FileStorage.py

##############################################################################
#
# Copyright (c) 2001, 2002 Zope Foundation and Contributors.
# All Rights Reserved.
#
# This software is subject to the provisions of the Zope Public License,
# Version 2.1 (ZPL).  A copy of the ZPL should accompany this distribution.
# THIS SOFTWARE IS PROVIDED "AS IS" AND ANY AND ALL EXPRESS OR IMPLIED
# WARRANTIES ARE DISCLAIMED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
# WARRANTIES OF TITLE, MERCHANTABILITY, AGAINST INFRINGEMENT, AND FITNESS
# FOR A PARTICULAR PURPOSE
#
##############################################################################
"""Storage implementation using a log written to a single file.
"""
from __future__ import print_function

import binascii
import contextlib
import errno
import logging
import os
import threading
import time
from struct import pack
from struct import unpack

from persistent.TimeStamp import TimeStamp
from six import string_types as STRING_TYPES
from zc.lockfile import LockFile
from zope.interface import alsoProvides
from zope.interface import implementer

from ZODB.blob import BlobStorageMixin
from ZODB.blob import link_or_copy
from ZODB.blob import remove_committed
from ZODB.blob import remove_committed_dir
from ZODB.BaseStorage import BaseStorage
from ZODB.BaseStorage import DataRecord as _DataRecord
from ZODB.BaseStorage import TransactionRecord as _TransactionRecord
from ZODB.ConflictResolution import ConflictResolvingStorage
from ZODB.ConflictResolution import ResolvedSerial
from ZODB.FileStorage.format import CorruptedDataError
from ZODB.FileStorage.format import CorruptedError
from ZODB.FileStorage.format import DATA_HDR
from ZODB.FileStorage.format import DATA_HDR_LEN
from ZODB.FileStorage.format import DataHeader
from ZODB.FileStorage.format import FileStorageFormatter
from ZODB.FileStorage.format import TRANS_HDR
from ZODB.FileStorage.format import TRANS_HDR_LEN
from ZODB.FileStorage.format import TxnHeader
from ZODB.FileStorage.fspack import FileStoragePacker
from ZODB.interfaces import IBlobStorageRestoreable
from ZODB.interfaces import IExternalGC
from ZODB.interfaces import IStorage
from ZODB.interfaces import IStorageCurrentRecordIteration
from ZODB.interfaces import IStorageIteration
from ZODB.interfaces import IStorageRestoreable
from ZODB.interfaces import IStorageUndoable
from ZODB.POSException import ConflictError
from ZODB.POSException import MultipleUndoErrors
from ZODB.POSException import POSKeyError
from ZODB.POSException import ReadOnlyError
from ZODB.POSException import StorageError
from ZODB.POSException import StorageSystemError
from ZODB.POSException import StorageTransactionError
from ZODB.POSException import UndoError
from ZODB.fsIndex import fsIndex
from ZODB.utils import as_bytes
from ZODB.utils import as_text
from ZODB.utils import cp
from ZODB.utils import mktemp
from ZODB.utils import p64
from ZODB.utils import u64
from ZODB.utils import z64
from ZODB._compat import Pickler
from ZODB._compat import loads
from ZODB._compat import decodebytes
from ZODB._compat import encodebytes
from ZODB._compat import _protocol
from ZODB._compat import FILESTORAGE_MAGIC


# Not all platforms have fsync
fsync = getattr(os, "fsync", None)

packed_version = FILESTORAGE_MAGIC

logger = logging.getLogger('ZODB.FileStorage')

def panic(message, *data):
    logger.critical(message, *data)
    raise CorruptedTransactionError(message % data)

class FileStorageError(StorageError):
    pass

class PackError(FileStorageError):
    pass

class FileStorageFormatError(FileStorageError):
    """Invalid file format

    The format of the given file is not valid.
    """

class CorruptedFileStorageError(FileStorageError,
                                StorageSystemError):
    """Corrupted file storage."""

class CorruptedTransactionError(CorruptedFileStorageError):
    pass

class FileStorageQuotaError(FileStorageError,
                            StorageSystemError):
    """File storage quota exceeded."""

# Intended to be raised only in fspack.py, and ignored here.
class RedundantPackWarning(FileStorageError):
    pass

class TempFormatter(FileStorageFormatter):
    """Helper class used to read formatted FileStorage data."""

    def __init__(self, afile):
        self._file = afile

@implementer(
        IStorage,
        IStorageRestoreable,
        IStorageIteration,
        IStorageUndoable,
        IStorageCurrentRecordIteration,
        IExternalGC,
        )
class FileStorage(
    FileStorageFormatter,
    BlobStorageMixin,
    ConflictResolvingStorage,
    BaseStorage,
    ):


    # Set True while a pack is in progress; undo is blocked for the duration.
    _pack_is_in_progress = False

    def __init__(self, file_name, create=False, read_only=False, stop=None,
                 quota=None, pack_gc=True, pack_keep_old=True, packer=None,
                 blob_dir=None):

        if read_only:
            self._is_read_only = True
            if create:
                raise ValueError("can't create a read-only file")
        elif stop is not None:
            raise ValueError("time-travel only supported in read-only mode")

        if stop is None:
            stop = b'\377'*8

        # Lock the database and set up the temp file.
        if not read_only:
            # Create the lock file
            self._lock_file = LockFile(file_name + '.lock')
            self._tfile = open(file_name + '.tmp', 'w+b')
            self._tfmt = TempFormatter(self._tfile)
        else:
            self._tfile = None

        self._file_name = os.path.abspath(file_name)

        self._pack_gc = pack_gc
        self.pack_keep_old = pack_keep_old
        if packer is not None:
            self.packer = packer

        BaseStorage.__init__(self, file_name)

        index, tindex = self._newIndexes()
        self._initIndex(index, tindex)

        # Now open the file

        self._file = None
        if not create:
            try:
                self._file = open(file_name, read_only and 'rb' or 'r+b')
            except IOError as exc:
                if exc.errno == errno.EFBIG:
                    # The file is too big to open.  Fail visibly.
                    raise
                if exc.errno == errno.ENOENT:
                    # The file doesn't exist.  Create it.
                    create = 1
                # If something else went wrong, it's hard to guess
                # what the problem was.  If the file does not exist,
                # create it.  Otherwise, fail.
                if os.path.exists(file_name):
                    raise
                else:
                    create = 1

        if self._file is None and create:
            if os.path.exists(file_name):
                os.remove(file_name)
            self._file = open(file_name, 'w+b')
            self._file.write(packed_version)

        self._files = FilePool(self._file_name)
        r = self._restore_index()
        if r is not None:
            self._used_index = 1 # Marker for testing
            index, start, ltid = r

            self._initIndex(index, tindex)
            self._pos, self._oid, tid = read_index(
                self._file, file_name, index, tindex, stop,
                ltid=ltid, start=start, read_only=read_only,
                )
        else:
            self._used_index = 0 # Marker for testing
            self._pos, self._oid, tid = read_index(
                self._file, file_name, index, tindex, stop,
                read_only=read_only,
                )
            self._save_index()

        self._ltid = tid

        # self._pos should always point just past the last
        # transaction.  During 2PC, data is written after _pos.
        # invariant is restored at tpc_abort() or tpc_finish().

