OLD | NEW |
| (Empty) |
1 /* | |
2 ** 2004 April 6 | |
3 ** | |
4 ** The author disclaims copyright to this source code. In place of | |
5 ** a legal notice, here is a blessing: | |
6 ** | |
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. | |
10 ** | |
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. | |
15 */ | |
16 #include "btreeInt.h" | |
17 | |
18 /* | |
19 ** The header string that appears at the beginning of every | |
20 ** SQLite database. | |
21 */ | |
22 static const char zMagicHeader[] = SQLITE_FILE_HEADER; | |
23 | |
24 /* | |
25 ** Set this global variable to 1 to enable tracing using the TRACE | |
26 ** macro. | |
27 */ | |
28 #if 0 | |
29 int sqlite3BtreeTrace=1; /* True to enable tracing */ | |
30 # define TRACE(X) if(sqlite3BtreeTrace){printf X;fflush(stdout);} | |
31 #else | |
32 # define TRACE(X) | |
33 #endif | |
34 | |
35 /* | |
36 ** Extract a 2-byte big-endian integer from an array of unsigned bytes. | |
37 ** But if the value is zero, make it 65536. | |
38 ** | |
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. | |
43 */ | |
44 #define get2byteNotZero(X) (((((int)get2byte(X))-1)&0xffff)+1) | |
45 | |
46 /* | |
47 ** Values passed as the 5th argument to allocateBtreePage() | |
48 */ | |
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 */ | |
52 | |
53 /* | |
54 ** Macro IfNotOmitAV(x) returns (x) if SQLITE_OMIT_AUTOVACUUM is not | |
55 ** defined, or 0 if it is. For example: | |
56 ** | |
57 ** bIncrVacuum = IfNotOmitAV(pBtShared->incrVacuum); | |
58 */ | |
59 #ifndef SQLITE_OMIT_AUTOVACUUM | |
60 #define IfNotOmitAV(expr) (expr) | |
61 #else | |
62 #define IfNotOmitAV(expr) 0 | |
63 #endif | |
64 | |
65 #ifndef SQLITE_OMIT_SHARED_CACHE | |
66 /* | |
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 | |
70 ** test builds. | |
71 ** | |
72 ** Access to this variable is protected by SQLITE_MUTEX_STATIC_MASTER. | |
73 */ | |
74 #ifdef SQLITE_TEST | |
75 BtShared *SQLITE_WSD sqlite3SharedCacheList = 0; | |
76 #else | |
77 static BtShared *SQLITE_WSD sqlite3SharedCacheList = 0; | |
78 #endif | |
79 #endif /* SQLITE_OMIT_SHARED_CACHE */ | |
80 | |
81 #ifndef SQLITE_OMIT_SHARED_CACHE | |
82 /* | |
83 ** Enable or disable the shared pager and schema features. | |
84 ** | |
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(). | |
88 */ | |
89 int sqlite3_enable_shared_cache(int enable){ | |
90 sqlite3GlobalConfig.sharedCacheEnabled = enable; | |
91 return SQLITE_OK; | |
92 } | |
93 #endif | |
94 | |
95 | |
96 | |
97 #ifdef SQLITE_OMIT_SHARED_CACHE | |
98 /* | |
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. | |
106 */ | |
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 | |
113 #endif | |
114 | |
115 #ifndef SQLITE_OMIT_SHARED_CACHE | |
116 | |
117 #ifdef SQLITE_DEBUG | |
118 /* | |
119 **** This function is only used as part of an assert() statement. *** | |
120 ** | |
121 ** Check to see if pBtree holds the required locks to read or write to the | |
122 ** table with root page iRoot. Return 1 if it does and 0 if not. | |
123 ** | |
124 ** For example, when writing to a table with root-page iRoot via | |
125 ** Btree connection pBtree: | |
126 ** | |
127 ** assert( hasSharedCacheTableLock(pBtree, iRoot, 0, WRITE_LOCK) ); | |
128 ** | |
129 ** When writing to an index that resides in a sharable database, the | |
130 ** caller should have first obtained a lock specifying the root page of | |
131 ** the corresponding table. This makes things a bit more complicated, | |
132 ** as this module treats each table as a separate structure. To determine | |
133 ** the table corresponding to the index being written, this | |
134 ** function has to search through the database schema. | |
135 ** | |
136 ** Instead of a lock on the table/index rooted at page iRoot, the caller may | |
137 ** hold a write-lock on the schema table (root page 1). This is also | |
138 ** acceptable. | |
139 */ | |
140 static int hasSharedCacheTableLock( | |
141 Btree *pBtree, /* Handle that must hold lock */ | |
142 Pgno iRoot, /* Root page of b-tree */ | |
143 int isIndex, /* True if iRoot is the root of an index b-tree */ | |
144 int eLockType /* Required lock type (READ_LOCK or WRITE_LOCK) */ | |
145 ){ | |
146 Schema *pSchema = (Schema *)pBtree->pBt->pSchema; | |
147 Pgno iTab = 0; | |
148 BtLock *pLock; | |
149 | |
150 /* If this database is not shareable, or if the client is reading | |
151 ** and has the read-uncommitted flag set, then no lock is required. | |
152 ** Return true immediately. | |
153 */ | |
154 if( (pBtree->sharable==0) | |
155 || (eLockType==READ_LOCK && (pBtree->db->flags & SQLITE_ReadUncommitted)) | |
156 ){ | |
157 return 1; | |
158 } | |
159 | |
160 /* If the client is reading or writing an index and the schema is | |
161 ** not loaded, then it is too difficult to actually check to see if | |
162 ** the correct locks are held. So do not bother - just return true. | |
163 ** This case does not come up very often anyhow. | |
164 */ | |
165 if( isIndex && (!pSchema || (pSchema->schemaFlags&DB_SchemaLoaded)==0) ){ | |
166 return 1; | |
167 } | |
168 | |
169 /* Figure out the root-page that the lock should be held on. For table | |
170 ** b-trees, this is just the root page of the b-tree being read or | |
171 ** written. For index b-trees, it is the root page of the associated | |
172 ** table. */ | |
173 if( isIndex ){ | |
174 HashElem *p; | |
175 for(p=sqliteHashFirst(&pSchema->idxHash); p; p=sqliteHashNext(p)){ | |
176 Index *pIdx = (Index *)sqliteHashData(p); | |
177 if( pIdx->tnum==(int)iRoot ){ | |
178 if( iTab ){ | |
179 /* Two or more indexes share the same root page. There must | |
180 ** be imposter tables. So just return true. The assert is not | |
181 ** useful in that case. */ | |
182 return 1; | |
183 } | |
184 iTab = pIdx->pTable->tnum; | |
185 } | |
186 } | |
187 }else{ | |
188 iTab = iRoot; | |
189 } | |
190 | |
191 /* Search for the required lock. Either a write-lock on root-page iTab, a | |
192 ** write-lock on the schema table, or (if the client is reading) a | |
193 ** read-lock on iTab will suffice. Return 1 if any of these are found. */ | |
194 for(pLock=pBtree->pBt->pLock; pLock; pLock=pLock->pNext){ | |
195 if( pLock->pBtree==pBtree | |
196 && (pLock->iTable==iTab || (pLock->eLock==WRITE_LOCK && pLock->iTable==1)) | |
197 && pLock->eLock>=eLockType | |
198 ){ | |
199 return 1; | |
200 } | |
201 } | |
202 | |
203 /* Failed to find the required lock. */ | |
204 return 0; | |
205 } | |
206 #endif /* SQLITE_DEBUG */ | |
207 | |
208 #ifdef SQLITE_DEBUG | |
209 /* | |
210 **** This function may be used as part of assert() statements only. **** | |
211 ** | |
212 ** Return true if it would be illegal for pBtree to write into the | |
213 ** table or index rooted at iRoot because other shared connections are | |
214 ** simultaneously reading that same table or index. | |
215 ** | |
216 ** It is illegal for pBtree to write if some other Btree object that | |
217 ** shares the same BtShared object is currently reading or writing | |
218 ** the iRoot table. Except, if the other Btree object has the | |
219 ** read-uncommitted flag set, then it is OK for the other object to | |
220 ** have a read cursor. | |
221 ** | |
222 ** For example, before writing to any part of the table or index | |
223 ** rooted at page iRoot, one should call: | |
224 ** | |
225 ** assert( !hasReadConflicts(pBtree, iRoot) ); | |
226 */ | |
227 static int hasReadConflicts(Btree *pBtree, Pgno iRoot){ | |
228 BtCursor *p; | |
229 for(p=pBtree->pBt->pCursor; p; p=p->pNext){ | |
230 if( p->pgnoRoot==iRoot | |
231 && p->pBtree!=pBtree | |
232 && 0==(p->pBtree->db->flags & SQLITE_ReadUncommitted) | |
233 ){ | |
234 return 1; | |
235 } | |
236 } | |
237 return 0; | |
238 } | |
239 #endif /* #ifdef SQLITE_DEBUG */ | |
240 | |
241 /* | |
242 ** Query to see if Btree handle p may obtain a lock of type eLock | |
243 ** (READ_LOCK or WRITE_LOCK) on the table with root-page iTab. Return | |
244 ** SQLITE_OK if the lock may be obtained (by calling | |
245 ** setSharedCacheTableLock()), or SQLITE_LOCKED if not. | |
246 */ | |
247 static int querySharedCacheTableLock(Btree *p, Pgno iTab, u8 eLock){ | |
248 BtShared *pBt = p->pBt; | |
249 BtLock *pIter; | |
250 | |
251 assert( sqlite3BtreeHoldsMutex(p) ); | |
252 assert( eLock==READ_LOCK || eLock==WRITE_LOCK ); | |
253 assert( p->db!=0 ); | |
254 assert( !(p->db->flags&SQLITE_ReadUncommitted)||eLock==WRITE_LOCK||iTab==1 ); | |
255 | |
256 /* If requesting a write-lock, then the Btree must have an open write | |
257 ** transaction on this file. And, obviously, for this to be so there | |
258 ** must be an open write transaction on the file itself. | |
259 */ | |
260 assert( eLock==READ_LOCK || (p==pBt->pWriter && p->inTrans==TRANS_WRITE) ); | |
261 assert( eLock==READ_LOCK || pBt->inTransaction==TRANS_WRITE ); | |
262 | |
263 /* This routine is a no-op if the shared-cache is not enabled */ | |
264 if( !p->sharable ){ | |
265 return SQLITE_OK; | |
266 } | |
267 | |
268 /* If some other connection is holding an exclusive lock, the | |
269 ** requested lock may not be obtained. | |
270 */ | |
271 if( pBt->pWriter!=p && (pBt->btsFlags & BTS_EXCLUSIVE)!=0 ){ | |
272 sqlite3ConnectionBlocked(p->db, pBt->pWriter->db); | |
273 return SQLITE_LOCKED_SHAREDCACHE; | |
274 } | |
275 | |
276 for(pIter=pBt->pLock; pIter; pIter=pIter->pNext){ | |
277 /* The condition (pIter->eLock!=eLock) in the following if(...) | |
278 ** statement is a simplification of: | |
279 ** | |
280 ** (eLock==WRITE_LOCK || pIter->eLock==WRITE_LOCK) | |
281 ** | |
282 ** since we know that if eLock==WRITE_LOCK, then no other connection | |
283 ** may hold a WRITE_LOCK on any table in this file (since there can | |
284 ** only be a single writer). | |
285 */ | |
286 assert( pIter->eLock==READ_LOCK || pIter->eLock==WRITE_LOCK ); | |
287 assert( eLock==READ_LOCK || pIter->pBtree==p || pIter->eLock==READ_LOCK); | |
288 if( pIter->pBtree!=p && pIter->iTable==iTab && pIter->eLock!=eLock ){ | |
289 sqlite3ConnectionBlocked(p->db, pIter->pBtree->db); | |
290 if( eLock==WRITE_LOCK ){ | |
291 assert( p==pBt->pWriter ); | |
292 pBt->btsFlags |= BTS_PENDING; | |
293 } | |
294 return SQLITE_LOCKED_SHAREDCACHE; | |
295 } | |
296 } | |
297 return SQLITE_OK; | |
298 } | |
299 #endif /* !SQLITE_OMIT_SHARED_CACHE */ | |
300 | |
301 #ifndef SQLITE_OMIT_SHARED_CACHE | |
302 /* | |
303 ** Add a lock on the table with root-page iTable to the shared-btree used | |
304 ** by Btree handle p. Parameter eLock must be either READ_LOCK or | |
305 ** WRITE_LOCK. | |
306 ** | |
307 ** This function assumes the following: | |
308 ** | |
309 ** (a) The specified Btree object p is connected to a sharable | |
310 ** database (one with the BtShared.sharable flag set), and | |
311 ** | |
312 ** (b) No other Btree objects hold a lock that conflicts | |
313 ** with the requested lock (i.e. querySharedCacheTableLock() has | |
314 ** already been called and returned SQLITE_OK). | |
315 ** | |
316 ** SQLITE_OK is returned if the lock is added successfully. SQLITE_NOMEM | |
317 ** is returned if a malloc attempt fails. | |
318 */ | |
319 static int setSharedCacheTableLock(Btree *p, Pgno iTable, u8 eLock){ | |
320 BtShared *pBt = p->pBt; | |
321 BtLock *pLock = 0; | |
322 BtLock *pIter; | |
323 | |
324 assert( sqlite3BtreeHoldsMutex(p) ); | |
325 assert( eLock==READ_LOCK || eLock==WRITE_LOCK ); | |
326 assert( p->db!=0 ); | |
327 | |
328 /* A connection with the read-uncommitted flag set will never try to | |
329 ** obtain a read-lock using this function. The only read-lock obtained | |
330 ** by a connection in read-uncommitted mode is on the sqlite_master | |
331 ** table, and that lock is obtained in BtreeBeginTrans(). */ | |
332 assert( 0==(p->db->flags&SQLITE_ReadUncommitted) || eLock==WRITE_LOCK ); | |
333 | |
334 /* This function should only be called on a sharable b-tree after it | |
335 ** has been determined that no other b-tree holds a conflicting lock. */ | |
336 assert( p->sharable ); | |
337 assert( SQLITE_OK==querySharedCacheTableLock(p, iTable, eLock) ); | |
338 | |
339 /* First search the list for an existing lock on this table. */ | |
340 for(pIter=pBt->pLock; pIter; pIter=pIter->pNext){ | |
341 if( pIter->iTable==iTable && pIter->pBtree==p ){ | |
342 pLock = pIter; | |
343 break; | |
344 } | |
345 } | |
346 | |
347 /* If the above search did not find a BtLock struct associating Btree p | |
348 ** with table iTable, allocate one and link it into the list. | |
349 */ | |
350 if( !pLock ){ | |
351 pLock = (BtLock *)sqlite3MallocZero(sizeof(BtLock)); | |
352 if( !pLock ){ | |
353 return SQLITE_NOMEM; | |
354 } | |
355 pLock->iTable = iTable; | |
356 pLock->pBtree = p; | |
357 pLock->pNext = pBt->pLock; | |
358 pBt->pLock = pLock; | |
359 } | |
360 | |
361 /* Set the BtLock.eLock variable to the maximum of the current lock | |
362 ** and the requested lock. This means if a write-lock was already held | |
363 ** and a read-lock requested, we don't incorrectly downgrade the lock. | |
364 */ | |
365 assert( WRITE_LOCK>READ_LOCK ); | |
366 if( eLock>pLock->eLock ){ | |
367 pLock->eLock = eLock; | |
368 } | |
369 | |
370 return SQLITE_OK; | |
371 } | |
372 #endif /* !SQLITE_OMIT_SHARED_CACHE */ | |
373 | |
374 #ifndef SQLITE_OMIT_SHARED_CACHE | |
375 /* | |
376 ** Release all the table locks (locks obtained via calls to | |
377 ** the setSharedCacheTableLock() procedure) held by Btree object p. | |
378 ** | |
379 ** This function assumes that Btree p has an open read or write | |
380 ** transaction. If it does not, then the BTS_PENDING flag | |
381 ** may be incorrectly cleared. | |
382 */ | |
383 static void clearAllSharedCacheTableLocks(Btree *p){ | |
384 BtShared *pBt = p->pBt; | |
385 BtLock **ppIter = &pBt->pLock; | |
386 | |
387 assert( sqlite3BtreeHoldsMutex(p) ); | |
388 assert( p->sharable || 0==*ppIter ); | |
389 assert( p->inTrans>0 ); | |
390 | |
391 while( *ppIter ){ | |
392 BtLock *pLock = *ppIter; | |
393 assert( (pBt->btsFlags & BTS_EXCLUSIVE)==0 || pBt->pWriter==pLock->pBtree ); | |
394 assert( pLock->pBtree->inTrans>=pLock->eLock ); | |
395 if( pLock->pBtree==p ){ | |
396 *ppIter = pLock->pNext; | |
397 assert( pLock->iTable!=1 || pLock==&p->lock ); | |
398 if( pLock->iTable!=1 ){ | |
399 sqlite3_free(pLock); | |
400 } | |
401 }else{ | |
402 ppIter = &pLock->pNext; | |
403 } | |
404 } | |
405 | |
406 assert( (pBt->btsFlags & BTS_PENDING)==0 || pBt->pWriter ); | |
407 if( pBt->pWriter==p ){ | |
408 pBt->pWriter = 0; | |
409 pBt->btsFlags &= ~(BTS_EXCLUSIVE|BTS_PENDING); | |
410 }else if( pBt->nTransaction==2 ){ | |
411 /* This function is called when Btree p is concluding its | |
412 ** transaction. If there currently exists a writer, and p is not | |
413 ** that writer, then the number of locks held by connections other | |
414 ** than the writer must be about to drop to zero. In this case | |
415 ** set the BTS_PENDING flag to 0. | |
416 ** | |
417 ** If there is not currently a writer, then BTS_PENDING must | |
418 ** be zero already. So this next line is harmless in that case. | |
419 */ | |
420 pBt->btsFlags &= ~BTS_PENDING; | |
421 } | |
422 } | |
423 | |
424 /* | |
425 ** This function changes all write-locks held by Btree p into read-locks. | |
426 */ | |
427 static void downgradeAllSharedCacheTableLocks(Btree *p){ | |
428 BtShared *pBt = p->pBt; | |
429 if( pBt->pWriter==p ){ | |
430 BtLock *pLock; | |
431 pBt->pWriter = 0; | |
432 pBt->btsFlags &= ~(BTS_EXCLUSIVE|BTS_PENDING); | |
433 for(pLock=pBt->pLock; pLock; pLock=pLock->pNext){ | |
434 assert( pLock->eLock==READ_LOCK || pLock->pBtree==p ); | |
435 pLock->eLock = READ_LOCK; | |
436 } | |
437 } | |
438 } | |
439 | |
440 #endif /* SQLITE_OMIT_SHARED_CACHE */ | |
441 | |
442 static void releasePage(MemPage *pPage); /* Forward reference */ | |
443 | |
444 /* | |
445 ***** This routine is used inside of assert() only **** | |
446 ** | |
447 ** Verify that the cursor holds the mutex on its BtShared | |
448 */ | |
449 #ifdef SQLITE_DEBUG | |
450 static int cursorHoldsMutex(BtCursor *p){ | |
451 return sqlite3_mutex_held(p->pBt->mutex); | |
452 } | |
453 #endif | |
454 | |
455 /* | |
456 ** Invalidate the overflow cache of the cursor passed as the first argument. | |
457 ** on the shared btree structure pBt. | |
458 */ | |
459 #define invalidateOverflowCache(pCur) (pCur->curFlags &= ~BTCF_ValidOvfl) | |
460 | |
461 /* | |
462 ** Invalidate the overflow page-list cache for all cursors opened | |
463 ** on the shared btree structure pBt. | |
464 */ | |
465 static void invalidateAllOverflowCache(BtShared *pBt){ | |
466 BtCursor *p; | |
467 assert( sqlite3_mutex_held(pBt->mutex) ); | |
468 for(p=pBt->pCursor; p; p=p->pNext){ | |
469 invalidateOverflowCache(p); | |
470 } | |
471 } | |
472 | |
473 #ifndef SQLITE_OMIT_INCRBLOB | |
474 /* | |
475 ** This function is called before modifying the contents of a table | |
476 ** to invalidate any incrblob cursors that are open on the | |
477 ** row or one of the rows being modified. | |
478 ** | |
479 ** If argument isClearTable is true, then the entire contents of the | |
480 ** table is about to be deleted. In this case invalidate all incrblob | |
481 ** cursors open on any row within the table with root-page pgnoRoot. | |
482 ** | |
483 ** Otherwise, if argument isClearTable is false, then the row with | |
484 ** rowid iRow is being replaced or deleted. In this case invalidate | |
485 ** only those incrblob cursors open on that specific row. | |
486 */ | |
487 static void invalidateIncrblobCursors( | |
488 Btree *pBtree, /* The database file to check */ | |
489 i64 iRow, /* The rowid that might be changing */ | |
490 int isClearTable /* True if all rows are being deleted */ | |
491 ){ | |
492 BtCursor *p; | |
493 if( pBtree->hasIncrblobCur==0 ) return; | |
494 assert( sqlite3BtreeHoldsMutex(pBtree) ); | |
495 pBtree->hasIncrblobCur = 0; | |
496 for(p=pBtree->pBt->pCursor; p; p=p->pNext){ | |
497 if( (p->curFlags & BTCF_Incrblob)!=0 ){ | |
498 pBtree->hasIncrblobCur = 1; | |
499 if( isClearTable || p->info.nKey==iRow ){ | |
500 p->eState = CURSOR_INVALID; | |
501 } | |
502 } | |
503 } | |
504 } | |
505 | |
506 #else | |
507 /* Stub function when INCRBLOB is omitted */ | |
508 #define invalidateIncrblobCursors(x,y,z) | |
509 #endif /* SQLITE_OMIT_INCRBLOB */ | |
510 | |
511 /* | |
512 ** Set bit pgno of the BtShared.pHasContent bitvec. This is called | |
513 ** when a page that previously contained data becomes a free-list leaf | |
514 ** page. | |
515 ** | |
516 ** The BtShared.pHasContent bitvec exists to work around an obscure | |
517 ** bug caused by the interaction of two useful IO optimizations surrounding | |
518 ** free-list leaf pages: | |
519 ** | |
520 ** 1) When all data is deleted from a page and the page becomes | |
521 ** a free-list leaf page, the page is not written to the database | |
522 ** (as free-list leaf pages contain no meaningful data). Sometimes | |
523 ** such a page is not even journalled (as it will not be modified, | |
524 ** why bother journalling it?). | |
525 ** | |
526 ** 2) When a free-list leaf page is reused, its content is not read | |
527 ** from the database or written to the journal file (why should it | |
528 ** be, if it is not at all meaningful?). | |
529 ** | |
530 ** By themselves, these optimizations work fine and provide a handy | |
531 ** performance boost to bulk delete or insert operations. However, if | |
532 ** a page is moved to the free-list and then reused within the same | |
533 ** transaction, a problem comes up. If the page is not journalled when | |
534 ** it is moved to the free-list and it is also not journalled when it | |
535 ** is extracted from the free-list and reused, then the original data | |
536 ** may be lost. In the event of a rollback, it may not be possible | |
537 ** to restore the database to its original configuration. | |
538 ** | |
539 ** The solution is the BtShared.pHasContent bitvec. Whenever a page is | |
540 ** moved to become a free-list leaf page, the corresponding bit is | |
541 ** set in the bitvec. Whenever a leaf page is extracted from the free-list, | |
542 ** optimization 2 above is omitted if the corresponding bit is already | |
543 ** set in BtShared.pHasContent. The contents of the bitvec are cleared | |
544 ** at the end of every transaction. | |
545 */ | |
546 static int btreeSetHasContent(BtShared *pBt, Pgno pgno){ | |
547 int rc = SQLITE_OK; | |
548 if( !pBt->pHasContent ){ | |
549 assert( pgno<=pBt->nPage ); | |
550 pBt->pHasContent = sqlite3BitvecCreate(pBt->nPage); | |
551 if( !pBt->pHasContent ){ | |
552 rc = SQLITE_NOMEM; | |
553 } | |
554 } | |
555 if( rc==SQLITE_OK && pgno<=sqlite3BitvecSize(pBt->pHasContent) ){ | |
556 rc = sqlite3BitvecSet(pBt->pHasContent, pgno); | |
557 } | |
558 return rc; | |
559 } | |
560 | |
561 /* | |
562 ** Query the BtShared.pHasContent vector. | |
563 ** | |
564 ** This function is called when a free-list leaf page is removed from the | |
565 ** free-list for reuse. It returns false if it is safe to retrieve the | |
566 ** page from the pager layer with the 'no-content' flag set. True otherwise. | |
567 */ | |
568 static int btreeGetHasContent(BtShared *pBt, Pgno pgno){ | |
569 Bitvec *p = pBt->pHasContent; | |
570 return (p && (pgno>sqlite3BitvecSize(p) || sqlite3BitvecTest(p, pgno))); | |
571 } | |
572 | |
573 /* | |
574 ** Clear (destroy) the BtShared.pHasContent bitvec. This should be | |
575 ** invoked at the conclusion of each write-transaction. | |
576 */ | |
577 static void btreeClearHasContent(BtShared *pBt){ | |
578 sqlite3BitvecDestroy(pBt->pHasContent); | |
579 pBt->pHasContent = 0; | |
580 } | |
581 | |
582 /* | |
583 ** Release all of the apPage[] pages for a cursor. | |
584 */ | |
585 static void btreeReleaseAllCursorPages(BtCursor *pCur){ | |
586 int i; | |
587 for(i=0; i<=pCur->iPage; i++){ | |
588 releasePage(pCur->apPage[i]); | |
589 pCur->apPage[i] = 0; | |
590 } | |
591 pCur->iPage = -1; | |
592 } | |
593 | |
594 /* | |
595 ** The cursor passed as the only argument must point to a valid entry | |
596 ** when this function is called (i.e. have eState==CURSOR_VALID). This | |
597 ** function saves the current cursor key in variables pCur->nKey and | |
598 ** pCur->pKey. SQLITE_OK is returned if successful or an SQLite error | |
599 ** code otherwise. | |
600 ** | |
601 ** If the cursor is open on an intkey table, then the integer key | |
602 ** (the rowid) is stored in pCur->nKey and pCur->pKey is left set to | |
603 ** NULL. If the cursor is open on a non-intkey table, then pCur->pKey is | |
604 ** set to point to a malloced buffer pCur->nKey bytes in size containing | |
605 ** the key. | |
606 */ | |
607 static int saveCursorKey(BtCursor *pCur){ | |
608 int rc; | |
609 assert( CURSOR_VALID==pCur->eState ); | |
610 assert( 0==pCur->pKey ); | |
611 assert( cursorHoldsMutex(pCur) ); | |
612 | |
613 rc = sqlite3BtreeKeySize(pCur, &pCur->nKey); | |
614 assert( rc==SQLITE_OK ); /* KeySize() cannot fail */ | |
615 | |
616 /* If this is an intKey table, then the above call to BtreeKeySize() | |
617 ** stores the integer key in pCur->nKey. In this case this value is | |
618 ** all that is required. Otherwise, if pCur is not open on an intKey | |
619 ** table, then malloc space for and store the pCur->nKey bytes of key | |
620 ** data. */ | |
621 if( 0==pCur->curIntKey ){ | |
622 void *pKey = sqlite3Malloc( pCur->nKey ); | |
623 if( pKey ){ | |
624 rc = sqlite3BtreeKey(pCur, 0, (int)pCur->nKey, pKey); | |
625 if( rc==SQLITE_OK ){ | |
626 pCur->pKey = pKey; | |
627 }else{ | |
628 sqlite3_free(pKey); | |
629 } | |
630 }else{ | |
631 rc = SQLITE_NOMEM; | |
632 } | |
633 } | |
634 assert( !pCur->curIntKey || !pCur->pKey ); | |
635 return rc; | |
636 } | |
637 | |
638 /* | |
639 ** Save the current cursor position in the variables BtCursor.nKey | |
640 ** and BtCursor.pKey. The cursor's state is set to CURSOR_REQUIRESEEK. | |
641 ** | |
642 ** The caller must ensure that the cursor is valid (has eState==CURSOR_VALID) | |
643 ** prior to calling this routine. | |
644 */ | |
645 static int saveCursorPosition(BtCursor *pCur){ | |
646 int rc; | |
647 | |
648 assert( CURSOR_VALID==pCur->eState || CURSOR_SKIPNEXT==pCur->eState ); | |
649 assert( 0==pCur->pKey ); | |
650 assert( cursorHoldsMutex(pCur) ); | |
651 | |
652 if( pCur->eState==CURSOR_SKIPNEXT ){ | |
653 pCur->eState = CURSOR_VALID; | |
654 }else{ | |
655 pCur->skipNext = 0; | |
656 } | |
657 | |
658 rc = saveCursorKey(pCur); | |
659 if( rc==SQLITE_OK ){ | |
660 btreeReleaseAllCursorPages(pCur); | |
661 pCur->eState = CURSOR_REQUIRESEEK; | |
662 } | |
663 | |
664 pCur->curFlags &= ~(BTCF_ValidNKey|BTCF_ValidOvfl|BTCF_AtLast); | |
665 return rc; | |
666 } | |
667 | |
668 /* Forward reference */ | |
669 static int SQLITE_NOINLINE saveCursorsOnList(BtCursor*,Pgno,BtCursor*); | |
670 | |
671 /* | |
672 ** Save the positions of all cursors (except pExcept) that are open on | |
673 ** the table with root-page iRoot. "Saving the cursor position" means that | |
674 ** the location in the btree is remembered in such a way that it can be | |
675 ** moved back to the same spot after the btree has been modified. This | |
676 ** routine is called just before cursor pExcept is used to modify the | |
677 ** table, for example in BtreeDelete() or BtreeInsert(). | |
678 ** | |
679 ** If there are two or more cursors on the same btree, then all such | |
680 ** cursors should have their BTCF_Multiple flag set. The btreeCursor() | |
681 ** routine enforces that rule. This routine only needs to be called in | |
682 ** the uncommon case when pExpect has the BTCF_Multiple flag set. | |
683 ** | |
684 ** If pExpect!=NULL and if no other cursors are found on the same root-page, | |
685 ** then the BTCF_Multiple flag on pExpect is cleared, to avoid another | |
686 ** pointless call to this routine. | |
687 ** | |
688 ** Implementation note: This routine merely checks to see if any cursors | |
689 ** need to be saved. It calls out to saveCursorsOnList() in the (unusual) | |
690 ** event that cursors are in need to being saved. | |
691 */ | |
692 static int saveAllCursors(BtShared *pBt, Pgno iRoot, BtCursor *pExcept){ | |
693 BtCursor *p; | |
694 assert( sqlite3_mutex_held(pBt->mutex) ); | |
695 assert( pExcept==0 || pExcept->pBt==pBt ); | |
696 for(p=pBt->pCursor; p; p=p->pNext){ | |
697 if( p!=pExcept && (0==iRoot || p->pgnoRoot==iRoot) ) break; | |
698 } | |
699 if( p ) return saveCursorsOnList(p, iRoot, pExcept); | |
700 if( pExcept ) pExcept->curFlags &= ~BTCF_Multiple; | |
701 return SQLITE_OK; | |
702 } | |
703 | |
704 /* This helper routine to saveAllCursors does the actual work of saving | |
705 ** the cursors if and when a cursor is found that actually requires saving. | |
706 ** The common case is that no cursors need to be saved, so this routine is | |
707 ** broken out from its caller to avoid unnecessary stack pointer movement. | |
708 */ | |
709 static int SQLITE_NOINLINE saveCursorsOnList( | |
710 BtCursor *p, /* The first cursor that needs saving */ | |
711 Pgno iRoot, /* Only save cursor with this iRoot. Save all if zero */ | |
712 BtCursor *pExcept /* Do not save this cursor */ | |
713 ){ | |
714 do{ | |
715 if( p!=pExcept && (0==iRoot || p->pgnoRoot==iRoot) ){ | |
716 if( p->eState==CURSOR_VALID || p->eState==CURSOR_SKIPNEXT ){ | |
717 int rc = saveCursorPosition(p); | |
718 if( SQLITE_OK!=rc ){ | |
719 return rc; | |
720 } | |
721 }else{ | |
722 testcase( p->iPage>0 ); | |
723 btreeReleaseAllCursorPages(p); | |
724 } | |
725 } | |
726 p = p->pNext; | |
727 }while( p ); | |
728 return SQLITE_OK; | |
729 } | |
730 | |
731 /* | |
732 ** Clear the current cursor position. | |
733 */ | |
734 void sqlite3BtreeClearCursor(BtCursor *pCur){ | |
735 assert( cursorHoldsMutex(pCur) ); | |
736 sqlite3_free(pCur->pKey); | |
737 pCur->pKey = 0; | |
738 pCur->eState = CURSOR_INVALID; | |
739 } | |
740 | |
741 /* | |
742 ** In this version of BtreeMoveto, pKey is a packed index record | |
743 ** such as is generated by the OP_MakeRecord opcode. Unpack the | |
744 ** record and then call BtreeMovetoUnpacked() to do the work. | |
745 */ | |
746 static int btreeMoveto( | |
747 BtCursor *pCur, /* Cursor open on the btree to be searched */ | |
748 const void *pKey, /* Packed key if the btree is an index */ | |
749 i64 nKey, /* Integer key for tables. Size of pKey for indices */ | |
750 int bias, /* Bias search to the high end */ | |
751 int *pRes /* Write search results here */ | |
752 ){ | |
753 int rc; /* Status code */ | |
754 UnpackedRecord *pIdxKey; /* Unpacked index key */ | |
755 char aSpace[200]; /* Temp space for pIdxKey - to avoid a malloc */ | |
756 char *pFree = 0; | |
757 | |
758 if( pKey ){ | |
759 assert( nKey==(i64)(int)nKey ); | |
760 pIdxKey = sqlite3VdbeAllocUnpackedRecord( | |
761 pCur->pKeyInfo, aSpace, sizeof(aSpace), &pFree | |
762 ); | |
763 if( pIdxKey==0 ) return SQLITE_NOMEM; | |
764 sqlite3VdbeRecordUnpack(pCur->pKeyInfo, (int)nKey, pKey, pIdxKey); | |
765 if( pIdxKey->nField==0 ){ | |
766 sqlite3DbFree(pCur->pKeyInfo->db, pFree); | |
767 return SQLITE_CORRUPT_BKPT; | |
768 } | |
769 }else{ | |
770 pIdxKey = 0; | |
771 } | |
772 rc = sqlite3BtreeMovetoUnpacked(pCur, pIdxKey, nKey, bias, pRes); | |
773 if( pFree ){ | |
774 sqlite3DbFree(pCur->pKeyInfo->db, pFree); | |
775 } | |
776 return rc; | |
777 } | |
778 | |
779 /* | |
780 ** Restore the cursor to the position it was in (or as close to as possible) | |
781 ** when saveCursorPosition() was called. Note that this call deletes the | |
782 ** saved position info stored by saveCursorPosition(), so there can be | |
783 ** at most one effective restoreCursorPosition() call after each | |
784 ** saveCursorPosition(). | |
785 */ | |
786 static int btreeRestoreCursorPosition(BtCursor *pCur){ | |
787 int rc; | |
788 int skipNext; | |
789 assert( cursorHoldsMutex(pCur) ); | |
790 assert( pCur->eState>=CURSOR_REQUIRESEEK ); | |
791 if( pCur->eState==CURSOR_FAULT ){ | |
792 return pCur->skipNext; | |
793 } | |
794 pCur->eState = CURSOR_INVALID; | |
795 rc = btreeMoveto(pCur, pCur->pKey, pCur->nKey, 0, &skipNext); | |
796 if( rc==SQLITE_OK ){ | |
797 sqlite3_free(pCur->pKey); | |
798 pCur->pKey = 0; | |
799 assert( pCur->eState==CURSOR_VALID || pCur->eState==CURSOR_INVALID ); | |
800 pCur->skipNext |= skipNext; | |
801 if( pCur->skipNext && pCur->eState==CURSOR_VALID ){ | |
802 pCur->eState = CURSOR_SKIPNEXT; | |
803 } | |
804 } | |
805 return rc; | |
806 } | |
807 | |
808 #define restoreCursorPosition(p) \ | |
809 (p->eState>=CURSOR_REQUIRESEEK ? \ | |
810 btreeRestoreCursorPosition(p) : \ | |
811 SQLITE_OK) | |
812 | |
813 /* | |
814 ** Determine whether or not a cursor has moved from the position where | |
815 ** it was last placed, or has been invalidated for any other reason. | |
816 ** Cursors can move when the row they are pointing at is deleted out | |
817 ** from under them, for example. Cursor might also move if a btree | |
818 ** is rebalanced. | |
819 ** | |
820 ** Calling this routine with a NULL cursor pointer returns false. | |
821 ** | |
822 ** Use the separate sqlite3BtreeCursorRestore() routine to restore a cursor | |
823 ** back to where it ought to be if this routine returns true. | |
824 */ | |
825 int sqlite3BtreeCursorHasMoved(BtCursor *pCur){ | |
826 return pCur->eState!=CURSOR_VALID; | |
827 } | |
828 | |
829 /* | |
830 ** This routine restores a cursor back to its original position after it | |
831 ** has been moved by some outside activity (such as a btree rebalance or | |
832 ** a row having been deleted out from under the cursor). | |
833 ** | |
834 ** On success, the *pDifferentRow parameter is false if the cursor is left | |
835 ** pointing at exactly the same row. *pDifferntRow is the row the cursor | |
836 ** was pointing to has been deleted, forcing the cursor to point to some | |
837 ** nearby row. | |
838 ** | |
839 ** This routine should only be called for a cursor that just returned | |
840 ** TRUE from sqlite3BtreeCursorHasMoved(). | |
841 */ | |
842 int sqlite3BtreeCursorRestore(BtCursor *pCur, int *pDifferentRow){ | |
843 int rc; | |
844 | |
845 assert( pCur!=0 ); | |
846 assert( pCur->eState!=CURSOR_VALID ); | |
847 rc = restoreCursorPosition(pCur); | |
848 if( rc ){ | |
849 *pDifferentRow = 1; | |
850 return rc; | |
851 } | |
852 if( pCur->eState!=CURSOR_VALID ){ | |
853 *pDifferentRow = 1; | |
854 }else{ | |
855 assert( pCur->skipNext==0 ); | |
856 *pDifferentRow = 0; | |
857 } | |
858 return SQLITE_OK; | |
859 } | |
860 | |
861 #ifdef SQLITE_ENABLE_CURSOR_HINTS | |
862 /* | |
863 ** Provide hints to the cursor. The particular hint given (and the type | |
864 ** and number of the varargs parameters) is determined by the eHintType | |
865 ** parameter. See the definitions of the BTREE_HINT_* macros for details. | |
866 */ | |
867 void sqlite3BtreeCursorHint(BtCursor *pCur, int eHintType, ...){ | |
868 /* Used only by system that substitute their own storage engine */ | |
869 } | |
870 #endif | |
871 | |
872 /* | |
873 ** Provide flag hints to the cursor. | |
874 */ | |
875 void sqlite3BtreeCursorHintFlags(BtCursor *pCur, unsigned x){ | |
876 assert( x==BTREE_SEEK_EQ || x==BTREE_BULKLOAD || x==0 ); | |
877 pCur->hints = x; | |
878 } | |
879 | |
880 | |
881 #ifndef SQLITE_OMIT_AUTOVACUUM | |
882 /* | |
883 ** Given a page number of a regular database page, return the page | |
884 ** number for the pointer-map page that contains the entry for the | |
885 ** input page number. | |
886 ** | |
887 ** Return 0 (not a valid page) for pgno==1 since there is | |
888 ** no pointer map associated with page 1. The integrity_check logic | |
889 ** requires that ptrmapPageno(*,1)!=1. | |
890 */ | |
891 static Pgno ptrmapPageno(BtShared *pBt, Pgno pgno){ | |
892 int nPagesPerMapPage; | |
893 Pgno iPtrMap, ret; | |
894 assert( sqlite3_mutex_held(pBt->mutex) ); | |
895 if( pgno<2 ) return 0; | |
896 nPagesPerMapPage = (pBt->usableSize/5)+1; | |
897 iPtrMap = (pgno-2)/nPagesPerMapPage; | |
898 ret = (iPtrMap*nPagesPerMapPage) + 2; | |
899 if( ret==PENDING_BYTE_PAGE(pBt) ){ | |
900 ret++; | |
901 } | |
902 return ret; | |
903 } | |
904 | |
905 /* | |
906 ** Write an entry into the pointer map. | |
907 ** | |
908 ** This routine updates the pointer map entry for page number 'key' | |
909 ** so that it maps to type 'eType' and parent page number 'pgno'. | |
910 ** | |
911 ** If *pRC is initially non-zero (non-SQLITE_OK) then this routine is | |
912 ** a no-op. If an error occurs, the appropriate error code is written | |
913 ** into *pRC. | |
914 */ | |
915 static void ptrmapPut(BtShared *pBt, Pgno key, u8 eType, Pgno parent, int *pRC){ | |
916 DbPage *pDbPage; /* The pointer map page */ | |
917 u8 *pPtrmap; /* The pointer map data */ | |
918 Pgno iPtrmap; /* The pointer map page number */ | |
919 int offset; /* Offset in pointer map page */ | |
920 int rc; /* Return code from subfunctions */ | |
921 | |
922 if( *pRC ) return; | |
923 | |
924 assert( sqlite3_mutex_held(pBt->mutex) ); | |
925 /* The master-journal page number must never be used as a pointer map page */ | |
926 assert( 0==PTRMAP_ISPAGE(pBt, PENDING_BYTE_PAGE(pBt)) ); | |
927 | |
928 assert( pBt->autoVacuum ); | |
929 if( key==0 ){ | |
930 *pRC = SQLITE_CORRUPT_BKPT; | |
931 return; | |
932 } | |
933 iPtrmap = PTRMAP_PAGENO(pBt, key); | |
934 rc = sqlite3PagerGet(pBt->pPager, iPtrmap, &pDbPage, 0); | |
935 if( rc!=SQLITE_OK ){ | |
936 *pRC = rc; | |
937 return; | |
938 } | |
939 offset = PTRMAP_PTROFFSET(iPtrmap, key); | |
940 if( offset<0 ){ | |
941 *pRC = SQLITE_CORRUPT_BKPT; | |
942 goto ptrmap_exit; | |
943 } | |
944 assert( offset <= (int)pBt->usableSize-5 ); | |
945 pPtrmap = (u8 *)sqlite3PagerGetData(pDbPage); | |
946 | |
947 if( eType!=pPtrmap[offset] || get4byte(&pPtrmap[offset+1])!=parent ){ | |
948 TRACE(("PTRMAP_UPDATE: %d->(%d,%d)\n", key, eType, parent)); | |
949 *pRC= rc = sqlite3PagerWrite(pDbPage); | |
950 if( rc==SQLITE_OK ){ | |
951 pPtrmap[offset] = eType; | |
952 put4byte(&pPtrmap[offset+1], parent); | |
953 } | |
954 } | |
955 | |
956 ptrmap_exit: | |
957 sqlite3PagerUnref(pDbPage); | |
958 } | |
959 | |
960 /* | |
961 ** Read an entry from the pointer map. | |
962 ** | |
963 ** This routine retrieves the pointer map entry for page 'key', writing | |
964 ** the type and parent page number to *pEType and *pPgno respectively. | |
965 ** An error code is returned if something goes wrong, otherwise SQLITE_OK. | |
966 */ | |
967 static int ptrmapGet(BtShared *pBt, Pgno key, u8 *pEType, Pgno *pPgno){ | |
968 DbPage *pDbPage; /* The pointer map page */ | |
969 int iPtrmap; /* Pointer map page index */ | |
970 u8 *pPtrmap; /* Pointer map page data */ | |
971 int offset; /* Offset of entry in pointer map */ | |
972 int rc; | |
973 | |
974 assert( sqlite3_mutex_held(pBt->mutex) ); | |
975 | |
976 iPtrmap = PTRMAP_PAGENO(pBt, key); | |
977 rc = sqlite3PagerGet(pBt->pPager, iPtrmap, &pDbPage, 0); | |
978 if( rc!=0 ){ | |
979 return rc; | |
980 } | |
981 pPtrmap = (u8 *)sqlite3PagerGetData(pDbPage); | |
982 | |
983 offset = PTRMAP_PTROFFSET(iPtrmap, key); | |
984 if( offset<0 ){ | |
985 sqlite3PagerUnref(pDbPage); | |
986 return SQLITE_CORRUPT_BKPT; | |
987 } | |
988 assert( offset <= (int)pBt->usableSize-5 ); | |
989 assert( pEType!=0 ); | |
990 *pEType = pPtrmap[offset]; | |
991 if( pPgno ) *pPgno = get4byte(&pPtrmap[offset+1]); | |
992 | |
993 sqlite3PagerUnref(pDbPage); | |
994 if( *pEType<1 || *pEType>5 ) return SQLITE_CORRUPT_BKPT; | |
995 return SQLITE_OK; | |
996 } | |
997 | |
998 #else /* if defined SQLITE_OMIT_AUTOVACUUM */ | |
999 #define ptrmapPut(w,x,y,z,rc) | |
1000 #define ptrmapGet(w,x,y,z) SQLITE_OK | |
1001 #define ptrmapPutOvflPtr(x, y, rc) | |
1002 #endif | |
1003 | |
1004 /* | |
1005 ** Given a btree page and a cell index (0 means the first cell on | |
1006 ** the page, 1 means the second cell, and so forth) return a pointer | |
1007 ** to the cell content. | |
1008 ** | |
1009 ** findCellPastPtr() does the same except it skips past the initial | |
1010 ** 4-byte child pointer found on interior pages, if there is one. | |
1011 ** | |
1012 ** This routine works only for pages that do not contain overflow cells. | |
1013 */ | |
1014 #define findCell(P,I) \ | |
1015 ((P)->aData + ((P)->maskPage & get2byteAligned(&(P)->aCellIdx[2*(I)]))) | |
1016 #define findCellPastPtr(P,I) \ | |
1017 ((P)->aDataOfst + ((P)->maskPage & get2byteAligned(&(P)->aCellIdx[2*(I)]))) | |
1018 | |
1019 | |
1020 /* | |
1021 ** This is common tail processing for btreeParseCellPtr() and | |
1022 ** btreeParseCellPtrIndex() for the case when the cell does not fit entirely | |
1023 ** on a single B-tree page. Make necessary adjustments to the CellInfo | |
1024 ** structure. | |
1025 */ | |
1026 static SQLITE_NOINLINE void btreeParseCellAdjustSizeForOverflow( | |
1027 MemPage *pPage, /* Page containing the cell */ | |
1028 u8 *pCell, /* Pointer to the cell text. */ | |
1029 CellInfo *pInfo /* Fill in this structure */ | |
1030 ){ | |
1031 /* If the payload will not fit completely on the local page, we have | |
1032 ** to decide how much to store locally and how much to spill onto | |
1033 ** overflow pages. The strategy is to minimize the amount of unused | |
1034 ** space on overflow pages while keeping the amount of local storage | |
1035 ** in between minLocal and maxLocal. | |
1036 ** | |
1037 ** Warning: changing the way overflow payload is distributed in any | |
1038 ** way will result in an incompatible file format. | |
1039 */ | |
1040 int minLocal; /* Minimum amount of payload held locally */ | |
1041 int maxLocal; /* Maximum amount of payload held locally */ | |
1042 int surplus; /* Overflow payload available for local storage */ | |
1043 | |
1044 minLocal = pPage->minLocal; | |
1045 maxLocal = pPage->maxLocal; | |
1046 surplus = minLocal + (pInfo->nPayload - minLocal)%(pPage->pBt->usableSize-4); | |
1047 testcase( surplus==maxLocal ); | |
1048 testcase( surplus==maxLocal+1 ); | |
1049 if( surplus <= maxLocal ){ | |
1050 pInfo->nLocal = (u16)surplus; | |
1051 }else{ | |
1052 pInfo->nLocal = (u16)minLocal; | |
1053 } | |
1054 pInfo->nSize = (u16)(&pInfo->pPayload[pInfo->nLocal] - pCell) + 4; | |
1055 } | |
1056 | |
1057 /* | |
1058 ** The following routines are implementations of the MemPage.xParseCell() | |
1059 ** method. | |
1060 ** | |
1061 ** Parse a cell content block and fill in the CellInfo structure. | |
1062 ** | |
1063 ** btreeParseCellPtr() => table btree leaf nodes | |
1064 ** btreeParseCellNoPayload() => table btree internal nodes | |
1065 ** btreeParseCellPtrIndex() => index btree nodes | |
1066 ** | |
1067 ** There is also a wrapper function btreeParseCell() that works for | |
1068 ** all MemPage types and that references the cell by index rather than | |
1069 ** by pointer. | |
1070 */ | |
1071 static void btreeParseCellPtrNoPayload( | |
1072 MemPage *pPage, /* Page containing the cell */ | |
1073 u8 *pCell, /* Pointer to the cell text. */ | |
1074 CellInfo *pInfo /* Fill in this structure */ | |
1075 ){ | |
1076 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
1077 assert( pPage->leaf==0 ); | |
1078 assert( pPage->noPayload ); | |
1079 assert( pPage->childPtrSize==4 ); | |
1080 #ifndef SQLITE_DEBUG | |
1081 UNUSED_PARAMETER(pPage); | |
1082 #endif | |
1083 pInfo->nSize = 4 + getVarint(&pCell[4], (u64*)&pInfo->nKey); | |
1084 pInfo->nPayload = 0; | |
1085 pInfo->nLocal = 0; | |
1086 pInfo->pPayload = 0; | |
1087 return; | |
1088 } | |
1089 static void btreeParseCellPtr( | |
1090 MemPage *pPage, /* Page containing the cell */ | |
1091 u8 *pCell, /* Pointer to the cell text. */ | |
1092 CellInfo *pInfo /* Fill in this structure */ | |
1093 ){ | |
1094 u8 *pIter; /* For scanning through pCell */ | |
1095 u32 nPayload; /* Number of bytes of cell payload */ | |
1096 u64 iKey; /* Extracted Key value */ | |
1097 | |
1098 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
1099 assert( pPage->leaf==0 || pPage->leaf==1 ); | |
1100 assert( pPage->intKeyLeaf || pPage->noPayload ); | |
1101 assert( pPage->noPayload==0 ); | |
1102 assert( pPage->intKeyLeaf ); | |
1103 assert( pPage->childPtrSize==0 ); | |
1104 pIter = pCell; | |
1105 | |
1106 /* The next block of code is equivalent to: | |
1107 ** | |
1108 ** pIter += getVarint32(pIter, nPayload); | |
1109 ** | |
1110 ** The code is inlined to avoid a function call. | |
1111 */ | |
1112 nPayload = *pIter; | |
1113 if( nPayload>=0x80 ){ | |
1114 u8 *pEnd = &pIter[8]; | |
1115 nPayload &= 0x7f; | |
1116 do{ | |
1117 nPayload = (nPayload<<7) | (*++pIter & 0x7f); | |
1118 }while( (*pIter)>=0x80 && pIter<pEnd ); | |
1119 } | |
1120 pIter++; | |
1121 | |
1122 /* The next block of code is equivalent to: | |
1123 ** | |
1124 ** pIter += getVarint(pIter, (u64*)&pInfo->nKey); | |
1125 ** | |
1126 ** The code is inlined to avoid a function call. | |
1127 */ | |
1128 iKey = *pIter; | |
1129 if( iKey>=0x80 ){ | |
1130 u8 *pEnd = &pIter[7]; | |
1131 iKey &= 0x7f; | |
1132 while(1){ | |
1133 iKey = (iKey<<7) | (*++pIter & 0x7f); | |
1134 if( (*pIter)<0x80 ) break; | |
1135 if( pIter>=pEnd ){ | |
1136 iKey = (iKey<<8) | *++pIter; | |
1137 break; | |
1138 } | |
1139 } | |
1140 } | |
1141 pIter++; | |
1142 | |
1143 pInfo->nKey = *(i64*)&iKey; | |
1144 pInfo->nPayload = nPayload; | |
1145 pInfo->pPayload = pIter; | |
1146 testcase( nPayload==pPage->maxLocal ); | |
1147 testcase( nPayload==pPage->maxLocal+1 ); | |
1148 if( nPayload<=pPage->maxLocal ){ | |
1149 /* This is the (easy) common case where the entire payload fits | |
1150 ** on the local page. No overflow is required. | |
1151 */ | |
1152 pInfo->nSize = nPayload + (u16)(pIter - pCell); | |
1153 if( pInfo->nSize<4 ) pInfo->nSize = 4; | |
1154 pInfo->nLocal = (u16)nPayload; | |
1155 }else{ | |
1156 btreeParseCellAdjustSizeForOverflow(pPage, pCell, pInfo); | |
1157 } | |
1158 } | |
1159 static void btreeParseCellPtrIndex( | |
1160 MemPage *pPage, /* Page containing the cell */ | |
1161 u8 *pCell, /* Pointer to the cell text. */ | |
1162 CellInfo *pInfo /* Fill in this structure */ | |
1163 ){ | |
1164 u8 *pIter; /* For scanning through pCell */ | |
1165 u32 nPayload; /* Number of bytes of cell payload */ | |
1166 | |
1167 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
1168 assert( pPage->leaf==0 || pPage->leaf==1 ); | |
1169 assert( pPage->intKeyLeaf==0 ); | |
1170 assert( pPage->noPayload==0 ); | |
1171 pIter = pCell + pPage->childPtrSize; | |
1172 nPayload = *pIter; | |
1173 if( nPayload>=0x80 ){ | |
1174 u8 *pEnd = &pIter[8]; | |
1175 nPayload &= 0x7f; | |
1176 do{ | |
1177 nPayload = (nPayload<<7) | (*++pIter & 0x7f); | |
1178 }while( *(pIter)>=0x80 && pIter<pEnd ); | |
1179 } | |
1180 pIter++; | |
1181 pInfo->nKey = nPayload; | |
1182 pInfo->nPayload = nPayload; | |
1183 pInfo->pPayload = pIter; | |
1184 testcase( nPayload==pPage->maxLocal ); | |
1185 testcase( nPayload==pPage->maxLocal+1 ); | |
1186 if( nPayload<=pPage->maxLocal ){ | |
1187 /* This is the (easy) common case where the entire payload fits | |
1188 ** on the local page. No overflow is required. | |
1189 */ | |
1190 pInfo->nSize = nPayload + (u16)(pIter - pCell); | |
1191 if( pInfo->nSize<4 ) pInfo->nSize = 4; | |
1192 pInfo->nLocal = (u16)nPayload; | |
1193 }else{ | |
1194 btreeParseCellAdjustSizeForOverflow(pPage, pCell, pInfo); | |
1195 } | |
1196 } | |
1197 static void btreeParseCell( | |
1198 MemPage *pPage, /* Page containing the cell */ | |
1199 int iCell, /* The cell index. First cell is 0 */ | |
1200 CellInfo *pInfo /* Fill in this structure */ | |
1201 ){ | |
1202 pPage->xParseCell(pPage, findCell(pPage, iCell), pInfo); | |
1203 } | |
1204 | |
1205 /* | |
1206 ** The following routines are implementations of the MemPage.xCellSize | |
1207 ** method. | |
1208 ** | |
1209 ** Compute the total number of bytes that a Cell needs in the cell | |
1210 ** data area of the btree-page. The return number includes the cell | |
1211 ** data header and the local payload, but not any overflow page or | |
1212 ** the space used by the cell pointer. | |
1213 ** | |
1214 ** cellSizePtrNoPayload() => table internal nodes | |
1215 ** cellSizePtr() => all index nodes & table leaf nodes | |
1216 */ | |
1217 static u16 cellSizePtr(MemPage *pPage, u8 *pCell){ | |
1218 u8 *pIter = pCell + pPage->childPtrSize; /* For looping over bytes of pCell */ | |
1219 u8 *pEnd; /* End mark for a varint */ | |
1220 u32 nSize; /* Size value to return */ | |
1221 | |
1222 #ifdef SQLITE_DEBUG | |
1223 /* The value returned by this function should always be the same as | |
1224 ** the (CellInfo.nSize) value found by doing a full parse of the | |
1225 ** cell. If SQLITE_DEBUG is defined, an assert() at the bottom of | |
1226 ** this function verifies that this invariant is not violated. */ | |
1227 CellInfo debuginfo; | |
1228 pPage->xParseCell(pPage, pCell, &debuginfo); | |
1229 #endif | |
1230 | |
1231 assert( pPage->noPayload==0 ); | |
1232 nSize = *pIter; | |
1233 if( nSize>=0x80 ){ | |
1234 pEnd = &pIter[8]; | |
1235 nSize &= 0x7f; | |
1236 do{ | |
1237 nSize = (nSize<<7) | (*++pIter & 0x7f); | |
1238 }while( *(pIter)>=0x80 && pIter<pEnd ); | |
1239 } | |
1240 pIter++; | |
1241 if( pPage->intKey ){ | |
1242 /* pIter now points at the 64-bit integer key value, a variable length | |
1243 ** integer. The following block moves pIter to point at the first byte | |
1244 ** past the end of the key value. */ | |
1245 pEnd = &pIter[9]; | |
1246 while( (*pIter++)&0x80 && pIter<pEnd ); | |
1247 } | |
1248 testcase( nSize==pPage->maxLocal ); | |
1249 testcase( nSize==pPage->maxLocal+1 ); | |
1250 if( nSize<=pPage->maxLocal ){ | |
1251 nSize += (u32)(pIter - pCell); | |
1252 if( nSize<4 ) nSize = 4; | |
1253 }else{ | |
1254 int minLocal = pPage->minLocal; | |
1255 nSize = minLocal + (nSize - minLocal) % (pPage->pBt->usableSize - 4); | |
1256 testcase( nSize==pPage->maxLocal ); | |
1257 testcase( nSize==pPage->maxLocal+1 ); | |
1258 if( nSize>pPage->maxLocal ){ | |
1259 nSize = minLocal; | |
1260 } | |
1261 nSize += 4 + (u16)(pIter - pCell); | |
1262 } | |
1263 assert( nSize==debuginfo.nSize || CORRUPT_DB ); | |
1264 return (u16)nSize; | |
1265 } | |
1266 static u16 cellSizePtrNoPayload(MemPage *pPage, u8 *pCell){ | |
1267 u8 *pIter = pCell + 4; /* For looping over bytes of pCell */ | |
1268 u8 *pEnd; /* End mark for a varint */ | |
1269 | |
1270 #ifdef SQLITE_DEBUG | |
1271 /* The value returned by this function should always be the same as | |
1272 ** the (CellInfo.nSize) value found by doing a full parse of the | |
1273 ** cell. If SQLITE_DEBUG is defined, an assert() at the bottom of | |
1274 ** this function verifies that this invariant is not violated. */ | |
1275 CellInfo debuginfo; | |
