src/pcache.c

   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.  Every
  18 ** entry in the cache holds a single page of the database file.  The
  19 ** btree layer only operates on the cached copy of the database pages.
  20 **
  21 ** A page cache entry is "clean" if it exactly matches what is currently
  22 ** on disk.  A page is "dirty" if it has been modified and needs to be
  23 ** persisted to disk.
  24 **
  25 ** pDirty, pDirtyTail, pSynced:
  26 **   All dirty pages are linked into the doubly linked list using
  27 **   PgHdr.pDirtyNext and pDirtyPrev. The list is maintained in LRU order
  28 **   such that p was added to the list more recently than p->pDirtyNext.
  29 **   PCache.pDirty points to the first (newest) element in the list and
  30 **   pDirtyTail to the last (oldest).
  31 **
  32 **   The PCache.pSynced variable is used to optimize searching for a dirty
  33 **   page to eject from the cache mid-transaction. It is better to eject
  34 **   a page that does not require a journal sync than one that does.
  35 **   Therefore, pSynced is maintained so that it *almost* always points
  36 **   to either the oldest page in the pDirty/pDirtyTail list that has a
  37 **   clear PGHDR_NEED_SYNC flag or to a page that is older than this one
  38 **   (so that the right page to eject can be found by following pDirtyPrev
  39 **   pointers).
  40 */
  41 struct PCache {
  42   PgHdr *pDirty, *pDirtyTail;         /* List of dirty pages in LRU order */
  43   PgHdr *pSynced;                     /* Last synced page in dirty page list */
  44   i64 nRefSum;                        /* Sum of ref counts over all pages */
  45   int szCache;                        /* Configured cache size */
  46   int szSpill;                        /* Size before spilling occurs */
  47   int szPage;                         /* Size of every page in this cache */
  48   int szExtra;                        /* Size of extra space for each page */
  49   u8 bPurgeable;                      /* True if pages are on backing store */
  50   u8 eCreate;                         /* eCreate value for for xFetch() */
  51   int (*xStress)(void*,PgHdr*);       /* Call to try make a page clean */
  52   void *pStress;                      /* Argument to xStress */
  53   sqlite3_pcache *pCache;             /* Pluggable cache module */
  54 };
  55
  56 /********************************** Test and Debug Logic **********************/
  57 /*
  58 ** Debug tracing macros.  Enable by by changing the "0" to "1" and
  59 ** recompiling.
  60 **
  61 ** When sqlite3PcacheTrace is 1, single line trace messages are issued.
  62 ** When sqlite3PcacheTrace is 2, a dump of the pcache showing all cache entries
  63 ** is displayed for many operations, resulting in a lot of output.
  64 */
  65 #if defined(SQLITE_DEBUG) && 0
  66   int sqlite3PcacheTrace = 2;       /* 0: off  1: simple  2: cache dumps */
  67   int sqlite3PcacheMxDump = 9999;   /* Max cache entries for pcacheDump() */
  68 # define pcacheTrace(X) if(sqlite3PcacheTrace){sqlite3DebugPrintf X;}
  69   static void pcachePageTrace(int i, sqlite3_pcache_page *pLower){
  70     PgHdr *pPg;
  71     unsigned char *a;
  72     int j;
  73     if( pLower==0 ){
  74       printf("%3d: NULL\n", i);
  75     }else{
  76       pPg = (PgHdr*)pLower->pExtra;
  77       printf("%3d: nRef %2lld flgs %02x data ", i, pPg->nRef, pPg->flags);
  78       a = (unsigned char *)pLower->pBuf;
  79       for(j=0; j<12; j++) printf("%02x", a[j]);
  80       printf(" ptr %p\n", pPg);
  81     }
  82   }
  83   static void pcacheDump(PCache *pCache){
  84     int N;
  85     int i;
  86     sqlite3_pcache_page *pLower;
  87
  88     if( sqlite3PcacheTrace<2 ) return;
  89     if( pCache->pCache==0 ) return;
  90     N = sqlite3PcachePagecount(pCache);
  91     if( N>sqlite3PcacheMxDump ) N = sqlite3PcacheMxDump;
  92     for(i=1; i<=N; i++){
  93        pLower = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache, i, 0);
  94        pcachePageTrace(i, pLower);
  95        if( pLower && ((PgHdr*)pLower)->pPage==0 ){
  96          sqlite3GlobalConfig.pcache2.xUnpin(pCache->pCache, pLower, 0);
  97        }
  98     }
  99   }
 100 #else
 101 # define pcacheTrace(X)
 102 # define pcachePageTrace(PGNO, X)
 103 # define pcacheDump(X)
 104 #endif
 105
 106 /*
 107 ** Return 1 if pPg is on the dirty list for pCache.  Return 0 if not.
 108 ** This routine runs inside of assert() statements only.
 109 */
 110 #ifdef SQLITE_DEBUG
 111 static int pageOnDirtyList(PCache *pCache, PgHdr *pPg){
 112   PgHdr *p;
 113   for(p=pCache->pDirty; p; p=p->pDirtyNext){
 114     if( p==pPg ) return 1;
 115   }
 116   return 0;
 117 }
 118 #endif
 119
 120 /*
 121 ** Check invariants on a PgHdr entry.  Return true if everything is OK.
 122 ** Return false if any invariant is violated.
