Use mmap() to read from the database file in rollback mode. This branch is unix only for now.

FossilOrigin-Name: 6f21d9cbf5d457e63a7282015a89ae785526cf6d

manifest

@@ -1,5 +1,5 @@
-C Enhance\stests\sfor\sticket\s[4dd95f6943].
-D 2013-03-13T07:02:04.083
+C Use\smmap()\sto\sread\sfrom\sthe\sdatabase\sfile\sin\srollback\smode.\sThis\sbranch\sis\sunix\sonly\sfor\snow.
+D 2013-03-14T18:34:37.796
 F Makefile.arm-wince-mingw32ce-gcc d6df77f1f48d690bd73162294bbba7f59507c72f
 F Makefile.in 9a804abbd3cae82d196e4d33aba13239e32522a5
 F Makefile.linux-gcc 91d710bdc4998cb015f39edf3cb314ec4f4d7e23
@@ -121,7 +121,7 @@ F src/auth.c 523da7fb4979469955d822ff9298352d6b31de34
 F src/backup.c b2cac9f7993f3f9588827b824b1501d0c820fa68
 F src/bitvec.c 26675fe8e431dc555e6f2d0e11e651d172234aa1
 F src/btmutex.c 976f45a12e37293e32cae0281b15a21d48a8aaa7
-F src/btree.c 746c4dafae8565b3be6fb9ce3bb1fa9f1e67cc59
+F src/btree.c c1a956c6762f2a45188c945e1070daec29f5253f
 F src/btree.h 3ad7964d6c5b1c7bff569aab6adfa075f8bf06cd
 F src/btreeInt.h eecc84f02375b2bb7a44abbcbbe3747dde73edb2
 F src/build.c 375e5df716e03b9343c5e1211be3b24e6d6dff05
@@ -160,13 +160,13 @@ F src/notify.c 976dd0f6171d4588e89e874fcc765e92914b6d30
 F src/os.c e1acdc09ff3ac2412945cca9766e2dcf4675f31c
 F src/os.h 027491c77d2404c0a678bb3fb06286f331eb9b57
 F src/os_common.h 92815ed65f805560b66166e3583470ff94478f04
-F src/os_unix.c f6387eef0cf8f6b808738f4f3aa47e6132af0940
+F src/os_unix.c 2a4cd96aabf413f39cf562baebb27aa9993f6f54
 F src/os_win.c f7da4dc0a2545c0a430080380809946ae4d676d6
-F src/pager.c 582f8da52d0bd4b43d3bdaeba0ea7702c1f23702
+F src/pager.c 4e7e66c2959ee43caf8a7000712d6a4121b0888a
 F src/pager.h 1109a06578ec5574dc2c74cf8d9f69daf36fe3e0
 F src/parse.y 5d5e12772845805fdfeb889163516b84fbb9ae95
 F src/pcache.c f8043b433a57aba85384a531e3937a804432a346
-F src/pcache.h 1b5dcc3dc8103d03e625b177023ee67764fa6b7c
+F src/pcache.h a5e4f5d9f5d592051d91212c5949517971ae6222
 F src/pcache1.c 9fd22671c270b35131ef480bbc00392b8b5f8ab9
 F src/pragma.c 9f0ee3d74a7f33eeeff40a4b014fc3abf8182ce2
 F src/prepare.c 78cd7ae8cd0d8de8ef8a8b5352fc5f38693a0852
@@ -176,7 +176,7 @@ F src/resolve.c 9079da7d59aed2bb14ec8315bc7f720dd85b5b65
 F src/rowset.c 64655f1a627c9c212d9ab497899e7424a34222e0
 F src/select.c e1c6f6abdf9f359f4e735cb8ae11d2f359bf52a9
 F src/shell.c 7c41bfcd9e5bf9d96b9215f79b03a5b2b44a3bca
-F src/sqlite.h.in f2fa32f440dda59ca47e22889966b2a6eb3b491c
+F src/sqlite.h.in 9a5c737a1feb4495d351c56883587d4fda52e81e
 F src/sqlite3.rc fea433eb0a59f4c9393c8e6d76a6e2596b1fe0c0
 F src/sqlite3ext.h 7183ab832e23db0f934494f16928da127a571d75
 F src/sqliteInt.h 0f8f05ee4db4ba9120b38f7a3992b325698f6e8a
@@ -656,7 +656,7 @@ F test/pageropt.test 9191867ed19a2b3db6c42d1b36b6fbc657cd1ab0
 F test/pagesize.test 1dd51367e752e742f58e861e65ed7390603827a0
 F test/pcache.test 065aa286e722ab24f2e51792c1f093bf60656b16
 F test/pcache2.test a83efe2dec0d392f814bfc998def1d1833942025
-F test/permutations.test 360b92859c0af814b3fe10b68746936389606501
+F test/permutations.test eb49937dca270b2c3f62d4c91fc7034ca905b7f1
 F test/pragma.test 60d29cd3d8098a2c20bf4c072810f99e3bf2757a
 F test/pragma2.test 3a55f82b954242c642f8342b17dffc8b47472947
 F test/printf.test ec9870c4dce8686a37818e0bf1aba6e6a1863552
@@ -1038,7 +1038,10 @@ F tool/vdbe-compress.tcl f12c884766bd14277f4fcedcae07078011717381
 F tool/warnings-clang.sh f6aa929dc20ef1f856af04a730772f59283631d4
 F tool/warnings.sh fbc018d67fd7395f440c28f33ef0f94420226381
 F tool/win/sqlite.vsix 97894c2790eda7b5bce3cc79cb2a8ec2fde9b3ac
-P 839aa91faf1db7025d90fa3c65e50efb829b053b
-R ad71abb6cd4a6ebc6857153126653bad
+P 0b452734faa0839c817f040322e7733e423bfce2
+R d4c281f6a900c9b41816881ffb3c2331
+T *branch * experimental-mmap
+T *sym-experimental-mmap *
+T -sym-trunk *
 U dan
-Z 69b39b8becc09fc4758d1b797bf5f4f2
+Z f16a594407ec9e95c26319dfbc751443