1276 pPage->xParseCell(pPage, pCell, &debuginfo); | |
1277 #else | |
1278 UNUSED_PARAMETER(pPage); | |
1279 #endif | |
1280 | |
1281 assert( pPage->childPtrSize==4 ); | |
1282 pEnd = pIter + 9; | |
1283 while( (*pIter++)&0x80 && pIter<pEnd ); | |
1284 assert( debuginfo.nSize==(u16)(pIter - pCell) || CORRUPT_DB ); | |
1285 return (u16)(pIter - pCell); | |
1286 } | |
1287 | |
1288 | |
1289 #ifdef SQLITE_DEBUG | |
1290 /* This variation on cellSizePtr() is used inside of assert() statements | |
1291 ** only. */ | |
1292 static u16 cellSize(MemPage *pPage, int iCell){ | |
1293 return pPage->xCellSize(pPage, findCell(pPage, iCell)); | |
1294 } | |
1295 #endif | |
1296 | |
1297 #ifndef SQLITE_OMIT_AUTOVACUUM | |
1298 /* | |
1299 ** If the cell pCell, part of page pPage contains a pointer | |
1300 ** to an overflow page, insert an entry into the pointer-map | |
1301 ** for the overflow page. | |
1302 */ | |
1303 static void ptrmapPutOvflPtr(MemPage *pPage, u8 *pCell, int *pRC){ | |
1304 CellInfo info; | |
1305 if( *pRC ) return; | |
1306 assert( pCell!=0 ); | |
1307 pPage->xParseCell(pPage, pCell, &info); | |
1308 if( info.nLocal<info.nPayload ){ | |
1309 Pgno ovfl = get4byte(&pCell[info.nSize-4]); | |
1310 ptrmapPut(pPage->pBt, ovfl, PTRMAP_OVERFLOW1, pPage->pgno, pRC); | |
1311 } | |
1312 } | |
1313 #endif | |
1314 | |
1315 | |
1316 /* | |
1317 ** Defragment the page given. All Cells are moved to the | |
1318 ** end of the page and all free space is collected into one | |
1319 ** big FreeBlk that occurs in between the header and cell | |
1320 ** pointer array and the cell content area. | |
1321 ** | |
1322 ** EVIDENCE-OF: R-44582-60138 SQLite may from time to time reorganize a | |
1323 ** b-tree page so that there are no freeblocks or fragment bytes, all | |
1324 ** unused bytes are contained in the unallocated space region, and all | |
1325 ** cells are packed tightly at the end of the page. | |
1326 */ | |
1327 static int defragmentPage(MemPage *pPage){ | |
1328 int i; /* Loop counter */ | |
1329 int pc; /* Address of the i-th cell */ | |
1330 int hdr; /* Offset to the page header */ | |
1331 int size; /* Size of a cell */ | |
1332 int usableSize; /* Number of usable bytes on a page */ | |
1333 int cellOffset; /* Offset to the cell pointer array */ | |
1334 int cbrk; /* Offset to the cell content area */ | |
1335 int nCell; /* Number of cells on the page */ | |
1336 unsigned char *data; /* The page data */ | |
1337 unsigned char *temp; /* Temp area for cell content */ | |
1338 unsigned char *src; /* Source of content */ | |
1339 int iCellFirst; /* First allowable cell index */ | |
1340 int iCellLast; /* Last possible cell index */ | |
1341 | |
1342 | |
1343 assert( sqlite3PagerIswriteable(pPage->pDbPage) ); | |
1344 assert( pPage->pBt!=0 ); | |
1345 assert( pPage->pBt->usableSize <= SQLITE_MAX_PAGE_SIZE ); | |
1346 assert( pPage->nOverflow==0 ); | |
1347 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
1348 temp = 0; | |
1349 src = data = pPage->aData; | |
1350 hdr = pPage->hdrOffset; | |
1351 cellOffset = pPage->cellOffset; | |
1352 nCell = pPage->nCell; | |
1353 assert( nCell==get2byte(&data[hdr+3]) ); | |
1354 usableSize = pPage->pBt->usableSize; | |
1355 cbrk = usableSize; | |
1356 iCellFirst = cellOffset + 2*nCell; | |
1357 iCellLast = usableSize - 4; | |
1358 for(i=0; i<nCell; i++){ | |
1359 u8 *pAddr; /* The i-th cell pointer */ | |
1360 pAddr = &data[cellOffset + i*2]; | |
1361 pc = get2byte(pAddr); | |
1362 testcase( pc==iCellFirst ); | |
1363 testcase( pc==iCellLast ); | |
1364 /* These conditions have already been verified in btreeInitPage() | |
1365 ** if PRAGMA cell_size_check=ON. | |
1366 */ | |
1367 if( pc<iCellFirst || pc>iCellLast ){ | |
1368 return SQLITE_CORRUPT_BKPT; | |
1369 } | |
1370 assert( pc>=iCellFirst && pc<=iCellLast ); | |
1371 size = pPage->xCellSize(pPage, &src[pc]); | |
1372 cbrk -= size; | |
1373 if( cbrk<iCellFirst || pc+size>usableSize ){ | |
1374 return SQLITE_CORRUPT_BKPT; | |
1375 } | |
1376 assert( cbrk+size<=usableSize && cbrk>=iCellFirst ); | |
1377 testcase( cbrk+size==usableSize ); | |
1378 testcase( pc+size==usableSize ); | |
1379 put2byte(pAddr, cbrk); | |
1380 if( temp==0 ){ | |
1381 int x; | |
1382 if( cbrk==pc ) continue; | |
1383 temp = sqlite3PagerTempSpace(pPage->pBt->pPager); | |
1384 x = get2byte(&data[hdr+5]); | |
1385 memcpy(&temp[x], &data[x], (cbrk+size) - x); | |
1386 src = temp; | |
1387 } | |
1388 memcpy(&data[cbrk], &src[pc], size); | |
1389 } | |
1390 assert( cbrk>=iCellFirst ); | |
1391 put2byte(&data[hdr+5], cbrk); | |
1392 data[hdr+1] = 0; | |
1393 data[hdr+2] = 0; | |
1394 data[hdr+7] = 0; | |
1395 memset(&data[iCellFirst], 0, cbrk-iCellFirst); | |
1396 assert( sqlite3PagerIswriteable(pPage->pDbPage) ); | |
1397 if( cbrk-iCellFirst!=pPage->nFree ){ | |
1398 return SQLITE_CORRUPT_BKPT; | |
1399 } | |
1400 return SQLITE_OK; | |
1401 } | |
1402 | |
1403 /* | |
1404 ** Search the free-list on page pPg for space to store a cell nByte bytes in | |
1405 ** size. If one can be found, return a pointer to the space and remove it | |
1406 ** from the free-list. | |
1407 ** | |
1408 ** If no suitable space can be found on the free-list, return NULL. | |
1409 ** | |
1410 ** This function may detect corruption within pPg. If corruption is | |
1411 ** detected then *pRc is set to SQLITE_CORRUPT and NULL is returned. | |
1412 ** | |
1413 ** Slots on the free list that are between 1 and 3 bytes larger than nByte | |
1414 ** will be ignored if adding the extra space to the fragmentation count | |
1415 ** causes the fragmentation count to exceed 60. | |
1416 */ | |
1417 static u8 *pageFindSlot(MemPage *pPg, int nByte, int *pRc){ | |
1418 const int hdr = pPg->hdrOffset; | |
1419 u8 * const aData = pPg->aData; | |
1420 int iAddr = hdr + 1; | |
1421 int pc = get2byte(&aData[iAddr]); | |
1422 int x; | |
1423 int usableSize = pPg->pBt->usableSize; | |
1424 | |
1425 assert( pc>0 ); | |
1426 do{ | |
1427 int size; /* Size of the free slot */ | |
1428 /* EVIDENCE-OF: R-06866-39125 Freeblocks are always connected in order of | |
1429 ** increasing offset. */ | |
1430 if( pc>usableSize-4 || pc<iAddr+4 ){ | |
1431 *pRc = SQLITE_CORRUPT_BKPT; | |
1432 return 0; | |
1433 } | |
1434 /* EVIDENCE-OF: R-22710-53328 The third and fourth bytes of each | |
1435 ** freeblock form a big-endian integer which is the size of the freeblock | |
1436 ** in bytes, including the 4-byte header. */ | |
1437 size = get2byte(&aData[pc+2]); | |
1438 if( (x = size - nByte)>=0 ){ | |
1439 testcase( x==4 ); | |
1440 testcase( x==3 ); | |
1441 if( pc < pPg->cellOffset+2*pPg->nCell || size+pc > usableSize ){ | |
1442 *pRc = SQLITE_CORRUPT_BKPT; | |
1443 return 0; | |
1444 }else if( x<4 ){ | |
1445 /* EVIDENCE-OF: R-11498-58022 In a well-formed b-tree page, the total | |
1446 ** number of bytes in fragments may not exceed 60. */ | |
1447 if( aData[hdr+7]>57 ) return 0; | |
1448 | |
1449 /* Remove the slot from the free-list. Update the number of | |
1450 ** fragmented bytes within the page. */ | |
1451 memcpy(&aData[iAddr], &aData[pc], 2); | |
1452 aData[hdr+7] += (u8)x; | |
1453 }else{ | |
1454 /* The slot remains on the free-list. Reduce its size to account | |
1455 ** for the portion used by the new allocation. */ | |
1456 put2byte(&aData[pc+2], x); | |
1457 } | |
1458 return &aData[pc + x]; | |
1459 } | |
1460 iAddr = pc; | |
1461 pc = get2byte(&aData[pc]); | |
1462 }while( pc ); | |
1463 | |
1464 return 0; | |
1465 } | |
1466 | |
1467 /* | |
1468 ** Allocate nByte bytes of space from within the B-Tree page passed | |
1469 ** as the first argument. Write into *pIdx the index into pPage->aData[] | |
1470 ** of the first byte of allocated space. Return either SQLITE_OK or | |
1471 ** an error code (usually SQLITE_CORRUPT). | |
1472 ** | |
1473 ** The caller guarantees that there is sufficient space to make the | |
1474 ** allocation. This routine might need to defragment in order to bring | |
1475 ** all the space together, however. This routine will avoid using | |
1476 ** the first two bytes past the cell pointer area since presumably this | |
1477 ** allocation is being made in order to insert a new cell, so we will | |
1478 ** also end up needing a new cell pointer. | |
1479 */ | |
1480 static int allocateSpace(MemPage *pPage, int nByte, int *pIdx){ | |
1481 const int hdr = pPage->hdrOffset; /* Local cache of pPage->hdrOffset */ | |
1482 u8 * const data = pPage->aData; /* Local cache of pPage->aData */ | |
1483 int top; /* First byte of cell content area */ | |
1484 int rc = SQLITE_OK; /* Integer return code */ | |
1485 int gap; /* First byte of gap between cell pointers and cell content */ | |
1486 | |
1487 assert( sqlite3PagerIswriteable(pPage->pDbPage) ); | |
1488 assert( pPage->pBt ); | |
1489 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
1490 assert( nByte>=0 ); /* Minimum cell size is 4 */ | |
1491 assert( pPage->nFree>=nByte ); | |
1492 assert( pPage->nOverflow==0 ); | |
1493 assert( nByte < (int)(pPage->pBt->usableSize-8) ); | |
1494 | |
1495 assert( pPage->cellOffset == hdr + 12 - 4*pPage->leaf ); | |
1496 gap = pPage->cellOffset + 2*pPage->nCell; | |
1497 assert( gap<=65536 ); | |
1498 /* EVIDENCE-OF: R-29356-02391 If the database uses a 65536-byte page size | |
1499 ** and the reserved space is zero (the usual value for reserved space) | |
1500 ** then the cell content offset of an empty page wants to be 65536. | |
1501 ** However, that integer is too large to be stored in a 2-byte unsigned | |
1502 ** integer, so a value of 0 is used in its place. */ | |
1503 top = get2byte(&data[hdr+5]); | |
1504 assert( top<=(int)pPage->pBt->usableSize ); /* Prevent by getAndInitPage() */ | |
1505 if( gap>top ){ | |
1506 if( top==0 && pPage->pBt->usableSize==65536 ){ | |
1507 top = 65536; | |
1508 }else{ | |
1509 return SQLITE_CORRUPT_BKPT; | |
1510 } | |
1511 } | |
1512 | |
1513 /* If there is enough space between gap and top for one more cell pointer | |
1514 ** array entry offset, and if the freelist is not empty, then search the | |
1515 ** freelist looking for a free slot big enough to satisfy the request. | |
1516 */ | |
1517 testcase( gap+2==top ); | |
1518 testcase( gap+1==top ); | |
1519 testcase( gap==top ); | |
1520 if( (data[hdr+2] || data[hdr+1]) && gap+2<=top ){ | |
1521 u8 *pSpace = pageFindSlot(pPage, nByte, &rc); | |
1522 if( pSpace ){ | |
1523 assert( pSpace>=data && (pSpace - data)<65536 ); | |
1524 *pIdx = (int)(pSpace - data); | |
1525 return SQLITE_OK; | |
1526 }else if( rc ){ | |
1527 return rc; | |
1528 } | |
1529 } | |
1530 | |
1531 /* The request could not be fulfilled using a freelist slot. Check | |
1532 ** to see if defragmentation is necessary. | |
1533 */ | |
1534 testcase( gap+2+nByte==top ); | |
1535 if( gap+2+nByte>top ){ | |
1536 assert( pPage->nCell>0 || CORRUPT_DB ); | |
1537 rc = defragmentPage(pPage); | |
1538 if( rc ) return rc; | |
1539 top = get2byteNotZero(&data[hdr+5]); | |
1540 assert( gap+nByte<=top ); | |
1541 } | |
1542 | |
1543 | |
1544 /* Allocate memory from the gap in between the cell pointer array | |
1545 ** and the cell content area. The btreeInitPage() call has already | |
1546 ** validated the freelist. Given that the freelist is valid, there | |
1547 ** is no way that the allocation can extend off the end of the page. | |
1548 ** The assert() below verifies the previous sentence. | |
1549 */ | |
1550 top -= nByte; | |
1551 put2byte(&data[hdr+5], top); | |
1552 assert( top+nByte <= (int)pPage->pBt->usableSize ); | |
1553 *pIdx = top; | |
1554 return SQLITE_OK; | |
1555 } | |
1556 | |
1557 /* | |
1558 ** Return a section of the pPage->aData to the freelist. | |
1559 ** The first byte of the new free block is pPage->aData[iStart] | |
1560 ** and the size of the block is iSize bytes. | |
1561 ** | |
1562 ** Adjacent freeblocks are coalesced. | |
1563 ** | |
1564 ** Note that even though the freeblock list was checked by btreeInitPage(), | |
1565 ** that routine will not detect overlap between cells or freeblocks. Nor | |
1566 ** does it detect cells or freeblocks that encrouch into the reserved bytes | |
1567 ** at the end of the page. So do additional corruption checks inside this | |
1568 ** routine and return SQLITE_CORRUPT if any problems are found. | |
1569 */ | |
1570 static int freeSpace(MemPage *pPage, u16 iStart, u16 iSize){ | |
1571 u16 iPtr; /* Address of ptr to next freeblock */ | |
1572 u16 iFreeBlk; /* Address of the next freeblock */ | |
1573 u8 hdr; /* Page header size. 0 or 100 */ | |
1574 u8 nFrag = 0; /* Reduction in fragmentation */ | |
1575 u16 iOrigSize = iSize; /* Original value of iSize */ | |
1576 u32 iLast = pPage->pBt->usableSize-4; /* Largest possible freeblock offset */ | |
1577 u32 iEnd = iStart + iSize; /* First byte past the iStart buffer */ | |
1578 unsigned char *data = pPage->aData; /* Page content */ | |
1579 | |
1580 assert( pPage->pBt!=0 ); | |
1581 assert( sqlite3PagerIswriteable(pPage->pDbPage) ); | |
1582 assert( CORRUPT_DB || iStart>=pPage->hdrOffset+6+pPage->childPtrSize ); | |
1583 assert( CORRUPT_DB || iEnd <= pPage->pBt->usableSize ); | |
1584 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
1585 assert( iSize>=4 ); /* Minimum cell size is 4 */ | |
1586 assert( iStart<=iLast ); | |
1587 | |
1588 /* Overwrite deleted information with zeros when the secure_delete | |
1589 ** option is enabled */ | |
1590 if( pPage->pBt->btsFlags & BTS_SECURE_DELETE ){ | |
1591 memset(&data[iStart], 0, iSize); | |
1592 } | |
1593 | |
1594 /* The list of freeblocks must be in ascending order. Find the | |
1595 ** spot on the list where iStart should be inserted. | |
1596 */ | |
1597 hdr = pPage->hdrOffset; | |
1598 iPtr = hdr + 1; | |
1599 if( data[iPtr+1]==0 && data[iPtr]==0 ){ | |
1600 iFreeBlk = 0; /* Shortcut for the case when the freelist is empty */ | |
1601 }else{ | |
1602 while( (iFreeBlk = get2byte(&data[iPtr]))>0 && iFreeBlk<iStart ){ | |
1603 if( iFreeBlk<iPtr+4 ) return SQLITE_CORRUPT_BKPT; | |
1604 iPtr = iFreeBlk; | |
1605 } | |
1606 if( iFreeBlk>iLast ) return SQLITE_CORRUPT_BKPT; | |
1607 assert( iFreeBlk>iPtr || iFreeBlk==0 ); | |
1608 | |
1609 /* At this point: | |
1610 ** iFreeBlk: First freeblock after iStart, or zero if none | |
1611 ** iPtr: The address of a pointer to iFreeBlk | |
1612 ** | |
1613 ** Check to see if iFreeBlk should be coalesced onto the end of iStart. | |
1614 */ | |
1615 if( iFreeBlk && iEnd+3>=iFreeBlk ){ | |
1616 nFrag = iFreeBlk - iEnd; | |
1617 if( iEnd>iFreeBlk ) return SQLITE_CORRUPT_BKPT; | |
1618 iEnd = iFreeBlk + get2byte(&data[iFreeBlk+2]); | |
1619 if( iEnd > pPage->pBt->usableSize ) return SQLITE_CORRUPT_BKPT; | |
1620 iSize = iEnd - iStart; | |
1621 iFreeBlk = get2byte(&data[iFreeBlk]); | |
1622 } | |
1623 | |
1624 /* If iPtr is another freeblock (that is, if iPtr is not the freelist | |
1625 ** pointer in the page header) then check to see if iStart should be | |
1626 ** coalesced onto the end of iPtr. | |
1627 */ | |
1628 if( iPtr>hdr+1 ){ | |
1629 int iPtrEnd = iPtr + get2byte(&data[iPtr+2]); | |
1630 if( iPtrEnd+3>=iStart ){ | |
1631 if( iPtrEnd>iStart ) return SQLITE_CORRUPT_BKPT; | |
1632 nFrag += iStart - iPtrEnd; | |
1633 iSize = iEnd - iPtr; | |
1634 iStart = iPtr; | |
1635 } | |
1636 } | |
1637 if( nFrag>data[hdr+7] ) return SQLITE_CORRUPT_BKPT; | |
1638 data[hdr+7] -= nFrag; | |
1639 } | |
1640 if( iStart==get2byte(&data[hdr+5]) ){ | |
1641 /* The new freeblock is at the beginning of the cell content area, | |
1642 ** so just extend the cell content area rather than create another | |
1643 ** freelist entry */ | |
1644 if( iPtr!=hdr+1 ) return SQLITE_CORRUPT_BKPT; | |
1645 put2byte(&data[hdr+1], iFreeBlk); | |
1646 put2byte(&data[hdr+5], iEnd); | |
1647 }else{ | |
1648 /* Insert the new freeblock into the freelist */ | |
1649 put2byte(&data[iPtr], iStart); | |
1650 put2byte(&data[iStart], iFreeBlk); | |
1651 put2byte(&data[iStart+2], iSize); | |
1652 } | |
1653 pPage->nFree += iOrigSize; | |
1654 return SQLITE_OK; | |
1655 } | |
1656 | |
1657 /* | |
1658 ** Decode the flags byte (the first byte of the header) for a page | |
1659 ** and initialize fields of the MemPage structure accordingly. | |
1660 ** | |
1661 ** Only the following combinations are supported. Anything different | |
1662 ** indicates a corrupt database files: | |
1663 ** | |
1664 ** PTF_ZERODATA | |
1665 ** PTF_ZERODATA | PTF_LEAF | |
1666 ** PTF_LEAFDATA | PTF_INTKEY | |
1667 ** PTF_LEAFDATA | PTF_INTKEY | PTF_LEAF | |
1668 */ | |
1669 static int decodeFlags(MemPage *pPage, int flagByte){ | |
1670 BtShared *pBt; /* A copy of pPage->pBt */ | |
1671 | |
1672 assert( pPage->hdrOffset==(pPage->pgno==1 ? 100 : 0) ); | |
1673 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
1674 pPage->leaf = (u8)(flagByte>>3); assert( PTF_LEAF == 1<<3 ); | |
1675 flagByte &= ~PTF_LEAF; | |
1676 pPage->childPtrSize = 4-4*pPage->leaf; | |
1677 pPage->xCellSize = cellSizePtr; | |
1678 pBt = pPage->pBt; | |
1679 if( flagByte==(PTF_LEAFDATA | PTF_INTKEY) ){ | |
1680 /* EVIDENCE-OF: R-03640-13415 A value of 5 means the page is an interior | |
1681 ** table b-tree page. */ | |
1682 assert( (PTF_LEAFDATA|PTF_INTKEY)==5 ); | |
1683 /* EVIDENCE-OF: R-20501-61796 A value of 13 means the page is a leaf | |
1684 ** table b-tree page. */ | |
1685 assert( (PTF_LEAFDATA|PTF_INTKEY|PTF_LEAF)==13 ); | |
1686 pPage->intKey = 1; | |
1687 if( pPage->leaf ){ | |
1688 pPage->intKeyLeaf = 1; | |
1689 pPage->noPayload = 0; | |
1690 pPage->xParseCell = btreeParseCellPtr; | |
1691 }else{ | |
1692 pPage->intKeyLeaf = 0; | |
1693 pPage->noPayload = 1; | |
1694 pPage->xCellSize = cellSizePtrNoPayload; | |
1695 pPage->xParseCell = btreeParseCellPtrNoPayload; | |
1696 } | |
1697 pPage->maxLocal = pBt->maxLeaf; | |
1698 pPage->minLocal = pBt->minLeaf; | |
1699 }else if( flagByte==PTF_ZERODATA ){ | |
1700 /* EVIDENCE-OF: R-27225-53936 A value of 2 means the page is an interior | |
1701 ** index b-tree page. */ | |
1702 assert( (PTF_ZERODATA)==2 ); | |
1703 /* EVIDENCE-OF: R-16571-11615 A value of 10 means the page is a leaf | |
1704 ** index b-tree page. */ | |
1705 assert( (PTF_ZERODATA|PTF_LEAF)==10 ); | |
1706 pPage->intKey = 0; | |
1707 pPage->intKeyLeaf = 0; | |
1708 pPage->noPayload = 0; | |
1709 pPage->xParseCell = btreeParseCellPtrIndex; | |
1710 pPage->maxLocal = pBt->maxLocal; | |
1711 pPage->minLocal = pBt->minLocal; | |
1712 }else{ | |
1713 /* EVIDENCE-OF: R-47608-56469 Any other value for the b-tree page type is | |
1714 ** an error. */ | |
1715 return SQLITE_CORRUPT_BKPT; | |
1716 } | |
1717 pPage->max1bytePayload = pBt->max1bytePayload; | |
1718 return SQLITE_OK; | |
1719 } | |
1720 | |
1721 /* | |
1722 ** Initialize the auxiliary information for a disk block. | |
1723 ** | |
1724 ** Return SQLITE_OK on success. If we see that the page does | |
1725 ** not contain a well-formed database page, then return | |
1726 ** SQLITE_CORRUPT. Note that a return of SQLITE_OK does not | |
1727 ** guarantee that the page is well-formed. It only shows that | |
1728 ** we failed to detect any corruption. | |
1729 */ | |
1730 static int btreeInitPage(MemPage *pPage){ | |
1731 | |
1732 assert( pPage->pBt!=0 ); | |
1733 assert( pPage->pBt->db!=0 ); | |
1734 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
1735 assert( pPage->pgno==sqlite3PagerPagenumber(pPage->pDbPage) ); | |
1736 assert( pPage == sqlite3PagerGetExtra(pPage->pDbPage) ); | |
1737 assert( pPage->aData == sqlite3PagerGetData(pPage->pDbPage) ); | |
1738 | |
1739 if( !pPage->isInit ){ | |
1740 u16 pc; /* Address of a freeblock within pPage->aData[] */ | |
1741 u8 hdr; /* Offset to beginning of page header */ | |
1742 u8 *data; /* Equal to pPage->aData */ | |
1743 BtShared *pBt; /* The main btree structure */ | |
1744 int usableSize; /* Amount of usable space on each page */ | |
1745 u16 cellOffset; /* Offset from start of page to first cell pointer */ | |
1746 int nFree; /* Number of unused bytes on the page */ | |
1747 int top; /* First byte of the cell content area */ | |
1748 int iCellFirst; /* First allowable cell or freeblock offset */ | |
1749 int iCellLast; /* Last possible cell or freeblock offset */ | |
1750 | |
1751 pBt = pPage->pBt; | |
1752 | |
1753 hdr = pPage->hdrOffset; | |
1754 data = pPage->aData; | |
1755 /* EVIDENCE-OF: R-28594-02890 The one-byte flag at offset 0 indicating | |
1756 ** the b-tree page type. */ | |
1757 if( decodeFlags(pPage, data[hdr]) ) return SQLITE_CORRUPT_BKPT; | |
1758 assert( pBt->pageSize>=512 && pBt->pageSize<=65536 ); | |
1759 pPage->maskPage = (u16)(pBt->pageSize - 1); | |
1760 pPage->nOverflow = 0; | |
1761 usableSize = pBt->usableSize; | |
1762 pPage->cellOffset = cellOffset = hdr + 8 + pPage->childPtrSize; | |
1763 pPage->aDataEnd = &data[usableSize]; | |
1764 pPage->aCellIdx = &data[cellOffset]; | |
1765 pPage->aDataOfst = &data[pPage->childPtrSize]; | |
1766 /* EVIDENCE-OF: R-58015-48175 The two-byte integer at offset 5 designates | |
1767 ** the start of the cell content area. A zero value for this integer is | |
1768 ** interpreted as 65536. */ | |
1769 top = get2byteNotZero(&data[hdr+5]); | |
1770 /* EVIDENCE-OF: R-37002-32774 The two-byte integer at offset 3 gives the | |
1771 ** number of cells on the page. */ | |
1772 pPage->nCell = get2byte(&data[hdr+3]); | |
1773 if( pPage->nCell>MX_CELL(pBt) ){ | |
1774 /* To many cells for a single page. The page must be corrupt */ | |
1775 return SQLITE_CORRUPT_BKPT; | |
1776 } | |
1777 testcase( pPage->nCell==MX_CELL(pBt) ); | |
1778 /* EVIDENCE-OF: R-24089-57979 If a page contains no cells (which is only | |
1779 ** possible for a root page of a table that contains no rows) then the | |
1780 ** offset to the cell content area will equal the page size minus the | |
1781 ** bytes of reserved space. */ | |
1782 assert( pPage->nCell>0 || top==usableSize || CORRUPT_DB ); | |
1783 | |
1784 /* A malformed database page might cause us to read past the end | |
1785 ** of page when parsing a cell. | |
1786 ** | |
1787 ** The following block of code checks early to see if a cell extends | |
1788 ** past the end of a page boundary and causes SQLITE_CORRUPT to be | |
1789 ** returned if it does. | |
1790 */ | |
1791 iCellFirst = cellOffset + 2*pPage->nCell; | |
1792 iCellLast = usableSize - 4; | |
1793 if( pBt->db->flags & SQLITE_CellSizeCk ){ | |
1794 int i; /* Index into the cell pointer array */ | |
1795 int sz; /* Size of a cell */ | |
1796 | |
1797 if( !pPage->leaf ) iCellLast--; | |
1798 for(i=0; i<pPage->nCell; i++){ | |
1799 pc = get2byteAligned(&data[cellOffset+i*2]); | |
1800 testcase( pc==iCellFirst ); | |
1801 testcase( pc==iCellLast ); | |
1802 if( pc<iCellFirst || pc>iCellLast ){ | |
1803 return SQLITE_CORRUPT_BKPT; | |
1804 } | |
1805 sz = pPage->xCellSize(pPage, &data[pc]); | |
1806 testcase( pc+sz==usableSize ); | |
1807 if( pc+sz>usableSize ){ | |
1808 return SQLITE_CORRUPT_BKPT; | |
1809 } | |
1810 } | |
1811 if( !pPage->leaf ) iCellLast++; | |
1812 } | |
1813 | |
1814 /* Compute the total free space on the page | |
1815 ** EVIDENCE-OF: R-23588-34450 The two-byte integer at offset 1 gives the | |
1816 ** start of the first freeblock on the page, or is zero if there are no | |
1817 ** freeblocks. */ | |
1818 pc = get2byte(&data[hdr+1]); | |
1819 nFree = data[hdr+7] + top; /* Init nFree to non-freeblock free space */ | |
1820 while( pc>0 ){ | |
1821 u16 next, size; | |
1822 if( pc<iCellFirst || pc>iCellLast ){ | |
1823 /* EVIDENCE-OF: R-55530-52930 In a well-formed b-tree page, there will | |
1824 ** always be at least one cell before the first freeblock. | |
1825 ** | |
1826 ** Or, the freeblock is off the end of the page | |
1827 */ | |
1828 return SQLITE_CORRUPT_BKPT; | |
1829 } | |
1830 next = get2byte(&data[pc]); | |
1831 size = get2byte(&data[pc+2]); | |
1832 if( (next>0 && next<=pc+size+3) || pc+size>usableSize ){ | |
1833 /* Free blocks must be in ascending order. And the last byte of | |
1834 ** the free-block must lie on the database page. */ | |
1835 return SQLITE_CORRUPT_BKPT; | |
1836 } | |
1837 nFree = nFree + size; | |
1838 pc = next; | |
1839 } | |
1840 | |
1841 /* At this point, nFree contains the sum of the offset to the start | |
1842 ** of the cell-content area plus the number of free bytes within | |
1843 ** the cell-content area. If this is greater than the usable-size | |
1844 ** of the page, then the page must be corrupted. This check also | |
1845 ** serves to verify that the offset to the start of the cell-content | |
1846 ** area, according to the page header, lies within the page. | |
1847 */ | |
1848 if( nFree>usableSize ){ | |
1849 return SQLITE_CORRUPT_BKPT; | |
1850 } | |
1851 pPage->nFree = (u16)(nFree - iCellFirst); | |
1852 pPage->isInit = 1; | |
1853 } | |
1854 return SQLITE_OK; | |
1855 } | |
1856 | |
1857 /* | |
1858 ** Set up a raw page so that it looks like a database page holding | |
1859 ** no entries. | |
1860 */ | |
1861 static void zeroPage(MemPage *pPage, int flags){ | |
1862 unsigned char *data = pPage->aData; | |
1863 BtShared *pBt = pPage->pBt; | |
1864 u8 hdr = pPage->hdrOffset; | |
1865 u16 first; | |
1866 | |
1867 assert( sqlite3PagerPagenumber(pPage->pDbPage)==pPage->pgno ); | |
1868 assert( sqlite3PagerGetExtra(pPage->pDbPage) == (void*)pPage ); | |
1869 assert( sqlite3PagerGetData(pPage->pDbPage) == data ); | |
1870 assert( sqlite3PagerIswriteable(pPage->pDbPage) ); | |
1871 assert( sqlite3_mutex_held(pBt->mutex) ); | |
1872 if( pBt->btsFlags & BTS_SECURE_DELETE ){ | |
1873 memset(&data[hdr], 0, pBt->usableSize - hdr); | |
1874 } | |
1875 data[hdr] = (char)flags; | |
1876 first = hdr + ((flags&PTF_LEAF)==0 ? 12 : 8); | |
1877 memset(&data[hdr+1], 0, 4); | |
1878 data[hdr+7] = 0; | |
1879 put2byte(&data[hdr+5], pBt->usableSize); | |
1880 pPage->nFree = (u16)(pBt->usableSize - first); | |
1881 decodeFlags(pPage, flags); | |
1882 pPage->cellOffset = first; | |
1883 pPage->aDataEnd = &data[pBt->usableSize]; | |
1884 pPage->aCellIdx = &data[first]; | |
1885 pPage->aDataOfst = &data[pPage->childPtrSize]; | |
1886 pPage->nOverflow = 0; | |
1887 assert( pBt->pageSize>=512 && pBt->pageSize<=65536 ); | |
1888 pPage->maskPage = (u16)(pBt->pageSize - 1); | |
1889 pPage->nCell = 0; | |
1890 pPage->isInit = 1; | |
1891 } | |
1892 | |
1893 | |
1894 /* | |
1895 ** Convert a DbPage obtained from the pager into a MemPage used by | |
1896 ** the btree layer. | |
1897 */ | |
1898 static MemPage *btreePageFromDbPage(DbPage *pDbPage, Pgno pgno, BtShared *pBt){ | |
1899 MemPage *pPage = (MemPage*)sqlite3PagerGetExtra(pDbPage); | |
1900 if( pgno!=pPage->pgno ){ | |
1901 pPage->aData = sqlite3PagerGetData(pDbPage); | |
1902 pPage->pDbPage = pDbPage; | |
1903 pPage->pBt = pBt; | |
1904 pPage->pgno = pgno; | |
1905 pPage->hdrOffset = pgno==1 ? 100 : 0; | |
1906 } | |
1907 assert( pPage->aData==sqlite3PagerGetData(pDbPage) ); | |
1908 return pPage; | |
1909 } | |
1910 | |
1911 /* | |
1912 ** Get a page from the pager. Initialize the MemPage.pBt and | |
1913 ** MemPage.aData elements if needed. See also: btreeGetUnusedPage(). | |
1914 ** | |
1915 ** If the PAGER_GET_NOCONTENT flag is set, it means that we do not care | |
1916 ** about the content of the page at this time. So do not go to the disk | |
1917 ** to fetch the content. Just fill in the content with zeros for now. | |
1918 ** If in the future we call sqlite3PagerWrite() on this page, that | |
1919 ** means we have started to be concerned about content and the disk | |
1920 ** read should occur at that point. | |
1921 */ | |
1922 static int btreeGetPage( | |
1923 BtShared *pBt, /* The btree */ | |
1924 Pgno pgno, /* Number of the page to fetch */ | |
1925 MemPage **ppPage, /* Return the page in this parameter */ | |
1926 int flags /* PAGER_GET_NOCONTENT or PAGER_GET_READONLY */ | |
1927 ){ | |
1928 int rc; | |
1929 DbPage *pDbPage; | |
1930 | |
1931 assert( flags==0 || flags==PAGER_GET_NOCONTENT || flags==PAGER_GET_READONLY ); | |
1932 assert( sqlite3_mutex_held(pBt->mutex) ); | |
1933 rc = sqlite3PagerGet(pBt->pPager, pgno, (DbPage**)&pDbPage, flags); | |
1934 if( rc ) return rc; | |
1935 *ppPage = btreePageFromDbPage(pDbPage, pgno, pBt); | |
1936 return SQLITE_OK; | |
1937 } | |
1938 | |
1939 /* | |
1940 ** Retrieve a page from the pager cache. If the requested page is not | |
1941 ** already in the pager cache return NULL. Initialize the MemPage.pBt and | |
1942 ** MemPage.aData elements if needed. | |
1943 */ | |
1944 static MemPage *btreePageLookup(BtShared *pBt, Pgno pgno){ | |
1945 DbPage *pDbPage; | |
1946 assert( sqlite3_mutex_held(pBt->mutex) ); | |
1947 pDbPage = sqlite3PagerLookup(pBt->pPager, pgno); | |
1948 if( pDbPage ){ | |
1949 return btreePageFromDbPage(pDbPage, pgno, pBt); | |
1950 } | |
1951 return 0; | |
1952 } | |
1953 | |
1954 /* | |
1955 ** Return the size of the database file in pages. If there is any kind of | |
1956 ** error, return ((unsigned int)-1). | |
1957 */ | |
1958 static Pgno btreePagecount(BtShared *pBt){ | |
1959 return pBt->nPage; | |
1960 } | |
1961 u32 sqlite3BtreeLastPage(Btree *p){ | |
1962 assert( sqlite3BtreeHoldsMutex(p) ); | |
1963 assert( ((p->pBt->nPage)&0x8000000)==0 ); | |
1964 return btreePagecount(p->pBt); | |
1965 } | |
1966 | |
1967 /* | |
1968 ** Get a page from the pager and initialize it. | |
1969 ** | |
1970 ** If pCur!=0 then the page is being fetched as part of a moveToChild() | |
1971 ** call. Do additional sanity checking on the page in this case. | |
1972 ** And if the fetch fails, this routine must decrement pCur->iPage. | |
1973 ** | |
1974 ** The page is fetched as read-write unless pCur is not NULL and is | |
1975 ** a read-only cursor. | |
1976 ** | |
1977 ** If an error occurs, then *ppPage is undefined. It | |
1978 ** may remain unchanged, or it may be set to an invalid value. | |
1979 */ | |
1980 static int getAndInitPage( | |
1981 BtShared *pBt, /* The database file */ | |
1982 Pgno pgno, /* Number of the page to get */ | |
1983 MemPage **ppPage, /* Write the page pointer here */ | |
1984 BtCursor *pCur, /* Cursor to receive the page, or NULL */ | |
1985 int bReadOnly /* True for a read-only page */ | |
1986 ){ | |
1987 int rc; | |
1988 DbPage *pDbPage; | |
1989 assert( sqlite3_mutex_held(pBt->mutex) ); | |
1990 assert( pCur==0 || ppPage==&pCur->apPage[pCur->iPage] ); | |
1991 assert( pCur==0 || bReadOnly==pCur->curPagerFlags ); | |
1992 assert( pCur==0 || pCur->iPage>0 ); | |
1993 | |
1994 if( pgno>btreePagecount(pBt) ){ | |
1995 rc = SQLITE_CORRUPT_BKPT; | |
1996 goto getAndInitPage_error; | |
1997 } | |
1998 rc = sqlite3PagerGet(pBt->pPager, pgno, (DbPage**)&pDbPage, bReadOnly); | |
1999 if( rc ){ | |
2000 goto getAndInitPage_error; | |
2001 } | |
2002 *ppPage = (MemPage*)sqlite3PagerGetExtra(pDbPage); | |
2003 if( (*ppPage)->isInit==0 ){ | |
2004 btreePageFromDbPage(pDbPage, pgno, pBt); | |
2005 rc = btreeInitPage(*ppPage); | |
2006 if( rc!=SQLITE_OK ){ | |
2007 releasePage(*ppPage); | |
2008 goto getAndInitPage_error; | |
2009 } | |
2010 } | |
2011 assert( (*ppPage)->pgno==pgno ); | |
2012 assert( (*ppPage)->aData==sqlite3PagerGetData(pDbPage) ); | |
2013 | |
2014 /* If obtaining a child page for a cursor, we must verify that the page is | |
2015 ** compatible with the root page. */ | |
2016 if( pCur && ((*ppPage)->nCell<1 || (*ppPage)->intKey!=pCur->curIntKey) ){ | |
2017 rc = SQLITE_CORRUPT_BKPT; | |
2018 releasePage(*ppPage); | |
2019 goto getAndInitPage_error; | |
2020 } | |
2021 return SQLITE_OK; | |
2022 | |
2023 getAndInitPage_error: | |
2024 if( pCur ) pCur->iPage--; | |
2025 testcase( pgno==0 ); | |
2026 assert( pgno!=0 || rc==SQLITE_CORRUPT ); | |
2027 return rc; | |
2028 } | |
2029 | |
2030 /* | |
2031 ** Release a MemPage. This should be called once for each prior | |
2032 ** call to btreeGetPage. | |
2033 */ | |
2034 static void releasePageNotNull(MemPage *pPage){ | |
2035 assert( pPage->aData ); | |
2036 assert( pPage->pBt ); | |
2037 assert( pPage->pDbPage!=0 ); | |
2038 assert( sqlite3PagerGetExtra(pPage->pDbPage) == (void*)pPage ); | |
2039 assert( sqlite3PagerGetData(pPage->pDbPage)==pPage->aData ); | |
2040 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
2041 sqlite3PagerUnrefNotNull(pPage->pDbPage); | |
2042 } | |
2043 static void releasePage(MemPage *pPage){ | |
2044 if( pPage ) releasePageNotNull(pPage); | |
2045 } | |
2046 | |
2047 /* | |
2048 ** Get an unused page. | |
2049 ** | |
2050 ** This works just like btreeGetPage() with the addition: | |
2051 ** | |
2052 ** * If the page is already in use for some other purpose, immediately | |
2053 ** release it and return an SQLITE_CURRUPT error. | |
2054 ** * Make sure the isInit flag is clear | |
2055 */ | |
2056 static int btreeGetUnusedPage( | |
2057 BtShared *pBt, /* The btree */ | |
2058 Pgno pgno, /* Number of the page to fetch */ | |
2059 MemPage **ppPage, /* Return the page in this parameter */ | |
2060 int flags /* PAGER_GET_NOCONTENT or PAGER_GET_READONLY */ | |
2061 ){ | |
2062 int rc = btreeGetPage(pBt, pgno, ppPage, flags); | |
2063 if( rc==SQLITE_OK ){ | |
2064 if( sqlite3PagerPageRefcount((*ppPage)->pDbPage)>1 ){ | |
2065 releasePage(*ppPage); | |
2066 *ppPage = 0; | |
2067 return SQLITE_CORRUPT_BKPT; | |
2068 } | |
2069 (*ppPage)->isInit = 0; | |
2070 }else{ | |
2071 *ppPage = 0; | |
2072 } | |
2073 return rc; | |
2074 } | |
2075 | |
2076 | |
2077 /* | |
2078 ** During a rollback, when the pager reloads information into the cache | |
2079 ** so that the cache is restored to its original state at the start of | |
2080 ** the transaction, for each page restored this routine is called. | |
2081 ** | |
2082 ** This routine needs to reset the extra data section at the end of the | |
2083 ** page to agree with the restored data. | |
2084 */ | |
2085 static void pageReinit(DbPage *pData){ | |
2086 MemPage *pPage; | |
2087 pPage = (MemPage *)sqlite3PagerGetExtra(pData); | |
2088 assert( sqlite3PagerPageRefcount(pData)>0 ); | |
2089 if( pPage->isInit ){ | |
2090 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
2091 pPage->isInit = 0; | |
2092 if( sqlite3PagerPageRefcount(pData)>1 ){ | |
2093 /* pPage might not be a btree page; it might be an overflow page | |
2094 ** or ptrmap page or a free page. In those cases, the following | |
2095 ** call to btreeInitPage() will likely return SQLITE_CORRUPT. | |
2096 ** But no harm is done by this. And it is very important that | |
2097 ** btreeInitPage() be called on every btree page so we make | |
2098 ** the call for every page that comes in for re-initing. */ | |
2099 btreeInitPage(pPage); | |
2100 } | |
2101 } | |
2102 } | |
2103 | |
2104 /* | |
2105 ** Invoke the busy handler for a btree. | |
2106 */ | |
2107 static int btreeInvokeBusyHandler(void *pArg){ | |
2108 BtShared *pBt = (BtShared*)pArg; | |
2109 assert( pBt->db ); | |
2110 assert( sqlite3_mutex_held(pBt->db->mutex) ); | |
2111 return sqlite3InvokeBusyHandler(&pBt->db->busyHandler); | |
2112 } | |
2113 | |
2114 /* | |
2115 ** Open a database file. | |
2116 ** | |
2117 ** zFilename is the name of the database file. If zFilename is NULL | |
2118 ** then an ephemeral database is created. The ephemeral database might | |
2119 ** be exclusively in memory, or it might use a disk-based memory cache. | |
2120 ** Either way, the ephemeral database will be automatically deleted | |
2121 ** when sqlite3BtreeClose() is called. | |
2122 ** | |
2123 ** If zFilename is ":memory:" then an in-memory database is created | |
2124 ** that is automatically destroyed when it is closed. | |
2125 ** | |
2126 ** The "flags" parameter is a bitmask that might contain bits like | |
2127 ** BTREE_OMIT_JOURNAL and/or BTREE_MEMORY. | |
2128 ** | |
2129 ** If the database is already opened in the same database connection | |
2130 ** and we are in shared cache mode, then the open will fail with an | |
2131 ** SQLITE_CONSTRAINT error. We cannot allow two or more BtShared | |
2132 ** objects in the same database connection since doing so will lead | |
2133 ** to problems with locking. | |
2134 */ | |
2135 int sqlite3BtreeOpen( | |
2136 sqlite3_vfs *pVfs, /* VFS to use for this b-tree */ | |
2137 const char *zFilename, /* Name of the file containing the BTree database */ | |
2138 sqlite3 *db, /* Associated database handle */ | |
2139 Btree **ppBtree, /* Pointer to new Btree object written here */ | |
2140 int flags, /* Options */ | |
2141 int vfsFlags /* Flags passed through to sqlite3_vfs.xOpen() */ | |
2142 ){ | |
2143 BtShared *pBt = 0; /* Shared part of btree structure */ | |
2144 Btree *p; /* Handle to return */ | |
2145 sqlite3_mutex *mutexOpen = 0; /* Prevents a race condition. Ticket #3537 */ | |
2146 int rc = SQLITE_OK; /* Result code from this function */ | |
2147 u8 nReserve; /* Byte of unused space on each page */ | |
2148 unsigned char zDbHeader[100]; /* Database header content */ | |
2149 | |
2150 /* True if opening an ephemeral, temporary database */ | |
2151 const int isTempDb = zFilename==0 || zFilename[0]==0; | |
2152 | |
2153 /* Set the variable isMemdb to true for an in-memory database, or | |
2154 ** false for a file-based database. | |
2155 */ | |
2156 #ifdef SQLITE_OMIT_MEMORYDB | |
2157 const int isMemdb = 0; | |
2158 #else | |
2159 const int isMemdb = (zFilename && strcmp(zFilename, ":memory:")==0) | |
2160 || (isTempDb && sqlite3TempInMemory(db)) | |
2161 || (vfsFlags & SQLITE_OPEN_MEMORY)!=0; | |
2162 #endif | |
2163 | |
2164 assert( db!=0 ); | |
2165 assert( pVfs!=0 ); | |
2166 assert( sqlite3_mutex_held(db->mutex) ); | |
2167 assert( (flags&0xff)==flags ); /* flags fit in 8 bits */ | |
2168 | |
2169 /* Only a BTREE_SINGLE database can be BTREE_UNORDERED */ | |
2170 assert( (flags & BTREE_UNORDERED)==0 || (flags & BTREE_SINGLE)!=0 ); | |
2171 | |
2172 /* A BTREE_SINGLE database is always a temporary and/or ephemeral */ | |
2173 assert( (flags & BTREE_SINGLE)==0 || isTempDb ); | |
2174 | |
2175 if( isMemdb ){ | |
2176 flags |= BTREE_MEMORY; | |
2177 } | |
2178 if( (vfsFlags & SQLITE_OPEN_MAIN_DB)!=0 && (isMemdb || isTempDb) ){ | |
2179 vfsFlags = (vfsFlags & ~SQLITE_OPEN_MAIN_DB) | SQLITE_OPEN_TEMP_DB; | |
2180 } | |
2181 p = sqlite3MallocZero(sizeof(Btree)); | |
2182 if( !p ){ | |
2183 return SQLITE_NOMEM; | |
2184 } | |
2185 p->inTrans = TRANS_NONE; | |
2186 p->db = db; | |
2187 #ifndef SQLITE_OMIT_SHARED_CACHE | |
2188 p->lock.pBtree = p; | |
2189 p->lock.iTable = 1; | |
2190 #endif | |
2191 | |
2192 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO) | |
2193 /* | |
2194 ** If this Btree is a candidate for shared cache, try to find an | |
2195 ** existing BtShared object that we can share with | |
2196 */ | |
2197 if( isTempDb==0 && (isMemdb==0 || (vfsFlags&SQLITE_OPEN_URI)!=0) ){ | |
2198 if( vfsFlags & SQLITE_OPEN_SHAREDCACHE ){ | |
2199 int nFilename = sqlite3Strlen30(zFilename)+1; | |
2200 int nFullPathname = pVfs->mxPathname+1; | |
2201 char *zFullPathname = sqlite3Malloc(MAX(nFullPathname,nFilename)); | |
2202 MUTEX_LOGIC( sqlite3_mutex *mutexShared; ) | |
2203 | |
2204 p->sharable = 1; | |
2205 if( !zFullPathname ){ | |
2206 sqlite3_free(p); | |
2207 return SQLITE_NOMEM; | |
2208 } | |
2209 if( isMemdb ){ | |
2210 memcpy(zFullPathname, zFilename, nFilename); | |
2211 }else{ | |
2212 rc = sqlite3OsFullPathname(pVfs, zFilename, | |
2213 nFullPathname, zFullPathname); | |
2214 if( rc ){ | |
2215 sqlite3_free(zFullPathname); | |
2216 sqlite3_free(p); | |
2217 return rc; | |
2218 } | |
2219 } | |
2220 #if SQLITE_THREADSAFE | |
2221 mutexOpen = sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_OPEN); | |
2222 sqlite3_mutex_enter(mutexOpen); | |
2223 mutexShared = sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER); | |
2224 sqlite3_mutex_enter(mutexShared); | |
2225 #endif | |
2226 for(pBt=GLOBAL(BtShared*,sqlite3SharedCacheList); pBt; pBt=pBt->pNext){ | |
2227 assert( pBt->nRef>0 ); | |
2228 if( 0==strcmp(zFullPathname, sqlite3PagerFilename(pBt->pPager, 0)) | |
2229 && sqlite3PagerVfs(pBt->pPager)==pVfs ){ | |
2230 int iDb; | |
2231 for(iDb=db->nDb-1; iDb>=0; iDb--){ | |
2232 Btree *pExisting = db->aDb[iDb].pBt; | |
2233 if( pExisting && pExisting->pBt==pBt ){ | |
2234 sqlite3_mutex_leave(mutexShared); | |
2235 sqlite3_mutex_leave(mutexOpen); | |
2236 sqlite3_free(zFullPathname); | |
2237 sqlite3_free(p); | |
2238 return SQLITE_CONSTRAINT; | |
2239 } | |
2240 } | |
2241 p->pBt = pBt; | |
2242 pBt->nRef++; | |
2243 break; | |
2244 } | |
2245 } | |
2246 sqlite3_mutex_leave(mutexShared); | |
2247 sqlite3_free(zFullPathname); | |
2248 } | |
2249 #ifdef SQLITE_DEBUG | |
2250 else{ | |
2251 /* In debug mode, we mark all persistent databases as sharable | |
2252 ** even when they are not. This exercises the locking code and | |
2253 ** gives more opportunity for asserts(sqlite3_mutex_held()) | |
2254 ** statements to find locking problems. | |
2255 */ | |
2256 p->sharable = 1; | |
2257 } | |
2258 #endif | |
2259 } | |
2260 #endif | |
2261 if( pBt==0 ){ | |
2262 /* | |
2263 ** The following asserts make sure that structures used by the btree are | |
2264 ** the right size. This is to guard against size changes that result | |
2265 ** when compiling on a different architecture. | |
2266 */ | |
2267 assert( sizeof(i64)==8 ); | |
2268 assert( sizeof(u64)==8 ); | |
2269 assert( sizeof(u32)==4 ); | |
2270 assert( sizeof(u16)==2 ); | |
2271 assert( sizeof(Pgno)==4 ); | |
2272 | |
2273 pBt = sqlite3MallocZero( sizeof(*pBt) ); | |
2274 if( pBt==0 ){ | |
2275 rc = SQLITE_NOMEM; | |
2276 goto btree_open_out; | |
2277 } | |
2278 rc = sqlite3PagerOpen(pVfs, &pBt->pPager, zFilename, | |
2279 EXTRA_SIZE, flags, vfsFlags, pageReinit); | |
2280 if( rc==SQLITE_OK ){ | |
2281 sqlite3PagerSetMmapLimit(pBt->pPager, db->szMmap); | |
2282 rc = sqlite3PagerReadFileheader(pBt->pPager,sizeof(zDbHeader),zDbHeader); | |
2283 } | |
2284 if( rc!=SQLITE_OK ){ | |
2285 goto btree_open_out; | |
2286 } | |
2287 pBt->openFlags = (u8)flags; | |
2288 pBt->db = db; | |
2289 sqlite3PagerSetBusyhandler(pBt->pPager, btreeInvokeBusyHandler, pBt); | |
2290 p->pBt = pBt; | |
2291 | |
2292 pBt->pCursor = 0; | |
2293 pBt->pPage1 = 0; | |
2294 if( sqlite3PagerIsreadonly(pBt->pPager) ) pBt->btsFlags |= BTS_READ_ONLY; | |
2295 #ifdef SQLITE_SECURE_DELETE | |
2296 pBt->btsFlags |= BTS_SECURE_DELETE; | |
2297 #endif | |
2298 /* EVIDENCE-OF: R-51873-39618 The page size for a database file is | |
2299 ** determined by the 2-byte integer located at an offset of 16 bytes from | |
2300 ** the beginning of the database file. */ | |
2301 pBt->pageSize = (zDbHeader[16]<<8) | (zDbHeader[17]<<16); | |
2302 if( pBt->pageSize<512 || pBt->pageSize>SQLITE_MAX_PAGE_SIZE | |
2303 || ((pBt->pageSize-1)&pBt->pageSize)!=0 ){ | |
2304 pBt->pageSize = 0; | |
2305 #ifndef SQLITE_OMIT_AUTOVACUUM | |
2306 /* If the magic name ":memory:" will create an in-memory database, then | |
2307 ** leave the autoVacuum mode at 0 (do not auto-vacuum), even if | |
2308 ** SQLITE_DEFAULT_AUTOVACUUM is true. On the other hand, if | |
2309 ** SQLITE_OMIT_MEMORYDB has been defined, then ":memory:" is just a | |
2310 ** regular file-name. In this case the auto-vacuum applies as per normal. | |
2311 */ | |
2312 if( zFilename && !isMemdb ){ | |
2313 pBt->autoVacuum = (SQLITE_DEFAULT_AUTOVACUUM ? 1 : 0); | |
2314 pBt->incrVacuum = (SQLITE_DEFAULT_AUTOVACUUM==2 ? 1 : 0); | |
2315 } | |
2316 #endif | |
2317 nReserve = 0; | |
2318 }else{ | |
2319 /* EVIDENCE-OF: R-37497-42412 The size of the reserved region is | |
2320 ** determined by the one-byte unsigned integer found at an offset of 20 | |
2321 ** into the database file header. */ | |
2322 nReserve = zDbHeader[20]; | |
2323 pBt->btsFlags |= BTS_PAGESIZE_FIXED; | |
2324 #ifndef SQLITE_OMIT_AUTOVACUUM | |
2325 pBt->autoVacuum = (get4byte(&zDbHeader[36 + 4*4])?1:0); | |
2326 pBt->incrVacuum = (get4byte(&zDbHeader[36 + 7*4])?1:0); | |
2327 #endif | |
2328 } | |
2329 rc = sqlite3PagerSetPagesize(pBt->pPager, &pBt->pageSize, nReserve); | |
2330 if( rc ) goto btree_open_out; | |
2331 pBt->usableSize = pBt->pageSize - nReserve; | |
2332 assert( (pBt->pageSize & 7)==0 ); /* 8-byte alignment of pageSize */ | |
2333 | |
2334 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO) | |
2335 /* Add the new BtShared object to the linked list sharable BtShareds. | |
2336 */ | |
2337 if( p->sharable ){ | |
2338 MUTEX_LOGIC( sqlite3_mutex *mutexShared; ) | |
2339 pBt->nRef = 1; | |
2340 MUTEX_LOGIC( mutexShared = sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER);) | |
2341 if( SQLITE_THREADSAFE && sqlite3GlobalConfig.bCoreMutex ){ | |
2342 pBt->mutex = sqlite3MutexAlloc(SQLITE_MUTEX_FAST); | |
2343 if( pBt->mutex==0 ){ | |
2344 rc = SQLITE_NOMEM; | |
2345 db->mallocFailed = 0; | |
2346 goto btree_open_out; | |
2347 } | |
2348 } | |
2349 sqlite3_mutex_enter(mutexShared); | |
2350 pBt->pNext = GLOBAL(BtShared*,sqlite3SharedCacheList); | |
2351 GLOBAL(BtShared*,sqlite3SharedCacheList) = pBt; | |
2352 sqlite3_mutex_leave(mutexShared); | |
2353 } | |
2354 #endif | |
2355 } | |
2356 | |
2357 #if !defined(SQLITE_OMIT_SHARED_CACHE) && !defined(SQLITE_OMIT_DISKIO) | |
2358 /* If the new Btree uses a sharable pBtShared, then link the new | |
2359 ** Btree into the list of all sharable Btrees for the same connection. | |
2360 ** The list is kept in ascending order by pBt address. | |
2361 */ | |
2362 if( p->sharable ){ | |
2363 int i; | |
2364 Btree *pSib; | |
2365 for(i=0; i<db->nDb; i++){ | |
2366 if( (pSib = db->aDb[i].pBt)!=0 && pSib->sharable ){ | |
2367 while( pSib->pPrev ){ pSib = pSib->pPrev; } | |
2368 if( p->pBt<pSib->pBt ){ | |
2369 p->pNext = pSib; | |
2370 p->pPrev = 0; | |
2371 pSib->pPrev = p; | |
2372 }else{ | |
2373 while( pSib->pNext && pSib->pNext->pBt<p->pBt ){ | |
2374 pSib = pSib->pNext; | |
2375 } | |
2376 p->pNext = pSib->pNext; | |
2377 p->pPrev = pSib; | |
2378 if( p->pNext ){ | |
2379 p->pNext->pPrev = p; | |
2380 } | |
2381 pSib->pNext = p; | |
2382 } | |
2383 break; | |
2384 } | |
2385 } | |
2386 } | |
2387 #endif | |
2388 *ppBtree = p; | |
2389 | |
2390 btree_open_out: | |
2391 if( rc!=SQLITE_OK ){ | |
2392 if( pBt && pBt->pPager ){ | |
2393 sqlite3PagerClose(pBt->pPager); | |
2394 } | |
2395 sqlite3_free(pBt); | |
2396 sqlite3_free(p); | |
2397 *ppBtree = 0; | |
2398 }else{ | |
2399 /* If the B-Tree was successfully opened, set the pager-cache size to the | |
2400 ** default value. Except, when opening on an existing shared pager-cache, | |
2401 ** do not change the pager-cache size. | |
2402 */ | |
2403 if( sqlite3BtreeSchema(p, 0, 0)==0 ){ | |
2404 sqlite3PagerSetCachesize(p->pBt->pPager, SQLITE_DEFAULT_CACHE_SIZE); | |
2405 } | |
2406 } | |
2407 if( mutexOpen ){ | |
2408 assert( sqlite3_mutex_held(mutexOpen) ); | |
2409 sqlite3_mutex_leave(mutexOpen); | |
2410 } | |
2411 return rc; | |
2412 } | |
2413 | |
2414 /* | |
2415 ** Decrement the BtShared.nRef counter. When it reaches zero, | |
2416 ** remove the BtShared structure from the sharing list. Return | |
2417 ** true if the BtShared.nRef counter reaches zero and return | |
2418 ** false if it is still positive. | |
2419 */ | |
2420 static int removeFromSharingList(BtShared *pBt){ | |
2421 #ifndef SQLITE_OMIT_SHARED_CACHE | |
2422 MUTEX_LOGIC( sqlite3_mutex *pMaster; ) | |
2423 BtShared *pList; | |
2424 int removed = 0; | |
2425 | |
2426 assert( sqlite3_mutex_notheld(pBt->mutex) ); | |
2427 MUTEX_LOGIC( pMaster = sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_MASTER); ) | |
2428 sqlite3_mutex_enter(pMaster); | |
2429 pBt->nRef--; | |
2430 if( pBt->nRef<=0 ){ | |
2431 if( GLOBAL(BtShared*,sqlite3SharedCacheList)==pBt ){ | |
2432 GLOBAL(BtShared*,sqlite3SharedCacheList) = pBt->pNext; | |
2433 }else{ | |
2434 pList = GLOBAL(BtShared*,sqlite3SharedCacheList); | |
2435 while( ALWAYS(pList) && pList->pNext!=pBt ){ | |
2436 pList=pList->pNext; | |
2437 } | |
2438 if( ALWAYS(pList) ){ | |
2439 pList->pNext = pBt->pNext; | |
2440 } | |
2441 } | |
2442 if( SQLITE_THREADSAFE ){ | |
2443 sqlite3_mutex_free(pBt->mutex); | |
2444 } | |
2445 removed = 1; | |
2446 } | |
2447 sqlite3_mutex_leave(pMaster); | |
2448 return removed; | |
2449 #else | |
2450 return 1; | |
2451 #endif | |
2452 } | |
2453 | |
2454 /* | |
2455 ** Make sure pBt->pTmpSpace points to an allocation of | |
2456 ** MX_CELL_SIZE(pBt) bytes with a 4-byte prefix for a left-child | |
2457 ** pointer. | |
2458 */ | |
2459 static void allocateTempSpace(BtShared *pBt){ | |
2460 if( !pBt->pTmpSpace ){ | |
2461 pBt->pTmpSpace = sqlite3PageMalloc( pBt->pageSize ); | |
2462 | |
2463 /* One of the uses of pBt->pTmpSpace is to format cells before | |
2464 ** inserting them into a leaf page (function fillInCell()). If | |
2465 ** a cell is less than 4 bytes in size, it is rounded up to 4 bytes | |
2466 ** by the various routines that manipulate binary cells. Which | |
