Go to file
trapexit b21172c609
Merge pull request #1321 from trapexit/allowmmap
Add support for 'direct-io-allow-mmap' if supported by kernel
2024-03-21 21:54:04 -04:00
.github Add CodeQL workflow for GitHub code scanning 2022-11-11 07:59:54 +00:00
buildtools Create ld_preload to reopen files on underlying filesystem (#1294) 2024-01-22 12:42:55 -06:00
debian fix rpm spec and chmod in deb 2020-06-30 11:07:39 -04:00
libfuse Add support for 'direct-io-allow-mmap' if supported by kernel 2024-03-21 20:48:42 -05:00
man Add support for 'direct-io-allow-mmap' if supported by kernel 2024-03-21 20:48:42 -05:00
src Add support for 'direct-io-allow-mmap' if supported by kernel 2024-03-21 20:48:42 -05:00
tests Align msg buffer memory to allow O_DIRECT to work 2023-07-09 15:32:11 -05:00
tools Special case O_TMPFILE in preload 2024-02-29 14:33:27 -06:00
.cirrus.yml Create ld_preload to reopen files on underlying filesystem (#1294) 2024-01-22 12:42:55 -06:00
.gitignore add artifacts to gitignore 2015-12-14 23:50:35 -05:00
DEPENDENCIES Remove use of pthread_getname_np to work with older musl versions 2023-10-14 01:15:29 -05:00
LICENSE Create ld_preload to reopen files on underlying filesystem (#1294) 2024-01-22 12:42:55 -06:00
Makefile Create ld_preload to reopen files on underlying filesystem (#1294) 2024-01-22 12:42:55 -06:00
README.md Add support for 'direct-io-allow-mmap' if supported by kernel 2024-03-21 20:48:42 -05:00
mergerfs.spec Create ld_preload to reopen files on underlying filesystem (#1294) 2024-01-22 12:42:55 -06:00

README.md

% mergerfs(1) mergerfs user manual

NAME

mergerfs - a featureful union filesystem

SYNOPSIS

mergerfs -o<options> <branches> <mountpoint>

DESCRIPTION

mergerfs is a union filesystem geared towards simplifying storage and management of files across numerous commodity storage devices. It is similar to mhddfs, unionfs, and aufs.

FEATURES

  • Configurable behaviors / file placement
  • Ability to add or remove filesystems at will
  • Resistance to individual filesystem failure
  • Support for extended attributes (xattrs)
  • Support for file attributes (chattr)
  • Runtime configurable (via xattrs)
  • Works with heterogeneous filesystem types
  • Moving of file when filesystem runs out of space while writing
  • Ignore read-only filesystems when creating files
  • Turn read-only files into symlinks to underlying file
  • Hard link copy-on-write / CoW
  • Support for POSIX ACLs
  • Misc other things

HOW IT WORKS

mergerfs logically merges multiple paths together. Think a union of sets. The file/s or directory/s acted on or presented through mergerfs are based on the policy chosen for that particular action. Read more about policies below.

A         +      B        =       C
/disk1           /disk2           /merged
|                |                |
+-- /dir1        +-- /dir1        +-- /dir1
|   |            |   |            |   |
|   +-- file1    |   +-- file2    |   +-- file1
|                |   +-- file3    |   +-- file2
+-- /dir2        |                |   +-- file3
|   |            +-- /dir3        |
|   +-- file4        |            +-- /dir2
|                     +-- file5   |   |
+-- file6                         |   +-- file4
                                  |
                                  +-- /dir3
                                  |   |
                                  |   +-- file5
                                  |
                                  +-- file6

mergerfs does not support the copy-on-write (CoW) or whiteout behaviors found in aufs and overlayfs. You can not mount a read-only filesystem and write to it. However, mergerfs will ignore read-only filesystems when creating new files so you can mix read-write and read-only filesystems. It also does not split data across filesystems. It is not RAID0 / striping. It is simply a union of other filesystems.

TERMINOLOGY

  • branch: A base path used in the pool.
  • pool: The mergerfs mount. The union of the branches.
  • relative path: The path in the pool relative to the branch and mount.
  • function: A filesystem call (open, unlink, create, getattr, rmdir, etc.)
  • category: A collection of functions based on basic behavior (action, create, search).
  • policy: The algorithm used to select a file when performing a function.
  • path preservation: Aspect of some policies which includes checking the path for which a file would be created.

BASIC SETUP

If you don't already know that you have a special use case then just start with one of the following option sets.

You need mmap (used by rtorrent and many sqlite3 base software)

cache.files=auto-full,dropcacheonclose=true,category.create=mfs

or if you are on a Linux kernel >= 6.6.x mergerfs will enable a mode that allows shared mmap when cache.files=off. To be sure of the best performance between cache.files=off and cache.files=auto-full you'll need to do your own benchmarking but often off is faster.

You don't need mmap

cache.files=off,dropcacheonclose=true,category.create=mfs

Command Line

mergerfs -o cache.files=auto-full,dropcacheonclose=true,category.create=mfs /mnt/hdd0:/mnt/hdd1 /media

/etc/fstab

/mnt/hdd0:/mnt/hdd1 /media mergerfs cache.files=auto-full,dropcacheonclose=true,category.create=mfs 0 0

systemd mount

https://github.com/trapexit/mergerfs/wiki/systemd

[Unit]
Description=mergerfs service

[Service]
Type=simple
KillMode=none
ExecStart=/usr/bin/mergerfs \
  -f \
  -o cache.files=auto-full \
  -o dropcacheonclose=true \
  -o category.create=mfs \
  /mnt/hdd0:/mnt/hdd1 \
  /media
ExecStop=/bin/fusermount -uz /media
Restart=on-failure

[Install]
WantedBy=default.target

See the mergerfs wiki for real world deployments for comparisons / ideas.

OPTIONS

These options are the same regardless of whether you use them with the mergerfs commandline program, in fstab, or in a config file.

