src/pcache1.c

   1 /*
   2 ** 2008 November 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 **
  13 ** This file implements the default page cache implementation (the
  14 ** sqlite3_pcache interface). It also contains part of the implementation
  15 ** of the SQLITE_CONFIG_PAGECACHE and sqlite3_release_memory() features.
  16 ** If the default page cache implementation is overridden, then neither of
  17 ** these two features are available.
  18 **
  19 ** A Page cache line looks like this:
  20 **
  21 **  -------------------------------------------------------------
  22 **  |  database page content   |  PgHdr1  |  MemPage  |  PgHdr  |
  23 **  -------------------------------------------------------------
  24 **
  25 ** The database page content is up front (so that buffer overreads tend to
  26 ** flow harmlessly into the PgHdr1, MemPage, and PgHdr extensions).   MemPage
  27 ** is the extension added by the btree.c module containing information such
  28 ** as the database page number and how that database page is used.  PgHdr
  29 ** is added by the pcache.c layer and contains information used to keep track
  30 ** of which pages are "dirty".  PgHdr1 is an extension added by this
  31 ** module (pcache1.c).  The PgHdr1 header is a subclass of sqlite3_pcache_page.
  32 ** PgHdr1 contains information needed to look up a page by its page number.
  33 ** The superclass sqlite3_pcache_page.pBuf points to the start of the
  34 ** database page content and sqlite3_pcache_page.pExtra points to PgHdr.
  35 **
  36 ** The size of the extension (MemPage+PgHdr+PgHdr1) can be determined at
  37 ** runtime using sqlite3_config(SQLITE_CONFIG_PCACHE_HDRSZ, &size).  The
  38 ** sizes of the extensions sum to 272 bytes on x64 for 3.8.10, but this
  39 ** size can vary according to architecture, compile-time options, and
  40 ** SQLite library version number.
  41 **
  42 ** If SQLITE_PCACHE_SEPARATE_HEADER is defined, then the extension is obtained
  43 ** using a separate memory allocation from the database page content.  This
  44 ** seeks to overcome the "clownshoe" problem (also called "internal
  45 ** fragmentation" in academic literature) of allocating a few bytes more
  46 ** than a power of two with the memory allocator rounding up to the next
  47 ** power of two, and leaving the rounded-up space unused.
  48 **
  49 ** This module tracks pointers to PgHdr1 objects.  Only pcache.c communicates
  50 ** with this module.  Information is passed back and forth as PgHdr1 pointers.
  51 **
  52 ** The pcache.c and pager.c modules deal pointers to PgHdr objects.
  53 ** The btree.c module deals with pointers to MemPage objects.
  54 **
  55 ** SOURCE OF PAGE CACHE MEMORY:
  56 **
  57 ** Memory for a page might come from any of three sources:
  58 **
  59 **    (1)  The general-purpose memory allocator - sqlite3Malloc()
  60 **    (2)  Global page-cache memory provided using sqlite3_config() with
  61 **         SQLITE_CONFIG_PAGECACHE.
  62 **    (3)  PCache-local bulk allocation.
  63 **
  64 ** The third case is a chunk of heap memory (defaulting to 100 pages worth)
  65 ** that is allocated when the page cache is created.  The size of the local
  66 ** bulk allocation can be adjusted using
  67 **
  68 **     sqlite3_config(SQLITE_CONFIG_PAGECACHE, (void*)0, 0, N).
  69 **
  70 ** If N is positive, then N pages worth of memory are allocated using a single
  71 ** sqlite3Malloc() call and that memory is used for the first N pages allocated.
  72 ** Or if N is negative, then -1024*N bytes of memory are allocated and used
  73 ** for as many pages as can be accomodated.
  74 **
  75 ** Only one of (2) or (3) can be used.  Once the memory available to (2) or
  76 ** (3) is exhausted, subsequent allocations fail over to the general-purpose
  77 ** memory allocator (1).
  78 **
  79 ** Earlier versions of SQLite used only methods (1) and (2).  But experiments
  80 ** show that method (3) with N==100 provides about a 5% performance boost for
  81 ** common workloads.
  82 */
  83 #include "sqliteInt.h"
  84
  85 typedef struct PCache1 PCache1;
  86 typedef struct PgHdr1 PgHdr1;
  87 typedef struct PgFreeslot PgFreeslot;
  88 typedef struct PGroup PGroup;
  89
  90 /*
  91 ** Each cache entry is represented by an instance of the following
  92 ** structure. Unless SQLITE_PCACHE_SEPARATE_HEADER is defined, a buffer of
  93 ** PgHdr1.pCache->szPage bytes is allocated directly before this structure
  94 ** in memory.
  95 **
  96 ** Note: Variables isBulkLocal and isAnchor were once type "u8". That works,
  97 ** but causes a 2-byte gap in the structure for most architectures (since
  98 ** pointers must be either 4 or 8-byte aligned). As this structure is located
  99 ** in memory directly after the associated page data, if the database is
 100 ** corrupt, code at the b-tree layer may overread the page buffer and
 101 ** read part of this structure before the corruption is detected. This
 102 ** can cause a valgrind error if the unitialized gap is accessed. Using u16
 103 ** ensures there is no such gap, and therefore no bytes of unitialized memory
 104 ** in the structure.
 105 */
 106 struct PgHdr1 {
 107   sqlite3_pcache_page page;      /* Base class. Must be first. pBuf & pExtra */
 108   unsigned int iKey;             /* Key value (page number) */
 109   u16 isBulkLocal;               /* This page from bulk local storage */
 110   u16 isAnchor;                  /* This is the PGroup.lru element */
 111   PgHdr1 *pNext;                 /* Next in hash table chain */
 112   PCache1 *pCache;               /* Cache that currently owns this page */
 113   PgHdr1 *pLruNext;              /* Next in LRU list of unpinned pages */
 114   PgHdr1 *pLruPrev;              /* Previous in LRU list of unpinned pages */
 115                                  /* NB: pLruPrev is only valid if pLruNext!=0 */
 116 };
 117
 118 /*
 119 ** A page is pinned if it is not on the LRU list.  To be "pinned" means
 120 ** that the page is in active use and must not be deallocated.
 121 */
 122 #define PAGE_IS_PINNED(p)    ((p)->pLruNext==0)
 123 #define PAGE_IS_UNPINNED(p)  ((p)->pLruNext!=0)
 124
 125 /* Each page cache (or PCache) belongs to a PGroup.  A PGroup is a set
 126 ** of one or more PCaches that are able to recycle each other's unpinned
 127 ** pages when they are under memory pressure.  A PGroup is an instance of
 128 ** the following object.
 129 **
 130 ** This page cache implementation works in one of two modes:
 131 **
 132 **   (1)  Every PCache is the sole member of its own PGroup.  There is
 133 **        one PGroup per PCache.
 134 **
 135 **   (2)  There is a single global PGroup that all PCaches are a member
 136 **        of.
 137 **
 138 ** Mode 1 uses more memory (since PCache instances are not able to rob
 139 ** unused pages from other PCaches) but it also operates without a mutex,
 140 ** and is therefore often faster.  Mode 2 requires a mutex in order to be
 141 ** threadsafe, but recycles pages more efficiently.
 142 **
 143 ** For mode (1), PGroup.mutex is NULL.  For mode (2) there is only a single
 144 ** PGroup which is the pcache1.grp global variable and its mutex is
 145 ** SQLITE_MUTEX_STATIC_LRU.