manifest.uuid

@@ -1 +1 @@
-0b452734faa0839c817f040322e7733e423bfce2
+6f21d9cbf5d457e63a7282015a89ae785526cf6d

src/btree.c

@@ -2564,6 +2564,34 @@ int sqlite3BtreeNewDb(Btree *p){
   return rc;
 }
 
+/*
+** If the shared-btree passed as the only argument is holding references
+** to mmap pages, replace them with read/write pages. Return SQLITE_OK
+** if successful, or an error code otherwise.
+*/
+static int btreeSwapOutMmap(BtShared *pBt){
+  BtCursor *pCsr;
+  for(pCsr=pBt->pCursor; pCsr; pCsr=pCsr->pNext){
+    int i;
+    for(i=0; i<=pCsr->iPage; i++){
+      MemPage *pPg = pCsr->apPage[i];
+      if( pPg->pDbPage->flags & PGHDR_MMAP ){
+        int rc;
+        MemPage *pNew = 0;
+        rc = btreeGetPage(pBt, pPg->pgno, &pNew, 0);
+        if( rc==SQLITE_OK && i==pCsr->iPage ){
+          pCsr->info.pCell = pNew->aData + (pCsr->info.pCell - pPg->aData);
+        }
+        pCsr->apPage[i] = pNew;
+        releasePage(pPg);
+        if( rc!=SQLITE_OK ) return rc;
+      }
+    }
+  }
+
+  return SQLITE_OK;
+}
+
 /*
 ** Attempt to start a new transaction. A write-transaction
 ** is started if the second argument is nonzero, otherwise a read-
@@ -2670,6 +2698,9 @@ int sqlite3BtreeBeginTrans(Btree *p, int wrflag){
         rc = SQLITE_READONLY;
       }else{
         rc = sqlite3PagerBegin(pBt->pPager,wrflag>1,sqlite3TempInMemory(p->db));
+        if( rc==SQLITE_OK ){
+          rc = btreeSwapOutMmap(pBt);
+        }
         if( rc==SQLITE_OK ){
           rc = newDatabase(pBt);
         }

src/os_unix.c

@@ -3623,6 +3623,10 @@ static int unixFileControl(sqlite3_file *id, int op, void *pArg){
       }
       return SQLITE_OK;
     }
+    case SQLITE_FCNTL_GETFD: {
+      *(int*)pArg = pFile->h;
+      return SQLITE_OK;
+    }
 #ifdef SQLITE_DEBUG
     /* The pager calls this method to signal that it has done
     ** a rollback and that the database is therefore unchanged and

src/pager.c

@@ -655,6 +655,11 @@ struct Pager {
   PagerSavepoint *aSavepoint; /* Array of active savepoints */
   int nSavepoint;             /* Number of elements in aSavepoint[] */
   char dbFileVers[16];        /* Changes whenever database file changes */
+
+  void *pMap;                 /* Memory mapped prefix of database file */
+  i64 nMap;                   /* Size of mapping at pMap in bytes */ 
+  int nMmapOut;               /* Number of mmap pages currently outstanding */
+  PgHdr *pFree;               /* List of free mmap page headers (pDirty) */
   /*
   ** End of the routinely-changing class members
   ***************************************************************************/
@@ -3803,6 +3808,105 @@ static int pagerSyncHotJournal(Pager *pPager){
   return rc;
 }
 
