4 ** The author disclaims copyright to this source code. In place of
5 ** a legal notice, here is a blessing:
7 ** May you do good and not evil.
8 ** May you find forgiveness for yourself and forgive others.
9 ** May you share freely, never taking more than you give.
11 *************************************************************************
12 ** This file implements an external (disk-based) database using BTrees.
13 ** See the header comment on "btreeInt.h" for additional information.
14 ** Including a description of file format and an overview of operation.
19 ** The header string that appears at the beginning of every
22 static const char zMagicHeader
[] = SQLITE_FILE_HEADER
;
25 ** Set this global variable to 1 to enable tracing using the TRACE
29 int sqlite3BtreeTrace
=1; /* True to enable tracing */
30 # define TRACE(X) if(sqlite3BtreeTrace){printf X;fflush(stdout);}
36 ** Extract a 2-byte big-endian integer from an array of unsigned bytes.
37 ** But if the value is zero, make it 65536.
39 ** This routine is used to extract the "offset to cell content area" value
40 ** from the header of a btree page. If the page size is 65536 and the page
41 ** is empty, the offset should be 65536, but the 2-byte value stores zero.
42 ** This routine makes the necessary adjustment to 65536.
44 #define get2byteNotZero(X) (((((int)get2byte(X))-1)&0xffff)+1)
47 ** Values passed as the 5th argument to allocateBtreePage()
49 #define BTALLOC_ANY 0 /* Allocate any page */
50 #define BTALLOC_EXACT 1 /* Allocate exact page if possible */
51 #define BTALLOC_LE 2 /* Allocate any page <= the parameter */
54 ** Macro IfNotOmitAV(x) returns (x) if SQLITE_OMIT_AUTOVACUUM is not
55 ** defined, or 0 if it is. For example:
57 ** bIncrVacuum = IfNotOmitAV(pBtShared->incrVacuum);
59 #ifndef SQLITE_OMIT_AUTOVACUUM
60 #define IfNotOmitAV(expr) (expr)
62 #define IfNotOmitAV(expr) 0
65 #ifndef SQLITE_OMIT_SHARED_CACHE
67 ** A list of BtShared objects that are eligible for participation
68 ** in shared cache. This variable has file scope during normal builds,
69 ** but the test harness needs to access it so we make it global for
72 ** Access to this variable is protected by SQLITE_MUTEX_STATIC_MASTER.
75 BtShared
*SQLITE_WSD sqlite3SharedCacheList
= 0;
77 static BtShared
*SQLITE_WSD sqlite3SharedCacheList
= 0;
79 #endif /* SQLITE_OMIT_SHARED_CACHE */
81 #ifndef SQLITE_OMIT_SHARED_CACHE
83 ** Enable or disable the shared pager and schema features.
85 ** This routine has no effect on existing database connections.
86 ** The shared cache setting effects only future calls to
87 ** sqlite3_open(), sqlite3_open16(), or sqlite3_open_v2().
89 int sqlite3_enable_shared_cache(int enable
){
90 sqlite3GlobalConfig
.sharedCacheEnabled
= enable
;
97 #ifdef SQLITE_OMIT_SHARED_CACHE
99 ** The functions querySharedCacheTableLock(), setSharedCacheTableLock(),
100 ** and clearAllSharedCacheTableLocks()
101 ** manipulate entries in the BtShared.pLock linked list used to store
102 ** shared-cache table level locks. If the library is compiled with the
103 ** shared-cache feature disabled, then there is only ever one user
104 ** of each BtShared structure and so this locking is not necessary.
105 ** So define the lock related functions as no-ops.
107 #define querySharedCacheTableLock(a,b,c) SQLITE_OK
108 #define setSharedCacheTableLock(a,b,c) SQLITE_OK
109 #define clearAllSharedCacheTableLocks(a)
110 #define downgradeAllSharedCacheTableLocks(a)
111 #define hasSharedCacheTableLock(a,b,c,d) 1
112 #define hasReadConflicts(a, b) 0
116 ** Implementation of the SQLITE_CORRUPT_PAGE() macro. Takes a single
117 ** (MemPage*) as an argument. The (MemPage*) must not be NULL.
119 ** If SQLITE_DEBUG is not defined, then this macro is equivalent to
120 ** SQLITE_CORRUPT_BKPT. Or, if SQLITE_DEBUG is set, then the log message
121 ** normally produced as a side-effect of SQLITE_CORRUPT_BKPT is augmented
122 ** with the page number and filename associated with the (MemPage*).
125 int corruptPageError(int lineno
, MemPage
*p
){
126 char *zMsg
= sqlite3_mprintf("database corruption page %d of %s",
127 (int)p
->pgno
, sqlite3PagerFilename(p
->pBt
->pPager
, 0)
130 sqlite3ReportError(SQLITE_CORRUPT
, lineno
, zMsg
);
133 return SQLITE_CORRUPT_BKPT
;
135 # define SQLITE_CORRUPT_PAGE(pMemPage) corruptPageError(__LINE__, pMemPage)
137 # define SQLITE_CORRUPT_PAGE(pMemPage) SQLITE_CORRUPT_PGNO(pMemPage->pgno)
140 #ifndef SQLITE_OMIT_SHARED_CACHE
144 **** This function is only used as part of an assert() statement. ***
146 ** Check to see if pBtree holds the required locks to read or write to the
147 ** table with root page iRoot. Return 1 if it does and 0 if not.
149 ** For example, when writing to a table with root-page iRoot via
150 ** Btree connection pBtree:
152 ** assert( hasSharedCacheTableLock(pBtree, iRoot, 0, WRITE_LOCK) );
154 ** When writing to an index that resides in a sharable database, the
155 ** caller should have first obtained a lock specifying the root page of
156 ** the corresponding table. This makes things a bit more complicated,
157 ** as this module treats each table as a separate structure. To determine
158 ** the table corresponding to the index being written, this
159 ** function has to search through the database schema.
161 ** Instead of a lock on the table/index rooted at page iRoot, the caller may
162 ** hold a write-lock on the schema table (root page 1). This is also
165 static int hasSharedCacheTableLock(
166 Btree
*pBtree
, /* Handle that must hold lock */
167 Pgno iRoot
, /* Root page of b-tree */
168 int isIndex
, /* True if iRoot is the root of an index b-tree */
169 int eLockType
/* Required lock type (READ_LOCK or WRITE_LOCK) */
171 Schema
*pSchema
= (Schema
*)pBtree
->pBt
->pSchema
;
175 /* If this database is not shareable, or if the client is reading
176 ** and has the read-uncommitted flag set, then no lock is required.
177 ** Return true immediately.
179 if( (pBtree
->sharable
==0)
180 || (eLockType
==READ_LOCK
&& (pBtree
->db
->flags
& SQLITE_ReadUncommit
))
185 /* If the client is reading or writing an index and the schema is
186 ** not loaded, then it is too difficult to actually check to see if
187 ** the correct locks are held. So do not bother - just return true.
188 ** This case does not come up very often anyhow.
190 if( isIndex
&& (!pSchema
|| (pSchema
->schemaFlags
&DB_SchemaLoaded
)==0) ){
194 /* Figure out the root-page that the lock should be held on. For table
195 ** b-trees, this is just the root page of the b-tree being read or
196 ** written. For index b-trees, it is the root page of the associated
200 for(p
=sqliteHashFirst(&pSchema
->idxHash
); p
; p
=sqliteHashNext(p
)){
201 Index
*pIdx
= (Index
*)sqliteHashData(p
);
202 if( pIdx
->tnum
==(int)iRoot
){
204 /* Two or more indexes share the same root page. There must
205 ** be imposter tables. So just return true. The assert is not
206 ** useful in that case. */
209 iTab
= pIdx
->pTable
->tnum
;
216 /* Search for the required lock. Either a write-lock on root-page iTab, a
217 ** write-lock on the schema table, or (if the client is reading) a
218 ** read-lock on iTab will suffice. Return 1 if any of these are found. */
219 for(pLock
=pBtree
->pBt
->pLock
; pLock
; pLock
=pLock
->pNext
){
220 if( pLock
->pBtree
==pBtree
221 && (pLock
->iTable
==iTab
|| (pLock
->eLock
==WRITE_LOCK
&& pLock
->iTable
==1))
222 && pLock
->eLock
>=eLockType
228 /* Failed to find the required lock. */
231 #endif /* SQLITE_DEBUG */
235 **** This function may be used as part of assert() statements only. ****
237 ** Return true if it would be illegal for pBtree to write into the
238 ** table or index rooted at iRoot because other shared connections are
239 ** simultaneously reading that same table or index.
241 ** It is illegal for pBtree to write if some other Btree object that
242 ** shares the same BtShared object is currently reading or writing
243 ** the iRoot table. Except, if the other Btree object has the
244 ** read-uncommitted flag set, then it is OK for the other object to
245 ** have a read cursor.
247 ** For example, before writing to any part of the table or index
248 ** rooted at page iRoot, one should call:
250 ** assert( !hasReadConflicts(pBtree, iRoot) );
252 static int hasReadConflicts(Btree
*pBtree
, Pgno iRoot
){
254 for(p
=pBtree
->pBt
->pCursor
; p
; p
=p
->pNext
){
255 if( p
->pgnoRoot
==iRoot
257 && 0==(p
->pBtree
->db
->flags
& SQLITE_ReadUncommit
)
264 #endif /* #ifdef SQLITE_DEBUG */
267 ** Query to see if Btree handle p may obtain a lock of type eLock
268 ** (READ_LOCK or WRITE_LOCK) on the table with root-page iTab. Return
269 ** SQLITE_OK if the lock may be obtained (by calling
270 ** setSharedCacheTableLock()), or SQLITE_LOCKED if not.
272 static int querySharedCacheTableLock(Btree
*p
, Pgno iTab
, u8 eLock
){
273 BtShared
*pBt
= p
->pBt
;
276 assert( sqlite3BtreeHoldsMutex(p
) );
277 assert( eLock
==READ_LOCK
|| eLock
==WRITE_LOCK
);
279 assert( !(p
->db
->flags
&SQLITE_ReadUncommit
)||eLock
==WRITE_LOCK
||iTab
==1 );
281 /* If requesting a write-lock, then the Btree must have an open write
282 ** transaction on this file. And, obviously, for this to be so there
283 ** must be an open write transaction on the file itself.
285 assert( eLock
==READ_LOCK
|| (p
==pBt
->pWriter
&& p
->inTrans
==TRANS_WRITE
) );
286 assert( eLock
==READ_LOCK
|| pBt
->inTransaction
==TRANS_WRITE
);
288 /* This routine is a no-op if the shared-cache is not enabled */
293 /* If some other connection is holding an exclusive lock, the
294 ** requested lock may not be obtained.
296 if( pBt
->pWriter
!=p
&& (pBt
->btsFlags
& BTS_EXCLUSIVE
)!=0 ){
297 sqlite3ConnectionBlocked(p
->db
, pBt
->pWriter
->db
);
298 return SQLITE_LOCKED_SHAREDCACHE
;
301 for(pIter
=pBt
->pLock
; pIter
; pIter
=pIter
->pNext
){
302 /* The condition (pIter->eLock!=eLock) in the following if(...)
303 ** statement is a simplification of:
305 ** (eLock==WRITE_LOCK || pIter->eLock==WRITE_LOCK)
307 ** since we know that if eLock==WRITE_LOCK, then no other connection
308 ** may hold a WRITE_LOCK on any table in this file (since there can
309 ** only be a single writer).
311 assert( pIter
->eLock
==READ_LOCK
|| pIter
->eLock
==WRITE_LOCK
);
312 assert( eLock
==READ_LOCK
|| pIter
->pBtree
==p
|| pIter
->eLock
==READ_LOCK
);
313 if( pIter
->pBtree
!=p
&& pIter
->iTable
==iTab
&& pIter
->eLock
!=eLock
){
314 sqlite3ConnectionBlocked(p
->db
, pIter
->pBtree
->db
);
315 if( eLock
==WRITE_LOCK
){
316 assert( p
==pBt
->pWriter
);
317 pBt
->btsFlags
|= BTS_PENDING
;
319 return SQLITE_LOCKED_SHAREDCACHE
;
324 #endif /* !SQLITE_OMIT_SHARED_CACHE */
326 #ifndef SQLITE_OMIT_SHARED_CACHE
328 ** Add a lock on the table with root-page iTable to the shared-btree used
329 ** by Btree handle p. Parameter eLock must be either READ_LOCK or
332 ** This function assumes the following:
334 ** (a) The specified Btree object p is connected to a sharable
335 ** database (one with the BtShared.sharable flag set), and
337 ** (b) No other Btree objects hold a lock that conflicts
338 ** with the requested lock (i.e. querySharedCacheTableLock() has
339 ** already been called and returned SQLITE_OK).
341 ** SQLITE_OK is returned if the lock is added successfully. SQLITE_NOMEM
342 ** is returned if a malloc attempt fails.
344 static int setSharedCacheTableLock(Btree
*p
, Pgno iTable
, u8 eLock
){
345 BtShared
*pBt
= p
->pBt
;
349 assert( sqlite3BtreeHoldsMutex(p
) );
350 assert( eLock
==READ_LOCK
|| eLock
==WRITE_LOCK
);
353 /* A connection with the read-uncommitted flag set will never try to
354 ** obtain a read-lock using this function. The only read-lock obtained
355 ** by a connection in read-uncommitted mode is on the sqlite_master
356 ** table, and that lock is obtained in BtreeBeginTrans(). */
357 assert( 0==(p
->db
->flags
&SQLITE_ReadUncommit
) || eLock
==WRITE_LOCK
);
359 /* This function should only be called on a sharable b-tree after it
360 ** has been determined that no other b-tree holds a conflicting lock. */
361 assert( p
->sharable
);
362 assert( SQLITE_OK
==querySharedCacheTableLock(p
, iTable
, eLock
) );
364 /* First search the list for an existing lock on this table. */
365 for(pIter
=pBt
->pLock
; pIter
; pIter
=pIter
->pNext
){
366 if( pIter
->iTable
==iTable
&& pIter
->pBtree
==p
){
372 /* If the above search did not find a BtLock struct associating Btree p
373 ** with table iTable, allocate one and link it into the list.
376 pLock
= (BtLock
*)sqlite3MallocZero(sizeof(BtLock
));
378 return SQLITE_NOMEM_BKPT
;
380 pLock
->iTable
= iTable
;
382 pLock
->pNext
= pBt
->pLock
;
386 /* Set the BtLock.eLock variable to the maximum of the current lock
387 ** and the requested lock. This means if a write-lock was already held
388 ** and a read-lock requested, we don't incorrectly downgrade the lock.
390 assert( WRITE_LOCK
>READ_LOCK
);
391 if( eLock
>pLock
->eLock
){
392 pLock
->eLock
= eLock
;
397 #endif /* !SQLITE_OMIT_SHARED_CACHE */
399 #ifndef SQLITE_OMIT_SHARED_CACHE
401 ** Release all the table locks (locks obtained via calls to
402 ** the setSharedCacheTableLock() procedure) held by Btree object p.
404 ** This function assumes that Btree p has an open read or write
405 ** transaction. If it does not, then the BTS_PENDING flag
406 ** may be incorrectly cleared.
408 static void clearAllSharedCacheTableLocks(Btree
*p
){
409 BtShared
*pBt
= p
->pBt
;
410 BtLock
**ppIter
= &pBt
->pLock
;
412 assert( sqlite3BtreeHoldsMutex(p
) );
413 assert( p
->sharable
|| 0==*ppIter
);
414 assert( p
->inTrans
>0 );
417 BtLock
*pLock
= *ppIter
;
418 assert( (pBt
->btsFlags
& BTS_EXCLUSIVE
)==0 || pBt
->pWriter
==pLock
->pBtree
);
419 assert( pLock
->pBtree
->inTrans
>=pLock
->eLock
);
420 if( pLock
->pBtree
==p
){
421 *ppIter
= pLock
->pNext
;
422 assert( pLock
->iTable
!=1 || pLock
==&p
->lock
);
423 if( pLock
->iTable
!=1 ){
427 ppIter
= &pLock
->pNext
;
431 assert( (pBt
->btsFlags
& BTS_PENDING
)==0 || pBt
->pWriter
);
432 if( pBt
->pWriter
==p
){
434 pBt
->btsFlags
&= ~(BTS_EXCLUSIVE
|BTS_PENDING
);
435 }else if( pBt
->nTransaction
==2 ){
436 /* This function is called when Btree p is concluding its
437 ** transaction. If there currently exists a writer, and p is not
438 ** that writer, then the number of locks held by connections other
439 ** than the writer must be about to drop to zero. In this case
440 ** set the BTS_PENDING flag to 0.
442 ** If there is not currently a writer, then BTS_PENDING must
443 ** be zero already. So this next line is harmless in that case.
445 pBt
->btsFlags
&= ~BTS_PENDING
;
450 ** This function changes all write-locks held by Btree p into read-locks.
452 static void downgradeAllSharedCacheTableLocks(Btree
*p
){
453 BtShared
*pBt
= p
->pBt
;
454 if( pBt
->pWriter
==p
){
457 pBt
->btsFlags
&= ~(BTS_EXCLUSIVE
|BTS_PENDING
);
458 for(pLock
=pBt
->pLock
; pLock
; pLock
=pLock
->pNext
){
459 assert( pLock
->eLock
==READ_LOCK
|| pLock
->pBtree
==p
);
460 pLock
->eLock
= READ_LOCK
;
465 #endif /* SQLITE_OMIT_SHARED_CACHE */
467 static void releasePage(MemPage
*pPage
); /* Forward reference */
468 static void releasePageOne(MemPage
*pPage
); /* Forward reference */
469 static void releasePageNotNull(MemPage
*pPage
); /* Forward reference */
472 ***** This routine is used inside of assert() only ****
474 ** Verify that the cursor holds the mutex on its BtShared
477 static int cursorHoldsMutex(BtCursor
*p
){
478 return sqlite3_mutex_held(p
->pBt
->mutex
);
481 /* Verify that the cursor and the BtShared agree about what is the current
482 ** database connetion. This is important in shared-cache mode. If the database
483 ** connection pointers get out-of-sync, it is possible for routines like
484 ** btreeInitPage() to reference an stale connection pointer that references a
485 ** a connection that has already closed. This routine is used inside assert()
486 ** statements only and for the purpose of double-checking that the btree code
487 ** does keep the database connection pointers up-to-date.
489 static int cursorOwnsBtShared(BtCursor
*p
){
490 assert( cursorHoldsMutex(p
) );
491 return (p
->pBtree
->db
==p
->pBt
->db
);
496 ** Invalidate the overflow cache of the cursor passed as the first argument.
497 ** on the shared btree structure pBt.
499 #define invalidateOverflowCache(pCur) (pCur->curFlags &= ~BTCF_ValidOvfl)
502 ** Invalidate the overflow page-list cache for all cursors opened
503 ** on the shared btree structure pBt.
505 static void invalidateAllOverflowCache(BtShared
*pBt
){
507 assert( sqlite3_mutex_held(pBt
->mutex
) );
508 for(p
=pBt
->pCursor
; p
; p
=p
->pNext
){
509 invalidateOverflowCache(p
);
513 #ifndef SQLITE_OMIT_INCRBLOB
515 ** This function is called before modifying the contents of a table
516 ** to invalidate any incrblob cursors that are open on the
517 ** row or one of the rows being modified.
519 ** If argument isClearTable is true, then the entire contents of the
520 ** table is about to be deleted. In this case invalidate all incrblob
521 ** cursors open on any row within the table with root-page pgnoRoot.
523 ** Otherwise, if argument isClearTable is false, then the row with
524 ** rowid iRow is being replaced or deleted. In this case invalidate
525 ** only those incrblob cursors open on that specific row.
527 static void invalidateIncrblobCursors(
528 Btree
*pBtree
, /* The database file to check */
529 Pgno pgnoRoot
, /* The table that might be changing */
530 i64 iRow
, /* The rowid that might be changing */
531 int isClearTable
/* True if all rows are being deleted */
534 if( pBtree
->hasIncrblobCur
==0 ) return;
535 assert( sqlite3BtreeHoldsMutex(pBtree
) );
536 pBtree
->hasIncrblobCur
= 0;
537 for(p
=pBtree
->pBt
->pCursor
; p
; p
=p
->pNext
){
538 if( (p
->curFlags
& BTCF_Incrblob
)!=0 ){
539 pBtree
->hasIncrblobCur
= 1;
540 if( p
->pgnoRoot
==pgnoRoot
&& (isClearTable
|| p
->info
.nKey
==iRow
) ){
541 p
->eState
= CURSOR_INVALID
;
548 /* Stub function when INCRBLOB is omitted */
549 #define invalidateIncrblobCursors(w,x,y,z)
550 #endif /* SQLITE_OMIT_INCRBLOB */
553 ** Set bit pgno of the BtShared.pHasContent bitvec. This is called
554 ** when a page that previously contained data becomes a free-list leaf
557 ** The BtShared.pHasContent bitvec exists to work around an obscure
558 ** bug caused by the interaction of two useful IO optimizations surrounding
559 ** free-list leaf pages:
561 ** 1) When all data is deleted from a page and the page becomes
562 ** a free-list leaf page, the page is not written to the database
563 ** (as free-list leaf pages contain no meaningful data). Sometimes
564 ** such a page is not even journalled (as it will not be modified,
565 ** why bother journalling it?).
567 ** 2) When a free-list leaf page is reused, its content is not read
568 ** from the database or written to the journal file (why should it
569 ** be, if it is not at all meaningful?).
571 ** By themselves, these optimizations work fine and provide a handy
572 ** performance boost to bulk delete or insert operations. However, if
573 ** a page is moved to the free-list and then reused within the same
574 ** transaction, a problem comes up. If the page is not journalled when
575 ** it is moved to the free-list and it is also not journalled when it
576 ** is extracted from the free-list and reused, then the original data
577 ** may be lost. In the event of a rollback, it may not be possible
578 ** to restore the database to its original configuration.
580 ** The solution is the BtShared.pHasContent bitvec. Whenever a page is
581 ** moved to become a free-list leaf page, the corresponding bit is
582 ** set in the bitvec. Whenever a leaf page is extracted from the free-list,
583 ** optimization 2 above is omitted if the corresponding bit is already
584 ** set in BtShared.pHasContent. The contents of the bitvec are cleared
585 ** at the end of every transaction.
587 static int btreeSetHasContent(BtShared
*pBt
, Pgno pgno
){
589 if( !pBt
->pHasContent
){
590 assert( pgno
<=pBt
->nPage
);
591 pBt
->pHasContent
= sqlite3BitvecCreate(pBt
->nPage
);
592 if( !pBt
->pHasContent
){
593 rc
= SQLITE_NOMEM_BKPT
;
596 if( rc
==SQLITE_OK
&& pgno
<=sqlite3BitvecSize(pBt
->pHasContent
) ){
597 rc
= sqlite3BitvecSet(pBt
->pHasContent
, pgno
);
603 ** Query the BtShared.pHasContent vector.
605 ** This function is called when a free-list leaf page is removed from the
606 ** free-list for reuse. It returns false if it is safe to retrieve the
607 ** page from the pager layer with the 'no-content' flag set. True otherwise.
609 static int btreeGetHasContent(BtShared
*pBt
, Pgno pgno
){
610 Bitvec
*p
= pBt
->pHasContent
;
611 return (p
&& (pgno
>sqlite3BitvecSize(p
) || sqlite3BitvecTest(p
, pgno
)));
615 ** Clear (destroy) the BtShared.pHasContent bitvec. This should be
616 ** invoked at the conclusion of each write-transaction.
618 static void btreeClearHasContent(BtShared
*pBt
){
619 sqlite3BitvecDestroy(pBt
->pHasContent
);
620 pBt
->pHasContent
= 0;
624 ** Release all of the apPage[] pages for a cursor.
626 static void btreeReleaseAllCursorPages(BtCursor
*pCur
){
628 if( pCur
->iPage
>=0 ){
629 for(i
=0; i
<pCur
->iPage
; i
++){
630 releasePageNotNull(pCur
->apPage
[i
]);
632 releasePageNotNull(pCur
->pPage
);
638 ** The cursor passed as the only argument must point to a valid entry
639 ** when this function is called (i.e. have eState==CURSOR_VALID). This
640 ** function saves the current cursor key in variables pCur->nKey and
641 ** pCur->pKey. SQLITE_OK is returned if successful or an SQLite error
644 ** If the cursor is open on an intkey table, then the integer key
645 ** (the rowid) is stored in pCur->nKey and pCur->pKey is left set to
646 ** NULL. If the cursor is open on a non-intkey table, then pCur->pKey is
647 ** set to point to a malloced buffer pCur->nKey bytes in size containing
650 static int saveCursorKey(BtCursor
*pCur
){
652 assert( CURSOR_VALID
==pCur
->eState
);
653 assert( 0==pCur
->pKey
);
654 assert( cursorHoldsMutex(pCur
) );
656 if( pCur
->curIntKey
){
657 /* Only the rowid is required for a table btree */
658 pCur
->nKey
= sqlite3BtreeIntegerKey(pCur
);
660 /* For an index btree, save the complete key content */
662 pCur
->nKey
= sqlite3BtreePayloadSize(pCur
);
663 pKey
= sqlite3Malloc( pCur
->nKey
);
665 rc
= sqlite3BtreePayload(pCur
, 0, (int)pCur
->nKey
, pKey
);
672 rc
= SQLITE_NOMEM_BKPT
;
675 assert( !pCur
->curIntKey
|| !pCur
->pKey
);
680 ** Save the current cursor position in the variables BtCursor.nKey
681 ** and BtCursor.pKey. The cursor's state is set to CURSOR_REQUIRESEEK.
683 ** The caller must ensure that the cursor is valid (has eState==CURSOR_VALID)
684 ** prior to calling this routine.
686 static int saveCursorPosition(BtCursor
*pCur
){
689 assert( CURSOR_VALID
==pCur
->eState
|| CURSOR_SKIPNEXT
==pCur
->eState
);
690 assert( 0==pCur
->pKey
);
691 assert( cursorHoldsMutex(pCur
) );
693 if( pCur
->eState
==CURSOR_SKIPNEXT
){
694 pCur
->eState
= CURSOR_VALID
;
699 rc
= saveCursorKey(pCur
);
701 btreeReleaseAllCursorPages(pCur
);
702 pCur
->eState
= CURSOR_REQUIRESEEK
;
705 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
|BTCF_AtLast
);
709 /* Forward reference */
710 static int SQLITE_NOINLINE
saveCursorsOnList(BtCursor
*,Pgno
,BtCursor
*);
713 ** Save the positions of all cursors (except pExcept) that are open on
714 ** the table with root-page iRoot. "Saving the cursor position" means that
715 ** the location in the btree is remembered in such a way that it can be
716 ** moved back to the same spot after the btree has been modified. This
717 ** routine is called just before cursor pExcept is used to modify the
718 ** table, for example in BtreeDelete() or BtreeInsert().
720 ** If there are two or more cursors on the same btree, then all such
721 ** cursors should have their BTCF_Multiple flag set. The btreeCursor()
722 ** routine enforces that rule. This routine only needs to be called in
723 ** the uncommon case when pExpect has the BTCF_Multiple flag set.
725 ** If pExpect!=NULL and if no other cursors are found on the same root-page,
726 ** then the BTCF_Multiple flag on pExpect is cleared, to avoid another
727 ** pointless call to this routine.
729 ** Implementation note: This routine merely checks to see if any cursors
730 ** need to be saved. It calls out to saveCursorsOnList() in the (unusual)
731 ** event that cursors are in need to being saved.
733 static int saveAllCursors(BtShared
*pBt
, Pgno iRoot
, BtCursor
*pExcept
){
735 assert( sqlite3_mutex_held(pBt
->mutex
) );
736 assert( pExcept
==0 || pExcept
->pBt
==pBt
);
737 for(p
=pBt
->pCursor
; p
; p
=p
->pNext
){
738 if( p
!=pExcept
&& (0==iRoot
|| p
->pgnoRoot
==iRoot
) ) break;
740 if( p
) return saveCursorsOnList(p
, iRoot
, pExcept
);
741 if( pExcept
) pExcept
->curFlags
&= ~BTCF_Multiple
;
745 /* This helper routine to saveAllCursors does the actual work of saving
746 ** the cursors if and when a cursor is found that actually requires saving.
747 ** The common case is that no cursors need to be saved, so this routine is
748 ** broken out from its caller to avoid unnecessary stack pointer movement.
750 static int SQLITE_NOINLINE
saveCursorsOnList(
751 BtCursor
*p
, /* The first cursor that needs saving */
752 Pgno iRoot
, /* Only save cursor with this iRoot. Save all if zero */
753 BtCursor
*pExcept
/* Do not save this cursor */
756 if( p
!=pExcept
&& (0==iRoot
|| p
->pgnoRoot
==iRoot
) ){
757 if( p
->eState
==CURSOR_VALID
|| p
->eState
==CURSOR_SKIPNEXT
){
758 int rc
= saveCursorPosition(p
);
763 testcase( p
->iPage
>=0 );
764 btreeReleaseAllCursorPages(p
);
773 ** Clear the current cursor position.
775 void sqlite3BtreeClearCursor(BtCursor
*pCur
){
776 assert( cursorHoldsMutex(pCur
) );
777 sqlite3_free(pCur
->pKey
);
779 pCur
->eState
= CURSOR_INVALID
;
783 ** In this version of BtreeMoveto, pKey is a packed index record
784 ** such as is generated by the OP_MakeRecord opcode. Unpack the
785 ** record and then call BtreeMovetoUnpacked() to do the work.
787 static int btreeMoveto(
788 BtCursor
*pCur
, /* Cursor open on the btree to be searched */
789 const void *pKey
, /* Packed key if the btree is an index */
790 i64 nKey
, /* Integer key for tables. Size of pKey for indices */
791 int bias
, /* Bias search to the high end */
792 int *pRes
/* Write search results here */
794 int rc
; /* Status code */
795 UnpackedRecord
*pIdxKey
; /* Unpacked index key */
798 assert( nKey
==(i64
)(int)nKey
);
799 pIdxKey
= sqlite3VdbeAllocUnpackedRecord(pCur
->pKeyInfo
);
800 if( pIdxKey
==0 ) return SQLITE_NOMEM_BKPT
;
801 sqlite3VdbeRecordUnpack(pCur
->pKeyInfo
, (int)nKey
, pKey
, pIdxKey
);
802 if( pIdxKey
->nField
==0 ){
803 rc
= SQLITE_CORRUPT_BKPT
;
809 rc
= sqlite3BtreeMovetoUnpacked(pCur
, pIdxKey
, nKey
, bias
, pRes
);
812 sqlite3DbFree(pCur
->pKeyInfo
->db
, pIdxKey
);
818 ** Restore the cursor to the position it was in (or as close to as possible)
819 ** when saveCursorPosition() was called. Note that this call deletes the
820 ** saved position info stored by saveCursorPosition(), so there can be
821 ** at most one effective restoreCursorPosition() call after each
822 ** saveCursorPosition().
824 static int btreeRestoreCursorPosition(BtCursor
*pCur
){
827 assert( cursorOwnsBtShared(pCur
) );
828 assert( pCur
->eState
>=CURSOR_REQUIRESEEK
);
829 if( pCur
->eState
==CURSOR_FAULT
){
830 return pCur
->skipNext
;
832 pCur
->eState
= CURSOR_INVALID
;
833 rc
= btreeMoveto(pCur
, pCur
->pKey
, pCur
->nKey
, 0, &skipNext
);
835 sqlite3_free(pCur
->pKey
);
837 assert( pCur
->eState
==CURSOR_VALID
|| pCur
->eState
==CURSOR_INVALID
);
838 pCur
->skipNext
|= skipNext
;
839 if( pCur
->skipNext
&& pCur
->eState
==CURSOR_VALID
){
840 pCur
->eState
= CURSOR_SKIPNEXT
;
846 #define restoreCursorPosition(p) \
847 (p->eState>=CURSOR_REQUIRESEEK ? \
848 btreeRestoreCursorPosition(p) : \
852 ** Determine whether or not a cursor has moved from the position where
853 ** it was last placed, or has been invalidated for any other reason.
854 ** Cursors can move when the row they are pointing at is deleted out
855 ** from under them, for example. Cursor might also move if a btree
858 ** Calling this routine with a NULL cursor pointer returns false.
860 ** Use the separate sqlite3BtreeCursorRestore() routine to restore a cursor
861 ** back to where it ought to be if this routine returns true.
863 int sqlite3BtreeCursorHasMoved(BtCursor
*pCur
){
864 return pCur
->eState
!=CURSOR_VALID
;
868 ** Return a pointer to a fake BtCursor object that will always answer
869 ** false to the sqlite3BtreeCursorHasMoved() routine above. The fake
870 ** cursor returned must not be used with any other Btree interface.
872 BtCursor
*sqlite3BtreeFakeValidCursor(void){
873 static u8 fakeCursor
= CURSOR_VALID
;
874 assert( offsetof(BtCursor
, eState
)==0 );
875 return (BtCursor
*)&fakeCursor
;
879 ** This routine restores a cursor back to its original position after it
880 ** has been moved by some outside activity (such as a btree rebalance or
881 ** a row having been deleted out from under the cursor).
883 ** On success, the *pDifferentRow parameter is false if the cursor is left
884 ** pointing at exactly the same row. *pDifferntRow is the row the cursor
885 ** was pointing to has been deleted, forcing the cursor to point to some
888 ** This routine should only be called for a cursor that just returned
889 ** TRUE from sqlite3BtreeCursorHasMoved().
891 int sqlite3BtreeCursorRestore(BtCursor
*pCur
, int *pDifferentRow
){
895 assert( pCur
->eState
!=CURSOR_VALID
);
896 rc
= restoreCursorPosition(pCur
);
901 if( pCur
->eState
!=CURSOR_VALID
){
904 assert( pCur
->skipNext
==0 );
910 #ifdef SQLITE_ENABLE_CURSOR_HINTS
912 ** Provide hints to the cursor. The particular hint given (and the type
913 ** and number of the varargs parameters) is determined by the eHintType
914 ** parameter. See the definitions of the BTREE_HINT_* macros for details.
916 void sqlite3BtreeCursorHint(BtCursor
*pCur
, int eHintType
, ...){
917 /* Used only by system that substitute their own storage engine */
922 ** Provide flag hints to the cursor.
924 void sqlite3BtreeCursorHintFlags(BtCursor
*pCur
, unsigned x
){
925 assert( x
==BTREE_SEEK_EQ
|| x
==BTREE_BULKLOAD
|| x
==0 );
930 #ifndef SQLITE_OMIT_AUTOVACUUM
932 ** Given a page number of a regular database page, return the page
933 ** number for the pointer-map page that contains the entry for the
934 ** input page number.
936 ** Return 0 (not a valid page) for pgno==1 since there is
937 ** no pointer map associated with page 1. The integrity_check logic
938 ** requires that ptrmapPageno(*,1)!=1.
940 static Pgno
ptrmapPageno(BtShared
*pBt
, Pgno pgno
){
941 int nPagesPerMapPage
;
943 assert( sqlite3_mutex_held(pBt
->mutex
) );
944 if( pgno
<2 ) return 0;
945 nPagesPerMapPage
= (pBt
->usableSize
/5)+1;
946 iPtrMap
= (pgno
-2)/nPagesPerMapPage
;
947 ret
= (iPtrMap
*nPagesPerMapPage
) + 2;
948 if( ret
==PENDING_BYTE_PAGE(pBt
) ){
955 ** Write an entry into the pointer map.
957 ** This routine updates the pointer map entry for page number 'key'
958 ** so that it maps to type 'eType' and parent page number 'pgno'.
960 ** If *pRC is initially non-zero (non-SQLITE_OK) then this routine is
961 ** a no-op. If an error occurs, the appropriate error code is written
964 static void ptrmapPut(BtShared
*pBt
, Pgno key
, u8 eType
, Pgno parent
, int *pRC
){
965 DbPage
*pDbPage
; /* The pointer map page */
966 u8
*pPtrmap
; /* The pointer map data */
967 Pgno iPtrmap
; /* The pointer map page number */
968 int offset
; /* Offset in pointer map page */
969 int rc
; /* Return code from subfunctions */
973 assert( sqlite3_mutex_held(pBt
->mutex
) );
974 /* The master-journal page number must never be used as a pointer map page */
975 assert( 0==PTRMAP_ISPAGE(pBt
, PENDING_BYTE_PAGE(pBt
)) );
977 assert( pBt
->autoVacuum
);
979 *pRC
= SQLITE_CORRUPT_BKPT
;
982 iPtrmap
= PTRMAP_PAGENO(pBt
, key
);
983 rc
= sqlite3PagerGet(pBt
->pPager
, iPtrmap
, &pDbPage
, 0);
988 offset
= PTRMAP_PTROFFSET(iPtrmap
, key
);
990 *pRC
= SQLITE_CORRUPT_BKPT
;
993 assert( offset
<= (int)pBt
->usableSize
-5 );
994 pPtrmap
= (u8
*)sqlite3PagerGetData(pDbPage
);
996 if( eType
!=pPtrmap
[offset
] || get4byte(&pPtrmap
[offset
+1])!=parent
){
997 TRACE(("PTRMAP_UPDATE: %d->(%d,%d)\n", key
, eType
, parent
));
998 *pRC
= rc
= sqlite3PagerWrite(pDbPage
);
1000 pPtrmap
[offset
] = eType
;
1001 put4byte(&pPtrmap
[offset
+1], parent
);
1006 sqlite3PagerUnref(pDbPage
);
1010 ** Read an entry from the pointer map.
1012 ** This routine retrieves the pointer map entry for page 'key', writing
1013 ** the type and parent page number to *pEType and *pPgno respectively.
1014 ** An error code is returned if something goes wrong, otherwise SQLITE_OK.
1016 static int ptrmapGet(BtShared
*pBt
, Pgno key
, u8
*pEType
, Pgno
*pPgno
){
1017 DbPage
*pDbPage
; /* The pointer map page */
1018 int iPtrmap
; /* Pointer map page index */
1019 u8
*pPtrmap
; /* Pointer map page data */
1020 int offset
; /* Offset of entry in pointer map */
1023 assert( sqlite3_mutex_held(pBt
->mutex
) );
1025 iPtrmap
= PTRMAP_PAGENO(pBt
, key
);
1026 rc
= sqlite3PagerGet(pBt
->pPager
, iPtrmap
, &pDbPage
, 0);
1030 pPtrmap
= (u8
*)sqlite3PagerGetData(pDbPage
);
1032 offset
= PTRMAP_PTROFFSET(iPtrmap
, key
);
1034 sqlite3PagerUnref(pDbPage
);
1035 return SQLITE_CORRUPT_BKPT
;
1037 assert( offset
<= (int)pBt
->usableSize
-5 );
1038 assert( pEType
!=0 );
1039 *pEType
= pPtrmap
[offset
];
1040 if( pPgno
) *pPgno
= get4byte(&pPtrmap
[offset
+1]);
1042 sqlite3PagerUnref(pDbPage
);
1043 if( *pEType
<1 || *pEType
>5 ) return SQLITE_CORRUPT_PGNO(iPtrmap
);
1047 #else /* if defined SQLITE_OMIT_AUTOVACUUM */
1048 #define ptrmapPut(w,x,y,z,rc)
1049 #define ptrmapGet(w,x,y,z) SQLITE_OK
1050 #define ptrmapPutOvflPtr(x, y, rc)
1054 ** Given a btree page and a cell index (0 means the first cell on
1055 ** the page, 1 means the second cell, and so forth) return a pointer
1056 ** to the cell content.
1058 ** findCellPastPtr() does the same except it skips past the initial
1059 ** 4-byte child pointer found on interior pages, if there is one.
1061 ** This routine works only for pages that do not contain overflow cells.
1063 #define findCell(P,I) \
1064 ((P)->aData + ((P)->maskPage & get2byteAligned(&(P)->aCellIdx[2*(I)])))
1065 #define findCellPastPtr(P,I) \
1066 ((P)->aDataOfst + ((P)->maskPage & get2byteAligned(&(P)->aCellIdx[2*(I)])))
1070 ** This is common tail processing for btreeParseCellPtr() and
1071 ** btreeParseCellPtrIndex() for the case when the cell does not fit entirely
1072 ** on a single B-tree page. Make necessary adjustments to the CellInfo
1075 static SQLITE_NOINLINE
void btreeParseCellAdjustSizeForOverflow(
1076 MemPage
*pPage
, /* Page containing the cell */
1077 u8
*pCell
, /* Pointer to the cell text. */
1078 CellInfo
*pInfo
/* Fill in this structure */
1080 /* If the payload will not fit completely on the local page, we have
1081 ** to decide how much to store locally and how much to spill onto
1082 ** overflow pages. The strategy is to minimize the amount of unused
1083 ** space on overflow pages while keeping the amount of local storage
1084 ** in between minLocal and maxLocal.
1086 ** Warning: changing the way overflow payload is distributed in any
1087 ** way will result in an incompatible file format.
1089 int minLocal
; /* Minimum amount of payload held locally */
1090 int maxLocal
; /* Maximum amount of payload held locally */
1091 int surplus
; /* Overflow payload available for local storage */
1093 minLocal
= pPage
->minLocal
;
1094 maxLocal
= pPage
->maxLocal
;
1095 surplus
= minLocal
+ (pInfo
->nPayload
- minLocal
)%(pPage
->pBt
->usableSize
-4);
1096 testcase( surplus
==maxLocal
);
1097 testcase( surplus
==maxLocal
+1 );
1098 if( surplus
<= maxLocal
){
1099 pInfo
->nLocal
= (u16
)surplus
;
1101 pInfo
->nLocal
= (u16
)minLocal
;
1103 pInfo
->nSize
= (u16
)(&pInfo
->pPayload
[pInfo
->nLocal
] - pCell
) + 4;
1107 ** The following routines are implementations of the MemPage.xParseCell()
1110 ** Parse a cell content block and fill in the CellInfo structure.
1112 ** btreeParseCellPtr() => table btree leaf nodes
1113 ** btreeParseCellNoPayload() => table btree internal nodes
1114 ** btreeParseCellPtrIndex() => index btree nodes
1116 ** There is also a wrapper function btreeParseCell() that works for
1117 ** all MemPage types and that references the cell by index rather than
1120 static void btreeParseCellPtrNoPayload(
1121 MemPage
*pPage
, /* Page containing the cell */
1122 u8
*pCell
, /* Pointer to the cell text. */
1123 CellInfo
*pInfo
/* Fill in this structure */
1125 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1126 assert( pPage
->leaf
==0 );
1127 assert( pPage
->childPtrSize
==4 );
1128 #ifndef SQLITE_DEBUG
1129 UNUSED_PARAMETER(pPage
);
1131 pInfo
->nSize
= 4 + getVarint(&pCell
[4], (u64
*)&pInfo
->nKey
);
1132 pInfo
->nPayload
= 0;
1134 pInfo
->pPayload
= 0;
1137 static void btreeParseCellPtr(
1138 MemPage
*pPage
, /* Page containing the cell */
1139 u8
*pCell
, /* Pointer to the cell text. */
1140 CellInfo
*pInfo
/* Fill in this structure */
1142 u8
*pIter
; /* For scanning through pCell */
1143 u32 nPayload
; /* Number of bytes of cell payload */
1144 u64 iKey
; /* Extracted Key value */
1146 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1147 assert( pPage
->leaf
==0 || pPage
->leaf
==1 );
1148 assert( pPage
->intKeyLeaf
);
1149 assert( pPage
->childPtrSize
==0 );
1152 /* The next block of code is equivalent to:
1154 ** pIter += getVarint32(pIter, nPayload);
1156 ** The code is inlined to avoid a function call.
1159 if( nPayload
>=0x80 ){
1160 u8
*pEnd
= &pIter
[8];
1163 nPayload
= (nPayload
<<7) | (*++pIter
& 0x7f);
1164 }while( (*pIter
)>=0x80 && pIter
<pEnd
);
1168 /* The next block of code is equivalent to:
1170 ** pIter += getVarint(pIter, (u64*)&pInfo->nKey);
1172 ** The code is inlined to avoid a function call.
1176 u8
*pEnd
= &pIter
[7];
1179 iKey
= (iKey
<<7) | (*++pIter
& 0x7f);
1180 if( (*pIter
)<0x80 ) break;
1182 iKey
= (iKey
<<8) | *++pIter
;
1189 pInfo
->nKey
= *(i64
*)&iKey
;
1190 pInfo
->nPayload
= nPayload
;
1191 pInfo
->pPayload
= pIter
;
1192 testcase( nPayload
==pPage
->maxLocal
);
1193 testcase( nPayload
==pPage
->maxLocal
+1 );
1194 if( nPayload
<=pPage
->maxLocal
){
1195 /* This is the (easy) common case where the entire payload fits
1196 ** on the local page. No overflow is required.
1198 pInfo
->nSize
= nPayload
+ (u16
)(pIter
- pCell
);
1199 if( pInfo
->nSize
<4 ) pInfo
->nSize
= 4;
1200 pInfo
->nLocal
= (u16
)nPayload
;
1202 btreeParseCellAdjustSizeForOverflow(pPage
, pCell
, pInfo
);
1205 static void btreeParseCellPtrIndex(
1206 MemPage
*pPage
, /* Page containing the cell */
1207 u8
*pCell
, /* Pointer to the cell text. */
1208 CellInfo
*pInfo
/* Fill in this structure */
1210 u8
*pIter
; /* For scanning through pCell */
1211 u32 nPayload
; /* Number of bytes of cell payload */
1213 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1214 assert( pPage
->leaf
==0 || pPage
->leaf
==1 );
1215 assert( pPage
->intKeyLeaf
==0 );
1216 pIter
= pCell
+ pPage
->childPtrSize
;
1218 if( nPayload
>=0x80 ){
1219 u8
*pEnd
= &pIter
[8];
1222 nPayload
= (nPayload
<<7) | (*++pIter
& 0x7f);
1223 }while( *(pIter
)>=0x80 && pIter
<pEnd
);
1226 pInfo
->nKey
= nPayload
;
1227 pInfo
->nPayload
= nPayload
;
1228 pInfo
->pPayload
= pIter
;
1229 testcase( nPayload
==pPage
->maxLocal
);
1230 testcase( nPayload
==pPage
->maxLocal
+1 );
1231 if( nPayload
<=pPage
->maxLocal
){
1232 /* This is the (easy) common case where the entire payload fits
1233 ** on the local page. No overflow is required.
1235 pInfo
->nSize
= nPayload
+ (u16
)(pIter
- pCell
);
1236 if( pInfo
->nSize
<4 ) pInfo
->nSize
= 4;
1237 pInfo
->nLocal
= (u16
)nPayload
;
1239 btreeParseCellAdjustSizeForOverflow(pPage
, pCell
, pInfo
);
1242 static void btreeParseCell(
1243 MemPage
*pPage
, /* Page containing the cell */
1244 int iCell
, /* The cell index. First cell is 0 */
1245 CellInfo
*pInfo
/* Fill in this structure */
1247 pPage
->xParseCell(pPage
, findCell(pPage
, iCell
), pInfo
);
1251 ** The following routines are implementations of the MemPage.xCellSize
1254 ** Compute the total number of bytes that a Cell needs in the cell
1255 ** data area of the btree-page. The return number includes the cell
1256 ** data header and the local payload, but not any overflow page or
1257 ** the space used by the cell pointer.
1259 ** cellSizePtrNoPayload() => table internal nodes
1260 ** cellSizePtr() => all index nodes & table leaf nodes
1262 static u16
cellSizePtr(MemPage
*pPage
, u8
*pCell
){
1263 u8
*pIter
= pCell
+ pPage
->childPtrSize
; /* For looping over bytes of pCell */
1264 u8
*pEnd
; /* End mark for a varint */
1265 u32 nSize
; /* Size value to return */
1268 /* The value returned by this function should always be the same as
1269 ** the (CellInfo.nSize) value found by doing a full parse of the
1270 ** cell. If SQLITE_DEBUG is defined, an assert() at the bottom of
1271 ** this function verifies that this invariant is not violated. */
1273 pPage
->xParseCell(pPage
, pCell
, &debuginfo
);
1281 nSize
= (nSize
<<7) | (*++pIter
& 0x7f);
1282 }while( *(pIter
)>=0x80 && pIter
<pEnd
);
1285 if( pPage
->intKey
){
1286 /* pIter now points at the 64-bit integer key value, a variable length
1287 ** integer. The following block moves pIter to point at the first byte
1288 ** past the end of the key value. */
1290 while( (*pIter
++)&0x80 && pIter
<pEnd
);
1292 testcase( nSize
==pPage
->maxLocal
);
1293 testcase( nSize
==pPage
->maxLocal
+1 );
1294 if( nSize
<=pPage
->maxLocal
){
1295 nSize
+= (u32
)(pIter
- pCell
);
1296 if( nSize
<4 ) nSize
= 4;
1298 int minLocal
= pPage
->minLocal
;
1299 nSize
= minLocal
+ (nSize
- minLocal
) % (pPage
->pBt
->usableSize
- 4);
1300 testcase( nSize
==pPage
->maxLocal
);
1301 testcase( nSize
==pPage
->maxLocal
+1 );
1302 if( nSize
>pPage
->maxLocal
){
1305 nSize
+= 4 + (u16
)(pIter
- pCell
);
1307 assert( nSize
==debuginfo
.nSize
|| CORRUPT_DB
);
1310 static u16
cellSizePtrNoPayload(MemPage
*pPage
, u8
*pCell
){
1311 u8
*pIter
= pCell
+ 4; /* For looping over bytes of pCell */
1312 u8
*pEnd
; /* End mark for a varint */
1315 /* The value returned by this function should always be the same as
1316 ** the (CellInfo.nSize) value found by doing a full parse of the
1317 ** cell. If SQLITE_DEBUG is defined, an assert() at the bottom of
1318 ** this function verifies that this invariant is not violated. */
1320 pPage
->xParseCell(pPage
, pCell
, &debuginfo
);
1322 UNUSED_PARAMETER(pPage
);
1325 assert( pPage
->childPtrSize
==4 );
1327 while( (*pIter
++)&0x80 && pIter
<pEnd
);
1328 assert( debuginfo
.nSize
==(u16
)(pIter
- pCell
) || CORRUPT_DB
);
1329 return (u16
)(pIter
- pCell
);
1334 /* This variation on cellSizePtr() is used inside of assert() statements
1336 static u16
cellSize(MemPage
*pPage
, int iCell
){
1337 return pPage
->xCellSize(pPage
, findCell(pPage
, iCell
));
1341 #ifndef SQLITE_OMIT_AUTOVACUUM
1343 ** If the cell pCell, part of page pPage contains a pointer
1344 ** to an overflow page, insert an entry into the pointer-map
1345 ** for the overflow page.