2467 ** can mean that fillInCell() only initializes the first 2 or 3 | |
2468 ** bytes of pTmpSpace, but that the first 4 bytes are copied from | |
2469 ** it into a database page. This is not actually a problem, but it | |
2470 ** does cause a valgrind error when the 1 or 2 bytes of unitialized | |
2471 ** data is passed to system call write(). So to avoid this error, | |
2472 ** zero the first 4 bytes of temp space here. | |
2473 ** | |
2474 ** Also: Provide four bytes of initialized space before the | |
2475 ** beginning of pTmpSpace as an area available to prepend the | |
2476 ** left-child pointer to the beginning of a cell. | |
2477 */ | |
2478 if( pBt->pTmpSpace ){ | |
2479 memset(pBt->pTmpSpace, 0, 8); | |
2480 pBt->pTmpSpace += 4; | |
2481 } | |
2482 } | |
2483 } | |
2484 | |
2485 /* | |
2486 ** Free the pBt->pTmpSpace allocation | |
2487 */ | |
2488 static void freeTempSpace(BtShared *pBt){ | |
2489 if( pBt->pTmpSpace ){ | |
2490 pBt->pTmpSpace -= 4; | |
2491 sqlite3PageFree(pBt->pTmpSpace); | |
2492 pBt->pTmpSpace = 0; | |
2493 } | |
2494 } | |
2495 | |
2496 /* | |
2497 ** Close an open database and invalidate all cursors. | |
2498 */ | |
2499 int sqlite3BtreeClose(Btree *p){ | |
2500 BtShared *pBt = p->pBt; | |
2501 BtCursor *pCur; | |
2502 | |
2503 /* Close all cursors opened via this handle. */ | |
2504 assert( sqlite3_mutex_held(p->db->mutex) ); | |
2505 sqlite3BtreeEnter(p); | |
2506 pCur = pBt->pCursor; | |
2507 while( pCur ){ | |
2508 BtCursor *pTmp = pCur; | |
2509 pCur = pCur->pNext; | |
2510 if( pTmp->pBtree==p ){ | |
2511 sqlite3BtreeCloseCursor(pTmp); | |
2512 } | |
2513 } | |
2514 | |
2515 /* Rollback any active transaction and free the handle structure. | |
2516 ** The call to sqlite3BtreeRollback() drops any table-locks held by | |
2517 ** this handle. | |
2518 */ | |
2519 sqlite3BtreeRollback(p, SQLITE_OK, 0); | |
2520 sqlite3BtreeLeave(p); | |
2521 | |
2522 /* If there are still other outstanding references to the shared-btree | |
2523 ** structure, return now. The remainder of this procedure cleans | |
2524 ** up the shared-btree. | |
2525 */ | |
2526 assert( p->wantToLock==0 && p->locked==0 ); | |
2527 if( !p->sharable || removeFromSharingList(pBt) ){ | |
2528 /* The pBt is no longer on the sharing list, so we can access | |
2529 ** it without having to hold the mutex. | |
2530 ** | |
2531 ** Clean out and delete the BtShared object. | |
2532 */ | |
2533 assert( !pBt->pCursor ); | |
2534 sqlite3PagerClose(pBt->pPager); | |
2535 if( pBt->xFreeSchema && pBt->pSchema ){ | |
2536 pBt->xFreeSchema(pBt->pSchema); | |
2537 } | |
2538 sqlite3DbFree(0, pBt->pSchema); | |
2539 freeTempSpace(pBt); | |
2540 sqlite3_free(pBt); | |
2541 } | |
2542 | |
2543 #ifndef SQLITE_OMIT_SHARED_CACHE | |
2544 assert( p->wantToLock==0 ); | |
2545 assert( p->locked==0 ); | |
2546 if( p->pPrev ) p->pPrev->pNext = p->pNext; | |
2547 if( p->pNext ) p->pNext->pPrev = p->pPrev; | |
2548 #endif | |
2549 | |
2550 sqlite3_free(p); | |
2551 return SQLITE_OK; | |
2552 } | |
2553 | |
2554 /* | |
2555 ** Change the "soft" limit on the number of pages in the cache. | |
2556 ** Unused and unmodified pages will be recycled when the number of | |
2557 ** pages in the cache exceeds this soft limit. But the size of the | |
2558 ** cache is allowed to grow larger than this limit if it contains | |
2559 ** dirty pages or pages still in active use. | |
2560 */ | |
2561 int sqlite3BtreeSetCacheSize(Btree *p, int mxPage){ | |
2562 BtShared *pBt = p->pBt; | |
2563 assert( sqlite3_mutex_held(p->db->mutex) ); | |
2564 sqlite3BtreeEnter(p); | |
2565 sqlite3PagerSetCachesize(pBt->pPager, mxPage); | |
2566 sqlite3BtreeLeave(p); | |
2567 return SQLITE_OK; | |
2568 } | |
2569 | |
2570 /* | |
2571 ** Change the "spill" limit on the number of pages in the cache. | |
2572 ** If the number of pages exceeds this limit during a write transaction, | |
2573 ** the pager might attempt to "spill" pages to the journal early in | |
2574 ** order to free up memory. | |
2575 ** | |
2576 ** The value returned is the current spill size. If zero is passed | |
2577 ** as an argument, no changes are made to the spill size setting, so | |
2578 ** using mxPage of 0 is a way to query the current spill size. | |
2579 */ | |
2580 int sqlite3BtreeSetSpillSize(Btree *p, int mxPage){ | |
2581 BtShared *pBt = p->pBt; | |
2582 int res; | |
2583 assert( sqlite3_mutex_held(p->db->mutex) ); | |
2584 sqlite3BtreeEnter(p); | |
2585 res = sqlite3PagerSetSpillsize(pBt->pPager, mxPage); | |
2586 sqlite3BtreeLeave(p); | |
2587 return res; | |
2588 } | |
2589 | |
2590 #if SQLITE_MAX_MMAP_SIZE>0 | |
2591 /* | |
2592 ** Change the limit on the amount of the database file that may be | |
2593 ** memory mapped. | |
2594 */ | |
2595 int sqlite3BtreeSetMmapLimit(Btree *p, sqlite3_int64 szMmap){ | |
2596 BtShared *pBt = p->pBt; | |
2597 assert( sqlite3_mutex_held(p->db->mutex) ); | |
2598 sqlite3BtreeEnter(p); | |
2599 sqlite3PagerSetMmapLimit(pBt->pPager, szMmap); | |
2600 sqlite3BtreeLeave(p); | |
2601 return SQLITE_OK; | |
2602 } | |
2603 #endif /* SQLITE_MAX_MMAP_SIZE>0 */ | |
2604 | |
2605 /* | |
2606 ** Change the way data is synced to disk in order to increase or decrease | |
2607 ** how well the database resists damage due to OS crashes and power | |
2608 ** failures. Level 1 is the same as asynchronous (no syncs() occur and | |
2609 ** there is a high probability of damage) Level 2 is the default. There | |
2610 ** is a very low but non-zero probability of damage. Level 3 reduces the | |
2611 ** probability of damage to near zero but with a write performance reduction. | |
2612 */ | |
2613 #ifndef SQLITE_OMIT_PAGER_PRAGMAS | |
2614 int sqlite3BtreeSetPagerFlags( | |
2615 Btree *p, /* The btree to set the safety level on */ | |
2616 unsigned pgFlags /* Various PAGER_* flags */ | |
2617 ){ | |
2618 BtShared *pBt = p->pBt; | |
2619 assert( sqlite3_mutex_held(p->db->mutex) ); | |
2620 sqlite3BtreeEnter(p); | |
2621 sqlite3PagerSetFlags(pBt->pPager, pgFlags); | |
2622 sqlite3BtreeLeave(p); | |
2623 return SQLITE_OK; | |
2624 } | |
2625 #endif | |
2626 | |
2627 /* | |
2628 ** Return TRUE if the given btree is set to safety level 1. In other | |
2629 ** words, return TRUE if no sync() occurs on the disk files. | |
2630 */ | |
2631 int sqlite3BtreeSyncDisabled(Btree *p){ | |
2632 BtShared *pBt = p->pBt; | |
2633 int rc; | |
2634 assert( sqlite3_mutex_held(p->db->mutex) ); | |
2635 sqlite3BtreeEnter(p); | |
2636 assert( pBt && pBt->pPager ); | |
2637 rc = sqlite3PagerNosync(pBt->pPager); | |
2638 sqlite3BtreeLeave(p); | |
2639 return rc; | |
2640 } | |
2641 | |
2642 /* | |
2643 ** Change the default pages size and the number of reserved bytes per page. | |
2644 ** Or, if the page size has already been fixed, return SQLITE_READONLY | |
2645 ** without changing anything. | |
2646 ** | |
2647 ** The page size must be a power of 2 between 512 and 65536. If the page | |
2648 ** size supplied does not meet this constraint then the page size is not | |
2649 ** changed. | |
2650 ** | |
2651 ** Page sizes are constrained to be a power of two so that the region | |
2652 ** of the database file used for locking (beginning at PENDING_BYTE, | |
2653 ** the first byte past the 1GB boundary, 0x40000000) needs to occur | |
2654 ** at the beginning of a page. | |
2655 ** | |
2656 ** If parameter nReserve is less than zero, then the number of reserved | |
2657 ** bytes per page is left unchanged. | |
2658 ** | |
2659 ** If the iFix!=0 then the BTS_PAGESIZE_FIXED flag is set so that the page size | |
2660 ** and autovacuum mode can no longer be changed. | |
2661 */ | |
2662 int sqlite3BtreeSetPageSize(Btree *p, int pageSize, int nReserve, int iFix){ | |
2663 int rc = SQLITE_OK; | |
2664 BtShared *pBt = p->pBt; | |
2665 assert( nReserve>=-1 && nReserve<=255 ); | |
2666 sqlite3BtreeEnter(p); | |
2667 #if SQLITE_HAS_CODEC | |
2668 if( nReserve>pBt->optimalReserve ) pBt->optimalReserve = (u8)nReserve; | |
2669 #endif | |
2670 if( pBt->btsFlags & BTS_PAGESIZE_FIXED ){ | |
2671 sqlite3BtreeLeave(p); | |
2672 return SQLITE_READONLY; | |
2673 } | |
2674 if( nReserve<0 ){ | |
2675 nReserve = pBt->pageSize - pBt->usableSize; | |
2676 } | |
2677 assert( nReserve>=0 && nReserve<=255 ); | |
2678 if( pageSize>=512 && pageSize<=SQLITE_MAX_PAGE_SIZE && | |
2679 ((pageSize-1)&pageSize)==0 ){ | |
2680 assert( (pageSize & 7)==0 ); | |
2681 assert( !pBt->pCursor ); | |
2682 pBt->pageSize = (u32)pageSize; | |
2683 freeTempSpace(pBt); | |
2684 } | |
2685 rc = sqlite3PagerSetPagesize(pBt->pPager, &pBt->pageSize, nReserve); | |
2686 pBt->usableSize = pBt->pageSize - (u16)nReserve; | |
2687 if( iFix ) pBt->btsFlags |= BTS_PAGESIZE_FIXED; | |
2688 sqlite3BtreeLeave(p); | |
2689 return rc; | |
2690 } | |
2691 | |
2692 /* | |
2693 ** Return the currently defined page size | |
2694 */ | |
2695 int sqlite3BtreeGetPageSize(Btree *p){ | |
2696 return p->pBt->pageSize; | |
2697 } | |
2698 | |
2699 /* | |
2700 ** This function is similar to sqlite3BtreeGetReserve(), except that it | |
2701 ** may only be called if it is guaranteed that the b-tree mutex is already | |
2702 ** held. | |
2703 ** | |
2704 ** This is useful in one special case in the backup API code where it is | |
2705 ** known that the shared b-tree mutex is held, but the mutex on the | |
2706 ** database handle that owns *p is not. In this case if sqlite3BtreeEnter() | |
2707 ** were to be called, it might collide with some other operation on the | |
2708 ** database handle that owns *p, causing undefined behavior. | |
2709 */ | |
2710 int sqlite3BtreeGetReserveNoMutex(Btree *p){ | |
2711 int n; | |
2712 assert( sqlite3_mutex_held(p->pBt->mutex) ); | |
2713 n = p->pBt->pageSize - p->pBt->usableSize; | |
2714 return n; | |
2715 } | |
2716 | |
2717 /* | |
2718 ** Return the number of bytes of space at the end of every page that | |
2719 ** are intentually left unused. This is the "reserved" space that is | |
2720 ** sometimes used by extensions. | |
2721 ** | |
2722 ** If SQLITE_HAS_MUTEX is defined then the number returned is the | |
2723 ** greater of the current reserved space and the maximum requested | |
2724 ** reserve space. | |
2725 */ | |
2726 int sqlite3BtreeGetOptimalReserve(Btree *p){ | |
2727 int n; | |
2728 sqlite3BtreeEnter(p); | |
2729 n = sqlite3BtreeGetReserveNoMutex(p); | |
2730 #ifdef SQLITE_HAS_CODEC | |
2731 if( n<p->pBt->optimalReserve ) n = p->pBt->optimalReserve; | |
2732 #endif | |
2733 sqlite3BtreeLeave(p); | |
2734 return n; | |
2735 } | |
2736 | |
2737 | |
2738 /* | |
2739 ** Set the maximum page count for a database if mxPage is positive. | |
2740 ** No changes are made if mxPage is 0 or negative. | |
2741 ** Regardless of the value of mxPage, return the maximum page count. | |
2742 */ | |
2743 int sqlite3BtreeMaxPageCount(Btree *p, int mxPage){ | |
2744 int n; | |
2745 sqlite3BtreeEnter(p); | |
2746 n = sqlite3PagerMaxPageCount(p->pBt->pPager, mxPage); | |
2747 sqlite3BtreeLeave(p); | |
2748 return n; | |
2749 } | |
2750 | |
2751 /* | |
2752 ** Set the BTS_SECURE_DELETE flag if newFlag is 0 or 1. If newFlag is -1, | |
2753 ** then make no changes. Always return the value of the BTS_SECURE_DELETE | |
2754 ** setting after the change. | |
2755 */ | |
2756 int sqlite3BtreeSecureDelete(Btree *p, int newFlag){ | |
2757 int b; | |
2758 if( p==0 ) return 0; | |
2759 sqlite3BtreeEnter(p); | |
2760 if( newFlag>=0 ){ | |
2761 p->pBt->btsFlags &= ~BTS_SECURE_DELETE; | |
2762 if( newFlag ) p->pBt->btsFlags |= BTS_SECURE_DELETE; | |
2763 } | |
2764 b = (p->pBt->btsFlags & BTS_SECURE_DELETE)!=0; | |
2765 sqlite3BtreeLeave(p); | |
2766 return b; | |
2767 } | |
2768 | |
2769 /* | |
2770 ** Change the 'auto-vacuum' property of the database. If the 'autoVacuum' | |
2771 ** parameter is non-zero, then auto-vacuum mode is enabled. If zero, it | |
2772 ** is disabled. The default value for the auto-vacuum property is | |
2773 ** determined by the SQLITE_DEFAULT_AUTOVACUUM macro. | |
2774 */ | |
2775 int sqlite3BtreeSetAutoVacuum(Btree *p, int autoVacuum){ | |
2776 #ifdef SQLITE_OMIT_AUTOVACUUM | |
2777 return SQLITE_READONLY; | |
2778 #else | |
2779 BtShared *pBt = p->pBt; | |
2780 int rc = SQLITE_OK; | |
2781 u8 av = (u8)autoVacuum; | |
2782 | |
2783 sqlite3BtreeEnter(p); | |
2784 if( (pBt->btsFlags & BTS_PAGESIZE_FIXED)!=0 && (av ?1:0)!=pBt->autoVacuum ){ | |
2785 rc = SQLITE_READONLY; | |
2786 }else{ | |
2787 pBt->autoVacuum = av ?1:0; | |
2788 pBt->incrVacuum = av==2 ?1:0; | |
2789 } | |
2790 sqlite3BtreeLeave(p); | |
2791 return rc; | |
2792 #endif | |
2793 } | |
2794 | |
2795 /* | |
2796 ** Return the value of the 'auto-vacuum' property. If auto-vacuum is | |
2797 ** enabled 1 is returned. Otherwise 0. | |
2798 */ | |
2799 int sqlite3BtreeGetAutoVacuum(Btree *p){ | |
2800 #ifdef SQLITE_OMIT_AUTOVACUUM | |
2801 return BTREE_AUTOVACUUM_NONE; | |
2802 #else | |
2803 int rc; | |
2804 sqlite3BtreeEnter(p); | |
2805 rc = ( | |
2806 (!p->pBt->autoVacuum)?BTREE_AUTOVACUUM_NONE: | |
2807 (!p->pBt->incrVacuum)?BTREE_AUTOVACUUM_FULL: | |
2808 BTREE_AUTOVACUUM_INCR | |
2809 ); | |
2810 sqlite3BtreeLeave(p); | |
2811 return rc; | |
2812 #endif | |
2813 } | |
2814 | |
2815 | |
2816 /* | |
2817 ** Get a reference to pPage1 of the database file. This will | |
2818 ** also acquire a readlock on that file. | |
2819 ** | |
2820 ** SQLITE_OK is returned on success. If the file is not a | |
2821 ** well-formed database file, then SQLITE_CORRUPT is returned. | |
2822 ** SQLITE_BUSY is returned if the database is locked. SQLITE_NOMEM | |
2823 ** is returned if we run out of memory. | |
2824 */ | |
2825 static int lockBtree(BtShared *pBt){ | |
2826 int rc; /* Result code from subfunctions */ | |
2827 MemPage *pPage1; /* Page 1 of the database file */ | |
2828 int nPage; /* Number of pages in the database */ | |
2829 int nPageFile = 0; /* Number of pages in the database file */ | |
2830 int nPageHeader; /* Number of pages in the database according to hdr */ | |
2831 | |
2832 assert( sqlite3_mutex_held(pBt->mutex) ); | |
2833 assert( pBt->pPage1==0 ); | |
2834 rc = sqlite3PagerSharedLock(pBt->pPager); | |
2835 if( rc!=SQLITE_OK ) return rc; | |
2836 rc = btreeGetPage(pBt, 1, &pPage1, 0); | |
2837 if( rc!=SQLITE_OK ) return rc; | |
2838 | |
2839 /* Do some checking to help insure the file we opened really is | |
2840 ** a valid database file. | |
2841 */ | |
2842 nPage = nPageHeader = get4byte(28+(u8*)pPage1->aData); | |
2843 sqlite3PagerPagecount(pBt->pPager, &nPageFile); | |
2844 if( nPage==0 || memcmp(24+(u8*)pPage1->aData, 92+(u8*)pPage1->aData,4)!=0 ){ | |
2845 nPage = nPageFile; | |
2846 } | |
2847 if( nPage>0 ){ | |
2848 u32 pageSize; | |
2849 u32 usableSize; | |
2850 u8 *page1 = pPage1->aData; | |
2851 rc = SQLITE_NOTADB; | |
2852 /* EVIDENCE-OF: R-43737-39999 Every valid SQLite database file begins | |
2853 ** with the following 16 bytes (in hex): 53 51 4c 69 74 65 20 66 6f 72 6d | |
2854 ** 61 74 20 33 00. */ | |
2855 if( memcmp(page1, zMagicHeader, 16)!=0 ){ | |
2856 goto page1_init_failed; | |
2857 } | |
2858 | |
2859 #ifdef SQLITE_OMIT_WAL | |
2860 if( page1[18]>1 ){ | |
2861 pBt->btsFlags |= BTS_READ_ONLY; | |
2862 } | |
2863 if( page1[19]>1 ){ | |
2864 goto page1_init_failed; | |
2865 } | |
2866 #else | |
2867 if( page1[18]>2 ){ | |
2868 pBt->btsFlags |= BTS_READ_ONLY; | |
2869 } | |
2870 if( page1[19]>2 ){ | |
2871 goto page1_init_failed; | |
2872 } | |
2873 | |
2874 /* If the write version is set to 2, this database should be accessed | |
2875 ** in WAL mode. If the log is not already open, open it now. Then | |
2876 ** return SQLITE_OK and return without populating BtShared.pPage1. | |
2877 ** The caller detects this and calls this function again. This is | |
2878 ** required as the version of page 1 currently in the page1 buffer | |
2879 ** may not be the latest version - there may be a newer one in the log | |
2880 ** file. | |
2881 */ | |
2882 if( page1[19]==2 && (pBt->btsFlags & BTS_NO_WAL)==0 ){ | |
2883 int isOpen = 0; | |
2884 rc = sqlite3PagerOpenWal(pBt->pPager, &isOpen); | |
2885 if( rc!=SQLITE_OK ){ | |
2886 goto page1_init_failed; | |
2887 }else if( isOpen==0 ){ | |
2888 releasePage(pPage1); | |
2889 return SQLITE_OK; | |
2890 } | |
2891 rc = SQLITE_NOTADB; | |
2892 } | |
2893 #endif | |
2894 | |
2895 /* EVIDENCE-OF: R-15465-20813 The maximum and minimum embedded payload | |
2896 ** fractions and the leaf payload fraction values must be 64, 32, and 32. | |
2897 ** | |
2898 ** The original design allowed these amounts to vary, but as of | |
2899 ** version 3.6.0, we require them to be fixed. | |
2900 */ | |
2901 if( memcmp(&page1[21], "\100\040\040",3)!=0 ){ | |
2902 goto page1_init_failed; | |
2903 } | |
2904 /* EVIDENCE-OF: R-51873-39618 The page size for a database file is | |
2905 ** determined by the 2-byte integer located at an offset of 16 bytes from | |
2906 ** the beginning of the database file. */ | |
2907 pageSize = (page1[16]<<8) | (page1[17]<<16); | |
2908 /* EVIDENCE-OF: R-25008-21688 The size of a page is a power of two | |
2909 ** between 512 and 65536 inclusive. */ | |
2910 if( ((pageSize-1)&pageSize)!=0 | |
2911 || pageSize>SQLITE_MAX_PAGE_SIZE | |
2912 || pageSize<=256 | |
2913 ){ | |
2914 goto page1_init_failed; | |
2915 } | |
2916 assert( (pageSize & 7)==0 ); | |
2917 /* EVIDENCE-OF: R-59310-51205 The "reserved space" size in the 1-byte | |
2918 ** integer at offset 20 is the number of bytes of space at the end of | |
2919 ** each page to reserve for extensions. | |
2920 ** | |
2921 ** EVIDENCE-OF: R-37497-42412 The size of the reserved region is | |
2922 ** determined by the one-byte unsigned integer found at an offset of 20 | |
2923 ** into the database file header. */ | |
2924 usableSize = pageSize - page1[20]; | |
2925 if( (u32)pageSize!=pBt->pageSize ){ | |
2926 /* After reading the first page of the database assuming a page size | |
2927 ** of BtShared.pageSize, we have discovered that the page-size is | |
2928 ** actually pageSize. Unlock the database, leave pBt->pPage1 at | |
2929 ** zero and return SQLITE_OK. The caller will call this function | |
2930 ** again with the correct page-size. | |
2931 */ | |
2932 releasePage(pPage1); | |
2933 pBt->usableSize = usableSize; | |
2934 pBt->pageSize = pageSize; | |
2935 freeTempSpace(pBt); | |
2936 rc = sqlite3PagerSetPagesize(pBt->pPager, &pBt->pageSize, | |
2937 pageSize-usableSize); | |
2938 return rc; | |
2939 } | |
2940 if( (pBt->db->flags & SQLITE_RecoveryMode)==0 && nPage>nPageFile ){ | |
2941 rc = SQLITE_CORRUPT_BKPT; | |
2942 goto page1_init_failed; | |
2943 } | |
2944 /* EVIDENCE-OF: R-28312-64704 However, the usable size is not allowed to | |
2945 ** be less than 480. In other words, if the page size is 512, then the | |
2946 ** reserved space size cannot exceed 32. */ | |
2947 if( usableSize<480 ){ | |
2948 goto page1_init_failed; | |
2949 } | |
2950 pBt->pageSize = pageSize; | |
2951 pBt->usableSize = usableSize; | |
2952 #ifndef SQLITE_OMIT_AUTOVACUUM | |
2953 pBt->autoVacuum = (get4byte(&page1[36 + 4*4])?1:0); | |
2954 pBt->incrVacuum = (get4byte(&page1[36 + 7*4])?1:0); | |
2955 #endif | |
2956 } | |
2957 | |
2958 /* maxLocal is the maximum amount of payload to store locally for | |
2959 ** a cell. Make sure it is small enough so that at least minFanout | |
2960 ** cells can will fit on one page. We assume a 10-byte page header. | |
2961 ** Besides the payload, the cell must store: | |
2962 ** 2-byte pointer to the cell | |
2963 ** 4-byte child pointer | |
2964 ** 9-byte nKey value | |
2965 ** 4-byte nData value | |
2966 ** 4-byte overflow page pointer | |
2967 ** So a cell consists of a 2-byte pointer, a header which is as much as | |
2968 ** 17 bytes long, 0 to N bytes of payload, and an optional 4 byte overflow | |
2969 ** page pointer. | |
2970 */ | |
2971 pBt->maxLocal = (u16)((pBt->usableSize-12)*64/255 - 23); | |
2972 pBt->minLocal = (u16)((pBt->usableSize-12)*32/255 - 23); | |
2973 pBt->maxLeaf = (u16)(pBt->usableSize - 35); | |
2974 pBt->minLeaf = (u16)((pBt->usableSize-12)*32/255 - 23); | |
2975 if( pBt->maxLocal>127 ){ | |
2976 pBt->max1bytePayload = 127; | |
2977 }else{ | |
2978 pBt->max1bytePayload = (u8)pBt->maxLocal; | |
2979 } | |
2980 assert( pBt->maxLeaf + 23 <= MX_CELL_SIZE(pBt) ); | |
2981 pBt->pPage1 = pPage1; | |
2982 pBt->nPage = nPage; | |
2983 return SQLITE_OK; | |
2984 | |
2985 page1_init_failed: | |
2986 releasePage(pPage1); | |
2987 pBt->pPage1 = 0; | |
2988 return rc; | |
2989 } | |
2990 | |
2991 #ifndef NDEBUG | |
2992 /* | |
2993 ** Return the number of cursors open on pBt. This is for use | |
2994 ** in assert() expressions, so it is only compiled if NDEBUG is not | |
2995 ** defined. | |
2996 ** | |
2997 ** Only write cursors are counted if wrOnly is true. If wrOnly is | |
2998 ** false then all cursors are counted. | |
2999 ** | |
3000 ** For the purposes of this routine, a cursor is any cursor that | |
3001 ** is capable of reading or writing to the database. Cursors that | |
3002 ** have been tripped into the CURSOR_FAULT state are not counted. | |
3003 */ | |
3004 static int countValidCursors(BtShared *pBt, int wrOnly){ | |
3005 BtCursor *pCur; | |
3006 int r = 0; | |
3007 for(pCur=pBt->pCursor; pCur; pCur=pCur->pNext){ | |
3008 if( (wrOnly==0 || (pCur->curFlags & BTCF_WriteFlag)!=0) | |
3009 && pCur->eState!=CURSOR_FAULT ) r++; | |
3010 } | |
3011 return r; | |
3012 } | |
3013 #endif | |
3014 | |
3015 /* | |
3016 ** If there are no outstanding cursors and we are not in the middle | |
3017 ** of a transaction but there is a read lock on the database, then | |
3018 ** this routine unrefs the first page of the database file which | |
3019 ** has the effect of releasing the read lock. | |
3020 ** | |
3021 ** If there is a transaction in progress, this routine is a no-op. | |
3022 */ | |
3023 static void unlockBtreeIfUnused(BtShared *pBt){ | |
3024 assert( sqlite3_mutex_held(pBt->mutex) ); | |
3025 assert( countValidCursors(pBt,0)==0 || pBt->inTransaction>TRANS_NONE ); | |
3026 if( pBt->inTransaction==TRANS_NONE && pBt->pPage1!=0 ){ | |
3027 MemPage *pPage1 = pBt->pPage1; | |
3028 assert( pPage1->aData ); | |
3029 assert( sqlite3PagerRefcount(pBt->pPager)==1 ); | |
3030 pBt->pPage1 = 0; | |
3031 releasePageNotNull(pPage1); | |
3032 } | |
3033 } | |
3034 | |
3035 /* | |
3036 ** If pBt points to an empty file then convert that empty file | |
3037 ** into a new empty database by initializing the first page of | |
3038 ** the database. | |
3039 */ | |
3040 static int newDatabase(BtShared *pBt){ | |
3041 MemPage *pP1; | |
3042 unsigned char *data; | |
3043 int rc; | |
3044 | |
3045 assert( sqlite3_mutex_held(pBt->mutex) ); | |
3046 if( pBt->nPage>0 ){ | |
3047 return SQLITE_OK; | |
3048 } | |
3049 pP1 = pBt->pPage1; | |
3050 assert( pP1!=0 ); | |
3051 data = pP1->aData; | |
3052 rc = sqlite3PagerWrite(pP1->pDbPage); | |
3053 if( rc ) return rc; | |
3054 memcpy(data, zMagicHeader, sizeof(zMagicHeader)); | |
3055 assert( sizeof(zMagicHeader)==16 ); | |
3056 data[16] = (u8)((pBt->pageSize>>8)&0xff); | |
3057 data[17] = (u8)((pBt->pageSize>>16)&0xff); | |
3058 data[18] = 1; | |
3059 data[19] = 1; | |
3060 assert( pBt->usableSize<=pBt->pageSize && pBt->usableSize+255>=pBt->pageSize); | |
3061 data[20] = (u8)(pBt->pageSize - pBt->usableSize); | |
3062 data[21] = 64; | |
3063 data[22] = 32; | |
3064 data[23] = 32; | |
3065 memset(&data[24], 0, 100-24); | |
3066 zeroPage(pP1, PTF_INTKEY|PTF_LEAF|PTF_LEAFDATA ); | |
3067 pBt->btsFlags |= BTS_PAGESIZE_FIXED; | |
3068 #ifndef SQLITE_OMIT_AUTOVACUUM | |
3069 assert( pBt->autoVacuum==1 || pBt->autoVacuum==0 ); | |
3070 assert( pBt->incrVacuum==1 || pBt->incrVacuum==0 ); | |
3071 put4byte(&data[36 + 4*4], pBt->autoVacuum); | |
3072 put4byte(&data[36 + 7*4], pBt->incrVacuum); | |
3073 #endif | |
3074 pBt->nPage = 1; | |
3075 data[31] = 1; | |
3076 return SQLITE_OK; | |
3077 } | |
3078 | |
3079 /* | |
3080 ** Initialize the first page of the database file (creating a database | |
3081 ** consisting of a single page and no schema objects). Return SQLITE_OK | |
3082 ** if successful, or an SQLite error code otherwise. | |
3083 */ | |
3084 int sqlite3BtreeNewDb(Btree *p){ | |
3085 int rc; | |
3086 sqlite3BtreeEnter(p); | |
3087 p->pBt->nPage = 0; | |
3088 rc = newDatabase(p->pBt); | |
3089 sqlite3BtreeLeave(p); | |
3090 return rc; | |
3091 } | |
3092 | |
3093 /* | |
3094 ** Attempt to start a new transaction. A write-transaction | |
3095 ** is started if the second argument is nonzero, otherwise a read- | |
3096 ** transaction. If the second argument is 2 or more and exclusive | |
3097 ** transaction is started, meaning that no other process is allowed | |
3098 ** to access the database. A preexisting transaction may not be | |
3099 ** upgraded to exclusive by calling this routine a second time - the | |
3100 ** exclusivity flag only works for a new transaction. | |
3101 ** | |
3102 ** A write-transaction must be started before attempting any | |
3103 ** changes to the database. None of the following routines | |
3104 ** will work unless a transaction is started first: | |
3105 ** | |
3106 ** sqlite3BtreeCreateTable() | |
3107 ** sqlite3BtreeCreateIndex() | |
3108 ** sqlite3BtreeClearTable() | |
3109 ** sqlite3BtreeDropTable() | |
3110 ** sqlite3BtreeInsert() | |
3111 ** sqlite3BtreeDelete() | |
3112 ** sqlite3BtreeUpdateMeta() | |
3113 ** | |
3114 ** If an initial attempt to acquire the lock fails because of lock contention | |
3115 ** and the database was previously unlocked, then invoke the busy handler | |
3116 ** if there is one. But if there was previously a read-lock, do not | |
3117 ** invoke the busy handler - just return SQLITE_BUSY. SQLITE_BUSY is | |
3118 ** returned when there is already a read-lock in order to avoid a deadlock. | |
3119 ** | |
3120 ** Suppose there are two processes A and B. A has a read lock and B has | |
3121 ** a reserved lock. B tries to promote to exclusive but is blocked because | |
3122 ** of A's read lock. A tries to promote to reserved but is blocked by B. | |
3123 ** One or the other of the two processes must give way or there can be | |
3124 ** no progress. By returning SQLITE_BUSY and not invoking the busy callback | |
3125 ** when A already has a read lock, we encourage A to give up and let B | |
3126 ** proceed. | |
3127 */ | |
3128 int sqlite3BtreeBeginTrans(Btree *p, int wrflag){ | |
3129 sqlite3 *pBlock = 0; | |
3130 BtShared *pBt = p->pBt; | |
3131 int rc = SQLITE_OK; | |
3132 | |
3133 sqlite3BtreeEnter(p); | |
3134 btreeIntegrity(p); | |
3135 | |
3136 /* If the btree is already in a write-transaction, or it | |
3137 ** is already in a read-transaction and a read-transaction | |
3138 ** is requested, this is a no-op. | |
3139 */ | |
3140 if( p->inTrans==TRANS_WRITE || (p->inTrans==TRANS_READ && !wrflag) ){ | |
3141 goto trans_begun; | |
3142 } | |
3143 assert( pBt->inTransaction==TRANS_WRITE || IfNotOmitAV(pBt->bDoTruncate)==0 ); | |
3144 | |
3145 /* Write transactions are not possible on a read-only database */ | |
3146 if( (pBt->btsFlags & BTS_READ_ONLY)!=0 && wrflag ){ | |
3147 rc = SQLITE_READONLY; | |
3148 goto trans_begun; | |
3149 } | |
3150 | |
3151 #ifndef SQLITE_OMIT_SHARED_CACHE | |
3152 /* If another database handle has already opened a write transaction | |
3153 ** on this shared-btree structure and a second write transaction is | |
3154 ** requested, return SQLITE_LOCKED. | |
3155 */ | |
3156 if( (wrflag && pBt->inTransaction==TRANS_WRITE) | |
3157 || (pBt->btsFlags & BTS_PENDING)!=0 | |
3158 ){ | |
3159 pBlock = pBt->pWriter->db; | |
3160 }else if( wrflag>1 ){ | |
3161 BtLock *pIter; | |
3162 for(pIter=pBt->pLock; pIter; pIter=pIter->pNext){ | |
3163 if( pIter->pBtree!=p ){ | |
3164 pBlock = pIter->pBtree->db; | |
3165 break; | |
3166 } | |
3167 } | |
3168 } | |
3169 if( pBlock ){ | |
3170 sqlite3ConnectionBlocked(p->db, pBlock); | |
3171 rc = SQLITE_LOCKED_SHAREDCACHE; | |
3172 goto trans_begun; | |
3173 } | |
3174 #endif | |
3175 | |
3176 /* Any read-only or read-write transaction implies a read-lock on | |
3177 ** page 1. So if some other shared-cache client already has a write-lock | |
3178 ** on page 1, the transaction cannot be opened. */ | |
3179 rc = querySharedCacheTableLock(p, MASTER_ROOT, READ_LOCK); | |
3180 if( SQLITE_OK!=rc ) goto trans_begun; | |
3181 | |
3182 pBt->btsFlags &= ~BTS_INITIALLY_EMPTY; | |
3183 if( pBt->nPage==0 ) pBt->btsFlags |= BTS_INITIALLY_EMPTY; | |
3184 do { | |
3185 /* Call lockBtree() until either pBt->pPage1 is populated or | |
3186 ** lockBtree() returns something other than SQLITE_OK. lockBtree() | |
3187 ** may return SQLITE_OK but leave pBt->pPage1 set to 0 if after | |
3188 ** reading page 1 it discovers that the page-size of the database | |
3189 ** file is not pBt->pageSize. In this case lockBtree() will update | |
3190 ** pBt->pageSize to the page-size of the file on disk. | |
3191 */ | |
3192 while( pBt->pPage1==0 && SQLITE_OK==(rc = lockBtree(pBt)) ); | |
3193 | |
3194 if( rc==SQLITE_OK && wrflag ){ | |
3195 if( (pBt->btsFlags & BTS_READ_ONLY)!=0 ){ | |
3196 rc = SQLITE_READONLY; | |
3197 }else{ | |
3198 rc = sqlite3PagerBegin(pBt->pPager,wrflag>1,sqlite3TempInMemory(p->db)); | |
3199 if( rc==SQLITE_OK ){ | |
3200 rc = newDatabase(pBt); | |
3201 } | |
3202 } | |
3203 } | |
3204 | |
3205 if( rc!=SQLITE_OK ){ | |
3206 unlockBtreeIfUnused(pBt); | |
3207 } | |
3208 }while( (rc&0xFF)==SQLITE_BUSY && pBt->inTransaction==TRANS_NONE && | |
3209 btreeInvokeBusyHandler(pBt) ); | |
3210 | |
3211 if( rc==SQLITE_OK ){ | |
3212 if( p->inTrans==TRANS_NONE ){ | |
3213 pBt->nTransaction++; | |
3214 #ifndef SQLITE_OMIT_SHARED_CACHE | |
3215 if( p->sharable ){ | |
3216 assert( p->lock.pBtree==p && p->lock.iTable==1 ); | |
3217 p->lock.eLock = READ_LOCK; | |
3218 p->lock.pNext = pBt->pLock; | |
3219 pBt->pLock = &p->lock; | |
3220 } | |
3221 #endif | |
3222 } | |
3223 p->inTrans = (wrflag?TRANS_WRITE:TRANS_READ); | |
3224 if( p->inTrans>pBt->inTransaction ){ | |
3225 pBt->inTransaction = p->inTrans; | |
3226 } | |
3227 if( wrflag ){ | |
3228 MemPage *pPage1 = pBt->pPage1; | |
3229 #ifndef SQLITE_OMIT_SHARED_CACHE | |
3230 assert( !pBt->pWriter ); | |
3231 pBt->pWriter = p; | |
3232 pBt->btsFlags &= ~BTS_EXCLUSIVE; | |
3233 if( wrflag>1 ) pBt->btsFlags |= BTS_EXCLUSIVE; | |
3234 #endif | |
3235 | |
3236 /* If the db-size header field is incorrect (as it may be if an old | |
3237 ** client has been writing the database file), update it now. Doing | |
3238 ** this sooner rather than later means the database size can safely | |
3239 ** re-read the database size from page 1 if a savepoint or transaction | |
3240 ** rollback occurs within the transaction. | |
3241 */ | |
3242 if( pBt->nPage!=get4byte(&pPage1->aData[28]) ){ | |
3243 rc = sqlite3PagerWrite(pPage1->pDbPage); | |
3244 if( rc==SQLITE_OK ){ | |
3245 put4byte(&pPage1->aData[28], pBt->nPage); | |
3246 } | |
3247 } | |
3248 } | |
3249 } | |
3250 | |
3251 | |
3252 trans_begun: | |
3253 if( rc==SQLITE_OK && wrflag ){ | |
3254 /* This call makes sure that the pager has the correct number of | |
3255 ** open savepoints. If the second parameter is greater than 0 and | |
3256 ** the sub-journal is not already open, then it will be opened here. | |
3257 */ | |
3258 rc = sqlite3PagerOpenSavepoint(pBt->pPager, p->db->nSavepoint); | |
3259 } | |
3260 | |
3261 btreeIntegrity(p); | |
3262 sqlite3BtreeLeave(p); | |
3263 return rc; | |
3264 } | |
3265 | |
3266 #ifndef SQLITE_OMIT_AUTOVACUUM | |
3267 | |
3268 /* | |
3269 ** Set the pointer-map entries for all children of page pPage. Also, if | |
3270 ** pPage contains cells that point to overflow pages, set the pointer | |
3271 ** map entries for the overflow pages as well. | |
3272 */ | |
3273 static int setChildPtrmaps(MemPage *pPage){ | |
3274 int i; /* Counter variable */ | |
3275 int nCell; /* Number of cells in page pPage */ | |
3276 int rc; /* Return code */ | |
3277 BtShared *pBt = pPage->pBt; | |
3278 u8 isInitOrig = pPage->isInit; | |
3279 Pgno pgno = pPage->pgno; | |
3280 | |
3281 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
3282 rc = btreeInitPage(pPage); | |
3283 if( rc!=SQLITE_OK ){ | |
3284 goto set_child_ptrmaps_out; | |
3285 } | |
3286 nCell = pPage->nCell; | |
3287 | |
3288 for(i=0; i<nCell; i++){ | |
3289 u8 *pCell = findCell(pPage, i); | |
3290 | |
3291 ptrmapPutOvflPtr(pPage, pCell, &rc); | |
3292 | |
3293 if( !pPage->leaf ){ | |
3294 Pgno childPgno = get4byte(pCell); | |
3295 ptrmapPut(pBt, childPgno, PTRMAP_BTREE, pgno, &rc); | |
3296 } | |
3297 } | |
3298 | |
3299 if( !pPage->leaf ){ | |
3300 Pgno childPgno = get4byte(&pPage->aData[pPage->hdrOffset+8]); | |
3301 ptrmapPut(pBt, childPgno, PTRMAP_BTREE, pgno, &rc); | |
3302 } | |
3303 | |
3304 set_child_ptrmaps_out: | |
3305 pPage->isInit = isInitOrig; | |
3306 return rc; | |
3307 } | |
3308 | |
3309 /* | |
3310 ** Somewhere on pPage is a pointer to page iFrom. Modify this pointer so | |
3311 ** that it points to iTo. Parameter eType describes the type of pointer to | |
3312 ** be modified, as follows: | |
3313 ** | |
3314 ** PTRMAP_BTREE: pPage is a btree-page. The pointer points at a child | |
3315 ** page of pPage. | |
3316 ** | |
3317 ** PTRMAP_OVERFLOW1: pPage is a btree-page. The pointer points at an overflow | |
3318 ** page pointed to by one of the cells on pPage. | |
3319 ** | |
3320 ** PTRMAP_OVERFLOW2: pPage is an overflow-page. The pointer points at the next | |
3321 ** overflow page in the list. | |
3322 */ | |
3323 static int modifyPagePointer(MemPage *pPage, Pgno iFrom, Pgno iTo, u8 eType){ | |
3324 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
3325 assert( sqlite3PagerIswriteable(pPage->pDbPage) ); | |
3326 if( eType==PTRMAP_OVERFLOW2 ){ | |
3327 /* The pointer is always the first 4 bytes of the page in this case. */ | |
3328 if( get4byte(pPage->aData)!=iFrom ){ | |
3329 return SQLITE_CORRUPT_BKPT; | |
3330 } | |
3331 put4byte(pPage->aData, iTo); | |
3332 }else{ | |
3333 u8 isInitOrig = pPage->isInit; | |
3334 int i; | |
3335 int nCell; | |
3336 int rc; | |
3337 | |
3338 rc = btreeInitPage(pPage); | |
3339 if( rc ) return rc; | |
3340 nCell = pPage->nCell; | |
3341 | |
3342 for(i=0; i<nCell; i++){ | |
3343 u8 *pCell = findCell(pPage, i); | |
3344 if( eType==PTRMAP_OVERFLOW1 ){ | |
3345 CellInfo info; | |
3346 pPage->xParseCell(pPage, pCell, &info); | |
3347 if( info.nLocal<info.nPayload | |
3348 && pCell+info.nSize-1<=pPage->aData+pPage->maskPage | |
3349 && iFrom==get4byte(pCell+info.nSize-4) | |
3350 ){ | |
3351 put4byte(pCell+info.nSize-4, iTo); | |
3352 break; | |
3353 } | |
3354 }else{ | |
3355 if( get4byte(pCell)==iFrom ){ | |
3356 put4byte(pCell, iTo); | |
3357 break; | |
3358 } | |
3359 } | |
3360 } | |
3361 | |
3362 if( i==nCell ){ | |
3363 if( eType!=PTRMAP_BTREE || | |
3364 get4byte(&pPage->aData[pPage->hdrOffset+8])!=iFrom ){ | |
3365 return SQLITE_CORRUPT_BKPT; | |
3366 } | |
3367 put4byte(&pPage->aData[pPage->hdrOffset+8], iTo); | |
3368 } | |
3369 | |
3370 pPage->isInit = isInitOrig; | |
3371 } | |
3372 return SQLITE_OK; | |
3373 } | |
3374 | |
3375 | |
3376 /* | |
3377 ** Move the open database page pDbPage to location iFreePage in the | |
3378 ** database. The pDbPage reference remains valid. | |
3379 ** | |
3380 ** The isCommit flag indicates that there is no need to remember that | |
3381 ** the journal needs to be sync()ed before database page pDbPage->pgno | |
3382 ** can be written to. The caller has already promised not to write to that | |
3383 ** page. | |
3384 */ | |
3385 static int relocatePage( | |
3386 BtShared *pBt, /* Btree */ | |
3387 MemPage *pDbPage, /* Open page to move */ | |
3388 u8 eType, /* Pointer map 'type' entry for pDbPage */ | |
3389 Pgno iPtrPage, /* Pointer map 'page-no' entry for pDbPage */ | |
3390 Pgno iFreePage, /* The location to move pDbPage to */ | |
3391 int isCommit /* isCommit flag passed to sqlite3PagerMovepage */ | |
3392 ){ | |
3393 MemPage *pPtrPage; /* The page that contains a pointer to pDbPage */ | |
3394 Pgno iDbPage = pDbPage->pgno; | |
3395 Pager *pPager = pBt->pPager; | |
3396 int rc; | |
3397 | |
3398 assert( eType==PTRMAP_OVERFLOW2 || eType==PTRMAP_OVERFLOW1 || | |
3399 eType==PTRMAP_BTREE || eType==PTRMAP_ROOTPAGE ); | |
3400 assert( sqlite3_mutex_held(pBt->mutex) ); | |
3401 assert( pDbPage->pBt==pBt ); | |
3402 | |
3403 /* Move page iDbPage from its current location to page number iFreePage */ | |
3404 TRACE(("AUTOVACUUM: Moving %d to free page %d (ptr page %d type %d)\n", | |
3405 iDbPage, iFreePage, iPtrPage, eType)); | |
3406 rc = sqlite3PagerMovepage(pPager, pDbPage->pDbPage, iFreePage, isCommit); | |
3407 if( rc!=SQLITE_OK ){ | |
3408 return rc; | |
3409 } | |
3410 pDbPage->pgno = iFreePage; | |
3411 | |
3412 /* If pDbPage was a btree-page, then it may have child pages and/or cells | |
3413 ** that point to overflow pages. The pointer map entries for all these | |
3414 ** pages need to be changed. | |
3415 ** | |
3416 ** If pDbPage is an overflow page, then the first 4 bytes may store a | |
3417 ** pointer to a subsequent overflow page. If this is the case, then | |
3418 ** the pointer map needs to be updated for the subsequent overflow page. | |
3419 */ | |
3420 if( eType==PTRMAP_BTREE || eType==PTRMAP_ROOTPAGE ){ | |
3421 rc = setChildPtrmaps(pDbPage); | |
3422 if( rc!=SQLITE_OK ){ | |
3423 return rc; | |
3424 } | |
3425 }else{ | |
3426 Pgno nextOvfl = get4byte(pDbPage->aData); | |
3427 if( nextOvfl!=0 ){ | |
3428 ptrmapPut(pBt, nextOvfl, PTRMAP_OVERFLOW2, iFreePage, &rc); | |
3429 if( rc!=SQLITE_OK ){ | |
3430 return rc; | |
3431 } | |
3432 } | |
3433 } | |
3434 | |
3435 /* Fix the database pointer on page iPtrPage that pointed at iDbPage so | |
3436 ** that it points at iFreePage. Also fix the pointer map entry for | |
3437 ** iPtrPage. | |
3438 */ | |
3439 if( eType!=PTRMAP_ROOTPAGE ){ | |
3440 rc = btreeGetPage(pBt, iPtrPage, &pPtrPage, 0); | |
3441 if( rc!=SQLITE_OK ){ | |
3442 return rc; | |
3443 } | |
3444 rc = sqlite3PagerWrite(pPtrPage->pDbPage); | |
3445 if( rc!=SQLITE_OK ){ | |
3446 releasePage(pPtrPage); | |
3447 return rc; | |
3448 } | |
3449 rc = modifyPagePointer(pPtrPage, iDbPage, iFreePage, eType); | |
3450 releasePage(pPtrPage); | |
3451 if( rc==SQLITE_OK ){ | |
3452 ptrmapPut(pBt, iFreePage, eType, iPtrPage, &rc); | |
3453 } | |
3454 } | |
3455 return rc; | |
3456 } | |
3457 | |
3458 /* Forward declaration required by incrVacuumStep(). */ | |
3459 static int allocateBtreePage(BtShared *, MemPage **, Pgno *, Pgno, u8); | |
3460 | |
3461 /* | |
3462 ** Perform a single step of an incremental-vacuum. If successful, return | |
3463 ** SQLITE_OK. If there is no work to do (and therefore no point in | |
3464 ** calling this function again), return SQLITE_DONE. Or, if an error | |
3465 ** occurs, return some other error code. | |
3466 ** | |
3467 ** More specifically, this function attempts to re-organize the database so | |
3468 ** that the last page of the file currently in use is no longer in use. | |
3469 ** | |
3470 ** Parameter nFin is the number of pages that this database would contain | |
3471 ** were this function called until it returns SQLITE_DONE. | |
3472 ** | |
3473 ** If the bCommit parameter is non-zero, this function assumes that the | |
3474 ** caller will keep calling incrVacuumStep() until it returns SQLITE_DONE | |
3475 ** or an error. bCommit is passed true for an auto-vacuum-on-commit | |
3476 ** operation, or false for an incremental vacuum. | |
3477 */ | |
3478 static int incrVacuumStep(BtShared *pBt, Pgno nFin, Pgno iLastPg, int bCommit){ | |
3479 Pgno nFreeList; /* Number of pages still on the free-list */ | |
3480 int rc; | |
3481 | |
3482 assert( sqlite3_mutex_held(pBt->mutex) ); | |
3483 assert( iLastPg>nFin ); | |
3484 | |
3485 if( !PTRMAP_ISPAGE(pBt, iLastPg) && iLastPg!=PENDING_BYTE_PAGE(pBt) ){ | |
3486 u8 eType; | |
3487 Pgno iPtrPage; | |
3488 | |
3489 nFreeList = get4byte(&pBt->pPage1->aData[36]); | |
3490 if( nFreeList==0 ){ | |
3491 return SQLITE_DONE; | |
3492 } | |
3493 | |
3494 rc = ptrmapGet(pBt, iLastPg, &eType, &iPtrPage); | |
3495 if( rc!=SQLITE_OK ){ | |
3496 return rc; | |
3497 } | |
3498 if( eType==PTRMAP_ROOTPAGE ){ | |
3499 return SQLITE_CORRUPT_BKPT; | |
3500 } | |
3501 | |
3502 if( eType==PTRMAP_FREEPAGE ){ | |
3503 if( bCommit==0 ){ | |
3504 /* Remove the page from the files free-list. This is not required | |
3505 ** if bCommit is non-zero. In that case, the free-list will be | |
3506 ** truncated to zero after this function returns, so it doesn't | |
3507 ** matter if it still contains some garbage entries. | |
3508 */ | |
3509 Pgno iFreePg; | |
3510 MemPage *pFreePg; | |
3511 rc = allocateBtreePage(pBt, &pFreePg, &iFreePg, iLastPg, BTALLOC_EXACT); | |
3512 if( rc!=SQLITE_OK ){ | |
3513 return rc; | |
3514 } | |
3515 assert( iFreePg==iLastPg ); | |
3516 releasePage(pFreePg); | |
3517 } | |
3518 } else { | |
3519 Pgno iFreePg; /* Index of free page to move pLastPg to */ | |
3520 MemPage *pLastPg; | |
3521 u8 eMode = BTALLOC_ANY; /* Mode parameter for allocateBtreePage() */ | |
3522 Pgno iNear = 0; /* nearby parameter for allocateBtreePage() */ | |
3523 | |
3524 rc = btreeGetPage(pBt, iLastPg, &pLastPg, 0); | |
3525 if( rc!=SQLITE_OK ){ | |
3526 return rc; | |
3527 } | |
3528 | |
3529 /* If bCommit is zero, this loop runs exactly once and page pLastPg | |
3530 ** is swapped with the first free page pulled off the free list. | |
3531 ** | |
3532 ** On the other hand, if bCommit is greater than zero, then keep | |
3533 ** looping until a free-page located within the first nFin pages | |
3534 ** of the file is found. | |
3535 */ | |
3536 if( bCommit==0 ){ | |
3537 eMode = BTALLOC_LE; | |
3538 iNear = nFin; | |
3539 } | |
3540 do { | |
3541 MemPage *pFreePg; | |
3542 rc = allocateBtreePage(pBt, &pFreePg, &iFreePg, iNear, eMode); | |
3543 if( rc!=SQLITE_OK ){ | |
3544 releasePage(pLastPg); | |
3545 return rc; | |
3546 } | |
3547 releasePage(pFreePg); | |
3548 }while( bCommit && iFreePg>nFin ); | |
3549 assert( iFreePg<iLastPg ); | |
3550 | |
3551 rc = relocatePage(pBt, pLastPg, eType, iPtrPage, iFreePg, bCommit); | |
3552 releasePage(pLastPg); | |
3553 if( rc!=SQLITE_OK ){ | |
3554 return rc; | |
3555 } | |
3556 } | |
3557 } | |
3558 | |
3559 if( bCommit==0 ){ | |
3560 do { | |
3561 iLastPg--; | |
3562 }while( iLastPg==PENDING_BYTE_PAGE(pBt) || PTRMAP_ISPAGE(pBt, iLastPg) ); | |
3563 pBt->bDoTruncate = 1; | |
3564 pBt->nPage = iLastPg; | |
3565 } | |
3566 return SQLITE_OK; | |
3567 } | |
3568 | |
3569 /* | |
3570 ** The database opened by the first argument is an auto-vacuum database | |
3571 ** nOrig pages in size containing nFree free pages. Return the expected | |
3572 ** size of the database in pages following an auto-vacuum operation. | |
3573 */ | |
3574 static Pgno finalDbSize(BtShared *pBt, Pgno nOrig, Pgno nFree){ | |
3575 int nEntry; /* Number of entries on one ptrmap page */ | |
3576 Pgno nPtrmap; /* Number of PtrMap pages to be freed */ | |
3577 Pgno nFin; /* Return value */ | |
3578 | |
3579 nEntry = pBt->usableSize/5; | |
3580 nPtrmap = (nFree-nOrig+PTRMAP_PAGENO(pBt, nOrig)+nEntry)/nEntry; | |
3581 nFin = nOrig - nFree - nPtrmap; | |
3582 if( nOrig>PENDING_BYTE_PAGE(pBt) && nFin<PENDING_BYTE_PAGE(pBt) ){ | |
3583 nFin--; | |
3584 } | |
3585 while( PTRMAP_ISPAGE(pBt, nFin) || nFin==PENDING_BYTE_PAGE(pBt) ){ | |
3586 nFin--; | |
3587 } | |
3588 | |
3589 return nFin; | |
3590 } | |
3591 | |
3592 /* | |
3593 ** A write-transaction must be opened before calling this function. | |
3594 ** It performs a single unit of work towards an incremental vacuum. | |
3595 ** | |
3596 ** If the incremental vacuum is finished after this function has run, | |
3597 ** SQLITE_DONE is returned. If it is not finished, but no error occurred, | |
3598 ** SQLITE_OK is returned. Otherwise an SQLite error code. | |
3599 */ | |
3600 int sqlite3BtreeIncrVacuum(Btree *p){ | |
3601 int rc; | |
3602 BtShared *pBt = p->pBt; | |
3603 | |
3604 sqlite3BtreeEnter(p); | |
3605 assert( pBt->inTransaction==TRANS_WRITE && p->inTrans==TRANS_WRITE ); | |
3606 if( !pBt->autoVacuum ){ | |
3607 rc = SQLITE_DONE; | |
3608 }else{ | |
3609 Pgno nOrig = btreePagecount(pBt); | |
3610 Pgno nFree = get4byte(&pBt->pPage1->aData[36]); | |
3611 Pgno nFin = finalDbSize(pBt, nOrig, nFree); | |
3612 | |
3613 if( nOrig<nFin ){ | |
3614 rc = SQLITE_CORRUPT_BKPT; | |
3615 }else if( nFree>0 ){ | |
3616 rc = saveAllCursors(pBt, 0, 0); | |
3617 if( rc==SQLITE_OK ){ | |
3618 invalidateAllOverflowCache(pBt); | |
3619 rc = incrVacuumStep(pBt, nFin, nOrig, 0); | |
3620 } | |
3621 if( rc==SQLITE_OK ){ | |
3622 rc = sqlite3PagerWrite(pBt->pPage1->pDbPage); | |
3623 put4byte(&pBt->pPage1->aData[28], pBt->nPage); | |
3624 } | |
3625 }else{ | |
3626 rc = SQLITE_DONE; | |
3627 } | |
3628 } | |
3629 sqlite3BtreeLeave(p); | |
3630 return rc; | |
3631 } | |
3632 | |
3633 /* | |
3634 ** This routine is called prior to sqlite3PagerCommit when a transaction | |
3635 ** is committed for an auto-vacuum database. | |
3636 ** | |
3637 ** If SQLITE_OK is returned, then *pnTrunc is set to the number of pages | |
3638 ** the database file should be truncated to during the commit process. | |
3639 ** i.e. the database has been reorganized so that only the first *pnTrunc | |
3640 ** pages are in use. | |
3641 */ | |
3642 static int autoVacuumCommit(BtShared *pBt){ | |
3643 int rc = SQLITE_OK; | |
3644 Pager *pPager = pBt->pPager; | |
3645 VVA_ONLY( int nRef = sqlite3PagerRefcount(pPager); ) | |
3646 | |
3647 assert( sqlite3_mutex_held(pBt->mutex) ); | |
3648 invalidateAllOverflowCache(pBt); | |
3649 assert(pBt->autoVacuum); | |
3650 if( !pBt->incrVacuum ){ | |
3651 Pgno nFin; /* Number of pages in database after autovacuuming */ | |
3652 Pgno nFree; /* Number of pages on the freelist initially */ | |
3653 Pgno iFree; /* The next page to be freed */ | |
3654 Pgno nOrig; /* Database size before freeing */ | |
3655 | |
3656 nOrig = btreePagecount(pBt); | |
3657 if( PTRMAP_ISPAGE(pBt, nOrig) || nOrig==PENDING_BYTE_PAGE(pBt) ){ | |
3658 /* It is not possible to create a database for which the final page | |
3659 ** is either a pointer-map page or the pending-byte page. If one | |
3660 ** is encountered, this indicates corruption. | |
3661 */ | |
3662 return SQLITE_CORRUPT_BKPT; | |
3663 } | |
3664 | |
3665 nFree = get4byte(&pBt->pPage1->aData[36]); | |
3666 nFin = finalDbSize(pBt, nOrig, nFree); | |
3667 if( nFin>nOrig ) return SQLITE_CORRUPT_BKPT; | |
3668 if( nFin<nOrig ){ | |
3669 rc = saveAllCursors(pBt, 0, 0); | |
3670 } | |
3671 for(iFree=nOrig; iFree>nFin && rc==SQLITE_OK; iFree--){ | |
3672 rc = incrVacuumStep(pBt, nFin, iFree, 1); | |
3673 } | |
3674 if( (rc==SQLITE_DONE || rc==SQLITE_OK) && nFree>0 ){ | |
3675 rc = sqlite3PagerWrite(pBt->pPage1->pDbPage); | |
3676 put4byte(&pBt->pPage1->aData[32], 0); | |
3677 put4byte(&pBt->pPage1->aData[36], 0); | |
3678 put4byte(&pBt->pPage1->aData[28], nFin); | |
3679 pBt->bDoTruncate = 1; | |
3680 pBt->nPage = nFin; | |
3681 } | |
3682 if( rc!=SQLITE_OK ){ | |
3683 sqlite3PagerRollback(pPager); | |
3684 } | |
3685 } | |
3686 | |
3687 assert( nRef>=sqlite3PagerRefcount(pPager) ); | |
3688 return rc; | |
3689 } | |
3690 | |
3691 #else /* ifndef SQLITE_OMIT_AUTOVACUUM */ | |
3692 # define setChildPtrmaps(x) SQLITE_OK | |
3693 #endif | |
3694 | |
3695 /* | |
3696 ** This routine does the first phase of a two-phase commit. This routine | |
3697 ** causes a rollback journal to be created (if it does not already exist) | |
3698 ** and populated with enough information so that if a power loss occurs | |
3699 ** the database can be restored to its original state by playing back | |
3700 ** the journal. Then the contents of the journal are flushed out to | |
3701 ** the disk. After the journal is safely on oxide, the changes to the | |
3702 ** database are written into the database file and flushed to oxide. | |
3703 ** At the end of this call, the rollback journal still exists on the | |
3704 ** disk and we are still holding all locks, so the transaction has not | |
3705 ** committed. See sqlite3BtreeCommitPhaseTwo() for the second phase of the | |
3706 ** commit process. | |
3707 ** | |
3708 ** This call is a no-op if no write-transaction is currently active on pBt. | |
3709 ** | |
3710 ** Otherwise, sync the database file for the btree pBt. zMaster points to | |
3711 ** the name of a master journal file that should be written into the | |
3712 ** individual journal file, or is NULL, indicating no master journal file | |
3713 ** (single database transaction). | |
3714 ** | |
3715 ** When this is called, the master journal should already have been | |
3716 ** created, populated with this journal pointer and synced to disk. | |
3717 ** | |
3718 ** Once this is routine has returned, the only thing required to commit | |