mount options

  • config: Path to a config file. Same arguments as below in key=val / ini style format.
  • branches: Colon delimited list of branches.
  • minfreespace=SIZE: The minimum space value used for creation policies. Can be overridden by branch specific option. Understands 'K', 'M', and 'G' to represent kilobyte, megabyte, and gigabyte respectively. (default: 4G)
  • moveonenospc=BOOL|POLICY: When enabled if a write fails with ENOSPC (no space left on device) or EDQUOT (disk quota exceeded) the policy selected will run to find a new location for the file. An attempt to move the file to that branch will occur (keeping all metadata possible) and if successful the original is unlinked and the write retried. (default: false, true = mfs)
  • inodecalc=passthrough|path-hash|devino-hash|hybrid-hash: Selects the inode calculation algorithm. (default: hybrid-hash)
  • dropcacheonclose=BOOL: When a file is requested to be closed call posix_fadvise on it first to instruct the kernel that we no longer need the data and it can drop its cache. Recommended when cache.files=partial|full|auto-full|per-process to limit double caching. (default: false)
  • direct-io-allow-mmap=BOOL: On newer kernels (>= 6.6) it is possible to disable file page caching while still allowing for shared mmap support. mergerfs will enable this feature if available but an option is provided to turn it off for testing and debugging purposes. (default: true)
  • symlinkify=BOOL: When enabled and a file is not writable and its mtime or ctime is older than symlinkify_timeout files will be reported as symlinks to the original files. Please read more below before using. (default: false)
  • symlinkify_timeout=UINT: Time to wait, in seconds, to activate the symlinkify behavior. (default: 3600)
  • nullrw=BOOL: Turns reads and writes into no-ops. The request will succeed but do nothing. Useful for benchmarking mergerfs. (default: false)
  • lazy-umount-mountpoint=BOOL: mergerfs will attempt to "lazy umount" the mountpoint before mounting itself. Useful when performing live upgrades of mergerfs. (default: false)
  • ignorepponrename=BOOL: Ignore path preserving on rename. Typically rename and link act differently depending on the policy of create (read below). Enabling this will cause rename and link to always use the non-path preserving behavior. This means files, when renamed or linked, will stay on the same filesystem. (default: false)
  • export-support=BOOL: Sets a low-level FUSE feature intended to indicate the filesystem can support being exported via NFS. (default: true)
  • security_capability=BOOL: If false return ENOATTR when xattr security.capability is queried. (default: true)
  • xattr=passthrough|noattr|nosys: Runtime control of xattrs. Default is to passthrough xattr requests. 'noattr' will short circuit as if nothing exists. 'nosys' will respond with ENOSYS as if xattrs are not supported or disabled. (default: passthrough)
  • link_cow=BOOL: When enabled if a regular file is opened which has a link count > 1 it will copy the file to a temporary file and rename over the original. Breaking the link and providing a basic copy-on-write function similar to cow-shell. (default: false)
  • statfs=base|full: Controls how statfs works. 'base' means it will always use all branches in statfs calculations. 'full' is in effect path preserving and only includes branches where the path exists. (default: base)
  • statfs_ignore=none|ro|nc: 'ro' will cause statfs calculations to ignore available space for branches mounted or tagged as 'read-only' or 'no create'. 'nc' will ignore available space for branches tagged as 'no create'. (default: none)
  • nfsopenhack=off|git|all: A workaround for exporting mergerfs over NFS where there are issues with creating files for write while setting the mode to read-only. (default: off)
  • branches-mount-timeout=UINT: Number of seconds to wait at startup for branches to be a mount other than the mountpoint's filesystem. (default: 0)
  • follow-symlinks=never|directory|regular|all: Turns symlinks into what they point to. (default: never)
  • link-exdev=passthrough|rel-symlink|abs-base-symlink|abs-pool-symlink: When a link fails with EXDEV optionally create a symlink to the file instead.
  • rename-exdev=passthrough|rel-symlink|abs-symlink: When a rename fails with EXDEV optionally move the file to a special directory and symlink to it.
  • readahead=UINT: Set readahead (in kilobytes) for mergerfs and branches if greater than 0. (default: 0)
  • posix_acl=BOOL: Enable POSIX ACL support (if supported by kernel and underlying filesystem). (default: false)
  • async_read=BOOL: Perform reads asynchronously. If disabled or unavailable the kernel will ensure there is at most one pending read request per file handle and will attempt to order requests by offset. (default: true)
  • fuse_msg_size=UINT: Set the max number of pages per FUSE message. Only available on Linux >= 4.20 and ignored otherwise. (min: 1; max: 256; default: 256)
  • threads=INT: Number of threads to use. When used alone (process-thread-count=-1) it sets the number of threads reading and processing FUSE messages. When used together it sets the number of threads reading from FUSE. When set to zero it will attempt to discover and use the number of logical cores. If the thread count is set negative it will look up the number of cores then divide by the absolute value. ie. threads=-2 on an 8 core machine will result in 8 / 2 = 4 threads. There will always be at least 1 thread. If set to -1 in combination with process-thread-count then it will try to pick reasonable values based on CPU thread count. NOTE: higher number of threads increases parallelism but usually decreases throughput. (default: 0)
  • read-thread-count=INT: Alias for threads.
  • process-thread-count=INT: Enables separate thread pool to asynchronously process FUSE requests. In this mode read-thread-count refers to the number of threads reading FUSE messages which are dispatched to process threads. -1 means disabled otherwise acts like read-thread-count. (default: -1)
  • process-thread-queue-depth=UINT: Sets the number of requests any single process thread can have queued up at one time. Meaning the total memory usage of the queues is queue depth multiplied by the number of process threads plus read thread count. 0 sets the depth to the same as the process thread count. (default: 0)
  • pin-threads=STR: Selects a strategy to pin threads to CPUs (default: unset)
  • flush-on-close=never|always|opened-for-write: Flush data cache on file close. Mostly for when writeback is enabled or merging network filesystems. (default: opened-for-write)
  • scheduling-priority=INT: Set mergerfs' scheduling priority. Valid values range from -20 to 19. See setpriority man page for more details. (default: -10)
  • fsname=STR: Sets the name of the filesystem as seen in mount, df, etc. Defaults to a list of the source paths concatenated together with the longest common prefix removed.
  • func.FUNC=POLICY: Sets the specific FUSE function's policy. See below for the list of value types. Example: func.getattr=newest
  • func.readdir=seq|cosr|cor|cosr:INT|cor:INT: Sets readdir policy. INT value sets the number of threads to use for concurrency. (default: seq)
  • category.action=POLICY: Sets policy of all FUSE functions in the action category. (default: epall)
  • category.create=POLICY: Sets policy of all FUSE functions in the create category. (default: epmfs)
  • category.search=POLICY: Sets policy of all FUSE functions in the search category. (default: ff)
  • cache.open=UINT: 'open' policy cache timeout in seconds. (default: 0)
  • cache.statfs=UINT: 'statfs' cache timeout in seconds. (default: 0)
  • cache.attr=UINT: File attribute cache timeout in seconds. (default: 1)
  • cache.entry=UINT: File name lookup cache timeout in seconds. (default: 1)
  • cache.negative_entry=UINT: Negative file name lookup cache timeout in seconds. (default: 0)
  • cache.files=libfuse|off|partial|full|auto-full|per-process: File page caching mode (default: libfuse)
  • cache.files.process-names=LIST: A pipe | delimited list of process comm names to enable page caching for when cache.files=per-process. (default: "rtorrent|qbittorrent-nox")
  • cache.writeback=BOOL: Enable kernel writeback caching (default: false)
  • cache.symlinks=BOOL: Cache symlinks (if supported by kernel) (default: false)
  • cache.readdir=BOOL: Cache readdir (if supported by kernel) (default: false)
  • parallel-direct-writes=BOOL: Allow the kernel to dispatch multiple, parallel (non-extending) write requests for files opened with cache.files=per-process (if the process is not in process-names) or cache.files=off. (This requires kernel support, and was added in v6.2)
  • direct_io: deprecated - Bypass page cache. Use cache.files=off instead. (default: false)
  • kernel_cache: deprecated - Do not invalidate data cache on file open. Use cache.files=full instead. (default: false)
  • auto_cache: deprecated - Invalidate data cache if file mtime or size change. Use cache.files=auto-full instead. (default: false)
  • async_read: deprecated - Perform reads asynchronously. Use async_read=true instead.
  • sync_read: deprecated - Perform reads synchronously. Use async_read=false instead.
  • splice_read: deprecated - Does nothing.
  • splice_write: deprecated - Does nothing.
  • splice_move: deprecated - Does nothing.
  • allow_other: deprecated - mergerfs v2.35.0 and newer sets this FUSE option automatically if running as root.
  • use_ino: deprecated - mergerfs should always control inode calculation so this is enabled all the time.

NOTE: Options are evaluated in the order listed so if the options are func.rmdir=rand,category.action=ff the action category setting will override the rmdir setting.

NOTE: Always look at the documentation for the version of mergerfs you're using. Not all features are available in older releases. Use man mergerfs or find the docs as linked in the release.

Value Types

  • BOOL = 'true' | 'false'
  • INT = [MIN_INT,MAX_INT]
  • UINT = [0,MAX_INT]
  • SIZE = 'NNM'; NN = INT, M = 'K' | 'M' | 'G' | 'T'
  • STR = string (may refer to an enumerated value, see details of argument)
  • FUNC = filesystem function
  • CATEGORY = function category
  • POLICY = mergerfs function policy

branches

The 'branches' argument is a colon (':') delimited list of paths to be pooled together. It does not matter if the paths are on the same or different filesystems nor does it matter the filesystem type (within reason). Used and available space will not be duplicated for paths on the same filesystem and any features which aren't supported by the underlying filesystem (such as file attributes or extended attributes) will return the appropriate errors.

Branches currently have two options which can be set. A type which impacts whether or not the branch is included in a policy calculation and a individual minfreespace value. The values are set by prepending an = at the end of a branch designation and using commas as delimiters. Example: /mnt/drive=RW,1234

branch mode

  • RW: (read/write) - Default behavior. Will be eligible in all policy categories.
  • RO: (read-only) - Will be excluded from create and action policies. Same as a read-only mounted filesystem would be (though faster to process).
  • NC: (no-create) - Will be excluded from create policies. You can't create on that branch but you can change or delete.

minfreespace

Same purpose and syntax as the global option but specific to the branch. If not set the global value is used.

globbing

To make it easier to include multiple branches mergerfs supports globbing. The globbing tokens MUST be escaped when using via the shell else the shell itself will apply the glob itself.

# mergerfs /mnt/hdd\*:/mnt/ssd /media

The above line will use all mount points in /mnt prefixed with hdd and ssd.

To have the pool mounted at boot or otherwise accessible from related tools use /etc/fstab.

# <file system>        <mount point>  <type>    <options>             <dump>  <pass>
/mnt/hdd*:/mnt/ssd    /media          mergerfs  minfreespace=16G      0       0

NOTE: the globbing is done at mount or when updated using the runtime API. If a new directory is added matching the glob after the fact it will not be automatically included.

NOTE: for mounting via fstab to work you must have mount.fuse installed. For Ubuntu/Debian it is included in the fuse package.

inodecalc

Inodes (st_ino) are unique identifiers within a filesystem. Each mounted filesystem has device ID (st_dev) as well and together they can uniquely identify a file on the whole of the system. Entries on the same device with the same inode are in fact references to the same underlying file. It is a many to one relationship between names and an inode. Directories, however, do not have multiple links on most systems due to the complexity they add.

FUSE allows the server (mergerfs) to set inode values but not device IDs. Creating an inode value is somewhat complex in mergerfs' case as files aren't really in its control. If a policy changes what directory or file is to be selected or something changes out of band it becomes unclear what value should be used. Most software does not to care what the values are but those that do often break if a value changes unexpectedly. The tool find will abort a directory walk if it sees a directory inode change. NFS can return stale handle errors if the inode changes out of band. File dedup tools will usually leverage device ids and inodes as a shortcut in searching for duplicate files and would resort to full file comparisons should it find different inode values.

mergerfs offers multiple ways to calculate the inode in hopes of covering different usecases.

  • passthrough: Passes through the underlying inode value. Mostly intended for testing as using this does not address any of the problems mentioned above and could confuse file deduplication software as inodes from different filesystems can be the same.
  • path-hash: Hashes the relative path of the entry in question. The underlying file's values are completely ignored. This means the inode value will always be the same for that file path. This is useful when using NFS and you make changes out of band such as copy data between branches. This also means that entries that do point to the same file will not be recognizable via inodes. That does not mean hard links don't work. They will.
  • path-hash32: 32bit version of path-hash.
  • devino-hash: Hashes the device id and inode of the underlying entry. This won't prevent issues with NFS should the policy pick a different file or files move out of band but will present the same inode for underlying files that do too.
  • devino-hash32: 32bit version of devino-hash.
  • hybrid-hash: Performs path-hash on directories and devino-hash on other file types. Since directories can't have hard links the static value won't make a difference and the files will get values useful for finding duplicates. Probably the best to use if not using NFS. As such it is the default.
  • hybrid-hash32: 32bit version of hybrid-hash.

32bit versions are provided as there is some software which does not handle 64bit inodes well.

While there is a risk of hash collision in tests of a couple million entries there were zero collisions. Unlike a typical filesystem FUSE filesystems can reuse inodes and not refer to the same entry. The internal identifier used to reference a file in FUSE is different from the inode value presented. The former is the nodeid and is actually a tuple of 2 64bit values: nodeid and generation. This tuple is not client facing. The inode that is presented to the client is passed through the kernel uninterpreted.

From FUSE docs for use_ino:

Honor the st_ino field in the functions getattr() and
fill_dir(). This value is used to fill in the st_ino field
in the stat(2), lstat(2), fstat(2) functions and the d_ino
field in the readdir(2) function. The filesystem does not
have to guarantee uniqueness, however some applications
rely on this value being unique for the whole filesystem.
Note that this does *not* affect the inode that libfuse
and the kernel use internally (also called the "nodeid").

As of version 2.35.0 the use_ino option has been removed. mergerfs should always be managing inode values.

pin-threads

Simple strategies for pinning read and/or process threads. If process threads are not enabled than the strategy simply works on the read threads. Invalid values are ignored.