1347 static void ptrmapPutOvflPtr(MemPage
*pPage
, u8
*pCell
, int *pRC
){
1351 pPage
->xParseCell(pPage
, pCell
, &info
);
1352 if( info
.nLocal
<info
.nPayload
){
1353 Pgno ovfl
= get4byte(&pCell
[info
.nSize
-4]);
1354 ptrmapPut(pPage
->pBt
, ovfl
, PTRMAP_OVERFLOW1
, pPage
->pgno
, pRC
);
1361 ** Defragment the page given. This routine reorganizes cells within the
1362 ** page so that there are no free-blocks on the free-block list.
1364 ** Parameter nMaxFrag is the maximum amount of fragmented space that may be
1365 ** present in the page after this routine returns.
1367 ** EVIDENCE-OF: R-44582-60138 SQLite may from time to time reorganize a
1368 ** b-tree page so that there are no freeblocks or fragment bytes, all
1369 ** unused bytes are contained in the unallocated space region, and all
1370 ** cells are packed tightly at the end of the page.
1372 static int defragmentPage(MemPage
*pPage
, int nMaxFrag
){
1373 int i
; /* Loop counter */
1374 int pc
; /* Address of the i-th cell */
1375 int hdr
; /* Offset to the page header */
1376 int size
; /* Size of a cell */
1377 int usableSize
; /* Number of usable bytes on a page */
1378 int cellOffset
; /* Offset to the cell pointer array */
1379 int cbrk
; /* Offset to the cell content area */
1380 int nCell
; /* Number of cells on the page */
1381 unsigned char *data
; /* The page data */
1382 unsigned char *temp
; /* Temp area for cell content */
1383 unsigned char *src
; /* Source of content */
1384 int iCellFirst
; /* First allowable cell index */
1385 int iCellLast
; /* Last possible cell index */
1387 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1388 assert( pPage
->pBt
!=0 );
1389 assert( pPage
->pBt
->usableSize
<= SQLITE_MAX_PAGE_SIZE
);
1390 assert( pPage
->nOverflow
==0 );
1391 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1393 src
= data
= pPage
->aData
;
1394 hdr
= pPage
->hdrOffset
;
1395 cellOffset
= pPage
->cellOffset
;
1396 nCell
= pPage
->nCell
;
1397 assert( nCell
==get2byte(&data
[hdr
+3]) );
1398 iCellFirst
= cellOffset
+ 2*nCell
;
1399 usableSize
= pPage
->pBt
->usableSize
;
1401 /* This block handles pages with two or fewer free blocks and nMaxFrag
1402 ** or fewer fragmented bytes. In this case it is faster to move the
1403 ** two (or one) blocks of cells using memmove() and add the required
1404 ** offsets to each pointer in the cell-pointer array than it is to
1405 ** reconstruct the entire page. */
1406 if( (int)data
[hdr
+7]<=nMaxFrag
){
1407 int iFree
= get2byte(&data
[hdr
+1]);
1409 int iFree2
= get2byte(&data
[iFree
]);
1411 /* pageFindSlot() has already verified that free blocks are sorted
1412 ** in order of offset within the page, and that no block extends
1413 ** past the end of the page. Provided the two free slots do not
1414 ** overlap, this guarantees that the memmove() calls below will not
1415 ** overwrite the usableSize byte buffer, even if the database page
1417 assert( iFree2
==0 || iFree2
>iFree
);
1418 assert( iFree
+get2byte(&data
[iFree
+2]) <= usableSize
);
1419 assert( iFree2
==0 || iFree2
+get2byte(&data
[iFree2
+2]) <= usableSize
);
1421 if( 0==iFree2
|| (data
[iFree2
]==0 && data
[iFree2
+1]==0) ){
1422 u8
*pEnd
= &data
[cellOffset
+ nCell
*2];
1425 int sz
= get2byte(&data
[iFree
+2]);
1426 int top
= get2byte(&data
[hdr
+5]);
1428 return SQLITE_CORRUPT_PAGE(pPage
);
1431 assert( iFree
+sz
<=iFree2
); /* Verified by pageFindSlot() */
1432 sz2
= get2byte(&data
[iFree2
+2]);
1433 assert( iFree
+sz
+sz2
+iFree2
-(iFree
+sz
) <= usableSize
);
1434 memmove(&data
[iFree
+sz
+sz2
], &data
[iFree
+sz
], iFree2
-(iFree
+sz
));
1438 assert( cbrk
+(iFree
-top
) <= usableSize
);
1439 memmove(&data
[cbrk
], &data
[top
], iFree
-top
);
1440 for(pAddr
=&data
[cellOffset
]; pAddr
<pEnd
; pAddr
+=2){
1441 pc
= get2byte(pAddr
);
1442 if( pc
<iFree
){ put2byte(pAddr
, pc
+sz
); }
1443 else if( pc
<iFree2
){ put2byte(pAddr
, pc
+sz2
); }
1445 goto defragment_out
;
1451 iCellLast
= usableSize
- 4;
1452 for(i
=0; i
<nCell
; i
++){
1453 u8
*pAddr
; /* The i-th cell pointer */
1454 pAddr
= &data
[cellOffset
+ i
*2];
1455 pc
= get2byte(pAddr
);
1456 testcase( pc
==iCellFirst
);
1457 testcase( pc
==iCellLast
);
1458 /* These conditions have already been verified in btreeInitPage()
1459 ** if PRAGMA cell_size_check=ON.
1461 if( pc
<iCellFirst
|| pc
>iCellLast
){
1462 return SQLITE_CORRUPT_PAGE(pPage
);
1464 assert( pc
>=iCellFirst
&& pc
<=iCellLast
);
1465 size
= pPage
->xCellSize(pPage
, &src
[pc
]);
1467 if( cbrk
<iCellFirst
|| pc
+size
>usableSize
){
1468 return SQLITE_CORRUPT_PAGE(pPage
);
1470 assert( cbrk
+size
<=usableSize
&& cbrk
>=iCellFirst
);
1471 testcase( cbrk
+size
==usableSize
);
1472 testcase( pc
+size
==usableSize
);
1473 put2byte(pAddr
, cbrk
);
1476 if( cbrk
==pc
) continue;
1477 temp
= sqlite3PagerTempSpace(pPage
->pBt
->pPager
);
1478 x
= get2byte(&data
[hdr
+5]);
1479 memcpy(&temp
[x
], &data
[x
], (cbrk
+size
) - x
);
1482 memcpy(&data
[cbrk
], &src
[pc
], size
);
1487 if( data
[hdr
+7]+cbrk
-iCellFirst
!=pPage
->nFree
){
1488 return SQLITE_CORRUPT_PAGE(pPage
);
1490 assert( cbrk
>=iCellFirst
);
1491 put2byte(&data
[hdr
+5], cbrk
);
1494 memset(&data
[iCellFirst
], 0, cbrk
-iCellFirst
);
1495 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1500 ** Search the free-list on page pPg for space to store a cell nByte bytes in
1501 ** size. If one can be found, return a pointer to the space and remove it
1502 ** from the free-list.
1504 ** If no suitable space can be found on the free-list, return NULL.
1506 ** This function may detect corruption within pPg. If corruption is
1507 ** detected then *pRc is set to SQLITE_CORRUPT and NULL is returned.
1509 ** Slots on the free list that are between 1 and 3 bytes larger than nByte
1510 ** will be ignored if adding the extra space to the fragmentation count
1511 ** causes the fragmentation count to exceed 60.
1513 static u8
*pageFindSlot(MemPage
*pPg
, int nByte
, int *pRc
){
1514 const int hdr
= pPg
->hdrOffset
;
1515 u8
* const aData
= pPg
->aData
;
1516 int iAddr
= hdr
+ 1;
1517 int pc
= get2byte(&aData
[iAddr
]);
1519 int usableSize
= pPg
->pBt
->usableSize
;
1520 int size
; /* Size of the free slot */
1523 while( pc
<=usableSize
-4 ){
1524 /* EVIDENCE-OF: R-22710-53328 The third and fourth bytes of each
1525 ** freeblock form a big-endian integer which is the size of the freeblock
1526 ** in bytes, including the 4-byte header. */
1527 size
= get2byte(&aData
[pc
+2]);
1528 if( (x
= size
- nByte
)>=0 ){
1531 if( size
+pc
> usableSize
){
1532 *pRc
= SQLITE_CORRUPT_PAGE(pPg
);
1535 /* EVIDENCE-OF: R-11498-58022 In a well-formed b-tree page, the total
1536 ** number of bytes in fragments may not exceed 60. */
1537 if( aData
[hdr
+7]>57 ) return 0;
1539 /* Remove the slot from the free-list. Update the number of
1540 ** fragmented bytes within the page. */
1541 memcpy(&aData
[iAddr
], &aData
[pc
], 2);
1542 aData
[hdr
+7] += (u8
)x
;
1544 /* The slot remains on the free-list. Reduce its size to account
1545 ** for the portion used by the new allocation. */
1546 put2byte(&aData
[pc
+2], x
);
1548 return &aData
[pc
+ x
];
1551 pc
= get2byte(&aData
[pc
]);
1552 if( pc
<iAddr
+size
) break;
1555 *pRc
= SQLITE_CORRUPT_PAGE(pPg
);
1562 ** Allocate nByte bytes of space from within the B-Tree page passed
1563 ** as the first argument. Write into *pIdx the index into pPage->aData[]
1564 ** of the first byte of allocated space. Return either SQLITE_OK or
1565 ** an error code (usually SQLITE_CORRUPT).
1567 ** The caller guarantees that there is sufficient space to make the
1568 ** allocation. This routine might need to defragment in order to bring
1569 ** all the space together, however. This routine will avoid using
1570 ** the first two bytes past the cell pointer area since presumably this
1571 ** allocation is being made in order to insert a new cell, so we will
1572 ** also end up needing a new cell pointer.
1574 static int allocateSpace(MemPage
*pPage
, int nByte
, int *pIdx
){
1575 const int hdr
= pPage
->hdrOffset
; /* Local cache of pPage->hdrOffset */
1576 u8
* const data
= pPage
->aData
; /* Local cache of pPage->aData */
1577 int top
; /* First byte of cell content area */
1578 int rc
= SQLITE_OK
; /* Integer return code */
1579 int gap
; /* First byte of gap between cell pointers and cell content */
1581 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1582 assert( pPage
->pBt
);
1583 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1584 assert( nByte
>=0 ); /* Minimum cell size is 4 */
1585 assert( pPage
->nFree
>=nByte
);
1586 assert( pPage
->nOverflow
==0 );
1587 assert( nByte
< (int)(pPage
->pBt
->usableSize
-8) );
1589 assert( pPage
->cellOffset
== hdr
+ 12 - 4*pPage
->leaf
);
1590 gap
= pPage
->cellOffset
+ 2*pPage
->nCell
;
1591 assert( gap
<=65536 );
1592 /* EVIDENCE-OF: R-29356-02391 If the database uses a 65536-byte page size
1593 ** and the reserved space is zero (the usual value for reserved space)
1594 ** then the cell content offset of an empty page wants to be 65536.
1595 ** However, that integer is too large to be stored in a 2-byte unsigned
1596 ** integer, so a value of 0 is used in its place. */
1597 top
= get2byte(&data
[hdr
+5]);
1598 assert( top
<=(int)pPage
->pBt
->usableSize
); /* Prevent by getAndInitPage() */
1600 if( top
==0 && pPage
->pBt
->usableSize
==65536 ){
1603 return SQLITE_CORRUPT_PAGE(pPage
);
1607 /* If there is enough space between gap and top for one more cell pointer
1608 ** array entry offset, and if the freelist is not empty, then search the
1609 ** freelist looking for a free slot big enough to satisfy the request.
1611 testcase( gap
+2==top
);
1612 testcase( gap
+1==top
);
1613 testcase( gap
==top
);
1614 if( (data
[hdr
+2] || data
[hdr
+1]) && gap
+2<=top
){
1615 u8
*pSpace
= pageFindSlot(pPage
, nByte
, &rc
);
1617 assert( pSpace
>=data
&& (pSpace
- data
)<65536 );
1618 *pIdx
= (int)(pSpace
- data
);
1625 /* The request could not be fulfilled using a freelist slot. Check
1626 ** to see if defragmentation is necessary.
1628 testcase( gap
+2+nByte
==top
);
1629 if( gap
+2+nByte
>top
){
1630 assert( pPage
->nCell
>0 || CORRUPT_DB
);
1631 rc
= defragmentPage(pPage
, MIN(4, pPage
->nFree
- (2+nByte
)));
1633 top
= get2byteNotZero(&data
[hdr
+5]);
1634 assert( gap
+2+nByte
<=top
);
1638 /* Allocate memory from the gap in between the cell pointer array
1639 ** and the cell content area. The btreeInitPage() call has already
1640 ** validated the freelist. Given that the freelist is valid, there
1641 ** is no way that the allocation can extend off the end of the page.
1642 ** The assert() below verifies the previous sentence.
1645 put2byte(&data
[hdr
+5], top
);
1646 assert( top
+nByte
<= (int)pPage
->pBt
->usableSize
);
1652 ** Return a section of the pPage->aData to the freelist.
1653 ** The first byte of the new free block is pPage->aData[iStart]
1654 ** and the size of the block is iSize bytes.
1656 ** Adjacent freeblocks are coalesced.
1658 ** Note that even though the freeblock list was checked by btreeInitPage(),
1659 ** that routine will not detect overlap between cells or freeblocks. Nor
1660 ** does it detect cells or freeblocks that encrouch into the reserved bytes
1661 ** at the end of the page. So do additional corruption checks inside this
1662 ** routine and return SQLITE_CORRUPT if any problems are found.
1664 static int freeSpace(MemPage
*pPage
, u16 iStart
, u16 iSize
){
1665 u16 iPtr
; /* Address of ptr to next freeblock */
1666 u16 iFreeBlk
; /* Address of the next freeblock */
1667 u8 hdr
; /* Page header size. 0 or 100 */
1668 u8 nFrag
= 0; /* Reduction in fragmentation */
1669 u16 iOrigSize
= iSize
; /* Original value of iSize */
1670 u16 x
; /* Offset to cell content area */
1671 u32 iEnd
= iStart
+ iSize
; /* First byte past the iStart buffer */
1672 unsigned char *data
= pPage
->aData
; /* Page content */
1674 assert( pPage
->pBt
!=0 );
1675 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1676 assert( CORRUPT_DB
|| iStart
>=pPage
->hdrOffset
+6+pPage
->childPtrSize
);
1677 assert( CORRUPT_DB
|| iEnd
<= pPage
->pBt
->usableSize
);
1678 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1679 assert( iSize
>=4 ); /* Minimum cell size is 4 */
1680 assert( iStart
<=pPage
->pBt
->usableSize
-4 );
1682 /* The list of freeblocks must be in ascending order. Find the
1683 ** spot on the list where iStart should be inserted.
1685 hdr
= pPage
->hdrOffset
;
1687 if( data
[iPtr
+1]==0 && data
[iPtr
]==0 ){
1688 iFreeBlk
= 0; /* Shortcut for the case when the freelist is empty */
1690 while( (iFreeBlk
= get2byte(&data
[iPtr
]))<iStart
){
1691 if( iFreeBlk
<iPtr
+4 ){
1692 if( iFreeBlk
==0 ) break;
1693 return SQLITE_CORRUPT_PAGE(pPage
);
1697 if( iFreeBlk
>pPage
->pBt
->usableSize
-4 ){
1698 return SQLITE_CORRUPT_PAGE(pPage
);
1700 assert( iFreeBlk
>iPtr
|| iFreeBlk
==0 );
1703 ** iFreeBlk: First freeblock after iStart, or zero if none
1704 ** iPtr: The address of a pointer to iFreeBlk
1706 ** Check to see if iFreeBlk should be coalesced onto the end of iStart.
1708 if( iFreeBlk
&& iEnd
+3>=iFreeBlk
){
1709 nFrag
= iFreeBlk
- iEnd
;
1710 if( iEnd
>iFreeBlk
) return SQLITE_CORRUPT_PAGE(pPage
);
1711 iEnd
= iFreeBlk
+ get2byte(&data
[iFreeBlk
+2]);
1712 if( iEnd
> pPage
->pBt
->usableSize
){
1713 return SQLITE_CORRUPT_PAGE(pPage
);
1715 iSize
= iEnd
- iStart
;
1716 iFreeBlk
= get2byte(&data
[iFreeBlk
]);
1719 /* If iPtr is another freeblock (that is, if iPtr is not the freelist
1720 ** pointer in the page header) then check to see if iStart should be
1721 ** coalesced onto the end of iPtr.
1724 int iPtrEnd
= iPtr
+ get2byte(&data
[iPtr
+2]);
1725 if( iPtrEnd
+3>=iStart
){
1726 if( iPtrEnd
>iStart
) return SQLITE_CORRUPT_PAGE(pPage
);
1727 nFrag
+= iStart
- iPtrEnd
;
1728 iSize
= iEnd
- iPtr
;
1732 if( nFrag
>data
[hdr
+7] ) return SQLITE_CORRUPT_PAGE(pPage
);
1733 data
[hdr
+7] -= nFrag
;
1735 x
= get2byte(&data
[hdr
+5]);
1737 /* The new freeblock is at the beginning of the cell content area,
1738 ** so just extend the cell content area rather than create another
1739 ** freelist entry */
1740 if( iStart
<x
|| iPtr
!=hdr
+1 ) return SQLITE_CORRUPT_PAGE(pPage
);
1741 put2byte(&data
[hdr
+1], iFreeBlk
);
1742 put2byte(&data
[hdr
+5], iEnd
);
1744 /* Insert the new freeblock into the freelist */
1745 put2byte(&data
[iPtr
], iStart
);
1747 if( pPage
->pBt
->btsFlags
& BTS_FAST_SECURE
){
1748 /* Overwrite deleted information with zeros when the secure_delete
1749 ** option is enabled */
1750 memset(&data
[iStart
], 0, iSize
);
1752 put2byte(&data
[iStart
], iFreeBlk
);
1753 put2byte(&data
[iStart
+2], iSize
);
1754 pPage
->nFree
+= iOrigSize
;
1759 ** Decode the flags byte (the first byte of the header) for a page
1760 ** and initialize fields of the MemPage structure accordingly.
1762 ** Only the following combinations are supported. Anything different
1763 ** indicates a corrupt database files:
1766 ** PTF_ZERODATA | PTF_LEAF
1767 ** PTF_LEAFDATA | PTF_INTKEY
1768 ** PTF_LEAFDATA | PTF_INTKEY | PTF_LEAF
1770 static int decodeFlags(MemPage
*pPage
, int flagByte
){
1771 BtShared
*pBt
; /* A copy of pPage->pBt */
1773 assert( pPage
->hdrOffset
==(pPage
->pgno
==1 ? 100 : 0) );
1774 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1775 pPage
->leaf
= (u8
)(flagByte
>>3); assert( PTF_LEAF
== 1<<3 );
1776 flagByte
&= ~PTF_LEAF
;
1777 pPage
->childPtrSize
= 4-4*pPage
->leaf
;
1778 pPage
->xCellSize
= cellSizePtr
;
1780 if( flagByte
==(PTF_LEAFDATA
| PTF_INTKEY
) ){
1781 /* EVIDENCE-OF: R-07291-35328 A value of 5 (0x05) means the page is an
1782 ** interior table b-tree page. */
1783 assert( (PTF_LEAFDATA
|PTF_INTKEY
)==5 );
1784 /* EVIDENCE-OF: R-26900-09176 A value of 13 (0x0d) means the page is a
1785 ** leaf table b-tree page. */
1786 assert( (PTF_LEAFDATA
|PTF_INTKEY
|PTF_LEAF
)==13 );
1789 pPage
->intKeyLeaf
= 1;
1790 pPage
->xParseCell
= btreeParseCellPtr
;
1792 pPage
->intKeyLeaf
= 0;
1793 pPage
->xCellSize
= cellSizePtrNoPayload
;
1794 pPage
->xParseCell
= btreeParseCellPtrNoPayload
;
1796 pPage
->maxLocal
= pBt
->maxLeaf
;
1797 pPage
->minLocal
= pBt
->minLeaf
;
1798 }else if( flagByte
==PTF_ZERODATA
){
1799 /* EVIDENCE-OF: R-43316-37308 A value of 2 (0x02) means the page is an
1800 ** interior index b-tree page. */
1801 assert( (PTF_ZERODATA
)==2 );
1802 /* EVIDENCE-OF: R-59615-42828 A value of 10 (0x0a) means the page is a
1803 ** leaf index b-tree page. */
1804 assert( (PTF_ZERODATA
|PTF_LEAF
)==10 );
1806 pPage
->intKeyLeaf
= 0;
1807 pPage
->xParseCell
= btreeParseCellPtrIndex
;
1808 pPage
->maxLocal
= pBt
->maxLocal
;
1809 pPage
->minLocal
= pBt
->minLocal
;
1811 /* EVIDENCE-OF: R-47608-56469 Any other value for the b-tree page type is
1813 return SQLITE_CORRUPT_PAGE(pPage
);
1815 pPage
->max1bytePayload
= pBt
->max1bytePayload
;
1820 ** Initialize the auxiliary information for a disk block.
1822 ** Return SQLITE_OK on success. If we see that the page does
1823 ** not contain a well-formed database page, then return
1824 ** SQLITE_CORRUPT. Note that a return of SQLITE_OK does not
1825 ** guarantee that the page is well-formed. It only shows that
1826 ** we failed to detect any corruption.
1828 static int btreeInitPage(MemPage
*pPage
){
1829 int pc
; /* Address of a freeblock within pPage->aData[] */
1830 u8 hdr
; /* Offset to beginning of page header */
1831 u8
*data
; /* Equal to pPage->aData */
1832 BtShared
*pBt
; /* The main btree structure */
1833 int usableSize
; /* Amount of usable space on each page */
1834 u16 cellOffset
; /* Offset from start of page to first cell pointer */
1835 int nFree
; /* Number of unused bytes on the page */
1836 int top
; /* First byte of the cell content area */
1837 int iCellFirst
; /* First allowable cell or freeblock offset */
1838 int iCellLast
; /* Last possible cell or freeblock offset */
1840 assert( pPage
->pBt
!=0 );
1841 assert( pPage
->pBt
->db
!=0 );
1842 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1843 assert( pPage
->pgno
==sqlite3PagerPagenumber(pPage
->pDbPage
) );
1844 assert( pPage
== sqlite3PagerGetExtra(pPage
->pDbPage
) );
1845 assert( pPage
->aData
== sqlite3PagerGetData(pPage
->pDbPage
) );
1846 assert( pPage
->isInit
==0 );
1849 hdr
= pPage
->hdrOffset
;
1850 data
= pPage
->aData
;
1851 /* EVIDENCE-OF: R-28594-02890 The one-byte flag at offset 0 indicating
1852 ** the b-tree page type. */
1853 if( decodeFlags(pPage
, data
[hdr
]) ){
1854 return SQLITE_CORRUPT_PAGE(pPage
);
1856 assert( pBt
->pageSize
>=512 && pBt
->pageSize
<=65536 );
1857 pPage
->maskPage
= (u16
)(pBt
->pageSize
- 1);
1858 pPage
->nOverflow
= 0;
1859 usableSize
= pBt
->usableSize
;
1860 pPage
->cellOffset
= cellOffset
= hdr
+ 8 + pPage
->childPtrSize
;
1861 pPage
->aDataEnd
= &data
[usableSize
];
1862 pPage
->aCellIdx
= &data
[cellOffset
];
1863 pPage
->aDataOfst
= &data
[pPage
->childPtrSize
];
1864 /* EVIDENCE-OF: R-58015-48175 The two-byte integer at offset 5 designates
1865 ** the start of the cell content area. A zero value for this integer is
1866 ** interpreted as 65536. */
1867 top
= get2byteNotZero(&data
[hdr
+5]);
1868 /* EVIDENCE-OF: R-37002-32774 The two-byte integer at offset 3 gives the
1869 ** number of cells on the page. */
1870 pPage
->nCell
= get2byte(&data
[hdr
+3]);
1871 if( pPage
->nCell
>MX_CELL(pBt
) ){
1872 /* To many cells for a single page. The page must be corrupt */
1873 return SQLITE_CORRUPT_PAGE(pPage
);
1875 testcase( pPage
->nCell
==MX_CELL(pBt
) );
1876 /* EVIDENCE-OF: R-24089-57979 If a page contains no cells (which is only
1877 ** possible for a root page of a table that contains no rows) then the
1878 ** offset to the cell content area will equal the page size minus the
1879 ** bytes of reserved space. */
1880 assert( pPage
->nCell
>0 || top
==usableSize
|| CORRUPT_DB
);
1882 /* A malformed database page might cause us to read past the end
1883 ** of page when parsing a cell.
1885 ** The following block of code checks early to see if a cell extends
1886 ** past the end of a page boundary and causes SQLITE_CORRUPT to be
1887 ** returned if it does.
1889 iCellFirst
= cellOffset
+ 2*pPage
->nCell
;
1890 iCellLast
= usableSize
- 4;
1891 if( pBt
->db
->flags
& SQLITE_CellSizeCk
){
1892 int i
; /* Index into the cell pointer array */
1893 int sz
; /* Size of a cell */
1895 if( !pPage
->leaf
) iCellLast
--;
1896 for(i
=0; i
<pPage
->nCell
; i
++){
1897 pc
= get2byteAligned(&data
[cellOffset
+i
*2]);
1898 testcase( pc
==iCellFirst
);
1899 testcase( pc
==iCellLast
);
1900 if( pc
<iCellFirst
|| pc
>iCellLast
){
1901 return SQLITE_CORRUPT_PAGE(pPage
);
1903 sz
= pPage
->xCellSize(pPage
, &data
[pc
]);
1904 testcase( pc
+sz
==usableSize
);
1905 if( pc
+sz
>usableSize
){
1906 return SQLITE_CORRUPT_PAGE(pPage
);
1909 if( !pPage
->leaf
) iCellLast
++;
1912 /* Compute the total free space on the page
1913 ** EVIDENCE-OF: R-23588-34450 The two-byte integer at offset 1 gives the
1914 ** start of the first freeblock on the page, or is zero if there are no
1916 pc
= get2byte(&data
[hdr
+1]);
1917 nFree
= data
[hdr
+7] + top
; /* Init nFree to non-freeblock free space */
1920 if( pc
<iCellFirst
){
1921 /* EVIDENCE-OF: R-55530-52930 In a well-formed b-tree page, there will
1922 ** always be at least one cell before the first freeblock.
1924 return SQLITE_CORRUPT_PAGE(pPage
);
1928 /* Freeblock off the end of the page */
1929 return SQLITE_CORRUPT_PAGE(pPage
);
1931 next
= get2byte(&data
[pc
]);
1932 size
= get2byte(&data
[pc
+2]);
1933 nFree
= nFree
+ size
;
1934 if( next
<=pc
+size
+3 ) break;
1938 /* Freeblock not in ascending order */
1939 return SQLITE_CORRUPT_PAGE(pPage
);
1941 if( pc
+size
>(unsigned int)usableSize
){
1942 /* Last freeblock extends past page end */
1943 return SQLITE_CORRUPT_PAGE(pPage
);
1947 /* At this point, nFree contains the sum of the offset to the start
1948 ** of the cell-content area plus the number of free bytes within
1949 ** the cell-content area. If this is greater than the usable-size
1950 ** of the page, then the page must be corrupted. This check also
1951 ** serves to verify that the offset to the start of the cell-content
1952 ** area, according to the page header, lies within the page.
1954 if( nFree
>usableSize
){
1955 return SQLITE_CORRUPT_PAGE(pPage
);
1957 pPage
->nFree
= (u16
)(nFree
- iCellFirst
);
1963 ** Set up a raw page so that it looks like a database page holding
1966 static void zeroPage(MemPage
*pPage
, int flags
){
1967 unsigned char *data
= pPage
->aData
;
1968 BtShared
*pBt
= pPage
->pBt
;
1969 u8 hdr
= pPage
->hdrOffset
;
1972 assert( sqlite3PagerPagenumber(pPage
->pDbPage
)==pPage
->pgno
);
1973 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
1974 assert( sqlite3PagerGetData(pPage
->pDbPage
) == data
);
1975 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1976 assert( sqlite3_mutex_held(pBt
->mutex
) );
1977 if( pBt
->btsFlags
& BTS_FAST_SECURE
){
1978 memset(&data
[hdr
], 0, pBt
->usableSize
- hdr
);
1980 data
[hdr
] = (char)flags
;
1981 first
= hdr
+ ((flags
&PTF_LEAF
)==0 ? 12 : 8);
1982 memset(&data
[hdr
+1], 0, 4);
1984 put2byte(&data
[hdr
+5], pBt
->usableSize
);
1985 pPage
->nFree
= (u16
)(pBt
->usableSize
- first
);
1986 decodeFlags(pPage
, flags
);
1987 pPage
->cellOffset
= first
;
1988 pPage
->aDataEnd
= &data
[pBt
->usableSize
];
1989 pPage
->aCellIdx
= &data
[first
];
1990 pPage
->aDataOfst
= &data
[pPage
->childPtrSize
];
1991 pPage
->nOverflow
= 0;
1992 assert( pBt
->pageSize
>=512 && pBt
->pageSize
<=65536 );
1993 pPage
->maskPage
= (u16
)(pBt
->pageSize
- 1);
2000 ** Convert a DbPage obtained from the pager into a MemPage used by
2003 static MemPage
*btreePageFromDbPage(DbPage
*pDbPage
, Pgno pgno
, BtShared
*pBt
){
2004 MemPage
*pPage
= (MemPage
*)sqlite3PagerGetExtra(pDbPage
);
2005 if( pgno
!=pPage
->pgno
){
2006 pPage
->aData
= sqlite3PagerGetData(pDbPage
);
2007 pPage
->pDbPage
= pDbPage
;
2010 pPage
->hdrOffset
= pgno
==1 ? 100 : 0;
2012 assert( pPage
->aData
==sqlite3PagerGetData(pDbPage
) );
2017 ** Get a page from the pager. Initialize the MemPage.pBt and
2018 ** MemPage.aData elements if needed. See also: btreeGetUnusedPage().
2020 ** If the PAGER_GET_NOCONTENT flag is set, it means that we do not care
2021 ** about the content of the page at this time. So do not go to the disk
2022 ** to fetch the content. Just fill in the content with zeros for now.
2023 ** If in the future we call sqlite3PagerWrite() on this page, that
2024 ** means we have started to be concerned about content and the disk
2025 ** read should occur at that point.
2027 static int btreeGetPage(
2028 BtShared
*pBt
, /* The btree */
2029 Pgno pgno
, /* Number of the page to fetch */
2030 MemPage
**ppPage
, /* Return the page in this parameter */
2031 int flags
/* PAGER_GET_NOCONTENT or PAGER_GET_READONLY */
2036 assert( flags
==0 || flags
==PAGER_GET_NOCONTENT
|| flags
==PAGER_GET_READONLY
);
2037 assert( sqlite3_mutex_held(pBt
->mutex
) );
2038 rc
= sqlite3PagerGet(pBt
->pPager
, pgno
, (DbPage
**)&pDbPage
, flags
);
2040 *ppPage
= btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2045 ** Retrieve a page from the pager cache. If the requested page is not
2046 ** already in the pager cache return NULL. Initialize the MemPage.pBt and
2047 ** MemPage.aData elements if needed.
2049 static MemPage
*btreePageLookup(BtShared
*pBt
, Pgno pgno
){
2051 assert( sqlite3_mutex_held(pBt
->mutex
) );
2052 pDbPage
= sqlite3PagerLookup(pBt
->pPager
, pgno
);
2054 return btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2060 ** Return the size of the database file in pages. If there is any kind of
2061 ** error, return ((unsigned int)-1).
2063 static Pgno
btreePagecount(BtShared
*pBt
){
2066 u32
sqlite3BtreeLastPage(Btree
*p
){
2067 assert( sqlite3BtreeHoldsMutex(p
) );
2068 assert( ((p
->pBt
->nPage
)&0x80000000)==0 );
2069 return btreePagecount(p
->pBt
);
2073 ** Get a page from the pager and initialize it.
2075 ** If pCur!=0 then the page is being fetched as part of a moveToChild()
2076 ** call. Do additional sanity checking on the page in this case.
2077 ** And if the fetch fails, this routine must decrement pCur->iPage.
2079 ** The page is fetched as read-write unless pCur is not NULL and is
2080 ** a read-only cursor.
2082 ** If an error occurs, then *ppPage is undefined. It
2083 ** may remain unchanged, or it may be set to an invalid value.
2085 static int getAndInitPage(
2086 BtShared
*pBt
, /* The database file */
2087 Pgno pgno
, /* Number of the page to get */
2088 MemPage
**ppPage
, /* Write the page pointer here */
2089 BtCursor
*pCur
, /* Cursor to receive the page, or NULL */
2090 int bReadOnly
/* True for a read-only page */
2094 assert( sqlite3_mutex_held(pBt
->mutex
) );
2095 assert( pCur
==0 || ppPage
==&pCur
->pPage
);
2096 assert( pCur
==0 || bReadOnly
==pCur
->curPagerFlags
);
2097 assert( pCur
==0 || pCur
->iPage
>0 );
2099 if( pgno
>btreePagecount(pBt
) ){
2100 rc
= SQLITE_CORRUPT_BKPT
;
2101 goto getAndInitPage_error
;
2103 rc
= sqlite3PagerGet(pBt
->pPager
, pgno
, (DbPage
**)&pDbPage
, bReadOnly
);
2105 goto getAndInitPage_error
;
2107 *ppPage
= (MemPage
*)sqlite3PagerGetExtra(pDbPage
);
2108 if( (*ppPage
)->isInit
==0 ){
2109 btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2110 rc
= btreeInitPage(*ppPage
);
2111 if( rc
!=SQLITE_OK
){
2112 releasePage(*ppPage
);
2113 goto getAndInitPage_error
;
2116 assert( (*ppPage
)->pgno
==pgno
);
2117 assert( (*ppPage
)->aData
==sqlite3PagerGetData(pDbPage
) );
2119 /* If obtaining a child page for a cursor, we must verify that the page is
2120 ** compatible with the root page. */
2121 if( pCur
&& ((*ppPage
)->nCell
<1 || (*ppPage
)->intKey
!=pCur
->curIntKey
) ){
2122 rc
= SQLITE_CORRUPT_PGNO(pgno
);
2123 releasePage(*ppPage
);
2124 goto getAndInitPage_error
;
2128 getAndInitPage_error
:
2131 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
2133 testcase( pgno
==0 );
2134 assert( pgno
!=0 || rc
==SQLITE_CORRUPT
);
2139 ** Release a MemPage. This should be called once for each prior
2140 ** call to btreeGetPage.
2142 ** Page1 is a special case and must be released using releasePageOne().
2144 static void releasePageNotNull(MemPage
*pPage
){
2145 assert( pPage
->aData
);
2146 assert( pPage
->pBt
);
2147 assert( pPage
->pDbPage
!=0 );
2148 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
2149 assert( sqlite3PagerGetData(pPage
->pDbPage
)==pPage
->aData
);
2150 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2151 sqlite3PagerUnrefNotNull(pPage
->pDbPage
);
2153 static void releasePage(MemPage
*pPage
){
2154 if( pPage
) releasePageNotNull(pPage
);
2156 static void releasePageOne(MemPage
*pPage
){
2158 assert( pPage
->aData
);
2159 assert( pPage
->pBt
);
2160 assert( pPage
->pDbPage
!=0 );
2161 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
2162 assert( sqlite3PagerGetData(pPage
->pDbPage
)==pPage
->aData
);
2163 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2164 sqlite3PagerUnrefPageOne(pPage
->pDbPage
);
2168 ** Get an unused page.
2170 ** This works just like btreeGetPage() with the addition:
2172 ** * If the page is already in use for some other purpose, immediately
2173 ** release it and return an SQLITE_CURRUPT error.
2174 ** * Make sure the isInit flag is clear
2176 static int btreeGetUnusedPage(
2177 BtShared
*pBt
, /* The btree */
2178 Pgno pgno
, /* Number of the page to fetch */
2179 MemPage
**ppPage
, /* Return the page in this parameter */
2180 int flags
/* PAGER_GET_NOCONTENT or PAGER_GET_READONLY */
2182 int rc
= btreeGetPage(pBt
, pgno
, ppPage
, flags
);
2183 if( rc
==SQLITE_OK
){
2184 if( sqlite3PagerPageRefcount((*ppPage
)->pDbPage
)>1 ){
2185 releasePage(*ppPage
);
2187 return SQLITE_CORRUPT_BKPT
;
2189 (*ppPage
)->isInit
= 0;
2198 ** During a rollback, when the pager reloads information into the cache
2199 ** so that the cache is restored to its original state at the start of
2200 ** the transaction, for each page restored this routine is called.
2202 ** This routine needs to reset the extra data section at the end of the
2203 ** page to agree with the restored data.
2205 static void pageReinit(DbPage
*pData
){
2207 pPage
= (MemPage
*)sqlite3PagerGetExtra(pData
);
2208 assert( sqlite3PagerPageRefcount(pData
)>0 );
2209 if( pPage
->isInit
){
2210 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2212 if( sqlite3PagerPageRefcount(pData
)>1 ){
2213 /* pPage might not be a btree page; it might be an overflow page
2214 ** or ptrmap page or a free page. In those cases, the following
2215 ** call to btreeInitPage() will likely return SQLITE_CORRUPT.
2216 ** But no harm is done by this. And it is very important that
2217 ** btreeInitPage() be called on every btree page so we make
2218 ** the call for every page that comes in for re-initing. */
2219 btreeInitPage(pPage
);
2225 ** Invoke the busy handler for a btree.
2227 static int btreeInvokeBusyHandler(void *pArg
){
2228 BtShared
*pBt
= (BtShared
*)pArg
;
2230 assert( sqlite3_mutex_held(pBt
->db
->mutex
) );
2231 return sqlite3InvokeBusyHandler(&pBt
->db
->busyHandler
);
2235 ** Open a database file.
2237 ** zFilename is the name of the database file. If zFilename is NULL
2238 ** then an ephemeral database is created. The ephemeral database might
2239 ** be exclusively in memory, or it might use a disk-based memory cache.
2240 ** Either way, the ephemeral database will be automatically deleted
2241 ** when sqlite3BtreeClose() is called.
2243 ** If zFilename is ":memory:" then an in-memory database is created
2244 ** that is automatically destroyed when it is closed.
2246 ** The "flags" parameter is a bitmask that might contain bits like
2247 ** BTREE_OMIT_JOURNAL and/or BTREE_MEMORY.
2249 ** If the database is already opened in the same database connection
2250 ** and we are in shared cache mode, then the open will fail with an
2251 ** SQLITE_CONSTRAINT error. We cannot allow two or more BtShared
2252 ** objects in the same database connection since doing so will lead
2253 ** to problems with locking.
2255 int sqlite3BtreeOpen(
2256 sqlite3_vfs
*pVfs
, /* VFS to use for this b-tree */
2257 const char *zFilename
, /* Name of the file containing the BTree database */
2258 sqlite3
*db
, /* Associated database handle */
2259 Btree
**ppBtree
, /* Pointer to new Btree object written here */
2260 int flags
, /* Options */
2261 int vfsFlags
/* Flags passed through to sqlite3_vfs.xOpen() */
2263 BtShared
*pBt
= 0; /* Shared part of btree structure */
2264 Btree
*p
; /* Handle to return */
2265 sqlite3_mutex
*mutexOpen
= 0; /* Prevents a race condition. Ticket #3537 */
2266 int rc
= SQLITE_OK
; /* Result code from this function */
2267 u8 nReserve
; /* Byte of unused space on each page */
2268 unsigned char zDbHeader
[100]; /* Database header content */
2270 /* True if opening an ephemeral, temporary database */
2271 const int isTempDb
= zFilename
==0 || zFilename
[0]==0;
2273 /* Set the variable isMemdb to true for an in-memory database, or
2274 ** false for a file-based database.
2276 #ifdef SQLITE_OMIT_MEMORYDB
2277 const int isMemdb
= 0;
2279 const int isMemdb
= (zFilename
&& strcmp(zFilename
, ":memory:")==0)
2280 || (isTempDb
&& sqlite3TempInMemory(db
))
2281 || (vfsFlags
& SQLITE_OPEN_MEMORY
)!=0;
2286 assert( sqlite3_mutex_held(db
->mutex
) );
2287 assert( (flags
&0xff)==flags
); /* flags fit in 8 bits */
2289 /* Only a BTREE_SINGLE database can be BTREE_UNORDERED */
2290 assert( (flags
& BTREE_UNORDERED
)==0 || (flags
& BTREE_SINGLE
)!=0 );
2292 /* A BTREE_SINGLE database is always a temporary and/or ephemeral */
2293 assert( (flags
& BTREE_SINGLE
)==0 || isTempDb
);
2296 flags
|= BTREE_MEMORY
;
2298 if( (vfsFlags
& SQLITE_OPEN_MAIN_DB
)!=0 && (isMemdb
|| isTempDb
) ){
2299 vfsFlags
= (vfsFlags
& ~SQLITE_OPEN_MAIN_DB
) | SQLITE_OPEN_TEMP_DB
;
2301 p
= sqlite3MallocZero(sizeof(Btree
));
2303 return SQLITE_NOMEM_BKPT
;
2305 p
->inTrans
= TRANS_NONE
;
2307 #ifndef SQLITE_OMIT_SHARED_CACHE
2312 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2314 ** If this Btree is a candidate for shared cache, try to find an
2315 ** existing BtShared object that we can share with
2317 if( isTempDb
==0 && (isMemdb
==0 || (vfsFlags
&SQLITE_OPEN_URI
)!=0) ){
2318 if( vfsFlags
& SQLITE_OPEN_SHAREDCACHE
){
2319 int nFilename
= sqlite3Strlen30(zFilename
)+1;
2320 int nFullPathname
= pVfs
->mxPathname
+1;
2321 char *zFullPathname
= sqlite3Malloc(MAX(nFullPathname
,nFilename
));
2322 MUTEX_LOGIC( sqlite3_mutex
*mutexShared
; )
2325 if( !zFullPathname
){
2327 return SQLITE_NOMEM_BKPT
;
2330 memcpy(zFullPathname
, zFilename
, nFilename
);
2332 rc
= sqlite3OsFullPathname(pVfs
, zFilename
,
2333 nFullPathname
, zFullPathname
);
2335 sqlite3_free(zFullPathname
);
2340 #if SQLITE_THREADSAFE
2341 mutexOpen
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_OPEN
);
2342 sqlite3_mutex_enter(mutexOpen
);
2343 mutexShared
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER
);
2344 sqlite3_mutex_enter(mutexShared
);
2346 for(pBt
=GLOBAL(BtShared
*,sqlite3SharedCacheList
); pBt
; pBt
=pBt
->pNext
){
2347 assert( pBt
->nRef
>0 );
2348 if( 0==strcmp(zFullPathname
, sqlite3PagerFilename(pBt
->pPager
, 0))
2349 && sqlite3PagerVfs(pBt
->pPager
)==pVfs
){
2351 for(iDb
=db
->nDb
-1; iDb
>=0; iDb
--){
2352 Btree
*pExisting
= db
->aDb
[iDb
].pBt
;
2353 if( pExisting
&& pExisting
->pBt
==pBt
){
2354 sqlite3_mutex_leave(mutexShared
);
2355 sqlite3_mutex_leave(mutexOpen
);
2356 sqlite3_free(zFullPathname
);
2358 return SQLITE_CONSTRAINT
;
2366 sqlite3_mutex_leave(mutexShared
);
2367 sqlite3_free(zFullPathname
);
2371 /* In debug mode, we mark all persistent databases as sharable
2372 ** even when they are not. This exercises the locking code and
2373 ** gives more opportunity for asserts(sqlite3_mutex_held())
2374 ** statements to find locking problems.
2383 ** The following asserts make sure that structures used by the btree are
2384 ** the right size. This is to guard against size changes that result
2385 ** when compiling on a different architecture.
2387 assert( sizeof(i64
)==8 );
2388 assert( sizeof(u64
)==8 );
2389 assert( sizeof(u32
)==4 );
2390 assert( sizeof(u16
)==2 );
2391 assert( sizeof(Pgno
)==4 );
2393 pBt
= sqlite3MallocZero( sizeof(*pBt
) );
2395 rc
= SQLITE_NOMEM_BKPT
;
2396 goto btree_open_out
;
2398 rc
= sqlite3PagerOpen(pVfs
, &pBt
->pPager
, zFilename
,
2399 sizeof(MemPage
), flags
, vfsFlags
, pageReinit
);
2400 if( rc
==SQLITE_OK
){
2401 sqlite3PagerSetMmapLimit(pBt
->pPager
, db
->szMmap
);
2402 rc
= sqlite3PagerReadFileheader(pBt
->pPager
,sizeof(zDbHeader
),zDbHeader
);
2404 if( rc
!=SQLITE_OK
){
2405 goto btree_open_out
;
2407 pBt
->openFlags
= (u8
)flags
;
2409 sqlite3PagerSetBusyhandler(pBt
->pPager
, btreeInvokeBusyHandler
, pBt
);
2414 if( sqlite3PagerIsreadonly(pBt
->pPager
) ) pBt
->btsFlags
|= BTS_READ_ONLY
;
2415 #if defined(SQLITE_SECURE_DELETE)
2416 pBt
->btsFlags
|= BTS_SECURE_DELETE
;
2417 #elif defined(SQLITE_FAST_SECURE_DELETE)
2418 pBt
->btsFlags
|= BTS_OVERWRITE
;
2420 /* EVIDENCE-OF: R-51873-39618 The page size for a database file is
2421 ** determined by the 2-byte integer located at an offset of 16 bytes from
2422 ** the beginning of the database file. */
2423 pBt
->pageSize
= (zDbHeader
[16]<<8) | (zDbHeader
[17]<<16);
2424 if( pBt
->pageSize
<512 || pBt
->pageSize
>SQLITE_MAX_PAGE_SIZE
2425 || ((pBt
->pageSize
-1)&pBt
->pageSize
)!=0 ){
2427 #ifndef SQLITE_OMIT_AUTOVACUUM
2428 /* If the magic name ":memory:" will create an in-memory database, then
2429 ** leave the autoVacuum mode at 0 (do not auto-vacuum), even if
2430 ** SQLITE_DEFAULT_AUTOVACUUM is true. On the other hand, if
2431 ** SQLITE_OMIT_MEMORYDB has been defined, then ":memory:" is just a
2432 ** regular file-name. In this case the auto-vacuum applies as per normal.
2434 if( zFilename
&& !isMemdb
){
2435 pBt
->autoVacuum
= (SQLITE_DEFAULT_AUTOVACUUM
? 1 : 0);
2436 pBt
->incrVacuum
= (SQLITE_DEFAULT_AUTOVACUUM
==2 ? 1 : 0);
2441 /* EVIDENCE-OF: R-37497-42412 The size of the reserved region is
2442 ** determined by the one-byte unsigned integer found at an offset of 20
2443 ** into the database file header. */
2444 nReserve
= zDbHeader
[20];
2445 pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
2446 #ifndef SQLITE_OMIT_AUTOVACUUM
2447 pBt
->autoVacuum
= (get4byte(&zDbHeader
[36 + 4*4])?1:0);
2448 pBt
->incrVacuum
= (get4byte(&zDbHeader
[36 + 7*4])?1:0);
2451 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
, nReserve
);
2452 if( rc
) goto btree_open_out
;
2453 pBt
->usableSize
= pBt
->pageSize
- nReserve
;
2454 assert( (pBt
->pageSize
& 7)==0 ); /* 8-byte alignment of pageSize */
2456 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2457 /* Add the new BtShared object to the linked list sharable BtShareds.
2461 MUTEX_LOGIC( sqlite3_mutex
*mutexShared
; )
2462 MUTEX_LOGIC( mutexShared
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER
);)
2463 if( SQLITE_THREADSAFE
&& sqlite3GlobalConfig
.bCoreMutex
){
2464 pBt
->mutex
= sqlite3MutexAlloc(SQLITE_MUTEX_FAST
);
2465 if( pBt
->mutex
==0 ){
2466 rc
= SQLITE_NOMEM_BKPT
;
2467 goto btree_open_out
;
2470 sqlite3_mutex_enter(mutexShared
);
2471 pBt
->pNext
= GLOBAL(BtShared
*,sqlite3SharedCacheList
);
2472 GLOBAL(BtShared
*,sqlite3SharedCacheList
) = pBt
;
2473 sqlite3_mutex_leave(mutexShared
);
2478 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2479 /* If the new Btree uses a sharable pBtShared, then link the new
2480 ** Btree into the list of all sharable Btrees for the same connection.
