Backport more SQLite changes to enable system-sqlite experiments.

third_party/sqlite/src/src/os_unix.c

Index: third_party/sqlite/src/src/os_unix.c
diff --git a/third_party/sqlite/src/src/os_unix.c b/third_party/sqlite/src/src/os_unix.c
index e5b2540ac5d4d2949aa43d3352280acb53730dfd..804c588e45e1b914a753a60ff2f76959d42213bc 100644
--- a/third_party/sqlite/src/src/os_unix.c
+++ b/third_party/sqlite/src/src/os_unix.c
@@ -204,7 +204,6 @@ struct unixFile {
   sqlite3_io_methods const *pMethod;  /* Always the first entry */
   unixInodeInfo *pInode;              /* Info about locks on this inode */
   int h;                              /* The file descriptor */
-  int dirfd;                          /* File descriptor for the directory */
   unsigned char eFileLock;            /* The type of lock held on this fd */
   unsigned char ctrlFlags;            /* Behavioral bits.  UNIXFILE_* flags */
   int lastErrno;                      /* The unix errno from last I/O error */
@@ -248,6 +247,7 @@ struct unixFile {
 */
 #define UNIXFILE_EXCL   0x01     /* Connections from one process only */
 #define UNIXFILE_RDONLY 0x02     /* Connection is read only */
+#define UNIXFILE_DIRSYNC 0x04    /* Directory sync needed */
 
 /*
 ** Include code that is common to all os_*.c files
@@ -281,6 +281,9 @@ struct unixFile {
 #define threadid 0
 #endif
 
+/* Forward reference */
+static int openDirectory(const char*, int*);
+
 /*
 ** Many system calls are accessed through pointer-to-functions so that
 ** they may be overridden at runtime to facilitate fault injection during
@@ -377,6 +380,12 @@ static struct unix_syscall {
 #endif
 #define osFallocate ((int(*)(int,off_t,off_t))aSyscall[15].pCurrent)
 
+  { "unlink",       (sqlite3_syscall_ptr)unlink,           0 },
+#define osUnlink    ((int(*)(const char*))aSyscall[16].pCurrent)
+
+  { "openDirectory",    (sqlite3_syscall_ptr)openDirectory,      0 },
+#define osOpenDirectory ((int(*)(const char*,int*))aSyscall[17].pCurrent)
+
 }; /* End of the overrideable system calls */
 
 /*
@@ -1731,10 +1740,6 @@ static int unixUnlock(sqlite3_file *id, int eFileLock){
 */
 static int closeUnixFile(sqlite3_file *id){
   unixFile *pFile = (unixFile*)id;
-  if( pFile->dirfd>=0 ){
-    robust_close(pFile, pFile->dirfd, __LINE__);
-    pFile->dirfd=-1;
-  }
   if( pFile->h>=0 ){
     robust_close(pFile, pFile->h, __LINE__);
     pFile->h = -1;
@@ -1742,7 +1747,7 @@ static int closeUnixFile(sqlite3_file *id){
 #if OS_VXWORKS
   if( pFile->pId ){
     if( pFile->isDelete ){
-      unlink(pFile->pId->zCanonicalName);
+      osUnlink(pFile->pId->zCanonicalName);
     }
     vxworksReleaseFileId(pFile->pId);
     pFile->pId = 0;
@@ -1989,7 +1994,7 @@ static int dotlockUnlock(sqlite3_file *id, int eFileLock) {
   
   /* To fully unlock the database, delete the lock file */
   assert( eFileLock==NO_LOCK );
-  if( unlink(zLockFile) ){
+  if( osUnlink(zLockFile) ){
     int rc = 0;
     int tErrno = errno;
     if( ENOENT != tErrno ){
@@ -3226,6 +3231,50 @@ static int full_fsync(int fd, int fullSync, int dataOnly){
 }
 