 123 **
 124 ** This routine is for use inside of assert() statements only.  For
 125 ** example:
 126 **
 127 **          assert( sqlite3PcachePageSanity(pPg) );
 128 */
 129 #ifdef SQLITE_DEBUG
 130 int sqlite3PcachePageSanity(PgHdr *pPg){
 131   PCache *pCache;
 132   assert( pPg!=0 );
 133   assert( pPg->pgno>0 || pPg->pPager==0 );    /* Page number is 1 or more */
 134   pCache = pPg->pCache;
 135   assert( pCache!=0 );      /* Every page has an associated PCache */
 136   if( pPg->flags & PGHDR_CLEAN ){
 137     assert( (pPg->flags & PGHDR_DIRTY)==0 );/* Cannot be both CLEAN and DIRTY */
 138     assert( !pageOnDirtyList(pCache, pPg) );/* CLEAN pages not on dirty list */
 139   }else{
 140     assert( (pPg->flags & PGHDR_DIRTY)!=0 );/* If not CLEAN must be DIRTY */
 141     assert( pPg->pDirtyNext==0 || pPg->pDirtyNext->pDirtyPrev==pPg );
 142     assert( pPg->pDirtyPrev==0 || pPg->pDirtyPrev->pDirtyNext==pPg );
 143     assert( pPg->pDirtyPrev!=0 || pCache->pDirty==pPg );
 144     assert( pageOnDirtyList(pCache, pPg) );
 145   }
 146   /* WRITEABLE pages must also be DIRTY */
 147   if( pPg->flags & PGHDR_WRITEABLE ){
 148     assert( pPg->flags & PGHDR_DIRTY );     /* WRITEABLE implies DIRTY */
 149   }
 150   /* NEED_SYNC can be set independently of WRITEABLE.  This can happen,
 151   ** for example, when using the sqlite3PagerDontWrite() optimization:
 152   **    (1)  Page X is journalled, and gets WRITEABLE and NEED_SEEK.
 153   **    (2)  Page X moved to freelist, WRITEABLE is cleared
 154   **    (3)  Page X reused, WRITEABLE is set again
 155   ** If NEED_SYNC had been cleared in step 2, then it would not be reset
 156   ** in step 3, and page might be written into the database without first
 157   ** syncing the rollback journal, which might cause corruption on a power
 158   ** loss.
 159   **
 160   ** Another example is when the database page size is smaller than the
 161   ** disk sector size.  When any page of a sector is journalled, all pages
 162   ** in that sector are marked NEED_SYNC even if they are still CLEAN, just
 163   ** in case they are later modified, since all pages in the same sector
 164   ** must be journalled and synced before any of those pages can be safely
 165   ** written.
 166   */
 167   return 1;
 168 }
 169 #endif /* SQLITE_DEBUG */
 170
 171
 172 /********************************** Linked List Management ********************/
 173
 174 /* Allowed values for second argument to pcacheManageDirtyList() */
 175 #define PCACHE_DIRTYLIST_REMOVE   1    /* Remove pPage from dirty list */
 176 #define PCACHE_DIRTYLIST_ADD      2    /* Add pPage to the dirty list */
 177 #define PCACHE_DIRTYLIST_FRONT    3    /* Move pPage to the front of the list */
 178
 179 /*
 180 ** Manage pPage's participation on the dirty list.  Bits of the addRemove
 181 ** argument determines what operation to do.  The 0x01 bit means first
 182 ** remove pPage from the dirty list.  The 0x02 means add pPage back to
 183 ** the dirty list.  Doing both moves pPage to the front of the dirty list.
 184 */
 185 static void pcacheManageDirtyList(PgHdr *pPage, u8 addRemove){
 186   PCache *p = pPage->pCache;
 187
 188   pcacheTrace(("%p.DIRTYLIST.%s %d\n", p,
 189                 addRemove==1 ? "REMOVE" : addRemove==2 ? "ADD" : "FRONT",
 190                 pPage->pgno));
 191   if( addRemove & PCACHE_DIRTYLIST_REMOVE ){
 192     assert( pPage->pDirtyNext || pPage==p->pDirtyTail );
 193     assert( pPage->pDirtyPrev || pPage==p->pDirty );
 194
 195     /* Update the PCache1.pSynced variable if necessary. */
 196     if( p->pSynced==pPage ){
 197       p->pSynced = pPage->pDirtyPrev;
 198     }
 199
 200     if( pPage->pDirtyNext ){
 201       pPage->pDirtyNext->pDirtyPrev = pPage->pDirtyPrev;
 202     }else{
 203       assert( pPage==p->pDirtyTail );
 204       p->pDirtyTail = pPage->pDirtyPrev;
 205     }
 206     if( pPage->pDirtyPrev ){
 207       pPage->pDirtyPrev->pDirtyNext = pPage->pDirtyNext;
 208     }else{
 209       /* If there are now no dirty pages in the cache, set eCreate to 2.
