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 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1406 assert( iFree
+sz
<=iFree2
); /* Verified by pageFindSlot() */
1407 sz2
= get2byte(&data
[iFree2
+2]);
1408 assert( iFree
+sz
+sz2
+iFree2
-(iFree
+sz
) <= usableSize
);
1409 memmove(&data
[iFree
+sz
+sz2
], &data
[iFree
+sz
], iFree2
-(iFree
+sz
));
1413 assert( cbrk
+(iFree
-top
) <= usableSize
);
1414 memmove(&data
[cbrk
], &data
[top
], iFree
-top
);
1415 for(pAddr
=&data
[cellOffset
]; pAddr
<pEnd
; pAddr
+=2){
1416 pc
= get2byte(pAddr
);
1417 if( pc
<iFree
){ put2byte(pAddr
, pc
+sz
); }
1418 else if( pc
<iFree2
){ put2byte(pAddr
, pc
+sz2
); }
1420 goto defragment_out
;
1426 iCellLast
= usableSize
- 4;
1427 for(i
=0; i
<nCell
; i
++){
1428 u8
*pAddr
; /* The i-th cell pointer */
1429 pAddr
= &data
[cellOffset
+ i
*2];
1430 pc
= get2byte(pAddr
);
1431 testcase( pc
==iCellFirst
);
1432 testcase( pc
==iCellLast
);
1433 /* These conditions have already been verified in btreeInitPage()
1434 ** if PRAGMA cell_size_check=ON.
1436 if( pc
<iCellFirst
|| pc
>iCellLast
){
1437 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1439 assert( pc
>=iCellFirst
&& pc
<=iCellLast
);
1440 size
= pPage
->xCellSize(pPage
, &src
[pc
]);
1442 if( cbrk
<iCellFirst
|| pc
+size
>usableSize
){
1443 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1445 assert( cbrk
+size
<=usableSize
&& cbrk
>=iCellFirst
);
1446 testcase( cbrk
+size
==usableSize
);
1447 testcase( pc
+size
==usableSize
);
1448 put2byte(pAddr
, cbrk
);
1451 if( cbrk
==pc
) continue;
1452 temp
= sqlite3PagerTempSpace(pPage
->pBt
->pPager
);
1453 x
= get2byte(&data
[hdr
+5]);
1454 memcpy(&temp
[x
], &data
[x
], (cbrk
+size
) - x
);
1457 memcpy(&data
[cbrk
], &src
[pc
], size
);
1462 if( data
[hdr
+7]+cbrk
-iCellFirst
!=pPage
->nFree
){
1463 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1465 assert( cbrk
>=iCellFirst
);
1466 put2byte(&data
[hdr
+5], cbrk
);
1469 memset(&data
[iCellFirst
], 0, cbrk
-iCellFirst
);
1470 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1475 ** Search the free-list on page pPg for space to store a cell nByte bytes in
1476 ** size. If one can be found, return a pointer to the space and remove it
1477 ** from the free-list.
1479 ** If no suitable space can be found on the free-list, return NULL.
1481 ** This function may detect corruption within pPg. If corruption is
1482 ** detected then *pRc is set to SQLITE_CORRUPT and NULL is returned.
1484 ** Slots on the free list that are between 1 and 3 bytes larger than nByte
1485 ** will be ignored if adding the extra space to the fragmentation count
1486 ** causes the fragmentation count to exceed 60.
1488 static u8
*pageFindSlot(MemPage
*pPg
, int nByte
, int *pRc
){
1489 const int hdr
= pPg
->hdrOffset
;
1490 u8
* const aData
= pPg
->aData
;
1491 int iAddr
= hdr
+ 1;
1492 int pc
= get2byte(&aData
[iAddr
]);
1494 int usableSize
= pPg
->pBt
->usableSize
;
1495 int size
; /* Size of the free slot */
1498 while( pc
<=usableSize
-4 ){
1499 /* EVIDENCE-OF: R-22710-53328 The third and fourth bytes of each
1500 ** freeblock form a big-endian integer which is the size of the freeblock
1501 ** in bytes, including the 4-byte header. */
1502 size
= get2byte(&aData
[pc
+2]);
1503 if( (x
= size
- nByte
)>=0 ){
1506 if( size
+pc
> usableSize
){
1507 *pRc
= SQLITE_CORRUPT_PGNO(pPg
->pgno
);
1510 /* EVIDENCE-OF: R-11498-58022 In a well-formed b-tree page, the total
1511 ** number of bytes in fragments may not exceed 60. */
1512 if( aData
[hdr
+7]>57 ) return 0;
1514 /* Remove the slot from the free-list. Update the number of
1515 ** fragmented bytes within the page. */
1516 memcpy(&aData
[iAddr
], &aData
[pc
], 2);
1517 aData
[hdr
+7] += (u8
)x
;
1519 /* The slot remains on the free-list. Reduce its size to account
1520 ** for the portion used by the new allocation. */
1521 put2byte(&aData
[pc
+2], x
);
1523 return &aData
[pc
+ x
];
1526 pc
= get2byte(&aData
[pc
]);
1527 if( pc
<iAddr
+size
) break;
1530 *pRc
= SQLITE_CORRUPT_PGNO(pPg
->pgno
);
1537 ** Allocate nByte bytes of space from within the B-Tree page passed
1538 ** as the first argument. Write into *pIdx the index into pPage->aData[]
1539 ** of the first byte of allocated space. Return either SQLITE_OK or
1540 ** an error code (usually SQLITE_CORRUPT).
1542 ** The caller guarantees that there is sufficient space to make the
1543 ** allocation. This routine might need to defragment in order to bring
1544 ** all the space together, however. This routine will avoid using
1545 ** the first two bytes past the cell pointer area since presumably this
1546 ** allocation is being made in order to insert a new cell, so we will
1547 ** also end up needing a new cell pointer.
1549 static int allocateSpace(MemPage
*pPage
, int nByte
, int *pIdx
){
1550 const int hdr
= pPage
->hdrOffset
; /* Local cache of pPage->hdrOffset */
1551 u8
* const data
= pPage
->aData
; /* Local cache of pPage->aData */
1552 int top
; /* First byte of cell content area */
1553 int rc
= SQLITE_OK
; /* Integer return code */
1554 int gap
; /* First byte of gap between cell pointers and cell content */
1556 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1557 assert( pPage
->pBt
);
1558 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1559 assert( nByte
>=0 ); /* Minimum cell size is 4 */
1560 assert( pPage
->nFree
>=nByte
);
1561 assert( pPage
->nOverflow
==0 );
1562 assert( nByte
< (int)(pPage
->pBt
->usableSize
-8) );
1564 assert( pPage
->cellOffset
== hdr
+ 12 - 4*pPage
->leaf
);
1565 gap
= pPage
->cellOffset
+ 2*pPage
->nCell
;
1566 assert( gap
<=65536 );
1567 /* EVIDENCE-OF: R-29356-02391 If the database uses a 65536-byte page size
1568 ** and the reserved space is zero (the usual value for reserved space)
1569 ** then the cell content offset of an empty page wants to be 65536.
1570 ** However, that integer is too large to be stored in a 2-byte unsigned
1571 ** integer, so a value of 0 is used in its place. */
1572 top
= get2byte(&data
[hdr
+5]);
1573 assert( top
<=(int)pPage
->pBt
->usableSize
); /* Prevent by getAndInitPage() */
1575 if( top
==0 && pPage
->pBt
->usableSize
==65536 ){
1578 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1582 /* If there is enough space between gap and top for one more cell pointer
1583 ** array entry offset, and if the freelist is not empty, then search the
1584 ** freelist looking for a free slot big enough to satisfy the request.
1586 testcase( gap
+2==top
);
1587 testcase( gap
+1==top
);
1588 testcase( gap
==top
);
1589 if( (data
[hdr
+2] || data
[hdr
+1]) && gap
+2<=top
){
1590 u8
*pSpace
= pageFindSlot(pPage
, nByte
, &rc
);
1592 assert( pSpace
>=data
&& (pSpace
- data
)<65536 );
1593 *pIdx
= (int)(pSpace
- data
);
1600 /* The request could not be fulfilled using a freelist slot. Check
1601 ** to see if defragmentation is necessary.
1603 testcase( gap
+2+nByte
==top
);
1604 if( gap
+2+nByte
>top
){
1605 assert( pPage
->nCell
>0 || CORRUPT_DB
);
1606 rc
= defragmentPage(pPage
, MIN(4, pPage
->nFree
- (2+nByte
)));
1608 top
= get2byteNotZero(&data
[hdr
+5]);
1609 assert( gap
+2+nByte
<=top
);
1613 /* Allocate memory from the gap in between the cell pointer array
1614 ** and the cell content area. The btreeInitPage() call has already
1615 ** validated the freelist. Given that the freelist is valid, there
1616 ** is no way that the allocation can extend off the end of the page.
1617 ** The assert() below verifies the previous sentence.
1620 put2byte(&data
[hdr
+5], top
);
1621 assert( top
+nByte
<= (int)pPage
->pBt
->usableSize
);
1627 ** Return a section of the pPage->aData to the freelist.
1628 ** The first byte of the new free block is pPage->aData[iStart]
1629 ** and the size of the block is iSize bytes.
1631 ** Adjacent freeblocks are coalesced.
1633 ** Note that even though the freeblock list was checked by btreeInitPage(),
1634 ** that routine will not detect overlap between cells or freeblocks. Nor
1635 ** does it detect cells or freeblocks that encrouch into the reserved bytes
1636 ** at the end of the page. So do additional corruption checks inside this
1637 ** routine and return SQLITE_CORRUPT if any problems are found.
1639 static int freeSpace(MemPage
*pPage
, u16 iStart
, u16 iSize
){
1640 u16 iPtr
; /* Address of ptr to next freeblock */
1641 u16 iFreeBlk
; /* Address of the next freeblock */
1642 u8 hdr
; /* Page header size. 0 or 100 */
1643 u8 nFrag
= 0; /* Reduction in fragmentation */
1644 u16 iOrigSize
= iSize
; /* Original value of iSize */
1645 u16 x
; /* Offset to cell content area */
1646 u32 iEnd
= iStart
+ iSize
; /* First byte past the iStart buffer */
1647 unsigned char *data
= pPage
->aData
; /* Page content */
1649 assert( pPage
->pBt
!=0 );
1650 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1651 assert( CORRUPT_DB
|| iStart
>=pPage
->hdrOffset
+6+pPage
->childPtrSize
);
1652 assert( CORRUPT_DB
|| iEnd
<= pPage
->pBt
->usableSize
);
1653 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1654 assert( iSize
>=4 ); /* Minimum cell size is 4 */
1655 assert( iStart
<=pPage
->pBt
->usableSize
-4 );
1657 /* The list of freeblocks must be in ascending order. Find the
1658 ** spot on the list where iStart should be inserted.
1660 hdr
= pPage
->hdrOffset
;
1662 if( data
[iPtr
+1]==0 && data
[iPtr
]==0 ){
1663 iFreeBlk
= 0; /* Shortcut for the case when the freelist is empty */
1665 while( (iFreeBlk
= get2byte(&data
[iPtr
]))<iStart
){
1666 if( iFreeBlk
<iPtr
+4 ){
1667 if( iFreeBlk
==0 ) break;
1668 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1672 if( iFreeBlk
>pPage
->pBt
->usableSize
-4 ){
1673 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1675 assert( iFreeBlk
>iPtr
|| iFreeBlk
==0 );
1678 ** iFreeBlk: First freeblock after iStart, or zero if none
1679 ** iPtr: The address of a pointer to iFreeBlk
1681 ** Check to see if iFreeBlk should be coalesced onto the end of iStart.
1683 if( iFreeBlk
&& iEnd
+3>=iFreeBlk
){
1684 nFrag
= iFreeBlk
- iEnd
;
1685 if( iEnd
>iFreeBlk
) return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1686 iEnd
= iFreeBlk
+ get2byte(&data
[iFreeBlk
+2]);
1687 if( iEnd
> pPage
->pBt
->usableSize
){
1688 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1690 iSize
= iEnd
- iStart
;
1691 iFreeBlk
= get2byte(&data
[iFreeBlk
]);
1694 /* If iPtr is another freeblock (that is, if iPtr is not the freelist
1695 ** pointer in the page header) then check to see if iStart should be
1696 ** coalesced onto the end of iPtr.
1699 int iPtrEnd
= iPtr
+ get2byte(&data
[iPtr
+2]);
1700 if( iPtrEnd
+3>=iStart
){
1701 if( iPtrEnd
>iStart
) return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1702 nFrag
+= iStart
- iPtrEnd
;
1703 iSize
= iEnd
- iPtr
;
1707 if( nFrag
>data
[hdr
+7] ) return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1708 data
[hdr
+7] -= nFrag
;
1710 x
= get2byte(&data
[hdr
+5]);
1712 /* The new freeblock is at the beginning of the cell content area,
1713 ** so just extend the cell content area rather than create another
1714 ** freelist entry */
1715 if( iStart
<x
|| iPtr
!=hdr
+1 ) return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1716 put2byte(&data
[hdr
+1], iFreeBlk
);
1717 put2byte(&data
[hdr
+5], iEnd
);
1719 /* Insert the new freeblock into the freelist */
1720 put2byte(&data
[iPtr
], iStart
);
1722 if( pPage
->pBt
->btsFlags
& BTS_FAST_SECURE
){
1723 /* Overwrite deleted information with zeros when the secure_delete
1724 ** option is enabled */
1725 memset(&data
[iStart
], 0, iSize
);
1727 put2byte(&data
[iStart
], iFreeBlk
);
1728 put2byte(&data
[iStart
+2], iSize
);
1729 pPage
->nFree
+= iOrigSize
;
1734 ** Decode the flags byte (the first byte of the header) for a page
1735 ** and initialize fields of the MemPage structure accordingly.
1737 ** Only the following combinations are supported. Anything different
1738 ** indicates a corrupt database files:
1741 ** PTF_ZERODATA | PTF_LEAF
1742 ** PTF_LEAFDATA | PTF_INTKEY
1743 ** PTF_LEAFDATA | PTF_INTKEY | PTF_LEAF
1745 static int decodeFlags(MemPage
*pPage
, int flagByte
){
1746 BtShared
*pBt
; /* A copy of pPage->pBt */
1748 assert( pPage
->hdrOffset
==(pPage
->pgno
==1 ? 100 : 0) );
1749 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1750 pPage
->leaf
= (u8
)(flagByte
>>3); assert( PTF_LEAF
== 1<<3 );
1751 flagByte
&= ~PTF_LEAF
;
1752 pPage
->childPtrSize
= 4-4*pPage
->leaf
;
1753 pPage
->xCellSize
= cellSizePtr
;
1755 if( flagByte
==(PTF_LEAFDATA
| PTF_INTKEY
) ){
1756 /* EVIDENCE-OF: R-07291-35328 A value of 5 (0x05) means the page is an
1757 ** interior table b-tree page. */
1758 assert( (PTF_LEAFDATA
|PTF_INTKEY
)==5 );
1759 /* EVIDENCE-OF: R-26900-09176 A value of 13 (0x0d) means the page is a
1760 ** leaf table b-tree page. */
1761 assert( (PTF_LEAFDATA
|PTF_INTKEY
|PTF_LEAF
)==13 );
1764 pPage
->intKeyLeaf
= 1;
1765 pPage
->xParseCell
= btreeParseCellPtr
;
1767 pPage
->intKeyLeaf
= 0;
1768 pPage
->xCellSize
= cellSizePtrNoPayload
;
1769 pPage
->xParseCell
= btreeParseCellPtrNoPayload
;
1771 pPage
->maxLocal
= pBt
->maxLeaf
;
1772 pPage
->minLocal
= pBt
->minLeaf
;
1773 }else if( flagByte
==PTF_ZERODATA
){
1774 /* EVIDENCE-OF: R-43316-37308 A value of 2 (0x02) means the page is an
1775 ** interior index b-tree page. */
1776 assert( (PTF_ZERODATA
)==2 );
1777 /* EVIDENCE-OF: R-59615-42828 A value of 10 (0x0a) means the page is a
1778 ** leaf index b-tree page. */
1779 assert( (PTF_ZERODATA
|PTF_LEAF
)==10 );
1781 pPage
->intKeyLeaf
= 0;
1782 pPage
->xParseCell
= btreeParseCellPtrIndex
;
1783 pPage
->maxLocal
= pBt
->maxLocal
;
1784 pPage
->minLocal
= pBt
->minLocal
;
1786 /* EVIDENCE-OF: R-47608-56469 Any other value for the b-tree page type is
1788 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1790 pPage
->max1bytePayload
= pBt
->max1bytePayload
;
1795 ** Initialize the auxiliary information for a disk block.
1797 ** Return SQLITE_OK on success. If we see that the page does
1798 ** not contain a well-formed database page, then return
1799 ** SQLITE_CORRUPT. Note that a return of SQLITE_OK does not
1800 ** guarantee that the page is well-formed. It only shows that
1801 ** we failed to detect any corruption.
1803 static int btreeInitPage(MemPage
*pPage
){
1804 int pc
; /* Address of a freeblock within pPage->aData[] */
1805 u8 hdr
; /* Offset to beginning of page header */
1806 u8
*data
; /* Equal to pPage->aData */
1807 BtShared
*pBt
; /* The main btree structure */
1808 int usableSize
; /* Amount of usable space on each page */
1809 u16 cellOffset
; /* Offset from start of page to first cell pointer */
1810 int nFree
; /* Number of unused bytes on the page */
1811 int top
; /* First byte of the cell content area */
1812 int iCellFirst
; /* First allowable cell or freeblock offset */
1813 int iCellLast
; /* Last possible cell or freeblock offset */
1815 assert( pPage
->pBt
!=0 );
1816 assert( pPage
->pBt
->db
!=0 );
1817 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1818 assert( pPage
->pgno
==sqlite3PagerPagenumber(pPage
->pDbPage
) );
1819 assert( pPage
== sqlite3PagerGetExtra(pPage
->pDbPage
) );
1820 assert( pPage
->aData
== sqlite3PagerGetData(pPage
->pDbPage
) );
1821 assert( pPage
->isInit
==0 );
1824 hdr
= pPage
->hdrOffset
;
1825 data
= pPage
->aData
;
1826 /* EVIDENCE-OF: R-28594-02890 The one-byte flag at offset 0 indicating
1827 ** the b-tree page type. */
1828 if( decodeFlags(pPage
, data
[hdr
]) ){
1829 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1831 assert( pBt
->pageSize
>=512 && pBt
->pageSize
<=65536 );
1832 pPage
->maskPage
= (u16
)(pBt
->pageSize
- 1);
1833 pPage
->nOverflow
= 0;
1834 usableSize
= pBt
->usableSize
;
1835 pPage
->cellOffset
= cellOffset
= hdr
+ 8 + pPage
->childPtrSize
;
1836 pPage
->aDataEnd
= &data
[usableSize
];
1837 pPage
->aCellIdx
= &data
[cellOffset
];
1838 pPage
->aDataOfst
= &data
[pPage
->childPtrSize
];
1839 /* EVIDENCE-OF: R-58015-48175 The two-byte integer at offset 5 designates
1840 ** the start of the cell content area. A zero value for this integer is
1841 ** interpreted as 65536. */
1842 top
= get2byteNotZero(&data
[hdr
+5]);
1843 /* EVIDENCE-OF: R-37002-32774 The two-byte integer at offset 3 gives the
1844 ** number of cells on the page. */
1845 pPage
->nCell
= get2byte(&data
[hdr
+3]);
1846 if( pPage
->nCell
>MX_CELL(pBt
) ){
1847 /* To many cells for a single page. The page must be corrupt */
1848 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1850 testcase( pPage
->nCell
==MX_CELL(pBt
) );
1851 /* EVIDENCE-OF: R-24089-57979 If a page contains no cells (which is only
1852 ** possible for a root page of a table that contains no rows) then the
1853 ** offset to the cell content area will equal the page size minus the
1854 ** bytes of reserved space. */
1855 assert( pPage
->nCell
>0 || top
==usableSize
|| CORRUPT_DB
);
1857 /* A malformed database page might cause us to read past the end
1858 ** of page when parsing a cell.
1860 ** The following block of code checks early to see if a cell extends
1861 ** past the end of a page boundary and causes SQLITE_CORRUPT to be
1862 ** returned if it does.
1864 iCellFirst
= cellOffset
+ 2*pPage
->nCell
;
1865 iCellLast
= usableSize
- 4;
1866 if( pBt
->db
->flags
& SQLITE_CellSizeCk
){
1867 int i
; /* Index into the cell pointer array */
1868 int sz
; /* Size of a cell */
1870 if( !pPage
->leaf
) iCellLast
--;
1871 for(i
=0; i
<pPage
->nCell
; i
++){
1872 pc
= get2byteAligned(&data
[cellOffset
+i
*2]);
1873 testcase( pc
==iCellFirst
);
1874 testcase( pc
==iCellLast
);
1875 if( pc
<iCellFirst
|| pc
>iCellLast
){
1876 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1878 sz
= pPage
->xCellSize(pPage
, &data
[pc
]);
1879 testcase( pc
+sz
==usableSize
);
1880 if( pc
+sz
>usableSize
){
1881 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1884 if( !pPage
->leaf
) iCellLast
++;
1887 /* Compute the total free space on the page
1888 ** EVIDENCE-OF: R-23588-34450 The two-byte integer at offset 1 gives the
1889 ** start of the first freeblock on the page, or is zero if there are no
1891 pc
= get2byte(&data
[hdr
+1]);
1892 nFree
= data
[hdr
+7] + top
; /* Init nFree to non-freeblock free space */
1895 if( pc
<iCellFirst
){
1896 /* EVIDENCE-OF: R-55530-52930 In a well-formed b-tree page, there will
1897 ** always be at least one cell before the first freeblock.
1899 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1903 /* Freeblock off the end of the page */
1904 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1906 next
= get2byte(&data
[pc
]);
1907 size
= get2byte(&data
[pc
+2]);
1908 nFree
= nFree
+ size
;
1909 if( next
<=pc
+size
+3 ) break;
1913 /* Freeblock not in ascending order */
1914 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1916 if( pc
+size
>(unsigned int)usableSize
){
1917 /* Last freeblock extends past page end */
1918 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1922 /* At this point, nFree contains the sum of the offset to the start
1923 ** of the cell-content area plus the number of free bytes within
1924 ** the cell-content area. If this is greater than the usable-size
1925 ** of the page, then the page must be corrupted. This check also
1926 ** serves to verify that the offset to the start of the cell-content
1927 ** area, according to the page header, lies within the page.
1929 if( nFree
>usableSize
){
1930 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
1932 pPage
->nFree
= (u16
)(nFree
- iCellFirst
);
1938 ** Set up a raw page so that it looks like a database page holding
1941 static void zeroPage(MemPage
*pPage
, int flags
){
1942 unsigned char *data
= pPage
->aData
;
1943 BtShared
*pBt
= pPage
->pBt
;
1944 u8 hdr
= pPage
->hdrOffset
;
1947 assert( sqlite3PagerPagenumber(pPage
->pDbPage
)==pPage
->pgno
);
1948 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
1949 assert( sqlite3PagerGetData(pPage
->pDbPage
) == data
);
1950 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1951 assert( sqlite3_mutex_held(pBt
->mutex
) );
1952 if( pBt
->btsFlags
& BTS_FAST_SECURE
){
1953 memset(&data
[hdr
], 0, pBt
->usableSize
- hdr
);
1955 data
[hdr
] = (char)flags
;
1956 first
= hdr
+ ((flags
&PTF_LEAF
)==0 ? 12 : 8);
1957 memset(&data
[hdr
+1], 0, 4);
1959 put2byte(&data
[hdr
+5], pBt
->usableSize
);
1960 pPage
->nFree
= (u16
)(pBt
->usableSize
- first
);
1961 decodeFlags(pPage
, flags
);
1962 pPage
->cellOffset
= first
;
1963 pPage
->aDataEnd
= &data
[pBt
->usableSize
];
1964 pPage
->aCellIdx
= &data
[first
];
1965 pPage
->aDataOfst
= &data
[pPage
->childPtrSize
];
1966 pPage
->nOverflow
= 0;
1967 assert( pBt
->pageSize
>=512 && pBt
->pageSize
<=65536 );
1968 pPage
->maskPage
= (u16
)(pBt
->pageSize
- 1);
1975 ** Convert a DbPage obtained from the pager into a MemPage used by
1978 static MemPage
*btreePageFromDbPage(DbPage
*pDbPage
, Pgno pgno
, BtShared
*pBt
){
1979 MemPage
*pPage
= (MemPage
*)sqlite3PagerGetExtra(pDbPage
);
1980 if( pgno
!=pPage
->pgno
){
1981 pPage
->aData
= sqlite3PagerGetData(pDbPage
);
1982 pPage
->pDbPage
= pDbPage
;
1985 pPage
->hdrOffset
= pgno
==1 ? 100 : 0;
1987 assert( pPage
->aData
==sqlite3PagerGetData(pDbPage
) );
1992 ** Get a page from the pager. Initialize the MemPage.pBt and
1993 ** MemPage.aData elements if needed. See also: btreeGetUnusedPage().
1995 ** If the PAGER_GET_NOCONTENT flag is set, it means that we do not care
1996 ** about the content of the page at this time. So do not go to the disk
1997 ** to fetch the content. Just fill in the content with zeros for now.
1998 ** If in the future we call sqlite3PagerWrite() on this page, that
1999 ** means we have started to be concerned about content and the disk
2000 ** read should occur at that point.
2002 static int btreeGetPage(
2003 BtShared
*pBt
, /* The btree */
2004 Pgno pgno
, /* Number of the page to fetch */
2005 MemPage
**ppPage
, /* Return the page in this parameter */
2006 int flags
/* PAGER_GET_NOCONTENT or PAGER_GET_READONLY */
2011 assert( flags
==0 || flags
==PAGER_GET_NOCONTENT
|| flags
==PAGER_GET_READONLY
);
2012 assert( sqlite3_mutex_held(pBt
->mutex
) );
2013 rc
= sqlite3PagerGet(pBt
->pPager
, pgno
, (DbPage
**)&pDbPage
, flags
);
2015 *ppPage
= btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2020 ** Retrieve a page from the pager cache. If the requested page is not
2021 ** already in the pager cache return NULL. Initialize the MemPage.pBt and
2022 ** MemPage.aData elements if needed.
2024 static MemPage
*btreePageLookup(BtShared
*pBt
, Pgno pgno
){
2026 assert( sqlite3_mutex_held(pBt
->mutex
) );
2027 pDbPage
= sqlite3PagerLookup(pBt
->pPager
, pgno
);
2029 return btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2035 ** Return the size of the database file in pages. If there is any kind of
2036 ** error, return ((unsigned int)-1).
2038 static Pgno
btreePagecount(BtShared
*pBt
){
2041 u32
sqlite3BtreeLastPage(Btree
*p
){
2042 assert( sqlite3BtreeHoldsMutex(p
) );
2043 assert( ((p
->pBt
->nPage
)&0x80000000)==0 );
2044 return btreePagecount(p
->pBt
);
2048 ** Get a page from the pager and initialize it.
2050 ** If pCur!=0 then the page is being fetched as part of a moveToChild()
2051 ** call. Do additional sanity checking on the page in this case.
2052 ** And if the fetch fails, this routine must decrement pCur->iPage.
2054 ** The page is fetched as read-write unless pCur is not NULL and is
2055 ** a read-only cursor.
2057 ** If an error occurs, then *ppPage is undefined. It
2058 ** may remain unchanged, or it may be set to an invalid value.
2060 static int getAndInitPage(
2061 BtShared
*pBt
, /* The database file */
2062 Pgno pgno
, /* Number of the page to get */
2063 MemPage
**ppPage
, /* Write the page pointer here */
2064 BtCursor
*pCur
, /* Cursor to receive the page, or NULL */
2065 int bReadOnly
/* True for a read-only page */
2069 assert( sqlite3_mutex_held(pBt
->mutex
) );
2070 assert( pCur
==0 || ppPage
==&pCur
->pPage
);
2071 assert( pCur
==0 || bReadOnly
==pCur
->curPagerFlags
);
2072 assert( pCur
==0 || pCur
->iPage
>0 );
2074 if( pgno
>btreePagecount(pBt
) ){
2075 rc
= SQLITE_CORRUPT_BKPT
;
2076 goto getAndInitPage_error
;
2078 rc
= sqlite3PagerGet(pBt
->pPager
, pgno
, (DbPage
**)&pDbPage
, bReadOnly
);
2080 goto getAndInitPage_error
;
2082 *ppPage
= (MemPage
*)sqlite3PagerGetExtra(pDbPage
);
2083 if( (*ppPage
)->isInit
==0 ){
2084 btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2085 rc
= btreeInitPage(*ppPage
);
2086 if( rc
!=SQLITE_OK
){
2087 releasePage(*ppPage
);
2088 goto getAndInitPage_error
;
2091 assert( (*ppPage
)->pgno
==pgno
);
2092 assert( (*ppPage
)->aData
==sqlite3PagerGetData(pDbPage
) );
2094 /* If obtaining a child page for a cursor, we must verify that the page is
2095 ** compatible with the root page. */
2096 if( pCur
&& ((*ppPage
)->nCell
<1 || (*ppPage
)->intKey
!=pCur
->curIntKey
) ){
2097 rc
= SQLITE_CORRUPT_PGNO(pgno
);
2098 releasePage(*ppPage
);
2099 goto getAndInitPage_error
;
2103 getAndInitPage_error
:
2106 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
2108 testcase( pgno
==0 );
2109 assert( pgno
!=0 || rc
==SQLITE_CORRUPT
);
2114 ** Release a MemPage. This should be called once for each prior
2115 ** call to btreeGetPage.
2117 ** Page1 is a special case and must be released using releasePageOne().
2119 static void releasePageNotNull(MemPage
*pPage
){
2120 assert( pPage
->aData
);
2121 assert( pPage
->pBt
);
2122 assert( pPage
->pDbPage
!=0 );
2123 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
2124 assert( sqlite3PagerGetData(pPage
->pDbPage
)==pPage
->aData
);
2125 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2126 sqlite3PagerUnrefNotNull(pPage
->pDbPage
);
2128 static void releasePage(MemPage
*pPage
){
2129 if( pPage
) releasePageNotNull(pPage
);
2131 static void releasePageOne(MemPage
*pPage
){
2133 assert( pPage
->aData
);
2134 assert( pPage
->pBt
);
2135 assert( pPage
->pDbPage
!=0 );
2136 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
2137 assert( sqlite3PagerGetData(pPage
->pDbPage
)==pPage
->aData
);
2138 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2139 sqlite3PagerUnrefPageOne(pPage
->pDbPage
);
2143 ** Get an unused page.
2145 ** This works just like btreeGetPage() with the addition:
2147 ** * If the page is already in use for some other purpose, immediately
2148 ** release it and return an SQLITE_CURRUPT error.
2149 ** * Make sure the isInit flag is clear
2151 static int btreeGetUnusedPage(
2152 BtShared
*pBt
, /* The btree */
2153 Pgno pgno
, /* Number of the page to fetch */
2154 MemPage
**ppPage
, /* Return the page in this parameter */
2155 int flags
/* PAGER_GET_NOCONTENT or PAGER_GET_READONLY */
2157 int rc
= btreeGetPage(pBt
, pgno
, ppPage
, flags
);
2158 if( rc
==SQLITE_OK
){
2159 if( sqlite3PagerPageRefcount((*ppPage
)->pDbPage
)>1 ){
2160 releasePage(*ppPage
);
2162 return SQLITE_CORRUPT_BKPT
;
2164 (*ppPage
)->isInit
= 0;
2173 ** During a rollback, when the pager reloads information into the cache
2174 ** so that the cache is restored to its original state at the start of
2175 ** the transaction, for each page restored this routine is called.
2177 ** This routine needs to reset the extra data section at the end of the
2178 ** page to agree with the restored data.
2180 static void pageReinit(DbPage
*pData
){
2182 pPage
= (MemPage
*)sqlite3PagerGetExtra(pData
);
2183 assert( sqlite3PagerPageRefcount(pData
)>0 );
2184 if( pPage
->isInit
){
2185 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2187 if( sqlite3PagerPageRefcount(pData
)>1 ){
2188 /* pPage might not be a btree page; it might be an overflow page
2189 ** or ptrmap page or a free page. In those cases, the following
2190 ** call to btreeInitPage() will likely return SQLITE_CORRUPT.
2191 ** But no harm is done by this. And it is very important that
2192 ** btreeInitPage() be called on every btree page so we make
2193 ** the call for every page that comes in for re-initing. */
2194 btreeInitPage(pPage
);
2200 ** Invoke the busy handler for a btree.
2202 static int btreeInvokeBusyHandler(void *pArg
){
2203 BtShared
*pBt
= (BtShared
*)pArg
;
2205 assert( sqlite3_mutex_held(pBt
->db
->mutex
) );
2206 return sqlite3InvokeBusyHandler(&pBt
->db
->busyHandler
);
2210 ** Open a database file.
2212 ** zFilename is the name of the database file. If zFilename is NULL
2213 ** then an ephemeral database is created. The ephemeral database might
2214 ** be exclusively in memory, or it might use a disk-based memory cache.
2215 ** Either way, the ephemeral database will be automatically deleted
2216 ** when sqlite3BtreeClose() is called.
2218 ** If zFilename is ":memory:" then an in-memory database is created
2219 ** that is automatically destroyed when it is closed.
2221 ** The "flags" parameter is a bitmask that might contain bits like
2222 ** BTREE_OMIT_JOURNAL and/or BTREE_MEMORY.
2224 ** If the database is already opened in the same database connection
2225 ** and we are in shared cache mode, then the open will fail with an
2226 ** SQLITE_CONSTRAINT error. We cannot allow two or more BtShared
2227 ** objects in the same database connection since doing so will lead
2228 ** to problems with locking.
2230 int sqlite3BtreeOpen(
2231 sqlite3_vfs
*pVfs
, /* VFS to use for this b-tree */
2232 const char *zFilename
, /* Name of the file containing the BTree database */
2233 sqlite3
*db
, /* Associated database handle */
2234 Btree
**ppBtree
, /* Pointer to new Btree object written here */
2235 int flags
, /* Options */
2236 int vfsFlags
/* Flags passed through to sqlite3_vfs.xOpen() */
2238 BtShared
*pBt
= 0; /* Shared part of btree structure */
2239 Btree
*p
; /* Handle to return */
2240 sqlite3_mutex
*mutexOpen
= 0; /* Prevents a race condition. Ticket #3537 */
2241 int rc
= SQLITE_OK
; /* Result code from this function */
2242 u8 nReserve
; /* Byte of unused space on each page */
2243 unsigned char zDbHeader
[100]; /* Database header content */
2245 /* True if opening an ephemeral, temporary database */
2246 const int isTempDb
= zFilename
==0 || zFilename
[0]==0;
2248 /* Set the variable isMemdb to true for an in-memory database, or
2249 ** false for a file-based database.
2251 #ifdef SQLITE_OMIT_MEMORYDB
2252 const int isMemdb
= 0;
2254 const int isMemdb
= (zFilename
&& strcmp(zFilename
, ":memory:")==0)
2255 || (isTempDb
&& sqlite3TempInMemory(db
))
2256 || (vfsFlags
& SQLITE_OPEN_MEMORY
)!=0;
2261 assert( sqlite3_mutex_held(db
->mutex
) );
2262 assert( (flags
&0xff)==flags
); /* flags fit in 8 bits */
2264 /* Only a BTREE_SINGLE database can be BTREE_UNORDERED */
2265 assert( (flags
& BTREE_UNORDERED
)==0 || (flags
& BTREE_SINGLE
)!=0 );
2267 /* A BTREE_SINGLE database is always a temporary and/or ephemeral */
2268 assert( (flags
& BTREE_SINGLE
)==0 || isTempDb
);
2271 flags
|= BTREE_MEMORY
;
2273 if( (vfsFlags
& SQLITE_OPEN_MAIN_DB
)!=0 && (isMemdb
|| isTempDb
) ){
2274 vfsFlags
= (vfsFlags
& ~SQLITE_OPEN_MAIN_DB
) | SQLITE_OPEN_TEMP_DB
;
2276 p
= sqlite3MallocZero(sizeof(Btree
));
2278 return SQLITE_NOMEM_BKPT
;
2280 p
->inTrans
= TRANS_NONE
;
2282 #ifndef SQLITE_OMIT_SHARED_CACHE
2287 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2289 ** If this Btree is a candidate for shared cache, try to find an
2290 ** existing BtShared object that we can share with
2292 if( isTempDb
==0 && (isMemdb
==0 || (vfsFlags
&SQLITE_OPEN_URI
)!=0) ){
2293 if( vfsFlags
& SQLITE_OPEN_SHAREDCACHE
){
2294 int nFilename
= sqlite3Strlen30(zFilename
)+1;
2295 int nFullPathname
= pVfs
->mxPathname
+1;
2296 char *zFullPathname
= sqlite3Malloc(MAX(nFullPathname
,nFilename
));
2297 MUTEX_LOGIC( sqlite3_mutex
*mutexShared
; )
2300 if( !zFullPathname
){
2302 return SQLITE_NOMEM_BKPT
;
2305 memcpy(zFullPathname
, zFilename
, nFilename
);
2307 rc
= sqlite3OsFullPathname(pVfs
, zFilename
,
2308 nFullPathname
, zFullPathname
);
2310 sqlite3_free(zFullPathname
);
2315 #if SQLITE_THREADSAFE
2316 mutexOpen
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_OPEN
);
2317 sqlite3_mutex_enter(mutexOpen
);
2318 mutexShared
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER
);
2319 sqlite3_mutex_enter(mutexShared
);
2321 for(pBt
=GLOBAL(BtShared
*,sqlite3SharedCacheList
); pBt
; pBt
=pBt
->pNext
){
2322 assert( pBt
->nRef
>0 );
2323 if( 0==strcmp(zFullPathname
, sqlite3PagerFilename(pBt
->pPager
, 0))
2324 && sqlite3PagerVfs(pBt
->pPager
)==pVfs
){
2326 for(iDb
=db
->nDb
-1; iDb
>=0; iDb
--){
2327 Btree
*pExisting
= db
->aDb
[iDb
].pBt
;
2328 if( pExisting
&& pExisting
->pBt
==pBt
){
2329 sqlite3_mutex_leave(mutexShared
);
2330 sqlite3_mutex_leave(mutexOpen
);
2331 sqlite3_free(zFullPathname
);
2333 return SQLITE_CONSTRAINT
;
2341 sqlite3_mutex_leave(mutexShared
);
2342 sqlite3_free(zFullPathname
);
2346 /* In debug mode, we mark all persistent databases as sharable
2347 ** even when they are not. This exercises the locking code and
2348 ** gives more opportunity for asserts(sqlite3_mutex_held())
2349 ** statements to find locking problems.
2358 ** The following asserts make sure that structures used by the btree are
2359 ** the right size. This is to guard against size changes that result
2360 ** when compiling on a different architecture.
2362 assert( sizeof(i64
)==8 );
2363 assert( sizeof(u64
)==8 );
2364 assert( sizeof(u32
)==4 );
2365 assert( sizeof(u16
)==2 );
2366 assert( sizeof(Pgno
)==4 );
2368 pBt
= sqlite3MallocZero( sizeof(*pBt
) );
2370 rc
= SQLITE_NOMEM_BKPT
;
2371 goto btree_open_out
;
2373 rc
= sqlite3PagerOpen(pVfs
, &pBt
->pPager
, zFilename
,
2374 sizeof(MemPage
), flags
, vfsFlags
, pageReinit
);
2375 if( rc
==SQLITE_OK
){
2376 sqlite3PagerSetMmapLimit(pBt
->pPager
, db
->szMmap
);
2377 rc
= sqlite3PagerReadFileheader(pBt
->pPager
,sizeof(zDbHeader
),zDbHeader
);
2379 if( rc
!=SQLITE_OK
){
2380 goto btree_open_out
;
2382 pBt
->openFlags
= (u8
)flags
;
2384 sqlite3PagerSetBusyhandler(pBt
->pPager
, btreeInvokeBusyHandler
, pBt
);
2389 if( sqlite3PagerIsreadonly(pBt
->pPager
) ) pBt
->btsFlags
|= BTS_READ_ONLY
;
2390 #if defined(SQLITE_SECURE_DELETE)
2391 pBt
->btsFlags
|= BTS_SECURE_DELETE
;
2392 #elif defined(SQLITE_FAST_SECURE_DELETE)
2393 pBt
->btsFlags
|= BTS_OVERWRITE
;
2395 /* EVIDENCE-OF: R-51873-39618 The page size for a database file is
2396 ** determined by the 2-byte integer located at an offset of 16 bytes from
2397 ** the beginning of the database file. */
2398 pBt
->pageSize
= (zDbHeader
[16]<<8) | (zDbHeader
[17]<<16);
2399 if( pBt
->pageSize
<512 || pBt
->pageSize
>SQLITE_MAX_PAGE_SIZE
2400 || ((pBt
->pageSize
-1)&pBt
->pageSize
)!=0 ){
2402 #ifndef SQLITE_OMIT_AUTOVACUUM
2403 /* If the magic name ":memory:" will create an in-memory database, then
2404 ** leave the autoVacuum mode at 0 (do not auto-vacuum), even if
2405 ** SQLITE_DEFAULT_AUTOVACUUM is true. On the other hand, if
2406 ** SQLITE_OMIT_MEMORYDB has been defined, then ":memory:" is just a
2407 ** regular file-name. In this case the auto-vacuum applies as per normal.
2409 if( zFilename
&& !isMemdb
){
2410 pBt
->autoVacuum
= (SQLITE_DEFAULT_AUTOVACUUM
? 1 : 0);
2411 pBt
->incrVacuum
= (SQLITE_DEFAULT_AUTOVACUUM
==2 ? 1 : 0);
2416 /* EVIDENCE-OF: R-37497-42412 The size of the reserved region is
2417 ** determined by the one-byte unsigned integer found at an offset of 20
2418 ** into the database file header. */
2419 nReserve
= zDbHeader
[20];
2420 pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
2421 #ifndef SQLITE_OMIT_AUTOVACUUM
2422 pBt
->autoVacuum
= (get4byte(&zDbHeader
[36 + 4*4])?1:0);
2423 pBt
->incrVacuum
= (get4byte(&zDbHeader
[36 + 7*4])?1:0);
2426 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
, nReserve
);
2427 if( rc
) goto btree_open_out
;
2428 pBt
->usableSize
= pBt
->pageSize
- nReserve
;
2429 assert( (pBt
->pageSize
& 7)==0 ); /* 8-byte alignment of pageSize */
2431 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2432 /* Add the new BtShared object to the linked list sharable BtShareds.