 146 */
 147 struct PGroup {
 148   sqlite3_mutex *mutex;          /* MUTEX_STATIC_LRU or NULL */
 149   unsigned int nMaxPage;         /* Sum of nMax for purgeable caches */
 150   unsigned int nMinPage;         /* Sum of nMin for purgeable caches */
 151   unsigned int mxPinned;         /* nMaxpage + 10 - nMinPage */
 152   unsigned int nPurgeable;       /* Number of purgeable pages allocated */
 153   PgHdr1 lru;                    /* The beginning and end of the LRU list */
 154 };
 155
 156 /* Each page cache is an instance of the following object.  Every
 157 ** open database file (including each in-memory database and each
 158 ** temporary or transient database) has a single page cache which
 159 ** is an instance of this object.
 160 **
 161 ** Pointers to structures of this type are cast and returned as
 162 ** opaque sqlite3_pcache* handles.
 163 */
 164 struct PCache1 {
 165   /* Cache configuration parameters. Page size (szPage) and the purgeable
 166   ** flag (bPurgeable) and the pnPurgeable pointer are all set when the
 167   ** cache is created and are never changed thereafter. nMax may be
 168   ** modified at any time by a call to the pcache1Cachesize() method.
 169   ** The PGroup mutex must be held when accessing nMax.
 170   */
 171   PGroup *pGroup;                     /* PGroup this cache belongs to */
 172   unsigned int *pnPurgeable;          /* Pointer to pGroup->nPurgeable */
 173   int szPage;                         /* Size of database content section */
 174   int szExtra;                        /* sizeof(MemPage)+sizeof(PgHdr) */
 175   int szAlloc;                        /* Total size of one pcache line */
 176   int bPurgeable;                     /* True if cache is purgeable */
 177   unsigned int nMin;                  /* Minimum number of pages reserved */
 178   unsigned int nMax;                  /* Configured "cache_size" value */
 179   unsigned int n90pct;                /* nMax*9/10 */
 180   unsigned int iMaxKey;               /* Largest key seen since xTruncate() */
 181   unsigned int nPurgeableDummy;       /* pnPurgeable points here when not used*/
 182
 183   /* Hash table of all pages. The following variables may only be accessed
 184   ** when the accessor is holding the PGroup mutex.
 185   */
 186   unsigned int nRecyclable;           /* Number of pages in the LRU list */
 187   unsigned int nPage;                 /* Total number of pages in apHash */
 188   unsigned int nHash;                 /* Number of slots in apHash[] */
 189   PgHdr1 **apHash;                    /* Hash table for fast lookup by key */
 190   PgHdr1 *pFree;                      /* List of unused pcache-local pages */
 191   void *pBulk;                        /* Bulk memory used by pcache-local */
 192 };
 193
 194 /*
 195 ** Free slots in the allocator used to divide up the global page cache
 196 ** buffer provided using the SQLITE_CONFIG_PAGECACHE mechanism.
 197 */
 198 struct PgFreeslot {
 199   PgFreeslot *pNext;  /* Next free slot */
 200 };
 201
 202 /*
 203 ** Global data used by this cache.
 204 */
 205 static SQLITE_WSD struct PCacheGlobal {
 206   PGroup grp;                    /* The global PGroup for mode (2) */
 207
 208   /* Variables related to SQLITE_CONFIG_PAGECACHE settings.  The
 209   ** szSlot, nSlot, pStart, pEnd, nReserve, and isInit values are all
 210   ** fixed at sqlite3_initialize() time and do not require mutex protection.
 211   ** The nFreeSlot and pFree values do require mutex protection.
 212   */
 213   int isInit;                    /* True if initialized */
 214   int separateCache;             /* Use a new PGroup for each PCache */
 215   int nInitPage;                 /* Initial bulk allocation size */
 216   int szSlot;                    /* Size of each free slot */
 217   int nSlot;                     /* The number of pcache slots */
 218   int nReserve;                  /* Try to keep nFreeSlot above this */
 219   void *pStart, *pEnd;           /* Bounds of global page cache memory */
 220   /* Above requires no mutex.  Use mutex below for variable that follow. */
 221   sqlite3_mutex *mutex;          /* Mutex for accessing the following: */
 222   PgFreeslot *pFree;             /* Free page blocks */
 223   int nFreeSlot;                 /* Number of unused pcache slots */
 224   /* The following value requires a mutex to change.  We skip the mutex on
 225   ** reading because (1) most platforms read a 32-bit integer atomically and
 226   ** (2) even if an incorrect value is read, no great harm is done since this
 227   ** is really just an optimization. */
 228   int bUnderPressure;            /* True if low on PAGECACHE memory */
 229 } pcache1_g;
 230
 231 /*
 232 ** All code in this file should access the global structure above via the
 233 ** alias "pcache1". This ensures that the WSD emulation is used when
 234 ** compiling for systems that do not support real WSD.
 235 */
 236 #define pcache1 (GLOBAL(struct PCacheGlobal, pcache1_g))
 237
 238 /*
 239 ** Macros to enter and leave the PCache LRU mutex.
 240 */
 241 #if !defined(SQLITE_ENABLE_MEMORY_MANAGEMENT) || SQLITE_THREADSAFE==0
 242 # define pcache1EnterMutex(X)  assert((X)->mutex==0)
 243 # define pcache1LeaveMutex(X)  assert((X)->mutex==0)
 244 # define PCACHE1_MIGHT_USE_GROUP_MUTEX 0
 245 #else
 246 # define pcache1EnterMutex(X) sqlite3_mutex_enter((X)->mutex)
 247 # define pcache1LeaveMutex(X) sqlite3_mutex_leave((X)->mutex)
 248 # define PCACHE1_MIGHT_USE_GROUP_MUTEX 1
 249 #endif
 250
 251 /******************************************************************************/
 252 /******** Page Allocation/SQLITE_CONFIG_PCACHE Related Functions **************/
 253
 254
 255 /*
 256 ** This function is called during initialization if a static buffer is
 257 ** supplied to use for the page-cache by passing the SQLITE_CONFIG_PAGECACHE
 258 ** verb to sqlite3_config(). Parameter pBuf points to an allocation large
 259 ** enough to contain 'n' buffers of 'sz' bytes each.
 260 **
 261 ** This routine is called from sqlite3_initialize() and so it is guaranteed
 262 ** to be serialized already.  There is no need for further mutexing.
 263 */
 264 void sqlite3PCacheBufferSetup(void *pBuf, int sz, int n){
 265   if( pcache1.isInit ){
 266     PgFreeslot *p;
 267     if( pBuf==0 ) sz = n = 0;
 268     if( n==0 ) sz = 0;
 269     sz = ROUNDDOWN8(sz);
 270     pcache1.szSlot = sz;
 271     pcache1.nSlot = pcache1.nFreeSlot = n;
 272     pcache1.nReserve = n>90 ? 10 : (n/10 + 1);
 273     pcache1.pStart = pBuf;
 274     pcache1.pFree = 0;
 275     pcache1.bUnderPressure = 0;
 276     while( n-- ){
 277       p = (PgFreeslot*)pBuf;
 278       p->pNext = pcache1.pFree;
 279       pcache1.pFree = p;
 280       pBuf = (void*)&((char*)pBuf)[sz];
 281     }
 282     pcache1.pEnd = pBuf;
 283   }
 284 }
 285
 286 /*
 287 ** Try to initialize the pCache->pFree and pCache->pBulk fields.  Return
 288 ** true if pCache->pFree ends up containing one or more free pages.