        self._ts = tid = TimeStamp(tid)
        t = time.time()
        t = TimeStamp(*time.gmtime(t)[:5] + (t % 60,))
        if tid > t:
            seconds = tid.timeTime() - t.timeTime()
            complainer = logger.warning
            if seconds > 30 * 60:   # 30 minutes -- way screwed up
                complainer = logger.critical
            complainer("%s Database records %d seconds in the future",
                       file_name, seconds)

        self._quota = quota

        if blob_dir:
            self.blob_dir = os.path.abspath(blob_dir)
            if create and os.path.exists(self.blob_dir):
                remove_committed_dir(self.blob_dir)

            self._blob_init(blob_dir)
            alsoProvides(self, IBlobStorageRestoreable)
        else:
            self.blob_dir = None
            self._blob_init_no_blobs()

    def copyTransactionsFrom(self, other):
        if self.blob_dir:
            return BlobStorageMixin.copyTransactionsFrom(self, other)
        else:
            return BaseStorage.copyTransactionsFrom(self, other)

    def _initIndex(self, index, tindex):
        self._index=index
        self._tindex=tindex
        self._index_get=index.get

    def __len__(self):
        return len(self._index)

    def _newIndexes(self):
        # hook to use something other than builtin dict
        return fsIndex(), {}

    _saved = 0
    def _save_index(self):
        """Write the database index to a file to support quick startup."""

        if self._is_read_only:
            return

        index_name = self.__name__ + '.index'
        tmp_name = index_name + '.index_tmp'

        self._index.save(self._pos, tmp_name)

        try:
            try:
                os.remove(index_name)
            except OSError:
                pass
            os.rename(tmp_name, index_name)
        except: pass

        self._saved += 1

    def _clear_index(self):
        index_name = self.__name__ + '.index'
        if os.path.exists(index_name):
            try:
                os.remove(index_name)
            except OSError:
                pass

    def _sane(self, index, pos):
        """Sanity check saved index data by reading the last undone trans

        Basically, we read the last not undone transaction and
        check to see that the included records are consistent
        with the index.  Any invalid record records or inconsistent
        object positions cause zero to be returned.
        """
        r = self._check_sanity(index, pos)
        if not r:
            logger.warning("Ignoring index for %s", self._file_name)
        return r

    def _check_sanity(self, index, pos):

        if pos < 100:
            return 0 # insane
        self._file.seek(0, 2)
        if self._file.tell() < pos:
            return 0 # insane
        ltid = None

        max_checked = 5
        checked = 0

        while checked < max_checked:
            self._file.seek(pos - 8)
            rstl = self._file.read(8)
            tl = u64(rstl)
            pos = pos - tl - 8
            if pos < 4:
                return 0 # insane
            h = self._read_txn_header(pos)
            if not ltid:
                ltid = h.tid
            if h.tlen != tl:
                return 0 # inconsistent lengths
            if h.status == 'u':
                continue # undone trans, search back
            if h.status not in ' p':
                return 0 # insane
            if tl < h.headerlen():
                return 0 # insane
            tend = pos + tl
            opos = pos + h.headerlen()
            if opos == tend:
                continue # empty trans

            while opos < tend and checked < max_checked:
                # Read the data records for this transaction
                h = self._read_data_header(opos)

                if opos + h.recordlen() > tend or h.tloc != pos:
                    return 0

                if index.get(h.oid, 0) != opos:
                    return 0 # insane

                checked += 1

                opos = opos + h.recordlen()

            return ltid

    def _restore_index(self):
        """Load database index to support quick startup."""
        # Returns (index, pos, tid), or None in case of error.
        # The index returned is always an instance of fsIndex.  If the
        # index cached in the file is a Python dict, it's converted to
        # fsIndex here, and, if we're not in read-only mode, the .index
        # file is rewritten with the converted fsIndex so we don't need to
        # convert it again the next time.
        file_name=self.__name__
        index_name=file_name+'.index'

        if os.path.exists(index_name):
            try:
                info = fsIndex.load(index_name)
            except:
                logger.exception('loading index')
                return None
        else:
            return None

        index = info.get('index')
        pos = info.get('pos')
        if index is None or pos is None:
            return None
        pos = int(pos)

        if (isinstance(index, dict) or
                (isinstance(index, fsIndex) and
                 isinstance(index._data, dict))):
            # Convert dictionary indexes to fsIndexes *or* convert fsIndexes
            # which have a dict `_data` attribute to a new fsIndex (newer
            # fsIndexes have an OOBTree as `_data`).
            newindex = fsIndex()
            newindex.update(index)
            index = newindex
            if not self._is_read_only:
                # Save the converted index.
                f = open(index_name, 'wb')
                p = Pickler(f, _protocol)
                info['index'] = index
                p.dump(info)
                f.close()
                # Now call this method again to get the new data.
                return self._restore_index()

        tid = self._sane(index, pos)
        if not tid:
            return None

        return index, pos, tid

    def close(self):
        self._file.close()
        self._files.close()
        if hasattr(self,'_lock_file'):
            self._lock_file.close()
        if self._tfile:
            self._tfile.close()
        try:
            self._save_index()
        except:
            # Log the error and continue
            logger.error("Error saving index on close()", exc_info=True)

    def getSize(self):
        return self._pos

    def _lookup_pos(self, oid):
        try:
            return self._index[oid]
        except KeyError:
            raise POSKeyError(oid)
        except TypeError:
            raise TypeError("invalid oid %r" % (oid,))

    def load(self, oid, version=''):
        """Return pickle data and serial number."""
        assert not version

        with self._files.get() as _file:
            pos = self._lookup_pos(oid)
            h = self._read_data_header(pos, oid, _file)
            if h.plen:
                data = _file.read(h.plen)
                return data, h.tid
            elif h.back:
                # Get the data from the backpointer, but tid from
                # current txn.
                data = self._loadBack_impl(oid, h.back, _file=_file)[0]
                return data, h.tid
            else:
                raise POSKeyError(oid)

    def loadSerial(self, oid, serial):
        with self._lock:
            pos = self._lookup_pos(oid)
            while 1:
                h = self._read_data_header(pos, oid)
                if h.tid == serial:
                    break
                pos = h.prev
                if not pos:
                    raise POSKeyError(oid)
            if h.plen:
                return self._file.read(h.plen)
            else:
                return self._loadBack_impl(oid, h.back)[0]

    def loadBefore(self, oid, tid):
        with self._files.get() as _file:
            pos = self._lookup_pos(oid)
            end_tid = None
            while True:
                h = self._read_data_header(pos, oid, _file)
                if h.tid < tid:
                    break

                pos = h.prev
                end_tid = h.tid
                if not pos:
                    return None

            if h.back:
                data, _, _, _ = self._loadBack_impl(oid, h.back, _file=_file)
                return data, h.tid, end_tid
            else:
                return _file.read(h.plen), h.tid, end_tid

    def store(self, oid, oldserial, data, version, transaction):
        if self._is_read_only:
            raise ReadOnlyError()
        if transaction is not self._transaction:
            raise StorageTransactionError(self, transaction)
        assert not version

        with self._lock:
            if oid > self._oid:
                self.set_max_oid(oid)
            old = self._index_get(oid, 0)
            committed_tid = None
            pnv = None
            if old:
                h = self._read_data_header(old, oid)
                committed_tid = h.tid

                if oldserial != committed_tid:
                    data = self.tryToResolveConflict(oid, committed_tid,
                                                     oldserial, data)

            pos = self._pos
            here = pos + self._tfile.tell() + self._thl
            self._tindex[oid] = here
            new = DataHeader(oid, self._tid, old, pos, 0, len(data))

            self._tfile.write(new.asString())
            self._tfile.write(data)

