1 files changed, 106 insertions, 46 deletions

diff --git a/src/pager.c b/src/pager.c
index aee50e531..51661044d 100644
--- a/src/pager.c
+++ b/src/pager.c
@@ -947,6 +947,7 @@ static int assert_pager_state(Pager *p){
       assert( isOpen(p->jfd) 
            || p->journalMode==PAGER_JOURNALMODE_OFF 
            || p->journalMode==PAGER_JOURNALMODE_WAL 
+           || (sqlite3OsDeviceCharacteristics(p->fd)&SQLITE_IOCAP_BATCH_ATOMIC)
       );
       assert( pPager->dbOrigSize<=pPager->dbHintSize );
       break;
@@ -958,6 +959,7 @@ static int assert_pager_state(Pager *p){
       assert( isOpen(p->jfd) 
            || p->journalMode==PAGER_JOURNALMODE_OFF 
            || p->journalMode==PAGER_JOURNALMODE_WAL 
+           || (sqlite3OsDeviceCharacteristics(p->fd)&SQLITE_IOCAP_BATCH_ATOMIC)
       );
       break;
 
@@ -1168,34 +1170,45 @@ static int pagerLockDb(Pager *pPager, int eLock){
 }
 
 /*
-** This function determines whether or not the atomic-write optimization
-** can be used with this pager. The optimization can be used if:
+** This function determines whether or not the atomic-write or
+** atomic-batch-write optimizations can be used with this pager. The
+** atomic-write optimization can be used if:
 **
 **  (a) the value returned by OsDeviceCharacteristics() indicates that
 **      a database page may be written atomically, and
 **  (b) the value returned by OsSectorSize() is less than or equal
 **      to the page size.
 **
-** The optimization is also always enabled for temporary files. It is
-** an error to call this function if pPager is opened on an in-memory
-** database.
+** If it can be used, then the value returned is the size of the journal 
+** file when it contains rollback data for exactly one page.
 **
-** If the optimization cannot be used, 0 is returned. If it can be used,
-** then the value returned is the size of the journal file when it
-** contains rollback data for exactly one page.
+** The atomic-batch-write optimization can be used if OsDeviceCharacteristics()
+** returns a value with the SQLITE_IOCAP_BATCH_ATOMIC bit set. -1 is
+** returned in this case.
+**
+** If neither optimization can be used, 0 is returned.
 */
-#ifdef SQLITE_ENABLE_ATOMIC_WRITE
 static int jrnlBufferSize(Pager *pPager){
   assert( !MEMDB );
-  if( !pPager->tempFile ){
-    int dc;                           /* Device characteristics */
-    int nSector;                      /* Sector size */
-    int szPage;                       /* Page size */
 
-    assert( isOpen(pPager->fd) );
-    dc = sqlite3OsDeviceCharacteristics(pPager->fd);
-    nSector = pPager->sectorSize;
-    szPage = pPager->pageSize;
+#if defined(SQLITE_ENABLE_ATOMIC_WRITE) \
+ || defined(SQLITE_ENABLE_BATCH_ATOMIC_WRITE)
+  int dc;                           /* Device characteristics */
+
+  assert( isOpen(pPager->fd) );
+  dc = sqlite3OsDeviceCharacteristics(pPager->fd);
+#endif
+
+#ifdef SQLITE_ENABLE_BATCH_ATOMIC_WRITE
+  if( dc&SQLITE_IOCAP_BATCH_ATOMIC ){
+    return -1;
+  }
+#endif
+
+#ifdef SQLITE_ENABLE_ATOMIC_WRITE
+  {
+    int nSector = pPager->sectorSize;
+    int szPage = pPager->pageSize;
 
     assert(SQLITE_IOCAP_ATOMIC512==(512>>8));
     assert(SQLITE_IOCAP_ATOMIC64K==(65536>>8));
@@ -1205,11 +1218,11 @@ static int jrnlBufferSize(Pager *pPager){
   }
 
   return JOURNAL_HDR_SZ(pPager) + JOURNAL_PG_SZ(pPager);
-}
-#else
-# define jrnlBufferSize(x) 0
 #endif
 
+  return 0;
+}
+
 /*
 ** If SQLITE_CHECK_PAGES is defined then we do some sanity checking
 ** on the cache using a hash function.  This is used for testing
@@ -2012,7 +2025,9 @@ static int pager_end_transaction(Pager *pPager, int hasMaster, int bCommit){
   }
 
   releaseAllSavepoints(pPager);
-  assert( isOpen(pPager->jfd) || pPager->pInJournal==0 );
+  assert( isOpen(pPager->jfd) || pPager->pInJournal==0 
+      || (sqlite3OsDeviceCharacteristics(pPager->fd)&SQLITE_IOCAP_BATCH_ATOMIC)
+  );
   if( isOpen(pPager->jfd) ){
     assert( !pagerUseWal(pPager) );
 
@@ -4567,6 +4582,13 @@ static int pagerStress(void *p, PgHdr *pPg){
       rc = pagerWalFrames(pPager, pPg, 0, 0);
     }
   }else{
+    
+#ifdef SQLITE_ENABLE_BATCH_ATOMIC_WRITE
+    if( pPager->tempFile==0 ){
+      rc = sqlite3JournalCreate(pPager->jfd);
+      if( rc!=SQLITE_OK ) return pager_error(pPager, rc);
+    }
+#endif
   
