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 szCache; /* 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.pcache2.xUnpin(pCache->pCache, p->pPage, 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.pcache2.xInit==0 ){ 145 /* IMPLEMENTATION-OF: R-26801-64137 If the xInit() method is NULL, then the 146 ** built-in default page cache is used instead of the application defined 147 ** page cache. */ 148 sqlite3PCacheSetDefault(); 149 } 150 return sqlite3GlobalConfig.pcache2.xInit(sqlite3GlobalConfig.pcache2.pArg); 151 } 152 void sqlite3PcacheShutdown(void){ 153 if( sqlite3GlobalConfig.pcache2.xShutdown ){ 154 /* IMPLEMENTATION-OF: R-26000-56589 The xShutdown() method may be NULL. */ 155 sqlite3GlobalConfig.pcache2.xShutdown(sqlite3GlobalConfig.pcache2.pArg); 156 } 157 } 158 159 /* 160 ** Return the size in bytes of a PCache object. 161 */ 162 int sqlite3PcacheSize(void){ return sizeof(PCache); } 163 164 /* 165 ** Create a new PCache object. Storage space to hold the object 166 ** has already been allocated and is passed in as the p pointer. 167 ** The caller discovers how much space needs to be allocated by 168 ** calling sqlite3PcacheSize(). 169 */ 170 void sqlite3PcacheOpen( 171 int szPage, /* Size of every page */ 172 int szExtra, /* Extra space associated with each page */ 173 int bPurgeable, /* True if pages are on backing store */ 174 int (*xStress)(void*,PgHdr*),/* Call to try to make pages clean */ 175 void *pStress, /* Argument to xStress */ 176 PCache *p /* Preallocated space for the PCache */ 177 ){ 178 memset(p, 0, sizeof(PCache)); 179 p->szPage = szPage; 180 p->szExtra = szExtra; 181 p->bPurgeable = bPurgeable; 182 p->xStress = xStress; 183 p->pStress = pStress; 184 p->szCache = 100; 185 } 186 187 /* 188 ** Change the page size for PCache object. The caller must ensure that there 189 ** are no outstanding page references when this function is called. 190 */ 191 void sqlite3PcacheSetPageSize(PCache *pCache, int szPage){ 192 assert( pCache->nRef==0 && pCache->pDirty==0 ); 193 if( pCache->pCache ){ 194 sqlite3GlobalConfig.pcache2.xDestroy(pCache->pCache); 195 pCache->pCache = 0; 196 pCache->pPage1 = 0; 197 } 198 pCache->szPage = szPage; 199 } 200 201 /* 202 ** Compute the number of pages of cache requested. 203 */ 204 static int numberOfCachePages(PCache *p){ 205 if( p->szCache>=0 ){ 206 return p->szCache; 207 }else{ 208 return (int)((-1024*(i64)p->szCache)/(p->szPage+p->szExtra)); 209 } 210 } 211 212 /* 213 ** Try to obtain a page from the cache. 214 */ 215 int sqlite3PcacheFetch( 216 PCache *pCache, /* Obtain the page from this cache */ 217 Pgno pgno, /* Page number to obtain */ 218 int createFlag, /* If true, create page if it does not exist already */ 219 PgHdr **ppPage /* Write the page here */ 220 ){ 221 sqlite3_pcache_page *pPage = 0; 222 PgHdr *pPgHdr = 0; 223 int eCreate; 224 225 assert( pCache!