            # Check quota
            if self._quota is not None and here > self._quota:
                raise FileStorageQuotaError(
                    "The storage quota has been exceeded.")

            if old and oldserial != committed_tid:
                return ResolvedSerial
            else:
                return self._tid

    def deleteObject(self, oid, oldserial, transaction):
        if self._is_read_only:
            raise ReadOnlyError()
        if transaction is not self._transaction:
            raise StorageTransactionError(self, transaction)

        with self._lock:
            old = self._index_get(oid, 0)
            if not old:
                raise POSKeyError(oid)
            h = self._read_data_header(old, oid)
            committed_tid = h.tid

            if oldserial != committed_tid:
                raise ConflictError(
                    oid=oid, serials=(committed_tid, oldserial))

            pos = self._pos
            here = pos + self._tfile.tell() + self._thl
            self._tindex[oid] = here
            new = DataHeader(oid, self._tid, old, pos, 0, 0)
            self._tfile.write(new.asString())
            self._tfile.write(z64)

            # Check quota
            if self._quota is not None and here > self._quota:
                raise FileStorageQuotaError(
                    "The storage quota has been exceeded.")

    def _data_find(self, tpos, oid, data):
        # Return backpointer for oid.  Must call with the lock held.
        # This is a file offset to oid's data record if found, else 0.
        # The data records in the transaction at tpos are searched for oid.
        # If a data record for oid isn't found, returns 0.
        # Else if oid's data record contains a backpointer, that
        # backpointer is returned.
        # Else oid's data record contains the data, and the file offset of
        # oid's data record is returned.  This data record should contain
        # a pickle identical to the 'data' argument.

        # Unclear:  If the length of the stored data doesn't match len(data),
        # an exception is raised.  If the lengths match but the data isn't
        # the same, 0 is returned.  Why the discrepancy?
        self._file.seek(tpos)
        h = self._file.read(TRANS_HDR_LEN)
        tid, tl, status, ul, dl, el = unpack(TRANS_HDR, h)
        status = as_text(status)
        self._file.read(ul + dl + el)
        tend = tpos + tl + 8
        pos = self._file.tell()
        while pos < tend:
            h = self._read_data_header(pos)
            if h.oid == oid:
                # Make sure this looks like the right data record
                if h.plen == 0:
                    # This is also a backpointer.  Gotta trust it.
                    return pos
                if h.plen != len(data):
                    # The expected data doesn't match what's in the
                    # backpointer.  Something is wrong.
                    logger.error("Mismatch between data and"
                                 " backpointer at %d", pos)
                    return 0
                _data = self._file.read(h.plen)
                if data != _data:
                    return 0
                return pos
            pos += h.recordlen()
            self._file.seek(pos)
        return 0

    def restore(self, oid, serial, data, version, prev_txn, transaction):
        # A lot like store() but without all the consistency checks.  This
        # should only be used when we /know/ the data is good, hence the
        # method name.  While the signature looks like store() there are some
        # differences:
        #
        # - serial is the serial number of /this/ revision, not of the
        #   previous revision.  It is used instead of self._tid, which is
        #   ignored.
        #
        # - Nothing is returned
        #
        # - data can be None, which indicates a George Bailey object
        #   (i.e. one who's creation has been transactionally undone).
        #
        # prev_txn is a backpointer.  In the original database, it's possible
        # that the data was actually living in a previous transaction.  This
        # can happen for transactional undo and other operations, and is used
        # as a space saving optimization.  Under some circumstances the
        # prev_txn may not actually exist in the target database (i.e. self)
        # for example, if it's been packed away.  In that case, the prev_txn
        # should be considered just a hint, and is ignored if the transaction
        # doesn't exist.
        if self._is_read_only:
            raise ReadOnlyError()
        if transaction is not self._transaction:
            raise StorageTransactionError(self, transaction)
        if version:
            raise TypeError("Versions are no-longer supported")

        with self._lock:
            if oid > self._oid:
                self.set_max_oid(oid)
            prev_pos = 0
            if prev_txn is not None:
                prev_txn_pos = self._txn_find(prev_txn, 0)
                if prev_txn_pos:
                    prev_pos = self._data_find(prev_txn_pos, oid, data)
            old = self._index_get(oid, 0)
            # Calculate the file position in the temporary file
            here = self._pos + self._tfile.tell() + self._thl
            # And update the temp file index
            self._tindex[oid] = here
            if prev_pos:
                # If there is a valid prev_pos, don't write data.
                data = None
            if data is None:
                dlen = 0
            else:
                dlen = len(data)

            # Write the recovery data record
            new = DataHeader(oid, serial, old, self._pos, 0, dlen)

            self._tfile.write(new.asString())

            # Finally, write the data or a backpointer.
            if data is None:
                if prev_pos:
                    self._tfile.write(p64(prev_pos))
                else:
                    # Write a zero backpointer, which indicates an
                    # un-creation transaction.
                    self._tfile.write(z64)
            else:
                self._tfile.write(data)

    def supportsUndo(self):
        return 1

    def _clear_temp(self):
        self._tindex.clear()
        if self._tfile is not None:
            self._tfile.seek(0)

    def _begin(self, tid, u, d, e):
        self._nextpos = 0
        self._thl = TRANS_HDR_LEN + len(u) + len(d) + len(e)
        if self._thl > 65535:
            # one of u, d, or e may be > 65535
            # We have to check lengths here because struct.pack
            # doesn't raise an exception on overflow!
            if len(u) > 65535:
                raise FileStorageError('user name too long')
            if len(d) > 65535:
                raise FileStorageError('description too long')
            if len(e) > 65535:
                raise FileStorageError('too much extension data')

    def tpc_vote(self, transaction):
        with self._lock:
            if transaction is not self._transaction:
                raise StorageTransactionError(
                    "tpc_vote called with wrong transaction")
            dlen = self._tfile.tell()
            if not dlen:
                return # No data in this trans
            self._tfile.seek(0)
            user, descr, ext = self._ude

            self._file.seek(self._pos)
            tl = self._thl + dlen

            try:
                h = TxnHeader(self._tid, tl, "c", len(user),
                              len(descr), len(ext))
                h.user = user
                h.descr = descr
                h.ext = ext
                self._file.write(h.asString())
                cp(self._tfile, self._file, dlen)
                self._file.write(p64(tl))
                self._file.flush()
            except:
                # Hm, an error occurred writing out the data. Maybe the
                # disk is full. We don't want any turd at the end.
                self._file.truncate(self._pos)
                raise
            self._nextpos = self._pos + (tl + 8)

    def tpc_finish(self, transaction, f=None):
        with self._files.write_lock():
            with self._lock:
                if transaction is not self._transaction:
                    raise StorageTransactionError(
                        "tpc_finish called with wrong transaction")
                try:
                    if f is not None:
                        f(self._tid)
                    u, d, e = self._ude
                    self._finish(self._tid, u, d, e)
                    self._clear_temp()
                finally:
                    self._ude = None
                    self._transaction = None
                    self._commit_lock_release()

    def _finish(self, tid, u, d, e):
        # If self._nextpos is 0, then the transaction didn't write any
        # data, so we don't bother writing anything to the file.
        if self._nextpos:
            # Clear the checkpoint flag
            self._file.seek(self._pos+16)
            self._file.write(as_bytes(self._tstatus))
            try:
                # At this point, we may have committed the data to disk.
                # If we fail from here, we're in bad shape.
                self._finish_finish(tid)
            except:
                # Ouch.  This is bad.  Let's try to get back to where we were
                # and then roll over and die
                logger.critical("Failure in _finish. Closing.", exc_info=True)
                self.close()
                raise