  • R1L: All read threads pinned to a single logical CPU.
  • R1P: All read threads pinned to a single physical CPU.
  • RP1L: All read and process threads pinned to a single logical CPU.
  • RP1P: All read and process threads pinned to a single physical CPU.
  • R1LP1L: All read threads pinned to a single logical CPU, all process threads pinned to a (if possible) different logical CPU.
  • R1PP1P: All read threads pinned to a single physical CPU, all process threads pinned to a (if possible) different logical CPU.
  • RPSL: All read and process threads are spread across all logical CPUs.
  • RPSP: All read and process threads are spread across all physical CPUs.
  • R1PPSP: All read threads are pinned to a single physical CPU while process threads are spread across all other phsycial CPUs.

fuse_msg_size

FUSE applications communicate with the kernel over a special character device: /dev/fuse. A large portion of the overhead associated with FUSE is the cost of going back and forth from user space and kernel space over that device. Generally speaking the fewer trips needed the better the performance will be. Reducing the number of trips can be done a number of ways. Kernel level caching and increasing message sizes being two significant ones. When it comes to reads and writes if the message size is doubled the number of trips are approximately halved.

In Linux 4.20 a new feature was added allowing the negotiation of the max message size. Since the size is in multiples of pages the feature is called max_pages. There is a maximum max_pages value of 256 (1MiB) and minimum of 1 (4KiB). The default used by Linux >=4.20, and hardcoded value used before 4.20, is 32 (128KiB). In mergerfs its referred to as fuse_msg_size to make it clear what it impacts and provide some abstraction.

Since there should be no downsides to increasing fuse_msg_size / max_pages, outside a minor bump in RAM usage due to larger message buffers, mergerfs defaults the value to 256. On kernels before 4.20 the value has no effect. The reason the value is configurable is to enable experimentation and benchmarking. See the BENCHMARKING section for examples.

This feature, when enabled, will cause symlinks to be interpreted by mergerfs as their target (depending on the mode).

When there is a getattr/stat request for a file mergerfs will check if the file is a symlink and depending on the follow-symlinks setting will replace the information about the symlink with that of that which it points to.

When unlink'ing or rmdir'ing the followed symlink it will remove the symlink itself and not that which it points to.

  • never: Behave as normal. Symlinks are treated as such.
  • directory: Resolve symlinks only which point to directories.
  • regular: Resolve symlinks only which point to regular files.
  • all: Resolve all symlinks to that which they point to.

Symlinks which do not point to anything are left as is.

WARNING: This feature works but there might be edge cases yet found. If you find any odd behaviors please file a ticket on github.

If using path preservation and a link fails with EXDEV make a call to symlink where the target is the oldlink and the linkpath is the newpath. The target value is determined by the value of link-exdev.

  • passthrough: Return EXDEV as normal.
  • rel-symlink: A relative path from the newpath.
  • abs-base-symlink: A absolute value using the underlying branch.
  • abs-pool-symlink: A absolute value using the mergerfs mount point.

NOTE: It is possible that some applications check the file they link. In those cases it is possible it will error or complain.

rename-exdev

If using path preservation and a rename fails with EXDEV:

  1. Move file from /branch/a/b/c to /branch/.mergerfs_rename_exdev/a/b/c.
  2. symlink the rename's newpath to the moved file.

The target value is determined by the value of rename-exdev.

  • passthrough: Return EXDEV as normal.
  • rel-symlink: A relative path from the newpath.
  • abs-symlink: A absolute value using the mergerfs mount point.

NOTE: It is possible that some applications check the file they rename. In those cases it is possible it will error or complain.

NOTE: The reason abs-symlink is not split into two like link-exdev is due to the complexities in managing absolute base symlinks when multiple oldpaths exist.

symlinkify

Due to the levels of indirection introduced by mergerfs and the underlying technology FUSE there can be varying levels of performance degradation. This feature will turn non-directories which are not writable into symlinks to the original file found by the readlink policy after the mtime and ctime are older than the timeout.

WARNING: The current implementation has a known issue in which if the file is open and being used when the file is converted to a symlink then the application which has that file open will receive an error when using it. This is unlikely to occur in practice but is something to keep in mind.

WARNING: Some backup solutions, such as CrashPlan, do not backup the target of a symlink. If using this feature it will be necessary to point any backup software to the original filesystems or configure the software to follow symlinks if such an option is available. Alternatively create two mounts. One for backup and one for general consumption.

nullrw

Due to how FUSE works there is an overhead to all requests made to a FUSE filesystem that wouldn't exist for an in kernel one. Meaning that even a simple passthrough will have some slowdown. However, generally the overhead is minimal in comparison to the cost of the underlying I/O. By disabling the underlying I/O we can test the theoretical performance boundaries.

By enabling nullrw mergerfs will work as it always does except that all reads and writes will be no-ops. A write will succeed (the size of the write will be returned as if it were successful) but mergerfs does nothing with the data it was given. Similarly a read will return the size requested but won't touch the buffer.

See the BENCHMARKING section for suggestions on how to test.

xattr

Runtime extended attribute support can be managed via the xattr option. By default it will passthrough any xattr calls. Given xattr support is rarely used and can have significant performance implications mergerfs allows it to be disabled at runtime. The performance problems mostly comes when file caching is enabled. The kernel will send a getxattr for security.capability before every single write. It doesn't cache the responses to any getxattr. This might be addressed in the future but for now mergerfs can really only offer the following workarounds.

noattr will cause mergerfs to short circuit all xattr calls and return ENOATTR where appropriate. mergerfs still gets all the requests but they will not be forwarded on to the underlying filesystems. The runtime control will still function in this mode.

nosys will cause mergerfs to return ENOSYS for any xattr call. The difference with noattr is that the kernel will cache this fact and itself short circuit future calls. This is more efficient than noattr but will cause mergerfs' runtime control via the hidden file to stop working.

nfsopenhack

NFS is not fully POSIX compliant and historically certain behaviors, such as opening files with O_EXCL, are not or not well supported. When mergerfs (or any FUSE filesystem) is exported over NFS some of these issues come up due to how NFS and FUSE interact.

This hack addresses the issue where the creation of a file with a read-only mode but with a read/write or write only flag. Normally this is perfectly valid but NFS chops the one open call into multiple calls. Exactly how it is translated depends on the configuration and versions of the NFS server and clients but it results in a permission error because a normal user is not allowed to open a read-only file as writable.

Even though it's a more niche situation this hack breaks normal security and behavior and as such is off by default. If set to git it will only perform the hack when the path in question includes /.git/. all will result it applying anytime a read-only file which is empty is opened for writing.

export-support

In theory this flag should not be exposed to the end user. It is a low-level FUSE flag which indicates whether or not the kernel can send certain kinds of messages to it for the purposes of using with NFS. mergerfs does support these messages but due bugs and quirks found in the kernel and mergerfs this option is provided just in case it is needed for debugging.

Given that this flag is set when the FUSE connection is first initiated it is not possible to change during run time.

FUNCTIONS, CATEGORIES and POLICIES

The POSIX filesystem API is made up of a number of functions. creat, stat, chown, etc. For ease of configuration in mergerfs most of the core functions are grouped into 3 categories: action, create, and search. These functions and categories can be assigned a policy which dictates which branch is chosen when performing that function.

Some functions, listed in the category N/A below, can not be assigned the normal policies. These functions work with file handles, rather than file paths, which were created by open or create. That said many times the current FUSE kernel driver will not always provide the file handle when a client calls fgetattr, fchown, fchmod, futimens, ftruncate, etc. This means it will call the regular, path based, versions. statfs's behavior can be modified via other options.

When using policies which are based on a branch's available space the base path provided is used. Not the full path to the file in question. Meaning that mounts in the branch won't be considered in the space calculations. The reason is that it doesn't really work for non-path preserving policies and can lead to non-obvious behaviors.

NOTE: While any policy can be assigned to a function or category, some may not be very useful in practice. For instance: rand (random) may be useful for file creation (create) but could lead to very odd behavior if used for chmod if there were more than one copy of the file.

Functions and their Category classifications

Category FUSE Functions
action chmod, chown, link, removexattr, rename, rmdir, setxattr, truncate, unlink, utimens
create create, mkdir, mknod, symlink
search access, getattr, getxattr, ioctl (directories), listxattr, open, readlink
N/A fchmod, fchown, futimens, ftruncate, fallocate, fgetattr, fsync, ioctl (files), read, readdir, release, statfs, write, copy_file_range

In cases where something may be searched for (such as a path to clone) getattr will usually be used.

Policies

A policy is the algorithm used to choose a branch or branches for a function to work on or generally how the function behaves.

Any function in the create category will clone the relative path if needed. Some other functions (rename,link,ioctl) have special requirements or behaviors which you can read more about below.

Filtering

Most policies basically search branches and create a list of files / paths for functions to work on. The policy is responsible for filtering and sorting the branches. Filters include minfreespace, whether or not a branch is mounted read-only, and the branch tagging (RO,NC,RW). These filters are applied across most policies.

  • No search function policies filter.
  • All action function policies filter out branches which are mounted read-only or tagged as RO (read-only).
  • All create function policies filter out branches which are mounted read-only, tagged RO (read-only) or NC (no create), or has available space less than minfreespace.

Policies may have their own additional filtering such as those that require existing paths to be present.

If all branches are filtered an error will be returned. Typically EROFS (read-only filesystem) or ENOSPC (no space left on device) depending on the most recent reason for filtering a branch. ENOENT will be returned if no eligible branch is found.

If create, mkdir, mknod, or symlink fail with EROFS or other fundimental errors then mergerfs will mark any branch found to be read-only as such (IE will set the mode RO) and will rerun the policy and try again. This is mostly for ext4 filesystems that can suddenly become read-only when it encounters an error.

Path Preservation

Policies, as described below, are of two basic classifications. path preserving and non-path preserving.

All policies which start with ep (epff, eplfs, eplus, epmfs, eprand) are path preserving. ep stands for existing path.