 289 */
 290 static int pcache1InitBulk(PCache1 *pCache){
 291   i64 szBulk;
 292   char *zBulk;
 293   if( pcache1.nInitPage==0 ) return 0;
 294   /* Do not bother with a bulk allocation if the cache size very small */
 295   if( pCache->nMax<3 ) return 0;
 296   sqlite3BeginBenignMalloc();
 297   if( pcache1.nInitPage>0 ){
 298     szBulk = pCache->szAlloc * (i64)pcache1.nInitPage;
 299   }else{
 300     szBulk = -1024 * (i64)pcache1.nInitPage;
 301   }
 302   if( szBulk > pCache->szAlloc*(i64)pCache->nMax ){
 303     szBulk = pCache->szAlloc*(i64)pCache->nMax;
 304   }
 305   zBulk = pCache->pBulk = sqlite3Malloc( szBulk );
 306   sqlite3EndBenignMalloc();
 307   if( zBulk ){
 308     int nBulk = sqlite3MallocSize(zBulk)/pCache->szAlloc;
 309     do{
 310       PgHdr1 *pX = (PgHdr1*)&zBulk[pCache->szPage];
 311       pX->page.pBuf = zBulk;
 312       pX->page.pExtra = &pX[1];
 313       pX->isBulkLocal = 1;
 314       pX->isAnchor = 0;
 315       pX->pNext = pCache->pFree;
 316       pX->pLruPrev = 0;           /* Initializing this saves a valgrind error */
 317       pCache->pFree = pX;
 318       zBulk += pCache->szAlloc;
 319     }while( --nBulk );
 320   }
 321   return pCache->pFree!=0;
 322 }
 323
 324 /*
 325 ** Malloc function used within this file to allocate space from the buffer
 326 ** configured using sqlite3_config(SQLITE_CONFIG_PAGECACHE) option. If no
 327 ** such buffer exists or there is no space left in it, this function falls
 328 ** back to sqlite3Malloc().
 329 **
 330 ** Multiple threads can run this routine at the same time.  Global variables
 331 ** in pcache1 need to be protected via mutex.
 332 */
 333 static void *pcache1Alloc(int nByte){
 334   void *p = 0;
 335   assert( sqlite3_mutex_notheld(pcache1.grp.mutex) );
 336   if( nByte<=pcache1.szSlot ){
 337     sqlite3_mutex_enter(pcache1.mutex);
 338     p = (PgHdr1 *)pcache1.pFree;
 339     if( p ){
 340       pcache1.pFree = pcache1.pFree->pNext;
 341       pcache1.nFreeSlot--;
 342       pcache1.bUnderPressure = pcache1.nFreeSlot<pcache1.nReserve;
 343       assert( pcache1.nFreeSlot>=0 );
 344       sqlite3StatusHighwater(SQLITE_STATUS_PAGECACHE_SIZE, nByte);
 345       sqlite3StatusUp(SQLITE_STATUS_PAGECACHE_USED, 1);
 346     }
 347     sqlite3_mutex_leave(pcache1.mutex);
 348   }
 349   if( p==0 ){
 350     /* Memory is not available in the SQLITE_CONFIG_PAGECACHE pool.  Get
 351     ** it from sqlite3Malloc instead.
 352     */
 353     p = sqlite3Malloc(nByte);
 354 #ifndef SQLITE_DISABLE_PAGECACHE_OVERFLOW_STATS
 355     if( p ){
 356       int sz = sqlite3MallocSize(p);
 357       sqlite3_mutex_enter(pcache1.mutex);
 358       sqlite3StatusHighwater(SQLITE_STATUS_PAGECACHE_SIZE, nByte);
 359       sqlite3StatusUp(SQLITE_STATUS_PAGECACHE_OVERFLOW, sz);
 360       sqlite3_mutex_leave(pcache1.mutex);
 361     }
 362 #endif
 363     sqlite3MemdebugSetType(p, MEMTYPE_PCACHE);
 364   }
 365   return p;
 366 }
 367
 368 /*
 369 ** Free an allocated buffer obtained from pcache1Alloc().
 370 */
 371 static void pcache1Free(void *p){
 372   if( p==0 ) return;
 373   if( SQLITE_WITHIN(p, pcache1.pStart, pcache1.pEnd) ){
 374     PgFreeslot *pSlot;
 375     sqlite3_mutex_enter(pcache1.mutex);
 376     sqlite3StatusDown(SQLITE_STATUS_PAGECACHE_USED, 1);
 377     pSlot = (PgFreeslot*)p;
 378     pSlot->pNext = pcache1.pFree;
 379     pcache1.pFree = pSlot;
 380     pcache1.nFreeSlot++;
 381     pcache1.bUnderPressure = pcache1.nFreeSlot<pcache1.nReserve;
 382     assert( pcache1.nFreeSlot<=pcache1.nSlot );
 383     sqlite3_mutex_leave(pcache1.mutex);
 384   }else{
 385     assert( sqlite3MemdebugHasType(p, MEMTYPE_PCACHE) );
 386     sqlite3MemdebugSetType(p, MEMTYPE_HEAP);
 387 #ifndef SQLITE_DISABLE_PAGECACHE_OVERFLOW_STATS
 388     {
 389       int nFreed = 0;
 390       nFreed = sqlite3MallocSize(p);
 391       sqlite3_mutex_enter(pcache1.mutex);
 392       sqlite3StatusDown(SQLITE_STATUS_PAGECACHE_OVERFLOW, nFreed);
 393       sqlite3_mutex_leave(pcache1.mutex);
 394     }
 395 #endif
 396     sqlite3_free(p);
 397   }
 398 }
 399
 400 #ifdef SQLITE_ENABLE_MEMORY_MANAGEMENT
 401 /*
 402 ** Return the size of a pcache allocation
 403 */
 404 static int pcache1MemSize(void *p){
 405   if( p>=pcache1.pStart && p<pcache1.pEnd ){
 406     return pcache1.szSlot;
 407   }else{
 408     int iSize;
 409     assert( sqlite3MemdebugHasType(p, MEMTYPE_PCACHE) );
 410     sqlite3MemdebugSetType(p, MEMTYPE_HEAP);
 411     iSize = sqlite3MallocSize(p);
 412     sqlite3MemdebugSetType(p, MEMTYPE_PCACHE);
 413     return iSize;
 414   }
 415 }
 416 #endif /* SQLITE_ENABLE_MEMORY_MANAGEMENT */
 417
 418 /*
 419 ** Allocate a new page object initially associated with cache pCache.