    def _finish_finish(self, tid):
        # This is a separate method to allow tests to replace it with
        # something broken. :)

        self._file.flush()
        if fsync is not None:
            fsync(self._file.fileno())

        self._pos = self._nextpos
        self._index.update(self._tindex)
        self._ltid = tid
        self._blob_tpc_finish()

    def _abort(self):
        if self._nextpos:
            self._file.truncate(self._pos)
            self._nextpos=0
            self._blob_tpc_abort()

    def _undoDataInfo(self, oid, pos, tpos):
        """Return the tid, data pointer, and data for the oid record at pos
        """
        if tpos:
            pos = tpos - self._pos - self._thl
            tpos = self._tfile.tell()
            h = self._tfmt._read_data_header(pos, oid)
            afile = self._tfile
        else:
            h = self._read_data_header(pos, oid)
            afile = self._file
        if h.oid != oid:
            raise UndoError("Invalid undo transaction id", oid)

        if h.plen:
            data = afile.read(h.plen)
        else:
            data = ''
            pos = h.back

        if tpos:
            self._tfile.seek(tpos) # Restore temp file to end

        return h.tid, pos, data

    def getTid(self, oid):
        with self._lock:
            pos = self._lookup_pos(oid)
            h = self._read_data_header(pos, oid)
            if h.plen == 0 and h.back == 0:
                # Undone creation
                raise POSKeyError(oid)
            return h.tid

    def _transactionalUndoRecord(self, oid, pos, tid, pre):
        """Get the undo information for a data record

        'pos' points to the data header for 'oid' in the transaction
        being undone.  'tid' refers to the transaction being undone.
        'pre' is the 'prev' field of the same data header.

        Return a 3-tuple consisting of a pickle, data pointer, and
        current position.  If the pickle is true, then the data
        pointer must be 0, but the pickle can be empty *and* the
        pointer 0.
        """

        copy = 1 # Can we just copy a data pointer

        # First check if it is possible to undo this record.
        tpos = self._tindex.get(oid, 0)
        ipos = self._index.get(oid, 0)
        tipos = tpos or ipos

        if tipos != pos:
            # Eek, a later transaction modified the data, but,
            # maybe it is pointing at the same data we are.
            ctid, cdataptr, cdata = self._undoDataInfo(oid, ipos, tpos)

            if cdataptr != pos:
                # We aren't sure if we are talking about the same data
                try:
                    if (
                        # The current record wrote a new pickle
                        cdataptr == tipos
                        or
                        # Backpointers are different
                        self._loadBackPOS(oid, pos) !=
                        self._loadBackPOS(oid, cdataptr)
                        ):
                        if pre and not tpos:
                            copy = 0 # we'll try to do conflict resolution
                        else:
                            # We bail if:
                            # - We don't have a previous record, which should
                            #   be impossible.
                            raise UndoError("no previous record", oid)
                except KeyError:
                    # LoadBack gave us a key error. Bail.
                    raise UndoError("_loadBack() failed", oid)

        # Return the data that should be written in the undo record.
        if not pre:
            # There is no previous revision, because the object creation
            # is being undone.
            return "", 0, ipos

        if copy:
            # we can just copy our previous-record pointer forward
            return "", pre, ipos

        try:
            bdata = self._loadBack_impl(oid, pre)[0]
        except KeyError:
            # couldn't find oid; what's the real explanation for this?
            raise UndoError("_loadBack() failed for %s", oid)
        try:
            data = self.tryToResolveConflict(oid, ctid, tid, bdata, cdata)
            return data, 0, ipos
        except ConflictError:
            pass

        raise UndoError("Some data were modified by a later transaction", oid)

    # undoLog() returns a description dict that includes an id entry.
    # The id is opaque to the client, but contains the transaction id.
    # The transactionalUndo() implementation does a simple linear
    # search through the file (from the end) to find the transaction.

    def undoLog(self, first=0, last=-20, filter=None):
        if last < 0:
            # -last is supposed to be the max # of transactions.  Convert to
            # a positive index.  Should have x - first = -last, which
            # means x = first - last.  This is spelled out here because
            # the normalization code was incorrect for years (used +1
            # instead -- off by 1), until ZODB 3.4.
            last = first - last
        with self._lock:
            if self._pack_is_in_progress:
                raise UndoError(
                    'Undo is currently disabled for database maintenance.<p>')
            us = UndoSearch(self._file, self._pos, first, last, filter)
            while not us.finished():
                # Hold lock for batches of 20 searches, so default search
                # parameters will finish without letting another thread run.
                for i in range(20):
                    if us.finished():
                        break
                    us.search()
                # Give another thread a chance, so that a long undoLog()
                # operation doesn't block all other activity.
                self._lock_release()
                self._lock_acquire()
            return us.results

    def undo(self, transaction_id, transaction):
        """Undo a transaction, given by transaction_id.

        Do so by writing new data that reverses the action taken by
        the transaction.

        Usually, we can get by with just copying a data pointer, by
        writing a file position rather than a pickle. Sometimes, we
        may do conflict resolution, in which case we actually copy
        new data that results from resolution.
        """

        if self._is_read_only:
            raise ReadOnlyError()
        if transaction is not self._transaction:
            raise StorageTransactionError(self, transaction)

        with self._lock:
          # Find the right transaction to undo and call _txn_undo_write().
          tid = decodebytes(transaction_id + b'\n')
          assert len(tid) == 8
          tpos = self._txn_find(tid, 1)
          tindex = self._txn_undo_write(tpos)
          self._tindex.update(tindex)
          return self._tid, tindex.keys()

    def _txn_find(self, tid, stop_at_pack):
        pos = self._pos
        while pos > 39:
            self._file.seek(pos - 8)
            pos = pos - u64(self._file.read(8)) - 8
            self._file.seek(pos)
            h = self._file.read(TRANS_HDR_LEN)
            _tid = h[:8]
            if _tid == tid:
                return pos
            if stop_at_pack:
                # check the status field of the transaction header
                if h[16] == b'p':
                    break
        raise UndoError("Invalid transaction id")

    def _txn_undo_write(self, tpos):
        # a helper function to write the data records for transactional undo

        otloc = self._pos
        here = self._pos + self._tfile.tell() + self._thl
        base = here - self._tfile.tell()
        # Let's move the file pointer back to the start of the txn record.
        th = self._read_txn_header(tpos)
        if th.status != " ":
            raise UndoError('non-undoable transaction')
        tend = tpos + th.tlen
        pos = tpos + th.headerlen()
        tindex = {}

        # keep track of failures, cause we may succeed later
        failures = {}
        # Read the data records for this transaction
        while pos < tend:
            h = self._read_data_header(pos)
            if h.oid in failures:
                del failures[h.oid] # second chance!

            assert base + self._tfile.tell() == here, (here, base,
                                                       self._tfile.tell())
            try:
                p, prev, ipos = self._transactionalUndoRecord(
                    h.oid, pos, h.tid, h.prev)
            except UndoError as v:
                # Don't fail right away. We may be redeemed later!
                failures[h.oid] = v
            else:

                if self.blob_dir and not p and prev:
                    try:
                        up, userial = self._loadBackTxn(h.oid, prev)
                    except POSKeyError:
                        pass # It was removed, so no need to copy data
                    else:
                        if self.is_blob_record(up):
                            # We're undoing a blob modification operation.
                            # We have to copy the blob data
                            tmp = mktemp(dir=self.fshelper.temp_dir)
                            with self.openCommittedBlobFile(h.oid, userial) as sfp:
                                with open(tmp, 'wb') as dfp:
                                    cp(sfp, dfp)
                            self._blob_storeblob(h.oid, self._tid, tmp)

                new = DataHeader(h.oid, self._tid, ipos, otloc, 0, len(p))

