fs.py

'''
fs.py - this file is part of S3QL.

Copyright © 2008 Nikolaus Rath <Nikolaus@rath.org>

This work can be distributed under the terms of the GNU GPLv3.
'''

from .logging import logging
from . import deltadump, CTRL_NAME, CTRL_INODE
from .backends.common import NoSuchObject, CorruptedObjectError
from .common import get_path, parse_literal, time_ns
from .database import NoSuchRowError
from io import BytesIO
import collections
import errno
import pyfuse3
from pyfuse3 import FUSEError
import math
import os
import stat
import struct
import time
import trio

# We work in bytes
CTRL_NAME = CTRL_NAME.encode('us-ascii')

# standard logger for this module
log = logging.getLogger(__name__)

# During long-running operations, yield to event loop with at least
# this interval.
CHECKPOINT_INTERVAL = 0.05

# ACL_ERRNO is the error code returned for requests that try
# to modify or access extendeda attributes associated with ACL.
# Since we currently don't know how to keep these in sync
# with permission bits, we cannot support ACLs despite having
# full support for extended attributes.
#
# What errno to use for ACL_ERRNO is a bit tricky. acl_set_fd(3)
# returns ENOTSUP if the file system does not support ACLs. However,
# this function uses the setxattr() syscall internally, and
# setxattr(3) states that ENOTSUP means that the file system does not
# support extended attributes at all. A test with btrfs mounted with
# -o noacl shows that the actual errno returned by setxattr() is
# EOPNOTSUPP. Also, some Python versions do not know about
# errno.ENOTSUPP and errno(3) says that on Linux, EOPNOTSUPP and ENOTSUP
# have the same value (despite this violating POSIX).
#
# All in all, the situation seems complicated, so we try to use
# EOPNOTSUPP with a fallback on ENOTSUP just in case.

if not hasattr(errno, 'EOPNOTSUPP'):
    ACL_ERRNO = errno.ENOTSUP
else:
    ACL_ERRNO = errno.EOPNOTSUPP


class Operations(pyfuse3.Operations):
    """A full-featured file system for online data storage

    This class implements low-level FUSE operations and is meant to be passed to
    pyfuse3.init().

    The ``access`` method of this class always gives full access, independent of
    file permissions. If the FUSE library is initialized with ``allow_other`` or
    ``allow_root``, the ``default_permissions`` option should therefore always
    be passed as well.


    Attributes:
    -----------

    :cache:       Holds information about cached blocks
    :inode_cache: A cache for the attributes of the currently opened inodes.
    :open_inodes: dict of currently opened inodes. This is used to not remove
                  the blocks of unlinked inodes that are still open.
    :upload_event: If set, triggers a metadata upload
    :failsafe: Set when backend problems are encountered. In that case, fs only
               allows read access.
    :broken_blocks: Caches information about corrupted blocks to avoid repeated (pointless)
                    attempts to retrieve them. This attribute is a dict (indexed by inodes)
                    of sets of block indices. Broken blocks are removed from the cache
                    when an inode is forgotten.

    Directory Entry Types
    ----------------------

    S3QL is quite agnostic when it comes to directory entry types. Every
    directory entry can contain other entries *and* have a associated data,
    size, link target and device number. However, S3QL makes some provisions for
    users relying on unlink()/rmdir() to fail for a directory/file. For that, it
    explicitly checks the st_mode attribute.
    """

    supports_dot_lookup = True
    enable_acl = False
    enable_writeback_cache = True

    def __init__(self, block_cache, db, max_obj_size, inode_cache,
                 upload_event=None):
        super().__init__()

        self.inodes = inode_cache
        self.db = db
        self.upload_event = upload_event
        self.open_inodes = collections.defaultdict(lambda: 0)
        self.max_obj_size = max_obj_size
        self.cache = block_cache
        self.failsafe = False
        self.broken_blocks = collections.defaultdict(set)

        # Root inode is always open
        self.open_inodes[pyfuse3.ROOT_INODE] += 1

    async def destroy(self):
        await self.forget(list(self.open_inodes.items()))
        self.inodes.destroy()

    async def lookup(self, id_p, name, ctx):
        return self._lookup(id_p, name, ctx).entry_attributes()

    def _lookup(self, id_p, name, ctx):
        log.debug('started with %d, %r', id_p, name)

        if name == CTRL_NAME:
            inode = self.inodes[CTRL_INODE]

            # Make sure the control file is only writable by the user
            # who mounted the file system (but don't mark inode as dirty)
            object.__setattr__(inode, 'uid', os.getuid())
            object.__setattr__(inode, 'gid', os.getgid())

        elif name == '.':
            inode = self.inodes[id_p]

        elif name == '..':
            id_ = self.db.get_val("SELECT parent_inode FROM contents WHERE inode=?",
                                  (id_p,))
            inode = self.inodes[id_]

        else:
            try:
                id_ = self.db.get_val("SELECT inode FROM contents_v WHERE name=? AND parent_inode=?",
                                      (name, id_p))
            except NoSuchRowError:
                raise FUSEError(errno.ENOENT)
            inode = self.inodes[id_]