2436 MUTEX_LOGIC( sqlite3_mutex
*mutexShared
; )
2437 MUTEX_LOGIC( mutexShared
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER
);)
2438 if( SQLITE_THREADSAFE
&& sqlite3GlobalConfig
.bCoreMutex
){
2439 pBt
->mutex
= sqlite3MutexAlloc(SQLITE_MUTEX_FAST
);
2440 if( pBt
->mutex
==0 ){
2441 rc
= SQLITE_NOMEM_BKPT
;
2442 goto btree_open_out
;
2445 sqlite3_mutex_enter(mutexShared
);
2446 pBt
->pNext
= GLOBAL(BtShared
*,sqlite3SharedCacheList
);
2447 GLOBAL(BtShared
*,sqlite3SharedCacheList
) = pBt
;
2448 sqlite3_mutex_leave(mutexShared
);
2453 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2454 /* If the new Btree uses a sharable pBtShared, then link the new
2455 ** Btree into the list of all sharable Btrees for the same connection.
2456 ** The list is kept in ascending order by pBt address.
2461 for(i
=0; i
<db
->nDb
; i
++){
2462 if( (pSib
= db
->aDb
[i
].pBt
)!=0 && pSib
->sharable
){
2463 while( pSib
->pPrev
){ pSib
= pSib
->pPrev
; }
2464 if( (uptr
)p
->pBt
<(uptr
)pSib
->pBt
){
2469 while( pSib
->pNext
&& (uptr
)pSib
->pNext
->pBt
<(uptr
)p
->pBt
){
2472 p
->pNext
= pSib
->pNext
;
2475 p
->pNext
->pPrev
= p
;
2487 if( rc
!=SQLITE_OK
){
2488 if( pBt
&& pBt
->pPager
){
2489 sqlite3PagerClose(pBt
->pPager
, 0);
2495 sqlite3_file
*pFile
;
2497 /* If the B-Tree was successfully opened, set the pager-cache size to the
2498 ** default value. Except, when opening on an existing shared pager-cache,
2499 ** do not change the pager-cache size.
2501 if( sqlite3BtreeSchema(p
, 0, 0)==0 ){
2502 sqlite3PagerSetCachesize(p
->pBt
->pPager
, SQLITE_DEFAULT_CACHE_SIZE
);
2505 pFile
= sqlite3PagerFile(pBt
->pPager
);
2506 if( pFile
->pMethods
){
2507 sqlite3OsFileControlHint(pFile
, SQLITE_FCNTL_PDB
, (void*)&pBt
->db
);
2511 assert( sqlite3_mutex_held(mutexOpen
) );
2512 sqlite3_mutex_leave(mutexOpen
);
2514 assert( rc
!=SQLITE_OK
|| sqlite3BtreeConnectionCount(*ppBtree
)>0 );
2519 ** Decrement the BtShared.nRef counter. When it reaches zero,
2520 ** remove the BtShared structure from the sharing list. Return
2521 ** true if the BtShared.nRef counter reaches zero and return
2522 ** false if it is still positive.
2524 static int removeFromSharingList(BtShared
*pBt
){
2525 #ifndef SQLITE_OMIT_SHARED_CACHE
2526 MUTEX_LOGIC( sqlite3_mutex
*pMaster
; )
2530 assert( sqlite3_mutex_notheld(pBt
->mutex
) );
2531 MUTEX_LOGIC( pMaster
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER
); )
2532 sqlite3_mutex_enter(pMaster
);
2535 if( GLOBAL(BtShared
*,sqlite3SharedCacheList
)==pBt
){
2536 GLOBAL(BtShared
*,sqlite3SharedCacheList
) = pBt
->pNext
;
2538 pList
= GLOBAL(BtShared
*,sqlite3SharedCacheList
);
2539 while( ALWAYS(pList
) && pList
->pNext
!=pBt
){
2542 if( ALWAYS(pList
) ){
2543 pList
->pNext
= pBt
->pNext
;
2546 if( SQLITE_THREADSAFE
){
2547 sqlite3_mutex_free(pBt
->mutex
);
2551 sqlite3_mutex_leave(pMaster
);
2559 ** Make sure pBt->pTmpSpace points to an allocation of
2560 ** MX_CELL_SIZE(pBt) bytes with a 4-byte prefix for a left-child
2563 static void allocateTempSpace(BtShared
*pBt
){
2564 if( !pBt
->pTmpSpace
){
2565 pBt
->pTmpSpace
= sqlite3PageMalloc( pBt
->pageSize
);
2567 /* One of the uses of pBt->pTmpSpace is to format cells before
2568 ** inserting them into a leaf page (function fillInCell()). If
2569 ** a cell is less than 4 bytes in size, it is rounded up to 4 bytes
2570 ** by the various routines that manipulate binary cells. Which
2571 ** can mean that fillInCell() only initializes the first 2 or 3
2572 ** bytes of pTmpSpace, but that the first 4 bytes are copied from
2573 ** it into a database page. This is not actually a problem, but it
2574 ** does cause a valgrind error when the 1 or 2 bytes of unitialized
2575 ** data is passed to system call write(). So to avoid this error,
2576 ** zero the first 4 bytes of temp space here.
2578 ** Also: Provide four bytes of initialized space before the
2579 ** beginning of pTmpSpace as an area available to prepend the
2580 ** left-child pointer to the beginning of a cell.
2582 if( pBt
->pTmpSpace
){
2583 memset(pBt
->pTmpSpace
, 0, 8);
2584 pBt
->pTmpSpace
+= 4;
2590 ** Free the pBt->pTmpSpace allocation
2592 static void freeTempSpace(BtShared
*pBt
){
2593 if( pBt
->pTmpSpace
){
2594 pBt
->pTmpSpace
-= 4;
2595 sqlite3PageFree(pBt
->pTmpSpace
);
2601 ** Close an open database and invalidate all cursors.
2603 int sqlite3BtreeClose(Btree
*p
){
2604 BtShared
*pBt
= p
->pBt
;
2607 /* Close all cursors opened via this handle. */
2608 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2609 sqlite3BtreeEnter(p
);
2610 pCur
= pBt
->pCursor
;
2612 BtCursor
*pTmp
= pCur
;
2614 if( pTmp
->pBtree
==p
){
2615 sqlite3BtreeCloseCursor(pTmp
);
2619 /* Rollback any active transaction and free the handle structure.
2620 ** The call to sqlite3BtreeRollback() drops any table-locks held by
2623 sqlite3BtreeRollback(p
, SQLITE_OK
, 0);
2624 sqlite3BtreeLeave(p
);
2626 /* If there are still other outstanding references to the shared-btree
2627 ** structure, return now. The remainder of this procedure cleans
2628 ** up the shared-btree.
2630 assert( p
->wantToLock
==0 && p
->locked
==0 );
2631 if( !p
->sharable
|| removeFromSharingList(pBt
) ){
2632 /* The pBt is no longer on the sharing list, so we can access
2633 ** it without having to hold the mutex.
2635 ** Clean out and delete the BtShared object.
2637 assert( !pBt
->pCursor
);
2638 sqlite3PagerClose(pBt
->pPager
, p
->db
);
2639 if( pBt
->xFreeSchema
&& pBt
->pSchema
){
2640 pBt
->xFreeSchema(pBt
->pSchema
);
2642 sqlite3DbFree(0, pBt
->pSchema
);
2647 #ifndef SQLITE_OMIT_SHARED_CACHE
2648 assert( p
->wantToLock
==0 );
2649 assert( p
->locked
==0 );
2650 if( p
->pPrev
) p
->pPrev
->pNext
= p
->pNext
;
2651 if( p
->pNext
) p
->pNext
->pPrev
= p
->pPrev
;
2659 ** Change the "soft" limit on the number of pages in the cache.
2660 ** Unused and unmodified pages will be recycled when the number of
2661 ** pages in the cache exceeds this soft limit. But the size of the
2662 ** cache is allowed to grow larger than this limit if it contains
2663 ** dirty pages or pages still in active use.
2665 int sqlite3BtreeSetCacheSize(Btree
*p
, int mxPage
){
2666 BtShared
*pBt
= p
->pBt
;
2667 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2668 sqlite3BtreeEnter(p
);
2669 sqlite3PagerSetCachesize(pBt
->pPager
, mxPage
);
2670 sqlite3BtreeLeave(p
);
2675 ** Change the "spill" limit on the number of pages in the cache.
2676 ** If the number of pages exceeds this limit during a write transaction,
2677 ** the pager might attempt to "spill" pages to the journal early in
2678 ** order to free up memory.
2680 ** The value returned is the current spill size. If zero is passed
2681 ** as an argument, no changes are made to the spill size setting, so
2682 ** using mxPage of 0 is a way to query the current spill size.
2684 int sqlite3BtreeSetSpillSize(Btree
*p
, int mxPage
){
2685 BtShared
*pBt
= p
->pBt
;
2687 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2688 sqlite3BtreeEnter(p
);
2689 res
= sqlite3PagerSetSpillsize(pBt
->pPager
, mxPage
);
2690 sqlite3BtreeLeave(p
);
2694 #if SQLITE_MAX_MMAP_SIZE>0
2696 ** Change the limit on the amount of the database file that may be
2699 int sqlite3BtreeSetMmapLimit(Btree
*p
, sqlite3_int64 szMmap
){
2700 BtShared
*pBt
= p
->pBt
;
2701 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2702 sqlite3BtreeEnter(p
);
2703 sqlite3PagerSetMmapLimit(pBt
->pPager
, szMmap
);
2704 sqlite3BtreeLeave(p
);
2707 #endif /* SQLITE_MAX_MMAP_SIZE>0 */
2710 ** Change the way data is synced to disk in order to increase or decrease
2711 ** how well the database resists damage due to OS crashes and power
2712 ** failures. Level 1 is the same as asynchronous (no syncs() occur and
2713 ** there is a high probability of damage) Level 2 is the default. There
2714 ** is a very low but non-zero probability of damage. Level 3 reduces the
2715 ** probability of damage to near zero but with a write performance reduction.
2717 #ifndef SQLITE_OMIT_PAGER_PRAGMAS
2718 int sqlite3BtreeSetPagerFlags(
2719 Btree
*p
, /* The btree to set the safety level on */
2720 unsigned pgFlags
/* Various PAGER_* flags */
2722 BtShared
*pBt
= p
->pBt
;
2723 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2724 sqlite3BtreeEnter(p
);
2725 sqlite3PagerSetFlags(pBt
->pPager
, pgFlags
);
2726 sqlite3BtreeLeave(p
);
2732 ** Change the default pages size and the number of reserved bytes per page.
2733 ** Or, if the page size has already been fixed, return SQLITE_READONLY
2734 ** without changing anything.
2736 ** The page size must be a power of 2 between 512 and 65536. If the page
2737 ** size supplied does not meet this constraint then the page size is not
2740 ** Page sizes are constrained to be a power of two so that the region
2741 ** of the database file used for locking (beginning at PENDING_BYTE,
2742 ** the first byte past the 1GB boundary, 0x40000000) needs to occur
2743 ** at the beginning of a page.
2745 ** If parameter nReserve is less than zero, then the number of reserved
2746 ** bytes per page is left unchanged.
2748 ** If the iFix!=0 then the BTS_PAGESIZE_FIXED flag is set so that the page size
2749 ** and autovacuum mode can no longer be changed.
2751 int sqlite3BtreeSetPageSize(Btree
*p
, int pageSize
, int nReserve
, int iFix
){
2753 BtShared
*pBt
= p
->pBt
;
2754 assert( nReserve
>=-1 && nReserve
<=255 );
2755 sqlite3BtreeEnter(p
);
2756 #if SQLITE_HAS_CODEC
2757 if( nReserve
>pBt
->optimalReserve
) pBt
->optimalReserve
= (u8
)nReserve
;
2759 if( pBt
->btsFlags
& BTS_PAGESIZE_FIXED
){
2760 sqlite3BtreeLeave(p
);
2761 return SQLITE_READONLY
;
2764 nReserve
= pBt
->pageSize
- pBt
->usableSize
;
2766 assert( nReserve
>=0 && nReserve
<=255 );
2767 if( pageSize
>=512 && pageSize
<=SQLITE_MAX_PAGE_SIZE
&&
2768 ((pageSize
-1)&pageSize
)==0 ){
2769 assert( (pageSize
& 7)==0 );
2770 assert( !pBt
->pCursor
);
2771 pBt
->pageSize
= (u32
)pageSize
;
2774 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
, nReserve
);
2775 pBt
->usableSize
= pBt
->pageSize
- (u16
)nReserve
;
2776 if( iFix
) pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
2777 sqlite3BtreeLeave(p
);
2782 ** Return the currently defined page size
2784 int sqlite3BtreeGetPageSize(Btree
*p
){
2785 return p
->pBt
->pageSize
;
2789 ** This function is similar to sqlite3BtreeGetReserve(), except that it
2790 ** may only be called if it is guaranteed that the b-tree mutex is already
2793 ** This is useful in one special case in the backup API code where it is
2794 ** known that the shared b-tree mutex is held, but the mutex on the
2795 ** database handle that owns *p is not. In this case if sqlite3BtreeEnter()
2796 ** were to be called, it might collide with some other operation on the
2797 ** database handle that owns *p, causing undefined behavior.
2799 int sqlite3BtreeGetReserveNoMutex(Btree
*p
){
2801 assert( sqlite3_mutex_held(p
->pBt
->mutex
) );
2802 n
= p
->pBt
->pageSize
- p
->pBt
->usableSize
;
2807 ** Return the number of bytes of space at the end of every page that
2808 ** are intentually left unused. This is the "reserved" space that is
2809 ** sometimes used by extensions.
2811 ** If SQLITE_HAS_MUTEX is defined then the number returned is the
2812 ** greater of the current reserved space and the maximum requested
2815 int sqlite3BtreeGetOptimalReserve(Btree
*p
){
2817 sqlite3BtreeEnter(p
);
2818 n
= sqlite3BtreeGetReserveNoMutex(p
);
2819 #ifdef SQLITE_HAS_CODEC
2820 if( n
<p
->pBt
->optimalReserve
) n
= p
->pBt
->optimalReserve
;
2822 sqlite3BtreeLeave(p
);
2828 ** Set the maximum page count for a database if mxPage is positive.
2829 ** No changes are made if mxPage is 0 or negative.
2830 ** Regardless of the value of mxPage, return the maximum page count.
2832 int sqlite3BtreeMaxPageCount(Btree
*p
, int mxPage
){
2834 sqlite3BtreeEnter(p
);
2835 n
= sqlite3PagerMaxPageCount(p
->pBt
->pPager
, mxPage
);
2836 sqlite3BtreeLeave(p
);
2841 ** Change the values for the BTS_SECURE_DELETE and BTS_OVERWRITE flags:
2843 ** newFlag==0 Both BTS_SECURE_DELETE and BTS_OVERWRITE are cleared
2844 ** newFlag==1 BTS_SECURE_DELETE set and BTS_OVERWRITE is cleared
2845 ** newFlag==2 BTS_SECURE_DELETE cleared and BTS_OVERWRITE is set
2846 ** newFlag==(-1) No changes
2848 ** This routine acts as a query if newFlag is less than zero
2850 ** With BTS_OVERWRITE set, deleted content is overwritten by zeros, but
2851 ** freelist leaf pages are not written back to the database. Thus in-page
2852 ** deleted content is cleared, but freelist deleted content is not.
2854 ** With BTS_SECURE_DELETE, operation is like BTS_OVERWRITE with the addition
2855 ** that freelist leaf pages are written back into the database, increasing
2856 ** the amount of disk I/O.
2858 int sqlite3BtreeSecureDelete(Btree
*p
, int newFlag
){
2860 if( p
==0 ) return 0;
2861 sqlite3BtreeEnter(p
);
2862 assert( BTS_OVERWRITE
==BTS_SECURE_DELETE
*2 );
2863 assert( BTS_FAST_SECURE
==(BTS_OVERWRITE
|BTS_SECURE_DELETE
) );
2865 p
->pBt
->btsFlags
&= ~BTS_FAST_SECURE
;
2866 p
->pBt
->btsFlags
|= BTS_SECURE_DELETE
*newFlag
;
2868 b
= (p
->pBt
->btsFlags
& BTS_FAST_SECURE
)/BTS_SECURE_DELETE
;
2869 sqlite3BtreeLeave(p
);
2874 ** Change the 'auto-vacuum' property of the database. If the 'autoVacuum'
2875 ** parameter is non-zero, then auto-vacuum mode is enabled. If zero, it
2876 ** is disabled. The default value for the auto-vacuum property is
2877 ** determined by the SQLITE_DEFAULT_AUTOVACUUM macro.
2879 int sqlite3BtreeSetAutoVacuum(Btree
*p
, int autoVacuum
){
2880 #ifdef SQLITE_OMIT_AUTOVACUUM
2881 return SQLITE_READONLY
;
2883 BtShared
*pBt
= p
->pBt
;
2885 u8 av
= (u8
)autoVacuum
;
2887 sqlite3BtreeEnter(p
);
2888 if( (pBt
->btsFlags
& BTS_PAGESIZE_FIXED
)!=0 && (av
?1:0)!=pBt
->autoVacuum
){
2889 rc
= SQLITE_READONLY
;
2891 pBt
->autoVacuum
= av
?1:0;
2892 pBt
->incrVacuum
= av
==2 ?1:0;
2894 sqlite3BtreeLeave(p
);
2900 ** Return the value of the 'auto-vacuum' property. If auto-vacuum is
2901 ** enabled 1 is returned. Otherwise 0.
2903 int sqlite3BtreeGetAutoVacuum(Btree
*p
){
2904 #ifdef SQLITE_OMIT_AUTOVACUUM
2905 return BTREE_AUTOVACUUM_NONE
;
2908 sqlite3BtreeEnter(p
);
2910 (!p
->pBt
->autoVacuum
)?BTREE_AUTOVACUUM_NONE
:
2911 (!p
->pBt
->incrVacuum
)?BTREE_AUTOVACUUM_FULL
:
2912 BTREE_AUTOVACUUM_INCR
2914 sqlite3BtreeLeave(p
);
2920 ** If the user has not set the safety-level for this database connection
2921 ** using "PRAGMA synchronous", and if the safety-level is not already
2922 ** set to the value passed to this function as the second parameter,
2925 #if SQLITE_DEFAULT_SYNCHRONOUS!=SQLITE_DEFAULT_WAL_SYNCHRONOUS \
2926 && !defined(SQLITE_OMIT_WAL)
2927 static void setDefaultSyncFlag(BtShared
*pBt
, u8 safety_level
){
2930 if( (db
=pBt
->db
)!=0 && (pDb
=db
->aDb
)!=0 ){
2931 while( pDb
->pBt
==0 || pDb
->pBt
->pBt
!=pBt
){ pDb
++; }
2932 if( pDb
->bSyncSet
==0
2933 && pDb
->safety_level
!=safety_level
2936 pDb
->safety_level
= safety_level
;
2937 sqlite3PagerSetFlags(pBt
->pPager
,
2938 pDb
->safety_level
| (db
->flags
& PAGER_FLAGS_MASK
));
2943 # define setDefaultSyncFlag(pBt,safety_level)
2947 ** Get a reference to pPage1 of the database file. This will
2948 ** also acquire a readlock on that file.
2950 ** SQLITE_OK is returned on success. If the file is not a
2951 ** well-formed database file, then SQLITE_CORRUPT is returned.
2952 ** SQLITE_BUSY is returned if the database is locked. SQLITE_NOMEM
2953 ** is returned if we run out of memory.
2955 static int lockBtree(BtShared
*pBt
){
2956 int rc
; /* Result code from subfunctions */
2957 MemPage
*pPage1
; /* Page 1 of the database file */
2958 int nPage
; /* Number of pages in the database */
2959 int nPageFile
= 0; /* Number of pages in the database file */
2960 int nPageHeader
; /* Number of pages in the database according to hdr */
2962 assert( sqlite3_mutex_held(pBt
->mutex
) );
2963 assert( pBt
->pPage1
==0 );
2964 rc
= sqlite3PagerSharedLock(pBt
->pPager
);
2965 if( rc
!=SQLITE_OK
) return rc
;
2966 rc
= btreeGetPage(pBt
, 1, &pPage1
, 0);
2967 if( rc
!=SQLITE_OK
) return rc
;
2969 /* Do some checking to help insure the file we opened really is
2970 ** a valid database file.
2972 nPage
= nPageHeader
= get4byte(28+(u8
*)pPage1
->aData
);
2973 sqlite3PagerPagecount(pBt
->pPager
, &nPageFile
);
2974 if( nPage
==0 || memcmp(24+(u8
*)pPage1
->aData
, 92+(u8
*)pPage1
->aData
,4)!=0 ){
2980 u8
*page1
= pPage1
->aData
;
2982 /* EVIDENCE-OF: R-43737-39999 Every valid SQLite database file begins
2983 ** with the following 16 bytes (in hex): 53 51 4c 69 74 65 20 66 6f 72 6d
2984 ** 61 74 20 33 00. */
2985 if( memcmp(page1
, zMagicHeader
, 16)!=0 ){
2986 goto page1_init_failed
;
2989 #ifdef SQLITE_OMIT_WAL
2991 pBt
->btsFlags
|= BTS_READ_ONLY
;
2994 goto page1_init_failed
;
2998 pBt
->btsFlags
|= BTS_READ_ONLY
;
3001 goto page1_init_failed
;
3004 /* If the write version is set to 2, this database should be accessed
3005 ** in WAL mode. If the log is not already open, open it now. Then
3006 ** return SQLITE_OK and return without populating BtShared.pPage1.
3007 ** The caller detects this and calls this function again. This is
3008 ** required as the version of page 1 currently in the page1 buffer
3009 ** may not be the latest version - there may be a newer one in the log
3012 if( page1
[19]==2 && (pBt
->btsFlags
& BTS_NO_WAL
)==0 ){
3014 rc
= sqlite3PagerOpenWal(pBt
->pPager
, &isOpen
);
3015 if( rc
!=SQLITE_OK
){
3016 goto page1_init_failed
;
3018 setDefaultSyncFlag(pBt
, SQLITE_DEFAULT_WAL_SYNCHRONOUS
+1);
3020 releasePageOne(pPage1
);
3026 setDefaultSyncFlag(pBt
, SQLITE_DEFAULT_SYNCHRONOUS
+1);
3030 /* EVIDENCE-OF: R-15465-20813 The maximum and minimum embedded payload
3031 ** fractions and the leaf payload fraction values must be 64, 32, and 32.
3033 ** The original design allowed these amounts to vary, but as of
3034 ** version 3.6.0, we require them to be fixed.
3036 if( memcmp(&page1
[21], "\100\040\040",3)!=0 ){
3037 goto page1_init_failed
;
3039 /* EVIDENCE-OF: R-51873-39618 The page size for a database file is
3040 ** determined by the 2-byte integer located at an offset of 16 bytes from
3041 ** the beginning of the database file. */
3042 pageSize
= (page1
[16]<<8) | (page1
[17]<<16);
3043 /* EVIDENCE-OF: R-25008-21688 The size of a page is a power of two
3044 ** between 512 and 65536 inclusive. */
3045 if( ((pageSize
-1)&pageSize
)!=0
3046 || pageSize
>SQLITE_MAX_PAGE_SIZE
3049 goto page1_init_failed
;
3051 assert( (pageSize
& 7)==0 );
3052 /* EVIDENCE-OF: R-59310-51205 The "reserved space" size in the 1-byte
3053 ** integer at offset 20 is the number of bytes of space at the end of
3054 ** each page to reserve for extensions.
3056 ** EVIDENCE-OF: R-37497-42412 The size of the reserved region is
3057 ** determined by the one-byte unsigned integer found at an offset of 20
3058 ** into the database file header. */
3059 usableSize
= pageSize
- page1
[20];
3060 if( (u32
)pageSize
!=pBt
->pageSize
){
3061 /* After reading the first page of the database assuming a page size
3062 ** of BtShared.pageSize, we have discovered that the page-size is
3063 ** actually pageSize. Unlock the database, leave pBt->pPage1 at
3064 ** zero and return SQLITE_OK. The caller will call this function
3065 ** again with the correct page-size.
3067 releasePageOne(pPage1
);
3068 pBt
->usableSize
= usableSize
;
3069 pBt
->pageSize
= pageSize
;
3071 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
,
3072 pageSize
-usableSize
);
3075 if( (pBt
->db
->flags
& SQLITE_WriteSchema
)==0 && nPage
>nPageFile
){
3076 rc
= SQLITE_CORRUPT_BKPT
;
3077 goto page1_init_failed
;
3079 /* EVIDENCE-OF: R-28312-64704 However, the usable size is not allowed to
3080 ** be less than 480. In other words, if the page size is 512, then the
3081 ** reserved space size cannot exceed 32. */
3082 if( usableSize
<480 ){
3083 goto page1_init_failed
;
3085 pBt
->pageSize
= pageSize
;
3086 pBt
->usableSize
= usableSize
;
3087 #ifndef SQLITE_OMIT_AUTOVACUUM
3088 pBt
->autoVacuum
= (get4byte(&page1
[36 + 4*4])?1:0);
3089 pBt
->incrVacuum
= (get4byte(&page1
[36 + 7*4])?1:0);
3093 /* maxLocal is the maximum amount of payload to store locally for
3094 ** a cell. Make sure it is small enough so that at least minFanout
3095 ** cells can will fit on one page. We assume a 10-byte page header.
3096 ** Besides the payload, the cell must store:
3097 ** 2-byte pointer to the cell
3098 ** 4-byte child pointer
3099 ** 9-byte nKey value
3100 ** 4-byte nData value
3101 ** 4-byte overflow page pointer
3102 ** So a cell consists of a 2-byte pointer, a header which is as much as
3103 ** 17 bytes long, 0 to N bytes of payload, and an optional 4 byte overflow
3106 pBt
->maxLocal
= (u16
)((pBt
->usableSize
-12)*64/255 - 23);
3107 pBt
->minLocal
= (u16
)((pBt
->usableSize
-12)*32/255 - 23);
3108 pBt
->maxLeaf
= (u16
)(pBt
->usableSize
- 35);
3109 pBt
->minLeaf
= (u16
)((pBt
->usableSize
-12)*32/255 - 23);
3110 if( pBt
->maxLocal
>127 ){
3111 pBt
->max1bytePayload
= 127;
3113 pBt
->max1bytePayload
= (u8
)pBt
->maxLocal
;
3115 assert( pBt
->maxLeaf
+ 23 <= MX_CELL_SIZE(pBt
) );
3116 pBt
->pPage1
= pPage1
;
3121 releasePageOne(pPage1
);
3128 ** Return the number of cursors open on pBt. This is for use
3129 ** in assert() expressions, so it is only compiled if NDEBUG is not
3132 ** Only write cursors are counted if wrOnly is true. If wrOnly is
3133 ** false then all cursors are counted.
3135 ** For the purposes of this routine, a cursor is any cursor that
3136 ** is capable of reading or writing to the database. Cursors that
3137 ** have been tripped into the CURSOR_FAULT state are not counted.
3139 static int countValidCursors(BtShared
*pBt
, int wrOnly
){
3142 for(pCur
=pBt
->pCursor
; pCur
; pCur
=pCur
->pNext
){
3143 if( (wrOnly
==0 || (pCur
->curFlags
& BTCF_WriteFlag
)!=0)
3144 && pCur
->eState
!=CURSOR_FAULT
) r
++;
3151 ** If there are no outstanding cursors and we are not in the middle
3152 ** of a transaction but there is a read lock on the database, then
3153 ** this routine unrefs the first page of the database file which
3154 ** has the effect of releasing the read lock.
3156 ** If there is a transaction in progress, this routine is a no-op.
3158 static void unlockBtreeIfUnused(BtShared
*pBt
){
3159 assert( sqlite3_mutex_held(pBt
->mutex
) );
3160 assert( countValidCursors(pBt
,0)==0 || pBt
->inTransaction
>TRANS_NONE
);
3161 if( pBt
->inTransaction
==TRANS_NONE
&& pBt
->pPage1
!=0 ){
3162 MemPage
*pPage1
= pBt
->pPage1
;
3163 assert( pPage1
->aData
);
3164 assert( sqlite3PagerRefcount(pBt
->pPager
)==1 );
3166 releasePageOne(pPage1
);
3171 ** If pBt points to an empty file then convert that empty file
3172 ** into a new empty database by initializing the first page of
3175 static int newDatabase(BtShared
*pBt
){
3177 unsigned char *data
;
3180 assert( sqlite3_mutex_held(pBt
->mutex
) );
3187 rc
= sqlite3PagerWrite(pP1
->pDbPage
);
3189 memcpy(data
, zMagicHeader
, sizeof(zMagicHeader
));
3190 assert( sizeof(zMagicHeader
)==16 );
3191 data
[16] = (u8
)((pBt
->pageSize
>>8)&0xff);
3192 data
[17] = (u8
)((pBt
->pageSize
>>16)&0xff);
3195 assert( pBt
->usableSize
<=pBt
->pageSize
&& pBt
->usableSize
+255>=pBt
->pageSize
);
3196 data
[20] = (u8
)(pBt
->pageSize
- pBt
->usableSize
);
3200 memset(&data
[24], 0, 100-24);
3201 zeroPage(pP1
, PTF_INTKEY
|PTF_LEAF
|PTF_LEAFDATA
);
3202 pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
3203 #ifndef SQLITE_OMIT_AUTOVACUUM
3204 assert( pBt
->autoVacuum
==1 || pBt
->autoVacuum
==0 );
3205 assert( pBt
->incrVacuum
==1 || pBt
->incrVacuum
==0 );
3206 put4byte(&data
[36 + 4*4], pBt
->autoVacuum
);
3207 put4byte(&data
[36 + 7*4], pBt
->incrVacuum
);
3215 ** Initialize the first page of the database file (creating a database
3216 ** consisting of a single page and no schema objects). Return SQLITE_OK
3217 ** if successful, or an SQLite error code otherwise.
3219 int sqlite3BtreeNewDb(Btree
*p
){
3221 sqlite3BtreeEnter(p
);
3223 rc
= newDatabase(p
->pBt
);
3224 sqlite3BtreeLeave(p
);
3229 ** Attempt to start a new transaction. A write-transaction
3230 ** is started if the second argument is nonzero, otherwise a read-
3231 ** transaction. If the second argument is 2 or more and exclusive
3232 ** transaction is started, meaning that no other process is allowed
3233 ** to access the database. A preexisting transaction may not be
3234 ** upgraded to exclusive by calling this routine a second time - the
3235 ** exclusivity flag only works for a new transaction.
3237 ** A write-transaction must be started before attempting any
3238 ** changes to the database. None of the following routines
3239 ** will work unless a transaction is started first:
3241 ** sqlite3BtreeCreateTable()
3242 ** sqlite3BtreeCreateIndex()
3243 ** sqlite3BtreeClearTable()
3244 ** sqlite3BtreeDropTable()
3245 ** sqlite3BtreeInsert()
3246 ** sqlite3BtreeDelete()
3247 ** sqlite3BtreeUpdateMeta()
3249 ** If an initial attempt to acquire the lock fails because of lock contention
3250 ** and the database was previously unlocked, then invoke the busy handler
3251 ** if there is one. But if there was previously a read-lock, do not
3252 ** invoke the busy handler - just return SQLITE_BUSY. SQLITE_BUSY is
3253 ** returned when there is already a read-lock in order to avoid a deadlock.
3255 ** Suppose there are two processes A and B. A has a read lock and B has
3256 ** a reserved lock. B tries to promote to exclusive but is blocked because
3257 ** of A's read lock. A tries to promote to reserved but is blocked by B.
3258 ** One or the other of the two processes must give way or there can be
3259 ** no progress. By returning SQLITE_BUSY and not invoking the busy callback
3260 ** when A already has a read lock, we encourage A to give up and let B
3263 int sqlite3BtreeBeginTrans(Btree
*p
, int wrflag
){
3264 BtShared
*pBt
= p
->pBt
;
3267 sqlite3BtreeEnter(p
);
3270 /* If the btree is already in a write-transaction, or it
3271 ** is already in a read-transaction and a read-transaction
3272 ** is requested, this is a no-op.
3274 if( p
->inTrans
==TRANS_WRITE
|| (p
->inTrans
==TRANS_READ
&& !wrflag
) ){
3277 assert( pBt
->inTransaction
==TRANS_WRITE
|| IfNotOmitAV(pBt
->bDoTruncate
)==0 );
3279 /* Write transactions are not possible on a read-only database */
3280 if( (pBt
->btsFlags
& BTS_READ_ONLY
)!=0 && wrflag
){
3281 rc
= SQLITE_READONLY
;
3285 #ifndef SQLITE_OMIT_SHARED_CACHE
3287 sqlite3
*pBlock
= 0;
3288 /* If another database handle has already opened a write transaction
3289 ** on this shared-btree structure and a second write transaction is
3290 ** requested, return SQLITE_LOCKED.
3292 if( (wrflag
&& pBt
->inTransaction
==TRANS_WRITE
)
3293 || (pBt
->btsFlags
& BTS_PENDING
)!=0
3295 pBlock
= pBt
->pWriter
->db
;
3296 }else if( wrflag
>1 ){
3298 for(pIter
=pBt
->pLock
; pIter
; pIter
=pIter
->pNext
){
3299 if( pIter
->pBtree
!=p
){
3300 pBlock
= pIter
->pBtree
->db
;
3306 sqlite3ConnectionBlocked(p
->db
, pBlock
);
3307 rc
= SQLITE_LOCKED_SHAREDCACHE
;
3313 /* Any read-only or read-write transaction implies a read-lock on
3314 ** page 1. So if some other shared-cache client already has a write-lock
3315 ** on page 1, the transaction cannot be opened. */
3316 rc
= querySharedCacheTableLock(p
, MASTER_ROOT
, READ_LOCK
);
3317 if( SQLITE_OK
!=rc
) goto trans_begun
;
3319 pBt
->btsFlags
&= ~BTS_INITIALLY_EMPTY
;
3320 if( pBt
->nPage
==0 ) pBt
->btsFlags
|= BTS_INITIALLY_EMPTY
;
3322 /* Call lockBtree() until either pBt->pPage1 is populated or
3323 ** lockBtree() returns something other than SQLITE_OK. lockBtree()
3324 ** may return SQLITE_OK but leave pBt->pPage1 set to 0 if after
3325 ** reading page 1 it discovers that the page-size of the database
3326 ** file is not pBt->pageSize. In this case lockBtree() will update
3327 ** pBt->pageSize to the page-size of the file on disk.
3329 while( pBt
->pPage1
==0 && SQLITE_OK
==(rc
= lockBtree(pBt
)) );
3331 if( rc
==SQLITE_OK
&& wrflag
){
3332 if( (pBt
->btsFlags
& BTS_READ_ONLY
)!=0 ){
3333 rc
= SQLITE_READONLY
;
3335 rc
= sqlite3PagerBegin(pBt
->pPager
,wrflag
>1,sqlite3TempInMemory(p
->db
));
3336 if( rc
==SQLITE_OK
){
3337 rc
= newDatabase(pBt
);
3342 if( rc
!=SQLITE_OK
){
3343 unlockBtreeIfUnused(pBt
);
3345 }while( (rc
&0xFF)==SQLITE_BUSY
&& pBt
->inTransaction
==TRANS_NONE
&&
3346 btreeInvokeBusyHandler(pBt
) );
3348 if( rc
==SQLITE_OK
){
3349 if( p
->inTrans
==TRANS_NONE
){
3350 pBt
->nTransaction
++;
3351 #ifndef SQLITE_OMIT_SHARED_CACHE
3353 assert( p
->lock
.pBtree
==p
&& p
->lock
.iTable
==1 );
3354 p
->lock
.eLock
= READ_LOCK
;
3355 p
->lock
.pNext
= pBt
->pLock
;
3356 pBt
->pLock
= &p
->lock
;
3360 p
->inTrans
= (wrflag
?TRANS_WRITE
:TRANS_READ
);
3361 if( p
->inTrans
>pBt
->inTransaction
){
3362 pBt
->inTransaction
= p
->inTrans
;
3365 MemPage
*pPage1
= pBt
->pPage1
;
3366 #ifndef SQLITE_OMIT_SHARED_CACHE
3367 assert( !pBt
->pWriter
);
3369 pBt
->btsFlags
&= ~BTS_EXCLUSIVE
;
3370 if( wrflag
>1 ) pBt
->btsFlags
|= BTS_EXCLUSIVE
;
3373 /* If the db-size header field is incorrect (as it may be if an old
3374 ** client has been writing the database file), update it now. Doing
3375 ** this sooner rather than later means the database size can safely
3376 ** re-read the database size from page 1 if a savepoint or transaction
3377 ** rollback occurs within the transaction.
3379 if( pBt
->nPage
!=get4byte(&pPage1
->aData
[28]) ){
3380 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
3381 if( rc
==SQLITE_OK
){
3382 put4byte(&pPage1
->aData
[28], pBt
->nPage
);
3390 if( rc
==SQLITE_OK
&& wrflag
){
3391 /* This call makes sure that the pager has the correct number of
3392 ** open savepoints. If the second parameter is greater than 0 and
3393 ** the sub-journal is not already open, then it will be opened here.
3395 rc
= sqlite3PagerOpenSavepoint(pBt
->pPager
, p
->db
->nSavepoint
);
3399 sqlite3BtreeLeave(p
);
3403 #ifndef SQLITE_OMIT_AUTOVACUUM
3406 ** Set the pointer-map entries for all children of page pPage. Also, if
3407 ** pPage contains cells that point to overflow pages, set the pointer
3408 ** map entries for the overflow pages as well.
3410 static int setChildPtrmaps(MemPage
*pPage
){
3411 int i
; /* Counter variable */
3412 int nCell
; /* Number of cells in page pPage */
3413 int rc
; /* Return code */
3414 BtShared
*pBt
= pPage
->pBt
;
3415 Pgno pgno
= pPage
->pgno
;
3417 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
3418 rc
= pPage
->isInit
? SQLITE_OK
: btreeInitPage(pPage
);
3419 if( rc
!=SQLITE_OK
) return rc
;
3420 nCell
= pPage
->nCell
;
3422 for(i
=0; i
<nCell
; i
++){
3423 u8
*pCell
= findCell(pPage
, i
);
3425 ptrmapPutOvflPtr(pPage
, pCell
, &rc
);
3428 Pgno childPgno
= get4byte(pCell
);
3429 ptrmapPut(pBt
, childPgno
, PTRMAP_BTREE
, pgno
, &rc
);
3434 Pgno childPgno
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
3435 ptrmapPut(pBt
, childPgno
, PTRMAP_BTREE
, pgno
, &rc
);
3442 ** Somewhere on pPage is a pointer to page iFrom. Modify this pointer so
3443 ** that it points to iTo. Parameter eType describes the type of pointer to
3444 ** be modified, as follows:
3446 ** PTRMAP_BTREE: pPage is a btree-page. The pointer points at a child
3449 ** PTRMAP_OVERFLOW1: pPage is a btree-page. The pointer points at an overflow
3450 ** page pointed to by one of the cells on pPage.
3452 ** PTRMAP_OVERFLOW2: pPage is an overflow-page. The pointer points at the next
3453 ** overflow page in the list.
3455 static int modifyPagePointer(MemPage
*pPage
, Pgno iFrom
, Pgno iTo
, u8 eType
){
3456 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
3457 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
3458 if( eType
==PTRMAP_OVERFLOW2
){
3459 /* The pointer is always the first 4 bytes of the page in this case. */
3460 if( get4byte(pPage
->aData
)!=iFrom
){
3461 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
3463 put4byte(pPage
->aData
, iTo
);
3469 rc
= pPage
->isInit
? SQLITE_OK
: btreeInitPage(pPage
);
3471 nCell
= pPage
->nCell
;
3473 for(i
=0; i
<nCell
; i
++){
3474 u8
*pCell
= findCell(pPage
, i
);
3475 if( eType
==PTRMAP_OVERFLOW1
){
3477 pPage
->xParseCell(pPage
, pCell
, &info
);
3478 if( info
.nLocal
<info
.nPayload
){
3479 if( pCell
+info
.nSize
> pPage
->aData
+pPage
->pBt
->usableSize
){
3480 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
3482 if( iFrom
==get4byte(pCell
+info
.nSize
-4) ){
3483 put4byte(pCell
+info
.nSize
-4, iTo
);
3488 if( get4byte(pCell
)==iFrom
){
3489 put4byte(pCell
, iTo
);
3496 if( eType
!=PTRMAP_BTREE
||
3497 get4byte(&pPage
->aData
[pPage
->hdrOffset
+8])!=iFrom
){
3498 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
3500 put4byte(&pPage
->aData
[pPage
->hdrOffset
+8], iTo
);
3508 ** Move the open database page pDbPage to location iFreePage in the
3509 ** database. The pDbPage reference remains valid.
3511 ** The isCommit flag indicates that there is no need to remember that
3512 ** the journal needs to be sync()ed before database page pDbPage->pgno
3513 ** can be written to. The caller has already promised not to write to that
3516 static int relocatePage(
3517 BtShared
*pBt
, /* Btree */
3518 MemPage
*pDbPage
, /* Open page to move */
3519 u8 eType
, /* Pointer map 'type' entry for pDbPage */
3520 Pgno iPtrPage
, /* Pointer map 'page-no' entry for pDbPage */
3521 Pgno iFreePage
, /* The location to move pDbPage to */
3522 int isCommit
/* isCommit flag passed to sqlite3PagerMovepage */
3524 MemPage
*pPtrPage
; /* The page that contains a pointer to pDbPage */
3525 Pgno iDbPage
= pDbPage
->pgno
;
3526 Pager
*pPager
= pBt
->pPager
;
3529 assert( eType
==PTRMAP_OVERFLOW2
|| eType
==PTRMAP_OVERFLOW1
||
3530 eType
==PTRMAP_BTREE
|| eType
==PTRMAP_ROOTPAGE
);
3531 assert( sqlite3_mutex_held(pBt
->mutex
) );
3532 assert( pDbPage
->pBt
==pBt
);
3534 /* Move page iDbPage from its current location to page number iFreePage */
3535 TRACE(("AUTOVACUUM: Moving %d to free page %d (ptr page %d type %d)\n",
3536 iDbPage
, iFreePage
, iPtrPage
, eType
));
3537 rc
= sqlite3PagerMovepage(pPager
, pDbPage
->pDbPage
, iFreePage
, isCommit
);
3538 if( rc
!=SQLITE_OK
){
3541 pDbPage
->pgno
= iFreePage
;
3543 /* If pDbPage was a btree-page, then it may have child pages and/or cells
3544 ** that point to overflow pages. The pointer map entries for all these
3545 ** pages need to be changed.