                # TODO:  This seek shouldn't be necessary, but some other
                # bit of code is messing with the file pointer.
                assert self._tfile.tell() == here - base, (here, base,
                                                           self._tfile.tell())
                self._tfile.write(new.asString())
                if p:
                    self._tfile.write(p)
                else:
                    self._tfile.write(p64(prev))
                tindex[h.oid] = here
                here += new.recordlen()

            pos += h.recordlen()
            if pos > tend:
                raise UndoError("non-undoable transaction")

        if failures:
            raise MultipleUndoErrors(list(failures.items()))

        return tindex

    def history(self, oid, size=1, filter=None):
        with self._lock:
            r = []
            pos = self._lookup_pos(oid)

            while 1:
                if len(r) >= size: return r
                h = self._read_data_header(pos)

                th = self._read_txn_header(h.tloc)
                if th.ext:
                    d = loads(th.ext)
                else:
                    d = {}

                d.update({"time": TimeStamp(h.tid).timeTime(),
                          "user_name": th.user,
                          "description": th.descr,
                          "tid": h.tid,
                          "size": h.plen,
                          })

                if filter is None or filter(d):
                    r.append(d)

                if h.prev:
                    pos = h.prev
                else:
                    return r

    def _redundant_pack(self, file, pos):
        assert pos > 8, pos
        file.seek(pos - 8)
        p = u64(file.read(8))
        file.seek(pos - p + 8)
        return file.read(1) not in ' u'

    @staticmethod
    def packer(storage, referencesf, stop, gc):
        # Our default packer is built around the original packer.  We
        # simply adapt the old interface to the new.  We don't really
        # want to invest much in the old packer, at least for now.
        assert referencesf is not None
        p = FileStoragePacker(storage, referencesf, stop, gc)
        try:
            opos = p.pack()
            if opos is None:
                return None
            return opos, p.index
        finally:
            p.close()

    def pack(self, t, referencesf, gc=None):
        """Copy data from the current database file to a packed file

        Non-current records from transactions with time-stamp strings less
        than packtss are ommitted. As are all undone records.

        Also, data back pointers that point before packtss are resolved and
        the associated data are copied, since the old records are not copied.
        """
        if self._is_read_only:
            raise ReadOnlyError()

        stop = TimeStamp(*time.gmtime(t)[:5]+(t%60,)).raw()
        if stop == z64:
            raise FileStorageError('Invalid pack time')

        # If the storage is empty, there's nothing to do.
        if not self._index:
            return

        with self._lock:
            if self._pack_is_in_progress:
                raise FileStorageError('Already packing')
            self._pack_is_in_progress = True

        if gc is None:
            gc = self._pack_gc

        oldpath = self._file_name + ".old"
        if os.path.exists(oldpath):
            os.remove(oldpath)
        if self.blob_dir and os.path.exists(self.blob_dir + ".old"):
            remove_committed_dir(self.blob_dir + ".old")

        cleanup = []

        have_commit_lock = False
        try:
            pack_result = None
            try:
                pack_result = self.packer(self, referencesf, stop, gc)
            except RedundantPackWarning as detail:
                logger.info(str(detail))
            if pack_result is None:
                return
            have_commit_lock = True
            opos, index = pack_result
            with self._files.write_lock():
                with self._lock:
                    self._files.empty()
                    self._file.close()
                    try:
                        os.rename(self._file_name, oldpath)
                    except Exception:
                        self._file = open(self._file_name, 'r+b')
                        raise

                    # OK, we're beyond the point of no return
                    os.rename(self._file_name + '.pack', self._file_name)
                    self._file = open(self._file_name, 'r+b')
                    self._initIndex(index, self._tindex)
                    self._pos = opos

            # We're basically done.  Now we need to deal with removed
            # blobs and removing the .old file (see further down).

            if self.blob_dir:
                self._commit_lock_release()
                have_commit_lock = False
                self._remove_blob_files_tagged_for_removal_during_pack()

        finally:
            if have_commit_lock:
                self._commit_lock_release()
            with self._lock:
                self._pack_is_in_progress = False

        if not self.pack_keep_old:
            os.remove(oldpath)

        with self._lock:
            self._save_index()

    def _remove_blob_files_tagged_for_removal_during_pack(self):
        lblob_dir = len(self.blob_dir)
        fshelper = self.fshelper
        old = self.blob_dir+'.old'

        # Helper to clean up dirs left empty after moving things to old
        def maybe_remove_empty_dir_containing(path, level=0):
            path = os.path.dirname(path)
            if len(path) <= lblob_dir or os.listdir(path):
                return

            # Path points to an empty dir.  There may be a race.  We
            # might have just removed the dir for an oid (or a parent
            # dir) and while we're cleaning up it's parent, another
            # thread is adding a new entry to it.

            # We don't have to worry about level 0, as this is just a
            # directory containing an object's revisions. If it is
            # enmpty, the object must have been garbage.

            # If the level is 1 or higher, we need to be more
            # careful.  We'll get the storage lock and double check
            # that the dir is still empty before removing it.

            removed = False
            if level:
                self._lock_acquire()
            try:
                if not os.listdir(path):
                    os.rmdir(path)
                    removed = True
            finally:
                if level:
                    self._lock_release()

            if removed:
                maybe_remove_empty_dir_containing(path, level+1)


        if self.pack_keep_old:
            # Helpers that move oid dir or revision file to the old dir.
            os.mkdir(old, 0o777)
            link_or_copy(os.path.join(self.blob_dir, '.layout'),
                         os.path.join(old, '.layout'))
            def handle_file(path):
                newpath = old+path[lblob_dir:]
                dest = os.path.dirname(newpath)
                if not os.path.exists(dest):
                    os.makedirs(dest, 0o700)
                os.rename(path, newpath)
            handle_dir = handle_file
        else:
            # Helpers that remove an oid dir or revision file.
            handle_file = remove_committed
            handle_dir = remove_committed_dir

        # Fist step: move or remove oids or revisions
        with open(os.path.join(self.blob_dir, '.removed'), 'rb') as fp:
            for line in fp:
                line = binascii.unhexlify(line.strip())

                if len(line) == 8:
                    # oid is garbage, re/move dir
                    path = fshelper.getPathForOID(line)
                    if not os.path.exists(path):
                        # Hm, already gone. Odd.
                        continue
                    handle_dir(path)
                    maybe_remove_empty_dir_containing(path, 1)
                    continue

                if len(line) != 16:
                    raise ValueError("Bad record in ", self.blob_dir, '.removed')

                oid, tid = line[:8], line[8:]
                path = fshelper.getBlobFilename(oid, tid)
                if not os.path.exists(path):
                    # Hm, already gone. Odd.
                    continue
                handle_file(path)
                assert not os.path.exists(path)
                maybe_remove_empty_dir_containing(path)

        os.remove(os.path.join(self.blob_dir, '.removed'))

        if not self.pack_keep_old:
            return

        # Second step, copy remaining files.
        for path, dir_names, file_names in os.walk(self.blob_dir):
            for file_name in file_names:
                if not file_name.endswith('.blob'):
                    continue
                file_path = os.path.join(path, file_name)
                dest = os.path.dirname(old+file_path[lblob_dir:])
                if not os.path.exists(dest):
                    os.makedirs(dest, 0o700)
                link_or_copy(file_path, old+file_path[lblob_dir:])

    def iterator(self, start=None, stop=None):
        return FileIterator(self._file_name, start, stop)

    def lastInvalidations(self, count):
        file = self._file
        seek = file.seek
        read = file.read
        with self._lock:
            pos = self._pos
            while count > 0 and pos > 4:
                count -= 1
                seek(pos-8)
                pos = pos - 8 - u64(read(8))

            seek(0)
            return [(trans.tid, [r.oid for r in trans])
                    for trans in FileIterator(self._file_name, pos=pos)]

    def lastTid(self, oid):
        """Return last serialno committed for object oid.

