(pBt
, pgnoRoot
, PTRMAP_ROOTPAGE
, 0, &rc
);
8628 /* When the new root page was allocated, page 1 was made writable in
8629 ** order either to increase the database filesize, or to decrement the
8630 ** freelist count. Hence, the sqlite3BtreeUpdateMeta() call cannot fail.
8632 assert( sqlite3PagerIswriteable(pBt
->pPage1
->pDbPage
) );
8633 rc
= sqlite3BtreeUpdateMeta(p
, 4, pgnoRoot
);
8640 rc
= allocateBtreePage(pBt
, &pRoot
, &pgnoRoot
, 1, 0);
8644 assert( sqlite3PagerIswriteable(pRoot
->pDbPage
) );
8645 if( createTabFlags
& BTREE_INTKEY
){
8646 ptfFlags
= PTF_INTKEY
| PTF_LEAFDATA
| PTF_LEAF
;
8648 ptfFlags
= PTF_ZERODATA
| PTF_LEAF
;
8650 zeroPage(pRoot
, ptfFlags
);
8651 sqlite3PagerUnref(pRoot
->pDbPage
);
8652 assert( (pBt
->openFlags
& BTREE_SINGLE
)==0 || pgnoRoot
==2 );
8653 *piTable
= (int)pgnoRoot
;
8656 int sqlite3BtreeCreateTable(Btree
*p
, int *piTable
, int flags
){
8658 sqlite3BtreeEnter(p
);
8659 rc
= btreeCreateTable(p
, piTable
, flags
);
8660 sqlite3BtreeLeave(p
);
8665 ** Erase the given database page and all its children. Return
8666 ** the page to the freelist.
8668 static int clearDatabasePage(
8669 BtShared
*pBt
, /* The BTree that contains the table */
8670 Pgno pgno
, /* Page number to clear */
8671 int freePageFlag
, /* Deallocate page if true */
8672 int *pnChange
/* Add number of Cells freed to this counter */
8676 unsigned char *pCell
;
8681 assert( sqlite3_mutex_held(pBt
->mutex
) );
8682 if( pgno
>btreePagecount(pBt
) ){
8683 return SQLITE_CORRUPT_BKPT
;
8685 rc
= getAndInitPage(pBt
, pgno
, &pPage
, 0, 0);
8688 rc
= SQLITE_CORRUPT_BKPT
;
8689 goto cleardatabasepage_out
;
8692 hdr
= pPage
->hdrOffset
;
8693 for(i
=0; i
<pPage
->nCell
; i
++){
8694 pCell
= findCell(pPage
, i
);
8696 rc
= clearDatabasePage(pBt
, get4byte(pCell
), 1, pnChange
);
8697 if( rc
) goto cleardatabasepage_out
;
8699 rc
= clearCell(pPage
, pCell
, &info
);
8700 if( rc
) goto cleardatabasepage_out
;
8703 rc
= clearDatabasePage(pBt
, get4byte(&pPage
->aData
[hdr
+8]), 1, pnChange
);
8704 if( rc
) goto cleardatabasepage_out
;
8705 }else if( pnChange
){
8706 assert( pPage
->intKey
|| CORRUPT_DB
);
8707 testcase( !pPage
->intKey
);
8708 *pnChange
+= pPage
->nCell
;
8711 freePage(pPage
, &rc
);
8712 }else if( (rc
= sqlite3PagerWrite(pPage
->pDbPage
))==0 ){
8713 zeroPage(pPage
, pPage
->aData
[hdr
] | PTF_LEAF
);
8716 cleardatabasepage_out
:
8723 ** Delete all information from a single table in the database. iTable is
8724 ** the page number of the root of the table. After this routine returns,
8725 ** the root page is empty, but still exists.
8727 ** This routine will fail with SQLITE_LOCKED if there are any open
8728 ** read cursors on the table. Open write cursors are moved to the
8729 ** root of the table.
8731 ** If pnChange is not NULL, then table iTable must be an intkey table. The
8732 ** integer value pointed to by pnChange is incremented by the number of
8733 ** entries in the table.
8735 int sqlite3BtreeClearTable(Btree
*p
, int iTable
, int *pnChange
){
8737 BtShared
*pBt
= p
->pBt
;
8738 sqlite3BtreeEnter(p
);
8739 assert( p
->inTrans
==TRANS_WRITE
);
8741 rc
= saveAllCursors(pBt
, (Pgno
)iTable
, 0);
8743 if( SQLITE_OK
==rc
){
8744 /* Invalidate all incrblob cursors open on table iTable (assuming iTable
8745 ** is the root of a table b-tree - if it is not, the following call is
8747 invalidateIncrblobCursors(p
, (Pgno
)iTable
, 0, 1);
8748 rc
= clearDatabasePage(pBt
, (Pgno
)iTable
, 0, pnChange
);
8750 sqlite3BtreeLeave(p
);
8755 ** Delete all information from the single table that pCur is open on.
8757 ** This routine only work for pCur on an ephemeral table.
8759 int sqlite3BtreeClearTableOfCursor(BtCursor
*pCur
){
8760 return sqlite3BtreeClearTable(pCur
->pBtree
, pCur
->pgnoRoot
, 0);
8764 ** Erase all information in a table and add the root of the table to
8765 ** the freelist. Except, the root of the principle table (the one on
8766 ** page 1) is never added to the freelist.
8768 ** This routine will fail with SQLITE_LOCKED if there are any open
8769 ** cursors on the table.
8771 ** If AUTOVACUUM is enabled and the page at iTable is not the last
8772 ** root page in the database file, then the last root page
8773 ** in the database file is moved into the slot formerly occupied by
8774 ** iTable and that last slot formerly occupied by the last root page
8775 ** is added to the freelist instead of iTable. In this say, all
8776 ** root pages are kept at the beginning of the database file, which
8777 ** is necessary for AUTOVACUUM to work right. *piMoved is set to the
8778 ** page number that used to be the last root page in the file before
8779 ** the move. If no page gets moved, *piMoved is set to 0.
8780 ** The last root page is recorded in meta[3] and the value of
8781 ** meta[3] is updated by this procedure.
