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_MAIN.
75 BtShared
*SQLITE_WSD sqlite3SharedCacheList
= 0;
77 static BtShared
*SQLITE_WSD sqlite3SharedCacheList
= 0;
79 #endif /* SQLITE_OMIT_SHARED_CACHE */
81 #ifndef SQLITE_OMIT_SHARED_CACHE
83 ** Enable or disable the shared pager and schema features.
85 ** This routine has no effect on existing database connections.
86 ** The shared cache setting effects only future calls to
87 ** sqlite3_open(), sqlite3_open16(), or sqlite3_open_v2().
89 int sqlite3_enable_shared_cache(int enable
){
90 sqlite3GlobalConfig
.sharedCacheEnabled
= enable
;
97 #ifdef SQLITE_OMIT_SHARED_CACHE
99 ** The functions querySharedCacheTableLock(), setSharedCacheTableLock(),
100 ** and clearAllSharedCacheTableLocks()
101 ** manipulate entries in the BtShared.pLock linked list used to store
102 ** shared-cache table level locks. If the library is compiled with the
103 ** shared-cache feature disabled, then there is only ever one user
104 ** of each BtShared structure and so this locking is not necessary.
105 ** So define the lock related functions as no-ops.
107 #define querySharedCacheTableLock(a,b,c) SQLITE_OK
108 #define setSharedCacheTableLock(a,b,c) SQLITE_OK
109 #define clearAllSharedCacheTableLocks(a)
110 #define downgradeAllSharedCacheTableLocks(a)
111 #define hasSharedCacheTableLock(a,b,c,d) 1
112 #define hasReadConflicts(a, b) 0
116 ** Implementation of the SQLITE_CORRUPT_PAGE() macro. Takes a single
117 ** (MemPage*) as an argument. The (MemPage*) must not be NULL.
119 ** If SQLITE_DEBUG is not defined, then this macro is equivalent to
120 ** SQLITE_CORRUPT_BKPT. Or, if SQLITE_DEBUG is set, then the log message
121 ** normally produced as a side-effect of SQLITE_CORRUPT_BKPT is augmented
122 ** with the page number and filename associated with the (MemPage*).
125 int corruptPageError(int lineno
, MemPage
*p
){
127 sqlite3BeginBenignMalloc();
128 zMsg
= sqlite3_mprintf("database corruption page %d of %s",
129 (int)p
->pgno
, sqlite3PagerFilename(p
->pBt
->pPager
, 0)
131 sqlite3EndBenignMalloc();
133 sqlite3ReportError(SQLITE_CORRUPT
, lineno
, zMsg
);
136 return SQLITE_CORRUPT_BKPT
;
138 # define SQLITE_CORRUPT_PAGE(pMemPage) corruptPageError(__LINE__, pMemPage)
140 # define SQLITE_CORRUPT_PAGE(pMemPage) SQLITE_CORRUPT_PGNO(pMemPage->pgno)
143 #ifndef SQLITE_OMIT_SHARED_CACHE
147 **** This function is only used as part of an assert() statement. ***
149 ** Check to see if pBtree holds the required locks to read or write to the
150 ** table with root page iRoot. Return 1 if it does and 0 if not.
152 ** For example, when writing to a table with root-page iRoot via
153 ** Btree connection pBtree:
155 ** assert( hasSharedCacheTableLock(pBtree, iRoot, 0, WRITE_LOCK) );
157 ** When writing to an index that resides in a sharable database, the
158 ** caller should have first obtained a lock specifying the root page of
159 ** the corresponding table. This makes things a bit more complicated,
160 ** as this module treats each table as a separate structure. To determine
161 ** the table corresponding to the index being written, this
162 ** function has to search through the database schema.
164 ** Instead of a lock on the table/index rooted at page iRoot, the caller may
165 ** hold a write-lock on the schema table (root page 1). This is also
168 static int hasSharedCacheTableLock(
169 Btree
*pBtree
, /* Handle that must hold lock */
170 Pgno iRoot
, /* Root page of b-tree */
171 int isIndex
, /* True if iRoot is the root of an index b-tree */
172 int eLockType
/* Required lock type (READ_LOCK or WRITE_LOCK) */
174 Schema
*pSchema
= (Schema
*)pBtree
->pBt
->pSchema
;
178 /* If this database is not shareable, or if the client is reading
179 ** and has the read-uncommitted flag set, then no lock is required.
180 ** Return true immediately.
182 if( (pBtree
->sharable
==0)
183 || (eLockType
==READ_LOCK
&& (pBtree
->db
->flags
& SQLITE_ReadUncommit
))
188 /* If the client is reading or writing an index and the schema is
189 ** not loaded, then it is too difficult to actually check to see if
190 ** the correct locks are held. So do not bother - just return true.
191 ** This case does not come up very often anyhow.
193 if( isIndex
&& (!pSchema
|| (pSchema
->schemaFlags
&DB_SchemaLoaded
)==0) ){
197 /* Figure out the root-page that the lock should be held on. For table
198 ** b-trees, this is just the root page of the b-tree being read or
199 ** written. For index b-trees, it is the root page of the associated
204 for(p
=sqliteHashFirst(&pSchema
->idxHash
); p
; p
=sqliteHashNext(p
)){
205 Index
*pIdx
= (Index
*)sqliteHashData(p
);
206 if( pIdx
->tnum
==(int)iRoot
){
208 /* Two or more indexes share the same root page. There must
209 ** be imposter tables. So just return true. The assert is not
210 ** useful in that case. */
213 iTab
= pIdx
->pTable
->tnum
;
221 /* Search for the required lock. Either a write-lock on root-page iTab, a
222 ** write-lock on the schema table, or (if the client is reading) a
223 ** read-lock on iTab will suffice. Return 1 if any of these are found. */
224 for(pLock
=pBtree
->pBt
->pLock
; pLock
; pLock
=pLock
->pNext
){
225 if( pLock
->pBtree
==pBtree
226 && (pLock
->iTable
==iTab
|| (pLock
->eLock
==WRITE_LOCK
&& pLock
->iTable
==1))
227 && pLock
->eLock
>=eLockType
233 /* Failed to find the required lock. */
236 #endif /* SQLITE_DEBUG */
240 **** This function may be used as part of assert() statements only. ****
242 ** Return true if it would be illegal for pBtree to write into the
243 ** table or index rooted at iRoot because other shared connections are
244 ** simultaneously reading that same table or index.
246 ** It is illegal for pBtree to write if some other Btree object that
247 ** shares the same BtShared object is currently reading or writing
248 ** the iRoot table. Except, if the other Btree object has the
249 ** read-uncommitted flag set, then it is OK for the other object to
250 ** have a read cursor.
252 ** For example, before writing to any part of the table or index
253 ** rooted at page iRoot, one should call:
255 ** assert( !hasReadConflicts(pBtree, iRoot) );
257 static int hasReadConflicts(Btree
*pBtree
, Pgno iRoot
){
259 for(p
=pBtree
->pBt
->pCursor
; p
; p
=p
->pNext
){
260 if( p
->pgnoRoot
==iRoot
262 && 0==(p
->pBtree
->db
->flags
& SQLITE_ReadUncommit
)
269 #endif /* #ifdef SQLITE_DEBUG */
272 ** Query to see if Btree handle p may obtain a lock of type eLock
273 ** (READ_LOCK or WRITE_LOCK) on the table with root-page iTab. Return
274 ** SQLITE_OK if the lock may be obtained (by calling
275 ** setSharedCacheTableLock()), or SQLITE_LOCKED if not.
277 static int querySharedCacheTableLock(Btree
*p
, Pgno iTab
, u8 eLock
){
278 BtShared
*pBt
= p
->pBt
;
281 assert( sqlite3BtreeHoldsMutex(p
) );
282 assert( eLock
==READ_LOCK
|| eLock
==WRITE_LOCK
);
284 assert( !(p
->db
->flags
&SQLITE_ReadUncommit
)||eLock
==WRITE_LOCK
||iTab
==1 );
286 /* If requesting a write-lock, then the Btree must have an open write
287 ** transaction on this file. And, obviously, for this to be so there
288 ** must be an open write transaction on the file itself.
290 assert( eLock
==READ_LOCK
|| (p
==pBt
->pWriter
&& p
->inTrans
==TRANS_WRITE
) );
291 assert( eLock
==READ_LOCK
|| pBt
->inTransaction
==TRANS_WRITE
);
293 /* This routine is a no-op if the shared-cache is not enabled */
298 /* If some other connection is holding an exclusive lock, the
299 ** requested lock may not be obtained.
301 if( pBt
->pWriter
!=p
&& (pBt
->btsFlags
& BTS_EXCLUSIVE
)!=0 ){
302 sqlite3ConnectionBlocked(p
->db
, pBt
->pWriter
->db
);
303 return SQLITE_LOCKED_SHAREDCACHE
;
306 for(pIter
=pBt
->pLock
; pIter
; pIter
=pIter
->pNext
){
307 /* The condition (pIter->eLock!=eLock) in the following if(...)
308 ** statement is a simplification of:
310 ** (eLock==WRITE_LOCK || pIter->eLock==WRITE_LOCK)
312 ** since we know that if eLock==WRITE_LOCK, then no other connection
313 ** may hold a WRITE_LOCK on any table in this file (since there can
314 ** only be a single writer).
316 assert( pIter
->eLock
==READ_LOCK
|| pIter
->eLock
==WRITE_LOCK
);
317 assert( eLock
==READ_LOCK
|| pIter
->pBtree
==p
|| pIter
->eLock
==READ_LOCK
);
318 if( pIter
->pBtree
!=p
&& pIter
->iTable
==iTab
&& pIter
->eLock
!=eLock
){
319 sqlite3ConnectionBlocked(p
->db
, pIter
->pBtree
->db
);
320 if( eLock
==WRITE_LOCK
){
321 assert( p
==pBt
->pWriter
);
322 pBt
->btsFlags
|= BTS_PENDING
;
324 return SQLITE_LOCKED_SHAREDCACHE
;
329 #endif /* !SQLITE_OMIT_SHARED_CACHE */
331 #ifndef SQLITE_OMIT_SHARED_CACHE
333 ** Add a lock on the table with root-page iTable to the shared-btree used
334 ** by Btree handle p. Parameter eLock must be either READ_LOCK or
337 ** This function assumes the following:
339 ** (a) The specified Btree object p is connected to a sharable
340 ** database (one with the BtShared.sharable flag set), and
342 ** (b) No other Btree objects hold a lock that conflicts
343 ** with the requested lock (i.e. querySharedCacheTableLock() has
344 ** already been called and returned SQLITE_OK).
346 ** SQLITE_OK is returned if the lock is added successfully. SQLITE_NOMEM
347 ** is returned if a malloc attempt fails.
349 static int setSharedCacheTableLock(Btree
*p
, Pgno iTable
, u8 eLock
){
350 BtShared
*pBt
= p
->pBt
;
354 assert( sqlite3BtreeHoldsMutex(p
) );
355 assert( eLock
==READ_LOCK
|| eLock
==WRITE_LOCK
);
358 /* A connection with the read-uncommitted flag set will never try to
359 ** obtain a read-lock using this function. The only read-lock obtained
360 ** by a connection in read-uncommitted mode is on the sqlite_schema
361 ** table, and that lock is obtained in BtreeBeginTrans(). */
362 assert( 0==(p
->db
->flags
&SQLITE_ReadUncommit
) || eLock
==WRITE_LOCK
);
364 /* This function should only be called on a sharable b-tree after it
365 ** has been determined that no other b-tree holds a conflicting lock. */
366 assert( p
->sharable
);
367 assert( SQLITE_OK
==querySharedCacheTableLock(p
, iTable
, eLock
) );
369 /* First search the list for an existing lock on this table. */
370 for(pIter
=pBt
->pLock
; pIter
; pIter
=pIter
->pNext
){
371 if( pIter
->iTable
==iTable
&& pIter
->pBtree
==p
){
377 /* If the above search did not find a BtLock struct associating Btree p
378 ** with table iTable, allocate one and link it into the list.
381 pLock
= (BtLock
*)sqlite3MallocZero(sizeof(BtLock
));
383 return SQLITE_NOMEM_BKPT
;
385 pLock
->iTable
= iTable
;
387 pLock
->pNext
= pBt
->pLock
;
391 /* Set the BtLock.eLock variable to the maximum of the current lock
392 ** and the requested lock. This means if a write-lock was already held
393 ** and a read-lock requested, we don't incorrectly downgrade the lock.
395 assert( WRITE_LOCK
>READ_LOCK
);
396 if( eLock
>pLock
->eLock
){
397 pLock
->eLock
= eLock
;
402 #endif /* !SQLITE_OMIT_SHARED_CACHE */
404 #ifndef SQLITE_OMIT_SHARED_CACHE
406 ** Release all the table locks (locks obtained via calls to
407 ** the setSharedCacheTableLock() procedure) held by Btree object p.
409 ** This function assumes that Btree p has an open read or write
410 ** transaction. If it does not, then the BTS_PENDING flag
411 ** may be incorrectly cleared.
413 static void clearAllSharedCacheTableLocks(Btree
*p
){
414 BtShared
*pBt
= p
->pBt
;
415 BtLock
**ppIter
= &pBt
->pLock
;
417 assert( sqlite3BtreeHoldsMutex(p
) );
418 assert( p
->sharable
|| 0==*ppIter
);
419 assert( p
->inTrans
>0 );
422 BtLock
*pLock
= *ppIter
;
423 assert( (pBt
->btsFlags
& BTS_EXCLUSIVE
)==0 || pBt
->pWriter
==pLock
->pBtree
);
424 assert( pLock
->pBtree
->inTrans
>=pLock
->eLock
);
425 if( pLock
->pBtree
==p
){
426 *ppIter
= pLock
->pNext
;
427 assert( pLock
->iTable
!=1 || pLock
==&p
->lock
);
428 if( pLock
->iTable
!=1 ){
432 ppIter
= &pLock
->pNext
;
436 assert( (pBt
->btsFlags
& BTS_PENDING
)==0 || pBt
->pWriter
);
437 if( pBt
->pWriter
==p
){
439 pBt
->btsFlags
&= ~(BTS_EXCLUSIVE
|BTS_PENDING
);
440 }else if( pBt
->nTransaction
==2 ){
441 /* This function is called when Btree p is concluding its
442 ** transaction. If there currently exists a writer, and p is not
443 ** that writer, then the number of locks held by connections other
444 ** than the writer must be about to drop to zero. In this case
445 ** set the BTS_PENDING flag to 0.
447 ** If there is not currently a writer, then BTS_PENDING must
448 ** be zero already. So this next line is harmless in that case.
450 pBt
->btsFlags
&= ~BTS_PENDING
;
455 ** This function changes all write-locks held by Btree p into read-locks.
457 static void downgradeAllSharedCacheTableLocks(Btree
*p
){
458 BtShared
*pBt
= p
->pBt
;
459 if( pBt
->pWriter
==p
){
462 pBt
->btsFlags
&= ~(BTS_EXCLUSIVE
|BTS_PENDING
);
463 for(pLock
=pBt
->pLock
; pLock
; pLock
=pLock
->pNext
){
464 assert( pLock
->eLock
==READ_LOCK
|| pLock
->pBtree
==p
);
465 pLock
->eLock
= READ_LOCK
;
470 #endif /* SQLITE_OMIT_SHARED_CACHE */
472 static void releasePage(MemPage
*pPage
); /* Forward reference */
473 static void releasePageOne(MemPage
*pPage
); /* Forward reference */
474 static void releasePageNotNull(MemPage
*pPage
); /* Forward reference */
477 ***** This routine is used inside of assert() only ****
479 ** Verify that the cursor holds the mutex on its BtShared
482 static int cursorHoldsMutex(BtCursor
*p
){
483 return sqlite3_mutex_held(p
->pBt
->mutex
);
486 /* Verify that the cursor and the BtShared agree about what is the current
487 ** database connetion. This is important in shared-cache mode. If the database
488 ** connection pointers get out-of-sync, it is possible for routines like
489 ** btreeInitPage() to reference an stale connection pointer that references a
490 ** a connection that has already closed. This routine is used inside assert()
491 ** statements only and for the purpose of double-checking that the btree code
492 ** does keep the database connection pointers up-to-date.
494 static int cursorOwnsBtShared(BtCursor
*p
){
495 assert( cursorHoldsMutex(p
) );
496 return (p
->pBtree
->db
==p
->pBt
->db
);
501 ** Invalidate the overflow cache of the cursor passed as the first argument.
502 ** on the shared btree structure pBt.
504 #define invalidateOverflowCache(pCur) (pCur->curFlags &= ~BTCF_ValidOvfl)
507 ** Invalidate the overflow page-list cache for all cursors opened
508 ** on the shared btree structure pBt.
510 static void invalidateAllOverflowCache(BtShared
*pBt
){
512 assert( sqlite3_mutex_held(pBt
->mutex
) );
513 for(p
=pBt
->pCursor
; p
; p
=p
->pNext
){
514 invalidateOverflowCache(p
);
518 #ifndef SQLITE_OMIT_INCRBLOB
520 ** This function is called before modifying the contents of a table
521 ** to invalidate any incrblob cursors that are open on the
522 ** row or one of the rows being modified.
524 ** If argument isClearTable is true, then the entire contents of the
525 ** table is about to be deleted. In this case invalidate all incrblob
526 ** cursors open on any row within the table with root-page pgnoRoot.
528 ** Otherwise, if argument isClearTable is false, then the row with
529 ** rowid iRow is being replaced or deleted. In this case invalidate
530 ** only those incrblob cursors open on that specific row.
532 static void invalidateIncrblobCursors(
533 Btree
*pBtree
, /* The database file to check */
534 Pgno pgnoRoot
, /* The table that might be changing */
535 i64 iRow
, /* The rowid that might be changing */
536 int isClearTable
/* True if all rows are being deleted */
539 if( pBtree
->hasIncrblobCur
==0 ) return;
540 assert( sqlite3BtreeHoldsMutex(pBtree
) );
541 pBtree
->hasIncrblobCur
= 0;
542 for(p
=pBtree
->pBt
->pCursor
; p
; p
=p
->pNext
){
543 if( (p
->curFlags
& BTCF_Incrblob
)!=0 ){
544 pBtree
->hasIncrblobCur
= 1;
545 if( p
->pgnoRoot
==pgnoRoot
&& (isClearTable
|| p
->info
.nKey
==iRow
) ){
546 p
->eState
= CURSOR_INVALID
;
553 /* Stub function when INCRBLOB is omitted */
554 #define invalidateIncrblobCursors(w,x,y,z)
555 #endif /* SQLITE_OMIT_INCRBLOB */
558 ** Set bit pgno of the BtShared.pHasContent bitvec. This is called
559 ** when a page that previously contained data becomes a free-list leaf
562 ** The BtShared.pHasContent bitvec exists to work around an obscure
563 ** bug caused by the interaction of two useful IO optimizations surrounding
564 ** free-list leaf pages:
566 ** 1) When all data is deleted from a page and the page becomes
567 ** a free-list leaf page, the page is not written to the database
568 ** (as free-list leaf pages contain no meaningful data). Sometimes
569 ** such a page is not even journalled (as it will not be modified,
570 ** why bother journalling it?).
572 ** 2) When a free-list leaf page is reused, its content is not read
573 ** from the database or written to the journal file (why should it
574 ** be, if it is not at all meaningful?).
576 ** By themselves, these optimizations work fine and provide a handy
577 ** performance boost to bulk delete or insert operations. However, if
578 ** a page is moved to the free-list and then reused within the same
579 ** transaction, a problem comes up. If the page is not journalled when
580 ** it is moved to the free-list and it is also not journalled when it
581 ** is extracted from the free-list and reused, then the original data
582 ** may be lost. In the event of a rollback, it may not be possible
583 ** to restore the database to its original configuration.
585 ** The solution is the BtShared.pHasContent bitvec. Whenever a page is
586 ** moved to become a free-list leaf page, the corresponding bit is
587 ** set in the bitvec. Whenever a leaf page is extracted from the free-list,
588 ** optimization 2 above is omitted if the corresponding bit is already
589 ** set in BtShared.pHasContent. The contents of the bitvec are cleared
590 ** at the end of every transaction.
592 static int btreeSetHasContent(BtShared
*pBt
, Pgno pgno
){
594 if( !pBt
->pHasContent
){
595 assert( pgno
<=pBt
->nPage
);
596 pBt
->pHasContent
= sqlite3BitvecCreate(pBt
->nPage
);
597 if( !pBt
->pHasContent
){
598 rc
= SQLITE_NOMEM_BKPT
;
601 if( rc
==SQLITE_OK
&& pgno
<=sqlite3BitvecSize(pBt
->pHasContent
) ){
602 rc
= sqlite3BitvecSet(pBt
->pHasContent
, pgno
);
608 ** Query the BtShared.pHasContent vector.
610 ** This function is called when a free-list leaf page is removed from the
611 ** free-list for reuse. It returns false if it is safe to retrieve the
612 ** page from the pager layer with the 'no-content' flag set. True otherwise.
614 static int btreeGetHasContent(BtShared
*pBt
, Pgno pgno
){
615 Bitvec
*p
= pBt
->pHasContent
;
616 return p
&& (pgno
>sqlite3BitvecSize(p
) || sqlite3BitvecTestNotNull(p
, pgno
));
620 ** Clear (destroy) the BtShared.pHasContent bitvec. This should be
621 ** invoked at the conclusion of each write-transaction.
623 static void btreeClearHasContent(BtShared
*pBt
){
624 sqlite3BitvecDestroy(pBt
->pHasContent
);
625 pBt
->pHasContent
= 0;
629 ** Release all of the apPage[] pages for a cursor.
631 static void btreeReleaseAllCursorPages(BtCursor
*pCur
){
633 if( pCur
->iPage
>=0 ){
634 for(i
=0; i
<pCur
->iPage
; i
++){
635 releasePageNotNull(pCur
->apPage
[i
]);
637 releasePageNotNull(pCur
->pPage
);
643 ** The cursor passed as the only argument must point to a valid entry
644 ** when this function is called (i.e. have eState==CURSOR_VALID). This
645 ** function saves the current cursor key in variables pCur->nKey and
646 ** pCur->pKey. SQLITE_OK is returned if successful or an SQLite error
649 ** If the cursor is open on an intkey table, then the integer key
650 ** (the rowid) is stored in pCur->nKey and pCur->pKey is left set to
651 ** NULL. If the cursor is open on a non-intkey table, then pCur->pKey is
652 ** set to point to a malloced buffer pCur->nKey bytes in size containing
655 static int saveCursorKey(BtCursor
*pCur
){
657 assert( CURSOR_VALID
==pCur
->eState
);
658 assert( 0==pCur
->pKey
);
659 assert( cursorHoldsMutex(pCur
) );
661 if( pCur
->curIntKey
){
662 /* Only the rowid is required for a table btree */
663 pCur
->nKey
= sqlite3BtreeIntegerKey(pCur
);
665 /* For an index btree, save the complete key content. It is possible
666 ** that the current key is corrupt. In that case, it is possible that
667 ** the sqlite3VdbeRecordUnpack() function may overread the buffer by
668 ** up to the size of 1 varint plus 1 8-byte value when the cursor
669 ** position is restored. Hence the 17 bytes of padding allocated
672 pCur
->nKey
= sqlite3BtreePayloadSize(pCur
);
673 pKey
= sqlite3Malloc( pCur
->nKey
+ 9 + 8 );
675 rc
= sqlite3BtreePayload(pCur
, 0, (int)pCur
->nKey
, pKey
);
677 memset(((u8
*)pKey
)+pCur
->nKey
, 0, 9+8);
683 rc
= SQLITE_NOMEM_BKPT
;
686 assert( !pCur
->curIntKey
|| !pCur
->pKey
);
691 ** Save the current cursor position in the variables BtCursor.nKey
692 ** and BtCursor.pKey. The cursor's state is set to CURSOR_REQUIRESEEK.
694 ** The caller must ensure that the cursor is valid (has eState==CURSOR_VALID)
695 ** prior to calling this routine.
697 static int saveCursorPosition(BtCursor
*pCur
){
700 assert( CURSOR_VALID
==pCur
->eState
|| CURSOR_SKIPNEXT
==pCur
->eState
);
701 assert( 0==pCur
->pKey
);
702 assert( cursorHoldsMutex(pCur
) );
704 if( pCur
->curFlags
& BTCF_Pinned
){
705 return SQLITE_CONSTRAINT_PINNED
;
707 if( pCur
->eState
==CURSOR_SKIPNEXT
){
708 pCur
->eState
= CURSOR_VALID
;
713 rc
= saveCursorKey(pCur
);
715 btreeReleaseAllCursorPages(pCur
);
716 pCur
->eState
= CURSOR_REQUIRESEEK
;
719 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
|BTCF_AtLast
);
723 /* Forward reference */
724 static int SQLITE_NOINLINE
saveCursorsOnList(BtCursor
*,Pgno
,BtCursor
*);
727 ** Save the positions of all cursors (except pExcept) that are open on
728 ** the table with root-page iRoot. "Saving the cursor position" means that
729 ** the location in the btree is remembered in such a way that it can be
730 ** moved back to the same spot after the btree has been modified. This
731 ** routine is called just before cursor pExcept is used to modify the
732 ** table, for example in BtreeDelete() or BtreeInsert().
734 ** If there are two or more cursors on the same btree, then all such
735 ** cursors should have their BTCF_Multiple flag set. The btreeCursor()
736 ** routine enforces that rule. This routine only needs to be called in
737 ** the uncommon case when pExpect has the BTCF_Multiple flag set.
739 ** If pExpect!=NULL and if no other cursors are found on the same root-page,
740 ** then the BTCF_Multiple flag on pExpect is cleared, to avoid another
741 ** pointless call to this routine.
743 ** Implementation note: This routine merely checks to see if any cursors
744 ** need to be saved. It calls out to saveCursorsOnList() in the (unusual)
745 ** event that cursors are in need to being saved.
747 static int saveAllCursors(BtShared
*pBt
, Pgno iRoot
, BtCursor
*pExcept
){
749 assert( sqlite3_mutex_held(pBt
->mutex
) );
750 assert( pExcept
==0 || pExcept
->pBt
==pBt
);
751 for(p
=pBt
->pCursor
; p
; p
=p
->pNext
){
752 if( p
!=pExcept
&& (0==iRoot
|| p
->pgnoRoot
==iRoot
) ) break;
754 if( p
) return saveCursorsOnList(p
, iRoot
, pExcept
);
755 if( pExcept
) pExcept
->curFlags
&= ~BTCF_Multiple
;
759 /* This helper routine to saveAllCursors does the actual work of saving
760 ** the cursors if and when a cursor is found that actually requires saving.
761 ** The common case is that no cursors need to be saved, so this routine is
762 ** broken out from its caller to avoid unnecessary stack pointer movement.
764 static int SQLITE_NOINLINE
saveCursorsOnList(
765 BtCursor
*p
, /* The first cursor that needs saving */
766 Pgno iRoot
, /* Only save cursor with this iRoot. Save all if zero */
767 BtCursor
*pExcept
/* Do not save this cursor */
770 if( p
!=pExcept
&& (0==iRoot
|| p
->pgnoRoot
==iRoot
) ){
771 if( p
->eState
==CURSOR_VALID
|| p
->eState
==CURSOR_SKIPNEXT
){
772 int rc
= saveCursorPosition(p
);
777 testcase( p
->iPage
>=0 );
778 btreeReleaseAllCursorPages(p
);
787 ** Clear the current cursor position.
789 void sqlite3BtreeClearCursor(BtCursor
*pCur
){
790 assert( cursorHoldsMutex(pCur
) );
791 sqlite3_free(pCur
->pKey
);
793 pCur
->eState
= CURSOR_INVALID
;
797 ** In this version of BtreeMoveto, pKey is a packed index record
798 ** such as is generated by the OP_MakeRecord opcode. Unpack the
799 ** record and then call BtreeMovetoUnpacked() to do the work.
801 static int btreeMoveto(
802 BtCursor
*pCur
, /* Cursor open on the btree to be searched */
803 const void *pKey
, /* Packed key if the btree is an index */
804 i64 nKey
, /* Integer key for tables. Size of pKey for indices */
805 int bias
, /* Bias search to the high end */
806 int *pRes
/* Write search results here */
808 int rc
; /* Status code */
809 UnpackedRecord
*pIdxKey
; /* Unpacked index key */
812 KeyInfo
*pKeyInfo
= pCur
->pKeyInfo
;
813 assert( nKey
==(i64
)(int)nKey
);
814 pIdxKey
= sqlite3VdbeAllocUnpackedRecord(pKeyInfo
);
815 if( pIdxKey
==0 ) return SQLITE_NOMEM_BKPT
;
816 sqlite3VdbeRecordUnpack(pKeyInfo
, (int)nKey
, pKey
, pIdxKey
);
817 if( pIdxKey
->nField
==0 || pIdxKey
->nField
>pKeyInfo
->nAllField
){
818 rc
= SQLITE_CORRUPT_BKPT
;
824 rc
= sqlite3BtreeMovetoUnpacked(pCur
, pIdxKey
, nKey
, bias
, pRes
);
827 sqlite3DbFree(pCur
->pKeyInfo
->db
, pIdxKey
);
833 ** Restore the cursor to the position it was in (or as close to as possible)
834 ** when saveCursorPosition() was called. Note that this call deletes the
835 ** saved position info stored by saveCursorPosition(), so there can be
836 ** at most one effective restoreCursorPosition() call after each
837 ** saveCursorPosition().
839 static int btreeRestoreCursorPosition(BtCursor
*pCur
){
842 assert( cursorOwnsBtShared(pCur
) );
843 assert( pCur
->eState
>=CURSOR_REQUIRESEEK
);
844 if( pCur
->eState
==CURSOR_FAULT
){
845 return pCur
->skipNext
;
847 pCur
->eState
= CURSOR_INVALID
;
848 if( sqlite3FaultSim(410) ){
851 rc
= btreeMoveto(pCur
, pCur
->pKey
, pCur
->nKey
, 0, &skipNext
);
854 sqlite3_free(pCur
->pKey
);
856 assert( pCur
->eState
==CURSOR_VALID
|| pCur
->eState
==CURSOR_INVALID
);
857 if( skipNext
) pCur
->skipNext
= skipNext
;
858 if( pCur
->skipNext
&& pCur
->eState
==CURSOR_VALID
){
859 pCur
->eState
= CURSOR_SKIPNEXT
;
865 #define restoreCursorPosition(p) \
866 (p->eState>=CURSOR_REQUIRESEEK ? \
867 btreeRestoreCursorPosition(p) : \
871 ** Determine whether or not a cursor has moved from the position where
872 ** it was last placed, or has been invalidated for any other reason.
873 ** Cursors can move when the row they are pointing at is deleted out
874 ** from under them, for example. Cursor might also move if a btree
877 ** Calling this routine with a NULL cursor pointer returns false.
879 ** Use the separate sqlite3BtreeCursorRestore() routine to restore a cursor
880 ** back to where it ought to be if this routine returns true.
882 int sqlite3BtreeCursorHasMoved(BtCursor
*pCur
){
883 assert( EIGHT_BYTE_ALIGNMENT(pCur
)
884 || pCur
==sqlite3BtreeFakeValidCursor() );
885 assert( offsetof(BtCursor
, eState
)==0 );
886 assert( sizeof(pCur
->eState
)==1 );
887 return CURSOR_VALID
!= *(u8
*)pCur
;
891 ** Return a pointer to a fake BtCursor object that will always answer
892 ** false to the sqlite3BtreeCursorHasMoved() routine above. The fake
893 ** cursor returned must not be used with any other Btree interface.
895 BtCursor
*sqlite3BtreeFakeValidCursor(void){
896 static u8 fakeCursor
= CURSOR_VALID
;
897 assert( offsetof(BtCursor
, eState
)==0 );
898 return (BtCursor
*)&fakeCursor
;
902 ** This routine restores a cursor back to its original position after it
903 ** has been moved by some outside activity (such as a btree rebalance or
904 ** a row having been deleted out from under the cursor).
906 ** On success, the *pDifferentRow parameter is false if the cursor is left
907 ** pointing at exactly the same row. *pDifferntRow is the row the cursor
908 ** was pointing to has been deleted, forcing the cursor to point to some
911 ** This routine should only be called for a cursor that just returned
912 ** TRUE from sqlite3BtreeCursorHasMoved().
914 int sqlite3BtreeCursorRestore(BtCursor
*pCur
, int *pDifferentRow
){
918 assert( pCur
->eState
!=CURSOR_VALID
);
919 rc
= restoreCursorPosition(pCur
);
924 if( pCur
->eState
!=CURSOR_VALID
){
932 #ifdef SQLITE_ENABLE_CURSOR_HINTS
934 ** Provide hints to the cursor. The particular hint given (and the type
935 ** and number of the varargs parameters) is determined by the eHintType
936 ** parameter. See the definitions of the BTREE_HINT_* macros for details.
938 void sqlite3BtreeCursorHint(BtCursor
*pCur
, int eHintType
, ...){
939 /* Used only by system that substitute their own storage engine */
944 ** Provide flag hints to the cursor.
946 void sqlite3BtreeCursorHintFlags(BtCursor
*pCur
, unsigned x
){
947 assert( x
==BTREE_SEEK_EQ
|| x
==BTREE_BULKLOAD
|| x
==0 );
952 #ifndef SQLITE_OMIT_AUTOVACUUM
954 ** Given a page number of a regular database page, return the page
955 ** number for the pointer-map page that contains the entry for the
956 ** input page number.
958 ** Return 0 (not a valid page) for pgno==1 since there is
959 ** no pointer map associated with page 1. The integrity_check logic
960 ** requires that ptrmapPageno(*,1)!=1.
962 static Pgno
ptrmapPageno(BtShared
*pBt
, Pgno pgno
){
963 int nPagesPerMapPage
;
965 assert( sqlite3_mutex_held(pBt
->mutex
) );
966 if( pgno
<2 ) return 0;
967 nPagesPerMapPage
= (pBt
->usableSize
/5)+1;
968 iPtrMap
= (pgno
-2)/nPagesPerMapPage
;
969 ret
= (iPtrMap
*nPagesPerMapPage
) + 2;
970 if( ret
==PENDING_BYTE_PAGE(pBt
) ){
977 ** Write an entry into the pointer map.
979 ** This routine updates the pointer map entry for page number 'key'
980 ** so that it maps to type 'eType' and parent page number 'pgno'.
982 ** If *pRC is initially non-zero (non-SQLITE_OK) then this routine is
983 ** a no-op. If an error occurs, the appropriate error code is written
986 static void ptrmapPut(BtShared
*pBt
, Pgno key
, u8 eType
, Pgno parent
, int *pRC
){
987 DbPage
*pDbPage
; /* The pointer map page */
988 u8
*pPtrmap
; /* The pointer map data */
989 Pgno iPtrmap
; /* The pointer map page number */
990 int offset
; /* Offset in pointer map page */
991 int rc
; /* Return code from subfunctions */
995 assert( sqlite3_mutex_held(pBt
->mutex
) );
996 /* The super-journal page number must never be used as a pointer map page */
997 assert( 0==PTRMAP_ISPAGE(pBt
, PENDING_BYTE_PAGE(pBt
)) );
999 assert( pBt
->autoVacuum
);
1001 *pRC
= SQLITE_CORRUPT_BKPT
;
1004 iPtrmap
= PTRMAP_PAGENO(pBt
, key
);
1005 rc
= sqlite3PagerGet(pBt
->pPager
, iPtrmap
, &pDbPage
, 0);
1006 if( rc
!=SQLITE_OK
){
1010 if( ((char*)sqlite3PagerGetExtra(pDbPage
))[0]!=0 ){
1011 /* The first byte of the extra data is the MemPage.isInit byte.
1012 ** If that byte is set, it means this page is also being used
1013 ** as a btree page. */
1014 *pRC
= SQLITE_CORRUPT_BKPT
;
1017 offset
= PTRMAP_PTROFFSET(iPtrmap
, key
);
1019 *pRC
= SQLITE_CORRUPT_BKPT
;
1022 assert( offset
<= (int)pBt
->usableSize
-5 );
1023 pPtrmap
= (u8
*)sqlite3PagerGetData(pDbPage
);
1025 if( eType
!=pPtrmap
[offset
] || get4byte(&pPtrmap
[offset
+1])!=parent
){
1026 TRACE(("PTRMAP_UPDATE: %d->(%d,%d)\n", key
, eType
, parent
));
1027 *pRC
= rc
= sqlite3PagerWrite(pDbPage
);
1028 if( rc
==SQLITE_OK
){
1029 pPtrmap
[offset
] = eType
;
1030 put4byte(&pPtrmap
[offset
+1], parent
);
1035 sqlite3PagerUnref(pDbPage
);
1039 ** Read an entry from the pointer map.
1041 ** This routine retrieves the pointer map entry for page 'key', writing
1042 ** the type and parent page number to *pEType and *pPgno respectively.
1043 ** An error code is returned if something goes wrong, otherwise SQLITE_OK.
1045 static int ptrmapGet(BtShared
*pBt
, Pgno key
, u8
*pEType
, Pgno
*pPgno
){
1046 DbPage
*pDbPage
; /* The pointer map page */
1047 int iPtrmap
; /* Pointer map page index */
1048 u8
*pPtrmap
; /* Pointer map page data */
1049 int offset
; /* Offset of entry in pointer map */
1052 assert( sqlite3_mutex_held(pBt
->mutex
) );
1054 iPtrmap
= PTRMAP_PAGENO(pBt
, key
);
1055 rc
= sqlite3PagerGet(pBt
->pPager
, iPtrmap
, &pDbPage
, 0);
1059 pPtrmap
= (u8
*)sqlite3PagerGetData(pDbPage
);
1061 offset
= PTRMAP_PTROFFSET(iPtrmap
, key
);
1063 sqlite3PagerUnref(pDbPage
);
1064 return SQLITE_CORRUPT_BKPT
;
1066 assert( offset
<= (int)pBt
->usableSize
-5 );
1067 assert( pEType
!=0 );
1068 *pEType
= pPtrmap
[offset
];
1069 if( pPgno
) *pPgno
= get4byte(&pPtrmap
[offset
+1]);
1071 sqlite3PagerUnref(pDbPage
);
1072 if( *pEType
<1 || *pEType
>5 ) return SQLITE_CORRUPT_PGNO(iPtrmap
);
1076 #else /* if defined SQLITE_OMIT_AUTOVACUUM */
1077 #define ptrmapPut(w,x,y,z,rc)
1078 #define ptrmapGet(w,x,y,z) SQLITE_OK
1079 #define ptrmapPutOvflPtr(x, y, z, rc)
1083 ** Given a btree page and a cell index (0 means the first cell on
1084 ** the page, 1 means the second cell, and so forth) return a pointer
1085 ** to the cell content.
1087 ** findCellPastPtr() does the same except it skips past the initial
1088 ** 4-byte child pointer found on interior pages, if there is one.
1090 ** This routine works only for pages that do not contain overflow cells.
1092 #define findCell(P,I) \
1093 ((P)->aData + ((P)->maskPage & get2byteAligned(&(P)->aCellIdx[2*(I)])))
1094 #define findCellPastPtr(P,I) \
1095 ((P)->aDataOfst + ((P)->maskPage & get2byteAligned(&(P)->aCellIdx[2*(I)])))
1099 ** This is common tail processing for btreeParseCellPtr() and
1100 ** btreeParseCellPtrIndex() for the case when the cell does not fit entirely
1101 ** on a single B-tree page. Make necessary adjustments to the CellInfo
1104 static SQLITE_NOINLINE
void btreeParseCellAdjustSizeForOverflow(
1105 MemPage
*pPage
, /* Page containing the cell */
1106 u8
*pCell
, /* Pointer to the cell text. */
1107 CellInfo
*pInfo
/* Fill in this structure */
1109 /* If the payload will not fit completely on the local page, we have
1110 ** to decide how much to store locally and how much to spill onto
1111 ** overflow pages. The strategy is to minimize the amount of unused
1112 ** space on overflow pages while keeping the amount of local storage
1113 ** in between minLocal and maxLocal.
1115 ** Warning: changing the way overflow payload is distributed in any
1116 ** way will result in an incompatible file format.
1118 int minLocal
; /* Minimum amount of payload held locally */
1119 int maxLocal
; /* Maximum amount of payload held locally */
1120 int surplus
; /* Overflow payload available for local storage */
1122 minLocal
= pPage
->minLocal
;
1123 maxLocal
= pPage
->maxLocal
;
1124 surplus
= minLocal
+ (pInfo
->nPayload
- minLocal
)%(pPage
->pBt
->usableSize
-4);
1125 testcase( surplus
==maxLocal
);
1126 testcase( surplus
==maxLocal
+1 );
1127 if( surplus
<= maxLocal
){
1128 pInfo
->nLocal
= (u16
)surplus
;
1130 pInfo
->nLocal
= (u16
)minLocal
;
1132 pInfo
->nSize
= (u16
)(&pInfo
->pPayload
[pInfo
->nLocal
] - pCell
) + 4;
1136 ** The following routines are implementations of the MemPage.xParseCell()
1139 ** Parse a cell content block and fill in the CellInfo structure.
1141 ** btreeParseCellPtr() => table btree leaf nodes
1142 ** btreeParseCellNoPayload() => table btree internal nodes
1143 ** btreeParseCellPtrIndex() => index btree nodes
1145 ** There is also a wrapper function btreeParseCell() that works for
1146 ** all MemPage types and that references the cell by index rather than
1149 static void btreeParseCellPtrNoPayload(
1150 MemPage
*pPage
, /* Page containing the cell */
1151 u8
*pCell
, /* Pointer to the cell text. */
1152 CellInfo
*pInfo
/* Fill in this structure */
1154 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1155 assert( pPage
->leaf
==0 );
1156 assert( pPage
->childPtrSize
==4 );
1157 #ifndef SQLITE_DEBUG
1158 UNUSED_PARAMETER(pPage
);
1160 pInfo
->nSize
= 4 + getVarint(&pCell
[4], (u64
*)&pInfo
->nKey
);
1161 pInfo
->nPayload
= 0;
1163 pInfo
->pPayload
= 0;
1166 static void btreeParseCellPtr(
1167 MemPage
*pPage
, /* Page containing the cell */
1168 u8
*pCell
, /* Pointer to the cell text. */
1169 CellInfo
*pInfo
/* Fill in this structure */
1171 u8
*pIter
; /* For scanning through pCell */
1172 u32 nPayload
; /* Number of bytes of cell payload */
1173 u64 iKey
; /* Extracted Key value */
1175 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1176 assert( pPage
->leaf
==0 || pPage
->leaf
==1 );
1177 assert( pPage
->intKeyLeaf
);
1178 assert( pPage
->childPtrSize
==0 );
1181 /* The next block of code is equivalent to:
1183 ** pIter += getVarint32(pIter, nPayload);
1185 ** The code is inlined to avoid a function call.
1188 if( nPayload
>=0x80 ){
1189 u8
*pEnd
= &pIter
[8];
1192 nPayload
= (nPayload
<<7) | (*++pIter
& 0x7f);
1193 }while( (*pIter
)>=0x80 && pIter
<pEnd
);
1197 /* The next block of code is equivalent to:
1199 ** pIter += getVarint(pIter, (u64*)&pInfo->nKey);
1201 ** The code is inlined to avoid a function call.
1205 u8
*pEnd
= &pIter
[7];
1208 iKey
= (iKey
<<7) | (*++pIter
& 0x7f);
1209 if( (*pIter
)<0x80 ) break;
1211 iKey
= (iKey
<<8) | *++pIter
;
1218 pInfo
->nKey
= *(i64
*)&iKey
;
1219 pInfo
->nPayload
= nPayload
;
1220 pInfo
->pPayload
= pIter
;
1221 testcase( nPayload
==pPage
->maxLocal
);
1222 testcase( nPayload
==pPage
->maxLocal
+1 );
1223 if( nPayload
<=pPage
->maxLocal
){
1224 /* This is the (easy) common case where the entire payload fits
1225 ** on the local page. No overflow is required.
1227 pInfo
->nSize
= nPayload
+ (u16
)(pIter
- pCell
);
1228 if( pInfo
->nSize
<4 ) pInfo
->nSize
= 4;
1229 pInfo
->nLocal
= (u16
)nPayload
;
1231 btreeParseCellAdjustSizeForOverflow(pPage
, pCell
, pInfo
);
1234 static void btreeParseCellPtrIndex(
1235 MemPage
*pPage
, /* Page containing the cell */
1236 u8
*pCell
, /* Pointer to the cell text. */
1237 CellInfo
*pInfo
/* Fill in this structure */
1239 u8
*pIter
; /* For scanning through pCell */
1240 u32 nPayload
; /* Number of bytes of cell payload */
1242 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1243 assert( pPage
->leaf
==0 || pPage
->leaf
==1 );
1244 assert( pPage
->intKeyLeaf
==0 );
1245 pIter
= pCell
+ pPage
->childPtrSize
;
1247 if( nPayload
>=0x80 ){
1248 u8
*pEnd
= &pIter
[8];
1251 nPayload
= (nPayload
<<7) | (*++pIter
& 0x7f);
1252 }while( *(pIter
)>=0x80 && pIter
<pEnd
);
1255 pInfo
->nKey
= nPayload
;
1256 pInfo
->nPayload
= nPayload
;
1257 pInfo
->pPayload
= pIter
;
1258 testcase( nPayload
==pPage
->maxLocal
);
1259 testcase( nPayload
==pPage
->maxLocal
+1 );
1260 if( nPayload
<=pPage
->maxLocal
){
1261 /* This is the (easy) common case where the entire payload fits
1262 ** on the local page. No overflow is required.
1264 pInfo
->nSize
= nPayload
+ (u16
)(pIter
- pCell
);
1265 if( pInfo
->nSize
<4 ) pInfo
->nSize
= 4;
1266 pInfo
->nLocal
= (u16
)nPayload
;
1268 btreeParseCellAdjustSizeForOverflow(pPage
, pCell
, pInfo
);
1271 static void btreeParseCell(
1272 MemPage
*pPage
, /* Page containing the cell */
1273 int iCell
, /* The cell index. First cell is 0 */
1274 CellInfo
*pInfo
/* Fill in this structure */
1276 pPage
->xParseCell(pPage
, findCell(pPage
, iCell
), pInfo
);
1280 ** The following routines are implementations of the MemPage.xCellSize
1283 ** Compute the total number of bytes that a Cell needs in the cell
1284 ** data area of the btree-page. The return number includes the cell
1285 ** data header and the local payload, but not any overflow page or
1286 ** the space used by the cell pointer.
1288 ** cellSizePtrNoPayload() => table internal nodes
1289 ** cellSizePtr() => all index nodes & table leaf nodes
1291 static u16
cellSizePtr(MemPage
*pPage
, u8
*pCell
){
1292 u8
*pIter
= pCell
+ pPage
->childPtrSize
; /* For looping over bytes of pCell */
1293 u8
*pEnd
; /* End mark for a varint */
1294 u32 nSize
; /* Size value to return */
1297 /* The value returned by this function should always be the same as
1298 ** the (CellInfo.nSize) value found by doing a full parse of the
1299 ** cell. If SQLITE_DEBUG is defined, an assert() at the bottom of
1300 ** this function verifies that this invariant is not violated. */
1302 pPage
->xParseCell(pPage
, pCell
, &debuginfo
);
1310 nSize
= (nSize
<<7) | (*++pIter
& 0x7f);
1311 }while( *(pIter
)>=0x80 && pIter
<pEnd
);
1314 if( pPage
->intKey
){
1315 /* pIter now points at the 64-bit integer key value, a variable length
1316 ** integer. The following block moves pIter to point at the first byte
1317 ** past the end of the key value. */
1319 while( (*pIter
++)&0x80 && pIter
<pEnd
);
1321 testcase( nSize
==pPage
->maxLocal
);
1322 testcase( nSize
==pPage
->maxLocal
+1 );
1323 if( nSize
<=pPage
->maxLocal
){
1324 nSize
+= (u32
)(pIter
- pCell
);
1325 if( nSize
<4 ) nSize
= 4;
1327 int minLocal
= pPage
->minLocal
;
1328 nSize
= minLocal
+ (nSize
- minLocal
) % (pPage
->pBt
->usableSize
- 4);
1329 testcase( nSize
==pPage
->maxLocal
);
1330 testcase( nSize
==pPage
->maxLocal
+1 );
1331 if( nSize
>pPage
->maxLocal
){
1334 nSize
+= 4 + (u16
)(pIter
- pCell
);
1336 assert( nSize
==debuginfo
.nSize
|| CORRUPT_DB
);
1339 static u16
cellSizePtrNoPayload(MemPage
*pPage
, u8
*pCell
){
1340 u8
*pIter
= pCell
+ 4; /* For looping over bytes of pCell */
1341 u8
*pEnd
; /* End mark for a varint */
1344 /* The value returned by this function should always be the same as
1345 ** the (CellInfo.nSize) value found by doing a full parse of the
1346 ** cell. If SQLITE_DEBUG is defined, an assert() at the bottom of
1347 ** this function verifies that this invariant is not violated. */
1349 pPage
->xParseCell(pPage
, pCell
, &debuginfo
);
1351 UNUSED_PARAMETER(pPage
);
1354 assert( pPage
->childPtrSize
==4 );
1356 while( (*pIter
++)&0x80 && pIter
<pEnd
);
1357 assert( debuginfo
.nSize
==(u16
)(pIter
- pCell
) || CORRUPT_DB
);
1358 return (u16
)(pIter
- pCell
);
1363 /* This variation on cellSizePtr() is used inside of assert() statements
1365 static u16
cellSize(MemPage
*pPage
, int iCell
){
1366 return pPage
->xCellSize(pPage
, findCell(pPage
, iCell
));
1370 #ifndef SQLITE_OMIT_AUTOVACUUM
1372 ** The cell pCell is currently part of page pSrc but will ultimately be part
1373 ** of pPage. (pSrc and pPager are often the same.) If pCell contains a
1374 ** pointer to an overflow page, insert an entry into the pointer-map for
1375 ** the overflow page that will be valid after pCell has been moved to pPage.
1377 static void ptrmapPutOvflPtr(MemPage
*pPage
, MemPage
*pSrc
, u8
*pCell
,int *pRC
){
1381 pPage
->xParseCell(pPage
, pCell
, &info
);
1382 if( info
.nLocal
<info
.nPayload
){
1384 if( SQLITE_WITHIN(pSrc
->aDataEnd
, pCell
, pCell
+info
.nLocal
) ){
1385 testcase( pSrc
!=pPage
);
1386 *pRC
= SQLITE_CORRUPT_BKPT
;
1389 ovfl
= get4byte(&pCell
[info
.nSize
-4]);
1390 ptrmapPut(pPage
->pBt
, ovfl
, PTRMAP_OVERFLOW1
, pPage
->pgno
, pRC
);
1397 ** Defragment the page given. This routine reorganizes cells within the
1398 ** page so that there are no free-blocks on the free-block list.
1400 ** Parameter nMaxFrag is the maximum amount of fragmented space that may be
1401 ** present in the page after this routine returns.
1403 ** EVIDENCE-OF: R-44582-60138 SQLite may from time to time reorganize a
1404 ** b-tree page so that there are no freeblocks or fragment bytes, all
1405 ** unused bytes are contained in the unallocated space region, and all
1406 ** cells are packed tightly at the end of the page.