 420 */
 421 static PgHdr1 *pcache1AllocPage(PCache1 *pCache, int benignMalloc){
 422   PgHdr1 *p = 0;
 423   void *pPg;
 424
 425   assert( sqlite3_mutex_held(pCache->pGroup->mutex) );
 426   if( pCache->pFree || (pCache->nPage==0 && pcache1InitBulk(pCache)) ){
 427     p = pCache->pFree;
 428     pCache->pFree = p->pNext;
 429     p->pNext = 0;
 430   }else{
 431 #ifdef SQLITE_ENABLE_MEMORY_MANAGEMENT
 432     /* The group mutex must be released before pcache1Alloc() is called. This
 433     ** is because it might call sqlite3_release_memory(), which assumes that
 434     ** this mutex is not held. */
 435     assert( pcache1.separateCache==0 );
 436     assert( pCache->pGroup==&pcache1.grp );
 437     pcache1LeaveMutex(pCache->pGroup);
 438 #endif
 439     if( benignMalloc ){ sqlite3BeginBenignMalloc(); }
 440 #ifdef SQLITE_PCACHE_SEPARATE_HEADER
 441     pPg = pcache1Alloc(pCache->szPage);
 442     p = sqlite3Malloc(sizeof(PgHdr1) + pCache->szExtra);
 443     if( !pPg || !p ){
 444       pcache1Free(pPg);
 445       sqlite3_free(p);
 446       pPg = 0;
 447     }
 448 #else
 449     pPg = pcache1Alloc(pCache->szAlloc);
 450     p = (PgHdr1 *)&((u8 *)pPg)[pCache->szPage];
 451 #endif
 452     if( benignMalloc ){ sqlite3EndBenignMalloc(); }
 453 #ifdef SQLITE_ENABLE_MEMORY_MANAGEMENT
 454     pcache1EnterMutex(pCache->pGroup);
 455 #endif
 456     if( pPg==0 ) return 0;
 457     p->page.pBuf = pPg;
 458     p->page.pExtra = &p[1];
 459     p->isBulkLocal = 0;
 460     p->isAnchor = 0;
 461   }
 462   (*pCache->pnPurgeable)++;
 463   return p;
 464 }
 465
 466 /*
 467 ** Free a page object allocated by pcache1AllocPage().
 468 */
 469 static void pcache1FreePage(PgHdr1 *p){
 470   PCache1 *pCache;
 471   assert( p!=0 );
 472   pCache = p->pCache;
 473   assert( sqlite3_mutex_held(p->pCache->pGroup->mutex) );
 474   if( p->isBulkLocal ){
 475     p->pNext = pCache->pFree;
 476     pCache->pFree = p;
 477   }else{
 478     pcache1Free(p->page.pBuf);
 479 #ifdef SQLITE_PCACHE_SEPARATE_HEADER
 480     sqlite3_free(p);
 481 #endif
 482   }
 483   (*pCache->pnPurgeable)--;
 484 }
 485
 486 /*
 487 ** Malloc function used by SQLite to obtain space from the buffer configured
 488 ** using sqlite3_config(SQLITE_CONFIG_PAGECACHE) option. If no such buffer
 489 ** exists, this function falls back to sqlite3Malloc().
 490 */
 491 void *sqlite3PageMalloc(int sz){
 492   /* During rebalance operations on a corrupt database file, it is sometimes
 493   ** (rarely) possible to overread the temporary page buffer by a few bytes.
 494   ** Enlarge the allocation slightly so that this does not cause problems. */
 495   return pcache1Alloc(sz);
 496 }
 497
 498 /*
 499 ** Free an allocated buffer obtained from sqlite3PageMalloc().
 500 */
 501 void sqlite3PageFree(void *p){
 502   pcache1Free(p);
 503 }
 504
 505
 506 /*
 507 ** Return true if it desirable to avoid allocating a new page cache
 508 ** entry.
 509 **
 510 ** If memory was allocated specifically to the page cache using
 511 ** SQLITE_CONFIG_PAGECACHE but that memory has all been used, then
 512 ** it is desirable to avoid allocating a new page cache entry because
 513 ** presumably SQLITE_CONFIG_PAGECACHE was suppose to be sufficient
 514 ** for all page cache needs and we should not need to spill the
 515 ** allocation onto the heap.
 516 **
 517 ** Or, the heap is used for all page cache memory but the heap is
 518 ** under memory pressure, then again it is desirable to avoid
 519 ** allocating a new page cache entry in order to avoid stressing
 520 ** the heap even further.
 521 */
 522 static int pcache1UnderMemoryPressure(PCache1 *pCache){
 523   if( pcache1.nSlot && (pCache->szPage+pCache->szExtra)<=pcache1.szSlot ){
 524     return pcache1.bUnderPressure;
 525   }else{
 526     return sqlite3HeapNearlyFull();
 527   }
 528 }
 529
 530 /******************************************************************************/
 531 /******** General Implementation Functions ************************************/
 532
 533 /*
 534 ** This function is used to resize the hash table used by the cache passed
 535 ** as the first argument.
 536 **
 537 ** The PCache mutex must be held when this function is called.
 538 */
 539 static void pcache1ResizeHash(PCache1 *p){
 540   PgHdr1 **apNew;
 541   unsigned int nNew;
 542   unsigned int i;
 543
 544   assert( sqlite3_mutex_held(p->pGroup->mutex) );
 545
 546   nNew = p->nHash*2;
 547   if( nNew<256 ){
 548     nNew = 256;
 549   }
 550
 551   pcache1LeaveMutex(p->pGroup);
 552   if( p->nHash ){ sqlite3BeginBenignMalloc(); }
 553   apNew = (PgHdr1 **)sqlite3MallocZero(sizeof(PgHdr1 *)*nNew);
 554   if( p->nHash ){ sqlite3EndBenignMalloc(); }
 555   pcache1EnterMutex(p->pGroup);
 556   if( apNew ){
 557     for(i=0; i<p->nHash; i++){
 558       PgHdr1 *pPage;
 559       PgHdr1 *pNext = p->apHash[i];
 560       while( (pPage = pNext)!=0 ){
 561         unsigned int h = pPage->iKey % nNew;
 562         pNext = pPage->pNext;
 563         pPage->pNext = apNew[h];
 564         apNew[h] = pPage;
 565       }
 566     }
 567     sqlite3_free(p->apHash);
 568     p->apHash = apNew;
 569     p->nHash = nNew;
 570   }
 571 }
 572
 573 /*
 574 ** This function is used internally to remove the page pPage from the
 575 ** PGroup LRU list, if is part of it. If pPage is not part of the PGroup
 576 ** LRU list, then this function is a no-op.
 577 **
 578 ** The PGroup mutex must be held when this function is called.
 579 */
 580 static PgHdr1 *pcache1PinPage(PgHdr1 *pPage){
 581   assert( pPage!=0 );
 582   assert( PAGE_IS_UNPINNED(pPage) );
 583   assert( pPage->pLruNext );
 584   assert( pPage->pLruPrev );
 585   assert( sqlite3_mutex_held(pPage->pCache->pGroup->mutex) );
 586   pPage->pLruPrev->pLruNext = pPage->pLruNext;
 587   pPage->pLruNext->pLruPrev = pPage->pLruPrev;
 588   pPage->pLruNext = 0;
 589   /* pPage->pLruPrev = 0;
 590   ** No need to clear pLruPrev as it is never accessed if pLruNext is 0 */
 591   assert( pPage->isAnchor==0 );
 592   assert( pPage->pCache->pGroup->lru.isAnchor==1 );
 593   pPage->pCache->nRecyclable--;
 594   return pPage;
 595 }
 596
 597
 598 /*
 599 ** Remove the page supplied as an argument from the hash table
 600 ** (PCache1.apHash structure) that it is currently stored in.
 601 ** Also free the page if freePage is true.
 602 **
 603 ** The PGroup mutex must be held when this function is called.