2481 ** The list is kept in ascending order by pBt address.
2486 for(i
=0; i
<db
->nDb
; i
++){
2487 if( (pSib
= db
->aDb
[i
].pBt
)!=0 && pSib
->sharable
){
2488 while( pSib
->pPrev
){ pSib
= pSib
->pPrev
; }
2489 if( (uptr
)p
->pBt
<(uptr
)pSib
->pBt
){
2494 while( pSib
->pNext
&& (uptr
)pSib
->pNext
->pBt
<(uptr
)p
->pBt
){
2497 p
->pNext
= pSib
->pNext
;
2500 p
->pNext
->pPrev
= p
;
2512 if( rc
!=SQLITE_OK
){
2513 if( pBt
&& pBt
->pPager
){
2514 sqlite3PagerClose(pBt
->pPager
, 0);
2520 sqlite3_file
*pFile
;
2522 /* If the B-Tree was successfully opened, set the pager-cache size to the
2523 ** default value. Except, when opening on an existing shared pager-cache,
2524 ** do not change the pager-cache size.
2526 if( sqlite3BtreeSchema(p
, 0, 0)==0 ){
2527 sqlite3PagerSetCachesize(p
->pBt
->pPager
, SQLITE_DEFAULT_CACHE_SIZE
);
2530 pFile
= sqlite3PagerFile(pBt
->pPager
);
2531 if( pFile
->pMethods
){
2532 sqlite3OsFileControlHint(pFile
, SQLITE_FCNTL_PDB
, (void*)&pBt
->db
);
2536 assert( sqlite3_mutex_held(mutexOpen
) );
2537 sqlite3_mutex_leave(mutexOpen
);
2539 assert( rc
!=SQLITE_OK
|| sqlite3BtreeConnectionCount(*ppBtree
)>0 );
2544 ** Decrement the BtShared.nRef counter. When it reaches zero,
2545 ** remove the BtShared structure from the sharing list. Return
2546 ** true if the BtShared.nRef counter reaches zero and return
2547 ** false if it is still positive.
2549 static int removeFromSharingList(BtShared
*pBt
){
2550 #ifndef SQLITE_OMIT_SHARED_CACHE
2551 MUTEX_LOGIC( sqlite3_mutex
*pMaster
; )
2555 assert( sqlite3_mutex_notheld(pBt
->mutex
) );
2556 MUTEX_LOGIC( pMaster
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER
); )
2557 sqlite3_mutex_enter(pMaster
);
2560 if( GLOBAL(BtShared
*,sqlite3SharedCacheList
)==pBt
){
2561 GLOBAL(BtShared
*,sqlite3SharedCacheList
) = pBt
->pNext
;
2563 pList
= GLOBAL(BtShared
*,sqlite3SharedCacheList
);
2564 while( ALWAYS(pList
) && pList
->pNext
!=pBt
){
2567 if( ALWAYS(pList
) ){
2568 pList
->pNext
= pBt
->pNext
;
2571 if( SQLITE_THREADSAFE
){
2572 sqlite3_mutex_free(pBt
->mutex
);
2576 sqlite3_mutex_leave(pMaster
);
2584 ** Make sure pBt->pTmpSpace points to an allocation of
2585 ** MX_CELL_SIZE(pBt) bytes with a 4-byte prefix for a left-child
2588 static void allocateTempSpace(BtShared
*pBt
){
2589 if( !pBt
->pTmpSpace
){
2590 pBt
->pTmpSpace
= sqlite3PageMalloc( pBt
->pageSize
);
2592 /* One of the uses of pBt->pTmpSpace is to format cells before
2593 ** inserting them into a leaf page (function fillInCell()). If
2594 ** a cell is less than 4 bytes in size, it is rounded up to 4 bytes
2595 ** by the various routines that manipulate binary cells. Which
2596 ** can mean that fillInCell() only initializes the first 2 or 3
2597 ** bytes of pTmpSpace, but that the first 4 bytes are copied from
2598 ** it into a database page. This is not actually a problem, but it
2599 ** does cause a valgrind error when the 1 or 2 bytes of unitialized
2600 ** data is passed to system call write(). So to avoid this error,
2601 ** zero the first 4 bytes of temp space here.
2603 ** Also: Provide four bytes of initialized space before the
2604 ** beginning of pTmpSpace as an area available to prepend the
2605 ** left-child pointer to the beginning of a cell.
2607 if( pBt
->pTmpSpace
){
2608 memset(pBt
->pTmpSpace
, 0, 8);
2609 pBt
->pTmpSpace
+= 4;
2615 ** Free the pBt->pTmpSpace allocation
2617 static void freeTempSpace(BtShared
*pBt
){
2618 if( pBt
->pTmpSpace
){
2619 pBt
->pTmpSpace
-= 4;
2620 sqlite3PageFree(pBt
->pTmpSpace
);
2626 ** Close an open database and invalidate all cursors.
2628 int sqlite3BtreeClose(Btree
*p
){
2629 BtShared
*pBt
= p
->pBt
;
2632 /* Close all cursors opened via this handle. */
2633 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2634 sqlite3BtreeEnter(p
);
2635 pCur
= pBt
->pCursor
;
2637 BtCursor
*pTmp
= pCur
;
2639 if( pTmp
->pBtree
==p
){
2640 sqlite3BtreeCloseCursor(pTmp
);
2644 /* Rollback any active transaction and free the handle structure.
2645 ** The call to sqlite3BtreeRollback() drops any table-locks held by
2648 sqlite3BtreeRollback(p
, SQLITE_OK
, 0);
2649 sqlite3BtreeLeave(p
);
2651 /* If there are still other outstanding references to the shared-btree
2652 ** structure, return now. The remainder of this procedure cleans
2653 ** up the shared-btree.
2655 assert( p
->wantToLock
==0 && p
->locked
==0 );
2656 if( !p
->sharable
|| removeFromSharingList(pBt
) ){
2657 /* The pBt is no longer on the sharing list, so we can access
2658 ** it without having to hold the mutex.
2660 ** Clean out and delete the BtShared object.
2662 assert( !pBt
->pCursor
);
2663 sqlite3PagerClose(pBt
->pPager
, p
->db
);
2664 if( pBt
->xFreeSchema
&& pBt
->pSchema
){
2665 pBt
->xFreeSchema(pBt
->pSchema
);
2667 sqlite3DbFree(0, pBt
->pSchema
);
2672 #ifndef SQLITE_OMIT_SHARED_CACHE
2673 assert( p
->wantToLock
==0 );
2674 assert( p
->locked
==0 );
2675 if( p
->pPrev
) p
->pPrev
->pNext
= p
->pNext
;
2676 if( p
->pNext
) p
->pNext
->pPrev
= p
->pPrev
;
2684 ** Change the "soft" limit on the number of pages in the cache.
2685 ** Unused and unmodified pages will be recycled when the number of
2686 ** pages in the cache exceeds this soft limit. But the size of the
2687 ** cache is allowed to grow larger than this limit if it contains
2688 ** dirty pages or pages still in active use.
2690 int sqlite3BtreeSetCacheSize(Btree
*p
, int mxPage
){
2691 BtShared
*pBt
= p
->pBt
;
2692 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2693 sqlite3BtreeEnter(p
);
2694 sqlite3PagerSetCachesize(pBt
->pPager
, mxPage
);
2695 sqlite3BtreeLeave(p
);
2700 ** Change the "spill" limit on the number of pages in the cache.
2701 ** If the number of pages exceeds this limit during a write transaction,
2702 ** the pager might attempt to "spill" pages to the journal early in
2703 ** order to free up memory.
2705 ** The value returned is the current spill size. If zero is passed
2706 ** as an argument, no changes are made to the spill size setting, so
2707 ** using mxPage of 0 is a way to query the current spill size.
2709 int sqlite3BtreeSetSpillSize(Btree
*p
, int mxPage
){
2710 BtShared
*pBt
= p
->pBt
;
2712 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2713 sqlite3BtreeEnter(p
);
2714 res
= sqlite3PagerSetSpillsize(pBt
->pPager
, mxPage
);
2715 sqlite3BtreeLeave(p
);
2719 #if SQLITE_MAX_MMAP_SIZE>0
2721 ** Change the limit on the amount of the database file that may be
2724 int sqlite3BtreeSetMmapLimit(Btree
*p
, sqlite3_int64 szMmap
){
2725 BtShared
*pBt
= p
->pBt
;
2726 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2727 sqlite3BtreeEnter(p
);
2728 sqlite3PagerSetMmapLimit(pBt
->pPager
, szMmap
);
2729 sqlite3BtreeLeave(p
);
2732 #endif /* SQLITE_MAX_MMAP_SIZE>0 */
2735 ** Change the way data is synced to disk in order to increase or decrease
2736 ** how well the database resists damage due to OS crashes and power
2737 ** failures. Level 1 is the same as asynchronous (no syncs() occur and
2738 ** there is a high probability of damage) Level 2 is the default. There
2739 ** is a very low but non-zero probability of damage. Level 3 reduces the
2740 ** probability of damage to near zero but with a write performance reduction.
2742 #ifndef SQLITE_OMIT_PAGER_PRAGMAS
2743 int sqlite3BtreeSetPagerFlags(
2744 Btree
*p
, /* The btree to set the safety level on */
2745 unsigned pgFlags
/* Various PAGER_* flags */
2747 BtShared
*pBt
= p
->pBt
;
2748 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2749 sqlite3BtreeEnter(p
);
2750 sqlite3PagerSetFlags(pBt
->pPager
, pgFlags
);
2751 sqlite3BtreeLeave(p
);
2757 ** Change the default pages size and the number of reserved bytes per page.
2758 ** Or, if the page size has already been fixed, return SQLITE_READONLY
2759 ** without changing anything.
2761 ** The page size must be a power of 2 between 512 and 65536. If the page
2762 ** size supplied does not meet this constraint then the page size is not
2765 ** Page sizes are constrained to be a power of two so that the region
2766 ** of the database file used for locking (beginning at PENDING_BYTE,
2767 ** the first byte past the 1GB boundary, 0x40000000) needs to occur
2768 ** at the beginning of a page.
2770 ** If parameter nReserve is less than zero, then the number of reserved
2771 ** bytes per page is left unchanged.
2773 ** If the iFix!=0 then the BTS_PAGESIZE_FIXED flag is set so that the page size
2774 ** and autovacuum mode can no longer be changed.
2776 int sqlite3BtreeSetPageSize(Btree
*p
, int pageSize
, int nReserve
, int iFix
){
2778 BtShared
*pBt
= p
->pBt
;
2779 assert( nReserve
>=-1 && nReserve
<=255 );
2780 sqlite3BtreeEnter(p
);
2781 #if SQLITE_HAS_CODEC
2782 if( nReserve
>pBt
->optimalReserve
) pBt
->optimalReserve
= (u8
)nReserve
;
2784 if( pBt
->btsFlags
& BTS_PAGESIZE_FIXED
){
2785 sqlite3BtreeLeave(p
);
2786 return SQLITE_READONLY
;
2789 nReserve
= pBt
->pageSize
- pBt
->usableSize
;
2791 assert( nReserve
>=0 && nReserve
<=255 );
2792 if( pageSize
>=512 && pageSize
<=SQLITE_MAX_PAGE_SIZE
&&
2793 ((pageSize
-1)&pageSize
)==0 ){
2794 assert( (pageSize
& 7)==0 );
2795 assert( !pBt
->pCursor
);
2796 pBt
->pageSize
= (u32
)pageSize
;
2799 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
, nReserve
);
2800 pBt
->usableSize
= pBt
->pageSize
- (u16
)nReserve
;
2801 if( iFix
) pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
2802 sqlite3BtreeLeave(p
);
2807 ** Return the currently defined page size
2809 int sqlite3BtreeGetPageSize(Btree
*p
){
2810 return p
->pBt
->pageSize
;
2814 ** This function is similar to sqlite3BtreeGetReserve(), except that it
2815 ** may only be called if it is guaranteed that the b-tree mutex is already
2818 ** This is useful in one special case in the backup API code where it is
2819 ** known that the shared b-tree mutex is held, but the mutex on the
2820 ** database handle that owns *p is not. In this case if sqlite3BtreeEnter()
2821 ** were to be called, it might collide with some other operation on the
2822 ** database handle that owns *p, causing undefined behavior.
2824 int sqlite3BtreeGetReserveNoMutex(Btree
*p
){
2826 assert( sqlite3_mutex_held(p
->pBt
->mutex
) );
2827 n
= p
->pBt
->pageSize
- p
->pBt
->usableSize
;
2832 ** Return the number of bytes of space at the end of every page that
2833 ** are intentually left unused. This is the "reserved" space that is
2834 ** sometimes used by extensions.
2836 ** If SQLITE_HAS_MUTEX is defined then the number returned is the
2837 ** greater of the current reserved space and the maximum requested
2840 int sqlite3BtreeGetOptimalReserve(Btree
*p
){
2842 sqlite3BtreeEnter(p
);
2843 n
= sqlite3BtreeGetReserveNoMutex(p
);
2844 #ifdef SQLITE_HAS_CODEC
2845 if( n
<p
->pBt
->optimalReserve
) n
= p
->pBt
->optimalReserve
;
2847 sqlite3BtreeLeave(p
);
2853 ** Set the maximum page count for a database if mxPage is positive.
2854 ** No changes are made if mxPage is 0 or negative.
2855 ** Regardless of the value of mxPage, return the maximum page count.
2857 int sqlite3BtreeMaxPageCount(Btree
*p
, int mxPage
){
2859 sqlite3BtreeEnter(p
);
2860 n
= sqlite3PagerMaxPageCount(p
->pBt
->pPager
, mxPage
);
2861 sqlite3BtreeLeave(p
);
2866 ** Change the values for the BTS_SECURE_DELETE and BTS_OVERWRITE flags:
2868 ** newFlag==0 Both BTS_SECURE_DELETE and BTS_OVERWRITE are cleared
2869 ** newFlag==1 BTS_SECURE_DELETE set and BTS_OVERWRITE is cleared
2870 ** newFlag==2 BTS_SECURE_DELETE cleared and BTS_OVERWRITE is set
2871 ** newFlag==(-1) No changes
2873 ** This routine acts as a query if newFlag is less than zero
2875 ** With BTS_OVERWRITE set, deleted content is overwritten by zeros, but
2876 ** freelist leaf pages are not written back to the database. Thus in-page
2877 ** deleted content is cleared, but freelist deleted content is not.
2879 ** With BTS_SECURE_DELETE, operation is like BTS_OVERWRITE with the addition
2880 ** that freelist leaf pages are written back into the database, increasing
2881 ** the amount of disk I/O.
2883 int sqlite3BtreeSecureDelete(Btree
*p
, int newFlag
){
2885 if( p
==0 ) return 0;
2886 sqlite3BtreeEnter(p
);
2887 assert( BTS_OVERWRITE
==BTS_SECURE_DELETE
*2 );
2888 assert( BTS_FAST_SECURE
==(BTS_OVERWRITE
|BTS_SECURE_DELETE
) );
2890 p
->pBt
->btsFlags
&= ~BTS_FAST_SECURE
;
2891 p
->pBt
->btsFlags
|= BTS_SECURE_DELETE
*newFlag
;
2893 b
= (p
->pBt
->btsFlags
& BTS_FAST_SECURE
)/BTS_SECURE_DELETE
;
2894 sqlite3BtreeLeave(p
);
2899 ** Change the 'auto-vacuum' property of the database. If the 'autoVacuum'
2900 ** parameter is non-zero, then auto-vacuum mode is enabled. If zero, it
2901 ** is disabled. The default value for the auto-vacuum property is
2902 ** determined by the SQLITE_DEFAULT_AUTOVACUUM macro.
2904 int sqlite3BtreeSetAutoVacuum(Btree
*p
, int autoVacuum
){
2905 #ifdef SQLITE_OMIT_AUTOVACUUM
2906 return SQLITE_READONLY
;
2908 BtShared
*pBt
= p
->pBt
;
2910 u8 av
= (u8
)autoVacuum
;
2912 sqlite3BtreeEnter(p
);
2913 if( (pBt
->btsFlags
& BTS_PAGESIZE_FIXED
)!=0 && (av
?1:0)!=pBt
->autoVacuum
){
2914 rc
= SQLITE_READONLY
;
2916 pBt
->autoVacuum
= av
?1:0;
2917 pBt
->incrVacuum
= av
==2 ?1:0;
2919 sqlite3BtreeLeave(p
);
2925 ** Return the value of the 'auto-vacuum' property. If auto-vacuum is
2926 ** enabled 1 is returned. Otherwise 0.
2928 int sqlite3BtreeGetAutoVacuum(Btree
*p
){
2929 #ifdef SQLITE_OMIT_AUTOVACUUM
2930 return BTREE_AUTOVACUUM_NONE
;
2933 sqlite3BtreeEnter(p
);
2935 (!p
->pBt
->autoVacuum
)?BTREE_AUTOVACUUM_NONE
:
2936 (!p
->pBt
->incrVacuum
)?BTREE_AUTOVACUUM_FULL
:
2937 BTREE_AUTOVACUUM_INCR
2939 sqlite3BtreeLeave(p
);
2945 ** If the user has not set the safety-level for this database connection
2946 ** using "PRAGMA synchronous", and if the safety-level is not already
2947 ** set to the value passed to this function as the second parameter,
2950 #if SQLITE_DEFAULT_SYNCHRONOUS!=SQLITE_DEFAULT_WAL_SYNCHRONOUS \
2951 && !defined(SQLITE_OMIT_WAL)
2952 static void setDefaultSyncFlag(BtShared
*pBt
, u8 safety_level
){
2955 if( (db
=pBt
->db
)!=0 && (pDb
=db
->aDb
)!=0 ){
2956 while( pDb
->pBt
==0 || pDb
->pBt
->pBt
!=pBt
){ pDb
++; }
2957 if( pDb
->bSyncSet
==0
2958 && pDb
->safety_level
!=safety_level
2961 pDb
->safety_level
= safety_level
;
2962 sqlite3PagerSetFlags(pBt
->pPager
,
2963 pDb
->safety_level
| (db
->flags
& PAGER_FLAGS_MASK
));
2968 # define setDefaultSyncFlag(pBt,safety_level)
2972 ** Get a reference to pPage1 of the database file. This will
2973 ** also acquire a readlock on that file.
2975 ** SQLITE_OK is returned on success. If the file is not a
2976 ** well-formed database file, then SQLITE_CORRUPT is returned.
2977 ** SQLITE_BUSY is returned if the database is locked. SQLITE_NOMEM
2978 ** is returned if we run out of memory.
2980 static int lockBtree(BtShared
*pBt
){
2981 int rc
; /* Result code from subfunctions */
2982 MemPage
*pPage1
; /* Page 1 of the database file */
2983 int nPage
; /* Number of pages in the database */
2984 int nPageFile
= 0; /* Number of pages in the database file */
2985 int nPageHeader
; /* Number of pages in the database according to hdr */
2987 assert( sqlite3_mutex_held(pBt
->mutex
) );
2988 assert( pBt
->pPage1
==0 );
2989 rc
= sqlite3PagerSharedLock(pBt
->pPager
);
2990 if( rc
!=SQLITE_OK
) return rc
;
2991 rc
= btreeGetPage(pBt
, 1, &pPage1
, 0);
2992 if( rc
!=SQLITE_OK
) return rc
;
2994 /* Do some checking to help insure the file we opened really is
2995 ** a valid database file.
2997 nPage
= nPageHeader
= get4byte(28+(u8
*)pPage1
->aData
);
2998 sqlite3PagerPagecount(pBt
->pPager
, &nPageFile
);
2999 if( nPage
==0 || memcmp(24+(u8
*)pPage1
->aData
, 92+(u8
*)pPage1
->aData
,4)!=0 ){
3005 u8
*page1
= pPage1
->aData
;
3007 /* EVIDENCE-OF: R-43737-39999 Every valid SQLite database file begins
3008 ** with the following 16 bytes (in hex): 53 51 4c 69 74 65 20 66 6f 72 6d
3009 ** 61 74 20 33 00. */
3010 if( memcmp(page1
, zMagicHeader
, 16)!=0 ){
3011 goto page1_init_failed
;
3014 #ifdef SQLITE_OMIT_WAL
3016 pBt
->btsFlags
|= BTS_READ_ONLY
;
3019 goto page1_init_failed
;
3023 pBt
->btsFlags
|= BTS_READ_ONLY
;
3026 goto page1_init_failed
;
3029 /* If the write version is set to 2, this database should be accessed
3030 ** in WAL mode. If the log is not already open, open it now. Then
3031 ** return SQLITE_OK and return without populating BtShared.pPage1.
3032 ** The caller detects this and calls this function again. This is
3033 ** required as the version of page 1 currently in the page1 buffer
3034 ** may not be the latest version - there may be a newer one in the log
3037 if( page1
[19]==2 && (pBt
->btsFlags
& BTS_NO_WAL
)==0 ){
3039 rc
= sqlite3PagerOpenWal(pBt
->pPager
, &isOpen
);
3040 if( rc
!=SQLITE_OK
){
3041 goto page1_init_failed
;
3043 setDefaultSyncFlag(pBt
, SQLITE_DEFAULT_WAL_SYNCHRONOUS
+1);
3045 releasePageOne(pPage1
);
3051 setDefaultSyncFlag(pBt
, SQLITE_DEFAULT_SYNCHRONOUS
+1);
3055 /* EVIDENCE-OF: R-15465-20813 The maximum and minimum embedded payload
3056 ** fractions and the leaf payload fraction values must be 64, 32, and 32.
3058 ** The original design allowed these amounts to vary, but as of
3059 ** version 3.6.0, we require them to be fixed.
3061 if( memcmp(&page1
[21], "\100\040\040",3)!=0 ){
3062 goto page1_init_failed
;
3064 /* EVIDENCE-OF: R-51873-39618 The page size for a database file is
3065 ** determined by the 2-byte integer located at an offset of 16 bytes from
3066 ** the beginning of the database file. */
3067 pageSize
= (page1
[16]<<8) | (page1
[17]<<16);
3068 /* EVIDENCE-OF: R-25008-21688 The size of a page is a power of two
3069 ** between 512 and 65536 inclusive. */
3070 if( ((pageSize
-1)&pageSize
)!=0
3071 || pageSize
>SQLITE_MAX_PAGE_SIZE
3074 goto page1_init_failed
;
3076 assert( (pageSize
& 7)==0 );
3077 /* EVIDENCE-OF: R-59310-51205 The "reserved space" size in the 1-byte
3078 ** integer at offset 20 is the number of bytes of space at the end of
3079 ** each page to reserve for extensions.
3081 ** EVIDENCE-OF: R-37497-42412 The size of the reserved region is
3082 ** determined by the one-byte unsigned integer found at an offset of 20
3083 ** into the database file header. */
3084 usableSize
= pageSize
- page1
[20];
3085 if( (u32
)pageSize
!=pBt
->pageSize
){
3086 /* After reading the first page of the database assuming a page size
3087 ** of BtShared.pageSize, we have discovered that the page-size is
3088 ** actually pageSize. Unlock the database, leave pBt->pPage1 at
3089 ** zero and return SQLITE_OK. The caller will call this function
3090 ** again with the correct page-size.
3092 releasePageOne(pPage1
);
3093 pBt
->usableSize
= usableSize
;
3094 pBt
->pageSize
= pageSize
;
3096 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
,
3097 pageSize
-usableSize
);
3100 if( (pBt
->db
->flags
& SQLITE_WriteSchema
)==0 && nPage
>nPageFile
){
3101 rc
= SQLITE_CORRUPT_BKPT
;
3102 goto page1_init_failed
;
3104 /* EVIDENCE-OF: R-28312-64704 However, the usable size is not allowed to
3105 ** be less than 480. In other words, if the page size is 512, then the
3106 ** reserved space size cannot exceed 32. */
3107 if( usableSize
<480 ){
3108 goto page1_init_failed
;
3110 pBt
->pageSize
= pageSize
;
3111 pBt
->usableSize
= usableSize
;
3112 #ifndef SQLITE_OMIT_AUTOVACUUM
3113 pBt
->autoVacuum
= (get4byte(&page1
[36 + 4*4])?1:0);
3114 pBt
->incrVacuum
= (get4byte(&page1
[36 + 7*4])?1:0);
3118 /* maxLocal is the maximum amount of payload to store locally for
3119 ** a cell. Make sure it is small enough so that at least minFanout
3120 ** cells can will fit on one page. We assume a 10-byte page header.
3121 ** Besides the payload, the cell must store:
3122 ** 2-byte pointer to the cell
3123 ** 4-byte child pointer
3124 ** 9-byte nKey value
3125 ** 4-byte nData value
3126 ** 4-byte overflow page pointer
3127 ** So a cell consists of a 2-byte pointer, a header which is as much as
3128 ** 17 bytes long, 0 to N bytes of payload, and an optional 4 byte overflow
3131 pBt
->maxLocal
= (u16
)((pBt
->usableSize
-12)*64/255 - 23);
3132 pBt
->minLocal
= (u16
)((pBt
->usableSize
-12)*32/255 - 23);
3133 pBt
->maxLeaf
= (u16
)(pBt
->usableSize
- 35);
3134 pBt
->minLeaf
= (u16
)((pBt
->usableSize
-12)*32/255 - 23);
3135 if( pBt
->maxLocal
>127 ){
3136 pBt
->max1bytePayload
= 127;
3138 pBt
->max1bytePayload
= (u8
)pBt
->maxLocal
;
3140 assert( pBt
->maxLeaf
+ 23 <= MX_CELL_SIZE(pBt
) );
3141 pBt
->pPage1
= pPage1
;
3146 releasePageOne(pPage1
);
3153 ** Return the number of cursors open on pBt. This is for use
3154 ** in assert() expressions, so it is only compiled if NDEBUG is not
3157 ** Only write cursors are counted if wrOnly is true. If wrOnly is
3158 ** false then all cursors are counted.
3160 ** For the purposes of this routine, a cursor is any cursor that
3161 ** is capable of reading or writing to the database. Cursors that
3162 ** have been tripped into the CURSOR_FAULT state are not counted.
3164 static int countValidCursors(BtShared
*pBt
, int wrOnly
){
3167 for(pCur
=pBt
->pCursor
; pCur
; pCur
=pCur
->pNext
){
3168 if( (wrOnly
==0 || (pCur
->curFlags
& BTCF_WriteFlag
)!=0)
3169 && pCur
->eState
!=CURSOR_FAULT
) r
++;
3176 ** If there are no outstanding cursors and we are not in the middle
3177 ** of a transaction but there is a read lock on the database, then
3178 ** this routine unrefs the first page of the database file which
3179 ** has the effect of releasing the read lock.
3181 ** If there is a transaction in progress, this routine is a no-op.
3183 static void unlockBtreeIfUnused(BtShared
*pBt
){
3184 assert( sqlite3_mutex_held(pBt
->mutex
) );
3185 assert( countValidCursors(pBt
,0)==0 || pBt
->inTransaction
>TRANS_NONE
);
3186 if( pBt
->inTransaction
==TRANS_NONE
&& pBt
->pPage1
!=0 ){
3187 MemPage
*pPage1
= pBt
->pPage1
;
3188 assert( pPage1
->aData
);
3189 assert( sqlite3PagerRefcount(pBt
->pPager
)==1 );
3191 releasePageOne(pPage1
);
3196 ** If pBt points to an empty file then convert that empty file
3197 ** into a new empty database by initializing the first page of
3200 static int newDatabase(BtShared
*pBt
){
3202 unsigned char *data
;
3205 assert( sqlite3_mutex_held(pBt
->mutex
) );
3212 rc
= sqlite3PagerWrite(pP1
->pDbPage
);
3214 memcpy(data
, zMagicHeader
, sizeof(zMagicHeader
));
3215 assert( sizeof(zMagicHeader
)==16 );
3216 data
[16] = (u8
)((pBt
->pageSize
>>8)&0xff);
3217 data
[17] = (u8
)((pBt
->pageSize
>>16)&0xff);
3220 assert( pBt
->usableSize
<=pBt
->pageSize
&& pBt
->usableSize
+255>=pBt
->pageSize
);
3221 data
[20] = (u8
)(pBt
->pageSize
- pBt
->usableSize
);
3225 memset(&data
[24], 0, 100-24);
3226 zeroPage(pP1
, PTF_INTKEY
|PTF_LEAF
|PTF_LEAFDATA
);
3227 pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
3228 #ifndef SQLITE_OMIT_AUTOVACUUM
3229 assert( pBt
->autoVacuum
==1 || pBt
->autoVacuum
==0 );
3230 assert( pBt
->incrVacuum
==1 || pBt
->incrVacuum
==0 );
3231 put4byte(&data
[36 + 4*4], pBt
->autoVacuum
);
3232 put4byte(&data
[36 + 7*4], pBt
->incrVacuum
);
3240 ** Initialize the first page of the database file (creating a database
3241 ** consisting of a single page and no schema objects). Return SQLITE_OK
3242 ** if successful, or an SQLite error code otherwise.
3244 int sqlite3BtreeNewDb(Btree
*p
){
3246 sqlite3BtreeEnter(p
);
3248 rc
= newDatabase(p
->pBt
);
3249 sqlite3BtreeLeave(p
);
3254 ** Attempt to start a new transaction. A write-transaction
3255 ** is started if the second argument is nonzero, otherwise a read-
3256 ** transaction. If the second argument is 2 or more and exclusive
3257 ** transaction is started, meaning that no other process is allowed
3258 ** to access the database. A preexisting transaction may not be
3259 ** upgraded to exclusive by calling this routine a second time - the
3260 ** exclusivity flag only works for a new transaction.
3262 ** A write-transaction must be started before attempting any
3263 ** changes to the database. None of the following routines
3264 ** will work unless a transaction is started first:
3266 ** sqlite3BtreeCreateTable()
3267 ** sqlite3BtreeCreateIndex()
3268 ** sqlite3BtreeClearTable()
3269 ** sqlite3BtreeDropTable()
3270 ** sqlite3BtreeInsert()
3271 ** sqlite3BtreeDelete()
3272 ** sqlite3BtreeUpdateMeta()
3274 ** If an initial attempt to acquire the lock fails because of lock contention
3275 ** and the database was previously unlocked, then invoke the busy handler
3276 ** if there is one. But if there was previously a read-lock, do not
3277 ** invoke the busy handler - just return SQLITE_BUSY. SQLITE_BUSY is
3278 ** returned when there is already a read-lock in order to avoid a deadlock.
3280 ** Suppose there are two processes A and B. A has a read lock and B has
3281 ** a reserved lock. B tries to promote to exclusive but is blocked because
3282 ** of A's read lock. A tries to promote to reserved but is blocked by B.
3283 ** One or the other of the two processes must give way or there can be
3284 ** no progress. By returning SQLITE_BUSY and not invoking the busy callback
3285 ** when A already has a read lock, we encourage A to give up and let B
3288 int sqlite3BtreeBeginTrans(Btree
*p
, int wrflag
){
3289 BtShared
*pBt
= p
->pBt
;
3292 sqlite3BtreeEnter(p
);
3295 /* If the btree is already in a write-transaction, or it
3296 ** is already in a read-transaction and a read-transaction
3297 ** is requested, this is a no-op.
3299 if( p
->inTrans
==TRANS_WRITE
|| (p
->inTrans
==TRANS_READ
&& !wrflag
) ){
3302 assert( pBt
->inTransaction
==TRANS_WRITE
|| IfNotOmitAV(pBt
->bDoTruncate
)==0 );
3304 /* Write transactions are not possible on a read-only database */
3305 if( (pBt
->btsFlags
& BTS_READ_ONLY
)!=0 && wrflag
){
3306 rc
= SQLITE_READONLY
;
3310 #ifndef SQLITE_OMIT_SHARED_CACHE
3312 sqlite3
*pBlock
= 0;
3313 /* If another database handle has already opened a write transaction
3314 ** on this shared-btree structure and a second write transaction is
3315 ** requested, return SQLITE_LOCKED.
3317 if( (wrflag
&& pBt
->inTransaction
==TRANS_WRITE
)
3318 || (pBt
->btsFlags
& BTS_PENDING
)!=0
3320 pBlock
= pBt
->pWriter
->db
;
3321 }else if( wrflag
>1 ){
3323 for(pIter
=pBt
->pLock
; pIter
; pIter
=pIter
->pNext
){
3324 if( pIter
->pBtree
!=p
){
3325 pBlock
= pIter
->pBtree
->db
;
3331 sqlite3ConnectionBlocked(p
->db
, pBlock
);
3332 rc
= SQLITE_LOCKED_SHAREDCACHE
;
3338 /* Any read-only or read-write transaction implies a read-lock on
3339 ** page 1. So if some other shared-cache client already has a write-lock
3340 ** on page 1, the transaction cannot be opened. */
3341 rc
= querySharedCacheTableLock(p
, MASTER_ROOT
, READ_LOCK
);
3342 if( SQLITE_OK
!=rc
) goto trans_begun
;
3344 pBt
->btsFlags
&= ~BTS_INITIALLY_EMPTY
;
3345 if( pBt
->nPage
==0 ) pBt
->btsFlags
|= BTS_INITIALLY_EMPTY
;
3347 /* Call lockBtree() until either pBt->pPage1 is populated or
3348 ** lockBtree() returns something other than SQLITE_OK. lockBtree()
3349 ** may return SQLITE_OK but leave pBt->pPage1 set to 0 if after
3350 ** reading page 1 it discovers that the page-size of the database
3351 ** file is not pBt->pageSize. In this case lockBtree() will update
3352 ** pBt->pageSize to the page-size of the file on disk.
3354 while( pBt
->pPage1
==0 && SQLITE_OK
==(rc
= lockBtree(pBt
)) );
3356 if( rc
==SQLITE_OK
&& wrflag
){
3357 if( (pBt
->btsFlags
& BTS_READ_ONLY
)!=0 ){
3358 rc
= SQLITE_READONLY
;
3360 rc
= sqlite3PagerBegin(pBt
->pPager
,wrflag
>1,sqlite3TempInMemory(p
->db
));
3361 if( rc
==SQLITE_OK
){
3362 rc
= newDatabase(pBt
);
3367 if( rc
!=SQLITE_OK
){
3368 unlockBtreeIfUnused(pBt
);
3370 }while( (rc
&0xFF)==SQLITE_BUSY
&& pBt
->inTransaction
==TRANS_NONE
&&
3371 btreeInvokeBusyHandler(pBt
) );
3373 if( rc
==SQLITE_OK
){
3374 if( p
->inTrans
==TRANS_NONE
){
3375 pBt
->nTransaction
++;
3376 #ifndef SQLITE_OMIT_SHARED_CACHE
3378 assert( p
->lock
.pBtree
==p
&& p
->lock
.iTable
==1 );
3379 p
->lock
.eLock
= READ_LOCK
;
3380 p
->lock
.pNext
= pBt
->pLock
;
3381 pBt
->pLock
= &p
->lock
;
3385 p
->inTrans
= (wrflag
?TRANS_WRITE
:TRANS_READ
);
3386 if( p
->inTrans
>pBt
->inTransaction
){
3387 pBt
->inTransaction
= p
->inTrans
;
3390 MemPage
*pPage1
= pBt
->pPage1
;
3391 #ifndef SQLITE_OMIT_SHARED_CACHE
3392 assert( !pBt
->pWriter
);
3394 pBt
->btsFlags
&= ~BTS_EXCLUSIVE
;
3395 if( wrflag
>1 ) pBt
->btsFlags
|= BTS_EXCLUSIVE
;
3398 /* If the db-size header field is incorrect (as it may be if an old
3399 ** client has been writing the database file), update it now. Doing
3400 ** this sooner rather than later means the database size can safely
3401 ** re-read the database size from page 1 if a savepoint or transaction
3402 ** rollback occurs within the transaction.
3404 if( pBt
->nPage
!=get4byte(&pPage1
->aData
[28]) ){
3405 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
3406 if( rc
==SQLITE_OK
){
3407 put4byte(&pPage1
->aData
[28], pBt
->nPage
);
3415 if( rc
==SQLITE_OK
&& wrflag
){
3416 /* This call makes sure that the pager has the correct number of
3417 ** open savepoints. If the second parameter is greater than 0 and
3418 ** the sub-journal is not already open, then it will be opened here.
3420 rc
= sqlite3PagerOpenSavepoint(pBt
->pPager
, p
->db
->nSavepoint
);
3424 sqlite3BtreeLeave(p
);
3428 #ifndef SQLITE_OMIT_AUTOVACUUM
3431 ** Set the pointer-map entries for all children of page pPage. Also, if
3432 ** pPage contains cells that point to overflow pages, set the pointer
3433 ** map entries for the overflow pages as well.
3435 static int setChildPtrmaps(MemPage
*pPage
){
3436 int i
; /* Counter variable */
3437 int nCell
; /* Number of cells in page pPage */
3438 int rc
; /* Return code */
3439 BtShared
*pBt
= pPage
->pBt
;
3440 Pgno pgno
= pPage
->pgno
;
3442 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
3443 rc
= pPage
->isInit
? SQLITE_OK
: btreeInitPage(pPage
);
3444 if( rc
!=SQLITE_OK
) return rc
;
3445 nCell
= pPage
->nCell
;
3447 for(i
=0; i
<nCell
; i
++){
3448 u8
*pCell
= findCell(pPage
, i
);
3450 ptrmapPutOvflPtr(pPage
, pCell
, &rc
);
3453 Pgno childPgno
= get4byte(pCell
);
3454 ptrmapPut(pBt
, childPgno
, PTRMAP_BTREE
, pgno
, &rc
);
3459 Pgno childPgno
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
3460 ptrmapPut(pBt
, childPgno
, PTRMAP_BTREE
, pgno
, &rc
);
3467 ** Somewhere on pPage is a pointer to page iFrom. Modify this pointer so
3468 ** that it points to iTo. Parameter eType describes the type of pointer to
3469 ** be modified, as follows:
3471 ** PTRMAP_BTREE: pPage is a btree-page. The pointer points at a child
3474 ** PTRMAP_OVERFLOW1: pPage is a btree-page. The pointer points at an overflow
3475 ** page pointed to by one of the cells on pPage.
3477 ** PTRMAP_OVERFLOW2: pPage is an overflow-page. The pointer points at the next
3478 ** overflow page in the list.
3480 static int modifyPagePointer(MemPage
*pPage
, Pgno iFrom
, Pgno iTo
, u8 eType
){
3481 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
3482 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
3483 if( eType
==PTRMAP_OVERFLOW2
){
3484 /* The pointer is always the first 4 bytes of the page in this case. */
3485 if( get4byte(pPage
->aData
)!=iFrom
){
3486 return SQLITE_CORRUPT_PAGE(pPage
);
3488 put4byte(pPage
->aData
, iTo
);
3494 rc
= pPage
->isInit
? SQLITE_OK
: btreeInitPage(pPage
);
3496 nCell
= pPage
->nCell
;
3498 for(i
=0; i
<nCell
; i
++){
3499 u8
*pCell
= findCell(pPage
, i
);
3500 if( eType
==PTRMAP_OVERFLOW1
){
3502 pPage
->xParseCell(pPage
, pCell
, &info
);
3503 if( info
.nLocal
<info
.nPayload
){
3504 if( pCell
+info
.nSize
> pPage
->aData
+pPage
->pBt
->usableSize
){
3505 return SQLITE_CORRUPT_PAGE(pPage
);
3507 if( iFrom
==get4byte(pCell
+info
.nSize
-4) ){
3508 put4byte(pCell
+info
.nSize
-4, iTo
);
3513 if( get4byte(pCell
)==iFrom
){
3514 put4byte(pCell
, iTo
);
3521 if( eType
!=PTRMAP_BTREE
||
3522 get4byte(&pPage
->aData
[pPage
->hdrOffset
+8])!=iFrom
){
3523 return SQLITE_CORRUPT_PAGE(pPage
);
3525 put4byte(&pPage
->aData
[pPage
->hdrOffset
+8], iTo
);
3533 ** Move the open database page pDbPage to location iFreePage in the
3534 ** database. The pDbPage reference remains valid.
3536 ** The isCommit flag indicates that there is no need to remember that
3537 ** the journal needs to be sync()ed before database page pDbPage->pgno
3538 ** can be written to. The caller has already promised not to write to that
3541 static int relocatePage(
3542 BtShared
*pBt
, /* Btree */
3543 MemPage
*pDbPage
, /* Open page to move */
3544 u8 eType
, /* Pointer map 'type' entry for pDbPage */
3545 Pgno iPtrPage
, /* Pointer map 'page-no' entry for pDbPage */
3546 Pgno iFreePage
, /* The location to move pDbPage to */
3547 int isCommit
/* isCommit flag passed to sqlite3PagerMovepage */
3549 MemPage
*pPtrPage
; /* The page that contains a pointer to pDbPage */
3550 Pgno iDbPage
= pDbPage
->pgno
;
3551 Pager
*pPager
= pBt
->pPager
;
3554 assert( eType
==PTRMAP_OVERFLOW2
|| eType
==PTRMAP_OVERFLOW1
||
3555 eType
==PTRMAP_BTREE
|| eType
==PTRMAP_ROOTPAGE
);
3556 assert( sqlite3_mutex_held(pBt
->mutex
) );
3557 assert( pDbPage
->pBt
==pBt
);
3559 /* Move page iDbPage from its current location to page number iFreePage */
3560 TRACE(("AUTOVACUUM: Moving %d to free page %d (ptr page %d type %d)\n",
3561 iDbPage
, iFreePage
, iPtrPage
, eType
));
3562 rc
= sqlite3PagerMovepage(pPager
, pDbPage
->pDbPage
, iFreePage
, isCommit
);
3563 if( rc
!=SQLITE_OK
){
3566 pDbPage
->pgno
= iFreePage
;
3568 /* If pDbPage was a btree-page, then it may have child pages and/or cells
3569 ** that point to overflow pages. The pointer map entries for all these
3570 ** pages need to be changed.
3572 ** If pDbPage is an overflow page, then the first 4 bytes may store a
3573 ** pointer to a subsequent overflow page. If this is the case, then
3574 ** the pointer map needs to be updated for the subsequent overflow page.
3576 if( eType
==PTRMAP_BTREE
|| eType
==PTRMAP_ROOTPAGE
){
3577 rc
= setChildPtrmaps(pDbPage
);
3578 if( rc
!=SQLITE_OK
){
3582 Pgno nextOvfl
= get4byte(pDbPage
->aData
);
3584 ptrmapPut(pBt
, nextOvfl
, PTRMAP_OVERFLOW2
, iFreePage
, &rc
);
3585 if( rc
!=SQLITE_OK
){
3591 /* Fix the database pointer on page iPtrPage that pointed at iDbPage so
3592 ** that it points at iFreePage. Also fix the pointer map entry for
3595 if( eType
!=PTRMAP_ROOTPAGE
){
3596 rc
= btreeGetPage(pBt
, iPtrPage
, &pPtrPage
, 0);
3597 if( rc
!=SQLITE_OK
){
3600 rc
= sqlite3PagerWrite(pPtrPage
->pDbPage
);
3601 if( rc
!=SQLITE_OK
){
3602 releasePage(pPtrPage
);
3605 rc
= modifyPagePointer(pPtrPage
, iDbPage
, iFreePage
, eType
);
3606 releasePage(pPtrPage
);
3607 if( rc
==SQLITE_OK
){
3608 ptrmapPut(pBt
, iFreePage
, eType
, iPtrPage
, &rc
);
3614 /* Forward declaration required by incrVacuumStep(). */
3615 static int allocateBtreePage(BtShared
*, MemPage
**, Pgno
*, Pgno
, u8
);
3618 ** Perform a single step of an incremental-vacuum. If successful, return
3619 ** SQLITE_OK. If there is no work to do (and therefore no point in
3620 ** calling this function again), return SQLITE_DONE. Or, if an error
3621 ** occurs, return some other error code.
3623 ** More specifically, this function attempts to re-organize the database so
3624 ** that the last page of the file currently in use is no longer in use.
3626 ** Parameter nFin is the number of pages that this database would contain
3627 ** were this function called until it returns SQLITE_DONE.
3629 ** If the bCommit parameter is non-zero, this function assumes that the
3630 ** caller will keep calling incrVacuumStep() until it returns SQLITE_DONE
3631 ** or an error. bCommit is passed true for an auto-vacuum-on-commit
3632 ** operation, or false for an incremental vacuum.
3634 static int incrVacuumStep(BtShared
*pBt
, Pgno nFin
, Pgno iLastPg
, int bCommit
){
3635 Pgno nFreeList
; /* Number of pages still on the free-list */
3638 assert( sqlite3_mutex_held(pBt
->mutex
) );
3639 assert( iLastPg
>nFin
);
3641 if( !PTRMAP_ISPAGE(pBt
, iLastPg
) && iLastPg
!=PENDING_BYTE_PAGE(pBt
) ){
3645 nFreeList
= get4byte(&pBt
->pPage1
->aData
[36]);
3650 rc
= ptrmapGet(pBt
, iLastPg
, &eType
, &iPtrPage
);
3651 if( rc
!=SQLITE_OK
){
3654 if( eType
==PTRMAP_ROOTPAGE
){
3655 return SQLITE_CORRUPT_BKPT
;
3658 if( eType
==PTRMAP_FREEPAGE
){
3660 /* Remove the page from the files free-list. This is not required
3661 ** if bCommit is non-zero. In that case, the free-list will be
3662 ** truncated to zero after this function returns, so it doesn't
3663 ** matter if it still contains some garbage entries.
3667 rc
= allocateBtreePage(pBt
, &pFreePg
, &iFreePg
, iLastPg
, BTALLOC_EXACT
);
3668 if( rc
!=SQLITE_OK
){
3671 assert( iFreePg
==iLastPg
);
3672 releasePage(pFreePg
);
3675 Pgno iFreePg
; /* Index of free page to move pLastPg to */
3677 u8 eMode
= BTALLOC_ANY
; /* Mode parameter for allocateBtreePage() */
3678 Pgno iNear
= 0; /* nearby parameter for allocateBtreePage() */
3680 rc
= btreeGetPage(pBt
, iLastPg
, &pLastPg
, 0);
3681 if( rc
!=SQLITE_OK
){
3685 /* If bCommit is zero, this loop runs exactly once and page pLastPg
3686 ** is swapped with the first free page pulled off the free list.
3688 ** On the other hand, if bCommit is greater than zero, then keep
3689 ** looping until a free-page located within the first nFin pages
3690 ** of the file is found.
3698 rc
= allocateBtreePage(pBt
, &pFreePg
, &iFreePg
, iNear
, eMode
);
3699 if( rc
!=SQLITE_OK
){
3700 releasePage(pLastPg
);
3703 releasePage(pFreePg
);
3704 }while( bCommit
&& iFreePg
>nFin
);
3705 assert( iFreePg
<iLastPg
);
3707 rc
= relocatePage(pBt
, pLastPg
, eType
, iPtrPage
, iFreePg
, bCommit
);
3708 releasePage(pLastPg
);
3709 if( rc
!=SQLITE_OK
){
3718 }while( iLastPg
==PENDING_BYTE_PAGE(pBt
) || PTRMAP_ISPAGE(pBt
, iLastPg
) );
3719 pBt
->bDoTruncate
= 1;
3720 pBt
->nPage
= iLastPg
;
3726 ** The database opened by the first argument is an auto-vacuum database
3727 ** nOrig pages in size containing nFree free pages. Return the expected
3728 ** size of the database in pages following an auto-vacuum operation.
3730 static Pgno
finalDbSize(BtShared
*pBt
, Pgno nOrig
, Pgno nFree
){
3731 int nEntry
; /* Number of entries on one ptrmap page */
3732 Pgno nPtrmap
; /* Number of PtrMap pages to be freed */
3733 Pgno nFin
; /* Return value */
3735 nEntry
= pBt
->usableSize
/5;
3736 nPtrmap
= (nFree
-nOrig
+PTRMAP_PAGENO(pBt
, nOrig
)+nEntry
)/nEntry
;
3737 nFin
= nOrig
- nFree
- nPtrmap
;
3738 if( nOrig
>PENDING_BYTE_PAGE(pBt
) && nFin
<PENDING_BYTE_PAGE(pBt
) ){
3741 while( PTRMAP_ISPAGE(pBt
, nFin
) || nFin
==PENDING_BYTE_PAGE(pBt
) ){
3749 ** A write-transaction must be opened before calling this function.
3750 ** It performs a single unit of work towards an incremental vacuum.
3752 ** If the incremental vacuum is finished after this function has run,
3753 ** SQLITE_DONE is returned. If it is not finished, but no error occurred,
3754 ** SQLITE_OK is returned. Otherwise an SQLite error code.
3756 int sqlite3BtreeIncrVacuum(Btree
*p
){
3758 BtShared
*pBt
= p
->pBt
;
3760 sqlite3BtreeEnter(p
);
3761 assert( pBt
->inTransaction
==TRANS_WRITE
&& p
->inTrans
==TRANS_WRITE
);
3762 if( !pBt
->autoVacuum
){
3765 Pgno nOrig
= btreePagecount(pBt
);
3766 Pgno nFree
= get4byte(&pBt
->pPage1
->aData
[36]);
3767 Pgno nFin
= finalDbSize(pBt
, nOrig
, nFree
);
3770 rc
= SQLITE_CORRUPT_BKPT
;
3771 }else if( nFree
>0 ){
3772 rc
= saveAllCursors(pBt
, 0, 0);
3773 if( rc
==SQLITE_OK
){
3774 invalidateAllOverflowCache(pBt
);
3775 rc
= incrVacuumStep(pBt
, nFin
, nOrig
, 0);
3777 if( rc
==SQLITE_OK
){
3778 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
3779 put4byte(&pBt
->pPage1
->aData
[28], pBt
->nPage
);
3785 sqlite3BtreeLeave(p
);
3790 ** This routine is called prior to sqlite3PagerCommit when a transaction
3791 ** is committed for an auto-vacuum database.