        self.open_inodes[inode.id] += 1
        return inode

    async def getattr(self, id_, ctx):
        log.debug('started with %d', id_)
        if id_ == CTRL_INODE:
            # Make sure the control file is only writable by the user
            # who mounted the file system (but don't mark inode as dirty)
            inode = self.inodes[CTRL_INODE]
            object.__setattr__(inode, 'uid', os.getuid())
            object.__setattr__(inode, 'gid', os.getgid())
            return inode.entry_attributes()

        return self.inodes[id_].entry_attributes()

    async def readlink(self, id_, ctx):
        log.debug('started with %d', id_)
        now_ns = time_ns()
        inode = self.inodes[id_]
        if inode.atime_ns < inode.ctime_ns or inode.atime_ns < inode.mtime_ns:
            inode.atime_ns = now_ns
        try:
            return self.db.get_val("SELECT target FROM symlink_targets WHERE inode=?", (id_,))
        except NoSuchRowError:
            log.warning('Inode does not have symlink target: %d', id_)
            raise FUSEError(errno.EINVAL)

    async def opendir(self, id_, ctx):
        log.debug('started with %d', id_)
        return id_

    async def readdir(self, id_, off, token):
        log.debug('started with %d, %d', id_, off)
        if off == 0:
            off = -1

        inode = self.inodes[id_]
        if inode.atime_ns < inode.ctime_ns or inode.atime_ns < inode.mtime_ns:
            inode.atime_ns = time_ns()

        # NFS treats offsets 1 and 2 special, so we have to exclude
        # them.
        with self.db.query("SELECT name_id, name, inode FROM contents_v "
                           'WHERE parent_inode=? AND name_id > ? ORDER BY name_id',
                           (id_, off-3)) as res:
            for (next_, name, cid_) in res:
                if not pyfuse3.readdir_reply(
                    token, name, self.inodes[cid_].entry_attributes(), next_+3):
                    break
                self.open_inodes[cid_] += 1

    async def getxattr(self, id_, name, ctx):
        log.debug('started with %d, %r', id_, name)
        # Handle S3QL commands
        if id_ == CTRL_INODE:
            if name == b's3ql_pid?':
                return ('%d' % os.getpid()).encode()

            elif name == b's3qlstat':
                return self.extstat()

            raise FUSEError(errno.EINVAL)

        # http://code.google.com/p/s3ql/issues/detail?id=385
        elif name in (b'system.posix_acl_access',
                      b'system.posix_acl_default'):
            raise FUSEError(ACL_ERRNO)

        else:
            try:
                value = self.db.get_val('SELECT value FROM ext_attributes_v WHERE inode=? AND name=?',
                                          (id_, name))
            except NoSuchRowError:
                raise FUSEError(pyfuse3.ENOATTR)
            return value

    async def listxattr(self, id_, ctx):
        log.debug('started with %d', id_)
        names = list()
        for (name,) in self.db.query('SELECT name FROM ext_attributes_v WHERE inode=?', (id_,)):
            names.append(name)
        return names

    async def setxattr(self, id_, name, value, ctx):
        log.debug('started with %d, %r, %r', id_, name, value)

        # Handle S3QL commands
        if id_ == CTRL_INODE:
            if name == b's3ql_flushcache!':
                self.inodes.flush()
                await self.cache.flush()

            elif name == b's3ql_dropcache!':
                self.inodes.drop()
                await self.cache.drop()

            elif name == b'copy':
                try:
                    tup = parse_literal(value, (int, int))
                except ValueError:
                    log.warning('Received malformed command via control inode')
                    raise FUSEError.EINVAL()
                await self.copy_tree(*tup)

            elif name == b'upload-meta':
                if self.upload_event is not None:
                    self.inodes.flush()
                    self.upload_event.set()
                else:
                    raise FUSEError(errno.ENOTTY)

            elif name == b'lock':
                try:
                    id_ = parse_literal(value, int)
                except ValueError:
                    log.warning('Received malformed command via control inode')
                    raise FUSEError.EINVAL()
                await self.lock_tree(id_)

            elif name == b'rmtree':
                try:
                    tup = parse_literal(value, (int, bytes))
                except ValueError:
                    log.warning('Received malformed command via control inode')
                    raise FUSEError.EINVAL()
                await self.remove_tree(*tup)

            elif name == b'logging':
                try:
                    (lvl, modules)= parse_literal(value, (int, str))
                except (ValueError, KeyError):
                    log.warning('Received malformed command via control inode')
                    raise FUSEError.EINVAL()
                update_logging(lvl, modules.split(',') if modules else None)

            elif name == b'cachesize':
                try:
                    self.cache.cache.max_size = parse_literal(value, int)
                except ValueError:
                    log.warning('Received malformed command via control inode')
                    raise FUSEError.EINVAL()
                log.debug('updated cache size to %d bytes', self.cache.cache.max_size)

            else:
                log.warning('Received unknown command via control inode')
                raise FUSEError(errno.EINVAL)