 210       ** This is an optimization that allows sqlite3PcacheFetch() to skip
 211       ** searching for a dirty page to eject from the cache when it might
 212       ** otherwise have to.  */
 213       assert( pPage==p->pDirty );
 214       p->pDirty = pPage->pDirtyNext;
 215       assert( p->bPurgeable || p->eCreate==2 );
 216       if( p->pDirty==0 ){         /*OPTIMIZATION-IF-TRUE*/
 217         assert( p->bPurgeable==0 || p->eCreate==1 );
 218         p->eCreate = 2;
 219       }
 220     }
 221   }
 222   if( addRemove & PCACHE_DIRTYLIST_ADD ){
 223     pPage->pDirtyPrev = 0;
 224     pPage->pDirtyNext = p->pDirty;
 225     if( pPage->pDirtyNext ){
 226       assert( pPage->pDirtyNext->pDirtyPrev==0 );
 227       pPage->pDirtyNext->pDirtyPrev = pPage;
 228     }else{
 229       p->pDirtyTail = pPage;
 230       if( p->bPurgeable ){
 231         assert( p->eCreate==2 );
 232         p->eCreate = 1;
 233       }
 234     }
 235     p->pDirty = pPage;
 236
 237     /* If pSynced is NULL and this page has a clear NEED_SYNC flag, set
 238     ** pSynced to point to it. Checking the NEED_SYNC flag is an
 239     ** optimization, as if pSynced points to a page with the NEED_SYNC
 240     ** flag set sqlite3PcacheFetchStress() searches through all newer
 241     ** entries of the dirty-list for a page with NEED_SYNC clear anyway.  */
 242     if( !p->pSynced
 243      && 0==(pPage->flags&PGHDR_NEED_SYNC)   /*OPTIMIZATION-IF-FALSE*/
 244     ){
 245       p->pSynced = pPage;
 246     }
 247   }
 248   pcacheDump(p);
 249 }
 250
 251 /*
 252 ** Wrapper around the pluggable caches xUnpin method. If the cache is
 253 ** being used for an in-memory database, this function is a no-op.
 254 */
 255 static void pcacheUnpin(PgHdr *p){
 256   if( p->pCache->bPurgeable ){
 257     pcacheTrace(("%p.UNPIN %d\n", p->pCache, p->pgno));
 258     sqlite3GlobalConfig.pcache2.xUnpin(p->pCache->pCache, p->pPage, 0);
 259     pcacheDump(p->pCache);
 260   }
 261 }
 262
 263 /*
 264 ** Compute the number of pages of cache requested.   p->szCache is the
 265 ** cache size requested by the "PRAGMA cache_size" statement.
 266 */
 267 static int numberOfCachePages(PCache *p){
 268   if( p->szCache>=0 ){
 269     /* IMPLEMENTATION-OF: R-42059-47211 If the argument N is positive then the
 270     ** suggested cache size is set to N. */
 271     return p->szCache;
 272   }else{
 273     i64 n;
 274     /* IMPLEMENTATION-OF: R-59858-46238 If the argument N is negative, then the
 275     ** number of cache pages is adjusted to be a number of pages that would
 276     ** use approximately abs(N*1024) bytes of memory based on the current
 277     ** page size. */
 278     n = ((-1024*(i64)p->szCache)/(p->szPage+p->szExtra));
 279     if( n>1000000000 ) n = 1000000000;
 280     return (int)n;
 281   }
 282 }
 283
 284 /*************************************************** General Interfaces ******
 285 **
 286 ** Initialize and shutdown the page cache subsystem. Neither of these
 287 ** functions are threadsafe.
 288 */
 289 int sqlite3PcacheInitialize(void){
 290   if( sqlite3GlobalConfig.pcache2.xInit==0 ){
 291     /* IMPLEMENTATION-OF: R-26801-64137 If the xInit() method is NULL, then the
 292     ** built-in default page cache is used instead of the application defined
 293     ** page cache. */
 294     sqlite3PCacheSetDefault();
 295     assert( sqlite3GlobalConfig.pcache2.xInit!=0 );
 296   }
 297   return sqlite3GlobalConfig.pcache2.xInit(sqlite3GlobalConfig.pcache2.pArg);
 298 }
 299 void sqlite3PcacheShutdown(void){
 300   if( sqlite3GlobalConfig.pcache2.xShutdown ){
 301     /* IMPLEMENTATION-OF: R-26000-56589 The xShutdown() method may be NULL. */
 302     sqlite3GlobalConfig.pcache2.xShutdown(sqlite3GlobalConfig.pcache2.pArg);
 303   }
 304 }
 305
 306 /*
 307 ** Return the size in bytes of a PCache object.
 308 */
 309 int sqlite3PcacheSize(void){ return sizeof(PCache); }
 310
 311 /*
 312 ** Create a new PCache object. Storage space to hold the object
 313 ** has already been allocated and is passed in as the p pointer.
 314 ** The caller discovers how much space needs to be allocated by
 315 ** calling sqlite3PcacheSize().
 316 **
 317 ** szExtra is some extra space allocated for each page.  The first
 318 ** 8 bytes of the extra space will be zeroed as the page is allocated,
 319 ** but remaining content will be uninitialized.  Though it is opaque
 320 ** to this module, the extra space really ends up being the MemPage
 321 ** structure in the pager.
