borg/borg/archive.py

from datetime import datetime
from getpass import getuser
from itertools import groupby
import errno
import shutil
import tempfile
import threading
from .key import key_factory
from .remote import cache_if_remote
import msgpack
from multiprocessing import cpu_count
import os
import socket
import stat
import sys
import time
from io import BytesIO
from . import xattr
from .platform import acl_get, acl_set
from .chunker import Chunker
from .hashindex import ChunkIndex
from .helpers import parse_timestamp, Error, uid2user, user2uid, gid2group, group2gid, \
    Manifest, Statistics, decode_dict, st_mtime_ns, make_path_safe, StableDict, int_to_bigint, bigint_to_int, \
    make_queue, TerminatedQueue

ITEMS_BUFFER = 1024 * 1024
CHUNK_MIN = 1024
CHUNK_MAX = 10 * 1024 * 1024
WINDOW_SIZE = 0xfff
CHUNK_MASK = 0xffff

ZEROS = b'\0' * CHUNK_MAX

utime_supports_fd = os.utime in getattr(os, 'supports_fd', {})
utime_supports_follow_symlinks = os.utime in getattr(os, 'supports_follow_symlinks', {})
has_mtime_ns = sys.version >= '3.3'
has_lchmod = hasattr(os, 'lchmod')
has_lchflags = hasattr(os, 'lchflags')

# Python <= 3.2 raises OSError instead of PermissionError (See #164)
try:
    PermissionError = PermissionError
except NameError:
    PermissionError = OSError


class DownloadPipeline:

    def __init__(self, repository, key):
        self.repository = repository
        self.key = key

    def unpack_many(self, ids, filter=None, preload=False):
        unpacker = msgpack.Unpacker(use_list=False)
        for data in self.fetch_many(ids):
            unpacker.feed(data)
            items = [decode_dict(item, (b'path', b'source', b'user', b'group')) for item in unpacker]
            if filter:
                items = [item for item in items if filter(item)]
            if preload:
                for item in items:
                    if b'chunks' in item:
                        self.repository.preload([c[0] for c in item[b'chunks']])
            for item in items:
                yield item

    def fetch_many(self, ids, is_preloaded=False):
        for id_, data in zip(ids, self.repository.get_many(ids, is_preloaded=is_preloaded)):
            yield self.key.decrypt(id_, data)


class ChunkBuffer:
    BUFFER_SIZE = 1 * 1024 * 1024

    def __init__(self, key):
        self.buffer = BytesIO()
        self.packer = msgpack.Packer(unicode_errors='surrogateescape')
        self.chunks = []
        self.key = key
        self.chunker = Chunker(WINDOW_SIZE, CHUNK_MASK, CHUNK_MIN, CHUNK_MAX,self.key.chunk_seed)

    def add(self, item):
        self.buffer.write(self.packer.pack(StableDict(item)))
        if self.is_full():
            self.flush()

    def write_chunk(self, chunk):
        raise NotImplementedError

    def flush(self, flush=False):
        if self.buffer.tell() == 0:
            return
        self.buffer.seek(0)
        chunks = list(bytes(s) for s in self.chunker.chunkify(self.buffer))
        self.buffer.seek(0)
        self.buffer.truncate(0)
        # Leave the last partial chunk in the buffer unless flush is True
        end = None if flush or len(chunks) == 1 else -1
        for chunk in chunks[:end]:
            self.chunks.append(self.write_chunk(chunk))
        if end == -1:
            self.buffer.write(chunks[-1])

    def is_full(self):
        return self.buffer.tell() > self.BUFFER_SIZE


class CacheChunkBuffer(ChunkBuffer):

    def __init__(self, cache, key, stats):
        super(CacheChunkBuffer, self).__init__(key)
        self.cache = cache
        self.stats = stats

    def write_chunk(self, chunk):
        id_, _, _ = self.cache.add_chunk(self.key.id_hash(chunk), chunk, self.stats)
        return id_


class ParallelProcessor:
    def __init__(self, archive, ncrypters=None):
        self.archive = archive
        if ncrypters is None:
            # note: cpu_count for 2 cores with HT is 4
            # put load on all logical cores and avoid idle cores
            ncrypters = cpu_count()
        self.ncrypters = ncrypters
        self.start_threads()