1408 static int defragmentPage(MemPage
*pPage
, int nMaxFrag
){
1409 int i
; /* Loop counter */
1410 int pc
; /* Address of the i-th cell */
1411 int hdr
; /* Offset to the page header */
1412 int size
; /* Size of a cell */
1413 int usableSize
; /* Number of usable bytes on a page */
1414 int cellOffset
; /* Offset to the cell pointer array */
1415 int cbrk
; /* Offset to the cell content area */
1416 int nCell
; /* Number of cells on the page */
1417 unsigned char *data
; /* The page data */
1418 unsigned char *temp
; /* Temp area for cell content */
1419 unsigned char *src
; /* Source of content */
1420 int iCellFirst
; /* First allowable cell index */
1421 int iCellLast
; /* Last possible cell index */
1423 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1424 assert( pPage
->pBt
!=0 );
1425 assert( pPage
->pBt
->usableSize
<= SQLITE_MAX_PAGE_SIZE
);
1426 assert( pPage
->nOverflow
==0 );
1427 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1429 src
= data
= pPage
->aData
;
1430 hdr
= pPage
->hdrOffset
;
1431 cellOffset
= pPage
->cellOffset
;
1432 nCell
= pPage
->nCell
;
1433 assert( nCell
==get2byte(&data
[hdr
+3]) || CORRUPT_DB
);
1434 iCellFirst
= cellOffset
+ 2*nCell
;
1435 usableSize
= pPage
->pBt
->usableSize
;
1437 /* This block handles pages with two or fewer free blocks and nMaxFrag
1438 ** or fewer fragmented bytes. In this case it is faster to move the
1439 ** two (or one) blocks of cells using memmove() and add the required
1440 ** offsets to each pointer in the cell-pointer array than it is to
1441 ** reconstruct the entire page. */
1442 if( (int)data
[hdr
+7]<=nMaxFrag
){
1443 int iFree
= get2byte(&data
[hdr
+1]);
1444 if( iFree
>usableSize
-4 ) return SQLITE_CORRUPT_PAGE(pPage
);
1446 int iFree2
= get2byte(&data
[iFree
]);
1447 if( iFree2
>usableSize
-4 ) return SQLITE_CORRUPT_PAGE(pPage
);
1448 if( 0==iFree2
|| (data
[iFree2
]==0 && data
[iFree2
+1]==0) ){
1449 u8
*pEnd
= &data
[cellOffset
+ nCell
*2];
1452 int sz
= get2byte(&data
[iFree
+2]);
1453 int top
= get2byte(&data
[hdr
+5]);
1455 return SQLITE_CORRUPT_PAGE(pPage
);
1458 if( iFree
+sz
>iFree2
) return SQLITE_CORRUPT_PAGE(pPage
);
1459 sz2
= get2byte(&data
[iFree2
+2]);
1460 if( iFree2
+sz2
> usableSize
) return SQLITE_CORRUPT_PAGE(pPage
);
1461 memmove(&data
[iFree
+sz
+sz2
], &data
[iFree
+sz
], iFree2
-(iFree
+sz
));
1463 }else if( NEVER(iFree
+sz
>usableSize
) ){
1464 return SQLITE_CORRUPT_PAGE(pPage
);
1468 assert( cbrk
+(iFree
-top
) <= usableSize
);
1469 memmove(&data
[cbrk
], &data
[top
], iFree
-top
);
1470 for(pAddr
=&data
[cellOffset
]; pAddr
<pEnd
; pAddr
+=2){
1471 pc
= get2byte(pAddr
);
1472 if( pc
<iFree
){ put2byte(pAddr
, pc
+sz
); }
1473 else if( pc
<iFree2
){ put2byte(pAddr
, pc
+sz2
); }
1475 goto defragment_out
;
1481 iCellLast
= usableSize
- 4;
1482 for(i
=0; i
<nCell
; i
++){
1483 u8
*pAddr
; /* The i-th cell pointer */
1484 pAddr
= &data
[cellOffset
+ i
*2];
1485 pc
= get2byte(pAddr
);
1486 testcase( pc
==iCellFirst
);
1487 testcase( pc
==iCellLast
);
1488 /* These conditions have already been verified in btreeInitPage()
1489 ** if PRAGMA cell_size_check=ON.
1491 if( pc
<iCellFirst
|| pc
>iCellLast
){
1492 return SQLITE_CORRUPT_PAGE(pPage
);
1494 assert( pc
>=iCellFirst
&& pc
<=iCellLast
);
1495 size
= pPage
->xCellSize(pPage
, &src
[pc
]);
1497 if( cbrk
<iCellFirst
|| pc
+size
>usableSize
){
1498 return SQLITE_CORRUPT_PAGE(pPage
);
1500 assert( cbrk
+size
<=usableSize
&& cbrk
>=iCellFirst
);
1501 testcase( cbrk
+size
==usableSize
);
1502 testcase( pc
+size
==usableSize
);
1503 put2byte(pAddr
, cbrk
);
1506 if( cbrk
==pc
) continue;
1507 temp
= sqlite3PagerTempSpace(pPage
->pBt
->pPager
);
1508 x
= get2byte(&data
[hdr
+5]);
1509 memcpy(&temp
[x
], &data
[x
], (cbrk
+size
) - x
);
1512 memcpy(&data
[cbrk
], &src
[pc
], size
);
1517 assert( pPage
->nFree
>=0 );
1518 if( data
[hdr
+7]+cbrk
-iCellFirst
!=pPage
->nFree
){
1519 return SQLITE_CORRUPT_PAGE(pPage
);
1521 assert( cbrk
>=iCellFirst
);
1522 put2byte(&data
[hdr
+5], cbrk
);
1525 memset(&data
[iCellFirst
], 0, cbrk
-iCellFirst
);
1526 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1531 ** Search the free-list on page pPg for space to store a cell nByte bytes in
1532 ** size. If one can be found, return a pointer to the space and remove it
1533 ** from the free-list.
1535 ** If no suitable space can be found on the free-list, return NULL.
1537 ** This function may detect corruption within pPg. If corruption is
1538 ** detected then *pRc is set to SQLITE_CORRUPT and NULL is returned.
1540 ** Slots on the free list that are between 1 and 3 bytes larger than nByte
1541 ** will be ignored if adding the extra space to the fragmentation count
1542 ** causes the fragmentation count to exceed 60.
1544 static u8
*pageFindSlot(MemPage
*pPg
, int nByte
, int *pRc
){
1545 const int hdr
= pPg
->hdrOffset
; /* Offset to page header */
1546 u8
* const aData
= pPg
->aData
; /* Page data */
1547 int iAddr
= hdr
+ 1; /* Address of ptr to pc */
1548 int pc
= get2byte(&aData
[iAddr
]); /* Address of a free slot */
1549 int x
; /* Excess size of the slot */
1550 int maxPC
= pPg
->pBt
->usableSize
- nByte
; /* Max address for a usable slot */
1551 int size
; /* Size of the free slot */
1555 /* EVIDENCE-OF: R-22710-53328 The third and fourth bytes of each
1556 ** freeblock form a big-endian integer which is the size of the freeblock
1557 ** in bytes, including the 4-byte header. */
1558 size
= get2byte(&aData
[pc
+2]);
1559 if( (x
= size
- nByte
)>=0 ){
1563 /* EVIDENCE-OF: R-11498-58022 In a well-formed b-tree page, the total
1564 ** number of bytes in fragments may not exceed 60. */
1565 if( aData
[hdr
+7]>57 ) return 0;
1567 /* Remove the slot from the free-list. Update the number of
1568 ** fragmented bytes within the page. */
1569 memcpy(&aData
[iAddr
], &aData
[pc
], 2);
1570 aData
[hdr
+7] += (u8
)x
;
1571 }else if( x
+pc
> maxPC
){
1572 /* This slot extends off the end of the usable part of the page */
1573 *pRc
= SQLITE_CORRUPT_PAGE(pPg
);
1576 /* The slot remains on the free-list. Reduce its size to account
1577 ** for the portion used by the new allocation. */
1578 put2byte(&aData
[pc
+2], x
);
1580 return &aData
[pc
+ x
];
1583 pc
= get2byte(&aData
[pc
]);
1584 if( pc
<=iAddr
+size
){
1586 /* The next slot in the chain is not past the end of the current slot */
1587 *pRc
= SQLITE_CORRUPT_PAGE(pPg
);
1592 if( pc
>maxPC
+nByte
-4 ){
1593 /* The free slot chain extends off the end of the page */
1594 *pRc
= SQLITE_CORRUPT_PAGE(pPg
);
1600 ** Allocate nByte bytes of space from within the B-Tree page passed
1601 ** as the first argument. Write into *pIdx the index into pPage->aData[]
1602 ** of the first byte of allocated space. Return either SQLITE_OK or
1603 ** an error code (usually SQLITE_CORRUPT).
1605 ** The caller guarantees that there is sufficient space to make the
1606 ** allocation. This routine might need to defragment in order to bring
1607 ** all the space together, however. This routine will avoid using
1608 ** the first two bytes past the cell pointer area since presumably this
1609 ** allocation is being made in order to insert a new cell, so we will
1610 ** also end up needing a new cell pointer.
1612 static int allocateSpace(MemPage
*pPage
, int nByte
, int *pIdx
){
1613 const int hdr
= pPage
->hdrOffset
; /* Local cache of pPage->hdrOffset */
1614 u8
* const data
= pPage
->aData
; /* Local cache of pPage->aData */
1615 int top
; /* First byte of cell content area */
1616 int rc
= SQLITE_OK
; /* Integer return code */
1617 int gap
; /* First byte of gap between cell pointers and cell content */
1619 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1620 assert( pPage
->pBt
);
1621 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1622 assert( nByte
>=0 ); /* Minimum cell size is 4 */
1623 assert( pPage
->nFree
>=nByte
);
1624 assert( pPage
->nOverflow
==0 );
1625 assert( nByte
< (int)(pPage
->pBt
->usableSize
-8) );
1627 assert( pPage
->cellOffset
== hdr
+ 12 - 4*pPage
->leaf
);
1628 gap
= pPage
->cellOffset
+ 2*pPage
->nCell
;
1629 assert( gap
<=65536 );
1630 /* EVIDENCE-OF: R-29356-02391 If the database uses a 65536-byte page size
1631 ** and the reserved space is zero (the usual value for reserved space)
1632 ** then the cell content offset of an empty page wants to be 65536.
1633 ** However, that integer is too large to be stored in a 2-byte unsigned
1634 ** integer, so a value of 0 is used in its place. */
1635 top
= get2byte(&data
[hdr
+5]);
1636 assert( top
<=(int)pPage
->pBt
->usableSize
); /* by btreeComputeFreeSpace() */
1638 if( top
==0 && pPage
->pBt
->usableSize
==65536 ){
1641 return SQLITE_CORRUPT_PAGE(pPage
);
1645 /* If there is enough space between gap and top for one more cell pointer,
1646 ** and if the freelist is not empty, then search the
1647 ** freelist looking for a slot big enough to satisfy the request.
1649 testcase( gap
+2==top
);
1650 testcase( gap
+1==top
);
1651 testcase( gap
==top
);
1652 if( (data
[hdr
+2] || data
[hdr
+1]) && gap
+2<=top
){
1653 u8
*pSpace
= pageFindSlot(pPage
, nByte
, &rc
);
1656 assert( pSpace
+nByte
<=data
+pPage
->pBt
->usableSize
);
1657 *pIdx
= g2
= (int)(pSpace
-data
);
1658 if( NEVER(g2
<=gap
) ){
1659 return SQLITE_CORRUPT_PAGE(pPage
);
1668 /* The request could not be fulfilled using a freelist slot. Check
1669 ** to see if defragmentation is necessary.
1671 testcase( gap
+2+nByte
==top
);
1672 if( gap
+2+nByte
>top
){
1673 assert( pPage
->nCell
>0 || CORRUPT_DB
);
1674 assert( pPage
->nFree
>=0 );
1675 rc
= defragmentPage(pPage
, MIN(4, pPage
->nFree
- (2+nByte
)));
1677 top
= get2byteNotZero(&data
[hdr
+5]);
1678 assert( gap
+2+nByte
<=top
);
1682 /* Allocate memory from the gap in between the cell pointer array
1683 ** and the cell content area. The btreeComputeFreeSpace() call has already
1684 ** validated the freelist. Given that the freelist is valid, there
1685 ** is no way that the allocation can extend off the end of the page.
1686 ** The assert() below verifies the previous sentence.
1689 put2byte(&data
[hdr
+5], top
);
1690 assert( top
+nByte
<= (int)pPage
->pBt
->usableSize
);
1696 ** Return a section of the pPage->aData to the freelist.
1697 ** The first byte of the new free block is pPage->aData[iStart]
1698 ** and the size of the block is iSize bytes.
1700 ** Adjacent freeblocks are coalesced.
1702 ** Even though the freeblock list was checked by btreeComputeFreeSpace(),
1703 ** that routine will not detect overlap between cells or freeblocks. Nor
1704 ** does it detect cells or freeblocks that encrouch into the reserved bytes
1705 ** at the end of the page. So do additional corruption checks inside this
1706 ** routine and return SQLITE_CORRUPT if any problems are found.
1708 static int freeSpace(MemPage
*pPage
, u16 iStart
, u16 iSize
){
1709 u16 iPtr
; /* Address of ptr to next freeblock */
1710 u16 iFreeBlk
; /* Address of the next freeblock */
1711 u8 hdr
; /* Page header size. 0 or 100 */
1712 u8 nFrag
= 0; /* Reduction in fragmentation */
1713 u16 iOrigSize
= iSize
; /* Original value of iSize */
1714 u16 x
; /* Offset to cell content area */
1715 u32 iEnd
= iStart
+ iSize
; /* First byte past the iStart buffer */
1716 unsigned char *data
= pPage
->aData
; /* Page content */
1718 assert( pPage
->pBt
!=0 );
1719 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
1720 assert( CORRUPT_DB
|| iStart
>=pPage
->hdrOffset
+6+pPage
->childPtrSize
);
1721 assert( CORRUPT_DB
|| iEnd
<= pPage
->pBt
->usableSize
);
1722 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1723 assert( iSize
>=4 ); /* Minimum cell size is 4 */
1724 assert( iStart
<=pPage
->pBt
->usableSize
-4 );
1726 /* The list of freeblocks must be in ascending order. Find the
1727 ** spot on the list where iStart should be inserted.
1729 hdr
= pPage
->hdrOffset
;
1731 if( data
[iPtr
+1]==0 && data
[iPtr
]==0 ){
1732 iFreeBlk
= 0; /* Shortcut for the case when the freelist is empty */
1734 while( (iFreeBlk
= get2byte(&data
[iPtr
]))<iStart
){
1735 if( iFreeBlk
<iPtr
+4 ){
1736 if( iFreeBlk
==0 ) break; /* TH3: corrupt082.100 */
1737 return SQLITE_CORRUPT_PAGE(pPage
);
1741 if( iFreeBlk
>pPage
->pBt
->usableSize
-4 ){ /* TH3: corrupt081.100 */
1742 return SQLITE_CORRUPT_PAGE(pPage
);
1744 assert( iFreeBlk
>iPtr
|| iFreeBlk
==0 );
1747 ** iFreeBlk: First freeblock after iStart, or zero if none
1748 ** iPtr: The address of a pointer to iFreeBlk
1750 ** Check to see if iFreeBlk should be coalesced onto the end of iStart.
1752 if( iFreeBlk
&& iEnd
+3>=iFreeBlk
){
1753 nFrag
= iFreeBlk
- iEnd
;
1754 if( iEnd
>iFreeBlk
) return SQLITE_CORRUPT_PAGE(pPage
);
1755 iEnd
= iFreeBlk
+ get2byte(&data
[iFreeBlk
+2]);
1756 if( iEnd
> pPage
->pBt
->usableSize
){
1757 return SQLITE_CORRUPT_PAGE(pPage
);
1759 iSize
= iEnd
- iStart
;
1760 iFreeBlk
= get2byte(&data
[iFreeBlk
]);
1763 /* If iPtr is another freeblock (that is, if iPtr is not the freelist
1764 ** pointer in the page header) then check to see if iStart should be
1765 ** coalesced onto the end of iPtr.
1768 int iPtrEnd
= iPtr
+ get2byte(&data
[iPtr
+2]);
1769 if( iPtrEnd
+3>=iStart
){
1770 if( iPtrEnd
>iStart
) return SQLITE_CORRUPT_PAGE(pPage
);
1771 nFrag
+= iStart
- iPtrEnd
;
1772 iSize
= iEnd
- iPtr
;
1776 if( nFrag
>data
[hdr
+7] ) return SQLITE_CORRUPT_PAGE(pPage
);
1777 data
[hdr
+7] -= nFrag
;
1779 x
= get2byte(&data
[hdr
+5]);
1781 /* The new freeblock is at the beginning of the cell content area,
1782 ** so just extend the cell content area rather than create another
1783 ** freelist entry */
1784 if( iStart
<x
) return SQLITE_CORRUPT_PAGE(pPage
);
1785 if( iPtr
!=hdr
+1 ) return SQLITE_CORRUPT_PAGE(pPage
);
1786 put2byte(&data
[hdr
+1], iFreeBlk
);
1787 put2byte(&data
[hdr
+5], iEnd
);
1789 /* Insert the new freeblock into the freelist */
1790 put2byte(&data
[iPtr
], iStart
);
1792 if( pPage
->pBt
->btsFlags
& BTS_FAST_SECURE
){
1793 /* Overwrite deleted information with zeros when the secure_delete
1794 ** option is enabled */
1795 memset(&data
[iStart
], 0, iSize
);
1797 put2byte(&data
[iStart
], iFreeBlk
);
1798 put2byte(&data
[iStart
+2], iSize
);
1799 pPage
->nFree
+= iOrigSize
;
1804 ** Decode the flags byte (the first byte of the header) for a page
1805 ** and initialize fields of the MemPage structure accordingly.
1807 ** Only the following combinations are supported. Anything different
1808 ** indicates a corrupt database files:
1811 ** PTF_ZERODATA | PTF_LEAF
1812 ** PTF_LEAFDATA | PTF_INTKEY
1813 ** PTF_LEAFDATA | PTF_INTKEY | PTF_LEAF
1815 static int decodeFlags(MemPage
*pPage
, int flagByte
){
1816 BtShared
*pBt
; /* A copy of pPage->pBt */
1818 assert( pPage
->hdrOffset
==(pPage
->pgno
==1 ? 100 : 0) );
1819 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1820 pPage
->leaf
= (u8
)(flagByte
>>3); assert( PTF_LEAF
== 1<<3 );
1821 flagByte
&= ~PTF_LEAF
;
1822 pPage
->childPtrSize
= 4-4*pPage
->leaf
;
1823 pPage
->xCellSize
= cellSizePtr
;
1825 if( flagByte
==(PTF_LEAFDATA
| PTF_INTKEY
) ){
1826 /* EVIDENCE-OF: R-07291-35328 A value of 5 (0x05) means the page is an
1827 ** interior table b-tree page. */
1828 assert( (PTF_LEAFDATA
|PTF_INTKEY
)==5 );
1829 /* EVIDENCE-OF: R-26900-09176 A value of 13 (0x0d) means the page is a
1830 ** leaf table b-tree page. */
1831 assert( (PTF_LEAFDATA
|PTF_INTKEY
|PTF_LEAF
)==13 );
1834 pPage
->intKeyLeaf
= 1;
1835 pPage
->xParseCell
= btreeParseCellPtr
;
1837 pPage
->intKeyLeaf
= 0;
1838 pPage
->xCellSize
= cellSizePtrNoPayload
;
1839 pPage
->xParseCell
= btreeParseCellPtrNoPayload
;
1841 pPage
->maxLocal
= pBt
->maxLeaf
;
1842 pPage
->minLocal
= pBt
->minLeaf
;
1843 }else if( flagByte
==PTF_ZERODATA
){
1844 /* EVIDENCE-OF: R-43316-37308 A value of 2 (0x02) means the page is an
1845 ** interior index b-tree page. */
1846 assert( (PTF_ZERODATA
)==2 );
1847 /* EVIDENCE-OF: R-59615-42828 A value of 10 (0x0a) means the page is a
1848 ** leaf index b-tree page. */
1849 assert( (PTF_ZERODATA
|PTF_LEAF
)==10 );
1851 pPage
->intKeyLeaf
= 0;
1852 pPage
->xParseCell
= btreeParseCellPtrIndex
;
1853 pPage
->maxLocal
= pBt
->maxLocal
;
1854 pPage
->minLocal
= pBt
->minLocal
;
1856 /* EVIDENCE-OF: R-47608-56469 Any other value for the b-tree page type is
1858 return SQLITE_CORRUPT_PAGE(pPage
);
1860 pPage
->max1bytePayload
= pBt
->max1bytePayload
;
1865 ** Compute the amount of freespace on the page. In other words, fill
1866 ** in the pPage->nFree field.
1868 static int btreeComputeFreeSpace(MemPage
*pPage
){
1869 int pc
; /* Address of a freeblock within pPage->aData[] */
1870 u8 hdr
; /* Offset to beginning of page header */
1871 u8
*data
; /* Equal to pPage->aData */
1872 int usableSize
; /* Amount of usable space on each page */
1873 int nFree
; /* Number of unused bytes on the page */
1874 int top
; /* First byte of the cell content area */
1875 int iCellFirst
; /* First allowable cell or freeblock offset */
1876 int iCellLast
; /* Last possible cell or freeblock offset */
1878 assert( pPage
->pBt
!=0 );
1879 assert( pPage
->pBt
->db
!=0 );
1880 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1881 assert( pPage
->pgno
==sqlite3PagerPagenumber(pPage
->pDbPage
) );
1882 assert( pPage
== sqlite3PagerGetExtra(pPage
->pDbPage
) );
1883 assert( pPage
->aData
== sqlite3PagerGetData(pPage
->pDbPage
) );
1884 assert( pPage
->isInit
==1 );
1885 assert( pPage
->nFree
<0 );
1887 usableSize
= pPage
->pBt
->usableSize
;
1888 hdr
= pPage
->hdrOffset
;
1889 data
= pPage
->aData
;
1890 /* EVIDENCE-OF: R-58015-48175 The two-byte integer at offset 5 designates
1891 ** the start of the cell content area. A zero value for this integer is
1892 ** interpreted as 65536. */
1893 top
= get2byteNotZero(&data
[hdr
+5]);
1894 iCellFirst
= hdr
+ 8 + pPage
->childPtrSize
+ 2*pPage
->nCell
;
1895 iCellLast
= usableSize
- 4;
1897 /* Compute the total free space on the page
1898 ** EVIDENCE-OF: R-23588-34450 The two-byte integer at offset 1 gives the
1899 ** start of the first freeblock on the page, or is zero if there are no
1901 pc
= get2byte(&data
[hdr
+1]);
1902 nFree
= data
[hdr
+7] + top
; /* Init nFree to non-freeblock free space */
1906 /* EVIDENCE-OF: R-55530-52930 In a well-formed b-tree page, there will
1907 ** always be at least one cell before the first freeblock.
1909 return SQLITE_CORRUPT_PAGE(pPage
);
1913 /* Freeblock off the end of the page */
1914 return SQLITE_CORRUPT_PAGE(pPage
);
1916 next
= get2byte(&data
[pc
]);
1917 size
= get2byte(&data
[pc
+2]);
1918 nFree
= nFree
+ size
;
1919 if( next
<=pc
+size
+3 ) break;
1923 /* Freeblock not in ascending order */
1924 return SQLITE_CORRUPT_PAGE(pPage
);
1926 if( pc
+size
>(unsigned int)usableSize
){
1927 /* Last freeblock extends past page end */
1928 return SQLITE_CORRUPT_PAGE(pPage
);
1932 /* At this point, nFree contains the sum of the offset to the start
1933 ** of the cell-content area plus the number of free bytes within
1934 ** the cell-content area. If this is greater than the usable-size
1935 ** of the page, then the page must be corrupted. This check also
1936 ** serves to verify that the offset to the start of the cell-content
1937 ** area, according to the page header, lies within the page.
1939 if( nFree
>usableSize
|| nFree
<iCellFirst
){
1940 return SQLITE_CORRUPT_PAGE(pPage
);
1942 pPage
->nFree
= (u16
)(nFree
- iCellFirst
);
1947 ** Do additional sanity check after btreeInitPage() if
1948 ** PRAGMA cell_size_check=ON
1950 static SQLITE_NOINLINE
int btreeCellSizeCheck(MemPage
*pPage
){
1951 int iCellFirst
; /* First allowable cell or freeblock offset */
1952 int iCellLast
; /* Last possible cell or freeblock offset */
1953 int i
; /* Index into the cell pointer array */
1954 int sz
; /* Size of a cell */
1955 int pc
; /* Address of a freeblock within pPage->aData[] */
1956 u8
*data
; /* Equal to pPage->aData */
1957 int usableSize
; /* Maximum usable space on the page */
1958 int cellOffset
; /* Start of cell content area */
1960 iCellFirst
= pPage
->cellOffset
+ 2*pPage
->nCell
;
1961 usableSize
= pPage
->pBt
->usableSize
;
1962 iCellLast
= usableSize
- 4;
1963 data
= pPage
->aData
;
1964 cellOffset
= pPage
->cellOffset
;
1965 if( !pPage
->leaf
) iCellLast
--;
1966 for(i
=0; i
<pPage
->nCell
; i
++){
1967 pc
= get2byteAligned(&data
[cellOffset
+i
*2]);
1968 testcase( pc
==iCellFirst
);
1969 testcase( pc
==iCellLast
);
1970 if( pc
<iCellFirst
|| pc
>iCellLast
){
1971 return SQLITE_CORRUPT_PAGE(pPage
);
1973 sz
= pPage
->xCellSize(pPage
, &data
[pc
]);
1974 testcase( pc
+sz
==usableSize
);
1975 if( pc
+sz
>usableSize
){
1976 return SQLITE_CORRUPT_PAGE(pPage
);
1983 ** Initialize the auxiliary information for a disk block.
1985 ** Return SQLITE_OK on success. If we see that the page does
1986 ** not contain a well-formed database page, then return
1987 ** SQLITE_CORRUPT. Note that a return of SQLITE_OK does not
1988 ** guarantee that the page is well-formed. It only shows that
1989 ** we failed to detect any corruption.
1991 static int btreeInitPage(MemPage
*pPage
){
1992 u8
*data
; /* Equal to pPage->aData */
1993 BtShared
*pBt
; /* The main btree structure */
1995 assert( pPage
->pBt
!=0 );
1996 assert( pPage
->pBt
->db
!=0 );
1997 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
1998 assert( pPage
->pgno
==sqlite3PagerPagenumber(pPage
->pDbPage
) );
1999 assert( pPage
== sqlite3PagerGetExtra(pPage
->pDbPage
) );
2000 assert( pPage
->aData
== sqlite3PagerGetData(pPage
->pDbPage
) );
2001 assert( pPage
->isInit
==0 );
2004 data
= pPage
->aData
+ pPage
->hdrOffset
;
2005 /* EVIDENCE-OF: R-28594-02890 The one-byte flag at offset 0 indicating
2006 ** the b-tree page type. */
2007 if( decodeFlags(pPage
, data
[0]) ){
2008 return SQLITE_CORRUPT_PAGE(pPage
);
2010 assert( pBt
->pageSize
>=512 && pBt
->pageSize
<=65536 );
2011 pPage
->maskPage
= (u16
)(pBt
->pageSize
- 1);
2012 pPage
->nOverflow
= 0;
2013 pPage
->cellOffset
= pPage
->hdrOffset
+ 8 + pPage
->childPtrSize
;
2014 pPage
->aCellIdx
= data
+ pPage
->childPtrSize
+ 8;
2015 pPage
->aDataEnd
= pPage
->aData
+ pBt
->usableSize
;
2016 pPage
->aDataOfst
= pPage
->aData
+ pPage
->childPtrSize
;
2017 /* EVIDENCE-OF: R-37002-32774 The two-byte integer at offset 3 gives the
2018 ** number of cells on the page. */
2019 pPage
->nCell
= get2byte(&data
[3]);
2020 if( pPage
->nCell
>MX_CELL(pBt
) ){
2021 /* To many cells for a single page. The page must be corrupt */
2022 return SQLITE_CORRUPT_PAGE(pPage
);
2024 testcase( pPage
->nCell
==MX_CELL(pBt
) );
2025 /* EVIDENCE-OF: R-24089-57979 If a page contains no cells (which is only
2026 ** possible for a root page of a table that contains no rows) then the
2027 ** offset to the cell content area will equal the page size minus the
2028 ** bytes of reserved space. */
2029 assert( pPage
->nCell
>0
2030 || get2byteNotZero(&data
[5])==(int)pBt
->usableSize
2032 pPage
->nFree
= -1; /* Indicate that this value is yet uncomputed */
2034 if( pBt
->db
->flags
& SQLITE_CellSizeCk
){
2035 return btreeCellSizeCheck(pPage
);
2041 ** Set up a raw page so that it looks like a database page holding
2044 static void zeroPage(MemPage
*pPage
, int flags
){
2045 unsigned char *data
= pPage
->aData
;
2046 BtShared
*pBt
= pPage
->pBt
;
2047 u8 hdr
= pPage
->hdrOffset
;
2050 assert( sqlite3PagerPagenumber(pPage
->pDbPage
)==pPage
->pgno
);
2051 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
2052 assert( sqlite3PagerGetData(pPage
->pDbPage
) == data
);
2053 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
2054 assert( sqlite3_mutex_held(pBt
->mutex
) );
2055 if( pBt
->btsFlags
& BTS_FAST_SECURE
){
2056 memset(&data
[hdr
], 0, pBt
->usableSize
- hdr
);
2058 data
[hdr
] = (char)flags
;
2059 first
= hdr
+ ((flags
&PTF_LEAF
)==0 ? 12 : 8);
2060 memset(&data
[hdr
+1], 0, 4);
2062 put2byte(&data
[hdr
+5], pBt
->usableSize
);
2063 pPage
->nFree
= (u16
)(pBt
->usableSize
- first
);
2064 decodeFlags(pPage
, flags
);
2065 pPage
->cellOffset
= first
;
2066 pPage
->aDataEnd
= &data
[pBt
->usableSize
];
2067 pPage
->aCellIdx
= &data
[first
];
2068 pPage
->aDataOfst
= &data
[pPage
->childPtrSize
];
2069 pPage
->nOverflow
= 0;
2070 assert( pBt
->pageSize
>=512 && pBt
->pageSize
<=65536 );
2071 pPage
->maskPage
= (u16
)(pBt
->pageSize
- 1);
2078 ** Convert a DbPage obtained from the pager into a MemPage used by
2081 static MemPage
*btreePageFromDbPage(DbPage
*pDbPage
, Pgno pgno
, BtShared
*pBt
){
2082 MemPage
*pPage
= (MemPage
*)sqlite3PagerGetExtra(pDbPage
);
2083 if( pgno
!=pPage
->pgno
){
2084 pPage
->aData
= sqlite3PagerGetData(pDbPage
);
2085 pPage
->pDbPage
= pDbPage
;
2088 pPage
->hdrOffset
= pgno
==1 ? 100 : 0;
2090 assert( pPage
->aData
==sqlite3PagerGetData(pDbPage
) );
2095 ** Get a page from the pager. Initialize the MemPage.pBt and
2096 ** MemPage.aData elements if needed. See also: btreeGetUnusedPage().
2098 ** If the PAGER_GET_NOCONTENT flag is set, it means that we do not care
2099 ** about the content of the page at this time. So do not go to the disk
2100 ** to fetch the content. Just fill in the content with zeros for now.
2101 ** If in the future we call sqlite3PagerWrite() on this page, that
2102 ** means we have started to be concerned about content and the disk
2103 ** read should occur at that point.
2105 static int btreeGetPage(
2106 BtShared
*pBt
, /* The btree */
2107 Pgno pgno
, /* Number of the page to fetch */
2108 MemPage
**ppPage
, /* Return the page in this parameter */
2109 int flags
/* PAGER_GET_NOCONTENT or PAGER_GET_READONLY */
2114 assert( flags
==0 || flags
==PAGER_GET_NOCONTENT
|| flags
==PAGER_GET_READONLY
);
2115 assert( sqlite3_mutex_held(pBt
->mutex
) );
2116 rc
= sqlite3PagerGet(pBt
->pPager
, pgno
, (DbPage
**)&pDbPage
, flags
);
2118 *ppPage
= btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2123 ** Retrieve a page from the pager cache. If the requested page is not
2124 ** already in the pager cache return NULL. Initialize the MemPage.pBt and
2125 ** MemPage.aData elements if needed.
2127 static MemPage
*btreePageLookup(BtShared
*pBt
, Pgno pgno
){
2129 assert( sqlite3_mutex_held(pBt
->mutex
) );
2130 pDbPage
= sqlite3PagerLookup(pBt
->pPager
, pgno
);
2132 return btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2138 ** Return the size of the database file in pages. If there is any kind of
2139 ** error, return ((unsigned int)-1).
2141 static Pgno
btreePagecount(BtShared
*pBt
){
2144 Pgno
sqlite3BtreeLastPage(Btree
*p
){
2145 assert( sqlite3BtreeHoldsMutex(p
) );
2146 return btreePagecount(p
->pBt
);
2150 ** Get a page from the pager and initialize it.
2152 ** If pCur!=0 then the page is being fetched as part of a moveToChild()
2153 ** call. Do additional sanity checking on the page in this case.
2154 ** And if the fetch fails, this routine must decrement pCur->iPage.
2156 ** The page is fetched as read-write unless pCur is not NULL and is
2157 ** a read-only cursor.
2159 ** If an error occurs, then *ppPage is undefined. It
2160 ** may remain unchanged, or it may be set to an invalid value.
2162 static int getAndInitPage(
2163 BtShared
*pBt
, /* The database file */
2164 Pgno pgno
, /* Number of the page to get */
2165 MemPage
**ppPage
, /* Write the page pointer here */
2166 BtCursor
*pCur
, /* Cursor to receive the page, or NULL */
2167 int bReadOnly
/* True for a read-only page */
2171 assert( sqlite3_mutex_held(pBt
->mutex
) );
2172 assert( pCur
==0 || ppPage
==&pCur
->pPage
);
2173 assert( pCur
==0 || bReadOnly
==pCur
->curPagerFlags
);
2174 assert( pCur
==0 || pCur
->iPage
>0 );
2176 if( pgno
>btreePagecount(pBt
) ){
2177 rc
= SQLITE_CORRUPT_BKPT
;
2178 goto getAndInitPage_error1
;
2180 rc
= sqlite3PagerGet(pBt
->pPager
, pgno
, (DbPage
**)&pDbPage
, bReadOnly
);
2182 goto getAndInitPage_error1
;
2184 *ppPage
= (MemPage
*)sqlite3PagerGetExtra(pDbPage
);
2185 if( (*ppPage
)->isInit
==0 ){
2186 btreePageFromDbPage(pDbPage
, pgno
, pBt
);
2187 rc
= btreeInitPage(*ppPage
);
2188 if( rc
!=SQLITE_OK
){
2189 goto getAndInitPage_error2
;
2192 assert( (*ppPage
)->pgno
==pgno
);
2193 assert( (*ppPage
)->aData
==sqlite3PagerGetData(pDbPage
) );
2195 /* If obtaining a child page for a cursor, we must verify that the page is
2196 ** compatible with the root page. */
2197 if( pCur
&& ((*ppPage
)->nCell
<1 || (*ppPage
)->intKey
!=pCur
->curIntKey
) ){
2198 rc
= SQLITE_CORRUPT_PGNO(pgno
);
2199 goto getAndInitPage_error2
;
2203 getAndInitPage_error2
:
2204 releasePage(*ppPage
);
2205 getAndInitPage_error1
:
2208 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
2210 testcase( pgno
==0 );
2211 assert( pgno
!=0 || rc
==SQLITE_CORRUPT
);
2216 ** Release a MemPage. This should be called once for each prior
2217 ** call to btreeGetPage.
2219 ** Page1 is a special case and must be released using releasePageOne().
2221 static void releasePageNotNull(MemPage
*pPage
){
2222 assert( pPage
->aData
);
2223 assert( pPage
->pBt
);
2224 assert( pPage
->pDbPage
!=0 );
2225 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
2226 assert( sqlite3PagerGetData(pPage
->pDbPage
)==pPage
->aData
);
2227 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2228 sqlite3PagerUnrefNotNull(pPage
->pDbPage
);
2230 static void releasePage(MemPage
*pPage
){
2231 if( pPage
) releasePageNotNull(pPage
);
2233 static void releasePageOne(MemPage
*pPage
){
2235 assert( pPage
->aData
);
2236 assert( pPage
->pBt
);
2237 assert( pPage
->pDbPage
!=0 );
2238 assert( sqlite3PagerGetExtra(pPage
->pDbPage
) == (void*)pPage
);
2239 assert( sqlite3PagerGetData(pPage
->pDbPage
)==pPage
->aData
);
2240 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2241 sqlite3PagerUnrefPageOne(pPage
->pDbPage
);
2245 ** Get an unused page.
2247 ** This works just like btreeGetPage() with the addition:
2249 ** * If the page is already in use for some other purpose, immediately
2250 ** release it and return an SQLITE_CURRUPT error.
2251 ** * Make sure the isInit flag is clear
2253 static int btreeGetUnusedPage(
2254 BtShared
*pBt
, /* The btree */
2255 Pgno pgno
, /* Number of the page to fetch */
2256 MemPage
**ppPage
, /* Return the page in this parameter */
2257 int flags
/* PAGER_GET_NOCONTENT or PAGER_GET_READONLY */
2259 int rc
= btreeGetPage(pBt
, pgno
, ppPage
, flags
);
2260 if( rc
==SQLITE_OK
){
2261 if( sqlite3PagerPageRefcount((*ppPage
)->pDbPage
)>1 ){
2262 releasePage(*ppPage
);
2264 return SQLITE_CORRUPT_BKPT
;
2266 (*ppPage
)->isInit
= 0;
2275 ** During a rollback, when the pager reloads information into the cache
2276 ** so that the cache is restored to its original state at the start of
2277 ** the transaction, for each page restored this routine is called.
2279 ** This routine needs to reset the extra data section at the end of the
2280 ** page to agree with the restored data.
2282 static void pageReinit(DbPage
*pData
){
2284 pPage
= (MemPage
*)sqlite3PagerGetExtra(pData
);
2285 assert( sqlite3PagerPageRefcount(pData
)>0 );
2286 if( pPage
->isInit
){
2287 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
2289 if( sqlite3PagerPageRefcount(pData
)>1 ){
2290 /* pPage might not be a btree page; it might be an overflow page
2291 ** or ptrmap page or a free page. In those cases, the following
2292 ** call to btreeInitPage() will likely return SQLITE_CORRUPT.
2293 ** But no harm is done by this. And it is very important that
2294 ** btreeInitPage() be called on every btree page so we make
2295 ** the call for every page that comes in for re-initing. */
2296 btreeInitPage(pPage
);
2302 ** Invoke the busy handler for a btree.
2304 static int btreeInvokeBusyHandler(void *pArg
){
2305 BtShared
*pBt
= (BtShared
*)pArg
;
2307 assert( sqlite3_mutex_held(pBt
->db
->mutex
) );
2308 return sqlite3InvokeBusyHandler(&pBt
->db
->busyHandler
);
2312 ** Open a database file.
2314 ** zFilename is the name of the database file. If zFilename is NULL
2315 ** then an ephemeral database is created. The ephemeral database might
2316 ** be exclusively in memory, or it might use a disk-based memory cache.
2317 ** Either way, the ephemeral database will be automatically deleted
2318 ** when sqlite3BtreeClose() is called.
2320 ** If zFilename is ":memory:" then an in-memory database is created
2321 ** that is automatically destroyed when it is closed.
2323 ** The "flags" parameter is a bitmask that might contain bits like
2324 ** BTREE_OMIT_JOURNAL and/or BTREE_MEMORY.
2326 ** If the database is already opened in the same database connection
2327 ** and we are in shared cache mode, then the open will fail with an
2328 ** SQLITE_CONSTRAINT error. We cannot allow two or more BtShared
2329 ** objects in the same database connection since doing so will lead
2330 ** to problems with locking.
2332 int sqlite3BtreeOpen(
2333 sqlite3_vfs
*pVfs
, /* VFS to use for this b-tree */
2334 const char *zFilename
, /* Name of the file containing the BTree database */
2335 sqlite3
*db
, /* Associated database handle */
2336 Btree
**ppBtree
, /* Pointer to new Btree object written here */
2337 int flags
, /* Options */
2338 int vfsFlags
/* Flags passed through to sqlite3_vfs.xOpen() */
2340 BtShared
*pBt
= 0; /* Shared part of btree structure */
2341 Btree
*p
; /* Handle to return */
2342 sqlite3_mutex
*mutexOpen
= 0; /* Prevents a race condition. Ticket #3537 */
2343 int rc
= SQLITE_OK
; /* Result code from this function */
2344 u8 nReserve
; /* Byte of unused space on each page */
2345 unsigned char zDbHeader
[100]; /* Database header content */
2347 /* True if opening an ephemeral, temporary database */
2348 const int isTempDb
= zFilename
==0 || zFilename
[0]==0;
2350 /* Set the variable isMemdb to true for an in-memory database, or
2351 ** false for a file-based database.
2353 #ifdef SQLITE_OMIT_MEMORYDB
2354 const int isMemdb
= 0;
2356 const int isMemdb
= (zFilename
&& strcmp(zFilename
, ":memory:")==0)
2357 || (isTempDb
&& sqlite3TempInMemory(db
))
2358 || (vfsFlags
& SQLITE_OPEN_MEMORY
)!=0;
2363 assert( sqlite3_mutex_held(db
->mutex
) );
2364 assert( (flags
&0xff)==flags
); /* flags fit in 8 bits */
2366 /* Only a BTREE_SINGLE database can be BTREE_UNORDERED */
2367 assert( (flags
& BTREE_UNORDERED
)==0 || (flags
& BTREE_SINGLE
)!=0 );
2369 /* A BTREE_SINGLE database is always a temporary and/or ephemeral */
2370 assert( (flags
& BTREE_SINGLE
)==0 || isTempDb
);
2373 flags
|= BTREE_MEMORY
;
2375 if( (vfsFlags
& SQLITE_OPEN_MAIN_DB
)!=0 && (isMemdb
|| isTempDb
) ){
2376 vfsFlags
= (vfsFlags
& ~SQLITE_OPEN_MAIN_DB
) | SQLITE_OPEN_TEMP_DB
;
2378 p
= sqlite3MallocZero(sizeof(Btree
));
2380 return SQLITE_NOMEM_BKPT
;
2382 p
->inTrans
= TRANS_NONE
;
2384 #ifndef SQLITE_OMIT_SHARED_CACHE
2389 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2391 ** If this Btree is a candidate for shared cache, try to find an
2392 ** existing BtShared object that we can share with
2394 if( isTempDb
==0 && (isMemdb
==0 || (vfsFlags
&SQLITE_OPEN_URI
)!=0) ){
2395 if( vfsFlags
& SQLITE_OPEN_SHAREDCACHE
){
2396 int nFilename
= sqlite3Strlen30(zFilename
)+1;
2397 int nFullPathname
= pVfs
->mxPathname
+1;
2398 char *zFullPathname
= sqlite3Malloc(MAX(nFullPathname
,nFilename
));
2399 MUTEX_LOGIC( sqlite3_mutex
*mutexShared
; )
2402 if( !zFullPathname
){
2404 return SQLITE_NOMEM_BKPT
;
2407 memcpy(zFullPathname
, zFilename
, nFilename
);
2409 rc
= sqlite3OsFullPathname(pVfs
, zFilename
,
2410 nFullPathname
, zFullPathname
);
2412 if( rc
==SQLITE_OK_SYMLINK
){
2415 sqlite3_free(zFullPathname
);
2421 #if SQLITE_THREADSAFE
2422 mutexOpen
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_OPEN
);
2423 sqlite3_mutex_enter(mutexOpen
);
2424 mutexShared
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MAIN
);
2425 sqlite3_mutex_enter(mutexShared
);
2427 for(pBt
=GLOBAL(BtShared
*,sqlite3SharedCacheList
); pBt
; pBt
=pBt
->pNext
){
2428 assert( pBt
->nRef
>0 );
2429 if( 0==strcmp(zFullPathname
, sqlite3PagerFilename(pBt
->pPager
, 0))
2430 && sqlite3PagerVfs(pBt
->pPager
)==pVfs
){
2432 for(iDb
=db
->nDb
-1; iDb
>=0; iDb
--){
2433 Btree
*pExisting
= db
->aDb
[iDb
].pBt
;
2434 if( pExisting
&& pExisting
->pBt
==pBt
){
2435 sqlite3_mutex_leave(mutexShared
);
2436 sqlite3_mutex_leave(mutexOpen
);
2437 sqlite3_free(zFullPathname
);
2439 return SQLITE_CONSTRAINT
;
2447 sqlite3_mutex_leave(mutexShared
);
2448 sqlite3_free(zFullPathname
);
2452 /* In debug mode, we mark all persistent databases as sharable
2453 ** even when they are not. This exercises the locking code and
2454 ** gives more opportunity for asserts(sqlite3_mutex_held())
2455 ** statements to find locking problems.
2464 ** The following asserts make sure that structures used by the btree are
2465 ** the right size. This is to guard against size changes that result
2466 ** when compiling on a different architecture.
2468 assert( sizeof(i64
)==8 );
2469 assert( sizeof(u64
)==8 );
2470 assert( sizeof(u32
)==4 );
2471 assert( sizeof(u16
)==2 );
2472 assert( sizeof(Pgno
)==4 );
2474 pBt
= sqlite3MallocZero( sizeof(*pBt
) );
2476 rc
= SQLITE_NOMEM_BKPT
;
2477 goto btree_open_out
;
2479 rc
= sqlite3PagerOpen(pVfs
, &pBt
->pPager
, zFilename
,
2480 sizeof(MemPage
), flags
, vfsFlags
, pageReinit
);
2481 if( rc
==SQLITE_OK
){
2482 sqlite3PagerSetMmapLimit(pBt
->pPager
, db
->szMmap
);
2483 rc
= sqlite3PagerReadFileheader(pBt
->pPager
,sizeof(zDbHeader
),zDbHeader
);
2485 if( rc
!=SQLITE_OK
){
2486 goto btree_open_out
;
2488 pBt
->openFlags
= (u8
)flags
;
2490 sqlite3PagerSetBusyHandler(pBt
->pPager
, btreeInvokeBusyHandler
, pBt
);
2495 if( sqlite3PagerIsreadonly(pBt
->pPager
) ) pBt
->btsFlags
|= BTS_READ_ONLY
;
2496 #if defined(SQLITE_SECURE_DELETE)
2497 pBt
->btsFlags
|= BTS_SECURE_DELETE
;
2498 #elif defined(SQLITE_FAST_SECURE_DELETE)
2499 pBt
->btsFlags
|= BTS_OVERWRITE
;
2501 /* EVIDENCE-OF: R-51873-39618 The page size for a database file is
2502 ** determined by the 2-byte integer located at an offset of 16 bytes from
2503 ** the beginning of the database file. */
2504 pBt
->pageSize
= (zDbHeader
[16]<<8) | (zDbHeader
[17]<<16);
2505 if( pBt
->pageSize
<512 || pBt
->pageSize
>SQLITE_MAX_PAGE_SIZE
2506 || ((pBt
->pageSize
-1)&pBt
->pageSize
)!=0 ){
2508 #ifndef SQLITE_OMIT_AUTOVACUUM
2509 /* If the magic name ":memory:" will create an in-memory database, then
2510 ** leave the autoVacuum mode at 0 (do not auto-vacuum), even if
2511 ** SQLITE_DEFAULT_AUTOVACUUM is true. On the other hand, if
2512 ** SQLITE_OMIT_MEMORYDB has been defined, then ":memory:" is just a
2513 ** regular file-name. In this case the auto-vacuum applies as per normal.
2515 if( zFilename
&& !isMemdb
){
2516 pBt
->autoVacuum
= (SQLITE_DEFAULT_AUTOVACUUM
? 1 : 0);
2517 pBt
->incrVacuum
= (SQLITE_DEFAULT_AUTOVACUUM
==2 ? 1 : 0);
2522 /* EVIDENCE-OF: R-37497-42412 The size of the reserved region is
2523 ** determined by the one-byte unsigned integer found at an offset of 20
2524 ** into the database file header. */
2525 nReserve
= zDbHeader
[20];
2526 pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
2527 #ifndef SQLITE_OMIT_AUTOVACUUM
2528 pBt
->autoVacuum
= (get4byte(&zDbHeader
[36 + 4*4])?1:0);
2529 pBt
->incrVacuum
= (get4byte(&zDbHeader
[36 + 7*4])?1:0);
2532 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
, nReserve
);
2533 if( rc
) goto btree_open_out
;
2534 pBt
->usableSize
= pBt
->pageSize
- nReserve
;
2535 assert( (pBt
->pageSize
& 7)==0 ); /* 8-byte alignment of pageSize */
2537 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2538 /* Add the new BtShared object to the linked list sharable BtShareds.
2542 MUTEX_LOGIC( sqlite3_mutex
*mutexShared
; )
2543 MUTEX_LOGIC( mutexShared
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MAIN
);)
2544 if( SQLITE_THREADSAFE
&& sqlite3GlobalConfig
.bCoreMutex
){
2545 pBt
->mutex
= sqlite3MutexAlloc(SQLITE_MUTEX_FAST
);
2546 if( pBt
->mutex
==0 ){
2547 rc
= SQLITE_NOMEM_BKPT
;
2548 goto btree_open_out
;
2551 sqlite3_mutex_enter(mutexShared
);
2552 pBt
->pNext
= GLOBAL(BtShared
*,sqlite3SharedCacheList
);
2553 GLOBAL(BtShared
*,sqlite3SharedCacheList
) = pBt
;
2554 sqlite3_mutex_leave(mutexShared
);
2559 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO)
2560 /* If the new Btree uses a sharable pBtShared, then link the new
2561 ** Btree into the list of all sharable Btrees for the same connection.
2562 ** The list is kept in ascending order by pBt address.
2567 for(i
=0; i
<db
->nDb
; i
++){
2568 if( (pSib
= db
->aDb
[i
].pBt
)!=0 && pSib
->sharable
){
2569 while( pSib
->pPrev
){ pSib
= pSib
->pPrev
; }
2570 if( (uptr
)p
->pBt
<(uptr
)pSib
->pBt
){
2575 while( pSib
->pNext
&& (uptr
)pSib
->pNext
->pBt
<(uptr
)p
->pBt
){
2578 p
->pNext
= pSib
->pNext
;
2581 p
->pNext
->pPrev
= p
;
2593 if( rc
!=SQLITE_OK
){
2594 if( pBt
&& pBt
->pPager
){
2595 sqlite3PagerClose(pBt
->pPager
, 0);
2601 sqlite3_file
*pFile
;
2603 /* If the B-Tree was successfully opened, set the pager-cache size to the
2604 ** default value. Except, when opening on an existing shared pager-cache,
2605 ** do not change the pager-cache size.
2607 if( sqlite3BtreeSchema(p
, 0, 0)==0 ){
2608 sqlite3PagerSetCachesize(p
->pBt
->pPager
, SQLITE_DEFAULT_CACHE_SIZE
);
2611 pFile
= sqlite3PagerFile(pBt
->pPager
);
2612 if( pFile
->pMethods
){
2613 sqlite3OsFileControlHint(pFile
, SQLITE_FCNTL_PDB
, (void*)&pBt
->db
);
2617 assert( sqlite3_mutex_held(mutexOpen
) );
2618 sqlite3_mutex_leave(mutexOpen
);
2620 assert( rc
!=SQLITE_OK
|| sqlite3BtreeConnectionCount(*ppBtree
)>0 );
2625 ** Decrement the BtShared.nRef counter. When it reaches zero,
2626 ** remove the BtShared structure from the sharing list. Return
2627 ** true if the BtShared.nRef counter reaches zero and return
2628 ** false if it is still positive.
2630 static int removeFromSharingList(BtShared
*pBt
){
2631 #ifndef SQLITE_OMIT_SHARED_CACHE
2632 MUTEX_LOGIC( sqlite3_mutex
*pMainMtx
; )
2636 assert( sqlite3_mutex_notheld(pBt
->mutex
) );
2637 MUTEX_LOGIC( pMainMtx
= sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MAIN
); )
2638 sqlite3_mutex_enter(pMainMtx
);
2641 if( GLOBAL(BtShared
*,sqlite3SharedCacheList
)==pBt
){
2642 GLOBAL(BtShared
*,sqlite3SharedCacheList
) = pBt
->pNext
;
2644 pList
= GLOBAL(BtShared
*,sqlite3SharedCacheList
);
2645 while( ALWAYS(pList
) && pList
->pNext
!=pBt
){
2648 if( ALWAYS(pList
) ){
2649 pList
->pNext
= pBt
->pNext
;
2652 if( SQLITE_THREADSAFE
){
2653 sqlite3_mutex_free(pBt
->mutex
);
2657 sqlite3_mutex_leave(pMainMtx
);
2665 ** Make sure pBt->pTmpSpace points to an allocation of
2666 ** MX_CELL_SIZE(pBt) bytes with a 4-byte prefix for a left-child
2669 static void allocateTempSpace(BtShared
*pBt
){
2670 if( !pBt
->pTmpSpace
){
2671 pBt
->pTmpSpace
= sqlite3PageMalloc( pBt
->pageSize
);
2673 /* One of the uses of pBt->pTmpSpace is to format cells before
2674 ** inserting them into a leaf page (function fillInCell()). If
2675 ** a cell is less than 4 bytes in size, it is rounded up to 4 bytes
2676 ** by the various routines that manipulate binary cells. Which
2677 ** can mean that fillInCell() only initializes the first 2 or 3
2678 ** bytes of pTmpSpace, but that the first 4 bytes are copied from
2679 ** it into a database page. This is not actually a problem, but it
2680 ** does cause a valgrind error when the 1 or 2 bytes of unitialized
2681 ** data is passed to system call write(). So to avoid this error,
2682 ** zero the first 4 bytes of temp space here.
2684 ** Also: Provide four bytes of initialized space before the
2685 ** beginning of pTmpSpace as an area available to prepend the
2686 ** left-child pointer to the beginning of a cell.