 322 */
 323 int sqlite3PcacheOpen(
 324   int szPage,                  /* Size of every page */
 325   int szExtra,                 /* Extra space associated with each page */
 326   int bPurgeable,              /* True if pages are on backing store */
 327   int (*xStress)(void*,PgHdr*),/* Call to try to make pages clean */
 328   void *pStress,               /* Argument to xStress */
 329   PCache *p                    /* Preallocated space for the PCache */
 330 ){
 331   memset(p, 0, sizeof(PCache));
 332   p->szPage = 1;
 333   p->szExtra = szExtra;
 334   assert( szExtra>=8 );  /* First 8 bytes will be zeroed */
 335   p->bPurgeable = bPurgeable;
 336   p->eCreate = 2;
 337   p->xStress = xStress;
 338   p->pStress = pStress;
 339   p->szCache = 100;
 340   p->szSpill = 1;
 341   pcacheTrace(("%p.OPEN szPage %d bPurgeable %d\n",p,szPage,bPurgeable));
 342   return sqlite3PcacheSetPageSize(p, szPage);
 343 }
 344
 345 /*
 346 ** Change the page size for PCache object. The caller must ensure that there
 347 ** are no outstanding page references when this function is called.
 348 */
 349 int sqlite3PcacheSetPageSize(PCache *pCache, int szPage){
 350   assert( pCache->nRefSum==0 && pCache->pDirty==0 );
 351   if( pCache->szPage ){
 352     sqlite3_pcache *pNew;
 353     pNew = sqlite3GlobalConfig.pcache2.xCreate(
 354                 szPage, pCache->szExtra + ROUND8(sizeof(PgHdr)),
 355                 pCache->bPurgeable
 356     );
 357     if( pNew==0 ) return SQLITE_NOMEM_BKPT;
 358     sqlite3GlobalConfig.pcache2.xCachesize(pNew, numberOfCachePages(pCache));
 359     if( pCache->pCache ){
 360       sqlite3GlobalConfig.pcache2.xDestroy(pCache->pCache);
 361     }
 362     pCache->pCache = pNew;
 363     pCache->szPage = szPage;
 364     pcacheTrace(("%p.PAGESIZE %d\n",pCache,szPage));
 365   }
 366   return SQLITE_OK;
 367 }
 368
 369 /*
 370 ** Try to obtain a page from the cache.
 371 **
 372 ** This routine returns a pointer to an sqlite3_pcache_page object if
 373 ** such an object is already in cache, or if a new one is created.
 374 ** This routine returns a NULL pointer if the object was not in cache
 375 ** and could not be created.
 376 **
 377 ** The createFlags should be 0 to check for existing pages and should
 378 ** be 3 (not 1, but 3) to try to create a new page.
 379 **
 380 ** If the createFlag is 0, then NULL is always returned if the page
 381 ** is not already in the cache.  If createFlag is 1, then a new page
 382 ** is created only if that can be done without spilling dirty pages
 383 ** and without exceeding the cache size limit.
 384 **
 385 ** The caller needs to invoke sqlite3PcacheFetchFinish() to properly
 386 ** initialize the sqlite3_pcache_page object and convert it into a
 387 ** PgHdr object.  The sqlite3PcacheFetch() and sqlite3PcacheFetchFinish()
 388 ** routines are split this way for performance reasons. When separated
 389 ** they can both (usually) operate without having to push values to
 390 ** the stack on entry and pop them back off on exit, which saves a
 391 ** lot of pushing and popping.
 392 */
 393 sqlite3_pcache_page *sqlite3PcacheFetch(
 394   PCache *pCache,       /* Obtain the page from this cache */
 395   Pgno pgno,            /* Page number to obtain */
 396   int createFlag        /* If true, create page if it does not exist already */
 397 ){
 398   int eCreate;
 399   sqlite3_pcache_page *pRes;
 400
 401   assert( pCache!=0 );
 402   assert( pCache->pCache!=0 );
 403   assert( createFlag==3 || createFlag==0 );
 404   assert( pCache->eCreate==((pCache->bPurgeable && pCache->pDirty) ? 1 : 2) );
 405
 406   /* eCreate defines what to do if the page does not exist.
 407   **    0     Do not allocate a new page.  (createFlag==0)
 408   **    1     Allocate a new page if doing so is inexpensive.
 409   **          (createFlag==1 AND bPurgeable AND pDirty)
 410   **    2     Allocate a new page even it doing so is difficult.
 411   **          (createFlag==1 AND !(bPurgeable AND pDirty)
 412   */
 413   eCreate = createFlag & pCache->eCreate;
 414   assert( eCreate==0 || eCreate==1 || eCreate==2 );
 415   assert( createFlag==0 || pCache->eCreate==eCreate );
 416   assert( createFlag==0 || eCreate==1+(!pCache->bPurgeable||!pCache->pDirty) );
 417   pRes = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache, pgno, eCreate);
 418   pcacheTrace(("%p.FETCH %d%s (result: %p) ",pCache,pgno,
 419                createFlag?" create":"",pRes));
 420   pcachePageTrace(pgno, pRes);
 421   return pRes;
 422 }
 423
 424 /*
 425 ** If the sqlite3PcacheFetch() routine is unable to allocate a new
 426 ** page because no clean pages are available for reuse and the cache
 427 ** size limit has been reached, then this routine can be invoked to
 428 ** try harder to allocate a page.  This routine might invoke the stress
 429 ** callback to spill dirty pages to the journal.  It will then try to
 430 ** allocate the new page and will only fail to allocate a new page on
 431 ** an OOM error.
 432 **
 433 ** This routine should be invoked only after sqlite3PcacheFetch() fails.
