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   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     i64 n;
 247     /* IMPLEMANTATION-OF: R-59858-46238 If the argument N is negative, then the
 248     ** number of cache pages is adjusted to be a number of pages that would
 249     ** use approximately abs(N*1024) bytes of memory based on the current
 250     ** page size. */
 251     n = ((-1024*(i64)p->szCache)/(p->szPage+p->szExtra));
 252     if( n>1000000000 ) n = 1000000000;
 253     return (int)n;
 254   }
 255 }
 256
 257 /*************************************************** General Interfaces ******
 258 **
 259 ** Initialize and shutdown the page cache subsystem. Neither of these
 260 ** functions are threadsafe.
 261 */
 262 int sqlite3PcacheInitialize(void){
 263   if( sqlite3GlobalConfig.pcache2.xInit==0 ){
 264     /* IMPLEMENTATION-OF: R-26801-64137 If the xInit() method is NULL, then the
 265     ** built-in default page cache is used instead of the application defined
 266     ** page cache. */
 267     sqlite3PCacheSetDefault();
 268     assert( sqlite3GlobalConfig.pcache2.xInit!=0 );
 269   }
 270   return sqlite3GlobalConfig.pcache2.xInit(sqlite3GlobalConfig.pcache2.pArg);
 271 }
 272 void sqlite3PcacheShutdown(void){
 273   if( sqlite3GlobalConfig.pcache2.xShutdown ){
 274     /* IMPLEMENTATION-OF: R-26000-56589 The xShutdown() method may be NULL. */
 275     sqlite3GlobalConfig.pcache2.xShutdown(sqlite3GlobalConfig.pcache2.pArg);
 276   }
 277 }
 278
 279 /*
 280 ** Return the size in bytes of a PCache object.
 281 */
 282 int sqlite3PcacheSize(void){ return sizeof(PCache); }
 283
 284 /*
 285 ** Create a new PCache object. Storage space to hold the object
 286 ** has already been allocated and is passed in as the p pointer.
 287 ** The caller discovers how much space needs to be allocated by
 288 ** calling sqlite3PcacheSize().
 289 **
 290 ** szExtra is some extra space allocated for each page.  The first
 291 ** 8 bytes of the extra space will be zeroed as the page is allocated,
 292 ** but remaining content will be uninitialized.  Though it is opaque
 293 ** to this module, the extra space really ends up being the MemPage
 294 ** structure in the pager.
 295 */
 296 int sqlite3PcacheOpen(
 297   int szPage,                  /* Size of every page */
 298   int szExtra,                 /* Extra space associated with each page */
 299   int bPurgeable,              /* True if pages are on backing store */
 300   int (*xStress)(void*,PgHdr*),/* Call to try to make pages clean */
 301   void *pStress,               /* Argument to xStress */
 302   PCache *p                    /* Preallocated space for the PCache */
 303 ){
 304   memset(p, 0, sizeof(PCache));
 305   p->szPage = 1;
 306   p->szExtra = szExtra;
 307   assert( szExtra>=8 );  /* First 8 bytes will be zeroed */
 308   p->bPurgeable = bPurgeable;
 309   p->eCreate = 2;
 310   p->xStress = xStress;
 311   p->pStress = pStress;
 312   p->szCache = 100;
 313   p->szSpill = 1;
 314   pcacheTrace(("%p.OPEN szPage %d bPurgeable %d\n",p,szPage,bPurgeable));
 315   return sqlite3PcacheSetPageSize(p, szPage);
 316 }
 317
 318 /*
 319 ** Change the page size for PCache object. The caller must ensure that there
 320 ** are no outstanding page references when this function is called.
 321 */
 322 int sqlite3PcacheSetPageSize(PCache *pCache, int szPage){
 323   assert( pCache->nRefSum==0 && pCache->pDirty==0 );
 324   if( pCache->szPage ){
 325     sqlite3_pcache *pNew;
 326     pNew = sqlite3GlobalConfig.pcache2.xCreate(
 327                 szPage, pCache->szExtra + ROUND8(sizeof(PgHdr)),
 328                 pCache->bPurgeable
 329     );
 330     if( pNew==0 ) return SQLITE_NOMEM_BKPT;
 331     sqlite3GlobalConfig.pcache2.xCachesize(pNew, numberOfCachePages(pCache));
 332     if( pCache->pCache ){
 333       sqlite3GlobalConfig.pcache2.xDestroy(pCache->pCache);
 334     }
 335     pCache->pCache = pNew;
 336     pCache->szPage = szPage;
 337     pcacheTrace(("%p.PAGESIZE %d\n",pCache,szPage));
 338   }
 339   return SQLITE_OK;
 340 }
 341
 342 /*
 343 ** Try to obtain a page from the cache.