 604 */
 605 static void pcache1RemoveFromHash(PgHdr1 *pPage, int freeFlag){
 606   unsigned int h;
 607   PCache1 *pCache = pPage->pCache;
 608   PgHdr1 **pp;
 609
 610   assert( sqlite3_mutex_held(pCache->pGroup->mutex) );
 611   h = pPage->iKey % pCache->nHash;
 612   for(pp=&pCache->apHash[h]; (*pp)!=pPage; pp=&(*pp)->pNext);
 613   *pp = (*pp)->pNext;
 614
 615   pCache->nPage--;
 616   if( freeFlag ) pcache1FreePage(pPage);
 617 }
 618
 619 /*
 620 ** If there are currently more than nMaxPage pages allocated, try
 621 ** to recycle pages to reduce the number allocated to nMaxPage.
 622 */
 623 static void pcache1EnforceMaxPage(PCache1 *pCache){
 624   PGroup *pGroup = pCache->pGroup;
 625   PgHdr1 *p;
 626   assert( sqlite3_mutex_held(pGroup->mutex) );
 627   while( pGroup->nPurgeable>pGroup->nMaxPage
 628       && (p=pGroup->lru.pLruPrev)->isAnchor==0
 629   ){
 630     assert( p->pCache->pGroup==pGroup );
 631     assert( PAGE_IS_UNPINNED(p) );
 632     pcache1PinPage(p);
 633     pcache1RemoveFromHash(p, 1);
 634   }
 635   if( pCache->nPage==0 && pCache->pBulk ){
 636     sqlite3_free(pCache->pBulk);
 637     pCache->pBulk = pCache->pFree = 0;
 638   }
 639 }
 640
 641 /*
 642 ** Discard all pages from cache pCache with a page number (key value)
 643 ** greater than or equal to iLimit. Any pinned pages that meet this
 644 ** criteria are unpinned before they are discarded.
 645 **
 646 ** The PCache mutex must be held when this function is called.
 647 */
 648 static void pcache1TruncateUnsafe(
 649   PCache1 *pCache,             /* The cache to truncate */
 650   unsigned int iLimit          /* Drop pages with this pgno or larger */
 651 ){
 652   TESTONLY( int nPage = 0; )  /* To assert pCache->nPage is correct */
 653   unsigned int h, iStop;
 654   assert( sqlite3_mutex_held(pCache->pGroup->mutex) );
 655   assert( pCache->iMaxKey >= iLimit );
 656   assert( pCache->nHash > 0 );
 657   if( pCache->iMaxKey - iLimit < pCache->nHash ){
 658     /* If we are just shaving the last few pages off the end of the
 659     ** cache, then there is no point in scanning the entire hash table.
 660     ** Only scan those hash slots that might contain pages that need to
 661     ** be removed. */
 662     h = iLimit % pCache->nHash;
 663     iStop = pCache->iMaxKey % pCache->nHash;
 664     TESTONLY( nPage = -10; )  /* Disable the pCache->nPage validity check */
 665   }else{
 666     /* This is the general case where many pages are being removed.
 667     ** It is necessary to scan the entire hash table */
 668     h = pCache->nHash/2;
 669     iStop = h - 1;
 670   }
 671   for(;;){
 672     PgHdr1 **pp;
 673     PgHdr1 *pPage;
 674     assert( h<pCache->nHash );
 675     pp = &pCache->apHash[h];
 676     while( (pPage = *pp)!=0 ){
 677       if( pPage->iKey>=iLimit ){
 678         pCache->nPage--;
 679         *pp = pPage->pNext;
 680         if( PAGE_IS_UNPINNED(pPage) ) pcache1PinPage(pPage);
 681         pcache1FreePage(pPage);
 682       }else{
 683         pp = &pPage->pNext;
 684         TESTONLY( if( nPage>=0 ) nPage++; )
 685       }
 686     }
 687     if( h==iStop ) break;
 688     h = (h+1) % pCache->nHash;
 689   }
 690   assert( nPage<0 || pCache->nPage==(unsigned)nPage );
 691 }
 692
 693 /******************************************************************************/
 694 /******** sqlite3_pcache Methods **********************************************/
 695
 696 /*
 697 ** Implementation of the sqlite3_pcache.xInit method.
 698 */
 699 static int pcache1Init(void *NotUsed){
 700   UNUSED_PARAMETER(NotUsed);
 701   assert( pcache1.isInit==0 );
 702   memset(&pcache1, 0, sizeof(pcache1));
 703
 704
 705   /*
 706   ** The pcache1.separateCache variable is true if each PCache has its own
 707   ** private PGroup (mode-1).  pcache1.separateCache is false if the single
 708   ** PGroup in pcache1.grp is used for all page caches (mode-2).
 709   **
 710   **   *  Always use a unified cache (mode-2) if ENABLE_MEMORY_MANAGEMENT
 711   **
 712   **   *  Use a unified cache in single-threaded applications that have
 713   **      configured a start-time buffer for use as page-cache memory using
 714   **      sqlite3_config(SQLITE_CONFIG_PAGECACHE, pBuf, sz, N) with non-NULL
 715   **      pBuf argument.
 716   **
 717   **   *  Otherwise use separate caches (mode-1)
 718   */
 719 #if defined(SQLITE_ENABLE_MEMORY_MANAGEMENT)
 720   pcache1.separateCache = 0;
 721 #elif SQLITE_THREADSAFE
 722   pcache1.separateCache = sqlite3GlobalConfig.pPage==0
 723                           || sqlite3GlobalConfig.bCoreMutex>0;
 724 #else
 725   pcache1.separateCache = sqlite3GlobalConfig.pPage==0;
 726 #endif
 727
 728 #if SQLITE_THREADSAFE
 729   if( sqlite3GlobalConfig.bCoreMutex ){
 730     pcache1.grp.mutex = sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_LRU);
 731     pcache1.mutex = sqlite3MutexAlloc(SQLITE_MUTEX_STATIC_PMEM);
 732   }
 733 #endif
 734   if( pcache1.separateCache
 735    && sqlite3GlobalConfig.nPage!=0
 736    && sqlite3GlobalConfig.pPage==0
 737   ){
 738     pcache1.nInitPage = sqlite3GlobalConfig.nPage;
 739   }else{
 740     pcache1.nInitPage = 0;
 741   }
 742   pcache1.grp.mxPinned = 10;
 743   pcache1.isInit = 1;
 744   return SQLITE_OK;
 745 }
 746
 747 /*
 748 ** Implementation of the sqlite3_pcache.xShutdown method.
 749 ** Note that the static mutex allocated in xInit does
 750 ** not need to be freed.
 751 */
 752 static void pcache1Shutdown(void *NotUsed){
 753   UNUSED_PARAMETER(NotUsed);
 754   assert( pcache1.isInit!=0 );
 755   memset(&pcache1, 0, sizeof(pcache1));
 756 }
 757
 758 /* forward declaration */
 759 static void pcache1Destroy(sqlite3_pcache *p);
 760
 761 /*
 762 ** Implementation of the sqlite3_pcache.xCreate method.
 763 **
 764 ** Allocate a new cache.