3547 ** If pDbPage is an overflow page, then the first 4 bytes may store a
3548 ** pointer to a subsequent overflow page. If this is the case, then
3549 ** the pointer map needs to be updated for the subsequent overflow page.
3551 if( eType
==PTRMAP_BTREE
|| eType
==PTRMAP_ROOTPAGE
){
3552 rc
= setChildPtrmaps(pDbPage
);
3553 if( rc
!=SQLITE_OK
){
3557 Pgno nextOvfl
= get4byte(pDbPage
->aData
);
3559 ptrmapPut(pBt
, nextOvfl
, PTRMAP_OVERFLOW2
, iFreePage
, &rc
);
3560 if( rc
!=SQLITE_OK
){
3566 /* Fix the database pointer on page iPtrPage that pointed at iDbPage so
3567 ** that it points at iFreePage. Also fix the pointer map entry for
3570 if( eType
!=PTRMAP_ROOTPAGE
){
3571 rc
= btreeGetPage(pBt
, iPtrPage
, &pPtrPage
, 0);
3572 if( rc
!=SQLITE_OK
){
3575 rc
= sqlite3PagerWrite(pPtrPage
->pDbPage
);
3576 if( rc
!=SQLITE_OK
){
3577 releasePage(pPtrPage
);
3580 rc
= modifyPagePointer(pPtrPage
, iDbPage
, iFreePage
, eType
);
3581 releasePage(pPtrPage
);
3582 if( rc
==SQLITE_OK
){
3583 ptrmapPut(pBt
, iFreePage
, eType
, iPtrPage
, &rc
);
3589 /* Forward declaration required by incrVacuumStep(). */
3590 static int allocateBtreePage(BtShared
*, MemPage
**, Pgno
*, Pgno
, u8
);
3593 ** Perform a single step of an incremental-vacuum. If successful, return
3594 ** SQLITE_OK. If there is no work to do (and therefore no point in
3595 ** calling this function again), return SQLITE_DONE. Or, if an error
3596 ** occurs, return some other error code.
3598 ** More specifically, this function attempts to re-organize the database so
3599 ** that the last page of the file currently in use is no longer in use.
3601 ** Parameter nFin is the number of pages that this database would contain
3602 ** were this function called until it returns SQLITE_DONE.
3604 ** If the bCommit parameter is non-zero, this function assumes that the
3605 ** caller will keep calling incrVacuumStep() until it returns SQLITE_DONE
3606 ** or an error. bCommit is passed true for an auto-vacuum-on-commit
3607 ** operation, or false for an incremental vacuum.
3609 static int incrVacuumStep(BtShared
*pBt
, Pgno nFin
, Pgno iLastPg
, int bCommit
){
3610 Pgno nFreeList
; /* Number of pages still on the free-list */
3613 assert( sqlite3_mutex_held(pBt
->mutex
) );
3614 assert( iLastPg
>nFin
);
3616 if( !PTRMAP_ISPAGE(pBt
, iLastPg
) && iLastPg
!=PENDING_BYTE_PAGE(pBt
) ){
3620 nFreeList
= get4byte(&pBt
->pPage1
->aData
[36]);
3625 rc
= ptrmapGet(pBt
, iLastPg
, &eType
, &iPtrPage
);
3626 if( rc
!=SQLITE_OK
){
3629 if( eType
==PTRMAP_ROOTPAGE
){
3630 return SQLITE_CORRUPT_BKPT
;
3633 if( eType
==PTRMAP_FREEPAGE
){
3635 /* Remove the page from the files free-list. This is not required
3636 ** if bCommit is non-zero. In that case, the free-list will be
3637 ** truncated to zero after this function returns, so it doesn't
3638 ** matter if it still contains some garbage entries.
3642 rc
= allocateBtreePage(pBt
, &pFreePg
, &iFreePg
, iLastPg
, BTALLOC_EXACT
);
3643 if( rc
!=SQLITE_OK
){
3646 assert( iFreePg
==iLastPg
);
3647 releasePage(pFreePg
);
3650 Pgno iFreePg
; /* Index of free page to move pLastPg to */
3652 u8 eMode
= BTALLOC_ANY
; /* Mode parameter for allocateBtreePage() */
3653 Pgno iNear
= 0; /* nearby parameter for allocateBtreePage() */
3655 rc
= btreeGetPage(pBt
, iLastPg
, &pLastPg
, 0);
3656 if( rc
!=SQLITE_OK
){
3660 /* If bCommit is zero, this loop runs exactly once and page pLastPg
3661 ** is swapped with the first free page pulled off the free list.
3663 ** On the other hand, if bCommit is greater than zero, then keep
3664 ** looping until a free-page located within the first nFin pages
3665 ** of the file is found.
3673 rc
= allocateBtreePage(pBt
, &pFreePg
, &iFreePg
, iNear
, eMode
);
3674 if( rc
!=SQLITE_OK
){
3675 releasePage(pLastPg
);
3678 releasePage(pFreePg
);
3679 }while( bCommit
&& iFreePg
>nFin
);
3680 assert( iFreePg
<iLastPg
);
3682 rc
= relocatePage(pBt
, pLastPg
, eType
, iPtrPage
, iFreePg
, bCommit
);
3683 releasePage(pLastPg
);
3684 if( rc
!=SQLITE_OK
){
3693 }while( iLastPg
==PENDING_BYTE_PAGE(pBt
) || PTRMAP_ISPAGE(pBt
, iLastPg
) );
3694 pBt
->bDoTruncate
= 1;
3695 pBt
->nPage
= iLastPg
;
3701 ** The database opened by the first argument is an auto-vacuum database
3702 ** nOrig pages in size containing nFree free pages. Return the expected
3703 ** size of the database in pages following an auto-vacuum operation.
3705 static Pgno
finalDbSize(BtShared
*pBt
, Pgno nOrig
, Pgno nFree
){
3706 int nEntry
; /* Number of entries on one ptrmap page */
3707 Pgno nPtrmap
; /* Number of PtrMap pages to be freed */
3708 Pgno nFin
; /* Return value */
3710 nEntry
= pBt
->usableSize
/5;
3711 nPtrmap
= (nFree
-nOrig
+PTRMAP_PAGENO(pBt
, nOrig
)+nEntry
)/nEntry
;
3712 nFin
= nOrig
- nFree
- nPtrmap
;
3713 if( nOrig
>PENDING_BYTE_PAGE(pBt
) && nFin
<PENDING_BYTE_PAGE(pBt
) ){
3716 while( PTRMAP_ISPAGE(pBt
, nFin
) || nFin
==PENDING_BYTE_PAGE(pBt
) ){
3724 ** A write-transaction must be opened before calling this function.
3725 ** It performs a single unit of work towards an incremental vacuum.
3727 ** If the incremental vacuum is finished after this function has run,
3728 ** SQLITE_DONE is returned. If it is not finished, but no error occurred,
3729 ** SQLITE_OK is returned. Otherwise an SQLite error code.
3731 int sqlite3BtreeIncrVacuum(Btree
*p
){
3733 BtShared
*pBt
= p
->pBt
;
3735 sqlite3BtreeEnter(p
);
3736 assert( pBt
->inTransaction
==TRANS_WRITE
&& p
->inTrans
==TRANS_WRITE
);
3737 if( !pBt
->autoVacuum
){
3740 Pgno nOrig
= btreePagecount(pBt
);
3741 Pgno nFree
= get4byte(&pBt
->pPage1
->aData
[36]);
3742 Pgno nFin
= finalDbSize(pBt
, nOrig
, nFree
);
3745 rc
= SQLITE_CORRUPT_BKPT
;
3746 }else if( nFree
>0 ){
3747 rc
= saveAllCursors(pBt
, 0, 0);
3748 if( rc
==SQLITE_OK
){
3749 invalidateAllOverflowCache(pBt
);
3750 rc
= incrVacuumStep(pBt
, nFin
, nOrig
, 0);
3752 if( rc
==SQLITE_OK
){
3753 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
3754 put4byte(&pBt
->pPage1
->aData
[28], pBt
->nPage
);
3760 sqlite3BtreeLeave(p
);
3765 ** This routine is called prior to sqlite3PagerCommit when a transaction
3766 ** is committed for an auto-vacuum database.
3768 ** If SQLITE_OK is returned, then *pnTrunc is set to the number of pages
3769 ** the database file should be truncated to during the commit process.
3770 ** i.e. the database has been reorganized so that only the first *pnTrunc
3771 ** pages are in use.
3773 static int autoVacuumCommit(BtShared
*pBt
){
3775 Pager
*pPager
= pBt
->pPager
;
3776 VVA_ONLY( int nRef
= sqlite3PagerRefcount(pPager
); )
3778 assert( sqlite3_mutex_held(pBt
->mutex
) );
3779 invalidateAllOverflowCache(pBt
);
3780 assert(pBt
->autoVacuum
);
3781 if( !pBt
->incrVacuum
){
3782 Pgno nFin
; /* Number of pages in database after autovacuuming */
3783 Pgno nFree
; /* Number of pages on the freelist initially */
3784 Pgno iFree
; /* The next page to be freed */
3785 Pgno nOrig
; /* Database size before freeing */
3787 nOrig
= btreePagecount(pBt
);
3788 if( PTRMAP_ISPAGE(pBt
, nOrig
) || nOrig
==PENDING_BYTE_PAGE(pBt
) ){
3789 /* It is not possible to create a database for which the final page
3790 ** is either a pointer-map page or the pending-byte page. If one
3791 ** is encountered, this indicates corruption.
3793 return SQLITE_CORRUPT_BKPT
;
3796 nFree
= get4byte(&pBt
->pPage1
->aData
[36]);
3797 nFin
= finalDbSize(pBt
, nOrig
, nFree
);
3798 if( nFin
>nOrig
) return SQLITE_CORRUPT_BKPT
;
3800 rc
= saveAllCursors(pBt
, 0, 0);
3802 for(iFree
=nOrig
; iFree
>nFin
&& rc
==SQLITE_OK
; iFree
--){
3803 rc
= incrVacuumStep(pBt
, nFin
, iFree
, 1);
3805 if( (rc
==SQLITE_DONE
|| rc
==SQLITE_OK
) && nFree
>0 ){
3806 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
3807 put4byte(&pBt
->pPage1
->aData
[32], 0);
3808 put4byte(&pBt
->pPage1
->aData
[36], 0);
3809 put4byte(&pBt
->pPage1
->aData
[28], nFin
);
3810 pBt
->bDoTruncate
= 1;
3813 if( rc
!=SQLITE_OK
){
3814 sqlite3PagerRollback(pPager
);
3818 assert( nRef
>=sqlite3PagerRefcount(pPager
) );
3822 #else /* ifndef SQLITE_OMIT_AUTOVACUUM */
3823 # define setChildPtrmaps(x) SQLITE_OK
3827 ** This routine does the first phase of a two-phase commit. This routine
3828 ** causes a rollback journal to be created (if it does not already exist)
3829 ** and populated with enough information so that if a power loss occurs
3830 ** the database can be restored to its original state by playing back
3831 ** the journal. Then the contents of the journal are flushed out to
3832 ** the disk. After the journal is safely on oxide, the changes to the
3833 ** database are written into the database file and flushed to oxide.
3834 ** At the end of this call, the rollback journal still exists on the
3835 ** disk and we are still holding all locks, so the transaction has not
3836 ** committed. See sqlite3BtreeCommitPhaseTwo() for the second phase of the
3839 ** This call is a no-op if no write-transaction is currently active on pBt.
3841 ** Otherwise, sync the database file for the btree pBt. zMaster points to
3842 ** the name of a master journal file that should be written into the
3843 ** individual journal file, or is NULL, indicating no master journal file
3844 ** (single database transaction).
3846 ** When this is called, the master journal should already have been
3847 ** created, populated with this journal pointer and synced to disk.
3849 ** Once this is routine has returned, the only thing required to commit
3850 ** the write-transaction for this database file is to delete the journal.
3852 int sqlite3BtreeCommitPhaseOne(Btree
*p
, const char *zMaster
){
3854 if( p
->inTrans
==TRANS_WRITE
){
3855 BtShared
*pBt
= p
->pBt
;
3856 sqlite3BtreeEnter(p
);
3857 #ifndef SQLITE_OMIT_AUTOVACUUM
3858 if( pBt
->autoVacuum
){
3859 rc
= autoVacuumCommit(pBt
);
3860 if( rc
!=SQLITE_OK
){
3861 sqlite3BtreeLeave(p
);
3865 if( pBt
->bDoTruncate
){
3866 sqlite3PagerTruncateImage(pBt
->pPager
, pBt
->nPage
);
3869 rc
= sqlite3PagerCommitPhaseOne(pBt
->pPager
, zMaster
, 0);
3870 sqlite3BtreeLeave(p
);
3876 ** This function is called from both BtreeCommitPhaseTwo() and BtreeRollback()
3877 ** at the conclusion of a transaction.
3879 static void btreeEndTransaction(Btree
*p
){
3880 BtShared
*pBt
= p
->pBt
;
3881 sqlite3
*db
= p
->db
;
3882 assert( sqlite3BtreeHoldsMutex(p
) );
3884 #ifndef SQLITE_OMIT_AUTOVACUUM
3885 pBt
->bDoTruncate
= 0;
3887 if( p
->inTrans
>TRANS_NONE
&& db
->nVdbeRead
>1 ){
3888 /* If there are other active statements that belong to this database
3889 ** handle, downgrade to a read-only transaction. The other statements
3890 ** may still be reading from the database. */
3891 downgradeAllSharedCacheTableLocks(p
);
3892 p
->inTrans
= TRANS_READ
;
3894 /* If the handle had any kind of transaction open, decrement the
3895 ** transaction count of the shared btree. If the transaction count
3896 ** reaches 0, set the shared state to TRANS_NONE. The unlockBtreeIfUnused()
3897 ** call below will unlock the pager. */
3898 if( p
->inTrans
!=TRANS_NONE
){
3899 clearAllSharedCacheTableLocks(p
);
3900 pBt
->nTransaction
--;
3901 if( 0==pBt
->nTransaction
){
3902 pBt
->inTransaction
= TRANS_NONE
;
3906 /* Set the current transaction state to TRANS_NONE and unlock the
3907 ** pager if this call closed the only read or write transaction. */
3908 p
->inTrans
= TRANS_NONE
;
3909 unlockBtreeIfUnused(pBt
);
3916 ** Commit the transaction currently in progress.
3918 ** This routine implements the second phase of a 2-phase commit. The
3919 ** sqlite3BtreeCommitPhaseOne() routine does the first phase and should
3920 ** be invoked prior to calling this routine. The sqlite3BtreeCommitPhaseOne()
3921 ** routine did all the work of writing information out to disk and flushing the
3922 ** contents so that they are written onto the disk platter. All this
3923 ** routine has to do is delete or truncate or zero the header in the
3924 ** the rollback journal (which causes the transaction to commit) and
3927 ** Normally, if an error occurs while the pager layer is attempting to
3928 ** finalize the underlying journal file, this function returns an error and
3929 ** the upper layer will attempt a rollback. However, if the second argument
3930 ** is non-zero then this b-tree transaction is part of a multi-file
3931 ** transaction. In this case, the transaction has already been committed
3932 ** (by deleting a master journal file) and the caller will ignore this
3933 ** functions return code. So, even if an error occurs in the pager layer,
3934 ** reset the b-tree objects internal state to indicate that the write
3935 ** transaction has been closed. This is quite safe, as the pager will have
3936 ** transitioned to the error state.
3938 ** This will release the write lock on the database file. If there
3939 ** are no active cursors, it also releases the read lock.
3941 int sqlite3BtreeCommitPhaseTwo(Btree
*p
, int bCleanup
){
3943 if( p
->inTrans
==TRANS_NONE
) return SQLITE_OK
;
3944 sqlite3BtreeEnter(p
);
3947 /* If the handle has a write-transaction open, commit the shared-btrees
3948 ** transaction and set the shared state to TRANS_READ.
3950 if( p
->inTrans
==TRANS_WRITE
){
3952 BtShared
*pBt
= p
->pBt
;
3953 assert( pBt
->inTransaction
==TRANS_WRITE
);
3954 assert( pBt
->nTransaction
>0 );
3955 rc
= sqlite3PagerCommitPhaseTwo(pBt
->pPager
);
3956 if( rc
!=SQLITE_OK
&& bCleanup
==0 ){
3957 sqlite3BtreeLeave(p
);
3960 p
->iDataVersion
--; /* Compensate for pPager->iDataVersion++; */
3961 pBt
->inTransaction
= TRANS_READ
;
3962 btreeClearHasContent(pBt
);
3965 btreeEndTransaction(p
);
3966 sqlite3BtreeLeave(p
);
3971 ** Do both phases of a commit.
3973 int sqlite3BtreeCommit(Btree
*p
){
3975 sqlite3BtreeEnter(p
);
3976 rc
= sqlite3BtreeCommitPhaseOne(p
, 0);
3977 if( rc
==SQLITE_OK
){
3978 rc
= sqlite3BtreeCommitPhaseTwo(p
, 0);
3980 sqlite3BtreeLeave(p
);
3985 ** This routine sets the state to CURSOR_FAULT and the error
3986 ** code to errCode for every cursor on any BtShared that pBtree
3987 ** references. Or if the writeOnly flag is set to 1, then only
3988 ** trip write cursors and leave read cursors unchanged.
3990 ** Every cursor is a candidate to be tripped, including cursors
3991 ** that belong to other database connections that happen to be
3992 ** sharing the cache with pBtree.
3994 ** This routine gets called when a rollback occurs. If the writeOnly
3995 ** flag is true, then only write-cursors need be tripped - read-only
3996 ** cursors save their current positions so that they may continue
3997 ** following the rollback. Or, if writeOnly is false, all cursors are
3998 ** tripped. In general, writeOnly is false if the transaction being
3999 ** rolled back modified the database schema. In this case b-tree root
4000 ** pages may be moved or deleted from the database altogether, making
4001 ** it unsafe for read cursors to continue.
4003 ** If the writeOnly flag is true and an error is encountered while
4004 ** saving the current position of a read-only cursor, all cursors,
4005 ** including all read-cursors are tripped.
4007 ** SQLITE_OK is returned if successful, or if an error occurs while
4008 ** saving a cursor position, an SQLite error code.
4010 int sqlite3BtreeTripAllCursors(Btree
*pBtree
, int errCode
, int writeOnly
){
4014 assert( (writeOnly
==0 || writeOnly
==1) && BTCF_WriteFlag
==1 );
4016 sqlite3BtreeEnter(pBtree
);
4017 for(p
=pBtree
->pBt
->pCursor
; p
; p
=p
->pNext
){
4018 if( writeOnly
&& (p
->curFlags
& BTCF_WriteFlag
)==0 ){
4019 if( p
->eState
==CURSOR_VALID
|| p
->eState
==CURSOR_SKIPNEXT
){
4020 rc
= saveCursorPosition(p
);
4021 if( rc
!=SQLITE_OK
){
4022 (void)sqlite3BtreeTripAllCursors(pBtree
, rc
, 0);
4027 sqlite3BtreeClearCursor(p
);
4028 p
->eState
= CURSOR_FAULT
;
4029 p
->skipNext
= errCode
;
4031 btreeReleaseAllCursorPages(p
);
4033 sqlite3BtreeLeave(pBtree
);
4039 ** Rollback the transaction in progress.
4041 ** If tripCode is not SQLITE_OK then cursors will be invalidated (tripped).
4042 ** Only write cursors are tripped if writeOnly is true but all cursors are
4043 ** tripped if writeOnly is false. Any attempt to use
4044 ** a tripped cursor will result in an error.
4046 ** This will release the write lock on the database file. If there
4047 ** are no active cursors, it also releases the read lock.
4049 int sqlite3BtreeRollback(Btree
*p
, int tripCode
, int writeOnly
){
4051 BtShared
*pBt
= p
->pBt
;
4054 assert( writeOnly
==1 || writeOnly
==0 );
4055 assert( tripCode
==SQLITE_ABORT_ROLLBACK
|| tripCode
==SQLITE_OK
);
4056 sqlite3BtreeEnter(p
);
4057 if( tripCode
==SQLITE_OK
){
4058 rc
= tripCode
= saveAllCursors(pBt
, 0, 0);
4059 if( rc
) writeOnly
= 0;
4064 int rc2
= sqlite3BtreeTripAllCursors(p
, tripCode
, writeOnly
);
4065 assert( rc
==SQLITE_OK
|| (writeOnly
==0 && rc2
==SQLITE_OK
) );
4066 if( rc2
!=SQLITE_OK
) rc
= rc2
;
4070 if( p
->inTrans
==TRANS_WRITE
){
4073 assert( TRANS_WRITE
==pBt
->inTransaction
);
4074 rc2
= sqlite3PagerRollback(pBt
->pPager
);
4075 if( rc2
!=SQLITE_OK
){
4079 /* The rollback may have destroyed the pPage1->aData value. So
4080 ** call btreeGetPage() on page 1 again to make
4081 ** sure pPage1->aData is set correctly. */
4082 if( btreeGetPage(pBt
, 1, &pPage1
, 0)==SQLITE_OK
){
4083 int nPage
= get4byte(28+(u8
*)pPage1
->aData
);
4084 testcase( nPage
==0 );
4085 if( nPage
==0 ) sqlite3PagerPagecount(pBt
->pPager
, &nPage
);
4086 testcase( pBt
->nPage
!=nPage
);
4088 releasePageOne(pPage1
);
4090 assert( countValidCursors(pBt
, 1)==0 );
4091 pBt
->inTransaction
= TRANS_READ
;
4092 btreeClearHasContent(pBt
);
4095 btreeEndTransaction(p
);
4096 sqlite3BtreeLeave(p
);
4101 ** Start a statement subtransaction. The subtransaction can be rolled
4102 ** back independently of the main transaction. You must start a transaction
4103 ** before starting a subtransaction. The subtransaction is ended automatically
4104 ** if the main transaction commits or rolls back.
4106 ** Statement subtransactions are used around individual SQL statements
4107 ** that are contained within a BEGIN...COMMIT block. If a constraint
4108 ** error occurs within the statement, the effect of that one statement
4109 ** can be rolled back without having to rollback the entire transaction.
4111 ** A statement sub-transaction is implemented as an anonymous savepoint. The
4112 ** value passed as the second parameter is the total number of savepoints,
4113 ** including the new anonymous savepoint, open on the B-Tree. i.e. if there
4114 ** are no active savepoints and no other statement-transactions open,
4115 ** iStatement is 1. This anonymous savepoint can be released or rolled back
4116 ** using the sqlite3BtreeSavepoint() function.
4118 int sqlite3BtreeBeginStmt(Btree
*p
, int iStatement
){
4120 BtShared
*pBt
= p
->pBt
;
4121 sqlite3BtreeEnter(p
);
4122 assert( p
->inTrans
==TRANS_WRITE
);
4123 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
4124 assert( iStatement
>0 );
4125 assert( iStatement
>p
->db
->nSavepoint
);
4126 assert( pBt
->inTransaction
==TRANS_WRITE
);
4127 /* At the pager level, a statement transaction is a savepoint with
4128 ** an index greater than all savepoints created explicitly using
4129 ** SQL statements. It is illegal to open, release or rollback any
4130 ** such savepoints while the statement transaction savepoint is active.
4132 rc
= sqlite3PagerOpenSavepoint(pBt
->pPager
, iStatement
);
4133 sqlite3BtreeLeave(p
);
4138 ** The second argument to this function, op, is always SAVEPOINT_ROLLBACK
4139 ** or SAVEPOINT_RELEASE. This function either releases or rolls back the
4140 ** savepoint identified by parameter iSavepoint, depending on the value
4143 ** Normally, iSavepoint is greater than or equal to zero. However, if op is
4144 ** SAVEPOINT_ROLLBACK, then iSavepoint may also be -1. In this case the
4145 ** contents of the entire transaction are rolled back. This is different
4146 ** from a normal transaction rollback, as no locks are released and the
4147 ** transaction remains open.
4149 int sqlite3BtreeSavepoint(Btree
*p
, int op
, int iSavepoint
){
4151 if( p
&& p
->inTrans
==TRANS_WRITE
){
4152 BtShared
*pBt
= p
->pBt
;
4153 assert( op
==SAVEPOINT_RELEASE
|| op
==SAVEPOINT_ROLLBACK
);
4154 assert( iSavepoint
>=0 || (iSavepoint
==-1 && op
==SAVEPOINT_ROLLBACK
) );
4155 sqlite3BtreeEnter(p
);
4156 if( op
==SAVEPOINT_ROLLBACK
){
4157 rc
= saveAllCursors(pBt
, 0, 0);
4159 if( rc
==SQLITE_OK
){
4160 rc
= sqlite3PagerSavepoint(pBt
->pPager
, op
, iSavepoint
);
4162 if( rc
==SQLITE_OK
){
4163 if( iSavepoint
<0 && (pBt
->btsFlags
& BTS_INITIALLY_EMPTY
)!=0 ){
4166 rc
= newDatabase(pBt
);
4167 pBt
->nPage
= get4byte(28 + pBt
->pPage1
->aData
);
4169 /* The database size was written into the offset 28 of the header
4170 ** when the transaction started, so we know that the value at offset
4171 ** 28 is nonzero. */
4172 assert( pBt
->nPage
>0 );
4174 sqlite3BtreeLeave(p
);
4180 ** Create a new cursor for the BTree whose root is on the page
4181 ** iTable. If a read-only cursor is requested, it is assumed that
4182 ** the caller already has at least a read-only transaction open
4183 ** on the database already. If a write-cursor is requested, then
4184 ** the caller is assumed to have an open write transaction.
4186 ** If the BTREE_WRCSR bit of wrFlag is clear, then the cursor can only
4187 ** be used for reading. If the BTREE_WRCSR bit is set, then the cursor
4188 ** can be used for reading or for writing if other conditions for writing
4189 ** are also met. These are the conditions that must be met in order
4190 ** for writing to be allowed:
4192 ** 1: The cursor must have been opened with wrFlag containing BTREE_WRCSR
4194 ** 2: Other database connections that share the same pager cache
4195 ** but which are not in the READ_UNCOMMITTED state may not have
4196 ** cursors open with wrFlag==0 on the same table. Otherwise
4197 ** the changes made by this write cursor would be visible to
4198 ** the read cursors in the other database connection.
4200 ** 3: The database must be writable (not on read-only media)
4202 ** 4: There must be an active transaction.
4204 ** The BTREE_FORDELETE bit of wrFlag may optionally be set if BTREE_WRCSR
4205 ** is set. If FORDELETE is set, that is a hint to the implementation that
4206 ** this cursor will only be used to seek to and delete entries of an index
4207 ** as part of a larger DELETE statement. The FORDELETE hint is not used by
4208 ** this implementation. But in a hypothetical alternative storage engine
4209 ** in which index entries are automatically deleted when corresponding table
4210 ** rows are deleted, the FORDELETE flag is a hint that all SEEK and DELETE
4211 ** operations on this cursor can be no-ops and all READ operations can
4212 ** return a null row (2-bytes: 0x01 0x00).
4214 ** No checking is done to make sure that page iTable really is the
4215 ** root page of a b-tree. If it is not, then the cursor acquired
4216 ** will not work correctly.
4218 ** It is assumed that the sqlite3BtreeCursorZero() has been called
4219 ** on pCur to initialize the memory space prior to invoking this routine.
4221 static int btreeCursor(
4222 Btree
*p
, /* The btree */
4223 int iTable
, /* Root page of table to open */
4224 int wrFlag
, /* 1 to write. 0 read-only */
4225 struct KeyInfo
*pKeyInfo
, /* First arg to comparison function */
4226 BtCursor
*pCur
/* Space for new cursor */
4228 BtShared
*pBt
= p
->pBt
; /* Shared b-tree handle */
4229 BtCursor
*pX
; /* Looping over other all cursors */
4231 assert( sqlite3BtreeHoldsMutex(p
) );
4233 || wrFlag
==BTREE_WRCSR
4234 || wrFlag
==(BTREE_WRCSR
|BTREE_FORDELETE
)
4237 /* The following assert statements verify that if this is a sharable
4238 ** b-tree database, the connection is holding the required table locks,
4239 ** and that no other connection has any open cursor that conflicts with
4241 assert( hasSharedCacheTableLock(p
, iTable
, pKeyInfo
!=0, (wrFlag
?2:1)) );
4242 assert( wrFlag
==0 || !hasReadConflicts(p
, iTable
) );
4244 /* Assert that the caller has opened the required transaction. */
4245 assert( p
->inTrans
>TRANS_NONE
);
4246 assert( wrFlag
==0 || p
->inTrans
==TRANS_WRITE
);
4247 assert( pBt
->pPage1
&& pBt
->pPage1
->aData
);
4248 assert( wrFlag
==0 || (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
4251 allocateTempSpace(pBt
);
4252 if( pBt
->pTmpSpace
==0 ) return SQLITE_NOMEM_BKPT
;
4254 if( iTable
==1 && btreePagecount(pBt
)==0 ){
4255 assert( wrFlag
==0 );
4259 /* Now that no other errors can occur, finish filling in the BtCursor
4260 ** variables and link the cursor into the BtShared list. */
4261 pCur
->pgnoRoot
= (Pgno
)iTable
;
4263 pCur
->pKeyInfo
= pKeyInfo
;
4266 pCur
->curFlags
= wrFlag
? BTCF_WriteFlag
: 0;
4267 pCur
->curPagerFlags
= wrFlag
? 0 : PAGER_GET_READONLY
;
4268 /* If there are two or more cursors on the same btree, then all such
4269 ** cursors *must* have the BTCF_Multiple flag set. */
4270 for(pX
=pBt
->pCursor
; pX
; pX
=pX
->pNext
){
4271 if( pX
->pgnoRoot
==(Pgno
)iTable
){
4272 pX
->curFlags
|= BTCF_Multiple
;
4273 pCur
->curFlags
|= BTCF_Multiple
;
4276 pCur
->pNext
= pBt
->pCursor
;
4277 pBt
->pCursor
= pCur
;
4278 pCur
->eState
= CURSOR_INVALID
;
4281 int sqlite3BtreeCursor(
4282 Btree
*p
, /* The btree */
4283 int iTable
, /* Root page of table to open */
4284 int wrFlag
, /* 1 to write. 0 read-only */
4285 struct KeyInfo
*pKeyInfo
, /* First arg to xCompare() */
4286 BtCursor
*pCur
/* Write new cursor here */
4290 rc
= SQLITE_CORRUPT_BKPT
;
4292 sqlite3BtreeEnter(p
);
4293 rc
= btreeCursor(p
, iTable
, wrFlag
, pKeyInfo
, pCur
);
4294 sqlite3BtreeLeave(p
);
4300 ** Return the size of a BtCursor object in bytes.
4302 ** This interfaces is needed so that users of cursors can preallocate
4303 ** sufficient storage to hold a cursor. The BtCursor object is opaque
4304 ** to users so they cannot do the sizeof() themselves - they must call
4307 int sqlite3BtreeCursorSize(void){
4308 return ROUND8(sizeof(BtCursor
));
4312 ** Initialize memory that will be converted into a BtCursor object.
4314 ** The simple approach here would be to memset() the entire object
4315 ** to zero. But it turns out that the apPage[] and aiIdx[] arrays
4316 ** do not need to be zeroed and they are large, so we can save a lot
4317 ** of run-time by skipping the initialization of those elements.
4319 void sqlite3BtreeCursorZero(BtCursor
*p
){
4320 memset(p
, 0, offsetof(BtCursor
, iPage
));
4324 ** Close a cursor. The read lock on the database file is released
4325 ** when the last cursor is closed.
4327 int sqlite3BtreeCloseCursor(BtCursor
*pCur
){
4328 Btree
*pBtree
= pCur
->pBtree
;
4330 BtShared
*pBt
= pCur
->pBt
;
4331 sqlite3BtreeEnter(pBtree
);
4332 assert( pBt
->pCursor
!=0 );
4333 if( pBt
->pCursor
==pCur
){
4334 pBt
->pCursor
= pCur
->pNext
;
4336 BtCursor
*pPrev
= pBt
->pCursor
;
4338 if( pPrev
->pNext
==pCur
){
4339 pPrev
->pNext
= pCur
->pNext
;
4342 pPrev
= pPrev
->pNext
;
4343 }while( ALWAYS(pPrev
) );
4345 btreeReleaseAllCursorPages(pCur
);
4346 unlockBtreeIfUnused(pBt
);
4347 sqlite3_free(pCur
->aOverflow
);
4348 sqlite3_free(pCur
->pKey
);
4349 sqlite3BtreeLeave(pBtree
);
4355 ** Make sure the BtCursor* given in the argument has a valid
4356 ** BtCursor.info structure. If it is not already valid, call
4357 ** btreeParseCell() to fill it in.
4359 ** BtCursor.info is a cache of the information in the current cell.
4360 ** Using this cache reduces the number of calls to btreeParseCell().
4363 static void assertCellInfo(BtCursor
*pCur
){
4365 memset(&info
, 0, sizeof(info
));
4366 btreeParseCell(pCur
->pPage
, pCur
->ix
, &info
);
4367 assert( CORRUPT_DB
|| memcmp(&info
, &pCur
->info
, sizeof(info
))==0 );
4370 #define assertCellInfo(x)
4372 static SQLITE_NOINLINE
void getCellInfo(BtCursor
*pCur
){
4373 if( pCur
->info
.nSize
==0 ){
4374 pCur
->curFlags
|= BTCF_ValidNKey
;
4375 btreeParseCell(pCur
->pPage
,pCur
->ix
,&pCur
->info
);
4377 assertCellInfo(pCur
);
4381 #ifndef NDEBUG /* The next routine used only within assert() statements */
4383 ** Return true if the given BtCursor is valid. A valid cursor is one
4384 ** that is currently pointing to a row in a (non-empty) table.
4385 ** This is a verification routine is used only within assert() statements.
4387 int sqlite3BtreeCursorIsValid(BtCursor
*pCur
){
4388 return pCur
&& pCur
->eState
==CURSOR_VALID
;
4391 int sqlite3BtreeCursorIsValidNN(BtCursor
*pCur
){
4393 return pCur
->eState
==CURSOR_VALID
;
4397 ** Return the value of the integer key or "rowid" for a table btree.
4398 ** This routine is only valid for a cursor that is pointing into a
4399 ** ordinary table btree. If the cursor points to an index btree or
4400 ** is invalid, the result of this routine is undefined.
4402 i64
sqlite3BtreeIntegerKey(BtCursor
*pCur
){
4403 assert( cursorHoldsMutex(pCur
) );
4404 assert( pCur
->eState
==CURSOR_VALID
);
4405 assert( pCur
->curIntKey
);
4407 return pCur
->info
.nKey
;
4411 ** Return the number of bytes of payload for the entry that pCur is
4412 ** currently pointing to. For table btrees, this will be the amount
4413 ** of data. For index btrees, this will be the size of the key.
4415 ** The caller must guarantee that the cursor is pointing to a non-NULL
4416 ** valid entry. In other words, the calling procedure must guarantee
4417 ** that the cursor has Cursor.eState==CURSOR_VALID.
4419 u32
sqlite3BtreePayloadSize(BtCursor
*pCur
){
4420 assert( cursorHoldsMutex(pCur
) );
4421 assert( pCur
->eState
==CURSOR_VALID
);
4423 return pCur
->info
.nPayload
;
4427 ** Given the page number of an overflow page in the database (parameter
4428 ** ovfl), this function finds the page number of the next page in the
4429 ** linked list of overflow pages. If possible, it uses the auto-vacuum
4430 ** pointer-map data instead of reading the content of page ovfl to do so.
4432 ** If an error occurs an SQLite error code is returned. Otherwise:
4434 ** The page number of the next overflow page in the linked list is
4435 ** written to *pPgnoNext. If page ovfl is the last page in its linked
4436 ** list, *pPgnoNext is set to zero.
4438 ** If ppPage is not NULL, and a reference to the MemPage object corresponding
4439 ** to page number pOvfl was obtained, then *ppPage is set to point to that
4440 ** reference. It is the responsibility of the caller to call releasePage()
4441 ** on *ppPage to free the reference. In no reference was obtained (because
4442 ** the pointer-map was used to obtain the value for *pPgnoNext), then
4443 ** *ppPage is set to zero.
4445 static int getOverflowPage(
4446 BtShared
*pBt
, /* The database file */
4447 Pgno ovfl
, /* Current overflow page number */
4448 MemPage
**ppPage
, /* OUT: MemPage handle (may be NULL) */
4449 Pgno
*pPgnoNext
/* OUT: Next overflow page number */
4455 assert( sqlite3_mutex_held(pBt
->mutex
) );
4458 #ifndef SQLITE_OMIT_AUTOVACUUM
4459 /* Try to find the next page in the overflow list using the
4460 ** autovacuum pointer-map pages. Guess that the next page in
4461 ** the overflow list is page number (ovfl+1). If that guess turns
4462 ** out to be wrong, fall back to loading the data of page
4463 ** number ovfl to determine the next page number.
4465 if( pBt
->autoVacuum
){
4467 Pgno iGuess
= ovfl
+1;
4470 while( PTRMAP_ISPAGE(pBt
, iGuess
) || iGuess
==PENDING_BYTE_PAGE(pBt
) ){
4474 if( iGuess
<=btreePagecount(pBt
) ){
4475 rc
= ptrmapGet(pBt
, iGuess
, &eType
, &pgno
);
4476 if( rc
==SQLITE_OK
&& eType
==PTRMAP_OVERFLOW2
&& pgno
==ovfl
){
4484 assert( next
==0 || rc
==SQLITE_DONE
);
4485 if( rc
==SQLITE_OK
){
4486 rc
= btreeGetPage(pBt
, ovfl
, &pPage
, (ppPage
==0) ? PAGER_GET_READONLY
: 0);
4487 assert( rc
==SQLITE_OK
|| pPage
==0 );
4488 if( rc
==SQLITE_OK
){
4489 next
= get4byte(pPage
->aData
);
4499 return (rc
==SQLITE_DONE
? SQLITE_OK
: rc
);
4503 ** Copy data from a buffer to a page, or from a page to a buffer.
4505 ** pPayload is a pointer to data stored on database page pDbPage.
4506 ** If argument eOp is false, then nByte bytes of data are copied
4507 ** from pPayload to the buffer pointed at by pBuf. If eOp is true,
4508 ** then sqlite3PagerWrite() is called on pDbPage and nByte bytes
4509 ** of data are copied from the buffer pBuf to pPayload.
4511 ** SQLITE_OK is returned on success, otherwise an error code.
4513 static int copyPayload(
4514 void *pPayload
, /* Pointer to page data */
4515 void *pBuf
, /* Pointer to buffer */
4516 int nByte
, /* Number of bytes to copy */
4517 int eOp
, /* 0 -> copy from page, 1 -> copy to page */
4518 DbPage
*pDbPage
/* Page containing pPayload */
4521 /* Copy data from buffer to page (a write operation) */
4522 int rc
= sqlite3PagerWrite(pDbPage
);
4523 if( rc
!=SQLITE_OK
){
4526 memcpy(pPayload
, pBuf
, nByte
);
4528 /* Copy data from page to buffer (a read operation) */
4529 memcpy(pBuf
, pPayload
, nByte
);
4535 ** This function is used to read or overwrite payload information
4536 ** for the entry that the pCur cursor is pointing to. The eOp
4537 ** argument is interpreted as follows:
4539 ** 0: The operation is a read. Populate the overflow cache.
4540 ** 1: The operation is a write. Populate the overflow cache.
4542 ** A total of "amt" bytes are read or written beginning at "offset".
4543 ** Data is read to or from the buffer pBuf.
4545 ** The content being read or written might appear on the main page
4546 ** or be scattered out on multiple overflow pages.
4548 ** If the current cursor entry uses one or more overflow pages
4549 ** this function may allocate space for and lazily populate
4550 ** the overflow page-list cache array (BtCursor.aOverflow).
4551 ** Subsequent calls use this cache to make seeking to the supplied offset
4554 ** Once an overflow page-list cache has been allocated, it must be
4555 ** invalidated if some other cursor writes to the same table, or if
4556 ** the cursor is moved to a different row. Additionally, in auto-vacuum
4557 ** mode, the following events may invalidate an overflow page-list cache.
4559 ** * An incremental vacuum,
4560 ** * A commit in auto_vacuum="full" mode,
4561 ** * Creating a table (may require moving an overflow page).
4563 static int accessPayload(
4564 BtCursor
*pCur
, /* Cursor pointing to entry to read from */
4565 u32 offset
, /* Begin reading this far into payload */
4566 u32 amt
, /* Read this many bytes */
4567 unsigned char *pBuf
, /* Write the bytes into this buffer */
4568 int eOp
/* zero to read. non-zero to write. */
4570 unsigned char *aPayload
;
4573 MemPage
*pPage
= pCur
->pPage
; /* Btree page of current entry */
4574 BtShared
*pBt
= pCur
->pBt
; /* Btree this cursor belongs to */
4575 #ifdef SQLITE_DIRECT_OVERFLOW_READ
4576 unsigned char * const pBufStart
= pBuf
; /* Start of original out buffer */
4580 assert( eOp
==0 || eOp
==1 );
4581 assert( pCur
->eState
==CURSOR_VALID
);
4582 assert( pCur
->ix
<pPage
->nCell
);
4583 assert( cursorHoldsMutex(pCur
) );
4586 aPayload
= pCur
->info
.pPayload
;
4587 assert( offset
+amt
<= pCur
->info
.nPayload
);
4589 assert( aPayload
> pPage
->aData
);
4590 if( (uptr
)(aPayload
- pPage
->aData
) > (pBt
->usableSize
- pCur
->info
.nLocal
) ){
4591 /* Trying to read or write past the end of the data is an error. The
4592 ** conditional above is really:
4593 ** &aPayload[pCur->info.nLocal] > &pPage->aData[pBt->usableSize]
4594 ** but is recast into its current form to avoid integer overflow problems
4596 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
4599 /* Check if data must be read/written to/from the btree page itself. */
4600 if( offset
<pCur
->info
.nLocal
){
4602 if( a
+offset
>pCur
->info
.nLocal
){
4603 a
= pCur
->info
.nLocal
- offset
;
4605 rc
= copyPayload(&aPayload
[offset
], pBuf
, a
, eOp
, pPage
->pDbPage
);
4610 offset
-= pCur
->info
.nLocal
;
4614 if( rc
==SQLITE_OK
&& amt
>0 ){
4615 const u32 ovflSize
= pBt
->usableSize
- 4; /* Bytes content per ovfl page */
4618 nextPage
= get4byte(&aPayload
[pCur
->info
.nLocal
]);
4620 /* If the BtCursor.aOverflow[] has not been allocated, allocate it now.