8783 static int btreeDropTable(Btree
*p
, Pgno iTable
, int *piMoved
){
8786 BtShared
*pBt
= p
->pBt
;
8788 assert( sqlite3BtreeHoldsMutex(p
) );
8789 assert( p
->inTrans
==TRANS_WRITE
);
8790 assert( iTable
>=2 );
8792 rc
= btreeGetPage(pBt
, (Pgno
)iTable
, &pPage
, 0);
8794 rc
= sqlite3BtreeClearTable(p
, iTable
, 0);
8802 #ifdef SQLITE_OMIT_AUTOVACUUM
8803 freePage(pPage
, &rc
);
8806 if( pBt
->autoVacuum
){
8808 sqlite3BtreeGetMeta(p
, BTREE_LARGEST_ROOT_PAGE
, &maxRootPgno
);
8810 if( iTable
==maxRootPgno
){
8811 /* If the table being dropped is the table with the largest root-page
8812 ** number in the database, put the root page on the free list.
8814 freePage(pPage
, &rc
);
8816 if( rc
!=SQLITE_OK
){
8820 /* The table being dropped does not have the largest root-page
8821 ** number in the database. So move the page that does into the
8822 ** gap left by the deleted root-page.
8826 rc
= btreeGetPage(pBt
, maxRootPgno
, &pMove
, 0);
8827 if( rc
!=SQLITE_OK
){
8830 rc
= relocatePage(pBt
, pMove
, PTRMAP_ROOTPAGE
, 0, iTable
, 0);
8832 if( rc
!=SQLITE_OK
){
8836 rc
= btreeGetPage(pBt
, maxRootPgno
, &pMove
, 0);
8837 freePage(pMove
, &rc
);
8839 if( rc
!=SQLITE_OK
){
8842 *piMoved
= maxRootPgno
;
8845 /* Set the new 'max-root-page' value in the database header. This
8846 ** is the old value less one, less one more if that happens to
8847 ** be a root-page number, less one again if that is the
8848 ** PENDING_BYTE_PAGE.
8851 while( maxRootPgno
==PENDING_BYTE_PAGE(pBt
)
8852 || PTRMAP_ISPAGE(pBt
, maxRootPgno
) ){
8855 assert( maxRootPgno
!=PENDING_BYTE_PAGE(pBt
) );
8857 rc
= sqlite3BtreeUpdateMeta(p
, 4, maxRootPgno
);
8859 freePage(pPage
, &rc
);
8865 int sqlite3BtreeDropTable(Btree
*p
, int iTable
, int *piMoved
){
8867 sqlite3BtreeEnter(p
);
8868 rc
= btreeDropTable(p
, iTable
, piMoved
);
8869 sqlite3BtreeLeave(p
);
8875 ** This function may only be called if the b-tree connection already
8876 ** has a read or write transaction open on the database.
8878 ** Read the meta-information out of a database file. Meta[0]
8879 ** is the number of free pages currently in the database. Meta[1]
8880 ** through meta[15] are available for use by higher layers. Meta[0]
8881 ** is read-only, the others are read/write.
8883 ** The schema layer numbers meta values differently. At the schema
8884 ** layer (and the SetCookie and ReadCookie opcodes) the number of
8885 ** free pages is not visible. So Cookie[0] is the same as Meta[1].
8887 ** This routine treats Meta[BTREE_DATA_VERSION] as a special case. Instead
8888 ** of reading the value out of the header, it instead loads the "DataVersion"
8889 ** from the pager. The BTREE_DATA_VERSION value is not actually stored in the
8890 ** database file. It is a number computed by the pager. But its access
8891 ** pattern is the same as header meta values, and so it is convenient to
8892 ** read it from this routine.
8894 void sqlite3BtreeGetMeta(Btree
*p
, int idx
, u32
*pMeta
){
8895 BtShared
*pBt
= p
->pBt
;
8897 sqlite3BtreeEnter(p
);
8898 assert( p
->inTrans
>TRANS_NONE
);
8899 assert( SQLITE_OK
==querySharedCacheTableLock(p
, MASTER_ROOT
, READ_LOCK
) );
8900 assert( pBt
->pPage1
);
8901 assert( idx
>=0 && idx
<=15 );
8903 if( idx
==BTREE_DATA_VERSION
){
8904 *pMeta
= sqlite3PagerDataVersion(pBt
->pPager
) + p
->iDataVersion
;
8906 *pMeta
= get4byte(&pBt
->pPage1
->aData
[36 + idx
*4]);
8909 /* If auto-vacuum is disabled in this build and this is an auto-vacuum
8910 ** database, mark the database as read-only. */
8911 #ifdef SQLITE_OMIT_AUTOVACUUM
8912 if( idx
==BTREE_LARGEST_ROOT_PAGE
&& *pMeta
>0 ){
8913 pBt
->btsFlags
|= BTS_READ_ONLY
;
8917 sqlite3BtreeLeave(p
);
8921 ** Write meta-information back into the database. Meta[0] is
8922 ** read-only and may not be written.
8924 int sqlite3BtreeUpdateMeta(Btree
*p
, int idx
, u32 iMeta
){
8925 BtShared
*pBt
= p
->pBt
;
8928 assert( idx
>=1 && idx
<=15 );
8929 sqlite3BtreeEnter(p
);
8930 assert( p
->inTrans
==TRANS_WRITE
);
8931 assert( pBt
->pPage1
!=0 );
8932 pP1
= pBt
->pPage1
->aData
;
8933 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
8934 if( rc
==SQLITE_OK
){
8935 put4byte(&pP1
[36 + idx
*4], iMeta
);
8936 #ifndef SQLITE_OMIT_AUTOVACUUM
8937 if( idx
==BTREE_INCR_VACUUM
){
8938 assert( pBt
->autoVacuum
|| iMeta
==0 );
8939 assert( iMeta
==0 || iMeta
==1 );
8940 pBt
->incrVacuum
= (u8
)iMeta
;
8944 sqlite3BtreeLeave(p
);
8948 #ifndef SQLITE_OMIT_BTREECOUNT
8950 ** The first argument, pCur, is a cursor opened on some b-tree. Count the
8951 ** number of entries in the b-tree and write the result to *pnEntry.
8953 ** SQLITE_OK is returned if the operation is successfully executed.
8954 ** Otherwise, if an error is encountered (i.e. an IO error or database
8955 ** corruption) an SQLite error code is returned.
8957 int sqlite3BtreeCount(BtCursor
*pCur
, i64
*pnEntry
){
8958 i64 nEntry
= 0; /* Value to return in *pnEntry */
8959 int rc
; /* Return code */
8961 rc
= moveToRoot(pCur
);
8962 if( rc
==SQLITE_EMPTY
){
8967 /* Unless an error occurs, the following loop runs one iteration for each
8968 ** page in the B-Tree structure (not including overflow pages).