        If there is no serialno for this oid -- which can only occur
        if it is a new object -- return None.
        """
        try:
            return self.getTid(oid)
        except KeyError:
            return None

    def cleanup(self):
        """Remove all files created by this storage."""
        for ext in '', '.old', '.tmp', '.lock', '.index', '.pack':
            try:
                os.remove(self._file_name + ext)
            except OSError as e:
                if e.errno != errno.ENOENT:
                    raise

    def record_iternext(self, next=None):
        index = self._index
        oid = index.minKey(next)

        oid_as_long, = unpack(">Q", oid)
        next_oid = pack(">Q", oid_as_long + 1)
        try:
            next_oid = index.minKey(next_oid)
        except ValueError: # "empty tree" error
            next_oid = None

        data, tid = self.load(oid, "")

        return oid, tid, data, next_oid

    ######################################################################
    # The following 2 methods are for testing a ZEO extension mechanism
    def getExtensionMethods(self):
        return dict(answer_to_the_ultimate_question=None)

    def answer_to_the_ultimate_question(self):
        return 42
    #
    ######################################################################

def shift_transactions_forward(index, tindex, file, pos, opos):
    """Copy transactions forward in the data file

    This might be done as part of a recovery effort
    """

    # Cache a bunch of methods
    seek=file.seek
    read=file.read
    write=file.write

    index_get=index.get

    # Initialize,
    pv=z64
    p1=opos
    p2=pos
    offset=p2-p1

    # Copy the data in two stages.  In the packing stage,
    # we skip records that are non-current or that are for
    # unreferenced objects. We also skip undone transactions.
    #
    # After the packing stage, we copy everything but undone
    # transactions, however, we have to update various back pointers.
    # We have to have the storage lock in the second phase to keep
    # data from being changed while we're copying.
    pnv=None
    while 1:

        # Read the transaction record
        seek(pos)
        h=read(TRANS_HDR_LEN)
        if len(h) < TRANS_HDR_LEN: break
        tid, stl, status, ul, dl, el = unpack(TRANS_HDR,h)
        status = as_text(status)
        if status=='c': break # Oops. we found a checkpoint flag.
        tl=u64(stl)
        tpos=pos
        tend=tpos+tl

        otpos=opos # start pos of output trans

        thl=ul+dl+el
        h2=read(thl)
        if len(h2) != thl:
            raise PackError(opos)

        # write out the transaction record
        seek(opos)
        write(h)
        write(h2)

        thl=TRANS_HDR_LEN+thl
        pos=tpos+thl
        opos=otpos+thl

        while pos < tend:
            # Read the data records for this transaction
            seek(pos)
            h=read(DATA_HDR_LEN)
            oid,serial,sprev,stloc,vlen,splen = unpack(DATA_HDR, h)
            assert not vlen
            plen=u64(splen)
            dlen=DATA_HDR_LEN+(plen or 8)

            tindex[oid]=opos

            if plen: p=read(plen)
            else:
                p=read(8)
                p=u64(p)
                if p >= p2: p=p-offset
                elif p >= p1:
                    # Ick, we're in trouble. Let's bail
                    # to the index and hope for the best
                    p=index_get(oid, 0)
                p=p64(p)

            # WRITE
            seek(opos)
            sprev=p64(index_get(oid, 0))
            write(pack(DATA_HDR,
                       oid, serial, sprev, p64(otpos), 0, splen))

            write(p)

            opos=opos+dlen
            pos=pos+dlen

        # skip the (intentionally redundant) transaction length
        pos=pos+8

        if status != 'u':
            index.update(tindex) # Record the position

        tindex.clear()

        write(stl)
        opos=opos+8

    return opos

def search_back(file, pos):
    seek=file.seek
    read=file.read
    seek(0,2)
    s=p=file.tell()
    while p > pos:
        seek(p-8)
        l=u64(read(8))
        if l <= 0: break
        p=p-l-8

    return p, s

def recover(file_name):
    file=open(file_name, 'r+b')
    index={}
    tindex={}

    pos, oid, tid = read_index(file, file_name, index, tindex, recover=1)
    if oid is not None:
        print("Nothing to recover")
        return

    opos=pos
    pos, sz = search_back(file, pos)
    if pos < sz:
        npos = shift_transactions_forward(index, tindex, file, pos, opos)

    file.truncate(npos)

    print("Recovered file, lost %s, ended up with %s bytes" % (
        pos-opos, npos))



def read_index(file, name, index, tindex, stop=b'\377'*8,
               ltid=z64, start=4, maxoid=z64, recover=0, read_only=0):
    """Scan the file storage and update the index.

    Returns file position, max oid, and last transaction id.  It also
    stores index information in the three dictionary arguments.

    Arguments:
    file -- a file object (the Data.fs)
    name -- the name of the file (presumably file.name)
    index -- fsIndex, oid -> data record file offset
    tindex -- dictionary, oid -> data record offset
              tindex is cleared before return

    There are several default arguments that affect the scan or the
    return values.  TODO:  document them.

    start -- the file position at which to start scanning for oids added
             beyond the ones the passed-in indices know about.  The .index
             file caches the highest ._pos FileStorage knew about when the
             the .index file was last saved, and that's the intended value
             to pass in for start; accept the default (and pass empty
             indices) to recreate the index from scratch
    maxoid -- ignored (it meant something prior to ZODB 3.2.6; the argument
              still exists just so the signature of read_index() stayed the
              same)

    The file position returned is the position just after the last
    valid transaction record.  The oid returned is the maximum object
    id in `index`, or z64 if the index is empty.  The transaction id is the
    tid of the last transaction, or ltid if the index is empty.
    """

    read = file.read
    seek = file.seek
    seek(0, 2)
    file_size = file.tell()
    fmt = TempFormatter(file)

    if file_size:
        if file_size < start:
            raise FileStorageFormatError(file.name)
        seek(0)
        if read(4) != packed_version:
            raise FileStorageFormatError(name)
    else:
        if not read_only:
            file.write(packed_version)
        return 4, z64, ltid

    index_get = index.get

    pos = start
    seek(start)
    tid = b'\0' * 7 + b'\1'

    while 1:
        # Read the transaction record
        h = read(TRANS_HDR_LEN)
        if not h:
            break
        if len(h) != TRANS_HDR_LEN:
            if not read_only:
                logger.warning('%s truncated at %s', name, pos)
                seek(pos)
                file.truncate()
            break

        tid, tl, status, ul, dl, el = unpack(TRANS_HDR, h)
        status = as_text(status)

        if tid <= ltid:
            logger.warning("%s time-stamp reduction at %s", name, pos)
        ltid = tid