4622 ** The aOverflow[] array is sized at one entry for each overflow page
4623 ** in the overflow chain. The page number of the first overflow page is
4624 ** stored in aOverflow[0], etc. A value of 0 in the aOverflow[] array
4625 ** means "not yet known" (the cache is lazily populated).
4627 if( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 ){
4628 int nOvfl
= (pCur
->info
.nPayload
-pCur
->info
.nLocal
+ovflSize
-1)/ovflSize
;
4629 if( nOvfl
>pCur
->nOvflAlloc
){
4630 Pgno
*aNew
= (Pgno
*)sqlite3Realloc(
4631 pCur
->aOverflow
, nOvfl
*2*sizeof(Pgno
)
4634 return SQLITE_NOMEM_BKPT
;
4636 pCur
->nOvflAlloc
= nOvfl
*2;
4637 pCur
->aOverflow
= aNew
;
4640 memset(pCur
->aOverflow
, 0, nOvfl
*sizeof(Pgno
));
4641 pCur
->curFlags
|= BTCF_ValidOvfl
;
4643 /* If the overflow page-list cache has been allocated and the
4644 ** entry for the first required overflow page is valid, skip
4647 if( pCur
->aOverflow
[offset
/ovflSize
] ){
4648 iIdx
= (offset
/ovflSize
);
4649 nextPage
= pCur
->aOverflow
[iIdx
];
4650 offset
= (offset
%ovflSize
);
4654 assert( rc
==SQLITE_OK
&& amt
>0 );
4656 /* If required, populate the overflow page-list cache. */
4657 assert( pCur
->aOverflow
[iIdx
]==0
4658 || pCur
->aOverflow
[iIdx
]==nextPage
4660 pCur
->aOverflow
[iIdx
] = nextPage
;
4662 if( offset
>=ovflSize
){
4663 /* The only reason to read this page is to obtain the page
4664 ** number for the next page in the overflow chain. The page
4665 ** data is not required. So first try to lookup the overflow
4666 ** page-list cache, if any, then fall back to the getOverflowPage()
4669 assert( pCur
->curFlags
& BTCF_ValidOvfl
);
4670 assert( pCur
->pBtree
->db
==pBt
->db
);
4671 if( pCur
->aOverflow
[iIdx
+1] ){
4672 nextPage
= pCur
->aOverflow
[iIdx
+1];
4674 rc
= getOverflowPage(pBt
, nextPage
, 0, &nextPage
);
4678 /* Need to read this page properly. It contains some of the
4679 ** range of data that is being read (eOp==0) or written (eOp!=0).
4681 #ifdef SQLITE_DIRECT_OVERFLOW_READ
4682 sqlite3_file
*fd
; /* File from which to do direct overflow read */
4685 if( a
+ offset
> ovflSize
){
4686 a
= ovflSize
- offset
;
4689 #ifdef SQLITE_DIRECT_OVERFLOW_READ
4690 /* If all the following are true:
4692 ** 1) this is a read operation, and
4693 ** 2) data is required from the start of this overflow page, and
4694 ** 3) there is no open write-transaction, and
4695 ** 4) the database is file-backed, and
4696 ** 5) the page is not in the WAL file
4697 ** 6) at least 4 bytes have already been read into the output buffer
4699 ** then data can be read directly from the database file into the
4700 ** output buffer, bypassing the page-cache altogether. This speeds
4701 ** up loading large records that span many overflow pages.
4703 if( eOp
==0 /* (1) */
4704 && offset
==0 /* (2) */
4705 && pBt
->inTransaction
==TRANS_READ
/* (3) */
4706 && (fd
= sqlite3PagerFile(pBt
->pPager
))->pMethods
/* (4) */
4707 && 0==sqlite3PagerUseWal(pBt
->pPager
, nextPage
) /* (5) */
4708 && &pBuf
[-4]>=pBufStart
/* (6) */
4711 u8
*aWrite
= &pBuf
[-4];
4712 assert( aWrite
>=pBufStart
); /* due to (6) */
4713 memcpy(aSave
, aWrite
, 4);
4714 rc
= sqlite3OsRead(fd
, aWrite
, a
+4, (i64
)pBt
->pageSize
*(nextPage
-1));
4715 nextPage
= get4byte(aWrite
);
4716 memcpy(aWrite
, aSave
, 4);
4722 rc
= sqlite3PagerGet(pBt
->pPager
, nextPage
, &pDbPage
,
4723 (eOp
==0 ? PAGER_GET_READONLY
: 0)
4725 if( rc
==SQLITE_OK
){
4726 aPayload
= sqlite3PagerGetData(pDbPage
);
4727 nextPage
= get4byte(aPayload
);
4728 rc
= copyPayload(&aPayload
[offset
+4], pBuf
, a
, eOp
, pDbPage
);
4729 sqlite3PagerUnref(pDbPage
);
4734 if( amt
==0 ) return rc
;
4742 if( rc
==SQLITE_OK
&& amt
>0 ){
4743 /* Overflow chain ends prematurely */
4744 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
4750 ** Read part of the payload for the row at which that cursor pCur is currently
4751 ** pointing. "amt" bytes will be transferred into pBuf[]. The transfer
4752 ** begins at "offset".
4754 ** pCur can be pointing to either a table or an index b-tree.
4755 ** If pointing to a table btree, then the content section is read. If
4756 ** pCur is pointing to an index b-tree then the key section is read.
4758 ** For sqlite3BtreePayload(), the caller must ensure that pCur is pointing
4759 ** to a valid row in the table. For sqlite3BtreePayloadChecked(), the
4760 ** cursor might be invalid or might need to be restored before being read.
4762 ** Return SQLITE_OK on success or an error code if anything goes
4763 ** wrong. An error is returned if "offset+amt" is larger than
4764 ** the available payload.
4766 int sqlite3BtreePayload(BtCursor
*pCur
, u32 offset
, u32 amt
, void *pBuf
){
4767 assert( cursorHoldsMutex(pCur
) );
4768 assert( pCur
->eState
==CURSOR_VALID
);
4769 assert( pCur
->iPage
>=0 && pCur
->pPage
);
4770 assert( pCur
->ix
<pCur
->pPage
->nCell
);
4771 return accessPayload(pCur
, offset
, amt
, (unsigned char*)pBuf
, 0);
4775 ** This variant of sqlite3BtreePayload() works even if the cursor has not
4776 ** in the CURSOR_VALID state. It is only used by the sqlite3_blob_read()
4779 #ifndef SQLITE_OMIT_INCRBLOB
4780 static SQLITE_NOINLINE
int accessPayloadChecked(
4787 if ( pCur
->eState
==CURSOR_INVALID
){
4788 return SQLITE_ABORT
;
4790 assert( cursorOwnsBtShared(pCur
) );
4791 rc
= btreeRestoreCursorPosition(pCur
);
4792 return rc
? rc
: accessPayload(pCur
, offset
, amt
, pBuf
, 0);
4794 int sqlite3BtreePayloadChecked(BtCursor
*pCur
, u32 offset
, u32 amt
, void *pBuf
){
4795 if( pCur
->eState
==CURSOR_VALID
){
4796 assert( cursorOwnsBtShared(pCur
) );
4797 return accessPayload(pCur
, offset
, amt
, pBuf
, 0);
4799 return accessPayloadChecked(pCur
, offset
, amt
, pBuf
);
4802 #endif /* SQLITE_OMIT_INCRBLOB */
4805 ** Return a pointer to payload information from the entry that the
4806 ** pCur cursor is pointing to. The pointer is to the beginning of
4807 ** the key if index btrees (pPage->intKey==0) and is the data for
4808 ** table btrees (pPage->intKey==1). The number of bytes of available
4809 ** key/data is written into *pAmt. If *pAmt==0, then the value
4810 ** returned will not be a valid pointer.
4812 ** This routine is an optimization. It is common for the entire key
4813 ** and data to fit on the local page and for there to be no overflow
4814 ** pages. When that is so, this routine can be used to access the
4815 ** key and data without making a copy. If the key and/or data spills
4816 ** onto overflow pages, then accessPayload() must be used to reassemble
4817 ** the key/data and copy it into a preallocated buffer.
4819 ** The pointer returned by this routine looks directly into the cached
4820 ** page of the database. The data might change or move the next time
4821 ** any btree routine is called.
4823 static const void *fetchPayload(
4824 BtCursor
*pCur
, /* Cursor pointing to entry to read from */
4825 u32
*pAmt
/* Write the number of available bytes here */
4828 assert( pCur
!=0 && pCur
->iPage
>=0 && pCur
->pPage
);
4829 assert( pCur
->eState
==CURSOR_VALID
);
4830 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
4831 assert( cursorOwnsBtShared(pCur
) );
4832 assert( pCur
->ix
<pCur
->pPage
->nCell
);
4833 assert( pCur
->info
.nSize
>0 );
4834 assert( pCur
->info
.pPayload
>pCur
->pPage
->aData
|| CORRUPT_DB
);
4835 assert( pCur
->info
.pPayload
<pCur
->pPage
->aDataEnd
||CORRUPT_DB
);
4836 amt
= pCur
->info
.nLocal
;
4837 if( amt
>(int)(pCur
->pPage
->aDataEnd
- pCur
->info
.pPayload
) ){
4838 /* There is too little space on the page for the expected amount
4839 ** of local content. Database must be corrupt. */
4840 assert( CORRUPT_DB
);
4841 amt
= MAX(0, (int)(pCur
->pPage
->aDataEnd
- pCur
->info
.pPayload
));
4844 return (void*)pCur
->info
.pPayload
;
4849 ** For the entry that cursor pCur is point to, return as
4850 ** many bytes of the key or data as are available on the local
4851 ** b-tree page. Write the number of available bytes into *pAmt.
4853 ** The pointer returned is ephemeral. The key/data may move
4854 ** or be destroyed on the next call to any Btree routine,
4855 ** including calls from other threads against the same cache.
4856 ** Hence, a mutex on the BtShared should be held prior to calling
4859 ** These routines is used to get quick access to key and data
4860 ** in the common case where no overflow pages are used.
4862 const void *sqlite3BtreePayloadFetch(BtCursor
*pCur
, u32
*pAmt
){
4863 return fetchPayload(pCur
, pAmt
);
4868 ** Move the cursor down to a new child page. The newPgno argument is the
4869 ** page number of the child page to move to.
4871 ** This function returns SQLITE_CORRUPT if the page-header flags field of
4872 ** the new child page does not match the flags field of the parent (i.e.
4873 ** if an intkey page appears to be the parent of a non-intkey page, or
4876 static int moveToChild(BtCursor
*pCur
, u32 newPgno
){
4877 BtShared
*pBt
= pCur
->pBt
;
4879 assert( cursorOwnsBtShared(pCur
) );
4880 assert( pCur
->eState
==CURSOR_VALID
);
4881 assert( pCur
->iPage
<BTCURSOR_MAX_DEPTH
);
4882 assert( pCur
->iPage
>=0 );
4883 if( pCur
->iPage
>=(BTCURSOR_MAX_DEPTH
-1) ){
4884 return SQLITE_CORRUPT_BKPT
;
4886 pCur
->info
.nSize
= 0;
4887 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
4888 pCur
->aiIdx
[pCur
->iPage
] = pCur
->ix
;
4889 pCur
->apPage
[pCur
->iPage
] = pCur
->pPage
;
4892 return getAndInitPage(pBt
, newPgno
, &pCur
->pPage
, pCur
, pCur
->curPagerFlags
);
4897 ** Page pParent is an internal (non-leaf) tree page. This function
4898 ** asserts that page number iChild is the left-child if the iIdx'th
4899 ** cell in page pParent. Or, if iIdx is equal to the total number of
4900 ** cells in pParent, that page number iChild is the right-child of
4903 static void assertParentIndex(MemPage
*pParent
, int iIdx
, Pgno iChild
){
4904 if( CORRUPT_DB
) return; /* The conditions tested below might not be true
4905 ** in a corrupt database */
4906 assert( iIdx
<=pParent
->nCell
);
4907 if( iIdx
==pParent
->nCell
){
4908 assert( get4byte(&pParent
->aData
[pParent
->hdrOffset
+8])==iChild
);
4910 assert( get4byte(findCell(pParent
, iIdx
))==iChild
);
4914 # define assertParentIndex(x,y,z)
4918 ** Move the cursor up to the parent page.
4920 ** pCur->idx is set to the cell index that contains the pointer
4921 ** to the page we are coming from. If we are coming from the
4922 ** right-most child page then pCur->idx is set to one more than
4923 ** the largest cell index.
4925 static void moveToParent(BtCursor
*pCur
){
4927 assert( cursorOwnsBtShared(pCur
) );
4928 assert( pCur
->eState
==CURSOR_VALID
);
4929 assert( pCur
->iPage
>0 );
4930 assert( pCur
->pPage
);
4932 pCur
->apPage
[pCur
->iPage
-1],
4933 pCur
->aiIdx
[pCur
->iPage
-1],
4936 testcase( pCur
->aiIdx
[pCur
->iPage
-1] > pCur
->apPage
[pCur
->iPage
-1]->nCell
);
4937 pCur
->info
.nSize
= 0;
4938 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
4939 pCur
->ix
= pCur
->aiIdx
[pCur
->iPage
-1];
4940 pLeaf
= pCur
->pPage
;
4941 pCur
->pPage
= pCur
->apPage
[--pCur
->iPage
];
4942 releasePageNotNull(pLeaf
);
4946 ** Move the cursor to point to the root page of its b-tree structure.
4948 ** If the table has a virtual root page, then the cursor is moved to point
4949 ** to the virtual root page instead of the actual root page. A table has a
4950 ** virtual root page when the actual root page contains no cells and a
4951 ** single child page. This can only happen with the table rooted at page 1.
4953 ** If the b-tree structure is empty, the cursor state is set to
4954 ** CURSOR_INVALID and this routine returns SQLITE_EMPTY. Otherwise,
4955 ** the cursor is set to point to the first cell located on the root
4956 ** (or virtual root) page and the cursor state is set to CURSOR_VALID.
4958 ** If this function returns successfully, it may be assumed that the
4959 ** page-header flags indicate that the [virtual] root-page is the expected
4960 ** kind of b-tree page (i.e. if when opening the cursor the caller did not
4961 ** specify a KeyInfo structure the flags byte is set to 0x05 or 0x0D,
4962 ** indicating a table b-tree, or if the caller did specify a KeyInfo
4963 ** structure the flags byte is set to 0x02 or 0x0A, indicating an index
4966 static int moveToRoot(BtCursor
*pCur
){
4970 assert( cursorOwnsBtShared(pCur
) );
4971 assert( CURSOR_INVALID
< CURSOR_REQUIRESEEK
);
4972 assert( CURSOR_VALID
< CURSOR_REQUIRESEEK
);
4973 assert( CURSOR_FAULT
> CURSOR_REQUIRESEEK
);
4974 assert( pCur
->eState
< CURSOR_REQUIRESEEK
|| pCur
->iPage
<0 );
4975 assert( pCur
->pgnoRoot
>0 || pCur
->iPage
<0 );
4977 if( pCur
->iPage
>=0 ){
4979 releasePageNotNull(pCur
->pPage
);
4980 while( --pCur
->iPage
){
4981 releasePageNotNull(pCur
->apPage
[pCur
->iPage
]);
4983 pCur
->pPage
= pCur
->apPage
[0];
4986 }else if( pCur
->pgnoRoot
==0 ){
4987 pCur
->eState
= CURSOR_INVALID
;
4988 return SQLITE_EMPTY
;
4990 assert( pCur
->iPage
==(-1) );
4991 if( pCur
->eState
>=CURSOR_REQUIRESEEK
){
4992 if( pCur
->eState
==CURSOR_FAULT
){
4993 assert( pCur
->skipNext
!=SQLITE_OK
);
4994 return pCur
->skipNext
;
4996 sqlite3BtreeClearCursor(pCur
);
4998 rc
= getAndInitPage(pCur
->pBtree
->pBt
, pCur
->pgnoRoot
, &pCur
->pPage
,
4999 0, pCur
->curPagerFlags
);
5000 if( rc
!=SQLITE_OK
){
5001 pCur
->eState
= CURSOR_INVALID
;
5005 pCur
->curIntKey
= pCur
->pPage
->intKey
;
5007 pRoot
= pCur
->pPage
;
5008 assert( pRoot
->pgno
==pCur
->pgnoRoot
);
5010 /* If pCur->pKeyInfo is not NULL, then the caller that opened this cursor
5011 ** expected to open it on an index b-tree. Otherwise, if pKeyInfo is
5012 ** NULL, the caller expects a table b-tree. If this is not the case,
5013 ** return an SQLITE_CORRUPT error.
5015 ** Earlier versions of SQLite assumed that this test could not fail
5016 ** if the root page was already loaded when this function was called (i.e.
5017 ** if pCur->iPage>=0). But this is not so if the database is corrupted
5018 ** in such a way that page pRoot is linked into a second b-tree table
5019 ** (or the freelist). */
5020 assert( pRoot
->intKey
==1 || pRoot
->intKey
==0 );
5021 if( pRoot
->isInit
==0 || (pCur
->pKeyInfo
==0)!=pRoot
->intKey
){
5022 return SQLITE_CORRUPT_PGNO(pCur
->pPage
->pgno
);
5027 pCur
->info
.nSize
= 0;
5028 pCur
->curFlags
&= ~(BTCF_AtLast
|BTCF_ValidNKey
|BTCF_ValidOvfl
);
5030 pRoot
= pCur
->pPage
;
5031 if( pRoot
->nCell
>0 ){
5032 pCur
->eState
= CURSOR_VALID
;
5033 }else if( !pRoot
->leaf
){
5035 if( pRoot
->pgno
!=1 ) return SQLITE_CORRUPT_BKPT
;
5036 subpage
= get4byte(&pRoot
->aData
[pRoot
->hdrOffset
+8]);
5037 pCur
->eState
= CURSOR_VALID
;
5038 rc
= moveToChild(pCur
, subpage
);
5040 pCur
->eState
= CURSOR_INVALID
;
5047 ** Move the cursor down to the left-most leaf entry beneath the
5048 ** entry to which it is currently pointing.
5050 ** The left-most leaf is the one with the smallest key - the first
5051 ** in ascending order.
5053 static int moveToLeftmost(BtCursor
*pCur
){
5058 assert( cursorOwnsBtShared(pCur
) );
5059 assert( pCur
->eState
==CURSOR_VALID
);
5060 while( rc
==SQLITE_OK
&& !(pPage
= pCur
->pPage
)->leaf
){
5061 assert( pCur
->ix
<pPage
->nCell
);
5062 pgno
= get4byte(findCell(pPage
, pCur
->ix
));
5063 rc
= moveToChild(pCur
, pgno
);
5069 ** Move the cursor down to the right-most leaf entry beneath the
5070 ** page to which it is currently pointing. Notice the difference
5071 ** between moveToLeftmost() and moveToRightmost(). moveToLeftmost()
5072 ** finds the left-most entry beneath the *entry* whereas moveToRightmost()
5073 ** finds the right-most entry beneath the *page*.
5075 ** The right-most entry is the one with the largest key - the last
5076 ** key in ascending order.
5078 static int moveToRightmost(BtCursor
*pCur
){
5083 assert( cursorOwnsBtShared(pCur
) );
5084 assert( pCur
->eState
==CURSOR_VALID
);
5085 while( !(pPage
= pCur
->pPage
)->leaf
){
5086 pgno
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
5087 pCur
->ix
= pPage
->nCell
;
5088 rc
= moveToChild(pCur
, pgno
);
5091 pCur
->ix
= pPage
->nCell
-1;
5092 assert( pCur
->info
.nSize
==0 );
5093 assert( (pCur
->curFlags
& BTCF_ValidNKey
)==0 );
5097 /* Move the cursor to the first entry in the table. Return SQLITE_OK
5098 ** on success. Set *pRes to 0 if the cursor actually points to something
5099 ** or set *pRes to 1 if the table is empty.
5101 int sqlite3BtreeFirst(BtCursor
*pCur
, int *pRes
){
5104 assert( cursorOwnsBtShared(pCur
) );
5105 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5106 rc
= moveToRoot(pCur
);
5107 if( rc
==SQLITE_OK
){
5108 assert( pCur
->pPage
->nCell
>0 );
5110 rc
= moveToLeftmost(pCur
);
5111 }else if( rc
==SQLITE_EMPTY
){
5112 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5119 /* Move the cursor to the last entry in the table. Return SQLITE_OK
5120 ** on success. Set *pRes to 0 if the cursor actually points to something
5121 ** or set *pRes to 1 if the table is empty.
5123 int sqlite3BtreeLast(BtCursor
*pCur
, int *pRes
){
5126 assert( cursorOwnsBtShared(pCur
) );
5127 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5129 /* If the cursor already points to the last entry, this is a no-op. */
5130 if( CURSOR_VALID
==pCur
->eState
&& (pCur
->curFlags
& BTCF_AtLast
)!=0 ){
5132 /* This block serves to assert() that the cursor really does point
5133 ** to the last entry in the b-tree. */
5135 for(ii
=0; ii
<pCur
->iPage
; ii
++){
5136 assert( pCur
->aiIdx
[ii
]==pCur
->apPage
[ii
]->nCell
);
5138 assert( pCur
->ix
==pCur
->pPage
->nCell
-1 );
5139 assert( pCur
->pPage
->leaf
);
5144 rc
= moveToRoot(pCur
);
5145 if( rc
==SQLITE_OK
){
5146 assert( pCur
->eState
==CURSOR_VALID
);
5148 rc
= moveToRightmost(pCur
);
5149 if( rc
==SQLITE_OK
){
5150 pCur
->curFlags
|= BTCF_AtLast
;
5152 pCur
->curFlags
&= ~BTCF_AtLast
;
5154 }else if( rc
==SQLITE_EMPTY
){
5155 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5162 /* Move the cursor so that it points to an entry near the key
5163 ** specified by pIdxKey or intKey. Return a success code.
5165 ** For INTKEY tables, the intKey parameter is used. pIdxKey
5166 ** must be NULL. For index tables, pIdxKey is used and intKey
5169 ** If an exact match is not found, then the cursor is always
5170 ** left pointing at a leaf page which would hold the entry if it
5171 ** were present. The cursor might point to an entry that comes
5172 ** before or after the key.
5174 ** An integer is written into *pRes which is the result of
5175 ** comparing the key with the entry to which the cursor is
5176 ** pointing. The meaning of the integer written into
5177 ** *pRes is as follows:
5179 ** *pRes<0 The cursor is left pointing at an entry that
5180 ** is smaller than intKey/pIdxKey or if the table is empty
5181 ** and the cursor is therefore left point to nothing.
5183 ** *pRes==0 The cursor is left pointing at an entry that
5184 ** exactly matches intKey/pIdxKey.
5186 ** *pRes>0 The cursor is left pointing at an entry that
5187 ** is larger than intKey/pIdxKey.
5189 ** For index tables, the pIdxKey->eqSeen field is set to 1 if there
5190 ** exists an entry in the table that exactly matches pIdxKey.
5192 int sqlite3BtreeMovetoUnpacked(
5193 BtCursor
*pCur
, /* The cursor to be moved */
5194 UnpackedRecord
*pIdxKey
, /* Unpacked index key */
5195 i64 intKey
, /* The table key */
5196 int biasRight
, /* If true, bias the search to the high end */
5197 int *pRes
/* Write search results here */
5200 RecordCompare xRecordCompare
;
5202 assert( cursorOwnsBtShared(pCur
) );
5203 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5205 assert( (pIdxKey
==0)==(pCur
->pKeyInfo
==0) );
5206 assert( pCur
->eState
!=CURSOR_VALID
|| (pIdxKey
==0)==(pCur
->curIntKey
!=0) );
5208 /* If the cursor is already positioned at the point we are trying
5209 ** to move to, then just return without doing any work */
5211 && pCur
->eState
==CURSOR_VALID
&& (pCur
->curFlags
& BTCF_ValidNKey
)!=0
5213 if( pCur
->info
.nKey
==intKey
){
5217 if( pCur
->info
.nKey
<intKey
){
5218 if( (pCur
->curFlags
& BTCF_AtLast
)!=0 ){
5222 /* If the requested key is one more than the previous key, then
5223 ** try to get there using sqlite3BtreeNext() rather than a full
5224 ** binary search. This is an optimization only. The correct answer
5225 ** is still obtained without this case, only a little more slowely */
5226 if( pCur
->info
.nKey
+1==intKey
&& !pCur
->skipNext
){
5228 rc
= sqlite3BtreeNext(pCur
, 0);
5229 if( rc
==SQLITE_OK
){
5231 if( pCur
->info
.nKey
==intKey
){
5234 }else if( rc
==SQLITE_DONE
){
5244 xRecordCompare
= sqlite3VdbeFindCompare(pIdxKey
);
5245 pIdxKey
->errCode
= 0;
5246 assert( pIdxKey
->default_rc
==1
5247 || pIdxKey
->default_rc
==0
5248 || pIdxKey
->default_rc
==-1
5251 xRecordCompare
= 0; /* All keys are integers */
5254 rc
= moveToRoot(pCur
);
5256 if( rc
==SQLITE_EMPTY
){
5257 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5263 assert( pCur
->pPage
);
5264 assert( pCur
->pPage
->isInit
);
5265 assert( pCur
->eState
==CURSOR_VALID
);
5266 assert( pCur
->pPage
->nCell
> 0 );
5267 assert( pCur
->iPage
==0 || pCur
->apPage
[0]->intKey
==pCur
->curIntKey
);
5268 assert( pCur
->curIntKey
|| pIdxKey
);
5270 int lwr
, upr
, idx
, c
;
5272 MemPage
*pPage
= pCur
->pPage
;
5273 u8
*pCell
; /* Pointer to current cell in pPage */
5275 /* pPage->nCell must be greater than zero. If this is the root-page
5276 ** the cursor would have been INVALID above and this for(;;) loop
5277 ** not run. If this is not the root-page, then the moveToChild() routine
5278 ** would have already detected db corruption. Similarly, pPage must
5279 ** be the right kind (index or table) of b-tree page. Otherwise
5280 ** a moveToChild() or moveToRoot() call would have detected corruption. */
5281 assert( pPage
->nCell
>0 );
5282 assert( pPage
->intKey
==(pIdxKey
==0) );
5284 upr
= pPage
->nCell
-1;
5285 assert( biasRight
==0 || biasRight
==1 );
5286 idx
= upr
>>(1-biasRight
); /* idx = biasRight ? upr : (lwr+upr)/2; */
5287 pCur
->ix
= (u16
)idx
;
5288 if( xRecordCompare
==0 ){
5291 pCell
= findCellPastPtr(pPage
, idx
);
5292 if( pPage
->intKeyLeaf
){
5293 while( 0x80 <= *(pCell
++) ){
5294 if( pCell
>=pPage
->aDataEnd
){
5295 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
5299 getVarint(pCell
, (u64
*)&nCellKey
);
5300 if( nCellKey
<intKey
){
5302 if( lwr
>upr
){ c
= -1; break; }
5303 }else if( nCellKey
>intKey
){
5305 if( lwr
>upr
){ c
= +1; break; }
5307 assert( nCellKey
==intKey
);
5308 pCur
->ix
= (u16
)idx
;
5311 goto moveto_next_layer
;
5313 pCur
->curFlags
|= BTCF_ValidNKey
;
5314 pCur
->info
.nKey
= nCellKey
;
5315 pCur
->info
.nSize
= 0;
5320 assert( lwr
+upr
>=0 );
5321 idx
= (lwr
+upr
)>>1; /* idx = (lwr+upr)/2; */
5325 int nCell
; /* Size of the pCell cell in bytes */
5326 pCell
= findCellPastPtr(pPage
, idx
);
5328 /* The maximum supported page-size is 65536 bytes. This means that
5329 ** the maximum number of record bytes stored on an index B-Tree
5330 ** page is less than 16384 bytes and may be stored as a 2-byte
5331 ** varint. This information is used to attempt to avoid parsing
5332 ** the entire cell by checking for the cases where the record is
5333 ** stored entirely within the b-tree page by inspecting the first
5334 ** 2 bytes of the cell.
5337 if( nCell
<=pPage
->max1bytePayload
){
5338 /* This branch runs if the record-size field of the cell is a
5339 ** single byte varint and the record fits entirely on the main
5341 testcase( pCell
+nCell
+1==pPage
->aDataEnd
);
5342 c
= xRecordCompare(nCell
, (void*)&pCell
[1], pIdxKey
);
5343 }else if( !(pCell
[1] & 0x80)
5344 && (nCell
= ((nCell
&0x7f)<<7) + pCell
[1])<=pPage
->maxLocal
5346 /* The record-size field is a 2 byte varint and the record
5347 ** fits entirely on the main b-tree page. */
5348 testcase( pCell
+nCell
+2==pPage
->aDataEnd
);
5349 c
= xRecordCompare(nCell
, (void*)&pCell
[2], pIdxKey
);
5351 /* The record flows over onto one or more overflow pages. In
5352 ** this case the whole cell needs to be parsed, a buffer allocated
5353 ** and accessPayload() used to retrieve the record into the
5354 ** buffer before VdbeRecordCompare() can be called.
5356 ** If the record is corrupt, the xRecordCompare routine may read
5357 ** up to two varints past the end of the buffer. An extra 18
5358 ** bytes of padding is allocated at the end of the buffer in
5359 ** case this happens. */
5361 u8
* const pCellBody
= pCell
- pPage
->childPtrSize
;
5362 pPage
->xParseCell(pPage
, pCellBody
, &pCur
->info
);
5363 nCell
= (int)pCur
->info
.nKey
;
5364 testcase( nCell
<0 ); /* True if key size is 2^32 or more */
5365 testcase( nCell
==0 ); /* Invalid key size: 0x80 0x80 0x00 */
5366 testcase( nCell
==1 ); /* Invalid key size: 0x80 0x80 0x01 */
5367 testcase( nCell
==2 ); /* Minimum legal index key size */
5369 rc
= SQLITE_CORRUPT_PGNO(pPage
->pgno
);
5372 pCellKey
= sqlite3Malloc( nCell
+18 );
5374 rc
= SQLITE_NOMEM_BKPT
;
5377 pCur
->ix
= (u16
)idx
;
5378 rc
= accessPayload(pCur
, 0, nCell
, (unsigned char*)pCellKey
, 0);
5379 pCur
->curFlags
&= ~BTCF_ValidOvfl
;
5381 sqlite3_free(pCellKey
);
5384 c
= xRecordCompare(nCell
, pCellKey
, pIdxKey
);
5385 sqlite3_free(pCellKey
);
5388 (pIdxKey
->errCode
!=SQLITE_CORRUPT
|| c
==0)
5389 && (pIdxKey
->errCode
!=SQLITE_NOMEM
|| pCur
->pBtree
->db
->mallocFailed
)
5399 pCur
->ix
= (u16
)idx
;
5400 if( pIdxKey
->errCode
) rc
= SQLITE_CORRUPT_BKPT
;
5403 if( lwr
>upr
) break;
5404 assert( lwr
+upr
>=0 );
5405 idx
= (lwr
+upr
)>>1; /* idx = (lwr+upr)/2 */
5408 assert( lwr
==upr
+1 || (pPage
->intKey
&& !pPage
->leaf
) );
5409 assert( pPage
->isInit
);
5411 assert( pCur
->ix
<pCur
->pPage
->nCell
);
5412 pCur
->ix
= (u16
)idx
;
5418 if( lwr
>=pPage
->nCell
){
5419 chldPg
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
5421 chldPg
= get4byte(findCell(pPage
, lwr
));
5423 pCur
->ix
= (u16
)lwr
;
5424 rc
= moveToChild(pCur
, chldPg
);
5428 pCur
->info
.nSize
= 0;
5429 assert( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 );
5435 ** Return TRUE if the cursor is not pointing at an entry of the table.
5437 ** TRUE will be returned after a call to sqlite3BtreeNext() moves
5438 ** past the last entry in the table or sqlite3BtreePrev() moves past
5439 ** the first entry. TRUE is also returned if the table is empty.
5441 int sqlite3BtreeEof(BtCursor
*pCur
){
5442 /* TODO: What if the cursor is in CURSOR_REQUIRESEEK but all table entries
5443 ** have been deleted? This API will need to change to return an error code
5444 ** as well as the boolean result value.
5446 return (CURSOR_VALID
!=pCur
->eState
);
5450 ** Return an estimate for the number of rows in the table that pCur is
5451 ** pointing to. Return a negative number if no estimate is currently
5454 i64
sqlite3BtreeRowCountEst(BtCursor
*pCur
){
5458 assert( cursorOwnsBtShared(pCur
) );
5459 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5461 /* Currently this interface is only called by the OP_IfSmaller
5462 ** opcode, and it that case the cursor will always be valid and
5463 ** will always point to a leaf node. */
5464 if( NEVER(pCur
->eState
!=CURSOR_VALID
) ) return -1;
5465 if( NEVER(pCur
->pPage
->leaf
==0) ) return -1;
5467 n
= pCur
->pPage
->nCell
;
5468 for(i
=0; i
<pCur
->iPage
; i
++){
5469 n
*= pCur
->apPage
[i
]->nCell
;
5475 ** Advance the cursor to the next entry in the database.
5478 ** SQLITE_OK success
5479 ** SQLITE_DONE cursor is already pointing at the last element
5480 ** otherwise some kind of error occurred
5482 ** The main entry point is sqlite3BtreeNext(). That routine is optimized
5483 ** for the common case of merely incrementing the cell counter BtCursor.aiIdx
5484 ** to the next cell on the current page. The (slower) btreeNext() helper
5485 ** routine is called when it is necessary to move to a different page or
5486 ** to restore the cursor.
5488 ** If bit 0x01 of the F argument in sqlite3BtreeNext(C,F) is 1, then the
5489 ** cursor corresponds to an SQL index and this routine could have been
5490 ** skipped if the SQL index had been a unique index. The F argument
5491 ** is a hint to the implement. SQLite btree implementation does not use
5492 ** this hint, but COMDB2 does.
5494 static SQLITE_NOINLINE
int btreeNext(BtCursor
*pCur
){
5499 assert( cursorOwnsBtShared(pCur
) );
5500 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5501 if( pCur
->eState
!=CURSOR_VALID
){
5502 assert( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 );
5503 rc
= restoreCursorPosition(pCur
);
5504 if( rc
!=SQLITE_OK
){
5507 if( CURSOR_INVALID
==pCur
->eState
){
5510 if( pCur
->skipNext
){
5511 assert( pCur
->eState
==CURSOR_VALID
|| pCur
->eState
==CURSOR_SKIPNEXT
);
5512 pCur
->eState
= CURSOR_VALID
;
5513 if( pCur
->skipNext
>0 ){
5521 pPage
= pCur
->pPage
;
5523 assert( pPage
->isInit
);
5525 /* If the database file is corrupt, it is possible for the value of idx
5526 ** to be invalid here. This can only occur if a second cursor modifies
5527 ** the page while cursor pCur is holding a reference to it. Which can
5528 ** only happen if the database is corrupt in such a way as to link the
5529 ** page into more than one b-tree structure. */
5530 testcase( idx
>pPage
->nCell
);
5532 if( idx
>=pPage
->nCell
){
5534 rc
= moveToChild(pCur
, get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]));
5536 return moveToLeftmost(pCur
);
5539 if( pCur
->iPage
==0 ){
5540 pCur
->eState
= CURSOR_INVALID
;
5544 pPage
= pCur
->pPage
;
5545 }while( pCur
->ix
>=pPage
->nCell
);
5546 if( pPage
->intKey
){
5547 return sqlite3BtreeNext(pCur
, 0);
5555 return moveToLeftmost(pCur
);
5558 int sqlite3BtreeNext(BtCursor
*pCur
, int flags
){
5560 UNUSED_PARAMETER( flags
); /* Used in COMDB2 but not native SQLite */
5561 assert( cursorOwnsBtShared(pCur
) );
5562 assert( flags
==0 || flags
==1 );
5563 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5564 pCur
->info
.nSize
= 0;
5565 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
5566 if( pCur
->eState
!=CURSOR_VALID
) return btreeNext(pCur
);
5567 pPage
= pCur
->pPage
;
5568 if( (++pCur
->ix
)>=pPage
->nCell
){
5570 return btreeNext(pCur
);
5575 return moveToLeftmost(pCur
);
5580 ** Step the cursor to the back to the previous entry in the database.
5583 ** SQLITE_OK success
5584 ** SQLITE_DONE the cursor is already on the first element of the table
5585 ** otherwise some kind of error occurred
5587 ** The main entry point is sqlite3BtreePrevious(). That routine is optimized
5588 ** for the common case of merely decrementing the cell counter BtCursor.aiIdx
5589 ** to the previous cell on the current page. The (slower) btreePrevious()
5590 ** helper routine is called when it is necessary to move to a different page
5591 ** or to restore the cursor.
5593 ** If bit 0x01 of the F argument to sqlite3BtreePrevious(C,F) is 1, then
5594 ** the cursor corresponds to an SQL index and this routine could have been
5595 ** skipped if the SQL index had been a unique index. The F argument is a
5596 ** hint to the implement. The native SQLite btree implementation does not
5597 ** use this hint, but COMDB2 does.
5599 static SQLITE_NOINLINE
int btreePrevious(BtCursor
*pCur
){
5603 assert( cursorOwnsBtShared(pCur
) );
5604 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5605 assert( (pCur
->curFlags
& (BTCF_AtLast
|BTCF_ValidOvfl
|BTCF_ValidNKey
))==0 );
5606 assert( pCur
->info
.nSize
==0 );
5607 if( pCur
->eState
!=CURSOR_VALID
){
5608 rc
= restoreCursorPosition(pCur
);
5609 if( rc
!=SQLITE_OK
){
5612 if( CURSOR_INVALID
==pCur
->eState
){
5615 if( pCur
->skipNext
){
5616 assert( pCur
->eState
==CURSOR_VALID
|| pCur
->eState
==CURSOR_SKIPNEXT
);
5617 pCur
->eState
= CURSOR_VALID
;
5618 if( pCur
->skipNext
<0 ){
5626 pPage
= pCur
->pPage
;
5627 assert( pPage
->isInit
);
5630 rc
= moveToChild(pCur
, get4byte(findCell(pPage
, idx
)));
5632 rc
= moveToRightmost(pCur
);
5634 while( pCur
->ix
==0 ){
5635 if( pCur
->iPage
==0 ){
5636 pCur
->eState
= CURSOR_INVALID
;
5641 assert( pCur
->info
.nSize
==0 );
5642 assert( (pCur
->curFlags
& (BTCF_ValidOvfl
))==0 );
5645 pPage
= pCur
->pPage
;
5646 if( pPage
->intKey
&& !pPage
->leaf
){
5647 rc
= sqlite3BtreePrevious(pCur
, 0);
5654 int sqlite3BtreePrevious(BtCursor
*pCur
, int flags
){
5655 assert( cursorOwnsBtShared(pCur
) );
5656 assert( flags
==0 || flags
==1 );
5657 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5658 UNUSED_PARAMETER( flags
); /* Used in COMDB2 but not native SQLite */
5659 pCur
->curFlags
&= ~(BTCF_AtLast
|BTCF_ValidOvfl
|BTCF_ValidNKey
);
5660 pCur
->info
.nSize
= 0;
5661 if( pCur
->eState
!=CURSOR_VALID
5663 || pCur
->pPage
->leaf
==0
5665 return btreePrevious(pCur
);
5672 ** Allocate a new page from the database file.
5674 ** The new page is marked as dirty. (In other words, sqlite3PagerWrite()
5675 ** has already been called on the new page.) The new page has also
5676 ** been referenced and the calling routine is responsible for calling
5677 ** sqlite3PagerUnref() on the new page when it is done.
5679 ** SQLITE_OK is returned on success. Any other return value indicates
5680 ** an error. *ppPage is set to NULL in the event of an error.
5682 ** If the "nearby" parameter is not 0, then an effort is made to
5683 ** locate a page close to the page number "nearby". This can be used in an
5684 ** attempt to keep related pages close to each other in the database file,
5685 ** which in turn can make database access faster.
5687 ** If the eMode parameter is BTALLOC_EXACT and the nearby page exists
5688 ** anywhere on the free-list, then it is guaranteed to be returned. If
5689 ** eMode is BTALLOC_LT then the page returned will be less than or equal
5690 ** to nearby if any such page exists. If eMode is BTALLOC_ANY then there
5691 ** are no restrictions on which page is returned.
5693 static int allocateBtreePage(
5694 BtShared
*pBt
, /* The btree */
5695 MemPage
**ppPage
, /* Store pointer to the allocated page here */
5696 Pgno
*pPgno
, /* Store the page number here */
5697 Pgno nearby
, /* Search for a page near this one */
5698 u8 eMode
/* BTALLOC_EXACT, BTALLOC_LT, or BTALLOC_ANY */
5702 u32 n
; /* Number of pages on the freelist */
5703 u32 k
; /* Number of leaves on the trunk of the freelist */
5704 MemPage
*pTrunk
= 0;
5705 MemPage
*pPrevTrunk
= 0;
5706 Pgno mxPage
; /* Total size of the database file */
5708 assert( sqlite3_mutex_held(pBt
->mutex
) );
5709 assert( eMode
==BTALLOC_ANY
|| (nearby
>0 && IfNotOmitAV(pBt
->autoVacuum
)) );
5710 pPage1
= pBt
->pPage1
;
5711 mxPage
= btreePagecount(pBt
);
5712 /* EVIDENCE-OF: R-05119-02637 The 4-byte big-endian integer at offset 36
5713 ** stores stores the total number of pages on the freelist. */
5714 n
= get4byte(&pPage1
->aData
[36]);
5715 testcase( n
==mxPage
-1 );
5717 return SQLITE_CORRUPT_BKPT
;
5720 /* There are pages on the freelist. Reuse one of those pages. */
5722 u8 searchList
= 0; /* If the free-list must be searched for 'nearby' */
5723 u32 nSearch
= 0; /* Count of the number of search attempts */
5725 /* If eMode==BTALLOC_EXACT and a query of the pointer-map
5726 ** shows that the page 'nearby' is somewhere on the free-list, then
5727 ** the entire-list will be searched for that page.
