def _write_entire_chunk(self, chunk): bytes_per_sync = self._disk_file._mgr.bytes_per_sync while chunk: written = do_write(self._fd, chunk) chunk = chunk[written:] self._upload_size += written # For large files sync every 512MB (by default) written diff = self._upload_size - self._last_sync if diff >= bytes_per_sync: do_fdatasync(self._fd) do_fadvise64(self._fd, self._last_sync, diff) self._last_sync = self._upload_size
def _drop_cache(self, offset, length): """Method for no-oping buffer cache drop method.""" if not self._keep_cache and self._fd > -1: do_fadvise64(self._fd, offset, length)
def _finalize_put(self, metadata): # Write out metadata before fsync() to ensure it is also forced to # disk. write_metadata(self._fd, metadata) # We call fsync() before calling drop_cache() to lower the # amount of redundant work the drop cache code will perform on # the pages (now that after fsync the pages will be all # clean). do_fsync(self._fd) # From the Department of the Redundancy Department, make sure # we call drop_cache() after fsync() to avoid redundant work # (pages all clean). do_fadvise64(self._fd, self._last_sync, self._upload_size) # At this point we know that the object's full directory path # exists, so we can just rename it directly without using Swift's # swift.common.utils.renamer(), which makes the directory path and # adds extra stat() calls. df = self._disk_file attempts = 1 while True: try: do_rename(self._tmppath, df._data_file) except OSError as err: if err.errno in (errno.ENOENT, errno.EIO) \ and attempts < MAX_RENAME_ATTEMPTS: # FIXME: Why either of these two error conditions is # happening is unknown at this point. This might be a # FUSE issue of some sort or a possible race # condition. So let's sleep on it, and double check # the environment after a good nap. _random_sleep() # Tease out why this error occurred. The man page for # rename reads: # "The link named by tmppath does not exist; or, a # directory component in data_file does not exist; # or, tmppath or data_file is an empty string." assert len(self._tmppath) > 0 and len(df._data_file) > 0 tpstats = do_stat(self._tmppath) tfstats = do_fstat(self._fd) assert tfstats if not tpstats or tfstats.st_ino != tpstats.st_ino: # Temporary file name conflict raise DiskFileError( 'DiskFile.put(): temporary file, %s, was' ' already renamed (targeted for %s)' % ( self._tmppath, df._data_file)) else: # Data file target name now has a bad path! dfstats = do_stat(df._put_datadir) if not dfstats: raise DiskFileError( 'DiskFile.put(): path to object, %s, no' ' longer exists (targeted for %s)' % ( df._put_datadir, df._data_file)) else: is_dir = stat.S_ISDIR(dfstats.st_mode) if not is_dir: raise DiskFileError( 'DiskFile.put(): path to object, %s,' ' no longer a directory (targeted for' ' %s)' % (self._put_datadir, df._data_file)) else: # Let's retry since everything looks okay logging.warn( "DiskFile.put(): os.rename('%s','%s')" " initially failed (%s) but a" " stat('%s') following that succeeded:" " %r" % ( self._tmppath, df._data_file, str(err), df._put_datadir, dfstats)) attempts += 1 continue else: raise GlusterFileSystemOSError( err.errno, "%s, os.rename('%s', '%s')" % ( err.strerror, self._tmppath, df._data_file)) else: # Success! break # Close here so the calling context does not have to perform this # in a thread. self.close()
def _finalize_put(self, metadata): # Write out metadata before fsync() to ensure it is also forced to # disk. write_metadata(self._fd, metadata) # We call fsync() before calling drop_cache() to lower the # amount of redundant work the drop cache code will perform on # the pages (now that after fsync the pages will be all # clean). do_fsync(self._fd) # From the Department of the Redundancy Department, make sure # we call drop_cache() after fsync() to avoid redundant work # (pages all clean). do_fadvise64(self._fd, self._last_sync, self._upload_size) # At this point we know that the object's full directory path # exists, so we can just rename it directly without using Swift's # swift.common.utils.renamer(), which makes the directory path and # adds extra stat() calls. df = self._disk_file attempts = 1 while True: try: do_rename(self._tmppath, df._data_file) except OSError as err: if err.errno in (errno.ENOENT, errno.EIO) \ and attempts < MAX_RENAME_ATTEMPTS: # FIXME: Why either of these two error conditions is # happening is unknown at this point. This might be a # FUSE issue of some sort or a possible race # condition. So let's sleep on it, and double check # the environment after a good nap. _random_sleep() # Tease out why this error occurred. The man page for # rename reads: # "The link named by tmppath does not exist; or, a # directory component in data_file does not exist; # or, tmppath or data_file is an empty string." assert len(self._tmppath) > 0 and len(df._data_file) > 0 tpstats = do_stat(self._tmppath) tfstats = do_fstat(self._fd) assert tfstats if not tpstats or tfstats.st_ino != tpstats.st_ino: # Temporary file name conflict raise DiskFileError( 'DiskFile.put(): temporary file, %s, was' ' already renamed (targeted for %s)' % (self._tmppath, df._data_file)) else: # Data file target name now has a bad path! dfstats = do_stat(df._put_datadir) if not dfstats: raise DiskFileError( 'DiskFile.put(): path to object, %s, no' ' longer exists (targeted for %s)' % (df._put_datadir, df._data_file)) else: is_dir = stat.S_ISDIR(dfstats.st_mode) if not is_dir: raise DiskFileError( 'DiskFile.put(): path to object, %s,' ' no longer a directory (targeted for' ' %s)' % (self._put_datadir, df._data_file)) else: # Let's retry since everything looks okay logging.warn( "DiskFile.put(): rename('%s','%s')" " initially failed (%s) but a" " stat('%s') following that succeeded:" " %r" % (self._tmppath, df._data_file, str(err), df._put_datadir, dfstats)) attempts += 1 continue else: raise GlusterFileSystemOSError( err.errno, "%s, rename('%s', '%s')" % (err.strerror, self._tmppath, df._data_file)) else: # Success! break # Close here so the calling context does not have to perform this # in a thread. self.close()