        # http://code.google.com/p/s3ql/issues/detail?id=385
        elif name in (b'system.posix_acl_access',
                      b'system.posix_acl_default'):
            raise FUSEError(ACL_ERRNO)

        else:
            if self.failsafe or self.inodes[id_].locked:
                raise FUSEError(errno.EPERM)

            if len(value) > deltadump.MAX_BLOB_SIZE:
                raise FUSEError(errno.EINVAL)

            self.db.execute('INSERT OR REPLACE INTO ext_attributes (inode, name_id, value) '
                            'VALUES(?, ?, ?)', (id_, self._add_name(name), value))
            self.inodes[id_].ctime_ns = time_ns()

    async def removexattr(self, id_, name, ctx):
        log.debug('started with %d, %r', id_, name)

        if self.failsafe or self.inodes[id_].locked:
            raise FUSEError(errno.EPERM)

        try:
            name_id = self._del_name(name)
        except NoSuchRowError:
            raise FUSEError(pyfuse3.ENOATTR)

        changes = self.db.execute('DELETE FROM ext_attributes WHERE inode=? AND name_id=?',
                                  (id_, name_id))
        if changes == 0:
            raise FUSEError(pyfuse3.ENOATTR)

        self.inodes[id_].ctime_ns = time_ns()

    async def lock_tree(self, id0):
        '''Lock directory tree'''

        if self.failsafe:
            raise FUSEError(errno.EPERM)

        log.debug('started with %d', id0)
        queue = [ (id0, -1) ]
        self.inodes[id0].locked = True
        while queue:
            (id_p, off) = queue.pop()
            log.debug('Processing directory (%d, %d)', id_p, off)
            processed = 0
            with self.db.query('SELECT name_id, inode FROM contents WHERE parent_inode=? '
                               'AND name_id > ? ORDER BY name_id', (id_p, off)) as res:
                for (name_id, id_) in res:
                    self.inodes[id_].locked = True

                    if self.db.has_val('SELECT 1 FROM contents WHERE parent_inode=?', (id_,)):
                        queue.append((id_, -1))

                    # Break every once in a while - note that we can't yield
                    # right here because there's an active DB query.
                    processed += 1
                    if processed > 200:
                        queue.append((id_p, name_id))
                        break

            await trio.hazmat.checkpoint()

        log.debug('finished')

    async def remove_tree(self, id_p0, name0):
        '''Remove directory tree'''

        if self.failsafe:
            raise FUSEError(errno.EPERM)

        log.debug('started with %d, %s', id_p0, name0)

        if self.inodes[id_p0].locked:
            raise FUSEError(errno.EPERM)

        id0 = self._lookup(id_p0, name0, ctx=None).id
        queue = [ id0 ]  # Directories that we still need to delete
        batch_size = 200 # Entries to process before checkpointing
        stamp = time.time() # Time of last checkpoint
        while queue: # For every directory
            id_p = queue.pop()
            is_open = id_p in self.open_inodes

            # Per https://sqlite.org/isolation.html, results of removing rows
            # during select are undefined. Therefore, process data in chunks.
            # This is also a nice opportunity to release the GIL...
            query_chunk = self.db.get_list(
                'SELECT name, name_id, inode FROM contents_v WHERE '
                'parent_inode=? LIMIT %d' % batch_size, (id_p,))
            reinserted = False
            for (name, name_id, id_) in query_chunk:
                if self.db.has_val('SELECT 1 FROM contents WHERE parent_inode=?', (id_,)):
                    # First delete subdirectories
                    if not reinserted:
                        queue.append(id_p)
                        reinserted = True
                    queue.append(id_)
                else:
                    if is_open:
                        pyfuse3.invalidate_entry_async(id_p, name)
                    await self._remove(id_p, name, id_, force=True)

            if query_chunk and not reinserted:
                # Make sure to re-insert the directory to process the remaining
                # contents and delete the directory itself.
                queue.append(id_p)

            dt = time.time() - stamp
            batch_size = int(batch_size * CHECKPOINT_INTERVAL / dt)
            batch_size = min(batch_size, 200) # somewhat arbitrary...
            batch_size = max(batch_size, 20000)
            log.debug('Adjusting batch_size to %d and yielding', batch_size)
            await trio.hazmat.checkpoint()
            log.debug('re-acquired lock')
            stamp = time.time()

        if id_p0 in self.open_inodes:
            log.debug('invalidate_entry(%d, %r)', id_p0, name0)
            pyfuse3.invalidate_entry_async(id_p0, name0)
        await self._remove(id_p0, name0, id0, force=True)

        await self.forget([(id0, 1)])
        log.debug('finished')


    async def copy_tree(self, src_id, target_id):
        '''Efficiently copy directory tree'''

        if self.failsafe:
            raise FUSEError(errno.EPERM)

        log.debug('started with %d, %d', src_id, target_id)

        # To avoid lookups and make code tidier
        make_inode = self.inodes.create_inode
        db = self.db

        # Copy target attributes
        # These come from setxattr, so they may have been deleted
        # without being in open_inodes
        try:
            src_inode = self.inodes[src_id]
            target_inode = self.inodes[target_id]
        except KeyError:
            raise FUSEError(errno.ENOENT)
        for attr in ('atime_ns', 'ctime_ns', 'mtime_ns', 'mode', 'uid', 'gid'):
            setattr(target_inode, attr, getattr(src_inode, attr))

        # We first replicate into a dummy inode, so that we
        # need to invalidate only once.
        now_ns = time_ns()
        tmp = make_inode(mtime_ns=now_ns, ctime_ns=now_ns, atime_ns=now_ns,
                         uid=0, gid=0, mode=0, refcount=0)

        queue = [ (src_id, tmp.id, -1) ]
        id_cache = dict()
        while queue:
            (src_id, target_id, off) = queue.pop()
            log.debug('Processing directory (%d, %d, %d)', src_id, target_id, off)
            processed = 0
            with db.query('SELECT name_id, inode FROM contents WHERE parent_inode=? '
                          'AND name_id > ? ORDER BY name_id', (src_id, off)) as res:
                for (name_id, id_) in res:

                    # Make sure that all blocks are in the database
                    if id_ in self.open_inodes:
                        await self.cache.start_flush(id_)

                    if id_ not in id_cache:
                        inode = self.inodes[id_]
                        inode_new = make_inode(refcount=1, mode=inode.mode, size=inode.size,
                                               uid=inode.uid, gid=inode.gid,
                                               mtime_ns=inode.mtime_ns, atime_ns=inode.atime_ns,
                                               ctime_ns=inode.ctime_ns, rdev=inode.rdev)

                        id_new = inode_new.id

                        if inode.refcount != 1:
                            id_cache[id_] = id_new

                        db.execute('INSERT INTO symlink_targets (inode, target) '
                                   'SELECT ?, target FROM symlink_targets WHERE inode=?',
                                   (id_new, id_))

                        db.execute('INSERT INTO ext_attributes (inode, name_id, value) '
                                   'SELECT ?, name_id, value FROM ext_attributes WHERE inode=?',
                                   (id_new, id_))
                        db.execute('UPDATE names SET refcount = refcount + 1 WHERE '
                                   'id IN (SELECT name_id FROM ext_attributes WHERE inode=?)',
                                   (id_,))

                        processed += db.execute('INSERT INTO inode_blocks (inode, blockno, block_id) '
                                                'SELECT ?, blockno, block_id FROM inode_blocks '
                                                'WHERE inode=?', (id_new, id_))
                        db.execute('REPLACE INTO blocks (id, hash, refcount, size, obj_id) '
                                   'SELECT id, hash, refcount+COUNT(id), size, obj_id '
                                   'FROM inode_blocks JOIN blocks ON block_id = id '
                                   'WHERE inode = ? GROUP BY id', (id_new,))

                        if db.has_val('SELECT 1 FROM contents WHERE parent_inode=?', (id_,)):
                            queue.append((id_, id_new, -1))
                    else:
                        id_new = id_cache[id_]
                        self.inodes[id_new].refcount += 1

                    db.execute('INSERT INTO contents (name_id, inode, parent_inode) VALUES(?, ?, ?)',
                               (name_id, id_new, target_id))
                    db.execute('UPDATE names SET refcount=refcount+1 WHERE id=?', (name_id,))

                    # Break every once in a while - note that we can't yield
                    # right here because there's an active DB query.
                    processed += 1
                    if processed > 200:
                        queue.append((src_id, target_id, name_id))
                        break

            await trio.hazmat.checkpoint()

        # Make replication visible
        self.db.execute('UPDATE contents SET parent_inode=? WHERE parent_inode=?',
                        (target_inode.id, tmp.id))
        del self.inodes[tmp.id]
        pyfuse3.invalidate_inode(target_inode.id)

        log.debug('finished')

    async def unlink(self, id_p, name, ctx):
        log.debug('started with %d, %r', id_p, name)
        if self.failsafe:
            raise FUSEError(errno.EPERM)

        inode = self._lookup(id_p, name, ctx)

        if stat.S_ISDIR(inode.mode):
            raise FUSEError(errno.EISDIR)

        await self._remove(id_p, name, inode.id)

        await self.forget([(inode.id, 1)])

    async def rmdir(self, id_p, name, ctx):
        log.debug('started with %d, %r', id_p, name)
        if self.failsafe:
            raise FUSEError(errno.EPERM)

        inode = self._lookup(id_p, name, ctx)

        if self.inodes[id_p].locked:
            raise FUSEError(errno.EPERM)

        if not stat.S_ISDIR(inode.mode):
            raise FUSEError(errno.ENOTDIR)

        await self._remove(id_p, name, inode.id)

        await self.forget([(inode.id, 1)])

    async def _remove(self, id_p, name, id_, force=False):
        '''Remove entry `name` with parent inode `id_p`

        `id_` must be the inode of `name`. If `force` is True, then
        the `locked` attribute is ignored.
        '''

        log.debug('started with %d, %r', id_p, name)

        now_ns = time_ns()

        # Check that there are no child entries
        if self.db.has_val("SELECT 1 FROM contents WHERE parent_inode=?", (id_,)):
            log.debug("Attempted to remove entry with children: %s",
                      get_path(id_p, self.db, name))
            raise FUSEError(errno.ENOTEMPTY)

        if self.inodes[id_p].locked and not force:
            raise FUSEError(errno.EPERM)

        name_id = self._del_name(name)
        self.db.execute("DELETE FROM contents WHERE name_id=? AND parent_inode=?",
                        (name_id, id_p))

        inode = self.inodes[id_]
        inode.refcount -= 1
        inode.ctime_ns = now_ns

        inode_p = self.inodes[id_p]
        inode_p.mtime_ns = now_ns
        inode_p.ctime_ns = now_ns