 /*
+** Open a file descriptor to the directory containing file zFilename.
+** If successful, *pFd is set to the opened file descriptor and
+** SQLITE_OK is returned. If an error occurs, either SQLITE_NOMEM
+** or SQLITE_CANTOPEN is returned and *pFd is set to an undefined
+** value.
+**
+** The directory file descriptor is used for only one thing - to
+** fsync() a directory to make sure file creation and deletion events
+** are flushed to disk.  Such fsyncs are not needed on newer
+** journaling filesystems, but are required on older filesystems.
+**
+** This routine can be overridden using the xSetSysCall interface.
+** The ability to override this routine was added in support of the
+** chromium sandbox.  Opening a directory is a security risk (we are
+** told) so making it overrideable allows the chromium sandbox to
+** replace this routine with a harmless no-op.  To make this routine
+** a no-op, replace it with a stub that returns SQLITE_OK but leaves
+** *pFd set to a negative number.
+**
+** If SQLITE_OK is returned, the caller is responsible for closing
+** the file descriptor *pFd using close().
+*/
+static int openDirectory(const char *zFilename, int *pFd){
+  int ii;
+  int fd = -1;
+  char zDirname[MAX_PATHNAME+1];
+
+  sqlite3_snprintf(MAX_PATHNAME, zDirname, "%s", zFilename);
+  for(ii=(int)strlen(zDirname); ii>1 && zDirname[ii]!='/'; ii--);
+  if( ii>0 ){
+    zDirname[ii] = '\0';
+    fd = robust_open(zDirname, O_RDONLY|O_BINARY, 0);
+    if( fd>=0 ){
+#ifdef FD_CLOEXEC
+      osFcntl(fd, F_SETFD, osFcntl(fd, F_GETFD, 0) | FD_CLOEXEC);
+#endif
+      OSTRACE(("OPENDIR %-3d %s\n", fd, zDirname));
+    }
+  }
+  *pFd = fd;
+  return (fd>=0?SQLITE_OK:unixLogError(SQLITE_CANTOPEN_BKPT, "open", zDirname));
+}
+
+/*
 ** Make sure all writes to a particular file are committed to disk.
 **
 ** If dataOnly==0 then both the file itself and its metadata (file
@@ -3265,28 +3314,23 @@ static int unixSync(sqlite3_file *id, int flags){
     pFile->lastErrno = errno;
     return unixLogError(SQLITE_IOERR_FSYNC, "full_fsync", pFile->zPath);
   }
-  if( pFile->dirfd>=0 ){
-    OSTRACE(("DIRSYNC %-3d (have_fullfsync=%d fullsync=%d)\n", pFile->dirfd,
+
+  /* Also fsync the directory containing the file if the DIRSYNC flag
+  ** is set.  This is a one-time occurrance.  Many systems (examples: AIX)
+  ** are unable to fsync a directory, so ignore errors on the fsync.
+  */
+  if( pFile->ctrlFlags & UNIXFILE_DIRSYNC ){
+    int dirfd;
+    OSTRACE(("DIRSYNC %s (have_fullfsync=%d fullsync=%d)\n", pFile->zPath,
             HAVE_FULLFSYNC, isFullsync));
-#ifndef SQLITE_DISABLE_DIRSYNC
-    /* The directory sync is only attempted if full_fsync is
-    ** turned off or unavailable.  If a full_fsync occurred above,
-    ** then the directory sync is superfluous.
-    */
-    if( (!HAVE_FULLFSYNC || !isFullsync) && full_fsync(pFile->dirfd,0,0) ){
-       /*
-       ** We have received multiple reports of fsync() returning
-       ** errors when applied to directories on certain file systems.
-       ** A failed directory sync is not a big deal.  So it seems
-       ** better to ignore the error.  Ticket #1657
-       */
-       /* pFile->lastErrno = errno; */
-       /* return SQLITE_IOERR; */
+    rc = osOpenDirectory(pFile->zPath, &dirfd);
+    if( rc==SQLITE_OK && dirfd>=0 ){
+      full_fsync(dirfd, 0, 0);
+      robust_close(pFile, dirfd, __LINE__);
+    }else if( rc==SQLITE_CANTOPEN ){
+      rc = SQLITE_OK;
     }
-#endif
-    /* Only need to sync once, so close the  directory when we are done */
-    robust_close(pFile, pFile->dirfd, __LINE__);
-    pFile->dirfd = -1;
+    pFile->ctrlFlags &= ~UNIXFILE_DIRSYNC;
   }
   return rc;
 }
@@ -4110,7 +4154,7 @@ static int unixShmUnmap(
   assert( pShmNode->nRef>0 );
   pShmNode->nRef--;
   if( pShmNode->nRef==0 ){
-    if( deleteFlag && pShmNode->h>=0 ) unlink(pShmNode->zFilename);
+    if( deleteFlag && pShmNode->h>=0 ) osUnlink(pShmNode->zFilename);
     unixShmPurge(pDbFd);
   }
   unixLeaveMutex();
@@ -4430,7 +4474,7 @@ void initUnixFile(sqlite3_file* file) {
 int fillInUnixFile(
   sqlite3_vfs *pVfs,      /* Pointer to vfs object */
   int h,                  /* Open file descriptor of file being opened */
-  int dirfd,              /* Directory file descriptor */
+  int syncDir,            /* True to sync directory on first sync */
   sqlite3_file *pId,      /* Write to the unixFile structure here */
   const char *zFilename,  /* Name of the file being opened */
   int noLock,             /* Omit locking if true */
@@ -4461,7 +4505,6 @@ int fillInUnixFile(
 