8970 while( rc
==SQLITE_OK
){
8971 int iIdx
; /* Index of child node in parent */
8972 MemPage
*pPage
; /* Current page of the b-tree */
8974 /* If this is a leaf page or the tree is not an int-key tree, then
8975 ** this page contains countable entries. Increment the entry counter
8978 pPage
= pCur
->pPage
;
8979 if( pPage
->leaf
|| !pPage
->intKey
){
8980 nEntry
+= pPage
->nCell
;
8983 /* pPage is a leaf node. This loop navigates the cursor so that it
8984 ** points to the first interior cell that it points to the parent of
8985 ** the next page in the tree that has not yet been visited. The
8986 ** pCur->aiIdx[pCur->iPage] value is set to the index of the parent cell
8987 ** of the page, or to the number of cells in the page if the next page
8988 ** to visit is the right-child of its parent.
8990 ** If all pages in the tree have been visited, return SQLITE_OK to the
8995 if( pCur
->iPage
==0 ){
8996 /* All pages of the b-tree have been visited. Return successfully. */
8998 return moveToRoot(pCur
);
9001 }while ( pCur
->ix
>=pCur
->pPage
->nCell
);
9004 pPage
= pCur
->pPage
;
9007 /* Descend to the child node of the cell that the cursor currently
9008 ** points at. This is the right-child if (iIdx==pPage->nCell).
9011 if( iIdx
==pPage
->nCell
){
9012 rc
= moveToChild(pCur
, get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]));
9014 rc
= moveToChild(pCur
, get4byte(findCell(pPage
, iIdx
)));
9018 /* An error has occurred. Return an error code. */
9024 ** Return the pager associated with a BTree. This routine is used for
9025 ** testing and debugging only.
9027 Pager
*sqlite3BtreePager(Btree
*p
){
9028 return p
->pBt
->pPager
;
9031 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9033 ** Append a message to the error message string.
9035 static void checkAppendMsg(
9036 IntegrityCk
*pCheck
,
9037 const char *zFormat
,
9041 if( !pCheck
->mxErr
) return;
9044 va_start(ap
, zFormat
);
9045 if( pCheck
->errMsg
.nChar
){
9046 sqlite3StrAccumAppend(&pCheck
->errMsg
, "\n", 1);
9049 sqlite3XPrintf(&pCheck
->errMsg
, pCheck
->zPfx
, pCheck
->v1
, pCheck
->v2
);
9051 sqlite3VXPrintf(&pCheck
->errMsg
, zFormat
, ap
);
9053 if( pCheck
->errMsg
.accError
==STRACCUM_NOMEM
){
9054 pCheck
->mallocFailed
= 1;
9057 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9059 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9062 ** Return non-zero if the bit in the IntegrityCk.aPgRef[] array that
9063 ** corresponds to page iPg is already set.
9065 static int getPageReferenced(IntegrityCk
*pCheck
, Pgno iPg
){
9066 assert( iPg
<=pCheck
->nPage
&& sizeof(pCheck
->aPgRef
[0])==1 );
9067 return (pCheck
->aPgRef
[iPg
/8] & (1 << (iPg
& 0x07)));
9071 ** Set the bit in the IntegrityCk.aPgRef[] array that corresponds to page iPg.
9073 static void setPageReferenced(IntegrityCk
*pCheck
, Pgno iPg
){
9074 assert( iPg
<=pCheck
->nPage
&& sizeof(pCheck
->aPgRef
[0])==1 );
9075 pCheck
->aPgRef
[iPg
/8] |= (1 << (iPg
& 0x07));
9080 ** Add 1 to the reference count for page iPage. If this is the second
9081 ** reference to the page, add an error message to pCheck->zErrMsg.
9082 ** Return 1 if there are 2 or more references to the page and 0 if
9083 ** if this is the first reference to the page.
9085 ** Also check that the page number is in bounds.
9087 static int checkRef(IntegrityCk
*pCheck
, Pgno iPage
){
9088 if( iPage
==0 ) return 1;
9089 if( iPage
>pCheck
->nPage
){
9090 checkAppendMsg(pCheck
, "invalid page number %d", iPage
);
9093 if( getPageReferenced(pCheck
, iPage
) ){
9094 checkAppendMsg(pCheck
, "2nd reference to page %d", iPage
);
9097 setPageReferenced(pCheck
, iPage
);
9101 #ifndef SQLITE_OMIT_AUTOVACUUM
9103 ** Check that the entry in the pointer-map for page iChild maps to
9104 ** page iParent, pointer type ptrType. If not, append an error message
9107 static void checkPtrmap(
9108 IntegrityCk
*pCheck
, /* Integrity check context */
9109 Pgno iChild
, /* Child page number */
9110 u8 eType
, /* Expected pointer map type */
9111 Pgno iParent
/* Expected pointer map parent page number */
9117 rc
= ptrmapGet(pCheck
->pBt
, iChild
, &ePtrmapType
, &iPtrmapParent
);
9118 if( rc
!=SQLITE_OK
){
9119 if( rc
==SQLITE_NOMEM
|| rc
==SQLITE_IOERR_NOMEM
) pCheck
->mallocFailed
= 1;
9120 checkAppendMsg(pCheck
, "Failed to read ptrmap key=%d", iChild
);
9124 if( ePtrmapType
!=eType
|| iPtrmapParent
!=iParent
){
9125 checkAppendMsg(pCheck
,
9126 "Bad ptr map entry key=%d expected=(%d,%d) got=(%d,%d)",
9127 iChild
, eType
, iParent
, ePtrmapType
, iPtrmapParent
);
9133 ** Check the integrity of the freelist or of an overflow page list.
9134 ** Verify that the number of pages on the list is N.