3719 ** the write-transaction for this database file is to delete the journal. | |
3720 */ | |
3721 int sqlite3BtreeCommitPhaseOne(Btree *p, const char *zMaster){ | |
3722 int rc = SQLITE_OK; | |
3723 if( p->inTrans==TRANS_WRITE ){ | |
3724 BtShared *pBt = p->pBt; | |
3725 sqlite3BtreeEnter(p); | |
3726 #ifndef SQLITE_OMIT_AUTOVACUUM | |
3727 if( pBt->autoVacuum ){ | |
3728 rc = autoVacuumCommit(pBt); | |
3729 if( rc!=SQLITE_OK ){ | |
3730 sqlite3BtreeLeave(p); | |
3731 return rc; | |
3732 } | |
3733 } | |
3734 if( pBt->bDoTruncate ){ | |
3735 sqlite3PagerTruncateImage(pBt->pPager, pBt->nPage); | |
3736 } | |
3737 #endif | |
3738 rc = sqlite3PagerCommitPhaseOne(pBt->pPager, zMaster, 0); | |
3739 sqlite3BtreeLeave(p); | |
3740 } | |
3741 return rc; | |
3742 } | |
3743 | |
3744 /* | |
3745 ** This function is called from both BtreeCommitPhaseTwo() and BtreeRollback() | |
3746 ** at the conclusion of a transaction. | |
3747 */ | |
3748 static void btreeEndTransaction(Btree *p){ | |
3749 BtShared *pBt = p->pBt; | |
3750 sqlite3 *db = p->db; | |
3751 assert( sqlite3BtreeHoldsMutex(p) ); | |
3752 | |
3753 #ifndef SQLITE_OMIT_AUTOVACUUM | |
3754 pBt->bDoTruncate = 0; | |
3755 #endif | |
3756 if( p->inTrans>TRANS_NONE && db->nVdbeRead>1 ){ | |
3757 /* If there are other active statements that belong to this database | |
3758 ** handle, downgrade to a read-only transaction. The other statements | |
3759 ** may still be reading from the database. */ | |
3760 downgradeAllSharedCacheTableLocks(p); | |
3761 p->inTrans = TRANS_READ; | |
3762 }else{ | |
3763 /* If the handle had any kind of transaction open, decrement the | |
3764 ** transaction count of the shared btree. If the transaction count | |
3765 ** reaches 0, set the shared state to TRANS_NONE. The unlockBtreeIfUnused() | |
3766 ** call below will unlock the pager. */ | |
3767 if( p->inTrans!=TRANS_NONE ){ | |
3768 clearAllSharedCacheTableLocks(p); | |
3769 pBt->nTransaction--; | |
3770 if( 0==pBt->nTransaction ){ | |
3771 pBt->inTransaction = TRANS_NONE; | |
3772 } | |
3773 } | |
3774 | |
3775 /* Set the current transaction state to TRANS_NONE and unlock the | |
3776 ** pager if this call closed the only read or write transaction. */ | |
3777 p->inTrans = TRANS_NONE; | |
3778 unlockBtreeIfUnused(pBt); | |
3779 } | |
3780 | |
3781 btreeIntegrity(p); | |
3782 } | |
3783 | |
3784 /* | |
3785 ** Commit the transaction currently in progress. | |
3786 ** | |
3787 ** This routine implements the second phase of a 2-phase commit. The | |
3788 ** sqlite3BtreeCommitPhaseOne() routine does the first phase and should | |
3789 ** be invoked prior to calling this routine. The sqlite3BtreeCommitPhaseOne() | |
3790 ** routine did all the work of writing information out to disk and flushing the | |
3791 ** contents so that they are written onto the disk platter. All this | |
3792 ** routine has to do is delete or truncate or zero the header in the | |
3793 ** the rollback journal (which causes the transaction to commit) and | |
3794 ** drop locks. | |
3795 ** | |
3796 ** Normally, if an error occurs while the pager layer is attempting to | |
3797 ** finalize the underlying journal file, this function returns an error and | |
3798 ** the upper layer will attempt a rollback. However, if the second argument | |
3799 ** is non-zero then this b-tree transaction is part of a multi-file | |
3800 ** transaction. In this case, the transaction has already been committed | |
3801 ** (by deleting a master journal file) and the caller will ignore this | |
3802 ** functions return code. So, even if an error occurs in the pager layer, | |
3803 ** reset the b-tree objects internal state to indicate that the write | |
3804 ** transaction has been closed. This is quite safe, as the pager will have | |
3805 ** transitioned to the error state. | |
3806 ** | |
3807 ** This will release the write lock on the database file. If there | |
3808 ** are no active cursors, it also releases the read lock. | |
3809 */ | |
3810 int sqlite3BtreeCommitPhaseTwo(Btree *p, int bCleanup){ | |
3811 | |
3812 if( p->inTrans==TRANS_NONE ) return SQLITE_OK; | |
3813 sqlite3BtreeEnter(p); | |
3814 btreeIntegrity(p); | |
3815 | |
3816 /* If the handle has a write-transaction open, commit the shared-btrees | |
3817 ** transaction and set the shared state to TRANS_READ. | |
3818 */ | |
3819 if( p->inTrans==TRANS_WRITE ){ | |
3820 int rc; | |
3821 BtShared *pBt = p->pBt; | |
3822 assert( pBt->inTransaction==TRANS_WRITE ); | |
3823 assert( pBt->nTransaction>0 ); | |
3824 rc = sqlite3PagerCommitPhaseTwo(pBt->pPager); | |
3825 if( rc!=SQLITE_OK && bCleanup==0 ){ | |
3826 sqlite3BtreeLeave(p); | |
3827 return rc; | |
3828 } | |
3829 p->iDataVersion--; /* Compensate for pPager->iDataVersion++; */ | |
3830 pBt->inTransaction = TRANS_READ; | |
3831 btreeClearHasContent(pBt); | |
3832 } | |
3833 | |
3834 btreeEndTransaction(p); | |
3835 sqlite3BtreeLeave(p); | |
3836 return SQLITE_OK; | |
3837 } | |
3838 | |
3839 /* | |
3840 ** Do both phases of a commit. | |
3841 */ | |
3842 int sqlite3BtreeCommit(Btree *p){ | |
3843 int rc; | |
3844 sqlite3BtreeEnter(p); | |
3845 rc = sqlite3BtreeCommitPhaseOne(p, 0); | |
3846 if( rc==SQLITE_OK ){ | |
3847 rc = sqlite3BtreeCommitPhaseTwo(p, 0); | |
3848 } | |
3849 sqlite3BtreeLeave(p); | |
3850 return rc; | |
3851 } | |
3852 | |
3853 /* | |
3854 ** This routine sets the state to CURSOR_FAULT and the error | |
3855 ** code to errCode for every cursor on any BtShared that pBtree | |
3856 ** references. Or if the writeOnly flag is set to 1, then only | |
3857 ** trip write cursors and leave read cursors unchanged. | |
3858 ** | |
3859 ** Every cursor is a candidate to be tripped, including cursors | |
3860 ** that belong to other database connections that happen to be | |
3861 ** sharing the cache with pBtree. | |
3862 ** | |
3863 ** This routine gets called when a rollback occurs. If the writeOnly | |
3864 ** flag is true, then only write-cursors need be tripped - read-only | |
3865 ** cursors save their current positions so that they may continue | |
3866 ** following the rollback. Or, if writeOnly is false, all cursors are | |
3867 ** tripped. In general, writeOnly is false if the transaction being | |
3868 ** rolled back modified the database schema. In this case b-tree root | |
3869 ** pages may be moved or deleted from the database altogether, making | |
3870 ** it unsafe for read cursors to continue. | |
3871 ** | |
3872 ** If the writeOnly flag is true and an error is encountered while | |
3873 ** saving the current position of a read-only cursor, all cursors, | |
3874 ** including all read-cursors are tripped. | |
3875 ** | |
3876 ** SQLITE_OK is returned if successful, or if an error occurs while | |
3877 ** saving a cursor position, an SQLite error code. | |
3878 */ | |
3879 int sqlite3BtreeTripAllCursors(Btree *pBtree, int errCode, int writeOnly){ | |
3880 BtCursor *p; | |
3881 int rc = SQLITE_OK; | |
3882 | |
3883 assert( (writeOnly==0 || writeOnly==1) && BTCF_WriteFlag==1 ); | |
3884 if( pBtree ){ | |
3885 sqlite3BtreeEnter(pBtree); | |
3886 for(p=pBtree->pBt->pCursor; p; p=p->pNext){ | |
3887 int i; | |
3888 if( writeOnly && (p->curFlags & BTCF_WriteFlag)==0 ){ | |
3889 if( p->eState==CURSOR_VALID || p->eState==CURSOR_SKIPNEXT ){ | |
3890 rc = saveCursorPosition(p); | |
3891 if( rc!=SQLITE_OK ){ | |
3892 (void)sqlite3BtreeTripAllCursors(pBtree, rc, 0); | |
3893 break; | |
3894 } | |
3895 } | |
3896 }else{ | |
3897 sqlite3BtreeClearCursor(p); | |
3898 p->eState = CURSOR_FAULT; | |
3899 p->skipNext = errCode; | |
3900 } | |
3901 for(i=0; i<=p->iPage; i++){ | |
3902 releasePage(p->apPage[i]); | |
3903 p->apPage[i] = 0; | |
3904 } | |
3905 } | |
3906 sqlite3BtreeLeave(pBtree); | |
3907 } | |
3908 return rc; | |
3909 } | |
3910 | |
3911 /* | |
3912 ** Rollback the transaction in progress. | |
3913 ** | |
3914 ** If tripCode is not SQLITE_OK then cursors will be invalidated (tripped). | |
3915 ** Only write cursors are tripped if writeOnly is true but all cursors are | |
3916 ** tripped if writeOnly is false. Any attempt to use | |
3917 ** a tripped cursor will result in an error. | |
3918 ** | |
3919 ** This will release the write lock on the database file. If there | |
3920 ** are no active cursors, it also releases the read lock. | |
3921 */ | |
3922 int sqlite3BtreeRollback(Btree *p, int tripCode, int writeOnly){ | |
3923 int rc; | |
3924 BtShared *pBt = p->pBt; | |
3925 MemPage *pPage1; | |
3926 | |
3927 assert( writeOnly==1 || writeOnly==0 ); | |
3928 assert( tripCode==SQLITE_ABORT_ROLLBACK || tripCode==SQLITE_OK ); | |
3929 sqlite3BtreeEnter(p); | |
3930 if( tripCode==SQLITE_OK ){ | |
3931 rc = tripCode = saveAllCursors(pBt, 0, 0); | |
3932 if( rc ) writeOnly = 0; | |
3933 }else{ | |
3934 rc = SQLITE_OK; | |
3935 } | |
3936 if( tripCode ){ | |
3937 int rc2 = sqlite3BtreeTripAllCursors(p, tripCode, writeOnly); | |
3938 assert( rc==SQLITE_OK || (writeOnly==0 && rc2==SQLITE_OK) ); | |
3939 if( rc2!=SQLITE_OK ) rc = rc2; | |
3940 } | |
3941 btreeIntegrity(p); | |
3942 | |
3943 if( p->inTrans==TRANS_WRITE ){ | |
3944 int rc2; | |
3945 | |
3946 assert( TRANS_WRITE==pBt->inTransaction ); | |
3947 rc2 = sqlite3PagerRollback(pBt->pPager); | |
3948 if( rc2!=SQLITE_OK ){ | |
3949 rc = rc2; | |
3950 } | |
3951 | |
3952 /* The rollback may have destroyed the pPage1->aData value. So | |
3953 ** call btreeGetPage() on page 1 again to make | |
3954 ** sure pPage1->aData is set correctly. */ | |
3955 if( btreeGetPage(pBt, 1, &pPage1, 0)==SQLITE_OK ){ | |
3956 int nPage = get4byte(28+(u8*)pPage1->aData); | |
3957 testcase( nPage==0 ); | |
3958 if( nPage==0 ) sqlite3PagerPagecount(pBt->pPager, &nPage); | |
3959 testcase( pBt->nPage!=nPage ); | |
3960 pBt->nPage = nPage; | |
3961 releasePage(pPage1); | |
3962 } | |
3963 assert( countValidCursors(pBt, 1)==0 ); | |
3964 pBt->inTransaction = TRANS_READ; | |
3965 btreeClearHasContent(pBt); | |
3966 } | |
3967 | |
3968 btreeEndTransaction(p); | |
3969 sqlite3BtreeLeave(p); | |
3970 return rc; | |
3971 } | |
3972 | |
3973 /* | |
3974 ** Start a statement subtransaction. The subtransaction can be rolled | |
3975 ** back independently of the main transaction. You must start a transaction | |
3976 ** before starting a subtransaction. The subtransaction is ended automatically | |
3977 ** if the main transaction commits or rolls back. | |
3978 ** | |
3979 ** Statement subtransactions are used around individual SQL statements | |
3980 ** that are contained within a BEGIN...COMMIT block. If a constraint | |
3981 ** error occurs within the statement, the effect of that one statement | |
3982 ** can be rolled back without having to rollback the entire transaction. | |
3983 ** | |
3984 ** A statement sub-transaction is implemented as an anonymous savepoint. The | |
3985 ** value passed as the second parameter is the total number of savepoints, | |
3986 ** including the new anonymous savepoint, open on the B-Tree. i.e. if there | |
3987 ** are no active savepoints and no other statement-transactions open, | |
3988 ** iStatement is 1. This anonymous savepoint can be released or rolled back | |
3989 ** using the sqlite3BtreeSavepoint() function. | |
3990 */ | |
3991 int sqlite3BtreeBeginStmt(Btree *p, int iStatement){ | |
3992 int rc; | |
3993 BtShared *pBt = p->pBt; | |
3994 sqlite3BtreeEnter(p); | |
3995 assert( p->inTrans==TRANS_WRITE ); | |
3996 assert( (pBt->btsFlags & BTS_READ_ONLY)==0 ); | |
3997 assert( iStatement>0 ); | |
3998 assert( iStatement>p->db->nSavepoint ); | |
3999 assert( pBt->inTransaction==TRANS_WRITE ); | |
4000 /* At the pager level, a statement transaction is a savepoint with | |
4001 ** an index greater than all savepoints created explicitly using | |
4002 ** SQL statements. It is illegal to open, release or rollback any | |
4003 ** such savepoints while the statement transaction savepoint is active. | |
4004 */ | |
4005 rc = sqlite3PagerOpenSavepoint(pBt->pPager, iStatement); | |
4006 sqlite3BtreeLeave(p); | |
4007 return rc; | |
4008 } | |
4009 | |
4010 /* | |
4011 ** The second argument to this function, op, is always SAVEPOINT_ROLLBACK | |
4012 ** or SAVEPOINT_RELEASE. This function either releases or rolls back the | |
4013 ** savepoint identified by parameter iSavepoint, depending on the value | |
4014 ** of op. | |
4015 ** | |
4016 ** Normally, iSavepoint is greater than or equal to zero. However, if op is | |
4017 ** SAVEPOINT_ROLLBACK, then iSavepoint may also be -1. In this case the | |
4018 ** contents of the entire transaction are rolled back. This is different | |
4019 ** from a normal transaction rollback, as no locks are released and the | |
4020 ** transaction remains open. | |
4021 */ | |
4022 int sqlite3BtreeSavepoint(Btree *p, int op, int iSavepoint){ | |
4023 int rc = SQLITE_OK; | |
4024 if( p && p->inTrans==TRANS_WRITE ){ | |
4025 BtShared *pBt = p->pBt; | |
4026 assert( op==SAVEPOINT_RELEASE || op==SAVEPOINT_ROLLBACK ); | |
4027 assert( iSavepoint>=0 || (iSavepoint==-1 && op==SAVEPOINT_ROLLBACK) ); | |
4028 sqlite3BtreeEnter(p); | |
4029 rc = sqlite3PagerSavepoint(pBt->pPager, op, iSavepoint); | |
4030 if( rc==SQLITE_OK ){ | |
4031 if( iSavepoint<0 && (pBt->btsFlags & BTS_INITIALLY_EMPTY)!=0 ){ | |
4032 pBt->nPage = 0; | |
4033 } | |
4034 rc = newDatabase(pBt); | |
4035 pBt->nPage = get4byte(28 + pBt->pPage1->aData); | |
4036 | |
4037 /* The database size was written into the offset 28 of the header | |
4038 ** when the transaction started, so we know that the value at offset | |
4039 ** 28 is nonzero. */ | |
4040 assert( pBt->nPage>0 ); | |
4041 } | |
4042 sqlite3BtreeLeave(p); | |
4043 } | |
4044 return rc; | |
4045 } | |
4046 | |
4047 /* | |
4048 ** Create a new cursor for the BTree whose root is on the page | |
4049 ** iTable. If a read-only cursor is requested, it is assumed that | |
4050 ** the caller already has at least a read-only transaction open | |
4051 ** on the database already. If a write-cursor is requested, then | |
4052 ** the caller is assumed to have an open write transaction. | |
4053 ** | |
4054 ** If wrFlag==0, then the cursor can only be used for reading. | |
4055 ** If wrFlag==1, then the cursor can be used for reading or for | |
4056 ** writing if other conditions for writing are also met. These | |
4057 ** are the conditions that must be met in order for writing to | |
4058 ** be allowed: | |
4059 ** | |
4060 ** 1: The cursor must have been opened with wrFlag==1 | |
4061 ** | |
4062 ** 2: Other database connections that share the same pager cache | |
4063 ** but which are not in the READ_UNCOMMITTED state may not have | |
4064 ** cursors open with wrFlag==0 on the same table. Otherwise | |
4065 ** the changes made by this write cursor would be visible to | |
4066 ** the read cursors in the other database connection. | |
4067 ** | |
4068 ** 3: The database must be writable (not on read-only media) | |
4069 ** | |
4070 ** 4: There must be an active transaction. | |
4071 ** | |
4072 ** No checking is done to make sure that page iTable really is the | |
4073 ** root page of a b-tree. If it is not, then the cursor acquired | |
4074 ** will not work correctly. | |
4075 ** | |
4076 ** It is assumed that the sqlite3BtreeCursorZero() has been called | |
4077 ** on pCur to initialize the memory space prior to invoking this routine. | |
4078 */ | |
4079 static int btreeCursor( | |
4080 Btree *p, /* The btree */ | |
4081 int iTable, /* Root page of table to open */ | |
4082 int wrFlag, /* 1 to write. 0 read-only */ | |
4083 struct KeyInfo *pKeyInfo, /* First arg to comparison function */ | |
4084 BtCursor *pCur /* Space for new cursor */ | |
4085 ){ | |
4086 BtShared *pBt = p->pBt; /* Shared b-tree handle */ | |
4087 BtCursor *pX; /* Looping over other all cursors */ | |
4088 | |
4089 assert( sqlite3BtreeHoldsMutex(p) ); | |
4090 assert( wrFlag==0 | |
4091 || wrFlag==BTREE_WRCSR | |
4092 || wrFlag==(BTREE_WRCSR|BTREE_FORDELETE) | |
4093 ); | |
4094 | |
4095 /* The following assert statements verify that if this is a sharable | |
4096 ** b-tree database, the connection is holding the required table locks, | |
4097 ** and that no other connection has any open cursor that conflicts with | |
4098 ** this lock. */ | |
4099 assert( hasSharedCacheTableLock(p, iTable, pKeyInfo!=0, (wrFlag?2:1)) ); | |
4100 assert( wrFlag==0 || !hasReadConflicts(p, iTable) ); | |
4101 | |
4102 /* Assert that the caller has opened the required transaction. */ | |
4103 assert( p->inTrans>TRANS_NONE ); | |
4104 assert( wrFlag==0 || p->inTrans==TRANS_WRITE ); | |
4105 assert( pBt->pPage1 && pBt->pPage1->aData ); | |
4106 assert( wrFlag==0 || (pBt->btsFlags & BTS_READ_ONLY)==0 ); | |
4107 | |
4108 if( wrFlag ){ | |
4109 allocateTempSpace(pBt); | |
4110 if( pBt->pTmpSpace==0 ) return SQLITE_NOMEM; | |
4111 } | |
4112 if( iTable==1 && btreePagecount(pBt)==0 ){ | |
4113 assert( wrFlag==0 ); | |
4114 iTable = 0; | |
4115 } | |
4116 | |
4117 /* Now that no other errors can occur, finish filling in the BtCursor | |
4118 ** variables and link the cursor into the BtShared list. */ | |
4119 pCur->pgnoRoot = (Pgno)iTable; | |
4120 pCur->iPage = -1; | |
4121 pCur->pKeyInfo = pKeyInfo; | |
4122 pCur->pBtree = p; | |
4123 pCur->pBt = pBt; | |
4124 pCur->curFlags = wrFlag ? BTCF_WriteFlag : 0; | |
4125 pCur->curPagerFlags = wrFlag ? 0 : PAGER_GET_READONLY; | |
4126 /* If there are two or more cursors on the same btree, then all such | |
4127 ** cursors *must* have the BTCF_Multiple flag set. */ | |
4128 for(pX=pBt->pCursor; pX; pX=pX->pNext){ | |
4129 if( pX->pgnoRoot==(Pgno)iTable ){ | |
4130 pX->curFlags |= BTCF_Multiple; | |
4131 pCur->curFlags |= BTCF_Multiple; | |
4132 } | |
4133 } | |
4134 pCur->pNext = pBt->pCursor; | |
4135 pBt->pCursor = pCur; | |
4136 pCur->eState = CURSOR_INVALID; | |
4137 return SQLITE_OK; | |
4138 } | |
4139 int sqlite3BtreeCursor( | |
4140 Btree *p, /* The btree */ | |
4141 int iTable, /* Root page of table to open */ | |
4142 int wrFlag, /* 1 to write. 0 read-only */ | |
4143 struct KeyInfo *pKeyInfo, /* First arg to xCompare() */ | |
4144 BtCursor *pCur /* Write new cursor here */ | |
4145 ){ | |
4146 int rc; | |
4147 if( iTable<1 ){ | |
4148 rc = SQLITE_CORRUPT_BKPT; | |
4149 }else{ | |
4150 sqlite3BtreeEnter(p); | |
4151 rc = btreeCursor(p, iTable, wrFlag, pKeyInfo, pCur); | |
4152 sqlite3BtreeLeave(p); | |
4153 } | |
4154 return rc; | |
4155 } | |
4156 | |
4157 /* | |
4158 ** Return the size of a BtCursor object in bytes. | |
4159 ** | |
4160 ** This interfaces is needed so that users of cursors can preallocate | |
4161 ** sufficient storage to hold a cursor. The BtCursor object is opaque | |
4162 ** to users so they cannot do the sizeof() themselves - they must call | |
4163 ** this routine. | |
4164 */ | |
4165 int sqlite3BtreeCursorSize(void){ | |
4166 return ROUND8(sizeof(BtCursor)); | |
4167 } | |
4168 | |
4169 /* | |
4170 ** Initialize memory that will be converted into a BtCursor object. | |
4171 ** | |
4172 ** The simple approach here would be to memset() the entire object | |
4173 ** to zero. But it turns out that the apPage[] and aiIdx[] arrays | |
4174 ** do not need to be zeroed and they are large, so we can save a lot | |
4175 ** of run-time by skipping the initialization of those elements. | |
4176 */ | |
4177 void sqlite3BtreeCursorZero(BtCursor *p){ | |
4178 memset(p, 0, offsetof(BtCursor, iPage)); | |
4179 } | |
4180 | |
4181 /* | |
4182 ** Close a cursor. The read lock on the database file is released | |
4183 ** when the last cursor is closed. | |
4184 */ | |
4185 int sqlite3BtreeCloseCursor(BtCursor *pCur){ | |
4186 Btree *pBtree = pCur->pBtree; | |
4187 if( pBtree ){ | |
4188 int i; | |
4189 BtShared *pBt = pCur->pBt; | |
4190 sqlite3BtreeEnter(pBtree); | |
4191 sqlite3BtreeClearCursor(pCur); | |
4192 assert( pBt->pCursor!=0 ); | |
4193 if( pBt->pCursor==pCur ){ | |
4194 pBt->pCursor = pCur->pNext; | |
4195 }else{ | |
4196 BtCursor *pPrev = pBt->pCursor; | |
4197 do{ | |
4198 if( pPrev->pNext==pCur ){ | |
4199 pPrev->pNext = pCur->pNext; | |
4200 break; | |
4201 } | |
4202 pPrev = pPrev->pNext; | |
4203 }while( ALWAYS(pPrev) ); | |
4204 } | |
4205 for(i=0; i<=pCur->iPage; i++){ | |
4206 releasePage(pCur->apPage[i]); | |
4207 } | |
4208 unlockBtreeIfUnused(pBt); | |
4209 sqlite3_free(pCur->aOverflow); | |
4210 /* sqlite3_free(pCur); */ | |
4211 sqlite3BtreeLeave(pBtree); | |
4212 } | |
4213 return SQLITE_OK; | |
4214 } | |
4215 | |
4216 /* | |
4217 ** Make sure the BtCursor* given in the argument has a valid | |
4218 ** BtCursor.info structure. If it is not already valid, call | |
4219 ** btreeParseCell() to fill it in. | |
4220 ** | |
4221 ** BtCursor.info is a cache of the information in the current cell. | |
4222 ** Using this cache reduces the number of calls to btreeParseCell(). | |
4223 */ | |
4224 #ifndef NDEBUG | |
4225 static void assertCellInfo(BtCursor *pCur){ | |
4226 CellInfo info; | |
4227 int iPage = pCur->iPage; | |
4228 memset(&info, 0, sizeof(info)); | |
4229 btreeParseCell(pCur->apPage[iPage], pCur->aiIdx[iPage], &info); | |
4230 assert( CORRUPT_DB || memcmp(&info, &pCur->info, sizeof(info))==0 ); | |
4231 } | |
4232 #else | |
4233 #define assertCellInfo(x) | |
4234 #endif | |
4235 static SQLITE_NOINLINE void getCellInfo(BtCursor *pCur){ | |
4236 if( pCur->info.nSize==0 ){ | |
4237 int iPage = pCur->iPage; | |
4238 pCur->curFlags |= BTCF_ValidNKey; | |
4239 btreeParseCell(pCur->apPage[iPage],pCur->aiIdx[iPage],&pCur->info); | |
4240 }else{ | |
4241 assertCellInfo(pCur); | |
4242 } | |
4243 } | |
4244 | |
4245 #ifndef NDEBUG /* The next routine used only within assert() statements */ | |
4246 /* | |
4247 ** Return true if the given BtCursor is valid. A valid cursor is one | |
4248 ** that is currently pointing to a row in a (non-empty) table. | |
4249 ** This is a verification routine is used only within assert() statements. | |
4250 */ | |
4251 int sqlite3BtreeCursorIsValid(BtCursor *pCur){ | |
4252 return pCur && pCur->eState==CURSOR_VALID; | |
4253 } | |
4254 #endif /* NDEBUG */ | |
4255 | |
4256 /* | |
4257 ** Set *pSize to the size of the buffer needed to hold the value of | |
4258 ** the key for the current entry. If the cursor is not pointing | |
4259 ** to a valid entry, *pSize is set to 0. | |
4260 ** | |
4261 ** For a table with the INTKEY flag set, this routine returns the key | |
4262 ** itself, not the number of bytes in the key. | |
4263 ** | |
4264 ** The caller must position the cursor prior to invoking this routine. | |
4265 ** | |
4266 ** This routine cannot fail. It always returns SQLITE_OK. | |
4267 */ | |
4268 int sqlite3BtreeKeySize(BtCursor *pCur, i64 *pSize){ | |
4269 assert( cursorHoldsMutex(pCur) ); | |
4270 assert( pCur->eState==CURSOR_VALID ); | |
4271 getCellInfo(pCur); | |
4272 *pSize = pCur->info.nKey; | |
4273 return SQLITE_OK; | |
4274 } | |
4275 | |
4276 /* | |
4277 ** Set *pSize to the number of bytes of data in the entry the | |
4278 ** cursor currently points to. | |
4279 ** | |
4280 ** The caller must guarantee that the cursor is pointing to a non-NULL | |
4281 ** valid entry. In other words, the calling procedure must guarantee | |
4282 ** that the cursor has Cursor.eState==CURSOR_VALID. | |
4283 ** | |
4284 ** Failure is not possible. This function always returns SQLITE_OK. | |
4285 ** It might just as well be a procedure (returning void) but we continue | |
4286 ** to return an integer result code for historical reasons. | |
4287 */ | |
4288 int sqlite3BtreeDataSize(BtCursor *pCur, u32 *pSize){ | |
4289 assert( cursorHoldsMutex(pCur) ); | |
4290 assert( pCur->eState==CURSOR_VALID ); | |
4291 assert( pCur->iPage>=0 ); | |
4292 assert( pCur->iPage<BTCURSOR_MAX_DEPTH ); | |
4293 assert( pCur->apPage[pCur->iPage]->intKeyLeaf==1 ); | |
4294 getCellInfo(pCur); | |
4295 *pSize = pCur->info.nPayload; | |
4296 return SQLITE_OK; | |
4297 } | |
4298 | |
4299 /* | |
4300 ** Given the page number of an overflow page in the database (parameter | |
4301 ** ovfl), this function finds the page number of the next page in the | |
4302 ** linked list of overflow pages. If possible, it uses the auto-vacuum | |
4303 ** pointer-map data instead of reading the content of page ovfl to do so. | |
4304 ** | |
4305 ** If an error occurs an SQLite error code is returned. Otherwise: | |
4306 ** | |
4307 ** The page number of the next overflow page in the linked list is | |
4308 ** written to *pPgnoNext. If page ovfl is the last page in its linked | |
4309 ** list, *pPgnoNext is set to zero. | |
4310 ** | |
4311 ** If ppPage is not NULL, and a reference to the MemPage object corresponding | |
4312 ** to page number pOvfl was obtained, then *ppPage is set to point to that | |
4313 ** reference. It is the responsibility of the caller to call releasePage() | |
4314 ** on *ppPage to free the reference. In no reference was obtained (because | |
4315 ** the pointer-map was used to obtain the value for *pPgnoNext), then | |
4316 ** *ppPage is set to zero. | |
4317 */ | |
4318 static int getOverflowPage( | |
4319 BtShared *pBt, /* The database file */ | |
4320 Pgno ovfl, /* Current overflow page number */ | |
4321 MemPage **ppPage, /* OUT: MemPage handle (may be NULL) */ | |
4322 Pgno *pPgnoNext /* OUT: Next overflow page number */ | |
4323 ){ | |
4324 Pgno next = 0; | |
4325 MemPage *pPage = 0; | |
4326 int rc = SQLITE_OK; | |
4327 | |
4328 assert( sqlite3_mutex_held(pBt->mutex) ); | |
4329 assert(pPgnoNext); | |
4330 | |
4331 #ifndef SQLITE_OMIT_AUTOVACUUM | |
4332 /* Try to find the next page in the overflow list using the | |
4333 ** autovacuum pointer-map pages. Guess that the next page in | |
4334 ** the overflow list is page number (ovfl+1). If that guess turns | |
4335 ** out to be wrong, fall back to loading the data of page | |
4336 ** number ovfl to determine the next page number. | |
4337 */ | |
4338 if( pBt->autoVacuum ){ | |
4339 Pgno pgno; | |
4340 Pgno iGuess = ovfl+1; | |
4341 u8 eType; | |
4342 | |
4343 while( PTRMAP_ISPAGE(pBt, iGuess) || iGuess==PENDING_BYTE_PAGE(pBt) ){ | |
4344 iGuess++; | |
4345 } | |
4346 | |
4347 if( iGuess<=btreePagecount(pBt) ){ | |
4348 rc = ptrmapGet(pBt, iGuess, &eType, &pgno); | |
4349 if( rc==SQLITE_OK && eType==PTRMAP_OVERFLOW2 && pgno==ovfl ){ | |
4350 next = iGuess; | |
4351 rc = SQLITE_DONE; | |
4352 } | |
4353 } | |
4354 } | |
4355 #endif | |
4356 | |
4357 assert( next==0 || rc==SQLITE_DONE ); | |
4358 if( rc==SQLITE_OK ){ | |
4359 rc = btreeGetPage(pBt, ovfl, &pPage, (ppPage==0) ? PAGER_GET_READONLY : 0); | |
4360 assert( rc==SQLITE_OK || pPage==0 ); | |
4361 if( rc==SQLITE_OK ){ | |
4362 next = get4byte(pPage->aData); | |
4363 } | |
4364 } | |
4365 | |
4366 *pPgnoNext = next; | |
4367 if( ppPage ){ | |
4368 *ppPage = pPage; | |
4369 }else{ | |
4370 releasePage(pPage); | |
4371 } | |
4372 return (rc==SQLITE_DONE ? SQLITE_OK : rc); | |
4373 } | |
4374 | |
4375 /* | |
4376 ** Copy data from a buffer to a page, or from a page to a buffer. | |
4377 ** | |
4378 ** pPayload is a pointer to data stored on database page pDbPage. | |
4379 ** If argument eOp is false, then nByte bytes of data are copied | |
4380 ** from pPayload to the buffer pointed at by pBuf. If eOp is true, | |
4381 ** then sqlite3PagerWrite() is called on pDbPage and nByte bytes | |
4382 ** of data are copied from the buffer pBuf to pPayload. | |
4383 ** | |
4384 ** SQLITE_OK is returned on success, otherwise an error code. | |
4385 */ | |
4386 static int copyPayload( | |
4387 void *pPayload, /* Pointer to page data */ | |
4388 void *pBuf, /* Pointer to buffer */ | |
4389 int nByte, /* Number of bytes to copy */ | |
4390 int eOp, /* 0 -> copy from page, 1 -> copy to page */ | |
4391 DbPage *pDbPage /* Page containing pPayload */ | |
4392 ){ | |
4393 if( eOp ){ | |
4394 /* Copy data from buffer to page (a write operation) */ | |
4395 int rc = sqlite3PagerWrite(pDbPage); | |
4396 if( rc!=SQLITE_OK ){ | |
4397 return rc; | |
4398 } | |
4399 memcpy(pPayload, pBuf, nByte); | |
4400 }else{ | |
4401 /* Copy data from page to buffer (a read operation) */ | |
4402 memcpy(pBuf, pPayload, nByte); | |
4403 } | |
4404 return SQLITE_OK; | |
4405 } | |
4406 | |
4407 /* | |
4408 ** This function is used to read or overwrite payload information | |
4409 ** for the entry that the pCur cursor is pointing to. The eOp | |
4410 ** argument is interpreted as follows: | |
4411 ** | |
4412 ** 0: The operation is a read. Populate the overflow cache. | |
4413 ** 1: The operation is a write. Populate the overflow cache. | |
4414 ** 2: The operation is a read. Do not populate the overflow cache. | |
4415 ** | |
4416 ** A total of "amt" bytes are read or written beginning at "offset". | |
4417 ** Data is read to or from the buffer pBuf. | |
4418 ** | |
4419 ** The content being read or written might appear on the main page | |
4420 ** or be scattered out on multiple overflow pages. | |
4421 ** | |
4422 ** If the current cursor entry uses one or more overflow pages and the | |
4423 ** eOp argument is not 2, this function may allocate space for and lazily | |
4424 ** populates the overflow page-list cache array (BtCursor.aOverflow). | |
4425 ** Subsequent calls use this cache to make seeking to the supplied offset | |
4426 ** more efficient. | |
4427 ** | |
4428 ** Once an overflow page-list cache has been allocated, it may be | |
4429 ** invalidated if some other cursor writes to the same table, or if | |
4430 ** the cursor is moved to a different row. Additionally, in auto-vacuum | |
4431 ** mode, the following events may invalidate an overflow page-list cache. | |
4432 ** | |
4433 ** * An incremental vacuum, | |
4434 ** * A commit in auto_vacuum="full" mode, | |
4435 ** * Creating a table (may require moving an overflow page). | |
4436 */ | |
4437 static int accessPayload( | |
4438 BtCursor *pCur, /* Cursor pointing to entry to read from */ | |
4439 u32 offset, /* Begin reading this far into payload */ | |
4440 u32 amt, /* Read this many bytes */ | |
4441 unsigned char *pBuf, /* Write the bytes into this buffer */ | |
4442 int eOp /* zero to read. non-zero to write. */ | |
4443 ){ | |
4444 unsigned char *aPayload; | |
4445 int rc = SQLITE_OK; | |
4446 int iIdx = 0; | |
4447 MemPage *pPage = pCur->apPage[pCur->iPage]; /* Btree page of current entry */ | |
4448 BtShared *pBt = pCur->pBt; /* Btree this cursor belongs to */ | |
4449 #ifdef SQLITE_DIRECT_OVERFLOW_READ | |
4450 unsigned char * const pBufStart = pBuf; | |
4451 int bEnd; /* True if reading to end of data */ | |
4452 #endif | |
4453 | |
4454 assert( pPage ); | |
4455 assert( pCur->eState==CURSOR_VALID ); | |
4456 assert( pCur->aiIdx[pCur->iPage]<pPage->nCell ); | |
4457 assert( cursorHoldsMutex(pCur) ); | |
4458 assert( eOp!=2 || offset==0 ); /* Always start from beginning for eOp==2 */ | |
4459 | |
4460 getCellInfo(pCur); | |
4461 aPayload = pCur->info.pPayload; | |
4462 #ifdef SQLITE_DIRECT_OVERFLOW_READ | |
4463 bEnd = offset+amt==pCur->info.nPayload; | |
4464 #endif | |
4465 assert( offset+amt <= pCur->info.nPayload ); | |
4466 | |
4467 if( &aPayload[pCur->info.nLocal] > &pPage->aData[pBt->usableSize] ){ | |
4468 /* Trying to read or write past the end of the data is an error */ | |
4469 return SQLITE_CORRUPT_BKPT; | |
4470 } | |
4471 | |
4472 /* Check if data must be read/written to/from the btree page itself. */ | |
4473 if( offset<pCur->info.nLocal ){ | |
4474 int a = amt; | |
4475 if( a+offset>pCur->info.nLocal ){ | |
4476 a = pCur->info.nLocal - offset; | |
4477 } | |
4478 rc = copyPayload(&aPayload[offset], pBuf, a, (eOp & 0x01), pPage->pDbPage); | |
4479 offset = 0; | |
4480 pBuf += a; | |
4481 amt -= a; | |
4482 }else{ | |
4483 offset -= pCur->info.nLocal; | |
4484 } | |
4485 | |
4486 | |
4487 if( rc==SQLITE_OK && amt>0 ){ | |
4488 const u32 ovflSize = pBt->usableSize - 4; /* Bytes content per ovfl page */ | |
4489 Pgno nextPage; | |
4490 | |
4491 nextPage = get4byte(&aPayload[pCur->info.nLocal]); | |
4492 | |
4493 /* If the BtCursor.aOverflow[] has not been allocated, allocate it now. | |
4494 ** Except, do not allocate aOverflow[] for eOp==2. | |
4495 ** | |
4496 ** The aOverflow[] array is sized at one entry for each overflow page | |
4497 ** in the overflow chain. The page number of the first overflow page is | |
4498 ** stored in aOverflow[0], etc. A value of 0 in the aOverflow[] array | |
4499 ** means "not yet known" (the cache is lazily populated). | |
4500 */ | |
4501 if( eOp!=2 && (pCur->curFlags & BTCF_ValidOvfl)==0 ){ | |
4502 int nOvfl = (pCur->info.nPayload-pCur->info.nLocal+ovflSize-1)/ovflSize; | |
4503 if( nOvfl>pCur->nOvflAlloc ){ | |
4504 Pgno *aNew = (Pgno*)sqlite3Realloc( | |
4505 pCur->aOverflow, nOvfl*2*sizeof(Pgno) | |
4506 ); | |
4507 if( aNew==0 ){ | |
4508 rc = SQLITE_NOMEM; | |
4509 }else{ | |
4510 pCur->nOvflAlloc = nOvfl*2; | |
4511 pCur->aOverflow = aNew; | |
4512 } | |
4513 } | |
4514 if( rc==SQLITE_OK ){ | |
4515 memset(pCur->aOverflow, 0, nOvfl*sizeof(Pgno)); | |
4516 pCur->curFlags |= BTCF_ValidOvfl; | |
4517 } | |
4518 } | |
4519 | |
4520 /* If the overflow page-list cache has been allocated and the | |
4521 ** entry for the first required overflow page is valid, skip | |
4522 ** directly to it. | |
4523 */ | |
4524 if( (pCur->curFlags & BTCF_ValidOvfl)!=0 | |
4525 && pCur->aOverflow[offset/ovflSize] | |
4526 ){ | |
4527 iIdx = (offset/ovflSize); | |
4528 nextPage = pCur->aOverflow[iIdx]; | |
4529 offset = (offset%ovflSize); | |
4530 } | |
4531 | |
4532 for( ; rc==SQLITE_OK && amt>0 && nextPage; iIdx++){ | |
4533 | |
4534 /* If required, populate the overflow page-list cache. */ | |
4535 if( (pCur->curFlags & BTCF_ValidOvfl)!=0 ){ | |
4536 assert( pCur->aOverflow[iIdx]==0 | |
4537 || pCur->aOverflow[iIdx]==nextPage | |
4538 || CORRUPT_DB ); | |
4539 pCur->aOverflow[iIdx] = nextPage; | |
4540 } | |
4541 | |
4542 if( offset>=ovflSize ){ | |
4543 /* The only reason to read this page is to obtain the page | |
4544 ** number for the next page in the overflow chain. The page | |
4545 ** data is not required. So first try to lookup the overflow | |
4546 ** page-list cache, if any, then fall back to the getOverflowPage() | |
4547 ** function. | |
4548 ** | |
4549 ** Note that the aOverflow[] array must be allocated because eOp!=2 | |
4550 ** here. If eOp==2, then offset==0 and this branch is never taken. | |
4551 */ | |
4552 assert( eOp!=2 ); | |
4553 assert( pCur->curFlags & BTCF_ValidOvfl ); | |
4554 assert( pCur->pBtree->db==pBt->db ); | |
4555 if( pCur->aOverflow[iIdx+1] ){ | |
4556 nextPage = pCur->aOverflow[iIdx+1]; | |
4557 }else{ | |
4558 rc = getOverflowPage(pBt, nextPage, 0, &nextPage); | |
4559 } | |
4560 offset -= ovflSize; | |
4561 }else{ | |
4562 /* Need to read this page properly. It contains some of the | |
4563 ** range of data that is being read (eOp==0) or written (eOp!=0). | |
4564 */ | |
4565 #ifdef SQLITE_DIRECT_OVERFLOW_READ | |
4566 sqlite3_file *fd; | |
4567 #endif | |
4568 int a = amt; | |
4569 if( a + offset > ovflSize ){ | |
4570 a = ovflSize - offset; | |
4571 } | |
4572 | |
4573 #ifdef SQLITE_DIRECT_OVERFLOW_READ | |
4574 /* If all the following are true: | |
4575 ** | |
4576 ** 1) this is a read operation, and | |
4577 ** 2) data is required from the start of this overflow page, and | |
4578 ** 3) the database is file-backed, and | |
4579 ** 4) there is no open write-transaction, and | |
4580 ** 5) the database is not a WAL database, | |
4581 ** 6) all data from the page is being read. | |
4582 ** 7) at least 4 bytes have already been read into the output buffer | |
4583 ** | |
4584 ** then data can be read directly from the database file into the | |
4585 ** output buffer, bypassing the page-cache altogether. This speeds | |
4586 ** up loading large records that span many overflow pages. | |
4587 */ | |
4588 if( (eOp&0x01)==0 /* (1) */ | |
4589 && offset==0 /* (2) */ | |
4590 && (bEnd || a==ovflSize) /* (6) */ | |
4591 && pBt->inTransaction==TRANS_READ /* (4) */ | |
4592 && (fd = sqlite3PagerFile(pBt->pPager))->pMethods /* (3) */ | |
4593 && pBt->pPage1->aData[19]==0x01 /* (5) */ | |
4594 && &pBuf[-4]>=pBufStart /* (7) */ | |
4595 ){ | |
4596 u8 aSave[4]; | |
4597 u8 *aWrite = &pBuf[-4]; | |
4598 assert( aWrite>=pBufStart ); /* hence (7) */ | |
4599 memcpy(aSave, aWrite, 4); | |
4600 rc = sqlite3OsRead(fd, aWrite, a+4, (i64)pBt->pageSize*(nextPage-1)); | |
4601 nextPage = get4byte(aWrite); | |
4602 memcpy(aWrite, aSave, 4); | |
4603 }else | |
4604 #endif | |
4605 | |
4606 { | |
4607 DbPage *pDbPage; | |
4608 rc = sqlite3PagerGet(pBt->pPager, nextPage, &pDbPage, | |
4609 ((eOp&0x01)==0 ? PAGER_GET_READONLY : 0) | |
4610 ); | |
4611 if( rc==SQLITE_OK ){ | |
4612 aPayload = sqlite3PagerGetData(pDbPage); | |
4613 nextPage = get4byte(aPayload); | |
4614 rc = copyPayload(&aPayload[offset+4], pBuf, a, (eOp&0x01), pDbPage); | |
4615 sqlite3PagerUnref(pDbPage); | |
4616 offset = 0; | |
4617 } | |
4618 } | |
4619 amt -= a; | |
4620 pBuf += a; | |
4621 } | |
4622 } | |
4623 } | |
4624 | |
4625 if( rc==SQLITE_OK && amt>0 ){ | |
4626 return SQLITE_CORRUPT_BKPT; | |
4627 } | |
4628 return rc; | |
4629 } | |
4630 | |
4631 /* | |
4632 ** Read part of the key associated with cursor pCur. Exactly | |
4633 ** "amt" bytes will be transferred into pBuf[]. The transfer | |
4634 ** begins at "offset". | |
4635 ** | |
4636 ** The caller must ensure that pCur is pointing to a valid row | |
4637 ** in the table. | |
4638 ** | |
4639 ** Return SQLITE_OK on success or an error code if anything goes | |
4640 ** wrong. An error is returned if "offset+amt" is larger than | |
4641 ** the available payload. | |
4642 */ | |
4643 int sqlite3BtreeKey(BtCursor *pCur, u32 offset, u32 amt, void *pBuf){ | |
4644 assert( cursorHoldsMutex(pCur) ); | |
4645 assert( pCur->eState==CURSOR_VALID ); | |
4646 assert( pCur->iPage>=0 && pCur->apPage[pCur->iPage] ); | |
4647 assert( pCur->aiIdx[pCur->iPage]<pCur->apPage[pCur->iPage]->nCell ); | |
4648 return accessPayload(pCur, offset, amt, (unsigned char*)pBuf, 0); | |
4649 } | |
4650 | |
4651 /* | |
4652 ** Read part of the data associated with cursor pCur. Exactly | |
4653 ** "amt" bytes will be transfered into pBuf[]. The transfer | |
4654 ** begins at "offset". | |
4655 ** | |
4656 ** Return SQLITE_OK on success or an error code if anything goes | |
4657 ** wrong. An error is returned if "offset+amt" is larger than | |
4658 ** the available payload. | |
4659 */ | |
4660 int sqlite3BtreeData(BtCursor *pCur, u32 offset, u32 amt, void *pBuf){ | |
4661 int rc; | |
4662 | |
4663 #ifndef SQLITE_OMIT_INCRBLOB | |
4664 if ( pCur->eState==CURSOR_INVALID ){ | |
4665 return SQLITE_ABORT; | |
4666 } | |
4667 #endif | |
4668 | |
4669 assert( cursorHoldsMutex(pCur) ); | |
4670 rc = restoreCursorPosition(pCur); | |
4671 if( rc==SQLITE_OK ){ | |
4672 assert( pCur->eState==CURSOR_VALID ); | |
4673 assert( pCur->iPage>=0 && pCur->apPage[pCur->iPage] ); | |
4674 assert( pCur->aiIdx[pCur->iPage]<pCur->apPage[pCur->iPage]->nCell ); | |
4675 rc = accessPayload(pCur, offset, amt, pBuf, 0); | |
4676 } | |
4677 return rc; | |
4678 } | |
4679 | |
4680 /* | |
4681 ** Return a pointer to payload information from the entry that the | |
4682 ** pCur cursor is pointing to. The pointer is to the beginning of | |
4683 ** the key if index btrees (pPage->intKey==0) and is the data for | |
4684 ** table btrees (pPage->intKey==1). The number of bytes of available | |
4685 ** key/data is written into *pAmt. If *pAmt==0, then the value | |
4686 ** returned will not be a valid pointer. | |
4687 ** | |
4688 ** This routine is an optimization. It is common for the entire key | |
4689 ** and data to fit on the local page and for there to be no overflow | |
4690 ** pages. When that is so, this routine can be used to access the | |
4691 ** key and data without making a copy. If the key and/or data spills | |
4692 ** onto overflow pages, then accessPayload() must be used to reassemble | |
4693 ** the key/data and copy it into a preallocated buffer. | |
4694 ** | |
4695 ** The pointer returned by this routine looks directly into the cached | |
4696 ** page of the database. The data might change or move the next time | |
4697 ** any btree routine is called. | |
4698 */ | |
4699 static const void *fetchPayload( | |
4700 BtCursor *pCur, /* Cursor pointing to entry to read from */ | |
4701 u32 *pAmt /* Write the number of available bytes here */ | |
4702 ){ | |
4703 u32 amt; | |
4704 assert( pCur!=0 && pCur->iPage>=0 && pCur->apPage[pCur->iPage]); | |
4705 assert( pCur->eState==CURSOR_VALID ); | |
4706 assert( sqlite3_mutex_held(pCur->pBtree->db->mutex) ); | |
4707 assert( cursorHoldsMutex(pCur) ); | |
4708 assert( pCur->aiIdx[pCur->iPage]<pCur->apPage[pCur->iPage]->nCell ); | |
4709 assert( pCur->info.nSize>0 ); | |
4710 assert( pCur->info.pPayload>pCur->apPage[pCur->iPage]->aData || CORRUPT_DB ); | |
4711 assert( pCur->info.pPayload<pCur->apPage[pCur->iPage]->aDataEnd ||CORRUPT_DB); | |
4712 amt = (int)(pCur->apPage[pCur->iPage]->aDataEnd - pCur->info.pPayload); | |
4713 if( pCur->info.nLocal<amt ) amt = pCur->info.nLocal; | |
4714 *pAmt = amt; | |
4715 return (void*)pCur->info.pPayload; | |
4716 } | |
4717 | |
4718 | |
4719 /* | |
4720 ** For the entry that cursor pCur is point to, return as | |
4721 ** many bytes of the key or data as are available on the local | |
4722 ** b-tree page. Write the number of available bytes into *pAmt. | |
4723 ** | |
4724 ** The pointer returned is ephemeral. The key/data may move | |
4725 ** or be destroyed on the next call to any Btree routine, | |
4726 ** including calls from other threads against the same cache. | |
4727 ** Hence, a mutex on the BtShared should be held prior to calling | |
4728 ** this routine. | |
4729 ** | |
4730 ** These routines is used to get quick access to key and data | |
4731 ** in the common case where no overflow pages are used. | |
4732 */ | |
4733 const void *sqlite3BtreeKeyFetch(BtCursor *pCur, u32 *pAmt){ | |
4734 return fetchPayload(pCur, pAmt); | |
4735 } | |
4736 const void *sqlite3BtreeDataFetch(BtCursor *pCur, u32 *pAmt){ | |
4737 return fetchPayload(pCur, pAmt); | |
4738 } | |
4739 | |
4740 | |
4741 /* | |
4742 ** Move the cursor down to a new child page. The newPgno argument is the | |
4743 ** page number of the child page to move to. | |
4744 ** | |
4745 ** This function returns SQLITE_CORRUPT if the page-header flags field of | |
4746 ** the new child page does not match the flags field of the parent (i.e. | |
4747 ** if an intkey page appears to be the parent of a non-intkey page, or | |
4748 ** vice-versa). | |
4749 */ | |
4750 static int moveToChild(BtCursor *pCur, u32 newPgno){ | |
4751 BtShared *pBt = pCur->pBt; | |
4752 | |
4753 assert( cursorHoldsMutex(pCur) ); | |
4754 assert( pCur->eState==CURSOR_VALID ); | |
4755 assert( pCur->iPage<BTCURSOR_MAX_DEPTH ); | |
4756 assert( pCur->iPage>=0 ); | |
4757 if( pCur->iPage>=(BTCURSOR_MAX_DEPTH-1) ){ | |
4758 return SQLITE_CORRUPT_BKPT; | |
4759 } | |
4760 pCur->info.nSize = 0; | |
4761 pCur->curFlags &= ~(BTCF_ValidNKey|BTCF_ValidOvfl); | |
4762 pCur->iPage++; | |
4763 pCur->aiIdx[pCur->iPage] = 0; | |
4764 return getAndInitPage(pBt, newPgno, &pCur->apPage[pCur->iPage], | |
4765 pCur, pCur->curPagerFlags); | |
4766 } | |
4767 | |
4768 #if SQLITE_DEBUG | |
4769 /* | |
4770 ** Page pParent is an internal (non-leaf) tree page. This function | |
4771 ** asserts that page number iChild is the left-child if the iIdx'th | |
4772 ** cell in page pParent. Or, if iIdx is equal to the total number of | |
4773 ** cells in pParent, that page number iChild is the right-child of | |
4774 ** the page. | |
4775 */ | |
4776 static void assertParentIndex(MemPage *pParent, int iIdx, Pgno iChild){ | |
4777 if( CORRUPT_DB ) return; /* The conditions tested below might not be true | |
4778 ** in a corrupt database */ | |
4779 assert( iIdx<=pParent->nCell ); | |
4780 if( iIdx==pParent->nCell ){ | |
4781 assert( get4byte(&pParent->aData[pParent->hdrOffset+8])==iChild ); | |
4782 }else{ | |
4783 assert( get4byte(findCell(pParent, iIdx))==iChild ); | |
4784 } | |
4785 } | |
4786 #else | |
4787 # define assertParentIndex(x,y,z) | |
4788 #endif | |
4789 | |
4790 /* | |
4791 ** Move the cursor up to the parent page. | |
4792 ** | |
4793 ** pCur->idx is set to the cell index that contains the pointer | |
4794 ** to the page we are coming from. If we are coming from the | |
4795 ** right-most child page then pCur->idx is set to one more than | |
4796 ** the largest cell index. | |
4797 */ | |
4798 static void moveToParent(BtCursor *pCur){ | |
4799 assert( cursorHoldsMutex(pCur) ); | |
4800 assert( pCur->eState==CURSOR_VALID ); | |
4801 assert( pCur->iPage>0 ); | |
4802 assert( pCur->apPage[pCur->iPage] ); | |
4803 assertParentIndex( | |
4804 pCur->apPage[pCur->iPage-1], | |
4805 pCur->aiIdx[pCur->iPage-1], | |
4806 pCur->apPage[pCur->iPage]->pgno | |
4807 ); | |
4808 testcase( pCur->aiIdx[pCur->iPage-1] > pCur->apPage[pCur->iPage-1]->nCell ); | |
4809 pCur->info.nSize = 0; | |
4810 pCur->curFlags &= ~(BTCF_ValidNKey|BTCF_ValidOvfl); | |
4811 releasePageNotNull(pCur->apPage[pCur->iPage--]); | |
4812 } | |
4813 | |
4814 /* | |
4815 ** Move the cursor to point to the root page of its b-tree structure. | |
4816 ** | |
4817 ** If the table has a virtual root page, then the cursor is moved to point | |
4818 ** to the virtual root page instead of the actual root page. A table has a | |
4819 ** virtual root page when the actual root page contains no cells and a | |
4820 ** single child page. This can only happen with the table rooted at page 1. | |
4821 ** | |
4822 ** If the b-tree structure is empty, the cursor state is set to | |
4823 ** CURSOR_INVALID. Otherwise, the cursor is set to point to the first | |
4824 ** cell located on the root (or virtual root) page and the cursor state | |
4825 ** is set to CURSOR_VALID. | |
4826 ** | |
4827 ** If this function returns successfully, it may be assumed that the | |
4828 ** page-header flags indicate that the [virtual] root-page is the expected | |
4829 ** kind of b-tree page (i.e. if when opening the cursor the caller did not | |
4830 ** specify a KeyInfo structure the flags byte is set to 0x05 or 0x0D, | |
4831 ** indicating a table b-tree, or if the caller did specify a KeyInfo | |
4832 ** structure the flags byte is set to 0x02 or 0x0A, indicating an index | |
4833 ** b-tree). | |
4834 */ | |
4835 static int moveToRoot(BtCursor *pCur){ | |
4836 MemPage *pRoot; | |
4837 int rc = SQLITE_OK; | |
4838 | |
4839 assert( cursorHoldsMutex(pCur) ); | |
4840 assert( CURSOR_INVALID < CURSOR_REQUIRESEEK ); | |
4841 assert( CURSOR_VALID < CURSOR_REQUIRESEEK ); | |
4842 assert( CURSOR_FAULT > CURSOR_REQUIRESEEK ); | |
4843 if( pCur->eState>=CURSOR_REQUIRESEEK ){ | |
4844 if( pCur->eState==CURSOR_FAULT ){ | |
4845 assert( pCur->skipNext!=SQLITE_OK ); | |
4846 return pCur->skipNext; | |
4847 } | |
4848 sqlite3BtreeClearCursor(pCur); | |
4849 } | |
4850 | |
4851 if( pCur->iPage>=0 ){ | |
4852 while( pCur->iPage ){ | |
4853 assert( pCur->apPage[pCur->iPage]!=0 ); | |
4854 releasePageNotNull(pCur->apPage[pCur->iPage--]); | |
4855 } | |
4856 }else if( pCur->pgnoRoot==0 ){ | |
4857 pCur->eState = CURSOR_INVALID; | |
4858 return SQLITE_OK; | |
4859 }else{ | |
4860 assert( pCur->iPage==(-1) ); | |
4861 rc = getAndInitPage(pCur->pBtree->pBt, pCur->pgnoRoot, &pCur->apPage[0], | |
4862 0, pCur->curPagerFlags); | |
4863 if( rc!=SQLITE_OK ){ | |
4864 pCur->eState = CURSOR_INVALID; | |
4865 return rc; | |
4866 } | |
4867 pCur->iPage = 0; | |
4868 pCur->curIntKey = pCur->apPage[0]->intKey; | |
4869 } | |
4870 pRoot = pCur->apPage[0]; | |
4871 assert( pRoot->pgno==pCur->pgnoRoot ); | |
4872 | |
4873 /* If pCur->pKeyInfo is not NULL, then the caller that opened this cursor | |
4874 ** expected to open it on an index b-tree. Otherwise, if pKeyInfo is | |
4875 ** NULL, the caller expects a table b-tree. If this is not the case, | |
4876 ** return an SQLITE_CORRUPT error. | |
4877 ** | |
4878 ** Earlier versions of SQLite assumed that this test could not fail | |
4879 ** if the root page was already loaded when this function was called (i.e. | |
4880 ** if pCur->iPage>=0). But this is not so if the database is corrupted | |
4881 ** in such a way that page pRoot is linked into a second b-tree table | |
4882 ** (or the freelist). */ | |
4883 assert( pRoot->intKey==1 || pRoot->intKey==0 ); | |
4884 if( pRoot->isInit==0 || (pCur->pKeyInfo==0)!=pRoot->intKey ){ | |
4885 return SQLITE_CORRUPT_BKPT; | |
4886 } | |
4887 | |
4888 pCur->aiIdx[0] = 0; | |
4889 pCur->info.nSize = 0; | |
4890 pCur->curFlags &= ~(BTCF_AtLast|BTCF_ValidNKey|BTCF_ValidOvfl); | |
4891 | |
4892 if( pRoot->nCell>0 ){ | |
4893 pCur->eState = CURSOR_VALID; | |
