1 /* 2 ** 2008 August 05 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 that page cache. 13 */ 14 #include "sqliteInt.h" 15 16 /* 17 ** A complete page cache is an instance of this structure. 18 */ 19 struct PCache { 20 PgHdr *pDirty, *pDirtyTail; /* List of dirty pages in LRU order */ 21 PgHdr *pSynced; /* Last synced page in dirty page list */ 22 int nRef; /* Number of referenced pages */ 23 int nMax; /* Configured cache size */ 24 int szPage; /* Size of every page in this cache */ 25 int szExtra; /* Size of extra space for each page */ 26 int bPurgeable; /* True if pages are on backing store */ 27 int (*xStress)(void*,PgHdr*); /* Call to try make a page clean */ 28 void *pStress; /* Argument to xStress */ 29 sqlite3_pcache *pCache; /* Pluggable cache module */ 30 PgHdr *pPage1; /* Reference to page 1 */ 31 }; 32 33 /* 34 ** Some of the assert() macros in this code are too expensive to run 35 ** even during normal debugging. Use them only rarely on long-running 36 ** tests. Enable the expensive asserts using the 37 ** -DSQLITE_ENABLE_EXPENSIVE_ASSERT=1 compile-time option. 38 */ 39 #ifdef SQLITE_ENABLE_EXPENSIVE_ASSERT 40 # define expensive_assert(X) assert(X) 41 #else 42 # define expensive_assert(X) 43 #endif 44 45 /********************************** Linked List Management ********************/ 46 47 #if !defined(NDEBUG) && defined(SQLITE_ENABLE_EXPENSIVE_ASSERT) 48 /* 49 ** Check that the pCache->pSynced variable is set correctly. If it 50 ** is not, either fail an assert or return zero. Otherwise, return 51 ** non-zero. This is only used in debugging builds, as follows: 52 ** 53 ** expensive_assert( pcacheCheckSynced(pCache) ); 54 */ 55 static int pcacheCheckSynced(PCache *pCache){ 56 PgHdr *p; 57 for(p=pCache->pDirtyTail; p!=pCache->pSynced; p=p->pDirtyPrev){ 58 assert( p->nRef || (p->flags&PGHDR_NEED_SYNC) ); 59 } 60 return (p==0 || p->nRef || (p->flags&PGHDR_NEED_SYNC)==0); 61 } 62 #endif /* !NDEBUG && SQLITE_ENABLE_EXPENSIVE_ASSERT */ 63 64 /* 65 ** Remove page pPage from the list of dirty pages. 66 */ 67 static void pcacheRemoveFromDirtyList(PgHdr *pPage){ 68 PCache *p = pPage->pCache; 69 70 assert( pPage->pDirtyNext || pPage==p->pDirtyTail ); 71 assert( pPage->pDirtyPrev || pPage==p->pDirty ); 72 73 /* Update the PCache1.pSynced variable if necessary. */ 74 if( p->pSynced==pPage ){ 75 PgHdr *pSynced = pPage->pDirtyPrev; 76 while( pSynced && (pSynced->flags&PGHDR_NEED_SYNC) ){ 77 pSynced = pSynced->pDirtyPrev; 78 } 79 p->pSynced = pSynced; 80 } 81 82 if( pPage->pDirtyNext ){ 83 pPage->pDirtyNext->pDirtyPrev = pPage->pDirtyPrev; 84 }else{ 85 assert( pPage==p->pDirtyTail ); 86 p->pDirtyTail = pPage->pDirtyPrev; 87 } 88 if( pPage->pDirtyPrev ){ 89 pPage->pDirtyPrev->pDirtyNext = pPage->pDirtyNext; 90 }else{ 91 assert( pPage==p->pDirty ); 92 p->pDirty = pPage->pDirtyNext; 93 } 94 pPage->pDirtyNext = 0; 95 pPage->pDirtyPrev = 0; 96 97 expensive_assert( pcacheCheckSynced(p) ); 98 } 99 100 /* 101 ** Add page pPage to the head of the dirty list (PCache1.pDirty is set to 102 ** pPage). 103 */ 104 static void pcacheAddToDirtyList(PgHdr *pPage){ 105 PCache *p = pPage->pCache; 106 107 assert( pPage->pDirtyNext==0 && pPage->pDirtyPrev==0 && p->pDirty!