    def reader(self):
        while True:
            elem = self.reader_queue.get()
            if elem is None:
                self.reader_queue.task_done()
                break
            item = elem
            n = 0
            # Only chunkify the file if needed
            if b'chunks' in item and item[b'chunks'] is None:
                fd, fh = item.pop(b'fd', None), -1
                if fd is None:
                    fh = Archive._open_rb(item.pop(b'path_name'), item[b'st'])
                    fd = os.fdopen(fh, 'rb')
                with fd:
                    for chunk in self.archive.chunker.chunkify(fd, fh):
                        # important: chunk is a memoryview - make a copy or it will
                        # have changed when we use it!
                        chunk = bytes(chunk)
                        self.crypter_queue.put((item, n, chunk))
                        n += 1
            self.writer_queue.put((item, n, None, None, None, None))  # signal EOF via id == None , give number of chunks
            self.reader_queue.task_done()

    def crypter(self):
        while True:
            elem = self.crypter_queue.get()
            if elem is None:
                self.crypter_queue.task_done()
                break
            item, n, chunk = elem
            size = len(chunk)
            id = self.archive.key.id_hash(chunk)
            seen = self.archive.cache.seen_or_announce_chunk(id, size)
            if not seen:
                # we have never seen this id before, so we need to process it
                # TODO check if this creates duplicate IV/CTR values for AES
                cchunk = self.archive.key.encrypt(chunk)
                csize = len(cchunk)
            else:
                cchunk, csize = None, None
            self.writer_queue.put((item, n, cchunk, id, size, csize))
            self.crypter_queue.task_done()

    def writer(self):
        item_infos = {}  # item path -> info dict
        size_infos = {}  # chunk id -> sizes
        dying = False
        while True:
            elem = self.writer_queue.get()
            if elem is None:
                if not dying:
                    # received poison from stop_threads, start dying,
                    # but still do work the delayer thread might give us.
                    dying = True
                    # give poison to the delayer thread
                    self.delayer_queue.put(None)
                    self.writer_queue.task_done()
                    continue
                else:
                    # we received the final poison from the dying delayer
                    self.writer_queue.task_done()
                    # we are dead now
                    break
            item, n, cchunk, id, size, csize = elem
            path = item[b'path']
            info = item_infos.setdefault(path, dict(count=None, chunks=[]))
            if id is None:
                if n is not None:  # note: n == None is a retry
                    # EOF signalled, n is the total count of chunks
                    info['count'] = n
            else:
                size, csize, new_chunk = self.archive.cache.add_chunk_nostats(cchunk, id, size, csize)
                info['chunks'].append((n, id, new_chunk))
                if csize != 0:
                    size_infos[id] = (size, csize)
            if len(info['chunks']) == info['count']:
                # we have processed all chunks or no chunks needed processing
                if b'chunks' in item:
                    chunks = item[b'chunks']
                    if chunks is None:
                        # we want chunks, but we have no chunk id list yet, compute them
                        try:
                            chunks = self.archive.cache.postprocess_results(
                                size_infos, info['chunks'], self.archive.stats)
                        except self.archive.cache.ChunkSizeNotReady:
                            # we looked up a chunk id, but do not have the size info yet. retry later.
                            self.delayer_queue.put((item, None, None, None, None, None))
                            self.writer_queue.task_done()
                            continue
                    else:
                        # we have a chunk id list already, increase the ref counters, compute sizes
                        chunks = [self.archive.cache.chunk_incref(id_, self.archive.stats) for id_ in chunks]
                    item[b'chunks'] = chunks
                path_hash = item.pop(b'path_hash', None)
                if path_hash and chunks is not None:  # a fs object (not stdin) and a regular file
                    st = item.pop(b'st', None)
                    self.archive.cache.memorize_file(path_hash, st, [c[0] for c in chunks])
                del item_infos[path]
                self.archive.stats.nfiles += 1
                self.archive.add_item(item)
            self.writer_queue.task_done()

    def delayer(self):
        # it is a pain that we need the compressed size for the chunks cache as it is not
        # available for duplicate chunks until the original chunk has finished processing.
        # this loop of (writer, delayer) with pipes connecting them is a hack to address
        # this, but it makes thread teardown complicated. Rather get rid of csize?
        while True:
            elem = self.delayer_queue.get()
            if elem is None:
                # we received poison from dying writer thread, kill the writer, too.
                self.writer_queue.put(None)
                self.delayer_queue.task_done()
                # we are dead now
                break
            time.sleep(0.001)  # reschedule, avoid data circulating too fast
            self.writer_queue.put(elem)
            self.delayer_queue.task_done()

    def start_threads(self):
        def run_thread(func, name=None, daemon=False):
            t = threading.Thread(target=func, name=name)
            t.daemon = daemon
            t.start()
            return t

