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 to 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   int 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   void pcacheDump(PCache *pCache){
  70     int N;
  71     int i, j;
  72     sqlite3_pcache_page *pLower;
  73     PgHdr *pPg;
  74     unsigned char *a;
  75
  76     if( sqlite3PcacheTrace<2 ) return;
  77     if( pCache->pCache==0 ) return;
  78     N = sqlite3PcachePagecount(pCache);
  79     if( N>sqlite3PcacheMxDump ) N = sqlite3PcacheMxDump;
  80     for(i=1; i<=N; i++){
  81        pLower = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache, i, 0);
  82        if( pLower==0 ) continue;
  83        pPg = (PgHdr*)pLower->pExtra;
  84        printf("%3d: nRef %2d flgs %02x data ", i, pPg->nRef, pPg->flags);
  85        a = (unsigned char *)pLower->pBuf;
  86        for(j=0; j<12; j++) printf("%02x", a[j]);
  87        printf("\n");
  88        if( pPg->pPage==0 ){
  89          sqlite3GlobalConfig.pcache2.xUnpin(pCache->pCache, pLower, 0);
  90        }
  91     }
  92   }
  93   #else
  94 # define pcacheTrace(X)
  95 # define pcacheDump(X)
  96 #endif
  97
  98 /*
  99 ** Check invariants on a PgHdr entry.  Return true if everything is OK.
 100 ** Return false if any invariant is violated.
 101 **
 102 ** This routine is for use inside of assert() statements only.  For
 103 ** example:
 104 **
 105 **          assert( sqlite3PcachePageSanity(pPg) );
 106 */
 107 #ifdef SQLITE_DEBUG
 108 int sqlite3PcachePageSanity(PgHdr *pPg){
 109   PCache *pCache;
 110   assert( pPg!=0 );
 111   assert( pPg->pgno>0 || pPg->pPager==0 );    /* Page number is 1 or more */
 112   pCache = pPg->pCache;
 113   assert( pCache!=0 );      /* Every page has an associated PCache */
 114   if( pPg->flags & PGHDR_CLEAN ){
 115     assert( (pPg->flags & PGHDR_DIRTY)==0 );/* Cannot be both CLEAN and DIRTY */
 116     assert( pCache->pDirty!=pPg );          /* CLEAN pages not on dirty list */
 117     assert( pCache->pDirtyTail!=pPg );
 118   }
 119   /* WRITEABLE pages must also be DIRTY */
 120   if( pPg->flags & PGHDR_WRITEABLE ){
 121     assert( pPg->flags & PGHDR_DIRTY );     /* WRITEABLE implies DIRTY */
 122   }
 123   /* NEED_SYNC can be set independently of WRITEABLE.  This can happen,
 124   ** for example, when using the sqlite3PagerDontWrite() optimization:
 125   **    (1)  Page X is journalled, and gets WRITEABLE and NEED_SEEK.
 126   **    (2)  Page X moved to freelist, WRITEABLE is cleared
 127   **    (3)  Page X reused, WRITEABLE is set again
 128   ** If NEED_SYNC had been cleared in step 2, then it would not be reset
 129   ** in step 3, and page might be written into the database without first
 130   ** syncing the rollback journal, which might cause corruption on a power
 131   ** loss.
 132   **
 133   ** Another example is when the database page size is smaller than the
 134   ** disk sector size.  When any page of a sector is journalled, all pages
 135   ** in that sector are marked NEED_SYNC even if they are still CLEAN, just
 136   ** in case they are later modified, since all pages in the same sector
 137   ** must be journalled and synced before any of those pages can be safely
 138   ** written.
 139   */
 140   return 1;
 141 }
 142 #endif /* SQLITE_DEBUG */
 143
 144
 145 /********************************** Linked List Management ********************/
 146
 147 /* Allowed values for second argument to pcacheManageDirtyList() */
 148 #define PCACHE_DIRTYLIST_REMOVE   1    /* Remove pPage from dirty list */
 149 #define PCACHE_DIRTYLIST_ADD      2    /* Add pPage to the dirty list */
 150 #define PCACHE_DIRTYLIST_FRONT    3    /* Move pPage to the front of the list */
 151
 152 /*
 153 ** Manage pPage's participation on the dirty list.  Bits of the addRemove
 154 ** argument determines what operation to do.  The 0x01 bit means first
 155 ** remove pPage from the dirty list.  The 0x02 means add pPage back to
 156 ** the dirty list.  Doing both moves pPage to the front of the dirty list.