=0 ); 226 assert( createFlag==1 || createFlag==0 ); 227 assert( pgno>0 ); 228 229 /* If the pluggable cache (sqlite3_pcache*) has not been allocated, 230 ** allocate it now. 231 */ 232 if( !pCache->pCache && createFlag ){ 233 sqlite3_pcache *p; 234 p = sqlite3GlobalConfig.pcache2.xCreate( 235 pCache->szPage, pCache->szExtra + sizeof(PgHdr), pCache->bPurgeable 236 ); 237 if( !p ){ 238 return SQLITE_NOMEM; 239 } 240 sqlite3GlobalConfig.pcache2.xCachesize(p, numberOfCachePages(pCache)); 241 pCache->pCache = p; 242 } 243 244 eCreate = createFlag * (1 + (!pCache->bPurgeable || !pCache->pDirty)); 245 if( pCache->pCache ){ 246 pPage = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache, pgno, eCreate); 247 } 248 249 if( !pPage && eCreate==1 ){ 250 PgHdr *pPg; 251 252 /* Find a dirty page to write-out and recycle. First try to find a 253 ** page that does not require a journal-sync (one with PGHDR_NEED_SYNC 254 ** cleared), but if that is not possible settle for any other 255 ** unreferenced dirty page. 256 */ 257 expensive_assert( pcacheCheckSynced(pCache) ); 258 for(pPg=pCache->pSynced; 259 pPg && (pPg->nRef || (pPg->flags&PGHDR_NEED_SYNC)); 260 pPg=pPg->pDirtyPrev 261 ); 262 pCache->pSynced = pPg; 263 if( !pPg ){ 264 for(pPg=pCache->pDirtyTail; pPg && pPg->nRef; pPg=pPg->pDirtyPrev); 265 } 266 if( pPg ){ 267 int rc; 268 #ifdef SQLITE_LOG_CACHE_SPILL 269 sqlite3_log(SQLITE_FULL, 270 "spill page %d making room for %d - cache used: %d/%d", 271 pPg->pgno, pgno, 272 sqlite3GlobalConfig.pcache.xPagecount(pCache->pCache), 273 numberOfCachePages(pCache)); 274 #endif 275 rc = pCache->xStress(pCache->pStress, pPg); 276 if( rc!=SQLITE_OK && rc!=SQLITE_BUSY ){ 277 return rc; 278 } 279 } 280 281 pPage = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache, pgno, 2); 282 } 283 284 if( pPage ){ 285 pPgHdr = (PgHdr *)pPage->pExtra; 286 287 if( !pPgHdr->pPage ){ 288 memset(pPgHdr, 0, sizeof(PgHdr)); 289 pPgHdr->pPage = pPage; 290 pPgHdr->pData = pPage->pBuf; 291 pPgHdr->pExtra = (void *)&pPgHdr[1]; 292 memset(pPgHdr->pExtra, 0, pCache->szExtra); 293 pPgHdr->pCache = pCache; 294 pPgHdr->pgno = pgno; 295 } 296 assert( pPgHdr->pCache==pCache ); 297 assert( pPgHdr->pgno==pgno ); 298 assert( pPgHdr->pData==pPage->pBuf ); 299 assert( pPgHdr->pExtra==(void *)&pPgHdr[1] ); 300 301 if( 0==pPgHdr->nRef ){ 302 pCache->nRef++; 303 } 304 pPgHdr->nRef++; 305 if( pgno==1 ){ 306 pCache->pPage1 = pPgHdr; 307 } 308 } 309 *ppPage = pPgHdr; 310 return (pPgHdr==0 && eCreate) ? SQLITE_NOMEM : SQLITE_OK; 311 } 312 313 /* 314 ** Decrement the reference count on a page. If the page is clean and the 315 ** reference count drops to 0, then it is made elible for recycling. 316 */ 317 void sqlite3PcacheRelease(PgHdr *p){ 318 assert( p->nRef>0 ); 319 p->nRef--; 320 if( p->nRef==0 ){ 321 PCache *pCache = p->pCache; 322 pCache->nRef--; 323 if( (p->flags&PGHDR_DIRTY)==0 ){ 324 pcacheUnpin(p); 325 }else{ 326 /* Move the page to the head of the dirty list. */ 327 pcacheRemoveFromDirtyList(p); 328 pcacheAddToDirtyList(p); 329 } 330 } 331 } 332 333 /* 334 ** Increase the reference count of a supplied page by 1. 335 */ 336 void sqlite3PcacheRef(PgHdr *p){ 337 assert(p->nRef>0); 338 p->nRef++; 339 } 340 341 /* 342 ** Drop a page from the cache. There must be exactly one reference to the 343 ** page. This function deletes that reference, so after it returns the 344 ** page pointed to by p is invalid. 