5729 #ifndef SQLITE_OMIT_AUTOVACUUM
5730 if( eMode
==BTALLOC_EXACT
){
5731 if( nearby
<=mxPage
){
5734 assert( pBt
->autoVacuum
);
5735 rc
= ptrmapGet(pBt
, nearby
, &eType
, 0);
5737 if( eType
==PTRMAP_FREEPAGE
){
5741 }else if( eMode
==BTALLOC_LE
){
5746 /* Decrement the free-list count by 1. Set iTrunk to the index of the
5747 ** first free-list trunk page. iPrevTrunk is initially 1.
5749 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
5751 put4byte(&pPage1
->aData
[36], n
-1);
5753 /* The code within this loop is run only once if the 'searchList' variable
5754 ** is not true. Otherwise, it runs once for each trunk-page on the
5755 ** free-list until the page 'nearby' is located (eMode==BTALLOC_EXACT)
5756 ** or until a page less than 'nearby' is located (eMode==BTALLOC_LT)
5759 pPrevTrunk
= pTrunk
;
5761 /* EVIDENCE-OF: R-01506-11053 The first integer on a freelist trunk page
5762 ** is the page number of the next freelist trunk page in the list or
5763 ** zero if this is the last freelist trunk page. */
5764 iTrunk
= get4byte(&pPrevTrunk
->aData
[0]);
5766 /* EVIDENCE-OF: R-59841-13798 The 4-byte big-endian integer at offset 32
5767 ** stores the page number of the first page of the freelist, or zero if
5768 ** the freelist is empty. */
5769 iTrunk
= get4byte(&pPage1
->aData
[32]);
5771 testcase( iTrunk
==mxPage
);
5772 if( iTrunk
>mxPage
|| nSearch
++ > n
){
5773 rc
= SQLITE_CORRUPT_PGNO(pPrevTrunk
? pPrevTrunk
->pgno
: 1);
5775 rc
= btreeGetUnusedPage(pBt
, iTrunk
, &pTrunk
, 0);
5779 goto end_allocate_page
;
5781 assert( pTrunk
!=0 );
5782 assert( pTrunk
->aData
!=0 );
5783 /* EVIDENCE-OF: R-13523-04394 The second integer on a freelist trunk page
5784 ** is the number of leaf page pointers to follow. */
5785 k
= get4byte(&pTrunk
->aData
[4]);
5786 if( k
==0 && !searchList
){
5787 /* The trunk has no leaves and the list is not being searched.
5788 ** So extract the trunk page itself and use it as the newly
5789 ** allocated page */
5790 assert( pPrevTrunk
==0 );
5791 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
5793 goto end_allocate_page
;
5796 memcpy(&pPage1
->aData
[32], &pTrunk
->aData
[0], 4);
5799 TRACE(("ALLOCATE: %d trunk - %d free pages left\n", *pPgno
, n
-1));
5800 }else if( k
>(u32
)(pBt
->usableSize
/4 - 2) ){
5801 /* Value of k is out of range. Database corruption */
5802 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
5803 goto end_allocate_page
;
5804 #ifndef SQLITE_OMIT_AUTOVACUUM
5805 }else if( searchList
5806 && (nearby
==iTrunk
|| (iTrunk
<nearby
&& eMode
==BTALLOC_LE
))
5808 /* The list is being searched and this trunk page is the page
5809 ** to allocate, regardless of whether it has leaves.
5814 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
5816 goto end_allocate_page
;
5820 memcpy(&pPage1
->aData
[32], &pTrunk
->aData
[0], 4);
5822 rc
= sqlite3PagerWrite(pPrevTrunk
->pDbPage
);
5823 if( rc
!=SQLITE_OK
){
5824 goto end_allocate_page
;
5826 memcpy(&pPrevTrunk
->aData
[0], &pTrunk
->aData
[0], 4);
5829 /* The trunk page is required by the caller but it contains
5830 ** pointers to free-list leaves. The first leaf becomes a trunk
5831 ** page in this case.
5834 Pgno iNewTrunk
= get4byte(&pTrunk
->aData
[8]);
5835 if( iNewTrunk
>mxPage
){
5836 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
5837 goto end_allocate_page
;
5839 testcase( iNewTrunk
==mxPage
);
5840 rc
= btreeGetUnusedPage(pBt
, iNewTrunk
, &pNewTrunk
, 0);
5841 if( rc
!=SQLITE_OK
){
5842 goto end_allocate_page
;
5844 rc
= sqlite3PagerWrite(pNewTrunk
->pDbPage
);
5845 if( rc
!=SQLITE_OK
){
5846 releasePage(pNewTrunk
);
5847 goto end_allocate_page
;
5849 memcpy(&pNewTrunk
->aData
[0], &pTrunk
->aData
[0], 4);
5850 put4byte(&pNewTrunk
->aData
[4], k
-1);
5851 memcpy(&pNewTrunk
->aData
[8], &pTrunk
->aData
[12], (k
-1)*4);
5852 releasePage(pNewTrunk
);
5854 assert( sqlite3PagerIswriteable(pPage1
->pDbPage
) );
5855 put4byte(&pPage1
->aData
[32], iNewTrunk
);
5857 rc
= sqlite3PagerWrite(pPrevTrunk
->pDbPage
);
5859 goto end_allocate_page
;
5861 put4byte(&pPrevTrunk
->aData
[0], iNewTrunk
);
5865 TRACE(("ALLOCATE: %d trunk - %d free pages left\n", *pPgno
, n
-1));
5868 /* Extract a leaf from the trunk */
5871 unsigned char *aData
= pTrunk
->aData
;
5875 if( eMode
==BTALLOC_LE
){
5877 iPage
= get4byte(&aData
[8+i
*4]);
5878 if( iPage
<=nearby
){
5885 dist
= sqlite3AbsInt32(get4byte(&aData
[8]) - nearby
);
5887 int d2
= sqlite3AbsInt32(get4byte(&aData
[8+i
*4]) - nearby
);
5898 iPage
= get4byte(&aData
[8+closest
*4]);
5899 testcase( iPage
==mxPage
);
5901 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
5902 goto end_allocate_page
;
5904 testcase( iPage
==mxPage
);
5906 || (iPage
==nearby
|| (iPage
<nearby
&& eMode
==BTALLOC_LE
))
5910 TRACE(("ALLOCATE: %d was leaf %d of %d on trunk %d"
5911 ": %d more free pages\n",
5912 *pPgno
, closest
+1, k
, pTrunk
->pgno
, n
-1));
5913 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
5914 if( rc
) goto end_allocate_page
;
5916 memcpy(&aData
[8+closest
*4], &aData
[4+k
*4], 4);
5918 put4byte(&aData
[4], k
-1);
5919 noContent
= !btreeGetHasContent(pBt
, *pPgno
)? PAGER_GET_NOCONTENT
: 0;
5920 rc
= btreeGetUnusedPage(pBt
, *pPgno
, ppPage
, noContent
);
5921 if( rc
==SQLITE_OK
){
5922 rc
= sqlite3PagerWrite((*ppPage
)->pDbPage
);
5923 if( rc
!=SQLITE_OK
){
5924 releasePage(*ppPage
);
5931 releasePage(pPrevTrunk
);
5933 }while( searchList
);
5935 /* There are no pages on the freelist, so append a new page to the
5938 ** Normally, new pages allocated by this block can be requested from the
5939 ** pager layer with the 'no-content' flag set. This prevents the pager
5940 ** from trying to read the pages content from disk. However, if the
5941 ** current transaction has already run one or more incremental-vacuum
5942 ** steps, then the page we are about to allocate may contain content
5943 ** that is required in the event of a rollback. In this case, do
5944 ** not set the no-content flag. This causes the pager to load and journal
5945 ** the current page content before overwriting it.
5947 ** Note that the pager will not actually attempt to load or journal
5948 ** content for any page that really does lie past the end of the database
5949 ** file on disk. So the effects of disabling the no-content optimization
5950 ** here are confined to those pages that lie between the end of the
5951 ** database image and the end of the database file.
5953 int bNoContent
= (0==IfNotOmitAV(pBt
->bDoTruncate
))? PAGER_GET_NOCONTENT
:0;
5955 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
5958 if( pBt
->nPage
==PENDING_BYTE_PAGE(pBt
) ) pBt
->nPage
++;
5960 #ifndef SQLITE_OMIT_AUTOVACUUM
5961 if( pBt
->autoVacuum
&& PTRMAP_ISPAGE(pBt
, pBt
->nPage
) ){
5962 /* If *pPgno refers to a pointer-map page, allocate two new pages
5963 ** at the end of the file instead of one. The first allocated page
5964 ** becomes a new pointer-map page, the second is used by the caller.
5967 TRACE(("ALLOCATE: %d from end of file (pointer-map page)\n", pBt
->nPage
));
5968 assert( pBt
->nPage
!=PENDING_BYTE_PAGE(pBt
) );
5969 rc
= btreeGetUnusedPage(pBt
, pBt
->nPage
, &pPg
, bNoContent
);
5970 if( rc
==SQLITE_OK
){
5971 rc
= sqlite3PagerWrite(pPg
->pDbPage
);
5976 if( pBt
->nPage
==PENDING_BYTE_PAGE(pBt
) ){ pBt
->nPage
++; }
5979 put4byte(28 + (u8
*)pBt
->pPage1
->aData
, pBt
->nPage
);
5980 *pPgno
= pBt
->nPage
;
5982 assert( *pPgno
!=PENDING_BYTE_PAGE(pBt
) );
5983 rc
= btreeGetUnusedPage(pBt
, *pPgno
, ppPage
, bNoContent
);
5985 rc
= sqlite3PagerWrite((*ppPage
)->pDbPage
);
5986 if( rc
!=SQLITE_OK
){
5987 releasePage(*ppPage
);
5990 TRACE(("ALLOCATE: %d from end of file\n", *pPgno
));
5993 assert( *pPgno
!=PENDING_BYTE_PAGE(pBt
) );
5996 releasePage(pTrunk
);
5997 releasePage(pPrevTrunk
);
5998 assert( rc
!=SQLITE_OK
|| sqlite3PagerPageRefcount((*ppPage
)->pDbPage
)<=1 );
5999 assert( rc
!=SQLITE_OK
|| (*ppPage
)->isInit
==0 );
6004 ** This function is used to add page iPage to the database file free-list.
6005 ** It is assumed that the page is not already a part of the free-list.
6007 ** The value passed as the second argument to this function is optional.
6008 ** If the caller happens to have a pointer to the MemPage object
6009 ** corresponding to page iPage handy, it may pass it as the second value.
6010 ** Otherwise, it may pass NULL.
6012 ** If a pointer to a MemPage object is passed as the second argument,
6013 ** its reference count is not altered by this function.
6015 static int freePage2(BtShared
*pBt
, MemPage
*pMemPage
, Pgno iPage
){
6016 MemPage
*pTrunk
= 0; /* Free-list trunk page */
6017 Pgno iTrunk
= 0; /* Page number of free-list trunk page */
6018 MemPage
*pPage1
= pBt
->pPage1
; /* Local reference to page 1 */
6019 MemPage
*pPage
; /* Page being freed. May be NULL. */
6020 int rc
; /* Return Code */
6021 int nFree
; /* Initial number of pages on free-list */
6023 assert( sqlite3_mutex_held(pBt
->mutex
) );
6024 assert( CORRUPT_DB
|| iPage
>1 );
6025 assert( !pMemPage
|| pMemPage
->pgno
==iPage
);
6027 if( iPage
<2 ) return SQLITE_CORRUPT_BKPT
;
6030 sqlite3PagerRef(pPage
->pDbPage
);
6032 pPage
= btreePageLookup(pBt
, iPage
);
6035 /* Increment the free page count on pPage1 */
6036 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
6037 if( rc
) goto freepage_out
;
6038 nFree
= get4byte(&pPage1
->aData
[36]);
6039 put4byte(&pPage1
->aData
[36], nFree
+1);
6041 if( pBt
->btsFlags
& BTS_SECURE_DELETE
){
6042 /* If the secure_delete option is enabled, then
6043 ** always fully overwrite deleted information with zeros.
6045 if( (!pPage
&& ((rc
= btreeGetPage(pBt
, iPage
, &pPage
, 0))!=0) )
6046 || ((rc
= sqlite3PagerWrite(pPage
->pDbPage
))!=0)
6050 memset(pPage
->aData
, 0, pPage
->pBt
->pageSize
);
6053 /* If the database supports auto-vacuum, write an entry in the pointer-map
6054 ** to indicate that the page is free.
6057 ptrmapPut(pBt
, iPage
, PTRMAP_FREEPAGE
, 0, &rc
);
6058 if( rc
) goto freepage_out
;
6061 /* Now manipulate the actual database free-list structure. There are two
6062 ** possibilities. If the free-list is currently empty, or if the first
6063 ** trunk page in the free-list is full, then this page will become a
6064 ** new free-list trunk page. Otherwise, it will become a leaf of the
6065 ** first trunk page in the current free-list. This block tests if it
6066 ** is possible to add the page as a new free-list leaf.
6069 u32 nLeaf
; /* Initial number of leaf cells on trunk page */
6071 iTrunk
= get4byte(&pPage1
->aData
[32]);
6072 rc
= btreeGetPage(pBt
, iTrunk
, &pTrunk
, 0);
6073 if( rc
!=SQLITE_OK
){
6077 nLeaf
= get4byte(&pTrunk
->aData
[4]);
6078 assert( pBt
->usableSize
>32 );
6079 if( nLeaf
> (u32
)pBt
->usableSize
/4 - 2 ){
6080 rc
= SQLITE_CORRUPT_BKPT
;
6083 if( nLeaf
< (u32
)pBt
->usableSize
/4 - 8 ){
6084 /* In this case there is room on the trunk page to insert the page
6085 ** being freed as a new leaf.
6087 ** Note that the trunk page is not really full until it contains
6088 ** usableSize/4 - 2 entries, not usableSize/4 - 8 entries as we have
6089 ** coded. But due to a coding error in versions of SQLite prior to
6090 ** 3.6.0, databases with freelist trunk pages holding more than
6091 ** usableSize/4 - 8 entries will be reported as corrupt. In order
6092 ** to maintain backwards compatibility with older versions of SQLite,
6093 ** we will continue to restrict the number of entries to usableSize/4 - 8
6094 ** for now. At some point in the future (once everyone has upgraded
6095 ** to 3.6.0 or later) we should consider fixing the conditional above
6096 ** to read "usableSize/4-2" instead of "usableSize/4-8".
6098 ** EVIDENCE-OF: R-19920-11576 However, newer versions of SQLite still
6099 ** avoid using the last six entries in the freelist trunk page array in
6100 ** order that database files created by newer versions of SQLite can be
6101 ** read by older versions of SQLite.
6103 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
6104 if( rc
==SQLITE_OK
){
6105 put4byte(&pTrunk
->aData
[4], nLeaf
+1);
6106 put4byte(&pTrunk
->aData
[8+nLeaf
*4], iPage
);
6107 if( pPage
&& (pBt
->btsFlags
& BTS_SECURE_DELETE
)==0 ){
6108 sqlite3PagerDontWrite(pPage
->pDbPage
);
6110 rc
= btreeSetHasContent(pBt
, iPage
);
6112 TRACE(("FREE-PAGE: %d leaf on trunk page %d\n",pPage
->pgno
,pTrunk
->pgno
));
6117 /* If control flows to this point, then it was not possible to add the
6118 ** the page being freed as a leaf page of the first trunk in the free-list.
6119 ** Possibly because the free-list is empty, or possibly because the
6120 ** first trunk in the free-list is full. Either way, the page being freed
6121 ** will become the new first trunk page in the free-list.
6123 if( pPage
==0 && SQLITE_OK
!=(rc
= btreeGetPage(pBt
, iPage
, &pPage
, 0)) ){
6126 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
6127 if( rc
!=SQLITE_OK
){
6130 put4byte(pPage
->aData
, iTrunk
);
6131 put4byte(&pPage
->aData
[4], 0);
6132 put4byte(&pPage1
->aData
[32], iPage
);
6133 TRACE(("FREE-PAGE: %d new trunk page replacing %d\n", pPage
->pgno
, iTrunk
));
6140 releasePage(pTrunk
);
6143 static void freePage(MemPage
*pPage
, int *pRC
){
6144 if( (*pRC
)==SQLITE_OK
){
6145 *pRC
= freePage2(pPage
->pBt
, pPage
, pPage
->pgno
);
6150 ** Free any overflow pages associated with the given Cell. Write the
6151 ** local Cell size (the number of bytes on the original page, omitting
6152 ** overflow) into *pnSize.
6154 static int clearCell(
6155 MemPage
*pPage
, /* The page that contains the Cell */
6156 unsigned char *pCell
, /* First byte of the Cell */
6157 CellInfo
*pInfo
/* Size information about the cell */
6165 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6166 pPage
->xParseCell(pPage
, pCell
, pInfo
);
6167 if( pInfo
->nLocal
==pInfo
->nPayload
){
6168 return SQLITE_OK
; /* No overflow pages. Return without doing anything */
6170 if( pCell
+pInfo
->nSize
-1 > pPage
->aData
+pPage
->maskPage
){
6171 /* Cell extends past end of page */
6172 return SQLITE_CORRUPT_PGNO(pPage
->pgno
);
6174 ovflPgno
= get4byte(pCell
+ pInfo
->nSize
- 4);
6176 assert( pBt
->usableSize
> 4 );
6177 ovflPageSize
= pBt
->usableSize
- 4;
6178 nOvfl
= (pInfo
->nPayload
- pInfo
->nLocal
+ ovflPageSize
- 1)/ovflPageSize
;
6180 (CORRUPT_DB
&& (pInfo
->nPayload
+ ovflPageSize
)<ovflPageSize
)
6185 if( ovflPgno
<2 || ovflPgno
>btreePagecount(pBt
) ){
6186 /* 0 is not a legal page number and page 1 cannot be an
6187 ** overflow page. Therefore if ovflPgno<2 or past the end of the
6188 ** file the database must be corrupt. */
6189 return SQLITE_CORRUPT_BKPT
;
6192 rc
= getOverflowPage(pBt
, ovflPgno
, &pOvfl
, &iNext
);
6196 if( ( pOvfl
|| ((pOvfl
= btreePageLookup(pBt
, ovflPgno
))!=0) )
6197 && sqlite3PagerPageRefcount(pOvfl
->pDbPage
)!=1
6199 /* There is no reason any cursor should have an outstanding reference
6200 ** to an overflow page belonging to a cell that is being deleted/updated.
6201 ** So if there exists more than one reference to this page, then it
6202 ** must not really be an overflow page and the database must be corrupt.
6203 ** It is helpful to detect this before calling freePage2(), as
6204 ** freePage2() may zero the page contents if secure-delete mode is
6205 ** enabled. If this 'overflow' page happens to be a page that the
6206 ** caller is iterating through or using in some other way, this
6207 ** can be problematic.
6209 rc
= SQLITE_CORRUPT_BKPT
;
6211 rc
= freePage2(pBt
, pOvfl
, ovflPgno
);
6215 sqlite3PagerUnref(pOvfl
->pDbPage
);
6224 ** Create the byte sequence used to represent a cell on page pPage
6225 ** and write that byte sequence into pCell[]. Overflow pages are
6226 ** allocated and filled in as necessary. The calling procedure
6227 ** is responsible for making sure sufficient space has been allocated
6230 ** Note that pCell does not necessary need to point to the pPage->aData
6231 ** area. pCell might point to some temporary storage. The cell will
6232 ** be constructed in this temporary area then copied into pPage->aData
6235 static int fillInCell(
6236 MemPage
*pPage
, /* The page that contains the cell */
6237 unsigned char *pCell
, /* Complete text of the cell */
6238 const BtreePayload
*pX
, /* Payload with which to construct the cell */
6239 int *pnSize
/* Write cell size here */
6243 int nSrc
, n
, rc
, mn
;
6245 MemPage
*pToRelease
;
6246 unsigned char *pPrior
;
6247 unsigned char *pPayload
;
6252 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6254 /* pPage is not necessarily writeable since pCell might be auxiliary
6255 ** buffer space that is separate from the pPage buffer area */
6256 assert( pCell
<pPage
->aData
|| pCell
>=&pPage
->aData
[pPage
->pBt
->pageSize
]
6257 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6259 /* Fill in the header. */
6260 nHeader
= pPage
->childPtrSize
;
6261 if( pPage
->intKey
){
6262 nPayload
= pX
->nData
+ pX
->nZero
;
6265 assert( pPage
->intKeyLeaf
); /* fillInCell() only called for leaves */
6266 nHeader
+= putVarint32(&pCell
[nHeader
], nPayload
);
6267 nHeader
+= putVarint(&pCell
[nHeader
], *(u64
*)&pX
->nKey
);
6269 assert( pX
->nKey
<=0x7fffffff && pX
->pKey
!=0 );
6270 nSrc
= nPayload
= (int)pX
->nKey
;
6272 nHeader
+= putVarint32(&pCell
[nHeader
], nPayload
);
6275 /* Fill in the payload */
6276 pPayload
= &pCell
[nHeader
];
6277 if( nPayload
<=pPage
->maxLocal
){
6278 /* This is the common case where everything fits on the btree page
6279 ** and no overflow pages are required. */
6280 n
= nHeader
+ nPayload
;
6285 assert( nSrc
<=nPayload
);
6286 testcase( nSrc
<nPayload
);
6287 memcpy(pPayload
, pSrc
, nSrc
);
6288 memset(pPayload
+nSrc
, 0, nPayload
-nSrc
);
6292 /* If we reach this point, it means that some of the content will need
6293 ** to spill onto overflow pages.
6295 mn
= pPage
->minLocal
;
6296 n
= mn
+ (nPayload
- mn
) % (pPage
->pBt
->usableSize
- 4);
6297 testcase( n
==pPage
->maxLocal
);
6298 testcase( n
==pPage
->maxLocal
+1 );
6299 if( n
> pPage
->maxLocal
) n
= mn
;
6301 *pnSize
= n
+ nHeader
+ 4;
6302 pPrior
= &pCell
[nHeader
+n
];
6307 /* At this point variables should be set as follows:
6309 ** nPayload Total payload size in bytes
6310 ** pPayload Begin writing payload here
6311 ** spaceLeft Space available at pPayload. If nPayload>spaceLeft,
6312 ** that means content must spill into overflow pages.
6313 ** *pnSize Size of the local cell (not counting overflow pages)
6314 ** pPrior Where to write the pgno of the first overflow page
6316 ** Use a call to btreeParseCellPtr() to verify that the values above
6317 ** were computed correctly.
6322 pPage
->xParseCell(pPage
, pCell
, &info
);
6323 assert( nHeader
==(int)(info
.pPayload
- pCell
) );
6324 assert( info
.nKey
==pX
->nKey
);
6325 assert( *pnSize
== info
.nSize
);
6326 assert( spaceLeft
== info
.nLocal
);
6330 /* Write the payload into the local Cell and any extra into overflow pages */
6333 if( n
>spaceLeft
) n
= spaceLeft
;
6335 /* If pToRelease is not zero than pPayload points into the data area
6336 ** of pToRelease. Make sure pToRelease is still writeable. */
6337 assert( pToRelease
==0 || sqlite3PagerIswriteable(pToRelease
->pDbPage
) );
6339 /* If pPayload is part of the data area of pPage, then make sure pPage
6340 ** is still writeable */
6341 assert( pPayload
<pPage
->aData
|| pPayload
>=&pPage
->aData
[pBt
->pageSize
]
6342 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6345 memcpy(pPayload
, pSrc
, n
);
6348 memcpy(pPayload
, pSrc
, n
);
6350 memset(pPayload
, 0, n
);
6353 if( nPayload
<=0 ) break;
6360 #ifndef SQLITE_OMIT_AUTOVACUUM
6361 Pgno pgnoPtrmap
= pgnoOvfl
; /* Overflow page pointer-map entry page */
6362 if( pBt
->autoVacuum
){
6366 PTRMAP_ISPAGE(pBt
, pgnoOvfl
) || pgnoOvfl
==PENDING_BYTE_PAGE(pBt
)
6370 rc
= allocateBtreePage(pBt
, &pOvfl
, &pgnoOvfl
, pgnoOvfl
, 0);
6371 #ifndef SQLITE_OMIT_AUTOVACUUM
6372 /* If the database supports auto-vacuum, and the second or subsequent
6373 ** overflow page is being allocated, add an entry to the pointer-map
6374 ** for that page now.
6376 ** If this is the first overflow page, then write a partial entry
6377 ** to the pointer-map. If we write nothing to this pointer-map slot,
6378 ** then the optimistic overflow chain processing in clearCell()
6379 ** may misinterpret the uninitialized values and delete the
6380 ** wrong pages from the database.
6382 if( pBt
->autoVacuum
&& rc
==SQLITE_OK
){
6383 u8 eType
= (pgnoPtrmap
?PTRMAP_OVERFLOW2
:PTRMAP_OVERFLOW1
);
6384 ptrmapPut(pBt
, pgnoOvfl
, eType
, pgnoPtrmap
, &rc
);
6391 releasePage(pToRelease
);
6395 /* If pToRelease is not zero than pPrior points into the data area
6396 ** of pToRelease. Make sure pToRelease is still writeable. */
6397 assert( pToRelease
==0 || sqlite3PagerIswriteable(pToRelease
->pDbPage
) );
6399 /* If pPrior is part of the data area of pPage, then make sure pPage
6400 ** is still writeable */
6401 assert( pPrior
<pPage
->aData
|| pPrior
>=&pPage
->aData
[pBt
->pageSize
]
6402 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6404 put4byte(pPrior
, pgnoOvfl
);
6405 releasePage(pToRelease
);
6407 pPrior
= pOvfl
->aData
;
6408 put4byte(pPrior
, 0);
6409 pPayload
= &pOvfl
->aData
[4];
6410 spaceLeft
= pBt
->usableSize
- 4;
6413 releasePage(pToRelease
);
6418 ** Remove the i-th cell from pPage. This routine effects pPage only.
6419 ** The cell content is not freed or deallocated. It is assumed that
6420 ** the cell content has been copied someplace else. This routine just
6421 ** removes the reference to the cell from pPage.
6423 ** "sz" must be the number of bytes in the cell.
6425 static void dropCell(MemPage
*pPage
, int idx
, int sz
, int *pRC
){
6426 u32 pc
; /* Offset to cell content of cell being deleted */
6427 u8
*data
; /* pPage->aData */
6428 u8
*ptr
; /* Used to move bytes around within data[] */
6429 int rc
; /* The return code */
6430 int hdr
; /* Beginning of the header. 0 most pages. 100 page 1 */
6433 assert( idx
>=0 && idx
<pPage
->nCell
);
6434 assert( CORRUPT_DB
|| sz
==cellSize(pPage
, idx
) );
6435 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
6436 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6437 data
= pPage
->aData
;
6438 ptr
= &pPage
->aCellIdx
[2*idx
];
6440 hdr
= pPage
->hdrOffset
;
6441 testcase( pc
==get2byte(&data
[hdr
+5]) );
6442 testcase( pc
+sz
==pPage
->pBt
->usableSize
);
6443 if( pc
+sz
> pPage
->pBt
->usableSize
){
6444 *pRC
= SQLITE_CORRUPT_BKPT
;
6447 rc
= freeSpace(pPage
, pc
, sz
);
6453 if( pPage
->nCell
==0 ){
6454 memset(&data
[hdr
+1], 0, 4);
6456 put2byte(&data
[hdr
+5], pPage
->pBt
->usableSize
);
6457 pPage
->nFree
= pPage
->pBt
->usableSize
- pPage
->hdrOffset
6458 - pPage
->childPtrSize
- 8;
6460 memmove(ptr
, ptr
+2, 2*(pPage
->nCell
- idx
));
6461 put2byte(&data
[hdr
+3], pPage
->nCell
);
6467 ** Insert a new cell on pPage at cell index "i". pCell points to the
6468 ** content of the cell.
6470 ** If the cell content will fit on the page, then put it there. If it
6471 ** will not fit, then make a copy of the cell content into pTemp if
6472 ** pTemp is not null. Regardless of pTemp, allocate a new entry
6473 ** in pPage->apOvfl[] and make it point to the cell content (either
6474 ** in pTemp or the original pCell) and also record its index.
6475 ** Allocating a new entry in pPage->aCell[] implies that
6476 ** pPage->nOverflow is incremented.
6478 ** *pRC must be SQLITE_OK when this routine is called.
6480 static void insertCell(
6481 MemPage
*pPage
, /* Page into which we are copying */
6482 int i
, /* New cell becomes the i-th cell of the page */
6483 u8
*pCell
, /* Content of the new cell */
6484 int sz
, /* Bytes of content in pCell */
6485 u8
*pTemp
, /* Temp storage space for pCell, if needed */
6486 Pgno iChild
, /* If non-zero, replace first 4 bytes with this value */
6487 int *pRC
/* Read and write return code from here */
6489 int idx
= 0; /* Where to write new cell content in data[] */
6490 int j
; /* Loop counter */
6491 u8
*data
; /* The content of the whole page */
6492 u8
*pIns
; /* The point in pPage->aCellIdx[] where no cell inserted */
6494 assert( *pRC
==SQLITE_OK
);
6495 assert( i
>=0 && i
<=pPage
->nCell
+pPage
->nOverflow
);
6496 assert( MX_CELL(pPage
->pBt
)<=10921 );
6497 assert( pPage
->nCell
<=MX_CELL(pPage
->pBt
) || CORRUPT_DB
);
6498 assert( pPage
->nOverflow
<=ArraySize(pPage
->apOvfl
) );
6499 assert( ArraySize(pPage
->apOvfl
)==ArraySize(pPage
->aiOvfl
) );
6500 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6501 /* The cell should normally be sized correctly. However, when moving a
6502 ** malformed cell from a leaf page to an interior page, if the cell size
6503 ** wanted to be less than 4 but got rounded up to 4 on the leaf, then size
6504 ** might be less than 8 (leaf-size + pointer) on the interior node. Hence
6505 ** the term after the || in the following assert(). */
6506 assert( sz
==pPage
->xCellSize(pPage
, pCell
) || (sz
==8 && iChild
>0) );
6507 if( pPage
->nOverflow
|| sz
+2>pPage
->nFree
){
6509 memcpy(pTemp
, pCell
, sz
);
6513 put4byte(pCell
, iChild
);
6515 j
= pPage
->nOverflow
++;
6516 /* Comparison against ArraySize-1 since we hold back one extra slot
6517 ** as a contingency. In other words, never need more than 3 overflow
6518 ** slots but 4 are allocated, just to be safe. */
6519 assert( j
< ArraySize(pPage
->apOvfl
)-1 );
6520 pPage
->apOvfl
[j
] = pCell
;
6521 pPage
->aiOvfl
[j
] = (u16
)i
;
6523 /* When multiple overflows occur, they are always sequential and in
6524 ** sorted order. This invariants arise because multiple overflows can
6525 ** only occur when inserting divider cells into the parent page during
6526 ** balancing, and the dividers are adjacent and sorted.
6528 assert( j
==0 || pPage
->aiOvfl
[j
-1]<(u16
)i
); /* Overflows in sorted order */
6529 assert( j
==0 || i
==pPage
->aiOvfl
[j
-1]+1 ); /* Overflows are sequential */
6531 int rc
= sqlite3PagerWrite(pPage
->pDbPage
);
6532 if( rc
!=SQLITE_OK
){
6536 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
6537 data
= pPage
->aData
;
6538 assert( &data
[pPage
->cellOffset
]==pPage
->aCellIdx
);
6539 rc
= allocateSpace(pPage
, sz
, &idx
);
6540 if( rc
){ *pRC
= rc
; return; }
6541 /* The allocateSpace() routine guarantees the following properties
6542 ** if it returns successfully */
6544 assert( idx
>= pPage
->cellOffset
+2*pPage
->nCell
+2 || CORRUPT_DB
);
6545 assert( idx
+sz
<= (int)pPage
->pBt
->usableSize
);
6546 pPage
->nFree
-= (u16
)(2 + sz
);
6547 memcpy(&data
[idx
], pCell
, sz
);
6549 put4byte(&data
[idx
], iChild
);
6551 pIns
= pPage
->aCellIdx
+ i
*2;
6552 memmove(pIns
+2, pIns
, 2*(pPage
->nCell
- i
));
6553 put2byte(pIns
, idx
);
6555 /* increment the cell count */
6556 if( (++data
[pPage
->hdrOffset
+4])==0 ) data
[pPage
->hdrOffset
+3]++;
6557 assert( get2byte(&data
[pPage
->hdrOffset
+3])==pPage
->nCell
);
6558 #ifndef SQLITE_OMIT_AUTOVACUUM
6559 if( pPage
->pBt
->autoVacuum
){
6560 /* The cell may contain a pointer to an overflow page. If so, write
6561 ** the entry for the overflow page into the pointer map.
6563 ptrmapPutOvflPtr(pPage
, pCell
, pRC
);
6570 ** A CellArray object contains a cache of pointers and sizes for a
6571 ** consecutive sequence of cells that might be held on multiple pages.
6573 typedef struct CellArray CellArray
;
6575 int nCell
; /* Number of cells in apCell[] */
6576 MemPage
*pRef
; /* Reference page */
6577 u8
**apCell
; /* All cells begin balanced */
6578 u16
*szCell
; /* Local size of all cells in apCell[] */
6582 ** Make sure the cell sizes at idx, idx+1, ..., idx+N-1 have been
6585 static void populateCellCache(CellArray
*p
, int idx
, int N
){
6586 assert( idx
>=0 && idx
+N
<=p
->nCell
);
6588 assert( p
->apCell
[idx
]!=0 );
6589 if( p
->szCell
[idx
]==0 ){
6590 p
->szCell
[idx
] = p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[idx
]);
6592 assert( CORRUPT_DB
||
6593 p
->szCell
[idx
]==p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[idx
]) );
6601 ** Return the size of the Nth element of the cell array
6603 static SQLITE_NOINLINE u16
computeCellSize(CellArray
*p
, int N
){
6604 assert( N
>=0 && N
<p
->nCell
);
6605 assert( p
->szCell
[N
]==0 );
6606 p
->szCell
[N
] = p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[N
]);
6607 return p
->szCell
[N
];
6609 static u16
cachedCellSize(CellArray
*p
, int N
){
6610 assert( N
>=0 && N
<p
->nCell
);
6611 if( p
->szCell
[N
] ) return p
->szCell
[N
];
6612 return computeCellSize(p
, N
);
6616 ** Array apCell[] contains pointers to nCell b-tree page cells. The
6617 ** szCell[] array contains the size in bytes of each cell. This function
6618 ** replaces the current contents of page pPg with the contents of the cell
6621 ** Some of the cells in apCell[] may currently be stored in pPg. This
6622 ** function works around problems caused by this by making a copy of any
6623 ** such cells before overwriting the page data.
6625 ** The MemPage.nFree field is invalidated by this function. It is the
6626 ** responsibility of the caller to set it correctly.
6628 static int rebuildPage(
6629 MemPage
*pPg
, /* Edit this page */
6630 int nCell
, /* Final number of cells on page */
6631 u8
**apCell
, /* Array of cells */
6632 u16
*szCell
/* Array of cell sizes */
6634 const int hdr
= pPg
->hdrOffset
; /* Offset of header on pPg */
6635 u8
* const aData
= pPg
->aData
; /* Pointer to data for pPg */
6636 const int usableSize
= pPg
->pBt
->usableSize
;
6637 u8
* const pEnd
= &aData
[usableSize
];
6639 u8
*pCellptr
= pPg
->aCellIdx
;
6640 u8
*pTmp
= sqlite3PagerTempSpace(pPg
->pBt
->pPager
);
6643 i
= get2byte(&aData
[hdr
+5]);
6644 memcpy(&pTmp
[i
], &aData
[i
], usableSize
- i
);
6647 for(i
=0; i
<nCell
; i
++){
6648 u8
*pCell
= apCell
[i
];
6649 if( SQLITE_WITHIN(pCell
,aData
,pEnd
) ){
6650 pCell
= &pTmp
[pCell
- aData
];
6653 put2byte(pCellptr
, (pData
- aData
));
6655 if( pData
< pCellptr
) return SQLITE_CORRUPT_BKPT
;
6656 memcpy(pData
, pCell
, szCell
[i
]);
6657 assert( szCell
[i
]==pPg
->xCellSize(pPg
, pCell
) || CORRUPT_DB
);
6658 testcase( szCell
[i
]!=pPg
->xCellSize(pPg
,pCell
) );
6661 /* The pPg->nFree field is now set incorrectly. The caller will fix it. */
6665 put2byte(&aData
[hdr
+1], 0);
6666 put2byte(&aData
[hdr
+3], pPg
->nCell
);
6667 put2byte(&aData
[hdr
+5], pData
- aData
);
6668 aData
[hdr
+7] = 0x00;
6673 ** Array apCell[] contains nCell pointers to b-tree cells. Array szCell
6674 ** contains the size in bytes of each such cell. This function attempts to
6675 ** add the cells stored in the array to page pPg. If it cannot (because
6676 ** the page needs to be defragmented before the cells will fit), non-zero
6677 ** is returned. Otherwise, if the cells are added successfully, zero is
6680 ** Argument pCellptr points to the first entry in the cell-pointer array
6681 ** (part of page pPg) to populate. After cell apCell[0] is written to the
6682 ** page body, a 16-bit offset is written to pCellptr. And so on, for each
6683 ** cell in the array. It is the responsibility of the caller to ensure
6684 ** that it is safe to overwrite this part of the cell-pointer array.
6686 ** When this function is called, *ppData points to the start of the
6687 ** content area on page pPg. If the size of the content area is extended,
6688 ** *ppData is updated to point to the new start of the content area
6689 ** before returning.
6691 ** Finally, argument pBegin points to the byte immediately following the
6692 ** end of the space required by this page for the cell-pointer area (for
6693 ** all cells - not just those inserted by the current call). If the content
6694 ** area must be extended to before this point in order to accomodate all
6695 ** cells in apCell[], then the cells do not fit and non-zero is returned.
6697 static int pageInsertArray(
6698 MemPage
*pPg
, /* Page to add cells to */
6699 u8
*pBegin
, /* End of cell-pointer array */
6700 u8
**ppData
, /* IN/OUT: Page content -area pointer */
6701 u8
*pCellptr
, /* Pointer to cell-pointer area */
6702 int iFirst
, /* Index of first cell to add */
6703 int nCell
, /* Number of cells to add to pPg */
6704 CellArray
*pCArray
/* Array of cells */
6707 u8
*aData
= pPg
->aData
;
6708 u8
*pData
= *ppData
;
6709 int iEnd
= iFirst
+ nCell
;
6710 assert( CORRUPT_DB
|| pPg
->hdrOffset
==0 ); /* Never called on page 1 */
6711 for(i
=iFirst
; i
<iEnd
; i
++){
6714 sz
= cachedCellSize(pCArray
, i
);
6715 if( (aData
[1]==0 && aData
[2]==0) || (pSlot
= pageFindSlot(pPg
,sz
,&rc
))==0 ){
6716 if( (pData
- pBegin
)<sz
) return 1;
6720 /* pSlot and pCArray->apCell[i] will never overlap on a well-formed
6721 ** database. But they might for a corrupt database. Hence use memmove()
6722 ** since memcpy() sends SIGABORT with overlapping buffers on OpenBSD */
6723 assert( (pSlot
+sz
)<=pCArray
->apCell
[i
]
6724 || pSlot
>=(pCArray
->apCell
[i
]+sz
)
6726 memmove(pSlot
, pCArray
->apCell
[i
], sz
);
6727 put2byte(pCellptr
, (pSlot
- aData
));
6735 ** Array apCell[] contains nCell pointers to b-tree cells. Array szCell
6736 ** contains the size in bytes of each such cell. This function adds the
6737 ** space associated with each cell in the array that is currently stored
6738 ** within the body of pPg to the pPg free-list. The cell-pointers and other
6739 ** fields of the page are not updated.
6741 ** This function returns the total number of cells added to the free-list.
6743 static int pageFreeArray(
6744 MemPage
*pPg
, /* Page to edit */
6745 int iFirst
, /* First cell to delete */
6746 int nCell
, /* Cells to delete */
6747 CellArray
*pCArray
/* Array of cells */
6749 u8
* const aData
= pPg
->aData
;
6750 u8
* const pEnd
= &aData
[pPg
->pBt
->usableSize
];
6751 u8
* const pStart
= &aData
[pPg
->hdrOffset
+ 8 + pPg
->childPtrSize
];
6754 int iEnd
= iFirst
+ nCell
;
6758 for(i
=iFirst
; i
<iEnd
; i
++){
6759 u8
*pCell
= pCArray
->apCell
[i
];
6760 if( SQLITE_WITHIN(pCell
, pStart
, pEnd
) ){
6762 /* No need to use cachedCellSize() here. The sizes of all cells that
6763 ** are to be freed have already been computing while deciding which
6764 ** cells need freeing */
6765 sz
= pCArray
->szCell
[i
]; assert( sz
>0 );
6766 if( pFree
!=(pCell
+ sz
) ){
6768 assert( pFree
>aData
&& (pFree
- aData
)<65536 );
6769 freeSpace(pPg
, (u16
)(pFree
- aData
), szFree
);
6773 if( pFree
+sz
>pEnd
) return 0;
6782 assert( pFree
>aData
&& (pFree
- aData
)<65536 );
6783 freeSpace(pPg
, (u16
)(pFree
- aData
), szFree
);
6789 ** apCell[] and szCell[] contains pointers to and sizes of all cells in the
6790 ** pages being balanced. The current page, pPg, has pPg->nCell cells starting
6791 ** with apCell[iOld]. After balancing, this page should hold nNew cells
6792 ** starting at apCell[iNew].
6794 ** This routine makes the necessary adjustments to pPg so that it contains
6795 ** the correct cells after being balanced.
6797 ** The pPg->nFree field is invalid when this function returns. It is the
6798 ** responsibility of the caller to set it correctly.