 157 */
 158 static void pcacheManageDirtyList(PgHdr *pPage, u8 addRemove){
 159   PCache *p = pPage->pCache;
 160
 161   pcacheTrace(("%p.DIRTYLIST.%s %d\n", p,
 162                 addRemove==1 ? "REMOVE" : addRemove==2 ? "ADD" : "FRONT",
 163                 pPage->pgno));
 164   if( addRemove & PCACHE_DIRTYLIST_REMOVE ){
 165     assert( pPage->pDirtyNext || pPage==p->pDirtyTail );
 166     assert( pPage->pDirtyPrev || pPage==p->pDirty );
 167
 168     /* Update the PCache1.pSynced variable if necessary. */
 169     if( p->pSynced==pPage ){
 170       p->pSynced = pPage->pDirtyPrev;
 171     }
 172
 173     if( pPage->pDirtyNext ){
 174       pPage->pDirtyNext->pDirtyPrev = pPage->pDirtyPrev;
 175     }else{
 176       assert( pPage==p->pDirtyTail );
 177       p->pDirtyTail = pPage->pDirtyPrev;
 178     }
 179     if( pPage->pDirtyPrev ){
 180       pPage->pDirtyPrev->pDirtyNext = pPage->pDirtyNext;
 181     }else{
 182       /* If there are now no dirty pages in the cache, set eCreate to 2.
 183       ** This is an optimization that allows sqlite3PcacheFetch() to skip
 184       ** searching for a dirty page to eject from the cache when it might
 185       ** otherwise have to.  */
 186       assert( pPage==p->pDirty );
 187       p->pDirty = pPage->pDirtyNext;
 188       assert( p->bPurgeable || p->eCreate==2 );
 189       if( p->pDirty==0 ){         /*OPTIMIZATION-IF-TRUE*/
 190         assert( p->bPurgeable==0 || p->eCreate==1 );
 191         p->eCreate = 2;
 192       }
 193     }
 194   }
 195   if( addRemove & PCACHE_DIRTYLIST_ADD ){
 196     pPage->pDirtyPrev = 0;
 197     pPage->pDirtyNext = p->pDirty;
 198     if( pPage->pDirtyNext ){
 199       assert( pPage->pDirtyNext->pDirtyPrev==0 );
 200       pPage->pDirtyNext->pDirtyPrev = pPage;
 201     }else{
 202       p->pDirtyTail = pPage;
 203       if( p->bPurgeable ){
 204         assert( p->eCreate==2 );
 205         p->eCreate = 1;
 206       }
 207     }
 208     p->pDirty = pPage;
 209
 210     /* If pSynced is NULL and this page has a clear NEED_SYNC flag, set
 211     ** pSynced to point to it. Checking the NEED_SYNC flag is an
 212     ** optimization, as if pSynced points to a page with the NEED_SYNC
 213     ** flag set sqlite3PcacheFetchStress() searches through all newer
 214     ** entries of the dirty-list for a page with NEED_SYNC clear anyway.  */
 215     if( !p->pSynced
 216      && 0==(pPage->flags&PGHDR_NEED_SYNC)   /*OPTIMIZATION-IF-FALSE*/
 217     ){
 218       p->pSynced = pPage;
 219     }
 220   }
 221   pcacheDump(p);
 222 }
 223
 224 /*
 225 ** Wrapper around the pluggable caches xUnpin method. If the cache is
 226 ** being used for an in-memory database, this function is a no-op.
 227 */
 228 static void pcacheUnpin(PgHdr *p){
 229   if( p->pCache->bPurgeable ){
 230     pcacheTrace(("%p.UNPIN %d\n", p->pCache, p->pgno));
 231     sqlite3GlobalConfig.pcache2.xUnpin(p->pCache->pCache, p->pPage, 0);
 232     pcacheDump(p->pCache);
 233   }
 234 }
 235
 236 /*
 237 ** Compute the number of pages of cache requested.   p->szCache is the
 238 ** cache size requested by the "PRAGMA cache_size" statement.
 239 */
 240 static int numberOfCachePages(PCache *p){
 241   if( p->szCache>=0 ){
 242     /* IMPLEMENTATION-OF: R-42059-47211 If the argument N is positive then the
 243     ** suggested cache size is set to N. */
 244     return p->szCache;
 245   }else{
 246     /* IMPLEMENTATION-OF: R-61436-13639 If the argument N is negative, then
 247     ** the number of cache pages is adjusted to use approximately abs(N*1024)
 248     ** bytes of memory. */
 249     return (int)((-1024*(i64)p->szCache)/(p->szPage+p->szExtra));
 250   }
 251 }
 252
 253 /*************************************************** General Interfaces ******
 254 **
 255 ** Initialize and shutdown the page cache subsystem. Neither of these
 256 ** functions are threadsafe.