        if inode.refcount == 0 and id_ not in self.open_inodes:
            log.debug('removing from cache')
            await self.cache.remove(id_, 0, int(math.ceil(inode.size / self.max_obj_size)))
            # Since the inode is not open, it's not possible that new blocks
            # get created at this point and we can safely delete the inode
            self.db.execute('UPDATE names SET refcount = refcount - 1 WHERE '
                            'id IN (SELECT name_id FROM ext_attributes WHERE inode=?)',
                            (id_,))
            self.db.execute('DELETE FROM names WHERE refcount=0 AND '
                            'id IN (SELECT name_id FROM ext_attributes WHERE inode=?)',
                            (id_,))
            self.db.execute('DELETE FROM ext_attributes WHERE inode=?', (id_,))
            self.db.execute('DELETE FROM symlink_targets WHERE inode=?', (id_,))
            del self.inodes[id_]

        log.debug('finished')

    async def symlink(self, id_p, name, target, ctx):
        log.debug('started with %d, %r, %r', id_p, name, target)

        if self.failsafe:
            raise FUSEError(errno.EPERM)

        mode = (stat.S_IFLNK | stat.S_IRUSR | stat.S_IWUSR | stat.S_IXUSR |
                stat.S_IRGRP | stat.S_IWGRP | stat.S_IXGRP |
                stat.S_IROTH | stat.S_IWOTH | stat.S_IXOTH)

        # Unix semantics require the size of a symlink to be the length
        # of its target. Therefore, we create symlink directory entries
        # with this size. If the kernel ever learns to open and read
        # symlinks directly, it will read the corresponding number of \0
        # bytes.
        inode = self._create(id_p, name, mode, ctx, size=len(target))
        self.db.execute('INSERT INTO symlink_targets (inode, target) VALUES(?,?)',
                        (inode.id, target))
        self.open_inodes[inode.id] += 1
        return inode.entry_attributes()

    async def rename(self, id_p_old, name_old, id_p_new, name_new, flags, ctx):
        if flags:
            raise FUSEError(errno.ENOTSUP)

        log.debug('started with %d, %r, %d, %r', id_p_old, name_old, id_p_new, name_new)
        if name_new == CTRL_NAME or name_old == CTRL_NAME:
            log.warning('Attempted to rename s3ql control file (%s -> %s)',
                      get_path(id_p_old, self.db, name_old),
                      get_path(id_p_new, self.db, name_new))
            raise FUSEError(errno.EACCES)


        if (self.failsafe or self.inodes[id_p_old].locked
            or self.inodes[id_p_new].locked):
            raise FUSEError(errno.EPERM)

        inode_old = self._lookup(id_p_old, name_old, ctx)

        try:
            inode_new = self._lookup(id_p_new, name_new, ctx)
        except FUSEError as exc:
            if exc.errno != errno.ENOENT:
                raise
            else:
                target_exists = False
        else:
            target_exists = True

        if target_exists:
            self._replace(id_p_old, name_old, id_p_new, name_new,
                          inode_old.id, inode_new.id)
            await self.forget([(inode_old.id, 1), (inode_new.id, 1)])
        else:
            self._rename(id_p_old, name_old, id_p_new, name_new)
            await self.forget([(inode_old.id, 1)])

    def _add_name(self, name):
        '''Get id for *name* and increase refcount

        Name is inserted in table if it does not yet exist.
        '''

        try:
            name_id = self.db.get_val('SELECT id FROM names WHERE name=?', (name,))
        except NoSuchRowError:
            name_id = self.db.rowid('INSERT INTO names (name, refcount) VALUES(?,?)',
                                    (name, 1))
        else:
            self.db.execute('UPDATE names SET refcount=refcount+1 WHERE id=?', (name_id,))
        return name_id

    def _del_name(self, name):
        '''Decrease refcount for *name*

        Name is removed from table if refcount drops to zero. Returns the
        (possibly former) id of the name.
        '''

        (name_id, refcount) = self.db.get_row('SELECT id, refcount FROM names WHERE name=?', (name,))

        if refcount > 1:
            self.db.execute('UPDATE names SET refcount=refcount-1 WHERE id=?', (name_id,))
        else:
            self.db.execute('DELETE FROM names WHERE id=?', (name_id,))

        return name_id

    def _rename(self, id_p_old, name_old, id_p_new, name_new):
        now_ns = time_ns()

        name_id_new = self._add_name(name_new)
        name_id_old = self._del_name(name_old)

        self.db.execute("UPDATE contents SET name_id=?, parent_inode=? WHERE name_id=? "
                        "AND parent_inode=?", (name_id_new, id_p_new,
                                               name_id_old, id_p_old))

        inode_p_old = self.inodes[id_p_old]
        inode_p_new = self.inodes[id_p_new]
        inode_p_old.mtime_ns = now_ns
        inode_p_new.mtime_ns = now_ns
        inode_p_old.ctime_ns = now_ns
        inode_p_new.ctime_ns = now_ns

    def _replace(self, id_p_old, name_old, id_p_new, name_new,
                 id_old, id_new):

        now_ns = time_ns()

        if self.db.has_val("SELECT 1 FROM contents WHERE parent_inode=?", (id_new,)):
            log.info("Attempted to overwrite entry with children: %s",
                      get_path(id_p_new, self.db, name_new))
            raise FUSEError(errno.EINVAL)

        # Replace target
        name_id_new = self.db.get_val('SELECT id FROM names WHERE name=?', (name_new,))
        self.db.execute("UPDATE contents SET inode=? WHERE name_id=? AND parent_inode=?",
                        (id_old, name_id_new, id_p_new))