 434 */
 435 int sqlite3PcacheFetchStress(
 436   PCache *pCache,                 /* Obtain the page from this cache */
 437   Pgno pgno,                      /* Page number to obtain */
 438   sqlite3_pcache_page **ppPage    /* Write result here */
 439 ){
 440   PgHdr *pPg;
 441   if( pCache->eCreate==2 ) return 0;
 442
 443   if( sqlite3PcachePagecount(pCache)>pCache->szSpill ){
 444     /* Find a dirty page to write-out and recycle. First try to find a
 445     ** page that does not require a journal-sync (one with PGHDR_NEED_SYNC
 446     ** cleared), but if that is not possible settle for any other
 447     ** unreferenced dirty page.
 448     **
 449     ** If the LRU page in the dirty list that has a clear PGHDR_NEED_SYNC
 450     ** flag is currently referenced, then the following may leave pSynced
 451     ** set incorrectly (pointing to other than the LRU page with NEED_SYNC
 452     ** cleared). This is Ok, as pSynced is just an optimization.  */
 453     for(pPg=pCache->pSynced;
 454         pPg && (pPg->nRef || (pPg->flags&PGHDR_NEED_SYNC));
 455         pPg=pPg->pDirtyPrev
 456     );
 457     pCache->pSynced = pPg;
 458     if( !pPg ){
 459       for(pPg=pCache->pDirtyTail; pPg && pPg->nRef; pPg=pPg->pDirtyPrev);
 460     }
 461     if( pPg ){
 462       int rc;
 463 #ifdef SQLITE_LOG_CACHE_SPILL
 464       sqlite3_log(SQLITE_FULL,
 465                   "spill page %d making room for %d - cache used: %d/%d",
 466                   pPg->pgno, pgno,
 467                   sqlite3GlobalConfig.pcache2.xPagecount(pCache->pCache),
 468                 numberOfCachePages(pCache));
 469 #endif
 470       pcacheTrace(("%p.SPILL %d\n",pCache,pPg->pgno));
 471       rc = pCache->xStress(pCache->pStress, pPg);
 472       pcacheDump(pCache);
 473       if( rc!=SQLITE_OK && rc!=SQLITE_BUSY ){
 474         return rc;
 475       }
 476     }
 477   }
 478   *ppPage = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache, pgno, 2);
 479   return *ppPage==0 ? SQLITE_NOMEM_BKPT : SQLITE_OK;
 480 }
 481
 482 /*
 483 ** This is a helper routine for sqlite3PcacheFetchFinish()
 484 **
 485 ** In the uncommon case where the page being fetched has not been
 486 ** initialized, this routine is invoked to do the initialization.
 487 ** This routine is broken out into a separate function since it
 488 ** requires extra stack manipulation that can be avoided in the common
 489 ** case.
 490 */
 491 static SQLITE_NOINLINE PgHdr *pcacheFetchFinishWithInit(
 492   PCache *pCache,             /* Obtain the page from this cache */
 493   Pgno pgno,                  /* Page number obtained */
 494   sqlite3_pcache_page *pPage  /* Page obtained by prior PcacheFetch() call */
 495 ){
 496   PgHdr *pPgHdr;
 497   assert( pPage!=0 );
 498   pPgHdr = (PgHdr*)pPage->pExtra;
 499   assert( pPgHdr->pPage==0 );
 500   memset(&pPgHdr->pDirty, 0, sizeof(PgHdr) - offsetof(PgHdr,pDirty));
 501   pPgHdr->pPage = pPage;
 502   pPgHdr->pData = pPage->pBuf;
 503   pPgHdr->pExtra = (void *)&pPgHdr[1];
 504   memset(pPgHdr->pExtra, 0, 8);
 505   pPgHdr->pCache = pCache;
 506   pPgHdr->pgno = pgno;
 507   pPgHdr->flags = PGHDR_CLEAN;
 508   return sqlite3PcacheFetchFinish(pCache,pgno,pPage);
 509 }
 510
 511 /*
 512 ** This routine converts the sqlite3_pcache_page object returned by
 513 ** sqlite3PcacheFetch() into an initialized PgHdr object.  This routine
 514 ** must be called after sqlite3PcacheFetch() in order to get a usable
 515 ** result.
 516 */
 517 PgHdr *sqlite3PcacheFetchFinish(
 518   PCache *pCache,             /* Obtain the page from this cache */
 519   Pgno pgno,                  /* Page number obtained */
 520   sqlite3_pcache_page *pPage  /* Page obtained by prior PcacheFetch() call */
 521 ){
 522   PgHdr *pPgHdr;
 523
 524   assert( pPage!=0 );
 525   pPgHdr = (PgHdr *)pPage->pExtra;
 526
 527   if( !pPgHdr->pPage ){
 528     return pcacheFetchFinishWithInit(pCache, pgno, pPage);
 529   }
 530   pCache->nRefSum++;
 531   pPgHdr->nRef++;
 532   assert( sqlite3PcachePageSanity(pPgHdr) );
 533   return pPgHdr;
 534 }
 535
 536 /*
 537 ** Decrement the reference count on a page. If the page is clean and the
 538 ** reference count drops to 0, then it is made eligible for recycling.
 539 */
 540 void SQLITE_NOINLINE sqlite3PcacheRelease(PgHdr *p){
 541   assert( p->nRef>0 );
 542   p->pCache->nRefSum--;
 543   if( (--p->nRef)==0 ){
 544     if( p->flags&PGHDR_CLEAN ){
 545       pcacheUnpin(p);
 546     }else{
 547       pcacheManageDirtyList(p, PCACHE_DIRTYLIST_FRONT);
 548       assert( sqlite3PcachePageSanity(p) );
 549     }
 550   }
 551 }
 552
 553 /*
 554 ** Increase the reference count of a supplied page by 1.
