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