        # Delete old name
        name_id_old = self._del_name(name_old)
        self.db.execute('DELETE FROM contents WHERE name_id=? AND parent_inode=?',
                        (name_id_old, id_p_old))

        inode_new = self.inodes[id_new]
        inode_new.refcount -= 1
        inode_new.ctime_ns = now_ns

        inode_p_old = self.inodes[id_p_old]
        inode_p_old.ctime_ns = now_ns
        inode_p_old.mtime_ns = now_ns

        inode_p_new = self.inodes[id_p_new]
        inode_p_new.ctime_ns = now_ns
        inode_p_new.mtime_ns = now_ns

        if inode_new.refcount == 0 and id_new not in self.open_inodes:
            self.cache.remove(id_new, 0,
                              int(math.ceil(inode_new.size / self.max_obj_size)))
            # Since the inode is not open, it's not possible that new blocks
            # get created at this point and we can safely delete the inode
            self.db.execute('UPDATE names SET refcount = refcount - 1 WHERE '
                            'id IN (SELECT name_id FROM ext_attributes WHERE inode=?)',
                            (id_new,))
            self.db.execute('DELETE FROM names WHERE refcount=0')
            self.db.execute('DELETE FROM ext_attributes WHERE inode=?', (id_new,))
            self.db.execute('DELETE FROM symlink_targets WHERE inode=?', (id_new,))
            del self.inodes[id_new]


    async def link(self, id_, new_id_p, new_name, ctx):
        log.debug('started with %d, %d, %r', id_, new_id_p, new_name)

        if new_name == CTRL_NAME or id_ == CTRL_INODE:
            log.warning('Attempted to create s3ql control file at %s',
                      get_path(new_id_p, self.db, new_name))
            raise FUSEError(errno.EACCES)

        now_ns = time_ns()
        inode_p = self.inodes[new_id_p]

        if inode_p.refcount == 0:
            log.warning('Attempted to create entry %s with unlinked parent %d',
                     new_name, new_id_p)
            raise FUSEError(errno.EINVAL)

        if self.failsafe or inode_p.locked:
            raise FUSEError(errno.EPERM)

        inode_p.ctime_ns = now_ns
        inode_p.mtime_ns = now_ns

        self.db.execute("INSERT INTO contents (name_id, inode, parent_inode) VALUES(?,?,?)",
                        (self._add_name(new_name), id_, new_id_p))
        inode = self.inodes[id_]
        inode.refcount += 1
        inode.ctime_ns = now_ns

        self.open_inodes[inode.id] += 1
        return inode.entry_attributes()

    async def setattr(self, id_, attr, fields, fh, ctx):
        """Handles FUSE setattr() requests"""
        inode = self.inodes[id_]
        if fh is not None:
            assert fh == id_
        now_ns = time_ns()

        if self.failsafe or inode.locked:
            raise FUSEError(errno.EPERM)

        if fields.update_mode:
            inode.mode = attr.st_mode

        if fields.update_uid:
            inode.uid = attr.st_uid

        if fields.update_gid:
            inode.gid = attr.st_gid

        if fields.update_atime:
            inode.atime_ns = attr.st_atime_ns

        if fields.update_mtime:
            inode.mtime_ns = attr.st_mtime_ns

        inode.ctime_ns = now_ns

        # This needs to go last, because the call to cache.remove and cache.get
        # are asynchorouns and may thus evict the *inode* object from the cache.
        if fields.update_size:
            len_ = attr.st_size

            # Determine blocks to delete
            last_block = len_ // self.max_obj_size
            cutoff = len_ % self.max_obj_size
            total_blocks = int(math.ceil(inode.size / self.max_obj_size))

            # Adjust file size
            inode.size = len_

            # Delete blocks and truncate last one if required
            if cutoff == 0:
                await self.cache.remove(id_, last_block, total_blocks)
            else:
                await self.cache.remove(id_, last_block + 1, total_blocks)
                try:
                    async with self.cache.get(id_, last_block) as fh:
                        fh.truncate(cutoff)
                except NoSuchObject as exc:
                    log.warning('Backend lost block %d of inode %d (id %s)!',
                             last_block, id_, exc.key)
                    raise

                except CorruptedObjectError as exc:
                    log.warning('Backend returned malformed data for block %d of inode %d (%s)',
                                last_block, id_, exc)
                    self.failsafe = True
                    self.broken_blocks[id_].add(last_block)
                    raise FUSEError(errno.EIO)


        return inode.entry_attributes()

    async def mknod(self, id_p, name, mode, rdev, ctx):
        log.debug('started with %d, %r', id_p, name)
        if self.failsafe:
            raise FUSEError(errno.EPERM)
        inode = self._create(id_p, name, mode, ctx, rdev=rdev)
        self.open_inodes[inode.id] += 1
        return inode.entry_attributes()

    async def mkdir(self, id_p, name, mode, ctx):
        log.debug('started with %d, %r', id_p, name)
        if self.failsafe:
            raise FUSEError(errno.EPERM)
        inode = self._create(id_p, name, mode, ctx)
        self.open_inodes[inode.id] += 1
        return inode.entry_attributes()

    def extstat(self):
        '''Return extended file system statistics'''

        log.debug('started')