4894 }else if( !pRoot->leaf ){ | |
4895 Pgno subpage; | |
4896 if( pRoot->pgno!=1 ) return SQLITE_CORRUPT_BKPT; | |
4897 subpage = get4byte(&pRoot->aData[pRoot->hdrOffset+8]); | |
4898 pCur->eState = CURSOR_VALID; | |
4899 rc = moveToChild(pCur, subpage); | |
4900 }else{ | |
4901 pCur->eState = CURSOR_INVALID; | |
4902 } | |
4903 return rc; | |
4904 } | |
4905 | |
4906 /* | |
4907 ** Move the cursor down to the left-most leaf entry beneath the | |
4908 ** entry to which it is currently pointing. | |
4909 ** | |
4910 ** The left-most leaf is the one with the smallest key - the first | |
4911 ** in ascending order. | |
4912 */ | |
4913 static int moveToLeftmost(BtCursor *pCur){ | |
4914 Pgno pgno; | |
4915 int rc = SQLITE_OK; | |
4916 MemPage *pPage; | |
4917 | |
4918 assert( cursorHoldsMutex(pCur) ); | |
4919 assert( pCur->eState==CURSOR_VALID ); | |
4920 while( rc==SQLITE_OK && !(pPage = pCur->apPage[pCur->iPage])->leaf ){ | |
4921 assert( pCur->aiIdx[pCur->iPage]<pPage->nCell ); | |
4922 pgno = get4byte(findCell(pPage, pCur->aiIdx[pCur->iPage])); | |
4923 rc = moveToChild(pCur, pgno); | |
4924 } | |
4925 return rc; | |
4926 } | |
4927 | |
4928 /* | |
4929 ** Move the cursor down to the right-most leaf entry beneath the | |
4930 ** page to which it is currently pointing. Notice the difference | |
4931 ** between moveToLeftmost() and moveToRightmost(). moveToLeftmost() | |
4932 ** finds the left-most entry beneath the *entry* whereas moveToRightmost() | |
4933 ** finds the right-most entry beneath the *page*. | |
4934 ** | |
4935 ** The right-most entry is the one with the largest key - the last | |
4936 ** key in ascending order. | |
4937 */ | |
4938 static int moveToRightmost(BtCursor *pCur){ | |
4939 Pgno pgno; | |
4940 int rc = SQLITE_OK; | |
4941 MemPage *pPage = 0; | |
4942 | |
4943 assert( cursorHoldsMutex(pCur) ); | |
4944 assert( pCur->eState==CURSOR_VALID ); | |
4945 while( !(pPage = pCur->apPage[pCur->iPage])->leaf ){ | |
4946 pgno = get4byte(&pPage->aData[pPage->hdrOffset+8]); | |
4947 pCur->aiIdx[pCur->iPage] = pPage->nCell; | |
4948 rc = moveToChild(pCur, pgno); | |
4949 if( rc ) return rc; | |
4950 } | |
4951 pCur->aiIdx[pCur->iPage] = pPage->nCell-1; | |
4952 assert( pCur->info.nSize==0 ); | |
4953 assert( (pCur->curFlags & BTCF_ValidNKey)==0 ); | |
4954 return SQLITE_OK; | |
4955 } | |
4956 | |
4957 /* Move the cursor to the first entry in the table. Return SQLITE_OK | |
4958 ** on success. Set *pRes to 0 if the cursor actually points to something | |
4959 ** or set *pRes to 1 if the table is empty. | |
4960 */ | |
4961 int sqlite3BtreeFirst(BtCursor *pCur, int *pRes){ | |
4962 int rc; | |
4963 | |
4964 assert( cursorHoldsMutex(pCur) ); | |
4965 assert( sqlite3_mutex_held(pCur->pBtree->db->mutex) ); | |
4966 rc = moveToRoot(pCur); | |
4967 if( rc==SQLITE_OK ){ | |
4968 if( pCur->eState==CURSOR_INVALID ){ | |
4969 assert( pCur->pgnoRoot==0 || pCur->apPage[pCur->iPage]->nCell==0 ); | |
4970 *pRes = 1; | |
4971 }else{ | |
4972 assert( pCur->apPage[pCur->iPage]->nCell>0 ); | |
4973 *pRes = 0; | |
4974 rc = moveToLeftmost(pCur); | |
4975 } | |
4976 } | |
4977 return rc; | |
4978 } | |
4979 | |
4980 /* Move the cursor to the last entry in the table. Return SQLITE_OK | |
4981 ** on success. Set *pRes to 0 if the cursor actually points to something | |
4982 ** or set *pRes to 1 if the table is empty. | |
4983 */ | |
4984 int sqlite3BtreeLast(BtCursor *pCur, int *pRes){ | |
4985 int rc; | |
4986 | |
4987 assert( cursorHoldsMutex(pCur) ); | |
4988 assert( sqlite3_mutex_held(pCur->pBtree->db->mutex) ); | |
4989 | |
4990 /* If the cursor already points to the last entry, this is a no-op. */ | |
4991 if( CURSOR_VALID==pCur->eState && (pCur->curFlags & BTCF_AtLast)!=0 ){ | |
4992 #ifdef SQLITE_DEBUG | |
4993 /* This block serves to assert() that the cursor really does point | |
4994 ** to the last entry in the b-tree. */ | |
4995 int ii; | |
4996 for(ii=0; ii<pCur->iPage; ii++){ | |
4997 assert( pCur->aiIdx[ii]==pCur->apPage[ii]->nCell ); | |
4998 } | |
4999 assert( pCur->aiIdx[pCur->iPage]==pCur->apPage[pCur->iPage]->nCell-1 ); | |
5000 assert( pCur->apPage[pCur->iPage]->leaf ); | |
5001 #endif | |
5002 return SQLITE_OK; | |
5003 } | |
5004 | |
5005 rc = moveToRoot(pCur); | |
5006 if( rc==SQLITE_OK ){ | |
5007 if( CURSOR_INVALID==pCur->eState ){ | |
5008 assert( pCur->pgnoRoot==0 || pCur->apPage[pCur->iPage]->nCell==0 ); | |
5009 *pRes = 1; | |
5010 }else{ | |
5011 assert( pCur->eState==CURSOR_VALID ); | |
5012 *pRes = 0; | |
5013 rc = moveToRightmost(pCur); | |
5014 if( rc==SQLITE_OK ){ | |
5015 pCur->curFlags |= BTCF_AtLast; | |
5016 }else{ | |
5017 pCur->curFlags &= ~BTCF_AtLast; | |
5018 } | |
5019 | |
5020 } | |
5021 } | |
5022 return rc; | |
5023 } | |
5024 | |
5025 /* Move the cursor so that it points to an entry near the key | |
5026 ** specified by pIdxKey or intKey. Return a success code. | |
5027 ** | |
5028 ** For INTKEY tables, the intKey parameter is used. pIdxKey | |
5029 ** must be NULL. For index tables, pIdxKey is used and intKey | |
5030 ** is ignored. | |
5031 ** | |
5032 ** If an exact match is not found, then the cursor is always | |
5033 ** left pointing at a leaf page which would hold the entry if it | |
5034 ** were present. The cursor might point to an entry that comes | |
5035 ** before or after the key. | |
5036 ** | |
5037 ** An integer is written into *pRes which is the result of | |
5038 ** comparing the key with the entry to which the cursor is | |
5039 ** pointing. The meaning of the integer written into | |
5040 ** *pRes is as follows: | |
5041 ** | |
5042 ** *pRes<0 The cursor is left pointing at an entry that | |
5043 ** is smaller than intKey/pIdxKey or if the table is empty | |
5044 ** and the cursor is therefore left point to nothing. | |
5045 ** | |
5046 ** *pRes==0 The cursor is left pointing at an entry that | |
5047 ** exactly matches intKey/pIdxKey. | |
5048 ** | |
5049 ** *pRes>0 The cursor is left pointing at an entry that | |
5050 ** is larger than intKey/pIdxKey. | |
5051 ** | |
5052 ** For index tables, the pIdxKey->eqSeen field is set to 1 if there | |
5053 ** exists an entry in the table that exactly matches pIdxKey. | |
5054 */ | |
5055 int sqlite3BtreeMovetoUnpacked( | |
5056 BtCursor *pCur, /* The cursor to be moved */ | |
5057 UnpackedRecord *pIdxKey, /* Unpacked index key */ | |
5058 i64 intKey, /* The table key */ | |
5059 int biasRight, /* If true, bias the search to the high end */ | |
5060 int *pRes /* Write search results here */ | |
5061 ){ | |
5062 int rc; | |
5063 RecordCompare xRecordCompare; | |
5064 | |
5065 assert( cursorHoldsMutex(pCur) ); | |
5066 assert( sqlite3_mutex_held(pCur->pBtree->db->mutex) ); | |
5067 assert( pRes ); | |
5068 assert( (pIdxKey==0)==(pCur->pKeyInfo==0) ); | |
5069 | |
5070 /* If the cursor is already positioned at the point we are trying | |
5071 ** to move to, then just return without doing any work */ | |
5072 if( pCur->eState==CURSOR_VALID && (pCur->curFlags & BTCF_ValidNKey)!=0 | |
5073 && pCur->curIntKey | |
5074 ){ | |
5075 if( pCur->info.nKey==intKey ){ | |
5076 *pRes = 0; | |
5077 return SQLITE_OK; | |
5078 } | |
5079 if( (pCur->curFlags & BTCF_AtLast)!=0 && pCur->info.nKey<intKey ){ | |
5080 *pRes = -1; | |
5081 return SQLITE_OK; | |
5082 } | |
5083 } | |
5084 | |
5085 if( pIdxKey ){ | |
5086 xRecordCompare = sqlite3VdbeFindCompare(pIdxKey); | |
5087 pIdxKey->errCode = 0; | |
5088 assert( pIdxKey->default_rc==1 | |
5089 || pIdxKey->default_rc==0 | |
5090 || pIdxKey->default_rc==-1 | |
5091 ); | |
5092 }else{ | |
5093 xRecordCompare = 0; /* All keys are integers */ | |
5094 } | |
5095 | |
5096 rc = moveToRoot(pCur); | |
5097 if( rc ){ | |
5098 return rc; | |
5099 } | |
5100 assert( pCur->pgnoRoot==0 || pCur->apPage[pCur->iPage] ); | |
5101 assert( pCur->pgnoRoot==0 || pCur->apPage[pCur->iPage]->isInit ); | |
5102 assert( pCur->eState==CURSOR_INVALID || pCur->apPage[pCur->iPage]->nCell>0 ); | |
5103 if( pCur->eState==CURSOR_INVALID ){ | |
5104 *pRes = -1; | |
5105 assert( pCur->pgnoRoot==0 || pCur->apPage[pCur->iPage]->nCell==0 ); | |
5106 return SQLITE_OK; | |
5107 } | |
5108 assert( pCur->apPage[0]->intKey==pCur->curIntKey ); | |
5109 assert( pCur->curIntKey || pIdxKey ); | |
5110 for(;;){ | |
5111 int lwr, upr, idx, c; | |
5112 Pgno chldPg; | |
5113 MemPage *pPage = pCur->apPage[pCur->iPage]; | |
5114 u8 *pCell; /* Pointer to current cell in pPage */ | |
5115 | |
5116 /* pPage->nCell must be greater than zero. If this is the root-page | |
5117 ** the cursor would have been INVALID above and this for(;;) loop | |
5118 ** not run. If this is not the root-page, then the moveToChild() routine | |
5119 ** would have already detected db corruption. Similarly, pPage must | |
5120 ** be the right kind (index or table) of b-tree page. Otherwise | |
5121 ** a moveToChild() or moveToRoot() call would have detected corruption. */ | |
5122 assert( pPage->nCell>0 ); | |
5123 assert( pPage->intKey==(pIdxKey==0) ); | |
5124 lwr = 0; | |
5125 upr = pPage->nCell-1; | |
5126 assert( biasRight==0 || biasRight==1 ); | |
5127 idx = upr>>(1-biasRight); /* idx = biasRight ? upr : (lwr+upr)/2; */ | |
5128 pCur->aiIdx[pCur->iPage] = (u16)idx; | |
5129 if( xRecordCompare==0 ){ | |
5130 for(;;){ | |
5131 i64 nCellKey; | |
5132 pCell = findCellPastPtr(pPage, idx); | |
5133 if( pPage->intKeyLeaf ){ | |
5134 while( 0x80 <= *(pCell++) ){ | |
5135 if( pCell>=pPage->aDataEnd ) return SQLITE_CORRUPT_BKPT; | |
5136 } | |
5137 } | |
5138 getVarint(pCell, (u64*)&nCellKey); | |
5139 if( nCellKey<intKey ){ | |
5140 lwr = idx+1; | |
5141 if( lwr>upr ){ c = -1; break; } | |
5142 }else if( nCellKey>intKey ){ | |
5143 upr = idx-1; | |
5144 if( lwr>upr ){ c = +1; break; } | |
5145 }else{ | |
5146 assert( nCellKey==intKey ); | |
5147 pCur->curFlags |= BTCF_ValidNKey; | |
5148 pCur->info.nKey = nCellKey; | |
5149 pCur->aiIdx[pCur->iPage] = (u16)idx; | |
5150 if( !pPage->leaf ){ | |
5151 lwr = idx; | |
5152 goto moveto_next_layer; | |
5153 }else{ | |
5154 *pRes = 0; | |
5155 rc = SQLITE_OK; | |
5156 goto moveto_finish; | |
5157 } | |
5158 } | |
5159 assert( lwr+upr>=0 ); | |
5160 idx = (lwr+upr)>>1; /* idx = (lwr+upr)/2; */ | |
5161 } | |
5162 }else{ | |
5163 for(;;){ | |
5164 int nCell; /* Size of the pCell cell in bytes */ | |
5165 pCell = findCellPastPtr(pPage, idx); | |
5166 | |
5167 /* The maximum supported page-size is 65536 bytes. This means that | |
5168 ** the maximum number of record bytes stored on an index B-Tree | |
5169 ** page is less than 16384 bytes and may be stored as a 2-byte | |
5170 ** varint. This information is used to attempt to avoid parsing | |
5171 ** the entire cell by checking for the cases where the record is | |
5172 ** stored entirely within the b-tree page by inspecting the first | |
5173 ** 2 bytes of the cell. | |
5174 */ | |
5175 nCell = pCell[0]; | |
5176 if( nCell<=pPage->max1bytePayload ){ | |
5177 /* This branch runs if the record-size field of the cell is a | |
5178 ** single byte varint and the record fits entirely on the main | |
5179 ** b-tree page. */ | |
5180 testcase( pCell+nCell+1==pPage->aDataEnd ); | |
5181 c = xRecordCompare(nCell, (void*)&pCell[1], pIdxKey); | |
5182 }else if( !(pCell[1] & 0x80) | |
5183 && (nCell = ((nCell&0x7f)<<7) + pCell[1])<=pPage->maxLocal | |
5184 ){ | |
5185 /* The record-size field is a 2 byte varint and the record | |
5186 ** fits entirely on the main b-tree page. */ | |
5187 testcase( pCell+nCell+2==pPage->aDataEnd ); | |
5188 c = xRecordCompare(nCell, (void*)&pCell[2], pIdxKey); | |
5189 }else{ | |
5190 /* The record flows over onto one or more overflow pages. In | |
5191 ** this case the whole cell needs to be parsed, a buffer allocated | |
5192 ** and accessPayload() used to retrieve the record into the | |
5193 ** buffer before VdbeRecordCompare() can be called. | |
5194 ** | |
5195 ** If the record is corrupt, the xRecordCompare routine may read | |
5196 ** up to two varints past the end of the buffer. An extra 18 | |
5197 ** bytes of padding is allocated at the end of the buffer in | |
5198 ** case this happens. */ | |
5199 void *pCellKey; | |
5200 u8 * const pCellBody = pCell - pPage->childPtrSize; | |
5201 pPage->xParseCell(pPage, pCellBody, &pCur->info); | |
5202 nCell = (int)pCur->info.nKey; | |
5203 testcase( nCell<0 ); /* True if key size is 2^32 or more */ | |
5204 testcase( nCell==0 ); /* Invalid key size: 0x80 0x80 0x00 */ | |
5205 testcase( nCell==1 ); /* Invalid key size: 0x80 0x80 0x01 */ | |
5206 testcase( nCell==2 ); /* Minimum legal index key size */ | |
5207 if( nCell<2 ){ | |
5208 rc = SQLITE_CORRUPT_BKPT; | |
5209 goto moveto_finish; | |
5210 } | |
5211 pCellKey = sqlite3Malloc( nCell+18 ); | |
5212 if( pCellKey==0 ){ | |
5213 rc = SQLITE_NOMEM; | |
5214 goto moveto_finish; | |
5215 } | |
5216 pCur->aiIdx[pCur->iPage] = (u16)idx; | |
5217 rc = accessPayload(pCur, 0, nCell, (unsigned char*)pCellKey, 2); | |
5218 if( rc ){ | |
5219 sqlite3_free(pCellKey); | |
5220 goto moveto_finish; | |
5221 } | |
5222 c = xRecordCompare(nCell, pCellKey, pIdxKey); | |
5223 sqlite3_free(pCellKey); | |
5224 } | |
5225 assert( | |
5226 (pIdxKey->errCode!=SQLITE_CORRUPT || c==0) | |
5227 && (pIdxKey->errCode!=SQLITE_NOMEM || pCur->pBtree->db->mallocFailed) | |
5228 ); | |
5229 if( c<0 ){ | |
5230 lwr = idx+1; | |
5231 }else if( c>0 ){ | |
5232 upr = idx-1; | |
5233 }else{ | |
5234 assert( c==0 ); | |
5235 *pRes = 0; | |
5236 rc = SQLITE_OK; | |
5237 pCur->aiIdx[pCur->iPage] = (u16)idx; | |
5238 if( pIdxKey->errCode ) rc = SQLITE_CORRUPT; | |
5239 goto moveto_finish; | |
5240 } | |
5241 if( lwr>upr ) break; | |
5242 assert( lwr+upr>=0 ); | |
5243 idx = (lwr+upr)>>1; /* idx = (lwr+upr)/2 */ | |
5244 } | |
5245 } | |
5246 assert( lwr==upr+1 || (pPage->intKey && !pPage->leaf) ); | |
5247 assert( pPage->isInit ); | |
5248 if( pPage->leaf ){ | |
5249 assert( pCur->aiIdx[pCur->iPage]<pCur->apPage[pCur->iPage]->nCell ); | |
5250 pCur->aiIdx[pCur->iPage] = (u16)idx; | |
5251 *pRes = c; | |
5252 rc = SQLITE_OK; | |
5253 goto moveto_finish; | |
5254 } | |
5255 moveto_next_layer: | |
5256 if( lwr>=pPage->nCell ){ | |
5257 chldPg = get4byte(&pPage->aData[pPage->hdrOffset+8]); | |
5258 }else{ | |
5259 chldPg = get4byte(findCell(pPage, lwr)); | |
5260 } | |
5261 pCur->aiIdx[pCur->iPage] = (u16)lwr; | |
5262 rc = moveToChild(pCur, chldPg); | |
5263 if( rc ) break; | |
5264 } | |
5265 moveto_finish: | |
5266 pCur->info.nSize = 0; | |
5267 pCur->curFlags &= ~(BTCF_ValidNKey|BTCF_ValidOvfl); | |
5268 return rc; | |
5269 } | |
5270 | |
5271 | |
5272 /* | |
5273 ** Return TRUE if the cursor is not pointing at an entry of the table. | |
5274 ** | |
5275 ** TRUE will be returned after a call to sqlite3BtreeNext() moves | |
5276 ** past the last entry in the table or sqlite3BtreePrev() moves past | |
5277 ** the first entry. TRUE is also returned if the table is empty. | |
5278 */ | |
5279 int sqlite3BtreeEof(BtCursor *pCur){ | |
5280 /* TODO: What if the cursor is in CURSOR_REQUIRESEEK but all table entries | |
5281 ** have been deleted? This API will need to change to return an error code | |
5282 ** as well as the boolean result value. | |
5283 */ | |
5284 return (CURSOR_VALID!=pCur->eState); | |
5285 } | |
5286 | |
5287 /* | |
5288 ** Advance the cursor to the next entry in the database. If | |
5289 ** successful then set *pRes=0. If the cursor | |
5290 ** was already pointing to the last entry in the database before | |
5291 ** this routine was called, then set *pRes=1. | |
5292 ** | |
5293 ** The main entry point is sqlite3BtreeNext(). That routine is optimized | |
5294 ** for the common case of merely incrementing the cell counter BtCursor.aiIdx | |
5295 ** to the next cell on the current page. The (slower) btreeNext() helper | |
5296 ** routine is called when it is necessary to move to a different page or | |
5297 ** to restore the cursor. | |
5298 ** | |
5299 ** The calling function will set *pRes to 0 or 1. The initial *pRes value | |
5300 ** will be 1 if the cursor being stepped corresponds to an SQL index and | |
5301 ** if this routine could have been skipped if that SQL index had been | |
5302 ** a unique index. Otherwise the caller will have set *pRes to zero. | |
5303 ** Zero is the common case. The btree implementation is free to use the | |
5304 ** initial *pRes value as a hint to improve performance, but the current | |
5305 ** SQLite btree implementation does not. (Note that the comdb2 btree | |
5306 ** implementation does use this hint, however.) | |
5307 */ | |
5308 static SQLITE_NOINLINE int btreeNext(BtCursor *pCur, int *pRes){ | |
5309 int rc; | |
5310 int idx; | |
5311 MemPage *pPage; | |
5312 | |
5313 assert( cursorHoldsMutex(pCur) ); | |
5314 assert( pCur->skipNext==0 || pCur->eState!=CURSOR_VALID ); | |
5315 assert( *pRes==0 ); | |
5316 if( pCur->eState!=CURSOR_VALID ){ | |
5317 assert( (pCur->curFlags & BTCF_ValidOvfl)==0 ); | |
5318 rc = restoreCursorPosition(pCur); | |
5319 if( rc!=SQLITE_OK ){ | |
5320 return rc; | |
5321 } | |
5322 if( CURSOR_INVALID==pCur->eState ){ | |
5323 *pRes = 1; | |
5324 return SQLITE_OK; | |
5325 } | |
5326 if( pCur->skipNext ){ | |
5327 assert( pCur->eState==CURSOR_VALID || pCur->eState==CURSOR_SKIPNEXT ); | |
5328 pCur->eState = CURSOR_VALID; | |
5329 if( pCur->skipNext>0 ){ | |
5330 pCur->skipNext = 0; | |
5331 return SQLITE_OK; | |
5332 } | |
5333 pCur->skipNext = 0; | |
5334 } | |
5335 } | |
5336 | |
5337 pPage = pCur->apPage[pCur->iPage]; | |
5338 idx = ++pCur->aiIdx[pCur->iPage]; | |
5339 assert( pPage->isInit ); | |
5340 | |
5341 /* If the database file is corrupt, it is possible for the value of idx | |
5342 ** to be invalid here. This can only occur if a second cursor modifies | |
5343 ** the page while cursor pCur is holding a reference to it. Which can | |
5344 ** only happen if the database is corrupt in such a way as to link the | |
5345 ** page into more than one b-tree structure. */ | |
5346 testcase( idx>pPage->nCell ); | |
5347 | |
5348 if( idx>=pPage->nCell ){ | |
5349 if( !pPage->leaf ){ | |
5350 rc = moveToChild(pCur, get4byte(&pPage->aData[pPage->hdrOffset+8])); | |
5351 if( rc ) return rc; | |
5352 return moveToLeftmost(pCur); | |
5353 } | |
5354 do{ | |
5355 if( pCur->iPage==0 ){ | |
5356 *pRes = 1; | |
5357 pCur->eState = CURSOR_INVALID; | |
5358 return SQLITE_OK; | |
5359 } | |
5360 moveToParent(pCur); | |
5361 pPage = pCur->apPage[pCur->iPage]; | |
5362 }while( pCur->aiIdx[pCur->iPage]>=pPage->nCell ); | |
5363 if( pPage->intKey ){ | |
5364 return sqlite3BtreeNext(pCur, pRes); | |
5365 }else{ | |
5366 return SQLITE_OK; | |
5367 } | |
5368 } | |
5369 if( pPage->leaf ){ | |
5370 return SQLITE_OK; | |
5371 }else{ | |
5372 return moveToLeftmost(pCur); | |
5373 } | |
5374 } | |
5375 int sqlite3BtreeNext(BtCursor *pCur, int *pRes){ | |
5376 MemPage *pPage; | |
5377 assert( cursorHoldsMutex(pCur) ); | |
5378 assert( pRes!=0 ); | |
5379 assert( *pRes==0 || *pRes==1 ); | |
5380 assert( pCur->skipNext==0 || pCur->eState!=CURSOR_VALID ); | |
5381 pCur->info.nSize = 0; | |
5382 pCur->curFlags &= ~(BTCF_ValidNKey|BTCF_ValidOvfl); | |
5383 *pRes = 0; | |
5384 if( pCur->eState!=CURSOR_VALID ) return btreeNext(pCur, pRes); | |
5385 pPage = pCur->apPage[pCur->iPage]; | |
5386 if( (++pCur->aiIdx[pCur->iPage])>=pPage->nCell ){ | |
5387 pCur->aiIdx[pCur->iPage]--; | |
5388 return btreeNext(pCur, pRes); | |
5389 } | |
5390 if( pPage->leaf ){ | |
5391 return SQLITE_OK; | |
5392 }else{ | |
5393 return moveToLeftmost(pCur); | |
5394 } | |
5395 } | |
5396 | |
5397 /* | |
5398 ** Step the cursor to the back to the previous entry in the database. If | |
5399 ** successful then set *pRes=0. If the cursor | |
5400 ** was already pointing to the first entry in the database before | |
5401 ** this routine was called, then set *pRes=1. | |
5402 ** | |
5403 ** The main entry point is sqlite3BtreePrevious(). That routine is optimized | |
5404 ** for the common case of merely decrementing the cell counter BtCursor.aiIdx | |
5405 ** to the previous cell on the current page. The (slower) btreePrevious() | |
5406 ** helper routine is called when it is necessary to move to a different page | |
5407 ** or to restore the cursor. | |
5408 ** | |
5409 ** The calling function will set *pRes to 0 or 1. The initial *pRes value | |
5410 ** will be 1 if the cursor being stepped corresponds to an SQL index and | |
5411 ** if this routine could have been skipped if that SQL index had been | |
5412 ** a unique index. Otherwise the caller will have set *pRes to zero. | |
5413 ** Zero is the common case. The btree implementation is free to use the | |
5414 ** initial *pRes value as a hint to improve performance, but the current | |
5415 ** SQLite btree implementation does not. (Note that the comdb2 btree | |
5416 ** implementation does use this hint, however.) | |
5417 */ | |
5418 static SQLITE_NOINLINE int btreePrevious(BtCursor *pCur, int *pRes){ | |
5419 int rc; | |
5420 MemPage *pPage; | |
5421 | |
5422 assert( cursorHoldsMutex(pCur) ); | |
5423 assert( pRes!=0 ); | |
5424 assert( *pRes==0 ); | |
5425 assert( pCur->skipNext==0 || pCur->eState!=CURSOR_VALID ); | |
5426 assert( (pCur->curFlags & (BTCF_AtLast|BTCF_ValidOvfl|BTCF_ValidNKey))==0 ); | |
5427 assert( pCur->info.nSize==0 ); | |
5428 if( pCur->eState!=CURSOR_VALID ){ | |
5429 rc = restoreCursorPosition(pCur); | |
5430 if( rc!=SQLITE_OK ){ | |
5431 return rc; | |
5432 } | |
5433 if( CURSOR_INVALID==pCur->eState ){ | |
5434 *pRes = 1; | |
5435 return SQLITE_OK; | |
5436 } | |
5437 if( pCur->skipNext ){ | |
5438 assert( pCur->eState==CURSOR_VALID || pCur->eState==CURSOR_SKIPNEXT ); | |
5439 pCur->eState = CURSOR_VALID; | |
5440 if( pCur->skipNext<0 ){ | |
5441 pCur->skipNext = 0; | |
5442 return SQLITE_OK; | |
5443 } | |
5444 pCur->skipNext = 0; | |
5445 } | |
5446 } | |
5447 | |
5448 pPage = pCur->apPage[pCur->iPage]; | |
5449 assert( pPage->isInit ); | |
5450 if( !pPage->leaf ){ | |
5451 int idx = pCur->aiIdx[pCur->iPage]; | |
5452 rc = moveToChild(pCur, get4byte(findCell(pPage, idx))); | |
5453 if( rc ) return rc; | |
5454 rc = moveToRightmost(pCur); | |
5455 }else{ | |
5456 while( pCur->aiIdx[pCur->iPage]==0 ){ | |
5457 if( pCur->iPage==0 ){ | |
5458 pCur->eState = CURSOR_INVALID; | |
5459 *pRes = 1; | |
5460 return SQLITE_OK; | |
5461 } | |
5462 moveToParent(pCur); | |
5463 } | |
5464 assert( pCur->info.nSize==0 ); | |
5465 assert( (pCur->curFlags & (BTCF_ValidNKey|BTCF_ValidOvfl))==0 ); | |
5466 | |
5467 pCur->aiIdx[pCur->iPage]--; | |
5468 pPage = pCur->apPage[pCur->iPage]; | |
5469 if( pPage->intKey && !pPage->leaf ){ | |
5470 rc = sqlite3BtreePrevious(pCur, pRes); | |
5471 }else{ | |
5472 rc = SQLITE_OK; | |
5473 } | |
5474 } | |
5475 return rc; | |
5476 } | |
5477 int sqlite3BtreePrevious(BtCursor *pCur, int *pRes){ | |
5478 assert( cursorHoldsMutex(pCur) ); | |
5479 assert( pRes!=0 ); | |
5480 assert( *pRes==0 || *pRes==1 ); | |
5481 assert( pCur->skipNext==0 || pCur->eState!=CURSOR_VALID ); | |
5482 *pRes = 0; | |
5483 pCur->curFlags &= ~(BTCF_AtLast|BTCF_ValidOvfl|BTCF_ValidNKey); | |
5484 pCur->info.nSize = 0; | |
5485 if( pCur->eState!=CURSOR_VALID | |
5486 || pCur->aiIdx[pCur->iPage]==0 | |
5487 || pCur->apPage[pCur->iPage]->leaf==0 | |
5488 ){ | |
5489 return btreePrevious(pCur, pRes); | |
5490 } | |
5491 pCur->aiIdx[pCur->iPage]--; | |
5492 return SQLITE_OK; | |
5493 } | |
5494 | |
5495 /* | |
5496 ** Allocate a new page from the database file. | |
5497 ** | |
5498 ** The new page is marked as dirty. (In other words, sqlite3PagerWrite() | |
5499 ** has already been called on the new page.) The new page has also | |
5500 ** been referenced and the calling routine is responsible for calling | |
5501 ** sqlite3PagerUnref() on the new page when it is done. | |
5502 ** | |
5503 ** SQLITE_OK is returned on success. Any other return value indicates | |
5504 ** an error. *ppPage is set to NULL in the event of an error. | |
5505 ** | |
5506 ** If the "nearby" parameter is not 0, then an effort is made to | |
5507 ** locate a page close to the page number "nearby". This can be used in an | |
5508 ** attempt to keep related pages close to each other in the database file, | |
5509 ** which in turn can make database access faster. | |
5510 ** | |
5511 ** If the eMode parameter is BTALLOC_EXACT and the nearby page exists | |
5512 ** anywhere on the free-list, then it is guaranteed to be returned. If | |
5513 ** eMode is BTALLOC_LT then the page returned will be less than or equal | |
5514 ** to nearby if any such page exists. If eMode is BTALLOC_ANY then there | |
5515 ** are no restrictions on which page is returned. | |
5516 */ | |
5517 static int allocateBtreePage( | |
5518 BtShared *pBt, /* The btree */ | |
5519 MemPage **ppPage, /* Store pointer to the allocated page here */ | |
5520 Pgno *pPgno, /* Store the page number here */ | |
5521 Pgno nearby, /* Search for a page near this one */ | |
5522 u8 eMode /* BTALLOC_EXACT, BTALLOC_LT, or BTALLOC_ANY */ | |
5523 ){ | |
5524 MemPage *pPage1; | |
5525 int rc; | |
5526 u32 n; /* Number of pages on the freelist */ | |
5527 u32 k; /* Number of leaves on the trunk of the freelist */ | |
5528 MemPage *pTrunk = 0; | |
5529 MemPage *pPrevTrunk = 0; | |
5530 Pgno mxPage; /* Total size of the database file */ | |
5531 | |
5532 assert( sqlite3_mutex_held(pBt->mutex) ); | |
5533 assert( eMode==BTALLOC_ANY || (nearby>0 && IfNotOmitAV(pBt->autoVacuum)) ); | |
5534 pPage1 = pBt->pPage1; | |
5535 mxPage = btreePagecount(pBt); | |
5536 /* EVIDENCE-OF: R-05119-02637 The 4-byte big-endian integer at offset 36 | |
5537 ** stores stores the total number of pages on the freelist. */ | |
5538 n = get4byte(&pPage1->aData[36]); | |
5539 testcase( n==mxPage-1 ); | |
5540 if( n>=mxPage ){ | |
5541 return SQLITE_CORRUPT_BKPT; | |
5542 } | |
5543 if( n>0 ){ | |
5544 /* There are pages on the freelist. Reuse one of those pages. */ | |
5545 Pgno iTrunk; | |
5546 u8 searchList = 0; /* If the free-list must be searched for 'nearby' */ | |
5547 u32 nSearch = 0; /* Count of the number of search attempts */ | |
5548 | |
5549 /* If eMode==BTALLOC_EXACT and a query of the pointer-map | |
5550 ** shows that the page 'nearby' is somewhere on the free-list, then | |
5551 ** the entire-list will be searched for that page. | |
5552 */ | |
5553 #ifndef SQLITE_OMIT_AUTOVACUUM | |
5554 if( eMode==BTALLOC_EXACT ){ | |
5555 if( nearby<=mxPage ){ | |
5556 u8 eType; | |
5557 assert( nearby>0 ); | |
5558 assert( pBt->autoVacuum ); | |
5559 rc = ptrmapGet(pBt, nearby, &eType, 0); | |
5560 if( rc ) return rc; | |
5561 if( eType==PTRMAP_FREEPAGE ){ | |
5562 searchList = 1; | |
5563 } | |
5564 } | |
5565 }else if( eMode==BTALLOC_LE ){ | |
5566 searchList = 1; | |
5567 } | |
5568 #endif | |
5569 | |
5570 /* Decrement the free-list count by 1. Set iTrunk to the index of the | |
5571 ** first free-list trunk page. iPrevTrunk is initially 1. | |
5572 */ | |
5573 rc = sqlite3PagerWrite(pPage1->pDbPage); | |
5574 if( rc ) return rc; | |
5575 put4byte(&pPage1->aData[36], n-1); | |
5576 | |
5577 /* The code within this loop is run only once if the 'searchList' variable | |
5578 ** is not true. Otherwise, it runs once for each trunk-page on the | |
5579 ** free-list until the page 'nearby' is located (eMode==BTALLOC_EXACT) | |
5580 ** or until a page less than 'nearby' is located (eMode==BTALLOC_LT) | |
5581 */ | |
5582 do { | |
5583 pPrevTrunk = pTrunk; | |
5584 if( pPrevTrunk ){ | |
5585 /* EVIDENCE-OF: R-01506-11053 The first integer on a freelist trunk page | |
5586 ** is the page number of the next freelist trunk page in the list or | |
5587 ** zero if this is the last freelist trunk page. */ | |
5588 iTrunk = get4byte(&pPrevTrunk->aData[0]); | |
5589 }else{ | |
5590 /* EVIDENCE-OF: R-59841-13798 The 4-byte big-endian integer at offset 32 | |
5591 ** stores the page number of the first page of the freelist, or zero if | |
5592 ** the freelist is empty. */ | |
5593 iTrunk = get4byte(&pPage1->aData[32]); | |
5594 } | |
5595 testcase( iTrunk==mxPage ); | |
5596 if( iTrunk>mxPage || nSearch++ > n ){ | |
5597 rc = SQLITE_CORRUPT_BKPT; | |
5598 }else{ | |
5599 rc = btreeGetUnusedPage(pBt, iTrunk, &pTrunk, 0); | |
5600 } | |
5601 if( rc ){ | |
5602 pTrunk = 0; | |
5603 goto end_allocate_page; | |
5604 } | |
5605 assert( pTrunk!=0 ); | |
5606 assert( pTrunk->aData!=0 ); | |
5607 /* EVIDENCE-OF: R-13523-04394 The second integer on a freelist trunk page | |
5608 ** is the number of leaf page pointers to follow. */ | |
5609 k = get4byte(&pTrunk->aData[4]); | |
5610 if( k==0 && !searchList ){ | |
5611 /* The trunk has no leaves and the list is not being searched. | |
5612 ** So extract the trunk page itself and use it as the newly | |
5613 ** allocated page */ | |
5614 assert( pPrevTrunk==0 ); | |
5615 rc = sqlite3PagerWrite(pTrunk->pDbPage); | |
5616 if( rc ){ | |
5617 goto end_allocate_page; | |
5618 } | |
5619 *pPgno = iTrunk; | |
5620 memcpy(&pPage1->aData[32], &pTrunk->aData[0], 4); | |
5621 *ppPage = pTrunk; | |
5622 pTrunk = 0; | |
5623 TRACE(("ALLOCATE: %d trunk - %d free pages left\n", *pPgno, n-1)); | |
5624 }else if( k>(u32)(pBt->usableSize/4 - 2) ){ | |
5625 /* Value of k is out of range. Database corruption */ | |
5626 rc = SQLITE_CORRUPT_BKPT; | |
5627 goto end_allocate_page; | |
5628 #ifndef SQLITE_OMIT_AUTOVACUUM | |
5629 }else if( searchList | |
5630 && (nearby==iTrunk || (iTrunk<nearby && eMode==BTALLOC_LE)) | |
5631 ){ | |
5632 /* The list is being searched and this trunk page is the page | |
5633 ** to allocate, regardless of whether it has leaves. | |
5634 */ | |
5635 *pPgno = iTrunk; | |
5636 *ppPage = pTrunk; | |
5637 searchList = 0; | |
5638 rc = sqlite3PagerWrite(pTrunk->pDbPage); | |
5639 if( rc ){ | |
5640 goto end_allocate_page; | |
5641 } | |
5642 if( k==0 ){ | |
5643 if( !pPrevTrunk ){ | |
5644 memcpy(&pPage1->aData[32], &pTrunk->aData[0], 4); | |
5645 }else{ | |
5646 rc = sqlite3PagerWrite(pPrevTrunk->pDbPage); | |
5647 if( rc!=SQLITE_OK ){ | |
5648 goto end_allocate_page; | |
5649 } | |
5650 memcpy(&pPrevTrunk->aData[0], &pTrunk->aData[0], 4); | |
5651 } | |
5652 }else{ | |
5653 /* The trunk page is required by the caller but it contains | |
5654 ** pointers to free-list leaves. The first leaf becomes a trunk | |
5655 ** page in this case. | |
5656 */ | |
5657 MemPage *pNewTrunk; | |
5658 Pgno iNewTrunk = get4byte(&pTrunk->aData[8]); | |
5659 if( iNewTrunk>mxPage ){ | |
5660 rc = SQLITE_CORRUPT_BKPT; | |
5661 goto end_allocate_page; | |
5662 } | |
5663 testcase( iNewTrunk==mxPage ); | |
5664 rc = btreeGetUnusedPage(pBt, iNewTrunk, &pNewTrunk, 0); | |
5665 if( rc!=SQLITE_OK ){ | |
5666 goto end_allocate_page; | |
5667 } | |
5668 rc = sqlite3PagerWrite(pNewTrunk->pDbPage); | |
5669 if( rc!=SQLITE_OK ){ | |
5670 releasePage(pNewTrunk); | |
5671 goto end_allocate_page; | |
5672 } | |
5673 memcpy(&pNewTrunk->aData[0], &pTrunk->aData[0], 4); | |
5674 put4byte(&pNewTrunk->aData[4], k-1); | |
5675 memcpy(&pNewTrunk->aData[8], &pTrunk->aData[12], (k-1)*4); | |
5676 releasePage(pNewTrunk); | |
5677 if( !pPrevTrunk ){ | |
5678 assert( sqlite3PagerIswriteable(pPage1->pDbPage) ); | |
5679 put4byte(&pPage1->aData[32], iNewTrunk); | |
5680 }else{ | |
5681 rc = sqlite3PagerWrite(pPrevTrunk->pDbPage); | |
5682 if( rc ){ | |
5683 goto end_allocate_page; | |
5684 } | |
5685 put4byte(&pPrevTrunk->aData[0], iNewTrunk); | |
5686 } | |
5687 } | |
5688 pTrunk = 0; | |
5689 TRACE(("ALLOCATE: %d trunk - %d free pages left\n", *pPgno, n-1)); | |
5690 #endif | |
5691 }else if( k>0 ){ | |
5692 /* Extract a leaf from the trunk */ | |
5693 u32 closest; | |
5694 Pgno iPage; | |
5695 unsigned char *aData = pTrunk->aData; | |
5696 if( nearby>0 ){ | |
5697 u32 i; | |
5698 closest = 0; | |
5699 if( eMode==BTALLOC_LE ){ | |
5700 for(i=0; i<k; i++){ | |
5701 iPage = get4byte(&aData[8+i*4]); | |
5702 if( iPage<=nearby ){ | |
5703 closest = i; | |
5704 break; | |
5705 } | |
5706 } | |
5707 }else{ | |
5708 int dist; | |
5709 dist = sqlite3AbsInt32(get4byte(&aData[8]) - nearby); | |
5710 for(i=1; i<k; i++){ | |
5711 int d2 = sqlite3AbsInt32(get4byte(&aData[8+i*4]) - nearby); | |
5712 if( d2<dist ){ | |
5713 closest = i; | |
5714 dist = d2; | |
5715 } | |
5716 } | |
5717 } | |
5718 }else{ | |
5719 closest = 0; | |
5720 } | |
5721 | |
5722 iPage = get4byte(&aData[8+closest*4]); | |
5723 testcase( iPage==mxPage ); | |
5724 if( iPage>mxPage ){ | |
5725 rc = SQLITE_CORRUPT_BKPT; | |
5726 goto end_allocate_page; | |
5727 } | |
5728 testcase( iPage==mxPage ); | |
5729 if( !searchList | |
5730 || (iPage==nearby || (iPage<nearby && eMode==BTALLOC_LE)) | |
5731 ){ | |
5732 int noContent; | |
5733 *pPgno = iPage; | |
5734 TRACE(("ALLOCATE: %d was leaf %d of %d on trunk %d" | |
5735 ": %d more free pages\n", | |
5736 *pPgno, closest+1, k, pTrunk->pgno, n-1)); | |
5737 rc = sqlite3PagerWrite(pTrunk->pDbPage); | |
5738 if( rc ) goto end_allocate_page; | |
5739 if( closest<k-1 ){ | |
5740 memcpy(&aData[8+closest*4], &aData[4+k*4], 4); | |
5741 } | |
5742 put4byte(&aData[4], k-1); | |
5743 noContent = !btreeGetHasContent(pBt, *pPgno)? PAGER_GET_NOCONTENT : 0; | |
5744 rc = btreeGetUnusedPage(pBt, *pPgno, ppPage, noContent); | |
5745 if( rc==SQLITE_OK ){ | |
5746 rc = sqlite3PagerWrite((*ppPage)->pDbPage); | |
5747 if( rc!=SQLITE_OK ){ | |
5748 releasePage(*ppPage); | |
5749 *ppPage = 0; | |
5750 } | |
5751 } | |
5752 searchList = 0; | |
5753 } | |
5754 } | |
5755 releasePage(pPrevTrunk); | |
5756 pPrevTrunk = 0; | |
5757 }while( searchList ); | |
5758 }else{ | |
5759 /* There are no pages on the freelist, so append a new page to the | |
5760 ** database image. | |
5761 ** | |
5762 ** Normally, new pages allocated by this block can be requested from the | |
5763 ** pager layer with the 'no-content' flag set. This prevents the pager | |
5764 ** from trying to read the pages content from disk. However, if the | |
5765 ** current transaction has already run one or more incremental-vacuum | |
5766 ** steps, then the page we are about to allocate may contain content | |
5767 ** that is required in the event of a rollback. In this case, do | |
5768 ** not set the no-content flag. This causes the pager to load and journal | |
5769 ** the current page content before overwriting it. | |
5770 ** | |
5771 ** Note that the pager will not actually attempt to load or journal | |
5772 ** content for any page that really does lie past the end of the database | |
5773 ** file on disk. So the effects of disabling the no-content optimization | |
5774 ** here are confined to those pages that lie between the end of the | |
5775 ** database image and the end of the database file. | |
5776 */ | |
5777 int bNoContent = (0==IfNotOmitAV(pBt->bDoTruncate))? PAGER_GET_NOCONTENT:0; | |
5778 | |
5779 rc = sqlite3PagerWrite(pBt->pPage1->pDbPage); | |
5780 if( rc ) return rc; | |
5781 pBt->nPage++; | |
5782 if( pBt->nPage==PENDING_BYTE_PAGE(pBt) ) pBt->nPage++; | |
5783 | |
5784 #ifndef SQLITE_OMIT_AUTOVACUUM | |
5785 if( pBt->autoVacuum && PTRMAP_ISPAGE(pBt, pBt->nPage) ){ | |
5786 /* If *pPgno refers to a pointer-map page, allocate two new pages | |
5787 ** at the end of the file instead of one. The first allocated page | |
5788 ** becomes a new pointer-map page, the second is used by the caller. | |
5789 */ | |
5790 MemPage *pPg = 0; | |
5791 TRACE(("ALLOCATE: %d from end of file (pointer-map page)\n", pBt->nPage)); | |
5792 assert( pBt->nPage!=PENDING_BYTE_PAGE(pBt) ); | |
5793 rc = btreeGetUnusedPage(pBt, pBt->nPage, &pPg, bNoContent); | |
5794 if( rc==SQLITE_OK ){ | |
5795 rc = sqlite3PagerWrite(pPg->pDbPage); | |
5796 releasePage(pPg); | |
5797 } | |
5798 if( rc ) return rc; | |
5799 pBt->nPage++; | |
5800 if( pBt->nPage==PENDING_BYTE_PAGE(pBt) ){ pBt->nPage++; } | |
5801 } | |
5802 #endif | |
5803 put4byte(28 + (u8*)pBt->pPage1->aData, pBt->nPage); | |
5804 *pPgno = pBt->nPage; | |
5805 | |
5806 assert( *pPgno!=PENDING_BYTE_PAGE(pBt) ); | |
5807 rc = btreeGetUnusedPage(pBt, *pPgno, ppPage, bNoContent); | |
5808 if( rc ) return rc; | |
5809 rc = sqlite3PagerWrite((*ppPage)->pDbPage); | |
5810 if( rc!=SQLITE_OK ){ | |
5811 releasePage(*ppPage); | |
5812 *ppPage = 0; | |
5813 } | |
5814 TRACE(("ALLOCATE: %d from end of file\n", *pPgno)); | |
5815 } | |
5816 | |
5817 assert( *pPgno!=PENDING_BYTE_PAGE(pBt) ); | |
5818 | |
5819 end_allocate_page: | |
5820 releasePage(pTrunk); | |
5821 releasePage(pPrevTrunk); | |
5822 assert( rc!=SQLITE_OK || sqlite3PagerPageRefcount((*ppPage)->pDbPage)<=1 ); | |
5823 assert( rc!=SQLITE_OK || (*ppPage)->isInit==0 ); | |
5824 return rc; | |
5825 } | |
5826 | |
5827 /* | |
5828 ** This function is used to add page iPage to the database file free-list. | |
5829 ** It is assumed that the page is not already a part of the free-list. | |
5830 ** | |
5831 ** The value passed as the second argument to this function is optional. | |
5832 ** If the caller happens to have a pointer to the MemPage object | |
5833 ** corresponding to page iPage handy, it may pass it as the second value. | |
5834 ** Otherwise, it may pass NULL. | |
5835 ** | |
5836 ** If a pointer to a MemPage object is passed as the second argument, | |
5837 ** its reference count is not altered by this function. | |
5838 */ | |
5839 static int freePage2(BtShared *pBt, MemPage *pMemPage, Pgno iPage){ | |
5840 MemPage *pTrunk = 0; /* Free-list trunk page */ | |
5841 Pgno iTrunk = 0; /* Page number of free-list trunk page */ | |
5842 MemPage *pPage1 = pBt->pPage1; /* Local reference to page 1 */ | |
5843 MemPage *pPage; /* Page being freed. May be NULL. */ | |
5844 int rc; /* Return Code */ | |
5845 int nFree; /* Initial number of pages on free-list */ | |
5846 | |
5847 assert( sqlite3_mutex_held(pBt->mutex) ); | |
5848 assert( CORRUPT_DB || iPage>1 ); | |
5849 assert( !pMemPage || pMemPage->pgno==iPage ); | |
5850 | |
5851 if( iPage<2 ) return SQLITE_CORRUPT_BKPT; | |
5852 if( pMemPage ){ | |
5853 pPage = pMemPage; | |
5854 sqlite3PagerRef(pPage->pDbPage); | |
5855 }else{ | |
5856 pPage = btreePageLookup(pBt, iPage); | |
5857 } | |
5858 | |
5859 /* Increment the free page count on pPage1 */ | |
5860 rc = sqlite3PagerWrite(pPage1->pDbPage); | |
5861 if( rc ) goto freepage_out; | |
5862 nFree = get4byte(&pPage1->aData[36]); | |
5863 put4byte(&pPage1->aData[36], nFree+1); | |
5864 | |
5865 if( pBt->btsFlags & BTS_SECURE_DELETE ){ | |
5866 /* If the secure_delete option is enabled, then | |
5867 ** always fully overwrite deleted information with zeros. | |
5868 */ | |
5869 if( (!pPage && ((rc = btreeGetPage(pBt, iPage, &pPage, 0))!=0) ) | |
5870 || ((rc = sqlite3PagerWrite(pPage->pDbPage))!=0) | |
5871 ){ | |
5872 goto freepage_out; | |
5873 } | |
5874 memset(pPage->aData, 0, pPage->pBt->pageSize); | |
5875 } | |
5876 | |
5877 /* If the database supports auto-vacuum, write an entry in the pointer-map | |
5878 ** to indicate that the page is free. | |
5879 */ | |
5880 if( ISAUTOVACUUM ){ | |
5881 ptrmapPut(pBt, iPage, PTRMAP_FREEPAGE, 0, &rc); | |
5882 if( rc ) goto freepage_out; | |
5883 } | |
5884 | |
5885 /* Now manipulate the actual database free-list structure. There are two | |
5886 ** possibilities. If the free-list is currently empty, or if the first | |
5887 ** trunk page in the free-list is full, then this page will become a | |
5888 ** new free-list trunk page. Otherwise, it will become a leaf of the | |
5889 ** first trunk page in the current free-list. This block tests if it | |
5890 ** is possible to add the page as a new free-list leaf. | |
5891 */ | |
5892 if( nFree!=0 ){ | |
5893 u32 nLeaf; /* Initial number of leaf cells on trunk page */ | |
5894 | |
5895 iTrunk = get4byte(&pPage1->aData[32]); | |
5896 rc = btreeGetPage(pBt, iTrunk, &pTrunk, 0); | |
5897 if( rc!=SQLITE_OK ){ | |
5898 goto freepage_out; | |
5899 } | |
5900 | |
5901 nLeaf = get4byte(&pTrunk->aData[4]); | |
5902 assert( pBt->usableSize>32 ); | |
5903 if( nLeaf > (u32)pBt->usableSize/4 - 2 ){ | |
5904 rc = SQLITE_CORRUPT_BKPT; | |
5905 goto freepage_out; | |
5906 } | |
5907 if( nLeaf < (u32)pBt->usableSize/4 - 8 ){ | |
5908 /* In this case there is room on the trunk page to insert the page | |
5909 ** being freed as a new leaf. | |
5910 ** | |
5911 ** Note that the trunk page is not really full until it contains | |
5912 ** usableSize/4 - 2 entries, not usableSize/4 - 8 entries as we have | |
5913 ** coded. But due to a coding error in versions of SQLite prior to | |
5914 ** 3.6.0, databases with freelist trunk pages holding more than | |
5915 ** usableSize/4 - 8 entries will be reported as corrupt. In order | |
5916 ** to maintain backwards compatibility with older versions of SQLite, | |
5917 ** we will continue to restrict the number of entries to usableSize/4 - 8 | |
5918 ** for now. At some point in the future (once everyone has upgraded | |
5919 ** to 3.6.0 or later) we should consider fixing the conditional above | |
5920 ** to read "usableSize/4-2" instead of "usableSize/4-8". | |
5921 ** | |
5922 ** EVIDENCE-OF: R-19920-11576 However, newer versions of SQLite still | |
5923 ** avoid using the last six entries in the freelist trunk page array in | |
5924 ** order that database files created by newer versions of SQLite can be | |
5925 ** read by older versions of SQLite. | |
5926 */ | |
5927 rc = sqlite3PagerWrite(pTrunk->pDbPage); | |
5928 if( rc==SQLITE_OK ){ | |
5929 put4byte(&pTrunk->aData[4], nLeaf+1); | |
5930 put4byte(&pTrunk->aData[8+nLeaf*4], iPage); | |
5931 if( pPage && (pBt->btsFlags & BTS_SECURE_DELETE)==0 ){ | |
5932 sqlite3PagerDontWrite(pPage->pDbPage); | |
5933 } | |
5934 rc = btreeSetHasContent(pBt, iPage); | |
5935 } | |
5936 TRACE(("FREE-PAGE: %d leaf on trunk page %d\n",pPage->pgno,pTrunk->pgno)); | |
5937 goto freepage_out; | |
5938 } | |
5939 } | |
5940 | |
5941 /* If control flows to this point, then it was not possible to add the | |
5942 ** the page being freed as a leaf page of the first trunk in the free-list. | |
5943 ** Possibly because the free-list is empty, or possibly because the | |
5944 ** first trunk in the free-list is full. Either way, the page being freed | |
5945 ** will become the new first trunk page in the free-list. | |
5946 */ | |
5947 if( pPage==0 && SQLITE_OK!=(rc = btreeGetPage(pBt, iPage, &pPage, 0)) ){ | |
5948 goto freepage_out; | |
5949 } | |
5950 rc = sqlite3PagerWrite(pPage->pDbPage); | |
5951 if( rc!=SQLITE_OK ){ | |
5952 goto freepage_out; | |
5953 } | |
5954 put4byte(pPage->aData, iTrunk); | |
5955 put4byte(&pPage->aData[4], 0); | |
5956 put4byte(&pPage1->aData[32], iPage); | |
5957 TRACE(("FREE-PAGE: %d new trunk page replacing %d\n", pPage->pgno, iTrunk)); | |
5958 | |
5959 freepage_out: | |
5960 if( pPage ){ | |
5961 pPage->isInit = 0; | |
5962 } | |
5963 releasePage(pPage); | |
5964 releasePage(pTrunk); | |
5965 return rc; | |
5966 } | |
5967 static void freePage(MemPage *pPage, int *pRC){ | |
5968 if( (*pRC)==SQLITE_OK ){ | |
5969 *pRC = freePage2(pPage->pBt, pPage, pPage->pgno); | |
5970 } | |
5971 } | |
5972 | |
5973 /* | |
5974 ** Free any overflow pages associated with the given Cell. Write the | |
5975 ** local Cell size (the number of bytes on the original page, omitting | |
5976 ** overflow) into *pnSize. | |
5977 */ | |
5978 static int clearCell( | |
5979 MemPage *pPage, /* The page that contains the Cell */ | |
5980 unsigned char *pCell, /* First byte of the Cell */ | |
5981 u16 *pnSize /* Write the size of the Cell here */ | |
5982 ){ | |
5983 BtShared *pBt = pPage->pBt; | |
5984 CellInfo info; | |
5985 Pgno ovflPgno; | |
5986 int rc; | |
5987 int nOvfl; | |
5988 u32 ovflPageSize; | |
5989 | |
5990 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
5991 pPage->xParseCell(pPage, pCell, &info); | |
5992 *pnSize = info.nSize; | |
5993 if( info.nLocal==info.nPayload ){ | |
5994 return SQLITE_OK; /* No overflow pages. Return without doing anything */ | |
5995 } | |
5996 if( pCell+info.nSize-1 > pPage->aData+pPage->maskPage ){ | |
5997 return SQLITE_CORRUPT_BKPT; /* Cell extends past end of page */ | |
5998 } | |
5999 ovflPgno = get4byte(pCell + info.nSize - 4); | |
6000 assert( pBt->usableSize > 4 ); | |
6001 ovflPageSize = pBt->usableSize - 4; | |
6002 nOvfl = (info.nPayload - info.nLocal + ovflPageSize - 1)/ovflPageSize; | |
6003 assert( nOvfl>0 || | |
6004 (CORRUPT_DB && (info.nPayload + ovflPageSize)<ovflPageSize) | |
6005 ); | |
6006 while( nOvfl-- ){ | |
6007 Pgno iNext = 0; | |
6008 MemPage *pOvfl = 0; | |
6009 if( ovflPgno<2 || ovflPgno>btreePagecount(pBt) ){ | |
6010 /* 0 is not a legal page number and page 1 cannot be an | |
6011 ** overflow page. Therefore if ovflPgno<2 or past the end of the | |
6012 ** file the database must be corrupt. */ | |
6013 return SQLITE_CORRUPT_BKPT; | |
6014 } | |
6015 if( nOvfl ){ | |
6016 rc = getOverflowPage(pBt, ovflPgno, &pOvfl, &iNext); | |
6017 if( rc ) return rc; | |
6018 } | |
6019 | |
6020 if( ( pOvfl || ((pOvfl = btreePageLookup(pBt, ovflPgno))!=0) ) | |
6021 && sqlite3PagerPageRefcount(pOvfl->pDbPage)!=1 | |
6022 ){ | |
6023 /* There is no reason any cursor should have an outstanding reference | |
6024 ** to an overflow page belonging to a cell that is being deleted/updated. | |
6025 ** So if there exists more than one reference to this page, then it | |
6026 ** must not really be an overflow page and the database must be corrupt. | |
6027 ** It is helpful to detect this before calling freePage2(), as | |
6028 ** freePage2() may zero the page contents if secure-delete mode is | |
6029 ** enabled. If this 'overflow' page happens to be a page that the | |
6030 ** caller is iterating through or using in some other way, this | |
6031 ** can be problematic. | |
6032 */ | |
6033 rc = SQLITE_CORRUPT_BKPT; | |
6034 }else{ | |
6035 rc = freePage2(pBt, pOvfl, ovflPgno); | |
6036 } | |
6037 | |
6038 if( pOvfl ){ | |
6039 sqlite3PagerUnref(pOvfl->pDbPage); | |
6040 } | |
6041 if( rc ) return rc; | |
6042 ovflPgno = iNext; | |
6043 } | |
6044 return SQLITE_OK; | |
6045 } | |
6046 | |
6047 /* | |
6048 ** Create the byte sequence used to represent a cell on page pPage | |
6049 ** and write that byte sequence into pCell[]. Overflow pages are | |
6050 ** allocated and filled in as necessary. The calling procedure | |
6051 ** is responsible for making sure sufficient space has been allocated | |
6052 ** for pCell[]. | |
6053 ** | |
6054 ** Note that pCell does not necessary need to point to the pPage->aData | |
6055 ** area. pCell might point to some temporary storage. The cell will | |
6056 ** be constructed in this temporary area then copied into pPage->aData | |
6057 ** later. | |
6058 */ | |
6059 static int fillInCell( | |
6060 MemPage *pPage, /* The page that contains the cell */ | |
6061 unsigned char *pCell, /* Complete text of the cell */ | |
6062 const void *pKey, i64 nKey, /* The key */ | |
6063 const void *pData,int nData, /* The data */ | |
6064 int nZero, /* Extra zero bytes to append to pData */ | |
6065 int *pnSize /* Write cell size here */ | |
6066 ){ | |
6067 int nPayload; | |
6068 const u8 *pSrc; | |
6069 int nSrc, n, rc; | |
6070 int spaceLeft; | |
6071 MemPage *pOvfl = 0; | |
6072 MemPage *pToRelease = 0; | |
6073 unsigned char *pPrior; | |
6074 unsigned char *pPayload; | |
6075 BtShared *pBt = pPage->pBt; | |
6076 Pgno pgnoOvfl = 0; | |
6077 int nHeader; | |
6078 | |
6079 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
6080 | |
6081 /* pPage is not necessarily writeable since pCell might be auxiliary | |
6082 ** buffer space that is separate from the pPage buffer area */ | |
6083 assert( pCell<pPage->aData || pCell>=&pPage->aData[pBt->pageSize] | |
6084 || sqlite3PagerIswriteable(pPage->pDbPage) ); | |
6085 | |
6086 /* Fill in the header. */ | |
6087 nHeader = pPage->childPtrSize; | |
6088 nPayload = nData + nZero; | |
6089 if( pPage->intKeyLeaf ){ | |
6090 nHeader += putVarint32(&pCell[nHeader], nPayload); | |
6091 }else{ | |
6092 assert( nData==0 ); | |
6093 assert( nZero==0 ); | |
6094 } | |
6095 nHeader += putVarint(&pCell[nHeader], *(u64*)&nKey); | |
6096 | |
6097 /* Fill in the payload size */ | |
6098 if( pPage->intKey ){ | |
6099 pSrc = pData; | |
6100 nSrc = nData; | |
6101 nData = 0; | |
6102 }else{ | |
6103 assert( nKey<=0x7fffffff && pKey!=0 ); | |
6104 nPayload = (int)nKey; | |
6105 pSrc = pKey; | |
6106 nSrc = (int)nKey; | |
6107 } | |
6108 if( nPayload<=pPage->maxLocal ){ | |
6109 n = nHeader + nPayload; | |
6110 testcase( n==3 ); | |
6111 testcase( n==4 ); | |
6112 if( n<4 ) n = 4; | |
6113 *pnSize = n; | |
6114 spaceLeft = nPayload; | |
6115 pPrior = pCell; | |
6116 }else{ | |
6117 int mn = pPage->minLocal; | |
6118 n = mn + (nPayload - mn) % (pPage->pBt->usableSize - 4); | |
6119 testcase( n==pPage->maxLocal ); | |
6120 testcase( n==pPage->maxLocal+1 ); | |
6121 if( n > pPage->maxLocal ) n = mn; | |