9136 static void checkList(
9137 IntegrityCk
*pCheck
, /* Integrity checking context */
9138 int isFreeList
, /* True for a freelist. False for overflow page list */
9139 int iPage
, /* Page number for first page in the list */
9140 int N
/* Expected number of pages in the list */
9145 while( N
-- > 0 && pCheck
->mxErr
){
9147 unsigned char *pOvflData
;
9149 checkAppendMsg(pCheck
,
9150 "%d of %d pages missing from overflow list starting at %d",
9151 N
+1, expected
, iFirst
);
9154 if( checkRef(pCheck
, iPage
) ) break;
9155 if( sqlite3PagerGet(pCheck
->pPager
, (Pgno
)iPage
, &pOvflPage
, 0) ){
9156 checkAppendMsg(pCheck
, "failed to get page %d", iPage
);
9159 pOvflData
= (unsigned char *)sqlite3PagerGetData(pOvflPage
);
9161 int n
= get4byte(&pOvflData
[4]);
9162 #ifndef SQLITE_OMIT_AUTOVACUUM
9163 if( pCheck
->pBt
->autoVacuum
){
9164 checkPtrmap(pCheck
, iPage
, PTRMAP_FREEPAGE
, 0);
9167 if( n
>(int)pCheck
->pBt
->usableSize
/4-2 ){
9168 checkAppendMsg(pCheck
,
9169 "freelist leaf count too big on page %d", iPage
);
9173 Pgno iFreePage
= get4byte(&pOvflData
[8+i
*4]);
9174 #ifndef SQLITE_OMIT_AUTOVACUUM
9175 if( pCheck
->pBt
->autoVacuum
){
9176 checkPtrmap(pCheck
, iFreePage
, PTRMAP_FREEPAGE
, 0);
9179 checkRef(pCheck
, iFreePage
);
9184 #ifndef SQLITE_OMIT_AUTOVACUUM
9186 /* If this database supports auto-vacuum and iPage is not the last
9187 ** page in this overflow list, check that the pointer-map entry for
9188 ** the following page matches iPage.
9190 if( pCheck
->pBt
->autoVacuum
&& N
>0 ){
9191 i
= get4byte(pOvflData
);
9192 checkPtrmap(pCheck
, i
, PTRMAP_OVERFLOW2
, iPage
);
9196 iPage
= get4byte(pOvflData
);
9197 sqlite3PagerUnref(pOvflPage
);
9199 if( isFreeList
&& N
<(iPage
!=0) ){
9200 checkAppendMsg(pCheck
, "free-page count in header is too small");
9204 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9207 ** An implementation of a min-heap.
9209 ** aHeap[0] is the number of elements on the heap. aHeap[1] is the
9210 ** root element. The daughter nodes of aHeap[N] are aHeap[N*2]
9211 ** and aHeap[N*2+1].
9213 ** The heap property is this: Every node is less than or equal to both
9214 ** of its daughter nodes. A consequence of the heap property is that the
9215 ** root node aHeap[1] is always the minimum value currently in the heap.
9217 ** The btreeHeapInsert() routine inserts an unsigned 32-bit number onto
9218 ** the heap, preserving the heap property. The btreeHeapPull() routine
9219 ** removes the root element from the heap (the minimum value in the heap)
9220 ** and then moves other nodes around as necessary to preserve the heap
9223 ** This heap is used for cell overlap and coverage testing. Each u32
9224 ** entry represents the span of a cell or freeblock on a btree page.
9225 ** The upper 16 bits are the index of the first byte of a range and the
9226 ** lower 16 bits are the index of the last byte of that range.
9228 static void btreeHeapInsert(u32
*aHeap
, u32 x
){
9229 u32 j
, i
= ++aHeap
[0];
9231 while( (j
= i
/2)>0 && aHeap
[j
]>aHeap
[i
] ){
9233 aHeap
[j
] = aHeap
[i
];
9238 static int btreeHeapPull(u32
*aHeap
, u32
*pOut
){
9240 if( (x
= aHeap
[0])==0 ) return 0;
9242 aHeap
[1] = aHeap
[x
];
9243 aHeap
[x
] = 0xffffffff;
9246 while( (j
= i
*2)<=aHeap
[0] ){
9247 if( aHeap
[j
]>aHeap
[j
+1] ) j
++;
9248 if( aHeap
[i
]<aHeap
[j
] ) break;
9250 aHeap
[i
] = aHeap
[j
];
9257 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9259 ** Do various sanity checks on a single page of a tree. Return
9260 ** the tree depth. Root pages return 0. Parents of root pages
9261 ** return 1, and so forth.
9263 ** These checks are done:
9265 ** 1. Make sure that cells and freeblocks do not overlap
9266 ** but combine to completely cover the page.
9267 ** 2. Make sure integer cell keys are in order.
9268 ** 3. Check the integrity of overflow pages.
9269 ** 4. Recursively call checkTreePage on all children.
9270 ** 5. Verify that the depth of all children is the same.
9272 static int checkTreePage(
9273 IntegrityCk
*pCheck
, /* Context for the sanity check */
9274 int iPage
, /* Page number of the page to check */
9275 i64
*piMinKey
, /* Write minimum integer primary key here */
9276 i64 maxKey
/* Error if integer primary key greater than this */
9278 MemPage
*pPage
= 0; /* The page being analyzed */
9279 int i
; /* Loop counter */
9280 int rc
; /* Result code from subroutine call */
9281 int depth
= -1, d2
; /* Depth of a subtree */
9282 int pgno
; /* Page number */
9283 int nFrag
; /* Number of fragmented bytes on the page */
9284 int hdr
; /* Offset to the page header */
9285 int cellStart
; /* Offset to the start of the cell pointer array */
9286 int nCell
; /* Number of cells */
9287 int doCoverageCheck
= 1; /* True if cell coverage checking should be done */
9288 int keyCanBeEqual
= 1; /* True if IPK can be equal to maxKey
9289 ** False if IPK must be strictly less than maxKey */
9290 u8
*data
; /* Page content */
9291 u8
*pCell
; /* Cell content */
9292 u8
*pCellIdx
; /* Next element of the cell pointer array */
9293 BtShared
*pBt
; /* The BtShared object that owns pPage */
9294 u32 pc
; /* Address of a cell */
9295 u32 usableSize
; /* Usable size of the page */
9296 u32 contentOffset
; /* Offset to the start of the cell content area */
9297 u32
*heap
= 0; /* Min-heap used for checking cell coverage */
9298 u32 x
, prev
= 0; /* Next and previous entry on the min-heap */
9299 const char *saved_zPfx
= pCheck
->zPfx
;
9300 int saved_v1
= pCheck
->v1
;
9301 int saved_v2
= pCheck
->v2
;
9304 /* Check that the page exists