        # Flush inode cache to get better estimate of total fs size
        self.inodes.flush()

        entries = self.db.get_val("SELECT COUNT(rowid) FROM contents")
        blocks = self.db.get_val("SELECT COUNT(id) FROM objects")
        inodes = self.db.get_val("SELECT COUNT(id) FROM inodes")
        fs_size = self.db.get_val('SELECT SUM(size) FROM inodes') or 0
        dedup_size = self.db.get_val('SELECT SUM(size) FROM blocks') or 0

        # Objects that are currently being uploaded/compressed have size == -1
        compr_size = self.db.get_val('SELECT SUM(size) FROM objects '
                                     'WHERE size > 0') or 0

        return struct.pack('QQQQQQQQQQQQ', entries, blocks, inodes, fs_size, dedup_size,
                           compr_size, self.db.get_size(), *self.cache.get_usage())

    async def statfs(self, ctx):
        log.debug('started')

        stat_ = pyfuse3.StatvfsData()

        # Get number of blocks & inodes
        blocks = self.db.get_val("SELECT COUNT(id) FROM objects")
        inodes = self.db.get_val("SELECT COUNT(id) FROM inodes")
        size = self.db.get_val('SELECT SUM(size) FROM blocks')

        if size is None:
            size = 0

        # file system block size, i.e. the minimum amount of space that can
        # be allocated. This doesn't make much sense for S3QL, so we just
        # return the average size of stored blocks.
        stat_.f_frsize = max(4096, size // blocks) if blocks != 0 else 4096

        # This should actually be the "preferred block size for doing IO.  However, `df` incorrectly
        # interprets f_blocks, f_bfree and f_bavail in terms of f_bsize rather than f_frsize as it
        # should (according to statvfs(3)), so the only way to return correct values *and* have df
        # print something sensible is to set f_bsize and f_frsize to the same value. (cf.
        # http://bugs.debian.org/671490)
        stat_.f_bsize = stat_.f_frsize

        # size of fs in f_frsize units. Since backend is supposed to be unlimited,
        # always return a half-full filesystem, but at least 1 TB)
        fs_size = max(2 * size, 1024 ** 4)

        stat_.f_blocks = fs_size // stat_.f_frsize
        stat_.f_bfree = (fs_size - size) // stat_.f_frsize
        stat_.f_bavail = stat_.f_bfree # free for non-root

        total_inodes = max(2 * inodes, 1000000)
        stat_.f_files = total_inodes
        stat_.f_ffree = total_inodes - inodes
        stat_.f_favail = total_inodes - inodes # free for non-root

        return stat_

    async def open(self, id_, flags, ctx):
        log.debug('started with %d', id_)
        if ((flags & os.O_RDWR or flags & os.O_WRONLY)
            and (self.failsafe or self.inodes[id_].locked)):
            raise FUSEError(errno.EPERM)

        return id_

    async def access(self, id_, mode, ctx):
        '''Check if requesting process has `mode` rights on `inode`.

        This method always returns true, since it should only be called
        when permission checking is disabled (if permission checking is
        enabled, the `default_permissions` FUSE option should be set).
        '''
        # Yeah, could be a function and has unused arguments
        #pylint: disable=R0201,W0613

        log.debug('started with %d', id_)
        return True

    async def create(self, id_p, name, mode, flags, ctx):
        log.debug('started with id_p=%d, %s', id_p, name)
        if self.failsafe:
            raise FUSEError(errno.EPERM)

        try:
            id_ = self.db.get_val("SELECT inode FROM contents_v WHERE name=? AND parent_inode=?",
                                  (name, id_p))
        except NoSuchRowError:
            inode = self._create(id_p, name, mode, ctx)
        else:
            await self.open(id_, flags, ctx)
            inode = self.inodes[id_]

        self.open_inodes[inode.id] += 1
        return (inode.id, inode.entry_attributes())

    def _create(self, id_p, name, mode, ctx, rdev=0, size=0):
        if name == CTRL_NAME:
            log.warning('Attempted to create s3ql control file at %s',
                     get_path(id_p, self.db, name))
            raise FUSEError(errno.EACCES)

        now_ns = time_ns()
        inode_p = self.inodes[id_p]

        if inode_p.locked:
            raise FUSEError(errno.EPERM)

        if inode_p.refcount == 0:
            log.warning('Attempted to create entry %s with unlinked parent %d',
                     name, id_p)
            raise FUSEError(errno.EINVAL)
        inode_p.mtime_ns = now_ns
        inode_p.ctime_ns = now_ns

        if inode_p.mode & stat.S_ISGID:
            gid = inode_p.gid
            if stat.S_ISDIR(mode):
                mode |= stat.S_ISGID
        else:
            gid = ctx.gid
        inode = self.inodes.create_inode(mtime_ns=now_ns, ctime_ns=now_ns, atime_ns=now_ns,
                                         uid=ctx.uid, gid=gid, mode=mode, refcount=1,
                                         rdev=rdev, size=size)

        self.db.execute("INSERT INTO contents(name_id, inode, parent_inode) VALUES(?,?,?)",
                        (self._add_name(name), inode.id, id_p))

        return inode


    async def read(self, fh, offset, length):
        '''Read `size` bytes from `fh` at position `off`

        Unless EOF is reached, returns exactly `size` bytes.
        '''
        #log.debug('started with %d, %d, %d', fh, offset, length)
        buf = BytesIO()
        inode = self.inodes[fh]

        # Make sure that we don't read beyond the file size. This
        # should not happen unless direct_io is activated, but it's
        # cheap and nice for testing.
        size = inode.size
        length = min(size - offset, length)

        while length > 0:
            tmp = await self._readwrite(fh, offset, length=length)
            buf.write(tmp)
            length -= len(tmp)
            offset += len(tmp)

        # Inode may have expired from cache
        inode = self.inodes[fh]

        if inode.atime_ns < inode.ctime_ns or inode.atime_ns < inode.mtime_ns:
            inode.atime_ns = time_ns()

        return buf.getvalue()


    async def write(self, fh, offset, buf):
        '''Handle FUSE write requests.'''