3793 ** If SQLITE_OK is returned, then *pnTrunc is set to the number of pages
3794 ** the database file should be truncated to during the commit process.
3795 ** i.e. the database has been reorganized so that only the first *pnTrunc
3796 ** pages are in use.
3798 static int autoVacuumCommit(BtShared
*pBt
){
3800 Pager
*pPager
= pBt
->pPager
;
3801 VVA_ONLY( int nRef
= sqlite3PagerRefcount(pPager
); )
3803 assert( sqlite3_mutex_held(pBt
->mutex
) );
3804 invalidateAllOverflowCache(pBt
);
3805 assert(pBt
->autoVacuum
);
3806 if( !pBt
->incrVacuum
){
3807 Pgno nFin
; /* Number of pages in database after autovacuuming */
3808 Pgno nFree
; /* Number of pages on the freelist initially */
3809 Pgno iFree
; /* The next page to be freed */
3810 Pgno nOrig
; /* Database size before freeing */
3812 nOrig
= btreePagecount(pBt
);
3813 if( PTRMAP_ISPAGE(pBt
, nOrig
) || nOrig
==PENDING_BYTE_PAGE(pBt
) ){
3814 /* It is not possible to create a database for which the final page
3815 ** is either a pointer-map page or the pending-byte page. If one
3816 ** is encountered, this indicates corruption.
3818 return SQLITE_CORRUPT_BKPT
;
3821 nFree
= get4byte(&pBt
->pPage1
->aData
[36]);
3822 nFin
= finalDbSize(pBt
, nOrig
, nFree
);
3823 if( nFin
>nOrig
) return SQLITE_CORRUPT_BKPT
;
3825 rc
= saveAllCursors(pBt
, 0, 0);
3827 for(iFree
=nOrig
; iFree
>nFin
&& rc
==SQLITE_OK
; iFree
--){
3828 rc
= incrVacuumStep(pBt
, nFin
, iFree
, 1);
3830 if( (rc
==SQLITE_DONE
|| rc
==SQLITE_OK
) && nFree
>0 ){
3831 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
3832 put4byte(&pBt
->pPage1
->aData
[32], 0);
3833 put4byte(&pBt
->pPage1
->aData
[36], 0);
3834 put4byte(&pBt
->pPage1
->aData
[28], nFin
);
3835 pBt
->bDoTruncate
= 1;
3838 if( rc
!=SQLITE_OK
){
3839 sqlite3PagerRollback(pPager
);
3843 assert( nRef
>=sqlite3PagerRefcount(pPager
) );
3847 #else /* ifndef SQLITE_OMIT_AUTOVACUUM */
3848 # define setChildPtrmaps(x) SQLITE_OK
3852 ** This routine does the first phase of a two-phase commit. This routine
3853 ** causes a rollback journal to be created (if it does not already exist)
3854 ** and populated with enough information so that if a power loss occurs
3855 ** the database can be restored to its original state by playing back
3856 ** the journal. Then the contents of the journal are flushed out to
3857 ** the disk. After the journal is safely on oxide, the changes to the
3858 ** database are written into the database file and flushed to oxide.
3859 ** At the end of this call, the rollback journal still exists on the
3860 ** disk and we are still holding all locks, so the transaction has not
3861 ** committed. See sqlite3BtreeCommitPhaseTwo() for the second phase of the
3864 ** This call is a no-op if no write-transaction is currently active on pBt.
3866 ** Otherwise, sync the database file for the btree pBt. zMaster points to
3867 ** the name of a master journal file that should be written into the
3868 ** individual journal file, or is NULL, indicating no master journal file
3869 ** (single database transaction).
3871 ** When this is called, the master journal should already have been
3872 ** created, populated with this journal pointer and synced to disk.
3874 ** Once this is routine has returned, the only thing required to commit
3875 ** the write-transaction for this database file is to delete the journal.
3877 int sqlite3BtreeCommitPhaseOne(Btree
*p
, const char *zMaster
){
3879 if( p
->inTrans
==TRANS_WRITE
){
3880 BtShared
*pBt
= p
->pBt
;
3881 sqlite3BtreeEnter(p
);
3882 #ifndef SQLITE_OMIT_AUTOVACUUM
3883 if( pBt
->autoVacuum
){
3884 rc
= autoVacuumCommit(pBt
);
3885 if( rc
!=SQLITE_OK
){
3886 sqlite3BtreeLeave(p
);
3890 if( pBt
->bDoTruncate
){
3891 sqlite3PagerTruncateImage(pBt
->pPager
, pBt
->nPage
);
3894 rc
= sqlite3PagerCommitPhaseOne(pBt
->pPager
, zMaster
, 0);
3895 sqlite3BtreeLeave(p
);
3901 ** This function is called from both BtreeCommitPhaseTwo() and BtreeRollback()
3902 ** at the conclusion of a transaction.
3904 static void btreeEndTransaction(Btree
*p
){
3905 BtShared
*pBt
= p
->pBt
;
3906 sqlite3
*db
= p
->db
;
3907 assert( sqlite3BtreeHoldsMutex(p
) );
3909 #ifndef SQLITE_OMIT_AUTOVACUUM
3910 pBt
->bDoTruncate
= 0;
3912 if( p
->inTrans
>TRANS_NONE
&& db
->nVdbeRead
>1 ){
3913 /* If there are other active statements that belong to this database
3914 ** handle, downgrade to a read-only transaction. The other statements
3915 ** may still be reading from the database. */
3916 downgradeAllSharedCacheTableLocks(p
);
3917 p
->inTrans
= TRANS_READ
;
3919 /* If the handle had any kind of transaction open, decrement the
3920 ** transaction count of the shared btree. If the transaction count
3921 ** reaches 0, set the shared state to TRANS_NONE. The unlockBtreeIfUnused()
3922 ** call below will unlock the pager. */
3923 if( p
->inTrans
!=TRANS_NONE
){
3924 clearAllSharedCacheTableLocks(p
);
3925 pBt
->nTransaction
--;
3926 if( 0==pBt
->nTransaction
){
3927 pBt
->inTransaction
= TRANS_NONE
;
3931 /* Set the current transaction state to TRANS_NONE and unlock the
3932 ** pager if this call closed the only read or write transaction. */
3933 p
->inTrans
= TRANS_NONE
;
3934 unlockBtreeIfUnused(pBt
);
3941 ** Commit the transaction currently in progress.
3943 ** This routine implements the second phase of a 2-phase commit. The
3944 ** sqlite3BtreeCommitPhaseOne() routine does the first phase and should
3945 ** be invoked prior to calling this routine. The sqlite3BtreeCommitPhaseOne()
3946 ** routine did all the work of writing information out to disk and flushing the
3947 ** contents so that they are written onto the disk platter. All this
3948 ** routine has to do is delete or truncate or zero the header in the
3949 ** the rollback journal (which causes the transaction to commit) and
3952 ** Normally, if an error occurs while the pager layer is attempting to
3953 ** finalize the underlying journal file, this function returns an error and
3954 ** the upper layer will attempt a rollback. However, if the second argument
3955 ** is non-zero then this b-tree transaction is part of a multi-file
3956 ** transaction. In this case, the transaction has already been committed
3957 ** (by deleting a master journal file) and the caller will ignore this
3958 ** functions return code. So, even if an error occurs in the pager layer,
3959 ** reset the b-tree objects internal state to indicate that the write
3960 ** transaction has been closed. This is quite safe, as the pager will have
3961 ** transitioned to the error state.
3963 ** This will release the write lock on the database file. If there
3964 ** are no active cursors, it also releases the read lock.
3966 int sqlite3BtreeCommitPhaseTwo(Btree
*p
, int bCleanup
){
3968 if( p
->inTrans
==TRANS_NONE
) return SQLITE_OK
;
3969 sqlite3BtreeEnter(p
);
3972 /* If the handle has a write-transaction open, commit the shared-btrees
3973 ** transaction and set the shared state to TRANS_READ.
3975 if( p
->inTrans
==TRANS_WRITE
){
3977 BtShared
*pBt
= p
->pBt
;
3978 assert( pBt
->inTransaction
==TRANS_WRITE
);
3979 assert( pBt
->nTransaction
>0 );
3980 rc
= sqlite3PagerCommitPhaseTwo(pBt
->pPager
);
3981 if( rc
!=SQLITE_OK
&& bCleanup
==0 ){
3982 sqlite3BtreeLeave(p
);
3985 p
->iDataVersion
--; /* Compensate for pPager->iDataVersion++; */
3986 pBt
->inTransaction
= TRANS_READ
;
3987 btreeClearHasContent(pBt
);
3990 btreeEndTransaction(p
);
3991 sqlite3BtreeLeave(p
);
3996 ** Do both phases of a commit.
3998 int sqlite3BtreeCommit(Btree
*p
){
4000 sqlite3BtreeEnter(p
);
4001 rc
= sqlite3BtreeCommitPhaseOne(p
, 0);
4002 if( rc
==SQLITE_OK
){
4003 rc
= sqlite3BtreeCommitPhaseTwo(p
, 0);
4005 sqlite3BtreeLeave(p
);
4010 ** This routine sets the state to CURSOR_FAULT and the error
4011 ** code to errCode for every cursor on any BtShared that pBtree
4012 ** references. Or if the writeOnly flag is set to 1, then only
4013 ** trip write cursors and leave read cursors unchanged.
4015 ** Every cursor is a candidate to be tripped, including cursors
4016 ** that belong to other database connections that happen to be
4017 ** sharing the cache with pBtree.
4019 ** This routine gets called when a rollback occurs. If the writeOnly
4020 ** flag is true, then only write-cursors need be tripped - read-only
4021 ** cursors save their current positions so that they may continue
4022 ** following the rollback. Or, if writeOnly is false, all cursors are
4023 ** tripped. In general, writeOnly is false if the transaction being
4024 ** rolled back modified the database schema. In this case b-tree root
4025 ** pages may be moved or deleted from the database altogether, making
4026 ** it unsafe for read cursors to continue.
4028 ** If the writeOnly flag is true and an error is encountered while
4029 ** saving the current position of a read-only cursor, all cursors,
4030 ** including all read-cursors are tripped.
4032 ** SQLITE_OK is returned if successful, or if an error occurs while
4033 ** saving a cursor position, an SQLite error code.
4035 int sqlite3BtreeTripAllCursors(Btree
*pBtree
, int errCode
, int writeOnly
){
4039 assert( (writeOnly
==0 || writeOnly
==1) && BTCF_WriteFlag
==1 );
4041 sqlite3BtreeEnter(pBtree
);
4042 for(p
=pBtree
->pBt
->pCursor
; p
; p
=p
->pNext
){
4043 if( writeOnly
&& (p
->curFlags
& BTCF_WriteFlag
)==0 ){
4044 if( p
->eState
==CURSOR_VALID
|| p
->eState
==CURSOR_SKIPNEXT
){
4045 rc
= saveCursorPosition(p
);
4046 if( rc
!=SQLITE_OK
){
4047 (void)sqlite3BtreeTripAllCursors(pBtree
, rc
, 0);
4052 sqlite3BtreeClearCursor(p
);
4053 p
->eState
= CURSOR_FAULT
;
4054 p
->skipNext
= errCode
;
4056 btreeReleaseAllCursorPages(p
);
4058 sqlite3BtreeLeave(pBtree
);
4064 ** Rollback the transaction in progress.
4066 ** If tripCode is not SQLITE_OK then cursors will be invalidated (tripped).
4067 ** Only write cursors are tripped if writeOnly is true but all cursors are
4068 ** tripped if writeOnly is false. Any attempt to use
4069 ** a tripped cursor will result in an error.
4071 ** This will release the write lock on the database file. If there
4072 ** are no active cursors, it also releases the read lock.
4074 int sqlite3BtreeRollback(Btree
*p
, int tripCode
, int writeOnly
){
4076 BtShared
*pBt
= p
->pBt
;
4079 assert( writeOnly
==1 || writeOnly
==0 );
4080 assert( tripCode
==SQLITE_ABORT_ROLLBACK
|| tripCode
==SQLITE_OK
);
4081 sqlite3BtreeEnter(p
);
4082 if( tripCode
==SQLITE_OK
){
4083 rc
= tripCode
= saveAllCursors(pBt
, 0, 0);
4084 if( rc
) writeOnly
= 0;
4089 int rc2
= sqlite3BtreeTripAllCursors(p
, tripCode
, writeOnly
);
4090 assert( rc
==SQLITE_OK
|| (writeOnly
==0 && rc2
==SQLITE_OK
) );
4091 if( rc2
!=SQLITE_OK
) rc
= rc2
;
4095 if( p
->inTrans
==TRANS_WRITE
){
4098 assert( TRANS_WRITE
==pBt
->inTransaction
);
4099 rc2
= sqlite3PagerRollback(pBt
->pPager
);
4100 if( rc2
!=SQLITE_OK
){
4104 /* The rollback may have destroyed the pPage1->aData value. So
4105 ** call btreeGetPage() on page 1 again to make
4106 ** sure pPage1->aData is set correctly. */
4107 if( btreeGetPage(pBt
, 1, &pPage1
, 0)==SQLITE_OK
){
4108 int nPage
= get4byte(28+(u8
*)pPage1
->aData
);
4109 testcase( nPage
==0 );
4110 if( nPage
==0 ) sqlite3PagerPagecount(pBt
->pPager
, &nPage
);
4111 testcase( pBt
->nPage
!=nPage
);
4113 releasePageOne(pPage1
);
4115 assert( countValidCursors(pBt
, 1)==0 );
4116 pBt
->inTransaction
= TRANS_READ
;
4117 btreeClearHasContent(pBt
);
4120 btreeEndTransaction(p
);
4121 sqlite3BtreeLeave(p
);
4126 ** Start a statement subtransaction. The subtransaction can be rolled
4127 ** back independently of the main transaction. You must start a transaction
4128 ** before starting a subtransaction. The subtransaction is ended automatically
4129 ** if the main transaction commits or rolls back.
4131 ** Statement subtransactions are used around individual SQL statements
4132 ** that are contained within a BEGIN...COMMIT block. If a constraint
4133 ** error occurs within the statement, the effect of that one statement
4134 ** can be rolled back without having to rollback the entire transaction.
4136 ** A statement sub-transaction is implemented as an anonymous savepoint. The
4137 ** value passed as the second parameter is the total number of savepoints,
4138 ** including the new anonymous savepoint, open on the B-Tree. i.e. if there
4139 ** are no active savepoints and no other statement-transactions open,
4140 ** iStatement is 1. This anonymous savepoint can be released or rolled back
4141 ** using the sqlite3BtreeSavepoint() function.
4143 int sqlite3BtreeBeginStmt(Btree
*p
, int iStatement
){
4145 BtShared
*pBt
= p
->pBt
;
4146 sqlite3BtreeEnter(p
);
4147 assert( p
->inTrans
==TRANS_WRITE
);
4148 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
4149 assert( iStatement
>0 );
4150 assert( iStatement
>p
->db
->nSavepoint
);
4151 assert( pBt
->inTransaction
==TRANS_WRITE
);
4152 /* At the pager level, a statement transaction is a savepoint with
4153 ** an index greater than all savepoints created explicitly using
4154 ** SQL statements. It is illegal to open, release or rollback any
4155 ** such savepoints while the statement transaction savepoint is active.
4157 rc
= sqlite3PagerOpenSavepoint(pBt
->pPager
, iStatement
);
4158 sqlite3BtreeLeave(p
);
4163 ** The second argument to this function, op, is always SAVEPOINT_ROLLBACK
4164 ** or SAVEPOINT_RELEASE. This function either releases or rolls back the
4165 ** savepoint identified by parameter iSavepoint, depending on the value
4168 ** Normally, iSavepoint is greater than or equal to zero. However, if op is
4169 ** SAVEPOINT_ROLLBACK, then iSavepoint may also be -1. In this case the
4170 ** contents of the entire transaction are rolled back. This is different
4171 ** from a normal transaction rollback, as no locks are released and the
4172 ** transaction remains open.
4174 int sqlite3BtreeSavepoint(Btree
*p
, int op
, int iSavepoint
){
4176 if( p
&& p
->inTrans
==TRANS_WRITE
){
4177 BtShared
*pBt
= p
->pBt
;
4178 assert( op
==SAVEPOINT_RELEASE
|| op
==SAVEPOINT_ROLLBACK
);
4179 assert( iSavepoint
>=0 || (iSavepoint
==-1 && op
==SAVEPOINT_ROLLBACK
) );
4180 sqlite3BtreeEnter(p
);
4181 if( op
==SAVEPOINT_ROLLBACK
){
4182 rc
= saveAllCursors(pBt
, 0, 0);
4184 if( rc
==SQLITE_OK
){
4185 rc
= sqlite3PagerSavepoint(pBt
->pPager
, op
, iSavepoint
);
4187 if( rc
==SQLITE_OK
){
4188 if( iSavepoint
<0 && (pBt
->btsFlags
& BTS_INITIALLY_EMPTY
)!=0 ){
4191 rc
= newDatabase(pBt
);
4192 pBt
->nPage
= get4byte(28 + pBt
->pPage1
->aData
);
4194 /* The database size was written into the offset 28 of the header
4195 ** when the transaction started, so we know that the value at offset
4196 ** 28 is nonzero. */
4197 assert( pBt
->nPage
>0 );
4199 sqlite3BtreeLeave(p
);
4205 ** Create a new cursor for the BTree whose root is on the page
4206 ** iTable. If a read-only cursor is requested, it is assumed that
4207 ** the caller already has at least a read-only transaction open
4208 ** on the database already. If a write-cursor is requested, then
4209 ** the caller is assumed to have an open write transaction.
4211 ** If the BTREE_WRCSR bit of wrFlag is clear, then the cursor can only
4212 ** be used for reading. If the BTREE_WRCSR bit is set, then the cursor
4213 ** can be used for reading or for writing if other conditions for writing
4214 ** are also met. These are the conditions that must be met in order
4215 ** for writing to be allowed:
4217 ** 1: The cursor must have been opened with wrFlag containing BTREE_WRCSR
4219 ** 2: Other database connections that share the same pager cache
4220 ** but which are not in the READ_UNCOMMITTED state may not have
4221 ** cursors open with wrFlag==0 on the same table. Otherwise
4222 ** the changes made by this write cursor would be visible to
4223 ** the read cursors in the other database connection.
4225 ** 3: The database must be writable (not on read-only media)
4227 ** 4: There must be an active transaction.
4229 ** The BTREE_FORDELETE bit of wrFlag may optionally be set if BTREE_WRCSR
4230 ** is set. If FORDELETE is set, that is a hint to the implementation that
4231 ** this cursor will only be used to seek to and delete entries of an index
4232 ** as part of a larger DELETE statement. The FORDELETE hint is not used by
4233 ** this implementation. But in a hypothetical alternative storage engine
4234 ** in which index entries are automatically deleted when corresponding table
4235 ** rows are deleted, the FORDELETE flag is a hint that all SEEK and DELETE
4236 ** operations on this cursor can be no-ops and all READ operations can
4237 ** return a null row (2-bytes: 0x01 0x00).
4239 ** No checking is done to make sure that page iTable really is the
4240 ** root page of a b-tree. If it is not, then the cursor acquired
4241 ** will not work correctly.
4243 ** It is assumed that the sqlite3BtreeCursorZero() has been called
4244 ** on pCur to initialize the memory space prior to invoking this routine.
4246 static int btreeCursor(
4247 Btree
*p
, /* The btree */
4248 int iTable
, /* Root page of table to open */
4249 int wrFlag
, /* 1 to write. 0 read-only */
4250 struct KeyInfo
*pKeyInfo
, /* First arg to comparison function */
4251 BtCursor
*pCur
/* Space for new cursor */
4253 BtShared
*pBt
= p
->pBt
; /* Shared b-tree handle */
4254 BtCursor
*pX
; /* Looping over other all cursors */
4256 assert( sqlite3BtreeHoldsMutex(p
) );
4258 || wrFlag
==BTREE_WRCSR
4259 || wrFlag
==(BTREE_WRCSR
|BTREE_FORDELETE
)
4262 /* The following assert statements verify that if this is a sharable
4263 ** b-tree database, the connection is holding the required table locks,
4264 ** and that no other connection has any open cursor that conflicts with
4266 assert( hasSharedCacheTableLock(p
, iTable
, pKeyInfo
!=0, (wrFlag
?2:1)) );
4267 assert( wrFlag
==0 || !hasReadConflicts(p
, iTable
) );
4269 /* Assert that the caller has opened the required transaction. */
4270 assert( p
->inTrans
>TRANS_NONE
);
4271 assert( wrFlag
==0 || p
->inTrans
==TRANS_WRITE
);
4272 assert( pBt
->pPage1
&& pBt
->pPage1
->aData
);
4273 assert( wrFlag
==0 || (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
4276 allocateTempSpace(pBt
);
4277 if( pBt
->pTmpSpace
==0 ) return SQLITE_NOMEM_BKPT
;
4279 if( iTable
==1 && btreePagecount(pBt
)==0 ){
4280 assert( wrFlag
==0 );
4284 /* Now that no other errors can occur, finish filling in the BtCursor
4285 ** variables and link the cursor into the BtShared list. */
4286 pCur
->pgnoRoot
= (Pgno
)iTable
;
4288 pCur
->pKeyInfo
= pKeyInfo
;
4291 pCur
->curFlags
= wrFlag
? BTCF_WriteFlag
: 0;
4292 pCur
->curPagerFlags
= wrFlag
? 0 : PAGER_GET_READONLY
;
4293 /* If there are two or more cursors on the same btree, then all such
4294 ** cursors *must* have the BTCF_Multiple flag set. */
4295 for(pX
=pBt
->pCursor
; pX
; pX
=pX
->pNext
){
4296 if( pX
->pgnoRoot
==(Pgno
)iTable
){
4297 pX
->curFlags
|= BTCF_Multiple
;
4298 pCur
->curFlags
|= BTCF_Multiple
;
4301 pCur
->pNext
= pBt
->pCursor
;
4302 pBt
->pCursor
= pCur
;
4303 pCur
->eState
= CURSOR_INVALID
;
4306 int sqlite3BtreeCursor(
4307 Btree
*p
, /* The btree */
4308 int iTable
, /* Root page of table to open */
4309 int wrFlag
, /* 1 to write. 0 read-only */
4310 struct KeyInfo
*pKeyInfo
, /* First arg to xCompare() */
4311 BtCursor
*pCur
/* Write new cursor here */
4315 rc
= SQLITE_CORRUPT_BKPT
;
4317 sqlite3BtreeEnter(p
);
4318 rc
= btreeCursor(p
, iTable
, wrFlag
, pKeyInfo
, pCur
);
4319 sqlite3BtreeLeave(p
);
4325 ** Return the size of a BtCursor object in bytes.
4327 ** This interfaces is needed so that users of cursors can preallocate
4328 ** sufficient storage to hold a cursor. The BtCursor object is opaque
4329 ** to users so they cannot do the sizeof() themselves - they must call
4332 int sqlite3BtreeCursorSize(void){
4333 return ROUND8(sizeof(BtCursor
));
4337 ** Initialize memory that will be converted into a BtCursor object.
4339 ** The simple approach here would be to memset() the entire object
4340 ** to zero. But it turns out that the apPage[] and aiIdx[] arrays
4341 ** do not need to be zeroed and they are large, so we can save a lot
4342 ** of run-time by skipping the initialization of those elements.
4344 void sqlite3BtreeCursorZero(BtCursor
*p
){
4345 memset(p
, 0, offsetof(BtCursor
, iPage
));
4349 ** Close a cursor. The read lock on the database file is released
4350 ** when the last cursor is closed.
4352 int sqlite3BtreeCloseCursor(BtCursor
*pCur
){
4353 Btree
*pBtree
= pCur
->pBtree
;
4355 BtShared
*pBt
= pCur
->pBt
;
4356 sqlite3BtreeEnter(pBtree
);
4357 assert( pBt
->pCursor
!=0 );
4358 if( pBt
->pCursor
==pCur
){
4359 pBt
->pCursor
= pCur
->pNext
;
4361 BtCursor
*pPrev
= pBt
->pCursor
;
4363 if( pPrev
->pNext
==pCur
){
4364 pPrev
->pNext
= pCur
->pNext
;
4367 pPrev
= pPrev
->pNext
;
4368 }while( ALWAYS(pPrev
) );
4370 btreeReleaseAllCursorPages(pCur
);
4371 unlockBtreeIfUnused(pBt
);
4372 sqlite3_free(pCur
->aOverflow
);
4373 sqlite3_free(pCur
->pKey
);
4374 sqlite3BtreeLeave(pBtree
);
4380 ** Make sure the BtCursor* given in the argument has a valid
4381 ** BtCursor.info structure. If it is not already valid, call
4382 ** btreeParseCell() to fill it in.
4384 ** BtCursor.info is a cache of the information in the current cell.
4385 ** Using this cache reduces the number of calls to btreeParseCell().
4388 static void assertCellInfo(BtCursor
*pCur
){
4390 memset(&info
, 0, sizeof(info
));
4391 btreeParseCell(pCur
->pPage
, pCur
->ix
, &info
);
4392 assert( CORRUPT_DB
|| memcmp(&info
, &pCur
->info
, sizeof(info
))==0 );
4395 #define assertCellInfo(x)
4397 static SQLITE_NOINLINE
void getCellInfo(BtCursor
*pCur
){
4398 if( pCur
->info
.nSize
==0 ){
4399 pCur
->curFlags
|= BTCF_ValidNKey
;
4400 btreeParseCell(pCur
->pPage
,pCur
->ix
,&pCur
->info
);
4402 assertCellInfo(pCur
);
4406 #ifndef NDEBUG /* The next routine used only within assert() statements */
4408 ** Return true if the given BtCursor is valid. A valid cursor is one
4409 ** that is currently pointing to a row in a (non-empty) table.
4410 ** This is a verification routine is used only within assert() statements.
4412 int sqlite3BtreeCursorIsValid(BtCursor
*pCur
){
4413 return pCur
&& pCur
->eState
==CURSOR_VALID
;
4416 int sqlite3BtreeCursorIsValidNN(BtCursor
*pCur
){
4418 return pCur
->eState
==CURSOR_VALID
;
4422 ** Return the value of the integer key or "rowid" for a table btree.
4423 ** This routine is only valid for a cursor that is pointing into a
4424 ** ordinary table btree. If the cursor points to an index btree or
4425 ** is invalid, the result of this routine is undefined.
4427 i64
sqlite3BtreeIntegerKey(BtCursor
*pCur
){
4428 assert( cursorHoldsMutex(pCur
) );
4429 assert( pCur
->eState
==CURSOR_VALID
);
4430 assert( pCur
->curIntKey
);
4432 return pCur
->info
.nKey
;
4435 #ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC
4437 ** Return the offset into the database file for the start of the
4438 ** payload to which the cursor is pointing.
4440 i64
sqlite3BtreeOffset(BtCursor
*pCur
){
4441 assert( cursorHoldsMutex(pCur
) );
4442 assert( pCur
->eState
==CURSOR_VALID
);
4444 return (i64
)pCur
->pBt
->pageSize
*((i64
)pCur
->pPage
->pgno
- 1) +
4445 (i64
)(pCur
->info
.pPayload
- pCur
->pPage
->aData
);
4447 #endif /* SQLITE_ENABLE_OFFSET_SQL_FUNC */
4450 ** Return the number of bytes of payload for the entry that pCur is
4451 ** currently pointing to. For table btrees, this will be the amount
4452 ** of data. For index btrees, this will be the size of the key.
4454 ** The caller must guarantee that the cursor is pointing to a non-NULL
4455 ** valid entry. In other words, the calling procedure must guarantee
4456 ** that the cursor has Cursor.eState==CURSOR_VALID.
4458 u32
sqlite3BtreePayloadSize(BtCursor
*pCur
){
4459 assert( cursorHoldsMutex(pCur
) );
4460 assert( pCur
->eState
==CURSOR_VALID
);
4462 return pCur
->info
.nPayload
;
4466 ** Given the page number of an overflow page in the database (parameter
4467 ** ovfl), this function finds the page number of the next page in the
4468 ** linked list of overflow pages. If possible, it uses the auto-vacuum
4469 ** pointer-map data instead of reading the content of page ovfl to do so.
4471 ** If an error occurs an SQLite error code is returned. Otherwise:
4473 ** The page number of the next overflow page in the linked list is
4474 ** written to *pPgnoNext. If page ovfl is the last page in its linked
4475 ** list, *pPgnoNext is set to zero.
4477 ** If ppPage is not NULL, and a reference to the MemPage object corresponding
4478 ** to page number pOvfl was obtained, then *ppPage is set to point to that
4479 ** reference. It is the responsibility of the caller to call releasePage()
4480 ** on *ppPage to free the reference. In no reference was obtained (because
4481 ** the pointer-map was used to obtain the value for *pPgnoNext), then
4482 ** *ppPage is set to zero.
4484 static int getOverflowPage(
4485 BtShared
*pBt
, /* The database file */
4486 Pgno ovfl
, /* Current overflow page number */
4487 MemPage
**ppPage
, /* OUT: MemPage handle (may be NULL) */
4488 Pgno
*pPgnoNext
/* OUT: Next overflow page number */
4494 assert( sqlite3_mutex_held(pBt
->mutex
) );
4497 #ifndef SQLITE_OMIT_AUTOVACUUM
4498 /* Try to find the next page in the overflow list using the
4499 ** autovacuum pointer-map pages. Guess that the next page in
4500 ** the overflow list is page number (ovfl+1). If that guess turns
4501 ** out to be wrong, fall back to loading the data of page
4502 ** number ovfl to determine the next page number.
4504 if( pBt
->autoVacuum
){
4506 Pgno iGuess
= ovfl
+1;
4509 while( PTRMAP_ISPAGE(pBt
, iGuess
) || iGuess
==PENDING_BYTE_PAGE(pBt
) ){
4513 if( iGuess
<=btreePagecount(pBt
) ){
4514 rc
= ptrmapGet(pBt
, iGuess
, &eType
, &pgno
);
4515 if( rc
==SQLITE_OK
&& eType
==PTRMAP_OVERFLOW2
&& pgno
==ovfl
){
4523 assert( next
==0 || rc
==SQLITE_DONE
);
4524 if( rc
==SQLITE_OK
){
4525 rc
= btreeGetPage(pBt
, ovfl
, &pPage
, (ppPage
==0) ? PAGER_GET_READONLY
: 0);
4526 assert( rc
==SQLITE_OK
|| pPage
==0 );
4527 if( rc
==SQLITE_OK
){
4528 next
= get4byte(pPage
->aData
);
4538 return (rc
==SQLITE_DONE
? SQLITE_OK
: rc
);
4542 ** Copy data from a buffer to a page, or from a page to a buffer.
4544 ** pPayload is a pointer to data stored on database page pDbPage.
4545 ** If argument eOp is false, then nByte bytes of data are copied
4546 ** from pPayload to the buffer pointed at by pBuf. If eOp is true,
4547 ** then sqlite3PagerWrite() is called on pDbPage and nByte bytes
4548 ** of data are copied from the buffer pBuf to pPayload.
4550 ** SQLITE_OK is returned on success, otherwise an error code.
4552 static int copyPayload(
4553 void *pPayload
, /* Pointer to page data */
4554 void *pBuf
, /* Pointer to buffer */
4555 int nByte
, /* Number of bytes to copy */
4556 int eOp
, /* 0 -> copy from page, 1 -> copy to page */
4557 DbPage
*pDbPage
/* Page containing pPayload */
4560 /* Copy data from buffer to page (a write operation) */
4561 int rc
= sqlite3PagerWrite(pDbPage
);
4562 if( rc
!=SQLITE_OK
){
4565 memcpy(pPayload
, pBuf
, nByte
);
4567 /* Copy data from page to buffer (a read operation) */
4568 memcpy(pBuf
, pPayload
, nByte
);
4574 ** This function is used to read or overwrite payload information
4575 ** for the entry that the pCur cursor is pointing to. The eOp
4576 ** argument is interpreted as follows:
4578 ** 0: The operation is a read. Populate the overflow cache.
4579 ** 1: The operation is a write. Populate the overflow cache.
4581 ** A total of "amt" bytes are read or written beginning at "offset".
4582 ** Data is read to or from the buffer pBuf.
4584 ** The content being read or written might appear on the main page
4585 ** or be scattered out on multiple overflow pages.
4587 ** If the current cursor entry uses one or more overflow pages
4588 ** this function may allocate space for and lazily populate
4589 ** the overflow page-list cache array (BtCursor.aOverflow).
4590 ** Subsequent calls use this cache to make seeking to the supplied offset
4593 ** Once an overflow page-list cache has been allocated, it must be
4594 ** invalidated if some other cursor writes to the same table, or if
4595 ** the cursor is moved to a different row. Additionally, in auto-vacuum
4596 ** mode, the following events may invalidate an overflow page-list cache.
4598 ** * An incremental vacuum,
4599 ** * A commit in auto_vacuum="full" mode,
4600 ** * Creating a table (may require moving an overflow page).
4602 static int accessPayload(
4603 BtCursor
*pCur
, /* Cursor pointing to entry to read from */
4604 u32 offset
, /* Begin reading this far into payload */
4605 u32 amt
, /* Read this many bytes */
4606 unsigned char *pBuf
, /* Write the bytes into this buffer */
4607 int eOp
/* zero to read. non-zero to write. */
4609 unsigned char *aPayload
;
4612 MemPage
*pPage
= pCur
->pPage
; /* Btree page of current entry */
4613 BtShared
*pBt
= pCur
->pBt
; /* Btree this cursor belongs to */
4614 #ifdef SQLITE_DIRECT_OVERFLOW_READ
4615 unsigned char * const pBufStart
= pBuf
; /* Start of original out buffer */
4619 assert( eOp
==0 || eOp
==1 );
4620 assert( pCur
->eState
==CURSOR_VALID
);
4621 assert( pCur
->ix
<pPage
->nCell
);
4622 assert( cursorHoldsMutex(pCur
) );
4625 aPayload
= pCur
->info
.pPayload
;
4626 assert( offset
+amt
<= pCur
->info
.nPayload
);
4628 assert( aPayload
> pPage
->aData
);
4629 if( (uptr
)(aPayload
- pPage
->aData
) > (pBt
->usableSize
- pCur
->info
.nLocal
) ){
4630 /* Trying to read or write past the end of the data is an error. The
4631 ** conditional above is really:
4632 ** &aPayload[pCur->info.nLocal] > &pPage->aData[pBt->usableSize]
4633 ** but is recast into its current form to avoid integer overflow problems
4635 return SQLITE_CORRUPT_PAGE(pPage
);
4638 /* Check if data must be read/written to/from the btree page itself. */
4639 if( offset
<pCur
->info
.nLocal
){
4641 if( a
+offset
>pCur
->info
.nLocal
){
4642 a
= pCur
->info
.nLocal
- offset
;
4644 rc
= copyPayload(&aPayload
[offset
], pBuf
, a
, eOp
, pPage
->pDbPage
);
4649 offset
-= pCur
->info
.nLocal
;
4653 if( rc
==SQLITE_OK
&& amt
>0 ){
4654 const u32 ovflSize
= pBt
->usableSize
- 4; /* Bytes content per ovfl page */
4657 nextPage
= get4byte(&aPayload
[pCur
->info
.nLocal
]);
4659 /* If the BtCursor.aOverflow[] has not been allocated, allocate it now.
4661 ** The aOverflow[] array is sized at one entry for each overflow page
4662 ** in the overflow chain. The page number of the first overflow page is
4663 ** stored in aOverflow[0], etc. A value of 0 in the aOverflow[] array
4664 ** means "not yet known" (the cache is lazily populated).
4666 if( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 ){
4667 int nOvfl
= (pCur
->info
.nPayload
-pCur
->info
.nLocal
+ovflSize
-1)/ovflSize
;
4668 if( nOvfl
>pCur
->nOvflAlloc
){
4669 Pgno
*aNew
= (Pgno
*)sqlite3Realloc(
4670 pCur
->aOverflow
, nOvfl
*2*sizeof(Pgno
)
4673 return SQLITE_NOMEM_BKPT
;
4675 pCur
->nOvflAlloc
= nOvfl
*2;
4676 pCur
->aOverflow
= aNew
;
4679 memset(pCur
->aOverflow
, 0, nOvfl
*sizeof(Pgno
));
4680 pCur
->curFlags
|= BTCF_ValidOvfl
;
4682 /* If the overflow page-list cache has been allocated and the
4683 ** entry for the first required overflow page is valid, skip
4686 if( pCur
->aOverflow
[offset
/ovflSize
] ){
4687 iIdx
= (offset
/ovflSize
);
4688 nextPage
= pCur
->aOverflow
[iIdx
];
4689 offset
= (offset
%ovflSize
);
4693 assert( rc
==SQLITE_OK
&& amt
>0 );
4695 /* If required, populate the overflow page-list cache. */
4696 assert( pCur
->aOverflow
[iIdx
]==0
4697 || pCur
->aOverflow
[iIdx
]==nextPage
4699 pCur
->aOverflow
[iIdx
] = nextPage
;
4701 if( offset
>=ovflSize
){
4702 /* The only reason to read this page is to obtain the page
4703 ** number for the next page in the overflow chain. The page
4704 ** data is not required. So first try to lookup the overflow
4705 ** page-list cache, if any, then fall back to the getOverflowPage()
4708 assert( pCur
->curFlags
& BTCF_ValidOvfl
);
4709 assert( pCur
->pBtree
->db
==pBt
->db
);
4710 if( pCur
->aOverflow
[iIdx
+1] ){
4711 nextPage
= pCur
->aOverflow
[iIdx
+1];
4713 rc
= getOverflowPage(pBt
, nextPage
, 0, &nextPage
);
4717 /* Need to read this page properly. It contains some of the
4718 ** range of data that is being read (eOp==0) or written (eOp!=0).
4720 #ifdef SQLITE_DIRECT_OVERFLOW_READ
4721 sqlite3_file
*fd
; /* File from which to do direct overflow read */
4724 if( a
+ offset
> ovflSize
){
4725 a
= ovflSize
- offset
;
4728 #ifdef SQLITE_DIRECT_OVERFLOW_READ
4729 /* If all the following are true:
4731 ** 1) this is a read operation, and
4732 ** 2) data is required from the start of this overflow page, and
4733 ** 3) there is no open write-transaction, and
4734 ** 4) the database is file-backed, and
4735 ** 5) the page is not in the WAL file
4736 ** 6) at least 4 bytes have already been read into the output buffer
4738 ** then data can be read directly from the database file into the
4739 ** output buffer, bypassing the page-cache altogether. This speeds
4740 ** up loading large records that span many overflow pages.
4742 if( eOp
==0 /* (1) */
4743 && offset
==0 /* (2) */
4744 && pBt
->inTransaction
==TRANS_READ
/* (3) */
4745 && (fd
= sqlite3PagerFile(pBt
->pPager
))->pMethods
/* (4) */
4746 && 0==sqlite3PagerUseWal(pBt
->pPager
, nextPage
) /* (5) */
4747 && &pBuf
[-4]>=pBufStart
/* (6) */
4750 u8
*aWrite
= &pBuf
[-4];
4751 assert( aWrite
>=pBufStart
); /* due to (6) */
4752 memcpy(aSave
, aWrite
, 4);
4753 rc
= sqlite3OsRead(fd
, aWrite
, a
+4, (i64
)pBt
->pageSize
*(nextPage
-1));
4754 nextPage
= get4byte(aWrite
);
4755 memcpy(aWrite
, aSave
, 4);
4761 rc
= sqlite3PagerGet(pBt
->pPager
, nextPage
, &pDbPage
,
4762 (eOp
==0 ? PAGER_GET_READONLY
: 0)
4764 if( rc
==SQLITE_OK
){
4765 aPayload
= sqlite3PagerGetData(pDbPage
);
4766 nextPage
= get4byte(aPayload
);
4767 rc
= copyPayload(&aPayload
[offset
+4], pBuf
, a
, eOp
, pDbPage
);
4768 sqlite3PagerUnref(pDbPage
);
4773 if( amt
==0 ) return rc
;
4781 if( rc
==SQLITE_OK
&& amt
>0 ){
4782 /* Overflow chain ends prematurely */
4783 return SQLITE_CORRUPT_PAGE(pPage
);
4789 ** Read part of the payload for the row at which that cursor pCur is currently
4790 ** pointing. "amt" bytes will be transferred into pBuf[]. The transfer
4791 ** begins at "offset".
4793 ** pCur can be pointing to either a table or an index b-tree.
4794 ** If pointing to a table btree, then the content section is read. If
4795 ** pCur is pointing to an index b-tree then the key section is read.
4797 ** For sqlite3BtreePayload(), the caller must ensure that pCur is pointing
4798 ** to a valid row in the table. For sqlite3BtreePayloadChecked(), the
4799 ** cursor might be invalid or might need to be restored before being read.
4801 ** Return SQLITE_OK on success or an error code if anything goes
4802 ** wrong. An error is returned if "offset+amt" is larger than
4803 ** the available payload.
4805 int sqlite3BtreePayload(BtCursor
*pCur
, u32 offset
, u32 amt
, void *pBuf
){
4806 assert( cursorHoldsMutex(pCur
) );
4807 assert( pCur
->eState
==CURSOR_VALID
);
4808 assert( pCur
->iPage
>=0 && pCur
->pPage
);
4809 assert( pCur
->ix
<pCur
->pPage
->nCell
);
4810 return accessPayload(pCur
, offset
, amt
, (unsigned char*)pBuf
, 0);
4814 ** This variant of sqlite3BtreePayload() works even if the cursor has not
4815 ** in the CURSOR_VALID state. It is only used by the sqlite3_blob_read()
4818 #ifndef SQLITE_OMIT_INCRBLOB
4819 static SQLITE_NOINLINE
int accessPayloadChecked(
4826 if ( pCur
->eState
==CURSOR_INVALID
){
4827 return SQLITE_ABORT
;
4829 assert( cursorOwnsBtShared(pCur
) );
4830 rc
= btreeRestoreCursorPosition(pCur
);
4831 return rc
? rc
: accessPayload(pCur
, offset
, amt
, pBuf
, 0);
4833 int sqlite3BtreePayloadChecked(BtCursor
*pCur
, u32 offset
, u32 amt
, void *pBuf
){
4834 if( pCur
->eState
==CURSOR_VALID
){
4835 assert( cursorOwnsBtShared(pCur
) );
4836 return accessPayload(pCur
, offset
, amt
, pBuf
, 0);
4838 return accessPayloadChecked(pCur
, offset
, amt
, pBuf
);
4841 #endif /* SQLITE_OMIT_INCRBLOB */
4844 ** Return a pointer to payload information from the entry that the
4845 ** pCur cursor is pointing to. The pointer is to the beginning of
4846 ** the key if index btrees (pPage->intKey==0) and is the data for
4847 ** table btrees (pPage->intKey==1). The number of bytes of available
4848 ** key/data is written into *pAmt. If *pAmt==0, then the value
4849 ** returned will not be a valid pointer.
4851 ** This routine is an optimization. It is common for the entire key
4852 ** and data to fit on the local page and for there to be no overflow
4853 ** pages. When that is so, this routine can be used to access the
4854 ** key and data without making a copy. If the key and/or data spills
4855 ** onto overflow pages, then accessPayload() must be used to reassemble
4856 ** the key/data and copy it into a preallocated buffer.
4858 ** The pointer returned by this routine looks directly into the cached
4859 ** page of the database. The data might change or move the next time
4860 ** any btree routine is called.
4862 static const void *fetchPayload(
4863 BtCursor
*pCur
, /* Cursor pointing to entry to read from */
4864 u32
*pAmt
/* Write the number of available bytes here */
4867 assert( pCur
!=0 && pCur
->iPage
>=0 && pCur
->pPage
);
4868 assert( pCur
->eState
==CURSOR_VALID
);
4869 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
4870 assert( cursorOwnsBtShared(pCur
) );
4871 assert( pCur
->ix
<pCur
->pPage
->nCell
);
4872 assert( pCur
->info
.nSize
>0 );
4873 assert( pCur
->info
.pPayload
>pCur
->pPage
->aData
|| CORRUPT_DB
);
4874 assert( pCur
->info
.pPayload
<pCur
->pPage
->aDataEnd
||CORRUPT_DB
);
4875 amt
= pCur
->info
.nLocal
;
4876 if( amt
>(int)(pCur
->pPage
->aDataEnd
- pCur
->info
.pPayload
) ){
4877 /* There is too little space on the page for the expected amount
4878 ** of local content. Database must be corrupt. */
4879 assert( CORRUPT_DB
);
4880 amt
= MAX(0, (int)(pCur
->pPage
->aDataEnd
- pCur
->info
.pPayload
));
4883 return (void*)pCur
->info
.pPayload
;
4888 ** For the entry that cursor pCur is point to, return as
4889 ** many bytes of the key or data as are available on the local
4890 ** b-tree page. Write the number of available bytes into *pAmt.
4892 ** The pointer returned is ephemeral. The key/data may move
4893 ** or be destroyed on the next call to any Btree routine,
4894 ** including calls from other threads against the same cache.
4895 ** Hence, a mutex on the BtShared should be held prior to calling
4898 ** These routines is used to get quick access to key and data
4899 ** in the common case where no overflow pages are used.
4901 const void *sqlite3BtreePayloadFetch(BtCursor
*pCur
, u32
*pAmt
){
4902 return fetchPayload(pCur
, pAmt
);
4907 ** Move the cursor down to a new child page. The newPgno argument is the
4908 ** page number of the child page to move to.
4910 ** This function returns SQLITE_CORRUPT if the page-header flags field of
4911 ** the new child page does not match the flags field of the parent (i.e.
4912 ** if an intkey page appears to be the parent of a non-intkey page, or
4915 static int moveToChild(BtCursor
*pCur
, u32 newPgno
){
4916 BtShared
*pBt
= pCur
->pBt
;
4918 assert( cursorOwnsBtShared(pCur
) );
4919 assert( pCur
->eState
==CURSOR_VALID
);
4920 assert( pCur
->iPage
<BTCURSOR_MAX_DEPTH
);
4921 assert( pCur
->iPage
>=0 );
4922 if( pCur
->iPage
>=(BTCURSOR_MAX_DEPTH
-1) ){
4923 return SQLITE_CORRUPT_BKPT
;
4925 pCur
->info
.nSize
= 0;
4926 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
4927 pCur
->aiIdx
[pCur
->iPage
] = pCur
->ix
;
4928 pCur
->apPage
[pCur
->iPage
] = pCur
->pPage
;
4931 return getAndInitPage(pBt
, newPgno
, &pCur
->pPage
, pCur
, pCur
->curPagerFlags
);
4936 ** Page pParent is an internal (non-leaf) tree page. This function
4937 ** asserts that page number iChild is the left-child if the iIdx'th
4938 ** cell in page pParent. Or, if iIdx is equal to the total number of
4939 ** cells in pParent, that page number iChild is the right-child of
4942 static void assertParentIndex(MemPage
*pParent
, int iIdx
, Pgno iChild
){
4943 if( CORRUPT_DB
) return; /* The conditions tested below might not be true
4944 ** in a corrupt database */
4945 assert( iIdx
<=pParent
->nCell
);
4946 if( iIdx
==pParent
->nCell
){
4947 assert( get4byte(&pParent
->aData
[pParent
->hdrOffset
+8])==iChild
);
4949 assert( get4byte(findCell(pParent
, iIdx
))==iChild
);
4953 # define assertParentIndex(x,y,z)
4957 ** Move the cursor up to the parent page.
4959 ** pCur->idx is set to the cell index that contains the pointer
4960 ** to the page we are coming from. If we are coming from the
4961 ** right-most child page then pCur->idx is set to one more than
4962 ** the largest cell index.
4964 static void moveToParent(BtCursor
*pCur
){
4966 assert( cursorOwnsBtShared(pCur
) );
4967 assert( pCur
->eState
==CURSOR_VALID
);
4968 assert( pCur
->iPage
>0 );
4969 assert( pCur
->pPage
);
4971 pCur
->apPage
[pCur
->iPage
-1],
4972 pCur
->aiIdx
[pCur
->iPage
-1],
4975 testcase( pCur
->aiIdx
[pCur
->iPage
-1] > pCur
->apPage
[pCur
->iPage
-1]->nCell
);
4976 pCur
->info
.nSize
= 0;
4977 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
4978 pCur
->ix
= pCur
->aiIdx
[pCur
->iPage
-1];
4979 pLeaf
= pCur
->pPage
;
4980 pCur
->pPage
= pCur
->apPage
[--pCur
->iPage
];
4981 releasePageNotNull(pLeaf
);
4985 ** Move the cursor to point to the root page of its b-tree structure.
4987 ** If the table has a virtual root page, then the cursor is moved to point
4988 ** to the virtual root page instead of the actual root page. A table has a
4989 ** virtual root page when the actual root page contains no cells and a
4990 ** single child page. This can only happen with the table rooted at page 1.
4992 ** If the b-tree structure is empty, the cursor state is set to
4993 ** CURSOR_INVALID and this routine returns SQLITE_EMPTY. Otherwise,
4994 ** the cursor is set to point to the first cell located on the root
4995 ** (or virtual root) page and the cursor state is set to CURSOR_VALID.