6122 spaceLeft = n; | |
6123 *pnSize = n + nHeader + 4; | |
6124 pPrior = &pCell[nHeader+n]; | |
6125 } | |
6126 pPayload = &pCell[nHeader]; | |
6127 | |
6128 /* At this point variables should be set as follows: | |
6129 ** | |
6130 ** nPayload Total payload size in bytes | |
6131 ** pPayload Begin writing payload here | |
6132 ** spaceLeft Space available at pPayload. If nPayload>spaceLeft, | |
6133 ** that means content must spill into overflow pages. | |
6134 ** *pnSize Size of the local cell (not counting overflow pages) | |
6135 ** pPrior Where to write the pgno of the first overflow page | |
6136 ** | |
6137 ** Use a call to btreeParseCellPtr() to verify that the values above | |
6138 ** were computed correctly. | |
6139 */ | |
6140 #if SQLITE_DEBUG | |
6141 { | |
6142 CellInfo info; | |
6143 pPage->xParseCell(pPage, pCell, &info); | |
6144 assert( nHeader=(int)(info.pPayload - pCell) ); | |
6145 assert( info.nKey==nKey ); | |
6146 assert( *pnSize == info.nSize ); | |
6147 assert( spaceLeft == info.nLocal ); | |
6148 } | |
6149 #endif | |
6150 | |
6151 /* Write the payload into the local Cell and any extra into overflow pages */ | |
6152 while( nPayload>0 ){ | |
6153 if( spaceLeft==0 ){ | |
6154 #ifndef SQLITE_OMIT_AUTOVACUUM | |
6155 Pgno pgnoPtrmap = pgnoOvfl; /* Overflow page pointer-map entry page */ | |
6156 if( pBt->autoVacuum ){ | |
6157 do{ | |
6158 pgnoOvfl++; | |
6159 } while( | |
6160 PTRMAP_ISPAGE(pBt, pgnoOvfl) || pgnoOvfl==PENDING_BYTE_PAGE(pBt) | |
6161 ); | |
6162 } | |
6163 #endif | |
6164 rc = allocateBtreePage(pBt, &pOvfl, &pgnoOvfl, pgnoOvfl, 0); | |
6165 #ifndef SQLITE_OMIT_AUTOVACUUM | |
6166 /* If the database supports auto-vacuum, and the second or subsequent | |
6167 ** overflow page is being allocated, add an entry to the pointer-map | |
6168 ** for that page now. | |
6169 ** | |
6170 ** If this is the first overflow page, then write a partial entry | |
6171 ** to the pointer-map. If we write nothing to this pointer-map slot, | |
6172 ** then the optimistic overflow chain processing in clearCell() | |
6173 ** may misinterpret the uninitialized values and delete the | |
6174 ** wrong pages from the database. | |
6175 */ | |
6176 if( pBt->autoVacuum && rc==SQLITE_OK ){ | |
6177 u8 eType = (pgnoPtrmap?PTRMAP_OVERFLOW2:PTRMAP_OVERFLOW1); | |
6178 ptrmapPut(pBt, pgnoOvfl, eType, pgnoPtrmap, &rc); | |
6179 if( rc ){ | |
6180 releasePage(pOvfl); | |
6181 } | |
6182 } | |
6183 #endif | |
6184 if( rc ){ | |
6185 releasePage(pToRelease); | |
6186 return rc; | |
6187 } | |
6188 | |
6189 /* If pToRelease is not zero than pPrior points into the data area | |
6190 ** of pToRelease. Make sure pToRelease is still writeable. */ | |
6191 assert( pToRelease==0 || sqlite3PagerIswriteable(pToRelease->pDbPage) ); | |
6192 | |
6193 /* If pPrior is part of the data area of pPage, then make sure pPage | |
6194 ** is still writeable */ | |
6195 assert( pPrior<pPage->aData || pPrior>=&pPage->aData[pBt->pageSize] | |
6196 || sqlite3PagerIswriteable(pPage->pDbPage) ); | |
6197 | |
6198 put4byte(pPrior, pgnoOvfl); | |
6199 releasePage(pToRelease); | |
6200 pToRelease = pOvfl; | |
6201 pPrior = pOvfl->aData; | |
6202 put4byte(pPrior, 0); | |
6203 pPayload = &pOvfl->aData[4]; | |
6204 spaceLeft = pBt->usableSize - 4; | |
6205 } | |
6206 n = nPayload; | |
6207 if( n>spaceLeft ) n = spaceLeft; | |
6208 | |
6209 /* If pToRelease is not zero than pPayload points into the data area | |
6210 ** of pToRelease. Make sure pToRelease is still writeable. */ | |
6211 assert( pToRelease==0 || sqlite3PagerIswriteable(pToRelease->pDbPage) ); | |
6212 | |
6213 /* If pPayload is part of the data area of pPage, then make sure pPage | |
6214 ** is still writeable */ | |
6215 assert( pPayload<pPage->aData || pPayload>=&pPage->aData[pBt->pageSize] | |
6216 || sqlite3PagerIswriteable(pPage->pDbPage) ); | |
6217 | |
6218 if( nSrc>0 ){ | |
6219 if( n>nSrc ) n = nSrc; | |
6220 assert( pSrc ); | |
6221 memcpy(pPayload, pSrc, n); | |
6222 }else{ | |
6223 memset(pPayload, 0, n); | |
6224 } | |
6225 nPayload -= n; | |
6226 pPayload += n; | |
6227 pSrc += n; | |
6228 nSrc -= n; | |
6229 spaceLeft -= n; | |
6230 if( nSrc==0 ){ | |
6231 nSrc = nData; | |
6232 pSrc = pData; | |
6233 } | |
6234 } | |
6235 releasePage(pToRelease); | |
6236 return SQLITE_OK; | |
6237 } | |
6238 | |
6239 /* | |
6240 ** Remove the i-th cell from pPage. This routine effects pPage only. | |
6241 ** The cell content is not freed or deallocated. It is assumed that | |
6242 ** the cell content has been copied someplace else. This routine just | |
6243 ** removes the reference to the cell from pPage. | |
6244 ** | |
6245 ** "sz" must be the number of bytes in the cell. | |
6246 */ | |
6247 static void dropCell(MemPage *pPage, int idx, int sz, int *pRC){ | |
6248 u32 pc; /* Offset to cell content of cell being deleted */ | |
6249 u8 *data; /* pPage->aData */ | |
6250 u8 *ptr; /* Used to move bytes around within data[] */ | |
6251 int rc; /* The return code */ | |
6252 int hdr; /* Beginning of the header. 0 most pages. 100 page 1 */ | |
6253 | |
6254 if( *pRC ) return; | |
6255 | |
6256 assert( idx>=0 && idx<pPage->nCell ); | |
6257 assert( CORRUPT_DB || sz==cellSize(pPage, idx) ); | |
6258 assert( sqlite3PagerIswriteable(pPage->pDbPage) ); | |
6259 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
6260 data = pPage->aData; | |
6261 ptr = &pPage->aCellIdx[2*idx]; | |
6262 pc = get2byte(ptr); | |
6263 hdr = pPage->hdrOffset; | |
6264 testcase( pc==get2byte(&data[hdr+5]) ); | |
6265 testcase( pc+sz==pPage->pBt->usableSize ); | |
6266 if( pc < (u32)get2byte(&data[hdr+5]) || pc+sz > pPage->pBt->usableSize ){ | |
6267 *pRC = SQLITE_CORRUPT_BKPT; | |
6268 return; | |
6269 } | |
6270 rc = freeSpace(pPage, pc, sz); | |
6271 if( rc ){ | |
6272 *pRC = rc; | |
6273 return; | |
6274 } | |
6275 pPage->nCell--; | |
6276 if( pPage->nCell==0 ){ | |
6277 memset(&data[hdr+1], 0, 4); | |
6278 data[hdr+7] = 0; | |
6279 put2byte(&data[hdr+5], pPage->pBt->usableSize); | |
6280 pPage->nFree = pPage->pBt->usableSize - pPage->hdrOffset | |
6281 - pPage->childPtrSize - 8; | |
6282 }else{ | |
6283 memmove(ptr, ptr+2, 2*(pPage->nCell - idx)); | |
6284 put2byte(&data[hdr+3], pPage->nCell); | |
6285 pPage->nFree += 2; | |
6286 } | |
6287 } | |
6288 | |
6289 /* | |
6290 ** Insert a new cell on pPage at cell index "i". pCell points to the | |
6291 ** content of the cell. | |
6292 ** | |
6293 ** If the cell content will fit on the page, then put it there. If it | |
6294 ** will not fit, then make a copy of the cell content into pTemp if | |
6295 ** pTemp is not null. Regardless of pTemp, allocate a new entry | |
6296 ** in pPage->apOvfl[] and make it point to the cell content (either | |
6297 ** in pTemp or the original pCell) and also record its index. | |
6298 ** Allocating a new entry in pPage->aCell[] implies that | |
6299 ** pPage->nOverflow is incremented. | |
6300 */ | |
6301 static void insertCell( | |
6302 MemPage *pPage, /* Page into which we are copying */ | |
6303 int i, /* New cell becomes the i-th cell of the page */ | |
6304 u8 *pCell, /* Content of the new cell */ | |
6305 int sz, /* Bytes of content in pCell */ | |
6306 u8 *pTemp, /* Temp storage space for pCell, if needed */ | |
6307 Pgno iChild, /* If non-zero, replace first 4 bytes with this value */ | |
6308 int *pRC /* Read and write return code from here */ | |
6309 ){ | |
6310 int idx = 0; /* Where to write new cell content in data[] */ | |
6311 int j; /* Loop counter */ | |
6312 u8 *data; /* The content of the whole page */ | |
6313 u8 *pIns; /* The point in pPage->aCellIdx[] where no cell inserted */ | |
6314 | |
6315 if( *pRC ) return; | |
6316 | |
6317 assert( i>=0 && i<=pPage->nCell+pPage->nOverflow ); | |
6318 assert( MX_CELL(pPage->pBt)<=10921 ); | |
6319 assert( pPage->nCell<=MX_CELL(pPage->pBt) || CORRUPT_DB ); | |
6320 assert( pPage->nOverflow<=ArraySize(pPage->apOvfl) ); | |
6321 assert( ArraySize(pPage->apOvfl)==ArraySize(pPage->aiOvfl) ); | |
6322 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
6323 /* The cell should normally be sized correctly. However, when moving a | |
6324 ** malformed cell from a leaf page to an interior page, if the cell size | |
6325 ** wanted to be less than 4 but got rounded up to 4 on the leaf, then size | |
6326 ** might be less than 8 (leaf-size + pointer) on the interior node. Hence | |
6327 ** the term after the || in the following assert(). */ | |
6328 assert( sz==pPage->xCellSize(pPage, pCell) || (sz==8 && iChild>0) ); | |
6329 if( pPage->nOverflow || sz+2>pPage->nFree ){ | |
6330 if( pTemp ){ | |
6331 memcpy(pTemp, pCell, sz); | |
6332 pCell = pTemp; | |
6333 } | |
6334 if( iChild ){ | |
6335 put4byte(pCell, iChild); | |
6336 } | |
6337 j = pPage->nOverflow++; | |
6338 assert( j<(int)(sizeof(pPage->apOvfl)/sizeof(pPage->apOvfl[0])) ); | |
6339 pPage->apOvfl[j] = pCell; | |
6340 pPage->aiOvfl[j] = (u16)i; | |
6341 | |
6342 /* When multiple overflows occur, they are always sequential and in | |
6343 ** sorted order. This invariants arise because multiple overflows can | |
6344 ** only occur when inserting divider cells into the parent page during | |
6345 ** balancing, and the dividers are adjacent and sorted. | |
6346 */ | |
6347 assert( j==0 || pPage->aiOvfl[j-1]<(u16)i ); /* Overflows in sorted order */ | |
6348 assert( j==0 || i==pPage->aiOvfl[j-1]+1 ); /* Overflows are sequential */ | |
6349 }else{ | |
6350 int rc = sqlite3PagerWrite(pPage->pDbPage); | |
6351 if( rc!=SQLITE_OK ){ | |
6352 *pRC = rc; | |
6353 return; | |
6354 } | |
6355 assert( sqlite3PagerIswriteable(pPage->pDbPage) ); | |
6356 data = pPage->aData; | |
6357 assert( &data[pPage->cellOffset]==pPage->aCellIdx ); | |
6358 rc = allocateSpace(pPage, sz, &idx); | |
6359 if( rc ){ *pRC = rc; return; } | |
6360 /* The allocateSpace() routine guarantees the following properties | |
6361 ** if it returns successfully */ | |
6362 assert( idx >= 0 ); | |
6363 assert( idx >= pPage->cellOffset+2*pPage->nCell+2 || CORRUPT_DB ); | |
6364 assert( idx+sz <= (int)pPage->pBt->usableSize ); | |
6365 pPage->nFree -= (u16)(2 + sz); | |
6366 memcpy(&data[idx], pCell, sz); | |
6367 if( iChild ){ | |
6368 put4byte(&data[idx], iChild); | |
6369 } | |
6370 pIns = pPage->aCellIdx + i*2; | |
6371 memmove(pIns+2, pIns, 2*(pPage->nCell - i)); | |
6372 put2byte(pIns, idx); | |
6373 pPage->nCell++; | |
6374 /* increment the cell count */ | |
6375 if( (++data[pPage->hdrOffset+4])==0 ) data[pPage->hdrOffset+3]++; | |
6376 assert( get2byte(&data[pPage->hdrOffset+3])==pPage->nCell ); | |
6377 #ifndef SQLITE_OMIT_AUTOVACUUM | |
6378 if( pPage->pBt->autoVacuum ){ | |
6379 /* The cell may contain a pointer to an overflow page. If so, write | |
6380 ** the entry for the overflow page into the pointer map. | |
6381 */ | |
6382 ptrmapPutOvflPtr(pPage, pCell, pRC); | |
6383 } | |
6384 #endif | |
6385 } | |
6386 } | |
6387 | |
6388 /* | |
6389 ** A CellArray object contains a cache of pointers and sizes for a | |
6390 ** consecutive sequence of cells that might be held multiple pages. | |
6391 */ | |
6392 typedef struct CellArray CellArray; | |
6393 struct CellArray { | |
6394 int nCell; /* Number of cells in apCell[] */ | |
6395 MemPage *pRef; /* Reference page */ | |
6396 u8 **apCell; /* All cells begin balanced */ | |
6397 u16 *szCell; /* Local size of all cells in apCell[] */ | |
6398 }; | |
6399 | |
6400 /* | |
6401 ** Make sure the cell sizes at idx, idx+1, ..., idx+N-1 have been | |
6402 ** computed. | |
6403 */ | |
6404 static void populateCellCache(CellArray *p, int idx, int N){ | |
6405 assert( idx>=0 && idx+N<=p->nCell ); | |
6406 while( N>0 ){ | |
6407 assert( p->apCell[idx]!=0 ); | |
6408 if( p->szCell[idx]==0 ){ | |
6409 p->szCell[idx] = p->pRef->xCellSize(p->pRef, p->apCell[idx]); | |
6410 }else{ | |
6411 assert( CORRUPT_DB || | |
6412 p->szCell[idx]==p->pRef->xCellSize(p->pRef, p->apCell[idx]) ); | |
6413 } | |
6414 idx++; | |
6415 N--; | |
6416 } | |
6417 } | |
6418 | |
6419 /* | |
6420 ** Return the size of the Nth element of the cell array | |
6421 */ | |
6422 static SQLITE_NOINLINE u16 computeCellSize(CellArray *p, int N){ | |
6423 assert( N>=0 && N<p->nCell ); | |
6424 assert( p->szCell[N]==0 ); | |
6425 p->szCell[N] = p->pRef->xCellSize(p->pRef, p->apCell[N]); | |
6426 return p->szCell[N]; | |
6427 } | |
6428 static u16 cachedCellSize(CellArray *p, int N){ | |
6429 assert( N>=0 && N<p->nCell ); | |
6430 if( p->szCell[N] ) return p->szCell[N]; | |
6431 return computeCellSize(p, N); | |
6432 } | |
6433 | |
6434 /* | |
6435 ** Array apCell[] contains pointers to nCell b-tree page cells. The | |
6436 ** szCell[] array contains the size in bytes of each cell. This function | |
6437 ** replaces the current contents of page pPg with the contents of the cell | |
6438 ** array. | |
6439 ** | |
6440 ** Some of the cells in apCell[] may currently be stored in pPg. This | |
6441 ** function works around problems caused by this by making a copy of any | |
6442 ** such cells before overwriting the page data. | |
6443 ** | |
6444 ** The MemPage.nFree field is invalidated by this function. It is the | |
6445 ** responsibility of the caller to set it correctly. | |
6446 */ | |
6447 static int rebuildPage( | |
6448 MemPage *pPg, /* Edit this page */ | |
6449 int nCell, /* Final number of cells on page */ | |
6450 u8 **apCell, /* Array of cells */ | |
6451 u16 *szCell /* Array of cell sizes */ | |
6452 ){ | |
6453 const int hdr = pPg->hdrOffset; /* Offset of header on pPg */ | |
6454 u8 * const aData = pPg->aData; /* Pointer to data for pPg */ | |
6455 const int usableSize = pPg->pBt->usableSize; | |
6456 u8 * const pEnd = &aData[usableSize]; | |
6457 int i; | |
6458 u8 *pCellptr = pPg->aCellIdx; | |
6459 u8 *pTmp = sqlite3PagerTempSpace(pPg->pBt->pPager); | |
6460 u8 *pData; | |
6461 | |
6462 i = get2byte(&aData[hdr+5]); | |
6463 memcpy(&pTmp[i], &aData[i], usableSize - i); | |
6464 | |
6465 pData = pEnd; | |
6466 for(i=0; i<nCell; i++){ | |
6467 u8 *pCell = apCell[i]; | |
6468 if( SQLITE_WITHIN(pCell,aData,pEnd) ){ | |
6469 pCell = &pTmp[pCell - aData]; | |
6470 } | |
6471 pData -= szCell[i]; | |
6472 put2byte(pCellptr, (pData - aData)); | |
6473 pCellptr += 2; | |
6474 if( pData < pCellptr ) return SQLITE_CORRUPT_BKPT; | |
6475 memcpy(pData, pCell, szCell[i]); | |
6476 assert( szCell[i]==pPg->xCellSize(pPg, pCell) || CORRUPT_DB ); | |
6477 testcase( szCell[i]!=pPg->xCellSize(pPg,pCell) ); | |
6478 } | |
6479 | |
6480 /* The pPg->nFree field is now set incorrectly. The caller will fix it. */ | |
6481 pPg->nCell = nCell; | |
6482 pPg->nOverflow = 0; | |
6483 | |
6484 put2byte(&aData[hdr+1], 0); | |
6485 put2byte(&aData[hdr+3], pPg->nCell); | |
6486 put2byte(&aData[hdr+5], pData - aData); | |
6487 aData[hdr+7] = 0x00; | |
6488 return SQLITE_OK; | |
6489 } | |
6490 | |
6491 /* | |
6492 ** Array apCell[] contains nCell pointers to b-tree cells. Array szCell | |
6493 ** contains the size in bytes of each such cell. This function attempts to | |
6494 ** add the cells stored in the array to page pPg. If it cannot (because | |
6495 ** the page needs to be defragmented before the cells will fit), non-zero | |
6496 ** is returned. Otherwise, if the cells are added successfully, zero is | |
6497 ** returned. | |
6498 ** | |
6499 ** Argument pCellptr points to the first entry in the cell-pointer array | |
6500 ** (part of page pPg) to populate. After cell apCell[0] is written to the | |
6501 ** page body, a 16-bit offset is written to pCellptr. And so on, for each | |
6502 ** cell in the array. It is the responsibility of the caller to ensure | |
6503 ** that it is safe to overwrite this part of the cell-pointer array. | |
6504 ** | |
6505 ** When this function is called, *ppData points to the start of the | |
6506 ** content area on page pPg. If the size of the content area is extended, | |
6507 ** *ppData is updated to point to the new start of the content area | |
6508 ** before returning. | |
6509 ** | |
6510 ** Finally, argument pBegin points to the byte immediately following the | |
6511 ** end of the space required by this page for the cell-pointer area (for | |
6512 ** all cells - not just those inserted by the current call). If the content | |
6513 ** area must be extended to before this point in order to accomodate all | |
6514 ** cells in apCell[], then the cells do not fit and non-zero is returned. | |
6515 */ | |
6516 static int pageInsertArray( | |
6517 MemPage *pPg, /* Page to add cells to */ | |
6518 u8 *pBegin, /* End of cell-pointer array */ | |
6519 u8 **ppData, /* IN/OUT: Page content -area pointer */ | |
6520 u8 *pCellptr, /* Pointer to cell-pointer area */ | |
6521 int iFirst, /* Index of first cell to add */ | |
6522 int nCell, /* Number of cells to add to pPg */ | |
6523 CellArray *pCArray /* Array of cells */ | |
6524 ){ | |
6525 int i; | |
6526 u8 *aData = pPg->aData; | |
6527 u8 *pData = *ppData; | |
6528 int iEnd = iFirst + nCell; | |
6529 assert( CORRUPT_DB || pPg->hdrOffset==0 ); /* Never called on page 1 */ | |
6530 for(i=iFirst; i<iEnd; i++){ | |
6531 int sz, rc; | |
6532 u8 *pSlot; | |
6533 sz = cachedCellSize(pCArray, i); | |
6534 if( (aData[1]==0 && aData[2]==0) || (pSlot = pageFindSlot(pPg,sz,&rc))==0 ){ | |
6535 pData -= sz; | |
6536 if( pData<pBegin ) return 1; | |
6537 pSlot = pData; | |
6538 } | |
6539 /* pSlot and pCArray->apCell[i] will never overlap on a well-formed | |
6540 ** database. But they might for a corrupt database. Hence use memmove() | |
6541 ** since memcpy() sends SIGABORT with overlapping buffers on OpenBSD */ | |
6542 assert( (pSlot+sz)<=pCArray->apCell[i] | |
6543 || pSlot>=(pCArray->apCell[i]+sz) | |
6544 || CORRUPT_DB ); | |
6545 memmove(pSlot, pCArray->apCell[i], sz); | |
6546 put2byte(pCellptr, (pSlot - aData)); | |
6547 pCellptr += 2; | |
6548 } | |
6549 *ppData = pData; | |
6550 return 0; | |
6551 } | |
6552 | |
6553 /* | |
6554 ** Array apCell[] contains nCell pointers to b-tree cells. Array szCell | |
6555 ** contains the size in bytes of each such cell. This function adds the | |
6556 ** space associated with each cell in the array that is currently stored | |
6557 ** within the body of pPg to the pPg free-list. The cell-pointers and other | |
6558 ** fields of the page are not updated. | |
6559 ** | |
6560 ** This function returns the total number of cells added to the free-list. | |
6561 */ | |
6562 static int pageFreeArray( | |
6563 MemPage *pPg, /* Page to edit */ | |
6564 int iFirst, /* First cell to delete */ | |
6565 int nCell, /* Cells to delete */ | |
6566 CellArray *pCArray /* Array of cells */ | |
6567 ){ | |
6568 u8 * const aData = pPg->aData; | |
6569 u8 * const pEnd = &aData[pPg->pBt->usableSize]; | |
6570 u8 * const pStart = &aData[pPg->hdrOffset + 8 + pPg->childPtrSize]; | |
6571 int nRet = 0; | |
6572 int i; | |
6573 int iEnd = iFirst + nCell; | |
6574 u8 *pFree = 0; | |
6575 int szFree = 0; | |
6576 | |
6577 for(i=iFirst; i<iEnd; i++){ | |
6578 u8 *pCell = pCArray->apCell[i]; | |
6579 if( SQLITE_WITHIN(pCell, pStart, pEnd) ){ | |
6580 int sz; | |
6581 /* No need to use cachedCellSize() here. The sizes of all cells that | |
6582 ** are to be freed have already been computing while deciding which | |
6583 ** cells need freeing */ | |
6584 sz = pCArray->szCell[i]; assert( sz>0 ); | |
6585 if( pFree!=(pCell + sz) ){ | |
6586 if( pFree ){ | |
6587 assert( pFree>aData && (pFree - aData)<65536 ); | |
6588 freeSpace(pPg, (u16)(pFree - aData), szFree); | |
6589 } | |
6590 pFree = pCell; | |
6591 szFree = sz; | |
6592 if( pFree+sz>pEnd ) return 0; | |
6593 }else{ | |
6594 pFree = pCell; | |
6595 szFree += sz; | |
6596 } | |
6597 nRet++; | |
6598 } | |
6599 } | |
6600 if( pFree ){ | |
6601 assert( pFree>aData && (pFree - aData)<65536 ); | |
6602 freeSpace(pPg, (u16)(pFree - aData), szFree); | |
6603 } | |
6604 return nRet; | |
6605 } | |
6606 | |
6607 /* | |
6608 ** apCell[] and szCell[] contains pointers to and sizes of all cells in the | |
6609 ** pages being balanced. The current page, pPg, has pPg->nCell cells starting | |
6610 ** with apCell[iOld]. After balancing, this page should hold nNew cells | |
6611 ** starting at apCell[iNew]. | |
6612 ** | |
6613 ** This routine makes the necessary adjustments to pPg so that it contains | |
6614 ** the correct cells after being balanced. | |
6615 ** | |
6616 ** The pPg->nFree field is invalid when this function returns. It is the | |
6617 ** responsibility of the caller to set it correctly. | |
6618 */ | |
6619 static int editPage( | |
6620 MemPage *pPg, /* Edit this page */ | |
6621 int iOld, /* Index of first cell currently on page */ | |
6622 int iNew, /* Index of new first cell on page */ | |
6623 int nNew, /* Final number of cells on page */ | |
6624 CellArray *pCArray /* Array of cells and sizes */ | |
6625 ){ | |
6626 u8 * const aData = pPg->aData; | |
6627 const int hdr = pPg->hdrOffset; | |
6628 u8 *pBegin = &pPg->aCellIdx[nNew * 2]; | |
6629 int nCell = pPg->nCell; /* Cells stored on pPg */ | |
6630 u8 *pData; | |
6631 u8 *pCellptr; | |
6632 int i; | |
6633 int iOldEnd = iOld + pPg->nCell + pPg->nOverflow; | |
6634 int iNewEnd = iNew + nNew; | |
6635 | |
6636 #ifdef SQLITE_DEBUG | |
6637 u8 *pTmp = sqlite3PagerTempSpace(pPg->pBt->pPager); | |
6638 memcpy(pTmp, aData, pPg->pBt->usableSize); | |
6639 #endif | |
6640 | |
6641 /* Remove cells from the start and end of the page */ | |
6642 if( iOld<iNew ){ | |
6643 int nShift = pageFreeArray(pPg, iOld, iNew-iOld, pCArray); | |
6644 memmove(pPg->aCellIdx, &pPg->aCellIdx[nShift*2], nCell*2); | |
6645 nCell -= nShift; | |
6646 } | |
6647 if( iNewEnd < iOldEnd ){ | |
6648 nCell -= pageFreeArray(pPg, iNewEnd, iOldEnd - iNewEnd, pCArray); | |
6649 } | |
6650 | |
6651 pData = &aData[get2byteNotZero(&aData[hdr+5])]; | |
6652 if( pData<pBegin ) goto editpage_fail; | |
6653 | |
6654 /* Add cells to the start of the page */ | |
6655 if( iNew<iOld ){ | |
6656 int nAdd = MIN(nNew,iOld-iNew); | |
6657 assert( (iOld-iNew)<nNew || nCell==0 || CORRUPT_DB ); | |
6658 pCellptr = pPg->aCellIdx; | |
6659 memmove(&pCellptr[nAdd*2], pCellptr, nCell*2); | |
6660 if( pageInsertArray( | |
6661 pPg, pBegin, &pData, pCellptr, | |
6662 iNew, nAdd, pCArray | |
6663 ) ) goto editpage_fail; | |
6664 nCell += nAdd; | |
6665 } | |
6666 | |
6667 /* Add any overflow cells */ | |
6668 for(i=0; i<pPg->nOverflow; i++){ | |
6669 int iCell = (iOld + pPg->aiOvfl[i]) - iNew; | |
6670 if( iCell>=0 && iCell<nNew ){ | |
6671 pCellptr = &pPg->aCellIdx[iCell * 2]; | |
6672 memmove(&pCellptr[2], pCellptr, (nCell - iCell) * 2); | |
6673 nCell++; | |
6674 if( pageInsertArray( | |
6675 pPg, pBegin, &pData, pCellptr, | |
6676 iCell+iNew, 1, pCArray | |
6677 ) ) goto editpage_fail; | |
6678 } | |
6679 } | |
6680 | |
6681 /* Append cells to the end of the page */ | |
6682 pCellptr = &pPg->aCellIdx[nCell*2]; | |
6683 if( pageInsertArray( | |
6684 pPg, pBegin, &pData, pCellptr, | |
6685 iNew+nCell, nNew-nCell, pCArray | |
6686 ) ) goto editpage_fail; | |
6687 | |
6688 pPg->nCell = nNew; | |
6689 pPg->nOverflow = 0; | |
6690 | |
6691 put2byte(&aData[hdr+3], pPg->nCell); | |
6692 put2byte(&aData[hdr+5], pData - aData); | |
6693 | |
6694 #ifdef SQLITE_DEBUG | |
6695 for(i=0; i<nNew && !CORRUPT_DB; i++){ | |
6696 u8 *pCell = pCArray->apCell[i+iNew]; | |
6697 int iOff = get2byteAligned(&pPg->aCellIdx[i*2]); | |
6698 if( pCell>=aData && pCell<&aData[pPg->pBt->usableSize] ){ | |
6699 pCell = &pTmp[pCell - aData]; | |
6700 } | |
6701 assert( 0==memcmp(pCell, &aData[iOff], | |
6702 pCArray->pRef->xCellSize(pCArray->pRef, pCArray->apCell[i+iNew])) ); | |
6703 } | |
6704 #endif | |
6705 | |
6706 return SQLITE_OK; | |
6707 editpage_fail: | |
6708 /* Unable to edit this page. Rebuild it from scratch instead. */ | |
6709 populateCellCache(pCArray, iNew, nNew); | |
6710 return rebuildPage(pPg, nNew, &pCArray->apCell[iNew], &pCArray->szCell[iNew]); | |
6711 } | |
6712 | |
6713 /* | |
6714 ** The following parameters determine how many adjacent pages get involved | |
6715 ** in a balancing operation. NN is the number of neighbors on either side | |
6716 ** of the page that participate in the balancing operation. NB is the | |
6717 ** total number of pages that participate, including the target page and | |
6718 ** NN neighbors on either side. | |
6719 ** | |
6720 ** The minimum value of NN is 1 (of course). Increasing NN above 1 | |
6721 ** (to 2 or 3) gives a modest improvement in SELECT and DELETE performance | |
6722 ** in exchange for a larger degradation in INSERT and UPDATE performance. | |
6723 ** The value of NN appears to give the best results overall. | |
6724 */ | |
6725 #define NN 1 /* Number of neighbors on either side of pPage */ | |
6726 #define NB (NN*2+1) /* Total pages involved in the balance */ | |
6727 | |
6728 | |
6729 #ifndef SQLITE_OMIT_QUICKBALANCE | |
6730 /* | |
6731 ** This version of balance() handles the common special case where | |
6732 ** a new entry is being inserted on the extreme right-end of the | |
6733 ** tree, in other words, when the new entry will become the largest | |
6734 ** entry in the tree. | |
6735 ** | |
6736 ** Instead of trying to balance the 3 right-most leaf pages, just add | |
6737 ** a new page to the right-hand side and put the one new entry in | |
6738 ** that page. This leaves the right side of the tree somewhat | |
6739 ** unbalanced. But odds are that we will be inserting new entries | |
6740 ** at the end soon afterwards so the nearly empty page will quickly | |
6741 ** fill up. On average. | |
6742 ** | |
6743 ** pPage is the leaf page which is the right-most page in the tree. | |
6744 ** pParent is its parent. pPage must have a single overflow entry | |
6745 ** which is also the right-most entry on the page. | |
6746 ** | |
6747 ** The pSpace buffer is used to store a temporary copy of the divider | |
6748 ** cell that will be inserted into pParent. Such a cell consists of a 4 | |
6749 ** byte page number followed by a variable length integer. In other | |
6750 ** words, at most 13 bytes. Hence the pSpace buffer must be at | |
6751 ** least 13 bytes in size. | |
6752 */ | |
6753 static int balance_quick(MemPage *pParent, MemPage *pPage, u8 *pSpace){ | |
6754 BtShared *const pBt = pPage->pBt; /* B-Tree Database */ | |
6755 MemPage *pNew; /* Newly allocated page */ | |
6756 int rc; /* Return Code */ | |
6757 Pgno pgnoNew; /* Page number of pNew */ | |
6758 | |
6759 assert( sqlite3_mutex_held(pPage->pBt->mutex) ); | |
6760 assert( sqlite3PagerIswriteable(pParent->pDbPage) ); | |
6761 assert( pPage->nOverflow==1 ); | |
6762 | |
6763 /* This error condition is now caught prior to reaching this function */ | |
6764 if( NEVER(pPage->nCell==0) ) return SQLITE_CORRUPT_BKPT; | |
6765 | |
6766 /* Allocate a new page. This page will become the right-sibling of | |
6767 ** pPage. Make the parent page writable, so that the new divider cell | |
6768 ** may be inserted. If both these operations are successful, proceed. | |
6769 */ | |
6770 rc = allocateBtreePage(pBt, &pNew, &pgnoNew, 0, 0); | |
6771 | |
6772 if( rc==SQLITE_OK ){ | |
6773 | |
6774 u8 *pOut = &pSpace[4]; | |
6775 u8 *pCell = pPage->apOvfl[0]; | |
6776 u16 szCell = pPage->xCellSize(pPage, pCell); | |
6777 u8 *pStop; | |
6778 | |
6779 assert( sqlite3PagerIswriteable(pNew->pDbPage) ); | |
6780 assert( pPage->aData[0]==(PTF_INTKEY|PTF_LEAFDATA|PTF_LEAF) ); | |
6781 zeroPage(pNew, PTF_INTKEY|PTF_LEAFDATA|PTF_LEAF); | |
6782 rc = rebuildPage(pNew, 1, &pCell, &szCell); | |
6783 if( NEVER(rc) ) return rc; | |
6784 pNew->nFree = pBt->usableSize - pNew->cellOffset - 2 - szCell; | |
6785 | |
6786 /* If this is an auto-vacuum database, update the pointer map | |
6787 ** with entries for the new page, and any pointer from the | |
6788 ** cell on the page to an overflow page. If either of these | |
6789 ** operations fails, the return code is set, but the contents | |
6790 ** of the parent page are still manipulated by thh code below. | |
6791 ** That is Ok, at this point the parent page is guaranteed to | |
6792 ** be marked as dirty. Returning an error code will cause a | |
6793 ** rollback, undoing any changes made to the parent page. | |
6794 */ | |
6795 if( ISAUTOVACUUM ){ | |
6796 ptrmapPut(pBt, pgnoNew, PTRMAP_BTREE, pParent->pgno, &rc); | |
6797 if( szCell>pNew->minLocal ){ | |
6798 ptrmapPutOvflPtr(pNew, pCell, &rc); | |
6799 } | |
6800 } | |
6801 | |
6802 /* Create a divider cell to insert into pParent. The divider cell | |
6803 ** consists of a 4-byte page number (the page number of pPage) and | |
6804 ** a variable length key value (which must be the same value as the | |
6805 ** largest key on pPage). | |
6806 ** | |
6807 ** To find the largest key value on pPage, first find the right-most | |
6808 ** cell on pPage. The first two fields of this cell are the | |
6809 ** record-length (a variable length integer at most 32-bits in size) | |
6810 ** and the key value (a variable length integer, may have any value). | |
6811 ** The first of the while(...) loops below skips over the record-length | |
6812 ** field. The second while(...) loop copies the key value from the | |
6813 ** cell on pPage into the pSpace buffer. | |
6814 */ | |
6815 pCell = findCell(pPage, pPage->nCell-1); | |
6816 pStop = &pCell[9]; | |
6817 while( (*(pCell++)&0x80) && pCell<pStop ); | |
6818 pStop = &pCell[9]; | |
6819 while( ((*(pOut++) = *(pCell++))&0x80) && pCell<pStop ); | |
6820 | |
6821 /* Insert the new divider cell into pParent. */ | |
6822 insertCell(pParent, pParent->nCell, pSpace, (int)(pOut-pSpace), | |
6823 0, pPage->pgno, &rc); | |
6824 | |
6825 /* Set the right-child pointer of pParent to point to the new page. */ | |
6826 put4byte(&pParent->aData[pParent->hdrOffset+8], pgnoNew); | |
6827 | |
6828 /* Release the reference to the new page. */ | |
6829 releasePage(pNew); | |
6830 } | |
6831 | |
6832 return rc; | |
6833 } | |
6834 #endif /* SQLITE_OMIT_QUICKBALANCE */ | |
6835 | |
6836 #if 0 | |
6837 /* | |
6838 ** This function does not contribute anything to the operation of SQLite. | |
6839 ** it is sometimes activated temporarily while debugging code responsible | |
6840 ** for setting pointer-map entries. | |
6841 */ | |
6842 static int ptrmapCheckPages(MemPage **apPage, int nPage){ | |
6843 int i, j; | |
6844 for(i=0; i<nPage; i++){ | |
6845 Pgno n; | |
6846 u8 e; | |
6847 MemPage *pPage = apPage[i]; | |
6848 BtShared *pBt = pPage->pBt; | |
6849 assert( pPage->isInit ); | |
6850 | |
6851 for(j=0; j<pPage->nCell; j++){ | |
6852 CellInfo info; | |
6853 u8 *z; | |
6854 | |
6855 z = findCell(pPage, j); | |
6856 pPage->xParseCell(pPage, z, &info); | |
6857 if( info.nLocal<info.nPayload ){ | |
6858 Pgno ovfl = get4byte(&z[info.nSize-4]); | |
6859 ptrmapGet(pBt, ovfl, &e, &n); | |
6860 assert( n==pPage->pgno && e==PTRMAP_OVERFLOW1 ); | |
6861 } | |
6862 if( !pPage->leaf ){ | |
6863 Pgno child = get4byte(z); | |
6864 ptrmapGet(pBt, child, &e, &n); | |
6865 assert( n==pPage->pgno && e==PTRMAP_BTREE ); | |
6866 } | |
6867 } | |
6868 if( !pPage->leaf ){ | |
6869 Pgno child = get4byte(&pPage->aData[pPage->hdrOffset+8]); | |
6870 ptrmapGet(pBt, child, &e, &n); | |
6871 assert( n==pPage->pgno && e==PTRMAP_BTREE ); | |
6872 } | |
6873 } | |
6874 return 1; | |
6875 } | |
6876 #endif | |
6877 | |
6878 /* | |
6879 ** This function is used to copy the contents of the b-tree node stored | |
6880 ** on page pFrom to page pTo. If page pFrom was not a leaf page, then | |
6881 ** the pointer-map entries for each child page are updated so that the | |
6882 ** parent page stored in the pointer map is page pTo. If pFrom contained | |
6883 ** any cells with overflow page pointers, then the corresponding pointer | |
6884 ** map entries are also updated so that the parent page is page pTo. | |
6885 ** | |
6886 ** If pFrom is currently carrying any overflow cells (entries in the | |
6887 ** MemPage.apOvfl[] array), they are not copied to pTo. | |
6888 ** | |
6889 ** Before returning, page pTo is reinitialized using btreeInitPage(). | |
6890 ** | |
6891 ** The performance of this function is not critical. It is only used by | |
6892 ** the balance_shallower() and balance_deeper() procedures, neither of | |
6893 ** which are called often under normal circumstances. | |
6894 */ | |
6895 static void copyNodeContent(MemPage *pFrom, MemPage *pTo, int *pRC){ | |
6896 if( (*pRC)==SQLITE_OK ){ | |
6897 BtShared * const pBt = pFrom->pBt; | |
6898 u8 * const aFrom = pFrom->aData; | |
6899 u8 * const aTo = pTo->aData; | |
6900 int const iFromHdr = pFrom->hdrOffset; | |
6901 int const iToHdr = ((pTo->pgno==1) ? 100 : 0); | |
6902 int rc; | |
6903 int iData; | |
6904 | |
6905 | |
6906 assert( pFrom->isInit ); | |
6907 assert( pFrom->nFree>=iToHdr ); | |
6908 assert( get2byte(&aFrom[iFromHdr+5]) <= (int)pBt->usableSize ); | |
6909 | |
6910 /* Copy the b-tree node content from page pFrom to page pTo. */ | |
6911 iData = get2byte(&aFrom[iFromHdr+5]); | |
6912 memcpy(&aTo[iData], &aFrom[iData], pBt->usableSize-iData); | |
6913 memcpy(&aTo[iToHdr], &aFrom[iFromHdr], pFrom->cellOffset + 2*pFrom->nCell); | |
6914 | |
6915 /* Reinitialize page pTo so that the contents of the MemPage structure | |
6916 ** match the new data. The initialization of pTo can actually fail under | |
6917 ** fairly obscure circumstances, even though it is a copy of initialized | |
6918 ** page pFrom. | |
6919 */ | |
6920 pTo->isInit = 0; | |
6921 rc = btreeInitPage(pTo); | |
6922 if( rc!=SQLITE_OK ){ | |
6923 *pRC = rc; | |
6924 return; | |
6925 } | |
6926 | |
6927 /* If this is an auto-vacuum database, update the pointer-map entries | |
6928 ** for any b-tree or overflow pages that pTo now contains the pointers to. | |
6929 */ | |
6930 if( ISAUTOVACUUM ){ | |
6931 *pRC = setChildPtrmaps(pTo); | |
6932 } | |
6933 } | |
6934 } | |
6935 | |
6936 /* | |
6937 ** This routine redistributes cells on the iParentIdx'th child of pParent | |
6938 ** (hereafter "the page") and up to 2 siblings so that all pages have about the | |
6939 ** same amount of free space. Usually a single sibling on either side of the | |
6940 ** page are used in the balancing, though both siblings might come from one | |
6941 ** side if the page is the first or last child of its parent. If the page | |
6942 ** has fewer than 2 siblings (something which can only happen if the page | |
6943 ** is a root page or a child of a root page) then all available siblings | |
6944 ** participate in the balancing. | |
6945 ** | |
6946 ** The number of siblings of the page might be increased or decreased by | |
6947 ** one or two in an effort to keep pages nearly full but not over full. | |
6948 ** | |
6949 ** Note that when this routine is called, some of the cells on the page | |
6950 ** might not actually be stored in MemPage.aData[]. This can happen | |
6951 ** if the page is overfull. This routine ensures that all cells allocated | |
6952 ** to the page and its siblings fit into MemPage.aData[] before returning. | |
6953 ** | |
6954 ** In the course of balancing the page and its siblings, cells may be | |
6955 ** inserted into or removed from the parent page (pParent). Doing so | |
6956 ** may cause the parent page to become overfull or underfull. If this | |
6957 ** happens, it is the responsibility of the caller to invoke the correct | |
6958 ** balancing routine to fix this problem (see the balance() routine). | |
6959 ** | |
6960 ** If this routine fails for any reason, it might leave the database | |
6961 ** in a corrupted state. So if this routine fails, the database should | |
6962 ** be rolled back. | |
6963 ** | |
6964 ** The third argument to this function, aOvflSpace, is a pointer to a | |
6965 ** buffer big enough to hold one page. If while inserting cells into the parent | |
6966 ** page (pParent) the parent page becomes overfull, this buffer is | |
6967 ** used to store the parent's overflow cells. Because this function inserts | |
6968 ** a maximum of four divider cells into the parent page, and the maximum | |
6969 ** size of a cell stored within an internal node is always less than 1/4 | |
6970 ** of the page-size, the aOvflSpace[] buffer is guaranteed to be large | |
6971 ** enough for all overflow cells. | |
6972 ** | |
6973 ** If aOvflSpace is set to a null pointer, this function returns | |
6974 ** SQLITE_NOMEM. | |
6975 */ | |
6976 static int balance_nonroot( | |
6977 MemPage *pParent, /* Parent page of siblings being balanced */ | |
6978 int iParentIdx, /* Index of "the page" in pParent */ | |
6979 u8 *aOvflSpace, /* page-size bytes of space for parent ovfl */ | |
6980 int isRoot, /* True if pParent is a root-page */ | |
6981 int bBulk /* True if this call is part of a bulk load */ | |
6982 ){ | |
6983 BtShared *pBt; /* The whole database */ | |
6984 int nMaxCells = 0; /* Allocated size of apCell, szCell, aFrom. */ | |
6985 int nNew = 0; /* Number of pages in apNew[] */ | |
6986 int nOld; /* Number of pages in apOld[] */ | |
6987 int i, j, k; /* Loop counters */ | |
6988 int nxDiv; /* Next divider slot in pParent->aCell[] */ | |
6989 int rc = SQLITE_OK; /* The return code */ | |
6990 u16 leafCorrection; /* 4 if pPage is a leaf. 0 if not */ | |
6991 int leafData; /* True if pPage is a leaf of a LEAFDATA tree */ | |
6992 int usableSpace; /* Bytes in pPage beyond the header */ | |
6993 int pageFlags; /* Value of pPage->aData[0] */ | |
6994 int iSpace1 = 0; /* First unused byte of aSpace1[] */ | |
6995 int iOvflSpace = 0; /* First unused byte of aOvflSpace[] */ | |
6996 int szScratch; /* Size of scratch memory requested */ | |
6997 MemPage *apOld[NB]; /* pPage and up to two siblings */ | |
6998 MemPage *apNew[NB+2]; /* pPage and up to NB siblings after balancing */ | |
6999 u8 *pRight; /* Location in parent of right-sibling pointer */ | |
7000 u8 *apDiv[NB-1]; /* Divider cells in pParent */ | |
7001 int cntNew[NB+2]; /* Index in b.paCell[] of cell after i-th page */ | |
7002 int cntOld[NB+2]; /* Old index in b.apCell[] */ | |
7003 int szNew[NB+2]; /* Combined size of cells placed on i-th page */ | |
7004 u8 *aSpace1; /* Space for copies of dividers cells */ | |
7005 Pgno pgno; /* Temp var to store a page number in */ | |
7006 u8 abDone[NB+2]; /* True after i'th new page is populated */ | |
7007 Pgno aPgno[NB+2]; /* Page numbers of new pages before shuffling */ | |
7008 Pgno aPgOrder[NB+2]; /* Copy of aPgno[] used for sorting pages */ | |
7009 u16 aPgFlags[NB+2]; /* flags field of new pages before shuffling */ | |
7010 CellArray b; /* Parsed information on cells being balanced */ | |
7011 | |
7012 memset(abDone, 0, sizeof(abDone)); | |
7013 b.nCell = 0; | |
7014 b.apCell = 0; | |
7015 pBt = pParent->pBt; | |
7016 assert( sqlite3_mutex_held(pBt->mutex) ); | |
7017 assert( sqlite3PagerIswriteable(pParent->pDbPage) ); | |
7018 | |
7019 #if 0 | |
7020 TRACE(("BALANCE: begin page %d child of %d\n", pPage->pgno, pParent->pgno)); | |
7021 #endif | |
7022 | |
7023 /* At this point pParent may have at most one overflow cell. And if | |
7024 ** this overflow cell is present, it must be the cell with | |
7025 ** index iParentIdx. This scenario comes about when this function | |
7026 ** is called (indirectly) from sqlite3BtreeDelete(). | |
7027 */ | |
7028 assert( pParent->nOverflow==0 || pParent->nOverflow==1 ); | |
7029 assert( pParent->nOverflow==0 || pParent->aiOvfl[0]==iParentIdx ); | |
7030 | |
7031 if( !aOvflSpace ){ | |
7032 return SQLITE_NOMEM; | |
7033 } | |
7034 | |
7035 /* Find the sibling pages to balance. Also locate the cells in pParent | |
7036 ** that divide the siblings. An attempt is made to find NN siblings on | |
7037 ** either side of pPage. More siblings are taken from one side, however, | |
7038 ** if there are fewer than NN siblings on the other side. If pParent | |
7039 ** has NB or fewer children then all children of pParent are taken. | |
7040 ** | |
7041 ** This loop also drops the divider cells from the parent page. This | |
7042 ** way, the remainder of the function does not have to deal with any | |
7043 ** overflow cells in the parent page, since if any existed they will | |
7044 ** have already been removed. | |
7045 */ | |
7046 i = pParent->nOverflow + pParent->nCell; | |
7047 if( i<2 ){ | |
7048 nxDiv = 0; | |
7049 }else{ | |
7050 assert( bBulk==0 || bBulk==1 ); | |
7051 if( iParentIdx==0 ){ | |
7052 nxDiv = 0; | |
7053 }else if( iParentIdx==i ){ | |
7054 nxDiv = i-2+bBulk; | |
7055 }else{ | |
7056 nxDiv = iParentIdx-1; | |
7057 } | |
7058 i = 2-bBulk; | |
7059 } | |
7060 nOld = i+1; | |
7061 if( (i+nxDiv-pParent->nOverflow)==pParent->nCell ){ | |
7062 pRight = &pParent->aData[pParent->hdrOffset+8]; | |
7063 }else{ | |
7064 pRight = findCell(pParent, i+nxDiv-pParent->nOverflow); | |
7065 } | |
7066 pgno = get4byte(pRight); | |
7067 while( 1 ){ | |
7068 rc = getAndInitPage(pBt, pgno, &apOld[i], 0, 0); | |
7069 if( rc ){ | |
7070 memset(apOld, 0, (i+1)*sizeof(MemPage*)); | |
7071 goto balance_cleanup; | |
7072 } | |
7073 nMaxCells += 1+apOld[i]->nCell+apOld[i]->nOverflow; | |
7074 if( (i--)==0 ) break; | |
7075 | |
7076 if( i+nxDiv==pParent->aiOvfl[0] && pParent->nOverflow ){ | |
7077 apDiv[i] = pParent->apOvfl[0]; | |
7078 pgno = get4byte(apDiv[i]); | |
7079 szNew[i] = pParent->xCellSize(pParent, apDiv[i]); | |
7080 pParent->nOverflow = 0; | |
7081 }else{ | |
7082 apDiv[i] = findCell(pParent, i+nxDiv-pParent->nOverflow); | |
7083 pgno = get4byte(apDiv[i]); | |
7084 szNew[i] = pParent->xCellSize(pParent, apDiv[i]); | |
7085 | |
7086 /* Drop the cell from the parent page. apDiv[i] still points to | |
7087 ** the cell within the parent, even though it has been dropped. | |
7088 ** This is safe because dropping a cell only overwrites the first | |
7089 ** four bytes of it, and this function does not need the first | |
7090 ** four bytes of the divider cell. So the pointer is safe to use | |
7091 ** later on. | |
7092 ** | |
7093 ** But not if we are in secure-delete mode. In secure-delete mode, | |
7094 ** the dropCell() routine will overwrite the entire cell with zeroes. | |
7095 ** In this case, temporarily copy the cell into the aOvflSpace[] | |
7096 ** buffer. It will be copied out again as soon as the aSpace[] buffer | |
7097 ** is allocated. */ | |
7098 if( pBt->btsFlags & BTS_SECURE_DELETE ){ | |
7099 int iOff; | |
7100 | |
7101 iOff = SQLITE_PTR_TO_INT(apDiv[i]) - SQLITE_PTR_TO_INT(pParent->aData); | |
7102 if( (iOff+szNew[i])>(int)pBt->usableSize ){ | |
7103 rc = SQLITE_CORRUPT_BKPT; | |
7104 memset(apOld, 0, (i+1)*sizeof(MemPage*)); | |
7105 goto balance_cleanup; | |
7106 }else{ | |
7107 memcpy(&aOvflSpace[iOff], apDiv[i], szNew[i]); | |
7108 apDiv[i] = &aOvflSpace[apDiv[i]-pParent->aData]; | |
7109 } | |
7110 } | |
7111 dropCell(pParent, i+nxDiv-pParent->nOverflow, szNew[i], &rc); | |
7112 } | |
7113 } | |
7114 | |
7115 /* Make nMaxCells a multiple of 4 in order to preserve 8-byte | |
7116 ** alignment */ | |
7117 nMaxCells = (nMaxCells + 3)&~3; | |
7118 | |
7119 /* | |
7120 ** Allocate space for memory structures | |
7121 */ | |
7122 szScratch = | |
7123 nMaxCells*sizeof(u8*) /* b.apCell */ | |
7124 + nMaxCells*sizeof(u16) /* b.szCell */ | |
7125 + pBt->pageSize; /* aSpace1 */ | |
7126 | |
7127 /* EVIDENCE-OF: R-28375-38319 SQLite will never request a scratch buffer | |
7128 ** that is more than 6 times the database page size. */ | |
7129 assert( szScratch<=6*(int)pBt->pageSize ); | |
7130 b.apCell = sqlite3ScratchMalloc( szScratch ); | |
7131 if( b.apCell==0 ){ | |
7132 rc = SQLITE_NOMEM; | |
7133 goto balance_cleanup; | |
7134 } | |
7135 b.szCell = (u16*)&b.apCell[nMaxCells]; | |
7136 aSpace1 = (u8*)&b.szCell[nMaxCells]; | |
7137 assert( EIGHT_BYTE_ALIGNMENT(aSpace1) ); | |
7138 | |
7139 /* | |
7140 ** Load pointers to all cells on sibling pages and the divider cells | |
7141 ** into the local b.apCell[] array. Make copies of the divider cells | |
7142 ** into space obtained from aSpace1[]. The divider cells have already | |
7143 ** been removed from pParent. | |
7144 ** | |
7145 ** If the siblings are on leaf pages, then the child pointers of the | |
7146 ** divider cells are stripped from the cells before they are copied | |
7147 ** into aSpace1[]. In this way, all cells in b.apCell[] are without | |
7148 ** child pointers. If siblings are not leaves, then all cell in | |
7149 ** b.apCell[] include child pointers. Either way, all cells in b.apCell[] | |
7150 ** are alike. | |
7151 ** | |
7152 ** leafCorrection: 4 if pPage is a leaf. 0 if pPage is not a leaf. | |
7153 ** leafData: 1 if pPage holds key+data and pParent holds only keys. | |
7154 */ | |
7155 b.pRef = apOld[0]; | |
7156 leafCorrection = b.pRef->leaf*4; | |
7157 leafData = b.pRef->intKeyLeaf; | |
7158 for(i=0; i<nOld; i++){ | |
7159 MemPage *pOld = apOld[i]; | |
7160 int limit = pOld->nCell; | |
7161 u8 *aData = pOld->aData; | |
7162 u16 maskPage = pOld->maskPage; | |
7163 u8 *piCell = aData + pOld->cellOffset; | |
7164 u8 *piEnd; | |
7165 | |
7166 /* Verify that all sibling pages are of the same "type" (table-leaf, | |
7167 ** table-interior, index-leaf, or index-interior). | |
7168 */ | |
7169 if( pOld->aData[0]!=apOld[0]->aData[0] ){ | |
7170 rc = SQLITE_CORRUPT_BKPT; | |
7171 goto balance_cleanup; | |
7172 } | |
7173 | |
7174 /* Load b.apCell[] with pointers to all cells in pOld. If pOld | |
7175 ** constains overflow cells, include them in the b.apCell[] array | |
7176 ** in the correct spot. | |
7177 ** | |
7178 ** Note that when there are multiple overflow cells, it is always the | |
7179 ** case that they are sequential and adjacent. This invariant arises | |
7180 ** because multiple overflows can only occurs when inserting divider | |
7181 ** cells into a parent on a prior balance, and divider cells are always | |
7182 ** adjacent and are inserted in order. There is an assert() tagged | |
7183 ** with "NOTE 1" in the overflow cell insertion loop to prove this | |
7184 ** invariant. | |
7185 ** | |
7186 ** This must be done in advance. Once the balance starts, the cell | |
7187 ** offset section of the btree page will be overwritten and we will no | |
7188 ** long be able to find the cells if a pointer to each cell is not saved | |
7189 ** first. | |
7190 */ | |
7191 memset(&b.szCell[b.nCell], 0, sizeof(b.szCell[0])*limit); | |
7192 if( pOld->nOverflow>0 ){ | |
7193 memset(&b.szCell[b.nCell+limit], 0, sizeof(b.szCell[0])*pOld->nOverflow); | |
7194 limit = pOld->aiOvfl[0]; | |
7195 for(j=0; j<limit; j++){ | |
7196 b.apCell[b.nCell] = aData + (maskPage & get2byteAligned(piCell)); | |
7197 piCell += 2; | |
7198 b.nCell++; | |
7199 } | |
7200 for(k=0; k<pOld->nOverflow; k++){ | |
7201 assert( k==0 || pOld->aiOvfl[k-1]+1==pOld->aiOvfl[k] );/* NOTE 1 */ | |
7202 b.apCell[b.nCell] = pOld->apOvfl[k]; | |
7203 b.nCell++; | |
7204 } | |
7205 } | |
7206 piEnd = aData + pOld->cellOffset + 2*pOld->nCell; | |
7207 while( piCell<piEnd ){ | |
7208 assert( b.nCell<nMaxCells ); | |
7209 b.apCell[b.nCell] = aData + (maskPage & get2byteAligned(piCell)); | |
7210 piCell += 2; | |
7211 b.nCell++; | |
7212 } | |
7213 | |
7214 cntOld[i] = b.nCell; | |
7215 if( i<nOld-1 && !leafData){ | |
7216 u16 sz = (u16)szNew[i]; | |
7217 u8 *pTemp; | |
7218 assert( b.nCell<nMaxCells ); | |
7219 b.szCell[b.nCell] = sz; | |
7220 pTemp = &aSpace1[iSpace1]; | |
7221 iSpace1 += sz; | |
7222 assert( sz<=pBt->maxLocal+23 ); | |
7223 assert( iSpace1 <= (int)pBt->pageSize ); | |
7224 memcpy(pTemp, apDiv[i], sz); | |
7225 b.apCell[b.nCell] = pTemp+leafCorrection; | |
7226 assert( leafCorrection==0 || leafCorrection==4 ); | |
7227 b.szCell[b.nCell] = b.szCell[b.nCell] - leafCorrection; | |
7228 if( !pOld->leaf ){ | |
7229 assert( leafCorrection==0 ); | |
7230 assert( pOld->hdrOffset==0 ); | |
7231 /* The right pointer of the child page pOld becomes the left | |
7232 ** pointer of the divider cell */ | |
7233 memcpy(b.apCell[b.nCell], &pOld->aData[8], 4); | |
7234 }else{ | |
7235 assert( leafCorrection==4 ); | |
7236 while( b.szCell[b.nCell]<4 ){ | |
7237 /* Do not allow any cells smaller than 4 bytes. If a smaller cell | |
7238 ** does exist, pad it with 0x00 bytes. */ | |
7239 assert( b.szCell[b.nCell]==3 || CORRUPT_DB ); | |
7240 assert( b.apCell[b.nCell]==&aSpace1[iSpace1-3] || CORRUPT_DB ); | |
7241 aSpace1[iSpace1++] = 0x00; | |
7242 b.szCell[b.nCell]++; | |
7243 } | |
7244 } | |
7245 b.nCell++; | |
7246 } | |
7247 } | |
7248 | |
7249 /* | |
7250 ** Figure out the number of pages needed to hold all b.nCell cells. | |
7251 ** Store this number in "k". Also compute szNew[] which is the total | |
7252 ** size of all cells on the i-th page and cntNew[] which is the index | |
7253 ** in b.apCell[] of the cell that divides page i from page i+1. | |
7254 ** cntNew[k] should equal b.nCell. | |
7255 ** | |