 765 */
 766 static sqlite3_pcache *pcache1Create(int szPage, int szExtra, int bPurgeable){
 767   PCache1 *pCache;      /* The newly created page cache */
 768   PGroup *pGroup;       /* The group the new page cache will belong to */
 769   int sz;               /* Bytes of memory required to allocate the new cache */
 770
 771   assert( (szPage & (szPage-1))==0 && szPage>=512 && szPage<=65536 );
 772   assert( szExtra < 300 );
 773
 774   sz = sizeof(PCache1) + sizeof(PGroup)*pcache1.separateCache;
 775   pCache = (PCache1 *)sqlite3MallocZero(sz);
 776   if( pCache ){
 777     if( pcache1.separateCache ){
 778       pGroup = (PGroup*)&pCache[1];
 779       pGroup->mxPinned = 10;
 780     }else{
 781       pGroup = &pcache1.grp;
 782     }
 783     if( pGroup->lru.isAnchor==0 ){
 784       pGroup->lru.isAnchor = 1;
 785       pGroup->lru.pLruPrev = pGroup->lru.pLruNext = &pGroup->lru;
 786     }
 787     pCache->pGroup = pGroup;
 788     pCache->szPage = szPage;
 789     pCache->szExtra = szExtra;
 790     pCache->szAlloc = szPage + szExtra + ROUND8(sizeof(PgHdr1));
 791     pCache->bPurgeable = (bPurgeable ? 1 : 0);
 792     pcache1EnterMutex(pGroup);
 793     pcache1ResizeHash(pCache);
 794     if( bPurgeable ){
 795       pCache->nMin = 10;
 796       pGroup->nMinPage += pCache->nMin;
 797       pGroup->mxPinned = pGroup->nMaxPage + 10 - pGroup->nMinPage;
 798       pCache->pnPurgeable = &pGroup->nPurgeable;
 799     }else{
 800       pCache->pnPurgeable = &pCache->nPurgeableDummy;
 801     }
 802     pcache1LeaveMutex(pGroup);
 803     if( pCache->nHash==0 ){
 804       pcache1Destroy((sqlite3_pcache*)pCache);
 805       pCache = 0;
 806     }
 807   }
 808   return (sqlite3_pcache *)pCache;
 809 }
 810
 811 /*
 812 ** Implementation of the sqlite3_pcache.xCachesize method.
 813 **
 814 ** Configure the cache_size limit for a cache.
 815 */
 816 static void pcache1Cachesize(sqlite3_pcache *p, int nMax){
 817   PCache1 *pCache = (PCache1 *)p;
 818   if( pCache->bPurgeable ){
 819     PGroup *pGroup = pCache->pGroup;
 820     pcache1EnterMutex(pGroup);
 821     pGroup->nMaxPage += (nMax - pCache->nMax);
 822     pGroup->mxPinned = pGroup->nMaxPage + 10 - pGroup->nMinPage;
 823     pCache->nMax = nMax;
 824     pCache->n90pct = pCache->nMax*9/10;
 825     pcache1EnforceMaxPage(pCache);
 826     pcache1LeaveMutex(pGroup);
 827   }
 828 }
 829
 830 /*
 831 ** Implementation of the sqlite3_pcache.xShrink method.
 832 **
 833 ** Free up as much memory as possible.
 834 */
 835 static void pcache1Shrink(sqlite3_pcache *p){
 836   PCache1 *pCache = (PCache1*)p;
 837   if( pCache->bPurgeable ){
 838     PGroup *pGroup = pCache->pGroup;
 839     int savedMaxPage;
 840     pcache1EnterMutex(pGroup);
 841     savedMaxPage = pGroup->nMaxPage;
 842     pGroup->nMaxPage = 0;
 843     pcache1EnforceMaxPage(pCache);
 844     pGroup->nMaxPage = savedMaxPage;
 845     pcache1LeaveMutex(pGroup);
 846   }
 847 }
 848
 849 /*
 850 ** Implementation of the sqlite3_pcache.xPagecount method.
 851 */
 852 static int pcache1Pagecount(sqlite3_pcache *p){
 853   int n;
 854   PCache1 *pCache = (PCache1*)p;
 855   pcache1EnterMutex(pCache->pGroup);
 856   n = pCache->nPage;
 857   pcache1LeaveMutex(pCache->pGroup);
 858   return n;
 859 }
 860
 861
 862 /*
 863 ** Implement steps 3, 4, and 5 of the pcache1Fetch() algorithm described
 864 ** in the header of the pcache1Fetch() procedure.
 865 **
 866 ** This steps are broken out into a separate procedure because they are
 867 ** usually not needed, and by avoiding the stack initialization required
 868 ** for these steps, the main pcache1Fetch() procedure can run faster.
 869 */
 870 static SQLITE_NOINLINE PgHdr1 *pcache1FetchStage2(
 871   PCache1 *pCache,
 872   unsigned int iKey,
 873   int createFlag
 874 ){
 875   unsigned int nPinned;
 876   PGroup *pGroup = pCache->pGroup;
 877   PgHdr1 *pPage = 0;
 878
 879   /* Step 3: Abort if createFlag is 1 but the cache is nearly full */
 880   assert( pCache->nPage >= pCache->nRecyclable );
 881   nPinned = pCache->nPage - pCache->nRecyclable;
 882   assert( pGroup->mxPinned == pGroup->nMaxPage + 10 - pGroup->nMinPage );
 883   assert( pCache->n90pct == pCache->nMax*9/10 );
 884   if( createFlag==1 && (
 885         nPinned>=pGroup->mxPinned
 886      || nPinned>=pCache->n90pct
 887      || (pcache1UnderMemoryPressure(pCache) && pCache->nRecyclable<nPinned)
 888   )){
 889     return 0;
 890   }
 891
 892   if( pCache->nPage>=pCache->nHash ) pcache1ResizeHash(pCache);
 893   assert( pCache->nHash>0 && pCache->apHash );
 894
 895   /* Step 4. Try to recycle a page. */
 896   if( pCache->bPurgeable
 897    && !pGroup->lru.pLruPrev->isAnchor
 898    && ((pCache->nPage+1>=pCache->nMax) || pcache1UnderMemoryPressure(pCache))
 899   ){
 900     PCache1 *pOther;
 901     pPage = pGroup->lru.pLruPrev;
 902     assert( PAGE_IS_UNPINNED(pPage) );
 903     pcache1RemoveFromHash(pPage, 0);
 904     pcache1PinPage(pPage);
 905     pOther = pPage->pCache;
 906     if( pOther->szAlloc != pCache->szAlloc ){
 907       pcache1FreePage(pPage);
 908       pPage = 0;
 909     }else{
 910       pGroup->nPurgeable -= (pOther->bPurgeable - pCache->bPurgeable);
 911     }
 912   }
 913
 914   /* Step 5. If a usable page buffer has still not been found,
 915   ** attempt to allocate a new one.
 916   */
 917   if( !pPage ){
 918     pPage = pcache1AllocPage(pCache, createFlag==1);
 919   }
 920
 921   if( pPage ){
 922     unsigned int h = iKey % pCache->nHash;
 923     pCache->nPage++;
 924     pPage->iKey = iKey;
 925     pPage->pNext = pCache->apHash[h];
 926     pPage->pCache = pCache;
 927     pPage->pLruNext = 0;
 928     /* pPage->pLruPrev = 0;
 929     ** No need to clear pLruPrev since it is not accessed when pLruNext==0 */
 930     *(void **)pPage->page.pExtra = 0;
 931     pCache->apHash[h] = pPage;
 932     if( iKey>pCache->iMaxKey ){
 933       pCache->iMaxKey = iKey;
 934     }
 935   }
 936   return pPage;
 937 }
 938
 939 /*
 940 ** Implementation of the sqlite3_pcache.xFetch method.
 941 **
 942 ** Fetch a page by key value.