 555 */
 556 void sqlite3PcacheRef(PgHdr *p){
 557   assert(p->nRef>0);
 558   assert( sqlite3PcachePageSanity(p) );
 559   p->nRef++;
 560   p->pCache->nRefSum++;
 561 }
 562
 563 /*
 564 ** Drop a page from the cache. There must be exactly one reference to the
 565 ** page. This function deletes that reference, so after it returns the
 566 ** page pointed to by p is invalid.
 567 */
 568 void sqlite3PcacheDrop(PgHdr *p){
 569   assert( p->nRef==1 );
 570   assert( sqlite3PcachePageSanity(p) );
 571   if( p->flags&PGHDR_DIRTY ){
 572     pcacheManageDirtyList(p, PCACHE_DIRTYLIST_REMOVE);
 573   }
 574   p->pCache->nRefSum--;
 575   sqlite3GlobalConfig.pcache2.xUnpin(p->pCache->pCache, p->pPage, 1);
 576 }
 577
 578 /*
 579 ** Make sure the page is marked as dirty. If it isn't dirty already,
 580 ** make it so.
 581 */
 582 void sqlite3PcacheMakeDirty(PgHdr *p){
 583   assert( p->nRef>0 );
 584   assert( sqlite3PcachePageSanity(p) );
 585   if( p->flags & (PGHDR_CLEAN|PGHDR_DONT_WRITE) ){    /*OPTIMIZATION-IF-FALSE*/
 586     p->flags &= ~PGHDR_DONT_WRITE;
 587     if( p->flags & PGHDR_CLEAN ){
 588       p->flags ^= (PGHDR_DIRTY|PGHDR_CLEAN);
 589       pcacheTrace(("%p.DIRTY %d\n",p->pCache,p->pgno));
 590       assert( (p->flags & (PGHDR_DIRTY|PGHDR_CLEAN))==PGHDR_DIRTY );
 591       pcacheManageDirtyList(p, PCACHE_DIRTYLIST_ADD);
 592       assert( sqlite3PcachePageSanity(p) );
 593     }
 594     assert( sqlite3PcachePageSanity(p) );
 595   }
 596 }
 597
 598 /*
 599 ** Make sure the page is marked as clean. If it isn't clean already,
 600 ** make it so.
 601 */
 602 void sqlite3PcacheMakeClean(PgHdr *p){
 603   assert( sqlite3PcachePageSanity(p) );
 604   assert( (p->flags & PGHDR_DIRTY)!=0 );
 605   assert( (p->flags & PGHDR_CLEAN)==0 );
 606   pcacheManageDirtyList(p, PCACHE_DIRTYLIST_REMOVE);
 607   p->flags &= ~(PGHDR_DIRTY|PGHDR_NEED_SYNC|PGHDR_WRITEABLE);
 608   p->flags |= PGHDR_CLEAN;
 609   pcacheTrace(("%p.CLEAN %d\n",p->pCache,p->pgno));
 610   assert( sqlite3PcachePageSanity(p) );
 611   if( p->nRef==0 ){
 612     pcacheUnpin(p);
 613   }
 614 }
 615
 616 /*
 617 ** Make every page in the cache clean.
 618 */
 619 void sqlite3PcacheCleanAll(PCache *pCache){
 620   PgHdr *p;
 621   pcacheTrace(("%p.CLEAN-ALL\n",pCache));
 622   while( (p = pCache->pDirty)!=0 ){
 623     sqlite3PcacheMakeClean(p);
 624   }
 625 }
 626
 627 /*
 628 ** Clear the PGHDR_NEED_SYNC and PGHDR_WRITEABLE flag from all dirty pages.
 629 */
 630 void sqlite3PcacheClearWritable(PCache *pCache){
 631   PgHdr *p;
 632   pcacheTrace(("%p.CLEAR-WRITEABLE\n",pCache));
 633   for(p=pCache->pDirty; p; p=p->pDirtyNext){
 634     p->flags &= ~(PGHDR_NEED_SYNC|PGHDR_WRITEABLE);
 635   }
 636   pCache->pSynced = pCache->pDirtyTail;
 637 }
 638
 639 /*
 640 ** Clear the PGHDR_NEED_SYNC flag from all dirty pages.
 641 */
 642 void sqlite3PcacheClearSyncFlags(PCache *pCache){
 643   PgHdr *p;
 644   for(p=pCache->pDirty; p; p=p->pDirtyNext){
 645     p->flags &= ~PGHDR_NEED_SYNC;
 646   }
 647   pCache->pSynced = pCache->pDirtyTail;
 648 }
 649
 650 /*
 651 ** Change the page number of page p to newPgno.