 257 */
 258 int sqlite3PcacheInitialize(void){
 259   if( sqlite3GlobalConfig.pcache2.xInit==0 ){
 260     /* IMPLEMENTATION-OF: R-26801-64137 If the xInit() method is NULL, then the
 261     ** built-in default page cache is used instead of the application defined
 262     ** page cache. */
 263     sqlite3PCacheSetDefault();
 264   }
 265   return sqlite3GlobalConfig.pcache2.xInit(sqlite3GlobalConfig.pcache2.pArg);
 266 }
 267 void sqlite3PcacheShutdown(void){
 268   if( sqlite3GlobalConfig.pcache2.xShutdown ){
 269     /* IMPLEMENTATION-OF: R-26000-56589 The xShutdown() method may be NULL. */
 270     sqlite3GlobalConfig.pcache2.xShutdown(sqlite3GlobalConfig.pcache2.pArg);
 271   }
 272 }
 273
 274 /*
 275 ** Return the size in bytes of a PCache object.
 276 */
 277 int sqlite3PcacheSize(void){ return sizeof(PCache); }
 278
 279 /*
 280 ** Create a new PCache object. Storage space to hold the object
 281 ** has already been allocated and is passed in as the p pointer.
 282 ** The caller discovers how much space needs to be allocated by
 283 ** calling sqlite3PcacheSize().
 284 **
 285 ** szExtra is some extra space allocated for each page.  The first
 286 ** 8 bytes of the extra space will be zeroed as the page is allocated,
 287 ** but remaining content will be uninitialized.  Though it is opaque
 288 ** to this module, the extra space really ends up being the MemPage
 289 ** structure in the pager.
 290 */
 291 int sqlite3PcacheOpen(
 292   int szPage,                  /* Size of every page */
 293   int szExtra,                 /* Extra space associated with each page */
 294   int bPurgeable,              /* True if pages are on backing store */
 295   int (*xStress)(void*,PgHdr*),/* Call to try to make pages clean */
 296   void *pStress,               /* Argument to xStress */
 297   PCache *p                    /* Preallocated space for the PCache */
 298 ){
 299   memset(p, 0, sizeof(PCache));
 300   p->szPage = 1;
 301   p->szExtra = szExtra;
 302   assert( szExtra>=8 );  /* First 8 bytes will be zeroed */
 303   p->bPurgeable = bPurgeable;
 304   p->eCreate = 2;
 305   p->xStress = xStress;
 306   p->pStress = pStress;
 307   p->szCache = 100;
 308   p->szSpill = 1;
 309   pcacheTrace(("%p.OPEN szPage %d bPurgeable %d\n",p,szPage,bPurgeable));
 310   return sqlite3PcacheSetPageSize(p, szPage);
 311 }
 312
 313 /*
 314 ** Change the page size for PCache object. The caller must ensure that there
 315 ** are no outstanding page references when this function is called.
 316 */
 317 int sqlite3PcacheSetPageSize(PCache *pCache, int szPage){
 318   assert( pCache->nRefSum==0 && pCache->pDirty==0 );
 319   if( pCache->szPage ){
 320     sqlite3_pcache *pNew;
 321     pNew = sqlite3GlobalConfig.pcache2.xCreate(
 322                 szPage, pCache->szExtra + ROUND8(sizeof(PgHdr)),
 323                 pCache->bPurgeable
 324     );
 325     if( pNew==0 ) return SQLITE_NOMEM_BKPT;
 326     sqlite3GlobalConfig.pcache2.xCachesize(pNew, numberOfCachePages(pCache));
 327     if( pCache->pCache ){
 328       sqlite3GlobalConfig.pcache2.xDestroy(pCache->pCache);
 329     }
 330     pCache->pCache = pNew;
 331     pCache->szPage = szPage;
 332     pcacheTrace(("%p.PAGESIZE %d\n",pCache,szPage));
 333   }
 334   return SQLITE_OK;
 335 }
 336
 337 /*
 338 ** Try to obtain a page from the cache.
 339 **
 340 ** This routine returns a pointer to an sqlite3_pcache_page object if
 341 ** such an object is already in cache, or if a new one is created.
 342 ** This routine returns a NULL pointer if the object was not in cache
 343 ** and could not be created.
 344 **
 345 ** The createFlags should be 0 to check for existing pages and should
 346 ** be 3 (not 1, but 3) to try to create a new page.
 347 **
 348 ** If the createFlag is 0, then NULL is always returned if the page
 349 ** is not already in the cache.  If createFlag is 1, then a new page
 350 ** is created only if that can be done without spilling dirty pages
 351 ** and without exceeding the cache size limit.