        # max. memory usage of a queue with chunk data is about queue_len * CHUNK_MAX
        queue_len = min(max(self.ncrypters, 4), 8)
        self.reader_queue = make_queue('reader', queue_len * 10)  # small items (no chunk data)
        self.crypter_queue = make_queue('crypter', queue_len)
        self.writer_queue = make_queue('writer', queue_len)
        self.delayer_queue = make_queue('delay', queue_len)
        self.reader_thread = run_thread(self.reader, 'reader')
        self.crypter_threads = []
        for i in range(self.ncrypters):
            self.crypter_threads.append(run_thread(self.crypter, name='crypter-%d' % i))
        self.delayer_thread = run_thread(self.delayer, name='delayer')
        self.writer_thread = run_thread(self.writer, name='writer')

    def wait_finish(self):
        self.reader_queue.join()
        self.crypter_queue.join()
        self.writer_queue.join()
        self.delayer_queue.join()
        self.writer_queue.join()

    def stop_threads(self):
        count_before = threading.active_count()
        # for every thread:
        #   put poison pill into its queue,
        #   wait until queue is processed (and thread has terminated itself)
        #   make queue unusable
        self.reader_queue.put(None)
        self.reader_queue.join()
        self.reader_thread.join()
        self.reader_queue = TerminatedQueue()
        for i in range(self.ncrypters):
            self.crypter_queue.put(None)
        self.crypter_queue.join()
        for t in self.crypter_threads:
            t.join()
        self.crypter_queue = TerminatedQueue()
        self.writer_queue.put(None)  # the writer will poison the delayer first
        self.delayer_thread.join()
        self.delayer_queue = TerminatedQueue()
        self.writer_thread.join()
        self.writer_queue = TerminatedQueue()
        count_after = threading.active_count()
        assert count_before - 3 - self.ncrypters == count_after
        if count_after > 1:
            print('They are alive!')
            tl = [t.name for t in threading.enumerate()]
            tl.remove('MainThread')
            assert tl == []


class Archive:

    class DoesNotExist(Error):
        """Archive {} does not exist"""

    class AlreadyExists(Error):
        """Archive {} already exists"""

    class IncompatibleFilesystemEncodingError(Error):
        """Failed to encode filename "{}" into file system encoding "{}". Consider configuring the LANG environment variable."""


    def __init__(self, repository, key, manifest, name, cache=None, create=False,
                 checkpoint_interval=300, numeric_owner=False, progress=False):
        self.cwd = os.getcwd()
        self.key = key
        self.repository = repository
        self.cache = cache
        self.manifest = manifest
        self.hard_links = {}
        self.stats = Statistics()
        self.show_progress = progress
        self.last_progress = time.time()
        self.name = name
        self.checkpoint_interval = checkpoint_interval
        self.numeric_owner = numeric_owner
        self.pipeline = DownloadPipeline(self.repository, self.key)
        if create:
            self.pp = ParallelProcessor(self)
            self.items_buffer = CacheChunkBuffer(self.cache, self.key, self.stats)
            self.chunker = Chunker(WINDOW_SIZE, CHUNK_MASK, CHUNK_MIN, CHUNK_MAX, self.key.chunk_seed)
            if name in manifest.archives:
                raise self.AlreadyExists(name)
            self.last_checkpoint = time.time()
            i = 0
            while True:
                self.checkpoint_name = '%s.checkpoint%s' % (name, i and ('.%d' % i) or '')
                if self.checkpoint_name not in manifest.archives:
                    break
                i += 1
        else:
            self.pp = None
            if name not in self.manifest.archives:
                raise self.DoesNotExist(name)
            info = self.manifest.archives[name]
            self.load(info[b'id'])

    def close(self):
        if self.pp:
            self.pp.stop_threads()

    def _load_meta(self, id):
        data = self.key.decrypt(id, self.repository.get(id))
        metadata = msgpack.unpackb(data)
        if metadata[b'version'] != 1:
            raise Exception('Unknown archive metadata version')
        return metadata

    def load(self, id):
        self.id = id
        self.metadata = self._load_meta(self.id)
        decode_dict(self.metadata, (b'name', b'hostname', b'username', b'time'))
        self.metadata[b'cmdline'] = [arg.decode('utf-8', 'surrogateescape') for arg in self.metadata[b'cmdline']]
        self.name = self.metadata[b'name']