=pPage ); 108 109 pPage->pDirtyNext = p->pDirty; 110 if( pPage->pDirtyNext ){ 111 assert( pPage->pDirtyNext->pDirtyPrev==0 ); 112 pPage->pDirtyNext->pDirtyPrev = pPage; 113 } 114 p->pDirty = pPage; 115 if( !p->pDirtyTail ){ 116 p->pDirtyTail = pPage; 117 } 118 if( !p->pSynced && 0==(pPage->flags&PGHDR_NEED_SYNC) ){ 119 p->pSynced = pPage; 120 } 121 expensive_assert( pcacheCheckSynced(p) ); 122 } 123 124 /* 125 ** Wrapper around the pluggable caches xUnpin method. If the cache is 126 ** being used for an in-memory database, this function is a no-op. 127 */ 128 static void pcacheUnpin(PgHdr *p){ 129 PCache *pCache = p->pCache; 130 if( pCache->bPurgeable ){ 131 if( p->pgno==1 ){ 132 pCache->pPage1 = 0; 133 } 134 sqlite3GlobalConfig.pcache.xUnpin(pCache->pCache, p, 0); 135 } 136 } 137 138 /*************************************************** General Interfaces ****** 139 ** 140 ** Initialize and shutdown the page cache subsystem. Neither of these 141 ** functions are threadsafe. 142 */ 143 int sqlite3PcacheInitialize(void){ 144 if( sqlite3GlobalConfig.pcache.xInit==0 ){ 145 sqlite3PCacheSetDefault(); 146 } 147 return sqlite3GlobalConfig.pcache.xInit(sqlite3GlobalConfig.pcache.pArg); 148 } 149 void sqlite3PcacheShutdown(void){ 150 if( sqlite3GlobalConfig.pcache.xShutdown ){ 151 sqlite3GlobalConfig.pcache.xShutdown(sqlite3GlobalConfig.pcache.pArg); 152 } 153 } 154 155 /* 156 ** Return the size in bytes of a PCache object. 157 */ 158 int sqlite3PcacheSize(void){ return sizeof(PCache); } 159 160 /* 161 ** Create a new PCache object. Storage space to hold the object 162 ** has already been allocated and is passed in as the p pointer. 163 ** The caller discovers how much space needs to be allocated by 164 ** calling sqlite3PcacheSize(). 165 */ 166 void sqlite3PcacheOpen( 167 int szPage, /* Size of every page */ 168 int szExtra, /* Extra space associated with each page */ 169 int bPurgeable, /* True if pages are on backing store */ 170 int (*xStress)(void*,PgHdr*),/* Call to try to make pages clean */ 171 void *pStress, /* Argument to xStress */ 172 PCache *p /* Preallocated space for the PCache */ 173 ){ 174 memset(p, 0, sizeof(PCache)); 175 p->szPage = szPage; 176 p->szExtra = szExtra; 177 p->bPurgeable = bPurgeable; 178 p->xStress = xStress; 179 p->pStress = pStress; 180 p->nMax = 100; 181 } 182 183 /* 184 ** Change the page size for PCache object. The caller must ensure that there 185 ** are no outstanding page references when this function is called. 186 */ 187 void sqlite3PcacheSetPageSize(PCache *pCache, int szPage){ 188 assert( pCache->nRef==0 && pCache->pDirty==0 ); 189 if( pCache->pCache ){ 190 sqlite3GlobalConfig.pcache.xDestroy(pCache->pCache); 191 pCache->pCache = 0; 192 } 193 pCache->szPage = szPage; 194 } 195 196 /* 197 ** Try to obtain a page from the cache. 198 */ 199 int sqlite3PcacheFetch( 200 PCache *pCache, /* Obtain the page from this cache */ 201 Pgno pgno, /* Page number to obtain */ 202 int createFlag, /* If true, create page if it does not exist already */ 203 PgHdr **ppPage /* Write the page here */ 204 ){ 205 PgHdr *pPage = 0; 206 int eCreate; 207 208 assert( pCache!