        if pos+(tl+8) > file_size or status=='c':
            # Hm, the data were truncated or the checkpoint flag wasn't
            # cleared.  They may also be corrupted,
            # in which case, we don't want to totally lose the data.
            if not read_only:
                logger.warning("%s truncated, possibly due to damaged"
                               " records at %s", name, pos)
                _truncate(file, name, pos)
            break

        if status not in ' up':
            logger.warning('%s has invalid status, %s, at %s',
                           name, status, pos)

        if tl < TRANS_HDR_LEN + ul + dl + el:
            # We're in trouble. Find out if this is bad data in the
            # middle of the file, or just a turd that Win 9x dropped
            # at the end when the system crashed.
            # Skip to the end and read what should be the transaction length
            # of the last transaction.
            seek(-8, 2)
            rtl = u64(read(8))
            # Now check to see if the redundant transaction length is
            # reasonable:
            if file_size - rtl < pos or rtl < TRANS_HDR_LEN:
                logger.critical('%s has invalid transaction header at %s',
                                name, pos)
                if not read_only:
                    logger.warning(
                         "It appears that there is invalid data at the end of "
                         "the file, possibly due to a system crash.  %s "
                         "truncated to recover from bad data at end." % name)
                    _truncate(file, name, pos)
                break
            else:
                if recover:
                    return pos, None, None
                panic('%s has invalid transaction header at %s', name, pos)

        if tid >= stop:
            break

        tpos = pos
        tend = tpos + tl

        if status == 'u':
            # Undone transaction, skip it
            seek(tend)
            h = u64(read(8))
            if h != tl:
                if recover:
                    return tpos, None, None
                panic('%s has inconsistent transaction length at %s',
                      name, pos)
            pos = tend + 8
            continue

        pos = tpos + TRANS_HDR_LEN + ul + dl + el
        while pos < tend:
            # Read the data records for this transaction
            h = fmt._read_data_header(pos)
            dlen = h.recordlen()
            tindex[h.oid] = pos

            if pos + dlen > tend or h.tloc != tpos:
                if recover:
                    return tpos, None, None
                panic("%s data record exceeds transaction record at %s",
                      name, pos)

            if index_get(h.oid, 0) != h.prev:
                if h.prev:
                    if recover:
                        return tpos, None, None
                    logger.error("%s incorrect previous pointer at %s",
                                 name, pos)
                else:
                    logger.warning("%s incorrect previous pointer at %s",
                                   name, pos)

            pos += dlen

        if pos != tend:
            if recover:
                return tpos, None, None
            panic("%s data records don't add up at %s",name,tpos)

        # Read the (intentionally redundant) transaction length
        seek(pos)
        h = u64(read(8))
        if h != tl:
            if recover:
                return tpos, None, None
            panic("%s redundant transaction length check failed at %s",
                  name, pos)
        pos += 8

        index.update(tindex)
        tindex.clear()

    # Caution:  fsIndex doesn't have an efficient __nonzero__ or __len__.
    # That's why we do try/except instead.  fsIndex.maxKey() is fast.
    try:
        maxoid = index.maxKey()
    except ValueError:
        # The index is empty.
        maxoid == z64

    return pos, maxoid, ltid


def _truncate(file, name, pos):
    file.seek(0, 2)
    file_size = file.tell()
    try:
        i = 0
        while 1:
            oname='%s.tr%s' % (name, i)
            if os.path.exists(oname):
                i += 1
            else:
                logger.warning("Writing truncated data from %s to %s",
                               name, oname)
                o = open(oname,'wb')
                file.seek(pos)
                cp(file, o, file_size-pos)
                o.close()
                break
    except:
        logger.error("couldn\'t write truncated data for %s", name,
              exc_info=True)
        raise StorageSystemError("Couldn't save truncated data")

    file.seek(pos)
    file.truncate()


class FileIterator(FileStorageFormatter):
    """Iterate over the transactions in a FileStorage file.
    """
    _ltid = z64
    _file = None

    def __init__(self, filename, start=None, stop=None, pos=4):
        assert isinstance(filename, STRING_TYPES)
        file = open(filename, 'rb')
        self._file = file
        self._file_name = filename
        if file.read(4) != packed_version:
            raise FileStorageFormatError(file.name)
        file.seek(0,2)
        self._file_size = file.tell()
        if (pos < 4) or pos > self._file_size:
            raise ValueError("Given position is greater than the file size",
                             pos, self._file_size)
        self._pos = pos
        assert start is None or isinstance(start, bytes)
        assert stop is None or isinstance(stop, bytes)
        self._start = start
        self._stop = stop
        if start:
            if self._file_size <= 4:
                return
            self._skip_to_start(start)

    def __len__(self):
        # Define a bogus __len__() to make the iterator work
        # with code like builtin list() and tuple() in Python 2.1.
        # There's a lot of C code that expects a sequence to have
        # an __len__() but can cope with any sort of mistake in its
        # implementation.  So just return 0.
        return 0

    # This allows us to pass an iterator as the `other` argument to
    # copyTransactionsFrom() in BaseStorage.  The advantage here is that we
    # can create the iterator manually, e.g. setting start and stop, and then
    # just let copyTransactionsFrom() do its thing.
    def iterator(self):
        return self

    def close(self):
        file = self._file
        if file is not None:
            self._file = None
            file.close()

    def _skip_to_start(self, start):
        file = self._file
        pos1 = self._pos
        file.seek(pos1)
        tid1 = file.read(8) # XXX bytes
        if len(tid1) < 8:
            raise CorruptedError("Couldn't read tid.")
        if start < tid1:
            pos2 = pos1
            tid2 = tid1
            file.seek(4)
            tid1 = file.read(8)
            if start <= tid1:
                self._pos = 4
                return
            pos1 = 4
        else:
            if start == tid1:
                return

            # Try to read the last transaction. We could be unlucky and
            # opened the file while committing a transaction.  In that
            # case, we'll just scan from the beginning if the file is
            # small enough, otherwise we'll fail.
            file.seek(self._file_size-8)
            l = u64(file.read(8))
            if not (l + 12 <= self._file_size and
                    self._read_num(self._file_size-l) == l):
                if self._file_size < (1<<20):
                    return self._scan_foreward(start)
                raise ValueError("Can't find last transaction in large file")
            pos2 = self._file_size-l-8
            file.seek(pos2)
            tid2 = file.read(8)
            if tid2 < tid1:
                raise CorruptedError("Tids out of order.")
            if tid2 <= start:
                if tid2 == start:
                    self._pos = pos2
                else:
                    self._pos = self._file_size
                return

        t1 = TimeStamp(tid1).timeTime()
        t2 = TimeStamp(tid2).timeTime()
        ts = TimeStamp(start).timeTime()
        if (ts - t1) < (t2 - ts):
            return self._scan_forward(pos1, start)
        else:
            return self._scan_backward(pos2, start)

    def _scan_forward(self, pos, start):
        logger.debug("Scan forward %s:%s looking for %r",
                     self._file_name, pos, start)
        file = self._file
        while 1:
            # Read the transaction record
            h = self._read_txn_header(pos)
            if h.tid >= start:
                self._pos = pos
                return

            pos += h.tlen + 8

    def _scan_backward(self, pos, start):
        logger.debug("Scan backward %s:%s looking for %r",
                     self._file_name, pos, start)
        file = self._file
        seek = file.seek
        read = file.read
        while 1:
            pos -= 8
            seek(pos)
            tlen = u64(read(8))
            pos -= tlen
            h = self._read_txn_header(pos)
            if h.tid <= start:
                if h.tid == start:
                    self._pos = pos
                else:
                    self._pos = pos + tlen + 8
                return