9307 usableSize
= pBt
->usableSize
;
9308 if( iPage
==0 ) return 0;
9309 if( checkRef(pCheck
, iPage
) ) return 0;
9310 pCheck
->zPfx
= "Page %d: ";
9312 if( (rc
= btreeGetPage(pBt
, (Pgno
)iPage
, &pPage
, 0))!=0 ){
9313 checkAppendMsg(pCheck
,
9314 "unable to get the page. error code=%d", rc
);
9318 /* Clear MemPage.isInit to make sure the corruption detection code in
9319 ** btreeInitPage() is executed. */
9320 savedIsInit
= pPage
->isInit
;
9322 if( (rc
= btreeInitPage(pPage
))!=0 ){
9323 assert( rc
==SQLITE_CORRUPT
); /* The only possible error from InitPage */
9324 checkAppendMsg(pCheck
,
9325 "btreeInitPage() returns error code %d", rc
);
9328 data
= pPage
->aData
;
9329 hdr
= pPage
->hdrOffset
;
9331 /* Set up for cell analysis */
9332 pCheck
->zPfx
= "On tree page %d cell %d: ";
9333 contentOffset
= get2byteNotZero(&data
[hdr
+5]);
9334 assert( contentOffset
<=usableSize
); /* Enforced by btreeInitPage() */
9336 /* EVIDENCE-OF: R-37002-32774 The two-byte integer at offset 3 gives the
9337 ** number of cells on the page. */
9338 nCell
= get2byte(&data
[hdr
+3]);
9339 assert( pPage
->nCell
==nCell
);
9341 /* EVIDENCE-OF: R-23882-45353 The cell pointer array of a b-tree page
9342 ** immediately follows the b-tree page header. */
9343 cellStart
= hdr
+ 12 - 4*pPage
->leaf
;
9344 assert( pPage
->aCellIdx
==&data
[cellStart
] );
9345 pCellIdx
= &data
[cellStart
+ 2*(nCell
-1)];
9348 /* Analyze the right-child page of internal pages */
9349 pgno
= get4byte(&data
[hdr
+8]);
9350 #ifndef SQLITE_OMIT_AUTOVACUUM
9351 if( pBt
->autoVacuum
){
9352 pCheck
->zPfx
= "On page %d at right child: ";
9353 checkPtrmap(pCheck
, pgno
, PTRMAP_BTREE
, iPage
);
9356 depth
= checkTreePage(pCheck
, pgno
, &maxKey
, maxKey
);
9359 /* For leaf pages, the coverage check will occur in the same loop
9360 ** as the other cell checks, so initialize the heap. */
9361 heap
= pCheck
->heap
;
9365 /* EVIDENCE-OF: R-02776-14802 The cell pointer array consists of K 2-byte
9366 ** integer offsets to the cell contents. */
9367 for(i
=nCell
-1; i
>=0 && pCheck
->mxErr
; i
--){
9370 /* Check cell size */
9372 assert( pCellIdx
==&data
[cellStart
+ i
*2] );
9373 pc
= get2byteAligned(pCellIdx
);
9375 if( pc
<contentOffset
|| pc
>usableSize
-4 ){
9376 checkAppendMsg(pCheck
, "Offset %d out of range %d..%d",
9377 pc
, contentOffset
, usableSize
-4);
9378 doCoverageCheck
= 0;
9382 pPage
->xParseCell(pPage
, pCell
, &info
);
9383 if( pc
+info
.nSize
>usableSize
){
9384 checkAppendMsg(pCheck
, "Extends off end of page");
9385 doCoverageCheck
= 0;
9389 /* Check for integer primary key out of range */
9390 if( pPage
->intKey
){
9391 if( keyCanBeEqual
? (info
.nKey
> maxKey
) : (info
.nKey
>= maxKey
) ){
9392 checkAppendMsg(pCheck
, "Rowid %lld out of order", info
.nKey
);
9395 keyCanBeEqual
= 0; /* Only the first key on the page may ==maxKey */
9398 /* Check the content overflow list */
9399 if( info
.nPayload
>info
.nLocal
){
9400 int nPage
; /* Number of pages on the overflow chain */
9401 Pgno pgnoOvfl
; /* First page of the overflow chain */
9402 assert( pc
+ info
.nSize
- 4 <= usableSize
);
9403 nPage
= (info
.nPayload
- info
.nLocal
+ usableSize
- 5)/(usableSize
- 4);
9404 pgnoOvfl
= get4byte(&pCell
[info
.nSize
- 4]);
9405 #ifndef SQLITE_OMIT_AUTOVACUUM
9406 if( pBt
->autoVacuum
){
9407 checkPtrmap(pCheck
, pgnoOvfl
, PTRMAP_OVERFLOW1
, iPage
);
9410 checkList(pCheck
, 0, pgnoOvfl
, nPage
);
9414 /* Check sanity of left child page for internal pages */
9415 pgno
= get4byte(pCell
);
9416 #ifndef SQLITE_OMIT_AUTOVACUUM
9417 if( pBt
->autoVacuum
){
9418 checkPtrmap(pCheck
, pgno
, PTRMAP_BTREE
, iPage
);
9421 d2
= checkTreePage(pCheck
, pgno
, &maxKey
, maxKey
);
9424 checkAppendMsg(pCheck
, "Child page depth differs");
9428 /* Populate the coverage-checking heap for leaf pages */
9429 btreeHeapInsert(heap
, (pc
<<16)|(pc
+info
.nSize
-1));
9434 /* Check for complete coverage of the page
9437 if( doCoverageCheck
&& pCheck
->mxErr
>0 ){
9438 /* For leaf pages, the min-heap has already been initialized and the
9439 ** cells have already been inserted. But for internal pages, that has
9440 ** not yet been done, so do it now */
9442 heap
= pCheck
->heap
;
9444 for(i
=nCell
-1; i
>=0; i
--){
9446 pc
= get2byteAligned(&data
[cellStart
+i
*2]);
9447 size
= pPage
->xCellSize(pPage
, &data
[pc
]);
9448 btreeHeapInsert(heap
, (pc
<<16)|(pc
+size
-1));
9451 /* Add the freeblocks to the min-heap
9453 ** EVIDENCE-OF: R-20690-50594 The second field of the b-tree page header
9454 ** is the offset of the first freeblock, or zero if there are no
9455 ** freeblocks on the page.
9457 i
= get2byte(&data
[hdr
+1]);
9460 assert( (u32
)i
<=usableSize
-4 ); /* Enforced by btreeInitPage() */
9461 size
= get2byte(&data
[i
+2]);
9462 assert( (u32
)(i
+size
)<=usableSize
); /* Enforced by btreeInitPage() */
9463 btreeHeapInsert(heap
, (((u32
)i
)<<16)|(i
+size
-1));
9464 /* EVIDENCE-OF: R-58208-19414 The first 2 bytes of a freeblock are a
9465 ** big-endian integer which is the offset in the b-tree page of the next
9466 ** freeblock in the chain, or zero if the freeblock is the last on the
9468 j
= get2byte(&data
[i
]);
9469 /* EVIDENCE-OF: R-06866-39125 Freeblocks are always connected in order of
9470 ** increasing offset. */
9471 assert( j
==0 || j
>i
+size
); /* Enforced by btreeInitPage() */
9472 assert( (u32
)j
<=usableSize
-4 ); /* Enforced by btreeInitPage() */
9475 /* Analyze the min-heap looking for overlap between cells and/or
9476 ** freeblocks, and counting the number of untracked bytes in nFrag.