   OSTRACE(("OPEN    %-3d %s\n", h, zFilename));
   pNew->h = h;
-  pNew->dirfd = dirfd;
   pNew->zPath = zFilename;
   if( memcmp(pVfs->zName,"unix-excl",10)==0 ){
     pNew->ctrlFlags = UNIXFILE_EXCL;
@@ -4471,6 +4514,9 @@ int fillInUnixFile(
   if( isReadOnly ){
     pNew->ctrlFlags |= UNIXFILE_RDONLY;
   }
+  if( syncDir ){
+    pNew->ctrlFlags |= UNIXFILE_DIRSYNC;
+  }
 
 #if OS_VXWORKS
   pNew->pId = vxworksFindFileId(zFilename);
@@ -4597,13 +4643,12 @@ int fillInUnixFile(
   if( rc!=SQLITE_OK ){
     if( h>=0 ) robust_close(pNew, h, __LINE__);
     h = -1;
-    unlink(zFilename);
+    osUnlink(zFilename);
     isDelete = 0;
   }
   pNew->isDelete = isDelete;
 #endif
   if( rc!=SQLITE_OK ){
-    if( dirfd>=0 ) robust_close(pNew, dirfd, __LINE__);
     if( h>=0 ) robust_close(pNew, h, __LINE__);
   }else{
     pNew->pMethod = pLockingStyle;
@@ -4613,37 +4658,6 @@ int fillInUnixFile(
 }
 