 652 */
 653 void sqlite3PcacheMove(PgHdr *p, Pgno newPgno){
 654   PCache *pCache = p->pCache;
 655   sqlite3_pcache_page *pOther;
 656   assert( p->nRef>0 );
 657   assert( newPgno>0 );
 658   assert( sqlite3PcachePageSanity(p) );
 659   pcacheTrace(("%p.MOVE %d -> %d\n",pCache,p->pgno,newPgno));
 660   pOther = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache, newPgno, 0);
 661   if( pOther ){
 662     PgHdr *pXPage = (PgHdr*)pOther->pExtra;
 663     assert( pXPage->nRef==0 );
 664     pXPage->nRef++;
 665     pCache->nRefSum++;
 666     sqlite3PcacheDrop(pXPage);
 667   }
 668   sqlite3GlobalConfig.pcache2.xRekey(pCache->pCache, p->pPage, p->pgno,newPgno);
 669   p->pgno = newPgno;
 670   if( (p->flags&PGHDR_DIRTY) && (p->flags&PGHDR_NEED_SYNC) ){
 671     pcacheManageDirtyList(p, PCACHE_DIRTYLIST_FRONT);
 672     assert( sqlite3PcachePageSanity(p) );
 673   }
 674 }
 675
 676 /*
 677 ** Drop every cache entry whose page number is greater than "pgno". The
 678 ** caller must ensure that there are no outstanding references to any pages
 679 ** other than page 1 with a page number greater than pgno.
 680 **
 681 ** If there is a reference to page 1 and the pgno parameter passed to this
 682 ** function is 0, then the data area associated with page 1 is zeroed, but
 683 ** the page object is not dropped.
 684 */
 685 void sqlite3PcacheTruncate(PCache *pCache, Pgno pgno){
 686   if( pCache->pCache ){
 687     PgHdr *p;
 688     PgHdr *pNext;
 689     pcacheTrace(("%p.TRUNCATE %d\n",pCache,pgno));
 690     for(p=pCache->pDirty; p; p=pNext){
 691       pNext = p->pDirtyNext;
 692       /* This routine never gets call with a positive pgno except right
 693       ** after sqlite3PcacheCleanAll().  So if there are dirty pages,
 694       ** it must be that pgno==0.
 695       */
 696       assert( p->pgno>0 );
 697       if( p->pgno>pgno ){
 698         assert( p->flags&PGHDR_DIRTY );
 699         sqlite3PcacheMakeClean(p);
 700       }
 701     }
 702     if( pgno==0 && pCache->nRefSum ){
 703       sqlite3_pcache_page *pPage1;
 704       pPage1 = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache,1,0);
 705       if( ALWAYS(pPage1) ){  /* Page 1 is always available in cache, because
 706                              ** pCache->nRefSum>0 */
 707         memset(pPage1->pBuf, 0, pCache->szPage);
 708         pgno = 1;
 709       }
 710     }
 711     sqlite3GlobalConfig.pcache2.xTruncate(pCache->pCache, pgno+1);
 712   }
 713 }
 714
 715 /*
 716 ** Close a cache.
 717 */
 718 void sqlite3PcacheClose(PCache *pCache){
 719   assert( pCache->pCache!=0 );
 720   pcacheTrace(("%p.CLOSE\n",pCache));
 721   sqlite3GlobalConfig.pcache2.xDestroy(pCache->pCache);
 722 }
 723
 724 /*
 725 ** Discard the contents of the cache.
 726 */
 727 void sqlite3PcacheClear(PCache *pCache){
 728   sqlite3PcacheTruncate(pCache, 0);
 729 }
 730
 731 /*
 732 ** Merge two lists of pages connected by pDirty and in pgno order.
 733 ** Do not bother fixing the pDirtyPrev pointers.
 734 */
 735 static PgHdr *pcacheMergeDirtyList(PgHdr *pA, PgHdr *pB){
 736   PgHdr result, *pTail;
 737   pTail = &result;
 738   assert( pA!=0 && pB!=0 );
 739   for(;;){
 740     if( pA->pgno<pB->pgno ){
 741       pTail->pDirty = pA;
 742       pTail = pA;
 743       pA = pA->pDirty;
 744       if( pA==0 ){
 745         pTail->pDirty = pB;
 746         break;
 747       }
 748     }else{
 749       pTail->pDirty = pB;
 750       pTail = pB;
 751       pB = pB->pDirty;
 752       if( pB==0 ){
 753         pTail->pDirty = pA;
 754         break;
 755       }
 756     }
 757   }
 758   return result.pDirty;
 759 }
 760
 761 /*
 762 ** Sort the list of pages in ascending order by pgno.  Pages are
 763 ** connected by pDirty pointers.  The pDirtyPrev pointers are
 764 ** corrupted by this sort.
 765 **
 766 ** Since there cannot be more than 2^31 distinct pages in a database,
 767 ** there cannot be more than 31 buckets required by the merge sorter.
 768 ** One extra bucket is added to catch overflow in case something
 769 ** ever changes to make the previous sentence incorrect.
 770 */
 771 #define N_SORT_BUCKET  32
 772 static PgHdr *pcacheSortDirtyList(PgHdr *pIn){
 773   PgHdr *a[N_SORT_BUCKET], *p;
 774   int i;
 775   memset(a, 0, sizeof(a));
 776   while( pIn ){
 777     p = pIn;
 778     pIn = p->pDirty;
 779     p->pDirty = 0;
 780     for(i=0; ALWAYS(i<N_SORT_BUCKET-1); i++){
 781       if( a[i]==0 ){
 782         a[i] = p;
 783         break;
 784       }else{
 785         p = pcacheMergeDirtyList(a[i], p);
 786         a[i] = 0;
 787       }
 788     }
 789     if( NEVER(i==N_SORT_BUCKET-1) ){
 790       /* To get here, there need to be 2^(N_SORT_BUCKET) elements in
 791       ** the input list.  But that is impossible.