2688 if( pBt
->pTmpSpace
){
2689 memset(pBt
->pTmpSpace
, 0, 8);
2690 pBt
->pTmpSpace
+= 4;
2696 ** Free the pBt->pTmpSpace allocation
2698 static void freeTempSpace(BtShared
*pBt
){
2699 if( pBt
->pTmpSpace
){
2700 pBt
->pTmpSpace
-= 4;
2701 sqlite3PageFree(pBt
->pTmpSpace
);
2707 ** Close an open database and invalidate all cursors.
2709 int sqlite3BtreeClose(Btree
*p
){
2710 BtShared
*pBt
= p
->pBt
;
2713 /* Close all cursors opened via this handle. */
2714 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2715 sqlite3BtreeEnter(p
);
2716 pCur
= pBt
->pCursor
;
2718 BtCursor
*pTmp
= pCur
;
2720 if( pTmp
->pBtree
==p
){
2721 sqlite3BtreeCloseCursor(pTmp
);
2725 /* Rollback any active transaction and free the handle structure.
2726 ** The call to sqlite3BtreeRollback() drops any table-locks held by
2729 sqlite3BtreeRollback(p
, SQLITE_OK
, 0);
2730 sqlite3BtreeLeave(p
);
2732 /* If there are still other outstanding references to the shared-btree
2733 ** structure, return now. The remainder of this procedure cleans
2734 ** up the shared-btree.
2736 assert( p
->wantToLock
==0 && p
->locked
==0 );
2737 if( !p
->sharable
|| removeFromSharingList(pBt
) ){
2738 /* The pBt is no longer on the sharing list, so we can access
2739 ** it without having to hold the mutex.
2741 ** Clean out and delete the BtShared object.
2743 assert( !pBt
->pCursor
);
2744 sqlite3PagerClose(pBt
->pPager
, p
->db
);
2745 if( pBt
->xFreeSchema
&& pBt
->pSchema
){
2746 pBt
->xFreeSchema(pBt
->pSchema
);
2748 sqlite3DbFree(0, pBt
->pSchema
);
2753 #ifndef SQLITE_OMIT_SHARED_CACHE
2754 assert( p
->wantToLock
==0 );
2755 assert( p
->locked
==0 );
2756 if( p
->pPrev
) p
->pPrev
->pNext
= p
->pNext
;
2757 if( p
->pNext
) p
->pNext
->pPrev
= p
->pPrev
;
2765 ** Change the "soft" limit on the number of pages in the cache.
2766 ** Unused and unmodified pages will be recycled when the number of
2767 ** pages in the cache exceeds this soft limit. But the size of the
2768 ** cache is allowed to grow larger than this limit if it contains
2769 ** dirty pages or pages still in active use.
2771 int sqlite3BtreeSetCacheSize(Btree
*p
, int mxPage
){
2772 BtShared
*pBt
= p
->pBt
;
2773 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2774 sqlite3BtreeEnter(p
);
2775 sqlite3PagerSetCachesize(pBt
->pPager
, mxPage
);
2776 sqlite3BtreeLeave(p
);
2781 ** Change the "spill" limit on the number of pages in the cache.
2782 ** If the number of pages exceeds this limit during a write transaction,
2783 ** the pager might attempt to "spill" pages to the journal early in
2784 ** order to free up memory.
2786 ** The value returned is the current spill size. If zero is passed
2787 ** as an argument, no changes are made to the spill size setting, so
2788 ** using mxPage of 0 is a way to query the current spill size.
2790 int sqlite3BtreeSetSpillSize(Btree
*p
, int mxPage
){
2791 BtShared
*pBt
= p
->pBt
;
2793 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2794 sqlite3BtreeEnter(p
);
2795 res
= sqlite3PagerSetSpillsize(pBt
->pPager
, mxPage
);
2796 sqlite3BtreeLeave(p
);
2800 #if SQLITE_MAX_MMAP_SIZE>0
2802 ** Change the limit on the amount of the database file that may be
2805 int sqlite3BtreeSetMmapLimit(Btree
*p
, sqlite3_int64 szMmap
){
2806 BtShared
*pBt
= p
->pBt
;
2807 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2808 sqlite3BtreeEnter(p
);
2809 sqlite3PagerSetMmapLimit(pBt
->pPager
, szMmap
);
2810 sqlite3BtreeLeave(p
);
2813 #endif /* SQLITE_MAX_MMAP_SIZE>0 */
2816 ** Change the way data is synced to disk in order to increase or decrease
2817 ** how well the database resists damage due to OS crashes and power
2818 ** failures. Level 1 is the same as asynchronous (no syncs() occur and
2819 ** there is a high probability of damage) Level 2 is the default. There
2820 ** is a very low but non-zero probability of damage. Level 3 reduces the
2821 ** probability of damage to near zero but with a write performance reduction.
2823 #ifndef SQLITE_OMIT_PAGER_PRAGMAS
2824 int sqlite3BtreeSetPagerFlags(
2825 Btree
*p
, /* The btree to set the safety level on */
2826 unsigned pgFlags
/* Various PAGER_* flags */
2828 BtShared
*pBt
= p
->pBt
;
2829 assert( sqlite3_mutex_held(p
->db
->mutex
) );
2830 sqlite3BtreeEnter(p
);
2831 sqlite3PagerSetFlags(pBt
->pPager
, pgFlags
);
2832 sqlite3BtreeLeave(p
);
2838 ** Change the default pages size and the number of reserved bytes per page.
2839 ** Or, if the page size has already been fixed, return SQLITE_READONLY
2840 ** without changing anything.
2842 ** The page size must be a power of 2 between 512 and 65536. If the page
2843 ** size supplied does not meet this constraint then the page size is not
2846 ** Page sizes are constrained to be a power of two so that the region
2847 ** of the database file used for locking (beginning at PENDING_BYTE,
2848 ** the first byte past the 1GB boundary, 0x40000000) needs to occur
2849 ** at the beginning of a page.
2851 ** If parameter nReserve is less than zero, then the number of reserved
2852 ** bytes per page is left unchanged.
2854 ** If the iFix!=0 then the BTS_PAGESIZE_FIXED flag is set so that the page size
2855 ** and autovacuum mode can no longer be changed.
2857 int sqlite3BtreeSetPageSize(Btree
*p
, int pageSize
, int nReserve
, int iFix
){
2860 BtShared
*pBt
= p
->pBt
;
2861 assert( nReserve
>=0 && nReserve
<=255 );
2862 sqlite3BtreeEnter(p
);
2863 pBt
->nReserveWanted
= nReserve
;
2864 x
= pBt
->pageSize
- pBt
->usableSize
;
2865 if( nReserve
<x
) nReserve
= x
;
2866 if( pBt
->btsFlags
& BTS_PAGESIZE_FIXED
){
2867 sqlite3BtreeLeave(p
);
2868 return SQLITE_READONLY
;
2870 assert( nReserve
>=0 && nReserve
<=255 );
2871 if( pageSize
>=512 && pageSize
<=SQLITE_MAX_PAGE_SIZE
&&
2872 ((pageSize
-1)&pageSize
)==0 ){
2873 assert( (pageSize
& 7)==0 );
2874 assert( !pBt
->pCursor
);
2875 pBt
->pageSize
= (u32
)pageSize
;
2878 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
, nReserve
);
2879 pBt
->usableSize
= pBt
->pageSize
- (u16
)nReserve
;
2880 if( iFix
) pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
2881 sqlite3BtreeLeave(p
);
2886 ** Return the currently defined page size
2888 int sqlite3BtreeGetPageSize(Btree
*p
){
2889 return p
->pBt
->pageSize
;
2893 ** This function is similar to sqlite3BtreeGetReserve(), except that it
2894 ** may only be called if it is guaranteed that the b-tree mutex is already
2897 ** This is useful in one special case in the backup API code where it is
2898 ** known that the shared b-tree mutex is held, but the mutex on the
2899 ** database handle that owns *p is not. In this case if sqlite3BtreeEnter()
2900 ** were to be called, it might collide with some other operation on the
2901 ** database handle that owns *p, causing undefined behavior.
2903 int sqlite3BtreeGetReserveNoMutex(Btree
*p
){
2905 assert( sqlite3_mutex_held(p
->pBt
->mutex
) );
2906 n
= p
->pBt
->pageSize
- p
->pBt
->usableSize
;
2911 ** Return the number of bytes of space at the end of every page that
2912 ** are intentually left unused. This is the "reserved" space that is
2913 ** sometimes used by extensions.
2915 ** The value returned is the larger of the current reserve size and
2916 ** the latest reserve size requested by SQLITE_FILECTRL_RESERVE_BYTES.
2917 ** The amount of reserve can only grow - never shrink.
2919 int sqlite3BtreeGetRequestedReserve(Btree
*p
){
2921 sqlite3BtreeEnter(p
);
2922 n1
= (int)p
->pBt
->nReserveWanted
;
2923 n2
= sqlite3BtreeGetReserveNoMutex(p
);
2924 sqlite3BtreeLeave(p
);
2925 return n1
>n2
? n1
: n2
;
2930 ** Set the maximum page count for a database if mxPage is positive.
2931 ** No changes are made if mxPage is 0 or negative.
2932 ** Regardless of the value of mxPage, return the maximum page count.
2934 Pgno
sqlite3BtreeMaxPageCount(Btree
*p
, Pgno mxPage
){
2936 sqlite3BtreeEnter(p
);
2937 n
= sqlite3PagerMaxPageCount(p
->pBt
->pPager
, mxPage
);
2938 sqlite3BtreeLeave(p
);
2943 ** Change the values for the BTS_SECURE_DELETE and BTS_OVERWRITE flags:
2945 ** newFlag==0 Both BTS_SECURE_DELETE and BTS_OVERWRITE are cleared
2946 ** newFlag==1 BTS_SECURE_DELETE set and BTS_OVERWRITE is cleared
2947 ** newFlag==2 BTS_SECURE_DELETE cleared and BTS_OVERWRITE is set
2948 ** newFlag==(-1) No changes
2950 ** This routine acts as a query if newFlag is less than zero
2952 ** With BTS_OVERWRITE set, deleted content is overwritten by zeros, but
2953 ** freelist leaf pages are not written back to the database. Thus in-page
2954 ** deleted content is cleared, but freelist deleted content is not.
2956 ** With BTS_SECURE_DELETE, operation is like BTS_OVERWRITE with the addition
2957 ** that freelist leaf pages are written back into the database, increasing
2958 ** the amount of disk I/O.
2960 int sqlite3BtreeSecureDelete(Btree
*p
, int newFlag
){
2962 if( p
==0 ) return 0;
2963 sqlite3BtreeEnter(p
);
2964 assert( BTS_OVERWRITE
==BTS_SECURE_DELETE
*2 );
2965 assert( BTS_FAST_SECURE
==(BTS_OVERWRITE
|BTS_SECURE_DELETE
) );
2967 p
->pBt
->btsFlags
&= ~BTS_FAST_SECURE
;
2968 p
->pBt
->btsFlags
|= BTS_SECURE_DELETE
*newFlag
;
2970 b
= (p
->pBt
->btsFlags
& BTS_FAST_SECURE
)/BTS_SECURE_DELETE
;
2971 sqlite3BtreeLeave(p
);
2976 ** Change the 'auto-vacuum' property of the database. If the 'autoVacuum'
2977 ** parameter is non-zero, then auto-vacuum mode is enabled. If zero, it
2978 ** is disabled. The default value for the auto-vacuum property is
2979 ** determined by the SQLITE_DEFAULT_AUTOVACUUM macro.
2981 int sqlite3BtreeSetAutoVacuum(Btree
*p
, int autoVacuum
){
2982 #ifdef SQLITE_OMIT_AUTOVACUUM
2983 return SQLITE_READONLY
;
2985 BtShared
*pBt
= p
->pBt
;
2987 u8 av
= (u8
)autoVacuum
;
2989 sqlite3BtreeEnter(p
);
2990 if( (pBt
->btsFlags
& BTS_PAGESIZE_FIXED
)!=0 && (av
?1:0)!=pBt
->autoVacuum
){
2991 rc
= SQLITE_READONLY
;
2993 pBt
->autoVacuum
= av
?1:0;
2994 pBt
->incrVacuum
= av
==2 ?1:0;
2996 sqlite3BtreeLeave(p
);
3002 ** Return the value of the 'auto-vacuum' property. If auto-vacuum is
3003 ** enabled 1 is returned. Otherwise 0.
3005 int sqlite3BtreeGetAutoVacuum(Btree
*p
){
3006 #ifdef SQLITE_OMIT_AUTOVACUUM
3007 return BTREE_AUTOVACUUM_NONE
;
3010 sqlite3BtreeEnter(p
);
3012 (!p
->pBt
->autoVacuum
)?BTREE_AUTOVACUUM_NONE
:
3013 (!p
->pBt
->incrVacuum
)?BTREE_AUTOVACUUM_FULL
:
3014 BTREE_AUTOVACUUM_INCR
3016 sqlite3BtreeLeave(p
);
3022 ** If the user has not set the safety-level for this database connection
3023 ** using "PRAGMA synchronous", and if the safety-level is not already
3024 ** set to the value passed to this function as the second parameter,
3027 #if SQLITE_DEFAULT_SYNCHRONOUS!=SQLITE_DEFAULT_WAL_SYNCHRONOUS \
3028 && !defined(SQLITE_OMIT_WAL)
3029 static void setDefaultSyncFlag(BtShared
*pBt
, u8 safety_level
){
3032 if( (db
=pBt
->db
)!=0 && (pDb
=db
->aDb
)!=0 ){
3033 while( pDb
->pBt
==0 || pDb
->pBt
->pBt
!=pBt
){ pDb
++; }
3034 if( pDb
->bSyncSet
==0
3035 && pDb
->safety_level
!=safety_level
3038 pDb
->safety_level
= safety_level
;
3039 sqlite3PagerSetFlags(pBt
->pPager
,
3040 pDb
->safety_level
| (db
->flags
& PAGER_FLAGS_MASK
));
3045 # define setDefaultSyncFlag(pBt,safety_level)
3048 /* Forward declaration */
3049 static int newDatabase(BtShared
*);
3053 ** Get a reference to pPage1 of the database file. This will
3054 ** also acquire a readlock on that file.
3056 ** SQLITE_OK is returned on success. If the file is not a
3057 ** well-formed database file, then SQLITE_CORRUPT is returned.
3058 ** SQLITE_BUSY is returned if the database is locked. SQLITE_NOMEM
3059 ** is returned if we run out of memory.
3061 static int lockBtree(BtShared
*pBt
){
3062 int rc
; /* Result code from subfunctions */
3063 MemPage
*pPage1
; /* Page 1 of the database file */
3064 u32 nPage
; /* Number of pages in the database */
3065 u32 nPageFile
= 0; /* Number of pages in the database file */
3066 u32 nPageHeader
; /* Number of pages in the database according to hdr */
3068 assert( sqlite3_mutex_held(pBt
->mutex
) );
3069 assert( pBt
->pPage1
==0 );
3070 rc
= sqlite3PagerSharedLock(pBt
->pPager
);
3071 if( rc
!=SQLITE_OK
) return rc
;
3072 rc
= btreeGetPage(pBt
, 1, &pPage1
, 0);
3073 if( rc
!=SQLITE_OK
) return rc
;
3075 /* Do some checking to help insure the file we opened really is
3076 ** a valid database file.
3078 nPage
= nPageHeader
= get4byte(28+(u8
*)pPage1
->aData
);
3079 sqlite3PagerPagecount(pBt
->pPager
, (int*)&nPageFile
);
3080 if( nPage
==0 || memcmp(24+(u8
*)pPage1
->aData
, 92+(u8
*)pPage1
->aData
,4)!=0 ){
3083 if( (pBt
->db
->flags
& SQLITE_ResetDatabase
)!=0 ){
3089 u8
*page1
= pPage1
->aData
;
3091 /* EVIDENCE-OF: R-43737-39999 Every valid SQLite database file begins
3092 ** with the following 16 bytes (in hex): 53 51 4c 69 74 65 20 66 6f 72 6d
3093 ** 61 74 20 33 00. */
3094 if( memcmp(page1
, zMagicHeader
, 16)!=0 ){
3095 goto page1_init_failed
;
3098 #ifdef SQLITE_OMIT_WAL
3100 pBt
->btsFlags
|= BTS_READ_ONLY
;
3103 goto page1_init_failed
;
3107 pBt
->btsFlags
|= BTS_READ_ONLY
;
3110 goto page1_init_failed
;
3113 /* If the write version is set to 2, this database should be accessed
3114 ** in WAL mode. If the log is not already open, open it now. Then
3115 ** return SQLITE_OK and return without populating BtShared.pPage1.
3116 ** The caller detects this and calls this function again. This is
3117 ** required as the version of page 1 currently in the page1 buffer
3118 ** may not be the latest version - there may be a newer one in the log
3121 if( page1
[19]==2 && (pBt
->btsFlags
& BTS_NO_WAL
)==0 ){
3123 rc
= sqlite3PagerOpenWal(pBt
->pPager
, &isOpen
);
3124 if( rc
!=SQLITE_OK
){
3125 goto page1_init_failed
;
3127 setDefaultSyncFlag(pBt
, SQLITE_DEFAULT_WAL_SYNCHRONOUS
+1);
3129 releasePageOne(pPage1
);
3135 setDefaultSyncFlag(pBt
, SQLITE_DEFAULT_SYNCHRONOUS
+1);
3139 /* EVIDENCE-OF: R-15465-20813 The maximum and minimum embedded payload
3140 ** fractions and the leaf payload fraction values must be 64, 32, and 32.
3142 ** The original design allowed these amounts to vary, but as of
3143 ** version 3.6.0, we require them to be fixed.
3145 if( memcmp(&page1
[21], "\100\040\040",3)!=0 ){
3146 goto page1_init_failed
;
3148 /* EVIDENCE-OF: R-51873-39618 The page size for a database file is
3149 ** determined by the 2-byte integer located at an offset of 16 bytes from
3150 ** the beginning of the database file. */
3151 pageSize
= (page1
[16]<<8) | (page1
[17]<<16);
3152 /* EVIDENCE-OF: R-25008-21688 The size of a page is a power of two
3153 ** between 512 and 65536 inclusive. */
3154 if( ((pageSize
-1)&pageSize
)!=0
3155 || pageSize
>SQLITE_MAX_PAGE_SIZE
3158 goto page1_init_failed
;
3160 pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
3161 assert( (pageSize
& 7)==0 );
3162 /* EVIDENCE-OF: R-59310-51205 The "reserved space" size in the 1-byte
3163 ** integer at offset 20 is the number of bytes of space at the end of
3164 ** each page to reserve for extensions.
3166 ** EVIDENCE-OF: R-37497-42412 The size of the reserved region is
3167 ** determined by the one-byte unsigned integer found at an offset of 20
3168 ** into the database file header. */
3169 usableSize
= pageSize
- page1
[20];
3170 if( (u32
)pageSize
!=pBt
->pageSize
){
3171 /* After reading the first page of the database assuming a page size
3172 ** of BtShared.pageSize, we have discovered that the page-size is
3173 ** actually pageSize. Unlock the database, leave pBt->pPage1 at
3174 ** zero and return SQLITE_OK. The caller will call this function
3175 ** again with the correct page-size.
3177 releasePageOne(pPage1
);
3178 pBt
->usableSize
= usableSize
;
3179 pBt
->pageSize
= pageSize
;
3181 rc
= sqlite3PagerSetPagesize(pBt
->pPager
, &pBt
->pageSize
,
3182 pageSize
-usableSize
);
3185 if( sqlite3WritableSchema(pBt
->db
)==0 && nPage
>nPageFile
){
3186 rc
= SQLITE_CORRUPT_BKPT
;
3187 goto page1_init_failed
;
3189 /* EVIDENCE-OF: R-28312-64704 However, the usable size is not allowed to
3190 ** be less than 480. In other words, if the page size is 512, then the
3191 ** reserved space size cannot exceed 32. */
3192 if( usableSize
<480 ){
3193 goto page1_init_failed
;
3195 pBt
->pageSize
= pageSize
;
3196 pBt
->usableSize
= usableSize
;
3197 #ifndef SQLITE_OMIT_AUTOVACUUM
3198 pBt
->autoVacuum
= (get4byte(&page1
[36 + 4*4])?1:0);
3199 pBt
->incrVacuum
= (get4byte(&page1
[36 + 7*4])?1:0);
3203 /* maxLocal is the maximum amount of payload to store locally for
3204 ** a cell. Make sure it is small enough so that at least minFanout
3205 ** cells can will fit on one page. We assume a 10-byte page header.
3206 ** Besides the payload, the cell must store:
3207 ** 2-byte pointer to the cell
3208 ** 4-byte child pointer
3209 ** 9-byte nKey value
3210 ** 4-byte nData value
3211 ** 4-byte overflow page pointer
3212 ** So a cell consists of a 2-byte pointer, a header which is as much as
3213 ** 17 bytes long, 0 to N bytes of payload, and an optional 4 byte overflow
3216 pBt
->maxLocal
= (u16
)((pBt
->usableSize
-12)*64/255 - 23);
3217 pBt
->minLocal
= (u16
)((pBt
->usableSize
-12)*32/255 - 23);
3218 pBt
->maxLeaf
= (u16
)(pBt
->usableSize
- 35);
3219 pBt
->minLeaf
= (u16
)((pBt
->usableSize
-12)*32/255 - 23);
3220 if( pBt
->maxLocal
>127 ){
3221 pBt
->max1bytePayload
= 127;
3223 pBt
->max1bytePayload
= (u8
)pBt
->maxLocal
;
3225 assert( pBt
->maxLeaf
+ 23 <= MX_CELL_SIZE(pBt
) );
3226 pBt
->pPage1
= pPage1
;
3231 releasePageOne(pPage1
);
3238 ** Return the number of cursors open on pBt. This is for use
3239 ** in assert() expressions, so it is only compiled if NDEBUG is not
3242 ** Only write cursors are counted if wrOnly is true. If wrOnly is
3243 ** false then all cursors are counted.
3245 ** For the purposes of this routine, a cursor is any cursor that
3246 ** is capable of reading or writing to the database. Cursors that
3247 ** have been tripped into the CURSOR_FAULT state are not counted.
3249 static int countValidCursors(BtShared
*pBt
, int wrOnly
){
3252 for(pCur
=pBt
->pCursor
; pCur
; pCur
=pCur
->pNext
){
3253 if( (wrOnly
==0 || (pCur
->curFlags
& BTCF_WriteFlag
)!=0)
3254 && pCur
->eState
!=CURSOR_FAULT
) r
++;
3261 ** If there are no outstanding cursors and we are not in the middle
3262 ** of a transaction but there is a read lock on the database, then
3263 ** this routine unrefs the first page of the database file which
3264 ** has the effect of releasing the read lock.
3266 ** If there is a transaction in progress, this routine is a no-op.
3268 static void unlockBtreeIfUnused(BtShared
*pBt
){
3269 assert( sqlite3_mutex_held(pBt
->mutex
) );
3270 assert( countValidCursors(pBt
,0)==0 || pBt
->inTransaction
>TRANS_NONE
);
3271 if( pBt
->inTransaction
==TRANS_NONE
&& pBt
->pPage1
!=0 ){
3272 MemPage
*pPage1
= pBt
->pPage1
;
3273 assert( pPage1
->aData
);
3274 assert( sqlite3PagerRefcount(pBt
->pPager
)==1 );
3276 releasePageOne(pPage1
);
3281 ** If pBt points to an empty file then convert that empty file
3282 ** into a new empty database by initializing the first page of
3285 static int newDatabase(BtShared
*pBt
){
3287 unsigned char *data
;
3290 assert( sqlite3_mutex_held(pBt
->mutex
) );
3297 rc
= sqlite3PagerWrite(pP1
->pDbPage
);
3299 memcpy(data
, zMagicHeader
, sizeof(zMagicHeader
));
3300 assert( sizeof(zMagicHeader
)==16 );
3301 data
[16] = (u8
)((pBt
->pageSize
>>8)&0xff);
3302 data
[17] = (u8
)((pBt
->pageSize
>>16)&0xff);
3305 assert( pBt
->usableSize
<=pBt
->pageSize
&& pBt
->usableSize
+255>=pBt
->pageSize
);
3306 data
[20] = (u8
)(pBt
->pageSize
- pBt
->usableSize
);
3310 memset(&data
[24], 0, 100-24);
3311 zeroPage(pP1
, PTF_INTKEY
|PTF_LEAF
|PTF_LEAFDATA
);
3312 pBt
->btsFlags
|= BTS_PAGESIZE_FIXED
;
3313 #ifndef SQLITE_OMIT_AUTOVACUUM
3314 assert( pBt
->autoVacuum
==1 || pBt
->autoVacuum
==0 );
3315 assert( pBt
->incrVacuum
==1 || pBt
->incrVacuum
==0 );
3316 put4byte(&data
[36 + 4*4], pBt
->autoVacuum
);
3317 put4byte(&data
[36 + 7*4], pBt
->incrVacuum
);
3325 ** Initialize the first page of the database file (creating a database
3326 ** consisting of a single page and no schema objects). Return SQLITE_OK
3327 ** if successful, or an SQLite error code otherwise.
3329 int sqlite3BtreeNewDb(Btree
*p
){
3331 sqlite3BtreeEnter(p
);
3333 rc
= newDatabase(p
->pBt
);
3334 sqlite3BtreeLeave(p
);
3339 ** Attempt to start a new transaction. A write-transaction
3340 ** is started if the second argument is nonzero, otherwise a read-
3341 ** transaction. If the second argument is 2 or more and exclusive
3342 ** transaction is started, meaning that no other process is allowed
3343 ** to access the database. A preexisting transaction may not be
3344 ** upgraded to exclusive by calling this routine a second time - the
3345 ** exclusivity flag only works for a new transaction.
3347 ** A write-transaction must be started before attempting any
3348 ** changes to the database. None of the following routines
3349 ** will work unless a transaction is started first:
3351 ** sqlite3BtreeCreateTable()
3352 ** sqlite3BtreeCreateIndex()
3353 ** sqlite3BtreeClearTable()
3354 ** sqlite3BtreeDropTable()
3355 ** sqlite3BtreeInsert()
3356 ** sqlite3BtreeDelete()
3357 ** sqlite3BtreeUpdateMeta()
3359 ** If an initial attempt to acquire the lock fails because of lock contention
3360 ** and the database was previously unlocked, then invoke the busy handler
3361 ** if there is one. But if there was previously a read-lock, do not
3362 ** invoke the busy handler - just return SQLITE_BUSY. SQLITE_BUSY is
3363 ** returned when there is already a read-lock in order to avoid a deadlock.
3365 ** Suppose there are two processes A and B. A has a read lock and B has
3366 ** a reserved lock. B tries to promote to exclusive but is blocked because
3367 ** of A's read lock. A tries to promote to reserved but is blocked by B.
3368 ** One or the other of the two processes must give way or there can be
3369 ** no progress. By returning SQLITE_BUSY and not invoking the busy callback
3370 ** when A already has a read lock, we encourage A to give up and let B
3373 int sqlite3BtreeBeginTrans(Btree
*p
, int wrflag
, int *pSchemaVersion
){
3374 BtShared
*pBt
= p
->pBt
;
3375 Pager
*pPager
= pBt
->pPager
;
3378 sqlite3BtreeEnter(p
);
3381 /* If the btree is already in a write-transaction, or it
3382 ** is already in a read-transaction and a read-transaction
3383 ** is requested, this is a no-op.
3385 if( p
->inTrans
==TRANS_WRITE
|| (p
->inTrans
==TRANS_READ
&& !wrflag
) ){
3388 assert( pBt
->inTransaction
==TRANS_WRITE
|| IfNotOmitAV(pBt
->bDoTruncate
)==0 );
3390 if( (p
->db
->flags
& SQLITE_ResetDatabase
)
3391 && sqlite3PagerIsreadonly(pPager
)==0
3393 pBt
->btsFlags
&= ~BTS_READ_ONLY
;
3396 /* Write transactions are not possible on a read-only database */
3397 if( (pBt
->btsFlags
& BTS_READ_ONLY
)!=0 && wrflag
){
3398 rc
= SQLITE_READONLY
;
3402 #ifndef SQLITE_OMIT_SHARED_CACHE
3404 sqlite3
*pBlock
= 0;
3405 /* If another database handle has already opened a write transaction
3406 ** on this shared-btree structure and a second write transaction is
3407 ** requested, return SQLITE_LOCKED.
3409 if( (wrflag
&& pBt
->inTransaction
==TRANS_WRITE
)
3410 || (pBt
->btsFlags
& BTS_PENDING
)!=0
3412 pBlock
= pBt
->pWriter
->db
;
3413 }else if( wrflag
>1 ){
3415 for(pIter
=pBt
->pLock
; pIter
; pIter
=pIter
->pNext
){
3416 if( pIter
->pBtree
!=p
){
3417 pBlock
= pIter
->pBtree
->db
;
3423 sqlite3ConnectionBlocked(p
->db
, pBlock
);
3424 rc
= SQLITE_LOCKED_SHAREDCACHE
;
3430 /* Any read-only or read-write transaction implies a read-lock on
3431 ** page 1. So if some other shared-cache client already has a write-lock
3432 ** on page 1, the transaction cannot be opened. */
3433 rc
= querySharedCacheTableLock(p
, SCHEMA_ROOT
, READ_LOCK
);
3434 if( SQLITE_OK
!=rc
) goto trans_begun
;
3436 pBt
->btsFlags
&= ~BTS_INITIALLY_EMPTY
;
3437 if( pBt
->nPage
==0 ) pBt
->btsFlags
|= BTS_INITIALLY_EMPTY
;
3439 sqlite3PagerWalDb(pPager
, p
->db
);
3441 #ifdef SQLITE_ENABLE_SETLK_TIMEOUT
3442 /* If transitioning from no transaction directly to a write transaction,
3443 ** block for the WRITER lock first if possible. */
3444 if( pBt
->pPage1
==0 && wrflag
){
3445 assert( pBt
->inTransaction
==TRANS_NONE
);
3446 rc
= sqlite3PagerWalWriteLock(pPager
, 1);
3447 if( rc
!=SQLITE_BUSY
&& rc
!=SQLITE_OK
) break;
3451 /* Call lockBtree() until either pBt->pPage1 is populated or
3452 ** lockBtree() returns something other than SQLITE_OK. lockBtree()
3453 ** may return SQLITE_OK but leave pBt->pPage1 set to 0 if after
3454 ** reading page 1 it discovers that the page-size of the database
3455 ** file is not pBt->pageSize. In this case lockBtree() will update
3456 ** pBt->pageSize to the page-size of the file on disk.
3458 while( pBt
->pPage1
==0 && SQLITE_OK
==(rc
= lockBtree(pBt
)) );
3460 if( rc
==SQLITE_OK
&& wrflag
){
3461 if( (pBt
->btsFlags
& BTS_READ_ONLY
)!=0 ){
3462 rc
= SQLITE_READONLY
;
3464 rc
= sqlite3PagerBegin(pPager
, wrflag
>1, sqlite3TempInMemory(p
->db
));
3465 if( rc
==SQLITE_OK
){
3466 rc
= newDatabase(pBt
);
3467 }else if( rc
==SQLITE_BUSY_SNAPSHOT
&& pBt
->inTransaction
==TRANS_NONE
){
3468 /* if there was no transaction opened when this function was
3469 ** called and SQLITE_BUSY_SNAPSHOT is returned, change the error
3470 ** code to SQLITE_BUSY. */
3476 if( rc
!=SQLITE_OK
){
3477 (void)sqlite3PagerWalWriteLock(pPager
, 0);
3478 unlockBtreeIfUnused(pBt
);
3480 }while( (rc
&0xFF)==SQLITE_BUSY
&& pBt
->inTransaction
==TRANS_NONE
&&
3481 btreeInvokeBusyHandler(pBt
) );
3482 sqlite3PagerWalDb(pPager
, 0);
3483 #ifdef SQLITE_ENABLE_SETLK_TIMEOUT
3484 if( rc
==SQLITE_BUSY_TIMEOUT
) rc
= SQLITE_BUSY
;
3487 if( rc
==SQLITE_OK
){
3488 if( p
->inTrans
==TRANS_NONE
){
3489 pBt
->nTransaction
++;
3490 #ifndef SQLITE_OMIT_SHARED_CACHE
3492 assert( p
->lock
.pBtree
==p
&& p
->lock
.iTable
==1 );
3493 p
->lock
.eLock
= READ_LOCK
;
3494 p
->lock
.pNext
= pBt
->pLock
;
3495 pBt
->pLock
= &p
->lock
;
3499 p
->inTrans
= (wrflag
?TRANS_WRITE
:TRANS_READ
);
3500 if( p
->inTrans
>pBt
->inTransaction
){
3501 pBt
->inTransaction
= p
->inTrans
;
3504 MemPage
*pPage1
= pBt
->pPage1
;
3505 #ifndef SQLITE_OMIT_SHARED_CACHE
3506 assert( !pBt
->pWriter
);
3508 pBt
->btsFlags
&= ~BTS_EXCLUSIVE
;
3509 if( wrflag
>1 ) pBt
->btsFlags
|= BTS_EXCLUSIVE
;
3512 /* If the db-size header field is incorrect (as it may be if an old
3513 ** client has been writing the database file), update it now. Doing
3514 ** this sooner rather than later means the database size can safely
3515 ** re-read the database size from page 1 if a savepoint or transaction
3516 ** rollback occurs within the transaction.
3518 if( pBt
->nPage
!=get4byte(&pPage1
->aData
[28]) ){
3519 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
3520 if( rc
==SQLITE_OK
){
3521 put4byte(&pPage1
->aData
[28], pBt
->nPage
);
3528 if( rc
==SQLITE_OK
){
3529 if( pSchemaVersion
){
3530 *pSchemaVersion
= get4byte(&pBt
->pPage1
->aData
[40]);
3533 /* This call makes sure that the pager has the correct number of
3534 ** open savepoints. If the second parameter is greater than 0 and
3535 ** the sub-journal is not already open, then it will be opened here.
3537 rc
= sqlite3PagerOpenSavepoint(pPager
, p
->db
->nSavepoint
);
3542 sqlite3BtreeLeave(p
);
3546 #ifndef SQLITE_OMIT_AUTOVACUUM
3549 ** Set the pointer-map entries for all children of page pPage. Also, if
3550 ** pPage contains cells that point to overflow pages, set the pointer
3551 ** map entries for the overflow pages as well.
3553 static int setChildPtrmaps(MemPage
*pPage
){
3554 int i
; /* Counter variable */
3555 int nCell
; /* Number of cells in page pPage */
3556 int rc
; /* Return code */
3557 BtShared
*pBt
= pPage
->pBt
;
3558 Pgno pgno
= pPage
->pgno
;
3560 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
3561 rc
= pPage
->isInit
? SQLITE_OK
: btreeInitPage(pPage
);
3562 if( rc
!=SQLITE_OK
) return rc
;
3563 nCell
= pPage
->nCell
;
3565 for(i
=0; i
<nCell
; i
++){
3566 u8
*pCell
= findCell(pPage
, i
);
3568 ptrmapPutOvflPtr(pPage
, pPage
, pCell
, &rc
);
3571 Pgno childPgno
= get4byte(pCell
);
3572 ptrmapPut(pBt
, childPgno
, PTRMAP_BTREE
, pgno
, &rc
);
3577 Pgno childPgno
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
3578 ptrmapPut(pBt
, childPgno
, PTRMAP_BTREE
, pgno
, &rc
);
3585 ** Somewhere on pPage is a pointer to page iFrom. Modify this pointer so
3586 ** that it points to iTo. Parameter eType describes the type of pointer to
3587 ** be modified, as follows:
3589 ** PTRMAP_BTREE: pPage is a btree-page. The pointer points at a child
3592 ** PTRMAP_OVERFLOW1: pPage is a btree-page. The pointer points at an overflow
3593 ** page pointed to by one of the cells on pPage.
3595 ** PTRMAP_OVERFLOW2: pPage is an overflow-page. The pointer points at the next
3596 ** overflow page in the list.
3598 static int modifyPagePointer(MemPage
*pPage
, Pgno iFrom
, Pgno iTo
, u8 eType
){
3599 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
3600 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
3601 if( eType
==PTRMAP_OVERFLOW2
){
3602 /* The pointer is always the first 4 bytes of the page in this case. */
3603 if( get4byte(pPage
->aData
)!=iFrom
){
3604 return SQLITE_CORRUPT_PAGE(pPage
);
3606 put4byte(pPage
->aData
, iTo
);
3612 rc
= pPage
->isInit
? SQLITE_OK
: btreeInitPage(pPage
);
3614 nCell
= pPage
->nCell
;
3616 for(i
=0; i
<nCell
; i
++){
3617 u8
*pCell
= findCell(pPage
, i
);
3618 if( eType
==PTRMAP_OVERFLOW1
){
3620 pPage
->xParseCell(pPage
, pCell
, &info
);
3621 if( info
.nLocal
<info
.nPayload
){
3622 if( pCell
+info
.nSize
> pPage
->aData
+pPage
->pBt
->usableSize
){
3623 return SQLITE_CORRUPT_PAGE(pPage
);
3625 if( iFrom
==get4byte(pCell
+info
.nSize
-4) ){
3626 put4byte(pCell
+info
.nSize
-4, iTo
);
3631 if( get4byte(pCell
)==iFrom
){
3632 put4byte(pCell
, iTo
);
3639 if( eType
!=PTRMAP_BTREE
||
3640 get4byte(&pPage
->aData
[pPage
->hdrOffset
+8])!=iFrom
){
3641 return SQLITE_CORRUPT_PAGE(pPage
);
3643 put4byte(&pPage
->aData
[pPage
->hdrOffset
+8], iTo
);
3651 ** Move the open database page pDbPage to location iFreePage in the
3652 ** database. The pDbPage reference remains valid.
3654 ** The isCommit flag indicates that there is no need to remember that
3655 ** the journal needs to be sync()ed before database page pDbPage->pgno
3656 ** can be written to. The caller has already promised not to write to that
3659 static int relocatePage(
3660 BtShared
*pBt
, /* Btree */
3661 MemPage
*pDbPage
, /* Open page to move */
3662 u8 eType
, /* Pointer map 'type' entry for pDbPage */
3663 Pgno iPtrPage
, /* Pointer map 'page-no' entry for pDbPage */
3664 Pgno iFreePage
, /* The location to move pDbPage to */
3665 int isCommit
/* isCommit flag passed to sqlite3PagerMovepage */
3667 MemPage
*pPtrPage
; /* The page that contains a pointer to pDbPage */
3668 Pgno iDbPage
= pDbPage
->pgno
;
3669 Pager
*pPager
= pBt
->pPager
;
3672 assert( eType
==PTRMAP_OVERFLOW2
|| eType
==PTRMAP_OVERFLOW1
||
3673 eType
==PTRMAP_BTREE
|| eType
==PTRMAP_ROOTPAGE
);
3674 assert( sqlite3_mutex_held(pBt
->mutex
) );
3675 assert( pDbPage
->pBt
==pBt
);
3676 if( iDbPage
<3 ) return SQLITE_CORRUPT_BKPT
;
3678 /* Move page iDbPage from its current location to page number iFreePage */
3679 TRACE(("AUTOVACUUM: Moving %d to free page %d (ptr page %d type %d)\n",
3680 iDbPage
, iFreePage
, iPtrPage
, eType
));
3681 rc
= sqlite3PagerMovepage(pPager
, pDbPage
->pDbPage
, iFreePage
, isCommit
);
3682 if( rc
!=SQLITE_OK
){
3685 pDbPage
->pgno
= iFreePage
;
3687 /* If pDbPage was a btree-page, then it may have child pages and/or cells
3688 ** that point to overflow pages. The pointer map entries for all these
3689 ** pages need to be changed.
3691 ** If pDbPage is an overflow page, then the first 4 bytes may store a
3692 ** pointer to a subsequent overflow page. If this is the case, then
3693 ** the pointer map needs to be updated for the subsequent overflow page.
3695 if( eType
==PTRMAP_BTREE
|| eType
==PTRMAP_ROOTPAGE
){
3696 rc
= setChildPtrmaps(pDbPage
);
3697 if( rc
!=SQLITE_OK
){
3701 Pgno nextOvfl
= get4byte(pDbPage
->aData
);
3703 ptrmapPut(pBt
, nextOvfl
, PTRMAP_OVERFLOW2
, iFreePage
, &rc
);
3704 if( rc
!=SQLITE_OK
){
3710 /* Fix the database pointer on page iPtrPage that pointed at iDbPage so
3711 ** that it points at iFreePage. Also fix the pointer map entry for
3714 if( eType
!=PTRMAP_ROOTPAGE
){
3715 rc
= btreeGetPage(pBt
, iPtrPage
, &pPtrPage
, 0);
3716 if( rc
!=SQLITE_OK
){
3719 rc
= sqlite3PagerWrite(pPtrPage
->pDbPage
);
3720 if( rc
!=SQLITE_OK
){
3721 releasePage(pPtrPage
);
3724 rc
= modifyPagePointer(pPtrPage
, iDbPage
, iFreePage
, eType
);
3725 releasePage(pPtrPage
);
3726 if( rc
==SQLITE_OK
){
3727 ptrmapPut(pBt
, iFreePage
, eType
, iPtrPage
, &rc
);
3733 /* Forward declaration required by incrVacuumStep(). */
3734 static int allocateBtreePage(BtShared
*, MemPage
**, Pgno
*, Pgno
, u8
);
3737 ** Perform a single step of an incremental-vacuum. If successful, return
3738 ** SQLITE_OK. If there is no work to do (and therefore no point in
3739 ** calling this function again), return SQLITE_DONE. Or, if an error
3740 ** occurs, return some other error code.
3742 ** More specifically, this function attempts to re-organize the database so
3743 ** that the last page of the file currently in use is no longer in use.
3745 ** Parameter nFin is the number of pages that this database would contain
3746 ** were this function called until it returns SQLITE_DONE.
3748 ** If the bCommit parameter is non-zero, this function assumes that the
3749 ** caller will keep calling incrVacuumStep() until it returns SQLITE_DONE
3750 ** or an error. bCommit is passed true for an auto-vacuum-on-commit
3751 ** operation, or false for an incremental vacuum.
3753 static int incrVacuumStep(BtShared
*pBt
, Pgno nFin
, Pgno iLastPg
, int bCommit
){
3754 Pgno nFreeList
; /* Number of pages still on the free-list */
3757 assert( sqlite3_mutex_held(pBt
->mutex
) );
3758 assert( iLastPg
>nFin
);
3760 if( !PTRMAP_ISPAGE(pBt
, iLastPg
) && iLastPg
!=PENDING_BYTE_PAGE(pBt
) ){
3764 nFreeList
= get4byte(&pBt
->pPage1
->aData
[36]);
3769 rc
= ptrmapGet(pBt
, iLastPg
, &eType
, &iPtrPage
);
3770 if( rc
!=SQLITE_OK
){
3773 if( eType
==PTRMAP_ROOTPAGE
){
3774 return SQLITE_CORRUPT_BKPT
;
3777 if( eType
==PTRMAP_FREEPAGE
){
3779 /* Remove the page from the files free-list. This is not required
3780 ** if bCommit is non-zero. In that case, the free-list will be
3781 ** truncated to zero after this function returns, so it doesn't
3782 ** matter if it still contains some garbage entries.
3786 rc
= allocateBtreePage(pBt
, &pFreePg
, &iFreePg
, iLastPg
, BTALLOC_EXACT
);
3787 if( rc
!=SQLITE_OK
){
3790 assert( iFreePg
==iLastPg
);
3791 releasePage(pFreePg
);
3794 Pgno iFreePg
; /* Index of free page to move pLastPg to */
3796 u8 eMode
= BTALLOC_ANY
; /* Mode parameter for allocateBtreePage() */
3797 Pgno iNear
= 0; /* nearby parameter for allocateBtreePage() */
3799 rc
= btreeGetPage(pBt
, iLastPg
, &pLastPg
, 0);
3800 if( rc
!=SQLITE_OK
){
3804 /* If bCommit is zero, this loop runs exactly once and page pLastPg
3805 ** is swapped with the first free page pulled off the free list.
3807 ** On the other hand, if bCommit is greater than zero, then keep
3808 ** looping until a free-page located within the first nFin pages
3809 ** of the file is found.
3817 rc
= allocateBtreePage(pBt
, &pFreePg
, &iFreePg
, iNear
, eMode
);
3818 if( rc
!=SQLITE_OK
){
3819 releasePage(pLastPg
);
3822 releasePage(pFreePg
);
3823 }while( bCommit
&& iFreePg
>nFin
);
3824 assert( iFreePg
<iLastPg
);
3826 rc
= relocatePage(pBt
, pLastPg
, eType
, iPtrPage
, iFreePg
, bCommit
);
3827 releasePage(pLastPg
);
3828 if( rc
!=SQLITE_OK
){
3837 }while( iLastPg
==PENDING_BYTE_PAGE(pBt
) || PTRMAP_ISPAGE(pBt
, iLastPg
) );
3838 pBt
->bDoTruncate
= 1;
3839 pBt
->nPage
= iLastPg
;
3845 ** The database opened by the first argument is an auto-vacuum database
3846 ** nOrig pages in size containing nFree free pages. Return the expected
3847 ** size of the database in pages following an auto-vacuum operation.
3849 static Pgno
finalDbSize(BtShared
*pBt
, Pgno nOrig
, Pgno nFree
){
3850 int nEntry
; /* Number of entries on one ptrmap page */
3851 Pgno nPtrmap
; /* Number of PtrMap pages to be freed */
3852 Pgno nFin
; /* Return value */
3854 nEntry
= pBt
->usableSize
/5;
3855 nPtrmap
= (nFree
-nOrig
+PTRMAP_PAGENO(pBt
, nOrig
)+nEntry
)/nEntry
;
3856 nFin
= nOrig
- nFree
- nPtrmap
;
3857 if( nOrig
>PENDING_BYTE_PAGE(pBt
) && nFin
<PENDING_BYTE_PAGE(pBt
) ){
3860 while( PTRMAP_ISPAGE(pBt
, nFin
) || nFin
==PENDING_BYTE_PAGE(pBt
) ){
3868 ** A write-transaction must be opened before calling this function.
3869 ** It performs a single unit of work towards an incremental vacuum.
3871 ** If the incremental vacuum is finished after this function has run,
3872 ** SQLITE_DONE is returned. If it is not finished, but no error occurred,
3873 ** SQLITE_OK is returned. Otherwise an SQLite error code.
3875 int sqlite3BtreeIncrVacuum(Btree
*p
){
3877 BtShared
*pBt
= p
->pBt
;
3879 sqlite3BtreeEnter(p
);
3880 assert( pBt
->inTransaction
==TRANS_WRITE
&& p
->inTrans
==TRANS_WRITE
);
3881 if( !pBt
->autoVacuum
){
3884 Pgno nOrig
= btreePagecount(pBt
);
3885 Pgno nFree
= get4byte(&pBt
->pPage1
->aData
[36]);
3886 Pgno nFin
= finalDbSize(pBt
, nOrig
, nFree
);
3888 if( nOrig
<nFin
|| nFree
>=nOrig
){
3889 rc
= SQLITE_CORRUPT_BKPT
;
3890 }else if( nFree
>0 ){
3891 rc
= saveAllCursors(pBt
, 0, 0);
3892 if( rc
==SQLITE_OK
){
3893 invalidateAllOverflowCache(pBt
);
3894 rc
= incrVacuumStep(pBt
, nFin
, nOrig
, 0);
3896 if( rc
==SQLITE_OK
){
3897 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
3898 put4byte(&pBt
->pPage1
->aData
[28], pBt
->nPage
);
3904 sqlite3BtreeLeave(p
);
3909 ** This routine is called prior to sqlite3PagerCommit when a transaction
3910 ** is committed for an auto-vacuum database.
3912 ** If SQLITE_OK is returned, then *pnTrunc is set to the number of pages
3913 ** the database file should be truncated to during the commit process.
3914 ** i.e. the database has been reorganized so that only the first *pnTrunc
3915 ** pages are in use.
3917 static int autoVacuumCommit(BtShared
*pBt
){
3919 Pager
*pPager
= pBt
->pPager
;
3920 VVA_ONLY( int nRef
= sqlite3PagerRefcount(pPager
); )
3922 assert( sqlite3_mutex_held(pBt
->mutex
) );
3923 invalidateAllOverflowCache(pBt
);
3924 assert(pBt
->autoVacuum
);
3925 if( !pBt
->incrVacuum
){
3926 Pgno nFin
; /* Number of pages in database after autovacuuming */
3927 Pgno nFree
; /* Number of pages on the freelist initially */
3928 Pgno iFree
; /* The next page to be freed */
3929 Pgno nOrig
; /* Database size before freeing */
3931 nOrig
= btreePagecount(pBt
);
3932 if( PTRMAP_ISPAGE(pBt
, nOrig
) || nOrig
==PENDING_BYTE_PAGE(pBt
) ){
3933 /* It is not possible to create a database for which the final page
3934 ** is either a pointer-map page or the pending-byte page. If one
3935 ** is encountered, this indicates corruption.
3937 return SQLITE_CORRUPT_BKPT
;
3940 nFree
= get4byte(&pBt
->pPage1
->aData
[36]);
3941 nFin
= finalDbSize(pBt
, nOrig
, nFree
);
3942 if( nFin
>nOrig
) return SQLITE_CORRUPT_BKPT
;
3944 rc
= saveAllCursors(pBt
, 0, 0);
3946 for(iFree
=nOrig
; iFree
>nFin
&& rc
==SQLITE_OK
; iFree
--){
3947 rc
= incrVacuumStep(pBt
, nFin
, iFree
, 1);
3949 if( (rc
==SQLITE_DONE
|| rc
==SQLITE_OK
) && nFree
>0 ){
3950 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
3951 put4byte(&pBt
->pPage1
->aData
[32], 0);
3952 put4byte(&pBt
->pPage1
->aData
[36], 0);
3953 put4byte(&pBt
->pPage1
->aData
[28], nFin
);
3954 pBt
->bDoTruncate
= 1;
3957 if( rc
!=SQLITE_OK
){
3958 sqlite3PagerRollback(pPager
);
3962 assert( nRef
>=sqlite3PagerRefcount(pPager
) );
3966 #else /* ifndef SQLITE_OMIT_AUTOVACUUM */
3967 # define setChildPtrmaps(x) SQLITE_OK
3971 ** This routine does the first phase of a two-phase commit. This routine
3972 ** causes a rollback journal to be created (if it does not already exist)
3973 ** and populated with enough information so that if a power loss occurs
3974 ** the database can be restored to its original state by playing back
3975 ** the journal. Then the contents of the journal are flushed out to
3976 ** the disk. After the journal is safely on oxide, the changes to the
3977 ** database are written into the database file and flushed to oxide.
3978 ** At the end of this call, the rollback journal still exists on the
3979 ** disk and we are still holding all locks, so the transaction has not
3980 ** committed. See sqlite3BtreeCommitPhaseTwo() for the second phase of the
3983 ** This call is a no-op if no write-transaction is currently active on pBt.
3985 ** Otherwise, sync the database file for the btree pBt. zSuperJrnl points to
3986 ** the name of a super-journal file that should be written into the
3987 ** individual journal file, or is NULL, indicating no super-journal file
3988 ** (single database transaction).
3990 ** When this is called, the super-journal should already have been
3991 ** created, populated with this journal pointer and synced to disk.
3993 ** Once this is routine has returned, the only thing required to commit
3994 ** the write-transaction for this database file is to delete the journal.
3996 int sqlite3BtreeCommitPhaseOne(Btree
*p
, const char *zSuperJrnl
){
3998 if( p
->inTrans
==TRANS_WRITE
){
3999 BtShared
*pBt
= p
->pBt
;
4000 sqlite3BtreeEnter(p
);
4001 #ifndef SQLITE_OMIT_AUTOVACUUM
4002 if( pBt
->autoVacuum
){
4003 rc
= autoVacuumCommit(pBt
);
4004 if( rc
!=SQLITE_OK
){
4005 sqlite3BtreeLeave(p
);
4009 if( pBt
->bDoTruncate
){
4010 sqlite3PagerTruncateImage(pBt
->pPager
, pBt
->nPage
);
4013 rc
= sqlite3PagerCommitPhaseOne(pBt
->pPager
, zSuperJrnl
, 0);
4014 sqlite3BtreeLeave(p
);
4020 ** This function is called from both BtreeCommitPhaseTwo() and BtreeRollback()
4021 ** at the conclusion of a transaction.