    @property
    def ts(self):
        """Timestamp of archive creation in UTC"""
        return parse_timestamp(self.metadata[b'time'])

    def __repr__(self):
        return 'Archive(%r)' % self.name

    def iter_items(self, filter=None, preload=False):
        for item in self.pipeline.unpack_many(self.metadata[b'items'], filter=filter, preload=preload):
            yield item

    def add_item_queued(self, item):
        self.pp.reader_queue.put(item)

    def add_item(self, item):
        if self.show_progress and time.time() - self.last_progress > 0.2:
            self.stats.show_progress(item=item)
            self.last_progress = time.time()
        self.items_buffer.add(item)
        if time.time() - self.last_checkpoint > self.checkpoint_interval:
            self.write_checkpoint()
            self.last_checkpoint = time.time()

    def write_checkpoint(self):
        self.save(self.checkpoint_name)
        del self.manifest.archives[self.checkpoint_name]
        self.cache.chunk_decref(self.id, self.stats)

    def save(self, name=None, timestamp=None):
        self.pp.wait_finish()
        name = name or self.name
        if name in self.manifest.archives:
            raise self.AlreadyExists(name)
        self.items_buffer.flush(flush=True)
        if timestamp is None:
            timestamp = datetime.utcnow()
        metadata = StableDict({
            'version': 1,
            'name': name,
            'items': self.items_buffer.chunks,
            'cmdline': sys.argv,
            'hostname': socket.gethostname(),
            'username': getuser(),
            'time': timestamp.isoformat(),
        })
        data = msgpack.packb(metadata, unicode_errors='surrogateescape')
        self.id = self.key.id_hash(data)
        self.cache.add_chunk(self.id, data, self.stats)
        self.manifest.archives[name] = {'id': self.id, 'time': metadata['time']}
        self.manifest.write()
        self.repository.commit()
        self.cache.commit()

    def calc_stats(self, cache):
        def add(id):
            count, size, csize = cache.chunks[id]
            stats.update(size, csize, count == 1)
            cache.chunks[id] = count - 1, size, csize
        def add_file_chunks(chunks):
            for id, _, _ in chunks:
                add(id)
        # This function is a bit evil since it abuses the cache to calculate
        # the stats. The cache transaction must be rolled back afterwards
        unpacker = msgpack.Unpacker(use_list=False)
        cache.begin_txn()
        stats = Statistics()
        add(self.id)
        for id, chunk in zip(self.metadata[b'items'], self.repository.get_many(self.metadata[b'items'])):
            add(id)
            unpacker.feed(self.key.decrypt(id, chunk))
            for item in unpacker:
                if b'chunks' in item:
                    stats.nfiles += 1
                    add_file_chunks(item[b'chunks'])
        cache.rollback()
        return stats

    def extract_item(self, item, restore_attrs=True, dry_run=False, stdout=False, sparse=False):
        if dry_run or stdout:
            if b'chunks' in item:
                for data in self.pipeline.fetch_many([c[0] for c in item[b'chunks']], is_preloaded=True):
                    if stdout:
                        sys.stdout.buffer.write(data)
                if stdout:
                    sys.stdout.buffer.flush()
            return

        dest = self.cwd
        if item[b'path'].startswith('/') or item[b'path'].startswith('..'):
            raise Exception('Path should be relative and local')
        path = os.path.join(dest, item[b'path'])
        # Attempt to remove existing files, ignore errors on failure
        try:
            st = os.lstat(path)
            if stat.S_ISDIR(st.st_mode):
                os.rmdir(path)
            else:
                # XXX do not remove a regular file, it could be the "source"
                # of a hardlink - a still empty inode that needs to be filled.
                pass
        except UnicodeEncodeError:
            raise self.IncompatibleFilesystemEncodingError(path, sys.getfilesystemencoding())
        except OSError:
            pass
        mode = item[b'mode']
        if stat.S_ISDIR(mode):
            if not os.path.exists(path):
                os.makedirs(path)
            if restore_attrs:
                self.restore_attrs(path, item)
        elif stat.S_ISREG(mode):
            if not os.path.exists(os.path.dirname(path)):
                os.makedirs(os.path.dirname(path))
            # Hard link?
            if b'source' in item:
                source = os.path.join(dest, item[b'source'])
                if os.path.exists(path):
                    os.unlink(path)
                if not os.path.exists(source):
                    # due to multithreaded nature and different processing time,
                    # the hardlink (without file content) often is in the archive
                    # BEFORE the "source" file (with content).
                    # we create an empty file that is filled with content when
                    # the "source" item is extracted:
                    with open(source, 'wb') as fd:
                        pass
                os.link(source, path)
            else:
                with open(path, 'wb') as fd:
                    ids = [c[0] for c in item[b'chunks']]
                    for data in self.pipeline.fetch_many(ids, is_preloaded=True):
                        if sparse and ZEROS.startswith(data):
                            # all-zero chunk: create a hole in a sparse file
                            fd.seek(len(data), 1)
                        else:
                            fd.write(data)
                    pos = fd.tell()
                    fd.truncate(pos)
                    fd.flush()
                    self.restore_attrs(path, item, fd=fd.fileno())
        elif stat.S_ISFIFO(mode):
            if not os.path.exists(os.path.dirname(path)):
                os.makedirs(os.path.dirname(path))
            os.mkfifo(path)
            self.restore_attrs(path, item)
        elif stat.S_ISLNK(mode):
            if not os.path.exists(os.path.dirname(path)):
                os.makedirs(os.path.dirname(path))
            source = item[b'source']
            if os.path.exists(path):
                os.unlink(path)
            os.symlink(source, path)
            self.restore_attrs(path, item, symlink=True)
        elif stat.S_ISCHR(mode) or stat.S_ISBLK(mode):
            os.mknod(path, item[b'mode'], item[b'rdev'])
            self.restore_attrs(path, item)
        else:
            raise Exception('Unknown archive item type %r' % item[b'mode'])