 /*
-** Open a file descriptor to the directory containing file zFilename.
-** If successful, *pFd is set to the opened file descriptor and
-** SQLITE_OK is returned. If an error occurs, either SQLITE_NOMEM
-** or SQLITE_CANTOPEN is returned and *pFd is set to an undefined
-** value.
-**
-** If SQLITE_OK is returned, the caller is responsible for closing
-** the file descriptor *pFd using close().
-*/
-static int openDirectory(const char *zFilename, int *pFd){
-  int ii;
-  int fd = -1;
-  char zDirname[MAX_PATHNAME+1];
-
-  sqlite3_snprintf(MAX_PATHNAME, zDirname, "%s", zFilename);
-  for(ii=(int)strlen(zDirname); ii>1 && zDirname[ii]!='/'; ii--);
-  if( ii>0 ){
-    zDirname[ii] = '\0';
-    fd = robust_open(zDirname, O_RDONLY|O_BINARY, 0);
-    if( fd>=0 ){
-#ifdef FD_CLOEXEC
-      osFcntl(fd, F_SETFD, osFcntl(fd, F_GETFD, 0) | FD_CLOEXEC);
-#endif
-      OSTRACE(("OPENDIR %-3d %s\n", fd, zDirname));
-    }
-  }
-  *pFd = fd;
-  return (fd>=0?SQLITE_OK:unixLogError(SQLITE_CANTOPEN_BKPT, "open", zDirname));
-}
-
-/*
 ** Return the name of a directory in which to put temporary files.
 ** If no suitable temporary file directory can be found, return NULL.
 */
@@ -4938,7 +4952,6 @@ static int unixOpen(
 ){
   unixFile *p = (unixFile *)pFile;
   int fd = -1;                   /* File descriptor returned by open() */
-  int dirfd = -1;                /* Directory file descriptor */
   int openFlags = 0;             /* Flags to pass to open() */
   int eType = flags&0xFFFFFF00;  /* Type of file to open */
   int noLock;                    /* True to omit locking primitives */
@@ -4957,7 +4970,7 @@ static int unixOpen(
   ** a file-descriptor on the directory too. The first time unixSync()
   ** is called the directory file descriptor will be fsync()ed and close()d.
   */
-  int isOpenDirectory = (isCreate && (
+  int syncDir = (isCreate && (
         eType==SQLITE_OPEN_MASTER_JOURNAL 
      || eType==SQLITE_OPEN_MAIN_JOURNAL 
      || eType==SQLITE_OPEN_WAL
@@ -5004,7 +5017,7 @@ static int unixOpen(
     }
   }else if( !zName ){
     /* If zName is NULL, the upper layer is requesting a temp file. */
-    assert(isDelete && !isOpenDirectory);
+    assert(isDelete && !syncDir);
     rc = unixGetTempname(MAX_PATHNAME+1, zTmpname);
     if( rc!=SQLITE_OK ){
       return rc;
@@ -5057,7 +5070,7 @@ static int unixOpen(
 #if OS_VXWORKS
     zPath = zName;
 #else
-    unlink(zName);
+    osUnlink(zName);
 #endif
   }
 #if SQLITE_ENABLE_LOCKING_STYLE
@@ -5066,19 +5079,6 @@ static int unixOpen(
   }
 #endif
 
-  if( isOpenDirectory ){
-    rc = openDirectory(zPath, &dirfd);
-    if( rc!=SQLITE_OK ){
-      /* It is safe to close fd at this point, because it is guaranteed not
-      ** to be open on a database file. If it were open on a database file,
-      ** it would not be safe to close as this would release any locks held
-      ** on the file by this process.  */
-      assert( eType!=SQLITE_OPEN_MAIN_DB );
-      robust_close(p, fd, __LINE__);
-      goto open_finished;
-    }
-  }
-
 #ifdef FD_CLOEXEC
   osFcntl(fd, F_SETFD, osFcntl(fd, F_GETFD, 0) | FD_CLOEXEC);
 #endif
@@ -5090,7 +5090,6 @@ static int unixOpen(
   struct statfs fsInfo;
   if( fstatfs(fd, &fsInfo) == -1 ){
     ((unixFile*)pFile)->lastErrno = errno;
-    if( dirfd>=0 ) robust_close(p, dirfd, __LINE__);
     robust_close(p, fd, __LINE__);
     return SQLITE_IOERR_ACCESS;
   }
@@ -5122,9 +5121,6 @@ static int unixOpen(
         ** not while other file descriptors opened by the same process on
         ** the same file are working.  */
         p->lastErrno = errno;
-        if( dirfd>=0 ){
-          robust_close(p, dirfd, __LINE__);
-        }
         robust_close(p, fd, __LINE__);
         rc = SQLITE_IOERR_ACCESS;
         goto open_finished;
@@ -5132,7 +5128,7 @@ static int unixOpen(
       useProxy = !(fsInfo.f_flags&MNT_LOCAL);
     }
     if( useProxy ){
-      rc = fillInUnixFile(pVfs, fd, dirfd, pFile, zPath, noLock,
+      rc = fillInUnixFile(pVfs, fd, syncDir, pFile, zPath, noLock,
                           isDelete, isReadonly);
       if( rc==SQLITE_OK ){
         rc = proxyTransformUnixFile((unixFile*)pFile, ":auto:");
@@ -5150,7 +5146,7 @@ static int unixOpen(
   }
 #endif
   