+#include <sys/mman.h>
+
+/*
+** Unmap any mapping of the database file.
+*/
+static int pagerUnmap(Pager *pPager){
+  if( pPager->pMap ){
+    munmap(pPager->pMap, pPager->nMap);
+    pPager->pMap = 0;
+    pPager->nMap = 0;
+  }
+  return SQLITE_OK;
+}
+
+static int pagerMap(Pager *pPager){
+  int rc;
+  i64 sz = 0;
+
+  assert( pPager->pMap==0 && pPager->nMap==0 );
+
+  rc = sqlite3OsFileSize(pPager->fd, &sz);
+  if( rc==SQLITE_OK && sz>0 ){
+    int fd;
+    rc = sqlite3OsFileControl(pPager->fd, SQLITE_FCNTL_GETFD, (void *)&fd);
+    if( rc==SQLITE_OK ){
+      void *pMap = mmap(0, sz, PROT_READ, MAP_SHARED, fd, 0);
+      if( pMap==MAP_FAILED ){
+      assert( 0 );
+        return SQLITE_IOERR;
+      }
+      pPager->pMap = pMap;
+      pPager->nMap = sz;
+    }
+  }
+
+  return rc;
+}
+
+static int pagerAcquireMapPage(Pager *pPager, Pgno pgno, PgHdr **ppPage){
+  int rc;
+  *ppPage = 0;
+
+  assert( pPager->pWal==0 );
+
+  if( MEMDB==0 && pPager->tempFile==0 ){
+    if( pPager->pMap==0 ){
+      rc = pagerMap(pPager);
+      if( rc!=SQLITE_OK ) return rc;
+    }
+
+    if( pgno!=1 && pPager->pMap && pPager->nMap>=((i64)pgno*pPager->pageSize) ){
+      PgHdr *p;
+      if( pPager->pFree ){
+        p = pPager->pFree;
+        pPager->pFree = p->pDirty;
+        p->pDirty = 0;
+        memset(p->pExtra, 0, pPager->nExtra);
+      }else{
+        p = (PgHdr *)sqlite3MallocZero(sizeof(PgHdr) + pPager->nExtra);
+        if( p==0 ) return SQLITE_NOMEM;
+        p->pExtra = (void *)&p[1];
+        p->flags = PGHDR_MMAP;
+        p->nRef = 1;
+        p->pPager = pPager;
+      }
+
+      assert( p->pExtra==(void *)&p[1] );
+      assert( p->pPage==0 );
+      assert( p->flags==PGHDR_MMAP );
+      assert( p->pPager==pPager );
+      assert( p->nRef==1 );
+
+      p->pData = &((u8 *)pPager->pMap)[(i64)(pgno-1) * pPager->pageSize];
+      p->pgno = pgno;
+      pPager->nMmapOut++;
+      *ppPage = p;
+    }
+  }
+
+  return SQLITE_OK;
+}
+
+static void pagerReleaseMapPage(PgHdr *pPg){
+  Pager *pPager = pPg->pPager;
+  pPager->nMmapOut--;
+  pPg->pDirty = pPager->pFree;
+  pPager->pFree = pPg;
+}
+
+static void pagerFreeMapHdrs(Pager *pPager){
+  PgHdr *p;
+  PgHdr *pNext;
+  for(p=pPager->pFree; p; p=pNext){
+    pNext = p->pDirty;
+    sqlite3_free(p);
+  }
+}
+
+
 /*
 ** Shutdown the page cache.  Free all memory and close all files.
 **
@@ -3823,6 +3927,8 @@ int sqlite3PagerClose(Pager *pPager){
   assert( assert_pager_state(pPager) );
   disable_simulated_io_errors();
   sqlite3BeginBenignMalloc();
+  pagerUnmap(pPager);
+  pagerFreeMapHdrs(pPager);
   /* pPager->errCode = 0; */
   pPager->exclusiveMode = 0;
 #ifndef SQLITE_OMIT_WAL
@@ -4966,6 +5072,7 @@ int sqlite3PagerSharedLock(Pager *pPager){
 
       if( memcmp(pPager->dbFileVers, dbFileVers, sizeof(dbFileVers))!=0 ){
         pager_reset(pPager);
+        pagerUnmap(pPager);
       }
     }
 