 943 **
 944 ** Whether or not a new page may be allocated by this function depends on
 945 ** the value of the createFlag argument.  0 means do not allocate a new
 946 ** page.  1 means allocate a new page if space is easily available.  2
 947 ** means to try really hard to allocate a new page.
 948 **
 949 ** For a non-purgeable cache (a cache used as the storage for an in-memory
 950 ** database) there is really no difference between createFlag 1 and 2.  So
 951 ** the calling function (pcache.c) will never have a createFlag of 1 on
 952 ** a non-purgeable cache.
 953 **
 954 ** There are three different approaches to obtaining space for a page,
 955 ** depending on the value of parameter createFlag (which may be 0, 1 or 2).
 956 **
 957 **   1. Regardless of the value of createFlag, the cache is searched for a
 958 **      copy of the requested page. If one is found, it is returned.
 959 **
 960 **   2. If createFlag==0 and the page is not already in the cache, NULL is
 961 **      returned.
 962 **
 963 **   3. If createFlag is 1, and the page is not already in the cache, then
 964 **      return NULL (do not allocate a new page) if any of the following
 965 **      conditions are true:
 966 **
 967 **       (a) the number of pages pinned by the cache is greater than
 968 **           PCache1.nMax, or
 969 **
 970 **       (b) the number of pages pinned by the cache is greater than
 971 **           the sum of nMax for all purgeable caches, less the sum of
 972 **           nMin for all other purgeable caches, or
 973 **
 974 **   4. If none of the first three conditions apply and the cache is marked
 975 **      as purgeable, and if one of the following is true:
 976 **
 977 **       (a) The number of pages allocated for the cache is already
 978 **           PCache1.nMax, or
 979 **
 980 **       (b) The number of pages allocated for all purgeable caches is
 981 **           already equal to or greater than the sum of nMax for all
 982 **           purgeable caches,
 983 **
 984 **       (c) The system is under memory pressure and wants to avoid
 985 **           unnecessary pages cache entry allocations
 986 **
 987 **      then attempt to recycle a page from the LRU list. If it is the right
 988 **      size, return the recycled buffer. Otherwise, free the buffer and
 989 **      proceed to step 5.
 990 **
 991 **   5. Otherwise, allocate and return a new page buffer.
 992 **
 993 ** There are two versions of this routine.  pcache1FetchWithMutex() is
 994 ** the general case.  pcache1FetchNoMutex() is a faster implementation for
 995 ** the common case where pGroup->mutex is NULL.  The pcache1Fetch() wrapper
 996 ** invokes the appropriate routine.
 997 */
 998 static PgHdr1 *pcache1FetchNoMutex(
 999   sqlite3_pcache *p,
1000   unsigned int iKey,
1001   int createFlag
1002 ){
1003   PCache1 *pCache = (PCache1 *)p;
1004   PgHdr1 *pPage = 0;
1005
1006   /* Step 1: Search the hash table for an existing entry. */
1007   pPage = pCache->apHash[iKey % pCache->nHash];
1008   while( pPage && pPage->iKey!=iKey ){ pPage = pPage->pNext; }
1009
1010   /* Step 2: If the page was found in the hash table, then return it.
1011   ** If the page was not in the hash table and createFlag is 0, abort.
1012   ** Otherwise (page not in hash and createFlag!=0) continue with
1013   ** subsequent steps to try to create the page. */
1014   if( pPage ){
1015     if( PAGE_IS_UNPINNED(pPage) ){
1016       return pcache1PinPage(pPage);
1017     }else{
1018       return pPage;
1019     }
1020   }else if( createFlag ){
1021     /* Steps 3, 4, and 5 implemented by this subroutine */
1022     return pcache1FetchStage2(pCache, iKey, createFlag);
1023   }else{
1024     return 0;
1025   }
1026 }
1027 #if PCACHE1_MIGHT_USE_GROUP_MUTEX
1028 static PgHdr1 *pcache1FetchWithMutex(
1029   sqlite3_pcache *p,
1030   unsigned int iKey,
1031   int createFlag
1032 ){
1033   PCache1 *pCache = (PCache1 *)p;
1034   PgHdr1 *pPage;
1035
1036   pcache1EnterMutex(pCache->pGroup);
1037   pPage = pcache1FetchNoMutex(p, iKey, createFlag);
1038   assert( pPage==0 || pCache->iMaxKey>=iKey );
1039   pcache1LeaveMutex(pCache->pGroup);
1040   return pPage;
1041 }
1042 #endif
1043 static sqlite3_pcache_page *pcache1Fetch(
1044   sqlite3_pcache *p,
1045   unsigned int iKey,
1046   int createFlag
1047 ){
1048 #if PCACHE1_MIGHT_USE_GROUP_MUTEX || defined(SQLITE_DEBUG)
1049   PCache1 *pCache = (PCache1 *)p;
1050 #endif
1051
1052   assert( offsetof(PgHdr1,page)==0 );
1053   assert( pCache->bPurgeable || createFlag!=1 );
1054   assert( pCache->bPurgeable || pCache->nMin==0 );
1055   assert( pCache->bPurgeable==0 || pCache->nMin==10 );
1056   assert( pCache->nMin==0 || pCache->bPurgeable );
1057   assert( pCache->nHash>0 );
1058 #if PCACHE1_MIGHT_USE_GROUP_MUTEX
1059   if( pCache->pGroup->mutex ){
1060     return (sqlite3_pcache_page*)pcache1FetchWithMutex(p, iKey, createFlag);
1061   }else
1062 #endif
1063   {
1064     return (sqlite3_pcache_page*)pcache1FetchNoMutex(p, iKey, createFlag);
1065   }
1066 }
1067
1068
1069 /*
1070 ** Implementation of the sqlite3_pcache.xUnpin method.
1071 **
1072 ** Mark a page as unpinned (eligible for asynchronous recycling).
1073 */
1074 static void pcache1Unpin(
1075   sqlite3_pcache *p,
1076   sqlite3_pcache_page *pPg,
1077   int reuseUnlikely
1078 ){
1079   PCache1 *pCache = (PCache1 *)p;
1080   PgHdr1 *pPage = (PgHdr1 *)pPg;
1081   PGroup *pGroup = pCache->pGroup;
1082
1083   assert( pPage->pCache==pCache );
1084   pcache1EnterMutex(pGroup);
1085
1086   /* It is an error to call this function if the page is already
1087   ** part of the PGroup LRU list.
1088   */
1089   assert( pPage->pLruNext==0 );
1090   assert( PAGE_IS_PINNED(pPage) );
1091
1092   if( reuseUnlikely || pGroup->nPurgeable>pGroup->nMaxPage ){
1093     pcache1RemoveFromHash(pPage, 1);
1094   }else{
1095     /* Add the page to the PGroup LRU list. */
1096     PgHdr1 **ppFirst = &pGroup->lru.pLruNext;
1097     pPage->pLruPrev = &pGroup->lru;
1098     (pPage->pLruNext = *ppFirst)->pLruPrev = pPage;
1099     *ppFirst = pPage;
1100     pCache->nRecyclable++;
1101   }
1102
1103   pcache1LeaveMutex(pCache->pGroup);
1104 }
1105
1106 /*
1107 ** Implementation of the sqlite3_pcache.xRekey method.