6800 static int editPage(
6801 MemPage
*pPg
, /* Edit this page */
6802 int iOld
, /* Index of first cell currently on page */
6803 int iNew
, /* Index of new first cell on page */
6804 int nNew
, /* Final number of cells on page */
6805 CellArray
*pCArray
/* Array of cells and sizes */
6807 u8
* const aData
= pPg
->aData
;
6808 const int hdr
= pPg
->hdrOffset
;
6809 u8
*pBegin
= &pPg
->aCellIdx
[nNew
* 2];
6810 int nCell
= pPg
->nCell
; /* Cells stored on pPg */
6814 int iOldEnd
= iOld
+ pPg
->nCell
+ pPg
->nOverflow
;
6815 int iNewEnd
= iNew
+ nNew
;
6818 u8
*pTmp
= sqlite3PagerTempSpace(pPg
->pBt
->pPager
);
6819 memcpy(pTmp
, aData
, pPg
->pBt
->usableSize
);
6822 /* Remove cells from the start and end of the page */
6824 int nShift
= pageFreeArray(pPg
, iOld
, iNew
-iOld
, pCArray
);
6825 memmove(pPg
->aCellIdx
, &pPg
->aCellIdx
[nShift
*2], nCell
*2);
6828 if( iNewEnd
< iOldEnd
){
6829 nCell
-= pageFreeArray(pPg
, iNewEnd
, iOldEnd
- iNewEnd
, pCArray
);
6832 pData
= &aData
[get2byteNotZero(&aData
[hdr
+5])];
6833 if( pData
<pBegin
) goto editpage_fail
;
6835 /* Add cells to the start of the page */
6837 int nAdd
= MIN(nNew
,iOld
-iNew
);
6838 assert( (iOld
-iNew
)<nNew
|| nCell
==0 || CORRUPT_DB
);
6839 pCellptr
= pPg
->aCellIdx
;
6840 memmove(&pCellptr
[nAdd
*2], pCellptr
, nCell
*2);
6841 if( pageInsertArray(
6842 pPg
, pBegin
, &pData
, pCellptr
,
6844 ) ) goto editpage_fail
;
6848 /* Add any overflow cells */
6849 for(i
=0; i
<pPg
->nOverflow
; i
++){
6850 int iCell
= (iOld
+ pPg
->aiOvfl
[i
]) - iNew
;
6851 if( iCell
>=0 && iCell
<nNew
){
6852 pCellptr
= &pPg
->aCellIdx
[iCell
* 2];
6853 memmove(&pCellptr
[2], pCellptr
, (nCell
- iCell
) * 2);
6855 if( pageInsertArray(
6856 pPg
, pBegin
, &pData
, pCellptr
,
6857 iCell
+iNew
, 1, pCArray
6858 ) ) goto editpage_fail
;
6862 /* Append cells to the end of the page */
6863 pCellptr
= &pPg
->aCellIdx
[nCell
*2];
6864 if( pageInsertArray(
6865 pPg
, pBegin
, &pData
, pCellptr
,
6866 iNew
+nCell
, nNew
-nCell
, pCArray
6867 ) ) goto editpage_fail
;
6872 put2byte(&aData
[hdr
+3], pPg
->nCell
);
6873 put2byte(&aData
[hdr
+5], pData
- aData
);
6876 for(i
=0; i
<nNew
&& !CORRUPT_DB
; i
++){
6877 u8
*pCell
= pCArray
->apCell
[i
+iNew
];
6878 int iOff
= get2byteAligned(&pPg
->aCellIdx
[i
*2]);
6879 if( SQLITE_WITHIN(pCell
, aData
, &aData
[pPg
->pBt
->usableSize
]) ){
6880 pCell
= &pTmp
[pCell
- aData
];
6882 assert( 0==memcmp(pCell
, &aData
[iOff
],
6883 pCArray
->pRef
->xCellSize(pCArray
->pRef
, pCArray
->apCell
[i
+iNew
])) );
6889 /* Unable to edit this page. Rebuild it from scratch instead. */
6890 populateCellCache(pCArray
, iNew
, nNew
);
6891 return rebuildPage(pPg
, nNew
, &pCArray
->apCell
[iNew
], &pCArray
->szCell
[iNew
]);
6895 ** The following parameters determine how many adjacent pages get involved
6896 ** in a balancing operation. NN is the number of neighbors on either side
6897 ** of the page that participate in the balancing operation. NB is the
6898 ** total number of pages that participate, including the target page and
6899 ** NN neighbors on either side.
6901 ** The minimum value of NN is 1 (of course). Increasing NN above 1
6902 ** (to 2 or 3) gives a modest improvement in SELECT and DELETE performance
6903 ** in exchange for a larger degradation in INSERT and UPDATE performance.
6904 ** The value of NN appears to give the best results overall.
6906 #define NN 1 /* Number of neighbors on either side of pPage */
6907 #define NB (NN*2+1) /* Total pages involved in the balance */
6910 #ifndef SQLITE_OMIT_QUICKBALANCE
6912 ** This version of balance() handles the common special case where
6913 ** a new entry is being inserted on the extreme right-end of the
6914 ** tree, in other words, when the new entry will become the largest
6915 ** entry in the tree.
6917 ** Instead of trying to balance the 3 right-most leaf pages, just add
6918 ** a new page to the right-hand side and put the one new entry in
6919 ** that page. This leaves the right side of the tree somewhat
6920 ** unbalanced. But odds are that we will be inserting new entries
6921 ** at the end soon afterwards so the nearly empty page will quickly
6922 ** fill up. On average.
6924 ** pPage is the leaf page which is the right-most page in the tree.
6925 ** pParent is its parent. pPage must have a single overflow entry
6926 ** which is also the right-most entry on the page.
6928 ** The pSpace buffer is used to store a temporary copy of the divider
6929 ** cell that will be inserted into pParent. Such a cell consists of a 4
6930 ** byte page number followed by a variable length integer. In other
6931 ** words, at most 13 bytes. Hence the pSpace buffer must be at
6932 ** least 13 bytes in size.
6934 static int balance_quick(MemPage
*pParent
, MemPage
*pPage
, u8
*pSpace
){
6935 BtShared
*const pBt
= pPage
->pBt
; /* B-Tree Database */
6936 MemPage
*pNew
; /* Newly allocated page */
6937 int rc
; /* Return Code */
6938 Pgno pgnoNew
; /* Page number of pNew */
6940 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6941 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
6942 assert( pPage
->nOverflow
==1 );
6944 /* This error condition is now caught prior to reaching this function */
6945 if( NEVER(pPage
->nCell
==0) ) return SQLITE_CORRUPT_BKPT
;
6947 /* Allocate a new page. This page will become the right-sibling of
6948 ** pPage. Make the parent page writable, so that the new divider cell
6949 ** may be inserted. If both these operations are successful, proceed.
6951 rc
= allocateBtreePage(pBt
, &pNew
, &pgnoNew
, 0, 0);
6953 if( rc
==SQLITE_OK
){
6955 u8
*pOut
= &pSpace
[4];
6956 u8
*pCell
= pPage
->apOvfl
[0];
6957 u16 szCell
= pPage
->xCellSize(pPage
, pCell
);
6960 assert( sqlite3PagerIswriteable(pNew
->pDbPage
) );
6961 assert( pPage
->aData
[0]==(PTF_INTKEY
|PTF_LEAFDATA
|PTF_LEAF
) );
6962 zeroPage(pNew
, PTF_INTKEY
|PTF_LEAFDATA
|PTF_LEAF
);
6963 rc
= rebuildPage(pNew
, 1, &pCell
, &szCell
);
6964 if( NEVER(rc
) ) return rc
;
6965 pNew
->nFree
= pBt
->usableSize
- pNew
->cellOffset
- 2 - szCell
;
6967 /* If this is an auto-vacuum database, update the pointer map
6968 ** with entries for the new page, and any pointer from the
6969 ** cell on the page to an overflow page. If either of these
6970 ** operations fails, the return code is set, but the contents
6971 ** of the parent page are still manipulated by thh code below.
6972 ** That is Ok, at this point the parent page is guaranteed to
6973 ** be marked as dirty. Returning an error code will cause a
6974 ** rollback, undoing any changes made to the parent page.
6977 ptrmapPut(pBt
, pgnoNew
, PTRMAP_BTREE
, pParent
->pgno
, &rc
);
6978 if( szCell
>pNew
->minLocal
){
6979 ptrmapPutOvflPtr(pNew
, pCell
, &rc
);
6983 /* Create a divider cell to insert into pParent. The divider cell
6984 ** consists of a 4-byte page number (the page number of pPage) and
6985 ** a variable length key value (which must be the same value as the
6986 ** largest key on pPage).
6988 ** To find the largest key value on pPage, first find the right-most
6989 ** cell on pPage. The first two fields of this cell are the
6990 ** record-length (a variable length integer at most 32-bits in size)
6991 ** and the key value (a variable length integer, may have any value).
6992 ** The first of the while(...) loops below skips over the record-length
6993 ** field. The second while(...) loop copies the key value from the
6994 ** cell on pPage into the pSpace buffer.
6996 pCell
= findCell(pPage
, pPage
->nCell
-1);
6998 while( (*(pCell
++)&0x80) && pCell
<pStop
);
7000 while( ((*(pOut
++) = *(pCell
++))&0x80) && pCell
<pStop
);
7002 /* Insert the new divider cell into pParent. */
7003 if( rc
==SQLITE_OK
){
7004 insertCell(pParent
, pParent
->nCell
, pSpace
, (int)(pOut
-pSpace
),
7005 0, pPage
->pgno
, &rc
);
7008 /* Set the right-child pointer of pParent to point to the new page. */
7009 put4byte(&pParent
->aData
[pParent
->hdrOffset
+8], pgnoNew
);
7011 /* Release the reference to the new page. */
7017 #endif /* SQLITE_OMIT_QUICKBALANCE */
7021 ** This function does not contribute anything to the operation of SQLite.
7022 ** it is sometimes activated temporarily while debugging code responsible
7023 ** for setting pointer-map entries.
7025 static int ptrmapCheckPages(MemPage
**apPage
, int nPage
){
7027 for(i
=0; i
<nPage
; i
++){
7030 MemPage
*pPage
= apPage
[i
];
7031 BtShared
*pBt
= pPage
->pBt
;
7032 assert( pPage
->isInit
);
7034 for(j
=0; j
<pPage
->nCell
; j
++){
7038 z
= findCell(pPage
, j
);
7039 pPage
->xParseCell(pPage
, z
, &info
);
7040 if( info
.nLocal
<info
.nPayload
){
7041 Pgno ovfl
= get4byte(&z
[info
.nSize
-4]);
7042 ptrmapGet(pBt
, ovfl
, &e
, &n
);
7043 assert( n
==pPage
->pgno
&& e
==PTRMAP_OVERFLOW1
);
7046 Pgno child
= get4byte(z
);
7047 ptrmapGet(pBt
, child
, &e
, &n
);
7048 assert( n
==pPage
->pgno
&& e
==PTRMAP_BTREE
);
7052 Pgno child
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
7053 ptrmapGet(pBt
, child
, &e
, &n
);
7054 assert( n
==pPage
->pgno
&& e
==PTRMAP_BTREE
);
7062 ** This function is used to copy the contents of the b-tree node stored
7063 ** on page pFrom to page pTo. If page pFrom was not a leaf page, then
7064 ** the pointer-map entries for each child page are updated so that the
7065 ** parent page stored in the pointer map is page pTo. If pFrom contained
7066 ** any cells with overflow page pointers, then the corresponding pointer
7067 ** map entries are also updated so that the parent page is page pTo.
7069 ** If pFrom is currently carrying any overflow cells (entries in the
7070 ** MemPage.apOvfl[] array), they are not copied to pTo.
7072 ** Before returning, page pTo is reinitialized using btreeInitPage().
7074 ** The performance of this function is not critical. It is only used by
7075 ** the balance_shallower() and balance_deeper() procedures, neither of
7076 ** which are called often under normal circumstances.
7078 static void copyNodeContent(MemPage
*pFrom
, MemPage
*pTo
, int *pRC
){
7079 if( (*pRC
)==SQLITE_OK
){
7080 BtShared
* const pBt
= pFrom
->pBt
;
7081 u8
* const aFrom
= pFrom
->aData
;
7082 u8
* const aTo
= pTo
->aData
;
7083 int const iFromHdr
= pFrom
->hdrOffset
;
7084 int const iToHdr
= ((pTo
->pgno
==1) ? 100 : 0);
7089 assert( pFrom
->isInit
);
7090 assert( pFrom
->nFree
>=iToHdr
);
7091 assert( get2byte(&aFrom
[iFromHdr
+5]) <= (int)pBt
->usableSize
);
7093 /* Copy the b-tree node content from page pFrom to page pTo. */
7094 iData
= get2byte(&aFrom
[iFromHdr
+5]);
7095 memcpy(&aTo
[iData
], &aFrom
[iData
], pBt
->usableSize
-iData
);
7096 memcpy(&aTo
[iToHdr
], &aFrom
[iFromHdr
], pFrom
->cellOffset
+ 2*pFrom
->nCell
);
7098 /* Reinitialize page pTo so that the contents of the MemPage structure
7099 ** match the new data. The initialization of pTo can actually fail under
7100 ** fairly obscure circumstances, even though it is a copy of initialized
7104 rc
= btreeInitPage(pTo
);
7105 if( rc
!=SQLITE_OK
){
7110 /* If this is an auto-vacuum database, update the pointer-map entries
7111 ** for any b-tree or overflow pages that pTo now contains the pointers to.
7114 *pRC
= setChildPtrmaps(pTo
);
7120 ** This routine redistributes cells on the iParentIdx'th child of pParent
7121 ** (hereafter "the page") and up to 2 siblings so that all pages have about the
7122 ** same amount of free space. Usually a single sibling on either side of the
7123 ** page are used in the balancing, though both siblings might come from one
7124 ** side if the page is the first or last child of its parent. If the page
7125 ** has fewer than 2 siblings (something which can only happen if the page
7126 ** is a root page or a child of a root page) then all available siblings
7127 ** participate in the balancing.
7129 ** The number of siblings of the page might be increased or decreased by
7130 ** one or two in an effort to keep pages nearly full but not over full.
7132 ** Note that when this routine is called, some of the cells on the page
7133 ** might not actually be stored in MemPage.aData[]. This can happen
7134 ** if the page is overfull. This routine ensures that all cells allocated
7135 ** to the page and its siblings fit into MemPage.aData[] before returning.
7137 ** In the course of balancing the page and its siblings, cells may be
7138 ** inserted into or removed from the parent page (pParent). Doing so
7139 ** may cause the parent page to become overfull or underfull. If this
7140 ** happens, it is the responsibility of the caller to invoke the correct
7141 ** balancing routine to fix this problem (see the balance() routine).
7143 ** If this routine fails for any reason, it might leave the database
7144 ** in a corrupted state. So if this routine fails, the database should
7147 ** The third argument to this function, aOvflSpace, is a pointer to a
7148 ** buffer big enough to hold one page. If while inserting cells into the parent
7149 ** page (pParent) the parent page becomes overfull, this buffer is
7150 ** used to store the parent's overflow cells. Because this function inserts
7151 ** a maximum of four divider cells into the parent page, and the maximum
7152 ** size of a cell stored within an internal node is always less than 1/4
7153 ** of the page-size, the aOvflSpace[] buffer is guaranteed to be large
7154 ** enough for all overflow cells.
7156 ** If aOvflSpace is set to a null pointer, this function returns
7159 static int balance_nonroot(
7160 MemPage
*pParent
, /* Parent page of siblings being balanced */
7161 int iParentIdx
, /* Index of "the page" in pParent */
7162 u8
*aOvflSpace
, /* page-size bytes of space for parent ovfl */
7163 int isRoot
, /* True if pParent is a root-page */
7164 int bBulk
/* True if this call is part of a bulk load */
7166 BtShared
*pBt
; /* The whole database */
7167 int nMaxCells
= 0; /* Allocated size of apCell, szCell, aFrom. */
7168 int nNew
= 0; /* Number of pages in apNew[] */
7169 int nOld
; /* Number of pages in apOld[] */
7170 int i
, j
, k
; /* Loop counters */
7171 int nxDiv
; /* Next divider slot in pParent->aCell[] */
7172 int rc
= SQLITE_OK
; /* The return code */
7173 u16 leafCorrection
; /* 4 if pPage is a leaf. 0 if not */
7174 int leafData
; /* True if pPage is a leaf of a LEAFDATA tree */
7175 int usableSpace
; /* Bytes in pPage beyond the header */
7176 int pageFlags
; /* Value of pPage->aData[0] */
7177 int iSpace1
= 0; /* First unused byte of aSpace1[] */
7178 int iOvflSpace
= 0; /* First unused byte of aOvflSpace[] */
7179 int szScratch
; /* Size of scratch memory requested */
7180 MemPage
*apOld
[NB
]; /* pPage and up to two siblings */
7181 MemPage
*apNew
[NB
+2]; /* pPage and up to NB siblings after balancing */
7182 u8
*pRight
; /* Location in parent of right-sibling pointer */
7183 u8
*apDiv
[NB
-1]; /* Divider cells in pParent */
7184 int cntNew
[NB
+2]; /* Index in b.paCell[] of cell after i-th page */
7185 int cntOld
[NB
+2]; /* Old index in b.apCell[] */
7186 int szNew
[NB
+2]; /* Combined size of cells placed on i-th page */
7187 u8
*aSpace1
; /* Space for copies of dividers cells */
7188 Pgno pgno
; /* Temp var to store a page number in */
7189 u8 abDone
[NB
+2]; /* True after i'th new page is populated */
7190 Pgno aPgno
[NB
+2]; /* Page numbers of new pages before shuffling */
7191 Pgno aPgOrder
[NB
+2]; /* Copy of aPgno[] used for sorting pages */
7192 u16 aPgFlags
[NB
+2]; /* flags field of new pages before shuffling */
7193 CellArray b
; /* Parsed information on cells being balanced */
7195 memset(abDone
, 0, sizeof(abDone
));
7199 assert( sqlite3_mutex_held(pBt
->mutex
) );
7200 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7203 TRACE(("BALANCE: begin page %d child of %d\n", pPage
->pgno
, pParent
->pgno
));
7206 /* At this point pParent may have at most one overflow cell. And if
7207 ** this overflow cell is present, it must be the cell with
7208 ** index iParentIdx. This scenario comes about when this function
7209 ** is called (indirectly) from sqlite3BtreeDelete().
7211 assert( pParent
->nOverflow
==0 || pParent
->nOverflow
==1 );
7212 assert( pParent
->nOverflow
==0 || pParent
->aiOvfl
[0]==iParentIdx
);
7215 return SQLITE_NOMEM_BKPT
;
7218 /* Find the sibling pages to balance. Also locate the cells in pParent
7219 ** that divide the siblings. An attempt is made to find NN siblings on
7220 ** either side of pPage. More siblings are taken from one side, however,
7221 ** if there are fewer than NN siblings on the other side. If pParent
7222 ** has NB or fewer children then all children of pParent are taken.
7224 ** This loop also drops the divider cells from the parent page. This
7225 ** way, the remainder of the function does not have to deal with any
7226 ** overflow cells in the parent page, since if any existed they will
7227 ** have already been removed.
7229 i
= pParent
->nOverflow
+ pParent
->nCell
;
7233 assert( bBulk
==0 || bBulk
==1 );
7234 if( iParentIdx
==0 ){
7236 }else if( iParentIdx
==i
){
7239 nxDiv
= iParentIdx
-1;
7244 if( (i
+nxDiv
-pParent
->nOverflow
)==pParent
->nCell
){
7245 pRight
= &pParent
->aData
[pParent
->hdrOffset
+8];
7247 pRight
= findCell(pParent
, i
+nxDiv
-pParent
->nOverflow
);
7249 pgno
= get4byte(pRight
);
7251 rc
= getAndInitPage(pBt
, pgno
, &apOld
[i
], 0, 0);
7253 memset(apOld
, 0, (i
+1)*sizeof(MemPage
*));
7254 goto balance_cleanup
;
7256 nMaxCells
+= 1+apOld
[i
]->nCell
+apOld
[i
]->nOverflow
;
7257 if( (i
--)==0 ) break;
7259 if( pParent
->nOverflow
&& i
+nxDiv
==pParent
->aiOvfl
[0] ){
7260 apDiv
[i
] = pParent
->apOvfl
[0];
7261 pgno
= get4byte(apDiv
[i
]);
7262 szNew
[i
] = pParent
->xCellSize(pParent
, apDiv
[i
]);
7263 pParent
->nOverflow
= 0;
7265 apDiv
[i
] = findCell(pParent
, i
+nxDiv
-pParent
->nOverflow
);
7266 pgno
= get4byte(apDiv
[i
]);
7267 szNew
[i
] = pParent
->xCellSize(pParent
, apDiv
[i
]);
7269 /* Drop the cell from the parent page. apDiv[i] still points to
7270 ** the cell within the parent, even though it has been dropped.
7271 ** This is safe because dropping a cell only overwrites the first
7272 ** four bytes of it, and this function does not need the first
7273 ** four bytes of the divider cell. So the pointer is safe to use
7276 ** But not if we are in secure-delete mode. In secure-delete mode,
7277 ** the dropCell() routine will overwrite the entire cell with zeroes.
7278 ** In this case, temporarily copy the cell into the aOvflSpace[]
7279 ** buffer. It will be copied out again as soon as the aSpace[] buffer
7281 if( pBt
->btsFlags
& BTS_FAST_SECURE
){
7284 iOff
= SQLITE_PTR_TO_INT(apDiv
[i
]) - SQLITE_PTR_TO_INT(pParent
->aData
);
7285 if( (iOff
+szNew
[i
])>(int)pBt
->usableSize
){
7286 rc
= SQLITE_CORRUPT_BKPT
;
7287 memset(apOld
, 0, (i
+1)*sizeof(MemPage
*));
7288 goto balance_cleanup
;
7290 memcpy(&aOvflSpace
[iOff
], apDiv
[i
], szNew
[i
]);
7291 apDiv
[i
] = &aOvflSpace
[apDiv
[i
]-pParent
->aData
];
7294 dropCell(pParent
, i
+nxDiv
-pParent
->nOverflow
, szNew
[i
], &rc
);
7298 /* Make nMaxCells a multiple of 4 in order to preserve 8-byte
7300 nMaxCells
= (nMaxCells
+ 3)&~3;
7303 ** Allocate space for memory structures
7306 nMaxCells
*sizeof(u8
*) /* b.apCell */
7307 + nMaxCells
*sizeof(u16
) /* b.szCell */
7308 + pBt
->pageSize
; /* aSpace1 */
7310 assert( szScratch
<=6*(int)pBt
->pageSize
);
7311 b
.apCell
= sqlite3StackAllocRaw(0, szScratch
);
7313 rc
= SQLITE_NOMEM_BKPT
;
7314 goto balance_cleanup
;
7316 b
.szCell
= (u16
*)&b
.apCell
[nMaxCells
];
7317 aSpace1
= (u8
*)&b
.szCell
[nMaxCells
];
7318 assert( EIGHT_BYTE_ALIGNMENT(aSpace1
) );
7321 ** Load pointers to all cells on sibling pages and the divider cells
7322 ** into the local b.apCell[] array. Make copies of the divider cells
7323 ** into space obtained from aSpace1[]. The divider cells have already
7324 ** been removed from pParent.
7326 ** If the siblings are on leaf pages, then the child pointers of the
7327 ** divider cells are stripped from the cells before they are copied
7328 ** into aSpace1[]. In this way, all cells in b.apCell[] are without
7329 ** child pointers. If siblings are not leaves, then all cell in
7330 ** b.apCell[] include child pointers. Either way, all cells in b.apCell[]
7333 ** leafCorrection: 4 if pPage is a leaf. 0 if pPage is not a leaf.
7334 ** leafData: 1 if pPage holds key+data and pParent holds only keys.
7337 leafCorrection
= b
.pRef
->leaf
*4;
7338 leafData
= b
.pRef
->intKeyLeaf
;
7339 for(i
=0; i
<nOld
; i
++){
7340 MemPage
*pOld
= apOld
[i
];
7341 int limit
= pOld
->nCell
;
7342 u8
*aData
= pOld
->aData
;
7343 u16 maskPage
= pOld
->maskPage
;
7344 u8
*piCell
= aData
+ pOld
->cellOffset
;
7347 /* Verify that all sibling pages are of the same "type" (table-leaf,
7348 ** table-interior, index-leaf, or index-interior).
7350 if( pOld
->aData
[0]!=apOld
[0]->aData
[0] ){
7351 rc
= SQLITE_CORRUPT_BKPT
;
7352 goto balance_cleanup
;
7355 /* Load b.apCell[] with pointers to all cells in pOld. If pOld
7356 ** constains overflow cells, include them in the b.apCell[] array
7357 ** in the correct spot.
7359 ** Note that when there are multiple overflow cells, it is always the
7360 ** case that they are sequential and adjacent. This invariant arises
7361 ** because multiple overflows can only occurs when inserting divider
7362 ** cells into a parent on a prior balance, and divider cells are always
7363 ** adjacent and are inserted in order. There is an assert() tagged
7364 ** with "NOTE 1" in the overflow cell insertion loop to prove this
7367 ** This must be done in advance. Once the balance starts, the cell
7368 ** offset section of the btree page will be overwritten and we will no
7369 ** long be able to find the cells if a pointer to each cell is not saved
7372 memset(&b
.szCell
[b
.nCell
], 0, sizeof(b
.szCell
[0])*(limit
+pOld
->nOverflow
));
7373 if( pOld
->nOverflow
>0 ){
7374 limit
= pOld
->aiOvfl
[0];
7375 for(j
=0; j
<limit
; j
++){
7376 b
.apCell
[b
.nCell
] = aData
+ (maskPage
& get2byteAligned(piCell
));
7380 for(k
=0; k
<pOld
->nOverflow
; k
++){
7381 assert( k
==0 || pOld
->aiOvfl
[k
-1]+1==pOld
->aiOvfl
[k
] );/* NOTE 1 */
7382 b
.apCell
[b
.nCell
] = pOld
->apOvfl
[k
];
7386 piEnd
= aData
+ pOld
->cellOffset
+ 2*pOld
->nCell
;
7387 while( piCell
<piEnd
){
7388 assert( b
.nCell
<nMaxCells
);
7389 b
.apCell
[b
.nCell
] = aData
+ (maskPage
& get2byteAligned(piCell
));
7394 cntOld
[i
] = b
.nCell
;
7395 if( i
<nOld
-1 && !leafData
){
7396 u16 sz
= (u16
)szNew
[i
];
7398 assert( b
.nCell
<nMaxCells
);
7399 b
.szCell
[b
.nCell
] = sz
;
7400 pTemp
= &aSpace1
[iSpace1
];
7402 assert( sz
<=pBt
->maxLocal
+23 );
7403 assert( iSpace1
<= (int)pBt
->pageSize
);
7404 memcpy(pTemp
, apDiv
[i
], sz
);
7405 b
.apCell
[b
.nCell
] = pTemp
+leafCorrection
;
7406 assert( leafCorrection
==0 || leafCorrection
==4 );
7407 b
.szCell
[b
.nCell
] = b
.szCell
[b
.nCell
] - leafCorrection
;
7409 assert( leafCorrection
==0 );
7410 assert( pOld
->hdrOffset
==0 );
7411 /* The right pointer of the child page pOld becomes the left
7412 ** pointer of the divider cell */
7413 memcpy(b
.apCell
[b
.nCell
], &pOld
->aData
[8], 4);
7415 assert( leafCorrection
==4 );
7416 while( b
.szCell
[b
.nCell
]<4 ){
7417 /* Do not allow any cells smaller than 4 bytes. If a smaller cell
7418 ** does exist, pad it with 0x00 bytes. */
7419 assert( b
.szCell
[b
.nCell
]==3 || CORRUPT_DB
);
7420 assert( b
.apCell
[b
.nCell
]==&aSpace1
[iSpace1
-3] || CORRUPT_DB
);
7421 aSpace1
[iSpace1
++] = 0x00;
7422 b
.szCell
[b
.nCell
]++;
7430 ** Figure out the number of pages needed to hold all b.nCell cells.
7431 ** Store this number in "k". Also compute szNew[] which is the total
7432 ** size of all cells on the i-th page and cntNew[] which is the index
7433 ** in b.apCell[] of the cell that divides page i from page i+1.
7434 ** cntNew[k] should equal b.nCell.
7436 ** Values computed by this block:
7438 ** k: The total number of sibling pages
7439 ** szNew[i]: Spaced used on the i-th sibling page.
7440 ** cntNew[i]: Index in b.apCell[] and b.szCell[] for the first cell to
7441 ** the right of the i-th sibling page.
7442 ** usableSpace: Number of bytes of space available on each sibling.
7445 usableSpace
= pBt
->usableSize
- 12 + leafCorrection
;
7446 for(i
=0; i
<nOld
; i
++){
7447 MemPage
*p
= apOld
[i
];
7448 szNew
[i
] = usableSpace
- p
->nFree
;
7449 for(j
=0; j
<p
->nOverflow
; j
++){
7450 szNew
[i
] += 2 + p
->xCellSize(p
, p
->apOvfl
[j
]);
7452 cntNew
[i
] = cntOld
[i
];
7457 while( szNew
[i
]>usableSpace
){
7460 if( k
>NB
+2 ){ rc
= SQLITE_CORRUPT_BKPT
; goto balance_cleanup
; }
7462 cntNew
[k
-1] = b
.nCell
;
7464 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]-1);
7467 if( cntNew
[i
]<b
.nCell
){
7468 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7476 while( cntNew
[i
]<b
.nCell
){
7477 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7478 if( szNew
[i
]+sz
>usableSpace
) break;
7482 if( cntNew
[i
]<b
.nCell
){
7483 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7490 if( cntNew
[i
]>=b
.nCell
){
7492 }else if( cntNew
[i
] <= (i
>0 ? cntNew
[i
-1] : 0) ){
7493 rc
= SQLITE_CORRUPT_BKPT
;
7494 goto balance_cleanup
;
7499 ** The packing computed by the previous block is biased toward the siblings
7500 ** on the left side (siblings with smaller keys). The left siblings are
7501 ** always nearly full, while the right-most sibling might be nearly empty.
7502 ** The next block of code attempts to adjust the packing of siblings to
7503 ** get a better balance.
7505 ** This adjustment is more than an optimization. The packing above might
7506 ** be so out of balance as to be illegal. For example, the right-most
7507 ** sibling might be completely empty. This adjustment is not optional.
7509 for(i
=k
-1; i
>0; i
--){
7510 int szRight
= szNew
[i
]; /* Size of sibling on the right */
7511 int szLeft
= szNew
[i
-1]; /* Size of sibling on the left */
7512 int r
; /* Index of right-most cell in left sibling */
7513 int d
; /* Index of first cell to the left of right sibling */
7515 r
= cntNew
[i
-1] - 1;
7516 d
= r
+ 1 - leafData
;
7517 (void)cachedCellSize(&b
, d
);
7519 assert( d
<nMaxCells
);
7520 assert( r
<nMaxCells
);
7521 (void)cachedCellSize(&b
, r
);
7523 && (bBulk
|| szRight
+b
.szCell
[d
]+2 > szLeft
-(b
.szCell
[r
]+(i
==k
-1?0:2)))){
7526 szRight
+= b
.szCell
[d
] + 2;
7527 szLeft
-= b
.szCell
[r
] + 2;
7533 szNew
[i
-1] = szLeft
;
7534 if( cntNew
[i
-1] <= (i
>1 ? cntNew
[i
-2] : 0) ){
7535 rc
= SQLITE_CORRUPT_BKPT
;
7536 goto balance_cleanup
;
7540 /* Sanity check: For a non-corrupt database file one of the follwing
7542 ** (1) We found one or more cells (cntNew[0])>0), or
7543 ** (2) pPage is a virtual root page. A virtual root page is when
7544 ** the real root page is page 1 and we are the only child of
7547 assert( cntNew
[0]>0 || (pParent
->pgno
==1 && pParent
->nCell
==0) || CORRUPT_DB
);
7548 TRACE(("BALANCE: old: %d(nc=%d) %d(nc=%d) %d(nc=%d)\n",
7549 apOld
[0]->pgno
, apOld
[0]->nCell
,
7550 nOld
>=2 ? apOld
[1]->pgno
: 0, nOld
>=2 ? apOld
[1]->nCell
: 0,
7551 nOld
>=3 ? apOld
[2]->pgno
: 0, nOld
>=3 ? apOld
[2]->nCell
: 0
7555 ** Allocate k new pages. Reuse old pages where possible.
7557 pageFlags
= apOld
[0]->aData
[0];
7561 pNew
= apNew
[i
] = apOld
[i
];
7563 rc
= sqlite3PagerWrite(pNew
->pDbPage
);
7565 if( rc
) goto balance_cleanup
;
7568 rc
= allocateBtreePage(pBt
, &pNew
, &pgno
, (bBulk
? 1 : pgno
), 0);
7569 if( rc
) goto balance_cleanup
;
7570 zeroPage(pNew
, pageFlags
);
7573 cntOld
[i
] = b
.nCell
;
7575 /* Set the pointer-map entry for the new sibling page. */
7577 ptrmapPut(pBt
, pNew
->pgno
, PTRMAP_BTREE
, pParent
->pgno
, &rc
);
7578 if( rc
!=SQLITE_OK
){
7579 goto balance_cleanup
;
7586 ** Reassign page numbers so that the new pages are in ascending order.
7587 ** This helps to keep entries in the disk file in order so that a scan
7588 ** of the table is closer to a linear scan through the file. That in turn
7589 ** helps the operating system to deliver pages from the disk more rapidly.
7591 ** An O(n^2) insertion sort algorithm is used, but since n is never more
7592 ** than (NB+2) (a small constant), that should not be a problem.
7594 ** When NB==3, this one optimization makes the database about 25% faster
7595 ** for large insertions and deletions.
7597 for(i
=0; i
<nNew
; i
++){
7598 aPgOrder
[i
] = aPgno
[i
] = apNew
[i
]->pgno
;
7599 aPgFlags
[i
] = apNew
[i
]->pDbPage
->flags
;
7601 if( aPgno
[j
]==aPgno
[i
] ){
7602 /* This branch is taken if the set of sibling pages somehow contains
7603 ** duplicate entries. This can happen if the database is corrupt.
7604 ** It would be simpler to detect this as part of the loop below, but
7605 ** we do the detection here in order to avoid populating the pager
7606 ** cache with two separate objects associated with the same
7608 assert( CORRUPT_DB
);
7609 rc
= SQLITE_CORRUPT_BKPT
;
7610 goto balance_cleanup
;
7614 for(i
=0; i
<nNew
; i
++){
7615 int iBest
= 0; /* aPgno[] index of page number to use */
7616 for(j
=1; j
<nNew
; j
++){
7617 if( aPgOrder
[j
]<aPgOrder
[iBest
] ) iBest
= j
;
7619 pgno
= aPgOrder
[iBest
];
7620 aPgOrder
[iBest
] = 0xffffffff;
7623 sqlite3PagerRekey(apNew
[iBest
]->pDbPage
, pBt
->nPage
+iBest
+1, 0);
7625 sqlite3PagerRekey(apNew
[i
]->pDbPage
, pgno
, aPgFlags
[iBest
]);
7626 apNew
[i
]->pgno
= pgno
;
7630 TRACE(("BALANCE: new: %d(%d nc=%d) %d(%d nc=%d) %d(%d nc=%d) "
7631 "%d(%d nc=%d) %d(%d nc=%d)\n",
7632 apNew
[0]->pgno
, szNew
[0], cntNew
[0],
7633 nNew
>=2 ? apNew
[1]->pgno
: 0, nNew
>=2 ? szNew
[1] : 0,
7634 nNew
>=2 ? cntNew
[1] - cntNew
[0] - !leafData
: 0,
7635 nNew
>=3 ? apNew
[2]->pgno
: 0, nNew
>=3 ? szNew
[2] : 0,
7636 nNew
>=3 ? cntNew
[2] - cntNew
[1] - !leafData
: 0,
7637 nNew
>=4 ? apNew
[3]->pgno
: 0, nNew
>=4 ? szNew
[3] : 0,
7638 nNew
>=4 ? cntNew
[3] - cntNew
[2] - !leafData
: 0,
7639 nNew
>=5 ? apNew
[4]->pgno
: 0, nNew
>=5 ? szNew
[4] : 0,
7640 nNew
>=5 ? cntNew
[4] - cntNew
[3] - !leafData
: 0
7643 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7644 put4byte(pRight
, apNew
[nNew
-1]->pgno
);
7646 /* If the sibling pages are not leaves, ensure that the right-child pointer
7647 ** of the right-most new sibling page is set to the value that was
7648 ** originally in the same field of the right-most old sibling page. */
7649 if( (pageFlags
& PTF_LEAF
)==0 && nOld
!=nNew
){
7650 MemPage
*pOld
= (nNew
>nOld
? apNew
: apOld
)[nOld
-1];
7651 memcpy(&apNew
[nNew
-1]->aData
[8], &pOld
->aData
[8], 4);
7654 /* Make any required updates to pointer map entries associated with
7655 ** cells stored on sibling pages following the balance operation. Pointer
7656 ** map entries associated with divider cells are set by the insertCell()
7657 ** routine. The associated pointer map entries are:
7659 ** a) if the cell contains a reference to an overflow chain, the
7660 ** entry associated with the first page in the overflow chain, and
7662 ** b) if the sibling pages are not leaves, the child page associated
7665 ** If the sibling pages are not leaves, then the pointer map entry
7666 ** associated with the right-child of each sibling may also need to be
7667 ** updated. This happens below, after the sibling pages have been
7668 ** populated, not here.
7671 MemPage
*pNew
= apNew
[0];
7672 u8
*aOld
= pNew
->aData
;
7673 int cntOldNext
= pNew
->nCell
+ pNew
->nOverflow
;
7674 int usableSize
= pBt
->usableSize
;
7678 for(i
=0; i
<b
.nCell
; i
++){
7679 u8
*pCell
= b
.apCell
[i
];
7680 if( i
==cntOldNext
){
7681 MemPage
*pOld
= (++iOld
)<nNew
? apNew
[iOld
] : apOld
[iOld
];
7682 cntOldNext
+= pOld
->nCell
+ pOld
->nOverflow
+ !leafData
;
7685 if( i
==cntNew
[iNew
] ){
7686 pNew
= apNew
[++iNew
];
7687 if( !leafData
) continue;
7690 /* Cell pCell is destined for new sibling page pNew. Originally, it
7691 ** was either part of sibling page iOld (possibly an overflow cell),
7692 ** or else the divider cell to the left of sibling page iOld. So,
7693 ** if sibling page iOld had the same page number as pNew, and if
7694 ** pCell really was a part of sibling page iOld (not a divider or
7695 ** overflow cell), we can skip updating the pointer map entries. */
7697 || pNew
->pgno
!=aPgno
[iOld
]
7698 || !SQLITE_WITHIN(pCell
,aOld
,&aOld
[usableSize
])
7700 if( !leafCorrection
){
7701 ptrmapPut(pBt
, get4byte(pCell
), PTRMAP_BTREE
, pNew
->pgno
, &rc
);
7703 if( cachedCellSize(&b
,i
)>pNew
->minLocal
){
7704 ptrmapPutOvflPtr(pNew
, pCell
, &rc
);
7706 if( rc
) goto balance_cleanup
;
7711 /* Insert new divider cells into pParent. */
7712 for(i
=0; i
<nNew
-1; i
++){
7716 MemPage
*pNew
= apNew
[i
];
7719 assert( j
<nMaxCells
);
7720 assert( b
.apCell
[j
]!=0 );
7721 pCell
= b
.apCell
[j
];
7722 sz
= b
.szCell
[j
] + leafCorrection
;
7723 pTemp
= &aOvflSpace
[iOvflSpace
];
7725 memcpy(&pNew
->aData
[8], pCell
, 4);
7726 }else if( leafData
){
7727 /* If the tree is a leaf-data tree, and the siblings are leaves,
7728 ** then there is no divider cell in b.apCell[]. Instead, the divider
7729 ** cell consists of the integer key for the right-most cell of
7730 ** the sibling-page assembled above only.
7734 pNew
->xParseCell(pNew
, b
.apCell
[j
], &info
);
7736 sz
= 4 + putVarint(&pCell
[4], info
.nKey
);
7740 /* Obscure case for non-leaf-data trees: If the cell at pCell was
7741 ** previously stored on a leaf node, and its reported size was 4
7742 ** bytes, then it may actually be smaller than this
7743 ** (see btreeParseCellPtr(), 4 bytes is the minimum size of
7744 ** any cell). But it is important to pass the correct size to
7745 ** insertCell(), so reparse the cell now.
7747 ** This can only happen for b-trees used to evaluate "IN (SELECT ...)"
7748 ** and WITHOUT ROWID tables with exactly one column which is the
7751 if( b
.szCell
[j
]==4 ){
7752 assert(leafCorrection
==4);
7753 sz
= pParent
->xCellSize(pParent
, pCell
);
7757 assert( sz
<=pBt
->maxLocal
+23 );
7758 assert( iOvflSpace
<= (int)pBt
->pageSize
);
7759 insertCell(pParent
, nxDiv
+i
, pCell
, sz
, pTemp
, pNew
->pgno
, &rc
);
7760 if( rc
!=SQLITE_OK
) goto balance_cleanup
;
7761 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7764 /* Now update the actual sibling pages. The order in which they are updated
7765 ** is important, as this code needs to avoid disrupting any page from which
7766 ** cells may still to be read. In practice, this means:
7768 ** (1) If cells are moving left (from apNew[iPg] to apNew[iPg-1])
7769 ** then it is not safe to update page apNew[iPg] until after
7770 ** the left-hand sibling apNew[iPg-1] has been updated.
7772 ** (2) If cells are moving right (from apNew[iPg] to apNew[iPg+1])
7773 ** then it is not safe to update page apNew[iPg] until after
7774 ** the right-hand sibling apNew[iPg+1] has been updated.
7776 ** If neither of the above apply, the page is safe to update.