7256 ** Values computed by this block: | |
7257 ** | |
7258 ** k: The total number of sibling pages | |
7259 ** szNew[i]: Spaced used on the i-th sibling page. | |
7260 ** cntNew[i]: Index in b.apCell[] and b.szCell[] for the first cell to | |
7261 ** the right of the i-th sibling page. | |
7262 ** usableSpace: Number of bytes of space available on each sibling. | |
7263 ** | |
7264 */ | |
7265 usableSpace = pBt->usableSize - 12 + leafCorrection; | |
7266 for(i=0; i<nOld; i++){ | |
7267 MemPage *p = apOld[i]; | |
7268 szNew[i] = usableSpace - p->nFree; | |
7269 if( szNew[i]<0 ){ rc = SQLITE_CORRUPT_BKPT; goto balance_cleanup; } | |
7270 for(j=0; j<p->nOverflow; j++){ | |
7271 szNew[i] += 2 + p->xCellSize(p, p->apOvfl[j]); | |
7272 } | |
7273 cntNew[i] = cntOld[i]; | |
7274 } | |
7275 k = nOld; | |
7276 for(i=0; i<k; i++){ | |
7277 int sz; | |
7278 while( szNew[i]>usableSpace ){ | |
7279 if( i+1>=k ){ | |
7280 k = i+2; | |
7281 if( k>NB+2 ){ rc = SQLITE_CORRUPT_BKPT; goto balance_cleanup; } | |
7282 szNew[k-1] = 0; | |
7283 cntNew[k-1] = b.nCell; | |
7284 } | |
7285 sz = 2 + cachedCellSize(&b, cntNew[i]-1); | |
7286 szNew[i] -= sz; | |
7287 if( !leafData ){ | |
7288 if( cntNew[i]<b.nCell ){ | |
7289 sz = 2 + cachedCellSize(&b, cntNew[i]); | |
7290 }else{ | |
7291 sz = 0; | |
7292 } | |
7293 } | |
7294 szNew[i+1] += sz; | |
7295 cntNew[i]--; | |
7296 } | |
7297 while( cntNew[i]<b.nCell ){ | |
7298 sz = 2 + cachedCellSize(&b, cntNew[i]); | |
7299 if( szNew[i]+sz>usableSpace ) break; | |
7300 szNew[i] += sz; | |
7301 cntNew[i]++; | |
7302 if( !leafData ){ | |
7303 if( cntNew[i]<b.nCell ){ | |
7304 sz = 2 + cachedCellSize(&b, cntNew[i]); | |
7305 }else{ | |
7306 sz = 0; | |
7307 } | |
7308 } | |
7309 szNew[i+1] -= sz; | |
7310 } | |
7311 if( cntNew[i]>=b.nCell ){ | |
7312 k = i+1; | |
7313 }else if( cntNew[i] <= (i>0 ? cntNew[i-1] : 0) ){ | |
7314 rc = SQLITE_CORRUPT_BKPT; | |
7315 goto balance_cleanup; | |
7316 } | |
7317 } | |
7318 | |
7319 /* | |
7320 ** The packing computed by the previous block is biased toward the siblings | |
7321 ** on the left side (siblings with smaller keys). The left siblings are | |
7322 ** always nearly full, while the right-most sibling might be nearly empty. | |
7323 ** The next block of code attempts to adjust the packing of siblings to | |
7324 ** get a better balance. | |
7325 ** | |
7326 ** This adjustment is more than an optimization. The packing above might | |
7327 ** be so out of balance as to be illegal. For example, the right-most | |
7328 ** sibling might be completely empty. This adjustment is not optional. | |
7329 */ | |
7330 for(i=k-1; i>0; i--){ | |
7331 int szRight = szNew[i]; /* Size of sibling on the right */ | |
7332 int szLeft = szNew[i-1]; /* Size of sibling on the left */ | |
7333 int r; /* Index of right-most cell in left sibling */ | |
7334 int d; /* Index of first cell to the left of right sibling */ | |
7335 | |
7336 r = cntNew[i-1] - 1; | |
7337 d = r + 1 - leafData; | |
7338 (void)cachedCellSize(&b, d); | |
7339 do{ | |
7340 assert( d<nMaxCells ); | |
7341 assert( r<nMaxCells ); | |
7342 (void)cachedCellSize(&b, r); | |
7343 if( szRight!=0 | |
7344 && (bBulk || szRight+b.szCell[d]+2 > szLeft-(b.szCell[r]+2)) ){ | |
7345 break; | |
7346 } | |
7347 szRight += b.szCell[d] + 2; | |
7348 szLeft -= b.szCell[r] + 2; | |
7349 cntNew[i-1] = r; | |
7350 r--; | |
7351 d--; | |
7352 }while( r>=0 ); | |
7353 szNew[i] = szRight; | |
7354 szNew[i-1] = szLeft; | |
7355 if( cntNew[i-1] <= (i>1 ? cntNew[i-2] : 0) ){ | |
7356 rc = SQLITE_CORRUPT_BKPT; | |
7357 goto balance_cleanup; | |
7358 } | |
7359 } | |
7360 | |
7361 /* Sanity check: For a non-corrupt database file one of the follwing | |
7362 ** must be true: | |
7363 ** (1) We found one or more cells (cntNew[0])>0), or | |
7364 ** (2) pPage is a virtual root page. A virtual root page is when | |
7365 ** the real root page is page 1 and we are the only child of | |
7366 ** that page. | |
7367 */ | |
7368 assert( cntNew[0]>0 || (pParent->pgno==1 && pParent->nCell==0) || CORRUPT_DB); | |
7369 TRACE(("BALANCE: old: %d(nc=%d) %d(nc=%d) %d(nc=%d)\n", | |
7370 apOld[0]->pgno, apOld[0]->nCell, | |
7371 nOld>=2 ? apOld[1]->pgno : 0, nOld>=2 ? apOld[1]->nCell : 0, | |
7372 nOld>=3 ? apOld[2]->pgno : 0, nOld>=3 ? apOld[2]->nCell : 0 | |
7373 )); | |
7374 | |
7375 /* | |
7376 ** Allocate k new pages. Reuse old pages where possible. | |
7377 */ | |
7378 pageFlags = apOld[0]->aData[0]; | |
7379 for(i=0; i<k; i++){ | |
7380 MemPage *pNew; | |
7381 if( i<nOld ){ | |
7382 pNew = apNew[i] = apOld[i]; | |
7383 apOld[i] = 0; | |
7384 rc = sqlite3PagerWrite(pNew->pDbPage); | |
7385 nNew++; | |
7386 if( rc ) goto balance_cleanup; | |
7387 }else{ | |
7388 assert( i>0 ); | |
7389 rc = allocateBtreePage(pBt, &pNew, &pgno, (bBulk ? 1 : pgno), 0); | |
7390 if( rc ) goto balance_cleanup; | |
7391 zeroPage(pNew, pageFlags); | |
7392 apNew[i] = pNew; | |
7393 nNew++; | |
7394 cntOld[i] = b.nCell; | |
7395 | |
7396 /* Set the pointer-map entry for the new sibling page. */ | |
7397 if( ISAUTOVACUUM ){ | |
7398 ptrmapPut(pBt, pNew->pgno, PTRMAP_BTREE, pParent->pgno, &rc); | |
7399 if( rc!=SQLITE_OK ){ | |
7400 goto balance_cleanup; | |
7401 } | |
7402 } | |
7403 } | |
7404 } | |
7405 | |
7406 /* | |
7407 ** Reassign page numbers so that the new pages are in ascending order. | |
7408 ** This helps to keep entries in the disk file in order so that a scan | |
7409 ** of the table is closer to a linear scan through the file. That in turn | |
7410 ** helps the operating system to deliver pages from the disk more rapidly. | |
7411 ** | |
7412 ** An O(n^2) insertion sort algorithm is used, but since n is never more | |
7413 ** than (NB+2) (a small constant), that should not be a problem. | |
7414 ** | |
7415 ** When NB==3, this one optimization makes the database about 25% faster | |
7416 ** for large insertions and deletions. | |
7417 */ | |
7418 for(i=0; i<nNew; i++){ | |
7419 aPgOrder[i] = aPgno[i] = apNew[i]->pgno; | |
7420 aPgFlags[i] = apNew[i]->pDbPage->flags; | |
7421 for(j=0; j<i; j++){ | |
7422 if( aPgno[j]==aPgno[i] ){ | |
7423 /* This branch is taken if the set of sibling pages somehow contains | |
7424 ** duplicate entries. This can happen if the database is corrupt. | |
7425 ** It would be simpler to detect this as part of the loop below, but | |
7426 ** we do the detection here in order to avoid populating the pager | |
7427 ** cache with two separate objects associated with the same | |
7428 ** page number. */ | |
7429 assert( CORRUPT_DB ); | |
7430 rc = SQLITE_CORRUPT_BKPT; | |
7431 goto balance_cleanup; | |
7432 } | |
7433 } | |
7434 } | |
7435 for(i=0; i<nNew; i++){ | |
7436 int iBest = 0; /* aPgno[] index of page number to use */ | |
7437 for(j=1; j<nNew; j++){ | |
7438 if( aPgOrder[j]<aPgOrder[iBest] ) iBest = j; | |
7439 } | |
7440 pgno = aPgOrder[iBest]; | |
7441 aPgOrder[iBest] = 0xffffffff; | |
7442 if( iBest!=i ){ | |
7443 if( iBest>i ){ | |
7444 sqlite3PagerRekey(apNew[iBest]->pDbPage, pBt->nPage+iBest+1, 0); | |
7445 } | |
7446 sqlite3PagerRekey(apNew[i]->pDbPage, pgno, aPgFlags[iBest]); | |
7447 apNew[i]->pgno = pgno; | |
7448 } | |
7449 } | |
7450 | |
7451 TRACE(("BALANCE: new: %d(%d nc=%d) %d(%d nc=%d) %d(%d nc=%d) " | |
7452 "%d(%d nc=%d) %d(%d nc=%d)\n", | |
7453 apNew[0]->pgno, szNew[0], cntNew[0], | |
7454 nNew>=2 ? apNew[1]->pgno : 0, nNew>=2 ? szNew[1] : 0, | |
7455 nNew>=2 ? cntNew[1] - cntNew[0] - !leafData : 0, | |
7456 nNew>=3 ? apNew[2]->pgno : 0, nNew>=3 ? szNew[2] : 0, | |
7457 nNew>=3 ? cntNew[2] - cntNew[1] - !leafData : 0, | |
7458 nNew>=4 ? apNew[3]->pgno : 0, nNew>=4 ? szNew[3] : 0, | |
7459 nNew>=4 ? cntNew[3] - cntNew[2] - !leafData : 0, | |
7460 nNew>=5 ? apNew[4]->pgno : 0, nNew>=5 ? szNew[4] : 0, | |
7461 nNew>=5 ? cntNew[4] - cntNew[3] - !leafData : 0 | |
7462 )); | |
7463 | |
7464 assert( sqlite3PagerIswriteable(pParent->pDbPage) ); | |
7465 put4byte(pRight, apNew[nNew-1]->pgno); | |
7466 | |
7467 /* If the sibling pages are not leaves, ensure that the right-child pointer | |
7468 ** of the right-most new sibling page is set to the value that was | |
7469 ** originally in the same field of the right-most old sibling page. */ | |
7470 if( (pageFlags & PTF_LEAF)==0 && nOld!=nNew ){ | |
7471 MemPage *pOld = (nNew>nOld ? apNew : apOld)[nOld-1]; | |
7472 memcpy(&apNew[nNew-1]->aData[8], &pOld->aData[8], 4); | |
7473 } | |
7474 | |
7475 /* Make any required updates to pointer map entries associated with | |
7476 ** cells stored on sibling pages following the balance operation. Pointer | |
7477 ** map entries associated with divider cells are set by the insertCell() | |
7478 ** routine. The associated pointer map entries are: | |
7479 ** | |
7480 ** a) if the cell contains a reference to an overflow chain, the | |
7481 ** entry associated with the first page in the overflow chain, and | |
7482 ** | |
7483 ** b) if the sibling pages are not leaves, the child page associated | |
7484 ** with the cell. | |
7485 ** | |
7486 ** If the sibling pages are not leaves, then the pointer map entry | |
7487 ** associated with the right-child of each sibling may also need to be | |
7488 ** updated. This happens below, after the sibling pages have been | |
7489 ** populated, not here. | |
7490 */ | |
7491 if( ISAUTOVACUUM ){ | |
7492 MemPage *pNew = apNew[0]; | |
7493 u8 *aOld = pNew->aData; | |
7494 int cntOldNext = pNew->nCell + pNew->nOverflow; | |
7495 int usableSize = pBt->usableSize; | |
7496 int iNew = 0; | |
7497 int iOld = 0; | |
7498 | |
7499 for(i=0; i<b.nCell; i++){ | |
7500 u8 *pCell = b.apCell[i]; | |
7501 if( i==cntOldNext ){ | |
7502 MemPage *pOld = (++iOld)<nNew ? apNew[iOld] : apOld[iOld]; | |
7503 cntOldNext += pOld->nCell + pOld->nOverflow + !leafData; | |
7504 aOld = pOld->aData; | |
7505 } | |
7506 if( i==cntNew[iNew] ){ | |
7507 pNew = apNew[++iNew]; | |
7508 if( !leafData ) continue; | |
7509 } | |
7510 | |
7511 /* Cell pCell is destined for new sibling page pNew. Originally, it | |
7512 ** was either part of sibling page iOld (possibly an overflow cell), | |
7513 ** or else the divider cell to the left of sibling page iOld. So, | |
7514 ** if sibling page iOld had the same page number as pNew, and if | |
7515 ** pCell really was a part of sibling page iOld (not a divider or | |
7516 ** overflow cell), we can skip updating the pointer map entries. */ | |
7517 if( iOld>=nNew | |
7518 || pNew->pgno!=aPgno[iOld] | |
7519 || !SQLITE_WITHIN(pCell,aOld,&aOld[usableSize]) | |
7520 ){ | |
7521 if( !leafCorrection ){ | |
7522 ptrmapPut(pBt, get4byte(pCell), PTRMAP_BTREE, pNew->pgno, &rc); | |
7523 } | |
7524 if( cachedCellSize(&b,i)>pNew->minLocal ){ | |
7525 ptrmapPutOvflPtr(pNew, pCell, &rc); | |
7526 } | |
7527 if( rc ) goto balance_cleanup; | |
7528 } | |
7529 } | |
7530 } | |
7531 | |
7532 /* Insert new divider cells into pParent. */ | |
7533 for(i=0; i<nNew-1; i++){ | |
7534 u8 *pCell; | |
7535 u8 *pTemp; | |
7536 int sz; | |
7537 MemPage *pNew = apNew[i]; | |
7538 j = cntNew[i]; | |
7539 | |
7540 assert( j<nMaxCells ); | |
7541 assert( b.apCell[j]!=0 ); | |
7542 pCell = b.apCell[j]; | |
7543 sz = b.szCell[j] + leafCorrection; | |
7544 pTemp = &aOvflSpace[iOvflSpace]; | |
7545 if( !pNew->leaf ){ | |
7546 memcpy(&pNew->aData[8], pCell, 4); | |
7547 }else if( leafData ){ | |
7548 /* If the tree is a leaf-data tree, and the siblings are leaves, | |
7549 ** then there is no divider cell in b.apCell[]. Instead, the divider | |
7550 ** cell consists of the integer key for the right-most cell of | |
7551 ** the sibling-page assembled above only. | |
7552 */ | |
7553 CellInfo info; | |
7554 j--; | |
7555 pNew->xParseCell(pNew, b.apCell[j], &info); | |
7556 pCell = pTemp; | |
7557 sz = 4 + putVarint(&pCell[4], info.nKey); | |
7558 pTemp = 0; | |
7559 }else{ | |
7560 pCell -= 4; | |
7561 /* Obscure case for non-leaf-data trees: If the cell at pCell was | |
7562 ** previously stored on a leaf node, and its reported size was 4 | |
7563 ** bytes, then it may actually be smaller than this | |
7564 ** (see btreeParseCellPtr(), 4 bytes is the minimum size of | |
7565 ** any cell). But it is important to pass the correct size to | |
7566 ** insertCell(), so reparse the cell now. | |
7567 ** | |
7568 ** Note that this can never happen in an SQLite data file, as all | |
7569 ** cells are at least 4 bytes. It only happens in b-trees used | |
7570 ** to evaluate "IN (SELECT ...)" and similar clauses. | |
7571 */ | |
7572 if( b.szCell[j]==4 ){ | |
7573 assert(leafCorrection==4); | |
7574 sz = pParent->xCellSize(pParent, pCell); | |
7575 } | |
7576 } | |
7577 iOvflSpace += sz; | |
7578 assert( sz<=pBt->maxLocal+23 ); | |
7579 assert( iOvflSpace <= (int)pBt->pageSize ); | |
7580 insertCell(pParent, nxDiv+i, pCell, sz, pTemp, pNew->pgno, &rc); | |
7581 if( rc!=SQLITE_OK ) goto balance_cleanup; | |
7582 assert( sqlite3PagerIswriteable(pParent->pDbPage) ); | |
7583 } | |
7584 | |
7585 /* Now update the actual sibling pages. The order in which they are updated | |
7586 ** is important, as this code needs to avoid disrupting any page from which | |
7587 ** cells may still to be read. In practice, this means: | |
7588 ** | |
7589 ** (1) If cells are moving left (from apNew[iPg] to apNew[iPg-1]) | |
7590 ** then it is not safe to update page apNew[iPg] until after | |
7591 ** the left-hand sibling apNew[iPg-1] has been updated. | |
7592 ** | |
7593 ** (2) If cells are moving right (from apNew[iPg] to apNew[iPg+1]) | |
7594 ** then it is not safe to update page apNew[iPg] until after | |
7595 ** the right-hand sibling apNew[iPg+1] has been updated. | |
7596 ** | |
7597 ** If neither of the above apply, the page is safe to update. | |
7598 ** | |
7599 ** The iPg value in the following loop starts at nNew-1 goes down | |
7600 ** to 0, then back up to nNew-1 again, thus making two passes over | |
7601 ** the pages. On the initial downward pass, only condition (1) above | |
7602 ** needs to be tested because (2) will always be true from the previous | |
7603 ** step. On the upward pass, both conditions are always true, so the | |
7604 ** upwards pass simply processes pages that were missed on the downward | |
7605 ** pass. | |
7606 */ | |
7607 for(i=1-nNew; i<nNew; i++){ | |
7608 int iPg = i<0 ? -i : i; | |
7609 assert( iPg>=0 && iPg<nNew ); | |
7610 if( abDone[iPg] ) continue; /* Skip pages already processed */ | |
7611 if( i>=0 /* On the upwards pass, or... */ | |
7612 || cntOld[iPg-1]>=cntNew[iPg-1] /* Condition (1) is true */ | |
7613 ){ | |
7614 int iNew; | |
7615 int iOld; | |
7616 int nNewCell; | |
7617 | |
7618 /* Verify condition (1): If cells are moving left, update iPg | |
7619 ** only after iPg-1 has already been updated. */ | |
7620 assert( iPg==0 || cntOld[iPg-1]>=cntNew[iPg-1] || abDone[iPg-1] ); | |
7621 | |
7622 /* Verify condition (2): If cells are moving right, update iPg | |
7623 ** only after iPg+1 has already been updated. */ | |
7624 assert( cntNew[iPg]>=cntOld[iPg] || abDone[iPg+1] ); | |
7625 | |
7626 if( iPg==0 ){ | |
7627 iNew = iOld = 0; | |
7628 nNewCell = cntNew[0]; | |
7629 }else{ | |
7630 iOld = iPg<nOld ? (cntOld[iPg-1] + !leafData) : b.nCell; | |
7631 iNew = cntNew[iPg-1] + !leafData; | |
7632 nNewCell = cntNew[iPg] - iNew; | |
7633 } | |
7634 | |
7635 rc = editPage(apNew[iPg], iOld, iNew, nNewCell, &b); | |
7636 if( rc ) goto balance_cleanup; | |
7637 abDone[iPg]++; | |
7638 apNew[iPg]->nFree = usableSpace-szNew[iPg]; | |
7639 assert( apNew[iPg]->nOverflow==0 ); | |
7640 assert( apNew[iPg]->nCell==nNewCell ); | |
7641 } | |
7642 } | |
7643 | |
7644 /* All pages have been processed exactly once */ | |
7645 assert( memcmp(abDone, "\01\01\01\01\01", nNew)==0 ); | |
7646 | |
7647 assert( nOld>0 ); | |
7648 assert( nNew>0 ); | |
7649 | |
7650 if( isRoot && pParent->nCell==0 && pParent->hdrOffset<=apNew[0]->nFree ){ | |
7651 /* The root page of the b-tree now contains no cells. The only sibling | |
7652 ** page is the right-child of the parent. Copy the contents of the | |
7653 ** child page into the parent, decreasing the overall height of the | |
7654 ** b-tree structure by one. This is described as the "balance-shallower" | |
7655 ** sub-algorithm in some documentation. | |
7656 ** | |
7657 ** If this is an auto-vacuum database, the call to copyNodeContent() | |
7658 ** sets all pointer-map entries corresponding to database image pages | |
7659 ** for which the pointer is stored within the content being copied. | |
7660 ** | |
7661 ** It is critical that the child page be defragmented before being | |
7662 ** copied into the parent, because if the parent is page 1 then it will | |
7663 ** by smaller than the child due to the database header, and so all the | |
7664 ** free space needs to be up front. | |
7665 */ | |
7666 assert( nNew==1 || CORRUPT_DB ); | |
7667 rc = defragmentPage(apNew[0]); | |
7668 testcase( rc!=SQLITE_OK ); | |
7669 assert( apNew[0]->nFree == | |
7670 (get2byte(&apNew[0]->aData[5])-apNew[0]->cellOffset-apNew[0]->nCell*2) | |
7671 || rc!=SQLITE_OK | |
7672 ); | |
7673 copyNodeContent(apNew[0], pParent, &rc); | |
7674 freePage(apNew[0], &rc); | |
7675 }else if( ISAUTOVACUUM && !leafCorrection ){ | |
7676 /* Fix the pointer map entries associated with the right-child of each | |
7677 ** sibling page. All other pointer map entries have already been taken | |
7678 ** care of. */ | |
7679 for(i=0; i<nNew; i++){ | |
7680 u32 key = get4byte(&apNew[i]->aData[8]); | |
7681 ptrmapPut(pBt, key, PTRMAP_BTREE, apNew[i]->pgno, &rc); | |
7682 } | |
7683 } | |
7684 | |
7685 assert( pParent->isInit ); | |
7686 TRACE(("BALANCE: finished: old=%d new=%d cells=%d\n", | |
7687 nOld, nNew, b.nCell)); | |
7688 | |
7689 /* Free any old pages that were not reused as new pages. | |
7690 */ | |
7691 for(i=nNew; i<nOld; i++){ | |
7692 freePage(apOld[i], &rc); | |
7693 } | |
7694 | |
7695 #if 0 | |
7696 if( ISAUTOVACUUM && rc==SQLITE_OK && apNew[0]->isInit ){ | |
7697 /* The ptrmapCheckPages() contains assert() statements that verify that | |
7698 ** all pointer map pages are set correctly. This is helpful while | |
7699 ** debugging. This is usually disabled because a corrupt database may | |
7700 ** cause an assert() statement to fail. */ | |
7701 ptrmapCheckPages(apNew, nNew); | |
7702 ptrmapCheckPages(&pParent, 1); | |
7703 } | |
7704 #endif | |
7705 | |
7706 /* | |
7707 ** Cleanup before returning. | |
7708 */ | |
7709 balance_cleanup: | |
7710 sqlite3ScratchFree(b.apCell); | |
7711 for(i=0; i<nOld; i++){ | |
7712 releasePage(apOld[i]); | |
7713 } | |
7714 for(i=0; i<nNew; i++){ | |
7715 releasePage(apNew[i]); | |
7716 } | |
7717 | |
7718 return rc; | |
7719 } | |
7720 | |
7721 | |
7722 /* | |
7723 ** This function is called when the root page of a b-tree structure is | |
7724 ** overfull (has one or more overflow pages). | |
7725 ** | |
7726 ** A new child page is allocated and the contents of the current root | |
7727 ** page, including overflow cells, are copied into the child. The root | |
7728 ** page is then overwritten to make it an empty page with the right-child | |
7729 ** pointer pointing to the new page. | |
7730 ** | |
7731 ** Before returning, all pointer-map entries corresponding to pages | |
7732 ** that the new child-page now contains pointers to are updated. The | |
7733 ** entry corresponding to the new right-child pointer of the root | |
7734 ** page is also updated. | |
7735 ** | |
7736 ** If successful, *ppChild is set to contain a reference to the child | |
7737 ** page and SQLITE_OK is returned. In this case the caller is required | |
7738 ** to call releasePage() on *ppChild exactly once. If an error occurs, | |
7739 ** an error code is returned and *ppChild is set to 0. | |
7740 */ | |
7741 static int balance_deeper(MemPage *pRoot, MemPage **ppChild){ | |
7742 int rc; /* Return value from subprocedures */ | |
7743 MemPage *pChild = 0; /* Pointer to a new child page */ | |
7744 Pgno pgnoChild = 0; /* Page number of the new child page */ | |
7745 BtShared *pBt = pRoot->pBt; /* The BTree */ | |
7746 | |
7747 assert( pRoot->nOverflow>0 ); | |
7748 assert( sqlite3_mutex_held(pBt->mutex) ); | |
7749 | |
7750 /* Make pRoot, the root page of the b-tree, writable. Allocate a new | |
7751 ** page that will become the new right-child of pPage. Copy the contents | |
7752 ** of the node stored on pRoot into the new child page. | |
7753 */ | |
7754 rc = sqlite3PagerWrite(pRoot->pDbPage); | |
7755 if( rc==SQLITE_OK ){ | |
7756 rc = allocateBtreePage(pBt,&pChild,&pgnoChild,pRoot->pgno,0); | |
7757 copyNodeContent(pRoot, pChild, &rc); | |
7758 if( ISAUTOVACUUM ){ | |
7759 ptrmapPut(pBt, pgnoChild, PTRMAP_BTREE, pRoot->pgno, &rc); | |
7760 } | |
7761 } | |
7762 if( rc ){ | |
7763 *ppChild = 0; | |
7764 releasePage(pChild); | |
7765 return rc; | |
7766 } | |
7767 assert( sqlite3PagerIswriteable(pChild->pDbPage) ); | |
7768 assert( sqlite3PagerIswriteable(pRoot->pDbPage) ); | |
7769 assert( pChild->nCell==pRoot->nCell ); | |
7770 | |
7771 TRACE(("BALANCE: copy root %d into %d\n", pRoot->pgno, pChild->pgno)); | |
7772 | |
7773 /* Copy the overflow cells from pRoot to pChild */ | |
7774 memcpy(pChild->aiOvfl, pRoot->aiOvfl, | |
7775 pRoot->nOverflow*sizeof(pRoot->aiOvfl[0])); | |
7776 memcpy(pChild->apOvfl, pRoot->apOvfl, | |
7777 pRoot->nOverflow*sizeof(pRoot->apOvfl[0])); | |
7778 pChild->nOverflow = pRoot->nOverflow; | |
7779 | |
7780 /* Zero the contents of pRoot. Then install pChild as the right-child. */ | |
7781 zeroPage(pRoot, pChild->aData[0] & ~PTF_LEAF); | |
7782 put4byte(&pRoot->aData[pRoot->hdrOffset+8], pgnoChild); | |
7783 | |
7784 *ppChild = pChild; | |
7785 return SQLITE_OK; | |
7786 } | |
7787 | |
7788 /* | |
7789 ** The page that pCur currently points to has just been modified in | |
7790 ** some way. This function figures out if this modification means the | |
7791 ** tree needs to be balanced, and if so calls the appropriate balancing | |
7792 ** routine. Balancing routines are: | |
7793 ** | |
7794 ** balance_quick() | |
7795 ** balance_deeper() | |
7796 ** balance_nonroot() | |
7797 */ | |
7798 static int balance(BtCursor *pCur){ | |
7799 int rc = SQLITE_OK; | |
7800 const int nMin = pCur->pBt->usableSize * 2 / 3; | |
7801 u8 aBalanceQuickSpace[13]; | |
7802 u8 *pFree = 0; | |
7803 | |
7804 TESTONLY( int balance_quick_called = 0 ); | |
7805 TESTONLY( int balance_deeper_called = 0 ); | |
7806 | |
7807 do { | |
7808 int iPage = pCur->iPage; | |
7809 MemPage *pPage = pCur->apPage[iPage]; | |
7810 | |
7811 if( iPage==0 ){ | |
7812 if( pPage->nOverflow ){ | |
7813 /* The root page of the b-tree is overfull. In this case call the | |
7814 ** balance_deeper() function to create a new child for the root-page | |
7815 ** and copy the current contents of the root-page to it. The | |
7816 ** next iteration of the do-loop will balance the child page. | |
7817 */ | |
7818 assert( (balance_deeper_called++)==0 ); | |
7819 rc = balance_deeper(pPage, &pCur->apPage[1]); | |
7820 if( rc==SQLITE_OK ){ | |
7821 pCur->iPage = 1; | |
7822 pCur->aiIdx[0] = 0; | |
7823 pCur->aiIdx[1] = 0; | |
7824 assert( pCur->apPage[1]->nOverflow ); | |
7825 } | |
7826 }else{ | |
7827 break; | |
7828 } | |
7829 }else if( pPage->nOverflow==0 && pPage->nFree<=nMin ){ | |
7830 break; | |
7831 }else{ | |
7832 MemPage * const pParent = pCur->apPage[iPage-1]; | |
7833 int const iIdx = pCur->aiIdx[iPage-1]; | |
7834 | |
7835 rc = sqlite3PagerWrite(pParent->pDbPage); | |
7836 if( rc==SQLITE_OK ){ | |
7837 #ifndef SQLITE_OMIT_QUICKBALANCE | |
7838 if( pPage->intKeyLeaf | |
7839 && pPage->nOverflow==1 | |
7840 && pPage->aiOvfl[0]==pPage->nCell | |
7841 && pParent->pgno!=1 | |
7842 && pParent->nCell==iIdx | |
7843 ){ | |
7844 /* Call balance_quick() to create a new sibling of pPage on which | |
7845 ** to store the overflow cell. balance_quick() inserts a new cell | |
7846 ** into pParent, which may cause pParent overflow. If this | |
7847 ** happens, the next iteration of the do-loop will balance pParent | |
7848 ** use either balance_nonroot() or balance_deeper(). Until this | |
7849 ** happens, the overflow cell is stored in the aBalanceQuickSpace[] | |
7850 ** buffer. | |
7851 ** | |
7852 ** The purpose of the following assert() is to check that only a | |
7853 ** single call to balance_quick() is made for each call to this | |
7854 ** function. If this were not verified, a subtle bug involving reuse | |
7855 ** of the aBalanceQuickSpace[] might sneak in. | |
7856 */ | |
7857 assert( (balance_quick_called++)==0 ); | |
7858 rc = balance_quick(pParent, pPage, aBalanceQuickSpace); | |
7859 }else | |
7860 #endif | |
7861 { | |
7862 /* In this case, call balance_nonroot() to redistribute cells | |
7863 ** between pPage and up to 2 of its sibling pages. This involves | |
7864 ** modifying the contents of pParent, which may cause pParent to | |
7865 ** become overfull or underfull. The next iteration of the do-loop | |
7866 ** will balance the parent page to correct this. | |
7867 ** | |
7868 ** If the parent page becomes overfull, the overflow cell or cells | |
7869 ** are stored in the pSpace buffer allocated immediately below. | |
7870 ** A subsequent iteration of the do-loop will deal with this by | |
7871 ** calling balance_nonroot() (balance_deeper() may be called first, | |
7872 ** but it doesn't deal with overflow cells - just moves them to a | |
7873 ** different page). Once this subsequent call to balance_nonroot() | |
7874 ** has completed, it is safe to release the pSpace buffer used by | |
7875 ** the previous call, as the overflow cell data will have been | |
7876 ** copied either into the body of a database page or into the new | |
7877 ** pSpace buffer passed to the latter call to balance_nonroot(). | |
7878 */ | |
7879 u8 *pSpace = sqlite3PageMalloc(pCur->pBt->pageSize); | |
7880 rc = balance_nonroot(pParent, iIdx, pSpace, iPage==1, | |
7881 pCur->hints&BTREE_BULKLOAD); | |
7882 if( pFree ){ | |
7883 /* If pFree is not NULL, it points to the pSpace buffer used | |
7884 ** by a previous call to balance_nonroot(). Its contents are | |
7885 ** now stored either on real database pages or within the | |
7886 ** new pSpace buffer, so it may be safely freed here. */ | |
7887 sqlite3PageFree(pFree); | |
7888 } | |
7889 | |
7890 /* The pSpace buffer will be freed after the next call to | |
7891 ** balance_nonroot(), or just before this function returns, whichever | |
7892 ** comes first. */ | |
7893 pFree = pSpace; | |
7894 } | |
7895 } | |
7896 | |
7897 pPage->nOverflow = 0; | |
7898 | |
7899 /* The next iteration of the do-loop balances the parent page. */ | |
7900 releasePage(pPage); | |
7901 pCur->iPage--; | |
7902 assert( pCur->iPage>=0 ); | |
7903 } | |
7904 }while( rc==SQLITE_OK ); | |
7905 | |
7906 if( pFree ){ | |
7907 sqlite3PageFree(pFree); | |
7908 } | |
7909 return rc; | |
7910 } | |
7911 | |
7912 | |
7913 /* | |
7914 ** Insert a new record into the BTree. The key is given by (pKey,nKey) | |
7915 ** and the data is given by (pData,nData). The cursor is used only to | |
7916 ** define what table the record should be inserted into. The cursor | |
7917 ** is left pointing at a random location. | |
7918 ** | |
7919 ** For an INTKEY table, only the nKey value of the key is used. pKey is | |
7920 ** ignored. For a ZERODATA table, the pData and nData are both ignored. | |
7921 ** | |
7922 ** If the seekResult parameter is non-zero, then a successful call to | |
7923 ** MovetoUnpacked() to seek cursor pCur to (pKey, nKey) has already | |
7924 ** been performed. seekResult is the search result returned (a negative | |
7925 ** number if pCur points at an entry that is smaller than (pKey, nKey), or | |
7926 ** a positive value if pCur points at an entry that is larger than | |
7927 ** (pKey, nKey)). | |
7928 ** | |
7929 ** If the seekResult parameter is non-zero, then the caller guarantees that | |
7930 ** cursor pCur is pointing at the existing copy of a row that is to be | |
7931 ** overwritten. If the seekResult parameter is 0, then cursor pCur may | |
7932 ** point to any entry or to no entry at all and so this function has to seek | |
7933 ** the cursor before the new key can be inserted. | |
7934 */ | |
7935 int sqlite3BtreeInsert( | |
7936 BtCursor *pCur, /* Insert data into the table of this cursor */ | |
7937 const void *pKey, i64 nKey, /* The key of the new record */ | |
7938 const void *pData, int nData, /* The data of the new record */ | |
7939 int nZero, /* Number of extra 0 bytes to append to data */ | |
7940 int appendBias, /* True if this is likely an append */ | |
7941 int seekResult /* Result of prior MovetoUnpacked() call */ | |
7942 ){ | |
7943 int rc; | |
7944 int loc = seekResult; /* -1: before desired location +1: after */ | |
7945 int szNew = 0; | |
7946 int idx; | |
7947 MemPage *pPage; | |
7948 Btree *p = pCur->pBtree; | |
7949 BtShared *pBt = p->pBt; | |
7950 unsigned char *oldCell; | |
7951 unsigned char *newCell = 0; | |
7952 | |
7953 if( pCur->eState==CURSOR_FAULT ){ | |
7954 assert( pCur->skipNext!=SQLITE_OK ); | |
7955 return pCur->skipNext; | |
7956 } | |
7957 | |
7958 assert( cursorHoldsMutex(pCur) ); | |
7959 assert( (pCur->curFlags & BTCF_WriteFlag)!=0 | |
7960 && pBt->inTransaction==TRANS_WRITE | |
7961 && (pBt->btsFlags & BTS_READ_ONLY)==0 ); | |
7962 assert( hasSharedCacheTableLock(p, pCur->pgnoRoot, pCur->pKeyInfo!=0, 2) ); | |
7963 | |
7964 /* Assert that the caller has been consistent. If this cursor was opened | |
7965 ** expecting an index b-tree, then the caller should be inserting blob | |
7966 ** keys with no associated data. If the cursor was opened expecting an | |
7967 ** intkey table, the caller should be inserting integer keys with a | |
7968 ** blob of associated data. */ | |
7969 assert( (pKey==0)==(pCur->pKeyInfo==0) ); | |
7970 | |
7971 /* Save the positions of any other cursors open on this table. | |
7972 ** | |
7973 ** In some cases, the call to btreeMoveto() below is a no-op. For | |
7974 ** example, when inserting data into a table with auto-generated integer | |
7975 ** keys, the VDBE layer invokes sqlite3BtreeLast() to figure out the | |
7976 ** integer key to use. It then calls this function to actually insert the | |
7977 ** data into the intkey B-Tree. In this case btreeMoveto() recognizes | |
7978 ** that the cursor is already where it needs to be and returns without | |
7979 ** doing any work. To avoid thwarting these optimizations, it is important | |
7980 ** not to clear the cursor here. | |
7981 */ | |
7982 if( pCur->curFlags & BTCF_Multiple ){ | |
7983 rc = saveAllCursors(pBt, pCur->pgnoRoot, pCur); | |
7984 if( rc ) return rc; | |
7985 } | |
7986 | |
7987 if( pCur->pKeyInfo==0 ){ | |
7988 assert( pKey==0 ); | |
7989 /* If this is an insert into a table b-tree, invalidate any incrblob | |
7990 ** cursors open on the row being replaced */ | |
7991 invalidateIncrblobCursors(p, nKey, 0); | |
7992 | |
7993 /* If the cursor is currently on the last row and we are appending a | |
7994 ** new row onto the end, set the "loc" to avoid an unnecessary | |
7995 ** btreeMoveto() call */ | |
7996 if( (pCur->curFlags&BTCF_ValidNKey)!=0 && nKey>0 | |
7997 && pCur->info.nKey==nKey-1 ){ | |
7998 loc = -1; | |
7999 }else if( loc==0 ){ | |
8000 rc = sqlite3BtreeMovetoUnpacked(pCur, 0, nKey, appendBias, &loc); | |
8001 if( rc ) return rc; | |
8002 } | |
8003 }else if( loc==0 ){ | |
8004 rc = btreeMoveto(pCur, pKey, nKey, appendBias, &loc); | |
8005 if( rc ) return rc; | |
8006 } | |
8007 assert( pCur->eState==CURSOR_VALID || (pCur->eState==CURSOR_INVALID && loc) ); | |
8008 | |
8009 pPage = pCur->apPage[pCur->iPage]; | |
8010 assert( pPage->intKey || nKey>=0 ); | |
8011 assert( pPage->leaf || !pPage->intKey ); | |
8012 | |
8013 TRACE(("INSERT: table=%d nkey=%lld ndata=%d page=%d %s\n", | |
8014 pCur->pgnoRoot, nKey, nData, pPage->pgno, | |
8015 loc==0 ? "overwrite" : "new entry")); | |
8016 assert( pPage->isInit ); | |
8017 newCell = pBt->pTmpSpace; | |
8018 assert( newCell!=0 ); | |
8019 rc = fillInCell(pPage, newCell, pKey, nKey, pData, nData, nZero, &szNew); | |
8020 if( rc ) goto end_insert; | |
8021 assert( szNew==pPage->xCellSize(pPage, newCell) ); | |
8022 assert( szNew <= MX_CELL_SIZE(pBt) ); | |
8023 idx = pCur->aiIdx[pCur->iPage]; | |
8024 if( loc==0 ){ | |
8025 u16 szOld; | |
8026 assert( idx<pPage->nCell ); | |
8027 rc = sqlite3PagerWrite(pPage->pDbPage); | |
8028 if( rc ){ | |
8029 goto end_insert; | |
8030 } | |
8031 oldCell = findCell(pPage, idx); | |
8032 if( !pPage->leaf ){ | |
8033 memcpy(newCell, oldCell, 4); | |
8034 } | |
8035 rc = clearCell(pPage, oldCell, &szOld); | |
8036 dropCell(pPage, idx, szOld, &rc); | |
8037 if( rc ) goto end_insert; | |
8038 }else if( loc<0 && pPage->nCell>0 ){ | |
8039 assert( pPage->leaf ); | |
8040 idx = ++pCur->aiIdx[pCur->iPage]; | |
8041 }else{ | |
8042 assert( pPage->leaf ); | |
8043 } | |
8044 insertCell(pPage, idx, newCell, szNew, 0, 0, &rc); | |
8045 assert( rc!=SQLITE_OK || pPage->nCell>0 || pPage->nOverflow>0 ); | |
8046 | |
8047 /* If no error has occurred and pPage has an overflow cell, call balance() | |
8048 ** to redistribute the cells within the tree. Since balance() may move | |
8049 ** the cursor, zero the BtCursor.info.nSize and BTCF_ValidNKey | |
8050 ** variables. | |
8051 ** | |
8052 ** Previous versions of SQLite called moveToRoot() to move the cursor | |
8053 ** back to the root page as balance() used to invalidate the contents | |
8054 ** of BtCursor.apPage[] and BtCursor.aiIdx[]. Instead of doing that, | |
8055 ** set the cursor state to "invalid". This makes common insert operations | |
8056 ** slightly faster. | |
8057 ** | |
8058 ** There is a subtle but important optimization here too. When inserting | |
8059 ** multiple records into an intkey b-tree using a single cursor (as can | |
8060 ** happen while processing an "INSERT INTO ... SELECT" statement), it | |
8061 ** is advantageous to leave the cursor pointing to the last entry in | |
8062 ** the b-tree if possible. If the cursor is left pointing to the last | |
8063 ** entry in the table, and the next row inserted has an integer key | |
8064 ** larger than the largest existing key, it is possible to insert the | |
8065 ** row without seeking the cursor. This can be a big performance boost. | |
8066 */ | |
8067 pCur->info.nSize = 0; | |
8068 if( rc==SQLITE_OK && pPage->nOverflow ){ | |
8069 pCur->curFlags &= ~(BTCF_ValidNKey); | |
8070 rc = balance(pCur); | |
8071 | |
8072 /* Must make sure nOverflow is reset to zero even if the balance() | |
8073 ** fails. Internal data structure corruption will result otherwise. | |
8074 ** Also, set the cursor state to invalid. This stops saveCursorPosition() | |
8075 ** from trying to save the current position of the cursor. */ | |
8076 pCur->apPage[pCur->iPage]->nOverflow = 0; | |
8077 pCur->eState = CURSOR_INVALID; | |
8078 } | |
8079 assert( pCur->apPage[pCur->iPage]->nOverflow==0 ); | |
8080 | |
8081 end_insert: | |
8082 return rc; | |
8083 } | |
8084 | |
8085 /* | |
8086 ** Delete the entry that the cursor is pointing to. | |
8087 ** | |
8088 ** If the second parameter is zero, then the cursor is left pointing at an | |
8089 ** arbitrary location after the delete. If it is non-zero, then the cursor | |
8090 ** is left in a state such that the next call to BtreeNext() or BtreePrev() | |
8091 ** moves it to the same row as it would if the call to BtreeDelete() had | |
8092 ** been omitted. | |
8093 */ | |
8094 int sqlite3BtreeDelete(BtCursor *pCur, int bPreserve){ | |
8095 Btree *p = pCur->pBtree; | |
8096 BtShared *pBt = p->pBt; | |
8097 int rc; /* Return code */ | |
8098 MemPage *pPage; /* Page to delete cell from */ | |
8099 unsigned char *pCell; /* Pointer to cell to delete */ | |
8100 int iCellIdx; /* Index of cell to delete */ | |
8101 int iCellDepth; /* Depth of node containing pCell */ | |
8102 u16 szCell; /* Size of the cell being deleted */ | |
8103 int bSkipnext = 0; /* Leaf cursor in SKIPNEXT state */ | |
8104 | |
8105 assert( cursorHoldsMutex(pCur) ); | |
8106 assert( pBt->inTransaction==TRANS_WRITE ); | |
8107 assert( (pBt->btsFlags & BTS_READ_ONLY)==0 ); | |
8108 assert( pCur->curFlags & BTCF_WriteFlag ); | |
8109 assert( hasSharedCacheTableLock(p, pCur->pgnoRoot, pCur->pKeyInfo!=0, 2) ); | |
8110 assert( !hasReadConflicts(p, pCur->pgnoRoot) ); | |
8111 assert( pCur->aiIdx[pCur->iPage]<pCur->apPage[pCur->iPage]->nCell ); | |
8112 assert( pCur->eState==CURSOR_VALID ); | |
8113 | |
8114 iCellDepth = pCur->iPage; | |
8115 iCellIdx = pCur->aiIdx[iCellDepth]; | |
8116 pPage = pCur->apPage[iCellDepth]; | |
8117 pCell = findCell(pPage, iCellIdx); | |
8118 | |
8119 /* If the page containing the entry to delete is not a leaf page, move | |
8120 ** the cursor to the largest entry in the tree that is smaller than | |
8121 ** the entry being deleted. This cell will replace the cell being deleted | |
8122 ** from the internal node. The 'previous' entry is used for this instead | |
8123 ** of the 'next' entry, as the previous entry is always a part of the | |
8124 ** sub-tree headed by the child page of the cell being deleted. This makes | |
8125 ** balancing the tree following the delete operation easier. */ | |
8126 if( !pPage->leaf ){ | |
8127 int notUsed = 0; | |
8128 rc = sqlite3BtreePrevious(pCur, ¬Used); | |
8129 if( rc ) return rc; | |
8130 } | |
8131 | |
8132 /* Save the positions of any other cursors open on this table before | |
8133 ** making any modifications. */ | |
8134 if( pCur->curFlags & BTCF_Multiple ){ | |
8135 rc = saveAllCursors(pBt, pCur->pgnoRoot, pCur); | |
8136 if( rc ) return rc; | |
8137 } | |
8138 | |
8139 /* If this is a delete operation to remove a row from a table b-tree, | |
8140 ** invalidate any incrblob cursors open on the row being deleted. */ | |
8141 if( pCur->pKeyInfo==0 ){ | |
8142 invalidateIncrblobCursors(p, pCur->info.nKey, 0); | |
8143 } | |
8144 | |
8145 /* If the bPreserve flag is set to true, then the cursor position must | |
8146 ** be preserved following this delete operation. If the current delete | |
8147 ** will cause a b-tree rebalance, then this is done by saving the cursor | |
8148 ** key and leaving the cursor in CURSOR_REQUIRESEEK state before | |
8149 ** returning. | |
8150 ** | |
8151 ** Or, if the current delete will not cause a rebalance, then the cursor | |
8152 ** will be left in CURSOR_SKIPNEXT state pointing to the entry immediately | |
8153 ** before or after the deleted entry. In this case set bSkipnext to true. */ | |
8154 if( bPreserve ){ | |
8155 if( !pPage->leaf | |
8156 || (pPage->nFree+cellSizePtr(pPage,pCell)+2)>(int)(pBt->usableSize*2/3) | |
8157 ){ | |
8158 /* A b-tree rebalance will be required after deleting this entry. | |
8159 ** Save the cursor key. */ | |
8160 rc = saveCursorKey(pCur); | |
8161 if( rc ) return rc; | |
8162 }else{ | |
8163 bSkipnext = 1; | |
8164 } | |
8165 } | |
8166 | |
8167 /* Make the page containing the entry to be deleted writable. Then free any | |
8168 ** overflow pages associated with the entry and finally remove the cell | |
8169 ** itself from within the page. */ | |
8170 rc = sqlite3PagerWrite(pPage->pDbPage); | |
8171 if( rc ) return rc; | |
8172 rc = clearCell(pPage, pCell, &szCell); | |
8173 dropCell(pPage, iCellIdx, szCell, &rc); | |
8174 if( rc ) return rc; | |
8175 | |
8176 /* If the cell deleted was not located on a leaf page, then the cursor | |
8177 ** is currently pointing to the largest entry in the sub-tree headed | |
8178 ** by the child-page of the cell that was just deleted from an internal | |
8179 ** node. The cell from the leaf node needs to be moved to the internal | |
8180 ** node to replace the deleted cell. */ | |
8181 if( !pPage->leaf ){ | |
8182 MemPage *pLeaf = pCur->apPage[pCur->iPage]; | |
8183 int nCell; | |
8184 Pgno n = pCur->apPage[iCellDepth+1]->pgno; | |
8185 unsigned char *pTmp; | |
8186 | |
8187 pCell = findCell(pLeaf, pLeaf->nCell-1); | |
8188 if( pCell<&pLeaf->aData[4] ) return SQLITE_CORRUPT_BKPT; | |
8189 nCell = pLeaf->xCellSize(pLeaf, pCell); | |
8190 assert( MX_CELL_SIZE(pBt) >= nCell ); | |
8191 pTmp = pBt->pTmpSpace; | |
8192 assert( pTmp!=0 ); | |
8193 rc = sqlite3PagerWrite(pLeaf->pDbPage); | |
8194 insertCell(pPage, iCellIdx, pCell-4, nCell+4, pTmp, n, &rc); | |
8195 dropCell(pLeaf, pLeaf->nCell-1, nCell, &rc); | |
8196 if( rc ) return rc; | |
8197 } | |
8198 | |
8199 /* Balance the tree. If the entry deleted was located on a leaf page, | |
8200 ** then the cursor still points to that page. In this case the first | |
8201 ** call to balance() repairs the tree, and the if(...) condition is | |
8202 ** never true. | |
8203 ** | |
8204 ** Otherwise, if the entry deleted was on an internal node page, then | |
8205 ** pCur is pointing to the leaf page from which a cell was removed to | |
8206 ** replace the cell deleted from the internal node. This is slightly | |
8207 ** tricky as the leaf node may be underfull, and the internal node may | |
8208 ** be either under or overfull. In this case run the balancing algorithm | |
8209 ** on the leaf node first. If the balance proceeds far enough up the | |
8210 ** tree that we can be sure that any problem in the internal node has | |
8211 ** been corrected, so be it. Otherwise, after balancing the leaf node, | |
8212 ** walk the cursor up the tree to the internal node and balance it as | |
8213 ** well. */ | |
8214 rc = balance(pCur); | |
8215 if( rc==SQLITE_OK && pCur->iPage>iCellDepth ){ | |
8216 while( pCur->iPage>iCellDepth ){ | |
8217 releasePage(pCur->apPage[pCur->iPage--]); | |
8218 } | |
8219 rc = balance(pCur); | |
8220 } | |
8221 | |
8222 if( rc==SQLITE_OK ){ | |
8223 if( bSkipnext ){ | |
8224 assert( bPreserve && (pCur->iPage==iCellDepth || CORRUPT_DB) ); | |
8225 assert( pPage==pCur->apPage[pCur->iPage] ); | |
8226 assert( (pPage->nCell>0 || CORRUPT_DB) && iCellIdx<=pPage->nCell ); | |
8227 pCur->eState = CURSOR_SKIPNEXT; | |
8228 if( iCellIdx>=pPage->nCell ){ | |
8229 pCur->skipNext = -1; | |
8230 pCur->aiIdx[iCellDepth] = pPage->nCell-1; | |
8231 }else{ | |
8232 pCur->skipNext = 1; | |
8233 } | |
8234 }else{ | |
8235 rc = moveToRoot(pCur); | |
8236 if( bPreserve ){ | |
8237 pCur->eState = CURSOR_REQUIRESEEK; | |
8238 } | |
8239 } | |
8240 } | |
8241 return rc; | |
8242 } | |
8243 | |
8244 /* | |
8245 ** Create a new BTree table. Write into *piTable the page | |
8246 ** number for the root page of the new table. | |
8247 ** | |
8248 ** The type of type is determined by the flags parameter. Only the | |
8249 ** following values of flags are currently in use. Other values for | |
8250 ** flags might not work: | |
8251 ** | |
8252 ** BTREE_INTKEY|BTREE_LEAFDATA Used for SQL tables with rowid keys | |
8253 ** BTREE_ZERODATA Used for SQL indices | |
8254 */ | |
8255 static int btreeCreateTable(Btree *p, int *piTable, int createTabFlags){ | |
8256 BtShared *pBt = p->pBt; | |
8257 MemPage *pRoot; | |
8258 Pgno pgnoRoot; | |
8259 int rc; | |
8260 int ptfFlags; /* Page-type flage for the root page of new table */ | |
8261 | |
8262 assert( sqlite3BtreeHoldsMutex(p) ); | |
8263 assert( pBt->inTransaction==TRANS_WRITE ); | |
8264 assert( (pBt->btsFlags & BTS_READ_ONLY)==0 ); | |
8265 | |
8266 #ifdef SQLITE_OMIT_AUTOVACUUM | |
8267 rc = allocateBtreePage(pBt, &pRoot, &pgnoRoot, 1, 0); | |
8268 if( rc ){ | |
8269 return rc; | |
8270 } | |
8271 #else | |
8272 if( pBt->autoVacuum ){ | |
8273 Pgno pgnoMove; /* Move a page here to make room for the root-page */ | |
8274 MemPage *pPageMove; /* The page to move to. */ | |
8275 | |
8276 /* Creating a new table may probably require moving an existing database | |
8277 ** to make room for the new tables root page. In case this page turns | |
8278 ** out to be an overflow page, delete all overflow page-map caches | |
8279 ** held by open cursors. | |
8280 */ | |
8281 invalidateAllOverflowCache(pBt); | |
8282 | |
8283 /* Read the value of meta[3] from the database to determine where the | |
8284 ** root page of the new table should go. meta[3] is the largest root-page | |
8285 ** created so far, so the new root-page is (meta[3]+1). | |
8286 */ | |
8287 sqlite3BtreeGetMeta(p, BTREE_LARGEST_ROOT_PAGE, &pgnoRoot); | |
8288 pgnoRoot++; | |
8289 | |
8290 /* The new root-page may not be allocated on a pointer-map page, or the | |
8291 ** PENDING_BYTE page. | |
8292 */ | |
8293 while( pgnoRoot==PTRMAP_PAGENO(pBt, pgnoRoot) || | |
8294 pgnoRoot==PENDING_BYTE_PAGE(pBt) ){ | |
8295 pgnoRoot++; | |
8296 } | |
8297 assert( pgnoRoot>=3 || CORRUPT_DB ); | |
8298 testcase( pgnoRoot<3 ); | |
8299 | |
8300 /* Allocate a page. The page that currently resides at pgnoRoot will | |
8301 ** be moved to the allocated page (unless the allocated page happens | |
8302 ** to reside at pgnoRoot). | |
8303 */ | |
8304 rc = allocateBtreePage(pBt, &pPageMove, &pgnoMove, pgnoRoot, BTALLOC_EXACT); | |
8305 if( rc!=SQLITE_OK ){ | |
8306 return rc; | |
8307 } | |
8308 | |
8309 if( pgnoMove!=pgnoRoot ){ | |
8310 /* pgnoRoot is the page that will be used for the root-page of | |
8311 ** the new table (assuming an error did not occur). But we were | |
8312 ** allocated pgnoMove. If required (i.e. if it was not allocated | |
8313 ** by extending the file), the current page at position pgnoMove | |
8314 ** is already journaled. | |
8315 */ | |
8316 u8 eType = 0; | |
8317 Pgno iPtrPage = 0; | |
8318 | |
8319 /* Save the positions of any open cursors. This is required in | |
8320 ** case they are holding a reference to an xFetch reference | |
8321 ** corresponding to page pgnoRoot. */ | |
8322 rc = saveAllCursors(pBt, 0, 0); | |
8323 releasePage(pPageMove); | |
8324 if( rc!=SQLITE_OK ){ | |
8325 return rc; | |
8326 } | |
8327 | |
8328 /* Move the page currently at pgnoRoot to pgnoMove. */ | |
8329 rc = btreeGetPage(pBt, pgnoRoot, &pRoot, 0); | |
8330 if( rc!=SQLITE_OK ){ | |
8331 return rc; | |
8332 } | |
8333 rc = ptrmapGet(pBt, pgnoRoot, &eType, &iPtrPage); | |
8334 if( eType==PTRMAP_ROOTPAGE || eType==PTRMAP_FREEPAGE ){ | |
8335 rc = SQLITE_CORRUPT_BKPT; | |
8336 } | |
8337 if( rc!=SQLITE_OK ){ | |
8338 releasePage(pRoot); | |
8339 return rc; | |
8340 } | |
8341 assert( eType!=PTRMAP_ROOTPAGE ); | |
8342 assert( eType!=PTRMAP_FREEPAGE ); | |
8343 rc = relocatePage(pBt, pRoot, eType, iPtrPage, pgnoMove, 0); | |
8344 releasePage(pRoot); | |
8345 | |
8346 /* Obtain the page at pgnoRoot */ | |
8347 if( rc!=SQLITE_OK ){ | |
8348 return rc; | |
8349 } | |
8350 rc = btreeGetPage(pBt, pgnoRoot, &pRoot, 0); | |
8351 if( rc!=SQLITE_OK ){ | |
8352 return rc; | |
8353 } | |
8354 rc = sqlite3PagerWrite(pRoot->pDbPage); | |
8355 if( rc!=SQLITE_OK ){ | |
8356 releasePage(pRoot); | |
8357 return rc; | |
8358 } | |
8359 }else{ | |
8360 pRoot = pPageMove; | |
8361 } | |
8362 | |
8363 /* Update the pointer-map and meta-data with the new root-page number. */ | |
8364 ptrmapPut(pBt, pgnoRoot, PTRMAP_ROOTPAGE, 0, &rc); | |
8365 if( rc ){ | |
8366 releasePage(pRoot); | |
8367 return rc; | |
8368 } | |
8369 | |
8370 /* When the new root page was allocated, page 1 was made writable in | |
8371 ** order either to increase the database filesize, or to decrement the | |
8372 ** freelist count. Hence, the sqlite3BtreeUpdateMeta() call cannot fail. | |
8373 */ | |
8374 assert( sqlite3PagerIswriteable(pBt->pPage1->pDbPage) ); | |
8375 rc = sqlite3BtreeUpdateMeta(p, 4, pgnoRoot); | |
8376 if( NEVER(rc) ){ | |
8377 releasePage(pRoot); | |
8378 return rc; | |
8379 } | |
8380 | |
8381 }else{ | |
8382 rc = allocateBtreePage(pBt, &pRoot, &pgnoRoot, 1, 0); | |
8383 if( rc ) return rc; | |
8384 } | |
8385 #endif | |
8386 assert( sqlite3PagerIswriteable(pRoot->pDbPage) ); | |
8387 if( createTabFlags & BTREE_INTKEY ){ | |
8388 ptfFlags = PTF_INTKEY | PTF_LEAFDATA | PTF_LEAF; | |
8389 }else{ | |
8390 ptfFlags = PTF_ZERODATA | PTF_LEAF; | |
8391 } | |
8392 zeroPage(pRoot, ptfFlags); | |
8393 sqlite3PagerUnref(pRoot->pDbPage); | |
8394 assert( (pBt->openFlags & BTREE_SINGLE)==0 || pgnoRoot==2 ); | |
8395 *piTable = (int)pgnoRoot; | |
8396 return SQLITE_OK; | |
8397 } | |
8398 int sqlite3BtreeCreateTable(Btree *p, int *piTable, int flags){ | |
8399 int rc; | |
8400 sqlite3BtreeEnter(p); | |
8401 rc = btreeCreateTable(p, piTable, flags); | |
8402 sqlite3BtreeLeave(p); | |
8403 return rc; | |
8404 } | |
8405 | |
8406 /* | |
8407 ** Erase the given database page and all its children. Return | |
8408 ** the page to the freelist. | |
8409 */ | |
8410 static int clearDatabasePage( | |
8411 BtShared *pBt, /* The BTree that contains the table */ | |
8412 Pgno pgno, /* Page number to clear */ | |
8413 int freePageFlag, /* Deallocate page if true */ | |
8414 int *pnChange /* Add number of Cells freed to this counter */ | |
8415 ){ | |
8416 MemPage *pPage; | |
8417 int rc; | |
8418 unsigned char *pCell; | |
8419 int i; | |