A path preserving policy will only consider branches where the relative path being accessed already exists.

When using non-path preserving policies paths will be cloned to target branches as necessary.

With the msp or most shared path policies they are defined as path preserving for the purpose of controlling link and rename's behaviors since ignorepponrename is available to disable that behavior.

Policy descriptions

A policy's behavior differs, as mentioned above, based on the function it is used with. Sometimes it really might not make sense to even offer certain policies because they are literally the same as others but it makes things a bit more uniform.

Policy Description
all Search: For mkdir, mknod, and symlink it will apply to all branches. create works like ff.
epall (existing path, all) For mkdir, mknod, and symlink it will apply to all found. create works like epff (but more expensive because it doesn't stop after finding a valid branch).
epff (existing path, first found) Given the order of the branches, as defined at mount time or configured at runtime, act on the first one found where the relative path exists.
eplfs (existing path, least free space) Of all the branches on which the relative path exists choose the branch with the least free space.
eplus (existing path, least used space) Of all the branches on which the relative path exists choose the branch with the least used space.
epmfs (existing path, most free space) Of all the branches on which the relative path exists choose the branch with the most free space.
eppfrd (existing path, percentage free random distribution) Like pfrd but limited to existing paths.
eprand (existing path, random) Calls epall and then randomizes. Returns 1.
ff (first found) Given the order of the branches, as defined at mount time or configured at runtime, act on the first one found.
lfs (least free space) Pick the branch with the least available free space.
lus (least used space) Pick the branch with the least used space.
mfs (most free space) Pick the branch with the most available free space.
msplfs (most shared path, least free space) Like eplfs but if it fails to find a branch it will try again with the parent directory. Continues this pattern till finding one.
msplus (most shared path, least used space) Like eplus but if it fails to find a branch it will try again with the parent directory. Continues this pattern till finding one.
mspmfs (most shared path, most free space) Like epmfs but if it fails to find a branch it will try again with the parent directory. Continues this pattern till finding one.
msppfrd (most shared path, percentage free random distribution) Like eppfrd but if it fails to find a branch it will try again with the parent directory. Continues this pattern till finding one.
newest Pick the file / directory with the largest mtime.
pfrd (percentage free random distribution) Chooses a branch at random with the likelihood of selection based on a branch's available space relative to the total.
rand (random) Calls all and then randomizes. Returns 1 branch.

NOTE: If you are using an underlying filesystem that reserves blocks such as ext2, ext3, or ext4 be aware that mergerfs respects the reservation by using f_bavail (number of free blocks for unprivileged users) rather than f_bfree (number of free blocks) in policy calculations. df does NOT use f_bavail, it uses f_bfree, so direct comparisons between df output and mergerfs' policies is not appropriate.

Defaults

Category Policy
action epall
create epmfs
search ff

func.readdir

examples: func.readdir=seq, func.readdir=cor:4

readdir has policies to control how it manages reading directory content.

Policy Description
seq "sequential" : Iterate over branches in the order defined. This is the default and traditional behavior found prior to the readdir policy introduction.
cosr "concurrent open, sequential read" : Concurrently open branch directories using a thread pool and process them in order of definition. This keeps memory and CPU usage low while also reducing the time spent waiting on branches to respond. Number of threads defaults to the number of logical cores. Can be overwritten via the syntax func.readdir=cosr:N where N is the number of threads.
cor "concurrent open and read" : Concurrently open branch directories and immediately start reading their contents using a thread pool. This will result in slightly higher memory and CPU usage but reduced latency. Particularly when using higher latency / slower speed network filesystem branches. Unlike seq and cosr the order of files could change due the async nature of the thread pool. Number of threads defaults to the number of logical cores. Can be overwritten via the syntax func.readdir=cor:N where N is the number of threads.

Keep in mind that readdir mostly just provides a list of file names in a directory and possibly some basic metadata about said files. To know details about the files, as one would see from commands like find or ls, it is required to call stat on the file which is controlled by fuse.getattr.

ioctl

When ioctl is used with an open file then it will use the file handle which was created at the original open call. However, when using ioctl with a directory mergerfs will use the open policy to find the directory to act on.

NOTE: If you're receiving errors from software when files are moved / renamed / linked then you should consider changing the create policy to one which is not path preserving, enabling ignorepponrename, or contacting the author of the offending software and requesting that EXDEV (cross device / improper link) be properly handled.

rename and link are tricky functions in a union filesystem. rename only works within a single filesystem or device. If a rename can't be done atomically due to the source and destination paths existing on different mount points it will return -1 with errno = EXDEV (cross device / improper link). So if a rename's source and target are on different filesystems within the pool it creates an issue.

Originally mergerfs would return EXDEV whenever a rename was requested which was cross directory in any way. This made the code simple and was technically compliant with POSIX requirements. However, many applications fail to handle EXDEV at all and treat it as a normal error or otherwise handle it poorly. Such apps include: gvfsd-fuse v1.20.3 and prior, Finder / CIFS/SMB client in Apple OSX 10.9+, NZBGet, Samba's recycling bin feature.

As a result a compromise was made in order to get most software to work while still obeying mergerfs' policies. Below is the basic logic.

  • If using a create policy which tries to preserve directory paths (epff,eplfs,eplus,epmfs)
    • Using the rename policy get the list of files to rename
    • For each file attempt rename:
      • If failure with ENOENT (no such file or directory) run create policy
      • If create policy returns the same branch as currently evaluating then clone the path
      • Re-attempt rename
    • If any of the renames succeed the higher level rename is considered a success
    • If no renames succeed the first error encountered will be returned
    • On success:
      • Remove the target from all branches with no source file
      • Remove the source from all branches which failed to rename
  • If using a create policy which does not try to preserve directory paths
    • Using the rename policy get the list of files to rename
    • Using the getattr policy get the target path
    • For each file attempt rename:
      • If the source branch != target branch:
        • Clone target path from target branch to source branch
      • Rename
    • If any of the renames succeed the higher level rename is considered a success
    • If no renames succeed the first error encountered will be returned
    • On success:
      • Remove the target from all branches with no source file
      • Remove the source from all branches which failed to rename

The the removals are subject to normal entitlement checks.

The above behavior will help minimize the likelihood of EXDEV being returned but it will still be possible.

link uses the same strategy but without the removals.

statfs / statvfs

statvfs normalizes the source filesystems based on the fragment size and sums the number of adjusted blocks and inodes. This means you will see the combined space of all sources. Total, used, and free. The sources however are dedupped based on the filesystem so multiple sources on the same drive will not result in double counting its space. Other filesystems mounted further down the tree of the branch will not be included when checking the mount's stats.

The options statfs and statfs_ignore can be used to modify statfs behavior.

flush-on-close

https://lkml.kernel.org/linux-fsdevel/20211024132607.1636952-1-amir73il@gmail.com/T/

By default FUSE would issue a flush before the release of a file descriptor. This was considered a bit aggressive and a feature added to give the FUSE server the ability to choose when that happens.

Options:

  • always
  • never
  • opened-for-write

For now it defaults to "opened-for-write" which is less aggressive than the behavior before this feature was added. It should not be a problem because the flush is really only relevant when a file is written to. Given flush is irrelevant for many filesystems in the future a branch specific flag may be added so only files opened on a specific branch would be flushed on close.

ERROR HANDLING

POSIX filesystem functions offer a single return code meaning that there is some complication regarding the handling of multiple branches as mergerfs does. It tries to handle errors in a way that would generally return meaningful values for that particular function.

chmod, chown, removexattr, setxattr, truncate, utimens

  1. if no error: return 0 (success)
  2. if no successes: return first error
  3. if one of the files acted on was the same as the related search function: return its value
  4. return 0 (success)

While doing this increases the complexity and cost of error handling, particularly step 3, this provides probably the most reasonable return value.

  1. if no errors: return 0 (success)
  2. return first error

Older version of mergerfs would return success if any success occurred but for unlink and rmdir there are downstream assumptions that, while not impossible to occur, can confuse some software.

others

For search functions there is always a single thing acted on and as such whatever return value that comes from the single function call is returned.

For create functions mkdir, mknod, and symlink which don't return a file descriptor and therefore can have all or epall policies it will return success if any of the calls succeed and an error otherwise.

INSTALL

https://github.com/trapexit/mergerfs/releases

If your distribution's package manager includes mergerfs check if the version is up to date. If out of date it is recommended to use the latest release found on the release page. Details for common distros are below.

Debian

Most Debian installs are of a stable branch and therefore do not have the most up to date software. While mergerfs is available via apt it is suggested that uses install the most recent version available from the releases page.

prebuilt deb

wget https://github.com/trapexit/mergerfs/releases/download/<ver>/mergerfs_<ver>.debian-<rel>_<arch>.deb
dpkg -i mergerfs_<ver>.debian-<rel>_<arch>.deb

apt

sudo apt install -y mergerfs

Ubuntu

Most Ubuntu installs are of a stable branch and therefore do not have the most up to date software. While mergerfs is available via apt it is suggested that uses install the most recent version available from the releases page.

prebuilt deb

wget https://github.com/trapexit/mergerfs/releases/download/<version>/mergerfs_<ver>.ubuntu-<rel>_<arch>.deb
dpkg -i mergerfs_<ver>.ubuntu-<rel>_<arch>.deb

apt

sudo apt install -y mergerfs

Raspberry Pi OS

Effectively the same as Debian or Ubuntu.

Fedora

wget https://github.com/trapexit/mergerfs/releases/download/<ver>/mergerfs-<ver>.fc<rel>.<arch>.rpm
sudo rpm -i mergerfs-<ver>.fc<rel>.<arch>.rpm

CentOS / Rocky

wget https://github.com/trapexit/mergerfs/releases/download/<ver>/mergerfs-<ver>.el<rel>.<arch>.rpm
sudo rpm -i mergerfs-<ver>.el<rel>.<arch>.rpm

ArchLinux

  1. Setup AUR
  2. Install mergerfs

Other

Static binaries are provided for situations where native packages are unavailable.

wget https://github.com/trapexit/mergerfs/releases/download/<ver>/mergerfs-static-linux_<arch>.tar.gz
sudo tar xvf mergerfs-static-linux_<arch>.tar.gz -C /

BUILD

NOTE: Prebuilt packages can be found at and recommended for most users: https://github.com/trapexit/mergerfs/releases

NOTE: Only tagged releases are supported. master and other branches should be considered works in progress.