 352 **
 353 ** The caller needs to invoke sqlite3PcacheFetchFinish() to properly
 354 ** initialize the sqlite3_pcache_page object and convert it into a
 355 ** PgHdr object.  The sqlite3PcacheFetch() and sqlite3PcacheFetchFinish()
 356 ** routines are split this way for performance reasons. When separated
 357 ** they can both (usually) operate without having to push values to
 358 ** the stack on entry and pop them back off on exit, which saves a
 359 ** lot of pushing and popping.
 360 */
 361 sqlite3_pcache_page *sqlite3PcacheFetch(
 362   PCache *pCache,       /* Obtain the page from this cache */
 363   Pgno pgno,            /* Page number to obtain */
 364   int createFlag        /* If true, create page if it does not exist already */
 365 ){
 366   int eCreate;
 367   sqlite3_pcache_page *pRes;
 368
 369   assert( pCache!=0 );
 370   assert( pCache->pCache!=0 );
 371   assert( createFlag==3 || createFlag==0 );
 372   assert( pCache->eCreate==((pCache->bPurgeable && pCache->pDirty) ? 1 : 2) );
 373
 374   /* eCreate defines what to do if the page does not exist.
 375   **    0     Do not allocate a new page.  (createFlag==0)
 376   **    1     Allocate a new page if doing so is inexpensive.
 377   **          (createFlag==1 AND bPurgeable AND pDirty)
 378   **    2     Allocate a new page even it doing so is difficult.
 379   **          (createFlag==1 AND !(bPurgeable AND pDirty)
 380   */
 381   eCreate = createFlag & pCache->eCreate;
 382   assert( eCreate==0 || eCreate==1 || eCreate==2 );
 383   assert( createFlag==0 || pCache->eCreate==eCreate );
 384   assert( createFlag==0 || eCreate==1+(!pCache->bPurgeable||!pCache->pDirty) );
 385   pRes = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache, pgno, eCreate);
 386   pcacheTrace(("%p.FETCH %d%s (result: %p)\n",pCache,pgno,
 387                createFlag?" create":"",pRes));
 388   return pRes;
 389 }
 390
 391 /*
 392 ** If the sqlite3PcacheFetch() routine is unable to allocate a new
 393 ** page because no clean pages are available for reuse and the cache
 394 ** size limit has been reached, then this routine can be invoked to
 395 ** try harder to allocate a page.  This routine might invoke the stress
 396 ** callback to spill dirty pages to the journal.  It will then try to
 397 ** allocate the new page and will only fail to allocate a new page on
 398 ** an OOM error.
 399 **
 400 ** This routine should be invoked only after sqlite3PcacheFetch() fails.
 401 */
 402 int sqlite3PcacheFetchStress(
 403   PCache *pCache,                 /* Obtain the page from this cache */
 404   Pgno pgno,                      /* Page number to obtain */
 405   sqlite3_pcache_page **ppPage    /* Write result here */
 406 ){
 407   PgHdr *pPg;
 408   if( pCache->eCreate==2 ) return 0;
 409
 410   if( sqlite3PcachePagecount(pCache)>pCache->szSpill ){
 411     /* Find a dirty page to write-out and recycle. First try to find a
 412     ** page that does not require a journal-sync (one with PGHDR_NEED_SYNC
 413     ** cleared), but if that is not possible settle for any other
 414     ** unreferenced dirty page.
 415     **
 416     ** If the LRU page in the dirty list that has a clear PGHDR_NEED_SYNC
 417     ** flag is currently referenced, then the following may leave pSynced
 418     ** set incorrectly (pointing to other than the LRU page with NEED_SYNC
 419     ** cleared). This is Ok, as pSynced is just an optimization.  */
 420     for(pPg=pCache->pSynced;
 421         pPg && (pPg->nRef || (pPg->flags&PGHDR_NEED_SYNC));
 422         pPg=pPg->pDirtyPrev
 423     );
 424     pCache->pSynced = pPg;
 425     if( !pPg ){
 426       for(pPg=pCache->pDirtyTail; pPg && pPg->nRef; pPg=pPg->pDirtyPrev);
 427     }
 428     if( pPg ){
 429       int rc;
 430 #ifdef SQLITE_LOG_CACHE_SPILL
 431       sqlite3_log(SQLITE_FULL,
 432                   "spill page %d making room for %d - cache used: %d/%d",
 433                   pPg->pgno, pgno,
 434                   sqlite3GlobalConfig.pcache.xPagecount(pCache->pCache),
 435                 numberOfCachePages(pCache));
 436 #endif
 437       pcacheTrace(("%p.SPILL %d\n",pCache,pPg->pgno));
 438       rc = pCache->xStress(pCache->pStress, pPg);
 439       pcacheDump(pCache);
 440       if( rc!=SQLITE_OK && rc!=SQLITE_BUSY ){
 441         return rc;
 442       }
 443     }
 444   }
 445   *ppPage = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache, pgno, 2);
 446   return *ppPage==0 ? SQLITE_NOMEM_BKPT : SQLITE_OK;
 447 }
 448
 449 /*
 450 ** This is a helper routine for sqlite3PcacheFetchFinish()
 451 **
 452 ** In the uncommon case where the page being fetched has not been
 453 ** initialized, this routine is invoked to do the initialization.