8420 int hdr; | |
8421 u16 szCell; | |
8422 | |
8423 assert( sqlite3_mutex_held(pBt->mutex) ); | |
8424 if( pgno>btreePagecount(pBt) ){ | |
8425 return SQLITE_CORRUPT_BKPT; | |
8426 } | |
8427 rc = getAndInitPage(pBt, pgno, &pPage, 0, 0); | |
8428 if( rc ) return rc; | |
8429 if( pPage->bBusy ){ | |
8430 rc = SQLITE_CORRUPT_BKPT; | |
8431 goto cleardatabasepage_out; | |
8432 } | |
8433 pPage->bBusy = 1; | |
8434 hdr = pPage->hdrOffset; | |
8435 for(i=0; i<pPage->nCell; i++){ | |
8436 pCell = findCell(pPage, i); | |
8437 if( !pPage->leaf ){ | |
8438 rc = clearDatabasePage(pBt, get4byte(pCell), 1, pnChange); | |
8439 if( rc ) goto cleardatabasepage_out; | |
8440 } | |
8441 rc = clearCell(pPage, pCell, &szCell); | |
8442 if( rc ) goto cleardatabasepage_out; | |
8443 } | |
8444 if( !pPage->leaf ){ | |
8445 rc = clearDatabasePage(pBt, get4byte(&pPage->aData[hdr+8]), 1, pnChange); | |
8446 if( rc ) goto cleardatabasepage_out; | |
8447 }else if( pnChange ){ | |
8448 assert( pPage->intKey || CORRUPT_DB ); | |
8449 testcase( !pPage->intKey ); | |
8450 *pnChange += pPage->nCell; | |
8451 } | |
8452 if( freePageFlag ){ | |
8453 freePage(pPage, &rc); | |
8454 }else if( (rc = sqlite3PagerWrite(pPage->pDbPage))==0 ){ | |
8455 zeroPage(pPage, pPage->aData[hdr] | PTF_LEAF); | |
8456 } | |
8457 | |
8458 cleardatabasepage_out: | |
8459 pPage->bBusy = 0; | |
8460 releasePage(pPage); | |
8461 return rc; | |
8462 } | |
8463 | |
8464 /* | |
8465 ** Delete all information from a single table in the database. iTable is | |
8466 ** the page number of the root of the table. After this routine returns, | |
8467 ** the root page is empty, but still exists. | |
8468 ** | |
8469 ** This routine will fail with SQLITE_LOCKED if there are any open | |
8470 ** read cursors on the table. Open write cursors are moved to the | |
8471 ** root of the table. | |
8472 ** | |
8473 ** If pnChange is not NULL, then table iTable must be an intkey table. The | |
8474 ** integer value pointed to by pnChange is incremented by the number of | |
8475 ** entries in the table. | |
8476 */ | |
8477 int sqlite3BtreeClearTable(Btree *p, int iTable, int *pnChange){ | |
8478 int rc; | |
8479 BtShared *pBt = p->pBt; | |
8480 sqlite3BtreeEnter(p); | |
8481 assert( p->inTrans==TRANS_WRITE ); | |
8482 | |
8483 rc = saveAllCursors(pBt, (Pgno)iTable, 0); | |
8484 | |
8485 if( SQLITE_OK==rc ){ | |
8486 /* Invalidate all incrblob cursors open on table iTable (assuming iTable | |
8487 ** is the root of a table b-tree - if it is not, the following call is | |
8488 ** a no-op). */ | |
8489 invalidateIncrblobCursors(p, 0, 1); | |
8490 rc = clearDatabasePage(pBt, (Pgno)iTable, 0, pnChange); | |
8491 } | |
8492 sqlite3BtreeLeave(p); | |
8493 return rc; | |
8494 } | |
8495 | |
8496 /* | |
8497 ** Delete all information from the single table that pCur is open on. | |
8498 ** | |
8499 ** This routine only work for pCur on an ephemeral table. | |
8500 */ | |
8501 int sqlite3BtreeClearTableOfCursor(BtCursor *pCur){ | |
8502 return sqlite3BtreeClearTable(pCur->pBtree, pCur->pgnoRoot, 0); | |
8503 } | |
8504 | |
8505 /* | |
8506 ** Erase all information in a table and add the root of the table to | |
8507 ** the freelist. Except, the root of the principle table (the one on | |
8508 ** page 1) is never added to the freelist. | |
8509 ** | |
8510 ** This routine will fail with SQLITE_LOCKED if there are any open | |
8511 ** cursors on the table. | |
8512 ** | |
8513 ** If AUTOVACUUM is enabled and the page at iTable is not the last | |
8514 ** root page in the database file, then the last root page | |
8515 ** in the database file is moved into the slot formerly occupied by | |
8516 ** iTable and that last slot formerly occupied by the last root page | |
8517 ** is added to the freelist instead of iTable. In this say, all | |
8518 ** root pages are kept at the beginning of the database file, which | |
8519 ** is necessary for AUTOVACUUM to work right. *piMoved is set to the | |
8520 ** page number that used to be the last root page in the file before | |
8521 ** the move. If no page gets moved, *piMoved is set to 0. | |
8522 ** The last root page is recorded in meta[3] and the value of | |
8523 ** meta[3] is updated by this procedure. | |
8524 */ | |
8525 static int btreeDropTable(Btree *p, Pgno iTable, int *piMoved){ | |
8526 int rc; | |
8527 MemPage *pPage = 0; | |
8528 BtShared *pBt = p->pBt; | |
8529 | |
8530 assert( sqlite3BtreeHoldsMutex(p) ); | |
8531 assert( p->inTrans==TRANS_WRITE ); | |
8532 | |
8533 /* It is illegal to drop a table if any cursors are open on the | |
8534 ** database. This is because in auto-vacuum mode the backend may | |
8535 ** need to move another root-page to fill a gap left by the deleted | |
8536 ** root page. If an open cursor was using this page a problem would | |
8537 ** occur. | |
8538 ** | |
8539 ** This error is caught long before control reaches this point. | |
8540 */ | |
8541 if( NEVER(pBt->pCursor) ){ | |
8542 sqlite3ConnectionBlocked(p->db, pBt->pCursor->pBtree->db); | |
8543 return SQLITE_LOCKED_SHAREDCACHE; | |
8544 } | |
8545 | |
8546 rc = btreeGetPage(pBt, (Pgno)iTable, &pPage, 0); | |
8547 if( rc ) return rc; | |
8548 rc = sqlite3BtreeClearTable(p, iTable, 0); | |
8549 if( rc ){ | |
8550 releasePage(pPage); | |
8551 return rc; | |
8552 } | |
8553 | |
8554 *piMoved = 0; | |
8555 | |
8556 if( iTable>1 ){ | |
8557 #ifdef SQLITE_OMIT_AUTOVACUUM | |
8558 freePage(pPage, &rc); | |
8559 releasePage(pPage); | |
8560 #else | |
8561 if( pBt->autoVacuum ){ | |
8562 Pgno maxRootPgno; | |
8563 sqlite3BtreeGetMeta(p, BTREE_LARGEST_ROOT_PAGE, &maxRootPgno); | |
8564 | |
8565 if( iTable==maxRootPgno ){ | |
8566 /* If the table being dropped is the table with the largest root-page | |
8567 ** number in the database, put the root page on the free list. | |
8568 */ | |
8569 freePage(pPage, &rc); | |
8570 releasePage(pPage); | |
8571 if( rc!=SQLITE_OK ){ | |
8572 return rc; | |
8573 } | |
8574 }else{ | |
8575 /* The table being dropped does not have the largest root-page | |
8576 ** number in the database. So move the page that does into the | |
8577 ** gap left by the deleted root-page. | |
8578 */ | |
8579 MemPage *pMove; | |
8580 releasePage(pPage); | |
8581 rc = btreeGetPage(pBt, maxRootPgno, &pMove, 0); | |
8582 if( rc!=SQLITE_OK ){ | |
8583 return rc; | |
8584 } | |
8585 rc = relocatePage(pBt, pMove, PTRMAP_ROOTPAGE, 0, iTable, 0); | |
8586 releasePage(pMove); | |
8587 if( rc!=SQLITE_OK ){ | |
8588 return rc; | |
8589 } | |
8590 pMove = 0; | |
8591 rc = btreeGetPage(pBt, maxRootPgno, &pMove, 0); | |
8592 freePage(pMove, &rc); | |
8593 releasePage(pMove); | |
8594 if( rc!=SQLITE_OK ){ | |
8595 return rc; | |
8596 } | |
8597 *piMoved = maxRootPgno; | |
8598 } | |
8599 | |
8600 /* Set the new 'max-root-page' value in the database header. This | |
8601 ** is the old value less one, less one more if that happens to | |
8602 ** be a root-page number, less one again if that is the | |
8603 ** PENDING_BYTE_PAGE. | |
8604 */ | |
8605 maxRootPgno--; | |
8606 while( maxRootPgno==PENDING_BYTE_PAGE(pBt) | |
8607 || PTRMAP_ISPAGE(pBt, maxRootPgno) ){ | |
8608 maxRootPgno--; | |
8609 } | |
8610 assert( maxRootPgno!=PENDING_BYTE_PAGE(pBt) ); | |
8611 | |
8612 rc = sqlite3BtreeUpdateMeta(p, 4, maxRootPgno); | |
8613 }else{ | |
8614 freePage(pPage, &rc); | |
8615 releasePage(pPage); | |
8616 } | |
8617 #endif | |
8618 }else{ | |
8619 /* If sqlite3BtreeDropTable was called on page 1. | |
8620 ** This really never should happen except in a corrupt | |
8621 ** database. | |
8622 */ | |
8623 zeroPage(pPage, PTF_INTKEY|PTF_LEAF ); | |
8624 releasePage(pPage); | |
8625 } | |
8626 return rc; | |
8627 } | |
8628 int sqlite3BtreeDropTable(Btree *p, int iTable, int *piMoved){ | |
8629 int rc; | |
8630 sqlite3BtreeEnter(p); | |
8631 rc = btreeDropTable(p, iTable, piMoved); | |
8632 sqlite3BtreeLeave(p); | |
8633 return rc; | |
8634 } | |
8635 | |
8636 | |
8637 /* | |
8638 ** This function may only be called if the b-tree connection already | |
8639 ** has a read or write transaction open on the database. | |
8640 ** | |
8641 ** Read the meta-information out of a database file. Meta[0] | |
8642 ** is the number of free pages currently in the database. Meta[1] | |
8643 ** through meta[15] are available for use by higher layers. Meta[0] | |
8644 ** is read-only, the others are read/write. | |
8645 ** | |
8646 ** The schema layer numbers meta values differently. At the schema | |
8647 ** layer (and the SetCookie and ReadCookie opcodes) the number of | |
8648 ** free pages is not visible. So Cookie[0] is the same as Meta[1]. | |
8649 ** | |
8650 ** This routine treats Meta[BTREE_DATA_VERSION] as a special case. Instead | |
8651 ** of reading the value out of the header, it instead loads the "DataVersion" | |
8652 ** from the pager. The BTREE_DATA_VERSION value is not actually stored in the | |
8653 ** database file. It is a number computed by the pager. But its access | |
8654 ** pattern is the same as header meta values, and so it is convenient to | |
8655 ** read it from this routine. | |
8656 */ | |
8657 void sqlite3BtreeGetMeta(Btree *p, int idx, u32 *pMeta){ | |
8658 BtShared *pBt = p->pBt; | |
8659 | |
8660 sqlite3BtreeEnter(p); | |
8661 assert( p->inTrans>TRANS_NONE ); | |
8662 assert( SQLITE_OK==querySharedCacheTableLock(p, MASTER_ROOT, READ_LOCK) ); | |
8663 assert( pBt->pPage1 ); | |
8664 assert( idx>=0 && idx<=15 ); | |
8665 | |
8666 if( idx==BTREE_DATA_VERSION ){ | |
8667 *pMeta = sqlite3PagerDataVersion(pBt->pPager) + p->iDataVersion; | |
8668 }else{ | |
8669 *pMeta = get4byte(&pBt->pPage1->aData[36 + idx*4]); | |
8670 } | |
8671 | |
8672 /* If auto-vacuum is disabled in this build and this is an auto-vacuum | |
8673 ** database, mark the database as read-only. */ | |
8674 #ifdef SQLITE_OMIT_AUTOVACUUM | |
8675 if( idx==BTREE_LARGEST_ROOT_PAGE && *pMeta>0 ){ | |
8676 pBt->btsFlags |= BTS_READ_ONLY; | |
8677 } | |
8678 #endif | |
8679 | |
8680 sqlite3BtreeLeave(p); | |
8681 } | |
8682 | |
8683 /* | |
8684 ** Write meta-information back into the database. Meta[0] is | |
8685 ** read-only and may not be written. | |
8686 */ | |
8687 int sqlite3BtreeUpdateMeta(Btree *p, int idx, u32 iMeta){ | |
8688 BtShared *pBt = p->pBt; | |
8689 unsigned char *pP1; | |
8690 int rc; | |
8691 assert( idx>=1 && idx<=15 ); | |
8692 sqlite3BtreeEnter(p); | |
8693 assert( p->inTrans==TRANS_WRITE ); | |
8694 assert( pBt->pPage1!=0 ); | |
8695 pP1 = pBt->pPage1->aData; | |
8696 rc = sqlite3PagerWrite(pBt->pPage1->pDbPage); | |
8697 if( rc==SQLITE_OK ){ | |
8698 put4byte(&pP1[36 + idx*4], iMeta); | |
8699 #ifndef SQLITE_OMIT_AUTOVACUUM | |
8700 if( idx==BTREE_INCR_VACUUM ){ | |
8701 assert( pBt->autoVacuum || iMeta==0 ); | |
8702 assert( iMeta==0 || iMeta==1 ); | |
8703 pBt->incrVacuum = (u8)iMeta; | |
8704 } | |
8705 #endif | |
8706 } | |
8707 sqlite3BtreeLeave(p); | |
8708 return rc; | |
8709 } | |
8710 | |
8711 #ifndef SQLITE_OMIT_BTREECOUNT | |
8712 /* | |
8713 ** The first argument, pCur, is a cursor opened on some b-tree. Count the | |
8714 ** number of entries in the b-tree and write the result to *pnEntry. | |
8715 ** | |
8716 ** SQLITE_OK is returned if the operation is successfully executed. | |
8717 ** Otherwise, if an error is encountered (i.e. an IO error or database | |
8718 ** corruption) an SQLite error code is returned. | |
8719 */ | |
8720 int sqlite3BtreeCount(BtCursor *pCur, i64 *pnEntry){ | |
8721 i64 nEntry = 0; /* Value to return in *pnEntry */ | |
8722 int rc; /* Return code */ | |
8723 | |
8724 if( pCur->pgnoRoot==0 ){ | |
8725 *pnEntry = 0; | |
8726 return SQLITE_OK; | |
8727 } | |
8728 rc = moveToRoot(pCur); | |
8729 | |
8730 /* Unless an error occurs, the following loop runs one iteration for each | |
8731 ** page in the B-Tree structure (not including overflow pages). | |
8732 */ | |
8733 while( rc==SQLITE_OK ){ | |
8734 int iIdx; /* Index of child node in parent */ | |
8735 MemPage *pPage; /* Current page of the b-tree */ | |
8736 | |
8737 /* If this is a leaf page or the tree is not an int-key tree, then | |
8738 ** this page contains countable entries. Increment the entry counter | |
8739 ** accordingly. | |
8740 */ | |
8741 pPage = pCur->apPage[pCur->iPage]; | |
8742 if( pPage->leaf || !pPage->intKey ){ | |
8743 nEntry += pPage->nCell; | |
8744 } | |
8745 | |
8746 /* pPage is a leaf node. This loop navigates the cursor so that it | |
8747 ** points to the first interior cell that it points to the parent of | |
8748 ** the next page in the tree that has not yet been visited. The | |
8749 ** pCur->aiIdx[pCur->iPage] value is set to the index of the parent cell | |
8750 ** of the page, or to the number of cells in the page if the next page | |
8751 ** to visit is the right-child of its parent. | |
8752 ** | |
8753 ** If all pages in the tree have been visited, return SQLITE_OK to the | |
8754 ** caller. | |
8755 */ | |
8756 if( pPage->leaf ){ | |
8757 do { | |
8758 if( pCur->iPage==0 ){ | |
8759 /* All pages of the b-tree have been visited. Return successfully. */ | |
8760 *pnEntry = nEntry; | |
8761 return moveToRoot(pCur); | |
8762 } | |
8763 moveToParent(pCur); | |
8764 }while ( pCur->aiIdx[pCur->iPage]>=pCur->apPage[pCur->iPage]->nCell ); | |
8765 | |
8766 pCur->aiIdx[pCur->iPage]++; | |
8767 pPage = pCur->apPage[pCur->iPage]; | |
8768 } | |
8769 | |
8770 /* Descend to the child node of the cell that the cursor currently | |
8771 ** points at. This is the right-child if (iIdx==pPage->nCell). | |
8772 */ | |
8773 iIdx = pCur->aiIdx[pCur->iPage]; | |
8774 if( iIdx==pPage->nCell ){ | |
8775 rc = moveToChild(pCur, get4byte(&pPage->aData[pPage->hdrOffset+8])); | |
8776 }else{ | |
8777 rc = moveToChild(pCur, get4byte(findCell(pPage, iIdx))); | |
8778 } | |
8779 } | |
8780 | |
8781 /* An error has occurred. Return an error code. */ | |
8782 return rc; | |
8783 } | |
8784 #endif | |
8785 | |
8786 /* | |
8787 ** Return the pager associated with a BTree. This routine is used for | |
8788 ** testing and debugging only. | |
8789 */ | |
8790 Pager *sqlite3BtreePager(Btree *p){ | |
8791 return p->pBt->pPager; | |
8792 } | |
8793 | |
8794 #ifndef SQLITE_OMIT_INTEGRITY_CHECK | |
8795 /* | |
8796 ** Append a message to the error message string. | |
8797 */ | |
8798 static void checkAppendMsg( | |
8799 IntegrityCk *pCheck, | |
8800 const char *zFormat, | |
8801 ... | |
8802 ){ | |
8803 va_list ap; | |
8804 if( !pCheck->mxErr ) return; | |
8805 pCheck->mxErr--; | |
8806 pCheck->nErr++; | |
8807 va_start(ap, zFormat); | |
8808 if( pCheck->errMsg.nChar ){ | |
8809 sqlite3StrAccumAppend(&pCheck->errMsg, "\n", 1); | |
8810 } | |
8811 if( pCheck->zPfx ){ | |
8812 sqlite3XPrintf(&pCheck->errMsg, 0, pCheck->zPfx, pCheck->v1, pCheck->v2); | |
8813 } | |
8814 sqlite3VXPrintf(&pCheck->errMsg, 1, zFormat, ap); | |
8815 va_end(ap); | |
8816 if( pCheck->errMsg.accError==STRACCUM_NOMEM ){ | |
8817 pCheck->mallocFailed = 1; | |
8818 } | |
8819 } | |
8820 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */ | |
8821 | |
8822 #ifndef SQLITE_OMIT_INTEGRITY_CHECK | |
8823 | |
8824 /* | |
8825 ** Return non-zero if the bit in the IntegrityCk.aPgRef[] array that | |
8826 ** corresponds to page iPg is already set. | |
8827 */ | |
8828 static int getPageReferenced(IntegrityCk *pCheck, Pgno iPg){ | |
8829 assert( iPg<=pCheck->nPage && sizeof(pCheck->aPgRef[0])==1 ); | |
8830 return (pCheck->aPgRef[iPg/8] & (1 << (iPg & 0x07))); | |
8831 } | |
8832 | |
8833 /* | |
8834 ** Set the bit in the IntegrityCk.aPgRef[] array that corresponds to page iPg. | |
8835 */ | |
8836 static void setPageReferenced(IntegrityCk *pCheck, Pgno iPg){ | |
8837 assert( iPg<=pCheck->nPage && sizeof(pCheck->aPgRef[0])==1 ); | |
8838 pCheck->aPgRef[iPg/8] |= (1 << (iPg & 0x07)); | |
8839 } | |
8840 | |
8841 | |
8842 /* | |
8843 ** Add 1 to the reference count for page iPage. If this is the second | |
8844 ** reference to the page, add an error message to pCheck->zErrMsg. | |
8845 ** Return 1 if there are 2 or more references to the page and 0 if | |
8846 ** if this is the first reference to the page. | |
8847 ** | |
8848 ** Also check that the page number is in bounds. | |
8849 */ | |
8850 static int checkRef(IntegrityCk *pCheck, Pgno iPage){ | |
8851 if( iPage==0 ) return 1; | |
8852 if( iPage>pCheck->nPage ){ | |
8853 checkAppendMsg(pCheck, "invalid page number %d", iPage); | |
8854 return 1; | |
8855 } | |
8856 if( getPageReferenced(pCheck, iPage) ){ | |
8857 checkAppendMsg(pCheck, "2nd reference to page %d", iPage); | |
8858 return 1; | |
8859 } | |
8860 setPageReferenced(pCheck, iPage); | |
8861 return 0; | |
8862 } | |
8863 | |
8864 #ifndef SQLITE_OMIT_AUTOVACUUM | |
8865 /* | |
8866 ** Check that the entry in the pointer-map for page iChild maps to | |
8867 ** page iParent, pointer type ptrType. If not, append an error message | |
8868 ** to pCheck. | |
8869 */ | |
8870 static void checkPtrmap( | |
8871 IntegrityCk *pCheck, /* Integrity check context */ | |
8872 Pgno iChild, /* Child page number */ | |
8873 u8 eType, /* Expected pointer map type */ | |
8874 Pgno iParent /* Expected pointer map parent page number */ | |
8875 ){ | |
8876 int rc; | |
8877 u8 ePtrmapType; | |
8878 Pgno iPtrmapParent; | |
8879 | |
8880 rc = ptrmapGet(pCheck->pBt, iChild, &ePtrmapType, &iPtrmapParent); | |
8881 if( rc!=SQLITE_OK ){ | |
8882 if( rc==SQLITE_NOMEM || rc==SQLITE_IOERR_NOMEM ) pCheck->mallocFailed = 1; | |
8883 checkAppendMsg(pCheck, "Failed to read ptrmap key=%d", iChild); | |
8884 return; | |
8885 } | |
8886 | |
8887 if( ePtrmapType!=eType || iPtrmapParent!=iParent ){ | |
8888 checkAppendMsg(pCheck, | |
8889 "Bad ptr map entry key=%d expected=(%d,%d) got=(%d,%d)", | |
8890 iChild, eType, iParent, ePtrmapType, iPtrmapParent); | |
8891 } | |
8892 } | |
8893 #endif | |
8894 | |
8895 /* | |
8896 ** Check the integrity of the freelist or of an overflow page list. | |
8897 ** Verify that the number of pages on the list is N. | |
8898 */ | |
8899 static void checkList( | |
8900 IntegrityCk *pCheck, /* Integrity checking context */ | |
8901 int isFreeList, /* True for a freelist. False for overflow page list */ | |
8902 int iPage, /* Page number for first page in the list */ | |
8903 int N /* Expected number of pages in the list */ | |
8904 ){ | |
8905 int i; | |
8906 int expected = N; | |
8907 int iFirst = iPage; | |
8908 while( N-- > 0 && pCheck->mxErr ){ | |
8909 DbPage *pOvflPage; | |
8910 unsigned char *pOvflData; | |
8911 if( iPage<1 ){ | |
8912 checkAppendMsg(pCheck, | |
8913 "%d of %d pages missing from overflow list starting at %d", | |
8914 N+1, expected, iFirst); | |
8915 break; | |
8916 } | |
8917 if( checkRef(pCheck, iPage) ) break; | |
8918 if( sqlite3PagerGet(pCheck->pPager, (Pgno)iPage, &pOvflPage, 0) ){ | |
8919 checkAppendMsg(pCheck, "failed to get page %d", iPage); | |
8920 break; | |
8921 } | |
8922 pOvflData = (unsigned char *)sqlite3PagerGetData(pOvflPage); | |
8923 if( isFreeList ){ | |
8924 int n = get4byte(&pOvflData[4]); | |
8925 #ifndef SQLITE_OMIT_AUTOVACUUM | |
8926 if( pCheck->pBt->autoVacuum ){ | |
8927 checkPtrmap(pCheck, iPage, PTRMAP_FREEPAGE, 0); | |
8928 } | |
8929 #endif | |
8930 if( n>(int)pCheck->pBt->usableSize/4-2 ){ | |
8931 checkAppendMsg(pCheck, | |
8932 "freelist leaf count too big on page %d", iPage); | |
8933 N--; | |
8934 }else{ | |
8935 for(i=0; i<n; i++){ | |
8936 Pgno iFreePage = get4byte(&pOvflData[8+i*4]); | |
8937 #ifndef SQLITE_OMIT_AUTOVACUUM | |
8938 if( pCheck->pBt->autoVacuum ){ | |
8939 checkPtrmap(pCheck, iFreePage, PTRMAP_FREEPAGE, 0); | |
8940 } | |
8941 #endif | |
8942 checkRef(pCheck, iFreePage); | |
8943 } | |
8944 N -= n; | |
8945 } | |
8946 } | |
8947 #ifndef SQLITE_OMIT_AUTOVACUUM | |
8948 else{ | |
8949 /* If this database supports auto-vacuum and iPage is not the last | |
8950 ** page in this overflow list, check that the pointer-map entry for | |
8951 ** the following page matches iPage. | |
8952 */ | |
8953 if( pCheck->pBt->autoVacuum && N>0 ){ | |
8954 i = get4byte(pOvflData); | |
8955 checkPtrmap(pCheck, i, PTRMAP_OVERFLOW2, iPage); | |
8956 } | |
8957 } | |
8958 #endif | |
8959 iPage = get4byte(pOvflData); | |
8960 sqlite3PagerUnref(pOvflPage); | |
8961 | |
8962 if( isFreeList && N<(iPage!=0) ){ | |
8963 checkAppendMsg(pCheck, "free-page count in header is too small"); | |
8964 } | |
8965 } | |
8966 } | |
8967 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */ | |
8968 | |
8969 /* | |
8970 ** An implementation of a min-heap. | |
8971 ** | |
8972 ** aHeap[0] is the number of elements on the heap. aHeap[1] is the | |
8973 ** root element. The daughter nodes of aHeap[N] are aHeap[N*2] | |
8974 ** and aHeap[N*2+1]. | |
8975 ** | |
8976 ** The heap property is this: Every node is less than or equal to both | |
8977 ** of its daughter nodes. A consequence of the heap property is that the | |
8978 ** root node aHeap[1] is always the minimum value currently in the heap. | |
8979 ** | |
8980 ** The btreeHeapInsert() routine inserts an unsigned 32-bit number onto | |
8981 ** the heap, preserving the heap property. The btreeHeapPull() routine | |
8982 ** removes the root element from the heap (the minimum value in the heap) | |
8983 ** and then moves other nodes around as necessary to preserve the heap | |
8984 ** property. | |
8985 ** | |
8986 ** This heap is used for cell overlap and coverage testing. Each u32 | |
8987 ** entry represents the span of a cell or freeblock on a btree page. | |
8988 ** The upper 16 bits are the index of the first byte of a range and the | |
8989 ** lower 16 bits are the index of the last byte of that range. | |
8990 */ | |
8991 static void btreeHeapInsert(u32 *aHeap, u32 x){ | |
8992 u32 j, i = ++aHeap[0]; | |
8993 aHeap[i] = x; | |
8994 while( (j = i/2)>0 && aHeap[j]>aHeap[i] ){ | |
8995 x = aHeap[j]; | |
8996 aHeap[j] = aHeap[i]; | |
8997 aHeap[i] = x; | |
8998 i = j; | |
8999 } | |
9000 } | |
9001 static int btreeHeapPull(u32 *aHeap, u32 *pOut){ | |
9002 u32 j, i, x; | |
9003 if( (x = aHeap[0])==0 ) return 0; | |
9004 *pOut = aHeap[1]; | |
9005 aHeap[1] = aHeap[x]; | |
9006 aHeap[x] = 0xffffffff; | |
9007 aHeap[0]--; | |
9008 i = 1; | |
9009 while( (j = i*2)<=aHeap[0] ){ | |
9010 if( aHeap[j]>aHeap[j+1] ) j++; | |
9011 if( aHeap[i]<aHeap[j] ) break; | |
9012 x = aHeap[i]; | |
9013 aHeap[i] = aHeap[j]; | |
9014 aHeap[j] = x; | |
9015 i = j; | |
9016 } | |
9017 return 1; | |
9018 } | |
9019 | |
9020 #ifndef SQLITE_OMIT_INTEGRITY_CHECK | |
9021 /* | |
9022 ** Do various sanity checks on a single page of a tree. Return | |
9023 ** the tree depth. Root pages return 0. Parents of root pages | |
9024 ** return 1, and so forth. | |
9025 ** | |
9026 ** These checks are done: | |
9027 ** | |
9028 ** 1. Make sure that cells and freeblocks do not overlap | |
9029 ** but combine to completely cover the page. | |
9030 ** 2. Make sure integer cell keys are in order. | |
9031 ** 3. Check the integrity of overflow pages. | |
9032 ** 4. Recursively call checkTreePage on all children. | |
9033 ** 5. Verify that the depth of all children is the same. | |
9034 */ | |
9035 static int checkTreePage( | |
9036 IntegrityCk *pCheck, /* Context for the sanity check */ | |
9037 int iPage, /* Page number of the page to check */ | |
9038 i64 *piMinKey, /* Write minimum integer primary key here */ | |
9039 i64 maxKey /* Error if integer primary key greater than this */ | |
9040 ){ | |
9041 MemPage *pPage = 0; /* The page being analyzed */ | |
9042 int i; /* Loop counter */ | |
9043 int rc; /* Result code from subroutine call */ | |
9044 int depth = -1, d2; /* Depth of a subtree */ | |
9045 int pgno; /* Page number */ | |
9046 int nFrag; /* Number of fragmented bytes on the page */ | |
9047 int hdr; /* Offset to the page header */ | |
9048 int cellStart; /* Offset to the start of the cell pointer array */ | |
9049 int nCell; /* Number of cells */ | |
9050 int doCoverageCheck = 1; /* True if cell coverage checking should be done */ | |
9051 int keyCanBeEqual = 1; /* True if IPK can be equal to maxKey | |
9052 ** False if IPK must be strictly less than maxKey */ | |
9053 u8 *data; /* Page content */ | |
9054 u8 *pCell; /* Cell content */ | |
9055 u8 *pCellIdx; /* Next element of the cell pointer array */ | |
9056 BtShared *pBt; /* The BtShared object that owns pPage */ | |
9057 u32 pc; /* Address of a cell */ | |
9058 u32 usableSize; /* Usable size of the page */ | |
9059 u32 contentOffset; /* Offset to the start of the cell content area */ | |
9060 u32 *heap = 0; /* Min-heap used for checking cell coverage */ | |
9061 u32 x, prev = 0; /* Next and previous entry on the min-heap */ | |
9062 const char *saved_zPfx = pCheck->zPfx; | |
9063 int saved_v1 = pCheck->v1; | |
9064 int saved_v2 = pCheck->v2; | |
9065 u8 savedIsInit = 0; | |
9066 | |
9067 /* Check that the page exists | |
9068 */ | |
9069 pBt = pCheck->pBt; | |
9070 usableSize = pBt->usableSize; | |
9071 if( iPage==0 ) return 0; | |
9072 if( checkRef(pCheck, iPage) ) return 0; | |
9073 pCheck->zPfx = "Page %d: "; | |
9074 pCheck->v1 = iPage; | |
9075 if( (rc = btreeGetPage(pBt, (Pgno)iPage, &pPage, 0))!=0 ){ | |
9076 checkAppendMsg(pCheck, | |
9077 "unable to get the page. error code=%d", rc); | |
9078 goto end_of_check; | |
9079 } | |
9080 | |
9081 /* Clear MemPage.isInit to make sure the corruption detection code in | |
9082 ** btreeInitPage() is executed. */ | |
9083 savedIsInit = pPage->isInit; | |
9084 pPage->isInit = 0; | |
9085 if( (rc = btreeInitPage(pPage))!=0 ){ | |
9086 assert( rc==SQLITE_CORRUPT ); /* The only possible error from InitPage */ | |
9087 checkAppendMsg(pCheck, | |
9088 "btreeInitPage() returns error code %d", rc); | |
9089 goto end_of_check; | |
9090 } | |
9091 data = pPage->aData; | |
9092 hdr = pPage->hdrOffset; | |
9093 | |
9094 /* Set up for cell analysis */ | |
9095 pCheck->zPfx = "On tree page %d cell %d: "; | |
9096 contentOffset = get2byteNotZero(&data[hdr+5]); | |
9097 assert( contentOffset<=usableSize ); /* Enforced by btreeInitPage() */ | |
9098 | |
9099 /* EVIDENCE-OF: R-37002-32774 The two-byte integer at offset 3 gives the | |
9100 ** number of cells on the page. */ | |
9101 nCell = get2byte(&data[hdr+3]); | |
9102 assert( pPage->nCell==nCell ); | |
9103 | |
9104 /* EVIDENCE-OF: R-23882-45353 The cell pointer array of a b-tree page | |
9105 ** immediately follows the b-tree page header. */ | |
9106 cellStart = hdr + 12 - 4*pPage->leaf; | |
9107 assert( pPage->aCellIdx==&data[cellStart] ); | |
9108 pCellIdx = &data[cellStart + 2*(nCell-1)]; | |
9109 | |
9110 if( !pPage->leaf ){ | |
9111 /* Analyze the right-child page of internal pages */ | |
9112 pgno = get4byte(&data[hdr+8]); | |
9113 #ifndef SQLITE_OMIT_AUTOVACUUM | |
9114 if( pBt->autoVacuum ){ | |
9115 pCheck->zPfx = "On page %d at right child: "; | |
9116 checkPtrmap(pCheck, pgno, PTRMAP_BTREE, iPage); | |
9117 } | |
9118 #endif | |
9119 depth = checkTreePage(pCheck, pgno, &maxKey, maxKey); | |
9120 keyCanBeEqual = 0; | |
9121 }else{ | |
9122 /* For leaf pages, the coverage check will occur in the same loop | |
9123 ** as the other cell checks, so initialize the heap. */ | |
9124 heap = pCheck->heap; | |
9125 heap[0] = 0; | |
9126 } | |
9127 | |
9128 /* EVIDENCE-OF: R-02776-14802 The cell pointer array consists of K 2-byte | |
9129 ** integer offsets to the cell contents. */ | |
9130 for(i=nCell-1; i>=0 && pCheck->mxErr; i--){ | |
9131 CellInfo info; | |
9132 | |
9133 /* Check cell size */ | |
9134 pCheck->v2 = i; | |
9135 assert( pCellIdx==&data[cellStart + i*2] ); | |
9136 pc = get2byteAligned(pCellIdx); | |
9137 pCellIdx -= 2; | |
9138 if( pc<contentOffset || pc>usableSize-4 ){ | |
9139 checkAppendMsg(pCheck, "Offset %d out of range %d..%d", | |
9140 pc, contentOffset, usableSize-4); | |
9141 doCoverageCheck = 0; | |
9142 continue; | |
9143 } | |
9144 pCell = &data[pc]; | |
9145 pPage->xParseCell(pPage, pCell, &info); | |
9146 if( pc+info.nSize>usableSize ){ | |
9147 checkAppendMsg(pCheck, "Extends off end of page"); | |
9148 doCoverageCheck = 0; | |
9149 continue; | |
9150 } | |
9151 | |
9152 /* Check for integer primary key out of range */ | |
9153 if( pPage->intKey ){ | |
9154 if( keyCanBeEqual ? (info.nKey > maxKey) : (info.nKey >= maxKey) ){ | |
9155 checkAppendMsg(pCheck, "Rowid %lld out of order", info.nKey); | |
9156 } | |
9157 maxKey = info.nKey; | |
9158 } | |
9159 | |
9160 /* Check the content overflow list */ | |
9161 if( info.nPayload>info.nLocal ){ | |
9162 int nPage; /* Number of pages on the overflow chain */ | |
9163 Pgno pgnoOvfl; /* First page of the overflow chain */ | |
9164 assert( pc + info.nSize - 4 <= usableSize ); | |
9165 nPage = (info.nPayload - info.nLocal + usableSize - 5)/(usableSize - 4); | |
9166 pgnoOvfl = get4byte(&pCell[info.nSize - 4]); | |
9167 #ifndef SQLITE_OMIT_AUTOVACUUM | |
9168 if( pBt->autoVacuum ){ | |
9169 checkPtrmap(pCheck, pgnoOvfl, PTRMAP_OVERFLOW1, iPage); | |
9170 } | |
9171 #endif | |
9172 checkList(pCheck, 0, pgnoOvfl, nPage); | |
9173 } | |
9174 | |
9175 if( !pPage->leaf ){ | |
9176 /* Check sanity of left child page for internal pages */ | |
9177 pgno = get4byte(pCell); | |
9178 #ifndef SQLITE_OMIT_AUTOVACUUM | |
9179 if( pBt->autoVacuum ){ | |
9180 checkPtrmap(pCheck, pgno, PTRMAP_BTREE, iPage); | |
9181 } | |
9182 #endif | |
9183 d2 = checkTreePage(pCheck, pgno, &maxKey, maxKey); | |
9184 keyCanBeEqual = 0; | |
9185 if( d2!=depth ){ | |
9186 checkAppendMsg(pCheck, "Child page depth differs"); | |
9187 depth = d2; | |
9188 } | |
9189 }else{ | |
9190 /* Populate the coverage-checking heap for leaf pages */ | |
9191 btreeHeapInsert(heap, (pc<<16)|(pc+info.nSize-1)); | |
9192 } | |
9193 } | |
9194 *piMinKey = maxKey; | |
9195 | |
9196 /* Check for complete coverage of the page | |
9197 */ | |
9198 pCheck->zPfx = 0; | |
9199 if( doCoverageCheck && pCheck->mxErr>0 ){ | |
9200 /* For leaf pages, the min-heap has already been initialized and the | |
9201 ** cells have already been inserted. But for internal pages, that has | |
9202 ** not yet been done, so do it now */ | |
9203 if( !pPage->leaf ){ | |
9204 heap = pCheck->heap; | |
9205 heap[0] = 0; | |
9206 for(i=nCell-1; i>=0; i--){ | |
9207 u32 size; | |
9208 pc = get2byteAligned(&data[cellStart+i*2]); | |
9209 size = pPage->xCellSize(pPage, &data[pc]); | |
9210 btreeHeapInsert(heap, (pc<<16)|(pc+size-1)); | |
9211 } | |
9212 } | |
9213 /* Add the freeblocks to the min-heap | |
9214 ** | |
9215 ** EVIDENCE-OF: R-20690-50594 The second field of the b-tree page header | |
9216 ** is the offset of the first freeblock, or zero if there are no | |
9217 ** freeblocks on the page. | |
9218 */ | |
9219 i = get2byte(&data[hdr+1]); | |
9220 while( i>0 ){ | |
9221 int size, j; | |
9222 assert( (u32)i<=usableSize-4 ); /* Enforced by btreeInitPage() */ | |
9223 size = get2byte(&data[i+2]); | |
9224 assert( (u32)(i+size)<=usableSize ); /* Enforced by btreeInitPage() */ | |
9225 btreeHeapInsert(heap, (((u32)i)<<16)|(i+size-1)); | |
9226 /* EVIDENCE-OF: R-58208-19414 The first 2 bytes of a freeblock are a | |
9227 ** big-endian integer which is the offset in the b-tree page of the next | |
9228 ** freeblock in the chain, or zero if the freeblock is the last on the | |
9229 ** chain. */ | |
9230 j = get2byte(&data[i]); | |
9231 /* EVIDENCE-OF: R-06866-39125 Freeblocks are always connected in order of | |
9232 ** increasing offset. */ | |
9233 assert( j==0 || j>i+size ); /* Enforced by btreeInitPage() */ | |
9234 assert( (u32)j<=usableSize-4 ); /* Enforced by btreeInitPage() */ | |
9235 i = j; | |
9236 } | |
9237 /* Analyze the min-heap looking for overlap between cells and/or | |
9238 ** freeblocks, and counting the number of untracked bytes in nFrag. | |
9239 ** | |
9240 ** Each min-heap entry is of the form: (start_address<<16)|end_address. | |
9241 ** There is an implied first entry the covers the page header, the cell | |
9242 ** pointer index, and the gap between the cell pointer index and the start | |
9243 ** of cell content. | |
9244 ** | |
9245 ** The loop below pulls entries from the min-heap in order and compares | |
9246 ** the start_address against the previous end_address. If there is an | |
9247 ** overlap, that means bytes are used multiple times. If there is a gap, | |
9248 ** that gap is added to the fragmentation count. | |
9249 */ | |
9250 nFrag = 0; | |
9251 prev = contentOffset - 1; /* Implied first min-heap entry */ | |
9252 while( btreeHeapPull(heap,&x) ){ | |
9253 if( (prev&0xffff)>=(x>>16) ){ | |
9254 checkAppendMsg(pCheck, | |
9255 "Multiple uses for byte %u of page %d", x>>16, iPage); | |
9256 break; | |
9257 }else{ | |
9258 nFrag += (x>>16) - (prev&0xffff) - 1; | |
9259 prev = x; | |
9260 } | |
9261 } | |
9262 nFrag += usableSize - (prev&0xffff) - 1; | |
9263 /* EVIDENCE-OF: R-43263-13491 The total number of bytes in all fragments | |
9264 ** is stored in the fifth field of the b-tree page header. | |
9265 ** EVIDENCE-OF: R-07161-27322 The one-byte integer at offset 7 gives the | |
9266 ** number of fragmented free bytes within the cell content area. | |
9267 */ | |
9268 if( heap[0]==0 && nFrag!=data[hdr+7] ){ | |
9269 checkAppendMsg(pCheck, | |
9270 "Fragmentation of %d bytes reported as %d on page %d", | |
9271 nFrag, data[hdr+7], iPage); | |
9272 } | |
9273 } | |
9274 | |
9275 end_of_check: | |
9276 if( !doCoverageCheck ) pPage->isInit = savedIsInit; | |
9277 releasePage(pPage); | |
9278 pCheck->zPfx = saved_zPfx; | |
9279 pCheck->v1 = saved_v1; | |
9280 pCheck->v2 = saved_v2; | |
9281 return depth+1; | |
9282 } | |
9283 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */ | |
9284 | |
9285 #ifndef SQLITE_OMIT_INTEGRITY_CHECK | |
9286 /* | |
9287 ** This routine does a complete check of the given BTree file. aRoot[] is | |
9288 ** an array of pages numbers were each page number is the root page of | |
9289 ** a table. nRoot is the number of entries in aRoot. | |
9290 ** | |
9291 ** A read-only or read-write transaction must be opened before calling | |
9292 ** this function. | |
9293 ** | |
9294 ** Write the number of error seen in *pnErr. Except for some memory | |
9295 ** allocation errors, an error message held in memory obtained from | |
9296 ** malloc is returned if *pnErr is non-zero. If *pnErr==0 then NULL is | |
9297 ** returned. If a memory allocation error occurs, NULL is returned. | |
9298 */ | |
9299 char *sqlite3BtreeIntegrityCheck( | |
9300 Btree *p, /* The btree to be checked */ | |
9301 int *aRoot, /* An array of root pages numbers for individual trees */ | |
9302 int nRoot, /* Number of entries in aRoot[] */ | |
9303 int mxErr, /* Stop reporting errors after this many */ | |
9304 int *pnErr /* Write number of errors seen to this variable */ | |
9305 ){ | |
9306 Pgno i; | |
9307 IntegrityCk sCheck; | |
9308 BtShared *pBt = p->pBt; | |
9309 int savedDbFlags = pBt->db->flags; | |
9310 char zErr[100]; | |
9311 VVA_ONLY( int nRef ); | |
9312 | |
9313 sqlite3BtreeEnter(p); | |
9314 assert( p->inTrans>TRANS_NONE && pBt->inTransaction>TRANS_NONE ); | |
9315 assert( (nRef = sqlite3PagerRefcount(pBt->pPager))>=0 ); | |
9316 sCheck.pBt = pBt; | |
9317 sCheck.pPager = pBt->pPager; | |
9318 sCheck.nPage = btreePagecount(sCheck.pBt); | |
9319 sCheck.mxErr = mxErr; | |
9320 sCheck.nErr = 0; | |
9321 sCheck.mallocFailed = 0; | |
9322 sCheck.zPfx = 0; | |
9323 sCheck.v1 = 0; | |
9324 sCheck.v2 = 0; | |
9325 sCheck.aPgRef = 0; | |
9326 sCheck.heap = 0; | |
9327 sqlite3StrAccumInit(&sCheck.errMsg, 0, zErr, sizeof(zErr), SQLITE_MAX_LENGTH); | |
9328 if( sCheck.nPage==0 ){ | |
9329 goto integrity_ck_cleanup; | |
9330 } | |
9331 | |
9332 sCheck.aPgRef = sqlite3MallocZero((sCheck.nPage / 8)+ 1); | |
9333 if( !sCheck.aPgRef ){ | |
9334 sCheck.mallocFailed = 1; | |
9335 goto integrity_ck_cleanup; | |
9336 } | |
9337 sCheck.heap = (u32*)sqlite3PageMalloc( pBt->pageSize ); | |
9338 if( sCheck.heap==0 ){ | |
9339 sCheck.mallocFailed = 1; | |
9340 goto integrity_ck_cleanup; | |
9341 } | |
9342 | |
9343 i = PENDING_BYTE_PAGE(pBt); | |
9344 if( i<=sCheck.nPage ) setPageReferenced(&sCheck, i); | |
9345 | |
9346 /* Check the integrity of the freelist | |
9347 */ | |
9348 sCheck.zPfx = "Main freelist: "; | |
9349 checkList(&sCheck, 1, get4byte(&pBt->pPage1->aData[32]), | |
9350 get4byte(&pBt->pPage1->aData[36])); | |
9351 sCheck.zPfx = 0; | |
9352 | |
9353 /* Check all the tables. | |
9354 */ | |
9355 testcase( pBt->db->flags & SQLITE_CellSizeCk ); | |
9356 pBt->db->flags &= ~SQLITE_CellSizeCk; | |
9357 for(i=0; (int)i<nRoot && sCheck.mxErr; i++){ | |
9358 i64 notUsed; | |
9359 if( aRoot[i]==0 ) continue; | |
9360 #ifndef SQLITE_OMIT_AUTOVACUUM | |
9361 if( pBt->autoVacuum && aRoot[i]>1 ){ | |
9362 checkPtrmap(&sCheck, aRoot[i], PTRMAP_ROOTPAGE, 0); | |
9363 } | |
9364 #endif | |
9365 checkTreePage(&sCheck, aRoot[i], ¬Used, LARGEST_INT64); | |
9366 } | |
9367 pBt->db->flags = savedDbFlags; | |
9368 | |
9369 /* Make sure every page in the file is referenced | |
9370 */ | |
9371 for(i=1; i<=sCheck.nPage && sCheck.mxErr; i++){ | |
9372 #ifdef SQLITE_OMIT_AUTOVACUUM | |
9373 if( getPageReferenced(&sCheck, i)==0 ){ | |
9374 checkAppendMsg(&sCheck, "Page %d is never used", i); | |
9375 } | |
9376 #else | |
9377 /* If the database supports auto-vacuum, make sure no tables contain | |
9378 ** references to pointer-map pages. | |
9379 */ | |
9380 if( getPageReferenced(&sCheck, i)==0 && | |
9381 (PTRMAP_PAGENO(pBt, i)!=i || !pBt->autoVacuum) ){ | |
9382 checkAppendMsg(&sCheck, "Page %d is never used", i); | |
9383 } | |
9384 if( getPageReferenced(&sCheck, i)!=0 && | |
9385 (PTRMAP_PAGENO(pBt, i)==i && pBt->autoVacuum) ){ | |
9386 checkAppendMsg(&sCheck, "Pointer map page %d is referenced", i); | |
9387 } | |
9388 #endif | |
9389 } | |
9390 | |
9391 /* Clean up and report errors. | |
9392 */ | |
9393 integrity_ck_cleanup: | |
9394 sqlite3PageFree(sCheck.heap); | |
9395 sqlite3_free(sCheck.aPgRef); | |
9396 if( sCheck.mallocFailed ){ | |
9397 sqlite3StrAccumReset(&sCheck.errMsg); | |
9398 sCheck.nErr++; | |
9399 } | |
9400 *pnErr = sCheck.nErr; | |
9401 if( sCheck.nErr==0 ) sqlite3StrAccumReset(&sCheck.errMsg); | |
9402 /* Make sure this analysis did not leave any unref() pages. */ | |
9403 assert( nRef==sqlite3PagerRefcount(pBt->pPager) ); | |
9404 sqlite3BtreeLeave(p); | |
9405 return sqlite3StrAccumFinish(&sCheck.errMsg); | |
9406 } | |
9407 #endif /* SQLITE_OMIT_INTEGRITY_CHECK */ | |
9408 | |
9409 /* | |
9410 ** Return the full pathname of the underlying database file. Return | |
9411 ** an empty string if the database is in-memory or a TEMP database. | |
9412 ** | |
9413 ** The pager filename is invariant as long as the pager is | |
9414 ** open so it is safe to access without the BtShared mutex. | |
9415 */ | |
9416 const char *sqlite3BtreeGetFilename(Btree *p){ | |
9417 assert( p->pBt->pPager!=0 ); | |
9418 return sqlite3PagerFilename(p->pBt->pPager, 1); | |
9419 } | |
9420 | |
9421 /* | |
9422 ** Return the pathname of the journal file for this database. The return | |
9423 ** value of this routine is the same regardless of whether the journal file | |
9424 ** has been created or not. | |
9425 ** | |
9426 ** The pager journal filename is invariant as long as the pager is | |
9427 ** open so it is safe to access without the BtShared mutex. | |
9428 */ | |
9429 const char *sqlite3BtreeGetJournalname(Btree *p){ | |
9430 assert( p->pBt->pPager!=0 ); | |
9431 return sqlite3PagerJournalname(p->pBt->pPager); | |
9432 } | |
9433 | |
9434 /* | |
9435 ** Return non-zero if a transaction is active. | |
9436 */ | |
9437 int sqlite3BtreeIsInTrans(Btree *p){ | |
9438 assert( p==0 || sqlite3_mutex_held(p->db->mutex) ); | |
9439 return (p && (p->inTrans==TRANS_WRITE)); | |
9440 } | |
9441 | |
9442 #ifndef SQLITE_OMIT_WAL | |
9443 /* | |
9444 ** Run a checkpoint on the Btree passed as the first argument. | |
9445 ** | |
9446 ** Return SQLITE_LOCKED if this or any other connection has an open | |
9447 ** transaction on the shared-cache the argument Btree is connected to. | |
9448 ** | |
9449 ** Parameter eMode is one of SQLITE_CHECKPOINT_PASSIVE, FULL or RESTART. | |
9450 */ | |
9451 int sqlite3BtreeCheckpoint(Btree *p, int eMode, int *pnLog, int *pnCkpt){ | |
9452 int rc = SQLITE_OK; | |
9453 if( p ){ | |
9454 BtShared *pBt = p->pBt; | |
9455 sqlite3BtreeEnter(p); | |
9456 if( pBt->inTransaction!=TRANS_NONE ){ | |
9457 rc = SQLITE_LOCKED; | |
9458 }else{ | |
9459 rc = sqlite3PagerCheckpoint(pBt->pPager, eMode, pnLog, pnCkpt); | |
9460 } | |
9461 sqlite3BtreeLeave(p); | |
9462 } | |
9463 return rc; | |
9464 } | |
9465 #endif | |
9466 | |
9467 /* | |
9468 ** Return non-zero if a read (or write) transaction is active. | |
9469 */ | |
9470 int sqlite3BtreeIsInReadTrans(Btree *p){ | |
9471 assert( p ); | |
9472 assert( sqlite3_mutex_held(p->db->mutex) ); | |
9473 return p->inTrans!=TRANS_NONE; | |
9474 } | |
9475 | |
9476 int sqlite3BtreeIsInBackup(Btree *p){ | |
9477 assert( p ); | |
9478 assert( sqlite3_mutex_held(p->db->mutex) ); | |
9479 return p->nBackup!=0; | |
9480 } | |
9481 | |
9482 /* | |
9483 ** This function returns a pointer to a blob of memory associated with | |
9484 ** a single shared-btree. The memory is used by client code for its own | |
9485 ** purposes (for example, to store a high-level schema associated with | |
9486 ** the shared-btree). The btree layer manages reference counting issues. | |
9487 ** | |
9488 ** The first time this is called on a shared-btree, nBytes bytes of memory | |
9489 ** are allocated, zeroed, and returned to the caller. For each subsequent | |
9490 ** call the nBytes parameter is ignored and a pointer to the same blob | |
9491 ** of memory returned. | |
9492 ** | |
9493 ** If the nBytes parameter is 0 and the blob of memory has not yet been | |
9494 ** allocated, a null pointer is returned. If the blob has already been | |
9495 ** allocated, it is returned as normal. | |
9496 ** | |
9497 ** Just before the shared-btree is closed, the function passed as the | |
9498 ** xFree argument when the memory allocation was made is invoked on the | |
9499 ** blob of allocated memory. The xFree function should not call sqlite3_free() | |
9500 ** on the memory, the btree layer does that. | |
9501 */ | |
9502 void *sqlite3BtreeSchema(Btree *p, int nBytes, void(*xFree)(void *)){ | |
9503 BtShared *pBt = p->pBt; | |
9504 sqlite3BtreeEnter(p); | |
9505 if( !pBt->pSchema && nBytes ){ | |
9506 pBt->pSchema = sqlite3DbMallocZero(0, nBytes); | |
9507 pBt->xFreeSchema = xFree; | |
9508 } | |
9509 sqlite3BtreeLeave(p); | |
9510 return pBt->pSchema; | |
9511 } | |
9512 | |
9513 /* | |
9514 ** Return SQLITE_LOCKED_SHAREDCACHE if another user of the same shared | |
9515 ** btree as the argument handle holds an exclusive lock on the | |
9516 ** sqlite_master table. Otherwise SQLITE_OK. | |
9517 */ | |
9518 int sqlite3BtreeSchemaLocked(Btree *p){ | |
9519 int rc; | |
9520 assert( sqlite3_mutex_held(p->db->mutex) ); | |
9521 sqlite3BtreeEnter(p); | |
9522 rc = querySharedCacheTableLock(p, MASTER_ROOT, READ_LOCK); | |
9523 assert( rc==SQLITE_OK || rc==SQLITE_LOCKED_SHAREDCACHE ); | |
9524 sqlite3BtreeLeave(p); | |
9525 return rc; | |
9526 } | |
9527 | |
9528 | |
9529 #ifndef SQLITE_OMIT_SHARED_CACHE | |
9530 /* | |
9531 ** Obtain a lock on the table whose root page is iTab. The | |
9532 ** lock is a write lock if isWritelock is true or a read lock | |
9533 ** if it is false. | |
9534 */ | |
9535 int sqlite3BtreeLockTable(Btree *p, int iTab, u8 isWriteLock){ | |
9536 int rc = SQLITE_OK; | |
9537 assert( p->inTrans!=TRANS_NONE ); | |
9538 if( p->sharable ){ | |
9539 u8 lockType = READ_LOCK + isWriteLock; | |
9540 assert( READ_LOCK+1==WRITE_LOCK ); | |
9541 assert( isWriteLock==0 || isWriteLock==1 ); | |
9542 | |
9543 sqlite3BtreeEnter(p); | |
9544 rc = querySharedCacheTableLock(p, iTab, lockType); | |
9545 if( rc==SQLITE_OK ){ | |
9546 rc = setSharedCacheTableLock(p, iTab, lockType); | |
9547 } | |
9548 sqlite3BtreeLeave(p); | |
9549 } | |
9550 return rc; | |
9551 } | |
9552 #endif | |
9553 | |
9554 #ifndef SQLITE_OMIT_INCRBLOB | |
9555 /* | |
9556 ** Argument pCsr must be a cursor opened for writing on an | |
9557 ** INTKEY table currently pointing at a valid table entry. | |
9558 ** This function modifies the data stored as part of that entry. | |
9559 ** | |
9560 ** Only the data content may only be modified, it is not possible to | |
9561 ** change the length of the data stored. If this function is called with | |
9562 ** parameters that attempt to write past the end of the existing data, | |
9563 ** no modifications are made and SQLITE_CORRUPT is returned. | |
9564 */ | |
9565 int sqlite3BtreePutData(BtCursor *pCsr, u32 offset, u32 amt, void *z){ | |
9566 int rc; | |
9567 assert( cursorHoldsMutex(pCsr) ); | |
9568 assert( sqlite3_mutex_held(pCsr->pBtree->db->mutex) ); | |
9569 assert( pCsr->curFlags & BTCF_Incrblob ); | |
9570 | |
9571 rc = restoreCursorPosition(pCsr); | |
9572 if( rc!=SQLITE_OK ){ | |
9573 return rc; | |
9574 } | |
9575 assert( pCsr->eState!=CURSOR_REQUIRESEEK ); | |
9576 if( pCsr->eState!=CURSOR_VALID ){ | |
9577 return SQLITE_ABORT; | |
9578 } | |
9579 | |
9580 /* Save the positions of all other cursors open on this table. This is | |
9581 ** required in case any of them are holding references to an xFetch | |
9582 ** version of the b-tree page modified by the accessPayload call below. | |
9583 ** | |
9584 ** Note that pCsr must be open on a INTKEY table and saveCursorPosition() | |
9585 ** and hence saveAllCursors() cannot fail on a BTREE_INTKEY table, hence | |
9586 ** saveAllCursors can only return SQLITE_OK. | |
9587 */ | |
9588 VVA_ONLY(rc =) saveAllCursors(pCsr->pBt, pCsr->pgnoRoot, pCsr); | |
9589 assert( rc==SQLITE_OK ); | |
9590 | |
9591 /* Check some assumptions: | |
9592 ** (a) the cursor is open for writing, | |
9593 ** (b) there is a read/write transaction open, | |
9594 ** (c) the connection holds a write-lock on the table (if required), | |
9595 ** (d) there are no conflicting read-locks, and | |
9596 ** (e) the cursor points at a valid row of an intKey table. | |
9597 */ | |
9598 if( (pCsr->curFlags & BTCF_WriteFlag)==0 ){ | |
9599 return SQLITE_READONLY; | |
9600 } | |
9601 assert( (pCsr->pBt->btsFlags & BTS_READ_ONLY)==0 | |
9602 && pCsr->pBt->inTransaction==TRANS_WRITE ); | |
9603 assert( hasSharedCacheTableLock(pCsr->pBtree, pCsr->pgnoRoot, 0, 2) ); | |
9604 assert( !hasReadConflicts(pCsr->pBtree, pCsr->pgnoRoot) ); | |
9605 assert( pCsr->apPage[pCsr->iPage]->intKey ); | |
9606 | |
9607 return accessPayload(pCsr, offset, amt, (unsigned char *)z, 1); | |
9608 } | |
9609 | |
9610 /* | |
9611 ** Mark this cursor as an incremental blob cursor. | |
9612 */ | |
9613 void sqlite3BtreeIncrblobCursor(BtCursor *pCur){ | |
9614 pCur->curFlags |= BTCF_Incrblob; | |
9615 pCur->pBtree->hasIncrblobCur = 1; | |
9616 } | |
9617 #endif | |
9618 | |
9619 /* | |
9620 ** Set both the "read version" (single byte at byte offset 18) and | |
9621 ** "write version" (single byte at byte offset 19) fields in the database | |
9622 ** header to iVersion. | |
9623 */ | |
9624 int sqlite3BtreeSetVersion(Btree *pBtree, int iVersion){ | |
9625 BtShared *pBt = pBtree->pBt; | |
9626 int rc; /* Return code */ | |
9627 | |
9628 assert( iVersion==1 || iVersion==2 ); | |
9629 | |
9630 /* If setting the version fields to 1, do not automatically open the | |
9631 ** WAL connection, even if the version fields are currently set to 2. | |
9632 */ | |
9633 pBt->btsFlags &= ~BTS_NO_WAL; | |
9634 if( iVersion==1 ) pBt->btsFlags |= BTS_NO_WAL; | |
9635 | |
9636 rc = sqlite3BtreeBeginTrans(pBtree, 0); | |
9637 if( rc==SQLITE_OK ){ | |
9638 u8 *aData = pBt->pPage1->aData; | |
9639 if( aData[18]!=(u8)iVersion || aData[19]!=(u8)iVersion ){ | |
9640 rc = sqlite3BtreeBeginTrans(pBtree, 2); | |
9641 if( rc==SQLITE_OK ){ | |
9642 rc = sqlite3PagerWrite(pBt->pPage1->pDbPage); | |
9643 if( rc==SQLITE_OK ){ | |
9644 aData[18] = (u8)iVersion; | |
9645 aData[19] = (u8)iVersion; | |
9646 } | |
9647 } | |
9648 } | |
9649 } | |
9650 | |
9651 pBt->btsFlags &= ~BTS_NO_WAL; | |
9652 return rc; | |
9653 } | |
9654 | |
9655 /* | |
9656 ** Return true if the cursor has a hint specified. This routine is | |
9657 ** only used from within assert() statements | |
9658 */ | |
9659 int sqlite3BtreeCursorHasHint(BtCursor *pCsr, unsigned int mask){ | |
9660 return (pCsr->hints & mask)!=0; | |
9661 } | |
9662 | |
9663 /* | |
9664 ** Return true if the given Btree is read-only. | |
9665 */ | |
9666 int sqlite3BtreeIsReadonly(Btree *p){ | |
9667 return (p->pBt->btsFlags & BTS_READ_ONLY)!=0; | |
9668 } | |
9669 | |
9670 /* | |
9671 ** Return the size of the header added to each page by this module. | |
9672 */ | |
9673 int sqlite3HeaderSizeBtree(void){ return ROUND8(sizeof(MemPage)); } | |
OLD | NEW |