First get the code from github.

$ git clone https://github.com/trapexit/mergerfs.git
$ # or
$ wget https://github.com/trapexit/mergerfs/releases/download/<ver>/mergerfs-<ver>.tar.gz

Debian / Ubuntu

$ cd mergerfs
$ sudo tools/install-build-pkgs
$ make deb
$ sudo dpkg -i ../mergerfs_<version>_<arch>.deb

RHEL / CentOS / Rocky / Fedora

$ su -
# cd mergerfs
# tools/install-build-pkgs
# make rpm
# rpm -i rpmbuild/RPMS/<arch>/mergerfs-<version>.<arch>.rpm

Generic

Have git, g++, make, python installed.

$ cd mergerfs
$ make
$ sudo make install

Build options

$ make help
usage: make

make USE_XATTR=0      - build program without xattrs functionality
make STATIC=1         - build static binary
make LTO=1            - build with link time optimization

UPGRADE

mergerfs can be upgraded live by mounting on top of the previous instance. Simply install the new version of mergerfs and follow the instructions below.

Run mergerfs again or if using /etc/fstab call for it to mount again. Existing open files and such will continue to work fine though they won't see runtime changes since any such change would be the new mount. If you plan on changing settings with the new mount you should / could apply those before mounting the new version.

$ sudo mount /mnt/mergerfs
$ mount | grep mergerfs
media on /mnt/mergerfs type mergerfs (rw,relatime,user_id=0,group_id=0,default_permissions,allow_other)
media on /mnt/mergerfs type mergerfs (rw,relatime,user_id=0,group_id=0,default_permissions,allow_other)

A problem with this approach is that the underlying instance will continue to run even if the software using it stop or are restarted. To work around this you can use a "lazy umount". Before mounting over top the mount point with the new instance of mergerfs issue: umount -l <mergerfs_mountpoint>. Or you can let mergerfs do it by setting the option lazy-umount-mountpoint=true.

RUNTIME INTERFACES

RUNTIME CONFIG

.mergerfs pseudo file

<mountpoint>/.mergerfs

There is a pseudo file available at the mount point which allows for the runtime modification of certain mergerfs options. The file will not show up in readdir but can be stat'ed and manipulated via {list,get,set}xattrs calls.

Any changes made at runtime are not persisted. If you wish for values to persist they must be included as options wherever you configure the mounting of mergerfs (/etc/fstab).

Keys

Use getfattr -d /mountpoint/.mergerfs or xattr -l /mountpoint/.mergerfs to see all supported keys. Some are informational and therefore read-only. setxattr will return EINVAL (invalid argument) on read-only keys.

Values

Same as the command line.

user.mergerfs.branches

Used to query or modify the list of branches. When modifying there are several shortcuts to easy manipulation of the list.

Value Description
[list] set
+<[list] prepend
+>[list] append
-[list] remove all values provided
-< remove first in list
-> remove last in list

xattr -w user.mergerfs.branches +</mnt/drive3 /mnt/pool/.mergerfs

The =NC, =RO, =RW syntax works just as on the command line.

Example
[trapexit:/mnt/mergerfs] $ getfattr -d .mergerfs
user.mergerfs.branches="/mnt/a=RW:/mnt/b=RW"
user.mergerfs.minfreespace="4294967295"
user.mergerfs.moveonenospc="false"
...

[trapexit:/mnt/mergerfs] $ getfattr -n user.mergerfs.category.search .mergerfs
user.mergerfs.category.search="ff"

[trapexit:/mnt/mergerfs] $ setfattr -n user.mergerfs.category.search -v newest .mergerfs
[trapexit:/mnt/mergerfs] $ getfattr -n user.mergerfs.category.search .mergerfs
user.mergerfs.category.search="newest"

file / directory xattrs

While they won't show up when using getfattr mergerfs offers a number of special xattrs to query information about the files served. To access the values you will need to issue a getxattr for one of the following:

  • user.mergerfs.basepath: the base mount point for the file given the current getattr policy
  • user.mergerfs.relpath: the relative path of the file from the perspective of the mount point
  • user.mergerfs.fullpath: the full path of the original file given the getattr policy
  • user.mergerfs.allpaths: a NUL ('\0') separated list of full paths to all files found

SIGNALS

  • USR1: This will cause mergerfs to send invalidation notifications to the kernel for all files. This will cause all unused files to be released from memory.
  • USR2: Trigger a general cleanup of currently unused memory. A more thorough version of what happens every ~15 minutes.

IOCTLS

Found in fuse_ioctl.cpp:

typedef char IOCTL_BUF[4096];
#define IOCTL_APP_TYPE  0xDF
#define IOCTL_FILE_INFO            _IOWR(IOCTL_APP_TYPE,0,IOCTL_BUF)
#define IOCTL_GC                   _IO(IOCTL_APP_TYPE,1)
#define IOCTL_GC1                  _IO(IOCTL_APP_TYPE,2)
#define IOCTL_INVALIDATE_ALL_NODES _IO(IOCTL_APP_TYPE,3)
  • IOCTL_FILE_INFO: Same as the "file / directory xattrs" mentioned above. Use a buffer size of 4096 bytes. Pass in a string of "basepath", "relpath", "fullpath", or "allpaths". Receive details in same buffer.
  • IOCTL_GC: Triggers a thorough garbage collection of excess memory. Same as SIGUSR2.
  • IOCTL_GC1: Triggers a simple garbage collection of excess memory. Same as what happens every 15 minutes normally.
  • IOCTL_INVALIDATE_ALL_NODES: Same as SIGUSR1. Send invalidation notifications to the kernel for all files causing unused files to be released from memory.

TOOLING

preload.so

EXPERIMENTAL

For some time there has been work to enable passthrough IO in FUSE. Passthrough IO would allow for near native performance with regards to reads and writes (at the expense of certain mergerfs features.) However, there have been several complications which have kept the feature from making it into the mainline Linux kernel. Until that feature is available there are two methods to provide similar functionality. One method is using the LD_PRELOAD feature of the dynamic linker. The other leveraging ptrace to intercept syscalls. Each has their disadvantages. At the moment only a preload based tool is available. A ptrace based tool may be developed later if there is a need.

/usr/lib/mergerfs/preload.so

This preloadable library overrides the creation and opening of files in order to simulate passthrough file IO. It catches the open/creat/fopen calls, has mergerfs do the call, queries mergerfs for the branch the file exists on, reopens the file on the underlying filesystem and returns that instead. Meaning that you will get native read/write performance because mergerfs is no longer part of the workflow. Keep in mind that this also means certain mergerfs features that work by interrupting the read/write workflow, such as moveonenospc, will no longer work.

Also understand that this will only work on dynamically linked software. Anything statically compiled will not work. Many GoLang and Rust apps are statically compiled.

The library will not interfere with non-mergerfs filesystems. The library is written to always fallback to returning the mergerfs opened file on error.

While the library was written to account for a number of edgecases there could be some yet accounted for so please report any oddities.

Thank you to nohajc for prototyping the idea.

general usage

LD_PRELOAD=/usr/lib/mergerfs/preload.so touch /mnt/mergerfs/filename

Docker usage

Assume /mnt/fs0 and /mnt/fs1 are pooled with mergerfs at /media.

All mergerfs branch paths must be bind mounted into the container at the same path as found on the host so the preload library can see them.

docker run \
  -e LD_PRELOAD=/usr/lib/mergerfs/preload.so \
  -v /usr/lib/mergerfs/preload.so:/usr/lib/mergerfs/preload.so:ro \
  -v /media:/data \
  -v /mnt:/mnt \
  ubuntu:latest \
  bash

or more explicitly

docker run \
  -e LD_PRELOAD=/usr/lib/mergerfs/preload.so \
  -v /usr/lib/mergerfs/preload.so:/usr/lib/mergerfs/preload.so:ro \
  -v /media:/data \
  -v /mnt/fs0:/mnt/fs0 \
  -v /mnt/fs1:/mnt/fs1 \
  ubuntu:latest \
  bash

systemd unit

Use the Environment option to set the LD_PRELOAD variable.

[Service]
Environment=LD_PRELOAD=/usr/lib/mergerfs/preload.so

Misc

  • https://github.com/trapexit/mergerfs-tools
    • mergerfs.ctl: A tool to make it easier to query and configure mergerfs at runtime
    • mergerfs.fsck: Provides permissions and ownership auditing and the ability to fix them
    • mergerfs.dedup: Will help identify and optionally remove duplicate files
    • mergerfs.dup: Ensure there are at least N copies of a file across the pool
    • mergerfs.balance: Rebalance files across filesystems by moving them from the most filled to the least filled
    • mergerfs.consolidate: move files within a single mergerfs directory to the filesystem with most free space
  • https://github.com/trapexit/scorch
    • scorch: A tool to help discover silent corruption of files and keep track of files
  • https://github.com/trapexit/bbf
    • bbf (bad block finder): a tool to scan for and 'fix' hard drive bad blocks and find the files using those blocks

CACHING

page caching

https://en.wikipedia.org/wiki/Page_cache

  • cache.files=off: Disables page caching. Underlying files cached, mergerfs files are not.
  • cache.files=partial: Enables page caching. Underlying files cached, mergerfs files cached while open.
  • cache.files=full: Enables page caching. Underlying files cached, mergerfs files cached across opens.
  • cache.files=auto-full: Enables page caching. Underlying files cached, mergerfs files cached across opens if mtime and size are unchanged since previous open.
  • cache.files=libfuse: follow traditional libfuse direct_io, kernel_cache, and auto_cache arguments.
  • cache.files=per-process: Enable page caching (equivalent to cache.files=partial) only for processes whose 'comm' name matches one of the values defined in cache.files.process-names. If the name does not match the file open is equivalent to cache.files=off.