4997 ** If this function returns successfully, it may be assumed that the
4998 ** page-header flags indicate that the [virtual] root-page is the expected
4999 ** kind of b-tree page (i.e. if when opening the cursor the caller did not
5000 ** specify a KeyInfo structure the flags byte is set to 0x05 or 0x0D,
5001 ** indicating a table b-tree, or if the caller did specify a KeyInfo
5002 ** structure the flags byte is set to 0x02 or 0x0A, indicating an index
5005 static int moveToRoot(BtCursor
*pCur
){
5009 assert( cursorOwnsBtShared(pCur
) );
5010 assert( CURSOR_INVALID
< CURSOR_REQUIRESEEK
);
5011 assert( CURSOR_VALID
< CURSOR_REQUIRESEEK
);
5012 assert( CURSOR_FAULT
> CURSOR_REQUIRESEEK
);
5013 assert( pCur
->eState
< CURSOR_REQUIRESEEK
|| pCur
->iPage
<0 );
5014 assert( pCur
->pgnoRoot
>0 || pCur
->iPage
<0 );
5016 if( pCur
->iPage
>=0 ){
5018 releasePageNotNull(pCur
->pPage
);
5019 while( --pCur
->iPage
){
5020 releasePageNotNull(pCur
->apPage
[pCur
->iPage
]);
5022 pCur
->pPage
= pCur
->apPage
[0];
5025 }else if( pCur
->pgnoRoot
==0 ){
5026 pCur
->eState
= CURSOR_INVALID
;
5027 return SQLITE_EMPTY
;
5029 assert( pCur
->iPage
==(-1) );
5030 if( pCur
->eState
>=CURSOR_REQUIRESEEK
){
5031 if( pCur
->eState
==CURSOR_FAULT
){
5032 assert( pCur
->skipNext
!=SQLITE_OK
);
5033 return pCur
->skipNext
;
5035 sqlite3BtreeClearCursor(pCur
);
5037 rc
= getAndInitPage(pCur
->pBtree
->pBt
, pCur
->pgnoRoot
, &pCur
->pPage
,
5038 0, pCur
->curPagerFlags
);
5039 if( rc
!=SQLITE_OK
){
5040 pCur
->eState
= CURSOR_INVALID
;
5044 pCur
->curIntKey
= pCur
->pPage
->intKey
;
5046 pRoot
= pCur
->pPage
;
5047 assert( pRoot
->pgno
==pCur
->pgnoRoot
);
5049 /* If pCur->pKeyInfo is not NULL, then the caller that opened this cursor
5050 ** expected to open it on an index b-tree. Otherwise, if pKeyInfo is
5051 ** NULL, the caller expects a table b-tree. If this is not the case,
5052 ** return an SQLITE_CORRUPT error.
5054 ** Earlier versions of SQLite assumed that this test could not fail
5055 ** if the root page was already loaded when this function was called (i.e.
5056 ** if pCur->iPage>=0). But this is not so if the database is corrupted
5057 ** in such a way that page pRoot is linked into a second b-tree table
5058 ** (or the freelist). */
5059 assert( pRoot
->intKey
==1 || pRoot
->intKey
==0 );
5060 if( pRoot
->isInit
==0 || (pCur
->pKeyInfo
==0)!=pRoot
->intKey
){
5061 return SQLITE_CORRUPT_PAGE(pCur
->pPage
);
5066 pCur
->info
.nSize
= 0;
5067 pCur
->curFlags
&= ~(BTCF_AtLast
|BTCF_ValidNKey
|BTCF_ValidOvfl
);
5069 pRoot
= pCur
->pPage
;
5070 if( pRoot
->nCell
>0 ){
5071 pCur
->eState
= CURSOR_VALID
;
5072 }else if( !pRoot
->leaf
){
5074 if( pRoot
->pgno
!=1 ) return SQLITE_CORRUPT_BKPT
;
5075 subpage
= get4byte(&pRoot
->aData
[pRoot
->hdrOffset
+8]);
5076 pCur
->eState
= CURSOR_VALID
;
5077 rc
= moveToChild(pCur
, subpage
);
5079 pCur
->eState
= CURSOR_INVALID
;
5086 ** Move the cursor down to the left-most leaf entry beneath the
5087 ** entry to which it is currently pointing.
5089 ** The left-most leaf is the one with the smallest key - the first
5090 ** in ascending order.
5092 static int moveToLeftmost(BtCursor
*pCur
){
5097 assert( cursorOwnsBtShared(pCur
) );
5098 assert( pCur
->eState
==CURSOR_VALID
);
5099 while( rc
==SQLITE_OK
&& !(pPage
= pCur
->pPage
)->leaf
){
5100 assert( pCur
->ix
<pPage
->nCell
);
5101 pgno
= get4byte(findCell(pPage
, pCur
->ix
));
5102 rc
= moveToChild(pCur
, pgno
);
5108 ** Move the cursor down to the right-most leaf entry beneath the
5109 ** page to which it is currently pointing. Notice the difference
5110 ** between moveToLeftmost() and moveToRightmost(). moveToLeftmost()
5111 ** finds the left-most entry beneath the *entry* whereas moveToRightmost()
5112 ** finds the right-most entry beneath the *page*.
5114 ** The right-most entry is the one with the largest key - the last
5115 ** key in ascending order.
5117 static int moveToRightmost(BtCursor
*pCur
){
5122 assert( cursorOwnsBtShared(pCur
) );
5123 assert( pCur
->eState
==CURSOR_VALID
);
5124 while( !(pPage
= pCur
->pPage
)->leaf
){
5125 pgno
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
5126 pCur
->ix
= pPage
->nCell
;
5127 rc
= moveToChild(pCur
, pgno
);
5130 pCur
->ix
= pPage
->nCell
-1;
5131 assert( pCur
->info
.nSize
==0 );
5132 assert( (pCur
->curFlags
& BTCF_ValidNKey
)==0 );
5136 /* Move the cursor to the first entry in the table. Return SQLITE_OK
5137 ** on success. Set *pRes to 0 if the cursor actually points to something
5138 ** or set *pRes to 1 if the table is empty.
5140 int sqlite3BtreeFirst(BtCursor
*pCur
, int *pRes
){
5143 assert( cursorOwnsBtShared(pCur
) );
5144 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5145 rc
= moveToRoot(pCur
);
5146 if( rc
==SQLITE_OK
){
5147 assert( pCur
->pPage
->nCell
>0 );
5149 rc
= moveToLeftmost(pCur
);
5150 }else if( rc
==SQLITE_EMPTY
){
5151 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5158 /* Move the cursor to the last entry in the table. Return SQLITE_OK
5159 ** on success. Set *pRes to 0 if the cursor actually points to something
5160 ** or set *pRes to 1 if the table is empty.
5162 int sqlite3BtreeLast(BtCursor
*pCur
, int *pRes
){
5165 assert( cursorOwnsBtShared(pCur
) );
5166 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5168 /* If the cursor already points to the last entry, this is a no-op. */
5169 if( CURSOR_VALID
==pCur
->eState
&& (pCur
->curFlags
& BTCF_AtLast
)!=0 ){
5171 /* This block serves to assert() that the cursor really does point
5172 ** to the last entry in the b-tree. */
5174 for(ii
=0; ii
<pCur
->iPage
; ii
++){
5175 assert( pCur
->aiIdx
[ii
]==pCur
->apPage
[ii
]->nCell
);
5177 assert( pCur
->ix
==pCur
->pPage
->nCell
-1 );
5178 assert( pCur
->pPage
->leaf
);
5183 rc
= moveToRoot(pCur
);
5184 if( rc
==SQLITE_OK
){
5185 assert( pCur
->eState
==CURSOR_VALID
);
5187 rc
= moveToRightmost(pCur
);
5188 if( rc
==SQLITE_OK
){
5189 pCur
->curFlags
|= BTCF_AtLast
;
5191 pCur
->curFlags
&= ~BTCF_AtLast
;
5193 }else if( rc
==SQLITE_EMPTY
){
5194 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5201 /* Move the cursor so that it points to an entry near the key
5202 ** specified by pIdxKey or intKey. Return a success code.
5204 ** For INTKEY tables, the intKey parameter is used. pIdxKey
5205 ** must be NULL. For index tables, pIdxKey is used and intKey
5208 ** If an exact match is not found, then the cursor is always
5209 ** left pointing at a leaf page which would hold the entry if it
5210 ** were present. The cursor might point to an entry that comes
5211 ** before or after the key.
5213 ** An integer is written into *pRes which is the result of
5214 ** comparing the key with the entry to which the cursor is
5215 ** pointing. The meaning of the integer written into
5216 ** *pRes is as follows:
5218 ** *pRes<0 The cursor is left pointing at an entry that
5219 ** is smaller than intKey/pIdxKey or if the table is empty
5220 ** and the cursor is therefore left point to nothing.
5222 ** *pRes==0 The cursor is left pointing at an entry that
5223 ** exactly matches intKey/pIdxKey.
5225 ** *pRes>0 The cursor is left pointing at an entry that
5226 ** is larger than intKey/pIdxKey.
5228 ** For index tables, the pIdxKey->eqSeen field is set to 1 if there
5229 ** exists an entry in the table that exactly matches pIdxKey.
5231 int sqlite3BtreeMovetoUnpacked(
5232 BtCursor
*pCur
, /* The cursor to be moved */
5233 UnpackedRecord
*pIdxKey
, /* Unpacked index key */
5234 i64 intKey
, /* The table key */
5235 int biasRight
, /* If true, bias the search to the high end */
5236 int *pRes
/* Write search results here */
5239 RecordCompare xRecordCompare
;
5241 assert( cursorOwnsBtShared(pCur
) );
5242 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5244 assert( (pIdxKey
==0)==(pCur
->pKeyInfo
==0) );
5245 assert( pCur
->eState
!=CURSOR_VALID
|| (pIdxKey
==0)==(pCur
->curIntKey
!=0) );
5247 /* If the cursor is already positioned at the point we are trying
5248 ** to move to, then just return without doing any work */
5250 && pCur
->eState
==CURSOR_VALID
&& (pCur
->curFlags
& BTCF_ValidNKey
)!=0
5252 if( pCur
->info
.nKey
==intKey
){
5256 if( pCur
->info
.nKey
<intKey
){
5257 if( (pCur
->curFlags
& BTCF_AtLast
)!=0 ){
5261 /* If the requested key is one more than the previous key, then
5262 ** try to get there using sqlite3BtreeNext() rather than a full
5263 ** binary search. This is an optimization only. The correct answer
5264 ** is still obtained without this case, only a little more slowely */
5265 if( pCur
->info
.nKey
+1==intKey
&& !pCur
->skipNext
){
5267 rc
= sqlite3BtreeNext(pCur
, 0);
5268 if( rc
==SQLITE_OK
){
5270 if( pCur
->info
.nKey
==intKey
){
5273 }else if( rc
==SQLITE_DONE
){
5283 xRecordCompare
= sqlite3VdbeFindCompare(pIdxKey
);
5284 pIdxKey
->errCode
= 0;
5285 assert( pIdxKey
->default_rc
==1
5286 || pIdxKey
->default_rc
==0
5287 || pIdxKey
->default_rc
==-1
5290 xRecordCompare
= 0; /* All keys are integers */
5293 rc
= moveToRoot(pCur
);
5295 if( rc
==SQLITE_EMPTY
){
5296 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5302 assert( pCur
->pPage
);
5303 assert( pCur
->pPage
->isInit
);
5304 assert( pCur
->eState
==CURSOR_VALID
);
5305 assert( pCur
->pPage
->nCell
> 0 );
5306 assert( pCur
->iPage
==0 || pCur
->apPage
[0]->intKey
==pCur
->curIntKey
);
5307 assert( pCur
->curIntKey
|| pIdxKey
);
5309 int lwr
, upr
, idx
, c
;
5311 MemPage
*pPage
= pCur
->pPage
;
5312 u8
*pCell
; /* Pointer to current cell in pPage */
5314 /* pPage->nCell must be greater than zero. If this is the root-page
5315 ** the cursor would have been INVALID above and this for(;;) loop
5316 ** not run. If this is not the root-page, then the moveToChild() routine
5317 ** would have already detected db corruption. Similarly, pPage must
5318 ** be the right kind (index or table) of b-tree page. Otherwise
5319 ** a moveToChild() or moveToRoot() call would have detected corruption. */
5320 assert( pPage
->nCell
>0 );
5321 assert( pPage
->intKey
==(pIdxKey
==0) );
5323 upr
= pPage
->nCell
-1;
5324 assert( biasRight
==0 || biasRight
==1 );
5325 idx
= upr
>>(1-biasRight
); /* idx = biasRight ? upr : (lwr+upr)/2; */
5326 pCur
->ix
= (u16
)idx
;
5327 if( xRecordCompare
==0 ){
5330 pCell
= findCellPastPtr(pPage
, idx
);
5331 if( pPage
->intKeyLeaf
){
5332 while( 0x80 <= *(pCell
++) ){
5333 if( pCell
>=pPage
->aDataEnd
){
5334 return SQLITE_CORRUPT_PAGE(pPage
);
5338 getVarint(pCell
, (u64
*)&nCellKey
);
5339 if( nCellKey
<intKey
){
5341 if( lwr
>upr
){ c
= -1; break; }
5342 }else if( nCellKey
>intKey
){
5344 if( lwr
>upr
){ c
= +1; break; }
5346 assert( nCellKey
==intKey
);
5347 pCur
->ix
= (u16
)idx
;
5350 goto moveto_next_layer
;
5352 pCur
->curFlags
|= BTCF_ValidNKey
;
5353 pCur
->info
.nKey
= nCellKey
;
5354 pCur
->info
.nSize
= 0;
5359 assert( lwr
+upr
>=0 );
5360 idx
= (lwr
+upr
)>>1; /* idx = (lwr+upr)/2; */
5364 int nCell
; /* Size of the pCell cell in bytes */
5365 pCell
= findCellPastPtr(pPage
, idx
);
5367 /* The maximum supported page-size is 65536 bytes. This means that
5368 ** the maximum number of record bytes stored on an index B-Tree
5369 ** page is less than 16384 bytes and may be stored as a 2-byte
5370 ** varint. This information is used to attempt to avoid parsing
5371 ** the entire cell by checking for the cases where the record is
5372 ** stored entirely within the b-tree page by inspecting the first
5373 ** 2 bytes of the cell.
5376 if( nCell
<=pPage
->max1bytePayload
){
5377 /* This branch runs if the record-size field of the cell is a
5378 ** single byte varint and the record fits entirely on the main
5380 testcase( pCell
+nCell
+1==pPage
->aDataEnd
);
5381 c
= xRecordCompare(nCell
, (void*)&pCell
[1], pIdxKey
);
5382 }else if( !(pCell
[1] & 0x80)
5383 && (nCell
= ((nCell
&0x7f)<<7) + pCell
[1])<=pPage
->maxLocal
5385 /* The record-size field is a 2 byte varint and the record
5386 ** fits entirely on the main b-tree page. */
5387 testcase( pCell
+nCell
+2==pPage
->aDataEnd
);
5388 c
= xRecordCompare(nCell
, (void*)&pCell
[2], pIdxKey
);
5390 /* The record flows over onto one or more overflow pages. In
5391 ** this case the whole cell needs to be parsed, a buffer allocated
5392 ** and accessPayload() used to retrieve the record into the
5393 ** buffer before VdbeRecordCompare() can be called.
5395 ** If the record is corrupt, the xRecordCompare routine may read
5396 ** up to two varints past the end of the buffer. An extra 18
5397 ** bytes of padding is allocated at the end of the buffer in
5398 ** case this happens. */
5400 u8
* const pCellBody
= pCell
- pPage
->childPtrSize
;
5401 pPage
->xParseCell(pPage
, pCellBody
, &pCur
->info
);
5402 nCell
= (int)pCur
->info
.nKey
;
5403 testcase( nCell
<0 ); /* True if key size is 2^32 or more */
5404 testcase( nCell
==0 ); /* Invalid key size: 0x80 0x80 0x00 */
5405 testcase( nCell
==1 ); /* Invalid key size: 0x80 0x80 0x01 */
5406 testcase( nCell
==2 ); /* Minimum legal index key size */
5408 rc
= SQLITE_CORRUPT_PAGE(pPage
);
5411 pCellKey
= sqlite3Malloc( nCell
+18 );
5413 rc
= SQLITE_NOMEM_BKPT
;
5416 pCur
->ix
= (u16
)idx
;
5417 rc
= accessPayload(pCur
, 0, nCell
, (unsigned char*)pCellKey
, 0);
5418 pCur
->curFlags
&= ~BTCF_ValidOvfl
;
5420 sqlite3_free(pCellKey
);
5423 c
= xRecordCompare(nCell
, pCellKey
, pIdxKey
);
5424 sqlite3_free(pCellKey
);
5427 (pIdxKey
->errCode
!=SQLITE_CORRUPT
|| c
==0)
5428 && (pIdxKey
->errCode
!=SQLITE_NOMEM
|| pCur
->pBtree
->db
->mallocFailed
)
5438 pCur
->ix
= (u16
)idx
;
5439 if( pIdxKey
->errCode
) rc
= SQLITE_CORRUPT_BKPT
;
5442 if( lwr
>upr
) break;
5443 assert( lwr
+upr
>=0 );
5444 idx
= (lwr
+upr
)>>1; /* idx = (lwr+upr)/2 */
5447 assert( lwr
==upr
+1 || (pPage
->intKey
&& !pPage
->leaf
) );
5448 assert( pPage
->isInit
);
5450 assert( pCur
->ix
<pCur
->pPage
->nCell
);
5451 pCur
->ix
= (u16
)idx
;
5457 if( lwr
>=pPage
->nCell
){
5458 chldPg
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
5460 chldPg
= get4byte(findCell(pPage
, lwr
));
5462 pCur
->ix
= (u16
)lwr
;
5463 rc
= moveToChild(pCur
, chldPg
);
5467 pCur
->info
.nSize
= 0;
5468 assert( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 );
5474 ** Return TRUE if the cursor is not pointing at an entry of the table.
5476 ** TRUE will be returned after a call to sqlite3BtreeNext() moves
5477 ** past the last entry in the table or sqlite3BtreePrev() moves past
5478 ** the first entry. TRUE is also returned if the table is empty.
5480 int sqlite3BtreeEof(BtCursor
*pCur
){
5481 /* TODO: What if the cursor is in CURSOR_REQUIRESEEK but all table entries
5482 ** have been deleted? This API will need to change to return an error code
5483 ** as well as the boolean result value.
5485 return (CURSOR_VALID
!=pCur
->eState
);
5489 ** Return an estimate for the number of rows in the table that pCur is
5490 ** pointing to. Return a negative number if no estimate is currently
5493 i64
sqlite3BtreeRowCountEst(BtCursor
*pCur
){
5497 assert( cursorOwnsBtShared(pCur
) );
5498 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5500 /* Currently this interface is only called by the OP_IfSmaller
5501 ** opcode, and it that case the cursor will always be valid and
5502 ** will always point to a leaf node. */
5503 if( NEVER(pCur
->eState
!=CURSOR_VALID
) ) return -1;
5504 if( NEVER(pCur
->pPage
->leaf
==0) ) return -1;
5506 n
= pCur
->pPage
->nCell
;
5507 for(i
=0; i
<pCur
->iPage
; i
++){
5508 n
*= pCur
->apPage
[i
]->nCell
;
5514 ** Advance the cursor to the next entry in the database.
5517 ** SQLITE_OK success
5518 ** SQLITE_DONE cursor is already pointing at the last element
5519 ** otherwise some kind of error occurred
5521 ** The main entry point is sqlite3BtreeNext(). That routine is optimized
5522 ** for the common case of merely incrementing the cell counter BtCursor.aiIdx
5523 ** to the next cell on the current page. The (slower) btreeNext() helper
5524 ** routine is called when it is necessary to move to a different page or
5525 ** to restore the cursor.
5527 ** If bit 0x01 of the F argument in sqlite3BtreeNext(C,F) is 1, then the
5528 ** cursor corresponds to an SQL index and this routine could have been
5529 ** skipped if the SQL index had been a unique index. The F argument
5530 ** is a hint to the implement. SQLite btree implementation does not use
5531 ** this hint, but COMDB2 does.
5533 static SQLITE_NOINLINE
int btreeNext(BtCursor
*pCur
){
5538 assert( cursorOwnsBtShared(pCur
) );
5539 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5540 if( pCur
->eState
!=CURSOR_VALID
){
5541 assert( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 );
5542 rc
= restoreCursorPosition(pCur
);
5543 if( rc
!=SQLITE_OK
){
5546 if( CURSOR_INVALID
==pCur
->eState
){
5549 if( pCur
->skipNext
){
5550 assert( pCur
->eState
==CURSOR_VALID
|| pCur
->eState
==CURSOR_SKIPNEXT
);
5551 pCur
->eState
= CURSOR_VALID
;
5552 if( pCur
->skipNext
>0 ){
5560 pPage
= pCur
->pPage
;
5562 assert( pPage
->isInit
);
5564 /* If the database file is corrupt, it is possible for the value of idx
5565 ** to be invalid here. This can only occur if a second cursor modifies
5566 ** the page while cursor pCur is holding a reference to it. Which can
5567 ** only happen if the database is corrupt in such a way as to link the
5568 ** page into more than one b-tree structure. */
5569 testcase( idx
>pPage
->nCell
);
5571 if( idx
>=pPage
->nCell
){
5573 rc
= moveToChild(pCur
, get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]));
5575 return moveToLeftmost(pCur
);
5578 if( pCur
->iPage
==0 ){
5579 pCur
->eState
= CURSOR_INVALID
;
5583 pPage
= pCur
->pPage
;
5584 }while( pCur
->ix
>=pPage
->nCell
);
5585 if( pPage
->intKey
){
5586 return sqlite3BtreeNext(pCur
, 0);
5594 return moveToLeftmost(pCur
);
5597 int sqlite3BtreeNext(BtCursor
*pCur
, int flags
){
5599 UNUSED_PARAMETER( flags
); /* Used in COMDB2 but not native SQLite */
5600 assert( cursorOwnsBtShared(pCur
) );
5601 assert( flags
==0 || flags
==1 );
5602 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5603 pCur
->info
.nSize
= 0;
5604 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
5605 if( pCur
->eState
!=CURSOR_VALID
) return btreeNext(pCur
);
5606 pPage
= pCur
->pPage
;
5607 if( (++pCur
->ix
)>=pPage
->nCell
){
5609 return btreeNext(pCur
);
5614 return moveToLeftmost(pCur
);
5619 ** Step the cursor to the back to the previous entry in the database.
5622 ** SQLITE_OK success
5623 ** SQLITE_DONE the cursor is already on the first element of the table
5624 ** otherwise some kind of error occurred
5626 ** The main entry point is sqlite3BtreePrevious(). That routine is optimized
5627 ** for the common case of merely decrementing the cell counter BtCursor.aiIdx
5628 ** to the previous cell on the current page. The (slower) btreePrevious()
5629 ** helper routine is called when it is necessary to move to a different page
5630 ** or to restore the cursor.
5632 ** If bit 0x01 of the F argument to sqlite3BtreePrevious(C,F) is 1, then
5633 ** the cursor corresponds to an SQL index and this routine could have been
5634 ** skipped if the SQL index had been a unique index. The F argument is a
5635 ** hint to the implement. The native SQLite btree implementation does not
5636 ** use this hint, but COMDB2 does.
5638 static SQLITE_NOINLINE
int btreePrevious(BtCursor
*pCur
){
5642 assert( cursorOwnsBtShared(pCur
) );
5643 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5644 assert( (pCur
->curFlags
& (BTCF_AtLast
|BTCF_ValidOvfl
|BTCF_ValidNKey
))==0 );
5645 assert( pCur
->info
.nSize
==0 );
5646 if( pCur
->eState
!=CURSOR_VALID
){
5647 rc
= restoreCursorPosition(pCur
);
5648 if( rc
!=SQLITE_OK
){
5651 if( CURSOR_INVALID
==pCur
->eState
){
5654 if( pCur
->skipNext
){
5655 assert( pCur
->eState
==CURSOR_VALID
|| pCur
->eState
==CURSOR_SKIPNEXT
);
5656 pCur
->eState
= CURSOR_VALID
;
5657 if( pCur
->skipNext
<0 ){
5665 pPage
= pCur
->pPage
;
5666 assert( pPage
->isInit
);
5669 rc
= moveToChild(pCur
, get4byte(findCell(pPage
, idx
)));
5671 rc
= moveToRightmost(pCur
);
5673 while( pCur
->ix
==0 ){
5674 if( pCur
->iPage
==0 ){
5675 pCur
->eState
= CURSOR_INVALID
;
5680 assert( pCur
->info
.nSize
==0 );
5681 assert( (pCur
->curFlags
& (BTCF_ValidOvfl
))==0 );
5684 pPage
= pCur
->pPage
;
5685 if( pPage
->intKey
&& !pPage
->leaf
){
5686 rc
= sqlite3BtreePrevious(pCur
, 0);
5693 int sqlite3BtreePrevious(BtCursor
*pCur
, int flags
){
5694 assert( cursorOwnsBtShared(pCur
) );
5695 assert( flags
==0 || flags
==1 );
5696 assert( pCur
->skipNext
==0 || pCur
->eState
!=CURSOR_VALID
);
5697 UNUSED_PARAMETER( flags
); /* Used in COMDB2 but not native SQLite */
5698 pCur
->curFlags
&= ~(BTCF_AtLast
|BTCF_ValidOvfl
|BTCF_ValidNKey
);
5699 pCur
->info
.nSize
= 0;
5700 if( pCur
->eState
!=CURSOR_VALID
5702 || pCur
->pPage
->leaf
==0
5704 return btreePrevious(pCur
);
5711 ** Allocate a new page from the database file.
5713 ** The new page is marked as dirty. (In other words, sqlite3PagerWrite()
5714 ** has already been called on the new page.) The new page has also
5715 ** been referenced and the calling routine is responsible for calling
5716 ** sqlite3PagerUnref() on the new page when it is done.
5718 ** SQLITE_OK is returned on success. Any other return value indicates
5719 ** an error. *ppPage is set to NULL in the event of an error.
5721 ** If the "nearby" parameter is not 0, then an effort is made to
5722 ** locate a page close to the page number "nearby". This can be used in an
5723 ** attempt to keep related pages close to each other in the database file,
5724 ** which in turn can make database access faster.
5726 ** If the eMode parameter is BTALLOC_EXACT and the nearby page exists
5727 ** anywhere on the free-list, then it is guaranteed to be returned. If
5728 ** eMode is BTALLOC_LT then the page returned will be less than or equal
5729 ** to nearby if any such page exists. If eMode is BTALLOC_ANY then there
5730 ** are no restrictions on which page is returned.
5732 static int allocateBtreePage(
5733 BtShared
*pBt
, /* The btree */
5734 MemPage
**ppPage
, /* Store pointer to the allocated page here */
5735 Pgno
*pPgno
, /* Store the page number here */
5736 Pgno nearby
, /* Search for a page near this one */
5737 u8 eMode
/* BTALLOC_EXACT, BTALLOC_LT, or BTALLOC_ANY */
5741 u32 n
; /* Number of pages on the freelist */
5742 u32 k
; /* Number of leaves on the trunk of the freelist */
5743 MemPage
*pTrunk
= 0;
5744 MemPage
*pPrevTrunk
= 0;
5745 Pgno mxPage
; /* Total size of the database file */
5747 assert( sqlite3_mutex_held(pBt
->mutex
) );
5748 assert( eMode
==BTALLOC_ANY
|| (nearby
>0 && IfNotOmitAV(pBt
->autoVacuum
)) );
5749 pPage1
= pBt
->pPage1
;
5750 mxPage
= btreePagecount(pBt
);
5751 /* EVIDENCE-OF: R-05119-02637 The 4-byte big-endian integer at offset 36
5752 ** stores stores the total number of pages on the freelist. */
5753 n
= get4byte(&pPage1
->aData
[36]);
5754 testcase( n
==mxPage
-1 );
5756 return SQLITE_CORRUPT_BKPT
;
5759 /* There are pages on the freelist. Reuse one of those pages. */
5761 u8 searchList
= 0; /* If the free-list must be searched for 'nearby' */
5762 u32 nSearch
= 0; /* Count of the number of search attempts */
5764 /* If eMode==BTALLOC_EXACT and a query of the pointer-map
5765 ** shows that the page 'nearby' is somewhere on the free-list, then
5766 ** the entire-list will be searched for that page.
5768 #ifndef SQLITE_OMIT_AUTOVACUUM
5769 if( eMode
==BTALLOC_EXACT
){
5770 if( nearby
<=mxPage
){
5773 assert( pBt
->autoVacuum
);
5774 rc
= ptrmapGet(pBt
, nearby
, &eType
, 0);
5776 if( eType
==PTRMAP_FREEPAGE
){
5780 }else if( eMode
==BTALLOC_LE
){
5785 /* Decrement the free-list count by 1. Set iTrunk to the index of the
5786 ** first free-list trunk page. iPrevTrunk is initially 1.
5788 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
5790 put4byte(&pPage1
->aData
[36], n
-1);
5792 /* The code within this loop is run only once if the 'searchList' variable
5793 ** is not true. Otherwise, it runs once for each trunk-page on the
5794 ** free-list until the page 'nearby' is located (eMode==BTALLOC_EXACT)
5795 ** or until a page less than 'nearby' is located (eMode==BTALLOC_LT)
5798 pPrevTrunk
= pTrunk
;
5800 /* EVIDENCE-OF: R-01506-11053 The first integer on a freelist trunk page
5801 ** is the page number of the next freelist trunk page in the list or
5802 ** zero if this is the last freelist trunk page. */
5803 iTrunk
= get4byte(&pPrevTrunk
->aData
[0]);
5805 /* EVIDENCE-OF: R-59841-13798 The 4-byte big-endian integer at offset 32
5806 ** stores the page number of the first page of the freelist, or zero if
5807 ** the freelist is empty. */
5808 iTrunk
= get4byte(&pPage1
->aData
[32]);
5810 testcase( iTrunk
==mxPage
);
5811 if( iTrunk
>mxPage
|| nSearch
++ > n
){
5812 rc
= SQLITE_CORRUPT_PGNO(pPrevTrunk
? pPrevTrunk
->pgno
: 1);
5814 rc
= btreeGetUnusedPage(pBt
, iTrunk
, &pTrunk
, 0);
5818 goto end_allocate_page
;
5820 assert( pTrunk
!=0 );
5821 assert( pTrunk
->aData
!=0 );
5822 /* EVIDENCE-OF: R-13523-04394 The second integer on a freelist trunk page
5823 ** is the number of leaf page pointers to follow. */
5824 k
= get4byte(&pTrunk
->aData
[4]);
5825 if( k
==0 && !searchList
){
5826 /* The trunk has no leaves and the list is not being searched.
5827 ** So extract the trunk page itself and use it as the newly
5828 ** allocated page */
5829 assert( pPrevTrunk
==0 );
5830 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
5832 goto end_allocate_page
;
5835 memcpy(&pPage1
->aData
[32], &pTrunk
->aData
[0], 4);
5838 TRACE(("ALLOCATE: %d trunk - %d free pages left\n", *pPgno
, n
-1));
5839 }else if( k
>(u32
)(pBt
->usableSize
/4 - 2) ){
5840 /* Value of k is out of range. Database corruption */
5841 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
5842 goto end_allocate_page
;
5843 #ifndef SQLITE_OMIT_AUTOVACUUM
5844 }else if( searchList
5845 && (nearby
==iTrunk
|| (iTrunk
<nearby
&& eMode
==BTALLOC_LE
))
5847 /* The list is being searched and this trunk page is the page
5848 ** to allocate, regardless of whether it has leaves.
5853 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
5855 goto end_allocate_page
;
5859 memcpy(&pPage1
->aData
[32], &pTrunk
->aData
[0], 4);
5861 rc
= sqlite3PagerWrite(pPrevTrunk
->pDbPage
);
5862 if( rc
!=SQLITE_OK
){
5863 goto end_allocate_page
;
5865 memcpy(&pPrevTrunk
->aData
[0], &pTrunk
->aData
[0], 4);
5868 /* The trunk page is required by the caller but it contains
5869 ** pointers to free-list leaves. The first leaf becomes a trunk
5870 ** page in this case.
5873 Pgno iNewTrunk
= get4byte(&pTrunk
->aData
[8]);
5874 if( iNewTrunk
>mxPage
){
5875 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
5876 goto end_allocate_page
;
5878 testcase( iNewTrunk
==mxPage
);
5879 rc
= btreeGetUnusedPage(pBt
, iNewTrunk
, &pNewTrunk
, 0);
5880 if( rc
!=SQLITE_OK
){
5881 goto end_allocate_page
;
5883 rc
= sqlite3PagerWrite(pNewTrunk
->pDbPage
);
5884 if( rc
!=SQLITE_OK
){
5885 releasePage(pNewTrunk
);
5886 goto end_allocate_page
;
5888 memcpy(&pNewTrunk
->aData
[0], &pTrunk
->aData
[0], 4);
5889 put4byte(&pNewTrunk
->aData
[4], k
-1);
5890 memcpy(&pNewTrunk
->aData
[8], &pTrunk
->aData
[12], (k
-1)*4);
5891 releasePage(pNewTrunk
);
5893 assert( sqlite3PagerIswriteable(pPage1
->pDbPage
) );
5894 put4byte(&pPage1
->aData
[32], iNewTrunk
);
5896 rc
= sqlite3PagerWrite(pPrevTrunk
->pDbPage
);
5898 goto end_allocate_page
;
5900 put4byte(&pPrevTrunk
->aData
[0], iNewTrunk
);
5904 TRACE(("ALLOCATE: %d trunk - %d free pages left\n", *pPgno
, n
-1));
5907 /* Extract a leaf from the trunk */
5910 unsigned char *aData
= pTrunk
->aData
;
5914 if( eMode
==BTALLOC_LE
){
5916 iPage
= get4byte(&aData
[8+i
*4]);
5917 if( iPage
<=nearby
){
5924 dist
= sqlite3AbsInt32(get4byte(&aData
[8]) - nearby
);
5926 int d2
= sqlite3AbsInt32(get4byte(&aData
[8+i
*4]) - nearby
);
5937 iPage
= get4byte(&aData
[8+closest
*4]);
5938 testcase( iPage
==mxPage
);
5940 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
5941 goto end_allocate_page
;
5943 testcase( iPage
==mxPage
);
5945 || (iPage
==nearby
|| (iPage
<nearby
&& eMode
==BTALLOC_LE
))
5949 TRACE(("ALLOCATE: %d was leaf %d of %d on trunk %d"
5950 ": %d more free pages\n",
5951 *pPgno
, closest
+1, k
, pTrunk
->pgno
, n
-1));
5952 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
5953 if( rc
) goto end_allocate_page
;
5955 memcpy(&aData
[8+closest
*4], &aData
[4+k
*4], 4);
5957 put4byte(&aData
[4], k
-1);
5958 noContent
= !btreeGetHasContent(pBt
, *pPgno
)? PAGER_GET_NOCONTENT
: 0;
5959 rc
= btreeGetUnusedPage(pBt
, *pPgno
, ppPage
, noContent
);
5960 if( rc
==SQLITE_OK
){
5961 rc
= sqlite3PagerWrite((*ppPage
)->pDbPage
);
5962 if( rc
!=SQLITE_OK
){
5963 releasePage(*ppPage
);
5970 releasePage(pPrevTrunk
);
5972 }while( searchList
);
5974 /* There are no pages on the freelist, so append a new page to the
5977 ** Normally, new pages allocated by this block can be requested from the
5978 ** pager layer with the 'no-content' flag set. This prevents the pager
5979 ** from trying to read the pages content from disk. However, if the
5980 ** current transaction has already run one or more incremental-vacuum
5981 ** steps, then the page we are about to allocate may contain content
5982 ** that is required in the event of a rollback. In this case, do
5983 ** not set the no-content flag. This causes the pager to load and journal
5984 ** the current page content before overwriting it.
5986 ** Note that the pager will not actually attempt to load or journal
5987 ** content for any page that really does lie past the end of the database
5988 ** file on disk. So the effects of disabling the no-content optimization
5989 ** here are confined to those pages that lie between the end of the
5990 ** database image and the end of the database file.
5992 int bNoContent
= (0==IfNotOmitAV(pBt
->bDoTruncate
))? PAGER_GET_NOCONTENT
:0;
5994 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
5997 if( pBt
->nPage
==PENDING_BYTE_PAGE(pBt
) ) pBt
->nPage
++;
5999 #ifndef SQLITE_OMIT_AUTOVACUUM
6000 if( pBt
->autoVacuum
&& PTRMAP_ISPAGE(pBt
, pBt
->nPage
) ){
6001 /* If *pPgno refers to a pointer-map page, allocate two new pages
6002 ** at the end of the file instead of one. The first allocated page
6003 ** becomes a new pointer-map page, the second is used by the caller.
6006 TRACE(("ALLOCATE: %d from end of file (pointer-map page)\n", pBt
->nPage
));
6007 assert( pBt
->nPage
!=PENDING_BYTE_PAGE(pBt
) );
6008 rc
= btreeGetUnusedPage(pBt
, pBt
->nPage
, &pPg
, bNoContent
);
6009 if( rc
==SQLITE_OK
){
6010 rc
= sqlite3PagerWrite(pPg
->pDbPage
);
6015 if( pBt
->nPage
==PENDING_BYTE_PAGE(pBt
) ){ pBt
->nPage
++; }
6018 put4byte(28 + (u8
*)pBt
->pPage1
->aData
, pBt
->nPage
);
6019 *pPgno
= pBt
->nPage
;
6021 assert( *pPgno
!=PENDING_BYTE_PAGE(pBt
) );
6022 rc
= btreeGetUnusedPage(pBt
, *pPgno
, ppPage
, bNoContent
);
6024 rc
= sqlite3PagerWrite((*ppPage
)->pDbPage
);
6025 if( rc
!=SQLITE_OK
){
6026 releasePage(*ppPage
);
6029 TRACE(("ALLOCATE: %d from end of file\n", *pPgno
));
6032 assert( *pPgno
!=PENDING_BYTE_PAGE(pBt
) );
6035 releasePage(pTrunk
);
6036 releasePage(pPrevTrunk
);
6037 assert( rc
!=SQLITE_OK
|| sqlite3PagerPageRefcount((*ppPage
)->pDbPage
)<=1 );
6038 assert( rc
!=SQLITE_OK
|| (*ppPage
)->isInit
==0 );
6043 ** This function is used to add page iPage to the database file free-list.
6044 ** It is assumed that the page is not already a part of the free-list.
6046 ** The value passed as the second argument to this function is optional.
6047 ** If the caller happens to have a pointer to the MemPage object
6048 ** corresponding to page iPage handy, it may pass it as the second value.
6049 ** Otherwise, it may pass NULL.
6051 ** If a pointer to a MemPage object is passed as the second argument,
6052 ** its reference count is not altered by this function.
6054 static int freePage2(BtShared
*pBt
, MemPage
*pMemPage
, Pgno iPage
){
6055 MemPage
*pTrunk
= 0; /* Free-list trunk page */
6056 Pgno iTrunk
= 0; /* Page number of free-list trunk page */
6057 MemPage
*pPage1
= pBt
->pPage1
; /* Local reference to page 1 */
6058 MemPage
*pPage
; /* Page being freed. May be NULL. */
6059 int rc
; /* Return Code */
6060 int nFree
; /* Initial number of pages on free-list */
6062 assert( sqlite3_mutex_held(pBt
->mutex
) );
6063 assert( CORRUPT_DB
|| iPage
>1 );
6064 assert( !pMemPage
|| pMemPage
->pgno
==iPage
);
6066 if( iPage
<2 ) return SQLITE_CORRUPT_BKPT
;
6069 sqlite3PagerRef(pPage
->pDbPage
);
6071 pPage
= btreePageLookup(pBt
, iPage
);
6074 /* Increment the free page count on pPage1 */
6075 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
6076 if( rc
) goto freepage_out
;
6077 nFree
= get4byte(&pPage1
->aData
[36]);
6078 put4byte(&pPage1
->aData
[36], nFree
+1);
6080 if( pBt
->btsFlags
& BTS_SECURE_DELETE
){
6081 /* If the secure_delete option is enabled, then
6082 ** always fully overwrite deleted information with zeros.
6084 if( (!pPage
&& ((rc
= btreeGetPage(pBt
, iPage
, &pPage
, 0))!=0) )
6085 || ((rc
= sqlite3PagerWrite(pPage
->pDbPage
))!=0)
6089 memset(pPage
->aData
, 0, pPage
->pBt
->pageSize
);
6092 /* If the database supports auto-vacuum, write an entry in the pointer-map
6093 ** to indicate that the page is free.
6096 ptrmapPut(pBt
, iPage
, PTRMAP_FREEPAGE
, 0, &rc
);
6097 if( rc
) goto freepage_out
;
6100 /* Now manipulate the actual database free-list structure. There are two
6101 ** possibilities. If the free-list is currently empty, or if the first
6102 ** trunk page in the free-list is full, then this page will become a
6103 ** new free-list trunk page. Otherwise, it will become a leaf of the
6104 ** first trunk page in the current free-list. This block tests if it
6105 ** is possible to add the page as a new free-list leaf.
6108 u32 nLeaf
; /* Initial number of leaf cells on trunk page */
6110 iTrunk
= get4byte(&pPage1
->aData
[32]);
6111 rc
= btreeGetPage(pBt
, iTrunk
, &pTrunk
, 0);
6112 if( rc
!=SQLITE_OK
){
6116 nLeaf
= get4byte(&pTrunk
->aData
[4]);
6117 assert( pBt
->usableSize
>32 );
6118 if( nLeaf
> (u32
)pBt
->usableSize
/4 - 2 ){
6119 rc
= SQLITE_CORRUPT_BKPT
;
6122 if( nLeaf
< (u32
)pBt
->usableSize
/4 - 8 ){
6123 /* In this case there is room on the trunk page to insert the page
6124 ** being freed as a new leaf.
6126 ** Note that the trunk page is not really full until it contains
6127 ** usableSize/4 - 2 entries, not usableSize/4 - 8 entries as we have
6128 ** coded. But due to a coding error in versions of SQLite prior to
6129 ** 3.6.0, databases with freelist trunk pages holding more than
6130 ** usableSize/4 - 8 entries will be reported as corrupt. In order
6131 ** to maintain backwards compatibility with older versions of SQLite,
6132 ** we will continue to restrict the number of entries to usableSize/4 - 8
6133 ** for now. At some point in the future (once everyone has upgraded
6134 ** to 3.6.0 or later) we should consider fixing the conditional above
6135 ** to read "usableSize/4-2" instead of "usableSize/4-8".
6137 ** EVIDENCE-OF: R-19920-11576 However, newer versions of SQLite still
6138 ** avoid using the last six entries in the freelist trunk page array in
6139 ** order that database files created by newer versions of SQLite can be
6140 ** read by older versions of SQLite.
6142 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
6143 if( rc
==SQLITE_OK
){
6144 put4byte(&pTrunk
->aData
[4], nLeaf
+1);
6145 put4byte(&pTrunk
->aData
[8+nLeaf
*4], iPage
);
6146 if( pPage
&& (pBt
->btsFlags
& BTS_SECURE_DELETE
)==0 ){
6147 sqlite3PagerDontWrite(pPage
->pDbPage
);
6149 rc
= btreeSetHasContent(pBt
, iPage
);
6151 TRACE(("FREE-PAGE: %d leaf on trunk page %d\n",pPage
->pgno
,pTrunk
->pgno
));
6156 /* If control flows to this point, then it was not possible to add the
6157 ** the page being freed as a leaf page of the first trunk in the free-list.
6158 ** Possibly because the free-list is empty, or possibly because the
6159 ** first trunk in the free-list is full. Either way, the page being freed
6160 ** will become the new first trunk page in the free-list.
6162 if( pPage
==0 && SQLITE_OK
!=(rc
= btreeGetPage(pBt
, iPage
, &pPage
, 0)) ){
6165 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
6166 if( rc
!=SQLITE_OK
){
6169 put4byte(pPage
->aData
, iTrunk
);
6170 put4byte(&pPage
->aData
[4], 0);
6171 put4byte(&pPage1
->aData
[32], iPage
);
6172 TRACE(("FREE-PAGE: %d new trunk page replacing %d\n", pPage
->pgno
, iTrunk
));
6179 releasePage(pTrunk
);
6182 static void freePage(MemPage
*pPage
, int *pRC
){
6183 if( (*pRC
)==SQLITE_OK
){
6184 *pRC
= freePage2(pPage
->pBt
, pPage
, pPage
->pgno
);
6189 ** Free any overflow pages associated with the given Cell. Write the
6190 ** local Cell size (the number of bytes on the original page, omitting
6191 ** overflow) into *pnSize.
6193 static int clearCell(
6194 MemPage
*pPage
, /* The page that contains the Cell */
6195 unsigned char *pCell
, /* First byte of the Cell */
6196 CellInfo
*pInfo
/* Size information about the cell */
6204 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6205 pPage
->xParseCell(pPage
, pCell
, pInfo
);
6206 if( pInfo
->nLocal
==pInfo
->nPayload
){
6207 return SQLITE_OK
; /* No overflow pages. Return without doing anything */
6209 if( pCell
+pInfo
->nSize
-1 > pPage
->aData
+pPage
->maskPage
){
6210 /* Cell extends past end of page */
6211 return SQLITE_CORRUPT_PAGE(pPage
);
6213 ovflPgno
= get4byte(pCell
+ pInfo
->nSize
- 4);
6215 assert( pBt
->usableSize
> 4 );
6216 ovflPageSize
= pBt
->usableSize
- 4;
6217 nOvfl
= (pInfo
->nPayload
- pInfo
->nLocal
+ ovflPageSize
- 1)/ovflPageSize
;
6219 (CORRUPT_DB
&& (pInfo
->nPayload
+ ovflPageSize
)<ovflPageSize
)
6224 if( ovflPgno
<2 || ovflPgno
>btreePagecount(pBt
) ){
6225 /* 0 is not a legal page number and page 1 cannot be an
6226 ** overflow page. Therefore if ovflPgno<2 or past the end of the
6227 ** file the database must be corrupt. */
6228 return SQLITE_CORRUPT_BKPT
;
6231 rc
= getOverflowPage(pBt
, ovflPgno
, &pOvfl
, &iNext
);
6235 if( ( pOvfl
|| ((pOvfl
= btreePageLookup(pBt
, ovflPgno
))!=0) )
6236 && sqlite3PagerPageRefcount(pOvfl
->pDbPage
)!=1
6238 /* There is no reason any cursor should have an outstanding reference
6239 ** to an overflow page belonging to a cell that is being deleted/updated.
6240 ** So if there exists more than one reference to this page, then it
6241 ** must not really be an overflow page and the database must be corrupt.
6242 ** It is helpful to detect this before calling freePage2(), as
6243 ** freePage2() may zero the page contents if secure-delete mode is
6244 ** enabled. If this 'overflow' page happens to be a page that the
6245 ** caller is iterating through or using in some other way, this
6246 ** can be problematic.
6248 rc
= SQLITE_CORRUPT_BKPT
;
6250 rc
= freePage2(pBt
, pOvfl
, ovflPgno
);
6254 sqlite3PagerUnref(pOvfl
->pDbPage
);
6263 ** Create the byte sequence used to represent a cell on page pPage
6264 ** and write that byte sequence into pCell[]. Overflow pages are
6265 ** allocated and filled in as necessary. The calling procedure
6266 ** is responsible for making sure sufficient space has been allocated
6269 ** Note that pCell does not necessary need to point to the pPage->aData
6270 ** area. pCell might point to some temporary storage. The cell will
6271 ** be constructed in this temporary area then copied into pPage->aData
6274 static int fillInCell(
6275 MemPage
*pPage
, /* The page that contains the cell */
6276 unsigned char *pCell
, /* Complete text of the cell */
6277 const BtreePayload
*pX
, /* Payload with which to construct the cell */
6278 int *pnSize
/* Write cell size here */
6282 int nSrc
, n
, rc
, mn
;
6284 MemPage
*pToRelease
;
6285 unsigned char *pPrior
;
6286 unsigned char *pPayload
;
6291 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6293 /* pPage is not necessarily writeable since pCell might be auxiliary
6294 ** buffer space that is separate from the pPage buffer area */
6295 assert( pCell
<pPage
->aData
|| pCell
>=&pPage
->aData
[pPage
->pBt
->pageSize
]
6296 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6298 /* Fill in the header. */
6299 nHeader
= pPage
->childPtrSize
;
6300 if( pPage
->intKey
){
6301 nPayload
= pX
->nData
+ pX
->nZero
;
6304 assert( pPage
->intKeyLeaf
); /* fillInCell() only called for leaves */
6305 nHeader
+= putVarint32(&pCell
[nHeader
], nPayload
);
6306 nHeader
+= putVarint(&pCell
[nHeader
], *(u64
*)&pX
->nKey
);
6308 assert( pX
->nKey
<=0x7fffffff && pX
->pKey
!=0 );
6309 nSrc
= nPayload
= (int)pX
->nKey
;
6311 nHeader
+= putVarint32(&pCell
[nHeader
], nPayload
);
6314 /* Fill in the payload */
6315 pPayload
= &pCell
[nHeader
];
6316 if( nPayload
<=pPage
->maxLocal
){
6317 /* This is the common case where everything fits on the btree page
6318 ** and no overflow pages are required. */
6319 n
= nHeader
+ nPayload
;
6324 assert( nSrc
<=nPayload
);
6325 testcase( nSrc
<nPayload
);
6326 memcpy(pPayload
, pSrc
, nSrc
);
6327 memset(pPayload
+nSrc
, 0, nPayload
-nSrc
);
6331 /* If we reach this point, it means that some of the content will need
6332 ** to spill onto overflow pages.