4023 static void btreeEndTransaction(Btree
*p
){
4024 BtShared
*pBt
= p
->pBt
;
4025 sqlite3
*db
= p
->db
;
4026 assert( sqlite3BtreeHoldsMutex(p
) );
4028 #ifndef SQLITE_OMIT_AUTOVACUUM
4029 pBt
->bDoTruncate
= 0;
4031 if( p
->inTrans
>TRANS_NONE
&& db
->nVdbeRead
>1 ){
4032 /* If there are other active statements that belong to this database
4033 ** handle, downgrade to a read-only transaction. The other statements
4034 ** may still be reading from the database. */
4035 downgradeAllSharedCacheTableLocks(p
);
4036 p
->inTrans
= TRANS_READ
;
4038 /* If the handle had any kind of transaction open, decrement the
4039 ** transaction count of the shared btree. If the transaction count
4040 ** reaches 0, set the shared state to TRANS_NONE. The unlockBtreeIfUnused()
4041 ** call below will unlock the pager. */
4042 if( p
->inTrans
!=TRANS_NONE
){
4043 clearAllSharedCacheTableLocks(p
);
4044 pBt
->nTransaction
--;
4045 if( 0==pBt
->nTransaction
){
4046 pBt
->inTransaction
= TRANS_NONE
;
4050 /* Set the current transaction state to TRANS_NONE and unlock the
4051 ** pager if this call closed the only read or write transaction. */
4052 p
->inTrans
= TRANS_NONE
;
4053 unlockBtreeIfUnused(pBt
);
4060 ** Commit the transaction currently in progress.
4062 ** This routine implements the second phase of a 2-phase commit. The
4063 ** sqlite3BtreeCommitPhaseOne() routine does the first phase and should
4064 ** be invoked prior to calling this routine. The sqlite3BtreeCommitPhaseOne()
4065 ** routine did all the work of writing information out to disk and flushing the
4066 ** contents so that they are written onto the disk platter. All this
4067 ** routine has to do is delete or truncate or zero the header in the
4068 ** the rollback journal (which causes the transaction to commit) and
4071 ** Normally, if an error occurs while the pager layer is attempting to
4072 ** finalize the underlying journal file, this function returns an error and
4073 ** the upper layer will attempt a rollback. However, if the second argument
4074 ** is non-zero then this b-tree transaction is part of a multi-file
4075 ** transaction. In this case, the transaction has already been committed
4076 ** (by deleting a super-journal file) and the caller will ignore this
4077 ** functions return code. So, even if an error occurs in the pager layer,
4078 ** reset the b-tree objects internal state to indicate that the write
4079 ** transaction has been closed. This is quite safe, as the pager will have
4080 ** transitioned to the error state.
4082 ** This will release the write lock on the database file. If there
4083 ** are no active cursors, it also releases the read lock.
4085 int sqlite3BtreeCommitPhaseTwo(Btree
*p
, int bCleanup
){
4087 if( p
->inTrans
==TRANS_NONE
) return SQLITE_OK
;
4088 sqlite3BtreeEnter(p
);
4091 /* If the handle has a write-transaction open, commit the shared-btrees
4092 ** transaction and set the shared state to TRANS_READ.
4094 if( p
->inTrans
==TRANS_WRITE
){
4096 BtShared
*pBt
= p
->pBt
;
4097 assert( pBt
->inTransaction
==TRANS_WRITE
);
4098 assert( pBt
->nTransaction
>0 );
4099 rc
= sqlite3PagerCommitPhaseTwo(pBt
->pPager
);
4100 if( rc
!=SQLITE_OK
&& bCleanup
==0 ){
4101 sqlite3BtreeLeave(p
);
4104 p
->iDataVersion
--; /* Compensate for pPager->iDataVersion++; */
4105 pBt
->inTransaction
= TRANS_READ
;
4106 btreeClearHasContent(pBt
);
4109 btreeEndTransaction(p
);
4110 sqlite3BtreeLeave(p
);
4115 ** Do both phases of a commit.
4117 int sqlite3BtreeCommit(Btree
*p
){
4119 sqlite3BtreeEnter(p
);
4120 rc
= sqlite3BtreeCommitPhaseOne(p
, 0);
4121 if( rc
==SQLITE_OK
){
4122 rc
= sqlite3BtreeCommitPhaseTwo(p
, 0);
4124 sqlite3BtreeLeave(p
);
4129 ** This routine sets the state to CURSOR_FAULT and the error
4130 ** code to errCode for every cursor on any BtShared that pBtree
4131 ** references. Or if the writeOnly flag is set to 1, then only
4132 ** trip write cursors and leave read cursors unchanged.
4134 ** Every cursor is a candidate to be tripped, including cursors
4135 ** that belong to other database connections that happen to be
4136 ** sharing the cache with pBtree.
4138 ** This routine gets called when a rollback occurs. If the writeOnly
4139 ** flag is true, then only write-cursors need be tripped - read-only
4140 ** cursors save their current positions so that they may continue
4141 ** following the rollback. Or, if writeOnly is false, all cursors are
4142 ** tripped. In general, writeOnly is false if the transaction being
4143 ** rolled back modified the database schema. In this case b-tree root
4144 ** pages may be moved or deleted from the database altogether, making
4145 ** it unsafe for read cursors to continue.
4147 ** If the writeOnly flag is true and an error is encountered while
4148 ** saving the current position of a read-only cursor, all cursors,
4149 ** including all read-cursors are tripped.
4151 ** SQLITE_OK is returned if successful, or if an error occurs while
4152 ** saving a cursor position, an SQLite error code.
4154 int sqlite3BtreeTripAllCursors(Btree
*pBtree
, int errCode
, int writeOnly
){
4158 assert( (writeOnly
==0 || writeOnly
==1) && BTCF_WriteFlag
==1 );
4160 sqlite3BtreeEnter(pBtree
);
4161 for(p
=pBtree
->pBt
->pCursor
; p
; p
=p
->pNext
){
4162 if( writeOnly
&& (p
->curFlags
& BTCF_WriteFlag
)==0 ){
4163 if( p
->eState
==CURSOR_VALID
|| p
->eState
==CURSOR_SKIPNEXT
){
4164 rc
= saveCursorPosition(p
);
4165 if( rc
!=SQLITE_OK
){
4166 (void)sqlite3BtreeTripAllCursors(pBtree
, rc
, 0);
4171 sqlite3BtreeClearCursor(p
);
4172 p
->eState
= CURSOR_FAULT
;
4173 p
->skipNext
= errCode
;
4175 btreeReleaseAllCursorPages(p
);
4177 sqlite3BtreeLeave(pBtree
);
4183 ** Set the pBt->nPage field correctly, according to the current
4184 ** state of the database. Assume pBt->pPage1 is valid.
4186 static void btreeSetNPage(BtShared
*pBt
, MemPage
*pPage1
){
4187 int nPage
= get4byte(&pPage1
->aData
[28]);
4188 testcase( nPage
==0 );
4189 if( nPage
==0 ) sqlite3PagerPagecount(pBt
->pPager
, &nPage
);
4190 testcase( pBt
->nPage
!=nPage
);
4195 ** Rollback the transaction in progress.
4197 ** If tripCode is not SQLITE_OK then cursors will be invalidated (tripped).
4198 ** Only write cursors are tripped if writeOnly is true but all cursors are
4199 ** tripped if writeOnly is false. Any attempt to use
4200 ** a tripped cursor will result in an error.
4202 ** This will release the write lock on the database file. If there
4203 ** are no active cursors, it also releases the read lock.
4205 int sqlite3BtreeRollback(Btree
*p
, int tripCode
, int writeOnly
){
4207 BtShared
*pBt
= p
->pBt
;
4210 assert( writeOnly
==1 || writeOnly
==0 );
4211 assert( tripCode
==SQLITE_ABORT_ROLLBACK
|| tripCode
==SQLITE_OK
);
4212 sqlite3BtreeEnter(p
);
4213 if( tripCode
==SQLITE_OK
){
4214 rc
= tripCode
= saveAllCursors(pBt
, 0, 0);
4215 if( rc
) writeOnly
= 0;
4220 int rc2
= sqlite3BtreeTripAllCursors(p
, tripCode
, writeOnly
);
4221 assert( rc
==SQLITE_OK
|| (writeOnly
==0 && rc2
==SQLITE_OK
) );
4222 if( rc2
!=SQLITE_OK
) rc
= rc2
;
4226 if( p
->inTrans
==TRANS_WRITE
){
4229 assert( TRANS_WRITE
==pBt
->inTransaction
);
4230 rc2
= sqlite3PagerRollback(pBt
->pPager
);
4231 if( rc2
!=SQLITE_OK
){
4235 /* The rollback may have destroyed the pPage1->aData value. So
4236 ** call btreeGetPage() on page 1 again to make
4237 ** sure pPage1->aData is set correctly. */
4238 if( btreeGetPage(pBt
, 1, &pPage1
, 0)==SQLITE_OK
){
4239 btreeSetNPage(pBt
, pPage1
);
4240 releasePageOne(pPage1
);
4242 assert( countValidCursors(pBt
, 1)==0 );
4243 pBt
->inTransaction
= TRANS_READ
;
4244 btreeClearHasContent(pBt
);
4247 btreeEndTransaction(p
);
4248 sqlite3BtreeLeave(p
);
4253 ** Start a statement subtransaction. The subtransaction can be rolled
4254 ** back independently of the main transaction. You must start a transaction
4255 ** before starting a subtransaction. The subtransaction is ended automatically
4256 ** if the main transaction commits or rolls back.
4258 ** Statement subtransactions are used around individual SQL statements
4259 ** that are contained within a BEGIN...COMMIT block. If a constraint
4260 ** error occurs within the statement, the effect of that one statement
4261 ** can be rolled back without having to rollback the entire transaction.
4263 ** A statement sub-transaction is implemented as an anonymous savepoint. The
4264 ** value passed as the second parameter is the total number of savepoints,
4265 ** including the new anonymous savepoint, open on the B-Tree. i.e. if there
4266 ** are no active savepoints and no other statement-transactions open,
4267 ** iStatement is 1. This anonymous savepoint can be released or rolled back
4268 ** using the sqlite3BtreeSavepoint() function.
4270 int sqlite3BtreeBeginStmt(Btree
*p
, int iStatement
){
4272 BtShared
*pBt
= p
->pBt
;
4273 sqlite3BtreeEnter(p
);
4274 assert( p
->inTrans
==TRANS_WRITE
);
4275 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
4276 assert( iStatement
>0 );
4277 assert( iStatement
>p
->db
->nSavepoint
);
4278 assert( pBt
->inTransaction
==TRANS_WRITE
);
4279 /* At the pager level, a statement transaction is a savepoint with
4280 ** an index greater than all savepoints created explicitly using
4281 ** SQL statements. It is illegal to open, release or rollback any
4282 ** such savepoints while the statement transaction savepoint is active.
4284 rc
= sqlite3PagerOpenSavepoint(pBt
->pPager
, iStatement
);
4285 sqlite3BtreeLeave(p
);
4290 ** The second argument to this function, op, is always SAVEPOINT_ROLLBACK
4291 ** or SAVEPOINT_RELEASE. This function either releases or rolls back the
4292 ** savepoint identified by parameter iSavepoint, depending on the value
4295 ** Normally, iSavepoint is greater than or equal to zero. However, if op is
4296 ** SAVEPOINT_ROLLBACK, then iSavepoint may also be -1. In this case the
4297 ** contents of the entire transaction are rolled back. This is different
4298 ** from a normal transaction rollback, as no locks are released and the
4299 ** transaction remains open.
4301 int sqlite3BtreeSavepoint(Btree
*p
, int op
, int iSavepoint
){
4303 if( p
&& p
->inTrans
==TRANS_WRITE
){
4304 BtShared
*pBt
= p
->pBt
;
4305 assert( op
==SAVEPOINT_RELEASE
|| op
==SAVEPOINT_ROLLBACK
);
4306 assert( iSavepoint
>=0 || (iSavepoint
==-1 && op
==SAVEPOINT_ROLLBACK
) );
4307 sqlite3BtreeEnter(p
);
4308 if( op
==SAVEPOINT_ROLLBACK
){
4309 rc
= saveAllCursors(pBt
, 0, 0);
4311 if( rc
==SQLITE_OK
){
4312 rc
= sqlite3PagerSavepoint(pBt
->pPager
, op
, iSavepoint
);
4314 if( rc
==SQLITE_OK
){
4315 if( iSavepoint
<0 && (pBt
->btsFlags
& BTS_INITIALLY_EMPTY
)!=0 ){
4318 rc
= newDatabase(pBt
);
4319 btreeSetNPage(pBt
, pBt
->pPage1
);
4321 /* pBt->nPage might be zero if the database was corrupt when
4322 ** the transaction was started. Otherwise, it must be at least 1. */
4323 assert( CORRUPT_DB
|| pBt
->nPage
>0 );
4325 sqlite3BtreeLeave(p
);
4331 ** Create a new cursor for the BTree whose root is on the page
4332 ** iTable. If a read-only cursor is requested, it is assumed that
4333 ** the caller already has at least a read-only transaction open
4334 ** on the database already. If a write-cursor is requested, then
4335 ** the caller is assumed to have an open write transaction.
4337 ** If the BTREE_WRCSR bit of wrFlag is clear, then the cursor can only
4338 ** be used for reading. If the BTREE_WRCSR bit is set, then the cursor
4339 ** can be used for reading or for writing if other conditions for writing
4340 ** are also met. These are the conditions that must be met in order
4341 ** for writing to be allowed:
4343 ** 1: The cursor must have been opened with wrFlag containing BTREE_WRCSR
4345 ** 2: Other database connections that share the same pager cache
4346 ** but which are not in the READ_UNCOMMITTED state may not have
4347 ** cursors open with wrFlag==0 on the same table. Otherwise
4348 ** the changes made by this write cursor would be visible to
4349 ** the read cursors in the other database connection.
4351 ** 3: The database must be writable (not on read-only media)
4353 ** 4: There must be an active transaction.
4355 ** The BTREE_FORDELETE bit of wrFlag may optionally be set if BTREE_WRCSR
4356 ** is set. If FORDELETE is set, that is a hint to the implementation that
4357 ** this cursor will only be used to seek to and delete entries of an index
4358 ** as part of a larger DELETE statement. The FORDELETE hint is not used by
4359 ** this implementation. But in a hypothetical alternative storage engine
4360 ** in which index entries are automatically deleted when corresponding table
4361 ** rows are deleted, the FORDELETE flag is a hint that all SEEK and DELETE
4362 ** operations on this cursor can be no-ops and all READ operations can
4363 ** return a null row (2-bytes: 0x01 0x00).
4365 ** No checking is done to make sure that page iTable really is the
4366 ** root page of a b-tree. If it is not, then the cursor acquired
4367 ** will not work correctly.
4369 ** It is assumed that the sqlite3BtreeCursorZero() has been called
4370 ** on pCur to initialize the memory space prior to invoking this routine.
4372 static int btreeCursor(
4373 Btree
*p
, /* The btree */
4374 Pgno iTable
, /* Root page of table to open */
4375 int wrFlag
, /* 1 to write. 0 read-only */
4376 struct KeyInfo
*pKeyInfo
, /* First arg to comparison function */
4377 BtCursor
*pCur
/* Space for new cursor */
4379 BtShared
*pBt
= p
->pBt
; /* Shared b-tree handle */
4380 BtCursor
*pX
; /* Looping over other all cursors */
4382 assert( sqlite3BtreeHoldsMutex(p
) );
4384 || wrFlag
==BTREE_WRCSR
4385 || wrFlag
==(BTREE_WRCSR
|BTREE_FORDELETE
)
4388 /* The following assert statements verify that if this is a sharable
4389 ** b-tree database, the connection is holding the required table locks,
4390 ** and that no other connection has any open cursor that conflicts with
4391 ** this lock. The iTable<1 term disables the check for corrupt schemas. */
4392 assert( hasSharedCacheTableLock(p
, iTable
, pKeyInfo
!=0, (wrFlag
?2:1))
4394 assert( wrFlag
==0 || !hasReadConflicts(p
, iTable
) );
4396 /* Assert that the caller has opened the required transaction. */
4397 assert( p
->inTrans
>TRANS_NONE
);
4398 assert( wrFlag
==0 || p
->inTrans
==TRANS_WRITE
);
4399 assert( pBt
->pPage1
&& pBt
->pPage1
->aData
);
4400 assert( wrFlag
==0 || (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
4403 allocateTempSpace(pBt
);
4404 if( pBt
->pTmpSpace
==0 ) return SQLITE_NOMEM_BKPT
;
4408 return SQLITE_CORRUPT_BKPT
;
4409 }else if( btreePagecount(pBt
)==0 ){
4410 assert( wrFlag
==0 );
4415 /* Now that no other errors can occur, finish filling in the BtCursor
4416 ** variables and link the cursor into the BtShared list. */
4417 pCur
->pgnoRoot
= iTable
;
4419 pCur
->pKeyInfo
= pKeyInfo
;
4422 pCur
->curFlags
= wrFlag
? BTCF_WriteFlag
: 0;
4423 pCur
->curPagerFlags
= wrFlag
? 0 : PAGER_GET_READONLY
;
4424 /* If there are two or more cursors on the same btree, then all such
4425 ** cursors *must* have the BTCF_Multiple flag set. */
4426 for(pX
=pBt
->pCursor
; pX
; pX
=pX
->pNext
){
4427 if( pX
->pgnoRoot
==iTable
){
4428 pX
->curFlags
|= BTCF_Multiple
;
4429 pCur
->curFlags
|= BTCF_Multiple
;
4432 pCur
->pNext
= pBt
->pCursor
;
4433 pBt
->pCursor
= pCur
;
4434 pCur
->eState
= CURSOR_INVALID
;
4437 static int btreeCursorWithLock(
4438 Btree
*p
, /* The btree */
4439 Pgno iTable
, /* Root page of table to open */
4440 int wrFlag
, /* 1 to write. 0 read-only */
4441 struct KeyInfo
*pKeyInfo
, /* First arg to comparison function */
4442 BtCursor
*pCur
/* Space for new cursor */
4445 sqlite3BtreeEnter(p
);
4446 rc
= btreeCursor(p
, iTable
, wrFlag
, pKeyInfo
, pCur
);
4447 sqlite3BtreeLeave(p
);
4450 int sqlite3BtreeCursor(
4451 Btree
*p
, /* The btree */
4452 Pgno iTable
, /* Root page of table to open */
4453 int wrFlag
, /* 1 to write. 0 read-only */
4454 struct KeyInfo
*pKeyInfo
, /* First arg to xCompare() */
4455 BtCursor
*pCur
/* Write new cursor here */
4458 return btreeCursorWithLock(p
, iTable
, wrFlag
, pKeyInfo
, pCur
);
4460 return btreeCursor(p
, iTable
, wrFlag
, pKeyInfo
, pCur
);
4465 ** Return the size of a BtCursor object in bytes.
4467 ** This interfaces is needed so that users of cursors can preallocate
4468 ** sufficient storage to hold a cursor. The BtCursor object is opaque
4469 ** to users so they cannot do the sizeof() themselves - they must call
4472 int sqlite3BtreeCursorSize(void){
4473 return ROUND8(sizeof(BtCursor
));
4477 ** Initialize memory that will be converted into a BtCursor object.
4479 ** The simple approach here would be to memset() the entire object
4480 ** to zero. But it turns out that the apPage[] and aiIdx[] arrays
4481 ** do not need to be zeroed and they are large, so we can save a lot
4482 ** of run-time by skipping the initialization of those elements.
4484 void sqlite3BtreeCursorZero(BtCursor
*p
){
4485 memset(p
, 0, offsetof(BtCursor
, BTCURSOR_FIRST_UNINIT
));
4489 ** Close a cursor. The read lock on the database file is released
4490 ** when the last cursor is closed.
4492 int sqlite3BtreeCloseCursor(BtCursor
*pCur
){
4493 Btree
*pBtree
= pCur
->pBtree
;
4495 BtShared
*pBt
= pCur
->pBt
;
4496 sqlite3BtreeEnter(pBtree
);
4497 assert( pBt
->pCursor
!=0 );
4498 if( pBt
->pCursor
==pCur
){
4499 pBt
->pCursor
= pCur
->pNext
;
4501 BtCursor
*pPrev
= pBt
->pCursor
;
4503 if( pPrev
->pNext
==pCur
){
4504 pPrev
->pNext
= pCur
->pNext
;
4507 pPrev
= pPrev
->pNext
;
4508 }while( ALWAYS(pPrev
) );
4510 btreeReleaseAllCursorPages(pCur
);
4511 unlockBtreeIfUnused(pBt
);
4512 sqlite3_free(pCur
->aOverflow
);
4513 sqlite3_free(pCur
->pKey
);
4514 sqlite3BtreeLeave(pBtree
);
4521 ** Make sure the BtCursor* given in the argument has a valid
4522 ** BtCursor.info structure. If it is not already valid, call
4523 ** btreeParseCell() to fill it in.
4525 ** BtCursor.info is a cache of the information in the current cell.
4526 ** Using this cache reduces the number of calls to btreeParseCell().
4529 static int cellInfoEqual(CellInfo
*a
, CellInfo
*b
){
4530 if( a
->nKey
!=b
->nKey
) return 0;
4531 if( a
->pPayload
!=b
->pPayload
) return 0;
4532 if( a
->nPayload
!=b
->nPayload
) return 0;
4533 if( a
->nLocal
!=b
->nLocal
) return 0;
4534 if( a
->nSize
!=b
->nSize
) return 0;
4537 static void assertCellInfo(BtCursor
*pCur
){
4539 memset(&info
, 0, sizeof(info
));
4540 btreeParseCell(pCur
->pPage
, pCur
->ix
, &info
);
4541 assert( CORRUPT_DB
|| cellInfoEqual(&info
, &pCur
->info
) );
4544 #define assertCellInfo(x)
4546 static SQLITE_NOINLINE
void getCellInfo(BtCursor
*pCur
){
4547 if( pCur
->info
.nSize
==0 ){
4548 pCur
->curFlags
|= BTCF_ValidNKey
;
4549 btreeParseCell(pCur
->pPage
,pCur
->ix
,&pCur
->info
);
4551 assertCellInfo(pCur
);
4555 #ifndef NDEBUG /* The next routine used only within assert() statements */
4557 ** Return true if the given BtCursor is valid. A valid cursor is one
4558 ** that is currently pointing to a row in a (non-empty) table.
4559 ** This is a verification routine is used only within assert() statements.
4561 int sqlite3BtreeCursorIsValid(BtCursor
*pCur
){
4562 return pCur
&& pCur
->eState
==CURSOR_VALID
;
4565 int sqlite3BtreeCursorIsValidNN(BtCursor
*pCur
){
4567 return pCur
->eState
==CURSOR_VALID
;
4571 ** Return the value of the integer key or "rowid" for a table btree.
4572 ** This routine is only valid for a cursor that is pointing into a
4573 ** ordinary table btree. If the cursor points to an index btree or
4574 ** is invalid, the result of this routine is undefined.
4576 i64
sqlite3BtreeIntegerKey(BtCursor
*pCur
){
4577 assert( cursorHoldsMutex(pCur
) );
4578 assert( pCur
->eState
==CURSOR_VALID
);
4579 assert( pCur
->curIntKey
);
4581 return pCur
->info
.nKey
;
4585 ** Pin or unpin a cursor.
4587 void sqlite3BtreeCursorPin(BtCursor
*pCur
){
4588 assert( (pCur
->curFlags
& BTCF_Pinned
)==0 );
4589 pCur
->curFlags
|= BTCF_Pinned
;
4591 void sqlite3BtreeCursorUnpin(BtCursor
*pCur
){
4592 assert( (pCur
->curFlags
& BTCF_Pinned
)!=0 );
4593 pCur
->curFlags
&= ~BTCF_Pinned
;
4596 #ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC
4598 ** Return the offset into the database file for the start of the
4599 ** payload to which the cursor is pointing.
4601 i64
sqlite3BtreeOffset(BtCursor
*pCur
){
4602 assert( cursorHoldsMutex(pCur
) );
4603 assert( pCur
->eState
==CURSOR_VALID
);
4605 return (i64
)pCur
->pBt
->pageSize
*((i64
)pCur
->pPage
->pgno
- 1) +
4606 (i64
)(pCur
->info
.pPayload
- pCur
->pPage
->aData
);
4608 #endif /* SQLITE_ENABLE_OFFSET_SQL_FUNC */
4611 ** Return the number of bytes of payload for the entry that pCur is
4612 ** currently pointing to. For table btrees, this will be the amount
4613 ** of data. For index btrees, this will be the size of the key.
4615 ** The caller must guarantee that the cursor is pointing to a non-NULL
4616 ** valid entry. In other words, the calling procedure must guarantee
4617 ** that the cursor has Cursor.eState==CURSOR_VALID.
4619 u32
sqlite3BtreePayloadSize(BtCursor
*pCur
){
4620 assert( cursorHoldsMutex(pCur
) );
4621 assert( pCur
->eState
==CURSOR_VALID
);
4623 return pCur
->info
.nPayload
;
4627 ** Return an upper bound on the size of any record for the table
4628 ** that the cursor is pointing into.
4630 ** This is an optimization. Everything will still work if this
4631 ** routine always returns 2147483647 (which is the largest record
4632 ** that SQLite can handle) or more. But returning a smaller value might
4633 ** prevent large memory allocations when trying to interpret a
4634 ** corrupt datrabase.
4636 ** The current implementation merely returns the size of the underlying
4639 sqlite3_int64
sqlite3BtreeMaxRecordSize(BtCursor
*pCur
){
4640 assert( cursorHoldsMutex(pCur
) );
4641 assert( pCur
->eState
==CURSOR_VALID
);
4642 return pCur
->pBt
->pageSize
* (sqlite3_int64
)pCur
->pBt
->nPage
;
4646 ** Given the page number of an overflow page in the database (parameter
4647 ** ovfl), this function finds the page number of the next page in the
4648 ** linked list of overflow pages. If possible, it uses the auto-vacuum
4649 ** pointer-map data instead of reading the content of page ovfl to do so.
4651 ** If an error occurs an SQLite error code is returned. Otherwise:
4653 ** The page number of the next overflow page in the linked list is
4654 ** written to *pPgnoNext. If page ovfl is the last page in its linked
4655 ** list, *pPgnoNext is set to zero.
4657 ** If ppPage is not NULL, and a reference to the MemPage object corresponding
4658 ** to page number pOvfl was obtained, then *ppPage is set to point to that
4659 ** reference. It is the responsibility of the caller to call releasePage()
4660 ** on *ppPage to free the reference. In no reference was obtained (because
4661 ** the pointer-map was used to obtain the value for *pPgnoNext), then
4662 ** *ppPage is set to zero.
4664 static int getOverflowPage(
4665 BtShared
*pBt
, /* The database file */
4666 Pgno ovfl
, /* Current overflow page number */
4667 MemPage
**ppPage
, /* OUT: MemPage handle (may be NULL) */
4668 Pgno
*pPgnoNext
/* OUT: Next overflow page number */
4674 assert( sqlite3_mutex_held(pBt
->mutex
) );
4677 #ifndef SQLITE_OMIT_AUTOVACUUM
4678 /* Try to find the next page in the overflow list using the
4679 ** autovacuum pointer-map pages. Guess that the next page in
4680 ** the overflow list is page number (ovfl+1). If that guess turns
4681 ** out to be wrong, fall back to loading the data of page
4682 ** number ovfl to determine the next page number.
4684 if( pBt
->autoVacuum
){
4686 Pgno iGuess
= ovfl
+1;
4689 while( PTRMAP_ISPAGE(pBt
, iGuess
) || iGuess
==PENDING_BYTE_PAGE(pBt
) ){
4693 if( iGuess
<=btreePagecount(pBt
) ){
4694 rc
= ptrmapGet(pBt
, iGuess
, &eType
, &pgno
);
4695 if( rc
==SQLITE_OK
&& eType
==PTRMAP_OVERFLOW2
&& pgno
==ovfl
){
4703 assert( next
==0 || rc
==SQLITE_DONE
);
4704 if( rc
==SQLITE_OK
){
4705 rc
= btreeGetPage(pBt
, ovfl
, &pPage
, (ppPage
==0) ? PAGER_GET_READONLY
: 0);
4706 assert( rc
==SQLITE_OK
|| pPage
==0 );
4707 if( rc
==SQLITE_OK
){
4708 next
= get4byte(pPage
->aData
);
4718 return (rc
==SQLITE_DONE
? SQLITE_OK
: rc
);
4722 ** Copy data from a buffer to a page, or from a page to a buffer.
4724 ** pPayload is a pointer to data stored on database page pDbPage.
4725 ** If argument eOp is false, then nByte bytes of data are copied
4726 ** from pPayload to the buffer pointed at by pBuf. If eOp is true,
4727 ** then sqlite3PagerWrite() is called on pDbPage and nByte bytes
4728 ** of data are copied from the buffer pBuf to pPayload.
4730 ** SQLITE_OK is returned on success, otherwise an error code.
4732 static int copyPayload(
4733 void *pPayload
, /* Pointer to page data */
4734 void *pBuf
, /* Pointer to buffer */
4735 int nByte
, /* Number of bytes to copy */
4736 int eOp
, /* 0 -> copy from page, 1 -> copy to page */
4737 DbPage
*pDbPage
/* Page containing pPayload */
4740 /* Copy data from buffer to page (a write operation) */
4741 int rc
= sqlite3PagerWrite(pDbPage
);
4742 if( rc
!=SQLITE_OK
){
4745 memcpy(pPayload
, pBuf
, nByte
);
4747 /* Copy data from page to buffer (a read operation) */
4748 memcpy(pBuf
, pPayload
, nByte
);
4754 ** This function is used to read or overwrite payload information
4755 ** for the entry that the pCur cursor is pointing to. The eOp
4756 ** argument is interpreted as follows:
4758 ** 0: The operation is a read. Populate the overflow cache.
4759 ** 1: The operation is a write. Populate the overflow cache.
4761 ** A total of "amt" bytes are read or written beginning at "offset".
4762 ** Data is read to or from the buffer pBuf.
4764 ** The content being read or written might appear on the main page
4765 ** or be scattered out on multiple overflow pages.
4767 ** If the current cursor entry uses one or more overflow pages
4768 ** this function may allocate space for and lazily populate
4769 ** the overflow page-list cache array (BtCursor.aOverflow).
4770 ** Subsequent calls use this cache to make seeking to the supplied offset
4773 ** Once an overflow page-list cache has been allocated, it must be
4774 ** invalidated if some other cursor writes to the same table, or if
4775 ** the cursor is moved to a different row. Additionally, in auto-vacuum
4776 ** mode, the following events may invalidate an overflow page-list cache.
4778 ** * An incremental vacuum,
4779 ** * A commit in auto_vacuum="full" mode,
4780 ** * Creating a table (may require moving an overflow page).
4782 static int accessPayload(
4783 BtCursor
*pCur
, /* Cursor pointing to entry to read from */
4784 u32 offset
, /* Begin reading this far into payload */
4785 u32 amt
, /* Read this many bytes */
4786 unsigned char *pBuf
, /* Write the bytes into this buffer */
4787 int eOp
/* zero to read. non-zero to write. */
4789 unsigned char *aPayload
;
4792 MemPage
*pPage
= pCur
->pPage
; /* Btree page of current entry */
4793 BtShared
*pBt
= pCur
->pBt
; /* Btree this cursor belongs to */
4794 #ifdef SQLITE_DIRECT_OVERFLOW_READ
4795 unsigned char * const pBufStart
= pBuf
; /* Start of original out buffer */
4799 assert( eOp
==0 || eOp
==1 );
4800 assert( pCur
->eState
==CURSOR_VALID
);
4801 assert( pCur
->ix
<pPage
->nCell
);
4802 assert( cursorHoldsMutex(pCur
) );
4805 aPayload
= pCur
->info
.pPayload
;
4806 assert( offset
+amt
<= pCur
->info
.nPayload
);
4808 assert( aPayload
> pPage
->aData
);
4809 if( (uptr
)(aPayload
- pPage
->aData
) > (pBt
->usableSize
- pCur
->info
.nLocal
) ){
4810 /* Trying to read or write past the end of the data is an error. The
4811 ** conditional above is really:
4812 ** &aPayload[pCur->info.nLocal] > &pPage->aData[pBt->usableSize]
4813 ** but is recast into its current form to avoid integer overflow problems
4815 return SQLITE_CORRUPT_PAGE(pPage
);
4818 /* Check if data must be read/written to/from the btree page itself. */
4819 if( offset
<pCur
->info
.nLocal
){
4821 if( a
+offset
>pCur
->info
.nLocal
){
4822 a
= pCur
->info
.nLocal
- offset
;
4824 rc
= copyPayload(&aPayload
[offset
], pBuf
, a
, eOp
, pPage
->pDbPage
);
4829 offset
-= pCur
->info
.nLocal
;
4833 if( rc
==SQLITE_OK
&& amt
>0 ){
4834 const u32 ovflSize
= pBt
->usableSize
- 4; /* Bytes content per ovfl page */
4837 nextPage
= get4byte(&aPayload
[pCur
->info
.nLocal
]);
4839 /* If the BtCursor.aOverflow[] has not been allocated, allocate it now.
4841 ** The aOverflow[] array is sized at one entry for each overflow page
4842 ** in the overflow chain. The page number of the first overflow page is
4843 ** stored in aOverflow[0], etc. A value of 0 in the aOverflow[] array
4844 ** means "not yet known" (the cache is lazily populated).
4846 if( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 ){
4847 int nOvfl
= (pCur
->info
.nPayload
-pCur
->info
.nLocal
+ovflSize
-1)/ovflSize
;
4848 if( pCur
->aOverflow
==0
4849 || nOvfl
*(int)sizeof(Pgno
) > sqlite3MallocSize(pCur
->aOverflow
)
4851 Pgno
*aNew
= (Pgno
*)sqlite3Realloc(
4852 pCur
->aOverflow
, nOvfl
*2*sizeof(Pgno
)
4855 return SQLITE_NOMEM_BKPT
;
4857 pCur
->aOverflow
= aNew
;
4860 memset(pCur
->aOverflow
, 0, nOvfl
*sizeof(Pgno
));
4861 pCur
->curFlags
|= BTCF_ValidOvfl
;
4863 /* If the overflow page-list cache has been allocated and the
4864 ** entry for the first required overflow page is valid, skip
4867 if( pCur
->aOverflow
[offset
/ovflSize
] ){
4868 iIdx
= (offset
/ovflSize
);
4869 nextPage
= pCur
->aOverflow
[iIdx
];
4870 offset
= (offset
%ovflSize
);
4874 assert( rc
==SQLITE_OK
&& amt
>0 );
4876 /* If required, populate the overflow page-list cache. */
4877 if( nextPage
> pBt
->nPage
) return SQLITE_CORRUPT_BKPT
;
4878 assert( pCur
->aOverflow
[iIdx
]==0
4879 || pCur
->aOverflow
[iIdx
]==nextPage
4881 pCur
->aOverflow
[iIdx
] = nextPage
;
4883 if( offset
>=ovflSize
){
4884 /* The only reason to read this page is to obtain the page
4885 ** number for the next page in the overflow chain. The page
4886 ** data is not required. So first try to lookup the overflow
4887 ** page-list cache, if any, then fall back to the getOverflowPage()
4890 assert( pCur
->curFlags
& BTCF_ValidOvfl
);
4891 assert( pCur
->pBtree
->db
==pBt
->db
);
4892 if( pCur
->aOverflow
[iIdx
+1] ){
4893 nextPage
= pCur
->aOverflow
[iIdx
+1];
4895 rc
= getOverflowPage(pBt
, nextPage
, 0, &nextPage
);
4899 /* Need to read this page properly. It contains some of the
4900 ** range of data that is being read (eOp==0) or written (eOp!=0).
4903 if( a
+ offset
> ovflSize
){
4904 a
= ovflSize
- offset
;
4907 #ifdef SQLITE_DIRECT_OVERFLOW_READ
4908 /* If all the following are true:
4910 ** 1) this is a read operation, and
4911 ** 2) data is required from the start of this overflow page, and
4912 ** 3) there are no dirty pages in the page-cache
4913 ** 4) the database is file-backed, and
4914 ** 5) the page is not in the WAL file
4915 ** 6) at least 4 bytes have already been read into the output buffer
4917 ** then data can be read directly from the database file into the
4918 ** output buffer, bypassing the page-cache altogether. This speeds
4919 ** up loading large records that span many overflow pages.
4921 if( eOp
==0 /* (1) */
4922 && offset
==0 /* (2) */
4923 && sqlite3PagerDirectReadOk(pBt
->pPager
, nextPage
) /* (3,4,5) */
4924 && &pBuf
[-4]>=pBufStart
/* (6) */
4926 sqlite3_file
*fd
= sqlite3PagerFile(pBt
->pPager
);
4928 u8
*aWrite
= &pBuf
[-4];
4929 assert( aWrite
>=pBufStart
); /* due to (6) */
4930 memcpy(aSave
, aWrite
, 4);
4931 rc
= sqlite3OsRead(fd
, aWrite
, a
+4, (i64
)pBt
->pageSize
*(nextPage
-1));
4932 if( rc
&& nextPage
>pBt
->nPage
) rc
= SQLITE_CORRUPT_BKPT
;
4933 nextPage
= get4byte(aWrite
);
4934 memcpy(aWrite
, aSave
, 4);
4940 rc
= sqlite3PagerGet(pBt
->pPager
, nextPage
, &pDbPage
,
4941 (eOp
==0 ? PAGER_GET_READONLY
: 0)
4943 if( rc
==SQLITE_OK
){
4944 aPayload
= sqlite3PagerGetData(pDbPage
);
4945 nextPage
= get4byte(aPayload
);
4946 rc
= copyPayload(&aPayload
[offset
+4], pBuf
, a
, eOp
, pDbPage
);
4947 sqlite3PagerUnref(pDbPage
);
4952 if( amt
==0 ) return rc
;
4960 if( rc
==SQLITE_OK
&& amt
>0 ){
4961 /* Overflow chain ends prematurely */
4962 return SQLITE_CORRUPT_PAGE(pPage
);
4968 ** Read part of the payload for the row at which that cursor pCur is currently
4969 ** pointing. "amt" bytes will be transferred into pBuf[]. The transfer
4970 ** begins at "offset".
4972 ** pCur can be pointing to either a table or an index b-tree.
4973 ** If pointing to a table btree, then the content section is read. If
4974 ** pCur is pointing to an index b-tree then the key section is read.
4976 ** For sqlite3BtreePayload(), the caller must ensure that pCur is pointing
4977 ** to a valid row in the table. For sqlite3BtreePayloadChecked(), the
4978 ** cursor might be invalid or might need to be restored before being read.
4980 ** Return SQLITE_OK on success or an error code if anything goes
4981 ** wrong. An error is returned if "offset+amt" is larger than
4982 ** the available payload.
4984 int sqlite3BtreePayload(BtCursor
*pCur
, u32 offset
, u32 amt
, void *pBuf
){
4985 assert( cursorHoldsMutex(pCur
) );
4986 assert( pCur
->eState
==CURSOR_VALID
);
4987 assert( pCur
->iPage
>=0 && pCur
->pPage
);
4988 assert( pCur
->ix
<pCur
->pPage
->nCell
);
4989 return accessPayload(pCur
, offset
, amt
, (unsigned char*)pBuf
, 0);
4993 ** This variant of sqlite3BtreePayload() works even if the cursor has not
4994 ** in the CURSOR_VALID state. It is only used by the sqlite3_blob_read()
4997 #ifndef SQLITE_OMIT_INCRBLOB
4998 static SQLITE_NOINLINE
int accessPayloadChecked(
5005 if ( pCur
->eState
==CURSOR_INVALID
){
5006 return SQLITE_ABORT
;
5008 assert( cursorOwnsBtShared(pCur
) );
5009 rc
= btreeRestoreCursorPosition(pCur
);
5010 return rc
? rc
: accessPayload(pCur
, offset
, amt
, pBuf
, 0);
5012 int sqlite3BtreePayloadChecked(BtCursor
*pCur
, u32 offset
, u32 amt
, void *pBuf
){
5013 if( pCur
->eState
==CURSOR_VALID
){
5014 assert( cursorOwnsBtShared(pCur
) );
5015 return accessPayload(pCur
, offset
, amt
, pBuf
, 0);
5017 return accessPayloadChecked(pCur
, offset
, amt
, pBuf
);
5020 #endif /* SQLITE_OMIT_INCRBLOB */
5023 ** Return a pointer to payload information from the entry that the
5024 ** pCur cursor is pointing to. The pointer is to the beginning of
5025 ** the key if index btrees (pPage->intKey==0) and is the data for
5026 ** table btrees (pPage->intKey==1). The number of bytes of available
5027 ** key/data is written into *pAmt. If *pAmt==0, then the value
5028 ** returned will not be a valid pointer.
5030 ** This routine is an optimization. It is common for the entire key
5031 ** and data to fit on the local page and for there to be no overflow
5032 ** pages. When that is so, this routine can be used to access the
5033 ** key and data without making a copy. If the key and/or data spills
5034 ** onto overflow pages, then accessPayload() must be used to reassemble
5035 ** the key/data and copy it into a preallocated buffer.
5037 ** The pointer returned by this routine looks directly into the cached
5038 ** page of the database. The data might change or move the next time
5039 ** any btree routine is called.
5041 static const void *fetchPayload(
5042 BtCursor
*pCur
, /* Cursor pointing to entry to read from */
5043 u32
*pAmt
/* Write the number of available bytes here */
5046 assert( pCur
!=0 && pCur
->iPage
>=0 && pCur
->pPage
);
5047 assert( pCur
->eState
==CURSOR_VALID
);
5048 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5049 assert( cursorOwnsBtShared(pCur
) );
5050 assert( pCur
->ix
<pCur
->pPage
->nCell
);
5051 assert( pCur
->info
.nSize
>0 );
5052 assert( pCur
->info
.pPayload
>pCur
->pPage
->aData
|| CORRUPT_DB
);
5053 assert( pCur
->info
.pPayload
<pCur
->pPage
->aDataEnd
||CORRUPT_DB
);
5054 amt
= pCur
->info
.nLocal
;
5055 if( amt
>(int)(pCur
->pPage
->aDataEnd
- pCur
->info
.pPayload
) ){
5056 /* There is too little space on the page for the expected amount
5057 ** of local content. Database must be corrupt. */
5058 assert( CORRUPT_DB
);
5059 amt
= MAX(0, (int)(pCur
->pPage
->aDataEnd
- pCur
->info
.pPayload
));
5062 return (void*)pCur
->info
.pPayload
;
5067 ** For the entry that cursor pCur is point to, return as
5068 ** many bytes of the key or data as are available on the local
5069 ** b-tree page. Write the number of available bytes into *pAmt.
5071 ** The pointer returned is ephemeral. The key/data may move
5072 ** or be destroyed on the next call to any Btree routine,
5073 ** including calls from other threads against the same cache.
5074 ** Hence, a mutex on the BtShared should be held prior to calling
5077 ** These routines is used to get quick access to key and data
5078 ** in the common case where no overflow pages are used.
5080 const void *sqlite3BtreePayloadFetch(BtCursor
*pCur
, u32
*pAmt
){
5081 return fetchPayload(pCur
, pAmt
);
5086 ** Move the cursor down to a new child page. The newPgno argument is the
5087 ** page number of the child page to move to.
5089 ** This function returns SQLITE_CORRUPT if the page-header flags field of
5090 ** the new child page does not match the flags field of the parent (i.e.
5091 ** if an intkey page appears to be the parent of a non-intkey page, or
5094 static int moveToChild(BtCursor
*pCur
, u32 newPgno
){
5095 BtShared
*pBt
= pCur
->pBt
;
5097 assert( cursorOwnsBtShared(pCur
) );
5098 assert( pCur
->eState
==CURSOR_VALID
);
5099 assert( pCur
->iPage
<BTCURSOR_MAX_DEPTH
);
5100 assert( pCur
->iPage
>=0 );
5101 if( pCur
->iPage
>=(BTCURSOR_MAX_DEPTH
-1) ){
5102 return SQLITE_CORRUPT_BKPT
;
5104 pCur
->info
.nSize
= 0;
5105 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
5106 pCur
->aiIdx
[pCur
->iPage
] = pCur
->ix
;
5107 pCur
->apPage
[pCur
->iPage
] = pCur
->pPage
;
5110 return getAndInitPage(pBt
, newPgno
, &pCur
->pPage
, pCur
, pCur
->curPagerFlags
);
5115 ** Page pParent is an internal (non-leaf) tree page. This function
5116 ** asserts that page number iChild is the left-child if the iIdx'th
5117 ** cell in page pParent. Or, if iIdx is equal to the total number of
5118 ** cells in pParent, that page number iChild is the right-child of
5121 static void assertParentIndex(MemPage
*pParent
, int iIdx
, Pgno iChild
){
5122 if( CORRUPT_DB
) return; /* The conditions tested below might not be true
5123 ** in a corrupt database */
5124 assert( iIdx
<=pParent
->nCell
);
5125 if( iIdx
==pParent
->nCell
){
5126 assert( get4byte(&pParent
->aData
[pParent
->hdrOffset
+8])==iChild
);
5128 assert( get4byte(findCell(pParent
, iIdx
))==iChild
);
5132 # define assertParentIndex(x,y,z)
5136 ** Move the cursor up to the parent page.
5138 ** pCur->idx is set to the cell index that contains the pointer
5139 ** to the page we are coming from. If we are coming from the
5140 ** right-most child page then pCur->idx is set to one more than
5141 ** the largest cell index.
5143 static void moveToParent(BtCursor
*pCur
){
5145 assert( cursorOwnsBtShared(pCur
) );
5146 assert( pCur
->eState
==CURSOR_VALID
);
5147 assert( pCur
->iPage
>0 );
5148 assert( pCur
->pPage
);
5150 pCur
->apPage
[pCur
->iPage
-1],
5151 pCur
->aiIdx
[pCur
->iPage
-1],
5154 testcase( pCur
->aiIdx
[pCur
->iPage
-1] > pCur
->apPage
[pCur
->iPage
-1]->nCell
);
5155 pCur
->info
.nSize
= 0;
5156 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
5157 pCur
->ix
= pCur
->aiIdx
[pCur
->iPage
-1];
5158 pLeaf
= pCur
->pPage
;
5159 pCur
->pPage
= pCur
->apPage
[--pCur
->iPage
];
5160 releasePageNotNull(pLeaf
);
5164 ** Move the cursor to point to the root page of its b-tree structure.
5166 ** If the table has a virtual root page, then the cursor is moved to point
5167 ** to the virtual root page instead of the actual root page. A table has a
5168 ** virtual root page when the actual root page contains no cells and a
5169 ** single child page. This can only happen with the table rooted at page 1.
5171 ** If the b-tree structure is empty, the cursor state is set to
5172 ** CURSOR_INVALID and this routine returns SQLITE_EMPTY. Otherwise,
5173 ** the cursor is set to point to the first cell located on the root
5174 ** (or virtual root) page and the cursor state is set to CURSOR_VALID.
5176 ** If this function returns successfully, it may be assumed that the
5177 ** page-header flags indicate that the [virtual] root-page is the expected
5178 ** kind of b-tree page (i.e. if when opening the cursor the caller did not
5179 ** specify a KeyInfo structure the flags byte is set to 0x05 or 0x0D,
5180 ** indicating a table b-tree, or if the caller did specify a KeyInfo
5181 ** structure the flags byte is set to 0x02 or 0x0A, indicating an index
5184 static int moveToRoot(BtCursor
*pCur
){
5188 assert( cursorOwnsBtShared(pCur
) );
5189 assert( CURSOR_INVALID
< CURSOR_REQUIRESEEK
);
5190 assert( CURSOR_VALID
< CURSOR_REQUIRESEEK
);
5191 assert( CURSOR_FAULT
> CURSOR_REQUIRESEEK
);
5192 assert( pCur
->eState
< CURSOR_REQUIRESEEK
|| pCur
->iPage
<0 );
5193 assert( pCur
->pgnoRoot
>0 || pCur
->iPage
<0 );
5195 if( pCur
->iPage
>=0 ){
5197 releasePageNotNull(pCur
->pPage
);
5198 while( --pCur
->iPage
){
5199 releasePageNotNull(pCur
->apPage
[pCur
->iPage
]);
5201 pCur
->pPage
= pCur
->apPage
[0];
5204 }else if( pCur
->pgnoRoot
==0 ){
5205 pCur
->eState
= CURSOR_INVALID
;
5206 return SQLITE_EMPTY
;
5208 assert( pCur
->iPage
==(-1) );
5209 if( pCur
->eState
>=CURSOR_REQUIRESEEK
){
5210 if( pCur
->eState
==CURSOR_FAULT
){
5211 assert( pCur
->skipNext
!=SQLITE_OK
);
5212 return pCur
->skipNext
;
5214 sqlite3BtreeClearCursor(pCur
);
5216 rc
= getAndInitPage(pCur
->pBtree
->pBt
, pCur
->pgnoRoot
, &pCur
->pPage
,
5217 0, pCur
->curPagerFlags
);
5218 if( rc
!=SQLITE_OK
){
5219 pCur
->eState
= CURSOR_INVALID
;
5223 pCur
->curIntKey
= pCur
->pPage
->intKey
;
5225 pRoot
= pCur
->pPage
;
5226 assert( pRoot
->pgno
==pCur
->pgnoRoot
);
5228 /* If pCur->pKeyInfo is not NULL, then the caller that opened this cursor
5229 ** expected to open it on an index b-tree. Otherwise, if pKeyInfo is
5230 ** NULL, the caller expects a table b-tree. If this is not the case,
5231 ** return an SQLITE_CORRUPT error.
5233 ** Earlier versions of SQLite assumed that this test could not fail
5234 ** if the root page was already loaded when this function was called (i.e.
5235 ** if pCur->iPage>=0). But this is not so if the database is corrupted
5236 ** in such a way that page pRoot is linked into a second b-tree table
5237 ** (or the freelist). */
5238 assert( pRoot
->intKey
==1 || pRoot
->intKey
==0 );
5239 if( pRoot
->isInit
==0 || (pCur
->pKeyInfo
==0)!=pRoot
->intKey
){
5240 return SQLITE_CORRUPT_PAGE(pCur
->pPage
);
5245 pCur
->info
.nSize
= 0;
5246 pCur
->curFlags
&= ~(BTCF_AtLast
|BTCF_ValidNKey
|BTCF_ValidOvfl
);
5248 pRoot
= pCur
->pPage
;
5249 if( pRoot
->nCell
>0 ){
5250 pCur
->eState
= CURSOR_VALID
;
5251 }else if( !pRoot
->leaf
){
5253 if( pRoot
->pgno
!=1 ) return SQLITE_CORRUPT_BKPT
;
5254 subpage
= get4byte(&pRoot
->aData
[pRoot
->hdrOffset
+8]);
5255 pCur
->eState
= CURSOR_VALID
;
5256 rc
= moveToChild(pCur
, subpage
);
5258 pCur
->eState
= CURSOR_INVALID
;
5265 ** Move the cursor down to the left-most leaf entry beneath the
5266 ** entry to which it is currently pointing.
5268 ** The left-most leaf is the one with the smallest key - the first
5269 ** in ascending order.
5271 static int moveToLeftmost(BtCursor
*pCur
){
5276 assert( cursorOwnsBtShared(pCur
) );
5277 assert( pCur
->eState
==CURSOR_VALID
);
5278 while( rc
==SQLITE_OK
&& !(pPage
= pCur
->pPage
)->leaf
){
5279 assert( pCur
->ix
<pPage
->nCell
);
5280 pgno
= get4byte(findCell(pPage
, pCur
->ix
));
5281 rc
= moveToChild(pCur
, pgno
);
5287 ** Move the cursor down to the right-most leaf entry beneath the
5288 ** page to which it is currently pointing. Notice the difference
5289 ** between moveToLeftmost() and moveToRightmost(). moveToLeftmost()
5290 ** finds the left-most entry beneath the *entry* whereas moveToRightmost()
5291 ** finds the right-most entry beneath the *page*.
5293 ** The right-most entry is the one with the largest key - the last
5294 ** key in ascending order.
5296 static int moveToRightmost(BtCursor
*pCur
){
5301 assert( cursorOwnsBtShared(pCur
) );
5302 assert( pCur
->eState
==CURSOR_VALID
);
5303 while( !(pPage
= pCur
->pPage
)->leaf
){
5304 pgno
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
5305 pCur
->ix
= pPage
->nCell
;
5306 rc
= moveToChild(pCur
, pgno
);
5309 pCur
->ix
= pPage
->nCell
-1;
5310 assert( pCur
->info
.nSize
==0 );
5311 assert( (pCur
->curFlags
& BTCF_ValidNKey
)==0 );
5315 /* Move the cursor to the first entry in the table. Return SQLITE_OK
5316 ** on success. Set *pRes to 0 if the cursor actually points to something
5317 ** or set *pRes to 1 if the table is empty.
5319 int sqlite3BtreeFirst(BtCursor
*pCur
, int *pRes
){
5322 assert( cursorOwnsBtShared(pCur
) );
5323 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5324 rc
= moveToRoot(pCur
);
5325 if( rc
==SQLITE_OK
){
5326 assert( pCur
->pPage
->nCell
>0 );
5328 rc
= moveToLeftmost(pCur
);
5329 }else if( rc
==SQLITE_EMPTY
){
5330 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5337 /* Move the cursor to the last entry in the table. Return SQLITE_OK
5338 ** on success. Set *pRes to 0 if the cursor actually points to something
5339 ** or set *pRes to 1 if the table is empty.