FUSE, which mergerfs uses, offers a number of page caching modes. mergerfs tries to simplify their use via the cache.files option. It can and should replace usage of direct_io, kernel_cache, and auto_cache.

Due to mergerfs using FUSE and therefore being a userland process proxying existing filesystems the kernel will double cache the content being read and written through mergerfs. Once from the underlying filesystem and once from mergerfs (it sees them as two separate entities). Using cache.files=off will keep the double caching from happening by disabling caching of mergerfs but this has the side effect that all read and write calls will be passed to mergerfs which may be slower than enabling caching, you lose shared mmap support which can affect apps such as rtorrent, and no read-ahead will take place. The kernel will still cache the underlying filesystem data but that only helps so much given mergerfs will still process all requests.

If you do enable file page caching, cache.files=partial|full|auto-full, you should also enable dropcacheonclose which will cause mergerfs to instruct the kernel to flush the underlying file's page cache when the file is closed. This behavior is the same as the rsync fadvise / drop cache patch and Feh's nocache project.

If most files are read once through and closed (like media) it is best to enable dropcacheonclose regardless of caching mode in order to minimize buffer bloat.

It is difficult to balance memory usage, cache bloat & duplication, and performance. Ideally mergerfs would be able to disable caching for the files it reads/writes but allow page caching for itself. That would limit the FUSE overhead. However, there isn't a good way to achieve this. It would need to open all files with O_DIRECT which places limitations on the what underlying filesystems would be supported and complicates the code.

kernel documentation: https://www.kernel.org/doc/Documentation/filesystems/fuse-io.txt

entry & attribute caching

Given the relatively high cost of FUSE due to the kernel <-> userspace round trips there are kernel side caches for file entries and attributes. The entry cache limits the lookup calls to mergerfs which ask if a file exists. The attribute cache limits the need to make getattr calls to mergerfs which provide file attributes (mode, size, type, etc.). As with the page cache these should not be used if the underlying filesystems are being manipulated at the same time as it could lead to odd behavior or data corruption. The options for setting these are cache.entry and cache.negative_entry for the entry cache and cache.attr for the attributes cache. cache.negative_entry refers to the timeout for negative responses to lookups (non-existent files).

writeback caching

When cache.files is enabled the default is for it to perform writethrough caching. This behavior won't help improve performance as each write still goes one for one through the filesystem. By enabling the FUSE writeback cache small writes may be aggregated by the kernel and then sent to mergerfs as one larger request. This can greatly improve the throughput for apps which write to files inefficiently. The amount the kernel can aggregate is limited by the size of a FUSE message. Read the fuse_msg_size section for more details.

There is a small side effect as a result of enabling writeback caching. Underlying files won't ever be opened with O_APPEND or O_WRONLY. The former because the kernel then manages append mode and the latter because the kernel may request file data from mergerfs to populate the write cache. The O_APPEND change means that if a file is changed outside of mergerfs it could lead to corruption as the kernel won't know the end of the file has changed. That said any time you use caching you should keep from using the same file outside of mergerfs at the same time.

Note that if an application is properly sizing writes then writeback caching will have little or no effect. It will only help with writes of sizes below the FUSE message size (128K on older kernels, 1M on newer).

statfs caching

Of the syscalls used by mergerfs in policies the statfs / statvfs call is perhaps the most expensive. It's used to find out the available space of a filesystem and whether it is mounted read-only. Depending on the setup and usage pattern these queries can be relatively costly. When cache.statfs is enabled all calls to statfs by a policy will be cached for the number of seconds its set to.

Example: If the create policy is mfs and the timeout is 60 then for that 60 seconds the same filesystem will be returned as the target for creates because the available space won't be updated for that time.

As of version 4.20 Linux supports symlink caching. Significant performance increases can be had in workloads which use a lot of symlinks. Setting cache.symlinks=true will result in requesting symlink caching from the kernel only if supported. As a result its safe to enable it on systems prior to 4.20. That said it is disabled by default for now. You can see if caching is enabled by querying the xattr user.mergerfs.cache.symlinks but given it must be requested at startup you can not change it at runtime.

readdir caching

As of version 4.20 Linux supports readdir caching. This can have a significant impact on directory traversal. Especially when combined with entry (cache.entry) and attribute (cache.attr) caching. Setting cache.readdir=true will result in requesting readdir caching from the kernel on each opendir. If the kernel doesn't support readdir caching setting the option to true has no effect. This option is configurable at runtime via xattr user.mergerfs.cache.readdir.

tiered caching

Some storage technologies support what some call "tiered" caching. The placing of usually smaller, faster storage as a transparent cache to larger, slower storage. NVMe, SSD, Optane in front of traditional HDDs for instance.

mergerfs does not natively support any sort of tiered caching. Most users have no use for such a feature and its inclusion would complicate the code. However, there are a few situations where a cache filesystem could help with a typical mergerfs setup.

  1. Fast network, slow filesystems, many readers: You've a 10+Gbps network with many readers and your regular filesystems can't keep up.
  2. Fast network, slow filesystems, small'ish bursty writes: You have a 10+Gbps network and wish to transfer amounts of data less than your cache filesystem but wish to do so quickly.

With #1 it's arguable if you should be using mergerfs at all. RAID would probably be the better solution. If you're going to use mergerfs there are other tactics that may help: spreading the data across filesystems (see the mergerfs.dup tool) and setting func.open=rand, using symlinkify, or using dm-cache or a similar technology to add tiered cache to the underlying device.

With #2 one could use dm-cache as well but there is another solution which requires only mergerfs and a cronjob.

  1. Create 2 mergerfs pools. One which includes just the slow devices and one which has both the fast devices (SSD,NVME,etc.) and slow devices.
  2. The 'cache' pool should have the cache filesystems listed first.
  3. The best create policies to use for the 'cache' pool would probably be ff, epff, lfs, or eplfs. The latter two under the assumption that the cache filesystem(s) are far smaller than the backing filesystems. If using path preserving policies remember that you'll need to manually create the core directories of those paths you wish to be cached. Be sure the permissions are in sync. Use mergerfs.fsck to check / correct them. You could also set the slow filesystems mode to NC though that'd mean if the cache filesystems fill you'd get "out of space" errors.
  4. Enable moveonenospc and set minfreespace appropriately. To make sure there is enough room on the "slow" pool you might want to set minfreespace to at least as large as the size of the largest cache filesystem if not larger. This way in the worst case the whole of the cache filesystem(s) can be moved to the other drives.
  5. Set your programs to use the cache pool.
  6. Save one of the below scripts or create you're own.
  7. Use cron (as root) to schedule the command at whatever frequency is appropriate for your workflow.
time based expiring

Move files from cache to backing pool based only on the last time the file was accessed. Replace -atime with -amin if you want minutes rather than days. May want to use the fadvise / --drop-cache version of rsync or run rsync with the tool "nocache".

NOTE: The arguments to these scripts include the cache filesystem itself. Not the pool with the cache filesystem. You could have data loss if the source is the cache pool.

mergerfs.time-based-mover

percentage full expiring

Move the oldest file from the cache to the backing pool. Continue till below percentage threshold.

NOTE: The arguments to these scripts include the cache filesystem itself. Not the pool with the cache filesystem. You could have data loss if the source is the cache pool.

mergerfs.percent-full-mover

PERFORMANCE

mergerfs is at its core just a proxy and therefore its theoretical max performance is that of the underlying devices. However, given it is a FUSE filesystem working from userspace there is an increase in overhead relative to kernel based solutions. That said the performance can match the theoretical max but it depends greatly on the system's configuration. Especially when adding network filesystems into the mix there are many variables which can impact performance. Device speeds and latency, network speeds and latency, general concurrency, read/write sizes, etc. Unfortunately, given the number of variables it has been difficult to find a single set of settings which provide optimal performance. If you're having performance issues please look over the suggestions below (including the benchmarking section.)

NOTE: be sure to read about these features before changing them to understand what behaviors it may impact

  • disable security_capability and/or xattr
  • increase cache timeouts cache.attr, cache.entry, cache.negative_entry
  • enable (or disable) page caching (cache.files)
  • enable parallel-direct-writes
  • enable cache.writeback
  • enable cache.statfs
  • enable cache.symlinks
  • enable cache.readdir
  • change the number of worker threads
  • disable posix_acl
  • disable async_read
  • test theoretical performance using nullrw or mounting a ram disk
  • use symlinkify if your data is largely static and read-only
  • use tiered cache devices
  • use LVM and LVM cache to place a SSD in front of your HDDs
  • increase readahead: readahead=1024

If you come across a setting that significantly impacts performance please contact trapexit so he may investigate further. Please test both against your normal setup, a singular branch, and with nullrw=true

BENCHMARKING

Filesystems are complicated. They do many things and many of those are interconnected. Additionally, the OS, drivers, hardware, etc. all can impact performance. Therefore, when benchmarking, it is necessary that the test focus as narrowly as possible.

For most throughput is the key benchmark. To test throughput dd is useful but must be used with the correct settings in order to ensure the filesystem or device is actually being tested. The OS can and will cache data. Without forcing synchronous reads and writes and/or disabling caching the values returned will not be representative of the device's true performance.

When benchmarking through mergerfs ensure you only use 1 branch to remove any possibility of the policies complicating the situation. Benchmark the underlying filesystem first and then mount mergerfs over it and test again. If you're experience speeds below your expectation you will need to narrow down precisely which component is leading to the slowdown. Preferably test the following in the order listed (but not combined).

  1. Enable nullrw mode with nullrw=true. This will effectively make reads and writes no-ops. Removing the underlying device / filesystem from the equation. This will give us the top theoretical speeds.
  2. Mount mergerfs over tmpfs. tmpfs is a RAM disk. Extremely high speed and very low latency. This is a more realistic best case scenario. Example: mount -t tmpfs -o size=2G tmpfs /tmp/tmpfs
  3. Mount mergerfs over a local device. NVMe, SSD, HDD, etc. If you have more than one I'd suggest testing each of them as drives and/or controllers (their drivers) could impact performance.
  4. Finally, if you intend to use mergerfs with a network filesystem, either as the source of data or to combine with another through mergerfs, test each of those alone as above.