 454 ** This routine is broken out into a separate function since it
 455 ** requires extra stack manipulation that can be avoided in the common
 456 ** case.
 457 */
 458 static SQLITE_NOINLINE PgHdr *pcacheFetchFinishWithInit(
 459   PCache *pCache,             /* Obtain the page from this cache */
 460   Pgno pgno,                  /* Page number obtained */
 461   sqlite3_pcache_page *pPage  /* Page obtained by prior PcacheFetch() call */
 462 ){
 463   PgHdr *pPgHdr;
 464   assert( pPage!=0 );
 465   pPgHdr = (PgHdr*)pPage->pExtra;
 466   assert( pPgHdr->pPage==0 );
 467   memset(&pPgHdr->pDirty, 0, sizeof(PgHdr) - offsetof(PgHdr,pDirty));
 468   pPgHdr->pPage = pPage;
 469   pPgHdr->pData = pPage->pBuf;
 470   pPgHdr->pExtra = (void *)&pPgHdr[1];
 471   memset(pPgHdr->pExtra, 0, 8);
 472   pPgHdr->pCache = pCache;
 473   pPgHdr->pgno = pgno;
 474   pPgHdr->flags = PGHDR_CLEAN;
 475   return sqlite3PcacheFetchFinish(pCache,pgno,pPage);
 476 }
 477
 478 /*
 479 ** This routine converts the sqlite3_pcache_page object returned by
 480 ** sqlite3PcacheFetch() into an initialized PgHdr object.  This routine
 481 ** must be called after sqlite3PcacheFetch() in order to get a usable
 482 ** result.
 483 */
 484 PgHdr *sqlite3PcacheFetchFinish(
 485   PCache *pCache,             /* Obtain the page from this cache */
 486   Pgno pgno,                  /* Page number obtained */
 487   sqlite3_pcache_page *pPage  /* Page obtained by prior PcacheFetch() call */
 488 ){
 489   PgHdr *pPgHdr;
 490
 491   assert( pPage!=0 );
 492   pPgHdr = (PgHdr *)pPage->pExtra;
 493
 494   if( !pPgHdr->pPage ){
 495     return pcacheFetchFinishWithInit(pCache, pgno, pPage);
 496   }
 497   pCache->nRefSum++;
 498   pPgHdr->nRef++;
 499   assert( sqlite3PcachePageSanity(pPgHdr) );
 500   return pPgHdr;
 501 }
 502
 503 /*
 504 ** Decrement the reference count on a page. If the page is clean and the
 505 ** reference count drops to 0, then it is made eligible for recycling.
 506 */
 507 void SQLITE_NOINLINE sqlite3PcacheRelease(PgHdr *p){
 508   assert( p->nRef>0 );
 509   p->pCache->nRefSum--;
 510   if( (--p->nRef)==0 ){
 511     if( p->flags&PGHDR_CLEAN ){
 512       pcacheUnpin(p);
 513     }else{
 514       pcacheManageDirtyList(p, PCACHE_DIRTYLIST_FRONT);
 515     }
 516   }
 517 }
 518
 519 /*
 520 ** Increase the reference count of a supplied page by 1.
 521 */
 522 void sqlite3PcacheRef(PgHdr *p){
 523   assert(p->nRef>0);
 524   assert( sqlite3PcachePageSanity(p) );
 525   p->nRef++;
 526   p->pCache->nRefSum++;
 527 }
 528
 529 /*
 530 ** Drop a page from the cache. There must be exactly one reference to the
 531 ** page. This function deletes that reference, so after it returns the
 532 ** page pointed to by p is invalid.
 533 */
 534 void sqlite3PcacheDrop(PgHdr *p){
 535   assert( p->nRef==1 );
 536   assert( sqlite3PcachePageSanity(p) );
 537   if( p->flags&PGHDR_DIRTY ){
 538     pcacheManageDirtyList(p, PCACHE_DIRTYLIST_REMOVE);
 539   }
 540   p->pCache->nRefSum--;
 541   sqlite3GlobalConfig.pcache2.xUnpin(p->pCache->pCache, p->pPage, 1);
 542 }
 543
 544 /*
 545 ** Make sure the page is marked as dirty. If it isn't dirty already,
 546 ** make it so.