    def restore_attrs(self, path, item, symlink=False, fd=None):
        xattrs = item.get(b'xattrs')
        if xattrs:
                for k, v in xattrs.items():
                    try:
                        xattr.setxattr(fd or path, k, v, follow_symlinks=False)
                    except OSError as e:
                        if e.errno != errno.ENOTSUP:
                            raise
        uid = gid = None
        if not self.numeric_owner:
            uid = user2uid(item[b'user'])
            gid = group2gid(item[b'group'])
        uid = item[b'uid'] if uid is None else uid
        gid = item[b'gid'] if gid is None else gid
        # This code is a bit of a mess due to os specific differences
        try:
            if fd:
                os.fchown(fd, uid, gid)
            else:
                os.lchown(path, uid, gid)
        except OSError:
            pass
        if fd:
            os.fchmod(fd, item[b'mode'])
        elif not symlink:
            os.chmod(path, item[b'mode'])
        elif has_lchmod:  # Not available on Linux
            os.lchmod(path, item[b'mode'])
        mtime = bigint_to_int(item[b'mtime'])
        if fd and utime_supports_fd:  # Python >= 3.3
            os.utime(fd, None, ns=(mtime, mtime))
        elif utime_supports_follow_symlinks:  # Python >= 3.3
            os.utime(path, None, ns=(mtime, mtime), follow_symlinks=False)
        elif not symlink:
            os.utime(path, (mtime / 1e9, mtime / 1e9))
        acl_set(path, item, self.numeric_owner)
        # Only available on OS X and FreeBSD
        if has_lchflags and b'bsdflags' in item:
            try:
                os.lchflags(path, item[b'bsdflags'])
            except OSError:
                pass

    def rename(self, name):
        if name in self.manifest.archives:
            raise self.AlreadyExists(name)
        metadata = StableDict(self._load_meta(self.id))
        metadata[b'name'] = name
        data = msgpack.packb(metadata, unicode_errors='surrogateescape')
        new_id = self.key.id_hash(data)
        self.cache.add_chunk(new_id, data, self.stats)
        self.manifest.archives[name] = {'id': new_id, 'time': metadata[b'time']}
        self.cache.chunk_decref(self.id, self.stats)
        del self.manifest.archives[self.name]

    def delete(self, stats):
        unpacker = msgpack.Unpacker(use_list=False)
        for items_id, data in zip(self.metadata[b'items'], self.repository.get_many(self.metadata[b'items'])):
            unpacker.feed(self.key.decrypt(items_id, data))
            self.cache.chunk_decref(items_id, stats)
            for item in unpacker:
                if b'chunks' in item:
                    for chunk_id, size, csize in item[b'chunks']:
                        self.cache.chunk_decref(chunk_id, stats)

        self.cache.chunk_decref(self.id, stats)
        del self.manifest.archives[self.name]

    def stat_attrs(self, st, path):
        item = {
            b'mode': st.st_mode,
            b'uid': st.st_uid, b'user': uid2user(st.st_uid),
            b'gid': st.st_gid, b'group': gid2group(st.st_gid),
            b'mtime': int_to_bigint(st_mtime_ns(st))
        }
        if self.numeric_owner:
            item[b'user'] = item[b'group'] = None
        xattrs = xattr.get_all(path, follow_symlinks=False)
        if xattrs:
            item[b'xattrs'] = StableDict(xattrs)
        if has_lchflags and st.st_flags:
            item[b'bsdflags'] = st.st_flags
        acl_get(path, item, st, self.numeric_owner)
        return item

    def process_dir(self, path, st):
        item = {b'path': make_path_safe(path)}
        item.update(self.stat_attrs(st, path))
        self.add_item_queued(item)
        return 'd'  # directory