6334 mn
= pPage
->minLocal
;
6335 n
= mn
+ (nPayload
- mn
) % (pPage
->pBt
->usableSize
- 4);
6336 testcase( n
==pPage
->maxLocal
);
6337 testcase( n
==pPage
->maxLocal
+1 );
6338 if( n
> pPage
->maxLocal
) n
= mn
;
6340 *pnSize
= n
+ nHeader
+ 4;
6341 pPrior
= &pCell
[nHeader
+n
];
6346 /* At this point variables should be set as follows:
6348 ** nPayload Total payload size in bytes
6349 ** pPayload Begin writing payload here
6350 ** spaceLeft Space available at pPayload. If nPayload>spaceLeft,
6351 ** that means content must spill into overflow pages.
6352 ** *pnSize Size of the local cell (not counting overflow pages)
6353 ** pPrior Where to write the pgno of the first overflow page
6355 ** Use a call to btreeParseCellPtr() to verify that the values above
6356 ** were computed correctly.
6361 pPage
->xParseCell(pPage
, pCell
, &info
);
6362 assert( nHeader
==(int)(info
.pPayload
- pCell
) );
6363 assert( info
.nKey
==pX
->nKey
);
6364 assert( *pnSize
== info
.nSize
);
6365 assert( spaceLeft
== info
.nLocal
);
6369 /* Write the payload into the local Cell and any extra into overflow pages */
6372 if( n
>spaceLeft
) n
= spaceLeft
;
6374 /* If pToRelease is not zero than pPayload points into the data area
6375 ** of pToRelease. Make sure pToRelease is still writeable. */
6376 assert( pToRelease
==0 || sqlite3PagerIswriteable(pToRelease
->pDbPage
) );
6378 /* If pPayload is part of the data area of pPage, then make sure pPage
6379 ** is still writeable */
6380 assert( pPayload
<pPage
->aData
|| pPayload
>=&pPage
->aData
[pBt
->pageSize
]
6381 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6384 memcpy(pPayload
, pSrc
, n
);
6387 memcpy(pPayload
, pSrc
, n
);
6389 memset(pPayload
, 0, n
);
6392 if( nPayload
<=0 ) break;
6399 #ifndef SQLITE_OMIT_AUTOVACUUM
6400 Pgno pgnoPtrmap
= pgnoOvfl
; /* Overflow page pointer-map entry page */
6401 if( pBt
->autoVacuum
){
6405 PTRMAP_ISPAGE(pBt
, pgnoOvfl
) || pgnoOvfl
==PENDING_BYTE_PAGE(pBt
)
6409 rc
= allocateBtreePage(pBt
, &pOvfl
, &pgnoOvfl
, pgnoOvfl
, 0);
6410 #ifndef SQLITE_OMIT_AUTOVACUUM
6411 /* If the database supports auto-vacuum, and the second or subsequent
6412 ** overflow page is being allocated, add an entry to the pointer-map
6413 ** for that page now.
6415 ** If this is the first overflow page, then write a partial entry
6416 ** to the pointer-map. If we write nothing to this pointer-map slot,
6417 ** then the optimistic overflow chain processing in clearCell()
6418 ** may misinterpret the uninitialized values and delete the
6419 ** wrong pages from the database.
6421 if( pBt
->autoVacuum
&& rc
==SQLITE_OK
){
6422 u8 eType
= (pgnoPtrmap
?PTRMAP_OVERFLOW2
:PTRMAP_OVERFLOW1
);
6423 ptrmapPut(pBt
, pgnoOvfl
, eType
, pgnoPtrmap
, &rc
);
6430 releasePage(pToRelease
);
6434 /* If pToRelease is not zero than pPrior points into the data area
6435 ** of pToRelease. Make sure pToRelease is still writeable. */
6436 assert( pToRelease
==0 || sqlite3PagerIswriteable(pToRelease
->pDbPage
) );
6438 /* If pPrior is part of the data area of pPage, then make sure pPage
6439 ** is still writeable */
6440 assert( pPrior
<pPage
->aData
|| pPrior
>=&pPage
->aData
[pBt
->pageSize
]
6441 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6443 put4byte(pPrior
, pgnoOvfl
);
6444 releasePage(pToRelease
);
6446 pPrior
= pOvfl
->aData
;
6447 put4byte(pPrior
, 0);
6448 pPayload
= &pOvfl
->aData
[4];
6449 spaceLeft
= pBt
->usableSize
- 4;
6452 releasePage(pToRelease
);
6457 ** Remove the i-th cell from pPage. This routine effects pPage only.
6458 ** The cell content is not freed or deallocated. It is assumed that
6459 ** the cell content has been copied someplace else. This routine just
6460 ** removes the reference to the cell from pPage.
6462 ** "sz" must be the number of bytes in the cell.
6464 static void dropCell(MemPage
*pPage
, int idx
, int sz
, int *pRC
){
6465 u32 pc
; /* Offset to cell content of cell being deleted */
6466 u8
*data
; /* pPage->aData */
6467 u8
*ptr
; /* Used to move bytes around within data[] */
6468 int rc
; /* The return code */
6469 int hdr
; /* Beginning of the header. 0 most pages. 100 page 1 */
6472 assert( idx
>=0 && idx
<pPage
->nCell
);
6473 assert( CORRUPT_DB
|| sz
==cellSize(pPage
, idx
) );
6474 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
6475 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6476 data
= pPage
->aData
;
6477 ptr
= &pPage
->aCellIdx
[2*idx
];
6479 hdr
= pPage
->hdrOffset
;
6480 testcase( pc
==get2byte(&data
[hdr
+5]) );
6481 testcase( pc
+sz
==pPage
->pBt
->usableSize
);
6482 if( pc
+sz
> pPage
->pBt
->usableSize
){
6483 *pRC
= SQLITE_CORRUPT_BKPT
;
6486 rc
= freeSpace(pPage
, pc
, sz
);
6492 if( pPage
->nCell
==0 ){
6493 memset(&data
[hdr
+1], 0, 4);
6495 put2byte(&data
[hdr
+5], pPage
->pBt
->usableSize
);
6496 pPage
->nFree
= pPage
->pBt
->usableSize
- pPage
->hdrOffset
6497 - pPage
->childPtrSize
- 8;
6499 memmove(ptr
, ptr
+2, 2*(pPage
->nCell
- idx
));
6500 put2byte(&data
[hdr
+3], pPage
->nCell
);
6506 ** Insert a new cell on pPage at cell index "i". pCell points to the
6507 ** content of the cell.
6509 ** If the cell content will fit on the page, then put it there. If it
6510 ** will not fit, then make a copy of the cell content into pTemp if
6511 ** pTemp is not null. Regardless of pTemp, allocate a new entry
6512 ** in pPage->apOvfl[] and make it point to the cell content (either
6513 ** in pTemp or the original pCell) and also record its index.
6514 ** Allocating a new entry in pPage->aCell[] implies that
6515 ** pPage->nOverflow is incremented.
6517 ** *pRC must be SQLITE_OK when this routine is called.
6519 static void insertCell(
6520 MemPage
*pPage
, /* Page into which we are copying */
6521 int i
, /* New cell becomes the i-th cell of the page */
6522 u8
*pCell
, /* Content of the new cell */
6523 int sz
, /* Bytes of content in pCell */
6524 u8
*pTemp
, /* Temp storage space for pCell, if needed */
6525 Pgno iChild
, /* If non-zero, replace first 4 bytes with this value */
6526 int *pRC
/* Read and write return code from here */
6528 int idx
= 0; /* Where to write new cell content in data[] */
6529 int j
; /* Loop counter */
6530 u8
*data
; /* The content of the whole page */
6531 u8
*pIns
; /* The point in pPage->aCellIdx[] where no cell inserted */
6533 assert( *pRC
==SQLITE_OK
);
6534 assert( i
>=0 && i
<=pPage
->nCell
+pPage
->nOverflow
);
6535 assert( MX_CELL(pPage
->pBt
)<=10921 );
6536 assert( pPage
->nCell
<=MX_CELL(pPage
->pBt
) || CORRUPT_DB
);
6537 assert( pPage
->nOverflow
<=ArraySize(pPage
->apOvfl
) );
6538 assert( ArraySize(pPage
->apOvfl
)==ArraySize(pPage
->aiOvfl
) );
6539 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6540 /* The cell should normally be sized correctly. However, when moving a
6541 ** malformed cell from a leaf page to an interior page, if the cell size
6542 ** wanted to be less than 4 but got rounded up to 4 on the leaf, then size
6543 ** might be less than 8 (leaf-size + pointer) on the interior node. Hence
6544 ** the term after the || in the following assert(). */
6545 assert( sz
==pPage
->xCellSize(pPage
, pCell
) || (sz
==8 && iChild
>0) );
6546 if( pPage
->nOverflow
|| sz
+2>pPage
->nFree
){
6548 memcpy(pTemp
, pCell
, sz
);
6552 put4byte(pCell
, iChild
);
6554 j
= pPage
->nOverflow
++;
6555 /* Comparison against ArraySize-1 since we hold back one extra slot
6556 ** as a contingency. In other words, never need more than 3 overflow
6557 ** slots but 4 are allocated, just to be safe. */
6558 assert( j
< ArraySize(pPage
->apOvfl
)-1 );
6559 pPage
->apOvfl
[j
] = pCell
;
6560 pPage
->aiOvfl
[j
] = (u16
)i
;
6562 /* When multiple overflows occur, they are always sequential and in
6563 ** sorted order. This invariants arise because multiple overflows can
6564 ** only occur when inserting divider cells into the parent page during
6565 ** balancing, and the dividers are adjacent and sorted.
6567 assert( j
==0 || pPage
->aiOvfl
[j
-1]<(u16
)i
); /* Overflows in sorted order */
6568 assert( j
==0 || i
==pPage
->aiOvfl
[j
-1]+1 ); /* Overflows are sequential */
6570 int rc
= sqlite3PagerWrite(pPage
->pDbPage
);
6571 if( rc
!=SQLITE_OK
){
6575 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
6576 data
= pPage
->aData
;
6577 assert( &data
[pPage
->cellOffset
]==pPage
->aCellIdx
);
6578 rc
= allocateSpace(pPage
, sz
, &idx
);
6579 if( rc
){ *pRC
= rc
; return; }
6580 /* The allocateSpace() routine guarantees the following properties
6581 ** if it returns successfully */
6583 assert( idx
>= pPage
->cellOffset
+2*pPage
->nCell
+2 || CORRUPT_DB
);
6584 assert( idx
+sz
<= (int)pPage
->pBt
->usableSize
);
6585 pPage
->nFree
-= (u16
)(2 + sz
);
6586 memcpy(&data
[idx
], pCell
, sz
);
6588 put4byte(&data
[idx
], iChild
);
6590 pIns
= pPage
->aCellIdx
+ i
*2;
6591 memmove(pIns
+2, pIns
, 2*(pPage
->nCell
- i
));
6592 put2byte(pIns
, idx
);
6594 /* increment the cell count */
6595 if( (++data
[pPage
->hdrOffset
+4])==0 ) data
[pPage
->hdrOffset
+3]++;
6596 assert( get2byte(&data
[pPage
->hdrOffset
+3])==pPage
->nCell
);
6597 #ifndef SQLITE_OMIT_AUTOVACUUM
6598 if( pPage
->pBt
->autoVacuum
){
6599 /* The cell may contain a pointer to an overflow page. If so, write
6600 ** the entry for the overflow page into the pointer map.
6602 ptrmapPutOvflPtr(pPage
, pCell
, pRC
);
6609 ** A CellArray object contains a cache of pointers and sizes for a
6610 ** consecutive sequence of cells that might be held on multiple pages.
6612 typedef struct CellArray CellArray
;
6614 int nCell
; /* Number of cells in apCell[] */
6615 MemPage
*pRef
; /* Reference page */
6616 u8
**apCell
; /* All cells begin balanced */
6617 u16
*szCell
; /* Local size of all cells in apCell[] */
6621 ** Make sure the cell sizes at idx, idx+1, ..., idx+N-1 have been
6624 static void populateCellCache(CellArray
*p
, int idx
, int N
){
6625 assert( idx
>=0 && idx
+N
<=p
->nCell
);
6627 assert( p
->apCell
[idx
]!=0 );
6628 if( p
->szCell
[idx
]==0 ){
6629 p
->szCell
[idx
] = p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[idx
]);
6631 assert( CORRUPT_DB
||
6632 p
->szCell
[idx
]==p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[idx
]) );
6640 ** Return the size of the Nth element of the cell array
6642 static SQLITE_NOINLINE u16
computeCellSize(CellArray
*p
, int N
){
6643 assert( N
>=0 && N
<p
->nCell
);
6644 assert( p
->szCell
[N
]==0 );
6645 p
->szCell
[N
] = p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[N
]);
6646 return p
->szCell
[N
];
6648 static u16
cachedCellSize(CellArray
*p
, int N
){
6649 assert( N
>=0 && N
<p
->nCell
);
6650 if( p
->szCell
[N
] ) return p
->szCell
[N
];
6651 return computeCellSize(p
, N
);
6655 ** Array apCell[] contains pointers to nCell b-tree page cells. The
6656 ** szCell[] array contains the size in bytes of each cell. This function
6657 ** replaces the current contents of page pPg with the contents of the cell
6660 ** Some of the cells in apCell[] may currently be stored in pPg. This
6661 ** function works around problems caused by this by making a copy of any
6662 ** such cells before overwriting the page data.
6664 ** The MemPage.nFree field is invalidated by this function. It is the
6665 ** responsibility of the caller to set it correctly.
6667 static int rebuildPage(
6668 MemPage
*pPg
, /* Edit this page */
6669 int nCell
, /* Final number of cells on page */
6670 u8
**apCell
, /* Array of cells */
6671 u16
*szCell
/* Array of cell sizes */
6673 const int hdr
= pPg
->hdrOffset
; /* Offset of header on pPg */
6674 u8
* const aData
= pPg
->aData
; /* Pointer to data for pPg */
6675 const int usableSize
= pPg
->pBt
->usableSize
;
6676 u8
* const pEnd
= &aData
[usableSize
];
6678 u8
*pCellptr
= pPg
->aCellIdx
;
6679 u8
*pTmp
= sqlite3PagerTempSpace(pPg
->pBt
->pPager
);
6682 i
= get2byte(&aData
[hdr
+5]);
6683 memcpy(&pTmp
[i
], &aData
[i
], usableSize
- i
);
6686 for(i
=0; i
<nCell
; i
++){
6687 u8
*pCell
= apCell
[i
];
6688 if( SQLITE_WITHIN(pCell
,aData
,pEnd
) ){
6689 pCell
= &pTmp
[pCell
- aData
];
6692 put2byte(pCellptr
, (pData
- aData
));
6694 if( pData
< pCellptr
) return SQLITE_CORRUPT_BKPT
;
6695 memcpy(pData
, pCell
, szCell
[i
]);
6696 assert( szCell
[i
]==pPg
->xCellSize(pPg
, pCell
) || CORRUPT_DB
);
6697 testcase( szCell
[i
]!=pPg
->xCellSize(pPg
,pCell
) );
6700 /* The pPg->nFree field is now set incorrectly. The caller will fix it. */
6704 put2byte(&aData
[hdr
+1], 0);
6705 put2byte(&aData
[hdr
+3], pPg
->nCell
);
6706 put2byte(&aData
[hdr
+5], pData
- aData
);
6707 aData
[hdr
+7] = 0x00;
6712 ** Array apCell[] contains nCell pointers to b-tree cells. Array szCell
6713 ** contains the size in bytes of each such cell. This function attempts to
6714 ** add the cells stored in the array to page pPg. If it cannot (because
6715 ** the page needs to be defragmented before the cells will fit), non-zero
6716 ** is returned. Otherwise, if the cells are added successfully, zero is
6719 ** Argument pCellptr points to the first entry in the cell-pointer array
6720 ** (part of page pPg) to populate. After cell apCell[0] is written to the
6721 ** page body, a 16-bit offset is written to pCellptr. And so on, for each
6722 ** cell in the array. It is the responsibility of the caller to ensure
6723 ** that it is safe to overwrite this part of the cell-pointer array.
6725 ** When this function is called, *ppData points to the start of the
6726 ** content area on page pPg. If the size of the content area is extended,
6727 ** *ppData is updated to point to the new start of the content area
6728 ** before returning.
6730 ** Finally, argument pBegin points to the byte immediately following the
6731 ** end of the space required by this page for the cell-pointer area (for
6732 ** all cells - not just those inserted by the current call). If the content
6733 ** area must be extended to before this point in order to accomodate all
6734 ** cells in apCell[], then the cells do not fit and non-zero is returned.
6736 static int pageInsertArray(
6737 MemPage
*pPg
, /* Page to add cells to */
6738 u8
*pBegin
, /* End of cell-pointer array */
6739 u8
**ppData
, /* IN/OUT: Page content -area pointer */
6740 u8
*pCellptr
, /* Pointer to cell-pointer area */
6741 int iFirst
, /* Index of first cell to add */
6742 int nCell
, /* Number of cells to add to pPg */
6743 CellArray
*pCArray
/* Array of cells */
6746 u8
*aData
= pPg
->aData
;
6747 u8
*pData
= *ppData
;
6748 int iEnd
= iFirst
+ nCell
;
6749 assert( CORRUPT_DB
|| pPg
->hdrOffset
==0 ); /* Never called on page 1 */
6750 for(i
=iFirst
; i
<iEnd
; i
++){
6753 sz
= cachedCellSize(pCArray
, i
);
6754 if( (aData
[1]==0 && aData
[2]==0) || (pSlot
= pageFindSlot(pPg
,sz
,&rc
))==0 ){
6755 if( (pData
- pBegin
)<sz
) return 1;
6759 /* pSlot and pCArray->apCell[i] will never overlap on a well-formed
6760 ** database. But they might for a corrupt database. Hence use memmove()
6761 ** since memcpy() sends SIGABORT with overlapping buffers on OpenBSD */
6762 assert( (pSlot
+sz
)<=pCArray
->apCell
[i
]
6763 || pSlot
>=(pCArray
->apCell
[i
]+sz
)
6765 memmove(pSlot
, pCArray
->apCell
[i
], sz
);
6766 put2byte(pCellptr
, (pSlot
- aData
));
6774 ** Array apCell[] contains nCell pointers to b-tree cells. Array szCell
6775 ** contains the size in bytes of each such cell. This function adds the
6776 ** space associated with each cell in the array that is currently stored
6777 ** within the body of pPg to the pPg free-list. The cell-pointers and other
6778 ** fields of the page are not updated.
6780 ** This function returns the total number of cells added to the free-list.
6782 static int pageFreeArray(
6783 MemPage
*pPg
, /* Page to edit */
6784 int iFirst
, /* First cell to delete */
6785 int nCell
, /* Cells to delete */
6786 CellArray
*pCArray
/* Array of cells */
6788 u8
* const aData
= pPg
->aData
;
6789 u8
* const pEnd
= &aData
[pPg
->pBt
->usableSize
];
6790 u8
* const pStart
= &aData
[pPg
->hdrOffset
+ 8 + pPg
->childPtrSize
];
6793 int iEnd
= iFirst
+ nCell
;
6797 for(i
=iFirst
; i
<iEnd
; i
++){
6798 u8
*pCell
= pCArray
->apCell
[i
];
6799 if( SQLITE_WITHIN(pCell
, pStart
, pEnd
) ){
6801 /* No need to use cachedCellSize() here. The sizes of all cells that
6802 ** are to be freed have already been computing while deciding which
6803 ** cells need freeing */
6804 sz
= pCArray
->szCell
[i
]; assert( sz
>0 );
6805 if( pFree
!=(pCell
+ sz
) ){
6807 assert( pFree
>aData
&& (pFree
- aData
)<65536 );
6808 freeSpace(pPg
, (u16
)(pFree
- aData
), szFree
);
6812 if( pFree
+sz
>pEnd
) return 0;
6821 assert( pFree
>aData
&& (pFree
- aData
)<65536 );
6822 freeSpace(pPg
, (u16
)(pFree
- aData
), szFree
);
6828 ** apCell[] and szCell[] contains pointers to and sizes of all cells in the
6829 ** pages being balanced. The current page, pPg, has pPg->nCell cells starting
6830 ** with apCell[iOld]. After balancing, this page should hold nNew cells
6831 ** starting at apCell[iNew].
6833 ** This routine makes the necessary adjustments to pPg so that it contains
6834 ** the correct cells after being balanced.
6836 ** The pPg->nFree field is invalid when this function returns. It is the
6837 ** responsibility of the caller to set it correctly.
6839 static int editPage(
6840 MemPage
*pPg
, /* Edit this page */
6841 int iOld
, /* Index of first cell currently on page */
6842 int iNew
, /* Index of new first cell on page */
6843 int nNew
, /* Final number of cells on page */
6844 CellArray
*pCArray
/* Array of cells and sizes */
6846 u8
* const aData
= pPg
->aData
;
6847 const int hdr
= pPg
->hdrOffset
;
6848 u8
*pBegin
= &pPg
->aCellIdx
[nNew
* 2];
6849 int nCell
= pPg
->nCell
; /* Cells stored on pPg */
6853 int iOldEnd
= iOld
+ pPg
->nCell
+ pPg
->nOverflow
;
6854 int iNewEnd
= iNew
+ nNew
;
6857 u8
*pTmp
= sqlite3PagerTempSpace(pPg
->pBt
->pPager
);
6858 memcpy(pTmp
, aData
, pPg
->pBt
->usableSize
);
6861 /* Remove cells from the start and end of the page */
6863 int nShift
= pageFreeArray(pPg
, iOld
, iNew
-iOld
, pCArray
);
6864 memmove(pPg
->aCellIdx
, &pPg
->aCellIdx
[nShift
*2], nCell
*2);
6867 if( iNewEnd
< iOldEnd
){
6868 nCell
-= pageFreeArray(pPg
, iNewEnd
, iOldEnd
- iNewEnd
, pCArray
);
6871 pData
= &aData
[get2byteNotZero(&aData
[hdr
+5])];
6872 if( pData
<pBegin
) goto editpage_fail
;
6874 /* Add cells to the start of the page */
6876 int nAdd
= MIN(nNew
,iOld
-iNew
);
6877 assert( (iOld
-iNew
)<nNew
|| nCell
==0 || CORRUPT_DB
);
6878 pCellptr
= pPg
->aCellIdx
;
6879 memmove(&pCellptr
[nAdd
*2], pCellptr
, nCell
*2);
6880 if( pageInsertArray(
6881 pPg
, pBegin
, &pData
, pCellptr
,
6883 ) ) goto editpage_fail
;
6887 /* Add any overflow cells */
6888 for(i
=0; i
<pPg
->nOverflow
; i
++){
6889 int iCell
= (iOld
+ pPg
->aiOvfl
[i
]) - iNew
;
6890 if( iCell
>=0 && iCell
<nNew
){
6891 pCellptr
= &pPg
->aCellIdx
[iCell
* 2];
6892 memmove(&pCellptr
[2], pCellptr
, (nCell
- iCell
) * 2);
6894 if( pageInsertArray(
6895 pPg
, pBegin
, &pData
, pCellptr
,
6896 iCell
+iNew
, 1, pCArray
6897 ) ) goto editpage_fail
;
6901 /* Append cells to the end of the page */
6902 pCellptr
= &pPg
->aCellIdx
[nCell
*2];
6903 if( pageInsertArray(
6904 pPg
, pBegin
, &pData
, pCellptr
,
6905 iNew
+nCell
, nNew
-nCell
, pCArray
6906 ) ) goto editpage_fail
;
6911 put2byte(&aData
[hdr
+3], pPg
->nCell
);
6912 put2byte(&aData
[hdr
+5], pData
- aData
);
6915 for(i
=0; i
<nNew
&& !CORRUPT_DB
; i
++){
6916 u8
*pCell
= pCArray
->apCell
[i
+iNew
];
6917 int iOff
= get2byteAligned(&pPg
->aCellIdx
[i
*2]);
6918 if( SQLITE_WITHIN(pCell
, aData
, &aData
[pPg
->pBt
->usableSize
]) ){
6919 pCell
= &pTmp
[pCell
- aData
];
6921 assert( 0==memcmp(pCell
, &aData
[iOff
],
6922 pCArray
->pRef
->xCellSize(pCArray
->pRef
, pCArray
->apCell
[i
+iNew
])) );
6928 /* Unable to edit this page. Rebuild it from scratch instead. */
6929 populateCellCache(pCArray
, iNew
, nNew
);
6930 return rebuildPage(pPg
, nNew
, &pCArray
->apCell
[iNew
], &pCArray
->szCell
[iNew
]);
6934 ** The following parameters determine how many adjacent pages get involved
6935 ** in a balancing operation. NN is the number of neighbors on either side
6936 ** of the page that participate in the balancing operation. NB is the
6937 ** total number of pages that participate, including the target page and
6938 ** NN neighbors on either side.
6940 ** The minimum value of NN is 1 (of course). Increasing NN above 1
6941 ** (to 2 or 3) gives a modest improvement in SELECT and DELETE performance
6942 ** in exchange for a larger degradation in INSERT and UPDATE performance.
6943 ** The value of NN appears to give the best results overall.
6945 #define NN 1 /* Number of neighbors on either side of pPage */
6946 #define NB (NN*2+1) /* Total pages involved in the balance */
6949 #ifndef SQLITE_OMIT_QUICKBALANCE
6951 ** This version of balance() handles the common special case where
6952 ** a new entry is being inserted on the extreme right-end of the
6953 ** tree, in other words, when the new entry will become the largest
6954 ** entry in the tree.
6956 ** Instead of trying to balance the 3 right-most leaf pages, just add
6957 ** a new page to the right-hand side and put the one new entry in
6958 ** that page. This leaves the right side of the tree somewhat
6959 ** unbalanced. But odds are that we will be inserting new entries
6960 ** at the end soon afterwards so the nearly empty page will quickly
6961 ** fill up. On average.
6963 ** pPage is the leaf page which is the right-most page in the tree.
6964 ** pParent is its parent. pPage must have a single overflow entry
6965 ** which is also the right-most entry on the page.
6967 ** The pSpace buffer is used to store a temporary copy of the divider
6968 ** cell that will be inserted into pParent. Such a cell consists of a 4
6969 ** byte page number followed by a variable length integer. In other
6970 ** words, at most 13 bytes. Hence the pSpace buffer must be at
6971 ** least 13 bytes in size.
6973 static int balance_quick(MemPage
*pParent
, MemPage
*pPage
, u8
*pSpace
){
6974 BtShared
*const pBt
= pPage
->pBt
; /* B-Tree Database */
6975 MemPage
*pNew
; /* Newly allocated page */
6976 int rc
; /* Return Code */
6977 Pgno pgnoNew
; /* Page number of pNew */
6979 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6980 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
6981 assert( pPage
->nOverflow
==1 );
6983 /* This error condition is now caught prior to reaching this function */
6984 if( NEVER(pPage
->nCell
==0) ) return SQLITE_CORRUPT_BKPT
;
6986 /* Allocate a new page. This page will become the right-sibling of
6987 ** pPage. Make the parent page writable, so that the new divider cell
6988 ** may be inserted. If both these operations are successful, proceed.
6990 rc
= allocateBtreePage(pBt
, &pNew
, &pgnoNew
, 0, 0);
6992 if( rc
==SQLITE_OK
){
6994 u8
*pOut
= &pSpace
[4];
6995 u8
*pCell
= pPage
->apOvfl
[0];
6996 u16 szCell
= pPage
->xCellSize(pPage
, pCell
);
6999 assert( sqlite3PagerIswriteable(pNew
->pDbPage
) );
7000 assert( pPage
->aData
[0]==(PTF_INTKEY
|PTF_LEAFDATA
|PTF_LEAF
) );
7001 zeroPage(pNew
, PTF_INTKEY
|PTF_LEAFDATA
|PTF_LEAF
);
7002 rc
= rebuildPage(pNew
, 1, &pCell
, &szCell
);
7003 if( NEVER(rc
) ) return rc
;
7004 pNew
->nFree
= pBt
->usableSize
- pNew
->cellOffset
- 2 - szCell
;
7006 /* If this is an auto-vacuum database, update the pointer map
7007 ** with entries for the new page, and any pointer from the
7008 ** cell on the page to an overflow page. If either of these
7009 ** operations fails, the return code is set, but the contents
7010 ** of the parent page are still manipulated by thh code below.
7011 ** That is Ok, at this point the parent page is guaranteed to
7012 ** be marked as dirty. Returning an error code will cause a
7013 ** rollback, undoing any changes made to the parent page.
7016 ptrmapPut(pBt
, pgnoNew
, PTRMAP_BTREE
, pParent
->pgno
, &rc
);
7017 if( szCell
>pNew
->minLocal
){
7018 ptrmapPutOvflPtr(pNew
, pCell
, &rc
);
7022 /* Create a divider cell to insert into pParent. The divider cell
7023 ** consists of a 4-byte page number (the page number of pPage) and
7024 ** a variable length key value (which must be the same value as the
7025 ** largest key on pPage).
7027 ** To find the largest key value on pPage, first find the right-most
7028 ** cell on pPage. The first two fields of this cell are the
7029 ** record-length (a variable length integer at most 32-bits in size)
7030 ** and the key value (a variable length integer, may have any value).
7031 ** The first of the while(...) loops below skips over the record-length
7032 ** field. The second while(...) loop copies the key value from the
7033 ** cell on pPage into the pSpace buffer.
7035 pCell
= findCell(pPage
, pPage
->nCell
-1);
7037 while( (*(pCell
++)&0x80) && pCell
<pStop
);
7039 while( ((*(pOut
++) = *(pCell
++))&0x80) && pCell
<pStop
);
7041 /* Insert the new divider cell into pParent. */
7042 if( rc
==SQLITE_OK
){
7043 insertCell(pParent
, pParent
->nCell
, pSpace
, (int)(pOut
-pSpace
),
7044 0, pPage
->pgno
, &rc
);
7047 /* Set the right-child pointer of pParent to point to the new page. */
7048 put4byte(&pParent
->aData
[pParent
->hdrOffset
+8], pgnoNew
);
7050 /* Release the reference to the new page. */
7056 #endif /* SQLITE_OMIT_QUICKBALANCE */
7060 ** This function does not contribute anything to the operation of SQLite.
7061 ** it is sometimes activated temporarily while debugging code responsible
7062 ** for setting pointer-map entries.
7064 static int ptrmapCheckPages(MemPage
**apPage
, int nPage
){
7066 for(i
=0; i
<nPage
; i
++){
7069 MemPage
*pPage
= apPage
[i
];
7070 BtShared
*pBt
= pPage
->pBt
;
7071 assert( pPage
->isInit
);
7073 for(j
=0; j
<pPage
->nCell
; j
++){
7077 z
= findCell(pPage
, j
);
7078 pPage
->xParseCell(pPage
, z
, &info
);
7079 if( info
.nLocal
<info
.nPayload
){
7080 Pgno ovfl
= get4byte(&z
[info
.nSize
-4]);
7081 ptrmapGet(pBt
, ovfl
, &e
, &n
);
7082 assert( n
==pPage
->pgno
&& e
==PTRMAP_OVERFLOW1
);
7085 Pgno child
= get4byte(z
);
7086 ptrmapGet(pBt
, child
, &e
, &n
);
7087 assert( n
==pPage
->pgno
&& e
==PTRMAP_BTREE
);
7091 Pgno child
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
7092 ptrmapGet(pBt
, child
, &e
, &n
);
7093 assert( n
==pPage
->pgno
&& e
==PTRMAP_BTREE
);
7101 ** This function is used to copy the contents of the b-tree node stored
7102 ** on page pFrom to page pTo. If page pFrom was not a leaf page, then
7103 ** the pointer-map entries for each child page are updated so that the
7104 ** parent page stored in the pointer map is page pTo. If pFrom contained
7105 ** any cells with overflow page pointers, then the corresponding pointer
7106 ** map entries are also updated so that the parent page is page pTo.
7108 ** If pFrom is currently carrying any overflow cells (entries in the
7109 ** MemPage.apOvfl[] array), they are not copied to pTo.
7111 ** Before returning, page pTo is reinitialized using btreeInitPage().
7113 ** The performance of this function is not critical. It is only used by
7114 ** the balance_shallower() and balance_deeper() procedures, neither of
7115 ** which are called often under normal circumstances.
7117 static void copyNodeContent(MemPage
*pFrom
, MemPage
*pTo
, int *pRC
){
7118 if( (*pRC
)==SQLITE_OK
){
7119 BtShared
* const pBt
= pFrom
->pBt
;
7120 u8
* const aFrom
= pFrom
->aData
;
7121 u8
* const aTo
= pTo
->aData
;
7122 int const iFromHdr
= pFrom
->hdrOffset
;
7123 int const iToHdr
= ((pTo
->pgno
==1) ? 100 : 0);
7128 assert( pFrom
->isInit
);
7129 assert( pFrom
->nFree
>=iToHdr
);
7130 assert( get2byte(&aFrom
[iFromHdr
+5]) <= (int)pBt
->usableSize
);
7132 /* Copy the b-tree node content from page pFrom to page pTo. */
7133 iData
= get2byte(&aFrom
[iFromHdr
+5]);
7134 memcpy(&aTo
[iData
], &aFrom
[iData
], pBt
->usableSize
-iData
);
7135 memcpy(&aTo
[iToHdr
], &aFrom
[iFromHdr
], pFrom
->cellOffset
+ 2*pFrom
->nCell
);
7137 /* Reinitialize page pTo so that the contents of the MemPage structure
7138 ** match the new data. The initialization of pTo can actually fail under
7139 ** fairly obscure circumstances, even though it is a copy of initialized
7143 rc
= btreeInitPage(pTo
);
7144 if( rc
!=SQLITE_OK
){
7149 /* If this is an auto-vacuum database, update the pointer-map entries
7150 ** for any b-tree or overflow pages that pTo now contains the pointers to.
7153 *pRC
= setChildPtrmaps(pTo
);
7159 ** This routine redistributes cells on the iParentIdx'th child of pParent
7160 ** (hereafter "the page") and up to 2 siblings so that all pages have about the
7161 ** same amount of free space. Usually a single sibling on either side of the
7162 ** page are used in the balancing, though both siblings might come from one
7163 ** side if the page is the first or last child of its parent. If the page
7164 ** has fewer than 2 siblings (something which can only happen if the page
7165 ** is a root page or a child of a root page) then all available siblings
7166 ** participate in the balancing.
7168 ** The number of siblings of the page might be increased or decreased by
7169 ** one or two in an effort to keep pages nearly full but not over full.
7171 ** Note that when this routine is called, some of the cells on the page
7172 ** might not actually be stored in MemPage.aData[]. This can happen
7173 ** if the page is overfull. This routine ensures that all cells allocated
7174 ** to the page and its siblings fit into MemPage.aData[] before returning.
7176 ** In the course of balancing the page and its siblings, cells may be
7177 ** inserted into or removed from the parent page (pParent). Doing so
7178 ** may cause the parent page to become overfull or underfull. If this
7179 ** happens, it is the responsibility of the caller to invoke the correct
7180 ** balancing routine to fix this problem (see the balance() routine).
7182 ** If this routine fails for any reason, it might leave the database
7183 ** in a corrupted state. So if this routine fails, the database should
7186 ** The third argument to this function, aOvflSpace, is a pointer to a
7187 ** buffer big enough to hold one page. If while inserting cells into the parent
7188 ** page (pParent) the parent page becomes overfull, this buffer is
7189 ** used to store the parent's overflow cells. Because this function inserts
7190 ** a maximum of four divider cells into the parent page, and the maximum
7191 ** size of a cell stored within an internal node is always less than 1/4
7192 ** of the page-size, the aOvflSpace[] buffer is guaranteed to be large
7193 ** enough for all overflow cells.
7195 ** If aOvflSpace is set to a null pointer, this function returns
7198 static int balance_nonroot(
7199 MemPage
*pParent
, /* Parent page of siblings being balanced */
7200 int iParentIdx
, /* Index of "the page" in pParent */
7201 u8
*aOvflSpace
, /* page-size bytes of space for parent ovfl */
7202 int isRoot
, /* True if pParent is a root-page */
7203 int bBulk
/* True if this call is part of a bulk load */
7205 BtShared
*pBt
; /* The whole database */
7206 int nMaxCells
= 0; /* Allocated size of apCell, szCell, aFrom. */
7207 int nNew
= 0; /* Number of pages in apNew[] */
7208 int nOld
; /* Number of pages in apOld[] */
7209 int i
, j
, k
; /* Loop counters */
7210 int nxDiv
; /* Next divider slot in pParent->aCell[] */
7211 int rc
= SQLITE_OK
; /* The return code */
7212 u16 leafCorrection
; /* 4 if pPage is a leaf. 0 if not */
7213 int leafData
; /* True if pPage is a leaf of a LEAFDATA tree */
7214 int usableSpace
; /* Bytes in pPage beyond the header */
7215 int pageFlags
; /* Value of pPage->aData[0] */
7216 int iSpace1
= 0; /* First unused byte of aSpace1[] */
7217 int iOvflSpace
= 0; /* First unused byte of aOvflSpace[] */
7218 int szScratch
; /* Size of scratch memory requested */
7219 MemPage
*apOld
[NB
]; /* pPage and up to two siblings */
7220 MemPage
*apNew
[NB
+2]; /* pPage and up to NB siblings after balancing */
7221 u8
*pRight
; /* Location in parent of right-sibling pointer */
7222 u8
*apDiv
[NB
-1]; /* Divider cells in pParent */
7223 int cntNew
[NB
+2]; /* Index in b.paCell[] of cell after i-th page */
7224 int cntOld
[NB
+2]; /* Old index in b.apCell[] */
7225 int szNew
[NB
+2]; /* Combined size of cells placed on i-th page */
7226 u8
*aSpace1
; /* Space for copies of dividers cells */
7227 Pgno pgno
; /* Temp var to store a page number in */
7228 u8 abDone
[NB
+2]; /* True after i'th new page is populated */
7229 Pgno aPgno
[NB
+2]; /* Page numbers of new pages before shuffling */
7230 Pgno aPgOrder
[NB
+2]; /* Copy of aPgno[] used for sorting pages */
7231 u16 aPgFlags
[NB
+2]; /* flags field of new pages before shuffling */
7232 CellArray b
; /* Parsed information on cells being balanced */
7234 memset(abDone
, 0, sizeof(abDone
));
7238 assert( sqlite3_mutex_held(pBt
->mutex
) );
7239 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7242 TRACE(("BALANCE: begin page %d child of %d\n", pPage
->pgno
, pParent
->pgno
));
7245 /* At this point pParent may have at most one overflow cell. And if
7246 ** this overflow cell is present, it must be the cell with
7247 ** index iParentIdx. This scenario comes about when this function
7248 ** is called (indirectly) from sqlite3BtreeDelete().
7250 assert( pParent
->nOverflow
==0 || pParent
->nOverflow
==1 );
7251 assert( pParent
->nOverflow
==0 || pParent
->aiOvfl
[0]==iParentIdx
);
7254 return SQLITE_NOMEM_BKPT
;
7257 /* Find the sibling pages to balance. Also locate the cells in pParent
7258 ** that divide the siblings. An attempt is made to find NN siblings on
7259 ** either side of pPage. More siblings are taken from one side, however,
7260 ** if there are fewer than NN siblings on the other side. If pParent
7261 ** has NB or fewer children then all children of pParent are taken.
7263 ** This loop also drops the divider cells from the parent page. This
7264 ** way, the remainder of the function does not have to deal with any
7265 ** overflow cells in the parent page, since if any existed they will
7266 ** have already been removed.
7268 i
= pParent
->nOverflow
+ pParent
->nCell
;
7272 assert( bBulk
==0 || bBulk
==1 );
7273 if( iParentIdx
==0 ){
7275 }else if( iParentIdx
==i
){
7278 nxDiv
= iParentIdx
-1;
7283 if( (i
+nxDiv
-pParent
->nOverflow
)==pParent
->nCell
){
7284 pRight
= &pParent
->aData
[pParent
->hdrOffset
+8];
7286 pRight
= findCell(pParent
, i
+nxDiv
-pParent
->nOverflow
);
7288 pgno
= get4byte(pRight
);
7290 rc
= getAndInitPage(pBt
, pgno
, &apOld
[i
], 0, 0);
7292 memset(apOld
, 0, (i
+1)*sizeof(MemPage
*));
7293 goto balance_cleanup
;
7295 nMaxCells
+= 1+apOld
[i
]->nCell
+apOld
[i
]->nOverflow
;
7296 if( (i
--)==0 ) break;
7298 if( pParent
->nOverflow
&& i
+nxDiv
==pParent
->aiOvfl
[0] ){
7299 apDiv
[i
] = pParent
->apOvfl
[0];
7300 pgno
= get4byte(apDiv
[i
]);
7301 szNew
[i
] = pParent
->xCellSize(pParent
, apDiv
[i
]);
7302 pParent
->nOverflow
= 0;
7304 apDiv
[i
] = findCell(pParent
, i
+nxDiv
-pParent
->nOverflow
);
7305 pgno
= get4byte(apDiv
[i
]);
7306 szNew
[i
] = pParent
->xCellSize(pParent
, apDiv
[i
]);
7308 /* Drop the cell from the parent page. apDiv[i] still points to
7309 ** the cell within the parent, even though it has been dropped.
7310 ** This is safe because dropping a cell only overwrites the first
7311 ** four bytes of it, and this function does not need the first
7312 ** four bytes of the divider cell. So the pointer is safe to use
7315 ** But not if we are in secure-delete mode. In secure-delete mode,
7316 ** the dropCell() routine will overwrite the entire cell with zeroes.
7317 ** In this case, temporarily copy the cell into the aOvflSpace[]
7318 ** buffer. It will be copied out again as soon as the aSpace[] buffer
7320 if( pBt
->btsFlags
& BTS_FAST_SECURE
){
7323 iOff
= SQLITE_PTR_TO_INT(apDiv
[i
]) - SQLITE_PTR_TO_INT(pParent
->aData
);
7324 if( (iOff
+szNew
[i
])>(int)pBt
->usableSize
){
7325 rc
= SQLITE_CORRUPT_BKPT
;
7326 memset(apOld
, 0, (i
+1)*sizeof(MemPage
*));
7327 goto balance_cleanup
;
7329 memcpy(&aOvflSpace
[iOff
], apDiv
[i
], szNew
[i
]);
7330 apDiv
[i
] = &aOvflSpace
[apDiv
[i
]-pParent
->aData
];
7333 dropCell(pParent
, i
+nxDiv
-pParent
->nOverflow
, szNew
[i
], &rc
);
7337 /* Make nMaxCells a multiple of 4 in order to preserve 8-byte
7339 nMaxCells
= (nMaxCells
+ 3)&~3;
7342 ** Allocate space for memory structures
7345 nMaxCells
*sizeof(u8
*) /* b.apCell */
7346 + nMaxCells
*sizeof(u16
) /* b.szCell */
7347 + pBt
->pageSize
; /* aSpace1 */
7349 assert( szScratch
<=6*(int)pBt
->pageSize
);
7350 b
.apCell
= sqlite3StackAllocRaw(0, szScratch
);
7352 rc
= SQLITE_NOMEM_BKPT
;
7353 goto balance_cleanup
;
7355 b
.szCell
= (u16
*)&b
.apCell
[nMaxCells
];
7356 aSpace1
= (u8
*)&b
.szCell
[nMaxCells
];
7357 assert( EIGHT_BYTE_ALIGNMENT(aSpace1
) );
7360 ** Load pointers to all cells on sibling pages and the divider cells
7361 ** into the local b.apCell[] array. Make copies of the divider cells
7362 ** into space obtained from aSpace1[]. The divider cells have already
7363 ** been removed from pParent.
7365 ** If the siblings are on leaf pages, then the child pointers of the
7366 ** divider cells are stripped from the cells before they are copied
7367 ** into aSpace1[]. In this way, all cells in b.apCell[] are without
7368 ** child pointers. If siblings are not leaves, then all cell in
7369 ** b.apCell[] include child pointers. Either way, all cells in b.apCell[]
7372 ** leafCorrection: 4 if pPage is a leaf. 0 if pPage is not a leaf.
7373 ** leafData: 1 if pPage holds key+data and pParent holds only keys.
7376 leafCorrection
= b
.pRef
->leaf
*4;
7377 leafData
= b
.pRef
->intKeyLeaf
;
7378 for(i
=0; i
<nOld
; i
++){
7379 MemPage
*pOld
= apOld
[i
];
7380 int limit
= pOld
->nCell
;
7381 u8
*aData
= pOld
->aData
;
7382 u16 maskPage
= pOld
->maskPage
;
7383 u8
*piCell
= aData
+ pOld
->cellOffset
;
7386 /* Verify that all sibling pages are of the same "type" (table-leaf,
7387 ** table-interior, index-leaf, or index-interior).
7389 if( pOld
->aData
[0]!=apOld
[0]->aData
[0] ){
7390 rc
= SQLITE_CORRUPT_BKPT
;
7391 goto balance_cleanup
;
7394 /* Load b.apCell[] with pointers to all cells in pOld. If pOld
7395 ** constains overflow cells, include them in the b.apCell[] array
7396 ** in the correct spot.
7398 ** Note that when there are multiple overflow cells, it is always the
7399 ** case that they are sequential and adjacent. This invariant arises
7400 ** because multiple overflows can only occurs when inserting divider
7401 ** cells into a parent on a prior balance, and divider cells are always
7402 ** adjacent and are inserted in order. There is an assert() tagged
7403 ** with "NOTE 1" in the overflow cell insertion loop to prove this
7406 ** This must be done in advance. Once the balance starts, the cell
7407 ** offset section of the btree page will be overwritten and we will no
7408 ** long be able to find the cells if a pointer to each cell is not saved
7411 memset(&b
.szCell
[b
.nCell
], 0, sizeof(b
.szCell
[0])*(limit
+pOld
->nOverflow
));
7412 if( pOld
->nOverflow
>0 ){
7413 limit
= pOld
->aiOvfl
[0];
7414 for(j
=0; j
<limit
; j
++){
7415 b
.apCell
[b
.nCell
] = aData
+ (maskPage
& get2byteAligned(piCell
));
7419 for(k
=0; k
<pOld
->nOverflow
; k
++){
7420 assert( k
==0 || pOld
->aiOvfl
[k
-1]+1==pOld
->aiOvfl
[k
] );/* NOTE 1 */
7421 b
.apCell
[b
.nCell
] = pOld
->apOvfl
[k
];
7425 piEnd
= aData
+ pOld
->cellOffset
+ 2*pOld
->nCell
;
7426 while( piCell
<piEnd
){
7427 assert( b
.nCell
<nMaxCells
);
7428 b
.apCell
[b
.nCell
] = aData
+ (maskPage
& get2byteAligned(piCell
));
7433 cntOld
[i
] = b
.nCell
;
7434 if( i
<nOld
-1 && !leafData
){
7435 u16 sz
= (u16
)szNew
[i
];
7437 assert( b
.nCell
<nMaxCells
);
7438 b
.szCell
[b
.nCell
] = sz
;
7439 pTemp
= &aSpace1
[iSpace1
];
7441 assert( sz
<=pBt
->maxLocal
+23 );
7442 assert( iSpace1
<= (int)pBt
->pageSize
);
7443 memcpy(pTemp
, apDiv
[i
], sz
);
7444 b
.apCell
[b
.nCell
] = pTemp
+leafCorrection
;
7445 assert( leafCorrection
==0 || leafCorrection
==4 );
7446 b
.szCell
[b
.nCell
] = b
.szCell
[b
.nCell
] - leafCorrection
;
7448 assert( leafCorrection
==0 );
7449 assert( pOld
->hdrOffset
==0 );
7450 /* The right pointer of the child page pOld becomes the left
7451 ** pointer of the divider cell */
7452 memcpy(b
.apCell
[b
.nCell
], &pOld
->aData
[8], 4);
7454 assert( leafCorrection
==4 );
7455 while( b
.szCell
[b
.nCell
]<4 ){
7456 /* Do not allow any cells smaller than 4 bytes. If a smaller cell
7457 ** does exist, pad it with 0x00 bytes. */
7458 assert( b
.szCell
[b
.nCell
]==3 || CORRUPT_DB
);
7459 assert( b
.apCell
[b
.nCell
]==&aSpace1
[iSpace1
-3] || CORRUPT_DB
);
7460 aSpace1
[iSpace1
++] = 0x00;
7461 b
.szCell
[b
.nCell
]++;
7469 ** Figure out the number of pages needed to hold all b.nCell cells.
7470 ** Store this number in "k". Also compute szNew[] which is the total
7471 ** size of all cells on the i-th page and cntNew[] which is the index
7472 ** in b.apCell[] of the cell that divides page i from page i+1.
7473 ** cntNew[k] should equal b.nCell.
7475 ** Values computed by this block:
7477 ** k: The total number of sibling pages
7478 ** szNew[i]: Spaced used on the i-th sibling page.
7479 ** cntNew[i]: Index in b.apCell[] and b.szCell[] for the first cell to
7480 ** the right of the i-th sibling page.
7481 ** usableSpace: Number of bytes of space available on each sibling.