 344 **
 345 ** This routine returns a pointer to an sqlite3_pcache_page object if
 346 ** such an object is already in cache, or if a new one is created.
 347 ** This routine returns a NULL pointer if the object was not in cache
 348 ** and could not be created.
 349 **
 350 ** The createFlags should be 0 to check for existing pages and should
 351 ** be 3 (not 1, but 3) to try to create a new page.
 352 **
 353 ** If the createFlag is 0, then NULL is always returned if the page
 354 ** is not already in the cache.  If createFlag is 1, then a new page
 355 ** is created only if that can be done without spilling dirty pages
 356 ** and without exceeding the cache size limit.
 357 **
 358 ** The caller needs to invoke sqlite3PcacheFetchFinish() to properly
 359 ** initialize the sqlite3_pcache_page object and convert it into a
 360 ** PgHdr object.  The sqlite3PcacheFetch() and sqlite3PcacheFetchFinish()
 361 ** routines are split this way for performance reasons. When separated
 362 ** they can both (usually) operate without having to push values to
 363 ** the stack on entry and pop them back off on exit, which saves a
 364 ** lot of pushing and popping.
 365 */
 366 sqlite3_pcache_page *sqlite3PcacheFetch(
 367   PCache *pCache,       /* Obtain the page from this cache */
 368   Pgno pgno,            /* Page number to obtain */
 369   int createFlag        /* If true, create page if it does not exist already */
 370 ){
 371   int eCreate;
 372   sqlite3_pcache_page *pRes;
 373
 374   assert( pCache!=0 );
 375   assert( pCache->pCache!=0 );
 376   assert( createFlag==3 || createFlag==0 );
 377   assert( pCache->eCreate==((pCache->bPurgeable && pCache->pDirty) ? 1 : 2) );
 378
 379   /* eCreate defines what to do if the page does not exist.
 380   **    0     Do not allocate a new page.  (createFlag==0)
 381   **    1     Allocate a new page if doing so is inexpensive.
 382   **          (createFlag==1 AND bPurgeable AND pDirty)
 383   **    2     Allocate a new page even it doing so is difficult.
 384   **          (createFlag==1 AND !(bPurgeable AND pDirty)
 385   */
 386   eCreate = createFlag & pCache->eCreate;
 387   assert( eCreate==0 || eCreate==1 || eCreate==2 );
 388   assert( createFlag==0 || pCache->eCreate==eCreate );
 389   assert( createFlag==0 || eCreate==1+(!pCache->bPurgeable||!pCache->pDirty) );
 390   pRes = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache, pgno, eCreate);
 391   pcacheTrace(("%p.FETCH %d%s (result: %p)\n",pCache,pgno,
 392                createFlag?" create":"",pRes));
 393   return pRes;
 394 }
 395
 396 /*
 397 ** If the sqlite3PcacheFetch() routine is unable to allocate a new
 398 ** page because no clean pages are available for reuse and the cache
 399 ** size limit has been reached, then this routine can be invoked to
 400 ** try harder to allocate a page.  This routine might invoke the stress
 401 ** callback to spill dirty pages to the journal.  It will then try to
 402 ** allocate the new page and will only fail to allocate a new page on
 403 ** an OOM error.
 404 **
 405 ** This routine should be invoked only after sqlite3PcacheFetch() fails.
 406 */
 407 int sqlite3PcacheFetchStress(
 408   PCache *pCache,                 /* Obtain the page from this cache */
 409   Pgno pgno,                      /* Page number to obtain */
 410   sqlite3_pcache_page **ppPage    /* Write result here */
 411 ){
 412   PgHdr *pPg;
 413   if( pCache->eCreate==2 ) return 0;
 414
 415   if( sqlite3PcachePagecount(pCache)>pCache->szSpill ){
 416     /* Find a dirty page to write-out and recycle. First try to find a
 417     ** page that does not require a journal-sync (one with PGHDR_NEED_SYNC
 418     ** cleared), but if that is not possible settle for any other
 419     ** unreferenced dirty page.