5341 int sqlite3BtreeLast(BtCursor
*pCur
, int *pRes
){
5344 assert( cursorOwnsBtShared(pCur
) );
5345 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5347 /* If the cursor already points to the last entry, this is a no-op. */
5348 if( CURSOR_VALID
==pCur
->eState
&& (pCur
->curFlags
& BTCF_AtLast
)!=0 ){
5350 /* This block serves to assert() that the cursor really does point
5351 ** to the last entry in the b-tree. */
5353 for(ii
=0; ii
<pCur
->iPage
; ii
++){
5354 assert( pCur
->aiIdx
[ii
]==pCur
->apPage
[ii
]->nCell
);
5356 assert( pCur
->ix
==pCur
->pPage
->nCell
-1 );
5357 assert( pCur
->pPage
->leaf
);
5363 rc
= moveToRoot(pCur
);
5364 if( rc
==SQLITE_OK
){
5365 assert( pCur
->eState
==CURSOR_VALID
);
5367 rc
= moveToRightmost(pCur
);
5368 if( rc
==SQLITE_OK
){
5369 pCur
->curFlags
|= BTCF_AtLast
;
5371 pCur
->curFlags
&= ~BTCF_AtLast
;
5373 }else if( rc
==SQLITE_EMPTY
){
5374 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5381 /* Move the cursor so that it points to an entry near the key
5382 ** specified by pIdxKey or intKey. Return a success code.
5384 ** For INTKEY tables, the intKey parameter is used. pIdxKey
5385 ** must be NULL. For index tables, pIdxKey is used and intKey
5388 ** If an exact match is not found, then the cursor is always
5389 ** left pointing at a leaf page which would hold the entry if it
5390 ** were present. The cursor might point to an entry that comes
5391 ** before or after the key.
5393 ** An integer is written into *pRes which is the result of
5394 ** comparing the key with the entry to which the cursor is
5395 ** pointing. The meaning of the integer written into
5396 ** *pRes is as follows:
5398 ** *pRes<0 The cursor is left pointing at an entry that
5399 ** is smaller than intKey/pIdxKey or if the table is empty
5400 ** and the cursor is therefore left point to nothing.
5402 ** *pRes==0 The cursor is left pointing at an entry that
5403 ** exactly matches intKey/pIdxKey.
5405 ** *pRes>0 The cursor is left pointing at an entry that
5406 ** is larger than intKey/pIdxKey.
5408 ** For index tables, the pIdxKey->eqSeen field is set to 1 if there
5409 ** exists an entry in the table that exactly matches pIdxKey.
5411 int sqlite3BtreeMovetoUnpacked(
5412 BtCursor
*pCur
, /* The cursor to be moved */
5413 UnpackedRecord
*pIdxKey
, /* Unpacked index key */
5414 i64 intKey
, /* The table key */
5415 int biasRight
, /* If true, bias the search to the high end */
5416 int *pRes
/* Write search results here */
5419 RecordCompare xRecordCompare
;
5421 assert( cursorOwnsBtShared(pCur
) );
5422 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5424 assert( (pIdxKey
==0)==(pCur
->pKeyInfo
==0) );
5425 assert( pCur
->eState
!=CURSOR_VALID
|| (pIdxKey
==0)==(pCur
->curIntKey
!=0) );
5427 /* If the cursor is already positioned at the point we are trying
5428 ** to move to, then just return without doing any work */
5430 && pCur
->eState
==CURSOR_VALID
&& (pCur
->curFlags
& BTCF_ValidNKey
)!=0
5432 if( pCur
->info
.nKey
==intKey
){
5436 if( pCur
->info
.nKey
<intKey
){
5437 if( (pCur
->curFlags
& BTCF_AtLast
)!=0 ){
5441 /* If the requested key is one more than the previous key, then
5442 ** try to get there using sqlite3BtreeNext() rather than a full
5443 ** binary search. This is an optimization only. The correct answer
5444 ** is still obtained without this case, only a little more slowely */
5445 if( pCur
->info
.nKey
+1==intKey
){
5447 rc
= sqlite3BtreeNext(pCur
, 0);
5448 if( rc
==SQLITE_OK
){
5450 if( pCur
->info
.nKey
==intKey
){
5453 }else if( rc
==SQLITE_DONE
){
5463 xRecordCompare
= sqlite3VdbeFindCompare(pIdxKey
);
5464 pIdxKey
->errCode
= 0;
5465 assert( pIdxKey
->default_rc
==1
5466 || pIdxKey
->default_rc
==0
5467 || pIdxKey
->default_rc
==-1
5470 xRecordCompare
= 0; /* All keys are integers */
5473 rc
= moveToRoot(pCur
);
5475 if( rc
==SQLITE_EMPTY
){
5476 assert( pCur
->pgnoRoot
==0 || pCur
->pPage
->nCell
==0 );
5482 assert( pCur
->pPage
);
5483 assert( pCur
->pPage
->isInit
);
5484 assert( pCur
->eState
==CURSOR_VALID
);
5485 assert( pCur
->pPage
->nCell
> 0 );
5486 assert( pCur
->iPage
==0 || pCur
->apPage
[0]->intKey
==pCur
->curIntKey
);
5487 assert( pCur
->curIntKey
|| pIdxKey
);
5489 int lwr
, upr
, idx
, c
;
5491 MemPage
*pPage
= pCur
->pPage
;
5492 u8
*pCell
; /* Pointer to current cell in pPage */
5494 /* pPage->nCell must be greater than zero. If this is the root-page
5495 ** the cursor would have been INVALID above and this for(;;) loop
5496 ** not run. If this is not the root-page, then the moveToChild() routine
5497 ** would have already detected db corruption. Similarly, pPage must
5498 ** be the right kind (index or table) of b-tree page. Otherwise
5499 ** a moveToChild() or moveToRoot() call would have detected corruption. */
5500 assert( pPage
->nCell
>0 );
5501 assert( pPage
->intKey
==(pIdxKey
==0) );
5503 upr
= pPage
->nCell
-1;
5504 assert( biasRight
==0 || biasRight
==1 );
5505 idx
= upr
>>(1-biasRight
); /* idx = biasRight ? upr : (lwr+upr)/2; */
5506 pCur
->ix
= (u16
)idx
;
5507 if( xRecordCompare
==0 ){
5510 pCell
= findCellPastPtr(pPage
, idx
);
5511 if( pPage
->intKeyLeaf
){
5512 while( 0x80 <= *(pCell
++) ){
5513 if( pCell
>=pPage
->aDataEnd
){
5514 return SQLITE_CORRUPT_PAGE(pPage
);
5518 getVarint(pCell
, (u64
*)&nCellKey
);
5519 if( nCellKey
<intKey
){
5521 if( lwr
>upr
){ c
= -1; break; }
5522 }else if( nCellKey
>intKey
){
5524 if( lwr
>upr
){ c
= +1; break; }
5526 assert( nCellKey
==intKey
);
5527 pCur
->ix
= (u16
)idx
;
5530 goto moveto_next_layer
;
5532 pCur
->curFlags
|= BTCF_ValidNKey
;
5533 pCur
->info
.nKey
= nCellKey
;
5534 pCur
->info
.nSize
= 0;
5539 assert( lwr
+upr
>=0 );
5540 idx
= (lwr
+upr
)>>1; /* idx = (lwr+upr)/2; */
5544 int nCell
; /* Size of the pCell cell in bytes */
5545 pCell
= findCellPastPtr(pPage
, idx
);
5547 /* The maximum supported page-size is 65536 bytes. This means that
5548 ** the maximum number of record bytes stored on an index B-Tree
5549 ** page is less than 16384 bytes and may be stored as a 2-byte
5550 ** varint. This information is used to attempt to avoid parsing
5551 ** the entire cell by checking for the cases where the record is
5552 ** stored entirely within the b-tree page by inspecting the first
5553 ** 2 bytes of the cell.
5556 if( nCell
<=pPage
->max1bytePayload
){
5557 /* This branch runs if the record-size field of the cell is a
5558 ** single byte varint and the record fits entirely on the main
5560 testcase( pCell
+nCell
+1==pPage
->aDataEnd
);
5561 c
= xRecordCompare(nCell
, (void*)&pCell
[1], pIdxKey
);
5562 }else if( !(pCell
[1] & 0x80)
5563 && (nCell
= ((nCell
&0x7f)<<7) + pCell
[1])<=pPage
->maxLocal
5565 /* The record-size field is a 2 byte varint and the record
5566 ** fits entirely on the main b-tree page. */
5567 testcase( pCell
+nCell
+2==pPage
->aDataEnd
);
5568 c
= xRecordCompare(nCell
, (void*)&pCell
[2], pIdxKey
);
5570 /* The record flows over onto one or more overflow pages. In
5571 ** this case the whole cell needs to be parsed, a buffer allocated
5572 ** and accessPayload() used to retrieve the record into the
5573 ** buffer before VdbeRecordCompare() can be called.
5575 ** If the record is corrupt, the xRecordCompare routine may read
5576 ** up to two varints past the end of the buffer. An extra 18
5577 ** bytes of padding is allocated at the end of the buffer in
5578 ** case this happens. */
5580 u8
* const pCellBody
= pCell
- pPage
->childPtrSize
;
5581 const int nOverrun
= 18; /* Size of the overrun padding */
5582 pPage
->xParseCell(pPage
, pCellBody
, &pCur
->info
);
5583 nCell
= (int)pCur
->info
.nKey
;
5584 testcase( nCell
<0 ); /* True if key size is 2^32 or more */
5585 testcase( nCell
==0 ); /* Invalid key size: 0x80 0x80 0x00 */
5586 testcase( nCell
==1 ); /* Invalid key size: 0x80 0x80 0x01 */
5587 testcase( nCell
==2 ); /* Minimum legal index key size */
5588 if( nCell
<2 || nCell
/pCur
->pBt
->usableSize
>pCur
->pBt
->nPage
){
5589 rc
= SQLITE_CORRUPT_PAGE(pPage
);
5592 pCellKey
= sqlite3Malloc( nCell
+nOverrun
);
5594 rc
= SQLITE_NOMEM_BKPT
;
5597 pCur
->ix
= (u16
)idx
;
5598 rc
= accessPayload(pCur
, 0, nCell
, (unsigned char*)pCellKey
, 0);
5599 memset(((u8
*)pCellKey
)+nCell
,0,nOverrun
); /* Fix uninit warnings */
5600 pCur
->curFlags
&= ~BTCF_ValidOvfl
;
5602 sqlite3_free(pCellKey
);
5605 c
= sqlite3VdbeRecordCompare(nCell
, pCellKey
, pIdxKey
);
5606 sqlite3_free(pCellKey
);
5609 (pIdxKey
->errCode
!=SQLITE_CORRUPT
|| c
==0)
5610 && (pIdxKey
->errCode
!=SQLITE_NOMEM
|| pCur
->pBtree
->db
->mallocFailed
)
5620 pCur
->ix
= (u16
)idx
;
5621 if( pIdxKey
->errCode
) rc
= SQLITE_CORRUPT_BKPT
;
5624 if( lwr
>upr
) break;
5625 assert( lwr
+upr
>=0 );
5626 idx
= (lwr
+upr
)>>1; /* idx = (lwr+upr)/2 */
5629 assert( lwr
==upr
+1 || (pPage
->intKey
&& !pPage
->leaf
) );
5630 assert( pPage
->isInit
);
5632 assert( pCur
->ix
<pCur
->pPage
->nCell
);
5633 pCur
->ix
= (u16
)idx
;
5639 if( lwr
>=pPage
->nCell
){
5640 chldPg
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
5642 chldPg
= get4byte(findCell(pPage
, lwr
));
5644 pCur
->ix
= (u16
)lwr
;
5645 rc
= moveToChild(pCur
, chldPg
);
5649 pCur
->info
.nSize
= 0;
5650 assert( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 );
5656 ** Return TRUE if the cursor is not pointing at an entry of the table.
5658 ** TRUE will be returned after a call to sqlite3BtreeNext() moves
5659 ** past the last entry in the table or sqlite3BtreePrev() moves past
5660 ** the first entry. TRUE is also returned if the table is empty.
5662 int sqlite3BtreeEof(BtCursor
*pCur
){
5663 /* TODO: What if the cursor is in CURSOR_REQUIRESEEK but all table entries
5664 ** have been deleted? This API will need to change to return an error code
5665 ** as well as the boolean result value.
5667 return (CURSOR_VALID
!=pCur
->eState
);
5671 ** Return an estimate for the number of rows in the table that pCur is
5672 ** pointing to. Return a negative number if no estimate is currently
5675 i64
sqlite3BtreeRowCountEst(BtCursor
*pCur
){
5679 assert( cursorOwnsBtShared(pCur
) );
5680 assert( sqlite3_mutex_held(pCur
->pBtree
->db
->mutex
) );
5682 /* Currently this interface is only called by the OP_IfSmaller
5683 ** opcode, and it that case the cursor will always be valid and
5684 ** will always point to a leaf node. */
5685 if( NEVER(pCur
->eState
!=CURSOR_VALID
) ) return -1;
5686 if( NEVER(pCur
->pPage
->leaf
==0) ) return -1;
5688 n
= pCur
->pPage
->nCell
;
5689 for(i
=0; i
<pCur
->iPage
; i
++){
5690 n
*= pCur
->apPage
[i
]->nCell
;
5696 ** Advance the cursor to the next entry in the database.
5699 ** SQLITE_OK success
5700 ** SQLITE_DONE cursor is already pointing at the last element
5701 ** otherwise some kind of error occurred
5703 ** The main entry point is sqlite3BtreeNext(). That routine is optimized
5704 ** for the common case of merely incrementing the cell counter BtCursor.aiIdx
5705 ** to the next cell on the current page. The (slower) btreeNext() helper
5706 ** routine is called when it is necessary to move to a different page or
5707 ** to restore the cursor.
5709 ** If bit 0x01 of the F argument in sqlite3BtreeNext(C,F) is 1, then the
5710 ** cursor corresponds to an SQL index and this routine could have been
5711 ** skipped if the SQL index had been a unique index. The F argument
5712 ** is a hint to the implement. SQLite btree implementation does not use
5713 ** this hint, but COMDB2 does.
5715 static SQLITE_NOINLINE
int btreeNext(BtCursor
*pCur
){
5720 assert( cursorOwnsBtShared(pCur
) );
5721 if( pCur
->eState
!=CURSOR_VALID
){
5722 assert( (pCur
->curFlags
& BTCF_ValidOvfl
)==0 );
5723 rc
= restoreCursorPosition(pCur
);
5724 if( rc
!=SQLITE_OK
){
5727 if( CURSOR_INVALID
==pCur
->eState
){
5730 if( pCur
->eState
==CURSOR_SKIPNEXT
){
5731 pCur
->eState
= CURSOR_VALID
;
5732 if( pCur
->skipNext
>0 ) return SQLITE_OK
;
5736 pPage
= pCur
->pPage
;
5738 if( !pPage
->isInit
){
5739 /* The only known way for this to happen is for there to be a
5740 ** recursive SQL function that does a DELETE operation as part of a
5741 ** SELECT which deletes content out from under an active cursor
5742 ** in a corrupt database file where the table being DELETE-ed from
5743 ** has pages in common with the table being queried. See TH3
5744 ** module cov1/btree78.test testcase 220 (2018-06-08) for an
5746 return SQLITE_CORRUPT_BKPT
;
5749 /* If the database file is corrupt, it is possible for the value of idx
5750 ** to be invalid here. This can only occur if a second cursor modifies
5751 ** the page while cursor pCur is holding a reference to it. Which can
5752 ** only happen if the database is corrupt in such a way as to link the
5753 ** page into more than one b-tree structure.
5755 ** Update 2019-12-23: appears to long longer be possible after the
5756 ** addition of anotherValidCursor() condition on balance_deeper(). */
5757 harmless( idx
>pPage
->nCell
);
5759 if( idx
>=pPage
->nCell
){
5761 rc
= moveToChild(pCur
, get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]));
5763 return moveToLeftmost(pCur
);
5766 if( pCur
->iPage
==0 ){
5767 pCur
->eState
= CURSOR_INVALID
;
5771 pPage
= pCur
->pPage
;
5772 }while( pCur
->ix
>=pPage
->nCell
);
5773 if( pPage
->intKey
){
5774 return sqlite3BtreeNext(pCur
, 0);
5782 return moveToLeftmost(pCur
);
5785 int sqlite3BtreeNext(BtCursor
*pCur
, int flags
){
5787 UNUSED_PARAMETER( flags
); /* Used in COMDB2 but not native SQLite */
5788 assert( cursorOwnsBtShared(pCur
) );
5789 assert( flags
==0 || flags
==1 );
5790 pCur
->info
.nSize
= 0;
5791 pCur
->curFlags
&= ~(BTCF_ValidNKey
|BTCF_ValidOvfl
);
5792 if( pCur
->eState
!=CURSOR_VALID
) return btreeNext(pCur
);
5793 pPage
= pCur
->pPage
;
5794 if( (++pCur
->ix
)>=pPage
->nCell
){
5796 return btreeNext(pCur
);
5801 return moveToLeftmost(pCur
);
5806 ** Step the cursor to the back to the previous entry in the database.
5809 ** SQLITE_OK success
5810 ** SQLITE_DONE the cursor is already on the first element of the table
5811 ** otherwise some kind of error occurred
5813 ** The main entry point is sqlite3BtreePrevious(). That routine is optimized
5814 ** for the common case of merely decrementing the cell counter BtCursor.aiIdx
5815 ** to the previous cell on the current page. The (slower) btreePrevious()
5816 ** helper routine is called when it is necessary to move to a different page
5817 ** or to restore the cursor.
5819 ** If bit 0x01 of the F argument to sqlite3BtreePrevious(C,F) is 1, then
5820 ** the cursor corresponds to an SQL index and this routine could have been
5821 ** skipped if the SQL index had been a unique index. The F argument is a
5822 ** hint to the implement. The native SQLite btree implementation does not
5823 ** use this hint, but COMDB2 does.
5825 static SQLITE_NOINLINE
int btreePrevious(BtCursor
*pCur
){
5829 assert( cursorOwnsBtShared(pCur
) );
5830 assert( (pCur
->curFlags
& (BTCF_AtLast
|BTCF_ValidOvfl
|BTCF_ValidNKey
))==0 );
5831 assert( pCur
->info
.nSize
==0 );
5832 if( pCur
->eState
!=CURSOR_VALID
){
5833 rc
= restoreCursorPosition(pCur
);
5834 if( rc
!=SQLITE_OK
){
5837 if( CURSOR_INVALID
==pCur
->eState
){
5840 if( CURSOR_SKIPNEXT
==pCur
->eState
){
5841 pCur
->eState
= CURSOR_VALID
;
5842 if( pCur
->skipNext
<0 ) return SQLITE_OK
;
5846 pPage
= pCur
->pPage
;
5847 assert( pPage
->isInit
);
5850 rc
= moveToChild(pCur
, get4byte(findCell(pPage
, idx
)));
5852 rc
= moveToRightmost(pCur
);
5854 while( pCur
->ix
==0 ){
5855 if( pCur
->iPage
==0 ){
5856 pCur
->eState
= CURSOR_INVALID
;
5861 assert( pCur
->info
.nSize
==0 );
5862 assert( (pCur
->curFlags
& (BTCF_ValidOvfl
))==0 );
5865 pPage
= pCur
->pPage
;
5866 if( pPage
->intKey
&& !pPage
->leaf
){
5867 rc
= sqlite3BtreePrevious(pCur
, 0);
5874 int sqlite3BtreePrevious(BtCursor
*pCur
, int flags
){
5875 assert( cursorOwnsBtShared(pCur
) );
5876 assert( flags
==0 || flags
==1 );
5877 UNUSED_PARAMETER( flags
); /* Used in COMDB2 but not native SQLite */
5878 pCur
->curFlags
&= ~(BTCF_AtLast
|BTCF_ValidOvfl
|BTCF_ValidNKey
);
5879 pCur
->info
.nSize
= 0;
5880 if( pCur
->eState
!=CURSOR_VALID
5882 || pCur
->pPage
->leaf
==0
5884 return btreePrevious(pCur
);
5891 ** Allocate a new page from the database file.
5893 ** The new page is marked as dirty. (In other words, sqlite3PagerWrite()
5894 ** has already been called on the new page.) The new page has also
5895 ** been referenced and the calling routine is responsible for calling
5896 ** sqlite3PagerUnref() on the new page when it is done.
5898 ** SQLITE_OK is returned on success. Any other return value indicates
5899 ** an error. *ppPage is set to NULL in the event of an error.
5901 ** If the "nearby" parameter is not 0, then an effort is made to
5902 ** locate a page close to the page number "nearby". This can be used in an
5903 ** attempt to keep related pages close to each other in the database file,
5904 ** which in turn can make database access faster.
5906 ** If the eMode parameter is BTALLOC_EXACT and the nearby page exists
5907 ** anywhere on the free-list, then it is guaranteed to be returned. If
5908 ** eMode is BTALLOC_LT then the page returned will be less than or equal
5909 ** to nearby if any such page exists. If eMode is BTALLOC_ANY then there
5910 ** are no restrictions on which page is returned.
5912 static int allocateBtreePage(
5913 BtShared
*pBt
, /* The btree */
5914 MemPage
**ppPage
, /* Store pointer to the allocated page here */
5915 Pgno
*pPgno
, /* Store the page number here */
5916 Pgno nearby
, /* Search for a page near this one */
5917 u8 eMode
/* BTALLOC_EXACT, BTALLOC_LT, or BTALLOC_ANY */
5921 u32 n
; /* Number of pages on the freelist */
5922 u32 k
; /* Number of leaves on the trunk of the freelist */
5923 MemPage
*pTrunk
= 0;
5924 MemPage
*pPrevTrunk
= 0;
5925 Pgno mxPage
; /* Total size of the database file */
5927 assert( sqlite3_mutex_held(pBt
->mutex
) );
5928 assert( eMode
==BTALLOC_ANY
|| (nearby
>0 && IfNotOmitAV(pBt
->autoVacuum
)) );
5929 pPage1
= pBt
->pPage1
;
5930 mxPage
= btreePagecount(pBt
);
5931 /* EVIDENCE-OF: R-05119-02637 The 4-byte big-endian integer at offset 36
5932 ** stores stores the total number of pages on the freelist. */
5933 n
= get4byte(&pPage1
->aData
[36]);
5934 testcase( n
==mxPage
-1 );
5936 return SQLITE_CORRUPT_BKPT
;
5939 /* There are pages on the freelist. Reuse one of those pages. */
5941 u8 searchList
= 0; /* If the free-list must be searched for 'nearby' */
5942 u32 nSearch
= 0; /* Count of the number of search attempts */
5944 /* If eMode==BTALLOC_EXACT and a query of the pointer-map
5945 ** shows that the page 'nearby' is somewhere on the free-list, then
5946 ** the entire-list will be searched for that page.
5948 #ifndef SQLITE_OMIT_AUTOVACUUM
5949 if( eMode
==BTALLOC_EXACT
){
5950 if( nearby
<=mxPage
){
5953 assert( pBt
->autoVacuum
);
5954 rc
= ptrmapGet(pBt
, nearby
, &eType
, 0);
5956 if( eType
==PTRMAP_FREEPAGE
){
5960 }else if( eMode
==BTALLOC_LE
){
5965 /* Decrement the free-list count by 1. Set iTrunk to the index of the
5966 ** first free-list trunk page. iPrevTrunk is initially 1.
5968 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
5970 put4byte(&pPage1
->aData
[36], n
-1);
5972 /* The code within this loop is run only once if the 'searchList' variable
5973 ** is not true. Otherwise, it runs once for each trunk-page on the
5974 ** free-list until the page 'nearby' is located (eMode==BTALLOC_EXACT)
5975 ** or until a page less than 'nearby' is located (eMode==BTALLOC_LT)
5978 pPrevTrunk
= pTrunk
;
5980 /* EVIDENCE-OF: R-01506-11053 The first integer on a freelist trunk page
5981 ** is the page number of the next freelist trunk page in the list or
5982 ** zero if this is the last freelist trunk page. */
5983 iTrunk
= get4byte(&pPrevTrunk
->aData
[0]);
5985 /* EVIDENCE-OF: R-59841-13798 The 4-byte big-endian integer at offset 32
5986 ** stores the page number of the first page of the freelist, or zero if
5987 ** the freelist is empty. */
5988 iTrunk
= get4byte(&pPage1
->aData
[32]);
5990 testcase( iTrunk
==mxPage
);
5991 if( iTrunk
>mxPage
|| nSearch
++ > n
){
5992 rc
= SQLITE_CORRUPT_PGNO(pPrevTrunk
? pPrevTrunk
->pgno
: 1);
5994 rc
= btreeGetUnusedPage(pBt
, iTrunk
, &pTrunk
, 0);
5998 goto end_allocate_page
;
6000 assert( pTrunk
!=0 );
6001 assert( pTrunk
->aData
!=0 );
6002 /* EVIDENCE-OF: R-13523-04394 The second integer on a freelist trunk page
6003 ** is the number of leaf page pointers to follow. */
6004 k
= get4byte(&pTrunk
->aData
[4]);
6005 if( k
==0 && !searchList
){
6006 /* The trunk has no leaves and the list is not being searched.
6007 ** So extract the trunk page itself and use it as the newly
6008 ** allocated page */
6009 assert( pPrevTrunk
==0 );
6010 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
6012 goto end_allocate_page
;
6015 memcpy(&pPage1
->aData
[32], &pTrunk
->aData
[0], 4);
6018 TRACE(("ALLOCATE: %d trunk - %d free pages left\n", *pPgno
, n
-1));
6019 }else if( k
>(u32
)(pBt
->usableSize
/4 - 2) ){
6020 /* Value of k is out of range. Database corruption */
6021 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
6022 goto end_allocate_page
;
6023 #ifndef SQLITE_OMIT_AUTOVACUUM
6024 }else if( searchList
6025 && (nearby
==iTrunk
|| (iTrunk
<nearby
&& eMode
==BTALLOC_LE
))
6027 /* The list is being searched and this trunk page is the page
6028 ** to allocate, regardless of whether it has leaves.
6033 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
6035 goto end_allocate_page
;
6039 memcpy(&pPage1
->aData
[32], &pTrunk
->aData
[0], 4);
6041 rc
= sqlite3PagerWrite(pPrevTrunk
->pDbPage
);
6042 if( rc
!=SQLITE_OK
){
6043 goto end_allocate_page
;
6045 memcpy(&pPrevTrunk
->aData
[0], &pTrunk
->aData
[0], 4);
6048 /* The trunk page is required by the caller but it contains
6049 ** pointers to free-list leaves. The first leaf becomes a trunk
6050 ** page in this case.
6053 Pgno iNewTrunk
= get4byte(&pTrunk
->aData
[8]);
6054 if( iNewTrunk
>mxPage
){
6055 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
6056 goto end_allocate_page
;
6058 testcase( iNewTrunk
==mxPage
);
6059 rc
= btreeGetUnusedPage(pBt
, iNewTrunk
, &pNewTrunk
, 0);
6060 if( rc
!=SQLITE_OK
){
6061 goto end_allocate_page
;
6063 rc
= sqlite3PagerWrite(pNewTrunk
->pDbPage
);
6064 if( rc
!=SQLITE_OK
){
6065 releasePage(pNewTrunk
);
6066 goto end_allocate_page
;
6068 memcpy(&pNewTrunk
->aData
[0], &pTrunk
->aData
[0], 4);
6069 put4byte(&pNewTrunk
->aData
[4], k
-1);
6070 memcpy(&pNewTrunk
->aData
[8], &pTrunk
->aData
[12], (k
-1)*4);
6071 releasePage(pNewTrunk
);
6073 assert( sqlite3PagerIswriteable(pPage1
->pDbPage
) );
6074 put4byte(&pPage1
->aData
[32], iNewTrunk
);
6076 rc
= sqlite3PagerWrite(pPrevTrunk
->pDbPage
);
6078 goto end_allocate_page
;
6080 put4byte(&pPrevTrunk
->aData
[0], iNewTrunk
);
6084 TRACE(("ALLOCATE: %d trunk - %d free pages left\n", *pPgno
, n
-1));
6087 /* Extract a leaf from the trunk */
6090 unsigned char *aData
= pTrunk
->aData
;
6094 if( eMode
==BTALLOC_LE
){
6096 iPage
= get4byte(&aData
[8+i
*4]);
6097 if( iPage
<=nearby
){
6104 dist
= sqlite3AbsInt32(get4byte(&aData
[8]) - nearby
);
6106 int d2
= sqlite3AbsInt32(get4byte(&aData
[8+i
*4]) - nearby
);
6117 iPage
= get4byte(&aData
[8+closest
*4]);
6118 testcase( iPage
==mxPage
);
6120 rc
= SQLITE_CORRUPT_PGNO(iTrunk
);
6121 goto end_allocate_page
;
6123 testcase( iPage
==mxPage
);
6125 || (iPage
==nearby
|| (iPage
<nearby
&& eMode
==BTALLOC_LE
))
6129 TRACE(("ALLOCATE: %d was leaf %d of %d on trunk %d"
6130 ": %d more free pages\n",
6131 *pPgno
, closest
+1, k
, pTrunk
->pgno
, n
-1));
6132 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
6133 if( rc
) goto end_allocate_page
;
6135 memcpy(&aData
[8+closest
*4], &aData
[4+k
*4], 4);
6137 put4byte(&aData
[4], k
-1);
6138 noContent
= !btreeGetHasContent(pBt
, *pPgno
)? PAGER_GET_NOCONTENT
: 0;
6139 rc
= btreeGetUnusedPage(pBt
, *pPgno
, ppPage
, noContent
);
6140 if( rc
==SQLITE_OK
){
6141 rc
= sqlite3PagerWrite((*ppPage
)->pDbPage
);
6142 if( rc
!=SQLITE_OK
){
6143 releasePage(*ppPage
);
6150 releasePage(pPrevTrunk
);
6152 }while( searchList
);
6154 /* There are no pages on the freelist, so append a new page to the
6157 ** Normally, new pages allocated by this block can be requested from the
6158 ** pager layer with the 'no-content' flag set. This prevents the pager
6159 ** from trying to read the pages content from disk. However, if the
6160 ** current transaction has already run one or more incremental-vacuum
6161 ** steps, then the page we are about to allocate may contain content
6162 ** that is required in the event of a rollback. In this case, do
6163 ** not set the no-content flag. This causes the pager to load and journal
6164 ** the current page content before overwriting it.
6166 ** Note that the pager will not actually attempt to load or journal
6167 ** content for any page that really does lie past the end of the database
6168 ** file on disk. So the effects of disabling the no-content optimization
6169 ** here are confined to those pages that lie between the end of the
6170 ** database image and the end of the database file.
6172 int bNoContent
= (0==IfNotOmitAV(pBt
->bDoTruncate
))? PAGER_GET_NOCONTENT
:0;
6174 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
6177 if( pBt
->nPage
==PENDING_BYTE_PAGE(pBt
) ) pBt
->nPage
++;
6179 #ifndef SQLITE_OMIT_AUTOVACUUM
6180 if( pBt
->autoVacuum
&& PTRMAP_ISPAGE(pBt
, pBt
->nPage
) ){
6181 /* If *pPgno refers to a pointer-map page, allocate two new pages
6182 ** at the end of the file instead of one. The first allocated page
6183 ** becomes a new pointer-map page, the second is used by the caller.
6186 TRACE(("ALLOCATE: %d from end of file (pointer-map page)\n", pBt
->nPage
));
6187 assert( pBt
->nPage
!=PENDING_BYTE_PAGE(pBt
) );
6188 rc
= btreeGetUnusedPage(pBt
, pBt
->nPage
, &pPg
, bNoContent
);
6189 if( rc
==SQLITE_OK
){
6190 rc
= sqlite3PagerWrite(pPg
->pDbPage
);
6195 if( pBt
->nPage
==PENDING_BYTE_PAGE(pBt
) ){ pBt
->nPage
++; }
6198 put4byte(28 + (u8
*)pBt
->pPage1
->aData
, pBt
->nPage
);
6199 *pPgno
= pBt
->nPage
;
6201 assert( *pPgno
!=PENDING_BYTE_PAGE(pBt
) );
6202 rc
= btreeGetUnusedPage(pBt
, *pPgno
, ppPage
, bNoContent
);
6204 rc
= sqlite3PagerWrite((*ppPage
)->pDbPage
);
6205 if( rc
!=SQLITE_OK
){
6206 releasePage(*ppPage
);
6209 TRACE(("ALLOCATE: %d from end of file\n", *pPgno
));
6212 assert( CORRUPT_DB
|| *pPgno
!=PENDING_BYTE_PAGE(pBt
) );
6215 releasePage(pTrunk
);
6216 releasePage(pPrevTrunk
);
6217 assert( rc
!=SQLITE_OK
|| sqlite3PagerPageRefcount((*ppPage
)->pDbPage
)<=1 );
6218 assert( rc
!=SQLITE_OK
|| (*ppPage
)->isInit
==0 );
6223 ** This function is used to add page iPage to the database file free-list.
6224 ** It is assumed that the page is not already a part of the free-list.
6226 ** The value passed as the second argument to this function is optional.
6227 ** If the caller happens to have a pointer to the MemPage object
6228 ** corresponding to page iPage handy, it may pass it as the second value.
6229 ** Otherwise, it may pass NULL.
6231 ** If a pointer to a MemPage object is passed as the second argument,
6232 ** its reference count is not altered by this function.
6234 static int freePage2(BtShared
*pBt
, MemPage
*pMemPage
, Pgno iPage
){
6235 MemPage
*pTrunk
= 0; /* Free-list trunk page */
6236 Pgno iTrunk
= 0; /* Page number of free-list trunk page */
6237 MemPage
*pPage1
= pBt
->pPage1
; /* Local reference to page 1 */
6238 MemPage
*pPage
; /* Page being freed. May be NULL. */
6239 int rc
; /* Return Code */
6240 u32 nFree
; /* Initial number of pages on free-list */
6242 assert( sqlite3_mutex_held(pBt
->mutex
) );
6243 assert( CORRUPT_DB
|| iPage
>1 );
6244 assert( !pMemPage
|| pMemPage
->pgno
==iPage
);
6246 if( iPage
<2 || iPage
>pBt
->nPage
){
6247 return SQLITE_CORRUPT_BKPT
;
6251 sqlite3PagerRef(pPage
->pDbPage
);
6253 pPage
= btreePageLookup(pBt
, iPage
);
6256 /* Increment the free page count on pPage1 */
6257 rc
= sqlite3PagerWrite(pPage1
->pDbPage
);
6258 if( rc
) goto freepage_out
;
6259 nFree
= get4byte(&pPage1
->aData
[36]);
6260 put4byte(&pPage1
->aData
[36], nFree
+1);
6262 if( pBt
->btsFlags
& BTS_SECURE_DELETE
){
6263 /* If the secure_delete option is enabled, then
6264 ** always fully overwrite deleted information with zeros.
6266 if( (!pPage
&& ((rc
= btreeGetPage(pBt
, iPage
, &pPage
, 0))!=0) )
6267 || ((rc
= sqlite3PagerWrite(pPage
->pDbPage
))!=0)
6271 memset(pPage
->aData
, 0, pPage
->pBt
->pageSize
);
6274 /* If the database supports auto-vacuum, write an entry in the pointer-map
6275 ** to indicate that the page is free.
6278 ptrmapPut(pBt
, iPage
, PTRMAP_FREEPAGE
, 0, &rc
);
6279 if( rc
) goto freepage_out
;
6282 /* Now manipulate the actual database free-list structure. There are two
6283 ** possibilities. If the free-list is currently empty, or if the first
6284 ** trunk page in the free-list is full, then this page will become a
6285 ** new free-list trunk page. Otherwise, it will become a leaf of the
6286 ** first trunk page in the current free-list. This block tests if it
6287 ** is possible to add the page as a new free-list leaf.
6290 u32 nLeaf
; /* Initial number of leaf cells on trunk page */
6292 iTrunk
= get4byte(&pPage1
->aData
[32]);
6293 if( iTrunk
>btreePagecount(pBt
) ){
6294 rc
= SQLITE_CORRUPT_BKPT
;
6297 rc
= btreeGetPage(pBt
, iTrunk
, &pTrunk
, 0);
6298 if( rc
!=SQLITE_OK
){
6302 nLeaf
= get4byte(&pTrunk
->aData
[4]);
6303 assert( pBt
->usableSize
>32 );
6304 if( nLeaf
> (u32
)pBt
->usableSize
/4 - 2 ){
6305 rc
= SQLITE_CORRUPT_BKPT
;
6308 if( nLeaf
< (u32
)pBt
->usableSize
/4 - 8 ){
6309 /* In this case there is room on the trunk page to insert the page
6310 ** being freed as a new leaf.
6312 ** Note that the trunk page is not really full until it contains
6313 ** usableSize/4 - 2 entries, not usableSize/4 - 8 entries as we have
6314 ** coded. But due to a coding error in versions of SQLite prior to
6315 ** 3.6.0, databases with freelist trunk pages holding more than
6316 ** usableSize/4 - 8 entries will be reported as corrupt. In order
6317 ** to maintain backwards compatibility with older versions of SQLite,
6318 ** we will continue to restrict the number of entries to usableSize/4 - 8
6319 ** for now. At some point in the future (once everyone has upgraded
6320 ** to 3.6.0 or later) we should consider fixing the conditional above
6321 ** to read "usableSize/4-2" instead of "usableSize/4-8".
6323 ** EVIDENCE-OF: R-19920-11576 However, newer versions of SQLite still
6324 ** avoid using the last six entries in the freelist trunk page array in
6325 ** order that database files created by newer versions of SQLite can be
6326 ** read by older versions of SQLite.
6328 rc
= sqlite3PagerWrite(pTrunk
->pDbPage
);
6329 if( rc
==SQLITE_OK
){
6330 put4byte(&pTrunk
->aData
[4], nLeaf
+1);
6331 put4byte(&pTrunk
->aData
[8+nLeaf
*4], iPage
);
6332 if( pPage
&& (pBt
->btsFlags
& BTS_SECURE_DELETE
)==0 ){
6333 sqlite3PagerDontWrite(pPage
->pDbPage
);
6335 rc
= btreeSetHasContent(pBt
, iPage
);
6337 TRACE(("FREE-PAGE: %d leaf on trunk page %d\n",pPage
->pgno
,pTrunk
->pgno
));
6342 /* If control flows to this point, then it was not possible to add the
6343 ** the page being freed as a leaf page of the first trunk in the free-list.
6344 ** Possibly because the free-list is empty, or possibly because the
6345 ** first trunk in the free-list is full. Either way, the page being freed
6346 ** will become the new first trunk page in the free-list.
6348 if( pPage
==0 && SQLITE_OK
!=(rc
= btreeGetPage(pBt
, iPage
, &pPage
, 0)) ){
6351 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
6352 if( rc
!=SQLITE_OK
){
6355 put4byte(pPage
->aData
, iTrunk
);
6356 put4byte(&pPage
->aData
[4], 0);
6357 put4byte(&pPage1
->aData
[32], iPage
);
6358 TRACE(("FREE-PAGE: %d new trunk page replacing %d\n", pPage
->pgno
, iTrunk
));
6365 releasePage(pTrunk
);
6368 static void freePage(MemPage
*pPage
, int *pRC
){
6369 if( (*pRC
)==SQLITE_OK
){
6370 *pRC
= freePage2(pPage
->pBt
, pPage
, pPage
->pgno
);
6375 ** Free any overflow pages associated with the given Cell. Store
6376 ** size information about the cell in pInfo.
6378 static int clearCell(
6379 MemPage
*pPage
, /* The page that contains the Cell */
6380 unsigned char *pCell
, /* First byte of the Cell */
6381 CellInfo
*pInfo
/* Size information about the cell */
6389 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6390 pPage
->xParseCell(pPage
, pCell
, pInfo
);
6391 if( pInfo
->nLocal
==pInfo
->nPayload
){
6392 return SQLITE_OK
; /* No overflow pages. Return without doing anything */
6394 testcase( pCell
+ pInfo
->nSize
== pPage
->aDataEnd
);
6395 testcase( pCell
+ (pInfo
->nSize
-1) == pPage
->aDataEnd
);
6396 if( pCell
+ pInfo
->nSize
> pPage
->aDataEnd
){
6397 /* Cell extends past end of page */
6398 return SQLITE_CORRUPT_PAGE(pPage
);
6400 ovflPgno
= get4byte(pCell
+ pInfo
->nSize
- 4);
6402 assert( pBt
->usableSize
> 4 );
6403 ovflPageSize
= pBt
->usableSize
- 4;
6404 nOvfl
= (pInfo
->nPayload
- pInfo
->nLocal
+ ovflPageSize
- 1)/ovflPageSize
;
6406 (CORRUPT_DB
&& (pInfo
->nPayload
+ ovflPageSize
)<ovflPageSize
)
6411 if( ovflPgno
<2 || ovflPgno
>btreePagecount(pBt
) ){
6412 /* 0 is not a legal page number and page 1 cannot be an
6413 ** overflow page. Therefore if ovflPgno<2 or past the end of the
6414 ** file the database must be corrupt. */
6415 return SQLITE_CORRUPT_BKPT
;
6418 rc
= getOverflowPage(pBt
, ovflPgno
, &pOvfl
, &iNext
);
6422 if( ( pOvfl
|| ((pOvfl
= btreePageLookup(pBt
, ovflPgno
))!=0) )
6423 && sqlite3PagerPageRefcount(pOvfl
->pDbPage
)!=1
6425 /* There is no reason any cursor should have an outstanding reference
6426 ** to an overflow page belonging to a cell that is being deleted/updated.
6427 ** So if there exists more than one reference to this page, then it
6428 ** must not really be an overflow page and the database must be corrupt.
6429 ** It is helpful to detect this before calling freePage2(), as
6430 ** freePage2() may zero the page contents if secure-delete mode is
6431 ** enabled. If this 'overflow' page happens to be a page that the
6432 ** caller is iterating through or using in some other way, this
6433 ** can be problematic.
6435 rc
= SQLITE_CORRUPT_BKPT
;
6437 rc
= freePage2(pBt
, pOvfl
, ovflPgno
);
6441 sqlite3PagerUnref(pOvfl
->pDbPage
);
6450 ** Create the byte sequence used to represent a cell on page pPage
6451 ** and write that byte sequence into pCell[]. Overflow pages are
6452 ** allocated and filled in as necessary. The calling procedure
6453 ** is responsible for making sure sufficient space has been allocated
6456 ** Note that pCell does not necessary need to point to the pPage->aData
6457 ** area. pCell might point to some temporary storage. The cell will
6458 ** be constructed in this temporary area then copied into pPage->aData
6461 static int fillInCell(
6462 MemPage
*pPage
, /* The page that contains the cell */
6463 unsigned char *pCell
, /* Complete text of the cell */
6464 const BtreePayload
*pX
, /* Payload with which to construct the cell */
6465 int *pnSize
/* Write cell size here */
6469 int nSrc
, n
, rc
, mn
;
6471 MemPage
*pToRelease
;
6472 unsigned char *pPrior
;
6473 unsigned char *pPayload
;
6478 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6480 /* pPage is not necessarily writeable since pCell might be auxiliary
6481 ** buffer space that is separate from the pPage buffer area */
6482 assert( pCell
<pPage
->aData
|| pCell
>=&pPage
->aData
[pPage
->pBt
->pageSize
]
6483 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6485 /* Fill in the header. */
6486 nHeader
= pPage
->childPtrSize
;
6487 if( pPage
->intKey
){
6488 nPayload
= pX
->nData
+ pX
->nZero
;
6491 assert( pPage
->intKeyLeaf
); /* fillInCell() only called for leaves */
6492 nHeader
+= putVarint32(&pCell
[nHeader
], nPayload
);
6493 nHeader
+= putVarint(&pCell
[nHeader
], *(u64
*)&pX
->nKey
);
6495 assert( pX
->nKey
<=0x7fffffff && pX
->pKey
!=0 );
6496 nSrc
= nPayload
= (int)pX
->nKey
;
6498 nHeader
+= putVarint32(&pCell
[nHeader
], nPayload
);
6501 /* Fill in the payload */
6502 pPayload
= &pCell
[nHeader
];
6503 if( nPayload
<=pPage
->maxLocal
){
6504 /* This is the common case where everything fits on the btree page
6505 ** and no overflow pages are required. */
6506 n
= nHeader
+ nPayload
;
6511 assert( nSrc
<=nPayload
);
6512 testcase( nSrc
<nPayload
);
6513 memcpy(pPayload
, pSrc
, nSrc
);
6514 memset(pPayload
+nSrc
, 0, nPayload
-nSrc
);
6518 /* If we reach this point, it means that some of the content will need
6519 ** to spill onto overflow pages.
6521 mn
= pPage
->minLocal
;
6522 n
= mn
+ (nPayload
- mn
) % (pPage
->pBt
->usableSize
- 4);
6523 testcase( n
==pPage
->maxLocal
);
6524 testcase( n
==pPage
->maxLocal
+1 );
6525 if( n
> pPage
->maxLocal
) n
= mn
;
6527 *pnSize
= n
+ nHeader
+ 4;
6528 pPrior
= &pCell
[nHeader
+n
];
6533 /* At this point variables should be set as follows:
6535 ** nPayload Total payload size in bytes
6536 ** pPayload Begin writing payload here
6537 ** spaceLeft Space available at pPayload. If nPayload>spaceLeft,
6538 ** that means content must spill into overflow pages.
6539 ** *pnSize Size of the local cell (not counting overflow pages)
6540 ** pPrior Where to write the pgno of the first overflow page
6542 ** Use a call to btreeParseCellPtr() to verify that the values above
6543 ** were computed correctly.
6548 pPage
->xParseCell(pPage
, pCell
, &info
);
6549 assert( nHeader
==(int)(info
.pPayload
- pCell
) );
6550 assert( info
.nKey
==pX
->nKey
);
6551 assert( *pnSize
== info
.nSize
);
6552 assert( spaceLeft
== info
.nLocal
);
6556 /* Write the payload into the local Cell and any extra into overflow pages */
6559 if( n
>spaceLeft
) n
= spaceLeft
;
6561 /* If pToRelease is not zero than pPayload points into the data area
6562 ** of pToRelease. Make sure pToRelease is still writeable. */
6563 assert( pToRelease
==0 || sqlite3PagerIswriteable(pToRelease
->pDbPage
) );
6565 /* If pPayload is part of the data area of pPage, then make sure pPage
6566 ** is still writeable */
6567 assert( pPayload
<pPage
->aData
|| pPayload
>=&pPage
->aData
[pBt
->pageSize
]
6568 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6571 memcpy(pPayload
, pSrc
, n
);
6574 memcpy(pPayload
, pSrc
, n
);
6576 memset(pPayload
, 0, n
);
6579 if( nPayload
<=0 ) break;
6586 #ifndef SQLITE_OMIT_AUTOVACUUM
6587 Pgno pgnoPtrmap
= pgnoOvfl
; /* Overflow page pointer-map entry page */
6588 if( pBt
->autoVacuum
){
6592 PTRMAP_ISPAGE(pBt
, pgnoOvfl
) || pgnoOvfl
==PENDING_BYTE_PAGE(pBt
)
6596 rc
= allocateBtreePage(pBt
, &pOvfl
, &pgnoOvfl
, pgnoOvfl
, 0);
6597 #ifndef SQLITE_OMIT_AUTOVACUUM
6598 /* If the database supports auto-vacuum, and the second or subsequent
6599 ** overflow page is being allocated, add an entry to the pointer-map
6600 ** for that page now.
6602 ** If this is the first overflow page, then write a partial entry
6603 ** to the pointer-map. If we write nothing to this pointer-map slot,
6604 ** then the optimistic overflow chain processing in clearCell()
6605 ** may misinterpret the uninitialized values and delete the
6606 ** wrong pages from the database.
6608 if( pBt
->autoVacuum
&& rc
==SQLITE_OK
){
6609 u8 eType
= (pgnoPtrmap
?PTRMAP_OVERFLOW2
:PTRMAP_OVERFLOW1
);
6610 ptrmapPut(pBt
, pgnoOvfl
, eType
, pgnoPtrmap
, &rc
);
6617 releasePage(pToRelease
);
6621 /* If pToRelease is not zero than pPrior points into the data area
6622 ** of pToRelease. Make sure pToRelease is still writeable. */
6623 assert( pToRelease
==0 || sqlite3PagerIswriteable(pToRelease
->pDbPage
) );
6625 /* If pPrior is part of the data area of pPage, then make sure pPage
6626 ** is still writeable */
6627 assert( pPrior
<pPage
->aData
|| pPrior
>=&pPage
->aData
[pBt
->pageSize
]
6628 || sqlite3PagerIswriteable(pPage
->pDbPage
) );
6630 put4byte(pPrior
, pgnoOvfl
);
6631 releasePage(pToRelease
);
6633 pPrior
= pOvfl
->aData
;
6634 put4byte(pPrior
, 0);
6635 pPayload
= &pOvfl
->aData
[4];
6636 spaceLeft
= pBt
->usableSize
- 4;
6639 releasePage(pToRelease
);
6644 ** Remove the i-th cell from pPage. This routine effects pPage only.
6645 ** The cell content is not freed or deallocated. It is assumed that
6646 ** the cell content has been copied someplace else. This routine just
6647 ** removes the reference to the cell from pPage.
6649 ** "sz" must be the number of bytes in the cell.
6651 static void dropCell(MemPage
*pPage
, int idx
, int sz
, int *pRC
){
6652 u32 pc
; /* Offset to cell content of cell being deleted */
6653 u8
*data
; /* pPage->aData */
6654 u8
*ptr
; /* Used to move bytes around within data[] */
6655 int rc
; /* The return code */
6656 int hdr
; /* Beginning of the header. 0 most pages. 100 page 1 */
6659 assert( idx
>=0 && idx
<pPage
->nCell
);
6660 assert( CORRUPT_DB
|| sz
==cellSize(pPage
, idx
) );
6661 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
6662 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6663 assert( pPage
->nFree
>=0 );
6664 data
= pPage
->aData
;
6665 ptr
= &pPage
->aCellIdx
[2*idx
];
6667 hdr
= pPage
->hdrOffset
;
6668 testcase( pc
==get2byte(&data
[hdr
+5]) );
6669 testcase( pc
+sz
==pPage
->pBt
->usableSize
);
6670 if( pc
+sz
> pPage
->pBt
->usableSize
){
6671 *pRC
= SQLITE_CORRUPT_BKPT
;
6674 rc
= freeSpace(pPage
, pc
, sz
);
6680 if( pPage
->nCell
==0 ){
6681 memset(&data
[hdr
+1], 0, 4);
6683 put2byte(&data
[hdr
+5], pPage
->pBt
->usableSize
);
6684 pPage
->nFree
= pPage
->pBt
->usableSize
- pPage
->hdrOffset
6685 - pPage
->childPtrSize
- 8;
6687 memmove(ptr
, ptr
+2, 2*(pPage
->nCell
- idx
));
6688 put2byte(&data
[hdr
+3], pPage
->nCell
);
6694 ** Insert a new cell on pPage at cell index "i". pCell points to the
6695 ** content of the cell.
6697 ** If the cell content will fit on the page, then put it there. If it
6698 ** will not fit, then make a copy of the cell content into pTemp if
6699 ** pTemp is not null. Regardless of pTemp, allocate a new entry
6700 ** in pPage->apOvfl[] and make it point to the cell content (either
6701 ** in pTemp or the original pCell) and also record its index.
6702 ** Allocating a new entry in pPage->aCell[] implies that
6703 ** pPage->nOverflow is incremented.
6705 ** *pRC must be SQLITE_OK when this routine is called.