7484 usableSpace
= pBt
->usableSize
- 12 + leafCorrection
;
7485 for(i
=0; i
<nOld
; i
++){
7486 MemPage
*p
= apOld
[i
];
7487 szNew
[i
] = usableSpace
- p
->nFree
;
7488 for(j
=0; j
<p
->nOverflow
; j
++){
7489 szNew
[i
] += 2 + p
->xCellSize(p
, p
->apOvfl
[j
]);
7491 cntNew
[i
] = cntOld
[i
];
7496 while( szNew
[i
]>usableSpace
){
7499 if( k
>NB
+2 ){ rc
= SQLITE_CORRUPT_BKPT
; goto balance_cleanup
; }
7501 cntNew
[k
-1] = b
.nCell
;
7503 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]-1);
7506 if( cntNew
[i
]<b
.nCell
){
7507 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7515 while( cntNew
[i
]<b
.nCell
){
7516 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7517 if( szNew
[i
]+sz
>usableSpace
) break;
7521 if( cntNew
[i
]<b
.nCell
){
7522 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7529 if( cntNew
[i
]>=b
.nCell
){
7531 }else if( cntNew
[i
] <= (i
>0 ? cntNew
[i
-1] : 0) ){
7532 rc
= SQLITE_CORRUPT_BKPT
;
7533 goto balance_cleanup
;
7538 ** The packing computed by the previous block is biased toward the siblings
7539 ** on the left side (siblings with smaller keys). The left siblings are
7540 ** always nearly full, while the right-most sibling might be nearly empty.
7541 ** The next block of code attempts to adjust the packing of siblings to
7542 ** get a better balance.
7544 ** This adjustment is more than an optimization. The packing above might
7545 ** be so out of balance as to be illegal. For example, the right-most
7546 ** sibling might be completely empty. This adjustment is not optional.
7548 for(i
=k
-1; i
>0; i
--){
7549 int szRight
= szNew
[i
]; /* Size of sibling on the right */
7550 int szLeft
= szNew
[i
-1]; /* Size of sibling on the left */
7551 int r
; /* Index of right-most cell in left sibling */
7552 int d
; /* Index of first cell to the left of right sibling */
7554 r
= cntNew
[i
-1] - 1;
7555 d
= r
+ 1 - leafData
;
7556 (void)cachedCellSize(&b
, d
);
7558 assert( d
<nMaxCells
);
7559 assert( r
<nMaxCells
);
7560 (void)cachedCellSize(&b
, r
);
7562 && (bBulk
|| szRight
+b
.szCell
[d
]+2 > szLeft
-(b
.szCell
[r
]+(i
==k
-1?0:2)))){
7565 szRight
+= b
.szCell
[d
] + 2;
7566 szLeft
-= b
.szCell
[r
] + 2;
7572 szNew
[i
-1] = szLeft
;
7573 if( cntNew
[i
-1] <= (i
>1 ? cntNew
[i
-2] : 0) ){
7574 rc
= SQLITE_CORRUPT_BKPT
;
7575 goto balance_cleanup
;
7579 /* Sanity check: For a non-corrupt database file one of the follwing
7581 ** (1) We found one or more cells (cntNew[0])>0), or
7582 ** (2) pPage is a virtual root page. A virtual root page is when
7583 ** the real root page is page 1 and we are the only child of
7586 assert( cntNew
[0]>0 || (pParent
->pgno
==1 && pParent
->nCell
==0) || CORRUPT_DB
);
7587 TRACE(("BALANCE: old: %d(nc=%d) %d(nc=%d) %d(nc=%d)\n",
7588 apOld
[0]->pgno
, apOld
[0]->nCell
,
7589 nOld
>=2 ? apOld
[1]->pgno
: 0, nOld
>=2 ? apOld
[1]->nCell
: 0,
7590 nOld
>=3 ? apOld
[2]->pgno
: 0, nOld
>=3 ? apOld
[2]->nCell
: 0
7594 ** Allocate k new pages. Reuse old pages where possible.
7596 pageFlags
= apOld
[0]->aData
[0];
7600 pNew
= apNew
[i
] = apOld
[i
];
7602 rc
= sqlite3PagerWrite(pNew
->pDbPage
);
7604 if( rc
) goto balance_cleanup
;
7607 rc
= allocateBtreePage(pBt
, &pNew
, &pgno
, (bBulk
? 1 : pgno
), 0);
7608 if( rc
) goto balance_cleanup
;
7609 zeroPage(pNew
, pageFlags
);
7612 cntOld
[i
] = b
.nCell
;
7614 /* Set the pointer-map entry for the new sibling page. */
7616 ptrmapPut(pBt
, pNew
->pgno
, PTRMAP_BTREE
, pParent
->pgno
, &rc
);
7617 if( rc
!=SQLITE_OK
){
7618 goto balance_cleanup
;
7625 ** Reassign page numbers so that the new pages are in ascending order.
7626 ** This helps to keep entries in the disk file in order so that a scan
7627 ** of the table is closer to a linear scan through the file. That in turn
7628 ** helps the operating system to deliver pages from the disk more rapidly.
7630 ** An O(n^2) insertion sort algorithm is used, but since n is never more
7631 ** than (NB+2) (a small constant), that should not be a problem.
7633 ** When NB==3, this one optimization makes the database about 25% faster
7634 ** for large insertions and deletions.
7636 for(i
=0; i
<nNew
; i
++){
7637 aPgOrder
[i
] = aPgno
[i
] = apNew
[i
]->pgno
;
7638 aPgFlags
[i
] = apNew
[i
]->pDbPage
->flags
;
7640 if( aPgno
[j
]==aPgno
[i
] ){
7641 /* This branch is taken if the set of sibling pages somehow contains
7642 ** duplicate entries. This can happen if the database is corrupt.
7643 ** It would be simpler to detect this as part of the loop below, but
7644 ** we do the detection here in order to avoid populating the pager
7645 ** cache with two separate objects associated with the same
7647 assert( CORRUPT_DB
);
7648 rc
= SQLITE_CORRUPT_BKPT
;
7649 goto balance_cleanup
;
7653 for(i
=0; i
<nNew
; i
++){
7654 int iBest
= 0; /* aPgno[] index of page number to use */
7655 for(j
=1; j
<nNew
; j
++){
7656 if( aPgOrder
[j
]<aPgOrder
[iBest
] ) iBest
= j
;
7658 pgno
= aPgOrder
[iBest
];
7659 aPgOrder
[iBest
] = 0xffffffff;
7662 sqlite3PagerRekey(apNew
[iBest
]->pDbPage
, pBt
->nPage
+iBest
+1, 0);
7664 sqlite3PagerRekey(apNew
[i
]->pDbPage
, pgno
, aPgFlags
[iBest
]);
7665 apNew
[i
]->pgno
= pgno
;
7669 TRACE(("BALANCE: new: %d(%d nc=%d) %d(%d nc=%d) %d(%d nc=%d) "
7670 "%d(%d nc=%d) %d(%d nc=%d)\n",
7671 apNew
[0]->pgno
, szNew
[0], cntNew
[0],
7672 nNew
>=2 ? apNew
[1]->pgno
: 0, nNew
>=2 ? szNew
[1] : 0,
7673 nNew
>=2 ? cntNew
[1] - cntNew
[0] - !leafData
: 0,
7674 nNew
>=3 ? apNew
[2]->pgno
: 0, nNew
>=3 ? szNew
[2] : 0,
7675 nNew
>=3 ? cntNew
[2] - cntNew
[1] - !leafData
: 0,
7676 nNew
>=4 ? apNew
[3]->pgno
: 0, nNew
>=4 ? szNew
[3] : 0,
7677 nNew
>=4 ? cntNew
[3] - cntNew
[2] - !leafData
: 0,
7678 nNew
>=5 ? apNew
[4]->pgno
: 0, nNew
>=5 ? szNew
[4] : 0,
7679 nNew
>=5 ? cntNew
[4] - cntNew
[3] - !leafData
: 0
7682 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7683 put4byte(pRight
, apNew
[nNew
-1]->pgno
);
7685 /* If the sibling pages are not leaves, ensure that the right-child pointer
7686 ** of the right-most new sibling page is set to the value that was
7687 ** originally in the same field of the right-most old sibling page. */
7688 if( (pageFlags
& PTF_LEAF
)==0 && nOld
!=nNew
){
7689 MemPage
*pOld
= (nNew
>nOld
? apNew
: apOld
)[nOld
-1];
7690 memcpy(&apNew
[nNew
-1]->aData
[8], &pOld
->aData
[8], 4);
7693 /* Make any required updates to pointer map entries associated with
7694 ** cells stored on sibling pages following the balance operation. Pointer
7695 ** map entries associated with divider cells are set by the insertCell()
7696 ** routine. The associated pointer map entries are:
7698 ** a) if the cell contains a reference to an overflow chain, the
7699 ** entry associated with the first page in the overflow chain, and
7701 ** b) if the sibling pages are not leaves, the child page associated
7704 ** If the sibling pages are not leaves, then the pointer map entry
7705 ** associated with the right-child of each sibling may also need to be
7706 ** updated. This happens below, after the sibling pages have been
7707 ** populated, not here.
7710 MemPage
*pNew
= apNew
[0];
7711 u8
*aOld
= pNew
->aData
;
7712 int cntOldNext
= pNew
->nCell
+ pNew
->nOverflow
;
7713 int usableSize
= pBt
->usableSize
;
7717 for(i
=0; i
<b
.nCell
; i
++){
7718 u8
*pCell
= b
.apCell
[i
];
7719 if( i
==cntOldNext
){
7720 MemPage
*pOld
= (++iOld
)<nNew
? apNew
[iOld
] : apOld
[iOld
];
7721 cntOldNext
+= pOld
->nCell
+ pOld
->nOverflow
+ !leafData
;
7724 if( i
==cntNew
[iNew
] ){
7725 pNew
= apNew
[++iNew
];
7726 if( !leafData
) continue;
7729 /* Cell pCell is destined for new sibling page pNew. Originally, it
7730 ** was either part of sibling page iOld (possibly an overflow cell),
7731 ** or else the divider cell to the left of sibling page iOld. So,
7732 ** if sibling page iOld had the same page number as pNew, and if
7733 ** pCell really was a part of sibling page iOld (not a divider or
7734 ** overflow cell), we can skip updating the pointer map entries. */
7736 || pNew
->pgno
!=aPgno
[iOld
]
7737 || !SQLITE_WITHIN(pCell
,aOld
,&aOld
[usableSize
])
7739 if( !leafCorrection
){
7740 ptrmapPut(pBt
, get4byte(pCell
), PTRMAP_BTREE
, pNew
->pgno
, &rc
);
7742 if( cachedCellSize(&b
,i
)>pNew
->minLocal
){
7743 ptrmapPutOvflPtr(pNew
, pCell
, &rc
);
7745 if( rc
) goto balance_cleanup
;
7750 /* Insert new divider cells into pParent. */
7751 for(i
=0; i
<nNew
-1; i
++){
7755 MemPage
*pNew
= apNew
[i
];
7758 assert( j
<nMaxCells
);
7759 assert( b
.apCell
[j
]!=0 );
7760 pCell
= b
.apCell
[j
];
7761 sz
= b
.szCell
[j
] + leafCorrection
;
7762 pTemp
= &aOvflSpace
[iOvflSpace
];
7764 memcpy(&pNew
->aData
[8], pCell
, 4);
7765 }else if( leafData
){
7766 /* If the tree is a leaf-data tree, and the siblings are leaves,
7767 ** then there is no divider cell in b.apCell[]. Instead, the divider
7768 ** cell consists of the integer key for the right-most cell of
7769 ** the sibling-page assembled above only.
7773 pNew
->xParseCell(pNew
, b
.apCell
[j
], &info
);
7775 sz
= 4 + putVarint(&pCell
[4], info
.nKey
);
7779 /* Obscure case for non-leaf-data trees: If the cell at pCell was
7780 ** previously stored on a leaf node, and its reported size was 4
7781 ** bytes, then it may actually be smaller than this
7782 ** (see btreeParseCellPtr(), 4 bytes is the minimum size of
7783 ** any cell). But it is important to pass the correct size to
7784 ** insertCell(), so reparse the cell now.
7786 ** This can only happen for b-trees used to evaluate "IN (SELECT ...)"
7787 ** and WITHOUT ROWID tables with exactly one column which is the
7790 if( b
.szCell
[j
]==4 ){
7791 assert(leafCorrection
==4);
7792 sz
= pParent
->xCellSize(pParent
, pCell
);
7796 assert( sz
<=pBt
->maxLocal
+23 );
7797 assert( iOvflSpace
<= (int)pBt
->pageSize
);
7798 insertCell(pParent
, nxDiv
+i
, pCell
, sz
, pTemp
, pNew
->pgno
, &rc
);
7799 if( rc
!=SQLITE_OK
) goto balance_cleanup
;
7800 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7803 /* Now update the actual sibling pages. The order in which they are updated
7804 ** is important, as this code needs to avoid disrupting any page from which
7805 ** cells may still to be read. In practice, this means:
7807 ** (1) If cells are moving left (from apNew[iPg] to apNew[iPg-1])
7808 ** then it is not safe to update page apNew[iPg] until after
7809 ** the left-hand sibling apNew[iPg-1] has been updated.
7811 ** (2) If cells are moving right (from apNew[iPg] to apNew[iPg+1])
7812 ** then it is not safe to update page apNew[iPg] until after
7813 ** the right-hand sibling apNew[iPg+1] has been updated.
7815 ** If neither of the above apply, the page is safe to update.
7817 ** The iPg value in the following loop starts at nNew-1 goes down
7818 ** to 0, then back up to nNew-1 again, thus making two passes over
7819 ** the pages. On the initial downward pass, only condition (1) above
7820 ** needs to be tested because (2) will always be true from the previous
7821 ** step. On the upward pass, both conditions are always true, so the
7822 ** upwards pass simply processes pages that were missed on the downward
7825 for(i
=1-nNew
; i
<nNew
; i
++){
7826 int iPg
= i
<0 ? -i
: i
;
7827 assert( iPg
>=0 && iPg
<nNew
);
7828 if( abDone
[iPg
] ) continue; /* Skip pages already processed */
7829 if( i
>=0 /* On the upwards pass, or... */
7830 || cntOld
[iPg
-1]>=cntNew
[iPg
-1] /* Condition (1) is true */
7836 /* Verify condition (1): If cells are moving left, update iPg
7837 ** only after iPg-1 has already been updated. */
7838 assert( iPg
==0 || cntOld
[iPg
-1]>=cntNew
[iPg
-1] || abDone
[iPg
-1] );
7840 /* Verify condition (2): If cells are moving right, update iPg
7841 ** only after iPg+1 has already been updated. */
7842 assert( cntNew
[iPg
]>=cntOld
[iPg
] || abDone
[iPg
+1] );
7846 nNewCell
= cntNew
[0];
7848 iOld
= iPg
<nOld
? (cntOld
[iPg
-1] + !leafData
) : b
.nCell
;
7849 iNew
= cntNew
[iPg
-1] + !leafData
;
7850 nNewCell
= cntNew
[iPg
] - iNew
;
7853 rc
= editPage(apNew
[iPg
], iOld
, iNew
, nNewCell
, &b
);
7854 if( rc
) goto balance_cleanup
;
7856 apNew
[iPg
]->nFree
= usableSpace
-szNew
[iPg
];
7857 assert( apNew
[iPg
]->nOverflow
==0 );
7858 assert( apNew
[iPg
]->nCell
==nNewCell
);
7862 /* All pages have been processed exactly once */
7863 assert( memcmp(abDone
, "\01\01\01\01\01", nNew
)==0 );
7868 if( isRoot
&& pParent
->nCell
==0 && pParent
->hdrOffset
<=apNew
[0]->nFree
){
7869 /* The root page of the b-tree now contains no cells. The only sibling
7870 ** page is the right-child of the parent. Copy the contents of the
7871 ** child page into the parent, decreasing the overall height of the
7872 ** b-tree structure by one. This is described as the "balance-shallower"
7873 ** sub-algorithm in some documentation.
7875 ** If this is an auto-vacuum database, the call to copyNodeContent()
7876 ** sets all pointer-map entries corresponding to database image pages
7877 ** for which the pointer is stored within the content being copied.
7879 ** It is critical that the child page be defragmented before being
7880 ** copied into the parent, because if the parent is page 1 then it will
7881 ** by smaller than the child due to the database header, and so all the
7882 ** free space needs to be up front.
7884 assert( nNew
==1 || CORRUPT_DB
);
7885 rc
= defragmentPage(apNew
[0], -1);
7886 testcase( rc
!=SQLITE_OK
);
7887 assert( apNew
[0]->nFree
==
7888 (get2byte(&apNew
[0]->aData
[5])-apNew
[0]->cellOffset
-apNew
[0]->nCell
*2)
7891 copyNodeContent(apNew
[0], pParent
, &rc
);
7892 freePage(apNew
[0], &rc
);
7893 }else if( ISAUTOVACUUM
&& !leafCorrection
){
7894 /* Fix the pointer map entries associated with the right-child of each
7895 ** sibling page. All other pointer map entries have already been taken
7897 for(i
=0; i
<nNew
; i
++){
7898 u32 key
= get4byte(&apNew
[i
]->aData
[8]);
7899 ptrmapPut(pBt
, key
, PTRMAP_BTREE
, apNew
[i
]->pgno
, &rc
);
7903 assert( pParent
->isInit
);
7904 TRACE(("BALANCE: finished: old=%d new=%d cells=%d\n",
7905 nOld
, nNew
, b
.nCell
));
7907 /* Free any old pages that were not reused as new pages.
7909 for(i
=nNew
; i
<nOld
; i
++){
7910 freePage(apOld
[i
], &rc
);
7914 if( ISAUTOVACUUM
&& rc
==SQLITE_OK
&& apNew
[0]->isInit
){
7915 /* The ptrmapCheckPages() contains assert() statements that verify that
7916 ** all pointer map pages are set correctly. This is helpful while
7917 ** debugging. This is usually disabled because a corrupt database may
7918 ** cause an assert() statement to fail. */
7919 ptrmapCheckPages(apNew
, nNew
);
7920 ptrmapCheckPages(&pParent
, 1);
7925 ** Cleanup before returning.
7928 sqlite3StackFree(0, b
.apCell
);
7929 for(i
=0; i
<nOld
; i
++){
7930 releasePage(apOld
[i
]);
7932 for(i
=0; i
<nNew
; i
++){
7933 releasePage(apNew
[i
]);
7941 ** This function is called when the root page of a b-tree structure is
7942 ** overfull (has one or more overflow pages).
7944 ** A new child page is allocated and the contents of the current root
7945 ** page, including overflow cells, are copied into the child. The root
7946 ** page is then overwritten to make it an empty page with the right-child
7947 ** pointer pointing to the new page.
7949 ** Before returning, all pointer-map entries corresponding to pages
7950 ** that the new child-page now contains pointers to are updated. The
7951 ** entry corresponding to the new right-child pointer of the root
7952 ** page is also updated.
7954 ** If successful, *ppChild is set to contain a reference to the child
7955 ** page and SQLITE_OK is returned. In this case the caller is required
7956 ** to call releasePage() on *ppChild exactly once. If an error occurs,
7957 ** an error code is returned and *ppChild is set to 0.
7959 static int balance_deeper(MemPage
*pRoot
, MemPage
**ppChild
){
7960 int rc
; /* Return value from subprocedures */
7961 MemPage
*pChild
= 0; /* Pointer to a new child page */
7962 Pgno pgnoChild
= 0; /* Page number of the new child page */
7963 BtShared
*pBt
= pRoot
->pBt
; /* The BTree */
7965 assert( pRoot
->nOverflow
>0 );
7966 assert( sqlite3_mutex_held(pBt
->mutex
) );
7968 /* Make pRoot, the root page of the b-tree, writable. Allocate a new
7969 ** page that will become the new right-child of pPage. Copy the contents
7970 ** of the node stored on pRoot into the new child page.
7972 rc
= sqlite3PagerWrite(pRoot
->pDbPage
);
7973 if( rc
==SQLITE_OK
){
7974 rc
= allocateBtreePage(pBt
,&pChild
,&pgnoChild
,pRoot
->pgno
,0);
7975 copyNodeContent(pRoot
, pChild
, &rc
);
7977 ptrmapPut(pBt
, pgnoChild
, PTRMAP_BTREE
, pRoot
->pgno
, &rc
);
7982 releasePage(pChild
);
7985 assert( sqlite3PagerIswriteable(pChild
->pDbPage
) );
7986 assert( sqlite3PagerIswriteable(pRoot
->pDbPage
) );
7987 assert( pChild
->nCell
==pRoot
->nCell
);
7989 TRACE(("BALANCE: copy root %d into %d\n", pRoot
->pgno
, pChild
->pgno
));
7991 /* Copy the overflow cells from pRoot to pChild */
7992 memcpy(pChild
->aiOvfl
, pRoot
->aiOvfl
,
7993 pRoot
->nOverflow
*sizeof(pRoot
->aiOvfl
[0]));
7994 memcpy(pChild
->apOvfl
, pRoot
->apOvfl
,
7995 pRoot
->nOverflow
*sizeof(pRoot
->apOvfl
[0]));
7996 pChild
->nOverflow
= pRoot
->nOverflow
;
7998 /* Zero the contents of pRoot. Then install pChild as the right-child. */
7999 zeroPage(pRoot
, pChild
->aData
[0] & ~PTF_LEAF
);
8000 put4byte(&pRoot
->aData
[pRoot
->hdrOffset
+8], pgnoChild
);
8007 ** The page that pCur currently points to has just been modified in
8008 ** some way. This function figures out if this modification means the
8009 ** tree needs to be balanced, and if so calls the appropriate balancing
8010 ** routine. Balancing routines are:
8014 ** balance_nonroot()
8016 static int balance(BtCursor
*pCur
){
8018 const int nMin
= pCur
->pBt
->usableSize
* 2 / 3;
8019 u8 aBalanceQuickSpace
[13];
8022 VVA_ONLY( int balance_quick_called
= 0 );
8023 VVA_ONLY( int balance_deeper_called
= 0 );
8026 int iPage
= pCur
->iPage
;
8027 MemPage
*pPage
= pCur
->pPage
;
8030 if( pPage
->nOverflow
){
8031 /* The root page of the b-tree is overfull. In this case call the
8032 ** balance_deeper() function to create a new child for the root-page
8033 ** and copy the current contents of the root-page to it. The
8034 ** next iteration of the do-loop will balance the child page.
8036 assert( balance_deeper_called
==0 );
8037 VVA_ONLY( balance_deeper_called
++ );
8038 rc
= balance_deeper(pPage
, &pCur
->apPage
[1]);
8039 if( rc
==SQLITE_OK
){
8043 pCur
->apPage
[0] = pPage
;
8044 pCur
->pPage
= pCur
->apPage
[1];
8045 assert( pCur
->pPage
->nOverflow
);
8050 }else if( pPage
->nOverflow
==0 && pPage
->nFree
<=nMin
){
8053 MemPage
* const pParent
= pCur
->apPage
[iPage
-1];
8054 int const iIdx
= pCur
->aiIdx
[iPage
-1];
8056 rc
= sqlite3PagerWrite(pParent
->pDbPage
);
8057 if( rc
==SQLITE_OK
){
8058 #ifndef SQLITE_OMIT_QUICKBALANCE
8059 if( pPage
->intKeyLeaf
8060 && pPage
->nOverflow
==1
8061 && pPage
->aiOvfl
[0]==pPage
->nCell
8063 && pParent
->nCell
==iIdx
8065 /* Call balance_quick() to create a new sibling of pPage on which
8066 ** to store the overflow cell. balance_quick() inserts a new cell
8067 ** into pParent, which may cause pParent overflow. If this
8068 ** happens, the next iteration of the do-loop will balance pParent
8069 ** use either balance_nonroot() or balance_deeper(). Until this
8070 ** happens, the overflow cell is stored in the aBalanceQuickSpace[]
8073 ** The purpose of the following assert() is to check that only a
8074 ** single call to balance_quick() is made for each call to this
8075 ** function. If this were not verified, a subtle bug involving reuse
8076 ** of the aBalanceQuickSpace[] might sneak in.
8078 assert( balance_quick_called
==0 );
8079 VVA_ONLY( balance_quick_called
++ );
8080 rc
= balance_quick(pParent
, pPage
, aBalanceQuickSpace
);
8084 /* In this case, call balance_nonroot() to redistribute cells
8085 ** between pPage and up to 2 of its sibling pages. This involves
8086 ** modifying the contents of pParent, which may cause pParent to
8087 ** become overfull or underfull. The next iteration of the do-loop
8088 ** will balance the parent page to correct this.
8090 ** If the parent page becomes overfull, the overflow cell or cells
8091 ** are stored in the pSpace buffer allocated immediately below.
8092 ** A subsequent iteration of the do-loop will deal with this by
8093 ** calling balance_nonroot() (balance_deeper() may be called first,
8094 ** but it doesn't deal with overflow cells - just moves them to a
8095 ** different page). Once this subsequent call to balance_nonroot()
8096 ** has completed, it is safe to release the pSpace buffer used by
8097 ** the previous call, as the overflow cell data will have been
8098 ** copied either into the body of a database page or into the new
8099 ** pSpace buffer passed to the latter call to balance_nonroot().
8101 u8
*pSpace
= sqlite3PageMalloc(pCur
->pBt
->pageSize
);
8102 rc
= balance_nonroot(pParent
, iIdx
, pSpace
, iPage
==1,
8103 pCur
->hints
&BTREE_BULKLOAD
);
8105 /* If pFree is not NULL, it points to the pSpace buffer used
8106 ** by a previous call to balance_nonroot(). Its contents are
8107 ** now stored either on real database pages or within the
8108 ** new pSpace buffer, so it may be safely freed here. */
8109 sqlite3PageFree(pFree
);
8112 /* The pSpace buffer will be freed after the next call to
8113 ** balance_nonroot(), or just before this function returns, whichever
8119 pPage
->nOverflow
= 0;
8121 /* The next iteration of the do-loop balances the parent page. */
8124 assert( pCur
->iPage
>=0 );
8125 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
8127 }while( rc
==SQLITE_OK
);
8130 sqlite3PageFree(pFree
);
8137 ** Insert a new record into the BTree. The content of the new record
8138 ** is described by the pX object. The pCur cursor is used only to
8139 ** define what table the record should be inserted into, and is left
8140 ** pointing at a random location.
8142 ** For a table btree (used for rowid tables), only the pX.nKey value of
8143 ** the key is used. The pX.pKey value must be NULL. The pX.nKey is the
8144 ** rowid or INTEGER PRIMARY KEY of the row. The pX.nData,pData,nZero fields
8145 ** hold the content of the row.
8147 ** For an index btree (used for indexes and WITHOUT ROWID tables), the
8148 ** key is an arbitrary byte sequence stored in pX.pKey,nKey. The
8149 ** pX.pData,nData,nZero fields must be zero.
8151 ** If the seekResult parameter is non-zero, then a successful call to
8152 ** MovetoUnpacked() to seek cursor pCur to (pKey,nKey) has already
8153 ** been performed. In other words, if seekResult!=0 then the cursor
8154 ** is currently pointing to a cell that will be adjacent to the cell
8155 ** to be inserted. If seekResult<0 then pCur points to a cell that is
8156 ** smaller then (pKey,nKey). If seekResult>0 then pCur points to a cell
8157 ** that is larger than (pKey,nKey).
8159 ** If seekResult==0, that means pCur is pointing at some unknown location.
8160 ** In that case, this routine must seek the cursor to the correct insertion
8161 ** point for (pKey,nKey) before doing the insertion. For index btrees,
8162 ** if pX->nMem is non-zero, then pX->aMem contains pointers to the unpacked
8163 ** key values and pX->aMem can be used instead of pX->pKey to avoid having
8164 ** to decode the key.
8166 int sqlite3BtreeInsert(
8167 BtCursor
*pCur
, /* Insert data into the table of this cursor */
8168 const BtreePayload
*pX
, /* Content of the row to be inserted */
8169 int flags
, /* True if this is likely an append */
8170 int seekResult
/* Result of prior MovetoUnpacked() call */
8173 int loc
= seekResult
; /* -1: before desired location +1: after */
8177 Btree
*p
= pCur
->pBtree
;
8178 BtShared
*pBt
= p
->pBt
;
8179 unsigned char *oldCell
;
8180 unsigned char *newCell
= 0;
8182 assert( (flags
& (BTREE_SAVEPOSITION
|BTREE_APPEND
))==flags
);
8184 if( pCur
->eState
==CURSOR_FAULT
){
8185 assert( pCur
->skipNext
!=SQLITE_OK
);
8186 return pCur
->skipNext
;
8189 assert( cursorOwnsBtShared(pCur
) );
8190 assert( (pCur
->curFlags
& BTCF_WriteFlag
)!=0
8191 && pBt
->inTransaction
==TRANS_WRITE
8192 && (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
8193 assert( hasSharedCacheTableLock(p
, pCur
->pgnoRoot
, pCur
->pKeyInfo
!=0, 2) );
8195 /* Assert that the caller has been consistent. If this cursor was opened
8196 ** expecting an index b-tree, then the caller should be inserting blob
8197 ** keys with no associated data. If the cursor was opened expecting an
8198 ** intkey table, the caller should be inserting integer keys with a
8199 ** blob of associated data. */
8200 assert( (pX
->pKey
==0)==(pCur
->pKeyInfo
==0) );
8202 /* Save the positions of any other cursors open on this table.
8204 ** In some cases, the call to btreeMoveto() below is a no-op. For
8205 ** example, when inserting data into a table with auto-generated integer
8206 ** keys, the VDBE layer invokes sqlite3BtreeLast() to figure out the
8207 ** integer key to use. It then calls this function to actually insert the
8208 ** data into the intkey B-Tree. In this case btreeMoveto() recognizes
8209 ** that the cursor is already where it needs to be and returns without
8210 ** doing any work. To avoid thwarting these optimizations, it is important
8211 ** not to clear the cursor here.
8213 if( pCur
->curFlags
& BTCF_Multiple
){
8214 rc
= saveAllCursors(pBt
, pCur
->pgnoRoot
, pCur
);
8218 if( pCur
->pKeyInfo
==0 ){
8219 assert( pX
->pKey
==0 );
8220 /* If this is an insert into a table b-tree, invalidate any incrblob
8221 ** cursors open on the row being replaced */
8222 invalidateIncrblobCursors(p
, pCur
->pgnoRoot
, pX
->nKey
, 0);
8224 /* If BTREE_SAVEPOSITION is set, the cursor must already be pointing
8225 ** to a row with the same key as the new entry being inserted. */
8226 assert( (flags
& BTREE_SAVEPOSITION
)==0 ||
8227 ((pCur
->curFlags
&BTCF_ValidNKey
)!=0 && pX
->nKey
==pCur
->info
.nKey
) );
8229 /* If the cursor is currently on the last row and we are appending a
8230 ** new row onto the end, set the "loc" to avoid an unnecessary
8231 ** btreeMoveto() call */
8232 if( (pCur
->curFlags
&BTCF_ValidNKey
)!=0 && pX
->nKey
==pCur
->info
.nKey
){
8235 rc
= sqlite3BtreeMovetoUnpacked(pCur
, 0, pX
->nKey
, flags
!=0, &loc
);
8238 }else if( loc
==0 && (flags
& BTREE_SAVEPOSITION
)==0 ){
8241 r
.pKeyInfo
= pCur
->pKeyInfo
;
8243 r
.nField
= pX
->nMem
;
8249 rc
= sqlite3BtreeMovetoUnpacked(pCur
, &r
, 0, flags
!=0, &loc
);
8251 rc
= btreeMoveto(pCur
, pX
->pKey
, pX
->nKey
, flags
!=0, &loc
);
8255 assert( pCur
->eState
==CURSOR_VALID
|| (pCur
->eState
==CURSOR_INVALID
&& loc
) );
8257 pPage
= pCur
->pPage
;
8258 assert( pPage
->intKey
|| pX
->nKey
>=0 );
8259 assert( pPage
->leaf
|| !pPage
->intKey
);
8261 TRACE(("INSERT: table=%d nkey=%lld ndata=%d page=%d %s\n",
8262 pCur
->pgnoRoot
, pX
->nKey
, pX
->nData
, pPage
->pgno
,
8263 loc
==0 ? "overwrite" : "new entry"));
8264 assert( pPage
->isInit
);
8265 newCell
= pBt
->pTmpSpace
;
8266 assert( newCell
!=0 );
8267 rc
= fillInCell(pPage
, newCell
, pX
, &szNew
);
8268 if( rc
) goto end_insert
;
8269 assert( szNew
==pPage
->xCellSize(pPage
, newCell
) );
8270 assert( szNew
<= MX_CELL_SIZE(pBt
) );
8274 assert( idx
<pPage
->nCell
);
8275 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
8279 oldCell
= findCell(pPage
, idx
);
8281 memcpy(newCell
, oldCell
, 4);
8283 rc
= clearCell(pPage
, oldCell
, &info
);
8284 if( info
.nSize
==szNew
&& info
.nLocal
==info
.nPayload
8285 && (!ISAUTOVACUUM
|| szNew
<pPage
->minLocal
)
8287 /* Overwrite the old cell with the new if they are the same size.
8288 ** We could also try to do this if the old cell is smaller, then add
8289 ** the leftover space to the free list. But experiments show that
8290 ** doing that is no faster then skipping this optimization and just
8291 ** calling dropCell() and insertCell().
8293 ** This optimization cannot be used on an autovacuum database if the
8294 ** new entry uses overflow pages, as the insertCell() call below is
8295 ** necessary to add the PTRMAP_OVERFLOW1 pointer-map entry. */
8296 assert( rc
==SQLITE_OK
); /* clearCell never fails when nLocal==nPayload */
8297 if( oldCell
+szNew
> pPage
->aDataEnd
) return SQLITE_CORRUPT_BKPT
;
8298 memcpy(oldCell
, newCell
, szNew
);
8301 dropCell(pPage
, idx
, info
.nSize
, &rc
);
8302 if( rc
) goto end_insert
;
8303 }else if( loc
<0 && pPage
->nCell
>0 ){
8304 assert( pPage
->leaf
);
8306 pCur
->curFlags
&= ~BTCF_ValidNKey
;
8308 assert( pPage
->leaf
);
8310 insertCell(pPage
, idx
, newCell
, szNew
, 0, 0, &rc
);
8311 assert( pPage
->nOverflow
==0 || rc
==SQLITE_OK
);
8312 assert( rc
!=SQLITE_OK
|| pPage
->nCell
>0 || pPage
->nOverflow
>0 );
8314 /* If no error has occurred and pPage has an overflow cell, call balance()
8315 ** to redistribute the cells within the tree. Since balance() may move
8316 ** the cursor, zero the BtCursor.info.nSize and BTCF_ValidNKey
8319 ** Previous versions of SQLite called moveToRoot() to move the cursor
8320 ** back to the root page as balance() used to invalidate the contents
8321 ** of BtCursor.apPage[] and BtCursor.aiIdx[]. Instead of doing that,
8322 ** set the cursor state to "invalid". This makes common insert operations
8325 ** There is a subtle but important optimization here too. When inserting
8326 ** multiple records into an intkey b-tree using a single cursor (as can
8327 ** happen while processing an "INSERT INTO ... SELECT" statement), it
8328 ** is advantageous to leave the cursor pointing to the last entry in
8329 ** the b-tree if possible. If the cursor is left pointing to the last
8330 ** entry in the table, and the next row inserted has an integer key
8331 ** larger than the largest existing key, it is possible to insert the
8332 ** row without seeking the cursor. This can be a big performance boost.
8334 pCur
->info
.nSize
= 0;
8335 if( pPage
->nOverflow
){
8336 assert( rc
==SQLITE_OK
);
8337 pCur
->curFlags
&= ~(BTCF_ValidNKey
);
8340 /* Must make sure nOverflow is reset to zero even if the balance()
8341 ** fails. Internal data structure corruption will result otherwise.
8342 ** Also, set the cursor state to invalid. This stops saveCursorPosition()
8343 ** from trying to save the current position of the cursor. */
8344 pCur
->pPage
->nOverflow
= 0;
8345 pCur
->eState
= CURSOR_INVALID
;
8346 if( (flags
& BTREE_SAVEPOSITION
) && rc
==SQLITE_OK
){
8347 btreeReleaseAllCursorPages(pCur
);
8348 if( pCur
->pKeyInfo
){
8349 assert( pCur
->pKey
==0 );
8350 pCur
->pKey
= sqlite3Malloc( pX
->nKey
);
8351 if( pCur
->pKey
==0 ){
8354 memcpy(pCur
->pKey
, pX
->pKey
, pX
->nKey
);
8357 pCur
->eState
= CURSOR_REQUIRESEEK
;
8358 pCur
->nKey
= pX
->nKey
;
8361 assert( pCur
->iPage
<0 || pCur
->pPage
->nOverflow
==0 );
8368 ** Delete the entry that the cursor is pointing to.
8370 ** If the BTREE_SAVEPOSITION bit of the flags parameter is zero, then
8371 ** the cursor is left pointing at an arbitrary location after the delete.
8372 ** But if that bit is set, then the cursor is left in a state such that
8373 ** the next call to BtreeNext() or BtreePrev() moves it to the same row
8374 ** as it would have been on if the call to BtreeDelete() had been omitted.
8376 ** The BTREE_AUXDELETE bit of flags indicates that is one of several deletes
8377 ** associated with a single table entry and its indexes. Only one of those
8378 ** deletes is considered the "primary" delete. The primary delete occurs
8379 ** on a cursor that is not a BTREE_FORDELETE cursor. All but one delete
8380 ** operation on non-FORDELETE cursors is tagged with the AUXDELETE flag.
8381 ** The BTREE_AUXDELETE bit is a hint that is not used by this implementation,
8382 ** but which might be used by alternative storage engines.
8384 int sqlite3BtreeDelete(BtCursor
*pCur
, u8 flags
){
8385 Btree
*p
= pCur
->pBtree
;
8386 BtShared
*pBt
= p
->pBt
;
8387 int rc
; /* Return code */
8388 MemPage
*pPage
; /* Page to delete cell from */
8389 unsigned char *pCell
; /* Pointer to cell to delete */
8390 int iCellIdx
; /* Index of cell to delete */
8391 int iCellDepth
; /* Depth of node containing pCell */
8392 CellInfo info
; /* Size of the cell being deleted */
8393 int bSkipnext
= 0; /* Leaf cursor in SKIPNEXT state */
8394 u8 bPreserve
= flags
& BTREE_SAVEPOSITION
; /* Keep cursor valid */
8396 assert( cursorOwnsBtShared(pCur
) );
8397 assert( pBt
->inTransaction
==TRANS_WRITE
);
8398 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
8399 assert( pCur
->curFlags
& BTCF_WriteFlag
);
8400 assert( hasSharedCacheTableLock(p
, pCur
->pgnoRoot
, pCur
->pKeyInfo
!=0, 2) );
8401 assert( !hasReadConflicts(p
, pCur
->pgnoRoot
) );
8402 assert( pCur
->ix
<pCur
->pPage
->nCell
);
8403 assert( pCur
->eState
==CURSOR_VALID
);
8404 assert( (flags
& ~(BTREE_SAVEPOSITION
| BTREE_AUXDELETE
))==0 );
8406 iCellDepth
= pCur
->iPage
;
8407 iCellIdx
= pCur
->ix
;
8408 pPage
= pCur
->pPage
;
8409 pCell
= findCell(pPage
, iCellIdx
);
8411 /* If the bPreserve flag is set to true, then the cursor position must
8412 ** be preserved following this delete operation. If the current delete
8413 ** will cause a b-tree rebalance, then this is done by saving the cursor
8414 ** key and leaving the cursor in CURSOR_REQUIRESEEK state before
8417 ** Or, if the current delete will not cause a rebalance, then the cursor
8418 ** will be left in CURSOR_SKIPNEXT state pointing to the entry immediately
8419 ** before or after the deleted entry. In this case set bSkipnext to true. */
8422 || (pPage
->nFree
+cellSizePtr(pPage
,pCell
)+2)>(int)(pBt
->usableSize
*2/3)
8424 /* A b-tree rebalance will be required after deleting this entry.
8425 ** Save the cursor key. */
8426 rc
= saveCursorKey(pCur
);
8433 /* If the page containing the entry to delete is not a leaf page, move
8434 ** the cursor to the largest entry in the tree that is smaller than
8435 ** the entry being deleted. This cell will replace the cell being deleted
8436 ** from the internal node. The 'previous' entry is used for this instead
8437 ** of the 'next' entry, as the previous entry is always a part of the
8438 ** sub-tree headed by the child page of the cell being deleted. This makes
8439 ** balancing the tree following the delete operation easier. */
8441 rc
= sqlite3BtreePrevious(pCur
, 0);
8442 assert( rc
!=SQLITE_DONE
);
8446 /* Save the positions of any other cursors open on this table before
8447 ** making any modifications. */
8448 if( pCur
->curFlags
& BTCF_Multiple
){
8449 rc
= saveAllCursors(pBt
, pCur
->pgnoRoot
, pCur
);
8453 /* If this is a delete operation to remove a row from a table b-tree,
8454 ** invalidate any incrblob cursors open on the row being deleted. */
8455 if( pCur
->pKeyInfo
==0 ){
8456 invalidateIncrblobCursors(p
, pCur
->pgnoRoot
, pCur
->info
.nKey
, 0);
8459 /* Make the page containing the entry to be deleted writable. Then free any
8460 ** overflow pages associated with the entry and finally remove the cell
8461 ** itself from within the page. */
8462 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
8464 rc
= clearCell(pPage
, pCell
, &info
);
8465 dropCell(pPage
, iCellIdx
, info
.nSize
, &rc
);
8468 /* If the cell deleted was not located on a leaf page, then the cursor
8469 ** is currently pointing to the largest entry in the sub-tree headed
8470 ** by the child-page of the cell that was just deleted from an internal
8471 ** node. The cell from the leaf node needs to be moved to the internal
8472 ** node to replace the deleted cell. */
8474 MemPage
*pLeaf
= pCur
->pPage
;
8477 unsigned char *pTmp
;
8479 if( iCellDepth
<pCur
->iPage
-1 ){
8480 n
= pCur
->apPage
[iCellDepth
+1]->pgno
;
8482 n
= pCur
->pPage
->pgno
;
8484 pCell
= findCell(pLeaf
, pLeaf
->nCell
-1);
8485 if( pCell
<&pLeaf
->aData
[4] ) return SQLITE_CORRUPT_BKPT
;
8486 nCell
= pLeaf
->xCellSize(pLeaf
, pCell
);
8487 assert( MX_CELL_SIZE(pBt
) >= nCell
);
8488 pTmp
= pBt
->pTmpSpace
;
8490 rc
= sqlite3PagerWrite(pLeaf
->pDbPage
);
8491 if( rc
==SQLITE_OK
){
8492 insertCell(pPage
, iCellIdx
, pCell
-4, nCell
+4, pTmp
, n
, &rc
);
8494 dropCell(pLeaf
, pLeaf
->nCell
-1, nCell
, &rc
);
8498 /* Balance the tree. If the entry deleted was located on a leaf page,
8499 ** then the cursor still points to that page. In this case the first
8500 ** call to balance() repairs the tree, and the if(...) condition is
8503 ** Otherwise, if the entry deleted was on an internal node page, then
8504 ** pCur is pointing to the leaf page from which a cell was removed to
8505 ** replace the cell deleted from the internal node. This is slightly
8506 ** tricky as the leaf node may be underfull, and the internal node may
8507 ** be either under or overfull. In this case run the balancing algorithm
8508 ** on the leaf node first. If the balance proceeds far enough up the
8509 ** tree that we can be sure that any problem in the internal node has
8510 ** been corrected, so be it. Otherwise, after balancing the leaf node,
8511 ** walk the cursor up the tree to the internal node and balance it as
8514 if( rc
==SQLITE_OK
&& pCur
->iPage
>iCellDepth
){
8515 releasePageNotNull(pCur
->pPage
);
8517 while( pCur
->iPage
>iCellDepth
){
8518 releasePage(pCur
->apPage
[pCur
->iPage
--]);
8520 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
8524 if( rc
==SQLITE_OK
){
8526 assert( bPreserve
&& (pCur
->iPage
==iCellDepth
|| CORRUPT_DB
) );
8527 assert( pPage
==pCur
->pPage
|| CORRUPT_DB
);
8528 assert( (pPage
->nCell
>0 || CORRUPT_DB
) && iCellIdx
<=pPage
->nCell
);
8529 pCur
->eState
= CURSOR_SKIPNEXT
;
8530 if( iCellIdx
>=pPage
->nCell
){
8531 pCur
->skipNext
= -1;
8532 pCur
->ix
= pPage
->nCell
-1;
8537 rc
= moveToRoot(pCur
);
8539 btreeReleaseAllCursorPages(pCur
);
8540 pCur
->eState
= CURSOR_REQUIRESEEK
;
8542 if( rc
==SQLITE_EMPTY
) rc
= SQLITE_OK
;
8549 ** Create a new BTree table. Write into *piTable the page
8550 ** number for the root page of the new table.
8552 ** The type of type is determined by the flags parameter. Only the
8553 ** following values of flags are currently in use. Other values for
8554 ** flags might not work:
8556 ** BTREE_INTKEY|BTREE_LEAFDATA Used for SQL tables with rowid keys
8557 ** BTREE_ZERODATA Used for SQL indices
8559 static int btreeCreateTable(Btree
*p
, int *piTable
, int createTabFlags
){
8560 BtShared
*pBt
= p
->pBt
;
8564 int ptfFlags
; /* Page-type flage for the root page of new table */
8566 assert( sqlite3BtreeHoldsMutex(p
) );
8567 assert( pBt
->inTransaction
==TRANS_WRITE
);
8568 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
8570 #ifdef SQLITE_OMIT_AUTOVACUUM
8571 rc
= allocateBtreePage(pBt
, &pRoot
, &pgnoRoot
, 1, 0);
8576 if( pBt
->autoVacuum
){
8577 Pgno pgnoMove
; /* Move a page here to make room for the root-page */
8578 MemPage
*pPageMove
; /* The page to move to. */
8580 /* Creating a new table may probably require moving an existing database
8581 ** to make room for the new tables root page. In case this page turns
8582 ** out to be an overflow page, delete all overflow page-map caches
8583 ** held by open cursors.
8585 invalidateAllOverflowCache(pBt
);
8587 /* Read the value of meta[3] from the database to determine where the
8588 ** root page of the new table should go. meta[3] is the largest root-page
8589 ** created so far, so the new root-page is (meta[3]+1).
8591 sqlite3BtreeGetMeta(p
, BTREE_LARGEST_ROOT_PAGE
, &pgnoRoot
);
8594 /* The new root-page may not be allocated on a pointer-map page, or the
8595 ** PENDING_BYTE page.
8597 while( pgnoRoot
==PTRMAP_PAGENO(pBt
, pgnoRoot
) ||
8598 pgnoRoot
==PENDING_BYTE_PAGE(pBt
) ){
8601 assert( pgnoRoot
>=3 || CORRUPT_DB
);
8602 testcase( pgnoRoot
<3 );
8604 /* Allocate a page. The page that currently resides at pgnoRoot will
8605 ** be moved to the allocated page (unless the allocated page happens
8606 ** to reside at pgnoRoot).
8608 rc
= allocateBtreePage(pBt
, &pPageMove
, &pgnoMove
, pgnoRoot
, BTALLOC_EXACT
);
8609 if( rc
!=SQLITE_OK
){
8613 if( pgnoMove
!=pgnoRoot
){
8614 /* pgnoRoot is the page that will be used for the root-page of
8615 ** the new table (assuming an error did not occur). But we were
8616 ** allocated pgnoMove. If required (i.e. if it was not allocated
8617 ** by extending the file), the current page at position pgnoMove
8618 ** is already journaled.
8623 /* Save the positions of any open cursors. This is required in
8624 ** case they are holding a reference to an xFetch reference
8625 ** corresponding to page pgnoRoot. */
8626 rc
= saveAllCursors(pBt
, 0, 0);
8627 releasePage(pPageMove
);
8628 if( rc
!=SQLITE_OK
){
8632 /* Move the page currently at pgnoRoot to pgnoMove. */
8633 rc
= btreeGetPage(pBt
, pgnoRoot
, &pRoot
, 0);
8634 if( rc
!=SQLITE_OK
){
8637 rc
= ptrmapGet(pBt
, pgnoRoot
, &eType
, &iPtrPage
);
8638 if( eType
==PTRMAP_ROOTPAGE
|| eType
==PTRMAP_FREEPAGE
){
8639 rc
= SQLITE_CORRUPT_BKPT
;
8641 if( rc
!=SQLITE_OK
){
8645 assert( eType
!=PTRMAP_ROOTPAGE
);
8646 assert( eType
!=PTRMAP_FREEPAGE
);
8647 rc
= relocatePage(pBt
, pRoot
, eType
, iPtrPage
, pgnoMove
, 0);
8650 /* Obtain the page at pgnoRoot */
8651 if( rc
!=SQLITE_OK
){
8654 rc
= btreeGetPage(pBt
, pgnoRoot
, &pRoot
, 0);
8655 if( rc
!=SQLITE_OK
){
8658 rc
= sqlite3PagerWrite(pRoot
->pDbPage
);
8659 if( rc
!=SQLITE_OK
){
8667 /* Update the pointer-map and meta-data with the new root-page number. */
8668 ptrmapPut(pBt
, pgnoRoot
, PTRMAP_ROOTPAGE
, 0, &rc
);
8674 /* When the new root page was allocated, page 1 was made writable in
8675 ** order either to increase the database filesize, or to decrement the
8676 ** freelist count. Hence, the sqlite3BtreeUpdateMeta() call cannot fail.