Once you find the component which has the performance issue you can do further testing with different options to see if they impact performance. For reads and writes the most relevant would be: cache.files, async_read. Less likely but relevant when using NFS or with certain filesystems would be security_capability, xattr, and posix_acl. If you find a specific system, device, filesystem, controller, etc. that performs poorly contact trapexit so he may investigate further.

Sometimes the problem is really the application accessing or writing data through mergerfs. Some software use small buffer sizes which can lead to more requests and therefore greater overhead. You can test this out yourself by replace bs=1M in the examples below with ibs or obs and using a size of 512 instead of 1M. In one example test using nullrw the write speed dropped from 4.9GB/s to 69.7MB/s when moving from 1M to 512. Similar results were had when testing reads. Small writes overhead may be improved by leveraging a write cache but in casual tests little gain was found. More tests will need to be done before this feature would become available. If you have an app that appears slow with mergerfs it could be due to this. Contact trapexit so he may investigate further.

write benchmark

$ dd if=/dev/zero of=/mnt/mergerfs/1GB.file bs=1M count=1024 oflag=nocache conv=fdatasync status=progress

read benchmark

$ dd if=/mnt/mergerfs/1GB.file of=/dev/null bs=1M count=1024 iflag=nocache conv=fdatasync status=progress

other benchmarks

If you are attempting to benchmark other behaviors you must ensure you clear kernel caches before runs. In fact it would be a good deal to run before the read and write benchmarks as well just in case.

sync
echo 3 | sudo tee /proc/sys/vm/drop_caches

TIPS / NOTES

  • This document is literal and thorough. If a suspected feature isn't mentioned it doesn't exist. If certain libfuse arguments aren't listed they probably shouldn't be used.
  • Ensure you're using the latest version.
  • Run mergerfs as root. mergerfs is designed and intended to be run as root and may exibit incorrect behavior if run otherwise..
  • If you don't see some directories and files you expect, policies seem to skip branches, you get strange permission errors, etc. be sure the underlying filesystems' permissions are all the same. Use mergerfs.fsck to audit the filesystem for out of sync permissions.
  • If you still have permission issues be sure you are using POSIX ACL compliant filesystems. mergerfs doesn't generally make exceptions for FAT, NTFS, or other non-POSIX filesystem.
  • Do not use cache.files=off if you expect applications (such as rtorrent) to use mmap files. Shared mmap is not currently supported in FUSE w/ page caching disabled. Enabling dropcacheonclose is recommended when cache.files=partial|full|auto-full.
  • Kodi, Plex, Subsonic, etc. can use directory mtime to more efficiently determine whether to scan for new content rather than simply performing a full scan. If using the default getattr policy of ff it's possible those programs will miss an update on account of it returning the first directory found's stat info and it's a later directory on another mount which had the mtime recently updated. To fix this you will want to set func.getattr=newest. Remember though that this is just stat. If the file is later open'ed or unlink'ed and the policy is different for those then a completely different file or directory could be acted on.
  • Some policies mixed with some functions may result in strange behaviors. Not that some of these behaviors and race conditions couldn't happen outside mergerfs but that they are far more likely to occur on account of the attempt to merge together multiple sources of data which could be out of sync due to the different policies.
  • For consistency its generally best to set category wide policies rather than individual func's. This will help limit the confusion of tools such as rsync. However, the flexibility is there if needed.

KNOWN ISSUES / BUGS

kernel issues & bugs

https://github.com/trapexit/mergerfs/wiki/Kernel-Issues-&-Bugs

directory mtime is not being updated

Remember that the default policy for getattr is ff. The information for the first directory found will be returned. If it wasn't the directory which had been updated then it will appear outdated.

The reason this is the default is because any other policy would be more expensive and for many applications it is unnecessary. To always return the directory with the most recent mtime or a faked value based on all found would require a scan of all filesystems.

If you always want the directory information from the one with the most recent mtime then use the newest policy for getattr.

'mv /mnt/pool/foo /mnt/disk1/foo' removes 'foo'

This is not a bug.

Run in verbose mode to better understand what's happening:

$ mv -v /mnt/pool/foo /mnt/disk1/foo
copied '/mnt/pool/foo' -> '/mnt/disk1/foo'
removed '/mnt/pool/foo'
$ ls /mnt/pool/foo
ls: cannot access '/mnt/pool/foo': No such file or directory

mv, when working across devices, is copying the source to target and then removing the source. Since the source is the target in this case, depending on the unlink policy, it will remove the just copied file and other files across the branches.

If you want to move files to one filesystem just copy them there and use mergerfs.dedup to clean up the old paths or manually remove them from the branches directly.

cached memory appears greater than it should be

Use cache.files=off and/or dropcacheonclose=true. See the section on page caching.

NFS clients returning ESTALE / Stale file handle

NFS generally does not like out of band changes. Take a look at the section on NFS in the #remote-filesystems for more details.

rtorrent fails with ENODEV (No such device)

Be sure to set cache.files=partial|full|auto-full|per-processe. rtorrent and some other applications use mmap to read and write to files and offer no fallback to traditional methods. FUSE does not currently support mmap while using direct_io. There may be a performance penalty on writes with direct_io off as well as the problem of double caching but it's the only way to get such applications to work. If the performance loss is too high for other apps you can mount mergerfs twice. Once with direct_io enabled and one without it. Be sure to set dropcacheonclose=true if not using direct_io.

Plex doesn't work with mergerfs

It does. If you're trying to put Plex's config / metadata / database on mergerfs you can't set cache.files=off because Plex is using sqlite3 with mmap enabled. Shared mmap is not supported by Linux's FUSE implementation when page caching is disabled. To fix this place the data elsewhere (preferable) or enable cache.files (with dropcacheonclose=true). Sqlite3 does not need mmap but the developer needs to fall back to standard IO if mmap fails.

This applies to other software: Radarr, Sonarr, Lidarr, Jellyfin, etc.

I would recommend reaching out to the developers of the software you're having troubles with and asking them to add a fallback to regular file IO when mmap is unavailable.

If the issue is that scanning doesn't seem to pick up media then be sure to set func.getattr=newest though generally a full scan will pick up all media anyway.

When a program tries to move or rename a file it fails

Please read the section above regarding rename & link.

The problem is that many applications do not properly handle EXDEV errors which rename and link may return even though they are perfectly valid situations which do not indicate actual device, filesystem, or OS errors. The error will only be returned by mergerfs if using a path preserving policy as described in the policy section above. If you do not care about path preservation simply change the mergerfs policy to the non-path preserving version. For example: -o category.create=mfs Ideally the offending software would be fixed and it is recommended that if you run into this problem you contact the software's author and request proper handling of EXDEV errors.

my 32bit software has problems

Some software have problems with 64bit inode values. The symptoms can include EOVERFLOW errors when trying to list files. You can address this by setting inodecalc to one of the 32bit based algos as described in the relevant section.

Samba: Moving files / directories fails

Workaround: Copy the file/directory and then remove the original rather than move.

This isn't an issue with Samba but some SMB clients. GVFS-fuse v1.20.3 and prior (found in Ubuntu 14.04 among others) failed to handle certain error codes correctly. Particularly STATUS_NOT_SAME_DEVICE which comes from the EXDEV which is returned by rename when the call is crossing mount points. When a program gets an EXDEV it needs to explicitly take an alternate action to accomplish its goal. In the case of mv or similar it tries rename and on EXDEV falls back to a manual copying of data between the two locations and unlinking the source. In these older versions of GVFS-fuse if it received EXDEV it would translate that into EIO. This would cause mv or most any application attempting to move files around on that SMB share to fail with a IO error.

GVFS-fuse v1.22.0 and above fixed this issue but a large number of systems use the older release. On Ubuntu the version can be checked by issuing apt-cache showpkg gvfs-fuse. Most distros released in 2015 seem to have the updated release and will work fine but older systems may not. Upgrading gvfs-fuse or the distro in general will address the problem.

In Apple's MacOSX 10.9 they replaced Samba (client and server) with their own product. It appears their new client does not handle EXDEV either and responds similar to older release of gvfs on Linux.

Trashing files occasionally fails

This is the same issue as with Samba. rename returns EXDEV (in our case that will really only happen with path preserving policies like epmfs) and the software doesn't handle the situation well. This is unfortunately a common failure of software which moves files around. The standard indicates that an implementation MAY choose to support non-user home directory trashing of files (which is a MUST). The implementation MAY also support "top directory trashes" which many probably do.

To create a $topdir/.Trash directory as defined in the standard use the mergerfs-tools tool mergerfs.mktrash.

Supplemental user groups

Due to the overhead of getgroups/setgroups mergerfs utilizes a cache. This cache is opportunistic and per thread. Each thread will query the supplemental groups for a user when that particular thread needs to change credentials and will keep that data for the lifetime of the thread. This means that if a user is added to a group it may not be picked up without the restart of mergerfs. However, since the high level FUSE API's (at least the standard version) thread pool dynamically grows and shrinks it's possible that over time a thread will be killed and later a new thread with no cache will start and query the new data.

The gid cache uses fixed storage to simplify the design and be compatible with older systems which may not have C++11 compilers. There is enough storage for 256 users' supplemental groups. Each user is allowed up to 32 supplemental groups. Linux >= 2.6.3 allows up to 65535 groups per user but most other *nixs allow far less. NFS allowing only 16. The system does handle overflow gracefully. If the user has more than 32 supplemental groups only the first 32 will be used. If more than 256 users are using the system when an uncached user is found it will evict an existing user's cache at random. So long as there aren't more than 256 active users this should be fine. If either value is too low for your needs you will have to modify gidcache.hpp to increase the values. Note that doing so will increase the memory needed by each thread.

While not a bug some users have found when using containers that supplemental groups defined inside the container don't work properly with regard to permissions. This is expected as mergerfs lives outside the container and therefore is querying the host's group database. There might be a hack to work around this (make mergerfs read the /etc/group file in the container) but it is not yet implemented and would be limited to Linux and the /etc/group DB. Preferably users would mount in the host group file into the containers or use a standard shared user & groups technology like NIS or LDAP.