9478 ** Each min-heap entry is of the form: (start_address<<16)|end_address.
9479 ** There is an implied first entry the covers the page header, the cell
9480 ** pointer index, and the gap between the cell pointer index and the start
9483 ** The loop below pulls entries from the min-heap in order and compares
9484 ** the start_address against the previous end_address. If there is an
9485 ** overlap, that means bytes are used multiple times. If there is a gap,
9486 ** that gap is added to the fragmentation count.
9489 prev
= contentOffset
- 1; /* Implied first min-heap entry */
9490 while( btreeHeapPull(heap
,&x
) ){
9491 if( (prev
&0xffff)>=(x
>>16) ){
9492 checkAppendMsg(pCheck
,
9493 "Multiple uses for byte %u of page %d", x
>>16, iPage
);
9496 nFrag
+= (x
>>16) - (prev
&0xffff) - 1;
9500 nFrag
+= usableSize
- (prev
&0xffff) - 1;
9501 /* EVIDENCE-OF: R-43263-13491 The total number of bytes in all fragments
9502 ** is stored in the fifth field of the b-tree page header.
9503 ** EVIDENCE-OF: R-07161-27322 The one-byte integer at offset 7 gives the
9504 ** number of fragmented free bytes within the cell content area.
9506 if( heap
[0]==0 && nFrag
!=data
[hdr
+7] ){
9507 checkAppendMsg(pCheck
,
9508 "Fragmentation of %d bytes reported as %d on page %d",
9509 nFrag
, data
[hdr
+7], iPage
);
9514 if( !doCoverageCheck
) pPage
->isInit
= savedIsInit
;
9516 pCheck
->zPfx
= saved_zPfx
;
9517 pCheck
->v1
= saved_v1
;
9518 pCheck
->v2
= saved_v2
;
9521 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9523 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9525 ** This routine does a complete check of the given BTree file. aRoot[] is
9526 ** an array of pages numbers were each page number is the root page of
9527 ** a table. nRoot is the number of entries in aRoot.
9529 ** A read-only or read-write transaction must be opened before calling
9532 ** Write the number of error seen in *pnErr. Except for some memory
9533 ** allocation errors, an error message held in memory obtained from
9534 ** malloc is returned if *pnErr is non-zero. If *pnErr==0 then NULL is
9535 ** returned. If a memory allocation error occurs, NULL is returned.
9537 char *sqlite3BtreeIntegrityCheck(
9538 Btree
*p
, /* The btree to be checked */
9539 int *aRoot
, /* An array of root pages numbers for individual trees */
9540 int nRoot
, /* Number of entries in aRoot[] */
9541 int mxErr
, /* Stop reporting errors after this many */
9542 int *pnErr
/* Write number of errors seen to this variable */
9546 BtShared
*pBt
= p
->pBt
;
9547 int savedDbFlags
= pBt
->db
->flags
;
9549 VVA_ONLY( int nRef
);
9551 sqlite3BtreeEnter(p
);
9552 assert( p
->inTrans
>TRANS_NONE
&& pBt
->inTransaction
>TRANS_NONE
);
9553 VVA_ONLY( nRef
= sqlite3PagerRefcount(pBt
->pPager
) );
9556 sCheck
.pPager
= pBt
->pPager
;
9557 sCheck
.nPage
= btreePagecount(sCheck
.pBt
);
9558 sCheck
.mxErr
= mxErr
;
9560 sCheck
.mallocFailed
= 0;
9566 sqlite3StrAccumInit(&sCheck
.errMsg
, 0, zErr
, sizeof(zErr
), SQLITE_MAX_LENGTH
);
9567 sCheck
.errMsg
.printfFlags
= SQLITE_PRINTF_INTERNAL
;
9568 if( sCheck
.nPage
==0 ){
9569 goto integrity_ck_cleanup
;
9572 sCheck
.aPgRef
= sqlite3MallocZero((sCheck
.nPage
/ 8)+ 1);
9573 if( !sCheck
.aPgRef
){
9574 sCheck
.mallocFailed
= 1;
9575 goto integrity_ck_cleanup
;
9577 sCheck
.heap
= (u32
*)sqlite3PageMalloc( pBt
->pageSize
);
9578 if( sCheck
.heap
==0 ){
9579 sCheck
.mallocFailed
= 1;
9580 goto integrity_ck_cleanup
;
9583 i
= PENDING_BYTE_PAGE(pBt
);
9584 if( i
<=sCheck
.nPage
) setPageReferenced(&sCheck
, i
);
9586 /* Check the integrity of the freelist
9588 sCheck
.zPfx
= "Main freelist: ";
9589 checkList(&sCheck
, 1, get4byte(&pBt
->pPage1
->aData
[32]),
9590 get4byte(&pBt
->pPage1
->aData
[36]));
9593 /* Check all the tables.
9595 testcase( pBt
->db
->flags
& SQLITE_CellSizeCk
);
9596 pBt
->db
->flags
&= ~SQLITE_CellSizeCk
;
9597 for(i
=0; (int)i
<nRoot
&& sCheck
.mxErr
; i
++){
9599 if( aRoot
[i
]==0 ) continue;
9600 #ifndef SQLITE_OMIT_AUTOVACUUM
9601 if( pBt
->autoVacuum
&& aRoot
[i
]>1 ){
9602 checkPtrmap(&sCheck
, aRoot
[i
], PTRMAP_ROOTPAGE
, 0);
9605 checkTreePage(&sCheck
, aRoot
[i
], ¬Used
, LARGEST_INT64
);
9607 pBt
->db
->flags
= savedDbFlags
;
9609 /* Make sure every page in the file is referenced
9611 for(i
=1; i
<=sCheck
.nPage
&& sCheck
.mxErr
; i
++){
9612 #ifdef SQLITE_OMIT_AUTOVACUUM
9613 if( getPageReferenced(&sCheck
, i
)==0 ){
9614 checkAppendMsg(&sCheck
, "Page %d is never used", i
);
9617 /* If the database supports auto-vacuum, make sure no tables contain
9618 ** references to pointer-map pages.