=0 ); 209 assert( createFlag==1 || createFlag==0 ); 210 assert( pgno>0 ); 211 212 /* If the pluggable cache (sqlite3_pcache*) has not been allocated, 213 ** allocate it now. 214 */ 215 if( !pCache->pCache && createFlag ){ 216 sqlite3_pcache *p; 217 int nByte; 218 nByte = pCache->szPage + pCache->szExtra + sizeof(PgHdr); 219 p = sqlite3GlobalConfig.pcache.xCreate(nByte, pCache->bPurgeable); 220 if( !p ){ 221 return SQLITE_NOMEM; 222 } 223 sqlite3GlobalConfig.pcache.xCachesize(p, pCache->nMax); 224 pCache->pCache = p; 225 } 226 227 eCreate = createFlag * (1 + (!pCache->bPurgeable || !pCache->pDirty)); 228 if( pCache->pCache ){ 229 pPage = sqlite3GlobalConfig.pcache.xFetch(pCache->pCache, pgno, eCreate); 230 } 231 232 if( !pPage && eCreate==1 ){ 233 PgHdr *pPg; 234 235 /* Find a dirty page to write-out and recycle. First try to find a 236 ** page that does not require a journal-sync (one with PGHDR_NEED_SYNC 237 ** cleared), but if that is not possible settle for any other 238 ** unreferenced dirty page. 239 */ 240 expensive_assert( pcacheCheckSynced(pCache) ); 241 for(pPg=pCache->pSynced; 242 pPg && (pPg->nRef || (pPg->flags&PGHDR_NEED_SYNC)); 243 pPg=pPg->pDirtyPrev 244 ); 245 if( !pPg ){ 246 for(pPg=pCache->pDirtyTail; pPg && pPg->nRef; pPg=pPg->pDirtyPrev); 247 } 248 if( pPg ){ 249 int rc; 250 rc = pCache->xStress(pCache->pStress, pPg); 251 if( rc!=SQLITE_OK && rc!=SQLITE_BUSY ){ 252 return rc; 253 } 254 } 255 256 pPage = sqlite3GlobalConfig.pcache.xFetch(pCache->pCache, pgno, 2); 257 } 258 259 if( pPage ){ 260 if( !pPage->pData ){ 261 memset(pPage, 0, sizeof(PgHdr) + pCache->szExtra); 262 pPage->pExtra = (void*)&pPage[1]; 263 pPage->pData = (void *)&((char *)pPage)[sizeof(PgHdr) + pCache->szExtra]; 264 pPage->pCache = pCache; 265 pPage->pgno = pgno; 266 } 267 assert( pPage->pCache==pCache ); 268 assert( pPage->pgno==pgno ); 269 assert( pPage->pExtra==(void *)&pPage[1] ); 270 271 if( 0==pPage->nRef ){ 272 pCache->nRef++; 273 } 274 pPage->nRef++; 275 if( pgno==1 ){ 276 pCache->pPage1 = pPage; 277 } 278 } 279 *ppPage = pPage; 280 return (pPage==0 && eCreate) ? SQLITE_NOMEM : SQLITE_OK; 281 } 282 283 /* 284 ** Decrement the reference count on a page. If the page is clean and the 285 ** reference count drops to 0, then it is made elible for recycling. 286 */ 287 void sqlite3PcacheRelease(PgHdr *p){ 288 assert( p->nRef>0 ); 289 p->nRef--; 290 if( p->nRef==0 ){ 291 PCache *pCache = p->pCache; 292 pCache->nRef--; 293 if( (p->flags&PGHDR_DIRTY)==0 ){ 294 pcacheUnpin(p); 295 }else{ 296 /* Move the page to the head of the dirty list. */ 297 pcacheRemoveFromDirtyList(p); 298 pcacheAddToDirtyList(p); 299 } 300 } 301 } 302 303 /* 304 ** Increase the reference count of a supplied page by 1. 305 */ 306 void sqlite3PcacheRef(PgHdr *p){ 307 assert(p->nRef>0); 308 p->nRef++; 309 } 310 311 /* 312 ** Drop a page from the cache. There must be exactly one reference to the 313 ** page. This function deletes that reference, so after it returns the 314 ** page pointed to by p is invalid. 