     /* Sync the journal file if required. */
     if( pPg->flags&PGHDR_NEED_SYNC 
@@ -6347,6 +6369,21 @@ int sqlite3PagerCommitPhaseOne(
         sqlite3PcacheCleanAll(pPager->pPCache);
       }
     }else{
+      /* The bBatch boolean is true if the batch-atomic-write commit method
+      ** should be used.  No rollback journal is created if batch-atomic-write
+      ** is enabled.
+      */
+      sqlite3_file *fd = pPager->fd;
+#ifdef SQLITE_ENABLE_BATCH_ATOMIC_WRITE
+      const int bBatch = zMaster==0    /* An SQLITE_IOCAP_BATCH_ATOMIC commit */
+        && (sqlite3OsDeviceCharacteristics(fd) & SQLITE_IOCAP_BATCH_ATOMIC)
+        && !pPager->noSync
+        && sqlite3JournalIsInMemory(pPager->jfd);
+#else
+# define bBatch 0
+#endif
+
+#ifdef SQLITE_ENABLE_ATOMIC_WRITE
       /* The following block updates the change-counter. Exactly how it
       ** does this depends on whether or not the atomic-update optimization
       ** was enabled at compile time, and if this transaction meets the 
@@ -6370,33 +6407,40 @@ int sqlite3PagerCommitPhaseOne(
       ** in 'direct' mode. In this case the journal file will never be
       ** created for this transaction.
       */
-  #ifdef SQLITE_ENABLE_ATOMIC_WRITE
-      PgHdr *pPg;
-      assert( isOpen(pPager->jfd) 
-           || pPager->journalMode==PAGER_JOURNALMODE_OFF 
-           || pPager->journalMode==PAGER_JOURNALMODE_WAL 
-      );
-      if( !zMaster && isOpen(pPager->jfd) 
-       && pPager->journalOff==jrnlBufferSize(pPager) 
-       && pPager->dbSize>=pPager->dbOrigSize
-       && (0==(pPg = sqlite3PcacheDirtyList(pPager->pPCache)) || 0==pPg->pDirty)
-      ){
-        /* Update the db file change counter via the direct-write method. The 
-        ** following call will modify the in-memory representation of page 1 
-        ** to include the updated change counter and then write page 1 
-        ** directly to the database file. Because of the atomic-write 
-        ** property of the host file-system, this is safe.
-        */
-        rc = pager_incr_changecounter(pPager, 1);
-      }else{
-        rc = sqlite3JournalCreate(pPager->jfd);
-        if( rc==SQLITE_OK ){
-          rc = pager_incr_changecounter(pPager, 0);
+      if( bBatch==0 ){
+        PgHdr *pPg;
+        assert( isOpen(pPager->jfd) 
+            || pPager->journalMode==PAGER_JOURNALMODE_OFF 
+            || pPager->journalMode==PAGER_JOURNALMODE_WAL 
+            );
+        if( !zMaster && isOpen(pPager->jfd) 
+         && pPager->journalOff==jrnlBufferSize(pPager) 
+         && pPager->dbSize>=pPager->dbOrigSize
+         && (!(pPg = sqlite3PcacheDirtyList(pPager->pPCache)) || 0==pPg->pDirty)
+        ){
+          /* Update the db file change counter via the direct-write method. The 
+          ** following call will modify the in-memory representation of page 1 
+          ** to include the updated change counter and then write page 1 
+          ** directly to the database file. Because of the atomic-write 
+          ** property of the host file-system, this is safe.
+          */
+          rc = pager_incr_changecounter(pPager, 1);
+        }else{
+          rc = sqlite3JournalCreate(pPager->jfd);
+          if( rc==SQLITE_OK ){
+            rc = pager_incr_changecounter(pPager, 0);
+          }
         }
       }
-  #else
+#else 
+#ifdef SQLITE_ENABLE_BATCH_ATOMIC_WRITE
+      if( zMaster ){
+        rc = sqlite3JournalCreate(pPager->jfd);
+        if( rc!=SQLITE_OK ) goto commit_phase_one_exit;
+      }
+#endif
       rc = pager_incr_changecounter(pPager, 0);
-  #endif
+#endif
       if( rc!=SQLITE_OK ) goto commit_phase_one_exit;
   
       /* Write the master journal name into the journal file. If a master 
@@ -6419,8 +6463,24 @@ int sqlite3PagerCommitPhaseOne(
       */
       rc = syncJournal(pPager, 0);
       if( rc!=SQLITE_OK ) goto commit_phase_one_exit;
-  
+
+      if( bBatch ){
+        /* The pager is now in DBMOD state. But regardless of what happens
+        ** next, attempting to play the journal back into the database would
+        ** be unsafe. Close it now to make sure that does not happen.  */
+        sqlite3OsClose(pPager->jfd);
+        rc = sqlite3OsFileControl(fd, SQLITE_FCNTL_BEGIN_ATOMIC_WRITE, 0);
+        if( rc!=SQLITE_OK ) goto commit_phase_one_exit;
+      }
       rc = pager_write_pagelist(pPager,sqlite3PcacheDirtyList(pPager->pPCache));
+      if( bBatch ){
+        if( rc==SQLITE_OK ){
+          rc = sqlite3OsFileControl(fd, SQLITE_FCNTL_COMMIT_ATOMIC_WRITE, 0);
+        }else{
+          sqlite3OsFileControl(fd, SQLITE_FCNTL_ROLLBACK_ATOMIC_WRITE, 0);
+        }
+      }
+
       if( rc!=SQLITE_OK ){
         assert( rc!=SQLITE_IOERR_BLOCKED );
         goto commit_phase_one_exit;