9620 if( getPageReferenced(&sCheck
, i
)==0 &&
9621 (PTRMAP_PAGENO(pBt
, i
)!=i
|| !pBt
->autoVacuum
) ){
9622 checkAppendMsg(&sCheck
, "Page %d is never used", i
);
9624 if( getPageReferenced(&sCheck
, i
)!=0 &&
9625 (PTRMAP_PAGENO(pBt
, i
)==i
&& pBt
->autoVacuum
) ){
9626 checkAppendMsg(&sCheck
, "Pointer map page %d is referenced", i
);
9631 /* Clean up and report errors.
9633 integrity_ck_cleanup
:
9634 sqlite3PageFree(sCheck
.heap
);
9635 sqlite3_free(sCheck
.aPgRef
);
9636 if( sCheck
.mallocFailed
){
9637 sqlite3StrAccumReset(&sCheck
.errMsg
);
9640 *pnErr
= sCheck
.nErr
;
9641 if( sCheck
.nErr
==0 ) sqlite3StrAccumReset(&sCheck
.errMsg
);
9642 /* Make sure this analysis did not leave any unref() pages. */
9643 assert( nRef
==sqlite3PagerRefcount(pBt
->pPager
) );
9644 sqlite3BtreeLeave(p
);
9645 return sqlite3StrAccumFinish(&sCheck
.errMsg
);
9647 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9650 ** Return the full pathname of the underlying database file. Return
9651 ** an empty string if the database is in-memory or a TEMP database.
9653 ** The pager filename is invariant as long as the pager is
9654 ** open so it is safe to access without the BtShared mutex.
9656 const char *sqlite3BtreeGetFilename(Btree
*p
){
9657 assert( p
->pBt
->pPager
!=0 );
9658 return sqlite3PagerFilename(p
->pBt
->pPager
, 1);
9662 ** Return the pathname of the journal file for this database. The return
9663 ** value of this routine is the same regardless of whether the journal file
9664 ** has been created or not.
9666 ** The pager journal filename is invariant as long as the pager is
9667 ** open so it is safe to access without the BtShared mutex.
9669 const char *sqlite3BtreeGetJournalname(Btree
*p
){
9670 assert( p
->pBt
->pPager
!=0 );
9671 return sqlite3PagerJournalname(p
->pBt
->pPager
);
9675 ** Return non-zero if a transaction is active.
9677 int sqlite3BtreeIsInTrans(Btree
*p
){
9678 assert( p
==0 || sqlite3_mutex_held(p
->db
->mutex
) );
9679 return (p
&& (p
->inTrans
==TRANS_WRITE
));
9682 #ifndef SQLITE_OMIT_WAL
9684 ** Run a checkpoint on the Btree passed as the first argument.
9686 ** Return SQLITE_LOCKED if this or any other connection has an open
9687 ** transaction on the shared-cache the argument Btree is connected to.
9689 ** Parameter eMode is one of SQLITE_CHECKPOINT_PASSIVE, FULL or RESTART.
9691 int sqlite3BtreeCheckpoint(Btree
*p
, int eMode
, int *pnLog
, int *pnCkpt
){
9694 BtShared
*pBt
= p
->pBt
;
9695 sqlite3BtreeEnter(p
);
9696 if( pBt
->inTransaction
!=TRANS_NONE
){
9699 rc
= sqlite3PagerCheckpoint(pBt
->pPager
, p
->db
, eMode
, pnLog
, pnCkpt
);
9701 sqlite3BtreeLeave(p
);
9708 ** Return non-zero if a read (or write) transaction is active.
9710 int sqlite3BtreeIsInReadTrans(Btree
*p
){
9712 assert( sqlite3_mutex_held(p
->db
->mutex
) );
9713 return p
->inTrans
!=TRANS_NONE
;
9716 int sqlite3BtreeIsInBackup(Btree
*p
){
9718 assert( sqlite3_mutex_held(p
->db
->mutex
) );
9719 return p
->nBackup
!=0;
9723 ** This function returns a pointer to a blob of memory associated with
9724 ** a single shared-btree. The memory is used by client code for its own
9725 ** purposes (for example, to store a high-level schema associated with
9726 ** the shared-btree). The btree layer manages reference counting issues.
9728 ** The first time this is called on a shared-btree, nBytes bytes of memory
9729 ** are allocated, zeroed, and returned to the caller. For each subsequent
9730 ** call the nBytes parameter is ignored and a pointer to the same blob
9731 ** of memory returned.
9733 ** If the nBytes parameter is 0 and the blob of memory has not yet been
9734 ** allocated, a null pointer is returned. If the blob has already been
9735 ** allocated, it is returned as normal.
9737 ** Just before the shared-btree is closed, the function passed as the
9738 ** xFree argument when the memory allocation was made is invoked on the
9739 ** blob of allocated memory. The xFree function should not call sqlite3_free()
9740 ** on the memory, the btree layer does that.
9742 void *sqlite3BtreeSchema(Btree
*p
, int nBytes
, void(*xFree
)(void *)){
9743 BtShared
*pBt
= p
->pBt
;
9744 sqlite3BtreeEnter(p
);
9745 if( !pBt
->pSchema
&& nBytes
){
9746 pBt
->pSchema
= sqlite3DbMallocZero(0, nBytes
);
9747 pBt
->xFreeSchema
= xFree
;
9749 sqlite3BtreeLeave(p
);
9750 return pBt
->pSchema
;
9754 ** Return SQLITE_LOCKED_SHAREDCACHE if another user of the same shared
9755 ** btree as the argument handle holds an exclusive lock on the
9756 ** sqlite_master table. Otherwise SQLITE_OK.