 420     **
 421     ** If the LRU page in the dirty list that has a clear PGHDR_NEED_SYNC
 422     ** flag is currently referenced, then the following may leave pSynced
 423     ** set incorrectly (pointing to other than the LRU page with NEED_SYNC
 424     ** cleared). This is Ok, as pSynced is just an optimization.  */
 425     for(pPg=pCache->pSynced;
 426         pPg && (pPg->nRef || (pPg->flags&PGHDR_NEED_SYNC));
 427         pPg=pPg->pDirtyPrev
 428     );
 429     pCache->pSynced = pPg;
 430     if( !pPg ){
 431       for(pPg=pCache->pDirtyTail; pPg && pPg->nRef; pPg=pPg->pDirtyPrev);
 432     }
 433     if( pPg ){
 434       int rc;
 435 #ifdef SQLITE_LOG_CACHE_SPILL
 436       sqlite3_log(SQLITE_FULL,
 437                   "spill page %d making room for %d - cache used: %d/%d",
 438                   pPg->pgno, pgno,
 439                   sqlite3GlobalConfig.pcache2.xPagecount(pCache->pCache),
 440                 numberOfCachePages(pCache));
 441 #endif
 442       pcacheTrace(("%p.SPILL %d\n",pCache,pPg->pgno));
 443       rc = pCache->xStress(pCache->pStress, pPg);
 444       pcacheDump(pCache);
 445       if( rc!=SQLITE_OK && rc!=SQLITE_BUSY ){
 446         return rc;
 447       }
 448     }
 449   }
 450   *ppPage = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache, pgno, 2);
 451   return *ppPage==0 ? SQLITE_NOMEM_BKPT : SQLITE_OK;
 452 }
 453
 454 /*
 455 ** This is a helper routine for sqlite3PcacheFetchFinish()
 456 **
 457 ** In the uncommon case where the page being fetched has not been
 458 ** initialized, this routine is invoked to do the initialization.
 459 ** This routine is broken out into a separate function since it
 460 ** requires extra stack manipulation that can be avoided in the common
 461 ** case.
 462 */
 463 static SQLITE_NOINLINE PgHdr *pcacheFetchFinishWithInit(
 464   PCache *pCache,             /* Obtain the page from this cache */
 465   Pgno pgno,                  /* Page number obtained */
 466   sqlite3_pcache_page *pPage  /* Page obtained by prior PcacheFetch() call */
 467 ){
 468   PgHdr *pPgHdr;
 469   assert( pPage!=0 );
 470   pPgHdr = (PgHdr*)pPage->pExtra;
 471   assert( pPgHdr->pPage==0 );
 472   memset(&pPgHdr->pDirty, 0, sizeof(PgHdr) - offsetof(PgHdr,pDirty));
 473   pPgHdr->pPage = pPage;
 474   pPgHdr->pData = pPage->pBuf;
 475   pPgHdr->pExtra = (void *)&pPgHdr[1];
 476   memset(pPgHdr->pExtra, 0, 8);
 477   pPgHdr->pCache = pCache;
 478   pPgHdr->pgno = pgno;
 479   pPgHdr->flags = PGHDR_CLEAN;
 480   return sqlite3PcacheFetchFinish(pCache,pgno,pPage);
 481 }
 482
 483 /*
 484 ** This routine converts the sqlite3_pcache_page object returned by
 485 ** sqlite3PcacheFetch() into an initialized PgHdr object.  This routine
 486 ** must be called after sqlite3PcacheFetch() in order to get a usable
 487 ** result.