 547 */
 548 void sqlite3PcacheMakeDirty(PgHdr *p){
 549   assert( p->nRef>0 );
 550   assert( sqlite3PcachePageSanity(p) );
 551   if( p->flags & (PGHDR_CLEAN|PGHDR_DONT_WRITE) ){    /*OPTIMIZATION-IF-FALSE*/
 552     p->flags &= ~PGHDR_DONT_WRITE;
 553     if( p->flags & PGHDR_CLEAN ){
 554       p->flags ^= (PGHDR_DIRTY|PGHDR_CLEAN);
 555       pcacheTrace(("%p.DIRTY %d\n",p->pCache,p->pgno));
 556       assert( (p->flags & (PGHDR_DIRTY|PGHDR_CLEAN))==PGHDR_DIRTY );
 557       pcacheManageDirtyList(p, PCACHE_DIRTYLIST_ADD);
 558     }
 559     assert( sqlite3PcachePageSanity(p) );
 560   }
 561 }
 562
 563 /*
 564 ** Make sure the page is marked as clean. If it isn't clean already,
 565 ** make it so.
 566 */
 567 void sqlite3PcacheMakeClean(PgHdr *p){
 568   assert( sqlite3PcachePageSanity(p) );
 569   assert( (p->flags & PGHDR_DIRTY)!=0 );
 570   assert( (p->flags & PGHDR_CLEAN)==0 );
 571   pcacheManageDirtyList(p, PCACHE_DIRTYLIST_REMOVE);
 572   p->flags &= ~(PGHDR_DIRTY|PGHDR_NEED_SYNC|PGHDR_WRITEABLE);
 573   p->flags |= PGHDR_CLEAN;
 574   pcacheTrace(("%p.CLEAN %d\n",p->pCache,p->pgno));
 575   assert( sqlite3PcachePageSanity(p) );
 576   if( p->nRef==0 ){
 577     pcacheUnpin(p);
 578   }
 579 }
 580
 581 /*
 582 ** Make every page in the cache clean.
 583 */
 584 void sqlite3PcacheCleanAll(PCache *pCache){
 585   PgHdr *p;
 586   pcacheTrace(("%p.CLEAN-ALL\n",pCache));
 587   while( (p = pCache->pDirty)!=0 ){
 588     sqlite3PcacheMakeClean(p);
 589   }
 590 }
 591
 592 /*
 593 ** Clear the PGHDR_NEED_SYNC and PGHDR_WRITEABLE flag from all dirty pages.
 594 */
 595 void sqlite3PcacheClearWritable(PCache *pCache){
 596   PgHdr *p;
 597   pcacheTrace(("%p.CLEAR-WRITEABLE\n",pCache));
 598   for(p=pCache->pDirty; p; p=p->pDirtyNext){
 599     p->flags &= ~(PGHDR_NEED_SYNC|PGHDR_WRITEABLE);
 600   }
 601   pCache->pSynced = pCache->pDirtyTail;
 602 }
 603
 604 /*
 605 ** Clear the PGHDR_NEED_SYNC flag from all dirty pages.
 606 */
 607 void sqlite3PcacheClearSyncFlags(PCache *pCache){
 608   PgHdr *p;
 609   for(p=pCache->pDirty; p; p=p->pDirtyNext){
 610     p->flags &= ~PGHDR_NEED_SYNC;
 611   }
 612   pCache->pSynced = pCache->pDirtyTail;
 613 }
 614
 615 /*
 616 ** Change the page number of page p to newPgno.
 617 */
 618 void sqlite3PcacheMove(PgHdr *p, Pgno newPgno){
 619   PCache *pCache = p->pCache;
 620   assert( p->nRef>0 );
 621   assert( newPgno>0 );
 622   assert( sqlite3PcachePageSanity(p) );
 623   pcacheTrace(("%p.MOVE %d -> %d\n",pCache,p->pgno,newPgno));
 624   sqlite3GlobalConfig.pcache2.xRekey(pCache->pCache, p->pPage, p->pgno,newPgno);
 625   p->pgno = newPgno;
 626   if( (p->flags&PGHDR_DIRTY) && (p->flags&PGHDR_NEED_SYNC) ){
 627     pcacheManageDirtyList(p, PCACHE_DIRTYLIST_FRONT);
 628   }
 629 }
 630
 631 /*
 632 ** Drop every cache entry whose page number is greater than "pgno". The
 633 ** caller must ensure that there are no outstanding references to any pages
 634 ** other than page 1 with a page number greater than pgno.