345 */ 346 void sqlite3PcacheDrop(PgHdr *p){ 347 PCache *pCache; 348 assert( p->nRef==1 ); 349 if( p->flags&PGHDR_DIRTY ){ 350 pcacheRemoveFromDirtyList(p); 351 } 352 pCache = p->pCache; 353 pCache->nRef--; 354 if( p->pgno==1 ){ 355 pCache->pPage1 = 0; 356 } 357 sqlite3GlobalConfig.pcache2.xUnpin(pCache->pCache, p->pPage, 1); 358 } 359 360 /* 361 ** Make sure the page is marked as dirty. If it isn't dirty already, 362 ** make it so. 363 */ 364 void sqlite3PcacheMakeDirty(PgHdr *p){ 365 p->flags &= ~PGHDR_DONT_WRITE; 366 assert( p->nRef>0 ); 367 if( 0==(p->flags & PGHDR_DIRTY) ){ 368 p->flags |= PGHDR_DIRTY; 369 pcacheAddToDirtyList( p); 370 } 371 } 372 373 /* 374 ** Make sure the page is marked as clean. If it isn't clean already, 375 ** make it so. 376 */ 377 void sqlite3PcacheMakeClean(PgHdr *p){ 378 if( (p->flags & PGHDR_DIRTY) ){ 379 pcacheRemoveFromDirtyList(p); 380 p->flags &= ~(PGHDR_DIRTY|PGHDR_NEED_SYNC); 381 if( p->nRef==0 ){ 382 pcacheUnpin(p); 383 } 384 } 385 } 386 387 /* 388 ** Make every page in the cache clean. 389 */ 390 void sqlite3PcacheCleanAll(PCache *pCache){ 391 PgHdr *p; 392 while( (p = pCache->pDirty)!=0 ){ 393 sqlite3PcacheMakeClean(p); 394 } 395 } 396 397 /* 398 ** Clear the PGHDR_NEED_SYNC flag from all dirty pages. 399 */ 400 void sqlite3PcacheClearSyncFlags(PCache *pCache){ 401 PgHdr *p; 402 for(p=pCache->pDirty; p; p=p->pDirtyNext){ 403 p->flags &= ~PGHDR_NEED_SYNC; 404 } 405 pCache->pSynced = pCache->pDirtyTail; 406 } 407 408 /* 409 ** Change the page number of page p to newPgno. 410 */ 411 void sqlite3PcacheMove(PgHdr *p, Pgno newPgno){ 412 PCache *pCache = p->pCache; 413 assert( p->nRef>0 ); 414 assert( newPgno>0 ); 415 sqlite3GlobalConfig.pcache2.xRekey(pCache->pCache, p->pPage, p->pgno,newPgno); 416 p->pgno = newPgno; 417 if( (p->flags&PGHDR_DIRTY) && (p->flags&PGHDR_NEED_SYNC) ){ 418 pcacheRemoveFromDirtyList(p); 419 pcacheAddToDirtyList(p); 420 } 421 } 422 423 /* 424 ** Drop every cache entry whose page number is greater than "pgno". The 425 ** caller must ensure that there are no outstanding references to any pages 426 ** other than page 1 with a page number greater than pgno. 427 ** 428 ** If there is a reference to page 1 and the pgno parameter passed to this 429 ** function is 0, then the data area associated with page 1 is zeroed, but 430 ** the page object is not dropped. 431 */ 432 void sqlite3PcacheTruncate(PCache *pCache, Pgno pgno){ 433 if( pCache->pCache ){ 434 PgHdr *p; 435 PgHdr *pNext; 436 for(p=pCache->pDirty; p; p=pNext){ 437 pNext = p->pDirtyNext; 438 /* This routine never gets call with a positive pgno except right 439 ** after sqlite3PcacheCleanAll(). So if there are dirty pages, 440 ** it must be that pgno==0. 441 */ 442 assert( p->pgno>0 ); 443 if( ALWAYS(p->pgno>pgno) ){ 444 assert( p->flags&PGHDR_DIRTY ); 445 sqlite3PcacheMakeClean(p); 446 } 447 } 448 if( pgno==0 && pCache->pPage1 ){ 449 memset(pCache->pPage1->pData, 0, pCache->szPage); 450 pgno = 1; 451 } 452 sqlite3GlobalConfig.pcache2.xTruncate(pCache->pCache, pgno+1); 453 } 454 } 455 456 /* 457 ** Close a cache. 458 */ 459 void sqlite3PcacheClose(PCache *pCache){ 460 if( pCache->pCache ){ 461 sqlite3GlobalConfig.pcache2.xDestroy(pCache->pCache); 462 } 463 } 464 465 /* 466 ** Discard the contents of the cache. 467 */ 468 void sqlite3PcacheClear(PCache *pCache){ 469 sqlite3PcacheTruncate(pCache, 0); 470 } 471 472 /* 473 ** Merge two lists of pages connected by pDirty and in pgno order. 474 ** Do not both fixing the pDirtyPrev pointers. 