6707 static void insertCell(
6708 MemPage
*pPage
, /* Page into which we are copying */
6709 int i
, /* New cell becomes the i-th cell of the page */
6710 u8
*pCell
, /* Content of the new cell */
6711 int sz
, /* Bytes of content in pCell */
6712 u8
*pTemp
, /* Temp storage space for pCell, if needed */
6713 Pgno iChild
, /* If non-zero, replace first 4 bytes with this value */
6714 int *pRC
/* Read and write return code from here */
6716 int idx
= 0; /* Where to write new cell content in data[] */
6717 int j
; /* Loop counter */
6718 u8
*data
; /* The content of the whole page */
6719 u8
*pIns
; /* The point in pPage->aCellIdx[] where no cell inserted */
6721 assert( *pRC
==SQLITE_OK
);
6722 assert( i
>=0 && i
<=pPage
->nCell
+pPage
->nOverflow
);
6723 assert( MX_CELL(pPage
->pBt
)<=10921 );
6724 assert( pPage
->nCell
<=MX_CELL(pPage
->pBt
) || CORRUPT_DB
);
6725 assert( pPage
->nOverflow
<=ArraySize(pPage
->apOvfl
) );
6726 assert( ArraySize(pPage
->apOvfl
)==ArraySize(pPage
->aiOvfl
) );
6727 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
6728 assert( sz
==pPage
->xCellSize(pPage
, pCell
) || CORRUPT_DB
);
6729 assert( pPage
->nFree
>=0 );
6730 if( pPage
->nOverflow
|| sz
+2>pPage
->nFree
){
6732 memcpy(pTemp
, pCell
, sz
);
6736 put4byte(pCell
, iChild
);
6738 j
= pPage
->nOverflow
++;
6739 /* Comparison against ArraySize-1 since we hold back one extra slot
6740 ** as a contingency. In other words, never need more than 3 overflow
6741 ** slots but 4 are allocated, just to be safe. */
6742 assert( j
< ArraySize(pPage
->apOvfl
)-1 );
6743 pPage
->apOvfl
[j
] = pCell
;
6744 pPage
->aiOvfl
[j
] = (u16
)i
;
6746 /* When multiple overflows occur, they are always sequential and in
6747 ** sorted order. This invariants arise because multiple overflows can
6748 ** only occur when inserting divider cells into the parent page during
6749 ** balancing, and the dividers are adjacent and sorted.
6751 assert( j
==0 || pPage
->aiOvfl
[j
-1]<(u16
)i
); /* Overflows in sorted order */
6752 assert( j
==0 || i
==pPage
->aiOvfl
[j
-1]+1 ); /* Overflows are sequential */
6754 int rc
= sqlite3PagerWrite(pPage
->pDbPage
);
6755 if( rc
!=SQLITE_OK
){
6759 assert( sqlite3PagerIswriteable(pPage
->pDbPage
) );
6760 data
= pPage
->aData
;
6761 assert( &data
[pPage
->cellOffset
]==pPage
->aCellIdx
);
6762 rc
= allocateSpace(pPage
, sz
, &idx
);
6763 if( rc
){ *pRC
= rc
; return; }
6764 /* The allocateSpace() routine guarantees the following properties
6765 ** if it returns successfully */
6767 assert( idx
>= pPage
->cellOffset
+2*pPage
->nCell
+2 || CORRUPT_DB
);
6768 assert( idx
+sz
<= (int)pPage
->pBt
->usableSize
);
6769 pPage
->nFree
-= (u16
)(2 + sz
);
6771 /* In a corrupt database where an entry in the cell index section of
6772 ** a btree page has a value of 3 or less, the pCell value might point
6773 ** as many as 4 bytes in front of the start of the aData buffer for
6774 ** the source page. Make sure this does not cause problems by not
6775 ** reading the first 4 bytes */
6776 memcpy(&data
[idx
+4], pCell
+4, sz
-4);
6777 put4byte(&data
[idx
], iChild
);
6779 memcpy(&data
[idx
], pCell
, sz
);
6781 pIns
= pPage
->aCellIdx
+ i
*2;
6782 memmove(pIns
+2, pIns
, 2*(pPage
->nCell
- i
));
6783 put2byte(pIns
, idx
);
6785 /* increment the cell count */
6786 if( (++data
[pPage
->hdrOffset
+4])==0 ) data
[pPage
->hdrOffset
+3]++;
6787 assert( get2byte(&data
[pPage
->hdrOffset
+3])==pPage
->nCell
|| CORRUPT_DB
);
6788 #ifndef SQLITE_OMIT_AUTOVACUUM
6789 if( pPage
->pBt
->autoVacuum
){
6790 /* The cell may contain a pointer to an overflow page. If so, write
6791 ** the entry for the overflow page into the pointer map.
6793 ptrmapPutOvflPtr(pPage
, pPage
, pCell
, pRC
);
6800 ** The following parameters determine how many adjacent pages get involved
6801 ** in a balancing operation. NN is the number of neighbors on either side
6802 ** of the page that participate in the balancing operation. NB is the
6803 ** total number of pages that participate, including the target page and
6804 ** NN neighbors on either side.
6806 ** The minimum value of NN is 1 (of course). Increasing NN above 1
6807 ** (to 2 or 3) gives a modest improvement in SELECT and DELETE performance
6808 ** in exchange for a larger degradation in INSERT and UPDATE performance.
6809 ** The value of NN appears to give the best results overall.
6811 ** (Later:) The description above makes it seem as if these values are
6812 ** tunable - as if you could change them and recompile and it would all work.
6813 ** But that is unlikely. NB has been 3 since the inception of SQLite and
6814 ** we have never tested any other value.
6816 #define NN 1 /* Number of neighbors on either side of pPage */
6817 #define NB 3 /* (NN*2+1): Total pages involved in the balance */
6820 ** A CellArray object contains a cache of pointers and sizes for a
6821 ** consecutive sequence of cells that might be held on multiple pages.
6823 ** The cells in this array are the divider cell or cells from the pParent
6824 ** page plus up to three child pages. There are a total of nCell cells.
6826 ** pRef is a pointer to one of the pages that contributes cells. This is
6827 ** used to access information such as MemPage.intKey and MemPage.pBt->pageSize
6828 ** which should be common to all pages that contribute cells to this array.
6830 ** apCell[] and szCell[] hold, respectively, pointers to the start of each
6831 ** cell and the size of each cell. Some of the apCell[] pointers might refer
6832 ** to overflow cells. In other words, some apCel[] pointers might not point
6833 ** to content area of the pages.
6835 ** A szCell[] of zero means the size of that cell has not yet been computed.
6837 ** The cells come from as many as four different pages:
6844 ** --------- --------- ---------
6845 ** |Child-1| |Child-2| |Child-3|
6846 ** --------- --------- ---------
6848 ** The order of cells is in the array is for an index btree is:
6850 ** 1. All cells from Child-1 in order
6851 ** 2. The first divider cell from Parent
6852 ** 3. All cells from Child-2 in order
6853 ** 4. The second divider cell from Parent
6854 ** 5. All cells from Child-3 in order
6856 ** For a table-btree (with rowids) the items 2 and 4 are empty because
6857 ** content exists only in leaves and there are no divider cells.
6859 ** For an index btree, the apEnd[] array holds pointer to the end of page
6860 ** for Child-1, the Parent, Child-2, the Parent (again), and Child-3,
6861 ** respectively. The ixNx[] array holds the number of cells contained in
6862 ** each of these 5 stages, and all stages to the left. Hence:
6864 ** ixNx[0] = Number of cells in Child-1.
6865 ** ixNx[1] = Number of cells in Child-1 plus 1 for first divider.
6866 ** ixNx[2] = Number of cells in Child-1 and Child-2 + 1 for 1st divider.
6867 ** ixNx[3] = Number of cells in Child-1 and Child-2 + both divider cells
6868 ** ixNx[4] = Total number of cells.
6870 ** For a table-btree, the concept is similar, except only apEnd[0]..apEnd[2]
6871 ** are used and they point to the leaf pages only, and the ixNx value are:
6873 ** ixNx[0] = Number of cells in Child-1.
6874 ** ixNx[1] = Number of cells in Child-1 and Child-2.
6875 ** ixNx[2] = Total number of cells.
6877 ** Sometimes when deleting, a child page can have zero cells. In those
6878 ** cases, ixNx[] entries with higher indexes, and the corresponding apEnd[]
6879 ** entries, shift down. The end result is that each ixNx[] entry should
6880 ** be larger than the previous
6882 typedef struct CellArray CellArray
;
6884 int nCell
; /* Number of cells in apCell[] */
6885 MemPage
*pRef
; /* Reference page */
6886 u8
**apCell
; /* All cells begin balanced */
6887 u16
*szCell
; /* Local size of all cells in apCell[] */
6888 u8
*apEnd
[NB
*2]; /* MemPage.aDataEnd values */
6889 int ixNx
[NB
*2]; /* Index of at which we move to the next apEnd[] */
6893 ** Make sure the cell sizes at idx, idx+1, ..., idx+N-1 have been
6896 static void populateCellCache(CellArray
*p
, int idx
, int N
){
6897 assert( idx
>=0 && idx
+N
<=p
->nCell
);
6899 assert( p
->apCell
[idx
]!=0 );
6900 if( p
->szCell
[idx
]==0 ){
6901 p
->szCell
[idx
] = p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[idx
]);
6903 assert( CORRUPT_DB
||
6904 p
->szCell
[idx
]==p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[idx
]) );
6912 ** Return the size of the Nth element of the cell array
6914 static SQLITE_NOINLINE u16
computeCellSize(CellArray
*p
, int N
){
6915 assert( N
>=0 && N
<p
->nCell
);
6916 assert( p
->szCell
[N
]==0 );
6917 p
->szCell
[N
] = p
->pRef
->xCellSize(p
->pRef
, p
->apCell
[N
]);
6918 return p
->szCell
[N
];
6920 static u16
cachedCellSize(CellArray
*p
, int N
){
6921 assert( N
>=0 && N
<p
->nCell
);
6922 if( p
->szCell
[N
] ) return p
->szCell
[N
];
6923 return computeCellSize(p
, N
);
6927 ** Array apCell[] contains pointers to nCell b-tree page cells. The
6928 ** szCell[] array contains the size in bytes of each cell. This function
6929 ** replaces the current contents of page pPg with the contents of the cell
6932 ** Some of the cells in apCell[] may currently be stored in pPg. This
6933 ** function works around problems caused by this by making a copy of any
6934 ** such cells before overwriting the page data.
6936 ** The MemPage.nFree field is invalidated by this function. It is the
6937 ** responsibility of the caller to set it correctly.
6939 static int rebuildPage(
6940 CellArray
*pCArray
, /* Content to be added to page pPg */
6941 int iFirst
, /* First cell in pCArray to use */
6942 int nCell
, /* Final number of cells on page */
6943 MemPage
*pPg
/* The page to be reconstructed */
6945 const int hdr
= pPg
->hdrOffset
; /* Offset of header on pPg */
6946 u8
* const aData
= pPg
->aData
; /* Pointer to data for pPg */
6947 const int usableSize
= pPg
->pBt
->usableSize
;
6948 u8
* const pEnd
= &aData
[usableSize
];
6949 int i
= iFirst
; /* Which cell to copy from pCArray*/
6950 u32 j
; /* Start of cell content area */
6951 int iEnd
= i
+nCell
; /* Loop terminator */
6952 u8
*pCellptr
= pPg
->aCellIdx
;
6953 u8
*pTmp
= sqlite3PagerTempSpace(pPg
->pBt
->pPager
);
6955 int k
; /* Current slot in pCArray->apEnd[] */
6956 u8
*pSrcEnd
; /* Current pCArray->apEnd[k] value */
6959 j
= get2byte(&aData
[hdr
+5]);
6960 if( NEVER(j
>(u32
)usableSize
) ){ j
= 0; }
6961 memcpy(&pTmp
[j
], &aData
[j
], usableSize
- j
);
6963 for(k
=0; pCArray
->ixNx
[k
]<=i
&& ALWAYS(k
<NB
*2); k
++){}
6964 pSrcEnd
= pCArray
->apEnd
[k
];
6967 while( 1/*exit by break*/ ){
6968 u8
*pCell
= pCArray
->apCell
[i
];
6969 u16 sz
= pCArray
->szCell
[i
];
6971 if( SQLITE_WITHIN(pCell
,aData
,pEnd
) ){
6972 if( ((uptr
)(pCell
+sz
))>(uptr
)pEnd
) return SQLITE_CORRUPT_BKPT
;
6973 pCell
= &pTmp
[pCell
- aData
];
6974 }else if( (uptr
)(pCell
+sz
)>(uptr
)pSrcEnd
6975 && (uptr
)(pCell
)<(uptr
)pSrcEnd
6977 return SQLITE_CORRUPT_BKPT
;
6981 put2byte(pCellptr
, (pData
- aData
));
6983 if( pData
< pCellptr
) return SQLITE_CORRUPT_BKPT
;
6984 memcpy(pData
, pCell
, sz
);
6985 assert( sz
==pPg
->xCellSize(pPg
, pCell
) || CORRUPT_DB
);
6986 testcase( sz
!=pPg
->xCellSize(pPg
,pCell
) )
6988 if( i
>=iEnd
) break;
6989 if( pCArray
->ixNx
[k
]<=i
){
6991 pSrcEnd
= pCArray
->apEnd
[k
];
6995 /* The pPg->nFree field is now set incorrectly. The caller will fix it. */
6999 put2byte(&aData
[hdr
+1], 0);
7000 put2byte(&aData
[hdr
+3], pPg
->nCell
);
7001 put2byte(&aData
[hdr
+5], pData
- aData
);
7002 aData
[hdr
+7] = 0x00;
7007 ** The pCArray objects contains pointers to b-tree cells and the cell sizes.
7008 ** This function attempts to add the cells stored in the array to page pPg.
7009 ** If it cannot (because the page needs to be defragmented before the cells
7010 ** will fit), non-zero is returned. Otherwise, if the cells are added
7011 ** successfully, zero is returned.
7013 ** Argument pCellptr points to the first entry in the cell-pointer array
7014 ** (part of page pPg) to populate. After cell apCell[0] is written to the
7015 ** page body, a 16-bit offset is written to pCellptr. And so on, for each
7016 ** cell in the array. It is the responsibility of the caller to ensure
7017 ** that it is safe to overwrite this part of the cell-pointer array.
7019 ** When this function is called, *ppData points to the start of the
7020 ** content area on page pPg. If the size of the content area is extended,
7021 ** *ppData is updated to point to the new start of the content area
7022 ** before returning.
7024 ** Finally, argument pBegin points to the byte immediately following the
7025 ** end of the space required by this page for the cell-pointer area (for
7026 ** all cells - not just those inserted by the current call). If the content
7027 ** area must be extended to before this point in order to accomodate all
7028 ** cells in apCell[], then the cells do not fit and non-zero is returned.
7030 static int pageInsertArray(
7031 MemPage
*pPg
, /* Page to add cells to */
7032 u8
*pBegin
, /* End of cell-pointer array */
7033 u8
**ppData
, /* IN/OUT: Page content-area pointer */
7034 u8
*pCellptr
, /* Pointer to cell-pointer area */
7035 int iFirst
, /* Index of first cell to add */
7036 int nCell
, /* Number of cells to add to pPg */
7037 CellArray
*pCArray
/* Array of cells */
7039 int i
= iFirst
; /* Loop counter - cell index to insert */
7040 u8
*aData
= pPg
->aData
; /* Complete page */
7041 u8
*pData
= *ppData
; /* Content area. A subset of aData[] */
7042 int iEnd
= iFirst
+ nCell
; /* End of loop. One past last cell to ins */
7043 int k
; /* Current slot in pCArray->apEnd[] */
7044 u8
*pEnd
; /* Maximum extent of cell data */
7045 assert( CORRUPT_DB
|| pPg
->hdrOffset
==0 ); /* Never called on page 1 */
7046 if( iEnd
<=iFirst
) return 0;
7047 for(k
=0; pCArray
->ixNx
[k
]<=i
&& ALWAYS(k
<NB
*2); k
++){}
7048 pEnd
= pCArray
->apEnd
[k
];
7049 while( 1 /*Exit by break*/ ){
7052 assert( pCArray
->szCell
[i
]!=0 );
7053 sz
= pCArray
->szCell
[i
];
7054 if( (aData
[1]==0 && aData
[2]==0) || (pSlot
= pageFindSlot(pPg
,sz
,&rc
))==0 ){
7055 if( (pData
- pBegin
)<sz
) return 1;
7059 /* pSlot and pCArray->apCell[i] will never overlap on a well-formed
7060 ** database. But they might for a corrupt database. Hence use memmove()
7061 ** since memcpy() sends SIGABORT with overlapping buffers on OpenBSD */
7062 assert( (pSlot
+sz
)<=pCArray
->apCell
[i
]
7063 || pSlot
>=(pCArray
->apCell
[i
]+sz
)
7065 if( (uptr
)(pCArray
->apCell
[i
]+sz
)>(uptr
)pEnd
7066 && (uptr
)(pCArray
->apCell
[i
])<(uptr
)pEnd
7068 assert( CORRUPT_DB
);
7069 (void)SQLITE_CORRUPT_BKPT
;
7072 memmove(pSlot
, pCArray
->apCell
[i
], sz
);
7073 put2byte(pCellptr
, (pSlot
- aData
));
7076 if( i
>=iEnd
) break;
7077 if( pCArray
->ixNx
[k
]<=i
){
7079 pEnd
= pCArray
->apEnd
[k
];
7087 ** The pCArray object contains pointers to b-tree cells and their sizes.
7089 ** This function adds the space associated with each cell in the array
7090 ** that is currently stored within the body of pPg to the pPg free-list.
7091 ** The cell-pointers and other fields of the page are not updated.
7093 ** This function returns the total number of cells added to the free-list.
7095 static int pageFreeArray(
7096 MemPage
*pPg
, /* Page to edit */
7097 int iFirst
, /* First cell to delete */
7098 int nCell
, /* Cells to delete */
7099 CellArray
*pCArray
/* Array of cells */
7101 u8
* const aData
= pPg
->aData
;
7102 u8
* const pEnd
= &aData
[pPg
->pBt
->usableSize
];
7103 u8
* const pStart
= &aData
[pPg
->hdrOffset
+ 8 + pPg
->childPtrSize
];
7106 int iEnd
= iFirst
+ nCell
;
7110 for(i
=iFirst
; i
<iEnd
; i
++){
7111 u8
*pCell
= pCArray
->apCell
[i
];
7112 if( SQLITE_WITHIN(pCell
, pStart
, pEnd
) ){
7114 /* No need to use cachedCellSize() here. The sizes of all cells that
7115 ** are to be freed have already been computing while deciding which
7116 ** cells need freeing */
7117 sz
= pCArray
->szCell
[i
]; assert( sz
>0 );
7118 if( pFree
!=(pCell
+ sz
) ){
7120 assert( pFree
>aData
&& (pFree
- aData
)<65536 );
7121 freeSpace(pPg
, (u16
)(pFree
- aData
), szFree
);
7125 if( pFree
+sz
>pEnd
) return 0;
7134 assert( pFree
>aData
&& (pFree
- aData
)<65536 );
7135 freeSpace(pPg
, (u16
)(pFree
- aData
), szFree
);
7141 ** pCArray contains pointers to and sizes of all cells in the page being
7142 ** balanced. The current page, pPg, has pPg->nCell cells starting with
7143 ** pCArray->apCell[iOld]. After balancing, this page should hold nNew cells
7144 ** starting at apCell[iNew].
7146 ** This routine makes the necessary adjustments to pPg so that it contains
7147 ** the correct cells after being balanced.
7149 ** The pPg->nFree field is invalid when this function returns. It is the
7150 ** responsibility of the caller to set it correctly.
7152 static int editPage(
7153 MemPage
*pPg
, /* Edit this page */
7154 int iOld
, /* Index of first cell currently on page */
7155 int iNew
, /* Index of new first cell on page */
7156 int nNew
, /* Final number of cells on page */
7157 CellArray
*pCArray
/* Array of cells and sizes */
7159 u8
* const aData
= pPg
->aData
;
7160 const int hdr
= pPg
->hdrOffset
;
7161 u8
*pBegin
= &pPg
->aCellIdx
[nNew
* 2];
7162 int nCell
= pPg
->nCell
; /* Cells stored on pPg */
7166 int iOldEnd
= iOld
+ pPg
->nCell
+ pPg
->nOverflow
;
7167 int iNewEnd
= iNew
+ nNew
;
7170 u8
*pTmp
= sqlite3PagerTempSpace(pPg
->pBt
->pPager
);
7171 memcpy(pTmp
, aData
, pPg
->pBt
->usableSize
);
7174 /* Remove cells from the start and end of the page */
7177 int nShift
= pageFreeArray(pPg
, iOld
, iNew
-iOld
, pCArray
);
7178 if( NEVER(nShift
>nCell
) ) return SQLITE_CORRUPT_BKPT
;
7179 memmove(pPg
->aCellIdx
, &pPg
->aCellIdx
[nShift
*2], nCell
*2);
7182 if( iNewEnd
< iOldEnd
){
7183 int nTail
= pageFreeArray(pPg
, iNewEnd
, iOldEnd
- iNewEnd
, pCArray
);
7184 assert( nCell
>=nTail
);
7188 pData
= &aData
[get2byteNotZero(&aData
[hdr
+5])];
7189 if( pData
<pBegin
) goto editpage_fail
;
7191 /* Add cells to the start of the page */
7193 int nAdd
= MIN(nNew
,iOld
-iNew
);
7194 assert( (iOld
-iNew
)<nNew
|| nCell
==0 || CORRUPT_DB
);
7196 pCellptr
= pPg
->aCellIdx
;
7197 memmove(&pCellptr
[nAdd
*2], pCellptr
, nCell
*2);
7198 if( pageInsertArray(
7199 pPg
, pBegin
, &pData
, pCellptr
,
7201 ) ) goto editpage_fail
;
7205 /* Add any overflow cells */
7206 for(i
=0; i
<pPg
->nOverflow
; i
++){
7207 int iCell
= (iOld
+ pPg
->aiOvfl
[i
]) - iNew
;
7208 if( iCell
>=0 && iCell
<nNew
){
7209 pCellptr
= &pPg
->aCellIdx
[iCell
* 2];
7211 memmove(&pCellptr
[2], pCellptr
, (nCell
- iCell
) * 2);
7214 cachedCellSize(pCArray
, iCell
+iNew
);
7215 if( pageInsertArray(
7216 pPg
, pBegin
, &pData
, pCellptr
,
7217 iCell
+iNew
, 1, pCArray
7218 ) ) goto editpage_fail
;
7222 /* Append cells to the end of the page */
7224 pCellptr
= &pPg
->aCellIdx
[nCell
*2];
7225 if( pageInsertArray(
7226 pPg
, pBegin
, &pData
, pCellptr
,
7227 iNew
+nCell
, nNew
-nCell
, pCArray
7228 ) ) goto editpage_fail
;
7233 put2byte(&aData
[hdr
+3], pPg
->nCell
);
7234 put2byte(&aData
[hdr
+5], pData
- aData
);
7237 for(i
=0; i
<nNew
&& !CORRUPT_DB
; i
++){
7238 u8
*pCell
= pCArray
->apCell
[i
+iNew
];
7239 int iOff
= get2byteAligned(&pPg
->aCellIdx
[i
*2]);
7240 if( SQLITE_WITHIN(pCell
, aData
, &aData
[pPg
->pBt
->usableSize
]) ){
7241 pCell
= &pTmp
[pCell
- aData
];
7243 assert( 0==memcmp(pCell
, &aData
[iOff
],
7244 pCArray
->pRef
->xCellSize(pCArray
->pRef
, pCArray
->apCell
[i
+iNew
])) );
7250 /* Unable to edit this page. Rebuild it from scratch instead. */
7251 populateCellCache(pCArray
, iNew
, nNew
);
7252 return rebuildPage(pCArray
, iNew
, nNew
, pPg
);
7256 #ifndef SQLITE_OMIT_QUICKBALANCE
7258 ** This version of balance() handles the common special case where
7259 ** a new entry is being inserted on the extreme right-end of the
7260 ** tree, in other words, when the new entry will become the largest
7261 ** entry in the tree.
7263 ** Instead of trying to balance the 3 right-most leaf pages, just add
7264 ** a new page to the right-hand side and put the one new entry in
7265 ** that page. This leaves the right side of the tree somewhat
7266 ** unbalanced. But odds are that we will be inserting new entries
7267 ** at the end soon afterwards so the nearly empty page will quickly
7268 ** fill up. On average.
7270 ** pPage is the leaf page which is the right-most page in the tree.
7271 ** pParent is its parent. pPage must have a single overflow entry
7272 ** which is also the right-most entry on the page.
7274 ** The pSpace buffer is used to store a temporary copy of the divider
7275 ** cell that will be inserted into pParent. Such a cell consists of a 4
7276 ** byte page number followed by a variable length integer. In other
7277 ** words, at most 13 bytes. Hence the pSpace buffer must be at
7278 ** least 13 bytes in size.
7280 static int balance_quick(MemPage
*pParent
, MemPage
*pPage
, u8
*pSpace
){
7281 BtShared
*const pBt
= pPage
->pBt
; /* B-Tree Database */
7282 MemPage
*pNew
; /* Newly allocated page */
7283 int rc
; /* Return Code */
7284 Pgno pgnoNew
; /* Page number of pNew */
7286 assert( sqlite3_mutex_held(pPage
->pBt
->mutex
) );
7287 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7288 assert( pPage
->nOverflow
==1 );
7290 if( pPage
->nCell
==0 ) return SQLITE_CORRUPT_BKPT
; /* dbfuzz001.test */
7291 assert( pPage
->nFree
>=0 );
7292 assert( pParent
->nFree
>=0 );
7294 /* Allocate a new page. This page will become the right-sibling of
7295 ** pPage. Make the parent page writable, so that the new divider cell
7296 ** may be inserted. If both these operations are successful, proceed.
7298 rc
= allocateBtreePage(pBt
, &pNew
, &pgnoNew
, 0, 0);
7300 if( rc
==SQLITE_OK
){
7302 u8
*pOut
= &pSpace
[4];
7303 u8
*pCell
= pPage
->apOvfl
[0];
7304 u16 szCell
= pPage
->xCellSize(pPage
, pCell
);
7308 assert( sqlite3PagerIswriteable(pNew
->pDbPage
) );
7309 assert( CORRUPT_DB
|| pPage
->aData
[0]==(PTF_INTKEY
|PTF_LEAFDATA
|PTF_LEAF
) );
7310 zeroPage(pNew
, PTF_INTKEY
|PTF_LEAFDATA
|PTF_LEAF
);
7315 b
.apEnd
[0] = pPage
->aDataEnd
;
7317 rc
= rebuildPage(&b
, 0, 1, pNew
);
7322 pNew
->nFree
= pBt
->usableSize
- pNew
->cellOffset
- 2 - szCell
;
7324 /* If this is an auto-vacuum database, update the pointer map
7325 ** with entries for the new page, and any pointer from the
7326 ** cell on the page to an overflow page. If either of these
7327 ** operations fails, the return code is set, but the contents
7328 ** of the parent page are still manipulated by thh code below.
7329 ** That is Ok, at this point the parent page is guaranteed to
7330 ** be marked as dirty. Returning an error code will cause a
7331 ** rollback, undoing any changes made to the parent page.
7334 ptrmapPut(pBt
, pgnoNew
, PTRMAP_BTREE
, pParent
->pgno
, &rc
);
7335 if( szCell
>pNew
->minLocal
){
7336 ptrmapPutOvflPtr(pNew
, pNew
, pCell
, &rc
);
7340 /* Create a divider cell to insert into pParent. The divider cell
7341 ** consists of a 4-byte page number (the page number of pPage) and
7342 ** a variable length key value (which must be the same value as the
7343 ** largest key on pPage).
7345 ** To find the largest key value on pPage, first find the right-most
7346 ** cell on pPage. The first two fields of this cell are the
7347 ** record-length (a variable length integer at most 32-bits in size)
7348 ** and the key value (a variable length integer, may have any value).
7349 ** The first of the while(...) loops below skips over the record-length
7350 ** field. The second while(...) loop copies the key value from the
7351 ** cell on pPage into the pSpace buffer.
7353 pCell
= findCell(pPage
, pPage
->nCell
-1);
7355 while( (*(pCell
++)&0x80) && pCell
<pStop
);
7357 while( ((*(pOut
++) = *(pCell
++))&0x80) && pCell
<pStop
);
7359 /* Insert the new divider cell into pParent. */
7360 if( rc
==SQLITE_OK
){
7361 insertCell(pParent
, pParent
->nCell
, pSpace
, (int)(pOut
-pSpace
),
7362 0, pPage
->pgno
, &rc
);
7365 /* Set the right-child pointer of pParent to point to the new page. */
7366 put4byte(&pParent
->aData
[pParent
->hdrOffset
+8], pgnoNew
);
7368 /* Release the reference to the new page. */
7374 #endif /* SQLITE_OMIT_QUICKBALANCE */
7378 ** This function does not contribute anything to the operation of SQLite.
7379 ** it is sometimes activated temporarily while debugging code responsible
7380 ** for setting pointer-map entries.
7382 static int ptrmapCheckPages(MemPage
**apPage
, int nPage
){
7384 for(i
=0; i
<nPage
; i
++){
7387 MemPage
*pPage
= apPage
[i
];
7388 BtShared
*pBt
= pPage
->pBt
;
7389 assert( pPage
->isInit
);
7391 for(j
=0; j
<pPage
->nCell
; j
++){
7395 z
= findCell(pPage
, j
);
7396 pPage
->xParseCell(pPage
, z
, &info
);
7397 if( info
.nLocal
<info
.nPayload
){
7398 Pgno ovfl
= get4byte(&z
[info
.nSize
-4]);
7399 ptrmapGet(pBt
, ovfl
, &e
, &n
);
7400 assert( n
==pPage
->pgno
&& e
==PTRMAP_OVERFLOW1
);
7403 Pgno child
= get4byte(z
);
7404 ptrmapGet(pBt
, child
, &e
, &n
);
7405 assert( n
==pPage
->pgno
&& e
==PTRMAP_BTREE
);
7409 Pgno child
= get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]);
7410 ptrmapGet(pBt
, child
, &e
, &n
);
7411 assert( n
==pPage
->pgno
&& e
==PTRMAP_BTREE
);
7419 ** This function is used to copy the contents of the b-tree node stored
7420 ** on page pFrom to page pTo. If page pFrom was not a leaf page, then
7421 ** the pointer-map entries for each child page are updated so that the
7422 ** parent page stored in the pointer map is page pTo. If pFrom contained
7423 ** any cells with overflow page pointers, then the corresponding pointer
7424 ** map entries are also updated so that the parent page is page pTo.
7426 ** If pFrom is currently carrying any overflow cells (entries in the
7427 ** MemPage.apOvfl[] array), they are not copied to pTo.
7429 ** Before returning, page pTo is reinitialized using btreeInitPage().
7431 ** The performance of this function is not critical. It is only used by
7432 ** the balance_shallower() and balance_deeper() procedures, neither of
7433 ** which are called often under normal circumstances.
7435 static void copyNodeContent(MemPage
*pFrom
, MemPage
*pTo
, int *pRC
){
7436 if( (*pRC
)==SQLITE_OK
){
7437 BtShared
* const pBt
= pFrom
->pBt
;
7438 u8
* const aFrom
= pFrom
->aData
;
7439 u8
* const aTo
= pTo
->aData
;
7440 int const iFromHdr
= pFrom
->hdrOffset
;
7441 int const iToHdr
= ((pTo
->pgno
==1) ? 100 : 0);
7446 assert( pFrom
->isInit
);
7447 assert( pFrom
->nFree
>=iToHdr
);
7448 assert( get2byte(&aFrom
[iFromHdr
+5]) <= (int)pBt
->usableSize
);
7450 /* Copy the b-tree node content from page pFrom to page pTo. */
7451 iData
= get2byte(&aFrom
[iFromHdr
+5]);
7452 memcpy(&aTo
[iData
], &aFrom
[iData
], pBt
->usableSize
-iData
);
7453 memcpy(&aTo
[iToHdr
], &aFrom
[iFromHdr
], pFrom
->cellOffset
+ 2*pFrom
->nCell
);
7455 /* Reinitialize page pTo so that the contents of the MemPage structure
7456 ** match the new data. The initialization of pTo can actually fail under
7457 ** fairly obscure circumstances, even though it is a copy of initialized
7461 rc
= btreeInitPage(pTo
);
7462 if( rc
==SQLITE_OK
) rc
= btreeComputeFreeSpace(pTo
);
7463 if( rc
!=SQLITE_OK
){
7468 /* If this is an auto-vacuum database, update the pointer-map entries
7469 ** for any b-tree or overflow pages that pTo now contains the pointers to.
7472 *pRC
= setChildPtrmaps(pTo
);
7478 ** This routine redistributes cells on the iParentIdx'th child of pParent
7479 ** (hereafter "the page") and up to 2 siblings so that all pages have about the
7480 ** same amount of free space. Usually a single sibling on either side of the
7481 ** page are used in the balancing, though both siblings might come from one
7482 ** side if the page is the first or last child of its parent. If the page
7483 ** has fewer than 2 siblings (something which can only happen if the page
7484 ** is a root page or a child of a root page) then all available siblings
7485 ** participate in the balancing.
7487 ** The number of siblings of the page might be increased or decreased by
7488 ** one or two in an effort to keep pages nearly full but not over full.
7490 ** Note that when this routine is called, some of the cells on the page
7491 ** might not actually be stored in MemPage.aData[]. This can happen
7492 ** if the page is overfull. This routine ensures that all cells allocated
7493 ** to the page and its siblings fit into MemPage.aData[] before returning.
7495 ** In the course of balancing the page and its siblings, cells may be
7496 ** inserted into or removed from the parent page (pParent). Doing so
7497 ** may cause the parent page to become overfull or underfull. If this
7498 ** happens, it is the responsibility of the caller to invoke the correct
7499 ** balancing routine to fix this problem (see the balance() routine).
7501 ** If this routine fails for any reason, it might leave the database
7502 ** in a corrupted state. So if this routine fails, the database should
7505 ** The third argument to this function, aOvflSpace, is a pointer to a
7506 ** buffer big enough to hold one page. If while inserting cells into the parent
7507 ** page (pParent) the parent page becomes overfull, this buffer is
7508 ** used to store the parent's overflow cells. Because this function inserts
7509 ** a maximum of four divider cells into the parent page, and the maximum
7510 ** size of a cell stored within an internal node is always less than 1/4
7511 ** of the page-size, the aOvflSpace[] buffer is guaranteed to be large
7512 ** enough for all overflow cells.
7514 ** If aOvflSpace is set to a null pointer, this function returns
7517 static int balance_nonroot(
7518 MemPage
*pParent
, /* Parent page of siblings being balanced */
7519 int iParentIdx
, /* Index of "the page" in pParent */
7520 u8
*aOvflSpace
, /* page-size bytes of space for parent ovfl */
7521 int isRoot
, /* True if pParent is a root-page */
7522 int bBulk
/* True if this call is part of a bulk load */
7524 BtShared
*pBt
; /* The whole database */
7525 int nMaxCells
= 0; /* Allocated size of apCell, szCell, aFrom. */
7526 int nNew
= 0; /* Number of pages in apNew[] */
7527 int nOld
; /* Number of pages in apOld[] */
7528 int i
, j
, k
; /* Loop counters */
7529 int nxDiv
; /* Next divider slot in pParent->aCell[] */
7530 int rc
= SQLITE_OK
; /* The return code */
7531 u16 leafCorrection
; /* 4 if pPage is a leaf. 0 if not */
7532 int leafData
; /* True if pPage is a leaf of a LEAFDATA tree */
7533 int usableSpace
; /* Bytes in pPage beyond the header */
7534 int pageFlags
; /* Value of pPage->aData[0] */
7535 int iSpace1
= 0; /* First unused byte of aSpace1[] */
7536 int iOvflSpace
= 0; /* First unused byte of aOvflSpace[] */
7537 int szScratch
; /* Size of scratch memory requested */
7538 MemPage
*apOld
[NB
]; /* pPage and up to two siblings */
7539 MemPage
*apNew
[NB
+2]; /* pPage and up to NB siblings after balancing */
7540 u8
*pRight
; /* Location in parent of right-sibling pointer */
7541 u8
*apDiv
[NB
-1]; /* Divider cells in pParent */
7542 int cntNew
[NB
+2]; /* Index in b.paCell[] of cell after i-th page */
7543 int cntOld
[NB
+2]; /* Old index in b.apCell[] */
7544 int szNew
[NB
+2]; /* Combined size of cells placed on i-th page */
7545 u8
*aSpace1
; /* Space for copies of dividers cells */
7546 Pgno pgno
; /* Temp var to store a page number in */
7547 u8 abDone
[NB
+2]; /* True after i'th new page is populated */
7548 Pgno aPgno
[NB
+2]; /* Page numbers of new pages before shuffling */
7549 Pgno aPgOrder
[NB
+2]; /* Copy of aPgno[] used for sorting pages */
7550 u16 aPgFlags
[NB
+2]; /* flags field of new pages before shuffling */
7551 CellArray b
; /* Parsed information on cells being balanced */
7553 memset(abDone
, 0, sizeof(abDone
));
7557 assert( sqlite3_mutex_held(pBt
->mutex
) );
7558 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
7560 /* At this point pParent may have at most one overflow cell. And if
7561 ** this overflow cell is present, it must be the cell with
7562 ** index iParentIdx. This scenario comes about when this function
7563 ** is called (indirectly) from sqlite3BtreeDelete().
7565 assert( pParent
->nOverflow
==0 || pParent
->nOverflow
==1 );
7566 assert( pParent
->nOverflow
==0 || pParent
->aiOvfl
[0]==iParentIdx
);
7569 return SQLITE_NOMEM_BKPT
;
7571 assert( pParent
->nFree
>=0 );
7573 /* Find the sibling pages to balance. Also locate the cells in pParent
7574 ** that divide the siblings. An attempt is made to find NN siblings on
7575 ** either side of pPage. More siblings are taken from one side, however,
7576 ** if there are fewer than NN siblings on the other side. If pParent
7577 ** has NB or fewer children then all children of pParent are taken.
7579 ** This loop also drops the divider cells from the parent page. This
7580 ** way, the remainder of the function does not have to deal with any
7581 ** overflow cells in the parent page, since if any existed they will
7582 ** have already been removed.
7584 i
= pParent
->nOverflow
+ pParent
->nCell
;
7588 assert( bBulk
==0 || bBulk
==1 );
7589 if( iParentIdx
==0 ){
7591 }else if( iParentIdx
==i
){
7594 nxDiv
= iParentIdx
-1;
7599 if( (i
+nxDiv
-pParent
->nOverflow
)==pParent
->nCell
){
7600 pRight
= &pParent
->aData
[pParent
->hdrOffset
+8];
7602 pRight
= findCell(pParent
, i
+nxDiv
-pParent
->nOverflow
);
7604 pgno
= get4byte(pRight
);
7606 rc
= getAndInitPage(pBt
, pgno
, &apOld
[i
], 0, 0);
7608 memset(apOld
, 0, (i
+1)*sizeof(MemPage
*));
7609 goto balance_cleanup
;
7611 if( apOld
[i
]->nFree
<0 ){
7612 rc
= btreeComputeFreeSpace(apOld
[i
]);
7614 memset(apOld
, 0, (i
)*sizeof(MemPage
*));
7615 goto balance_cleanup
;
7618 if( (i
--)==0 ) break;
7620 if( pParent
->nOverflow
&& i
+nxDiv
==pParent
->aiOvfl
[0] ){
7621 apDiv
[i
] = pParent
->apOvfl
[0];
7622 pgno
= get4byte(apDiv
[i
]);
7623 szNew
[i
] = pParent
->xCellSize(pParent
, apDiv
[i
]);
7624 pParent
->nOverflow
= 0;
7626 apDiv
[i
] = findCell(pParent
, i
+nxDiv
-pParent
->nOverflow
);
7627 pgno
= get4byte(apDiv
[i
]);
7628 szNew
[i
] = pParent
->xCellSize(pParent
, apDiv
[i
]);
7630 /* Drop the cell from the parent page. apDiv[i] still points to
7631 ** the cell within the parent, even though it has been dropped.
7632 ** This is safe because dropping a cell only overwrites the first
7633 ** four bytes of it, and this function does not need the first
7634 ** four bytes of the divider cell. So the pointer is safe to use
7637 ** But not if we are in secure-delete mode. In secure-delete mode,
7638 ** the dropCell() routine will overwrite the entire cell with zeroes.
7639 ** In this case, temporarily copy the cell into the aOvflSpace[]
7640 ** buffer. It will be copied out again as soon as the aSpace[] buffer
7642 if( pBt
->btsFlags
& BTS_FAST_SECURE
){
7645 iOff
= SQLITE_PTR_TO_INT(apDiv
[i
]) - SQLITE_PTR_TO_INT(pParent
->aData
);
7646 if( (iOff
+szNew
[i
])>(int)pBt
->usableSize
){
7647 rc
= SQLITE_CORRUPT_BKPT
;
7648 memset(apOld
, 0, (i
+1)*sizeof(MemPage
*));
7649 goto balance_cleanup
;
7651 memcpy(&aOvflSpace
[iOff
], apDiv
[i
], szNew
[i
]);
7652 apDiv
[i
] = &aOvflSpace
[apDiv
[i
]-pParent
->aData
];
7655 dropCell(pParent
, i
+nxDiv
-pParent
->nOverflow
, szNew
[i
], &rc
);
7659 /* Make nMaxCells a multiple of 4 in order to preserve 8-byte
7661 nMaxCells
= nOld
*(MX_CELL(pBt
) + ArraySize(pParent
->apOvfl
));
7662 nMaxCells
= (nMaxCells
+ 3)&~3;
7665 ** Allocate space for memory structures
7668 nMaxCells
*sizeof(u8
*) /* b.apCell */
7669 + nMaxCells
*sizeof(u16
) /* b.szCell */
7670 + pBt
->pageSize
; /* aSpace1 */
7672 assert( szScratch
<=7*(int)pBt
->pageSize
);
7673 b
.apCell
= sqlite3StackAllocRaw(0, szScratch
);
7675 rc
= SQLITE_NOMEM_BKPT
;
7676 goto balance_cleanup
;
7678 b
.szCell
= (u16
*)&b
.apCell
[nMaxCells
];
7679 aSpace1
= (u8
*)&b
.szCell
[nMaxCells
];
7680 assert( EIGHT_BYTE_ALIGNMENT(aSpace1
) );
7683 ** Load pointers to all cells on sibling pages and the divider cells
7684 ** into the local b.apCell[] array. Make copies of the divider cells
7685 ** into space obtained from aSpace1[]. The divider cells have already
7686 ** been removed from pParent.
7688 ** If the siblings are on leaf pages, then the child pointers of the
7689 ** divider cells are stripped from the cells before they are copied
7690 ** into aSpace1[]. In this way, all cells in b.apCell[] are without
7691 ** child pointers. If siblings are not leaves, then all cell in
7692 ** b.apCell[] include child pointers. Either way, all cells in b.apCell[]
7695 ** leafCorrection: 4 if pPage is a leaf. 0 if pPage is not a leaf.
7696 ** leafData: 1 if pPage holds key+data and pParent holds only keys.
7699 leafCorrection
= b
.pRef
->leaf
*4;
7700 leafData
= b
.pRef
->intKeyLeaf
;
7701 for(i
=0; i
<nOld
; i
++){
7702 MemPage
*pOld
= apOld
[i
];
7703 int limit
= pOld
->nCell
;
7704 u8
*aData
= pOld
->aData
;
7705 u16 maskPage
= pOld
->maskPage
;
7706 u8
*piCell
= aData
+ pOld
->cellOffset
;
7708 VVA_ONLY( int nCellAtStart
= b
.nCell
; )
7710 /* Verify that all sibling pages are of the same "type" (table-leaf,
7711 ** table-interior, index-leaf, or index-interior).
7713 if( pOld
->aData
[0]!=apOld
[0]->aData
[0] ){
7714 rc
= SQLITE_CORRUPT_BKPT
;
7715 goto balance_cleanup
;
7718 /* Load b.apCell[] with pointers to all cells in pOld. If pOld
7719 ** contains overflow cells, include them in the b.apCell[] array
7720 ** in the correct spot.
7722 ** Note that when there are multiple overflow cells, it is always the
7723 ** case that they are sequential and adjacent. This invariant arises
7724 ** because multiple overflows can only occurs when inserting divider
7725 ** cells into a parent on a prior balance, and divider cells are always
7726 ** adjacent and are inserted in order. There is an assert() tagged
7727 ** with "NOTE 1" in the overflow cell insertion loop to prove this
7730 ** This must be done in advance. Once the balance starts, the cell
7731 ** offset section of the btree page will be overwritten and we will no
7732 ** long be able to find the cells if a pointer to each cell is not saved
7735 memset(&b
.szCell
[b
.nCell
], 0, sizeof(b
.szCell
[0])*(limit
+pOld
->nOverflow
));
7736 if( pOld
->nOverflow
>0 ){
7737 if( NEVER(limit
<pOld
->aiOvfl
[0]) ){
7738 rc
= SQLITE_CORRUPT_BKPT
;
7739 goto balance_cleanup
;
7741 limit
= pOld
->aiOvfl
[0];
7742 for(j
=0; j
<limit
; j
++){
7743 b
.apCell
[b
.nCell
] = aData
+ (maskPage
& get2byteAligned(piCell
));
7747 for(k
=0; k
<pOld
->nOverflow
; k
++){
7748 assert( k
==0 || pOld
->aiOvfl
[k
-1]+1==pOld
->aiOvfl
[k
] );/* NOTE 1 */
7749 b
.apCell
[b
.nCell
] = pOld
->apOvfl
[k
];
7753 piEnd
= aData
+ pOld
->cellOffset
+ 2*pOld
->nCell
;
7754 while( piCell
<piEnd
){
7755 assert( b
.nCell
<nMaxCells
);
7756 b
.apCell
[b
.nCell
] = aData
+ (maskPage
& get2byteAligned(piCell
));
7760 assert( (b
.nCell
-nCellAtStart
)==(pOld
->nCell
+pOld
->nOverflow
) );
7762 cntOld
[i
] = b
.nCell
;
7763 if( i
<nOld
-1 && !leafData
){
7764 u16 sz
= (u16
)szNew
[i
];
7766 assert( b
.nCell
<nMaxCells
);
7767 b
.szCell
[b
.nCell
] = sz
;
7768 pTemp
= &aSpace1
[iSpace1
];
7770 assert( sz
<=pBt
->maxLocal
+23 );
7771 assert( iSpace1
<= (int)pBt
->pageSize
);
7772 memcpy(pTemp
, apDiv
[i
], sz
);
7773 b
.apCell
[b
.nCell
] = pTemp
+leafCorrection
;
7774 assert( leafCorrection
==0 || leafCorrection
==4 );
7775 b
.szCell
[b
.nCell
] = b
.szCell
[b
.nCell
] - leafCorrection
;
7777 assert( leafCorrection
==0 );
7778 assert( pOld
->hdrOffset
==0 );
7779 /* The right pointer of the child page pOld becomes the left
7780 ** pointer of the divider cell */
7781 memcpy(b
.apCell
[b
.nCell
], &pOld
->aData
[8], 4);
7783 assert( leafCorrection
==4 );
7784 while( b
.szCell
[b
.nCell
]<4 ){
7785 /* Do not allow any cells smaller than 4 bytes. If a smaller cell
7786 ** does exist, pad it with 0x00 bytes. */
7787 assert( b
.szCell
[b
.nCell
]==3 || CORRUPT_DB
);
7788 assert( b
.apCell
[b
.nCell
]==&aSpace1
[iSpace1
-3] || CORRUPT_DB
);
7789 aSpace1
[iSpace1
++] = 0x00;
7790 b
.szCell
[b
.nCell
]++;
7798 ** Figure out the number of pages needed to hold all b.nCell cells.
7799 ** Store this number in "k". Also compute szNew[] which is the total
7800 ** size of all cells on the i-th page and cntNew[] which is the index
7801 ** in b.apCell[] of the cell that divides page i from page i+1.
7802 ** cntNew[k] should equal b.nCell.
7804 ** Values computed by this block:
7806 ** k: The total number of sibling pages
7807 ** szNew[i]: Spaced used on the i-th sibling page.
7808 ** cntNew[i]: Index in b.apCell[] and b.szCell[] for the first cell to
7809 ** the right of the i-th sibling page.
7810 ** usableSpace: Number of bytes of space available on each sibling.
7813 usableSpace
= pBt
->usableSize
- 12 + leafCorrection
;
7814 for(i
=k
=0; i
<nOld
; i
++, k
++){
7815 MemPage
*p
= apOld
[i
];
7816 b
.apEnd
[k
] = p
->aDataEnd
;
7817 b
.ixNx
[k
] = cntOld
[i
];
7818 if( k
&& b
.ixNx
[k
]==b
.ixNx
[k
-1] ){
7819 k
--; /* Omit b.ixNx[] entry for child pages with no cells */
7823 b
.apEnd
[k
] = pParent
->aDataEnd
;
7824 b
.ixNx
[k
] = cntOld
[i
]+1;
7826 assert( p
->nFree
>=0 );
7827 szNew
[i
] = usableSpace
- p
->nFree
;
7828 for(j
=0; j
<p
->nOverflow
; j
++){
7829 szNew
[i
] += 2 + p
->xCellSize(p
, p
->apOvfl
[j
]);
7831 cntNew
[i
] = cntOld
[i
];
7836 while( szNew
[i
]>usableSpace
){
7839 if( k
>NB
+2 ){ rc
= SQLITE_CORRUPT_BKPT
; goto balance_cleanup
; }
7841 cntNew
[k
-1] = b
.nCell
;
7843 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]-1);
7846 if( cntNew
[i
]<b
.nCell
){
7847 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7855 while( cntNew
[i
]<b
.nCell
){
7856 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7857 if( szNew
[i
]+sz
>usableSpace
) break;
7861 if( cntNew
[i
]<b
.nCell
){
7862 sz
= 2 + cachedCellSize(&b
, cntNew
[i
]);
7869 if( cntNew
[i
]>=b
.nCell
){
7871 }else if( cntNew
[i
] <= (i
>0 ? cntNew
[i
-1] : 0) ){
7872 rc
= SQLITE_CORRUPT_BKPT
;
7873 goto balance_cleanup
;
7878 ** The packing computed by the previous block is biased toward the siblings
7879 ** on the left side (siblings with smaller keys). The left siblings are
7880 ** always nearly full, while the right-most sibling might be nearly empty.
7881 ** The next block of code attempts to adjust the packing of siblings to
7882 ** get a better balance.
7884 ** This adjustment is more than an optimization. The packing above might
7885 ** be so out of balance as to be illegal. For example, the right-most
7886 ** sibling might be completely empty. This adjustment is not optional.
7888 for(i
=k
-1; i
>0; i
--){
7889 int szRight
= szNew
[i
]; /* Size of sibling on the right */
7890 int szLeft
= szNew
[i
-1]; /* Size of sibling on the left */
7891 int r
; /* Index of right-most cell in left sibling */
7892 int d
; /* Index of first cell to the left of right sibling */
7894 r
= cntNew
[i
-1] - 1;
7895 d
= r
+ 1 - leafData
;
7896 (void)cachedCellSize(&b
, d
);
7898 assert( d
<nMaxCells
);
7899 assert( r
<nMaxCells
);
7900 (void)cachedCellSize(&b
, r
);
7902 && (bBulk
|| szRight
+b
.szCell
[d
]+2 > szLeft
-(b
.szCell
[r
]+(i
==k
-1?0:2)))){
7905 szRight
+= b
.szCell
[d
] + 2;
7906 szLeft
-= b
.szCell
[r
] + 2;
7912 szNew
[i
-1] = szLeft
;
7913 if( cntNew
[i
-1] <= (i
>1 ? cntNew
[i
-2] : 0) ){
7914 rc
= SQLITE_CORRUPT_BKPT
;
7915 goto balance_cleanup
;
7919 /* Sanity check: For a non-corrupt database file one of the follwing
7921 ** (1) We found one or more cells (cntNew[0])>0), or
7922 ** (2) pPage is a virtual root page. A virtual root page is when
7923 ** the real root page is page 1 and we are the only child of
7926 assert( cntNew
[0]>0 || (pParent
->pgno
==1 && pParent
->nCell
==0) || CORRUPT_DB
);
7927 TRACE(("BALANCE: old: %d(nc=%d) %d(nc=%d) %d(nc=%d)\n",
7928 apOld
[0]->pgno
, apOld
[0]->nCell
,
7929 nOld
>=2 ? apOld
[1]->pgno
: 0, nOld
>=2 ? apOld
[1]->nCell
: 0,
7930 nOld
>=3 ? apOld
[2]->pgno
: 0, nOld
>=3 ? apOld
[2]->nCell
: 0
7934 ** Allocate k new pages. Reuse old pages where possible.
7936 pageFlags
= apOld
[0]->aData
[0];
7940 pNew
= apNew
[i
] = apOld
[i
];
7942 rc
= sqlite3PagerWrite(pNew
->pDbPage
);
7944 if( rc
) goto balance_cleanup
;
7947 rc
= allocateBtreePage(pBt
, &pNew
, &pgno
, (bBulk
? 1 : pgno
), 0);
7948 if( rc
) goto balance_cleanup
;
7949 zeroPage(pNew
, pageFlags
);
7952 cntOld
[i
] = b
.nCell
;
7954 /* Set the pointer-map entry for the new sibling page. */
7956 ptrmapPut(pBt
, pNew
->pgno
, PTRMAP_BTREE
, pParent
->pgno
, &rc
);
7957 if( rc
!=SQLITE_OK
){
7958 goto balance_cleanup
;
7965 ** Reassign page numbers so that the new pages are in ascending order.