    def process_fifo(self, path, st):
        item = {b'path': make_path_safe(path)}
        item.update(self.stat_attrs(st, path))
        self.add_item_queued(item)
        return 'f'  # fifo

    def process_dev(self, path, st):
        item = {b'path': make_path_safe(path), b'rdev': st.st_rdev}
        item.update(self.stat_attrs(st, path))
        self.add_item_queued(item)
        if stat.S_ISCHR(st.st_mode):
            return 'c'  # char device
        elif stat.S_ISBLK(st.st_mode):
            return 'b'  # block device

    def process_symlink(self, path, st):
        source = os.readlink(path)
        item = {b'path': make_path_safe(path), b'source': source}
        item.update(self.stat_attrs(st, path))
        self.add_item_queued(item)
        return 's'  # symlink

    def process_stdin(self, path, cache):
        uid, gid = 0, 0
        item = {
            b'path': path,
            b'fd': sys.stdin.buffer,  # binary
            b'mode': 0o100660,  # regular file, ug=rw
            b'uid': uid, b'user': uid2user(uid),
            b'gid': gid, b'group': gid2group(gid),
            b'mtime': int_to_bigint(int(time.time()) * 1000000000)
        }
        self.add_item_queued(item)
        return 'A'

    def process_file(self, path, st, cache):
        status = None
        safe_path = make_path_safe(path)
        # Is it a hard link?
        if st.st_nlink > 1:
            source = self.hard_links.get((st.st_ino, st.st_dev))
            if (st.st_ino, st.st_dev) in self.hard_links:
                item = self.stat_attrs(st, path)
                item.update({b'path': safe_path, b'source': source})
                self.add_item_queued(item)
                status = 'h'  # regular file, hardlink (to already seen inodes)
                return status
            else:
                self.hard_links[st.st_ino, st.st_dev] = safe_path
        path_hash = self.key.id_hash(os.path.join(self.cwd, path).encode('utf-8', 'surrogateescape'))
        ids = cache.file_known_and_unchanged(path_hash, st)
        chunks = None
        if ids is not None:
            # Make sure all ids are available
            for id_ in ids:
                if not cache.seen_chunk(id_):
                    break
            else:
                chunks = ids
                status = 'U'  # regular file, unchanged
        else:
            status = 'A'  # regular file, added
        if chunks is None:
            status = status or 'M'  # regular file, modified (if not 'A' already)

        item = {
            b'path': safe_path,
            b'path_name': path,
            b'path_hash': path_hash,
            b'chunks': chunks,
            b'st': st,
        }
        item.update(self.stat_attrs(st, path))
        self.add_item_queued(item)

        return status

    @staticmethod
    def list_archives(repository, key, manifest, cache=None):
        for name, info in manifest.archives.items():
            yield Archive(repository, key, manifest, name, cache=cache)

    @staticmethod
    def _open_rb(path, st):
        flags_normal = os.O_RDONLY | getattr(os, 'O_BINARY', 0)
        flags_noatime = flags_normal | getattr(os, 'NO_ATIME', 0)
        euid = None

        def open_simple(p, s):
            return os.open(p, flags_normal)

        def open_noatime(p, s):
            return os.open(p, flags_noatime)

        def open_noatime_if_owner(p, s):
            if euid == 0 or s.st_uid == euid:
                # we are root or owner of file
                return open_noatime(p, s)
            else:
                return open_simple(p, s)

        def open_noatime_with_fallback(p, s):
            try:
                fd = os.open(p, flags_noatime)
            except PermissionError:
                # Was this EPERM due to the O_NOATIME flag?
                fd = os.open(p, flags_normal)
                # Yes, it was -- otherwise the above line would have thrown
                # another exception.
                nonlocal euid
                euid = os.geteuid()
                # So in future, let's check whether the file is owned by us
                # before attempting to use O_NOATIME.
                Archive._open_rb = open_noatime_if_owner
            return fd

        if flags_noatime != flags_normal:
            # Always use O_NOATIME version.
            Archive._open_rb = open_noatime_with_fallback
        else:
            # Always use non-O_NOATIME version.
            Archive._open_rb = open_simple
        return Archive._open_rb(path, st)


class RobustUnpacker():
    """A restartable/robust version of the streaming msgpack unpacker
    """
    item_keys = [msgpack.packb(name) for name in ('path', 'mode', 'source', 'chunks', 'rdev', 'xattrs', 'user', 'group', 'uid', 'gid', 'mtime')]