        #log.debug('started with %d, %d, datalen=%d', fh, offset, len(buf))

        if self.failsafe or self.inodes[fh].locked:
            raise FUSEError(errno.EPERM)

        total = len(buf)
        minsize = offset + total
        while buf:
            written = await self._readwrite(fh, offset, buf=buf)
            offset += written
            buf = buf[written:]

        # Update file size if changed
        # Fuse does not ensure that we do not get concurrent write requests,
        # so we have to be careful not to undo a size extension made by
        # a concurrent write.
        now_ns = time_ns()
        inode = self.inodes[fh]
        inode.size = max(inode.size, minsize)
        inode.mtime_ns = now_ns
        inode.ctime_ns = now_ns

        return total

    async def _readwrite(self, id_, offset, *, buf=None, length=None):
        """Read or write as much as we can.

        If *buf* is None, read and return up to *length* bytes.

        If *length* is None, write from *buf* and return the number of
        bytes written.

        This is one method to reduce code duplication.
        """

        # Calculate required block
        blockno = offset // self.max_obj_size
        offset_rel = offset - blockno * self.max_obj_size

        if id_ in self.broken_blocks and blockno in self.broken_blocks[id_]:
            raise FUSEError(errno.EIO)

        if length is None:
            write = True
            length = len(buf)
        elif buf is None:
            write = False
        else:
            raise TypeError("Don't know what to do!")

        # Don't try to write/read into the next block
        if offset_rel + length > self.max_obj_size:
            length = self.max_obj_size - offset_rel

        try:
            async with self.cache.get(id_, blockno) as fh:
                fh.seek(offset_rel)
                if write:
                    fh.write(buf[:length])
                else:
                    buf = fh.read(length)

        except NoSuchObject as exc:
            log.error('Backend lost block %d of inode %d (id %s)!',
                      blockno, id_, exc.key)
            self.failsafe = True
            self.broken_blocks[id_].add(blockno)
            raise FUSEError(errno.EIO)

        except CorruptedObjectError as exc:
            log.error('Backend returned malformed data for block %d of inode %d (%s)',
                      blockno, id_, exc)
            self.failsafe = True
            self.broken_blocks[id_].add(blockno)
            raise FUSEError(errno.EIO)

        if write:
            return length
        elif len(buf) == length:
            return buf
        else:
            # If we can't read enough, add null bytes
            return buf + b"\0" * (length - len(buf))

    async def fsync(self, fh, datasync):
        log.debug('started with %d, %s', fh, datasync)
        if not datasync:
            self.inodes.flush_id(fh)

        for blockno in range(0, self.inodes[fh].size // self.max_obj_size + 1):
            await self.cache.flush_local(fh, blockno)

    async def forget(self, forget_list):
        log.debug('started with %s', forget_list)

        for (id_, nlookup) in forget_list:
            self.open_inodes[id_] -= nlookup

            if self.open_inodes[id_] == 0:
                del self.open_inodes[id_]
                if id_ in self.broken_blocks:
                    del self.broken_blocks[id_]

                inode = self.inodes[id_]
                if inode.refcount == 0:
                    log.debug('removing %d from cache', id_)
                    await self.cache.remove(id_, 0, inode.size // self.max_obj_size + 1)
                    # Since the inode is not open, it's not possible that new blocks
                    # get created at this point and we can safely delete the inode
                    self.db.execute('UPDATE names SET refcount = refcount - 1 WHERE '
                                    'id IN (SELECT name_id FROM ext_attributes WHERE inode=?)',
                                    (id_,))
                    self.db.execute('DELETE FROM names WHERE refcount=0 AND '
                                    'id IN (SELECT name_id FROM ext_attributes WHERE inode=?)',
                                    (id_,))
                    self.db.execute('DELETE FROM ext_attributes WHERE inode=?', (id_,))
                    self.db.execute('DELETE FROM symlink_targets WHERE inode=?', (id_,))
                    del self.inodes[id_]


    async def fsyncdir(self, fh, datasync):
        log.debug('started with %d, %s', fh, datasync)
        if not datasync:
            self.inodes.flush_id(fh)

    async def releasedir(self, fh):
        log.debug('started with %d', fh)

    async def release(self, fh):
        log.debug('started with %d', fh)

    async def flush(self, fh):
        log.debug('started with %d', fh)


def update_logging(level, modules):
    root_logger = logging.getLogger()
    root_logger.setLevel(logging.INFO)
    if level == logging.DEBUG:
        logging.disable(logging.NOTSET)
        if 'all' in modules:
            root_logger.setLevel(logging.DEBUG)
        else:
            for module in modules:
                logging.getLogger(module).setLevel(logging.DEBUG)

    else:
        logging.disable(logging.DEBUG)