8678 assert( sqlite3PagerIswriteable(pBt
->pPage1
->pDbPage
) );
8679 rc
= sqlite3BtreeUpdateMeta(p
, 4, pgnoRoot
);
8686 rc
= allocateBtreePage(pBt
, &pRoot
, &pgnoRoot
, 1, 0);
8690 assert( sqlite3PagerIswriteable(pRoot
->pDbPage
) );
8691 if( createTabFlags
& BTREE_INTKEY
){
8692 ptfFlags
= PTF_INTKEY
| PTF_LEAFDATA
| PTF_LEAF
;
8694 ptfFlags
= PTF_ZERODATA
| PTF_LEAF
;
8696 zeroPage(pRoot
, ptfFlags
);
8697 sqlite3PagerUnref(pRoot
->pDbPage
);
8698 assert( (pBt
->openFlags
& BTREE_SINGLE
)==0 || pgnoRoot
==2 );
8699 *piTable
= (int)pgnoRoot
;
8702 int sqlite3BtreeCreateTable(Btree
*p
, int *piTable
, int flags
){
8704 sqlite3BtreeEnter(p
);
8705 rc
= btreeCreateTable(p
, piTable
, flags
);
8706 sqlite3BtreeLeave(p
);
8711 ** Erase the given database page and all its children. Return
8712 ** the page to the freelist.
8714 static int clearDatabasePage(
8715 BtShared
*pBt
, /* The BTree that contains the table */
8716 Pgno pgno
, /* Page number to clear */
8717 int freePageFlag
, /* Deallocate page if true */
8718 int *pnChange
/* Add number of Cells freed to this counter */
8722 unsigned char *pCell
;
8727 assert( sqlite3_mutex_held(pBt
->mutex
) );
8728 if( pgno
>btreePagecount(pBt
) ){
8729 return SQLITE_CORRUPT_BKPT
;
8731 rc
= getAndInitPage(pBt
, pgno
, &pPage
, 0, 0);
8734 rc
= SQLITE_CORRUPT_BKPT
;
8735 goto cleardatabasepage_out
;
8738 hdr
= pPage
->hdrOffset
;
8739 for(i
=0; i
<pPage
->nCell
; i
++){
8740 pCell
= findCell(pPage
, i
);
8742 rc
= clearDatabasePage(pBt
, get4byte(pCell
), 1, pnChange
);
8743 if( rc
) goto cleardatabasepage_out
;
8745 rc
= clearCell(pPage
, pCell
, &info
);
8746 if( rc
) goto cleardatabasepage_out
;
8749 rc
= clearDatabasePage(pBt
, get4byte(&pPage
->aData
[hdr
+8]), 1, pnChange
);
8750 if( rc
) goto cleardatabasepage_out
;
8751 }else if( pnChange
){
8752 assert( pPage
->intKey
|| CORRUPT_DB
);
8753 testcase( !pPage
->intKey
);
8754 *pnChange
+= pPage
->nCell
;
8757 freePage(pPage
, &rc
);
8758 }else if( (rc
= sqlite3PagerWrite(pPage
->pDbPage
))==0 ){
8759 zeroPage(pPage
, pPage
->aData
[hdr
] | PTF_LEAF
);
8762 cleardatabasepage_out
:
8769 ** Delete all information from a single table in the database. iTable is
8770 ** the page number of the root of the table. After this routine returns,
8771 ** the root page is empty, but still exists.
8773 ** This routine will fail with SQLITE_LOCKED if there are any open
8774 ** read cursors on the table. Open write cursors are moved to the
8775 ** root of the table.
8777 ** If pnChange is not NULL, then table iTable must be an intkey table. The
8778 ** integer value pointed to by pnChange is incremented by the number of
8779 ** entries in the table.
8781 int sqlite3BtreeClearTable(Btree
*p
, int iTable
, int *pnChange
){
8783 BtShared
*pBt
= p
->pBt
;
8784 sqlite3BtreeEnter(p
);
8785 assert( p
->inTrans
==TRANS_WRITE
);
8787 rc
= saveAllCursors(pBt
, (Pgno
)iTable
, 0);
8789 if( SQLITE_OK
==rc
){
8790 /* Invalidate all incrblob cursors open on table iTable (assuming iTable
8791 ** is the root of a table b-tree - if it is not, the following call is
8793 invalidateIncrblobCursors(p
, (Pgno
)iTable
, 0, 1);
8794 rc
= clearDatabasePage(pBt
, (Pgno
)iTable
, 0, pnChange
);
8796 sqlite3BtreeLeave(p
);
8801 ** Delete all information from the single table that pCur is open on.
8803 ** This routine only work for pCur on an ephemeral table.
8805 int sqlite3BtreeClearTableOfCursor(BtCursor
*pCur
){
8806 return sqlite3BtreeClearTable(pCur
->pBtree
, pCur
->pgnoRoot
, 0);
8810 ** Erase all information in a table and add the root of the table to
8811 ** the freelist. Except, the root of the principle table (the one on
8812 ** page 1) is never added to the freelist.
8814 ** This routine will fail with SQLITE_LOCKED if there are any open
8815 ** cursors on the table.
8817 ** If AUTOVACUUM is enabled and the page at iTable is not the last
8818 ** root page in the database file, then the last root page
8819 ** in the database file is moved into the slot formerly occupied by
8820 ** iTable and that last slot formerly occupied by the last root page
8821 ** is added to the freelist instead of iTable. In this say, all
8822 ** root pages are kept at the beginning of the database file, which
8823 ** is necessary for AUTOVACUUM to work right. *piMoved is set to the
8824 ** page number that used to be the last root page in the file before
8825 ** the move. If no page gets moved, *piMoved is set to 0.
8826 ** The last root page is recorded in meta[3] and the value of
8827 ** meta[3] is updated by this procedure.
8829 static int btreeDropTable(Btree
*p
, Pgno iTable
, int *piMoved
){
8832 BtShared
*pBt
= p
->pBt
;
8834 assert( sqlite3BtreeHoldsMutex(p
) );
8835 assert( p
->inTrans
==TRANS_WRITE
);
8836 assert( iTable
>=2 );
8838 rc
= btreeGetPage(pBt
, (Pgno
)iTable
, &pPage
, 0);
8840 rc
= sqlite3BtreeClearTable(p
, iTable
, 0);
8848 #ifdef SQLITE_OMIT_AUTOVACUUM
8849 freePage(pPage
, &rc
);
8852 if( pBt
->autoVacuum
){
8854 sqlite3BtreeGetMeta(p
, BTREE_LARGEST_ROOT_PAGE
, &maxRootPgno
);
8856 if( iTable
==maxRootPgno
){
8857 /* If the table being dropped is the table with the largest root-page
8858 ** number in the database, put the root page on the free list.
8860 freePage(pPage
, &rc
);
8862 if( rc
!=SQLITE_OK
){
8866 /* The table being dropped does not have the largest root-page
8867 ** number in the database. So move the page that does into the
8868 ** gap left by the deleted root-page.
8872 rc
= btreeGetPage(pBt
, maxRootPgno
, &pMove
, 0);
8873 if( rc
!=SQLITE_OK
){
8876 rc
= relocatePage(pBt
, pMove
, PTRMAP_ROOTPAGE
, 0, iTable
, 0);
8878 if( rc
!=SQLITE_OK
){
8882 rc
= btreeGetPage(pBt
, maxRootPgno
, &pMove
, 0);
8883 freePage(pMove
, &rc
);
8885 if( rc
!=SQLITE_OK
){
8888 *piMoved
= maxRootPgno
;
8891 /* Set the new 'max-root-page' value in the database header. This
8892 ** is the old value less one, less one more if that happens to
8893 ** be a root-page number, less one again if that is the
8894 ** PENDING_BYTE_PAGE.
8897 while( maxRootPgno
==PENDING_BYTE_PAGE(pBt
)
8898 || PTRMAP_ISPAGE(pBt
, maxRootPgno
) ){
8901 assert( maxRootPgno
!=PENDING_BYTE_PAGE(pBt
) );
8903 rc
= sqlite3BtreeUpdateMeta(p
, 4, maxRootPgno
);
8905 freePage(pPage
, &rc
);
8911 int sqlite3BtreeDropTable(Btree
*p
, int iTable
, int *piMoved
){
8913 sqlite3BtreeEnter(p
);
8914 rc
= btreeDropTable(p
, iTable
, piMoved
);
8915 sqlite3BtreeLeave(p
);
8921 ** This function may only be called if the b-tree connection already
8922 ** has a read or write transaction open on the database.
8924 ** Read the meta-information out of a database file. Meta[0]
8925 ** is the number of free pages currently in the database. Meta[1]
8926 ** through meta[15] are available for use by higher layers. Meta[0]
8927 ** is read-only, the others are read/write.
8929 ** The schema layer numbers meta values differently. At the schema
8930 ** layer (and the SetCookie and ReadCookie opcodes) the number of
8931 ** free pages is not visible. So Cookie[0] is the same as Meta[1].
8933 ** This routine treats Meta[BTREE_DATA_VERSION] as a special case. Instead
8934 ** of reading the value out of the header, it instead loads the "DataVersion"
8935 ** from the pager. The BTREE_DATA_VERSION value is not actually stored in the
8936 ** database file. It is a number computed by the pager. But its access
8937 ** pattern is the same as header meta values, and so it is convenient to
8938 ** read it from this routine.
8940 void sqlite3BtreeGetMeta(Btree
*p
, int idx
, u32
*pMeta
){
8941 BtShared
*pBt
= p
->pBt
;
8943 sqlite3BtreeEnter(p
);
8944 assert( p
->inTrans
>TRANS_NONE
);
8945 assert( SQLITE_OK
==querySharedCacheTableLock(p
, MASTER_ROOT
, READ_LOCK
) );
8946 assert( pBt
->pPage1
);
8947 assert( idx
>=0 && idx
<=15 );
8949 if( idx
==BTREE_DATA_VERSION
){
8950 *pMeta
= sqlite3PagerDataVersion(pBt
->pPager
) + p
->iDataVersion
;
8952 *pMeta
= get4byte(&pBt
->pPage1
->aData
[36 + idx
*4]);
8955 /* If auto-vacuum is disabled in this build and this is an auto-vacuum
8956 ** database, mark the database as read-only. */
8957 #ifdef SQLITE_OMIT_AUTOVACUUM
8958 if( idx
==BTREE_LARGEST_ROOT_PAGE
&& *pMeta
>0 ){
8959 pBt
->btsFlags
|= BTS_READ_ONLY
;
8963 sqlite3BtreeLeave(p
);
8967 ** Write meta-information back into the database. Meta[0] is
8968 ** read-only and may not be written.
8970 int sqlite3BtreeUpdateMeta(Btree
*p
, int idx
, u32 iMeta
){
8971 BtShared
*pBt
= p
->pBt
;
8974 assert( idx
>=1 && idx
<=15 );
8975 sqlite3BtreeEnter(p
);
8976 assert( p
->inTrans
==TRANS_WRITE
);
8977 assert( pBt
->pPage1
!=0 );
8978 pP1
= pBt
->pPage1
->aData
;
8979 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
8980 if( rc
==SQLITE_OK
){
8981 put4byte(&pP1
[36 + idx
*4], iMeta
);
8982 #ifndef SQLITE_OMIT_AUTOVACUUM
8983 if( idx
==BTREE_INCR_VACUUM
){
8984 assert( pBt
->autoVacuum
|| iMeta
==0 );
8985 assert( iMeta
==0 || iMeta
==1 );
8986 pBt
->incrVacuum
= (u8
)iMeta
;
8990 sqlite3BtreeLeave(p
);
8994 #ifndef SQLITE_OMIT_BTREECOUNT
8996 ** The first argument, pCur, is a cursor opened on some b-tree. Count the
8997 ** number of entries in the b-tree and write the result to *pnEntry.
8999 ** SQLITE_OK is returned if the operation is successfully executed.
9000 ** Otherwise, if an error is encountered (i.e. an IO error or database
9001 ** corruption) an SQLite error code is returned.
9003 int sqlite3BtreeCount(BtCursor
*pCur
, i64
*pnEntry
){
9004 i64 nEntry
= 0; /* Value to return in *pnEntry */
9005 int rc
; /* Return code */
9007 rc
= moveToRoot(pCur
);
9008 if( rc
==SQLITE_EMPTY
){
9013 /* Unless an error occurs, the following loop runs one iteration for each
9014 ** page in the B-Tree structure (not including overflow pages).
9016 while( rc
==SQLITE_OK
){
9017 int iIdx
; /* Index of child node in parent */
9018 MemPage
*pPage
; /* Current page of the b-tree */
9020 /* If this is a leaf page or the tree is not an int-key tree, then
9021 ** this page contains countable entries. Increment the entry counter
9024 pPage
= pCur
->pPage
;
9025 if( pPage
->leaf
|| !pPage
->intKey
){
9026 nEntry
+= pPage
->nCell
;
9029 /* pPage is a leaf node. This loop navigates the cursor so that it
9030 ** points to the first interior cell that it points to the parent of
9031 ** the next page in the tree that has not yet been visited. The
9032 ** pCur->aiIdx[pCur->iPage] value is set to the index of the parent cell
9033 ** of the page, or to the number of cells in the page if the next page
9034 ** to visit is the right-child of its parent.
9036 ** If all pages in the tree have been visited, return SQLITE_OK to the
9041 if( pCur
->iPage
==0 ){
9042 /* All pages of the b-tree have been visited. Return successfully. */
9044 return moveToRoot(pCur
);
9047 }while ( pCur
->ix
>=pCur
->pPage
->nCell
);
9050 pPage
= pCur
->pPage
;
9053 /* Descend to the child node of the cell that the cursor currently
9054 ** points at. This is the right-child if (iIdx==pPage->nCell).
9057 if( iIdx
==pPage
->nCell
){
9058 rc
= moveToChild(pCur
, get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]));
9060 rc
= moveToChild(pCur
, get4byte(findCell(pPage
, iIdx
)));
9064 /* An error has occurred. Return an error code. */
9070 ** Return the pager associated with a BTree. This routine is used for
9071 ** testing and debugging only.
9073 Pager
*sqlite3BtreePager(Btree
*p
){
9074 return p
->pBt
->pPager
;
9077 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9079 ** Append a message to the error message string.
9081 static void checkAppendMsg(
9082 IntegrityCk
*pCheck
,
9083 const char *zFormat
,
9087 if( !pCheck
->mxErr
) return;
9090 va_start(ap
, zFormat
);
9091 if( pCheck
->errMsg
.nChar
){
9092 sqlite3StrAccumAppend(&pCheck
->errMsg
, "\n", 1);
9095 sqlite3XPrintf(&pCheck
->errMsg
, pCheck
->zPfx
, pCheck
->v1
, pCheck
->v2
);
9097 sqlite3VXPrintf(&pCheck
->errMsg
, zFormat
, ap
);
9099 if( pCheck
->errMsg
.accError
==STRACCUM_NOMEM
){
9100 pCheck
->mallocFailed
= 1;
9103 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9105 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9108 ** Return non-zero if the bit in the IntegrityCk.aPgRef[] array that
9109 ** corresponds to page iPg is already set.
9111 static int getPageReferenced(IntegrityCk
*pCheck
, Pgno iPg
){
9112 assert( iPg
<=pCheck
->nPage
&& sizeof(pCheck
->aPgRef
[0])==1 );
9113 return (pCheck
->aPgRef
[iPg
/8] & (1 << (iPg
& 0x07)));
9117 ** Set the bit in the IntegrityCk.aPgRef[] array that corresponds to page iPg.
9119 static void setPageReferenced(IntegrityCk
*pCheck
, Pgno iPg
){
9120 assert( iPg
<=pCheck
->nPage
&& sizeof(pCheck
->aPgRef
[0])==1 );
9121 pCheck
->aPgRef
[iPg
/8] |= (1 << (iPg
& 0x07));
9126 ** Add 1 to the reference count for page iPage. If this is the second
9127 ** reference to the page, add an error message to pCheck->zErrMsg.
9128 ** Return 1 if there are 2 or more references to the page and 0 if
9129 ** if this is the first reference to the page.
9131 ** Also check that the page number is in bounds.
9133 static int checkRef(IntegrityCk
*pCheck
, Pgno iPage
){
9134 if( iPage
==0 ) return 1;
9135 if( iPage
>pCheck
->nPage
){
9136 checkAppendMsg(pCheck
, "invalid page number %d", iPage
);
9139 if( getPageReferenced(pCheck
, iPage
) ){
9140 checkAppendMsg(pCheck
, "2nd reference to page %d", iPage
);
9143 setPageReferenced(pCheck
, iPage
);
9147 #ifndef SQLITE_OMIT_AUTOVACUUM
9149 ** Check that the entry in the pointer-map for page iChild maps to
9150 ** page iParent, pointer type ptrType. If not, append an error message
9153 static void checkPtrmap(
9154 IntegrityCk
*pCheck
, /* Integrity check context */
9155 Pgno iChild
, /* Child page number */
9156 u8 eType
, /* Expected pointer map type */
9157 Pgno iParent
/* Expected pointer map parent page number */
9163 rc
= ptrmapGet(pCheck
->pBt
, iChild
, &ePtrmapType
, &iPtrmapParent
);
9164 if( rc
!=SQLITE_OK
){
9165 if( rc
==SQLITE_NOMEM
|| rc
==SQLITE_IOERR_NOMEM
) pCheck
->mallocFailed
= 1;
9166 checkAppendMsg(pCheck
, "Failed to read ptrmap key=%d", iChild
);
9170 if( ePtrmapType
!=eType
|| iPtrmapParent
!=iParent
){
9171 checkAppendMsg(pCheck
,
9172 "Bad ptr map entry key=%d expected=(%d,%d) got=(%d,%d)",
9173 iChild
, eType
, iParent
, ePtrmapType
, iPtrmapParent
);
9179 ** Check the integrity of the freelist or of an overflow page list.
9180 ** Verify that the number of pages on the list is N.
9182 static void checkList(
9183 IntegrityCk
*pCheck
, /* Integrity checking context */
9184 int isFreeList
, /* True for a freelist. False for overflow page list */
9185 int iPage
, /* Page number for first page in the list */
9186 int N
/* Expected number of pages in the list */
9191 while( N
-- > 0 && pCheck
->mxErr
){
9193 unsigned char *pOvflData
;
9195 checkAppendMsg(pCheck
,
9196 "%d of %d pages missing from overflow list starting at %d",
9197 N
+1, expected
, iFirst
);
9200 if( checkRef(pCheck
, iPage
) ) break;
9201 if( sqlite3PagerGet(pCheck
->pPager
, (Pgno
)iPage
, &pOvflPage
, 0) ){
9202 checkAppendMsg(pCheck
, "failed to get page %d", iPage
);
9205 pOvflData
= (unsigned char *)sqlite3PagerGetData(pOvflPage
);
9207 int n
= get4byte(&pOvflData
[4]);
9208 #ifndef SQLITE_OMIT_AUTOVACUUM
9209 if( pCheck
->pBt
->autoVacuum
){
9210 checkPtrmap(pCheck
, iPage
, PTRMAP_FREEPAGE
, 0);
9213 if( n
>(int)pCheck
->pBt
->usableSize
/4-2 ){
9214 checkAppendMsg(pCheck
,
9215 "freelist leaf count too big on page %d", iPage
);
9219 Pgno iFreePage
= get4byte(&pOvflData
[8+i
*4]);
9220 #ifndef SQLITE_OMIT_AUTOVACUUM
9221 if( pCheck
->pBt
->autoVacuum
){
9222 checkPtrmap(pCheck
, iFreePage
, PTRMAP_FREEPAGE
, 0);
9225 checkRef(pCheck
, iFreePage
);
9230 #ifndef SQLITE_OMIT_AUTOVACUUM
9232 /* If this database supports auto-vacuum and iPage is not the last
9233 ** page in this overflow list, check that the pointer-map entry for
9234 ** the following page matches iPage.
9236 if( pCheck
->pBt
->autoVacuum
&& N
>0 ){
9237 i
= get4byte(pOvflData
);
9238 checkPtrmap(pCheck
, i
, PTRMAP_OVERFLOW2
, iPage
);
9242 iPage
= get4byte(pOvflData
);
9243 sqlite3PagerUnref(pOvflPage
);
9245 if( isFreeList
&& N
<(iPage
!=0) ){
9246 checkAppendMsg(pCheck
, "free-page count in header is too small");
9250 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9253 ** An implementation of a min-heap.
9255 ** aHeap[0] is the number of elements on the heap. aHeap[1] is the
9256 ** root element. The daughter nodes of aHeap[N] are aHeap[N*2]
9257 ** and aHeap[N*2+1].
9259 ** The heap property is this: Every node is less than or equal to both
9260 ** of its daughter nodes. A consequence of the heap property is that the
9261 ** root node aHeap[1] is always the minimum value currently in the heap.
9263 ** The btreeHeapInsert() routine inserts an unsigned 32-bit number onto
9264 ** the heap, preserving the heap property. The btreeHeapPull() routine
9265 ** removes the root element from the heap (the minimum value in the heap)
9266 ** and then moves other nodes around as necessary to preserve the heap
9269 ** This heap is used for cell overlap and coverage testing. Each u32
9270 ** entry represents the span of a cell or freeblock on a btree page.
9271 ** The upper 16 bits are the index of the first byte of a range and the
9272 ** lower 16 bits are the index of the last byte of that range.
9274 static void btreeHeapInsert(u32
*aHeap
, u32 x
){
9275 u32 j
, i
= ++aHeap
[0];
9277 while( (j
= i
/2)>0 && aHeap
[j
]>aHeap
[i
] ){
9279 aHeap
[j
] = aHeap
[i
];
9284 static int btreeHeapPull(u32
*aHeap
, u32
*pOut
){
9286 if( (x
= aHeap
[0])==0 ) return 0;
9288 aHeap
[1] = aHeap
[x
];
9289 aHeap
[x
] = 0xffffffff;
9292 while( (j
= i
*2)<=aHeap
[0] ){
9293 if( aHeap
[j
]>aHeap
[j
+1] ) j
++;
9294 if( aHeap
[i
]<aHeap
[j
] ) break;
9296 aHeap
[i
] = aHeap
[j
];
9303 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9305 ** Do various sanity checks on a single page of a tree. Return
9306 ** the tree depth. Root pages return 0. Parents of root pages
9307 ** return 1, and so forth.
9309 ** These checks are done:
9311 ** 1. Make sure that cells and freeblocks do not overlap
9312 ** but combine to completely cover the page.
9313 ** 2. Make sure integer cell keys are in order.
9314 ** 3. Check the integrity of overflow pages.
9315 ** 4. Recursively call checkTreePage on all children.
9316 ** 5. Verify that the depth of all children is the same.
9318 static int checkTreePage(
9319 IntegrityCk
*pCheck
, /* Context for the sanity check */
9320 int iPage
, /* Page number of the page to check */
9321 i64
*piMinKey
, /* Write minimum integer primary key here */
9322 i64 maxKey
/* Error if integer primary key greater than this */
9324 MemPage
*pPage
= 0; /* The page being analyzed */
9325 int i
; /* Loop counter */
9326 int rc
; /* Result code from subroutine call */
9327 int depth
= -1, d2
; /* Depth of a subtree */
9328 int pgno
; /* Page number */
9329 int nFrag
; /* Number of fragmented bytes on the page */
9330 int hdr
; /* Offset to the page header */
9331 int cellStart
; /* Offset to the start of the cell pointer array */
9332 int nCell
; /* Number of cells */
9333 int doCoverageCheck
= 1; /* True if cell coverage checking should be done */
9334 int keyCanBeEqual
= 1; /* True if IPK can be equal to maxKey
9335 ** False if IPK must be strictly less than maxKey */
9336 u8
*data
; /* Page content */
9337 u8
*pCell
; /* Cell content */
9338 u8
*pCellIdx
; /* Next element of the cell pointer array */
9339 BtShared
*pBt
; /* The BtShared object that owns pPage */
9340 u32 pc
; /* Address of a cell */
9341 u32 usableSize
; /* Usable size of the page */
9342 u32 contentOffset
; /* Offset to the start of the cell content area */
9343 u32
*heap
= 0; /* Min-heap used for checking cell coverage */
9344 u32 x
, prev
= 0; /* Next and previous entry on the min-heap */
9345 const char *saved_zPfx
= pCheck
->zPfx
;
9346 int saved_v1
= pCheck
->v1
;
9347 int saved_v2
= pCheck
->v2
;
9350 /* Check that the page exists
9353 usableSize
= pBt
->usableSize
;
9354 if( iPage
==0 ) return 0;
9355 if( checkRef(pCheck
, iPage
) ) return 0;
9356 pCheck
->zPfx
= "Page %d: ";
9358 if( (rc
= btreeGetPage(pBt
, (Pgno
)iPage
, &pPage
, 0))!=0 ){
9359 checkAppendMsg(pCheck
,
9360 "unable to get the page. error code=%d", rc
);
9364 /* Clear MemPage.isInit to make sure the corruption detection code in
9365 ** btreeInitPage() is executed. */
9366 savedIsInit
= pPage
->isInit
;
9368 if( (rc
= btreeInitPage(pPage
))!=0 ){
9369 assert( rc
==SQLITE_CORRUPT
); /* The only possible error from InitPage */
9370 checkAppendMsg(pCheck
,
9371 "btreeInitPage() returns error code %d", rc
);
9374 data
= pPage
->aData
;
9375 hdr
= pPage
->hdrOffset
;
9377 /* Set up for cell analysis */
9378 pCheck
->zPfx
= "On tree page %d cell %d: ";
9379 contentOffset
= get2byteNotZero(&data
[hdr
+5]);
9380 assert( contentOffset
<=usableSize
); /* Enforced by btreeInitPage() */
9382 /* EVIDENCE-OF: R-37002-32774 The two-byte integer at offset 3 gives the
9383 ** number of cells on the page. */
9384 nCell
= get2byte(&data
[hdr
+3]);
9385 assert( pPage
->nCell
==nCell
);
9387 /* EVIDENCE-OF: R-23882-45353 The cell pointer array of a b-tree page
9388 ** immediately follows the b-tree page header. */
9389 cellStart
= hdr
+ 12 - 4*pPage
->leaf
;
9390 assert( pPage
->aCellIdx
==&data
[cellStart
] );
9391 pCellIdx
= &data
[cellStart
+ 2*(nCell
-1)];
9394 /* Analyze the right-child page of internal pages */
9395 pgno
= get4byte(&data
[hdr
+8]);
9396 #ifndef SQLITE_OMIT_AUTOVACUUM
9397 if( pBt
->autoVacuum
){
9398 pCheck
->zPfx
= "On page %d at right child: ";
9399 checkPtrmap(pCheck
, pgno
, PTRMAP_BTREE
, iPage
);
9402 depth
= checkTreePage(pCheck
, pgno
, &maxKey
, maxKey
);
9405 /* For leaf pages, the coverage check will occur in the same loop
9406 ** as the other cell checks, so initialize the heap. */
9407 heap
= pCheck
->heap
;
9411 /* EVIDENCE-OF: R-02776-14802 The cell pointer array consists of K 2-byte
9412 ** integer offsets to the cell contents. */
9413 for(i
=nCell
-1; i
>=0 && pCheck
->mxErr
; i
--){
9416 /* Check cell size */
9418 assert( pCellIdx
==&data
[cellStart
+ i
*2] );
9419 pc
= get2byteAligned(pCellIdx
);
9421 if( pc
<contentOffset
|| pc
>usableSize
-4 ){
9422 checkAppendMsg(pCheck
, "Offset %d out of range %d..%d",
9423 pc
, contentOffset
, usableSize
-4);
9424 doCoverageCheck
= 0;
9428 pPage
->xParseCell(pPage
, pCell
, &info
);
9429 if( pc
+info
.nSize
>usableSize
){
9430 checkAppendMsg(pCheck
, "Extends off end of page");
9431 doCoverageCheck
= 0;
9435 /* Check for integer primary key out of range */
9436 if( pPage
->intKey
){
9437 if( keyCanBeEqual
? (info
.nKey
> maxKey
) : (info
.nKey
>= maxKey
) ){
9438 checkAppendMsg(pCheck
, "Rowid %lld out of order", info
.nKey
);
9441 keyCanBeEqual
= 0; /* Only the first key on the page may ==maxKey */
9444 /* Check the content overflow list */
9445 if( info
.nPayload
>info
.nLocal
){
9446 int nPage
; /* Number of pages on the overflow chain */
9447 Pgno pgnoOvfl
; /* First page of the overflow chain */
9448 assert( pc
+ info
.nSize
- 4 <= usableSize
);
9449 nPage
= (info
.nPayload
- info
.nLocal
+ usableSize
- 5)/(usableSize
- 4);
9450 pgnoOvfl
= get4byte(&pCell
[info
.nSize
- 4]);
9451 #ifndef SQLITE_OMIT_AUTOVACUUM
9452 if( pBt
->autoVacuum
){
9453 checkPtrmap(pCheck
, pgnoOvfl
, PTRMAP_OVERFLOW1
, iPage
);
9456 checkList(pCheck
, 0, pgnoOvfl
, nPage
);
9460 /* Check sanity of left child page for internal pages */
9461 pgno
= get4byte(pCell
);
9462 #ifndef SQLITE_OMIT_AUTOVACUUM
9463 if( pBt
->autoVacuum
){
9464 checkPtrmap(pCheck
, pgno
, PTRMAP_BTREE
, iPage
);
9467 d2
= checkTreePage(pCheck
, pgno
, &maxKey
, maxKey
);
9470 checkAppendMsg(pCheck
, "Child page depth differs");
9474 /* Populate the coverage-checking heap for leaf pages */
9475 btreeHeapInsert(heap
, (pc
<<16)|(pc
+info
.nSize
-1));
9480 /* Check for complete coverage of the page
9483 if( doCoverageCheck
&& pCheck
->mxErr
>0 ){
9484 /* For leaf pages, the min-heap has already been initialized and the
9485 ** cells have already been inserted. But for internal pages, that has
9486 ** not yet been done, so do it now */
9488 heap
= pCheck
->heap
;
9490 for(i
=nCell
-1; i
>=0; i
--){
9492 pc
= get2byteAligned(&data
[cellStart
+i
*2]);
9493 size
= pPage
->xCellSize(pPage
, &data
[pc
]);
9494 btreeHeapInsert(heap
, (pc
<<16)|(pc
+size
-1));
9497 /* Add the freeblocks to the min-heap
9499 ** EVIDENCE-OF: R-20690-50594 The second field of the b-tree page header
9500 ** is the offset of the first freeblock, or zero if there are no
9501 ** freeblocks on the page.
9503 i
= get2byte(&data
[hdr
+1]);
9506 assert( (u32
)i
<=usableSize
-4 ); /* Enforced by btreeInitPage() */
9507 size
= get2byte(&data
[i
+2]);
9508 assert( (u32
)(i
+size
)<=usableSize
); /* Enforced by btreeInitPage() */
9509 btreeHeapInsert(heap
, (((u32
)i
)<<16)|(i
+size
-1));
9510 /* EVIDENCE-OF: R-58208-19414 The first 2 bytes of a freeblock are a
9511 ** big-endian integer which is the offset in the b-tree page of the next
9512 ** freeblock in the chain, or zero if the freeblock is the last on the
9514 j
= get2byte(&data
[i
]);
9515 /* EVIDENCE-OF: R-06866-39125 Freeblocks are always connected in order of
9516 ** increasing offset. */
9517 assert( j
==0 || j
>i
+size
); /* Enforced by btreeInitPage() */
9518 assert( (u32
)j
<=usableSize
-4 ); /* Enforced by btreeInitPage() */
9521 /* Analyze the min-heap looking for overlap between cells and/or
9522 ** freeblocks, and counting the number of untracked bytes in nFrag.
9524 ** Each min-heap entry is of the form: (start_address<<16)|end_address.
9525 ** There is an implied first entry the covers the page header, the cell
9526 ** pointer index, and the gap between the cell pointer index and the start
9529 ** The loop below pulls entries from the min-heap in order and compares
9530 ** the start_address against the previous end_address. If there is an
9531 ** overlap, that means bytes are used multiple times. If there is a gap,
9532 ** that gap is added to the fragmentation count.
9535 prev
= contentOffset
- 1; /* Implied first min-heap entry */
9536 while( btreeHeapPull(heap
,&x
) ){
9537 if( (prev
&0xffff)>=(x
>>16) ){
9538 checkAppendMsg(pCheck
,
9539 "Multiple uses for byte %u of page %d", x
>>16, iPage
);
9542 nFrag
+= (x
>>16) - (prev
&0xffff) - 1;
9546 nFrag
+= usableSize
- (prev
&0xffff) - 1;
9547 /* EVIDENCE-OF: R-43263-13491 The total number of bytes in all fragments
9548 ** is stored in the fifth field of the b-tree page header.
9549 ** EVIDENCE-OF: R-07161-27322 The one-byte integer at offset 7 gives the
9550 ** number of fragmented free bytes within the cell content area.
9552 if( heap
[0]==0 && nFrag
!=data
[hdr
+7] ){
9553 checkAppendMsg(pCheck
,
9554 "Fragmentation of %d bytes reported as %d on page %d",
9555 nFrag
, data
[hdr
+7], iPage
);
9560 if( !doCoverageCheck
) pPage
->isInit
= savedIsInit
;
9562 pCheck
->zPfx
= saved_zPfx
;
9563 pCheck
->v1
= saved_v1
;
9564 pCheck
->v2
= saved_v2
;
9567 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9569 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9571 ** This routine does a complete check of the given BTree file. aRoot[] is
9572 ** an array of pages numbers were each page number is the root page of
9573 ** a table. nRoot is the number of entries in aRoot.
9575 ** A read-only or read-write transaction must be opened before calling
9578 ** Write the number of error seen in *pnErr. Except for some memory
9579 ** allocation errors, an error message held in memory obtained from
9580 ** malloc is returned if *pnErr is non-zero. If *pnErr==0 then NULL is
9581 ** returned. If a memory allocation error occurs, NULL is returned.
9583 char *sqlite3BtreeIntegrityCheck(
9584 Btree
*p
, /* The btree to be checked */
9585 int *aRoot
, /* An array of root pages numbers for individual trees */
9586 int nRoot
, /* Number of entries in aRoot[] */
9587 int mxErr
, /* Stop reporting errors after this many */
9588 int *pnErr
/* Write number of errors seen to this variable */
9592 BtShared
*pBt
= p
->pBt
;
9593 int savedDbFlags
= pBt
->db
->flags
;
9595 VVA_ONLY( int nRef
);
9597 sqlite3BtreeEnter(p
);
9598 assert( p
->inTrans
>TRANS_NONE
&& pBt
->inTransaction
>TRANS_NONE
);
9599 VVA_ONLY( nRef
= sqlite3PagerRefcount(pBt
->pPager
) );
9602 sCheck
.pPager
= pBt
->pPager
;
9603 sCheck
.nPage
= btreePagecount(sCheck
.pBt
);
9604 sCheck
.mxErr
= mxErr
;
9606 sCheck
.mallocFailed
= 0;
9612 sqlite3StrAccumInit(&sCheck
.errMsg
, 0, zErr
, sizeof(zErr
), SQLITE_MAX_LENGTH
);
9613 sCheck
.errMsg
.printfFlags
= SQLITE_PRINTF_INTERNAL
;
9614 if( sCheck
.nPage
==0 ){
9615 goto integrity_ck_cleanup
;
9618 sCheck
.aPgRef
= sqlite3MallocZero((sCheck
.nPage
/ 8)+ 1);
9619 if( !sCheck
.aPgRef
){
9620 sCheck
.mallocFailed
= 1;
9621 goto integrity_ck_cleanup
;
9623 sCheck
.heap
= (u32
*)sqlite3PageMalloc( pBt
->pageSize
);
9624 if( sCheck
.heap
==0 ){
9625 sCheck
.mallocFailed
= 1;
9626 goto integrity_ck_cleanup
;
9629 i
= PENDING_BYTE_PAGE(pBt
);
9630 if( i
<=sCheck
.nPage
) setPageReferenced(&sCheck
, i
);
9632 /* Check the integrity of the freelist
9634 sCheck
.zPfx
= "Main freelist: ";
9635 checkList(&sCheck
, 1, get4byte(&pBt
->pPage1
->aData
[32]),
9636 get4byte(&pBt
->pPage1
->aData
[36]));
9639 /* Check all the tables.
9641 testcase( pBt
->db
->flags
& SQLITE_CellSizeCk
);
9642 pBt
->db
->flags
&= ~SQLITE_CellSizeCk
;
9643 for(i
=0; (int)i
<nRoot
&& sCheck
.mxErr
; i
++){
9645 if( aRoot
[i
]==0 ) continue;
9646 #ifndef SQLITE_OMIT_AUTOVACUUM
9647 if( pBt
->autoVacuum
&& aRoot
[i
]>1 ){
9648 checkPtrmap(&sCheck
, aRoot
[i
], PTRMAP_ROOTPAGE
, 0);
9651 checkTreePage(&sCheck
, aRoot
[i
], ¬Used
, LARGEST_INT64
);
9653 pBt
->db
->flags
= savedDbFlags
;
9655 /* Make sure every page in the file is referenced
9657 for(i
=1; i
<=sCheck
.nPage
&& sCheck
.mxErr
; i
++){
9658 #ifdef SQLITE_OMIT_AUTOVACUUM
9659 if( getPageReferenced(&sCheck
, i
)==0 ){
9660 checkAppendMsg(&sCheck
, "Page %d is never used", i
);
9663 /* If the database supports auto-vacuum, make sure no tables contain
9664 ** references to pointer-map pages.
9666 if( getPageReferenced(&sCheck
, i
)==0 &&
9667 (PTRMAP_PAGENO(pBt
, i
)!=i
|| !pBt
->autoVacuum
) ){
9668 checkAppendMsg(&sCheck
, "Page %d is never used", i
);
9670 if( getPageReferenced(&sCheck
, i
)!=0 &&
9671 (PTRMAP_PAGENO(pBt
, i
)==i
&& pBt
->autoVacuum
) ){
9672 checkAppendMsg(&sCheck
, "Pointer map page %d is referenced", i
);
9677 /* Clean up and report errors.
9679 integrity_ck_cleanup
:
9680 sqlite3PageFree(sCheck
.heap
);
9681 sqlite3_free(sCheck
.aPgRef
);
9682 if( sCheck
.mallocFailed
){
9683 sqlite3StrAccumReset(&sCheck
.errMsg
);
9686 *pnErr
= sCheck
.nErr
;
9687 if( sCheck
.nErr
==0 ) sqlite3StrAccumReset(&sCheck
.errMsg
);
9688 /* Make sure this analysis did not leave any unref() pages. */
9689 assert( nRef
==sqlite3PagerRefcount(pBt
->pPager
) );
9690 sqlite3BtreeLeave(p
);
9691 return sqlite3StrAccumFinish(&sCheck
.errMsg
);
9693 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9696 ** Return the full pathname of the underlying database file. Return
9697 ** an empty string if the database is in-memory or a TEMP database.
9699 ** The pager filename is invariant as long as the pager is
9700 ** open so it is safe to access without the BtShared mutex.
9702 const char *sqlite3BtreeGetFilename(Btree
*p
){
9703 assert( p
->pBt
->pPager
!=0 );
9704 return sqlite3PagerFilename(p
->pBt
->pPager
, 1);
9708 ** Return the pathname of the journal file for this database. The return
9709 ** value of this routine is the same regardless of whether the journal file
9710 ** has been created or not.
9712 ** The pager journal filename is invariant as long as the pager is
9713 ** open so it is safe to access without the BtShared mutex.
9715 const char *sqlite3BtreeGetJournalname(Btree
*p
){
9716 assert( p
->pBt
->pPager
!=0 );
9717 return sqlite3PagerJournalname(p
->pBt
->pPager
);
9721 ** Return non-zero if a transaction is active.
9723 int sqlite3BtreeIsInTrans(Btree
*p
){
9724 assert( p
==0 || sqlite3_mutex_held(p
->db
->mutex
) );
9725 return (p
&& (p
->inTrans
==TRANS_WRITE
));
9728 #ifndef SQLITE_OMIT_WAL
9730 ** Run a checkpoint on the Btree passed as the first argument.
9732 ** Return SQLITE_LOCKED if this or any other connection has an open
9733 ** transaction on the shared-cache the argument Btree is connected to.
9735 ** Parameter eMode is one of SQLITE_CHECKPOINT_PASSIVE, FULL or RESTART.
9737 int sqlite3BtreeCheckpoint(Btree
*p
, int eMode
, int *pnLog
, int *pnCkpt
){
9740 BtShared
*pBt
= p
->pBt
;
9741 sqlite3BtreeEnter(p
);
9742 if( pBt
->inTransaction
!=TRANS_NONE
){
9745 rc
= sqlite3PagerCheckpoint(pBt
->pPager
, p
->db
, eMode
, pnLog
, pnCkpt
);
9747 sqlite3BtreeLeave(p
);
9754 ** Return non-zero if a read (or write) transaction is active.
9756 int sqlite3BtreeIsInReadTrans(Btree
*p
){
9758 assert( sqlite3_mutex_held(p
->db
->mutex
) );
9759 return p
->inTrans
!=TRANS_NONE
;
9762 int sqlite3BtreeIsInBackup(Btree
*p
){
9764 assert( sqlite3_mutex_held(p
->db
->mutex
) );
9765 return p
->nBackup
!=0;
9769 ** This function returns a pointer to a blob of memory associated with
9770 ** a single shared-btree. The memory is used by client code for its own
9771 ** purposes (for example, to store a high-level schema associated with
9772 ** the shared-btree). The btree layer manages reference counting issues.
9774 ** The first time this is called on a shared-btree, nBytes bytes of memory
9775 ** are allocated, zeroed, and returned to the caller. For each subsequent
9776 ** call the nBytes parameter is ignored and a pointer to the same blob
9777 ** of memory returned.
9779 ** If the nBytes parameter is 0 and the blob of memory has not yet been
9780 ** allocated, a null pointer is returned. If the blob has already been
9781 ** allocated, it is returned as normal.
9783 ** Just before the shared-btree is closed, the function passed as the
9784 ** xFree argument when the memory allocation was made is invoked on the
9785 ** blob of allocated memory. The xFree function should not call sqlite3_free()
9786 ** on the memory, the btree layer does that.
9788 void *sqlite3BtreeSchema(Btree
*p
, int nBytes
, void(*xFree
)(void *)){
9789 BtShared
*pBt
= p
->pBt
;
9790 sqlite3BtreeEnter(p
);
9791 if( !pBt
->pSchema
&& nBytes
){
9792 pBt
->pSchema
= sqlite3DbMallocZero(0, nBytes
);
9793 pBt
->xFreeSchema
= xFree
;
9795 sqlite3BtreeLeave(p
);
9796 return pBt
->pSchema
;
9800 ** Return SQLITE_LOCKED_SHAREDCACHE if another user of the same shared
9801 ** btree as the argument handle holds an exclusive lock on the
9802 ** sqlite_master table. Otherwise SQLITE_OK.
9804 int sqlite3BtreeSchemaLocked(Btree
*p
){
9806 assert( sqlite3_mutex_held(p
->db
->mutex
) );
9807 sqlite3BtreeEnter(p
);
9808 rc
= querySharedCacheTableLock(p
, MASTER_ROOT
, READ_LOCK
);
9809 assert( rc
==SQLITE_OK
|| rc
==SQLITE_LOCKED_SHAREDCACHE
);
9810 sqlite3BtreeLeave(p
);
9815 #ifndef SQLITE_OMIT_SHARED_CACHE
9817 ** Obtain a lock on the table whose root page is iTab. The
9818 ** lock is a write lock if isWritelock is true or a read lock
9821 int sqlite3BtreeLockTable(Btree
*p
, int iTab
, u8 isWriteLock
){
9823 assert( p
->inTrans
!=TRANS_NONE
);
9825 u8 lockType
= READ_LOCK
+ isWriteLock
;
9826 assert( READ_LOCK
+1==WRITE_LOCK
);
9827 assert( isWriteLock
==0 || isWriteLock
==1 );
9829 sqlite3BtreeEnter(p
);
9830 rc
= querySharedCacheTableLock(p
, iTab
, lockType
);
9831 if( rc
==SQLITE_OK
){
9832 rc
= setSharedCacheTableLock(p
, iTab
, lockType
);
9834 sqlite3BtreeLeave(p
);
9840 #ifndef SQLITE_OMIT_INCRBLOB
9842 ** Argument pCsr must be a cursor opened for writing on an
9843 ** INTKEY table currently pointing at a valid table entry.
9844 ** This function modifies the data stored as part of that entry.
9846 ** Only the data content may only be modified, it is not possible to
9847 ** change the length of the data stored. If this function is called with
9848 ** parameters that attempt to write past the end of the existing data,
9849 ** no modifications are made and SQLITE_CORRUPT is returned.
9851 int sqlite3BtreePutData(BtCursor
*pCsr
, u32 offset
, u32 amt
, void *z
){
9853 assert( cursorOwnsBtShared(pCsr
) );
9854 assert( sqlite3_mutex_held(pCsr
->pBtree
->db
->mutex
) );
9855 assert( pCsr
->curFlags
& BTCF_Incrblob
);
9857 rc
= restoreCursorPosition(pCsr
);
9858 if( rc
!=SQLITE_OK
){
9861 assert( pCsr
->eState
!=CURSOR_REQUIRESEEK
);
9862 if( pCsr
->eState
!=CURSOR_VALID
){
9863 return SQLITE_ABORT
;
9866 /* Save the positions of all other cursors open on this table. This is
9867 ** required in case any of them are holding references to an xFetch
9868 ** version of the b-tree page modified by the accessPayload call below.
9870 ** Note that pCsr must be open on a INTKEY table and saveCursorPosition()
9871 ** and hence saveAllCursors() cannot fail on a BTREE_INTKEY table, hence
9872 ** saveAllCursors can only return SQLITE_OK.
9874 VVA_ONLY(rc
=) saveAllCursors(pCsr
->pBt
, pCsr
->pgnoRoot
, pCsr
);
9875 assert( rc
==SQLITE_OK
);
9877 /* Check some assumptions:
9878 ** (a) the cursor is open for writing,
9879 ** (b) there is a read/write transaction open,
9880 ** (c) the connection holds a write-lock on the table (if required),
9881 ** (d) there are no conflicting read-locks, and
9882 ** (e) the cursor points at a valid row of an intKey table.
9884 if( (pCsr
->curFlags
& BTCF_WriteFlag
)==0 ){
9885 return SQLITE_READONLY
;
9887 assert( (pCsr
->pBt
->btsFlags
& BTS_READ_ONLY
)==0
9888 && pCsr
->pBt
->inTransaction
==TRANS_WRITE
);
9889 assert( hasSharedCacheTableLock(pCsr
->pBtree
, pCsr
->pgnoRoot
, 0, 2) );
9890 assert( !hasReadConflicts(pCsr
->pBtree
, pCsr
->pgnoRoot
) );
9891 assert( pCsr
->pPage
->intKey
);
9893 return accessPayload(pCsr
, offset
, amt
, (unsigned char *)z
, 1);
9897 ** Mark this cursor as an incremental blob cursor.
9899 void sqlite3BtreeIncrblobCursor(BtCursor
*pCur
){
9900 pCur
->curFlags
|= BTCF_Incrblob
;
9901 pCur
->pBtree
->hasIncrblobCur
= 1;
9906 ** Set both the "read version" (single byte at byte offset 18) and
9907 ** "write version" (single byte at byte offset 19) fields in the database
9908 ** header to iVersion.
9910 int sqlite3BtreeSetVersion(Btree
*pBtree
, int iVersion
){
9911 BtShared
*pBt
= pBtree
->pBt
;
9912 int rc
; /* Return code */
9914 assert( iVersion
==1 || iVersion
==2 );
9916 /* If setting the version fields to 1, do not automatically open the
9917 ** WAL connection, even if the version fields are currently set to 2.
9919 pBt
->btsFlags
&= ~BTS_NO_WAL
;
9920 if( iVersion
==1 ) pBt
->btsFlags
|= BTS_NO_WAL
;
9922 rc
= sqlite3BtreeBeginTrans(pBtree
, 0);
9923 if( rc
==SQLITE_OK
){
9924 u8
*aData
= pBt
->pPage1
->aData
;
9925 if( aData
[18]!=(u8
)iVersion
|| aData
[19]!=(u8
)iVersion
){
9926 rc
= sqlite3BtreeBeginTrans(pBtree
, 2);
9927 if( rc
==SQLITE_OK
){
9928 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
9929 if( rc
==SQLITE_OK
){
9930 aData
[18] = (u8
)iVersion
;
9931 aData
[19] = (u8
)iVersion
;
9937 pBt
->btsFlags
&= ~BTS_NO_WAL
;
9942 ** Return true if the cursor has a hint specified. This routine is
9943 ** only used from within assert() statements
9945 int sqlite3BtreeCursorHasHint(BtCursor
*pCsr
, unsigned int mask
){
9946 return (pCsr
->hints
& mask
)!=0;
9950 ** Return true if the given Btree is read-only.
9952 int sqlite3BtreeIsReadonly(Btree
*p
){
9953 return (p
->pBt
->btsFlags
& BTS_READ_ONLY
)!=0;
9957 ** Return the size of the header added to each page by this module.
9959 int sqlite3HeaderSizeBtree(void){ return ROUND8(sizeof(MemPage
)); }
9961 #if !defined(SQLITE_OMIT_SHARED_CACHE)
9963 ** Return true if the Btree passed as the only argument is sharable.
9965 int sqlite3BtreeSharable(Btree
*p
){
9970 ** Return the number of connections to the BtShared object accessed by
9971 ** the Btree handle passed as the only argument. For private caches
9972 ** this is always 1. For shared caches it may be 1 or greater.
9974 int sqlite3BtreeConnectionCount(Btree
*p
){
9975 testcase( p
->sharable
);
9976 return p
->pBt
->nRef
;