    def __init__(self, validator):
        super(RobustUnpacker, self).__init__()
        self.validator = validator
        self._buffered_data = []
        self._resync = False
        self._unpacker = msgpack.Unpacker(object_hook=StableDict)

    def resync(self):
        self._buffered_data = []
        self._resync = True

    def feed(self, data):
        if self._resync:
            self._buffered_data.append(data)
        else:
            self._unpacker.feed(data)

    def __iter__(self):
        return self

    def __next__(self):
        if self._resync:
            data = b''.join(self._buffered_data)
            while self._resync:
                if not data:
                    raise StopIteration
                # Abort early if the data does not look like a serialized dict
                if len(data) < 2 or ((data[0] & 0xf0) != 0x80) or ((data[1] & 0xe0) != 0xa0):
                    data = data[1:]
                    continue
                # Make sure it looks like an item dict
                for pattern in self.item_keys:
                    if data[1:].startswith(pattern):
                        break
                else:
                    data = data[1:]
                    continue

                self._unpacker = msgpack.Unpacker(object_hook=StableDict)
                self._unpacker.feed(data)
                try:
                    item = next(self._unpacker)
                    if self.validator(item):
                        self._resync = False
                        return item
                # Ignore exceptions that might be raised when feeding
                # msgpack with invalid data
                except (TypeError, ValueError, StopIteration):
                    pass
                data = data[1:]
        else:
            return next(self._unpacker)


class ArchiveChecker:

    def __init__(self):
        self.error_found = False
        self.possibly_superseded = set()
        self.tmpdir = tempfile.mkdtemp()

    def __del__(self):
        shutil.rmtree(self.tmpdir)

    def check(self, repository, repair=False, last=None):
        self.report_progress('Starting archive consistency check...')
        self.repair = repair
        self.repository = repository
        self.init_chunks()
        self.key = self.identify_key(repository)
        if Manifest.MANIFEST_ID not in self.chunks:
            self.manifest = self.rebuild_manifest()
        else:
            self.manifest, _ = Manifest.load(repository, key=self.key)
        self.rebuild_refcounts(last=last)
        if last is None:
            self.verify_chunks()
        else:
            self.report_progress('Orphaned objects check skipped (needs all archives checked)')
        if not self.error_found:
            self.report_progress('Archive consistency check complete, no problems found.')
        return self.repair or not self.error_found

    def init_chunks(self):
        """Fetch a list of all object keys from repository
        """
        # Explicity set the initial hash table capacity to avoid performance issues
        # due to hash table "resonance"
        capacity = int(len(self.repository) * 1.2)
        self.chunks = ChunkIndex(capacity)
        marker = None
        while True:
            result = self.repository.list(limit=10000, marker=marker)
            if not result:
                break
            marker = result[-1]
            for id_ in result:
                self.chunks[id_] = (0, 0, 0)

    def report_progress(self, msg, error=False):
        if error:
            self.error_found = True
        print(msg, file=sys.stderr if error else sys.stdout)

    def identify_key(self, repository):
        cdata = repository.get(next(self.chunks.iteritems())[0])
        return key_factory(repository, cdata)

    def rebuild_manifest(self):
        """Rebuild the manifest object if it is missing

        Iterates through all objects in the repository looking for archive metadata blocks.
        """
        self.report_progress('Rebuilding missing manifest, this might take some time...', error=True)
        manifest = Manifest(self.key, self.repository)
        for chunk_id, _ in self.chunks.iteritems():
            cdata = self.repository.get(chunk_id)
            data = self.key.decrypt(chunk_id, cdata)
            # Some basic sanity checks of the payload before feeding it into msgpack
            if len(data) < 2 or ((data[0] & 0xf0) != 0x80) or ((data[1] & 0xe0) != 0xa0):
                continue
            if b'cmdline' not in data or b'\xa7version\x01' not in data:
                continue
            try:
                archive = msgpack.unpackb(data)
            # Ignore exceptions that might be raised when feeding
            # msgpack with invalid data
            except (TypeError, ValueError, StopIteration):
                continue
            if isinstance(archive, dict) and b'items' in archive and b'cmdline' in archive:
                self.report_progress('Found archive ' + archive[b'name'].decode('utf-8'), error=True)
                manifest.archives[archive[b'name'].decode('utf-8')] = {b'id': chunk_id, b'time': archive[b'time']}
        self.report_progress('Manifest rebuild complete', error=True)
        return manifest

    def rebuild_refcounts(self, last=None):
        """Rebuild object reference counts by walking the metadata

