A S3QL file system is mounted with the mount.s3ql command. It has the following syntax:
mount.s3ql [options] <storage url> <mountpoint>
This command accepts the following options:
- --log <target>
Destination for log messages. Specify
nonefor standard output or
syslogfor the system logging daemon. Anything else will be interpreted as a file name. Log files will be rotated when they reach 1 MiB, and at most 5 old log files will be kept. Default:
- --cachedir <path>
Store cached data in this directory (default:
- --debug-modules <modules>
Activate debugging output from specified modules (use commas to separate multiple modules). Debug messages will be written to the target specified by the
Activate debugging output from all S3QL modules. Debug messages will be written to the target specified by the
be really quiet
- --backend-options <options>
Backend specific options (separate by commas). See backend documentation for available options.
just print program version and exit
- --authfile <path>
Read authentication credentials from this file (default:
- --compress <algorithm-lvl>
Compression algorithm and compression level to use when storing new data. algorithm may be any of
zlib, or none. lvl may be any integer from 0 (fastest) to 9 (slowest). Default:
- --cachesize <size>
Cache size in KiB (default: autodetect).
- --max-cache-entries <num>
Maximum number of entries in cache (default: autodetect). Each cache entry requires one file descriptor, so if you increase this number you have to make sure that your process file descriptor limit (as set with
ulimit -n) is high enough (at least the number of cache entries + 100).
Do not purge locally cached files on exit.
Normally, only the user who called
mount.s3qlcan access the mount point. This user then also has full access to it, independent of individual file permissions. If the
--allow-otheroption is specified, other users can access the mount point as well and individual file permissions are taken into account for all users.
--allow-other, but restrict access to the mounting user and the root user.
- --dirty-block-upload-delay <seconds>
Upload delay for dirty blocks in seconds (default: 10 seconds).
Do not daemonize, stay in foreground
- --fs-name FS_NAME
Mount name passed to fuse, the name will be shown in the first column of the system mount command output. If not specified your storage url is used.
Run as systemd unit. Consider specifying –log none as well to make use of journald.
- --metadata-upload-interval <seconds>
Interval in seconds between complete metadata uploads. Set to 0 to disable. Default: 24h.
- --threads <no>
Number of parallel upload threads to use (default: auto).
Enable some optimizations for exporting the file system over NFS. (default: False)
If the file system is mounted with neither the
allow-other option, the mounting user has full
permissions on the S3QL file system (he is effectively root). If one
(or both) of the options is used, standard unix permission checks
apply, i.e. only the real root user has full access and all other
users (including the mounting user) are subject to permission checks.
S3QL supports three compression algorithms, LZMA, Bzip2 and zlib (with LZMA being the default). The compression algorithm can be specified freely whenever the file system is mounted, since it affects only the compression of new data blocks.
Roughly speaking, LZMA is slower but achieves better compression ratios than Bzip2, while Bzip2 in turn is slower but achieves better compression ratios than zlib.
For maximum file system performance, the best algorithm therefore depends on your network connection speed: the compression algorithm should be fast enough to saturate your network connection.
To find the optimal algorithm and number of parallel compression
threads for your system, S3QL ships with a program called
benchmark.py in the
contrib directory. You should run this program
on a file that has a size that is roughly equal to the block size of
your file system and has similar contents. It will then determine the
compression speeds for the different algorithms and the upload speeds
for the specified backend and recommend the best algorithm that is
fast enough to saturate your network connection.
Obviously you should make sure that there is little other system load
when you run
benchmark.py (i.e., don’t compile software or encode
videos at the same time).
Notes about Caching¶
S3QL maintains a local cache of the file system data to speed up access. The cache is block based, so it is possible that only parts of a file are in the cache.
Maximum Number of Cache Entries¶
The maximum size of the cache can be configured with the
--cachesize option. In addition to that, the maximum number
of objects in the cache is limited by the
--max-cache-entries option, so it is possible that the cache
does not grow up to the maximum cache size because the maximum number
of cache elements has been reached. The reason for this limit is that
each cache entry requires one open file descriptor, and Linux
distributions usually limit the total number of file descriptors per
process to about a thousand.