7966 ** This helps to keep entries in the disk file in order so that a scan
7967 ** of the table is closer to a linear scan through the file. That in turn
7968 ** helps the operating system to deliver pages from the disk more rapidly.
7970 ** An O(n^2) insertion sort algorithm is used, but since n is never more
7971 ** than (NB+2) (a small constant), that should not be a problem.
7973 ** When NB==3, this one optimization makes the database about 25% faster
7974 ** for large insertions and deletions.
7976 for(i
=0; i
<nNew
; i
++){
7977 aPgOrder
[i
] = aPgno
[i
] = apNew
[i
]->pgno
;
7978 aPgFlags
[i
] = apNew
[i
]->pDbPage
->flags
;
7980 if( aPgno
[j
]==aPgno
[i
] ){
7981 /* This branch is taken if the set of sibling pages somehow contains
7982 ** duplicate entries. This can happen if the database is corrupt.
7983 ** It would be simpler to detect this as part of the loop below, but
7984 ** we do the detection here in order to avoid populating the pager
7985 ** cache with two separate objects associated with the same
7987 assert( CORRUPT_DB
);
7988 rc
= SQLITE_CORRUPT_BKPT
;
7989 goto balance_cleanup
;
7993 for(i
=0; i
<nNew
; i
++){
7994 int iBest
= 0; /* aPgno[] index of page number to use */
7995 for(j
=1; j
<nNew
; j
++){
7996 if( aPgOrder
[j
]<aPgOrder
[iBest
] ) iBest
= j
;
7998 pgno
= aPgOrder
[iBest
];
7999 aPgOrder
[iBest
] = 0xffffffff;
8002 sqlite3PagerRekey(apNew
[iBest
]->pDbPage
, pBt
->nPage
+iBest
+1, 0);
8004 sqlite3PagerRekey(apNew
[i
]->pDbPage
, pgno
, aPgFlags
[iBest
]);
8005 apNew
[i
]->pgno
= pgno
;
8009 TRACE(("BALANCE: new: %d(%d nc=%d) %d(%d nc=%d) %d(%d nc=%d) "
8010 "%d(%d nc=%d) %d(%d nc=%d)\n",
8011 apNew
[0]->pgno
, szNew
[0], cntNew
[0],
8012 nNew
>=2 ? apNew
[1]->pgno
: 0, nNew
>=2 ? szNew
[1] : 0,
8013 nNew
>=2 ? cntNew
[1] - cntNew
[0] - !leafData
: 0,
8014 nNew
>=3 ? apNew
[2]->pgno
: 0, nNew
>=3 ? szNew
[2] : 0,
8015 nNew
>=3 ? cntNew
[2] - cntNew
[1] - !leafData
: 0,
8016 nNew
>=4 ? apNew
[3]->pgno
: 0, nNew
>=4 ? szNew
[3] : 0,
8017 nNew
>=4 ? cntNew
[3] - cntNew
[2] - !leafData
: 0,
8018 nNew
>=5 ? apNew
[4]->pgno
: 0, nNew
>=5 ? szNew
[4] : 0,
8019 nNew
>=5 ? cntNew
[4] - cntNew
[3] - !leafData
: 0
8022 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
8023 assert( nNew
>=1 && nNew
<=ArraySize(apNew
) );
8024 assert( apNew
[nNew
-1]!=0 );
8025 put4byte(pRight
, apNew
[nNew
-1]->pgno
);
8027 /* If the sibling pages are not leaves, ensure that the right-child pointer
8028 ** of the right-most new sibling page is set to the value that was
8029 ** originally in the same field of the right-most old sibling page. */
8030 if( (pageFlags
& PTF_LEAF
)==0 && nOld
!=nNew
){
8031 MemPage
*pOld
= (nNew
>nOld
? apNew
: apOld
)[nOld
-1];
8032 memcpy(&apNew
[nNew
-1]->aData
[8], &pOld
->aData
[8], 4);
8035 /* Make any required updates to pointer map entries associated with
8036 ** cells stored on sibling pages following the balance operation. Pointer
8037 ** map entries associated with divider cells are set by the insertCell()
8038 ** routine. The associated pointer map entries are:
8040 ** a) if the cell contains a reference to an overflow chain, the
8041 ** entry associated with the first page in the overflow chain, and
8043 ** b) if the sibling pages are not leaves, the child page associated
8046 ** If the sibling pages are not leaves, then the pointer map entry
8047 ** associated with the right-child of each sibling may also need to be
8048 ** updated. This happens below, after the sibling pages have been
8049 ** populated, not here.
8053 MemPage
*pNew
= pOld
= apNew
[0];
8054 int cntOldNext
= pNew
->nCell
+ pNew
->nOverflow
;
8058 for(i
=0; i
<b
.nCell
; i
++){
8059 u8
*pCell
= b
.apCell
[i
];
8060 while( i
==cntOldNext
){
8062 assert( iOld
<nNew
|| iOld
<nOld
);
8063 assert( iOld
>=0 && iOld
<NB
);
8064 pOld
= iOld
<nNew
? apNew
[iOld
] : apOld
[iOld
];
8065 cntOldNext
+= pOld
->nCell
+ pOld
->nOverflow
+ !leafData
;
8067 if( i
==cntNew
[iNew
] ){
8068 pNew
= apNew
[++iNew
];
8069 if( !leafData
) continue;
8072 /* Cell pCell is destined for new sibling page pNew. Originally, it
8073 ** was either part of sibling page iOld (possibly an overflow cell),
8074 ** or else the divider cell to the left of sibling page iOld. So,
8075 ** if sibling page iOld had the same page number as pNew, and if
8076 ** pCell really was a part of sibling page iOld (not a divider or
8077 ** overflow cell), we can skip updating the pointer map entries. */
8079 || pNew
->pgno
!=aPgno
[iOld
]
8080 || !SQLITE_WITHIN(pCell
,pOld
->aData
,pOld
->aDataEnd
)
8082 if( !leafCorrection
){
8083 ptrmapPut(pBt
, get4byte(pCell
), PTRMAP_BTREE
, pNew
->pgno
, &rc
);
8085 if( cachedCellSize(&b
,i
)>pNew
->minLocal
){
8086 ptrmapPutOvflPtr(pNew
, pOld
, pCell
, &rc
);
8088 if( rc
) goto balance_cleanup
;
8093 /* Insert new divider cells into pParent. */
8094 for(i
=0; i
<nNew
-1; i
++){
8098 MemPage
*pNew
= apNew
[i
];
8101 assert( j
<nMaxCells
);
8102 assert( b
.apCell
[j
]!=0 );
8103 pCell
= b
.apCell
[j
];
8104 sz
= b
.szCell
[j
] + leafCorrection
;
8105 pTemp
= &aOvflSpace
[iOvflSpace
];
8107 memcpy(&pNew
->aData
[8], pCell
, 4);
8108 }else if( leafData
){
8109 /* If the tree is a leaf-data tree, and the siblings are leaves,
8110 ** then there is no divider cell in b.apCell[]. Instead, the divider
8111 ** cell consists of the integer key for the right-most cell of
8112 ** the sibling-page assembled above only.
8116 pNew
->xParseCell(pNew
, b
.apCell
[j
], &info
);
8118 sz
= 4 + putVarint(&pCell
[4], info
.nKey
);
8122 /* Obscure case for non-leaf-data trees: If the cell at pCell was
8123 ** previously stored on a leaf node, and its reported size was 4
8124 ** bytes, then it may actually be smaller than this
8125 ** (see btreeParseCellPtr(), 4 bytes is the minimum size of
8126 ** any cell). But it is important to pass the correct size to
8127 ** insertCell(), so reparse the cell now.
8129 ** This can only happen for b-trees used to evaluate "IN (SELECT ...)"
8130 ** and WITHOUT ROWID tables with exactly one column which is the
8133 if( b
.szCell
[j
]==4 ){
8134 assert(leafCorrection
==4);
8135 sz
= pParent
->xCellSize(pParent
, pCell
);
8139 assert( sz
<=pBt
->maxLocal
+23 );
8140 assert( iOvflSpace
<= (int)pBt
->pageSize
);
8141 insertCell(pParent
, nxDiv
+i
, pCell
, sz
, pTemp
, pNew
->pgno
, &rc
);
8142 if( rc
!=SQLITE_OK
) goto balance_cleanup
;
8143 assert( sqlite3PagerIswriteable(pParent
->pDbPage
) );
8146 /* Now update the actual sibling pages. The order in which they are updated
8147 ** is important, as this code needs to avoid disrupting any page from which
8148 ** cells may still to be read. In practice, this means:
8150 ** (1) If cells are moving left (from apNew[iPg] to apNew[iPg-1])
8151 ** then it is not safe to update page apNew[iPg] until after
8152 ** the left-hand sibling apNew[iPg-1] has been updated.
8154 ** (2) If cells are moving right (from apNew[iPg] to apNew[iPg+1])
8155 ** then it is not safe to update page apNew[iPg] until after
8156 ** the right-hand sibling apNew[iPg+1] has been updated.
8158 ** If neither of the above apply, the page is safe to update.
8160 ** The iPg value in the following loop starts at nNew-1 goes down
8161 ** to 0, then back up to nNew-1 again, thus making two passes over
8162 ** the pages. On the initial downward pass, only condition (1) above
8163 ** needs to be tested because (2) will always be true from the previous
8164 ** step. On the upward pass, both conditions are always true, so the
8165 ** upwards pass simply processes pages that were missed on the downward
8168 for(i
=1-nNew
; i
<nNew
; i
++){
8169 int iPg
= i
<0 ? -i
: i
;
8170 assert( iPg
>=0 && iPg
<nNew
);
8171 if( abDone
[iPg
] ) continue; /* Skip pages already processed */
8172 if( i
>=0 /* On the upwards pass, or... */
8173 || cntOld
[iPg
-1]>=cntNew
[iPg
-1] /* Condition (1) is true */
8179 /* Verify condition (1): If cells are moving left, update iPg
8180 ** only after iPg-1 has already been updated. */
8181 assert( iPg
==0 || cntOld
[iPg
-1]>=cntNew
[iPg
-1] || abDone
[iPg
-1] );
8183 /* Verify condition (2): If cells are moving right, update iPg
8184 ** only after iPg+1 has already been updated. */
8185 assert( cntNew
[iPg
]>=cntOld
[iPg
] || abDone
[iPg
+1] );
8189 nNewCell
= cntNew
[0];
8191 iOld
= iPg
<nOld
? (cntOld
[iPg
-1] + !leafData
) : b
.nCell
;
8192 iNew
= cntNew
[iPg
-1] + !leafData
;
8193 nNewCell
= cntNew
[iPg
] - iNew
;
8196 rc
= editPage(apNew
[iPg
], iOld
, iNew
, nNewCell
, &b
);
8197 if( rc
) goto balance_cleanup
;
8199 apNew
[iPg
]->nFree
= usableSpace
-szNew
[iPg
];
8200 assert( apNew
[iPg
]->nOverflow
==0 );
8201 assert( apNew
[iPg
]->nCell
==nNewCell
);
8205 /* All pages have been processed exactly once */
8206 assert( memcmp(abDone
, "\01\01\01\01\01", nNew
)==0 );
8211 if( isRoot
&& pParent
->nCell
==0 && pParent
->hdrOffset
<=apNew
[0]->nFree
){
8212 /* The root page of the b-tree now contains no cells. The only sibling
8213 ** page is the right-child of the parent. Copy the contents of the
8214 ** child page into the parent, decreasing the overall height of the
8215 ** b-tree structure by one. This is described as the "balance-shallower"
8216 ** sub-algorithm in some documentation.
8218 ** If this is an auto-vacuum database, the call to copyNodeContent()
8219 ** sets all pointer-map entries corresponding to database image pages
8220 ** for which the pointer is stored within the content being copied.
8222 ** It is critical that the child page be defragmented before being
8223 ** copied into the parent, because if the parent is page 1 then it will
8224 ** by smaller than the child due to the database header, and so all the
8225 ** free space needs to be up front.
8227 assert( nNew
==1 || CORRUPT_DB
);
8228 rc
= defragmentPage(apNew
[0], -1);
8229 testcase( rc
!=SQLITE_OK
);
8230 assert( apNew
[0]->nFree
==
8231 (get2byteNotZero(&apNew
[0]->aData
[5]) - apNew
[0]->cellOffset
8232 - apNew
[0]->nCell
*2)
8235 copyNodeContent(apNew
[0], pParent
, &rc
);
8236 freePage(apNew
[0], &rc
);
8237 }else if( ISAUTOVACUUM
&& !leafCorrection
){
8238 /* Fix the pointer map entries associated with the right-child of each
8239 ** sibling page. All other pointer map entries have already been taken
8241 for(i
=0; i
<nNew
; i
++){
8242 u32 key
= get4byte(&apNew
[i
]->aData
[8]);
8243 ptrmapPut(pBt
, key
, PTRMAP_BTREE
, apNew
[i
]->pgno
, &rc
);
8247 assert( pParent
->isInit
);
8248 TRACE(("BALANCE: finished: old=%d new=%d cells=%d\n",
8249 nOld
, nNew
, b
.nCell
));
8251 /* Free any old pages that were not reused as new pages.
8253 for(i
=nNew
; i
<nOld
; i
++){
8254 freePage(apOld
[i
], &rc
);
8258 if( ISAUTOVACUUM
&& rc
==SQLITE_OK
&& apNew
[0]->isInit
){
8259 /* The ptrmapCheckPages() contains assert() statements that verify that
8260 ** all pointer map pages are set correctly. This is helpful while
8261 ** debugging. This is usually disabled because a corrupt database may
8262 ** cause an assert() statement to fail. */
8263 ptrmapCheckPages(apNew
, nNew
);
8264 ptrmapCheckPages(&pParent
, 1);
8269 ** Cleanup before returning.
8272 sqlite3StackFree(0, b
.apCell
);
8273 for(i
=0; i
<nOld
; i
++){
8274 releasePage(apOld
[i
]);
8276 for(i
=0; i
<nNew
; i
++){
8277 releasePage(apNew
[i
]);
8285 ** This function is called when the root page of a b-tree structure is
8286 ** overfull (has one or more overflow pages).
8288 ** A new child page is allocated and the contents of the current root
8289 ** page, including overflow cells, are copied into the child. The root
8290 ** page is then overwritten to make it an empty page with the right-child
8291 ** pointer pointing to the new page.
8293 ** Before returning, all pointer-map entries corresponding to pages
8294 ** that the new child-page now contains pointers to are updated. The
8295 ** entry corresponding to the new right-child pointer of the root
8296 ** page is also updated.
8298 ** If successful, *ppChild is set to contain a reference to the child
8299 ** page and SQLITE_OK is returned. In this case the caller is required
8300 ** to call releasePage() on *ppChild exactly once. If an error occurs,
8301 ** an error code is returned and *ppChild is set to 0.
8303 static int balance_deeper(MemPage
*pRoot
, MemPage
**ppChild
){
8304 int rc
; /* Return value from subprocedures */
8305 MemPage
*pChild
= 0; /* Pointer to a new child page */
8306 Pgno pgnoChild
= 0; /* Page number of the new child page */
8307 BtShared
*pBt
= pRoot
->pBt
; /* The BTree */
8309 assert( pRoot
->nOverflow
>0 );
8310 assert( sqlite3_mutex_held(pBt
->mutex
) );
8312 /* Make pRoot, the root page of the b-tree, writable. Allocate a new
8313 ** page that will become the new right-child of pPage. Copy the contents
8314 ** of the node stored on pRoot into the new child page.
8316 rc
= sqlite3PagerWrite(pRoot
->pDbPage
);
8317 if( rc
==SQLITE_OK
){
8318 rc
= allocateBtreePage(pBt
,&pChild
,&pgnoChild
,pRoot
->pgno
,0);
8319 copyNodeContent(pRoot
, pChild
, &rc
);
8321 ptrmapPut(pBt
, pgnoChild
, PTRMAP_BTREE
, pRoot
->pgno
, &rc
);
8326 releasePage(pChild
);
8329 assert( sqlite3PagerIswriteable(pChild
->pDbPage
) );
8330 assert( sqlite3PagerIswriteable(pRoot
->pDbPage
) );
8331 assert( pChild
->nCell
==pRoot
->nCell
|| CORRUPT_DB
);
8333 TRACE(("BALANCE: copy root %d into %d\n", pRoot
->pgno
, pChild
->pgno
));
8335 /* Copy the overflow cells from pRoot to pChild */
8336 memcpy(pChild
->aiOvfl
, pRoot
->aiOvfl
,
8337 pRoot
->nOverflow
*sizeof(pRoot
->aiOvfl
[0]));
8338 memcpy(pChild
->apOvfl
, pRoot
->apOvfl
,
8339 pRoot
->nOverflow
*sizeof(pRoot
->apOvfl
[0]));
8340 pChild
->nOverflow
= pRoot
->nOverflow
;
8342 /* Zero the contents of pRoot. Then install pChild as the right-child. */
8343 zeroPage(pRoot
, pChild
->aData
[0] & ~PTF_LEAF
);
8344 put4byte(&pRoot
->aData
[pRoot
->hdrOffset
+8], pgnoChild
);
8351 ** Return SQLITE_CORRUPT if any cursor other than pCur is currently valid
8352 ** on the same B-tree as pCur.
8354 ** This can if a database is corrupt with two or more SQL tables
8355 ** pointing to the same b-tree. If an insert occurs on one SQL table
8356 ** and causes a BEFORE TRIGGER to do a secondary insert on the other SQL
8357 ** table linked to the same b-tree. If the secondary insert causes a
8358 ** rebalance, that can change content out from under the cursor on the
8359 ** first SQL table, violating invariants on the first insert.
8361 static int anotherValidCursor(BtCursor
*pCur
){
8363 for(pOther
=pCur
->pBt
->pCursor
; pOther
; pOther
=pOther
->pNext
){
8365 && pOther
->eState
==CURSOR_VALID
8366 && pOther
->pPage
==pCur
->pPage
8368 return SQLITE_CORRUPT_BKPT
;
8375 ** The page that pCur currently points to has just been modified in
8376 ** some way. This function figures out if this modification means the
8377 ** tree needs to be balanced, and if so calls the appropriate balancing
8378 ** routine. Balancing routines are:
8382 ** balance_nonroot()
8384 static int balance(BtCursor
*pCur
){
8386 const int nMin
= pCur
->pBt
->usableSize
* 2 / 3;
8387 u8 aBalanceQuickSpace
[13];
8390 VVA_ONLY( int balance_quick_called
= 0 );
8391 VVA_ONLY( int balance_deeper_called
= 0 );
8395 MemPage
*pPage
= pCur
->pPage
;
8397 if( NEVER(pPage
->nFree
<0) && btreeComputeFreeSpace(pPage
) ) break;
8398 if( pPage
->nOverflow
==0 && pPage
->nFree
<=nMin
){
8400 }else if( (iPage
= pCur
->iPage
)==0 ){
8401 if( pPage
->nOverflow
&& (rc
= anotherValidCursor(pCur
))==SQLITE_OK
){
8402 /* The root page of the b-tree is overfull. In this case call the
8403 ** balance_deeper() function to create a new child for the root-page
8404 ** and copy the current contents of the root-page to it. The
8405 ** next iteration of the do-loop will balance the child page.
8407 assert( balance_deeper_called
==0 );
8408 VVA_ONLY( balance_deeper_called
++ );
8409 rc
= balance_deeper(pPage
, &pCur
->apPage
[1]);
8410 if( rc
==SQLITE_OK
){
8414 pCur
->apPage
[0] = pPage
;
8415 pCur
->pPage
= pCur
->apPage
[1];
8416 assert( pCur
->pPage
->nOverflow
);
8422 MemPage
* const pParent
= pCur
->apPage
[iPage
-1];
8423 int const iIdx
= pCur
->aiIdx
[iPage
-1];
8425 rc
= sqlite3PagerWrite(pParent
->pDbPage
);
8426 if( rc
==SQLITE_OK
&& pParent
->nFree
<0 ){
8427 rc
= btreeComputeFreeSpace(pParent
);
8429 if( rc
==SQLITE_OK
){
8430 #ifndef SQLITE_OMIT_QUICKBALANCE
8431 if( pPage
->intKeyLeaf
8432 && pPage
->nOverflow
==1
8433 && pPage
->aiOvfl
[0]==pPage
->nCell
8435 && pParent
->nCell
==iIdx
8437 /* Call balance_quick() to create a new sibling of pPage on which
8438 ** to store the overflow cell. balance_quick() inserts a new cell
8439 ** into pParent, which may cause pParent overflow. If this
8440 ** happens, the next iteration of the do-loop will balance pParent
8441 ** use either balance_nonroot() or balance_deeper(). Until this
8442 ** happens, the overflow cell is stored in the aBalanceQuickSpace[]
8445 ** The purpose of the following assert() is to check that only a
8446 ** single call to balance_quick() is made for each call to this
8447 ** function. If this were not verified, a subtle bug involving reuse
8448 ** of the aBalanceQuickSpace[] might sneak in.
8450 assert( balance_quick_called
==0 );
8451 VVA_ONLY( balance_quick_called
++ );
8452 rc
= balance_quick(pParent
, pPage
, aBalanceQuickSpace
);
8456 /* In this case, call balance_nonroot() to redistribute cells
8457 ** between pPage and up to 2 of its sibling pages. This involves
8458 ** modifying the contents of pParent, which may cause pParent to
8459 ** become overfull or underfull. The next iteration of the do-loop
8460 ** will balance the parent page to correct this.
8462 ** If the parent page becomes overfull, the overflow cell or cells
8463 ** are stored in the pSpace buffer allocated immediately below.
8464 ** A subsequent iteration of the do-loop will deal with this by
8465 ** calling balance_nonroot() (balance_deeper() may be called first,
8466 ** but it doesn't deal with overflow cells - just moves them to a
8467 ** different page). Once this subsequent call to balance_nonroot()
8468 ** has completed, it is safe to release the pSpace buffer used by
8469 ** the previous call, as the overflow cell data will have been
8470 ** copied either into the body of a database page or into the new
8471 ** pSpace buffer passed to the latter call to balance_nonroot().
8473 u8
*pSpace
= sqlite3PageMalloc(pCur
->pBt
->pageSize
);
8474 rc
= balance_nonroot(pParent
, iIdx
, pSpace
, iPage
==1,
8475 pCur
->hints
&BTREE_BULKLOAD
);
8477 /* If pFree is not NULL, it points to the pSpace buffer used
8478 ** by a previous call to balance_nonroot(). Its contents are
8479 ** now stored either on real database pages or within the
8480 ** new pSpace buffer, so it may be safely freed here. */
8481 sqlite3PageFree(pFree
);
8484 /* The pSpace buffer will be freed after the next call to
8485 ** balance_nonroot(), or just before this function returns, whichever
8491 pPage
->nOverflow
= 0;
8493 /* The next iteration of the do-loop balances the parent page. */
8496 assert( pCur
->iPage
>=0 );
8497 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
8499 }while( rc
==SQLITE_OK
);
8502 sqlite3PageFree(pFree
);
8507 /* Overwrite content from pX into pDest. Only do the write if the
8508 ** content is different from what is already there.
8510 static int btreeOverwriteContent(
8511 MemPage
*pPage
, /* MemPage on which writing will occur */
8512 u8
*pDest
, /* Pointer to the place to start writing */
8513 const BtreePayload
*pX
, /* Source of data to write */
8514 int iOffset
, /* Offset of first byte to write */
8515 int iAmt
/* Number of bytes to be written */
8517 int nData
= pX
->nData
- iOffset
;
8519 /* Overwritting with zeros */
8521 for(i
=0; i
<iAmt
&& pDest
[i
]==0; i
++){}
8523 int rc
= sqlite3PagerWrite(pPage
->pDbPage
);
8525 memset(pDest
+ i
, 0, iAmt
- i
);
8529 /* Mixed read data and zeros at the end. Make a recursive call
8530 ** to write the zeros then fall through to write the real data */
8531 int rc
= btreeOverwriteContent(pPage
, pDest
+nData
, pX
, iOffset
+nData
,
8536 if( memcmp(pDest
, ((u8
*)pX
->pData
) + iOffset
, iAmt
)!=0 ){
8537 int rc
= sqlite3PagerWrite(pPage
->pDbPage
);
8539 /* In a corrupt database, it is possible for the source and destination
8540 ** buffers to overlap. This is harmless since the database is already
8541 ** corrupt but it does cause valgrind and ASAN warnings. So use
8543 memmove(pDest
, ((u8
*)pX
->pData
) + iOffset
, iAmt
);
8550 ** Overwrite the cell that cursor pCur is pointing to with fresh content
8553 static int btreeOverwriteCell(BtCursor
*pCur
, const BtreePayload
*pX
){
8554 int iOffset
; /* Next byte of pX->pData to write */
8555 int nTotal
= pX
->nData
+ pX
->nZero
; /* Total bytes of to write */
8556 int rc
; /* Return code */
8557 MemPage
*pPage
= pCur
->pPage
; /* Page being written */
8558 BtShared
*pBt
; /* Btree */
8559 Pgno ovflPgno
; /* Next overflow page to write */
8560 u32 ovflPageSize
; /* Size to write on overflow page */
8562 if( pCur
->info
.pPayload
+ pCur
->info
.nLocal
> pPage
->aDataEnd
8563 || pCur
->info
.pPayload
< pPage
->aData
+ pPage
->cellOffset
8565 return SQLITE_CORRUPT_BKPT
;
8567 /* Overwrite the local portion first */
8568 rc
= btreeOverwriteContent(pPage
, pCur
->info
.pPayload
, pX
,
8569 0, pCur
->info
.nLocal
);
8571 if( pCur
->info
.nLocal
==nTotal
) return SQLITE_OK
;
8573 /* Now overwrite the overflow pages */
8574 iOffset
= pCur
->info
.nLocal
;
8575 assert( nTotal
>=0 );
8576 assert( iOffset
>=0 );
8577 ovflPgno
= get4byte(pCur
->info
.pPayload
+ iOffset
);
8579 ovflPageSize
= pBt
->usableSize
- 4;
8581 rc
= btreeGetPage(pBt
, ovflPgno
, &pPage
, 0);
8583 if( sqlite3PagerPageRefcount(pPage
->pDbPage
)!=1 ){
8584 rc
= SQLITE_CORRUPT_BKPT
;
8586 if( iOffset
+ovflPageSize
<(u32
)nTotal
){
8587 ovflPgno
= get4byte(pPage
->aData
);
8589 ovflPageSize
= nTotal
- iOffset
;
8591 rc
= btreeOverwriteContent(pPage
, pPage
->aData
+4, pX
,
8592 iOffset
, ovflPageSize
);
8594 sqlite3PagerUnref(pPage
->pDbPage
);
8596 iOffset
+= ovflPageSize
;
8597 }while( iOffset
<nTotal
);
8603 ** Insert a new record into the BTree. The content of the new record
8604 ** is described by the pX object. The pCur cursor is used only to
8605 ** define what table the record should be inserted into, and is left
8606 ** pointing at a random location.
8608 ** For a table btree (used for rowid tables), only the pX.nKey value of
8609 ** the key is used. The pX.pKey value must be NULL. The pX.nKey is the
8610 ** rowid or INTEGER PRIMARY KEY of the row. The pX.nData,pData,nZero fields
8611 ** hold the content of the row.
8613 ** For an index btree (used for indexes and WITHOUT ROWID tables), the
8614 ** key is an arbitrary byte sequence stored in pX.pKey,nKey. The
8615 ** pX.pData,nData,nZero fields must be zero.
8617 ** If the seekResult parameter is non-zero, then a successful call to
8618 ** MovetoUnpacked() to seek cursor pCur to (pKey,nKey) has already
8619 ** been performed. In other words, if seekResult!=0 then the cursor
8620 ** is currently pointing to a cell that will be adjacent to the cell
8621 ** to be inserted. If seekResult<0 then pCur points to a cell that is
8622 ** smaller then (pKey,nKey). If seekResult>0 then pCur points to a cell
8623 ** that is larger than (pKey,nKey).
8625 ** If seekResult==0, that means pCur is pointing at some unknown location.
8626 ** In that case, this routine must seek the cursor to the correct insertion
8627 ** point for (pKey,nKey) before doing the insertion. For index btrees,
8628 ** if pX->nMem is non-zero, then pX->aMem contains pointers to the unpacked
8629 ** key values and pX->aMem can be used instead of pX->pKey to avoid having
8630 ** to decode the key.
8632 int sqlite3BtreeInsert(
8633 BtCursor
*pCur
, /* Insert data into the table of this cursor */
8634 const BtreePayload
*pX
, /* Content of the row to be inserted */
8635 int flags
, /* True if this is likely an append */
8636 int seekResult
/* Result of prior MovetoUnpacked() call */
8639 int loc
= seekResult
; /* -1: before desired location +1: after */
8643 Btree
*p
= pCur
->pBtree
;
8644 BtShared
*pBt
= p
->pBt
;
8645 unsigned char *oldCell
;
8646 unsigned char *newCell
= 0;
8648 assert( (flags
& (BTREE_SAVEPOSITION
|BTREE_APPEND
))==flags
);
8650 if( pCur
->eState
==CURSOR_FAULT
){
8651 assert( pCur
->skipNext
!=SQLITE_OK
);
8652 return pCur
->skipNext
;
8655 assert( cursorOwnsBtShared(pCur
) );
8656 assert( (pCur
->curFlags
& BTCF_WriteFlag
)!=0
8657 && pBt
->inTransaction
==TRANS_WRITE
8658 && (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
8659 assert( hasSharedCacheTableLock(p
, pCur
->pgnoRoot
, pCur
->pKeyInfo
!=0, 2) );
8661 /* Assert that the caller has been consistent. If this cursor was opened
8662 ** expecting an index b-tree, then the caller should be inserting blob
8663 ** keys with no associated data. If the cursor was opened expecting an
8664 ** intkey table, the caller should be inserting integer keys with a
8665 ** blob of associated data. */
8666 assert( (pX
->pKey
==0)==(pCur
->pKeyInfo
==0) );
8668 /* Save the positions of any other cursors open on this table.
8670 ** In some cases, the call to btreeMoveto() below is a no-op. For
8671 ** example, when inserting data into a table with auto-generated integer
8672 ** keys, the VDBE layer invokes sqlite3BtreeLast() to figure out the
8673 ** integer key to use. It then calls this function to actually insert the
8674 ** data into the intkey B-Tree. In this case btreeMoveto() recognizes
8675 ** that the cursor is already where it needs to be and returns without
8676 ** doing any work. To avoid thwarting these optimizations, it is important
8677 ** not to clear the cursor here.
8679 if( pCur
->curFlags
& BTCF_Multiple
){
8680 rc
= saveAllCursors(pBt
, pCur
->pgnoRoot
, pCur
);
8684 if( pCur
->pKeyInfo
==0 ){
8685 assert( pX
->pKey
==0 );
8686 /* If this is an insert into a table b-tree, invalidate any incrblob
8687 ** cursors open on the row being replaced */
8688 invalidateIncrblobCursors(p
, pCur
->pgnoRoot
, pX
->nKey
, 0);
8690 /* If BTREE_SAVEPOSITION is set, the cursor must already be pointing
8691 ** to a row with the same key as the new entry being inserted.
8694 if( flags
& BTREE_SAVEPOSITION
){
8695 assert( pCur
->curFlags
& BTCF_ValidNKey
);
8696 assert( pX
->nKey
==pCur
->info
.nKey
);
8701 /* On the other hand, BTREE_SAVEPOSITION==0 does not imply
8702 ** that the cursor is not pointing to a row to be overwritten.
8703 ** So do a complete check.
8705 if( (pCur
->curFlags
&BTCF_ValidNKey
)!=0 && pX
->nKey
==pCur
->info
.nKey
){
8706 /* The cursor is pointing to the entry that is to be
8708 assert( pX
->nData
>=0 && pX
->nZero
>=0 );
8709 if( pCur
->info
.nSize
!=0
8710 && pCur
->info
.nPayload
==(u32
)pX
->nData
+pX
->nZero
8712 /* New entry is the same size as the old. Do an overwrite */
8713 return btreeOverwriteCell(pCur
, pX
);
8717 /* The cursor is *not* pointing to the cell to be overwritten, nor
8718 ** to an adjacent cell. Move the cursor so that it is pointing either
8719 ** to the cell to be overwritten or an adjacent cell.
8721 rc
= sqlite3BtreeMovetoUnpacked(pCur
, 0, pX
->nKey
, flags
!=0, &loc
);
8725 /* This is an index or a WITHOUT ROWID table */
8727 /* If BTREE_SAVEPOSITION is set, the cursor must already be pointing
8728 ** to a row with the same key as the new entry being inserted.
8730 assert( (flags
& BTREE_SAVEPOSITION
)==0 || loc
==0 );
8732 /* If the cursor is not already pointing either to the cell to be
8733 ** overwritten, or if a new cell is being inserted, if the cursor is
8734 ** not pointing to an immediately adjacent cell, then move the cursor
8737 if( loc
==0 && (flags
& BTREE_SAVEPOSITION
)==0 ){
8740 r
.pKeyInfo
= pCur
->pKeyInfo
;
8742 r
.nField
= pX
->nMem
;
8748 rc
= sqlite3BtreeMovetoUnpacked(pCur
, &r
, 0, flags
!=0, &loc
);
8750 rc
= btreeMoveto(pCur
, pX
->pKey
, pX
->nKey
, flags
!=0, &loc
);
8755 /* If the cursor is currently pointing to an entry to be overwritten
8756 ** and the new content is the same as as the old, then use the
8757 ** overwrite optimization.
8761 if( pCur
->info
.nKey
==pX
->nKey
){
8763 x2
.pData
= pX
->pKey
;
8764 x2
.nData
= pX
->nKey
;
8766 return btreeOverwriteCell(pCur
, &x2
);
8771 assert( pCur
->eState
==CURSOR_VALID
8772 || (pCur
->eState
==CURSOR_INVALID
&& loc
)
8775 pPage
= pCur
->pPage
;
8776 assert( pPage
->intKey
|| pX
->nKey
>=0 );
8777 assert( pPage
->leaf
|| !pPage
->intKey
);
8778 if( pPage
->nFree
<0 ){
8779 if( pCur
->eState
>CURSOR_INVALID
){
8780 rc
= SQLITE_CORRUPT_BKPT
;
8782 rc
= btreeComputeFreeSpace(pPage
);
8787 TRACE(("INSERT: table=%d nkey=%lld ndata=%d page=%d %s\n",
8788 pCur
->pgnoRoot
, pX
->nKey
, pX
->nData
, pPage
->pgno
,
8789 loc
==0 ? "overwrite" : "new entry"));
8790 assert( pPage
->isInit
);
8791 newCell
= pBt
->pTmpSpace
;
8792 assert( newCell
!=0 );
8793 rc
= fillInCell(pPage
, newCell
, pX
, &szNew
);
8794 if( rc
) goto end_insert
;
8795 assert( szNew
==pPage
->xCellSize(pPage
, newCell
) );
8796 assert( szNew
<= MX_CELL_SIZE(pBt
) );
8800 assert( idx
<pPage
->nCell
);
8801 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
8805 oldCell
= findCell(pPage
, idx
);
8807 memcpy(newCell
, oldCell
, 4);
8809 rc
= clearCell(pPage
, oldCell
, &info
);
8810 testcase( pCur
->curFlags
& BTCF_ValidOvfl
);
8811 invalidateOverflowCache(pCur
);
8812 if( info
.nSize
==szNew
&& info
.nLocal
==info
.nPayload
8813 && (!ISAUTOVACUUM
|| szNew
<pPage
->minLocal
)
8815 /* Overwrite the old cell with the new if they are the same size.
8816 ** We could also try to do this if the old cell is smaller, then add
8817 ** the leftover space to the free list. But experiments show that
8818 ** doing that is no faster then skipping this optimization and just
8819 ** calling dropCell() and insertCell().
8821 ** This optimization cannot be used on an autovacuum database if the
8822 ** new entry uses overflow pages, as the insertCell() call below is
8823 ** necessary to add the PTRMAP_OVERFLOW1 pointer-map entry. */
8824 assert( rc
==SQLITE_OK
); /* clearCell never fails when nLocal==nPayload */
8825 if( oldCell
< pPage
->aData
+pPage
->hdrOffset
+10 ){
8826 return SQLITE_CORRUPT_BKPT
;
8828 if( oldCell
+szNew
> pPage
->aDataEnd
){
8829 return SQLITE_CORRUPT_BKPT
;
8831 memcpy(oldCell
, newCell
, szNew
);
8834 dropCell(pPage
, idx
, info
.nSize
, &rc
);
8835 if( rc
) goto end_insert
;
8836 }else if( loc
<0 && pPage
->nCell
>0 ){
8837 assert( pPage
->leaf
);
8839 pCur
->curFlags
&= ~BTCF_ValidNKey
;
8841 assert( pPage
->leaf
);
8843 insertCell(pPage
, idx
, newCell
, szNew
, 0, 0, &rc
);
8844 assert( pPage
->nOverflow
==0 || rc
==SQLITE_OK
);
8845 assert( rc
!=SQLITE_OK
|| pPage
->nCell
>0 || pPage
->nOverflow
>0 );
8847 /* If no error has occurred and pPage has an overflow cell, call balance()
8848 ** to redistribute the cells within the tree. Since balance() may move
8849 ** the cursor, zero the BtCursor.info.nSize and BTCF_ValidNKey
8852 ** Previous versions of SQLite called moveToRoot() to move the cursor
8853 ** back to the root page as balance() used to invalidate the contents
8854 ** of BtCursor.apPage[] and BtCursor.aiIdx[]. Instead of doing that,
8855 ** set the cursor state to "invalid". This makes common insert operations
8858 ** There is a subtle but important optimization here too. When inserting
8859 ** multiple records into an intkey b-tree using a single cursor (as can
8860 ** happen while processing an "INSERT INTO ... SELECT" statement), it
8861 ** is advantageous to leave the cursor pointing to the last entry in
8862 ** the b-tree if possible. If the cursor is left pointing to the last
8863 ** entry in the table, and the next row inserted has an integer key
8864 ** larger than the largest existing key, it is possible to insert the
8865 ** row without seeking the cursor. This can be a big performance boost.
8867 pCur
->info
.nSize
= 0;
8868 if( pPage
->nOverflow
){
8869 assert( rc
==SQLITE_OK
);
8870 pCur
->curFlags
&= ~(BTCF_ValidNKey
);
8873 /* Must make sure nOverflow is reset to zero even if the balance()
8874 ** fails. Internal data structure corruption will result otherwise.
8875 ** Also, set the cursor state to invalid. This stops saveCursorPosition()
8876 ** from trying to save the current position of the cursor. */
8877 pCur
->pPage
->nOverflow
= 0;
8878 pCur
->eState
= CURSOR_INVALID
;
8879 if( (flags
& BTREE_SAVEPOSITION
) && rc
==SQLITE_OK
){
8880 btreeReleaseAllCursorPages(pCur
);
8881 if( pCur
->pKeyInfo
){
8882 assert( pCur
->pKey
==0 );
8883 pCur
->pKey
= sqlite3Malloc( pX
->nKey
);
8884 if( pCur
->pKey
==0 ){
8887 memcpy(pCur
->pKey
, pX
->pKey
, pX
->nKey
);
8890 pCur
->eState
= CURSOR_REQUIRESEEK
;
8891 pCur
->nKey
= pX
->nKey
;
8894 assert( pCur
->iPage
<0 || pCur
->pPage
->nOverflow
==0 );
8901 ** Delete the entry that the cursor is pointing to.
8903 ** If the BTREE_SAVEPOSITION bit of the flags parameter is zero, then
8904 ** the cursor is left pointing at an arbitrary location after the delete.
8905 ** But if that bit is set, then the cursor is left in a state such that
8906 ** the next call to BtreeNext() or BtreePrev() moves it to the same row
8907 ** as it would have been on if the call to BtreeDelete() had been omitted.
8909 ** The BTREE_AUXDELETE bit of flags indicates that is one of several deletes
8910 ** associated with a single table entry and its indexes. Only one of those
8911 ** deletes is considered the "primary" delete. The primary delete occurs
8912 ** on a cursor that is not a BTREE_FORDELETE cursor. All but one delete
8913 ** operation on non-FORDELETE cursors is tagged with the AUXDELETE flag.
8914 ** The BTREE_AUXDELETE bit is a hint that is not used by this implementation,
8915 ** but which might be used by alternative storage engines.
8917 int sqlite3BtreeDelete(BtCursor
*pCur
, u8 flags
){
8918 Btree
*p
= pCur
->pBtree
;
8919 BtShared
*pBt
= p
->pBt
;
8920 int rc
; /* Return code */
8921 MemPage
*pPage
; /* Page to delete cell from */
8922 unsigned char *pCell
; /* Pointer to cell to delete */
8923 int iCellIdx
; /* Index of cell to delete */
8924 int iCellDepth
; /* Depth of node containing pCell */
8925 CellInfo info
; /* Size of the cell being deleted */
8926 int bSkipnext
= 0; /* Leaf cursor in SKIPNEXT state */
8927 u8 bPreserve
= flags
& BTREE_SAVEPOSITION
; /* Keep cursor valid */
8929 assert( cursorOwnsBtShared(pCur
) );
8930 assert( pBt
->inTransaction
==TRANS_WRITE
);
8931 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
8932 assert( pCur
->curFlags
& BTCF_WriteFlag
);
8933 assert( hasSharedCacheTableLock(p
, pCur
->pgnoRoot
, pCur
->pKeyInfo
!=0, 2) );
8934 assert( !hasReadConflicts(p
, pCur
->pgnoRoot
) );
8935 assert( (flags
& ~(BTREE_SAVEPOSITION
| BTREE_AUXDELETE
))==0 );
8936 if( pCur
->eState
==CURSOR_REQUIRESEEK
){
8937 rc
= btreeRestoreCursorPosition(pCur
);
8940 assert( pCur
->eState
==CURSOR_VALID
);
8942 iCellDepth
= pCur
->iPage
;
8943 iCellIdx
= pCur
->ix
;
8944 pPage
= pCur
->pPage
;
8945 pCell
= findCell(pPage
, iCellIdx
);
8946 if( pPage
->nFree
<0 && btreeComputeFreeSpace(pPage
) ) return SQLITE_CORRUPT
;
8948 /* If the bPreserve flag is set to true, then the cursor position must
8949 ** be preserved following this delete operation. If the current delete
8950 ** will cause a b-tree rebalance, then this is done by saving the cursor
8951 ** key and leaving the cursor in CURSOR_REQUIRESEEK state before
8954 ** Or, if the current delete will not cause a rebalance, then the cursor
8955 ** will be left in CURSOR_SKIPNEXT state pointing to the entry immediately
8956 ** before or after the deleted entry. In this case set bSkipnext to true. */
8959 || (pPage
->nFree
+cellSizePtr(pPage
,pCell
)+2)>(int)(pBt
->usableSize
*2/3)
8960 || pPage
->nCell
==1 /* See dbfuzz001.test for a test case */
8962 /* A b-tree rebalance will be required after deleting this entry.
8963 ** Save the cursor key. */
8964 rc
= saveCursorKey(pCur
);
8971 /* If the page containing the entry to delete is not a leaf page, move
8972 ** the cursor to the largest entry in the tree that is smaller than
8973 ** the entry being deleted. This cell will replace the cell being deleted
8974 ** from the internal node. The 'previous' entry is used for this instead
8975 ** of the 'next' entry, as the previous entry is always a part of the
8976 ** sub-tree headed by the child page of the cell being deleted. This makes
8977 ** balancing the tree following the delete operation easier. */
8979 rc
= sqlite3BtreePrevious(pCur
, 0);
8980 assert( rc
!=SQLITE_DONE
);
8984 /* Save the positions of any other cursors open on this table before
8985 ** making any modifications. */
8986 if( pCur
->curFlags
& BTCF_Multiple
){
8987 rc
= saveAllCursors(pBt
, pCur
->pgnoRoot
, pCur
);
8991 /* If this is a delete operation to remove a row from a table b-tree,
8992 ** invalidate any incrblob cursors open on the row being deleted. */
8993 if( pCur
->pKeyInfo
==0 ){
8994 invalidateIncrblobCursors(p
, pCur
->pgnoRoot
, pCur
->info
.nKey
, 0);
8997 /* Make the page containing the entry to be deleted writable. Then free any
8998 ** overflow pages associated with the entry and finally remove the cell
8999 ** itself from within the page. */
9000 rc
= sqlite3PagerWrite(pPage
->pDbPage
);
9002 rc
= clearCell(pPage
, pCell
, &info
);
9003 dropCell(pPage
, iCellIdx
, info
.nSize
, &rc
);
9006 /* If the cell deleted was not located on a leaf page, then the cursor
9007 ** is currently pointing to the largest entry in the sub-tree headed
9008 ** by the child-page of the cell that was just deleted from an internal
9009 ** node. The cell from the leaf node needs to be moved to the internal
9010 ** node to replace the deleted cell. */
9012 MemPage
*pLeaf
= pCur
->pPage
;
9015 unsigned char *pTmp
;
9017 if( pLeaf
->nFree
<0 ){
9018 rc
= btreeComputeFreeSpace(pLeaf
);
9021 if( iCellDepth
<pCur
->iPage
-1 ){
9022 n
= pCur
->apPage
[iCellDepth
+1]->pgno
;
9024 n
= pCur
->pPage
->pgno
;
9026 pCell
= findCell(pLeaf
, pLeaf
->nCell
-1);
9027 if( pCell
<&pLeaf
->aData
[4] ) return SQLITE_CORRUPT_BKPT
;
9028 nCell
= pLeaf
->xCellSize(pLeaf
, pCell
);
9029 assert( MX_CELL_SIZE(pBt
) >= nCell
);
9030 pTmp
= pBt
->pTmpSpace
;
9032 rc
= sqlite3PagerWrite(pLeaf
->pDbPage
);
9033 if( rc
==SQLITE_OK
){
9034 insertCell(pPage
, iCellIdx
, pCell
-4, nCell
+4, pTmp
, n
, &rc
);
9036 dropCell(pLeaf
, pLeaf
->nCell
-1, nCell
, &rc
);
9040 /* Balance the tree. If the entry deleted was located on a leaf page,
9041 ** then the cursor still points to that page. In this case the first
9042 ** call to balance() repairs the tree, and the if(...) condition is
9045 ** Otherwise, if the entry deleted was on an internal node page, then
9046 ** pCur is pointing to the leaf page from which a cell was removed to
9047 ** replace the cell deleted from the internal node. This is slightly
9048 ** tricky as the leaf node may be underfull, and the internal node may
9049 ** be either under or overfull. In this case run the balancing algorithm
9050 ** on the leaf node first. If the balance proceeds far enough up the
9051 ** tree that we can be sure that any problem in the internal node has
9052 ** been corrected, so be it. Otherwise, after balancing the leaf node,
9053 ** walk the cursor up the tree to the internal node and balance it as
9056 if( rc
==SQLITE_OK
&& pCur
->iPage
>iCellDepth
){
9057 releasePageNotNull(pCur
->pPage
);
9059 while( pCur
->iPage
>iCellDepth
){
9060 releasePage(pCur
->apPage
[pCur
->iPage
--]);
9062 pCur
->pPage
= pCur
->apPage
[pCur
->iPage
];
9066 if( rc
==SQLITE_OK
){
9068 assert( bPreserve
&& (pCur
->iPage
==iCellDepth
|| CORRUPT_DB
) );
9069 assert( pPage
==pCur
->pPage
|| CORRUPT_DB
);
9070 assert( (pPage
->nCell
>0 || CORRUPT_DB
) && iCellIdx
<=pPage
->nCell
);
9071 pCur
->eState
= CURSOR_SKIPNEXT
;
9072 if( iCellIdx
>=pPage
->nCell
){
9073 pCur
->skipNext
= -1;
9074 pCur
->ix
= pPage
->nCell
-1;
9079 rc
= moveToRoot(pCur
);
9081 btreeReleaseAllCursorPages(pCur
);
9082 pCur
->eState
= CURSOR_REQUIRESEEK
;
9084 if( rc
==SQLITE_EMPTY
) rc
= SQLITE_OK
;
9091 ** Create a new BTree table. Write into *piTable the page
9092 ** number for the root page of the new table.
9094 ** The type of type is determined by the flags parameter. Only the
9095 ** following values of flags are currently in use. Other values for
9096 ** flags might not work:
9098 ** BTREE_INTKEY|BTREE_LEAFDATA Used for SQL tables with rowid keys
9099 ** BTREE_ZERODATA Used for SQL indices
9101 static int btreeCreateTable(Btree
*p
, Pgno
*piTable
, int createTabFlags
){
9102 BtShared
*pBt
= p
->pBt
;
9106 int ptfFlags
; /* Page-type flage for the root page of new table */
9108 assert( sqlite3BtreeHoldsMutex(p
) );
9109 assert( pBt
->inTransaction
==TRANS_WRITE
);
9110 assert( (pBt
->btsFlags
& BTS_READ_ONLY
)==0 );
9112 #ifdef SQLITE_OMIT_AUTOVACUUM
9113 rc
= allocateBtreePage(pBt
, &pRoot
, &pgnoRoot
, 1, 0);
9118 if( pBt
->autoVacuum
){
9119 Pgno pgnoMove
; /* Move a page here to make room for the root-page */
9120 MemPage
*pPageMove
; /* The page to move to. */
9122 /* Creating a new table may probably require moving an existing database
9123 ** to make room for the new tables root page. In case this page turns
9124 ** out to be an overflow page, delete all overflow page-map caches
9125 ** held by open cursors.
9127 invalidateAllOverflowCache(pBt
);
9129 /* Read the value of meta[3] from the database to determine where the
9130 ** root page of the new table should go. meta[3] is the largest root-page
9131 ** created so far, so the new root-page is (meta[3]+1).
9133 sqlite3BtreeGetMeta(p
, BTREE_LARGEST_ROOT_PAGE
, &pgnoRoot
);
9134 if( pgnoRoot
>btreePagecount(pBt
) ){
9135 return SQLITE_CORRUPT_BKPT
;
9139 /* The new root-page may not be allocated on a pointer-map page, or the
9140 ** PENDING_BYTE page.
9142 while( pgnoRoot
==PTRMAP_PAGENO(pBt
, pgnoRoot
) ||
9143 pgnoRoot
==PENDING_BYTE_PAGE(pBt
) ){
9146 assert( pgnoRoot
>=3 );
9148 /* Allocate a page. The page that currently resides at pgnoRoot will
9149 ** be moved to the allocated page (unless the allocated page happens
9150 ** to reside at pgnoRoot).
9152 rc
= allocateBtreePage(pBt
, &pPageMove
, &pgnoMove
, pgnoRoot
, BTALLOC_EXACT
);
9153 if( rc
!=SQLITE_OK
){
9157 if( pgnoMove
!=pgnoRoot
){
9158 /* pgnoRoot is the page that will be used for the root-page of
9159 ** the new table (assuming an error did not occur). But we were
9160 ** allocated pgnoMove. If required (i.e. if it was not allocated
9161 ** by extending the file), the current page at position pgnoMove
9162 ** is already journaled.
9167 /* Save the positions of any open cursors. This is required in
9168 ** case they are holding a reference to an xFetch reference
9169 ** corresponding to page pgnoRoot. */
9170 rc
= saveAllCursors(pBt
, 0, 0);
9171 releasePage(pPageMove
);
9172 if( rc
!=SQLITE_OK
){
9176 /* Move the page currently at pgnoRoot to pgnoMove. */
9177 rc
= btreeGetPage(pBt
, pgnoRoot
, &pRoot
, 0);
9178 if( rc
!=SQLITE_OK
){
9181 rc
= ptrmapGet(pBt
, pgnoRoot
, &eType
, &iPtrPage
);
9182 if( eType
==PTRMAP_ROOTPAGE
|| eType
==PTRMAP_FREEPAGE
){
9183 rc
= SQLITE_CORRUPT_BKPT
;
9185 if( rc
!=SQLITE_OK
){
9189 assert( eType
!=PTRMAP_ROOTPAGE
);
9190 assert( eType
!=PTRMAP_FREEPAGE
);
9191 rc
= relocatePage(pBt
, pRoot
, eType
, iPtrPage
, pgnoMove
, 0);
9194 /* Obtain the page at pgnoRoot */
9195 if( rc
!=SQLITE_OK
){
9198 rc
= btreeGetPage(pBt
, pgnoRoot
, &pRoot
, 0);
9199 if( rc
!=SQLITE_OK
){
9202 rc
= sqlite3PagerWrite(pRoot
->pDbPage
);
9203 if( rc
!=SQLITE_OK
){
9211 /* Update the pointer-map and meta-data with the new root-page number. */
9212 ptrmapPut(pBt
, pgnoRoot
, PTRMAP_ROOTPAGE
, 0, &rc
);
9218 /* When the new root page was allocated, page 1 was made writable in
9219 ** order either to increase the database filesize, or to decrement the
9220 ** freelist count. Hence, the sqlite3BtreeUpdateMeta() call cannot fail.