 488 */
 489 PgHdr *sqlite3PcacheFetchFinish(
 490   PCache *pCache,             /* Obtain the page from this cache */
 491   Pgno pgno,                  /* Page number obtained */
 492   sqlite3_pcache_page *pPage  /* Page obtained by prior PcacheFetch() call */
 493 ){
 494   PgHdr *pPgHdr;
 495
 496   assert( pPage!=0 );
 497   pPgHdr = (PgHdr *)pPage->pExtra;
 498
 499   if( !pPgHdr->pPage ){
 500     return pcacheFetchFinishWithInit(pCache, pgno, pPage);
 501   }
 502   pCache->nRefSum++;
 503   pPgHdr->nRef++;
 504   assert( sqlite3PcachePageSanity(pPgHdr) );
 505   return pPgHdr;
 506 }
 507
 508 /*
 509 ** Decrement the reference count on a page. If the page is clean and the
 510 ** reference count drops to 0, then it is made eligible for recycling.
 511 */
 512 void SQLITE_NOINLINE sqlite3PcacheRelease(PgHdr *p){
 513   assert( p->nRef>0 );
 514   p->pCache->nRefSum--;
 515   if( (--p->nRef)==0 ){
 516     if( p->flags&PGHDR_CLEAN ){
 517       pcacheUnpin(p);
 518     }else{
 519       pcacheManageDirtyList(p, PCACHE_DIRTYLIST_FRONT);
 520     }
 521   }
 522 }
 523
 524 /*
 525 ** Increase the reference count of a supplied page by 1.
 526 */
 527 void sqlite3PcacheRef(PgHdr *p){
 528   assert(p->nRef>0);
 529   assert( sqlite3PcachePageSanity(p) );
 530   p->nRef++;
 531   p->pCache->nRefSum++;
 532 }
 533
 534 /*
 535 ** Drop a page from the cache. There must be exactly one reference to the
 536 ** page. This function deletes that reference, so after it returns the
 537 ** page pointed to by p is invalid.
 538 */
 539 void sqlite3PcacheDrop(PgHdr *p){
 540   assert( p->nRef==1 );
 541   assert( sqlite3PcachePageSanity(p) );
 542   if( p->flags&PGHDR_DIRTY ){
 543     pcacheManageDirtyList(p, PCACHE_DIRTYLIST_REMOVE);
 544   }
 545   p->pCache->nRefSum--;
 546   sqlite3GlobalConfig.pcache2.xUnpin(p->pCache->pCache, p->pPage, 1);
 547 }
 548
 549 /*
 550 ** Make sure the page is marked as dirty. If it isn't dirty already,
 551 ** make it so.
 552 */
 553 void sqlite3PcacheMakeDirty(PgHdr *p){
 554   assert( p->nRef>0 );
 555   assert( sqlite3PcachePageSanity(p) );
 556   if( p->flags & (PGHDR_CLEAN|PGHDR_DONT_WRITE) ){    /*OPTIMIZATION-IF-FALSE*/
 557     p->flags &= ~PGHDR_DONT_WRITE;
 558     if( p->flags & PGHDR_CLEAN ){
 559       p->flags ^= (PGHDR_DIRTY|PGHDR_CLEAN);
 560       pcacheTrace(("%p.DIRTY %d\n",p->pCache,p->pgno));
 561       assert( (p->flags & (PGHDR_DIRTY|PGHDR_CLEAN))==PGHDR_DIRTY );
 562       pcacheManageDirtyList(p, PCACHE_DIRTYLIST_ADD);
 563     }
 564     assert( sqlite3PcachePageSanity(p) );
 565   }
 566 }
 567
 568 /*
 569 ** Make sure the page is marked as clean. If it isn't clean already,
 570 ** make it so.
 571 */
 572 void sqlite3PcacheMakeClean(PgHdr *p){
 573   assert( sqlite3PcachePageSanity(p) );
 574   assert( (p->flags & PGHDR_DIRTY)!=0 );
 575   assert( (p->flags & PGHDR_CLEAN)==0 );
 576   pcacheManageDirtyList(p, PCACHE_DIRTYLIST_REMOVE);
 577   p->flags &= ~(PGHDR_DIRTY|PGHDR_NEED_SYNC|PGHDR_WRITEABLE);
 578   p->flags |= PGHDR_CLEAN;
 579   pcacheTrace(("%p.CLEAN %d\n",p->pCache,p->pgno));
 580   assert( sqlite3PcachePageSanity(p) );
 581   if( p->nRef==0 ){
 582     pcacheUnpin(p);
 583   }
 584 }
 585
 586 /*
 587 ** Make every page in the cache clean.