If you specify a value for
--max-cache-entries, you should
therefore make sure to also configure your system to increase the
maximum number of open file handles. This can be done temporarily with
the ulimit -n command. The method to permanently change this limit
system-wide depends on your distribution.
Cache Flushing and Expiration¶
S3QL flushes changed blocks in the cache to the backend whenever a block
has not been accessed for at least 10 seconds by default. This time can
be influenced using the
--dirty-block-upload-delay option. Note
that when a block is flushed, it still remains in the cache.
Cache expiration (i.e., removal of blocks from the cache) is only done when the maximum cache size is reached. S3QL always expires the least recently used blocks first.
S3QL filesystems can be exported over NFS. The
is recommended to improve performance when NFS is used, but no harm
will occur when it is not specified.
NFS supports persistence of client mounts across server restarts. This
means that if a client has mounted an S3QL file system over NFS, the
server may unmount and remount the S3QL filesystem (or even reboot)
without the client being affected beyond temporarily becoming
unavailable. This poses several challenges, but is supported by S3QL
as long as no
fsck.s3ql operation is run:
fsck.s3ql modifies a file system in any way, all NFS
clients must unmount and re-mount the NFS share before the
S3QL file system is re-mounted on the server.
Once an S3QL file system has been mounted, there is a multitude of
problems that can occur when communicating with the remote
server. Generally, mount.s3ql always tries to keep the file
system as accessible as possible under the circumstances. That means
that if network connectivity is lost, data can still be written as
long as there is space in the local cache. Attempts to read data not
already present in the cache, however, will block until connection is
re-established. If any sort of data corruption is detected, the file
system will switch to read-only mode. Attempting to read files that
are affected by the corruption will return an input/output error
(errno set to
In case of other unexpected or fatal problems, mount.s3ql
terminates, but does not unmount the file system. Any attempt to
access the mountpoint will result in a “Transport endpoint not
connected” error (errno set to
ESHUTDOWN). This ensures that a
mountpoint whose mount.s3ql process has terminated can not
be confused with a mountpoint containing an empty file system (which
would be fatal if e.g. the mountpoint is automatically mirrored). When
this has happened, the mountpoint can be cleared by using the
fusermount command (provided by FUSE) with the
mount.s3ql will automatically try to re-establish the connection to the server if network connectivity is lost, and retry sending a request when the connection is established but the remote server signals a temporary problem. These attempts will be made at increasing intervals for a period up to 24 hours, with retry intervals starting at 20 ms and increasing up to 5 minutes. After 24 hours, mount.s3ql will give up and terminate, leaving the mountpoint inaccessible as described above.
Generally, mount.s3ql will also emit log messages for any
unusual conditions that it encounters. The destination for these
messages can be set with the
--log parameter. It is highly
recommended to periodically check these logs, for example with a tool
like logcheck. Many potential issues that mount.s3ql may
encounter do not justify restricting access to the file system, but
should nevertheless be investigated if they occur. Checking the log
messages is the only way to find out about them.
If you want to mount and umount an S3QL file system automatically at
system startup and shutdown, you should do so with a dedicated S3QL
init job (instead of using
/etc/fstab. When using systemd,
mount.s3ql can be started with
--systemd to run
as a systemd service of type
In principle, it is also possible to automatically mount an S3QL
file system with an appropriate entry in
this is not recommended for several reasons:
file systems mounted in
/etc/fstabwill be unmounted with the umount command, so your system will not wait until all data has been uploaded but shutdown (or restart) immediately (this is a FUSE limitation, cf https://github.com/libfuse/libfuse/issues/1).
There is no way to tell the system that mounting S3QL requires a Python interpreter to be available, so it may attempt to run mount.s3ql before it has mounted the volume containing the Python interpreter.
There is no standard way to tell the system that internet connection has to be up before the S3QL file system can be mounted.