-  rc = fillInUnixFile(pVfs, fd, dirfd, pFile, zPath, noLock,
+  rc = fillInUnixFile(pVfs, fd, syncDir, pFile, zPath, noLock,
                       isDelete, isReadonly);
 open_finished:
   if( rc!=SQLITE_OK ){
@@ -5172,13 +5168,13 @@ static int unixDelete(
   int rc = SQLITE_OK;
   UNUSED_PARAMETER(NotUsed);
   SimulateIOError(return SQLITE_IOERR_DELETE);
-  if( unlink(zPath)==(-1) && errno!=ENOENT ){
+  if( osUnlink(zPath)==(-1) && errno!=ENOENT ){
     return unixLogError(SQLITE_IOERR_DELETE, "unlink", zPath);
   }
 #ifndef SQLITE_DISABLE_DIRSYNC
   if( dirSync ){
     int fd;
-    rc = openDirectory(zPath, &fd);
+    rc = osOpenDirectory(zPath, &fd);
     if( rc==SQLITE_OK ){
 #if OS_VXWORKS
       if( fsync(fd)==-1 )
@@ -5189,6 +5185,8 @@ static int unixDelete(
         rc = unixLogError(SQLITE_IOERR_DIR_FSYNC, "fsync", zPath);
       }
       robust_close(0, fd, __LINE__);
+    }else if( rc==SQLITE_CANTOPEN ){
+      rc = SQLITE_OK;
     }
   }
 #endif
@@ -5750,7 +5748,6 @@ static int proxyCreateUnixFile(
     int islockfile           /* if non zero missing dirs will be created */
 ) {
   int fd = -1;
-  int dirfd = -1;
   unixFile *pNew;
   int rc = SQLITE_OK;
   int openFlags = O_RDWR | O_CREAT;
@@ -5815,7 +5812,7 @@ static int proxyCreateUnixFile(
   pUnused->flags = openFlags;
   pNew->pUnused = pUnused;
   
-  rc = fillInUnixFile(&dummyVfs, fd, dirfd, (sqlite3_file*)pNew, path, 0, 0, 0);
+  rc = fillInUnixFile(&dummyVfs, fd, 0, (sqlite3_file*)pNew, path, 0, 0, 0);
   if( rc==SQLITE_OK ){
     *ppFile = pNew;
     return SQLITE_OK;
@@ -5929,7 +5926,7 @@ static int proxyBreakConchLock(unixFile *pFile, uuid_t myHostID){
 end_breaklock:
   if( rc ){
     if( fd>=0 ){
-      unlink(tPath);
+      osUnlink(tPath);
       robust_close(pFile, fd, __LINE__);
     }
     fprintf(stderr, "failed to break stale lock on %s, %s\n", cPath, errmsg);
@@ -6752,7 +6749,7 @@ int sqlite3_os_init(void){
 
   /* Double-check that the aSyscall[] array has been constructed
   ** correctly.  See ticket [bb3a86e890c8e96ab] */
-  assert( ArraySize(aSyscall)==16 );
+  assert( ArraySize(aSyscall)==18 );
 
   /* Register all VFSes defined in the aVfs[] array */
   for(i=0; i<(sizeof(aVfs)/sizeof(sqlite3_vfs)); i++){