 588 */
 589 void sqlite3PcacheCleanAll(PCache *pCache){
 590   PgHdr *p;
 591   pcacheTrace(("%p.CLEAN-ALL\n",pCache));
 592   while( (p = pCache->pDirty)!=0 ){
 593     sqlite3PcacheMakeClean(p);
 594   }
 595 }
 596
 597 /*
 598 ** Clear the PGHDR_NEED_SYNC and PGHDR_WRITEABLE flag from all dirty pages.
 599 */
 600 void sqlite3PcacheClearWritable(PCache *pCache){
 601   PgHdr *p;
 602   pcacheTrace(("%p.CLEAR-WRITEABLE\n",pCache));
 603   for(p=pCache->pDirty; p; p=p->pDirtyNext){
 604     p->flags &= ~(PGHDR_NEED_SYNC|PGHDR_WRITEABLE);
 605   }
 606   pCache->pSynced = pCache->pDirtyTail;
 607 }
 608
 609 /*
 610 ** Clear the PGHDR_NEED_SYNC flag from all dirty pages.
 611 */
 612 void sqlite3PcacheClearSyncFlags(PCache *pCache){
 613   PgHdr *p;
 614   for(p=pCache->pDirty; p; p=p->pDirtyNext){
 615     p->flags &= ~PGHDR_NEED_SYNC;
 616   }
 617   pCache->pSynced = pCache->pDirtyTail;
 618 }
 619
 620 /*
 621 ** Change the page number of page p to newPgno.
 622 */
 623 void sqlite3PcacheMove(PgHdr *p, Pgno newPgno){
 624   PCache *pCache = p->pCache;
 625   sqlite3_pcache_page *pOther;
 626   assert( p->nRef>0 );
 627   assert( newPgno>0 );
 628   assert( sqlite3PcachePageSanity(p) );
 629   pcacheTrace(("%p.MOVE %d -> %d\n",pCache,p->pgno,newPgno));
 630   pOther = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache, newPgno, 0);
 631   if( pOther ){
 632     PgHdr *pXPage = (PgHdr*)pOther->pExtra;
 633     assert( pXPage->nRef==0 );
 634     pXPage->nRef++;
 635     pCache->nRefSum++;
 636     sqlite3PcacheDrop(pXPage);
 637   }
 638   sqlite3GlobalConfig.pcache2.xRekey(pCache->pCache, p->pPage, p->pgno,newPgno);
 639   p->pgno = newPgno;
 640   if( (p->flags&PGHDR_DIRTY) && (p->flags&PGHDR_NEED_SYNC) ){
 641     pcacheManageDirtyList(p, PCACHE_DIRTYLIST_FRONT);
 642     assert( sqlite3PcachePageSanity(p) );
 643   }
 644 }
 645
 646 /*
 647 ** Drop every cache entry whose page number is greater than "pgno". The
 648 ** caller must ensure that there are no outstanding references to any pages
 649 ** other than page 1 with a page number greater than pgno.
 650 **
 651 ** If there is a reference to page 1 and the pgno parameter passed to this
 652 ** function is 0, then the data area associated with page 1 is zeroed, but
 653 ** the page object is not dropped.
 654 */
 655 void sqlite3PcacheTruncate(PCache *pCache, Pgno pgno){
 656   if( pCache->pCache ){
 657     PgHdr *p;
 658     PgHdr *pNext;
 659     pcacheTrace(("%p.TRUNCATE %d\n",pCache,pgno));
 660     for(p=pCache->pDirty; p; p=pNext){
 661       pNext = p->pDirtyNext;
 662       /* This routine never gets call with a positive pgno except right
 663       ** after sqlite3PcacheCleanAll().  So if there are dirty pages,
 664       ** it must be that pgno==0.
 665       */
 666       assert( p->pgno>0 );
 667       if( p->pgno>pgno ){
 668         assert( p->flags&PGHDR_DIRTY );
 669         sqlite3PcacheMakeClean(p);
 670       }
 671     }
 672     if( pgno==0 && pCache->nRefSum ){
 673       sqlite3_pcache_page *pPage1;
 674       pPage1 = sqlite3GlobalConfig.pcache2.xFetch(pCache->pCache,1,0);
 675       if( ALWAYS(pPage1) ){  /* Page 1 is always available in cache, because
 676                              ** pCache->nRefSum>0 */
 677         memset(pPage1->pBuf, 0, pCache->szPage);
 678         pgno = 1;
 679       }
 680     }
 681     sqlite3GlobalConfig.pcache2.xTruncate(pCache->pCache, pgno+1);
 682   }
 683 }
 684
 685 /*
 686 ** Close a cache.