    # Iterator protocol
    def __iter__(self):
        return self

    def __next__(self):
        if self._file is None:
            raise StopIteration()

        pos = self._pos
        while True:

            # Read the transaction record
            try:
                h = self._read_txn_header(pos)
            except CorruptedDataError as err:
                # If buf is empty, we've reached EOF.
                if not err.buf:
                    break
                raise

            if h.tid <= self._ltid:
                logger.warning("%s time-stamp reduction at %s",
                               self._file.name, pos)
            self._ltid = h.tid

            if self._stop is not None and h.tid > self._stop:
                break

            if h.status == "c":
                # Assume we've hit the last, in-progress transaction
                break

            if pos + h.tlen + 8 > self._file_size:
                # Hm, the data were truncated or the checkpoint flag wasn't
                # cleared.  They may also be corrupted,
                # in which case, we don't want to totally lose the data.
                logger.warning("%s truncated, possibly due to"
                               " damaged records at %s", self._file.name, pos)
                break

            if h.status not in " up":
                logger.warning('%s has invalid status,'
                               ' %s, at %s', self._file.name, h.status, pos)

            if h.tlen < h.headerlen():
                # We're in trouble. Find out if this is bad data in
                # the middle of the file, or just a turd that Win 9x
                # dropped at the end when the system crashed.  Skip to
                # the end and read what should be the transaction
                # length of the last transaction.
                self._file.seek(-8, 2)
                rtl = u64(self._file.read(8))
                # Now check to see if the redundant transaction length is
                # reasonable:
                if self._file_size - rtl < pos or rtl < TRANS_HDR_LEN:
                    logger.critical("%s has invalid transaction header at %s",
                                    self._file.name, pos)
                    logger.warning(
                         "It appears that there is invalid data at the end of "
                         "the file, possibly due to a system crash.  %s "
                         "truncated to recover from bad data at end."
                         % self._file.name)
                    break
                else:
                    logger.warning("%s has invalid transaction header at %s",
                                   self._file.name, pos)
                    break

            tpos = pos
            tend = tpos + h.tlen

            if h.status != "u":
                pos = tpos + h.headerlen()
                e = {}
                if h.elen:
                    try:
                        e = loads(h.ext)
                    except:
                        pass

                result = TransactionRecord(h.tid, h.status, h.user, h.descr,
                                           e, pos, tend, self._file, tpos)

            # Read the (intentionally redundant) transaction length
            self._file.seek(tend)
            rtl = u64(self._file.read(8))
            if rtl != h.tlen:
                logger.warning("%s redundant transaction length check"
                               " failed at %s", self._file.name, tend)
                break
            self._pos = tend + 8

            return result

        self.close()
        raise StopIteration()

    next = __next__


class TransactionRecord(_TransactionRecord):

    def __init__(self, tid, status, user, desc, ext, pos, tend, file, tpos):
        _TransactionRecord.__init__(
            self, tid, status, user, desc, ext)
        self._pos = pos
        self._tend = tend
        self._file = file
        self._tpos = tpos

    def __iter__(self):
        return TransactionRecordIterator(self)

class TransactionRecordIterator(FileStorageFormatter):
    """Iterate over the transactions in a FileStorage file."""

    def __init__(self, record):
        self._file = record._file
        self._pos = record._pos
        self._tpos = record._tpos
        self._tend = record._tend

    def __iter__(self):
        return self

    def __next__(self):
        pos = self._pos
        while pos < self._tend:
            # Read the data records for this transaction
            h = self._read_data_header(pos)
            dlen = h.recordlen()

            if pos + dlen > self._tend or h.tloc != self._tpos:
                logger.warning("%s data record exceeds transaction"
                               " record at %s", file.name, pos)
                break

            self._pos = pos + dlen
            prev_txn = None
            if h.plen:
                data = self._file.read(h.plen)
            else:
                if h.back == 0:
                    # If the backpointer is 0, then this transaction
                    # undoes the object creation.  It undid the
                    # transaction that created it.  Return None
                    # instead of a pickle to indicate this.
                    data = None
                else:
                    data, tid = self._loadBackTxn(h.oid, h.back, False)
                    # Caution:  :ooks like this only goes one link back.
                    # Should it go to the original data like BDBFullStorage?
                    prev_txn = self.getTxnFromData(h.oid, h.back)

            return Record(h.oid, h.tid, data, prev_txn, pos)

        raise StopIteration()

    next = __next__


class Record(_DataRecord):

    def __init__(self, oid, tid, data, prev, pos):
        super(Record, self).__init__(oid, tid, data, prev)
        self.pos = pos


class UndoSearch:

    def __init__(self, file, pos, first, last, filter=None):
        self.file = file
        self.pos = pos
        self.first = first
        self.last = last
        self.filter = filter
        # self.i is the index of the transaction we're _going_ to find
        # next.  When it reaches self.first, we should start appending
        # to self.results.  When it reaches self.last, we're done
        # (although we may finish earlier).
        self.i = 0
        self.results = []
        self.stop = False

    def finished(self):
        """Return True if UndoSearch has found enough records."""
        # BAW: Why 39 please?  This makes no sense (see also below).
        return self.i >= self.last or self.pos < 39 or self.stop

    def search(self):
        """Search for another record."""
        dict = self._readnext()
        if dict is not None and (self.filter is None or self.filter(dict)):
            if self.i >= self.first:
                self.results.append(dict)
            self.i += 1

    def _readnext(self):
        """Read the next record from the storage."""
        self.file.seek(self.pos - 8)
        self.pos -= u64(self.file.read(8)) + 8
        self.file.seek(self.pos)
        h = self.file.read(TRANS_HDR_LEN)
        tid, tl, status, ul, dl, el = unpack(TRANS_HDR, h)
        status = as_text(status)
        if status == 'p':
            self.stop = 1
            return None
        if status != ' ':
            return None
        d = u = b''
        if ul:
            u = self.file.read(ul)
        if dl:
            d = self.file.read(dl)
        e = {}
        if el:
            try:
                e = loads(self.file.read(el))
            except:
                pass
        d = {'id': encodebytes(tid).rstrip(),
             'time': TimeStamp(tid).timeTime(),
             'user_name': u,
             'size': tl,
             'description': d}
        d.update(e)
        return d

class FilePool:

    closed = False
    writing = False
    writers = 0

    def __init__(self, file_name):
        self.name = file_name
        self._files = []
        self._out = []
        self._cond = threading.Condition()

    @contextlib.contextmanager
    def write_lock(self):
        with self._cond:
            self.writers += 1
            while self.writing or self._out:
                self._cond.wait()
            if self.closed:
                raise ValueError('closed')
            self.writing = True

        try:
            yield None
        finally:
            with self._cond:
                self.writing = False
                if self.writers > 0:
                    self.writers -= 1
                self._cond.notifyAll()

    @contextlib.contextmanager
    def get(self):
        with self._cond:
            while self.writers:
                self._cond.wait()
            assert not self.writing
            if self.closed:
                raise ValueError('closed')

            try:
                f = self._files.pop()
            except IndexError:
                f = open(self.name, 'rb')
            self._out.append(f)

        try:
            yield f
        finally:
            self._out.remove(f)
            self._files.append(f)
            if not self._out:
                with self._cond:
                    if self.writers and not self._out:
                        self._cond.notifyAll()

    def empty(self):
        while self._files:
            self._files.pop().close()

    def close(self):
        with self._cond:
            self.closed = True
            while self._out:
                self._out.pop().close()
            self.empty()
            self.writing = self.writers = 0