1108 */
1109 static void pcache1Rekey(
1110   sqlite3_pcache *p,
1111   sqlite3_pcache_page *pPg,
1112   unsigned int iOld,
1113   unsigned int iNew
1114 ){
1115   PCache1 *pCache = (PCache1 *)p;
1116   PgHdr1 *pPage = (PgHdr1 *)pPg;
1117   PgHdr1 **pp;
1118   unsigned int h;
1119   assert( pPage->iKey==iOld );
1120   assert( pPage->pCache==pCache );
1121
1122   pcache1EnterMutex(pCache->pGroup);
1123
1124   h = iOld%pCache->nHash;
1125   pp = &pCache->apHash[h];
1126   while( (*pp)!=pPage ){
1127     pp = &(*pp)->pNext;
1128   }
1129   *pp = pPage->pNext;
1130
1131   h = iNew%pCache->nHash;
1132   pPage->iKey = iNew;
1133   pPage->pNext = pCache->apHash[h];
1134   pCache->apHash[h] = pPage;
1135   if( iNew>pCache->iMaxKey ){
1136     pCache->iMaxKey = iNew;
1137   }
1138
1139   pcache1LeaveMutex(pCache->pGroup);
1140 }
1141
1142 /*
1143 ** Implementation of the sqlite3_pcache.xTruncate method.
1144 **
1145 ** Discard all unpinned pages in the cache with a page number equal to
1146 ** or greater than parameter iLimit. Any pinned pages with a page number
1147 ** equal to or greater than iLimit are implicitly unpinned.
1148 */
1149 static void pcache1Truncate(sqlite3_pcache *p, unsigned int iLimit){
1150   PCache1 *pCache = (PCache1 *)p;
1151   pcache1EnterMutex(pCache->pGroup);
1152   if( iLimit<=pCache->iMaxKey ){
1153     pcache1TruncateUnsafe(pCache, iLimit);
1154     pCache->iMaxKey = iLimit-1;
1155   }
1156   pcache1LeaveMutex(pCache->pGroup);
1157 }
1158
1159 /*
1160 ** Implementation of the sqlite3_pcache.xDestroy method.
1161 **
1162 ** Destroy a cache allocated using pcache1Create().
1163 */
1164 static void pcache1Destroy(sqlite3_pcache *p){
1165   PCache1 *pCache = (PCache1 *)p;
1166   PGroup *pGroup = pCache->pGroup;
1167   assert( pCache->bPurgeable || (pCache->nMax==0 && pCache->nMin==0) );
1168   pcache1EnterMutex(pGroup);
1169   if( pCache->nPage ) pcache1TruncateUnsafe(pCache, 0);
1170   assert( pGroup->nMaxPage >= pCache->nMax );
1171   pGroup->nMaxPage -= pCache->nMax;
1172   assert( pGroup->nMinPage >= pCache->nMin );
1173   pGroup->nMinPage -= pCache->nMin;
1174   pGroup->mxPinned = pGroup->nMaxPage + 10 - pGroup->nMinPage;
1175   pcache1EnforceMaxPage(pCache);
1176   pcache1LeaveMutex(pGroup);
1177   sqlite3_free(pCache->pBulk);
1178   sqlite3_free(pCache->apHash);
1179   sqlite3_free(pCache);
1180 }
1181
1182 /*
1183 ** This function is called during initialization (sqlite3_initialize()) to
1184 ** install the default pluggable cache module, assuming the user has not
1185 ** already provided an alternative.
1186 */
1187 void sqlite3PCacheSetDefault(void){
1188   static const sqlite3_pcache_methods2 defaultMethods = {
1189     1,                       /* iVersion */
1190     0,                       /* pArg */
1191     pcache1Init,             /* xInit */
1192     pcache1Shutdown,         /* xShutdown */
1193     pcache1Create,           /* xCreate */
1194     pcache1Cachesize,        /* xCachesize */
1195     pcache1Pagecount,        /* xPagecount */
1196     pcache1Fetch,            /* xFetch */
1197     pcache1Unpin,            /* xUnpin */
1198     pcache1Rekey,            /* xRekey */
1199     pcache1Truncate,         /* xTruncate */
1200     pcache1Destroy,          /* xDestroy */
1201     pcache1Shrink            /* xShrink */
1202   };
1203   sqlite3_config(SQLITE_CONFIG_PCACHE2, &defaultMethods);
1204 }
1205
1206 /*
1207 ** Return the size of the header on each page of this PCACHE implementation.
1208 */
1209 int sqlite3HeaderSizePcache1(void){ return ROUND8(sizeof(PgHdr1)); }
1210
1211 /*
1212 ** Return the global mutex used by this PCACHE implementation.  The
1213 ** sqlite3_status() routine needs access to this mutex.
1214 */
1215 sqlite3_mutex *sqlite3Pcache1Mutex(void){
1216   return pcache1.mutex;
1217 }
1218
1219 #ifdef SQLITE_ENABLE_MEMORY_MANAGEMENT
1220 /*
1221 ** This function is called to free superfluous dynamically allocated memory
1222 ** held by the pager system. Memory in use by any SQLite pager allocated
1223 ** by the current thread may be sqlite3_free()ed.
1224 **
1225 ** nReq is the number of bytes of memory required. Once this much has
1226 ** been released, the function returns. The return value is the total number
1227 ** of bytes of memory released.
1228 */
1229 int sqlite3PcacheReleaseMemory(int nReq){
1230   int nFree = 0;
1231   assert( sqlite3_mutex_notheld(pcache1.grp.mutex) );
1232   assert( sqlite3_mutex_notheld(pcache1.mutex) );
1233   if( sqlite3GlobalConfig.pPage==0 ){
1234     PgHdr1 *p;
1235     pcache1EnterMutex(&pcache1.grp);
1236     while( (nReq<0 || nFree<nReq)
1237        &&  (p=pcache1.grp.lru.pLruPrev)!=0
1238        &&  p->isAnchor==0
1239     ){
1240       nFree += pcache1MemSize(p->page.pBuf);
1241 #ifdef SQLITE_PCACHE_SEPARATE_HEADER
1242       nFree += sqlite3MemSize(p);
1243 #endif
1244       assert( PAGE_IS_UNPINNED(p) );
1245       pcache1PinPage(p);
1246       pcache1RemoveFromHash(p, 1);
1247     }
1248     pcache1LeaveMutex(&pcache1.grp);
1249   }
1250   return nFree;
1251 }
1252 #endif /* SQLITE_ENABLE_MEMORY_MANAGEMENT */
1253
1254 #ifdef SQLITE_TEST
1255 /*
1256 ** This function is used by test procedures to inspect the internal state
1257 ** of the global cache.
1258 */
1259 void sqlite3PcacheStats(
1260   int *pnCurrent,      /* OUT: Total number of pages cached */
1261   int *pnMax,          /* OUT: Global maximum cache size */
1262   int *pnMin,          /* OUT: Sum of PCache1.nMin for purgeable caches */
1263   int *pnRecyclable    /* OUT: Total number of pages available for recycling */
1264 ){
1265   PgHdr1 *p;
1266   int nRecyclable = 0;
1267   for(p=pcache1.grp.lru.pLruNext; p && !p->isAnchor; p=p->pLruNext){
1268     assert( PAGE_IS_UNPINNED(p) );
1269     nRecyclable++;
1270   }
1271   *pnCurrent = pcache1.grp.nPurgeable;
1272   *pnMax = (int)pcache1.grp.nMaxPage;
1273   *pnMin = (int)pcache1.grp.nMinPage;
1274   *pnRecyclable = nRecyclable;
1275 }
1276 #endif