 687 */
 688 void sqlite3PcacheClose(PCache *pCache){
 689   assert( pCache->pCache!=0 );
 690   pcacheTrace(("%p.CLOSE\n",pCache));
 691   sqlite3GlobalConfig.pcache2.xDestroy(pCache->pCache);
 692 }
 693
 694 /*
 695 ** Discard the contents of the cache.
 696 */
 697 void sqlite3PcacheClear(PCache *pCache){
 698   sqlite3PcacheTruncate(pCache, 0);
 699 }
 700
 701 /*
 702 ** Merge two lists of pages connected by pDirty and in pgno order.
 703 ** Do not bother fixing the pDirtyPrev pointers.
 704 */
 705 static PgHdr *pcacheMergeDirtyList(PgHdr *pA, PgHdr *pB){
 706   PgHdr result, *pTail;
 707   pTail = &result;
 708   assert( pA!=0 && pB!=0 );
 709   for(;;){
 710     if( pA->pgno<pB->pgno ){
 711       pTail->pDirty = pA;
 712       pTail = pA;
 713       pA = pA->pDirty;
 714       if( pA==0 ){
 715         pTail->pDirty = pB;
 716         break;
 717       }
 718     }else{
 719       pTail->pDirty = pB;
 720       pTail = pB;
 721       pB = pB->pDirty;
 722       if( pB==0 ){
 723         pTail->pDirty = pA;
 724         break;
 725       }
 726     }
 727   }
 728   return result.pDirty;
 729 }
 730
 731 /*
 732 ** Sort the list of pages in accending order by pgno.  Pages are
 733 ** connected by pDirty pointers.  The pDirtyPrev pointers are
 734 ** corrupted by this sort.
 735 **
 736 ** Since there cannot be more than 2^31 distinct pages in a database,
 737 ** there cannot be more than 31 buckets required by the merge sorter.
 738 ** One extra bucket is added to catch overflow in case something
 739 ** ever changes to make the previous sentence incorrect.
 740 */
 741 #define N_SORT_BUCKET  32
 742 static PgHdr *pcacheSortDirtyList(PgHdr *pIn){
 743   PgHdr *a[N_SORT_BUCKET], *p;
 744   int i;
 745   memset(a, 0, sizeof(a));
 746   while( pIn ){
 747     p = pIn;
 748     pIn = p->pDirty;
 749     p->pDirty = 0;
 750     for(i=0; ALWAYS(i<N_SORT_BUCKET-1); i++){
 751       if( a[i]==0 ){
 752         a[i] = p;
 753         break;
 754       }else{
 755         p = pcacheMergeDirtyList(a[i], p);
 756         a[i] = 0;
 757       }
 758     }
 759     if( NEVER(i==N_SORT_BUCKET-1) ){
 760       /* To get here, there need to be 2^(N_SORT_BUCKET) elements in
 761       ** the input list.  But that is impossible.
 762       */
 763       a[i] = pcacheMergeDirtyList(a[i], p);
 764     }
 765   }
 766   p = a[0];
 767   for(i=1; i<N_SORT_BUCKET; i++){
 768     if( a[i]==0 ) continue;
 769     p = p ? pcacheMergeDirtyList(p, a[i]) : a[i];
 770   }
 771   return p;
 772 }
 773
 774 /*
 775 ** Return a list of all dirty pages in the cache, sorted by page number.
 776 */
 777 PgHdr *sqlite3PcacheDirtyList(PCache *pCache){
 778   PgHdr *p;
 779   for(p=pCache->pDirty; p; p=p->pDirtyNext){
 780     p->pDirty = p->pDirtyNext;
 781   }
 782   return pcacheSortDirtyList(pCache->pDirty);
 783 }
 784
 785 /*
 786 ** Return the total number of references to all pages held by the cache.