315 */ 316 void sqlite3PcacheDrop(PgHdr *p){ 317 PCache *pCache; 318 assert( p->nRef==1 ); 319 if( p->flags&PGHDR_DIRTY ){ 320 pcacheRemoveFromDirtyList(p); 321 } 322 pCache = p->pCache; 323 pCache->nRef--; 324 if( p->pgno==1 ){ 325 pCache->pPage1 = 0; 326 } 327 sqlite3GlobalConfig.pcache.xUnpin(pCache->pCache, p, 1); 328 } 329 330 /* 331 ** Make sure the page is marked as dirty. If it isn't dirty already, 332 ** make it so. 333 */ 334 void sqlite3PcacheMakeDirty(PgHdr *p){ 335 p->flags &= ~PGHDR_DONT_WRITE; 336 assert( p->nRef>0 ); 337 if( 0==(p->flags & PGHDR_DIRTY) ){ 338 p->flags |= PGHDR_DIRTY; 339 pcacheAddToDirtyList( p); 340 } 341 } 342 343 /* 344 ** Make sure the page is marked as clean. If it isn't clean already, 345 ** make it so. 346 */ 347 void sqlite3PcacheMakeClean(PgHdr *p){ 348 if( (p->flags & PGHDR_DIRTY) ){ 349 pcacheRemoveFromDirtyList(p); 350 p->flags &= ~(PGHDR_DIRTY|PGHDR_NEED_SYNC); 351 if( p->nRef==0 ){ 352 pcacheUnpin(p); 353 } 354 } 355 } 356 357 /* 358 ** Make every page in the cache clean. 359 */ 360 void sqlite3PcacheCleanAll(PCache *pCache){ 361 PgHdr *p; 362 while( (p = pCache->pDirty)!=0 ){ 363 sqlite3PcacheMakeClean(p); 364 } 365 } 366 367 /* 368 ** Clear the PGHDR_NEED_SYNC flag from all dirty pages. 369 */ 370 void sqlite3PcacheClearSyncFlags(PCache *pCache){ 371 PgHdr *p; 372 for(p=pCache->pDirty; p; p=p->pDirtyNext){ 373 p->flags &= ~PGHDR_NEED_SYNC; 374 } 375 pCache->pSynced = pCache->pDirtyTail; 376 } 377 378 /* 379 ** Change the page number of page p to newPgno. 380 */ 381 void sqlite3PcacheMove(PgHdr *p, Pgno newPgno){ 382 PCache *pCache = p->pCache; 383 assert( p->nRef>0 ); 384 assert( newPgno>0 ); 385 sqlite3GlobalConfig.pcache.xRekey(pCache->pCache, p, p->pgno, newPgno); 386 p->pgno = newPgno; 387 if( (p->flags&PGHDR_DIRTY) && (p->flags&PGHDR_NEED_SYNC) ){ 388 pcacheRemoveFromDirtyList(p); 389 pcacheAddToDirtyList(p); 390 } 391 } 392 393 /* 394 ** Drop every cache entry whose page number is greater than "pgno". The 395 ** caller must ensure that there are no outstanding references to any pages 396 ** other than page 1 with a page number greater than pgno. 397 ** 398 ** If there is a reference to page 1 and the pgno parameter passed to this 399 ** function is 0, then the data area associated with page 1 is zeroed, but 400 ** the page object is not dropped. 401 */ 402 void sqlite3PcacheTruncate(PCache *pCache, Pgno pgno){ 403 if( pCache->pCache ){ 404 PgHdr *p; 405 PgHdr *pNext; 406 for(p=pCache->pDirty; p; p=pNext){ 407 pNext = p->pDirtyNext; 408 if( p->pgno>pgno ){ 409 assert( p->flags&PGHDR_DIRTY ); 410 sqlite3PcacheMakeClean(p); 411 } 412 } 413 if( pgno==0 && pCache->pPage1 ){ 414 memset(pCache->pPage1->pData, 0, pCache->szPage); 415 pgno = 1; 416 } 417 sqlite3GlobalConfig.pcache.xTruncate(pCache->pCache, pgno+1); 418 } 419 } 420 421 /* 422 ** Close a cache. 423 */ 424 void sqlite3PcacheClose(PCache *pCache){ 425 if( pCache->pCache ){ 426 sqlite3GlobalConfig.pcache.xDestroy(pCache->pCache); 427 } 428 } 429 430 /* 431 ** Discard the contents of the cache. 432 */ 433 void sqlite3PcacheClear(PCache *pCache){ 434 sqlite3PcacheTruncate(pCache, 0); 435 } 436 437 /* 438 ** Merge two lists of pages connected by pDirty and in pgno order. 439 ** Do not both fixing the pDirtyPrev pointers. 