475 */ 476 static PgHdr *pcacheMergeDirtyList(PgHdr *pA, PgHdr *pB){ 477 PgHdr result, *pTail; 478 pTail = &result; 479 while( pA && pB ){ 480 if( pA->pgno<pB->pgno ){ 481 pTail->pDirty = pA; 482 pTail = pA; 483 pA = pA->pDirty; 484 }else{ 485 pTail->pDirty = pB; 486 pTail = pB; 487 pB = pB->pDirty; 488 } 489 } 490 if( pA ){ 491 pTail->pDirty = pA; 492 }else if( pB ){ 493 pTail->pDirty = pB; 494 }else{ 495 pTail->pDirty = 0; 496 } 497 return result.pDirty; 498 } 499 500 /* 501 ** Sort the list of pages in accending order by pgno. Pages are 502 ** connected by pDirty pointers. The pDirtyPrev pointers are 503 ** corrupted by this sort. 504 ** 505 ** Since there cannot be more than 2^31 distinct pages in a database, 506 ** there cannot be more than 31 buckets required by the merge sorter. 507 ** One extra bucket is added to catch overflow in case something 508 ** ever changes to make the previous sentence incorrect. 509 */ 510 #define N_SORT_BUCKET 32 511 static PgHdr *pcacheSortDirtyList(PgHdr *pIn){ 512 PgHdr *a[N_SORT_BUCKET], *p; 513 int i; 514 memset(a, 0, sizeof(a)); 515 while( pIn ){ 516 p = pIn; 517 pIn = p->pDirty; 518 p->pDirty = 0; 519 for(i=0; ALWAYS(i<N_SORT_BUCKET-1); i++){ 520 if( a[i]==0 ){ 521 a[i] = p; 522 break; 523 }else{ 524 p = pcacheMergeDirtyList(a[i], p); 525 a[i] = 0; 526 } 527 } 528 if( NEVER(i==N_SORT_BUCKET-1) ){ 529 /* To get here, there need to be 2^(N_SORT_BUCKET) elements in 530 ** the input list. But that is impossible. 531 */ 532 a[i] = pcacheMergeDirtyList(a[i], p); 533 } 534 } 535 p = a[0]; 536 for(i=1; i<N_SORT_BUCKET; i++){ 537 p = pcacheMergeDirtyList(p, a[i]); 538 } 539 return p; 540 } 541 542 /* 543 ** Return a list of all dirty pages in the cache, sorted by page number. 544 */ 545 PgHdr *sqlite3PcacheDirtyList(PCache *pCache){ 546 PgHdr *p; 547 for(p=pCache->pDirty; p; p=p->pDirtyNext){ 548 p->pDirty = p->pDirtyNext; 549 } 550 return pcacheSortDirtyList(pCache->pDirty); 551 } 552 553 /* 554 ** Return the total number of referenced pages held by the cache. 555 */ 556 int sqlite3PcacheRefCount(PCache *pCache){ 557 return pCache->nRef; 558 } 559 560 /* 561 ** Return the number of references to the page supplied as an argument. 562 */ 563 int sqlite3PcachePageRefcount(PgHdr *p){ 564 return p->nRef; 565 } 566 567 /* 568 ** Return the total number of pages in the cache. 569 */ 570 int sqlite3PcachePagecount(PCache *pCache){ 571 int nPage = 0; 572 if( pCache->pCache ){ 573 nPage = sqlite3GlobalConfig.pcache2.xPagecount(pCache->pCache); 574 } 575 return nPage; 576 } 577 578 #ifdef SQLITE_TEST 579 /* 580 ** Get the suggested cache-size value. 581 */ 582 int sqlite3PcacheGetCachesize(PCache *pCache){ 583 return numberOfCachePages(pCache); 584 } 585 #endif 586 587 /* 588 ** Set the suggested cache-size value. 589 */ 590 void sqlite3PcacheSetCachesize(PCache *pCache, int mxPage){ 591 pCache->szCache = mxPage; 592 if( pCache->pCache ){ 593 sqlite3GlobalConfig.pcache2.xCachesize(pCache->pCache, 594 numberOfCachePages(pCache)); 595 } 596 } 597 598 /* 599 ** Free up as much memory as possible from the page cache. 600 */ 601 void sqlite3PcacheShrink(PCache *pCache){ 602 if( pCache->pCache ){ 603 sqlite3GlobalConfig.pcache2.xShrink(pCache->pCache); 604 } 605 } 606 607 #if defined(SQLITE_CHECK_PAGES) || defined(SQLITE_DEBUG) 608 /* 609 ** For all dirty pages currently in the cache, invoke the specified 610 ** callback. This is only used if the SQLITE_CHECK_PAGES macro is 611 ** defined. 612 */ 613 void sqlite3PcacheIterateDirty(PCache *pCache, void (*xIter)(PgHdr *)){ 614 PgHdr *pDirty; 615 for(pDirty=pCache->pDirty; pDirty; pDirty=pDirty->pDirtyNext){ 616 xIter(pDirty); 617 } 618 } 619 #endif 620