7778 ** The iPg value in the following loop starts at nNew-1 goes down
7779 ** to 0, then back up to nNew-1 again, thus making two passes over
7780 ** the pages. On the initial downward pass, only condition (1) above
7781 ** needs to be tested because (2) will always be true from the previous
7782 ** step. On the upward pass, both conditions are always true, so the
7783 ** upwards pass simply processes pages that were missed on the downward
7786 for(i
=1-nNew
; i
<nNew
; i
++){
7787 int iPg
= i
<0 ? -i
: i
;
7788 assert( iPg
>=0 && iPg
<nNew
);
7789 if( abDone
[iPg
] ) continue; /* Skip pages already processed */
7790 if( i
>=0 /* On the upwards pass, or... */
7791 || cntOld
[iPg
-1]>=cntNew
[iPg
-1] /* Condition (1) is true */
7797 /* Verify condition (1): If cells are moving left, update iPg
7798 ** only after iPg-1 has already been updated. */
7799 assert( iPg
==0 || cntOld
[iPg
-1]>=cntNew
[iPg
-1] || abDone
[iPg
-1] );
7801 /* Verify condition (2): If cells are moving right, update iPg
7802 ** only after iPg+1 has already been updated. */
7803 assert( cntNew
[iPg
]>=cntOld
[iPg
] || abDone
[iPg
+1] );
7807 nNewCell
= cntNew
[0];
7809 iOld
= iPg
<nOld
? (cntOld
[iPg
-1] + !leafData
) : b
.nCell
;
7810 iNew
= cntNew
[iPg
-1] + !leafData
;
7811 nNewCell
= cntNew
[iPg
] - iNew
;
7814 rc
= editPage(apNew
[iPg
], iOld
, iNew
, nNewCell
, &b
);
7815 if( rc
) goto balance_cleanup
;
7817 apNew
[iPg
]->nFree
= usableSpace
-szNew
[iPg
];
7818 assert( apNew
[iPg
]->nOverflow
==0 );
7819 assert( apNew
[iPg
]->nCell
==nNewCell
);
7823 /* All pages have been processed exactly once */
7824 assert( memcmp(abDone
, "\01\01\01\01\01", nNew
)==0 );
7829 if( isRoot
&& pParent
->nCell
==0 && pParent
->hdrOffset
<=apNew
[0]->nFree
){
7830 /* The root page of the b-tree now contains no cells. The only sibling
7831 ** page is the right-child of the parent. Copy the contents of the
7832 ** child page into the parent, decreasing the overall height of the
7833 ** b-tree structure by one. This is described as the "balance-shallower"
7834 ** sub-algorithm in some documentation.
7836 ** If this is an auto-vacuum database, the call to copyNodeContent()
7837 ** sets all pointer-map entries corresponding to database image pages
7838 ** for which the pointer is stored within the content being copied.
7840 ** It is critical that the child page be defragmented before being
7841 ** copied into the parent, because if the parent is page 1 then it will
7842 ** by smaller than the child due to the database header, and so all the
7843 ** free space needs to be up front.
7845 assert( nNew
==1 || CORRUPT_DB
);
7846 rc
= defragmentPage(apNew
[0], -1);
7847 testcase( rc
!=SQLITE_OK
);
7848 assert( apNew
[0]->nFree
==
7849 (get2byte(&apNew
[0]->aData
[5])-apNew
[0]->cellOffset
-apNew
[0]->nCell
*2)
7852 copyNodeContent(apNew
[0], pParent
, &rc
);
7853 freePage(apNew
[0], &rc
);
7854 }else if( ISAUTOVACUUM
&& !leafCorrection
){
7855 /* Fix the pointer map entries associated with the right-child of each
7856 ** sibling page. All other pointer map entries have already been taken
7858 for(i
=0; i
<nNew
; i
++){
7859 u32 key
= get4byte(&apNew
[i
]->aData
[8]);
7860 ptrmapPut(pBt
, key
, PTRMAP_BTREE
, apNew
[i
]->pgno
, &rc
);
7864 assert( pParent
->isInit
);
7865 TRACE(("BALANCE: finished: old=%d new=%d cells=%d\n",
7866 nOld
, nNew
, b
.nCell
));
7868 /* Free any old pages that were not reused as new pages.
7870 for(i
=nNew
; i
<nOld
; i
++){
7871 freePage(apOld
[i
], &rc
);
7875 if( ISAUTOVACUUM
&& rc
==SQLITE_OK
&& apNew
[0]->isInit
){
7876 /* The ptrmapCheckPages() contains assert() statements that verify that
7877 ** all pointer map pages are set correctly. This is helpful while
7878 ** debugging. This is usually disabled because a corrupt database may
7879 ** cause an assert() statement to fail. */
7880 ptrmapCheckPages(apNew
, nNew
);
7881 ptrmapCheckPages(&pParent
, 1);
7886 ** Cleanup before returning.
7889 sqlite3StackFree(0, b
.apCell
);
7890 for(i
=0; i
<nOld
; i
++){
7891 releasePage(apOld
[i
]);
7893 for(i
=0; i
<nNew
; i
++){
7894 releasePage(apNew
[i
]);
7902 ** This function is called when the root page of a b-tree structure is
7903 ** overfull (has one or more overflow pages).
7905 ** A new child page is allocated and the contents of the current root
7906 ** page, including overflow cells, are copied into the child. The root
7907 ** page is then overwritten to make it an empty page with the right-child
7908 ** pointer pointing to the new page.
7910 ** Before returning, all pointer-map entries corresponding to pages
7911 ** that the new child-page now contains pointers to are updated. The
7912 ** entry corresponding to the new right-child pointer of the root
7913 ** page is also updated.
7915 ** If successful, *ppChild is set to contain a reference to the child
7916 ** page and SQLITE_OK is returned. In this case the caller is required
7917 ** to call releasePage() on *ppChild exactly once. If an error occurs,
7918 ** an error code is returned and *ppChild is set to 0.
7920 static int balance_deeper(MemPage
*pRoot
, MemPage
**ppChild
){
7921 int rc
; /* Return value from subprocedures */
7922 MemPage
*pChild
= 0; /* Pointer to a new child page */
7923 Pgno pgnoChild
= 0; /* Page number of the new child page */
7924 BtShared
*pBt
= pRoot
->pBt
; /* The BTree */
7926 assert( pRoot
->nOverflow
>0 );
7927 assert( sqlite3_mutex_held(pBt
->mutex
) );
7929 /* Make pRoot, the root page of the b-tree, writable. Allocate a new
7930 ** page that will become the new right-child of pPage. Copy the contents
7931 ** of the node stored on pRoot into the new child page.
7933 rc
= sqlite3PagerWrite(pRoot
->pDbPage
);
7934 if( rc
==SQLITE_OK
){
7935 rc
= allocateBtreePage(pBt
,&pChild
,&pgnoChild
,pRoot
->pgno
,0);
7936 copyNodeContent(pRoot
, pChild
, &rc
);
7938 ptrmapPut(pBt
, pgnoChild
, PTRMAP_BTREE
, pRoot
->pgno
, &rc
);
7943 releasePage(pChild
);
7946 assert( sqlite3PagerIswriteable(pChild
->pDbPage
) );
7947 assert( sqlite3PagerIswriteable(pRoot
->pDbPage
) );
7948 assert( pChild
->nCell
==pRoot
->nCell
);
7950 TRACE(("BALANCE: copy root %d into %d\n", pRoot
->pgno
, pChild
->pgno
));
7952 /* Copy the overflow cells from pRoot to pChild */
7953 memcpy(pChild
->aiOvfl
, pRoot
->aiOvfl
,
7954 pRoot
->nOverflow
*sizeof(pRoot
->aiOvfl
[0]));
7955 memcpy(pChild
->apOvfl
, pRoot
->apOvfl
,
7956 pRoot
->nOverflow
*sizeof(pRoot
->apOvfl
[0]));
7957 pChild
->nOverflow
= pRoot
->nOverflow
;
7959 /* Zero the contents of pRoot. Then install pChild as the right-child. */
7960 zeroPage(pRoot
, pChild
->aData
[0] & ~PTF_LEAF
);
7961 put4byte(&pRoot
->aData
[pRoot
->hdrOffset
+8], pgnoChild
);
7968 ** The page that pCur currently points to has just been modified in
7969 ** some way. This function figures out if this modification means the
7970 ** tree needs to be balanced, and if so calls the appropriate balancing
7971 ** routine. Balancing routines are:
7975 ** balance_nonroot()
7977 static int balance(BtCursor
*pCur
){
7979 const int nMin
= pCur
->pBt
->usableSize
* 2 / 3;
7980 u8 aBalanceQuickSpace
[13];
7983 VVA_ONLY( int balance_quick_called
= 0 );
7984 VVA_ONLY( int balance_deeper_called
= 0 );
7987 int iPage
= pCur
->iPage
;
7988 MemPage
*pPage
= pCur
->pPage
;
7991 if( pPage
->nOverflow
){
7992 /* The root page of the b-tree is overfull. In this case call the
7993 ** balance_deeper() function to create a new child for the root-page
7994 ** and copy the current contents of the root-page to it. The
7995 ** next iteration of the do-loop will balance the child page.
7997 assert( balance_deeper_called
==0 );
7998 VVA_ONLY( balance_deeper_called
++ );
7999 rc
= balance_deeper(pPage
, &pCur
->apPage
[1]);
8000 if( rc
==SQLITE_OK
){
8004 pCur
->apPage
[0] = pPage
;
8005 pCur
->pPage
= pCur
->apPage
[1];
8006 assert( pCur
->pPage
->nOverflow
);
8011 }else if( pPage
->nOverflow
==0 && pPage
->nFree
<=nMin
){
8014 MemPage
* const pParent
= pCur
->apPage
[iPage
-1];
8015 int const iIdx
= pCur
->aiIdx
[iPage
-1];
8017 rc
= sqlite3PagerWrite(pParent
->pDbPage
);
8018 if( rc
==SQLITE_OK
){
8019 #ifndef SQLITE_OMIT_QUICKBALANCE
8020 if( pPage
->intKeyLeaf
8021 && pPage
->nOverflow
==1
8022 && pPage
->aiOvfl
[0]==pPage
->nCell
8024 && pParent
->nCell
==iIdx
8026 /* Call balance_quick() to create a new sibling of pPage on which
8027 ** to store the overflow cell. balance_quick() inserts a new cell
8028 ** into pParent, which may cause pParent overflow. If this
8029 ** happens, the next iteration of the do-loop will balance pParent
8030 ** use either balance_nonroot() or balance_deeper(). Until this
8031 ** happens, the overflow cell is stored in the aBalanceQuickSpace[]
8034 ** The purpose of the following assert() is to check that only a
8035 ** single call to balance_quick() is made for each call to this
8036 ** function. If this were not verified, a subtle bug involving reuse
8037 ** of the aBalanceQuickSpace[] might sneak in.
8039 assert( balance_quick_called
==0 );
8040 VVA_ONLY( balance_quick_called
++ );
8041 rc
= balance_quick(pParent
, pPage
, aBalanceQuickSpace
);
8045 /* In this case, call balance_nonroot() to redistribute cells
8046 ** between pPage and up to 2 of its sibling pages. This involves
8047 ** modifying the contents of pParent, which may cause pParent to
8048 ** become overfull or underfull. The next iteration of the do-loop
8049 ** will balance the parent page to correct this.
8051 ** If the parent page becomes overfull, the overflow cell or cells
8052 ** are stored in the pSpace buffer allocated immediately below.
8053 ** A subsequent iteration of the do-loop will deal with this by
8054 ** calling balance_nonroot() (balance_deeper() may be called first,
8055 ** but it doesn't deal with overflow cells - just moves them to a
8056 ** different page). Once this subsequent call to balance_nonroot()
8057 ** has completed, it is safe to release the pSpace buffer used by
8058 ** the previous call, as the overflow cell data will have been
8059 ** copied either into the body of a database page or into the new
8060 ** pSpace buffer passed to the latter call to balance_nonroot().
8062 u8
*pSpace
= sqlite3PageMalloc(pCur
->pBt
->pageSize
);
8063 rc
= balance_nonroot(pParent
, iIdx
, pSpace
, iPage
==1,
8064 pCur
->hints
&BTREE_BULKLOAD
);
8066 /* If pFree is not NULL, it points to the pSpace buffer used
8067 ** by a previous call to balance_nonroot(). Its contents are
8068 ** now stored either on real database pages or within the
8069 ** new pSpace buffer, so it may be safely freed here. */
8070 sqlite3PageFree(pFree
);
8073 /* The pSpace buffer will be freed after the next call to
8074 ** balance_nonroot(), or just before this function returns, whichever
8080 pPage
->nOverflow
= 0;
8082 /* The next iteration of the do-loop balances the parent page. */
8085 assert( pCur
->iPage
>=0 );
8086 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
8088 }while( rc
==SQLITE_OK
);
8091 sqlite3PageFree(pFree
);
8098 ** Insert a new record into the BTree. The content of the new record
8099 ** is described by the pX object. The pCur cursor is used only to
8100 ** define what table the record should be inserted into, and is left
8101 ** pointing at a random location.
8103 ** For a table btree (used for rowid tables), only the pX.nKey value of
8104 ** the key is used. The pX.pKey value must be NULL. The pX.nKey is the
8105 ** rowid or INTEGER PRIMARY KEY of the row. The pX.nData,pData,nZero fields
8106 ** hold the content of the row.
8108 ** For an index btree (used for indexes and WITHOUT ROWID tables), the
8109 ** key is an arbitrary byte sequence stored in pX.pKey,nKey. The
8110 ** pX.pData,nData,nZero fields must be zero.
8112 ** If the seekResult parameter is non-zero, then a successful call to
8113 ** MovetoUnpacked() to seek cursor pCur to (pKey,nKey) has already
8114 ** been performed. In other words, if seekResult!=0 then the cursor
8115 ** is currently pointing to a cell that will be adjacent to the cell
8116 ** to be inserted. If seekResult<0 then pCur points to a cell that is
8117 ** smaller then (pKey,nKey). If seekResult>0 then pCur points to a cell
8118 ** that is larger than (pKey,nKey).
8120 ** If seekResult==0, that means pCur is pointing at some unknown location.
8121 ** In that case, this routine must seek the cursor to the correct insertion
8122 ** point for (pKey,nKey) before doing the insertion. For index btrees,
8123 ** if pX->nMem is non-zero, then pX->aMem contains pointers to the unpacked
8124 ** key values and pX->aMem can be used instead of pX->pKey to avoid having
8125 ** to decode the key.
8127 int sqlite3BtreeInsert(
8128 BtCursor
*pCur
, /* Insert data into the table of this cursor */
8129 const BtreePayload
*pX
, /* Content of the row to be inserted */
8130 int flags
, /* True if this is likely an append */
8131 int seekResult
/* Result of prior MovetoUnpacked() call */
8134 int loc
= seekResult
; /* -1: before desired location +1: after */
8138 Btree
*p
= pCur
->pBtree
;
8139 BtShared
*pBt
= p
->pBt
;
8140 unsigned char *oldCell
;
8141 unsigned char *newCell
= 0;
8143 assert( (flags
& (BTREE_SAVEPOSITION
|BTREE_APPEND
))==flags
);
8145 if( pCur
->eState
==CURSOR_FAULT
){
8146 assert( pCur
->skipNext
!=SQLITE_OK
);
8147 return pCur
->skipNext
;
8150 assert( cursorOwnsBtShared(pCur
) );
8151 assert( (pCur
->curFlags
& BTCF_WriteFlag
)!=0
8152 && pBt
->inTransaction
==TRANS_WRITE
8153 && (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
8154 assert( hasSharedCacheTableLock(p
, pCur
->pgnoRoot
, pCur
->pKeyInfo
!=0, 2) );
8156 /* Assert that the caller has been consistent. If this cursor was opened
8157 ** expecting an index b-tree, then the caller should be inserting blob
8158 ** keys with no associated data. If the cursor was opened expecting an
8159 ** intkey table, the caller should be inserting integer keys with a
8160 ** blob of associated data. */
8161 assert( (pX
->pKey
==0)==(pCur
->pKeyInfo
==0) );
8163 /* Save the positions of any other cursors open on this table.
8165 ** In some cases, the call to btreeMoveto() below is a no-op. For
8166 ** example, when inserting data into a table with auto-generated integer
8167 ** keys, the VDBE layer invokes sqlite3BtreeLast() to figure out the
8168 ** integer key to use. It then calls this function to actually insert the
8169 ** data into the intkey B-Tree. In this case btreeMoveto() recognizes
8170 ** that the cursor is already where it needs to be and returns without
8171 ** doing any work. To avoid thwarting these optimizations, it is important
8172 ** not to clear the cursor here.
8174 if( pCur
->curFlags
& BTCF_Multiple
){
8175 rc
= saveAllCursors(pBt
, pCur
->pgnoRoot
, pCur
);
8179 if( pCur
->pKeyInfo
==0 ){
8180 assert( pX
->pKey
==0 );
8181 /* If this is an insert into a table b-tree, invalidate any incrblob
8182 ** cursors open on the row being replaced */
8183 invalidateIncrblobCursors(p
, pCur
->pgnoRoot
, pX
->nKey
, 0);
8185 /* If BTREE_SAVEPOSITION is set, the cursor must already be pointing
8186 ** to a row with the same key as the new entry being inserted. */
8187 assert( (flags
& BTREE_SAVEPOSITION
)==0 ||
8188 ((pCur
->curFlags
&BTCF_ValidNKey
)!=0 && pX
->nKey
==pCur
->info
.nKey
) );
8190 /* If the cursor is currently on the last row and we are appending a
8191 ** new row onto the end, set the "loc" to avoid an unnecessary
8192 ** btreeMoveto() call */
8193 if( (pCur
->curFlags
&BTCF_ValidNKey
)!=0 && pX
->nKey
==pCur
->info
.nKey
){
8196 rc
= sqlite3BtreeMovetoUnpacked(pCur
, 0, pX
->nKey
, flags
!=0, &loc
);
8199 }else if( loc
==0 && (flags
& BTREE_SAVEPOSITION
)==0 ){
8202 r
.pKeyInfo
= pCur
->pKeyInfo
;
8204 r
.nField
= pX
->nMem
;
8210 rc
= sqlite3BtreeMovetoUnpacked(pCur
, &r
, 0, flags
!=0, &loc
);
8212 rc
= btreeMoveto(pCur
, pX
->pKey
, pX
->nKey
, flags
!=0, &loc
);
8216 assert( pCur
->eState
==CURSOR_VALID
|| (pCur
->eState
==CURSOR_INVALID
&& loc
) );
8218 pPage
= pCur
->pPage
;
8219 assert( pPage
->intKey
|| pX
->nKey
>=0 );
8220 assert( pPage
->leaf
|| !pPage
->intKey
);
8222 TRACE(("INSERT: table=%d nkey=%lld ndata=%d page=%d %s\n",
8223 pCur
->pgnoRoot
, pX
->nKey
, pX
->nData
, pPage
->pgno
,
8224 loc
==0 ? "overwrite" : "new entry"));
8225 assert( pPage
->isInit
);
8226 newCell
= pBt
->pTmpSpace
;
8227 assert( newCell
!=0 );
8228 rc
= fillInCell(pPage
, newCell
, pX
, &szNew
);
8229 if( rc
) goto end_insert
;
8230 assert( szNew
==pPage
->xCellSize(pPage
, newCell
) );
8231 assert( szNew
<= MX_CELL_SIZE(pBt
) );
8235 assert( idx
<pPage
->nCell
);
8236 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
8240 oldCell
= findCell(pPage
, idx
);
8242 memcpy(newCell
, oldCell
, 4);
8244 rc
= clearCell(pPage
, oldCell
, &info
);
8245 if( info
.nSize
==szNew
&& info
.nLocal
==info
.nPayload
8246 && (!ISAUTOVACUUM
|| szNew
<pPage
->minLocal
)
8248 /* Overwrite the old cell with the new if they are the same size.
8249 ** We could also try to do this if the old cell is smaller, then add
8250 ** the leftover space to the free list. But experiments show that
8251 ** doing that is no faster then skipping this optimization and just
8252 ** calling dropCell() and insertCell().
8254 ** This optimization cannot be used on an autovacuum database if the
8255 ** new entry uses overflow pages, as the insertCell() call below is
8256 ** necessary to add the PTRMAP_OVERFLOW1 pointer-map entry. */
8257 assert( rc
==SQLITE_OK
); /* clearCell never fails when nLocal==nPayload */
8258 if( oldCell
+szNew
> pPage
->aDataEnd
) return SQLITE_CORRUPT_BKPT
;
8259 memcpy(oldCell
, newCell
, szNew
);
8262 dropCell(pPage
, idx
, info
.nSize
, &rc
);
8263 if( rc
) goto end_insert
;
8264 }else if( loc
<0 && pPage
->nCell
>0 ){
8265 assert( pPage
->leaf
);
8267 pCur
->curFlags
&= ~BTCF_ValidNKey
;
8269 assert( pPage
->leaf
);
8271 insertCell(pPage
, idx
, newCell
, szNew
, 0, 0, &rc
);
8272 assert( pPage
->nOverflow
==0 || rc
==SQLITE_OK
);
8273 assert( rc
!=SQLITE_OK
|| pPage
->nCell
>0 || pPage
->nOverflow
>0 );
8275 /* If no error has occurred and pPage has an overflow cell, call balance()
8276 ** to redistribute the cells within the tree. Since balance() may move
8277 ** the cursor, zero the BtCursor.info.nSize and BTCF_ValidNKey
8280 ** Previous versions of SQLite called moveToRoot() to move the cursor
8281 ** back to the root page as balance() used to invalidate the contents
8282 ** of BtCursor.apPage[] and BtCursor.aiIdx[]. Instead of doing that,
8283 ** set the cursor state to "invalid". This makes common insert operations
8286 ** There is a subtle but important optimization here too. When inserting
8287 ** multiple records into an intkey b-tree using a single cursor (as can
8288 ** happen while processing an "INSERT INTO ... SELECT" statement), it
8289 ** is advantageous to leave the cursor pointing to the last entry in
8290 ** the b-tree if possible. If the cursor is left pointing to the last
8291 ** entry in the table, and the next row inserted has an integer key
8292 ** larger than the largest existing key, it is possible to insert the
8293 ** row without seeking the cursor. This can be a big performance boost.
8295 pCur
->info
.nSize
= 0;
8296 if( pPage
->nOverflow
){
8297 assert( rc
==SQLITE_OK
);
8298 pCur
->curFlags
&= ~(BTCF_ValidNKey
);
8301 /* Must make sure nOverflow is reset to zero even if the balance()
8302 ** fails. Internal data structure corruption will result otherwise.
8303 ** Also, set the cursor state to invalid. This stops saveCursorPosition()
8304 ** from trying to save the current position of the cursor. */
8305 pCur
->pPage
->nOverflow
= 0;
8306 pCur
->eState
= CURSOR_INVALID
;
8307 if( (flags
& BTREE_SAVEPOSITION
) && rc
==SQLITE_OK
){
8308 btreeReleaseAllCursorPages(pCur
);
8309 if( pCur
->pKeyInfo
){
8310 assert( pCur
->pKey
==0 );
8311 pCur
->pKey
= sqlite3Malloc( pX
->nKey
);
8312 if( pCur
->pKey
==0 ){
8315 memcpy(pCur
->pKey
, pX
->pKey
, pX
->nKey
);
8318 pCur
->eState
= CURSOR_REQUIRESEEK
;
8319 pCur
->nKey
= pX
->nKey
;
8322 assert( pCur
->iPage
<0 || pCur
->pPage
->nOverflow
==0 );
8329 ** Delete the entry that the cursor is pointing to.
8331 ** If the BTREE_SAVEPOSITION bit of the flags parameter is zero, then
8332 ** the cursor is left pointing at an arbitrary location after the delete.
8333 ** But if that bit is set, then the cursor is left in a state such that
8334 ** the next call to BtreeNext() or BtreePrev() moves it to the same row
8335 ** as it would have been on if the call to BtreeDelete() had been omitted.
8337 ** The BTREE_AUXDELETE bit of flags indicates that is one of several deletes
8338 ** associated with a single table entry and its indexes. Only one of those
8339 ** deletes is considered the "primary" delete. The primary delete occurs
8340 ** on a cursor that is not a BTREE_FORDELETE cursor. All but one delete
8341 ** operation on non-FORDELETE cursors is tagged with the AUXDELETE flag.
8342 ** The BTREE_AUXDELETE bit is a hint that is not used by this implementation,
8343 ** but which might be used by alternative storage engines.
8345 int sqlite3BtreeDelete(BtCursor
*pCur
, u8 flags
){
8346 Btree
*p
= pCur
->pBtree
;
8347 BtShared
*pBt
= p
->pBt
;
8348 int rc
; /* Return code */
8349 MemPage
*pPage
; /* Page to delete cell from */
8350 unsigned char *pCell
; /* Pointer to cell to delete */
8351 int iCellIdx
; /* Index of cell to delete */
8352 int iCellDepth
; /* Depth of node containing pCell */
8353 CellInfo info
; /* Size of the cell being deleted */
8354 int bSkipnext
= 0; /* Leaf cursor in SKIPNEXT state */
8355 u8 bPreserve
= flags
& BTREE_SAVEPOSITION
; /* Keep cursor valid */
8357 assert( cursorOwnsBtShared(pCur
) );
8358 assert( pBt
->inTransaction
==TRANS_WRITE
);
8359 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
8360 assert( pCur
->curFlags
& BTCF_WriteFlag
);
8361 assert( hasSharedCacheTableLock(p
, pCur
->pgnoRoot
, pCur
->pKeyInfo
!=0, 2) );
8362 assert( !hasReadConflicts(p
, pCur
->pgnoRoot
) );
8363 assert( pCur
->ix
<pCur
->pPage
->nCell
);
8364 assert( pCur
->eState
==CURSOR_VALID
);
8365 assert( (flags
& ~(BTREE_SAVEPOSITION
| BTREE_AUXDELETE
))==0 );
8367 iCellDepth
= pCur
->iPage
;
8368 iCellIdx
= pCur
->ix
;
8369 pPage
= pCur
->pPage
;
8370 pCell
= findCell(pPage
, iCellIdx
);
8372 /* If the bPreserve flag is set to true, then the cursor position must
8373 ** be preserved following this delete operation. If the current delete
8374 ** will cause a b-tree rebalance, then this is done by saving the cursor
8375 ** key and leaving the cursor in CURSOR_REQUIRESEEK state before
8378 ** Or, if the current delete will not cause a rebalance, then the cursor
8379 ** will be left in CURSOR_SKIPNEXT state pointing to the entry immediately
8380 ** before or after the deleted entry. In this case set bSkipnext to true. */
8383 || (pPage
->nFree
+cellSizePtr(pPage
,pCell
)+2)>(int)(pBt
->usableSize
*2/3)
8385 /* A b-tree rebalance will be required after deleting this entry.
8386 ** Save the cursor key. */
8387 rc
= saveCursorKey(pCur
);
8394 /* If the page containing the entry to delete is not a leaf page, move
8395 ** the cursor to the largest entry in the tree that is smaller than
8396 ** the entry being deleted. This cell will replace the cell being deleted
8397 ** from the internal node. The 'previous' entry is used for this instead
8398 ** of the 'next' entry, as the previous entry is always a part of the
8399 ** sub-tree headed by the child page of the cell being deleted. This makes
8400 ** balancing the tree following the delete operation easier. */
8402 rc
= sqlite3BtreePrevious(pCur
, 0);
8403 assert( rc
!=SQLITE_DONE
);
8407 /* Save the positions of any other cursors open on this table before
8408 ** making any modifications. */
8409 if( pCur
->curFlags
& BTCF_Multiple
){
8410 rc
= saveAllCursors(pBt
, pCur
->pgnoRoot
, pCur
);
8414 /* If this is a delete operation to remove a row from a table b-tree,
8415 ** invalidate any incrblob cursors open on the row being deleted. */
8416 if( pCur
->pKeyInfo
==0 ){
8417 invalidateIncrblobCursors(p
, pCur
->pgnoRoot
, pCur
->info
.nKey
, 0);
8420 /* Make the page containing the entry to be deleted writable. Then free any
8421 ** overflow pages associated with the entry and finally remove the cell
8422 ** itself from within the page. */
8423 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
8425 rc
= clearCell(pPage
, pCell
, &info
);
8426 dropCell(pPage
, iCellIdx
, info
.nSize
, &rc
);
8429 /* If the cell deleted was not located on a leaf page, then the cursor
8430 ** is currently pointing to the largest entry in the sub-tree headed
8431 ** by the child-page of the cell that was just deleted from an internal
8432 ** node. The cell from the leaf node needs to be moved to the internal
8433 ** node to replace the deleted cell. */
8435 MemPage
*pLeaf
= pCur
->pPage
;
8438 unsigned char *pTmp
;
8440 if( iCellDepth
<pCur
->iPage
-1 ){
8441 n
= pCur
->apPage
[iCellDepth
+1]->pgno
;
8443 n
= pCur
->pPage
->pgno
;
8445 pCell
= findCell(pLeaf
, pLeaf
->nCell
-1);
8446 if( pCell
<&pLeaf
->aData
[4] ) return SQLITE_CORRUPT_BKPT
;
8447 nCell
= pLeaf
->xCellSize(pLeaf
, pCell
);
8448 assert( MX_CELL_SIZE(pBt
) >= nCell
);
8449 pTmp
= pBt
->pTmpSpace
;
8451 rc
= sqlite3PagerWrite(pLeaf
->pDbPage
);
8452 if( rc
==SQLITE_OK
){
8453 insertCell(pPage
, iCellIdx
, pCell
-4, nCell
+4, pTmp
, n
, &rc
);
8455 dropCell(pLeaf
, pLeaf
->nCell
-1, nCell
, &rc
);
8459 /* Balance the tree. If the entry deleted was located on a leaf page,
8460 ** then the cursor still points to that page. In this case the first
8461 ** call to balance() repairs the tree, and the if(...) condition is
8464 ** Otherwise, if the entry deleted was on an internal node page, then
8465 ** pCur is pointing to the leaf page from which a cell was removed to
8466 ** replace the cell deleted from the internal node. This is slightly
8467 ** tricky as the leaf node may be underfull, and the internal node may
8468 ** be either under or overfull. In this case run the balancing algorithm
8469 ** on the leaf node first. If the balance proceeds far enough up the
8470 ** tree that we can be sure that any problem in the internal node has
8471 ** been corrected, so be it. Otherwise, after balancing the leaf node,
8472 ** walk the cursor up the tree to the internal node and balance it as
8475 if( rc
==SQLITE_OK
&& pCur
->iPage
>iCellDepth
){
8476 releasePageNotNull(pCur
->pPage
);
8478 while( pCur
->iPage
>iCellDepth
){
8479 releasePage(pCur
->apPage
[pCur
->iPage
--]);
8481 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
8485 if( rc
==SQLITE_OK
){
8487 assert( bPreserve
&& (pCur
->iPage
==iCellDepth
|| CORRUPT_DB
) );
8488 assert( pPage
==pCur
->pPage
|| CORRUPT_DB
);
8489 assert( (pPage
->nCell
>0 || CORRUPT_DB
) && iCellIdx
<=pPage
->nCell
);
8490 pCur
->eState
= CURSOR_SKIPNEXT
;
8491 if( iCellIdx
>=pPage
->nCell
){
8492 pCur
->skipNext
= -1;
8493 pCur
->ix
= pPage
->nCell
-1;
8498 rc
= moveToRoot(pCur
);
8500 btreeReleaseAllCursorPages(pCur
);
8501 pCur
->eState
= CURSOR_REQUIRESEEK
;
8503 if( rc
==SQLITE_EMPTY
) rc
= SQLITE_OK
;
8510 ** Create a new BTree table. Write into *piTable the page
8511 ** number for the root page of the new table.
8513 ** The type of type is determined by the flags parameter. Only the
8514 ** following values of flags are currently in use. Other values for
8515 ** flags might not work:
8517 ** BTREE_INTKEY|BTREE_LEAFDATA Used for SQL tables with rowid keys
8518 ** BTREE_ZERODATA Used for SQL indices
8520 static int btreeCreateTable(Btree
*p
, int *piTable
, int createTabFlags
){
8521 BtShared
*pBt
= p
->pBt
;
8525 int ptfFlags
; /* Page-type flage for the root page of new table */
8527 assert( sqlite3BtreeHoldsMutex(p
) );
8528 assert( pBt
->inTransaction
==TRANS_WRITE
);
8529 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
8531 #ifdef SQLITE_OMIT_AUTOVACUUM
8532 rc
= allocateBtreePage(pBt
, &pRoot
, &pgnoRoot
, 1, 0);
8537 if( pBt
->autoVacuum
){
8538 Pgno pgnoMove
; /* Move a page here to make room for the root-page */
8539 MemPage
*pPageMove
; /* The page to move to. */
8541 /* Creating a new table may probably require moving an existing database
8542 ** to make room for the new tables root page. In case this page turns
8543 ** out to be an overflow page, delete all overflow page-map caches
8544 ** held by open cursors.
8546 invalidateAllOverflowCache(pBt
);
8548 /* Read the value of meta[3] from the database to determine where the
8549 ** root page of the new table should go. meta[3] is the largest root-page
8550 ** created so far, so the new root-page is (meta[3]+1).
8552 sqlite3BtreeGetMeta(p
, BTREE_LARGEST_ROOT_PAGE
, &pgnoRoot
);
8555 /* The new root-page may not be allocated on a pointer-map page, or the
8556 ** PENDING_BYTE page.
8558 while( pgnoRoot
==PTRMAP_PAGENO(pBt
, pgnoRoot
) ||
8559 pgnoRoot
==PENDING_BYTE_PAGE(pBt
) ){
8562 assert( pgnoRoot
>=3 || CORRUPT_DB
);
8563 testcase( pgnoRoot
<3 );
8565 /* Allocate a page. The page that currently resides at pgnoRoot will
8566 ** be moved to the allocated page (unless the allocated page happens
8567 ** to reside at pgnoRoot).
8569 rc
= allocateBtreePage(pBt
, &pPageMove
, &pgnoMove
, pgnoRoot
, BTALLOC_EXACT
);
8570 if( rc
!=SQLITE_OK
){
8574 if( pgnoMove
!=pgnoRoot
){
8575 /* pgnoRoot is the page that will be used for the root-page of
8576 ** the new table (assuming an error did not occur). But we were
8577 ** allocated pgnoMove. If required (i.e. if it was not allocated
8578 ** by extending the file), the current page at position pgnoMove
8579 ** is already journaled.
8584 /* Save the positions of any open cursors. This is required in
8585 ** case they are holding a reference to an xFetch reference
8586 ** corresponding to page pgnoRoot. */
8587 rc
= saveAllCursors(pBt
, 0, 0);
8588 releasePage(pPageMove
);
8589 if( rc
!=SQLITE_OK
){
8593 /* Move the page currently at pgnoRoot to pgnoMove. */
8594 rc
= btreeGetPage(pBt
, pgnoRoot
, &pRoot
, 0);
8595 if( rc
!=SQLITE_OK
){
8598 rc
= ptrmapGet(pBt
, pgnoRoot
, &eType
, &iPtrPage
);
8599 if( eType
==PTRMAP_ROOTPAGE
|| eType
==PTRMAP_FREEPAGE
){
8600 rc
= SQLITE_CORRUPT_BKPT
;
8602 if( rc
!=SQLITE_OK
){
8606 assert( eType
!=PTRMAP_ROOTPAGE
);
8607 assert( eType
!=PTRMAP_FREEPAGE
);
8608 rc
= relocatePage(pBt
, pRoot
, eType
, iPtrPage
, pgnoMove
, 0);
8611 /* Obtain the page at pgnoRoot */
8612 if( rc
!=SQLITE_OK
){
8615 rc
= btreeGetPage(pBt
, pgnoRoot
, &pRoot
, 0);
8616 if( rc
!=SQLITE_OK
){
8619 rc
= sqlite3PagerWrite(pRoot
->pDbPage
);
8620 if( rc
!=SQLITE_OK
){
8628 /* Update the pointer-map and meta-data with the new root-page number. */
8629 ptrmapPut(pBt
, pgnoRoot
, PTRMAP_ROOTPAGE
, 0, &rc
);
8635 /* When the new root page was allocated, page 1 was made writable in
8636 ** order either to increase the database filesize, or to decrement the
8637 ** freelist count. Hence, the sqlite3BtreeUpdateMeta() call cannot fail.
8639 assert( sqlite3PagerIswriteable(pBt
->pPage1
->pDbPage
) );
8640 rc
= sqlite3BtreeUpdateMeta(p
, 4, pgnoRoot
);
8647 rc
= allocateBtreePage(pBt
, &pRoot
, &pgnoRoot
, 1, 0);
8651 assert( sqlite3PagerIswriteable(pRoot
->pDbPage
) );
8652 if( createTabFlags
& BTREE_INTKEY
){
8653 ptfFlags
= PTF_INTKEY
| PTF_LEAFDATA
| PTF_LEAF
;
8655 ptfFlags
= PTF_ZERODATA
| PTF_LEAF
;
8657 zeroPage(pRoot
, ptfFlags
);
8658 sqlite3PagerUnref(pRoot
->pDbPage
);
8659 assert( (pBt
->openFlags
& BTREE_SINGLE
)==0 || pgnoRoot
==2 );
8660 *piTable
= (int)pgnoRoot
;
8663 int sqlite3BtreeCreateTable(Btree
*p
, int *piTable
, int flags
){
8665 sqlite3BtreeEnter(p
);
8666 rc
= btreeCreateTable(p
, piTable
, flags
);
8667 sqlite3BtreeLeave(p
);
8672 ** Erase the given database page and all its children. Return
8673 ** the page to the freelist.
8675 static int clearDatabasePage(
8676 BtShared
*pBt
, /* The BTree that contains the table */
8677 Pgno pgno
, /* Page number to clear */
8678 int freePageFlag
, /* Deallocate page if true */
8679 int *pnChange
/* Add number of Cells freed to this counter */
8683 unsigned char *pCell
;
8688 assert( sqlite3_mutex_held(pBt
->mutex
) );
8689 if( pgno
>btreePagecount(pBt
) ){
8690 return SQLITE_CORRUPT_BKPT
;
8692 rc
= getAndInitPage(pBt
, pgno
, &pPage
, 0, 0);
8695 rc
= SQLITE_CORRUPT_BKPT
;
8696 goto cleardatabasepage_out
;
8699 hdr
= pPage
->hdrOffset
;
8700 for(i
=0; i
<pPage
->nCell
; i
++){
8701 pCell
= findCell(pPage
, i
);
8703 rc
= clearDatabasePage(pBt
, get4byte(pCell
), 1, pnChange
);
8704 if( rc
) goto cleardatabasepage_out
;
8706 rc
= clearCell(pPage
, pCell
, &info
);
8707 if( rc
) goto cleardatabasepage_out
;
8710 rc
= clearDatabasePage(pBt
, get4byte(&pPage
->aData
[hdr
+8]), 1, pnChange
);
8711 if( rc
) goto cleardatabasepage_out
;
8712 }else if( pnChange
){
8713 assert( pPage
->intKey
|| CORRUPT_DB
);
8714 testcase( !pPage
->intKey
);
8715 *pnChange
+= pPage
->nCell
;
8718 freePage(pPage
, &rc
);
8719 }else if( (rc
= sqlite3PagerWrite(pPage
->pDbPage
))==0 ){
8720 zeroPage(pPage
, pPage
->aData
[hdr
] | PTF_LEAF
);
8723 cleardatabasepage_out
:
8730 ** Delete all information from a single table in the database. iTable is
8731 ** the page number of the root of the table. After this routine returns,
8732 ** the root page is empty, but still exists.
8734 ** This routine will fail with SQLITE_LOCKED if there are any open
8735 ** read cursors on the table. Open write cursors are moved to the
8736 ** root of the table.
8738 ** If pnChange is not NULL, then table iTable must be an intkey table. The
8739 ** integer value pointed to by pnChange is incremented by the number of
8740 ** entries in the table.
8742 int sqlite3BtreeClearTable(Btree
*p
, int iTable
, int *pnChange
){
8744 BtShared
*pBt
= p
->pBt
;
8745 sqlite3BtreeEnter(p
);
8746 assert( p
->inTrans
==TRANS_WRITE
);
8748 rc
= saveAllCursors(pBt
, (Pgno
)iTable
, 0);
8750 if( SQLITE_OK
==rc
){
8751 /* Invalidate all incrblob cursors open on table iTable (assuming iTable
8752 ** is the root of a table b-tree - if it is not, the following call is
8754 invalidateIncrblobCursors(p
, (Pgno
)iTable
, 0, 1);
8755 rc
= clearDatabasePage(pBt
, (Pgno
)iTable
, 0, pnChange
);
8757 sqlite3BtreeLeave(p
);
8762 ** Delete all information from the single table that pCur is open on.
8764 ** This routine only work for pCur on an ephemeral table.
8766 int sqlite3BtreeClearTableOfCursor(BtCursor
*pCur
){
8767 return sqlite3BtreeClearTable(pCur
->pBtree
, pCur
->pgnoRoot
, 0);
8771 ** Erase all information in a table and add the root of the table to
8772 ** the freelist. Except, the root of the principle table (the one on
8773 ** page 1) is never added to the freelist.
8775 ** This routine will fail with SQLITE_LOCKED if there are any open
8776 ** cursors on the table.
8778 ** If AUTOVACUUM is enabled and the page at iTable is not the last
8779 ** root page in the database file, then the last root page
8780 ** in the database file is moved into the slot formerly occupied by
8781 ** iTable and that last slot formerly occupied by the last root page
8782 ** is added to the freelist instead of iTable. In this say, all
8783 ** root pages are kept at the beginning of the database file, which
8784 ** is necessary for AUTOVACUUM to work right. *piMoved is set to the
8785 ** page number that used to be the last root page in the file before
8786 ** the move. If no page gets moved, *piMoved is set to 0.
8787 ** The last root page is recorded in meta[3] and the value of
8788 ** meta[3] is updated by this procedure.
8790 static int btreeDropTable(Btree
*p
, Pgno iTable
, int *piMoved
){
8793 BtShared
*pBt
= p
->pBt
;
8795 assert( sqlite3BtreeHoldsMutex(p
) );
8796 assert( p
->inTrans
==TRANS_WRITE
);
8797 assert( iTable
>=2 );
8799 rc
= btreeGetPage(pBt
, (Pgno
)iTable
, &pPage
, 0);
8801 rc
= sqlite3BtreeClearTable(p
, iTable
, 0);
8809 #ifdef SQLITE_OMIT_AUTOVACUUM
8810 freePage(pPage
, &rc
);
8813 if( pBt
->autoVacuum
){
8815 sqlite3BtreeGetMeta(p
, BTREE_LARGEST_ROOT_PAGE
, &maxRootPgno
);
8817 if( iTable
==maxRootPgno
){
8818 /* If the table being dropped is the table with the largest root-page
8819 ** number in the database, put the root page on the free list.
8821 freePage(pPage
, &rc
);
8823 if( rc
!=SQLITE_OK
){
8827 /* The table being dropped does not have the largest root-page
8828 ** number in the database. So move the page that does into the
8829 ** gap left by the deleted root-page.
8833 rc
= btreeGetPage(pBt
, maxRootPgno
, &pMove
, 0);
8834 if( rc
!=SQLITE_OK
){
8837 rc
= relocatePage(pBt
, pMove
, PTRMAP_ROOTPAGE
, 0, iTable
, 0);
8839 if( rc
!=SQLITE_OK
){
8843 rc
= btreeGetPage(pBt
, maxRootPgno
, &pMove
, 0);
8844 freePage(pMove
, &rc
);
8846 if( rc
!=SQLITE_OK
){
8849 *piMoved
= maxRootPgno
;
8852 /* Set the new 'max-root-page' value in the database header. This
8853 ** is the old value less one, less one more if that happens to
8854 ** be a root-page number, less one again if that is the
8855 ** PENDING_BYTE_PAGE.