9758 int sqlite3BtreeSchemaLocked(Btree
*p
){
9760 assert( sqlite3_mutex_held(p
->db
->mutex
) );
9761 sqlite3BtreeEnter(p
);
9762 rc
= querySharedCacheTableLock(p
, MASTER_ROOT
, READ_LOCK
);
9763 assert( rc
==SQLITE_OK
|| rc
==SQLITE_LOCKED_SHAREDCACHE
);
9764 sqlite3BtreeLeave(p
);
9769 #ifndef SQLITE_OMIT_SHARED_CACHE
9771 ** Obtain a lock on the table whose root page is iTab. The
9772 ** lock is a write lock if isWritelock is true or a read lock
9775 int sqlite3BtreeLockTable(Btree
*p
, int iTab
, u8 isWriteLock
){
9777 assert( p
->inTrans
!=TRANS_NONE
);
9779 u8 lockType
= READ_LOCK
+ isWriteLock
;
9780 assert( READ_LOCK
+1==WRITE_LOCK
);
9781 assert( isWriteLock
==0 || isWriteLock
==1 );
9783 sqlite3BtreeEnter(p
);
9784 rc
= querySharedCacheTableLock(p
, iTab
, lockType
);
9785 if( rc
==SQLITE_OK
){
9786 rc
= setSharedCacheTableLock(p
, iTab
, lockType
);
9788 sqlite3BtreeLeave(p
);
9794 #ifndef SQLITE_OMIT_INCRBLOB
9796 ** Argument pCsr must be a cursor opened for writing on an
9797 ** INTKEY table currently pointing at a valid table entry.
9798 ** This function modifies the data stored as part of that entry.
9800 ** Only the data content may only be modified, it is not possible to
9801 ** change the length of the data stored. If this function is called with
9802 ** parameters that attempt to write past the end of the existing data,
9803 ** no modifications are made and SQLITE_CORRUPT is returned.
9805 int sqlite3BtreePutData(BtCursor
*pCsr
, u32 offset
, u32 amt
, void *z
){
9807 assert( cursorOwnsBtShared(pCsr
) );
9808 assert( sqlite3_mutex_held(pCsr
->pBtree
->db
->mutex
) );
9809 assert( pCsr
->curFlags
& BTCF_Incrblob
);
9811 rc
= restoreCursorPosition(pCsr
);
9812 if( rc
!=SQLITE_OK
){
9815 assert( pCsr
->eState
!=CURSOR_REQUIRESEEK
);
9816 if( pCsr
->eState
!=CURSOR_VALID
){
9817 return SQLITE_ABORT
;
9820 /* Save the positions of all other cursors open on this table. This is
9821 ** required in case any of them are holding references to an xFetch
9822 ** version of the b-tree page modified by the accessPayload call below.
9824 ** Note that pCsr must be open on a INTKEY table and saveCursorPosition()
9825 ** and hence saveAllCursors() cannot fail on a BTREE_INTKEY table, hence
9826 ** saveAllCursors can only return SQLITE_OK.
9828 VVA_ONLY(rc
=) saveAllCursors(pCsr
->pBt
, pCsr
->pgnoRoot
, pCsr
);
9829 assert( rc
==SQLITE_OK
);
9831 /* Check some assumptions:
9832 ** (a) the cursor is open for writing,
9833 ** (b) there is a read/write transaction open,
9834 ** (c) the connection holds a write-lock on the table (if required),
9835 ** (d) there are no conflicting read-locks, and
9836 ** (e) the cursor points at a valid row of an intKey table.
9838 if( (pCsr
->curFlags
& BTCF_WriteFlag
)==0 ){
9839 return SQLITE_READONLY
;
9841 assert( (pCsr
->pBt
->btsFlags
& BTS_READ_ONLY
)==0
9842 && pCsr
->pBt
->inTransaction
==TRANS_WRITE
);
9843 assert( hasSharedCacheTableLock(pCsr
->pBtree
, pCsr
->pgnoRoot
, 0, 2) );
9844 assert( !hasReadConflicts(pCsr
->pBtree
, pCsr
->pgnoRoot
) );
9845 assert( pCsr
->pPage
->intKey
);
9847 return accessPayload(pCsr
, offset
, amt
, (unsigned char *)z
, 1);
9851 ** Mark this cursor as an incremental blob cursor.
9853 void sqlite3BtreeIncrblobCursor(BtCursor
*pCur
){
9854 pCur
->curFlags
|= BTCF_Incrblob
;
9855 pCur
->pBtree
->hasIncrblobCur
= 1;
9860 ** Set both the "read version" (single byte at byte offset 18) and
9861 ** "write version" (single byte at byte offset 19) fields in the database
9862 ** header to iVersion.
9864 int sqlite3BtreeSetVersion(Btree
*pBtree
, int iVersion
){
9865 BtShared
*pBt
= pBtree
->pBt
;
9866 int rc
; /* Return code */
9868 assert( iVersion
==1 || iVersion
==2 );
9870 /* If setting the version fields to 1, do not automatically open the
9871 ** WAL connection, even if the version fields are currently set to 2.
9873 pBt
->btsFlags
&= ~BTS_NO_WAL
;
9874 if( iVersion
==1 ) pBt
->btsFlags
|= BTS_NO_WAL
;
9876 rc
= sqlite3BtreeBeginTrans(pBtree
, 0);
9877 if( rc
==SQLITE_OK
){
9878 u8
*aData
= pBt
->pPage1
->aData
;
9879 if( aData
[18]!=(u8
)iVersion
|| aData
[19]!=(u8
)iVersion
){
9880 rc
= sqlite3BtreeBeginTrans(pBtree
, 2);
9881 if( rc
==SQLITE_OK
){
9882 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
9883 if( rc
==SQLITE_OK
){
9884 aData
[18] = (u8
)iVersion
;
9885 aData
[19] = (u8
)iVersion
;
9891 pBt
->btsFlags
&= ~BTS_NO_WAL
;
9896 ** Return true if the cursor has a hint specified. This routine is
9897 ** only used from within assert() statements
9899 int sqlite3BtreeCursorHasHint(BtCursor
*pCsr
, unsigned int mask
){
9900 return (pCsr
->hints
& mask
)!=0;
9904 ** Return true if the given Btree is read-only.
9906 int sqlite3BtreeIsReadonly(Btree
*p
){
9907 return (p
->pBt
->btsFlags
& BTS_READ_ONLY
)!=0;
9911 ** Return the size of the header added to each page by this module.
9913 int sqlite3HeaderSizeBtree(void){ return ROUND8(sizeof(MemPage
)); }
9915 #if !defined(SQLITE_OMIT_SHARED_CACHE)
9917 ** Return true if the Btree passed as the only argument is sharable.
9919 int sqlite3BtreeSharable(Btree
*p
){
9924 ** Return the number of connections to the BtShared object accessed by
9925 ** the Btree handle passed as the only argument. For private caches
9926 ** this is always 1. For shared caches it may be 1 or greater.
9928 int sqlite3BtreeConnectionCount(Btree
*p
){
9929 testcase( p
->sharable
);
9930 return p
->pBt
->nRef
;