 635 **
 636 ** If there is a reference to page 1 and the pgno parameter passed to this
 637 ** function is 0, then the data area associated with page 1 is zeroed, but
 638 ** the page object is not dropped.
 639 */
 640 void sqlite3PcacheTruncate(PCache *pCache, Pgno pgno){
 641   if( pCache->pCache ){
 642     PgHdr *p;
 643     PgHdr *pNext;
 644     pcacheTrace(("%p.TRUNCATE %d\n",pCache,pgno));
 645     for(p=pCache->pDirty; p; p=pNext){
 646       pNext = p->pDirtyNext;
 647       /* This routine never gets call with a positive pgno except right
 648       ** after sqlite3PcacheCleanAll().  So if there are dirty pages,
 649       ** it must be that pgno==0.
 650       */
 651       assert( p->pgno>0 );
 652       if( p->pgno>pgno ){
 653         assert( p->flags&PGHDR_DIRTY );
 654         sqlite3PcacheMakeClean(p);
 655       }
 656     }
 657     if( pgno==0 && pCache->nRefSum ){
 658       sqlite3_pcache_page *pPage1;
 659       pPage1 = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache,1,0);
 660       if( ALWAYS(pPage1) ){  /* Page 1 is always available in cache, because
 661                              ** pCache->nRefSum>0 */
 662         memset(pPage1->pBuf, 0, pCache->szPage);
 663         pgno = 1;
 664       }
 665     }
 666     sqlite3GlobalConfig.pcache2.xTruncate(pCache->pCache, pgno+1);
 667   }
 668 }
 669
 670 /*
 671 ** Close a cache.
 672 */
 673 void sqlite3PcacheClose(PCache *pCache){
 674   assert( pCache->pCache!=0 );
 675   pcacheTrace(("%p.CLOSE\n",pCache));
 676   sqlite3GlobalConfig.pcache2.xDestroy(pCache->pCache);
 677 }
 678
 679 /*
 680 ** Discard the contents of the cache.
 681 */
 682 void sqlite3PcacheClear(PCache *pCache){
 683   sqlite3PcacheTruncate(pCache, 0);
 684 }
 685
 686 /*
 687 ** Merge two lists of pages connected by pDirty and in pgno order.
 688 ** Do not bother fixing the pDirtyPrev pointers.
 689 */
 690 static PgHdr *pcacheMergeDirtyList(PgHdr *pA, PgHdr *pB){
 691   PgHdr result, *pTail;
 692   pTail = &result;
 693   assert( pA!=0 && pB!=0 );
 694   for(;;){
 695     if( pA->pgno<pB->pgno ){
 696       pTail->pDirty = pA;
 697       pTail = pA;
 698       pA = pA->pDirty;
 699       if( pA==0 ){
 700         pTail->pDirty = pB;
 701         break;
 702       }
 703     }else{
 704       pTail->pDirty = pB;
 705       pTail = pB;
 706       pB = pB->pDirty;
 707       if( pB==0 ){
 708         pTail->pDirty = pA;
 709         break;
 710       }
 711     }
 712   }
 713   return result.pDirty;
 714 }
 715
 716 /*
 717 ** Sort the list of pages in accending order by pgno.  Pages are
 718 ** connected by pDirty pointers.  The pDirtyPrev pointers are
 719 ** corrupted by this sort.
 720 **
 721 ** Since there cannot be more than 2^31 distinct pages in a database,
 722 ** there cannot be more than 31 buckets required by the merge sorter.
 723 ** One extra bucket is added to catch overflow in case something
 724 ** ever changes to make the previous sentence incorrect.
 725 */
 726 #define N_SORT_BUCKET  32
 727 static PgHdr *pcacheSortDirtyList(PgHdr *pIn){
 728   PgHdr *a[N_SORT_BUCKET], *p;
 729   int i;
 730   memset(a, 0, sizeof(a));
 731   while( pIn ){
 732     p = pIn;
 733     pIn = p->pDirty;
 734     p->pDirty = 0;
 735     for(i=0; ALWAYS(i<N_SORT_BUCKET-1); i++){
 736       if( a[i]==0 ){
 737         a[i] = p;
 738         break;
 739       }else{
 740         p = pcacheMergeDirtyList(a[i], p);
 741         a[i] = 0;
 742       }
 743     }
 744     if( NEVER(i==N_SORT_BUCKET-1) ){
 745       /* To get here, there need to be 2^(N_SORT_BUCKET) elements in
 746       ** the input list.  But that is impossible.