        Missing and/or incorrect data is repaired when detected
        """
        # Exclude the manifest from chunks
        del self.chunks[Manifest.MANIFEST_ID]

        def mark_as_possibly_superseded(id_):
            if self.chunks.get(id_, (0,))[0] == 0:
                self.possibly_superseded.add(id_)

        def add_callback(chunk):
            id_ = self.key.id_hash(chunk)
            cdata = self.key.encrypt(chunk)
            add_reference(id_, len(chunk), len(cdata), cdata)
            return id_

        def add_reference(id_, size, csize, cdata=None):
            try:
                count, _, _ = self.chunks[id_]
                self.chunks[id_] = count + 1, size, csize
            except KeyError:
                assert cdata is not None
                self.chunks[id_] = 1, size, csize
                if self.repair:
                    self.repository.put(id_, cdata)

        def verify_file_chunks(item):
            """Verifies that all file chunks are present

            Missing file chunks will be replaced with new chunks of the same
            length containing all zeros.
            """
            offset = 0
            chunk_list = []
            for chunk_id, size, csize in item[b'chunks']:
                if chunk_id not in self.chunks:
                    # If a file chunk is missing, create an all empty replacement chunk
                    self.report_progress('{}: Missing file chunk detected (Byte {}-{})'.format(item[b'path'].decode('utf-8', 'surrogateescape'), offset, offset + size), error=True)
                    data = bytes(size)
                    chunk_id = self.key.id_hash(data)
                    cdata = self.key.encrypt(data)
                    csize = len(cdata)
                    add_reference(chunk_id, size, csize, cdata)
                else:
                    add_reference(chunk_id, size, csize)
                chunk_list.append((chunk_id, size, csize))
                offset += size
            item[b'chunks'] = chunk_list

        def robust_iterator(archive):
            """Iterates through all archive items

            Missing item chunks will be skipped and the msgpack stream will be restarted
            """
            unpacker = RobustUnpacker(lambda item: isinstance(item, dict) and b'path' in item)
            _state = 0

            def missing_chunk_detector(chunk_id):
                nonlocal _state
                if _state % 2 != int(chunk_id not in self.chunks):
                    _state += 1
                return _state

            for state, items in groupby(archive[b'items'], missing_chunk_detector):
                items = list(items)
                if state % 2:
                    self.report_progress('Archive metadata damage detected', error=True)
                    continue
                if state > 0:
                    unpacker.resync()
                for chunk_id, cdata in zip(items, repository.get_many(items)):
                    unpacker.feed(self.key.decrypt(chunk_id, cdata))
                    for item in unpacker:
                        yield item

        repository = cache_if_remote(self.repository)
        num_archives = len(self.manifest.archives)
        archive_items = sorted(self.manifest.archives.items(), reverse=True,
                               key=lambda name_info: name_info[1][b'time'])
        end = None if last is None else min(num_archives, last)
        for i, (name, info) in enumerate(archive_items[:end]):
            self.report_progress('Analyzing archive {} ({}/{})'.format(name, num_archives - i, num_archives))
            archive_id = info[b'id']
            if archive_id not in self.chunks:
                self.report_progress('Archive metadata block is missing', error=True)
                del self.manifest.archives[name]
                continue
            mark_as_possibly_superseded(archive_id)
            cdata = self.repository.get(archive_id)
            data = self.key.decrypt(archive_id, cdata)
            archive = StableDict(msgpack.unpackb(data))
            if archive[b'version'] != 1:
                raise Exception('Unknown archive metadata version')
            decode_dict(archive, (b'name', b'hostname', b'username', b'time'))  # fixme: argv
            items_buffer = ChunkBuffer(self.key)
            items_buffer.write_chunk = add_callback
            for item in robust_iterator(archive):
                if b'chunks' in item:
                    verify_file_chunks(item)
                items_buffer.add(item)
            items_buffer.flush(flush=True)
            for previous_item_id in archive[b'items']:
                mark_as_possibly_superseded(previous_item_id)
            archive[b'items'] = items_buffer.chunks
            data = msgpack.packb(archive, unicode_errors='surrogateescape')
            new_archive_id = self.key.id_hash(data)
            cdata = self.key.encrypt(data)
            add_reference(new_archive_id, len(data), len(cdata), cdata)
            info[b'id'] = new_archive_id

    def verify_chunks(self):
        unused = set()
        for id_, (count, size, csize) in self.chunks.iteritems():
            if count == 0:
                unused.add(id_)
        orphaned = unused - self.possibly_superseded
        if orphaned:
            self.report_progress('{} orphaned objects found'.format(len(orphaned)), error=True)
        if self.repair:
            for id_ in unused:
                self.repository.delete(id_)
            self.manifest.write()
            self.repository.commit()