440 */ 441 static PgHdr *pcacheMergeDirtyList(PgHdr *pA, PgHdr *pB){ 442 PgHdr result, *pTail; 443 pTail = &result; 444 while( pA && pB ){ 445 if( pA->pgno<pB->pgno ){ 446 pTail->pDirty = pA; 447 pTail = pA; 448 pA = pA->pDirty; 449 }else{ 450 pTail->pDirty = pB; 451 pTail = pB; 452 pB = pB->pDirty; 453 } 454 } 455 if( pA ){ 456 pTail->pDirty = pA; 457 }else if( pB ){ 458 pTail->pDirty = pB; 459 }else{ 460 pTail->pDirty = 0; 461 } 462 return result.pDirty; 463 } 464 465 /* 466 ** Sort the list of pages in accending order by pgno. Pages are 467 ** connected by pDirty pointers. The pDirtyPrev pointers are 468 ** corrupted by this sort. 469 ** 470 ** Since there cannot be more than 2^31 distinct pages in a database, 471 ** there cannot be more than 31 buckets required by the merge sorter. 472 ** One extra bucket is added to catch overflow in case something 473 ** ever changes to make the previous sentence incorrect. 474 */ 475 #define N_SORT_BUCKET 32 476 static PgHdr *pcacheSortDirtyList(PgHdr *pIn){ 477 PgHdr *a[N_SORT_BUCKET], *p; 478 int i; 479 memset(a, 0, sizeof(a)); 480 while( pIn ){ 481 p = pIn; 482 pIn = p->pDirty; 483 p->pDirty = 0; 484 for(i=0; ALWAYS(i<N_SORT_BUCKET-1); i++){ 485 if( a[i]==0 ){ 486 a[i] = p; 487 break; 488 }else{ 489 p = pcacheMergeDirtyList(a[i], p); 490 a[i] = 0; 491 } 492 } 493 if( NEVER(i==N_SORT_BUCKET-1) ){ 494 /* To get here, there need to be 2^(N_SORT_BUCKET) elements in 495 ** the input list. But that is impossible. 496 */ 497 a[i] = pcacheMergeDirtyList(a[i], p); 498 } 499 } 500 p = a[0]; 501 for(i=1; i<N_SORT_BUCKET; i++){ 502 p = pcacheMergeDirtyList(p, a[i]); 503 } 504 return p; 505 } 506 507 /* 508 ** Return a list of all dirty pages in the cache, sorted by page number. 509 */ 510 PgHdr *sqlite3PcacheDirtyList(PCache *pCache){ 511 PgHdr *p; 512 for(p=pCache->pDirty; p; p=p->pDirtyNext){ 513 p->pDirty = p->pDirtyNext; 514 } 515 return pcacheSortDirtyList(pCache->pDirty); 516 } 517 518 /* 519 ** Return the total number of referenced pages held by the cache. 520 */ 521 int sqlite3PcacheRefCount(PCache *pCache){ 522 return pCache->nRef; 523 } 524 525 /* 526 ** Return the number of references to the page supplied as an argument. 527 */ 528 int sqlite3PcachePageRefcount(PgHdr *p){ 529 return p->nRef; 530 } 531 532 /* 533 ** Return the total number of pages in the cache. 534 */ 535 int sqlite3PcachePagecount(PCache *pCache){ 536 int nPage = 0; 537 if( pCache->pCache ){ 538 nPage = sqlite3GlobalConfig.pcache.xPagecount(pCache->pCache); 539 } 540 return nPage; 541 } 542 543 #ifdef SQLITE_TEST 544 /* 545 ** Get the suggested cache-size value. 546 */ 547 int sqlite3PcacheGetCachesize(PCache *pCache){ 548 return pCache->nMax; 549 } 550 #endif 551 552 /* 553 ** Set the suggested cache-size value. 554 */ 555 void sqlite3PcacheSetCachesize(PCache *pCache, int mxPage){ 556 pCache->nMax = mxPage; 557 if( pCache->pCache ){ 558 sqlite3GlobalConfig.pcache.xCachesize(pCache->pCache, mxPage); 559 } 560 } 561 562 #if defined(SQLITE_CHECK_PAGES) || defined(SQLITE_DEBUG) 563 /* 564 ** For all dirty pages currently in the cache, invoke the specified 565 ** callback. This is only used if the SQLITE_CHECK_PAGES macro is 566 ** defined. 567 */ 568 void sqlite3PcacheIterateDirty(PCache *pCache, void (*xIter)(PgHdr *)){ 569 PgHdr *pDirty; 570 for(pDirty=pCache->pDirty; pDirty; pDirty=pDirty->pDirtyNext){ 571 xIter(pDirty); 572 } 573 } 574 #endif 575