@@ -5007,7 +5114,7 @@ int sqlite3PagerSharedLock(Pager *pPager){
 ** nothing to rollback, so this routine is a no-op.
 */ 
 static void pagerUnlockIfUnused(Pager *pPager){
-  if( (sqlite3PcacheRefCount(pPager->pPCache)==0) ){
+  if( (sqlite3PcacheRefCount(pPager->pPCache)==0) && pPager->nMmapOut==0 ){
     pagerUnlockAndRollback(pPager);
   }
 }
@@ -5083,6 +5190,15 @@ int sqlite3PagerAcquire(
   if( pPager->errCode!=SQLITE_OK ){
     rc = pPager->errCode;
   }else{
+    if( pPager->eState==PAGER_READER && pPager->pWal==0 ){
+      rc = pagerAcquireMapPage(pPager, pgno, &pPg);
+      if( rc!=SQLITE_OK ) goto pager_acquire_err;
+      if( pPg ){
+        *ppPage = pPg;
+        return SQLITE_OK;
+      }
+    }
+
     rc = sqlite3PcacheFetch(pPager->pPCache, pgno, 1, ppPage);
   }
 
@@ -5196,7 +5312,11 @@ DbPage *sqlite3PagerLookup(Pager *pPager, Pgno pgno){
 void sqlite3PagerUnref(DbPage *pPg){
   if( pPg ){
     Pager *pPager = pPg->pPager;
-    sqlite3PcacheRelease(pPg);
+    if( pPg->flags & PGHDR_MMAP ){
+      pagerReleaseMapPage(pPg);
+    }else{
+      sqlite3PcacheRelease(pPg);
+    }
     pagerUnlockIfUnused(pPager);
   }
 }
@@ -5313,6 +5433,8 @@ int sqlite3PagerBegin(Pager *pPager, int exFlag, int subjInMemory){
   assert( pPager->eState>=PAGER_READER && pPager->eState<PAGER_ERROR );
   pPager->subjInMemory = (u8)subjInMemory;
 
+  pagerUnmap(pPager);
+
   if( ALWAYS(pPager->eState==PAGER_READER) ){
     assert( pPager->pInJournal==0 );
 
@@ -5535,6 +5657,9 @@ int sqlite3PagerWrite(DbPage *pDbPage){
   assert( pPager->eState!=PAGER_ERROR );
   assert( assert_pager_state(pPager) );
 
+  /* There must not be any outstanding mmap pages at this point */
+  assert( pPager->nMmapOut==0 );
+
   if( nPagePerSector>1 ){
     Pgno nPageCount;          /* Total number of pages in database file */
     Pgno pg1;                 /* First page of the sector pPg is located on. */

src/pcache.h

@@ -53,6 +53,8 @@ struct PgHdr {
 #define PGHDR_REUSE_UNLIKELY    0x010  /* A hint that reuse is unlikely */
 #define PGHDR_DONT_WRITE        0x020  /* Do not write content to disk */
 
+#define PGHDR_MMAP              0x040  /* This is an mmap page object */
+
 /* Initialize and shutdown the page cache subsystem */
 int sqlite3PcacheInitialize(void);
 void sqlite3PcacheShutdown(void);

src/sqlite.h.in

@@ -900,6 +900,7 @@ struct sqlite3_io_methods {
 #define SQLITE_FCNTL_PRAGMA                 14
 #define SQLITE_FCNTL_BUSYHANDLER            15
 #define SQLITE_FCNTL_TEMPFILENAME           16
+#define SQLITE_FCNTL_GETFD                  17
 
 /*
 ** CAPI3REF: Mutex Handle

test/permutations.test

@@ -135,7 +135,8 @@ test_suite "veryquick" -prefix "" -description {
   This test suite is the same as the "quick" tests, except that some files
   that test malloc and IO errors are omitted.
 } -files [
-  test_set $allquicktests -exclude *malloc* *ioerr* *fault*
+  test_set $allquicktests -exclude *malloc* *ioerr* *fault* \
+  multiplex* server1.test shared2.test shared6.test
 ]
 
 test_suite "valgrind" -prefix "" -description {