 792       */
 793       a[i] = pcacheMergeDirtyList(a[i], p);
 794     }
 795   }
 796   p = a[0];
 797   for(i=1; i<N_SORT_BUCKET; i++){
 798     if( a[i]==0 ) continue;
 799     p = p ? pcacheMergeDirtyList(p, a[i]) : a[i];
 800   }
 801   return p;
 802 }
 803
 804 /*
 805 ** Return a list of all dirty pages in the cache, sorted by page number.
 806 */
 807 PgHdr *sqlite3PcacheDirtyList(PCache *pCache){
 808   PgHdr *p;
 809   for(p=pCache->pDirty; p; p=p->pDirtyNext){
 810     p->pDirty = p->pDirtyNext;
 811   }
 812   return pcacheSortDirtyList(pCache->pDirty);
 813 }
 814
 815 /*
 816 ** Return the total number of references to all pages held by the cache.
 817 **
 818 ** This is not the total number of pages referenced, but the sum of the
 819 ** reference count for all pages.
 820 */
 821 i64 sqlite3PcacheRefCount(PCache *pCache){
 822   return pCache->nRefSum;
 823 }
 824
 825 /*
 826 ** Return the number of references to the page supplied as an argument.
 827 */
 828 i64 sqlite3PcachePageRefcount(PgHdr *p){
 829   return p->nRef;
 830 }
 831
 832 /*
 833 ** Return the total number of pages in the cache.
 834 */
 835 int sqlite3PcachePagecount(PCache *pCache){
 836   assert( pCache->pCache!=0 );
 837   return sqlite3GlobalConfig.pcache2.xPagecount(pCache->pCache);
 838 }
 839
 840 #ifdef SQLITE_TEST
 841 /*
 842 ** Get the suggested cache-size value.
 843 */
 844 int sqlite3PcacheGetCachesize(PCache *pCache){
 845   return numberOfCachePages(pCache);
 846 }
 847 #endif
 848
 849 /*
 850 ** Set the suggested cache-size value.
 851 */
 852 void sqlite3PcacheSetCachesize(PCache *pCache, int mxPage){
 853   assert( pCache->pCache!=0 );
 854   pCache->szCache = mxPage;
 855   sqlite3GlobalConfig.pcache2.xCachesize(pCache->pCache,
 856                                          numberOfCachePages(pCache));
 857 }
 858
 859 /*
 860 ** Set the suggested cache-spill value.  Make no changes if if the
 861 ** argument is zero.  Return the effective cache-spill size, which will
 862 ** be the larger of the szSpill and szCache.
 863 */
 864 int sqlite3PcacheSetSpillsize(PCache *p, int mxPage){
 865   int res;
 866   assert( p->pCache!=0 );
 867   if( mxPage ){
 868     if( mxPage<0 ){
 869       mxPage = (int)((-1024*(i64)mxPage)/(p->szPage+p->szExtra));
 870     }
 871     p->szSpill = mxPage;
 872   }
 873   res = numberOfCachePages(p);
 874   if( res<p->szSpill ) res = p->szSpill;
 875   return res;
 876 }
 877
 878 /*
 879 ** Free up as much memory as possible from the page cache.
 880 */
 881 void sqlite3PcacheShrink(PCache *pCache){
 882   assert( pCache->pCache!=0 );
 883   sqlite3GlobalConfig.pcache2.xShrink(pCache->pCache);
 884 }
 885
 886 /*
 887 ** Return the size of the header added by this middleware layer
 888 ** in the page-cache hierarchy.
 889 */
 890 int sqlite3HeaderSizePcache(void){ return ROUND8(sizeof(PgHdr)); }
 891
 892 /*
 893 ** Return the number of dirty pages currently in the cache, as a percentage
 894 ** of the configured cache size.
 895 */
 896 int sqlite3PCachePercentDirty(PCache *pCache){
 897   PgHdr *pDirty;
 898   int nDirty = 0;
 899   int nCache = numberOfCachePages(pCache);
 900   for(pDirty=pCache->pDirty; pDirty; pDirty=pDirty->pDirtyNext) nDirty++;
 901   return nCache ? (int)(((i64)nDirty * 100) / nCache) : 0;
 902 }
 903
 904 #ifdef SQLITE_DIRECT_OVERFLOW_READ
 905 /*
 906 ** Return true if there are one or more dirty pages in the cache. Else false.
 907 */
 908 int sqlite3PCacheIsDirty(PCache *pCache){
 909   return (pCache->pDirty!=0);
 910 }
 911 #endif
 912
 913 #if defined(SQLITE_CHECK_PAGES) || defined(SQLITE_DEBUG)
 914 /*
 915 ** For all dirty pages currently in the cache, invoke the specified
 916 ** callback. This is only used if the SQLITE_CHECK_PAGES macro is
 917 ** defined.
 918 */
 919 void sqlite3PcacheIterateDirty(PCache *pCache, void (*xIter)(PgHdr *)){
 920   PgHdr *pDirty;
 921   for(pDirty=pCache->pDirty; pDirty; pDirty=pDirty->pDirtyNext){
 922     xIter(pDirty);
 923   }
 924 }
 925 #endif