 747       */
 748       a[i] = pcacheMergeDirtyList(a[i], p);
 749     }
 750   }
 751   p = a[0];
 752   for(i=1; i<N_SORT_BUCKET; i++){
 753     if( a[i]==0 ) continue;
 754     p = p ? pcacheMergeDirtyList(p, a[i]) : a[i];
 755   }
 756   return p;
 757 }
 758
 759 /*
 760 ** Return a list of all dirty pages in the cache, sorted by page number.
 761 */
 762 PgHdr *sqlite3PcacheDirtyList(PCache *pCache){
 763   PgHdr *p;
 764   for(p=pCache->pDirty; p; p=p->pDirtyNext){
 765     p->pDirty = p->pDirtyNext;
 766   }
 767   return pcacheSortDirtyList(pCache->pDirty);
 768 }
 769
 770 /*
 771 ** Return the total number of references to all pages held by the cache.
 772 **
 773 ** This is not the total number of pages referenced, but the sum of the
 774 ** reference count for all pages.
 775 */
 776 int sqlite3PcacheRefCount(PCache *pCache){
 777   return pCache->nRefSum;
 778 }
 779
 780 /*
 781 ** Return the number of references to the page supplied as an argument.
 782 */
 783 int sqlite3PcachePageRefcount(PgHdr *p){
 784   return p->nRef;
 785 }
 786
 787 /*
 788 ** Return the total number of pages in the cache.
 789 */
 790 int sqlite3PcachePagecount(PCache *pCache){
 791   assert( pCache->pCache!=0 );
 792   return sqlite3GlobalConfig.pcache2.xPagecount(pCache->pCache);
 793 }
 794
 795 #ifdef SQLITE_TEST
 796 /*
 797 ** Get the suggested cache-size value.
 798 */
 799 int sqlite3PcacheGetCachesize(PCache *pCache){
 800   return numberOfCachePages(pCache);
 801 }
 802 #endif
 803
 804 /*
 805 ** Set the suggested cache-size value.
 806 */
 807 void sqlite3PcacheSetCachesize(PCache *pCache, int mxPage){
 808   assert( pCache->pCache!=0 );
 809   pCache->szCache = mxPage;
 810   sqlite3GlobalConfig.pcache2.xCachesize(pCache->pCache,
 811                                          numberOfCachePages(pCache));
 812 }
 813
 814 /*
 815 ** Set the suggested cache-spill value.  Make no changes if if the
 816 ** argument is zero.  Return the effective cache-spill size, which will
 817 ** be the larger of the szSpill and szCache.
 818 */
 819 int sqlite3PcacheSetSpillsize(PCache *p, int mxPage){
 820   int res;
 821   assert( p->pCache!=0 );
 822   if( mxPage ){
 823     if( mxPage<0 ){
 824       mxPage = (int)((-1024*(i64)mxPage)/(p->szPage+p->szExtra));
 825     }
 826     p->szSpill = mxPage;
 827   }
 828   res = numberOfCachePages(p);
 829   if( res<p->szSpill ) res = p->szSpill;
 830   return res;
 831 }
 832
 833 /*
 834 ** Free up as much memory as possible from the page cache.
 835 */
 836 void sqlite3PcacheShrink(PCache *pCache){
 837   assert( pCache->pCache!=0 );
 838   sqlite3GlobalConfig.pcache2.xShrink(pCache->pCache);
 839 }
 840
 841 /*
 842 ** Return the size of the header added by this middleware layer
 843 ** in the page-cache hierarchy.
 844 */
 845 int sqlite3HeaderSizePcache(void){ return ROUND8(sizeof(PgHdr)); }
 846
 847 /*
 848 ** Return the number of dirty pages currently in the cache, as a percentage
 849 ** of the configured cache size.
 850 */
 851 int sqlite3PCachePercentDirty(PCache *pCache){
 852   PgHdr *pDirty;
 853   int nDirty = 0;
 854   int nCache = numberOfCachePages(pCache);
 855   for(pDirty=pCache->pDirty; pDirty; pDirty=pDirty->pDirtyNext) nDirty++;
 856   return nCache ? (int)(((i64)nDirty * 100) / nCache) : 0;
 857 }
 858
 859 #if defined(SQLITE_CHECK_PAGES) || defined(SQLITE_DEBUG)
 860 /*
 861 ** For all dirty pages currently in the cache, invoke the specified
 862 ** callback. This is only used if the SQLITE_CHECK_PAGES macro is
 863 ** defined.
 864 */
 865 void sqlite3PcacheIterateDirty(PCache *pCache, void (*xIter)(PgHdr *)){
 866   PgHdr *pDirty;
 867   for(pDirty=pCache->pDirty; pDirty; pDirty=pDirty->pDirtyNext){
 868     xIter(pDirty);
 869   }
 870 }
 871 #endif