8858 while( maxRootPgno
==PENDING_BYTE_PAGE(pBt
)
8859 || PTRMAP_ISPAGE(pBt
, maxRootPgno
) ){
8862 assert( maxRootPgno
!=PENDING_BYTE_PAGE(pBt
) );
8864 rc
= sqlite3BtreeUpdateMeta(p
, 4, maxRootPgno
);
8866 freePage(pPage
, &rc
);
8872 int sqlite3BtreeDropTable(Btree
*p
, int iTable
, int *piMoved
){
8874 sqlite3BtreeEnter(p
);
8875 rc
= btreeDropTable(p
, iTable
, piMoved
);
8876 sqlite3BtreeLeave(p
);
8882 ** This function may only be called if the b-tree connection already
8883 ** has a read or write transaction open on the database.
8885 ** Read the meta-information out of a database file. Meta[0]
8886 ** is the number of free pages currently in the database. Meta[1]
8887 ** through meta[15] are available for use by higher layers. Meta[0]
8888 ** is read-only, the others are read/write.
8890 ** The schema layer numbers meta values differently. At the schema
8891 ** layer (and the SetCookie and ReadCookie opcodes) the number of
8892 ** free pages is not visible. So Cookie[0] is the same as Meta[1].
8894 ** This routine treats Meta[BTREE_DATA_VERSION] as a special case. Instead
8895 ** of reading the value out of the header, it instead loads the "DataVersion"
8896 ** from the pager. The BTREE_DATA_VERSION value is not actually stored in the
8897 ** database file. It is a number computed by the pager. But its access
8898 ** pattern is the same as header meta values, and so it is convenient to
8899 ** read it from this routine.
8901 void sqlite3BtreeGetMeta(Btree
*p
, int idx
, u32
*pMeta
){
8902 BtShared
*pBt
= p
->pBt
;
8904 sqlite3BtreeEnter(p
);
8905 assert( p
->inTrans
>TRANS_NONE
);
8906 assert( SQLITE_OK
==querySharedCacheTableLock(p
, MASTER_ROOT
, READ_LOCK
) );
8907 assert( pBt
->pPage1
);
8908 assert( idx
>=0 && idx
<=15 );
8910 if( idx
==BTREE_DATA_VERSION
){
8911 *pMeta
= sqlite3PagerDataVersion(pBt
->pPager
) + p
->iDataVersion
;
8913 *pMeta
= get4byte(&pBt
->pPage1
->aData
[36 + idx
*4]);
8916 /* If auto-vacuum is disabled in this build and this is an auto-vacuum
8917 ** database, mark the database as read-only. */
8918 #ifdef SQLITE_OMIT_AUTOVACUUM
8919 if( idx
==BTREE_LARGEST_ROOT_PAGE
&& *pMeta
>0 ){
8920 pBt
->btsFlags
|= BTS_READ_ONLY
;
8924 sqlite3BtreeLeave(p
);
8928 ** Write meta-information back into the database. Meta[0] is
8929 ** read-only and may not be written.
8931 int sqlite3BtreeUpdateMeta(Btree
*p
, int idx
, u32 iMeta
){
8932 BtShared
*pBt
= p
->pBt
;
8935 assert( idx
>=1 && idx
<=15 );
8936 sqlite3BtreeEnter(p
);
8937 assert( p
->inTrans
==TRANS_WRITE
);
8938 assert( pBt
->pPage1
!=0 );
8939 pP1
= pBt
->pPage1
->aData
;
8940 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
8941 if( rc
==SQLITE_OK
){
8942 put4byte(&pP1
[36 + idx
*4], iMeta
);
8943 #ifndef SQLITE_OMIT_AUTOVACUUM
8944 if( idx
==BTREE_INCR_VACUUM
){
8945 assert( pBt
->autoVacuum
|| iMeta
==0 );
8946 assert( iMeta
==0 || iMeta
==1 );
8947 pBt
->incrVacuum
= (u8
)iMeta
;
8951 sqlite3BtreeLeave(p
);
8955 #ifndef SQLITE_OMIT_BTREECOUNT
8957 ** The first argument, pCur, is a cursor opened on some b-tree. Count the
8958 ** number of entries in the b-tree and write the result to *pnEntry.
8960 ** SQLITE_OK is returned if the operation is successfully executed.
8961 ** Otherwise, if an error is encountered (i.e. an IO error or database
8962 ** corruption) an SQLite error code is returned.
8964 int sqlite3BtreeCount(BtCursor
*pCur
, i64
*pnEntry
){
8965 i64 nEntry
= 0; /* Value to return in *pnEntry */
8966 int rc
; /* Return code */
8968 rc
= moveToRoot(pCur
);
8969 if( rc
==SQLITE_EMPTY
){
8974 /* Unless an error occurs, the following loop runs one iteration for each
8975 ** page in the B-Tree structure (not including overflow pages).
8977 while( rc
==SQLITE_OK
){
8978 int iIdx
; /* Index of child node in parent */
8979 MemPage
*pPage
; /* Current page of the b-tree */
8981 /* If this is a leaf page or the tree is not an int-key tree, then
8982 ** this page contains countable entries. Increment the entry counter
8985 pPage
= pCur
->pPage
;
8986 if( pPage
->leaf
|| !pPage
->intKey
){
8987 nEntry
+= pPage
->nCell
;
8990 /* pPage is a leaf node. This loop navigates the cursor so that it
8991 ** points to the first interior cell that it points to the parent of
8992 ** the next page in the tree that has not yet been visited. The
8993 ** pCur->aiIdx[pCur->iPage] value is set to the index of the parent cell
8994 ** of the page, or to the number of cells in the page if the next page
8995 ** to visit is the right-child of its parent.
8997 ** If all pages in the tree have been visited, return SQLITE_OK to the
9002 if( pCur
->iPage
==0 ){
9003 /* All pages of the b-tree have been visited. Return successfully. */
9005 return moveToRoot(pCur
);
9008 }while ( pCur
->ix
>=pCur
->pPage
->nCell
);
9011 pPage
= pCur
->pPage
;
9014 /* Descend to the child node of the cell that the cursor currently
9015 ** points at. This is the right-child if (iIdx==pPage->nCell).
9018 if( iIdx
==pPage
->nCell
){
9019 rc
= moveToChild(pCur
, get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]));
9021 rc
= moveToChild(pCur
, get4byte(findCell(pPage
, iIdx
)));
9025 /* An error has occurred. Return an error code. */
9031 ** Return the pager associated with a BTree. This routine is used for
9032 ** testing and debugging only.
9034 Pager
*sqlite3BtreePager(Btree
*p
){
9035 return p
->pBt
->pPager
;
9038 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9040 ** Append a message to the error message string.
9042 static void checkAppendMsg(
9043 IntegrityCk
*pCheck
,
9044 const char *zFormat
,
9048 if( !pCheck
->mxErr
) return;
9051 va_start(ap
, zFormat
);
9052 if( pCheck
->errMsg
.nChar
){
9053 sqlite3StrAccumAppend(&pCheck
->errMsg
, "\n", 1);
9056 sqlite3XPrintf(&pCheck
->errMsg
, pCheck
->zPfx
, pCheck
->v1
, pCheck
->v2
);
9058 sqlite3VXPrintf(&pCheck
->errMsg
, zFormat
, ap
);
9060 if( pCheck
->errMsg
.accError
==STRACCUM_NOMEM
){
9061 pCheck
->mallocFailed
= 1;
9064 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9066 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9069 ** Return non-zero if the bit in the IntegrityCk.aPgRef[] array that
9070 ** corresponds to page iPg is already set.
9072 static int getPageReferenced(IntegrityCk
*pCheck
, Pgno iPg
){
9073 assert( iPg
<=pCheck
->nPage
&& sizeof(pCheck
->aPgRef
[0])==1 );
9074 return (pCheck
->aPgRef
[iPg
/8] & (1 << (iPg
& 0x07)));
9078 ** Set the bit in the IntegrityCk.aPgRef[] array that corresponds to page iPg.
9080 static void setPageReferenced(IntegrityCk
*pCheck
, Pgno iPg
){
9081 assert( iPg
<=pCheck
->nPage
&& sizeof(pCheck
->aPgRef
[0])==1 );
9082 pCheck
->aPgRef
[iPg
/8] |= (1 << (iPg
& 0x07));
9087 ** Add 1 to the reference count for page iPage. If this is the second
9088 ** reference to the page, add an error message to pCheck->zErrMsg.
9089 ** Return 1 if there are 2 or more references to the page and 0 if
9090 ** if this is the first reference to the page.
9092 ** Also check that the page number is in bounds.
9094 static int checkRef(IntegrityCk
*pCheck
, Pgno iPage
){
9095 if( iPage
==0 ) return 1;
9096 if( iPage
>pCheck
->nPage
){
9097 checkAppendMsg(pCheck
, "invalid page number %d", iPage
);
9100 if( getPageReferenced(pCheck
, iPage
) ){
9101 checkAppendMsg(pCheck
, "2nd reference to page %d", iPage
);
9104 setPageReferenced(pCheck
, iPage
);
9108 #ifndef SQLITE_OMIT_AUTOVACUUM
9110 ** Check that the entry in the pointer-map for page iChild maps to
9111 ** page iParent, pointer type ptrType. If not, append an error message
9114 static void checkPtrmap(
9115 IntegrityCk
*pCheck
, /* Integrity check context */
9116 Pgno iChild
, /* Child page number */
9117 u8 eType
, /* Expected pointer map type */
9118 Pgno iParent
/* Expected pointer map parent page number */
9124 rc
= ptrmapGet(pCheck
->pBt
, iChild
, &ePtrmapType
, &iPtrmapParent
);
9125 if( rc
!=SQLITE_OK
){
9126 if( rc
==SQLITE_NOMEM
|| rc
==SQLITE_IOERR_NOMEM
) pCheck
->mallocFailed
= 1;
9127 checkAppendMsg(pCheck
, "Failed to read ptrmap key=%d", iChild
);
9131 if( ePtrmapType
!=eType
|| iPtrmapParent
!=iParent
){
9132 checkAppendMsg(pCheck
,
9133 "Bad ptr map entry key=%d expected=(%d,%d) got=(%d,%d)",
9134 iChild
, eType
, iParent
, ePtrmapType
, iPtrmapParent
);
9140 ** Check the integrity of the freelist or of an overflow page list.
9141 ** Verify that the number of pages on the list is N.
9143 static void checkList(
9144 IntegrityCk
*pCheck
, /* Integrity checking context */
9145 int isFreeList
, /* True for a freelist. False for overflow page list */
9146 int iPage
, /* Page number for first page in the list */
9147 int N
/* Expected number of pages in the list */
9152 while( N
-- > 0 && pCheck
->mxErr
){
9154 unsigned char *pOvflData
;
9156 checkAppendMsg(pCheck
,
9157 "%d of %d pages missing from overflow list starting at %d",
9158 N
+1, expected
, iFirst
);
9161 if( checkRef(pCheck
, iPage
) ) break;
9162 if( sqlite3PagerGet(pCheck
->pPager
, (Pgno
)iPage
, &pOvflPage
, 0) ){
9163 checkAppendMsg(pCheck
, "failed to get page %d", iPage
);
9166 pOvflData
= (unsigned char *)sqlite3PagerGetData(pOvflPage
);
9168 int n
= get4byte(&pOvflData
[4]);
9169 #ifndef SQLITE_OMIT_AUTOVACUUM
9170 if( pCheck
->pBt
->autoVacuum
){
9171 checkPtrmap(pCheck
, iPage
, PTRMAP_FREEPAGE
, 0);
9174 if( n
>(int)pCheck
->pBt
->usableSize
/4-2 ){
9175 checkAppendMsg(pCheck
,
9176 "freelist leaf count too big on page %d", iPage
);
9180 Pgno iFreePage
= get4byte(&pOvflData
[8+i
*4]);
9181 #ifndef SQLITE_OMIT_AUTOVACUUM
9182 if( pCheck
->pBt
->autoVacuum
){
9183 checkPtrmap(pCheck
, iFreePage
, PTRMAP_FREEPAGE
, 0);
9186 checkRef(pCheck
, iFreePage
);
9191 #ifndef SQLITE_OMIT_AUTOVACUUM
9193 /* If this database supports auto-vacuum and iPage is not the last
9194 ** page in this overflow list, check that the pointer-map entry for
9195 ** the following page matches iPage.
9197 if( pCheck
->pBt
->autoVacuum
&& N
>0 ){
9198 i
= get4byte(pOvflData
);
9199 checkPtrmap(pCheck
, i
, PTRMAP_OVERFLOW2
, iPage
);
9203 iPage
= get4byte(pOvflData
);
9204 sqlite3PagerUnref(pOvflPage
);
9206 if( isFreeList
&& N
<(iPage
!=0) ){
9207 checkAppendMsg(pCheck
, "free-page count in header is too small");
9211 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9214 ** An implementation of a min-heap.
9216 ** aHeap[0] is the number of elements on the heap. aHeap[1] is the
9217 ** root element. The daughter nodes of aHeap[N] are aHeap[N*2]
9218 ** and aHeap[N*2+1].
9220 ** The heap property is this: Every node is less than or equal to both
9221 ** of its daughter nodes. A consequence of the heap property is that the
9222 ** root node aHeap[1] is always the minimum value currently in the heap.
9224 ** The btreeHeapInsert() routine inserts an unsigned 32-bit number onto
9225 ** the heap, preserving the heap property. The btreeHeapPull() routine
9226 ** removes the root element from the heap (the minimum value in the heap)
9227 ** and then moves other nodes around as necessary to preserve the heap
9230 ** This heap is used for cell overlap and coverage testing. Each u32
9231 ** entry represents the span of a cell or freeblock on a btree page.
9232 ** The upper 16 bits are the index of the first byte of a range and the
9233 ** lower 16 bits are the index of the last byte of that range.
9235 static void btreeHeapInsert(u32
*aHeap
, u32 x
){
9236 u32 j
, i
= ++aHeap
[0];
9238 while( (j
= i
/2)>0 && aHeap
[j
]>aHeap
[i
] ){
9240 aHeap
[j
] = aHeap
[i
];
9245 static int btreeHeapPull(u32
*aHeap
, u32
*pOut
){
9247 if( (x
= aHeap
[0])==0 ) return 0;
9249 aHeap
[1] = aHeap
[x
];
9250 aHeap
[x
] = 0xffffffff;
9253 while( (j
= i
*2)<=aHeap
[0] ){
9254 if( aHeap
[j
]>aHeap
[j
+1] ) j
++;
9255 if( aHeap
[i
]<aHeap
[j
] ) break;
9257 aHeap
[i
] = aHeap
[j
];
9264 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9266 ** Do various sanity checks on a single page of a tree. Return
9267 ** the tree depth. Root pages return 0. Parents of root pages
9268 ** return 1, and so forth.
9270 ** These checks are done:
9272 ** 1. Make sure that cells and freeblocks do not overlap
9273 ** but combine to completely cover the page.
9274 ** 2. Make sure integer cell keys are in order.
9275 ** 3. Check the integrity of overflow pages.
9276 ** 4. Recursively call checkTreePage on all children.
9277 ** 5. Verify that the depth of all children is the same.
9279 static int checkTreePage(
9280 IntegrityCk
*pCheck
, /* Context for the sanity check */
9281 int iPage
, /* Page number of the page to check */
9282 i64
*piMinKey
, /* Write minimum integer primary key here */
9283 i64 maxKey
/* Error if integer primary key greater than this */
9285 MemPage
*pPage
= 0; /* The page being analyzed */
9286 int i
; /* Loop counter */
9287 int rc
; /* Result code from subroutine call */
9288 int depth
= -1, d2
; /* Depth of a subtree */
9289 int pgno
; /* Page number */
9290 int nFrag
; /* Number of fragmented bytes on the page */
9291 int hdr
; /* Offset to the page header */
9292 int cellStart
; /* Offset to the start of the cell pointer array */
9293 int nCell
; /* Number of cells */
9294 int doCoverageCheck
= 1; /* True if cell coverage checking should be done */
9295 int keyCanBeEqual
= 1; /* True if IPK can be equal to maxKey
9296 ** False if IPK must be strictly less than maxKey */
9297 u8
*data
; /* Page content */
9298 u8
*pCell
; /* Cell content */
9299 u8
*pCellIdx
; /* Next element of the cell pointer array */
9300 BtShared
*pBt
; /* The BtShared object that owns pPage */
9301 u32 pc
; /* Address of a cell */
9302 u32 usableSize
; /* Usable size of the page */
9303 u32 contentOffset
; /* Offset to the start of the cell content area */
9304 u32
*heap
= 0; /* Min-heap used for checking cell coverage */
9305 u32 x
, prev
= 0; /* Next and previous entry on the min-heap */
9306 const char *saved_zPfx
= pCheck
->zPfx
;
9307 int saved_v1
= pCheck
->v1
;
9308 int saved_v2
= pCheck
->v2
;
9311 /* Check that the page exists
9314 usableSize
= pBt
->usableSize
;
9315 if( iPage
==0 ) return 0;
9316 if( checkRef(pCheck
, iPage
) ) return 0;
9317 pCheck
->zPfx
= "Page %d: ";
9319 if( (rc
= btreeGetPage(pBt
, (Pgno
)iPage
, &pPage
, 0))!=0 ){
9320 checkAppendMsg(pCheck
,
9321 "unable to get the page. error code=%d", rc
);
9325 /* Clear MemPage.isInit to make sure the corruption detection code in
9326 ** btreeInitPage() is executed. */
9327 savedIsInit
= pPage
->isInit
;
9329 if( (rc
= btreeInitPage(pPage
))!=0 ){
9330 assert( rc
==SQLITE_CORRUPT
); /* The only possible error from InitPage */
9331 checkAppendMsg(pCheck
,
9332 "btreeInitPage() returns error code %d", rc
);
9335 data
= pPage
->aData
;
9336 hdr
= pPage
->hdrOffset
;
9338 /* Set up for cell analysis */
9339 pCheck
->zPfx
= "On tree page %d cell %d: ";
9340 contentOffset
= get2byteNotZero(&data
[hdr
+5]);
9341 assert( contentOffset
<=usableSize
); /* Enforced by btreeInitPage() */
9343 /* EVIDENCE-OF: R-37002-32774 The two-byte integer at offset 3 gives the
9344 ** number of cells on the page. */
9345 nCell
= get2byte(&data
[hdr
+3]);
9346 assert( pPage
->nCell
==nCell
);
9348 /* EVIDENCE-OF: R-23882-45353 The cell pointer array of a b-tree page
9349 ** immediately follows the b-tree page header. */
9350 cellStart
= hdr
+ 12 - 4*pPage
->leaf
;
9351 assert( pPage
->aCellIdx
==&data
[cellStart
] );
9352 pCellIdx
= &data
[cellStart
+ 2*(nCell
-1)];
9355 /* Analyze the right-child page of internal pages */
9356 pgno
= get4byte(&data
[hdr
+8]);
9357 #ifndef SQLITE_OMIT_AUTOVACUUM
9358 if( pBt
->autoVacuum
){
9359 pCheck
->zPfx
= "On page %d at right child: ";
9360 checkPtrmap(pCheck
, pgno
, PTRMAP_BTREE
, iPage
);
9363 depth
= checkTreePage(pCheck
, pgno
, &maxKey
, maxKey
);
9366 /* For leaf pages, the coverage check will occur in the same loop
9367 ** as the other cell checks, so initialize the heap. */
9368 heap
= pCheck
->heap
;
9372 /* EVIDENCE-OF: R-02776-14802 The cell pointer array consists of K 2-byte
9373 ** integer offsets to the cell contents. */
9374 for(i
=nCell
-1; i
>=0 && pCheck
->mxErr
; i
--){
9377 /* Check cell size */
9379 assert( pCellIdx
==&data
[cellStart
+ i
*2] );
9380 pc
= get2byteAligned(pCellIdx
);
9382 if( pc
<contentOffset
|| pc
>usableSize
-4 ){
9383 checkAppendMsg(pCheck
, "Offset %d out of range %d..%d",
9384 pc
, contentOffset
, usableSize
-4);
9385 doCoverageCheck
= 0;
9389 pPage
->xParseCell(pPage
, pCell
, &info
);
9390 if( pc
+info
.nSize
>usableSize
){
9391 checkAppendMsg(pCheck
, "Extends off end of page");
9392 doCoverageCheck
= 0;
9396 /* Check for integer primary key out of range */
9397 if( pPage
->intKey
){
9398 if( keyCanBeEqual
? (info
.nKey
> maxKey
) : (info
.nKey
>= maxKey
) ){
9399 checkAppendMsg(pCheck
, "Rowid %lld out of order", info
.nKey
);
9402 keyCanBeEqual
= 0; /* Only the first key on the page may ==maxKey */
9405 /* Check the content overflow list */
9406 if( info
.nPayload
>info
.nLocal
){
9407 int nPage
; /* Number of pages on the overflow chain */
9408 Pgno pgnoOvfl
; /* First page of the overflow chain */
9409 assert( pc
+ info
.nSize
- 4 <= usableSize
);
9410 nPage
= (info
.nPayload
- info
.nLocal
+ usableSize
- 5)/(usableSize
- 4);
9411 pgnoOvfl
= get4byte(&pCell
[info
.nSize
- 4]);
9412 #ifndef SQLITE_OMIT_AUTOVACUUM
9413 if( pBt
->autoVacuum
){
9414 checkPtrmap(pCheck
, pgnoOvfl
, PTRMAP_OVERFLOW1
, iPage
);
9417 checkList(pCheck
, 0, pgnoOvfl
, nPage
);
9421 /* Check sanity of left child page for internal pages */
9422 pgno
= get4byte(pCell
);
9423 #ifndef SQLITE_OMIT_AUTOVACUUM
9424 if( pBt
->autoVacuum
){
9425 checkPtrmap(pCheck
, pgno
, PTRMAP_BTREE
, iPage
);
9428 d2
= checkTreePage(pCheck
, pgno
, &maxKey
, maxKey
);
9431 checkAppendMsg(pCheck
, "Child page depth differs");
9435 /* Populate the coverage-checking heap for leaf pages */
9436 btreeHeapInsert(heap
, (pc
<<16)|(pc
+info
.nSize
-1));
9441 /* Check for complete coverage of the page
9444 if( doCoverageCheck
&& pCheck
->mxErr
>0 ){
9445 /* For leaf pages, the min-heap has already been initialized and the
9446 ** cells have already been inserted. But for internal pages, that has
9447 ** not yet been done, so do it now */
9449 heap
= pCheck
->heap
;
9451 for(i
=nCell
-1; i
>=0; i
--){
9453 pc
= get2byteAligned(&data
[cellStart
+i
*2]);
9454 size
= pPage
->xCellSize(pPage
, &data
[pc
]);
9455 btreeHeapInsert(heap
, (pc
<<16)|(pc
+size
-1));
9458 /* Add the freeblocks to the min-heap
9460 ** EVIDENCE-OF: R-20690-50594 The second field of the b-tree page header
9461 ** is the offset of the first freeblock, or zero if there are no
9462 ** freeblocks on the page.
9464 i
= get2byte(&data
[hdr
+1]);
9467 assert( (u32
)i
<=usableSize
-4 ); /* Enforced by btreeInitPage() */
9468 size
= get2byte(&data
[i
+2]);
9469 assert( (u32
)(i
+size
)<=usableSize
); /* Enforced by btreeInitPage() */
9470 btreeHeapInsert(heap
, (((u32
)i
)<<16)|(i
+size
-1));
9471 /* EVIDENCE-OF: R-58208-19414 The first 2 bytes of a freeblock are a
9472 ** big-endian integer which is the offset in the b-tree page of the next
9473 ** freeblock in the chain, or zero if the freeblock is the last on the
9475 j
= get2byte(&data
[i
]);
9476 /* EVIDENCE-OF: R-06866-39125 Freeblocks are always connected in order of
9477 ** increasing offset. */
9478 assert( j
==0 || j
>i
+size
); /* Enforced by btreeInitPage() */
9479 assert( (u32
)j
<=usableSize
-4 ); /* Enforced by btreeInitPage() */
9482 /* Analyze the min-heap looking for overlap between cells and/or
9483 ** freeblocks, and counting the number of untracked bytes in nFrag.
9485 ** Each min-heap entry is of the form: (start_address<<16)|end_address.
9486 ** There is an implied first entry the covers the page header, the cell
9487 ** pointer index, and the gap between the cell pointer index and the start
9490 ** The loop below pulls entries from the min-heap in order and compares
9491 ** the start_address against the previous end_address. If there is an
9492 ** overlap, that means bytes are used multiple times. If there is a gap,
9493 ** that gap is added to the fragmentation count.
9496 prev
= contentOffset
- 1; /* Implied first min-heap entry */
9497 while( btreeHeapPull(heap
,&x
) ){
9498 if( (prev
&0xffff)>=(x
>>16) ){
9499 checkAppendMsg(pCheck
,
9500 "Multiple uses for byte %u of page %d", x
>>16, iPage
);
9503 nFrag
+= (x
>>16) - (prev
&0xffff) - 1;
9507 nFrag
+= usableSize
- (prev
&0xffff) - 1;
9508 /* EVIDENCE-OF: R-43263-13491 The total number of bytes in all fragments
9509 ** is stored in the fifth field of the b-tree page header.
9510 ** EVIDENCE-OF: R-07161-27322 The one-byte integer at offset 7 gives the
9511 ** number of fragmented free bytes within the cell content area.
9513 if( heap
[0]==0 && nFrag
!=data
[hdr
+7] ){
9514 checkAppendMsg(pCheck
,
9515 "Fragmentation of %d bytes reported as %d on page %d",
9516 nFrag
, data
[hdr
+7], iPage
);
9521 if( !doCoverageCheck
) pPage
->isInit
= savedIsInit
;
9523 pCheck
->zPfx
= saved_zPfx
;
9524 pCheck
->v1
= saved_v1
;
9525 pCheck
->v2
= saved_v2
;
9528 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9530 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9532 ** This routine does a complete check of the given BTree file. aRoot[] is
9533 ** an array of pages numbers were each page number is the root page of
9534 ** a table. nRoot is the number of entries in aRoot.
9536 ** A read-only or read-write transaction must be opened before calling
9539 ** Write the number of error seen in *pnErr. Except for some memory
9540 ** allocation errors, an error message held in memory obtained from
9541 ** malloc is returned if *pnErr is non-zero. If *pnErr==0 then NULL is
9542 ** returned. If a memory allocation error occurs, NULL is returned.
9544 char *sqlite3BtreeIntegrityCheck(
9545 Btree
*p
, /* The btree to be checked */
9546 int *aRoot
, /* An array of root pages numbers for individual trees */
9547 int nRoot
, /* Number of entries in aRoot[] */
9548 int mxErr
, /* Stop reporting errors after this many */
9549 int *pnErr
/* Write number of errors seen to this variable */
9553 BtShared
*pBt
= p
->pBt
;
9554 int savedDbFlags
= pBt
->db
->flags
;
9556 VVA_ONLY( int nRef
);
9558 sqlite3BtreeEnter(p
);
9559 assert( p
->inTrans
>TRANS_NONE
&& pBt
->inTransaction
>TRANS_NONE
);
9560 VVA_ONLY( nRef
= sqlite3PagerRefcount(pBt
->pPager
) );
9563 sCheck
.pPager
= pBt
->pPager
;
9564 sCheck
.nPage
= btreePagecount(sCheck
.pBt
);
9565 sCheck
.mxErr
= mxErr
;
9567 sCheck
.mallocFailed
= 0;
9573 sqlite3StrAccumInit(&sCheck
.errMsg
, 0, zErr
, sizeof(zErr
), SQLITE_MAX_LENGTH
);
9574 sCheck
.errMsg
.printfFlags
= SQLITE_PRINTF_INTERNAL
;
9575 if( sCheck
.nPage
==0 ){
9576 goto integrity_ck_cleanup
;
9579 sCheck
.aPgRef
= sqlite3MallocZero((sCheck
.nPage
/ 8)+ 1);
9580 if( !sCheck
.aPgRef
){
9581 sCheck
.mallocFailed
= 1;
9582 goto integrity_ck_cleanup
;
9584 sCheck
.heap
= (u32
*)sqlite3PageMalloc( pBt
->pageSize
);
9585 if( sCheck
.heap
==0 ){
9586 sCheck
.mallocFailed
= 1;
9587 goto integrity_ck_cleanup
;
9590 i
= PENDING_BYTE_PAGE(pBt
);
9591 if( i
<=sCheck
.nPage
) setPageReferenced(&sCheck
, i
);
9593 /* Check the integrity of the freelist
9595 sCheck
.zPfx
= "Main freelist: ";
9596 checkList(&sCheck
, 1, get4byte(&pBt
->pPage1
->aData
[32]),
9597 get4byte(&pBt
->pPage1
->aData
[36]));
9600 /* Check all the tables.
9602 testcase( pBt
->db
->flags
& SQLITE_CellSizeCk
);
9603 pBt
->db
->flags
&= ~SQLITE_CellSizeCk
;
9604 for(i
=0; (int)i
<nRoot
&& sCheck
.mxErr
; i
++){
9606 if( aRoot
[i
]==0 ) continue;
9607 #ifndef SQLITE_OMIT_AUTOVACUUM
9608 if( pBt
->autoVacuum
&& aRoot
[i
]>1 ){
9609 checkPtrmap(&sCheck
, aRoot
[i
], PTRMAP_ROOTPAGE
, 0);
9612 checkTreePage(&sCheck
, aRoot
[i
], ¬Used
, LARGEST_INT64
);
9614 pBt
->db
->flags
= savedDbFlags
;
9616 /* Make sure every page in the file is referenced
9618 for(i
=1; i
<=sCheck
.nPage
&& sCheck
.mxErr
; i
++){
9619 #ifdef SQLITE_OMIT_AUTOVACUUM
9620 if( getPageReferenced(&sCheck
, i
)==0 ){
9621 checkAppendMsg(&sCheck
, "Page %d is never used", i
);
9624 /* If the database supports auto-vacuum, make sure no tables contain
9625 ** references to pointer-map pages.
9627 if( getPageReferenced(&sCheck
, i
)==0 &&
9628 (PTRMAP_PAGENO(pBt
, i
)!=i
|| !pBt
->autoVacuum
) ){
9629 checkAppendMsg(&sCheck
, "Page %d is never used", i
);
9631 if( getPageReferenced(&sCheck
, i
)!=0 &&
9632 (PTRMAP_PAGENO(pBt
, i
)==i
&& pBt
->autoVacuum
) ){
9633 checkAppendMsg(&sCheck
, "Pointer map page %d is referenced", i
);
9638 /* Clean up and report errors.
9640 integrity_ck_cleanup
:
9641 sqlite3PageFree(sCheck
.heap
);
9642 sqlite3_free(sCheck
.aPgRef
);
9643 if( sCheck
.mallocFailed
){
9644 sqlite3StrAccumReset(&sCheck
.errMsg
);
9647 *pnErr
= sCheck
.nErr
;
9648 if( sCheck
.nErr
==0 ) sqlite3StrAccumReset(&sCheck
.errMsg
);
9649 /* Make sure this analysis did not leave any unref() pages. */
9650 assert( nRef
==sqlite3PagerRefcount(pBt
->pPager
) );
9651 sqlite3BtreeLeave(p
);
9652 return sqlite3StrAccumFinish(&sCheck
.errMsg
);
9654 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9657 ** Return the full pathname of the underlying database file. Return
9658 ** an empty string if the database is in-memory or a TEMP database.
9660 ** The pager filename is invariant as long as the pager is
9661 ** open so it is safe to access without the BtShared mutex.
9663 const char *sqlite3BtreeGetFilename(Btree
*p
){
9664 assert( p
->pBt
->pPager
!=0 );
9665 return sqlite3PagerFilename(p
->pBt
->pPager
, 1);
9669 ** Return the pathname of the journal file for this database. The return
9670 ** value of this routine is the same regardless of whether the journal file
9671 ** has been created or not.
9673 ** The pager journal filename is invariant as long as the pager is
9674 ** open so it is safe to access without the BtShared mutex.
9676 const char *sqlite3BtreeGetJournalname(Btree
*p
){
9677 assert( p
->pBt
->pPager
!=0 );
9678 return sqlite3PagerJournalname(p
->pBt
->pPager
);
9682 ** Return non-zero if a transaction is active.
9684 int sqlite3BtreeIsInTrans(Btree
*p
){
9685 assert( p
==0 || sqlite3_mutex_held(p
->db
->mutex
) );
9686 return (p
&& (p
->inTrans
==TRANS_WRITE
));
9689 #ifndef SQLITE_OMIT_WAL
9691 ** Run a checkpoint on the Btree passed as the first argument.
9693 ** Return SQLITE_LOCKED if this or any other connection has an open
9694 ** transaction on the shared-cache the argument Btree is connected to.
9696 ** Parameter eMode is one of SQLITE_CHECKPOINT_PASSIVE, FULL or RESTART.
9698 int sqlite3BtreeCheckpoint(Btree
*p
, int eMode
, int *pnLog
, int *pnCkpt
){
9701 BtShared
*pBt
= p
->pBt
;
9702 sqlite3BtreeEnter(p
);
9703 if( pBt
->inTransaction
!=TRANS_NONE
){
9706 rc
= sqlite3PagerCheckpoint(pBt
->pPager
, p
->db
, eMode
, pnLog
, pnCkpt
);
9708 sqlite3BtreeLeave(p
);
9715 ** Return non-zero if a read (or write) transaction is active.
9717 int sqlite3BtreeIsInReadTrans(Btree
*p
){
9719 assert( sqlite3_mutex_held(p
->db
->mutex
) );
9720 return p
->inTrans
!=TRANS_NONE
;
9723 int sqlite3BtreeIsInBackup(Btree
*p
){
9725 assert( sqlite3_mutex_held(p
->db
->mutex
) );
9726 return p
->nBackup
!=0;
9730 ** This function returns a pointer to a blob of memory associated with
9731 ** a single shared-btree. The memory is used by client code for its own
9732 ** purposes (for example, to store a high-level schema associated with
9733 ** the shared-btree). The btree layer manages reference counting issues.
9735 ** The first time this is called on a shared-btree, nBytes bytes of memory
9736 ** are allocated, zeroed, and returned to the caller. For each subsequent
9737 ** call the nBytes parameter is ignored and a pointer to the same blob
9738 ** of memory returned.
9740 ** If the nBytes parameter is 0 and the blob of memory has not yet been
9741 ** allocated, a null pointer is returned. If the blob has already been
9742 ** allocated, it is returned as normal.
9744 ** Just before the shared-btree is closed, the function passed as the
9745 ** xFree argument when the memory allocation was made is invoked on the
9746 ** blob of allocated memory. The xFree function should not call sqlite3_free()
9747 ** on the memory, the btree layer does that.
9749 void *sqlite3BtreeSchema(Btree
*p
, int nBytes
, void(*xFree
)(void *)){
9750 BtShared
*pBt
= p
->pBt
;
9751 sqlite3BtreeEnter(p
);
9752 if( !pBt
->pSchema
&& nBytes
){
9753 pBt
->pSchema
= sqlite3DbMallocZero(0, nBytes
);
9754 pBt
->xFreeSchema
= xFree
;
9756 sqlite3BtreeLeave(p
);
9757 return pBt
->pSchema
;
9761 ** Return SQLITE_LOCKED_SHAREDCACHE if another user of the same shared
9762 ** btree as the argument handle holds an exclusive lock on the
9763 ** sqlite_master table. Otherwise SQLITE_OK.
9765 int sqlite3BtreeSchemaLocked(Btree
*p
){
9767 assert( sqlite3_mutex_held(p
->db
->mutex
) );
9768 sqlite3BtreeEnter(p
);
9769 rc
= querySharedCacheTableLock(p
, MASTER_ROOT
, READ_LOCK
);
9770 assert( rc
==SQLITE_OK
|| rc
==SQLITE_LOCKED_SHAREDCACHE
);
9771 sqlite3BtreeLeave(p
);
9776 #ifndef SQLITE_OMIT_SHARED_CACHE
9778 ** Obtain a lock on the table whose root page is iTab. The
9779 ** lock is a write lock if isWritelock is true or a read lock
9782 int sqlite3BtreeLockTable(Btree
*p
, int iTab
, u8 isWriteLock
){
9784 assert( p
->inTrans
!=TRANS_NONE
);
9786 u8 lockType
= READ_LOCK
+ isWriteLock
;
9787 assert( READ_LOCK
+1==WRITE_LOCK
);
9788 assert( isWriteLock
==0 || isWriteLock
==1 );
9790 sqlite3BtreeEnter(p
);
9791 rc
= querySharedCacheTableLock(p
, iTab
, lockType
);
9792 if( rc
==SQLITE_OK
){
9793 rc
= setSharedCacheTableLock(p
, iTab
, lockType
);
9795 sqlite3BtreeLeave(p
);
9801 #ifndef SQLITE_OMIT_INCRBLOB
9803 ** Argument pCsr must be a cursor opened for writing on an
9804 ** INTKEY table currently pointing at a valid table entry.
9805 ** This function modifies the data stored as part of that entry.
9807 ** Only the data content may only be modified, it is not possible to
9808 ** change the length of the data stored. If this function is called with
9809 ** parameters that attempt to write past the end of the existing data,
9810 ** no modifications are made and SQLITE_CORRUPT is returned.
9812 int sqlite3BtreePutData(BtCursor
*pCsr
, u32 offset
, u32 amt
, void *z
){
9814 assert( cursorOwnsBtShared(pCsr
) );
9815 assert( sqlite3_mutex_held(pCsr
->pBtree
->db
->mutex
) );
9816 assert( pCsr
->curFlags
& BTCF_Incrblob
);
9818 rc
= restoreCursorPosition(pCsr
);
9819 if( rc
!=SQLITE_OK
){
9822 assert( pCsr
->eState
!=CURSOR_REQUIRESEEK
);
9823 if( pCsr
->eState
!=CURSOR_VALID
){
9824 return SQLITE_ABORT
;
9827 /* Save the positions of all other cursors open on this table. This is
9828 ** required in case any of them are holding references to an xFetch
9829 ** version of the b-tree page modified by the accessPayload call below.
9831 ** Note that pCsr must be open on a INTKEY table and saveCursorPosition()
9832 ** and hence saveAllCursors() cannot fail on a BTREE_INTKEY table, hence
9833 ** saveAllCursors can only return SQLITE_OK.
9835 VVA_ONLY(rc
=) saveAllCursors(pCsr
->pBt
, pCsr
->pgnoRoot
, pCsr
);
9836 assert( rc
==SQLITE_OK
);
9838 /* Check some assumptions:
9839 ** (a) the cursor is open for writing,
9840 ** (b) there is a read/write transaction open,
9841 ** (c) the connection holds a write-lock on the table (if required),
9842 ** (d) there are no conflicting read-locks, and
9843 ** (e) the cursor points at a valid row of an intKey table.
9845 if( (pCsr
->curFlags
& BTCF_WriteFlag
)==0 ){
9846 return SQLITE_READONLY
;
9848 assert( (pCsr
->pBt
->btsFlags
& BTS_READ_ONLY
)==0
9849 && pCsr
->pBt
->inTransaction
==TRANS_WRITE
);
9850 assert( hasSharedCacheTableLock(pCsr
->pBtree
, pCsr
->pgnoRoot
, 0, 2) );
9851 assert( !hasReadConflicts(pCsr
->pBtree
, pCsr
->pgnoRoot
) );
9852 assert( pCsr
->pPage
->intKey
);
9854 return accessPayload(pCsr
, offset
, amt
, (unsigned char *)z
, 1);
9858 ** Mark this cursor as an incremental blob cursor.
9860 void sqlite3BtreeIncrblobCursor(BtCursor
*pCur
){
9861 pCur
->curFlags
|= BTCF_Incrblob
;
9862 pCur
->pBtree
->hasIncrblobCur
= 1;
9867 ** Set both the "read version" (single byte at byte offset 18) and
9868 ** "write version" (single byte at byte offset 19) fields in the database
9869 ** header to iVersion.
9871 int sqlite3BtreeSetVersion(Btree
*pBtree
, int iVersion
){
9872 BtShared
*pBt
= pBtree
->pBt
;
9873 int rc
; /* Return code */
9875 assert( iVersion
==1 || iVersion
==2 );
9877 /* If setting the version fields to 1, do not automatically open the
9878 ** WAL connection, even if the version fields are currently set to 2.
9880 pBt
->btsFlags
&= ~BTS_NO_WAL
;
9881 if( iVersion
==1 ) pBt
->btsFlags
|= BTS_NO_WAL
;
9883 rc
= sqlite3BtreeBeginTrans(pBtree
, 0);
9884 if( rc
==SQLITE_OK
){
9885 u8
*aData
= pBt
->pPage1
->aData
;
9886 if( aData
[18]!=(u8
)iVersion
|| aData
[19]!=(u8
)iVersion
){
9887 rc
= sqlite3BtreeBeginTrans(pBtree
, 2);
9888 if( rc
==SQLITE_OK
){
9889 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
9890 if( rc
==SQLITE_OK
){
9891 aData
[18] = (u8
)iVersion
;
9892 aData
[19] = (u8
)iVersion
;
9898 pBt
->btsFlags
&= ~BTS_NO_WAL
;
9903 ** Return true if the cursor has a hint specified. This routine is
9904 ** only used from within assert() statements
9906 int sqlite3BtreeCursorHasHint(BtCursor
*pCsr
, unsigned int mask
){
9907 return (pCsr
->hints
& mask
)!=0;
9911 ** Return true if the given Btree is read-only.
9913 int sqlite3BtreeIsReadonly(Btree
*p
){
9914 return (p
->pBt
->btsFlags
& BTS_READ_ONLY
)!=0;
9918 ** Return the size of the header added to each page by this module.
9920 int sqlite3HeaderSizeBtree(void){ return ROUND8(sizeof(MemPage
)); }
9922 #if !defined(SQLITE_OMIT_SHARED_CACHE)
9924 ** Return true if the Btree passed as the only argument is sharable.
9926 int sqlite3BtreeSharable(Btree
*p
){
9931 ** Return the number of connections to the BtShared object accessed by
9932 ** the Btree handle passed as the only argument. For private caches
9933 ** this is always 1. For shared caches it may be 1 or greater.
9935 int sqlite3BtreeConnectionCount(Btree
*p
){
9936 testcase( p
->sharable
);
9937 return p
->pBt
->nRef
;