9222 assert( sqlite3PagerIswriteable(pBt
->pPage1
->pDbPage
) );
9223 rc
= sqlite3BtreeUpdateMeta(p
, 4, pgnoRoot
);
9230 rc
= allocateBtreePage(pBt
, &pRoot
, &pgnoRoot
, 1, 0);
9234 assert( sqlite3PagerIswriteable(pRoot
->pDbPage
) );
9235 if( createTabFlags
& BTREE_INTKEY
){
9236 ptfFlags
= PTF_INTKEY
| PTF_LEAFDATA
| PTF_LEAF
;
9238 ptfFlags
= PTF_ZERODATA
| PTF_LEAF
;
9240 zeroPage(pRoot
, ptfFlags
);
9241 sqlite3PagerUnref(pRoot
->pDbPage
);
9242 assert( (pBt
->openFlags
& BTREE_SINGLE
)==0 || pgnoRoot
==2 );
9243 *piTable
= pgnoRoot
;
9246 int sqlite3BtreeCreateTable(Btree
*p
, Pgno
*piTable
, int flags
){
9248 sqlite3BtreeEnter(p
);
9249 rc
= btreeCreateTable(p
, piTable
, flags
);
9250 sqlite3BtreeLeave(p
);
9255 ** Erase the given database page and all its children. Return
9256 ** the page to the freelist.
9258 static int clearDatabasePage(
9259 BtShared
*pBt
, /* The BTree that contains the table */
9260 Pgno pgno
, /* Page number to clear */
9261 int freePageFlag
, /* Deallocate page if true */
9262 int *pnChange
/* Add number of Cells freed to this counter */
9266 unsigned char *pCell
;
9271 assert( sqlite3_mutex_held(pBt
->mutex
) );
9272 if( pgno
>btreePagecount(pBt
) ){
9273 return SQLITE_CORRUPT_BKPT
;
9275 rc
= getAndInitPage(pBt
, pgno
, &pPage
, 0, 0);
9278 rc
= SQLITE_CORRUPT_BKPT
;
9279 goto cleardatabasepage_out
;
9282 hdr
= pPage
->hdrOffset
;
9283 for(i
=0; i
<pPage
->nCell
; i
++){
9284 pCell
= findCell(pPage
, i
);
9286 rc
= clearDatabasePage(pBt
, get4byte(pCell
), 1, pnChange
);
9287 if( rc
) goto cleardatabasepage_out
;
9289 rc
= clearCell(pPage
, pCell
, &info
);
9290 if( rc
) goto cleardatabasepage_out
;
9293 rc
= clearDatabasePage(pBt
, get4byte(&pPage
->aData
[hdr
+8]), 1, pnChange
);
9294 if( rc
) goto cleardatabasepage_out
;
9295 }else if( pnChange
){
9296 assert( pPage
->intKey
|| CORRUPT_DB
);
9297 testcase( !pPage
->intKey
);
9298 *pnChange
+= pPage
->nCell
;
9301 freePage(pPage
, &rc
);
9302 }else if( (rc
= sqlite3PagerWrite(pPage
->pDbPage
))==0 ){
9303 zeroPage(pPage
, pPage
->aData
[hdr
] | PTF_LEAF
);
9306 cleardatabasepage_out
:
9313 ** Delete all information from a single table in the database. iTable is
9314 ** the page number of the root of the table. After this routine returns,
9315 ** the root page is empty, but still exists.
9317 ** This routine will fail with SQLITE_LOCKED if there are any open
9318 ** read cursors on the table. Open write cursors are moved to the
9319 ** root of the table.
9321 ** If pnChange is not NULL, then table iTable must be an intkey table. The
9322 ** integer value pointed to by pnChange is incremented by the number of
9323 ** entries in the table.
9325 int sqlite3BtreeClearTable(Btree
*p
, int iTable
, int *pnChange
){
9327 BtShared
*pBt
= p
->pBt
;
9328 sqlite3BtreeEnter(p
);
9329 assert( p
->inTrans
==TRANS_WRITE
);
9331 rc
= saveAllCursors(pBt
, (Pgno
)iTable
, 0);
9333 if( SQLITE_OK
==rc
){
9334 /* Invalidate all incrblob cursors open on table iTable (assuming iTable
9335 ** is the root of a table b-tree - if it is not, the following call is
9337 invalidateIncrblobCursors(p
, (Pgno
)iTable
, 0, 1);
9338 rc
= clearDatabasePage(pBt
, (Pgno
)iTable
, 0, pnChange
);
9340 sqlite3BtreeLeave(p
);
9345 ** Delete all information from the single table that pCur is open on.
9347 ** This routine only work for pCur on an ephemeral table.
9349 int sqlite3BtreeClearTableOfCursor(BtCursor
*pCur
){
9350 return sqlite3BtreeClearTable(pCur
->pBtree
, pCur
->pgnoRoot
, 0);
9354 ** Erase all information in a table and add the root of the table to
9355 ** the freelist. Except, the root of the principle table (the one on
9356 ** page 1) is never added to the freelist.
9358 ** This routine will fail with SQLITE_LOCKED if there are any open
9359 ** cursors on the table.
9361 ** If AUTOVACUUM is enabled and the page at iTable is not the last
9362 ** root page in the database file, then the last root page
9363 ** in the database file is moved into the slot formerly occupied by
9364 ** iTable and that last slot formerly occupied by the last root page
9365 ** is added to the freelist instead of iTable. In this say, all
9366 ** root pages are kept at the beginning of the database file, which
9367 ** is necessary for AUTOVACUUM to work right. *piMoved is set to the
9368 ** page number that used to be the last root page in the file before
9369 ** the move. If no page gets moved, *piMoved is set to 0.
9370 ** The last root page is recorded in meta[3] and the value of
9371 ** meta[3] is updated by this procedure.
9373 static int btreeDropTable(Btree
*p
, Pgno iTable
, int *piMoved
){
9376 BtShared
*pBt
= p
->pBt
;
9378 assert( sqlite3BtreeHoldsMutex(p
) );
9379 assert( p
->inTrans
==TRANS_WRITE
);
9380 assert( iTable
>=2 );
9381 if( iTable
>btreePagecount(pBt
) ){
9382 return SQLITE_CORRUPT_BKPT
;
9385 rc
= btreeGetPage(pBt
, (Pgno
)iTable
, &pPage
, 0);
9387 rc
= sqlite3BtreeClearTable(p
, iTable
, 0);
9395 #ifdef SQLITE_OMIT_AUTOVACUUM
9396 freePage(pPage
, &rc
);
9399 if( pBt
->autoVacuum
){
9401 sqlite3BtreeGetMeta(p
, BTREE_LARGEST_ROOT_PAGE
, &maxRootPgno
);
9403 if( iTable
==maxRootPgno
){
9404 /* If the table being dropped is the table with the largest root-page
9405 ** number in the database, put the root page on the free list.
9407 freePage(pPage
, &rc
);
9409 if( rc
!=SQLITE_OK
){
9413 /* The table being dropped does not have the largest root-page
9414 ** number in the database. So move the page that does into the
9415 ** gap left by the deleted root-page.
9419 rc
= btreeGetPage(pBt
, maxRootPgno
, &pMove
, 0);
9420 if( rc
!=SQLITE_OK
){
9423 rc
= relocatePage(pBt
, pMove
, PTRMAP_ROOTPAGE
, 0, iTable
, 0);
9425 if( rc
!=SQLITE_OK
){
9429 rc
= btreeGetPage(pBt
, maxRootPgno
, &pMove
, 0);
9430 freePage(pMove
, &rc
);
9432 if( rc
!=SQLITE_OK
){
9435 *piMoved
= maxRootPgno
;
9438 /* Set the new 'max-root-page' value in the database header. This
9439 ** is the old value less one, less one more if that happens to
9440 ** be a root-page number, less one again if that is the
9441 ** PENDING_BYTE_PAGE.
9444 while( maxRootPgno
==PENDING_BYTE_PAGE(pBt
)
9445 || PTRMAP_ISPAGE(pBt
, maxRootPgno
) ){
9448 assert( maxRootPgno
!=PENDING_BYTE_PAGE(pBt
) );
9450 rc
= sqlite3BtreeUpdateMeta(p
, 4, maxRootPgno
);
9452 freePage(pPage
, &rc
);
9458 int sqlite3BtreeDropTable(Btree
*p
, int iTable
, int *piMoved
){
9460 sqlite3BtreeEnter(p
);
9461 rc
= btreeDropTable(p
, iTable
, piMoved
);
9462 sqlite3BtreeLeave(p
);
9468 ** This function may only be called if the b-tree connection already
9469 ** has a read or write transaction open on the database.
9471 ** Read the meta-information out of a database file. Meta[0]
9472 ** is the number of free pages currently in the database. Meta[1]
9473 ** through meta[15] are available for use by higher layers. Meta[0]
9474 ** is read-only, the others are read/write.
9476 ** The schema layer numbers meta values differently. At the schema
9477 ** layer (and the SetCookie and ReadCookie opcodes) the number of
9478 ** free pages is not visible. So Cookie[0] is the same as Meta[1].
9480 ** This routine treats Meta[BTREE_DATA_VERSION] as a special case. Instead
9481 ** of reading the value out of the header, it instead loads the "DataVersion"
9482 ** from the pager. The BTREE_DATA_VERSION value is not actually stored in the
9483 ** database file. It is a number computed by the pager. But its access
9484 ** pattern is the same as header meta values, and so it is convenient to
9485 ** read it from this routine.
9487 void sqlite3BtreeGetMeta(Btree
*p
, int idx
, u32
*pMeta
){
9488 BtShared
*pBt
= p
->pBt
;
9490 sqlite3BtreeEnter(p
);
9491 assert( p
->inTrans
>TRANS_NONE
);
9492 assert( SQLITE_OK
==querySharedCacheTableLock(p
, SCHEMA_ROOT
, READ_LOCK
) );
9493 assert( pBt
->pPage1
);
9494 assert( idx
>=0 && idx
<=15 );
9496 if( idx
==BTREE_DATA_VERSION
){
9497 *pMeta
= sqlite3PagerDataVersion(pBt
->pPager
) + p
->iDataVersion
;
9499 *pMeta
= get4byte(&pBt
->pPage1
->aData
[36 + idx
*4]);
9502 /* If auto-vacuum is disabled in this build and this is an auto-vacuum
9503 ** database, mark the database as read-only. */
9504 #ifdef SQLITE_OMIT_AUTOVACUUM
9505 if( idx
==BTREE_LARGEST_ROOT_PAGE
&& *pMeta
>0 ){
9506 pBt
->btsFlags
|= BTS_READ_ONLY
;
9510 sqlite3BtreeLeave(p
);
9514 ** Write meta-information back into the database. Meta[0] is
9515 ** read-only and may not be written.
9517 int sqlite3BtreeUpdateMeta(Btree
*p
, int idx
, u32 iMeta
){
9518 BtShared
*pBt
= p
->pBt
;
9521 assert( idx
>=1 && idx
<=15 );
9522 sqlite3BtreeEnter(p
);
9523 assert( p
->inTrans
==TRANS_WRITE
);
9524 assert( pBt
->pPage1
!=0 );
9525 pP1
= pBt
->pPage1
->aData
;
9526 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
9527 if( rc
==SQLITE_OK
){
9528 put4byte(&pP1
[36 + idx
*4], iMeta
);
9529 #ifndef SQLITE_OMIT_AUTOVACUUM
9530 if( idx
==BTREE_INCR_VACUUM
){
9531 assert( pBt
->autoVacuum
|| iMeta
==0 );
9532 assert( iMeta
==0 || iMeta
==1 );
9533 pBt
->incrVacuum
= (u8
)iMeta
;
9537 sqlite3BtreeLeave(p
);
9542 ** The first argument, pCur, is a cursor opened on some b-tree. Count the
9543 ** number of entries in the b-tree and write the result to *pnEntry.
9545 ** SQLITE_OK is returned if the operation is successfully executed.
9546 ** Otherwise, if an error is encountered (i.e. an IO error or database
9547 ** corruption) an SQLite error code is returned.
9549 int sqlite3BtreeCount(sqlite3
*db
, BtCursor
*pCur
, i64
*pnEntry
){
9550 i64 nEntry
= 0; /* Value to return in *pnEntry */
9551 int rc
; /* Return code */
9553 rc
= moveToRoot(pCur
);
9554 if( rc
==SQLITE_EMPTY
){
9559 /* Unless an error occurs, the following loop runs one iteration for each
9560 ** page in the B-Tree structure (not including overflow pages).
9562 while( rc
==SQLITE_OK
&& !AtomicLoad(&db
->u1
.isInterrupted
) ){
9563 int iIdx
; /* Index of child node in parent */
9564 MemPage
*pPage
; /* Current page of the b-tree */
9566 /* If this is a leaf page or the tree is not an int-key tree, then
9567 ** this page contains countable entries. Increment the entry counter
9570 pPage
= pCur
->pPage
;
9571 if( pPage
->leaf
|| !pPage
->intKey
){
9572 nEntry
+= pPage
->nCell
;
9575 /* pPage is a leaf node. This loop navigates the cursor so that it
9576 ** points to the first interior cell that it points to the parent of
9577 ** the next page in the tree that has not yet been visited. The
9578 ** pCur->aiIdx[pCur->iPage] value is set to the index of the parent cell
9579 ** of the page, or to the number of cells in the page if the next page
9580 ** to visit is the right-child of its parent.
9582 ** If all pages in the tree have been visited, return SQLITE_OK to the
9587 if( pCur
->iPage
==0 ){
9588 /* All pages of the b-tree have been visited. Return successfully. */
9590 return moveToRoot(pCur
);
9593 }while ( pCur
->ix
>=pCur
->pPage
->nCell
);
9596 pPage
= pCur
->pPage
;
9599 /* Descend to the child node of the cell that the cursor currently
9600 ** points at. This is the right-child if (iIdx==pPage->nCell).
9603 if( iIdx
==pPage
->nCell
){
9604 rc
= moveToChild(pCur
, get4byte(&pPage
->aData
[pPage
->hdrOffset
+8]));
9606 rc
= moveToChild(pCur
, get4byte(findCell(pPage
, iIdx
)));
9610 /* An error has occurred. Return an error code. */
9615 ** Return the pager associated with a BTree. This routine is used for
9616 ** testing and debugging only.
9618 Pager
*sqlite3BtreePager(Btree
*p
){
9619 return p
->pBt
->pPager
;
9622 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9624 ** Append a message to the error message string.
9626 static void checkAppendMsg(
9627 IntegrityCk
*pCheck
,
9628 const char *zFormat
,
9632 if( !pCheck
->mxErr
) return;
9635 va_start(ap
, zFormat
);
9636 if( pCheck
->errMsg
.nChar
){
9637 sqlite3_str_append(&pCheck
->errMsg
, "\n", 1);
9640 sqlite3_str_appendf(&pCheck
->errMsg
, pCheck
->zPfx
, pCheck
->v1
, pCheck
->v2
);
9642 sqlite3_str_vappendf(&pCheck
->errMsg
, zFormat
, ap
);
9644 if( pCheck
->errMsg
.accError
==SQLITE_NOMEM
){
9645 pCheck
->bOomFault
= 1;
9648 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9650 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9653 ** Return non-zero if the bit in the IntegrityCk.aPgRef[] array that
9654 ** corresponds to page iPg is already set.
9656 static int getPageReferenced(IntegrityCk
*pCheck
, Pgno iPg
){
9657 assert( iPg
<=pCheck
->nPage
&& sizeof(pCheck
->aPgRef
[0])==1 );
9658 return (pCheck
->aPgRef
[iPg
/8] & (1 << (iPg
& 0x07)));
9662 ** Set the bit in the IntegrityCk.aPgRef[] array that corresponds to page iPg.
9664 static void setPageReferenced(IntegrityCk
*pCheck
, Pgno iPg
){
9665 assert( iPg
<=pCheck
->nPage
&& sizeof(pCheck
->aPgRef
[0])==1 );
9666 pCheck
->aPgRef
[iPg
/8] |= (1 << (iPg
& 0x07));
9671 ** Add 1 to the reference count for page iPage. If this is the second
9672 ** reference to the page, add an error message to pCheck->zErrMsg.
9673 ** Return 1 if there are 2 or more references to the page and 0 if
9674 ** if this is the first reference to the page.
9676 ** Also check that the page number is in bounds.
9678 static int checkRef(IntegrityCk
*pCheck
, Pgno iPage
){
9679 if( iPage
>pCheck
->nPage
|| iPage
==0 ){
9680 checkAppendMsg(pCheck
, "invalid page number %d", iPage
);
9683 if( getPageReferenced(pCheck
, iPage
) ){
9684 checkAppendMsg(pCheck
, "2nd reference to page %d", iPage
);
9687 if( AtomicLoad(&pCheck
->db
->u1
.isInterrupted
) ) return 1;
9688 setPageReferenced(pCheck
, iPage
);
9692 #ifndef SQLITE_OMIT_AUTOVACUUM
9694 ** Check that the entry in the pointer-map for page iChild maps to
9695 ** page iParent, pointer type ptrType. If not, append an error message
9698 static void checkPtrmap(
9699 IntegrityCk
*pCheck
, /* Integrity check context */
9700 Pgno iChild
, /* Child page number */
9701 u8 eType
, /* Expected pointer map type */
9702 Pgno iParent
/* Expected pointer map parent page number */
9708 rc
= ptrmapGet(pCheck
->pBt
, iChild
, &ePtrmapType
, &iPtrmapParent
);
9709 if( rc
!=SQLITE_OK
){
9710 if( rc
==SQLITE_NOMEM
|| rc
==SQLITE_IOERR_NOMEM
) pCheck
->bOomFault
= 1;
9711 checkAppendMsg(pCheck
, "Failed to read ptrmap key=%d", iChild
);
9715 if( ePtrmapType
!=eType
|| iPtrmapParent
!=iParent
){
9716 checkAppendMsg(pCheck
,
9717 "Bad ptr map entry key=%d expected=(%d,%d) got=(%d,%d)",
9718 iChild
, eType
, iParent
, ePtrmapType
, iPtrmapParent
);
9724 ** Check the integrity of the freelist or of an overflow page list.
9725 ** Verify that the number of pages on the list is N.
9727 static void checkList(
9728 IntegrityCk
*pCheck
, /* Integrity checking context */
9729 int isFreeList
, /* True for a freelist. False for overflow page list */
9730 Pgno iPage
, /* Page number for first page in the list */
9731 u32 N
/* Expected number of pages in the list */
9735 int nErrAtStart
= pCheck
->nErr
;
9736 while( iPage
!=0 && pCheck
->mxErr
){
9738 unsigned char *pOvflData
;
9739 if( checkRef(pCheck
, iPage
) ) break;
9741 if( sqlite3PagerGet(pCheck
->pPager
, (Pgno
)iPage
, &pOvflPage
, 0) ){
9742 checkAppendMsg(pCheck
, "failed to get page %d", iPage
);
9745 pOvflData
= (unsigned char *)sqlite3PagerGetData(pOvflPage
);
9747 u32 n
= (u32
)get4byte(&pOvflData
[4]);
9748 #ifndef SQLITE_OMIT_AUTOVACUUM
9749 if( pCheck
->pBt
->autoVacuum
){
9750 checkPtrmap(pCheck
, iPage
, PTRMAP_FREEPAGE
, 0);
9753 if( n
>pCheck
->pBt
->usableSize
/4-2 ){
9754 checkAppendMsg(pCheck
,
9755 "freelist leaf count too big on page %d", iPage
);
9758 for(i
=0; i
<(int)n
; i
++){
9759 Pgno iFreePage
= get4byte(&pOvflData
[8+i
*4]);
9760 #ifndef SQLITE_OMIT_AUTOVACUUM
9761 if( pCheck
->pBt
->autoVacuum
){
9762 checkPtrmap(pCheck
, iFreePage
, PTRMAP_FREEPAGE
, 0);
9765 checkRef(pCheck
, iFreePage
);
9770 #ifndef SQLITE_OMIT_AUTOVACUUM
9772 /* If this database supports auto-vacuum and iPage is not the last
9773 ** page in this overflow list, check that the pointer-map entry for
9774 ** the following page matches iPage.
9776 if( pCheck
->pBt
->autoVacuum
&& N
>0 ){
9777 i
= get4byte(pOvflData
);
9778 checkPtrmap(pCheck
, i
, PTRMAP_OVERFLOW2
, iPage
);
9782 iPage
= get4byte(pOvflData
);
9783 sqlite3PagerUnref(pOvflPage
);
9785 if( N
&& nErrAtStart
==pCheck
->nErr
){
9786 checkAppendMsg(pCheck
,
9787 "%s is %d but should be %d",
9788 isFreeList
? "size" : "overflow list length",
9789 expected
-N
, expected
);
9792 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
9795 ** An implementation of a min-heap.
9797 ** aHeap[0] is the number of elements on the heap. aHeap[1] is the
9798 ** root element. The daughter nodes of aHeap[N] are aHeap[N*2]
9799 ** and aHeap[N*2+1].
9801 ** The heap property is this: Every node is less than or equal to both
9802 ** of its daughter nodes. A consequence of the heap property is that the
9803 ** root node aHeap[1] is always the minimum value currently in the heap.
9805 ** The btreeHeapInsert() routine inserts an unsigned 32-bit number onto
9806 ** the heap, preserving the heap property. The btreeHeapPull() routine
9807 ** removes the root element from the heap (the minimum value in the heap)
9808 ** and then moves other nodes around as necessary to preserve the heap
9811 ** This heap is used for cell overlap and coverage testing. Each u32
9812 ** entry represents the span of a cell or freeblock on a btree page.
9813 ** The upper 16 bits are the index of the first byte of a range and the
9814 ** lower 16 bits are the index of the last byte of that range.
9816 static void btreeHeapInsert(u32
*aHeap
, u32 x
){
9817 u32 j
, i
= ++aHeap
[0];
9819 while( (j
= i
/2)>0 && aHeap
[j
]>aHeap
[i
] ){
9821 aHeap
[j
] = aHeap
[i
];
9826 static int btreeHeapPull(u32
*aHeap
, u32
*pOut
){
9828 if( (x
= aHeap
[0])==0 ) return 0;
9830 aHeap
[1] = aHeap
[x
];
9831 aHeap
[x
] = 0xffffffff;
9834 while( (j
= i
*2)<=aHeap
[0] ){
9835 if( aHeap
[j
]>aHeap
[j
+1] ) j
++;
9836 if( aHeap
[i
]<aHeap
[j
] ) break;
9838 aHeap
[i
] = aHeap
[j
];
9845 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
9847 ** Do various sanity checks on a single page of a tree. Return
9848 ** the tree depth. Root pages return 0. Parents of root pages
9849 ** return 1, and so forth.
9851 ** These checks are done:
9853 ** 1. Make sure that cells and freeblocks do not overlap
9854 ** but combine to completely cover the page.
9855 ** 2. Make sure integer cell keys are in order.
9856 ** 3. Check the integrity of overflow pages.
9857 ** 4. Recursively call checkTreePage on all children.
9858 ** 5. Verify that the depth of all children is the same.
9860 static int checkTreePage(
9861 IntegrityCk
*pCheck
, /* Context for the sanity check */
9862 Pgno iPage
, /* Page number of the page to check */
9863 i64
*piMinKey
, /* Write minimum integer primary key here */
9864 i64 maxKey
/* Error if integer primary key greater than this */
9866 MemPage
*pPage
= 0; /* The page being analyzed */
9867 int i
; /* Loop counter */
9868 int rc
; /* Result code from subroutine call */
9869 int depth
= -1, d2
; /* Depth of a subtree */
9870 int pgno
; /* Page number */
9871 int nFrag
; /* Number of fragmented bytes on the page */
9872 int hdr
; /* Offset to the page header */
9873 int cellStart
; /* Offset to the start of the cell pointer array */
9874 int nCell
; /* Number of cells */
9875 int doCoverageCheck
= 1; /* True if cell coverage checking should be done */
9876 int keyCanBeEqual
= 1; /* True if IPK can be equal to maxKey
9877 ** False if IPK must be strictly less than maxKey */
9878 u8
*data
; /* Page content */
9879 u8
*pCell
; /* Cell content */
9880 u8
*pCellIdx
; /* Next element of the cell pointer array */
9881 BtShared
*pBt
; /* The BtShared object that owns pPage */
9882 u32 pc
; /* Address of a cell */
9883 u32 usableSize
; /* Usable size of the page */
9884 u32 contentOffset
; /* Offset to the start of the cell content area */
9885 u32
*heap
= 0; /* Min-heap used for checking cell coverage */
9886 u32 x
, prev
= 0; /* Next and previous entry on the min-heap */
9887 const char *saved_zPfx
= pCheck
->zPfx
;
9888 int saved_v1
= pCheck
->v1
;
9889 int saved_v2
= pCheck
->v2
;
9892 /* Check that the page exists
9895 usableSize
= pBt
->usableSize
;
9896 if( iPage
==0 ) return 0;
9897 if( checkRef(pCheck
, iPage
) ) return 0;
9898 pCheck
->zPfx
= "Page %u: ";
9900 if( (rc
= btreeGetPage(pBt
, iPage
, &pPage
, 0))!=0 ){
9901 checkAppendMsg(pCheck
,
9902 "unable to get the page. error code=%d", rc
);
9906 /* Clear MemPage.isInit to make sure the corruption detection code in
9907 ** btreeInitPage() is executed. */
9908 savedIsInit
= pPage
->isInit
;
9910 if( (rc
= btreeInitPage(pPage
))!=0 ){
9911 assert( rc
==SQLITE_CORRUPT
); /* The only possible error from InitPage */
9912 checkAppendMsg(pCheck
,
9913 "btreeInitPage() returns error code %d", rc
);
9916 if( (rc
= btreeComputeFreeSpace(pPage
))!=0 ){
9917 assert( rc
==SQLITE_CORRUPT
);
9918 checkAppendMsg(pCheck
, "free space corruption", rc
);
9921 data
= pPage
->aData
;
9922 hdr
= pPage
->hdrOffset
;
9924 /* Set up for cell analysis */
9925 pCheck
->zPfx
= "On tree page %u cell %d: ";
9926 contentOffset
= get2byteNotZero(&data
[hdr
+5]);
9927 assert( contentOffset
<=usableSize
); /* Enforced by btreeInitPage() */
9929 /* EVIDENCE-OF: R-37002-32774 The two-byte integer at offset 3 gives the
9930 ** number of cells on the page. */
9931 nCell
= get2byte(&data
[hdr
+3]);
9932 assert( pPage
->nCell
==nCell
);
9934 /* EVIDENCE-OF: R-23882-45353 The cell pointer array of a b-tree page
9935 ** immediately follows the b-tree page header. */
9936 cellStart
= hdr
+ 12 - 4*pPage
->leaf
;
9937 assert( pPage
->aCellIdx
==&data
[cellStart
] );
9938 pCellIdx
= &data
[cellStart
+ 2*(nCell
-1)];
9941 /* Analyze the right-child page of internal pages */
9942 pgno
= get4byte(&data
[hdr
+8]);
9943 #ifndef SQLITE_OMIT_AUTOVACUUM
9944 if( pBt
->autoVacuum
){
9945 pCheck
->zPfx
= "On page %u at right child: ";
9946 checkPtrmap(pCheck
, pgno
, PTRMAP_BTREE
, iPage
);
9949 depth
= checkTreePage(pCheck
, pgno
, &maxKey
, maxKey
);
9952 /* For leaf pages, the coverage check will occur in the same loop
9953 ** as the other cell checks, so initialize the heap. */
9954 heap
= pCheck
->heap
;
9958 /* EVIDENCE-OF: R-02776-14802 The cell pointer array consists of K 2-byte
9959 ** integer offsets to the cell contents. */
9960 for(i
=nCell
-1; i
>=0 && pCheck
->mxErr
; i
--){
9963 /* Check cell size */
9965 assert( pCellIdx
==&data
[cellStart
+ i
*2] );
9966 pc
= get2byteAligned(pCellIdx
);
9968 if( pc
<contentOffset
|| pc
>usableSize
-4 ){
9969 checkAppendMsg(pCheck
, "Offset %d out of range %d..%d",
9970 pc
, contentOffset
, usableSize
-4);
9971 doCoverageCheck
= 0;
9975 pPage
->xParseCell(pPage
, pCell
, &info
);
9976 if( pc
+info
.nSize
>usableSize
){
9977 checkAppendMsg(pCheck
, "Extends off end of page");
9978 doCoverageCheck
= 0;
9982 /* Check for integer primary key out of range */
9983 if( pPage
->intKey
){
9984 if( keyCanBeEqual
? (info
.nKey
> maxKey
) : (info
.nKey
>= maxKey
) ){
9985 checkAppendMsg(pCheck
, "Rowid %lld out of order", info
.nKey
);
9988 keyCanBeEqual
= 0; /* Only the first key on the page may ==maxKey */
9991 /* Check the content overflow list */
9992 if( info
.nPayload
>info
.nLocal
){
9993 u32 nPage
; /* Number of pages on the overflow chain */
9994 Pgno pgnoOvfl
; /* First page of the overflow chain */
9995 assert( pc
+ info
.nSize
- 4 <= usableSize
);
9996 nPage
= (info
.nPayload
- info
.nLocal
+ usableSize
- 5)/(usableSize
- 4);
9997 pgnoOvfl
= get4byte(&pCell
[info
.nSize
- 4]);
9998 #ifndef SQLITE_OMIT_AUTOVACUUM
9999 if( pBt
->autoVacuum
){
10000 checkPtrmap(pCheck
, pgnoOvfl
, PTRMAP_OVERFLOW1
, iPage
);
10003 checkList(pCheck
, 0, pgnoOvfl
, nPage
);
10006 if( !pPage
->leaf
){
10007 /* Check sanity of left child page for internal pages */
10008 pgno
= get4byte(pCell
);
10009 #ifndef SQLITE_OMIT_AUTOVACUUM
10010 if( pBt
->autoVacuum
){
10011 checkPtrmap(pCheck
, pgno
, PTRMAP_BTREE
, iPage
);
10014 d2
= checkTreePage(pCheck
, pgno
, &maxKey
, maxKey
);
10017 checkAppendMsg(pCheck
, "Child page depth differs");
10021 /* Populate the coverage-checking heap for leaf pages */
10022 btreeHeapInsert(heap
, (pc
<<16)|(pc
+info
.nSize
-1));
10025 *piMinKey
= maxKey
;
10027 /* Check for complete coverage of the page
10030 if( doCoverageCheck
&& pCheck
->mxErr
>0 ){
10031 /* For leaf pages, the min-heap has already been initialized and the
10032 ** cells have already been inserted. But for internal pages, that has
10033 ** not yet been done, so do it now */
10034 if( !pPage
->leaf
){
10035 heap
= pCheck
->heap
;
10037 for(i
=nCell
-1; i
>=0; i
--){
10039 pc
= get2byteAligned(&data
[cellStart
+i
*2]);
10040 size
= pPage
->xCellSize(pPage
, &data
[pc
]);
10041 btreeHeapInsert(heap
, (pc
<<16)|(pc
+size
-1));
10044 /* Add the freeblocks to the min-heap
10046 ** EVIDENCE-OF: R-20690-50594 The second field of the b-tree page header
10047 ** is the offset of the first freeblock, or zero if there are no
10048 ** freeblocks on the page.
10050 i
= get2byte(&data
[hdr
+1]);
10053 assert( (u32
)i
<=usableSize
-4 ); /* Enforced by btreeComputeFreeSpace() */
10054 size
= get2byte(&data
[i
+2]);
10055 assert( (u32
)(i
+size
)<=usableSize
); /* due to btreeComputeFreeSpace() */
10056 btreeHeapInsert(heap
, (((u32
)i
)<<16)|(i
+size
-1));
10057 /* EVIDENCE-OF: R-58208-19414 The first 2 bytes of a freeblock are a
10058 ** big-endian integer which is the offset in the b-tree page of the next
10059 ** freeblock in the chain, or zero if the freeblock is the last on the
10061 j
= get2byte(&data
[i
]);
10062 /* EVIDENCE-OF: R-06866-39125 Freeblocks are always connected in order of
10063 ** increasing offset. */
10064 assert( j
==0 || j
>i
+size
); /* Enforced by btreeComputeFreeSpace() */
10065 assert( (u32
)j
<=usableSize
-4 ); /* Enforced by btreeComputeFreeSpace() */
10068 /* Analyze the min-heap looking for overlap between cells and/or
10069 ** freeblocks, and counting the number of untracked bytes in nFrag.
10071 ** Each min-heap entry is of the form: (start_address<<16)|end_address.
10072 ** There is an implied first entry the covers the page header, the cell
10073 ** pointer index, and the gap between the cell pointer index and the start
10074 ** of cell content.
10076 ** The loop below pulls entries from the min-heap in order and compares
10077 ** the start_address against the previous end_address. If there is an
10078 ** overlap, that means bytes are used multiple times. If there is a gap,
10079 ** that gap is added to the fragmentation count.
10082 prev
= contentOffset
- 1; /* Implied first min-heap entry */
10083 while( btreeHeapPull(heap
,&x
) ){
10084 if( (prev
&0xffff)>=(x
>>16) ){
10085 checkAppendMsg(pCheck
,
10086 "Multiple uses for byte %u of page %u", x
>>16, iPage
);
10089 nFrag
+= (x
>>16) - (prev
&0xffff) - 1;
10093 nFrag
+= usableSize
- (prev
&0xffff) - 1;
10094 /* EVIDENCE-OF: R-43263-13491 The total number of bytes in all fragments
10095 ** is stored in the fifth field of the b-tree page header.
10096 ** EVIDENCE-OF: R-07161-27322 The one-byte integer at offset 7 gives the
10097 ** number of fragmented free bytes within the cell content area.
10099 if( heap
[0]==0 && nFrag
!=data
[hdr
+7] ){
10100 checkAppendMsg(pCheck
,
10101 "Fragmentation of %d bytes reported as %d on page %u",
10102 nFrag
, data
[hdr
+7], iPage
);
10107 if( !doCoverageCheck
) pPage
->isInit
= savedIsInit
;
10108 releasePage(pPage
);
10109 pCheck
->zPfx
= saved_zPfx
;
10110 pCheck
->v1
= saved_v1
;
10111 pCheck
->v2
= saved_v2
;
10114 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
10116 #ifndef SQLITE_OMIT_INTEGRITY_CHECK
10118 ** This routine does a complete check of the given BTree file. aRoot[] is
10119 ** an array of pages numbers were each page number is the root page of
10120 ** a table. nRoot is the number of entries in aRoot.
10122 ** A read-only or read-write transaction must be opened before calling
10125 ** Write the number of error seen in *pnErr. Except for some memory
10126 ** allocation errors, an error message held in memory obtained from
10127 ** malloc is returned if *pnErr is non-zero. If *pnErr==0 then NULL is
10128 ** returned. If a memory allocation error occurs, NULL is returned.
10130 ** If the first entry in aRoot[] is 0, that indicates that the list of
10131 ** root pages is incomplete. This is a "partial integrity-check". This
10132 ** happens when performing an integrity check on a single table. The
10133 ** zero is skipped, of course. But in addition, the freelist checks
10134 ** and the checks to make sure every page is referenced are also skipped,
10135 ** since obviously it is not possible to know which pages are covered by
10136 ** the unverified btrees. Except, if aRoot[1] is 1, then the freelist
10137 ** checks are still performed.
10139 char *sqlite3BtreeIntegrityCheck(
10140 sqlite3
*db
, /* Database connection that is running the check */
10141 Btree
*p
, /* The btree to be checked */
10142 Pgno
*aRoot
, /* An array of root pages numbers for individual trees */
10143 int nRoot
, /* Number of entries in aRoot[] */
10144 int mxErr
, /* Stop reporting errors after this many */
10145 int *pnErr
/* Write number of errors seen to this variable */
10148 IntegrityCk sCheck
;
10149 BtShared
*pBt
= p
->pBt
;
10150 u64 savedDbFlags
= pBt
->db
->flags
;
10152 int bPartial
= 0; /* True if not checking all btrees */
10153 int bCkFreelist
= 1; /* True to scan the freelist */
10154 VVA_ONLY( int nRef
);
10157 /* aRoot[0]==0 means this is a partial check */
10161 if( aRoot
[1]!=1 ) bCkFreelist
= 0;
10164 sqlite3BtreeEnter(p
);
10165 assert( p
->inTrans
>TRANS_NONE
&& pBt
->inTransaction
>TRANS_NONE
);
10166 VVA_ONLY( nRef
= sqlite3PagerRefcount(pBt
->pPager
) );
10170 sCheck
.pPager
= pBt
->pPager
;
10171 sCheck
.nPage
= btreePagecount(sCheck
.pBt
);
10172 sCheck
.mxErr
= mxErr
;
10174 sCheck
.bOomFault
= 0;
10180 sqlite3StrAccumInit(&sCheck
.errMsg
, 0, zErr
, sizeof(zErr
), SQLITE_MAX_LENGTH
);
10181 sCheck
.errMsg
.printfFlags
= SQLITE_PRINTF_INTERNAL
;
10182 if( sCheck
.nPage
==0 ){
10183 goto integrity_ck_cleanup
;
10186 sCheck
.aPgRef
= sqlite3MallocZero((sCheck
.nPage
/ 8)+ 1);
10187 if( !sCheck
.aPgRef
){
10188 sCheck
.bOomFault
= 1;
10189 goto integrity_ck_cleanup
;
10191 sCheck
.heap
= (u32
*)sqlite3PageMalloc( pBt
->pageSize
);
10192 if( sCheck
.heap
==0 ){
10193 sCheck
.bOomFault
= 1;
10194 goto integrity_ck_cleanup
;
10197 i
= PENDING_BYTE_PAGE(pBt
);
10198 if( i
<=sCheck
.nPage
) setPageReferenced(&sCheck
, i
);
10200 /* Check the integrity of the freelist
10203 sCheck
.zPfx
= "Main freelist: ";
10204 checkList(&sCheck
, 1, get4byte(&pBt
->pPage1
->aData
[32]),
10205 get4byte(&pBt
->pPage1
->aData
[36]));
10209 /* Check all the tables.
10211 #ifndef SQLITE_OMIT_AUTOVACUUM
10213 if( pBt
->autoVacuum
){
10216 for(i
=0; (int)i
<nRoot
; i
++) if( mx
<aRoot
[i
] ) mx
= aRoot
[i
];
10217 mxInHdr
= get4byte(&pBt
->pPage1
->aData
[52]);
10219 checkAppendMsg(&sCheck
,
10220 "max rootpage (%d) disagrees with header (%d)",
10224 }else if( get4byte(&pBt
->pPage1
->aData
[64])!=0 ){
10225 checkAppendMsg(&sCheck
,
10226 "incremental_vacuum enabled with a max rootpage of zero"
10231 testcase( pBt
->db
->flags
& SQLITE_CellSizeCk
);
10232 pBt
->db
->flags
&= ~(u64
)SQLITE_CellSizeCk
;
10233 for(i
=0; (int)i
<nRoot
&& sCheck
.mxErr
; i
++){
10235 if( aRoot
[i
]==0 ) continue;
10236 #ifndef SQLITE_OMIT_AUTOVACUUM
10237 if( pBt
->autoVacuum
&& aRoot
[i
]>1 && !bPartial
){
10238 checkPtrmap(&sCheck
, aRoot
[i
], PTRMAP_ROOTPAGE
, 0);
10241 checkTreePage(&sCheck
, aRoot
[i
], ¬Used
, LARGEST_INT64
);
10243 pBt
->db
->flags
= savedDbFlags
;
10245 /* Make sure every page in the file is referenced
10248 for(i
=1; i
<=sCheck
.nPage
&& sCheck
.mxErr
; i
++){
10249 #ifdef SQLITE_OMIT_AUTOVACUUM
10250 if( getPageReferenced(&sCheck
, i
)==0 ){
10251 checkAppendMsg(&sCheck
, "Page %d is never used", i
);
10254 /* If the database supports auto-vacuum, make sure no tables contain
10255 ** references to pointer-map pages.
10257 if( getPageReferenced(&sCheck
, i
)==0 &&
10258 (PTRMAP_PAGENO(pBt
, i
)!=i
|| !pBt
->autoVacuum
) ){
10259 checkAppendMsg(&sCheck
, "Page %d is never used", i
);
10261 if( getPageReferenced(&sCheck
, i
)!=0 &&
10262 (PTRMAP_PAGENO(pBt
, i
)==i
&& pBt
->autoVacuum
) ){
10263 checkAppendMsg(&sCheck
, "Pointer map page %d is referenced", i
);
10269 /* Clean up and report errors.
10271 integrity_ck_cleanup
:
10272 sqlite3PageFree(sCheck
.heap
);
10273 sqlite3_free(sCheck
.aPgRef
);
10274 if( sCheck
.bOomFault
){
10275 sqlite3_str_reset(&sCheck
.errMsg
);
10278 *pnErr
= sCheck
.nErr
;
10279 if( sCheck
.nErr
==0 ) sqlite3_str_reset(&sCheck
.errMsg
);
10280 /* Make sure this analysis did not leave any unref() pages. */
10281 assert( nRef
==sqlite3PagerRefcount(pBt
->pPager
) );
10282 sqlite3BtreeLeave(p
);
10283 return sqlite3StrAccumFinish(&sCheck
.errMsg
);
10285 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */
10288 ** Return the full pathname of the underlying database file. Return
10289 ** an empty string if the database is in-memory or a TEMP database.
10291 ** The pager filename is invariant as long as the pager is
10292 ** open so it is safe to access without the BtShared mutex.
10294 const char *sqlite3BtreeGetFilename(Btree
*p
){
10295 assert( p
->pBt
->pPager
!=0 );
10296 return sqlite3PagerFilename(p
->pBt
->pPager
, 1);
10300 ** Return the pathname of the journal file for this database. The return
10301 ** value of this routine is the same regardless of whether the journal file
10302 ** has been created or not.
10304 ** The pager journal filename is invariant as long as the pager is
10305 ** open so it is safe to access without the BtShared mutex.
10307 const char *sqlite3BtreeGetJournalname(Btree
*p
){
10308 assert( p
->pBt
->pPager
!=0 );
10309 return sqlite3PagerJournalname(p
->pBt
->pPager
);
10313 ** Return non-zero if a transaction is active.
10315 int sqlite3BtreeIsInTrans(Btree
*p
){
10316 assert( p
==0 || sqlite3_mutex_held(p
->db
->mutex
) );
10317 return (p
&& (p
->inTrans
==TRANS_WRITE
));
10320 #ifndef SQLITE_OMIT_WAL
10322 ** Run a checkpoint on the Btree passed as the first argument.
10324 ** Return SQLITE_LOCKED if this or any other connection has an open
10325 ** transaction on the shared-cache the argument Btree is connected to.
10327 ** Parameter eMode is one of SQLITE_CHECKPOINT_PASSIVE, FULL or RESTART.
10329 int sqlite3BtreeCheckpoint(Btree
*p
, int eMode
, int *pnLog
, int *pnCkpt
){
10330 int rc
= SQLITE_OK
;
10332 BtShared
*pBt
= p
->pBt
;
10333 sqlite3BtreeEnter(p
);
10334 if( pBt
->inTransaction
!=TRANS_NONE
){
10335 rc
= SQLITE_LOCKED
;
10337 rc
= sqlite3PagerCheckpoint(pBt
->pPager
, p
->db
, eMode
, pnLog
, pnCkpt
);
10339 sqlite3BtreeLeave(p
);
10346 ** Return non-zero if a read (or write) transaction is active.
10348 int sqlite3BtreeIsInReadTrans(Btree
*p
){
10350 assert( sqlite3_mutex_held(p
->db
->mutex
) );
10351 return p
->inTrans
!=TRANS_NONE
;
10354 int sqlite3BtreeIsInBackup(Btree
*p
){
10356 assert( sqlite3_mutex_held(p
->db
->mutex
) );
10357 return p
->nBackup
!=0;
10361 ** This function returns a pointer to a blob of memory associated with
10362 ** a single shared-btree. The memory is used by client code for its own
10363 ** purposes (for example, to store a high-level schema associated with
10364 ** the shared-btree). The btree layer manages reference counting issues.
10366 ** The first time this is called on a shared-btree, nBytes bytes of memory
10367 ** are allocated, zeroed, and returned to the caller. For each subsequent
10368 ** call the nBytes parameter is ignored and a pointer to the same blob
10369 ** of memory returned.
10371 ** If the nBytes parameter is 0 and the blob of memory has not yet been
10372 ** allocated, a null pointer is returned. If the blob has already been
10373 ** allocated, it is returned as normal.
10375 ** Just before the shared-btree is closed, the function passed as the
10376 ** xFree argument when the memory allocation was made is invoked on the
10377 ** blob of allocated memory. The xFree function should not call sqlite3_free()
10378 ** on the memory, the btree layer does that.
10380 void *sqlite3BtreeSchema(Btree
*p
, int nBytes
, void(*xFree
)(void *)){
10381 BtShared
*pBt
= p
->pBt
;
10382 sqlite3BtreeEnter(p
);
10383 if( !pBt
->pSchema
&& nBytes
){
10384 pBt
->pSchema
= sqlite3DbMallocZero(0, nBytes
);
10385 pBt
->xFreeSchema
= xFree
;
10387 sqlite3BtreeLeave(p
);
10388 return pBt
->pSchema
;
10392 ** Return SQLITE_LOCKED_SHAREDCACHE if another user of the same shared
10393 ** btree as the argument handle holds an exclusive lock on the
10394 ** sqlite_schema table. Otherwise SQLITE_OK.
10396 int sqlite3BtreeSchemaLocked(Btree
*p
){
10398 assert( sqlite3_mutex_held(p
->db
->mutex
) );
10399 sqlite3BtreeEnter(p
);
10400 rc
= querySharedCacheTableLock(p
, SCHEMA_ROOT
, READ_LOCK
);
10401 assert( rc
==SQLITE_OK
|| rc
==SQLITE_LOCKED_SHAREDCACHE
);
10402 sqlite3BtreeLeave(p
);
10407 #ifndef SQLITE_OMIT_SHARED_CACHE
10409 ** Obtain a lock on the table whose root page is iTab. The
10410 ** lock is a write lock if isWritelock is true or a read lock
10413 int sqlite3BtreeLockTable(Btree
*p
, int iTab
, u8 isWriteLock
){
10414 int rc
= SQLITE_OK
;
10415 assert( p
->inTrans
!=TRANS_NONE
);
10417 u8 lockType
= READ_LOCK
+ isWriteLock
;
10418 assert( READ_LOCK
+1==WRITE_LOCK
);
10419 assert( isWriteLock
==0 || isWriteLock
==1 );
10421 sqlite3BtreeEnter(p
);
10422 rc
= querySharedCacheTableLock(p
, iTab
, lockType
);
10423 if( rc
==SQLITE_OK
){
10424 rc
= setSharedCacheTableLock(p
, iTab
, lockType
);
10426 sqlite3BtreeLeave(p
);
10432 #ifndef SQLITE_OMIT_INCRBLOB
10434 ** Argument pCsr must be a cursor opened for writing on an
10435 ** INTKEY table currently pointing at a valid table entry.
10436 ** This function modifies the data stored as part of that entry.
10438 ** Only the data content may only be modified, it is not possible to
10439 ** change the length of the data stored. If this function is called with
10440 ** parameters that attempt to write past the end of the existing data,
10441 ** no modifications are made and SQLITE_CORRUPT is returned.
10443 int sqlite3BtreePutData(BtCursor
*pCsr
, u32 offset
, u32 amt
, void *z
){
10445 assert( cursorOwnsBtShared(pCsr
) );
10446 assert( sqlite3_mutex_held(pCsr
->pBtree
->db
->mutex
) );
10447 assert( pCsr
->curFlags
& BTCF_Incrblob
);
10449 rc
= restoreCursorPosition(pCsr
);
10450 if( rc
!=SQLITE_OK
){
10453 assert( pCsr
->eState
!=CURSOR_REQUIRESEEK
);
10454 if( pCsr
->eState
!=CURSOR_VALID
){
10455 return SQLITE_ABORT
;
10458 /* Save the positions of all other cursors open on this table. This is
10459 ** required in case any of them are holding references to an xFetch
10460 ** version of the b-tree page modified by the accessPayload call below.
10462 ** Note that pCsr must be open on a INTKEY table and saveCursorPosition()
10463 ** and hence saveAllCursors() cannot fail on a BTREE_INTKEY table, hence
10464 ** saveAllCursors can only return SQLITE_OK.
10466 VVA_ONLY(rc
=) saveAllCursors(pCsr
->pBt
, pCsr
->pgnoRoot
, pCsr
);
10467 assert( rc
==SQLITE_OK
);
10469 /* Check some assumptions:
10470 ** (a) the cursor is open for writing,
10471 ** (b) there is a read/write transaction open,
10472 ** (c) the connection holds a write-lock on the table (if required),
10473 ** (d) there are no conflicting read-locks, and
10474 ** (e) the cursor points at a valid row of an intKey table.
10476 if( (pCsr
->curFlags
& BTCF_WriteFlag
)==0 ){
10477 return SQLITE_READONLY
;
10479 assert( (pCsr
->pBt
->btsFlags
& BTS_READ_ONLY
)==0
10480 && pCsr
->pBt
->inTransaction
==TRANS_WRITE
);
10481 assert( hasSharedCacheTableLock(pCsr
->pBtree
, pCsr
->pgnoRoot
, 0, 2) );
10482 assert( !hasReadConflicts(pCsr
->pBtree
, pCsr
->pgnoRoot
) );
10483 assert( pCsr
->pPage
->intKey
);
10485 return accessPayload(pCsr
, offset
, amt
, (unsigned char *)z
, 1);
10489 ** Mark this cursor as an incremental blob cursor.
10491 void sqlite3BtreeIncrblobCursor(BtCursor
*pCur
){
10492 pCur
->curFlags
|= BTCF_Incrblob
;
10493 pCur
->pBtree
->hasIncrblobCur
= 1;
10498 ** Set both the "read version" (single byte at byte offset 18) and
10499 ** "write version" (single byte at byte offset 19) fields in the database
10500 ** header to iVersion.
10502 int sqlite3BtreeSetVersion(Btree
*pBtree
, int iVersion
){
10503 BtShared
*pBt
= pBtree
->pBt
;
10504 int rc
; /* Return code */
10506 assert( iVersion
==1 || iVersion
==2 );
10508 /* If setting the version fields to 1, do not automatically open the
10509 ** WAL connection, even if the version fields are currently set to 2.
10511 pBt
->btsFlags
&= ~BTS_NO_WAL
;
10512 if( iVersion
==1 ) pBt
->btsFlags
|= BTS_NO_WAL
;
10514 rc
= sqlite3BtreeBeginTrans(pBtree
, 0, 0);
10515 if( rc
==SQLITE_OK
){
10516 u8
*aData
= pBt
->pPage1
->aData
;
10517 if( aData
[18]!=(u8
)iVersion
|| aData
[19]!=(u8
)iVersion
){
10518 rc
= sqlite3BtreeBeginTrans(pBtree
, 2, 0);
10519 if( rc
==SQLITE_OK
){
10520 rc
= sqlite3PagerWrite(pBt
->pPage1
->pDbPage
);
10521 if( rc
==SQLITE_OK
){
10522 aData
[18] = (u8
)iVersion
;
10523 aData
[19] = (u8
)iVersion
;
10529 pBt
->btsFlags
&= ~BTS_NO_WAL
;
10534 ** Return true if the cursor has a hint specified. This routine is
10535 ** only used from within assert() statements
10537 int sqlite3BtreeCursorHasHint(BtCursor
*pCsr
, unsigned int mask
){
10538 return (pCsr
->hints
& mask
)!=0;
10542 ** Return true if the given Btree is read-only.
10544 int sqlite3BtreeIsReadonly(Btree
*p
){
10545 return (p
->pBt
->btsFlags
& BTS_READ_ONLY
)!=0;
10549 ** Return the size of the header added to each page by this module.
10551 int sqlite3HeaderSizeBtree(void){ return ROUND8(sizeof(MemPage
)); }
10553 #if !defined(SQLITE_OMIT_SHARED_CACHE)
10555 ** Return true if the Btree passed as the only argument is sharable.
10557 int sqlite3BtreeSharable(Btree
*p
){
10558 return p
->sharable
;
10562 ** Return the number of connections to the BtShared object accessed by
10563 ** the Btree handle passed as the only argument. For private caches
10564 ** this is always 1. For shared caches it may be 1 or greater.
10566 int sqlite3BtreeConnectionCount(Btree
*p
){
10567 testcase( p
->sharable
);
10568 return p
->pBt
->nRef
;