 787 **
 788 ** This is not the total number of pages referenced, but the sum of the
 789 ** reference count for all pages.
 790 */
 791 int sqlite3PcacheRefCount(PCache *pCache){
 792   return pCache->nRefSum;
 793 }
 794
 795 /*
 796 ** Return the number of references to the page supplied as an argument.
 797 */
 798 int sqlite3PcachePageRefcount(PgHdr *p){
 799   return p->nRef;
 800 }
 801
 802 /*
 803 ** Return the total number of pages in the cache.
 804 */
 805 int sqlite3PcachePagecount(PCache *pCache){
 806   assert( pCache->pCache!=0 );
 807   return sqlite3GlobalConfig.pcache2.xPagecount(pCache->pCache);
 808 }
 809
 810 #ifdef SQLITE_TEST
 811 /*
 812 ** Get the suggested cache-size value.
 813 */
 814 int sqlite3PcacheGetCachesize(PCache *pCache){
 815   return numberOfCachePages(pCache);
 816 }
 817 #endif
 818
 819 /*
 820 ** Set the suggested cache-size value.
 821 */
 822 void sqlite3PcacheSetCachesize(PCache *pCache, int mxPage){
 823   assert( pCache->pCache!=0 );
 824   pCache->szCache = mxPage;
 825   sqlite3GlobalConfig.pcache2.xCachesize(pCache->pCache,
 826                                          numberOfCachePages(pCache));
 827 }
 828
 829 /*
 830 ** Set the suggested cache-spill value.  Make no changes if if the
 831 ** argument is zero.  Return the effective cache-spill size, which will
 832 ** be the larger of the szSpill and szCache.
 833 */
 834 int sqlite3PcacheSetSpillsize(PCache *p, int mxPage){
 835   int res;
 836   assert( p->pCache!=0 );
 837   if( mxPage ){
 838     if( mxPage<0 ){
 839       mxPage = (int)((-1024*(i64)mxPage)/(p->szPage+p->szExtra));
 840     }
 841     p->szSpill = mxPage;
 842   }
 843   res = numberOfCachePages(p);
 844   if( res<p->szSpill ) res = p->szSpill;
 845   return res;
 846 }
 847
 848 /*
 849 ** Free up as much memory as possible from the page cache.
 850 */
 851 void sqlite3PcacheShrink(PCache *pCache){
 852   assert( pCache->pCache!=0 );
 853   sqlite3GlobalConfig.pcache2.xShrink(pCache->pCache);
 854 }
 855
 856 /*
 857 ** Return the size of the header added by this middleware layer
 858 ** in the page-cache hierarchy.
 859 */
 860 int sqlite3HeaderSizePcache(void){ return ROUND8(sizeof(PgHdr)); }
 861
 862 /*
 863 ** Return the number of dirty pages currently in the cache, as a percentage
 864 ** of the configured cache size.
 865 */
 866 int sqlite3PCachePercentDirty(PCache *pCache){
 867   PgHdr *pDirty;
 868   int nDirty = 0;
 869   int nCache = numberOfCachePages(pCache);
 870   for(pDirty=pCache->pDirty; pDirty; pDirty=pDirty->pDirtyNext) nDirty++;
 871   return nCache ? (int)(((i64)nDirty * 100) / nCache) : 0;
 872 }
 873
 874 #ifdef SQLITE_DIRECT_OVERFLOW_READ
 875 /*
 876 ** Return true if there are one or more dirty pages in the cache. Else false.
 877 */
 878 int sqlite3PCacheIsDirty(PCache *pCache){
 879   return (pCache->pDirty!=0);
 880 }
 881 #endif
 882
 883 #if defined(SQLITE_CHECK_PAGES) || defined(SQLITE_DEBUG)
 884 /*
 885 ** For all dirty pages currently in the cache, invoke the specified
 886 ** callback. This is only used if the SQLITE_CHECK_PAGES macro is
 887 ** defined.
 888 */
 889 void sqlite3PcacheIterateDirty(PCache *pCache, void (*xIter)(PgHdr *)){
 890   PgHdr *pDirty;
 891   for(pDirty=pCache->pDirty; pDirty; pDirty=pDirty->pDirtyNext){
 892     xIter(pDirty);
 893   }
 894 }
 895 #endif