Remote Filesystems

Many users ask about compatibility with remote filesystems. This section is to describe any known issues or quirks when using mergerfs with common remote filesystems.

Keep in mind that, like with caching, it is not a good idea to change the contents of the remote filesystem out-of-band. Meaning that you really shouldn't change the contents of the underlying filesystems or mergerfs on the server hosting the remote filesystem. Doing so can lead to weird behavior, inconsistency, errors, and even data corruption should multiple programs try to write or read the same data at the same time. This isn't to say you can't do it or that data corruption is likely but it could happen. It is better to always use the remote filesystem. Even on the machine serving it.

NFS

NFS is a common remote filesystem on Unix/POSIX systems. Due to how NFS works there are some settings which need to be set in order for mergerfs to work with it.

It should be noted that NFS and FUSE (the technology mergerfs uses) do not work perfectly with one another due to certain design choices in FUSE (and mergerfs.) Due to these issues it is generally recommended to use SMB when possible till situations change. That said mergerfs should generally work as an export of NFS and issues discovered should still be reported.

To ensure compatibility between mergerfs and NFS use the following settings.

mergerfs settings:

  • noforget
  • inodecalc=path-hash

NFS export settings:

  • fsid=UUID
  • no_root_squash

noforget is needed because NFS uses the name_to_handle_at and open_by_handle_at functions which allow a program to keep a reference to a file without technically having it open in the typical sense. The problem is that FUSE has no way to know that NFS has a handle that it will later use to open the file again. As a result it is possible for the kernel to tell mergerfs to forget about the node and should NFS ever ask for that node's details in the future it would have nothing to respond with. Keeping nodes around forever is not ideal but at the moment the only way to manage the situation.

inodecalc=path-hash is needed because NFS is sensitive to out-of-band changes. FUSE doesn't care if a file's inode value changes but NFS, being stateful, does. So if you used the default inode calculation algorithm then it is possible that if you changed a file or updated a directory the file mergerfs will use will be on a different branch and therefore the inode would change. This isn't an ideal solution and others are being considered but it works for most situations.

fsid=UUID is needed because FUSE filesystems don't have different st_dev values which can cause issues when exporting. The easiest thing to do is set each mergerfs export fsid to some random value. An easy way to generate a random value is to use the command line tool uuid or uuidgen or through a website such as uuidgenerator.net.

no_root_squash is not strictly necessary but can lead to confusing permission and ownership issues if root squashing is enabled.

SMB / CIFS

SMB is a protocol most used by Microsoft Windows systems to share file shares, printers, etc. However, due to the popularity for Windows, it is also supported on many other platforms including Linux. The most popular way of supporting SMB on Linux is via the software Samba.

Samba, and other ways of serving Linux filesystems, via SMB should work fine with mergerfs. The services do not tend to use the same technologies which NFS uses and therefore don't have the same issues. There should not be an special settings required to use mergerfs with Samba. However, CIFSD and other programs have not been extensively tested. If you use mergerfs with CIFSD or other SMB servers please submit your experiences so these docs can be updated.

SSHFS

SSHFS is a FUSE filesystem leveraging SSH as the connection and transport layer. While often simpler to setup when compared to NFS or Samba the performance can be lacking and the project is very much in maintenance mode.

There are no known issues using sshfs with mergerfs. You may want to use the following arguments to improve performance but your millage may vary.

  • -o Ciphers=arcfour
  • -o Compression=no

More info can be found here.

Other

There are other remote filesystems but none popularly used to serve mergerfs. If you use something not listed above feel free to reach out and I will add it to the list.

FAQ

How well does mergerfs scale? Is it "production ready?"

Users have reported running mergerfs on everything from a Raspberry Pi to dual socket Xeon systems with >20 cores. I'm aware of at least a few companies which use mergerfs in production. Open Media Vault includes mergerfs as its sole solution for pooling filesystems. The author of mergerfs had it running for over 300 days managing 16+ devices with reasonably heavy 24/7 read and write usage. Stopping only after the machine's power supply died.

Most serious issues (crashes or data corruption) have been due to kernel bugs. All of which are fixed in stable releases.

Can mergerfs be used with filesystems which already have data / are in use?

Yes. mergerfs is really just a proxy and does NOT interfere with the normal form or function of the filesystems / mounts / paths it manages. It is just another userland application that is acting as a man-in-the-middle. It can't do anything that any other random piece of software can't do.

mergerfs is not a traditional filesystem that takes control over the underlying block device. mergerfs is not RAID. It does not manipulate the data that passes through it. It does not shard data across filesystems. It merely shards some behavior and aggregates others.

Can drives/filesystems be removed from the pool at will?

Yes. See previous question's answer.

Can mergerfs be removed without affecting the data?

Yes. See the previous question's answer.

Can drives/filesystems be moved to another pool?

Yes. See the previous question's answer.

How do I migrate data into or out of the pool when adding/removing drives/filesystems?

You don't need to. See the previous question's answer.

How do I remove a drive/filesystem but keep the data in the pool?

Nothing special needs to be done. Remove the branch from mergerfs' config and copy (rsync) the data from the removed filesystem into the pool. Effectively the same as if it were you transfering data from one filesystem to another.

If you wish to continue using the pool while performing the transfer simply create another, temporary pool without the filesystem in question and then copy the data. It would probably be a good idea to set the branch to RO prior to doing this to ensure no new content is written to the filesystem while performing the copy.

What policies should I use?

Unless you're doing something more niche the average user is probably best off using mfs for category.create. It will spread files out across your branches based on available space. Use mspmfs if you want to try to colocate the data a bit more. You may want to use lus if you prefer a slightly different distribution of data if you have a mix of smaller and larger filesystems. Generally though mfs, lus, or even rand are good for the general use case. If you are starting with an imbalanced pool you can use the tool mergerfs.balance to redistribute files across the pool.

If you really wish to try to colocate files based on directory you can set func.create to epmfs or similar and func.mkdir to rand or eprand depending on if you just want to colocate generally or on specific branches. Either way the need to colocate is rare. For instance: if you wish to remove the device regularly and want the data to predictably be on that device or if you don't use backup at all and don't wish to replace that data piecemeal. In which case using path preservation can help but will require some manual attention. Colocating after the fact can be accomplished using the mergerfs.consolidate tool. If you don't need strict colocation which the ep policies provide then you can use the msp based policies which will walk back the path till finding a branch that works.

Ultimately there is no correct answer. It is a preference or based on some particular need. mergerfs is very easy to test and experiment with. I suggest creating a test setup and experimenting to get a sense of what you want.

epmfs is the default category.create policy because ep policies are not going to change the general layout of the branches. It won't place files/dirs on branches that don't already have the relative branch. So it keeps the system in a known state. It's much easier to stop using epmfs or redistribute files around the filesystem than it is to consolidate them back.

What settings should I use?

Depends on what features you want. Generally speaking there are no "wrong" settings. All settings are performance or feature related. The best bet is to read over the available options and choose what fits your situation. If something isn't clear from the documentation please reach out and the documentation will be improved.

That said, for the average person, the following should be fine:

cache.files=off,dropcacheonclose=true,category.create=mfs

Why are all my files ending up on 1 filesystem?!

Did you start with empty filesystems? Did you explicitly configure a category.create policy? Are you using an existing path / path preserving policy?

The default create policy is epmfs. That is a path preserving algorithm. With such a policy for mkdir and create with a set of empty filesystems it will select only 1 filesystem when the first directory is created. Anything, files or directories, created in that first directory will be placed on the same branch because it is preserving paths.

This catches a lot of new users off guard but changing the default would break the setup for many existing users and this policy is the safest policy as it will not change the general layout of the existing filesystems. If you do not care about path preservation and wish your files to be spread across all your filesystems change to mfs or similar policy as described above. If you do want path preservation you'll need to perform the manual act of creating paths on the filesystems you want the data to land on before transferring your data. Setting func.mkdir=epall can simplify managing path preservation for create. Or use func.mkdir=rand if you're interested in just grouping together directory content by filesystem.

Yes. See also the option inodecalc for how inode values are calculated.

What mergerfs does not do is fake hard links across branches. Read the section "rename & link" for how it works.

Remember that hardlinks will NOT work across devices. That includes between the original filesystem and a mergerfs pool, between two separate pools of the same underlying filesystems, or bind mounts of paths within the mergerfs pool. The latter is common when using Docker or Podman. Multiple volumes (bind mounts) to the same underlying filesystem are considered different devices. There is no way to link between them. You should mount in the highest directory in the mergerfs pool that includes all the paths you need if you want links to work.

Does FICLONE or FICLONERANGE work?

Unfortunately not. FUSE, the technology mergerfs is based on, does not support the clone_file_range feature needed for it to work. mergerfs won't even know such a request is made. The kernel will simply return an error back to the application making the request.

Should FUSE gain the ability mergerfs will be updated to support it.

Can I use mergerfs without SnapRAID? SnapRAID without mergerfs?

Yes. They are completely unrelated pieces of software.

Can mergerfs run via Docker, Podman, Kubernetes, etc.

Yes. With Docker you'll need to include --cap-add=SYS_ADMIN --device=/dev/fuse --security-opt=apparmor:unconfined or similar with other container runtimes. You should also be running it as root or given sufficient caps to change user and group identity as well as have root like filesystem permissions.

Keep in mind that you MUST consider identity when using containers. For example: supplemental groups will be picked up from the container unless you properly manage users and groups by sharing relevant /etc files or by using some other means to share identity across containers. Similarly if you use "rootless" containers and user namespaces to do uid/gid translations you MUST consider that while managing shared files.

Also, as mentioned by hotio, with Docker you should probably be mounting with bind-propagation set to slave.

Does mergerfs support CoW / copy-on-write / writes to read-only filesystems?

Not in the sense of a filesystem like BTRFS or ZFS nor in the overlayfs or aufs sense. It does offer a cow-shell like hard link breaking (copy to temp file then rename over original) which can be useful when wanting to save space by hardlinking duplicate files but wish to treat each name as if it were a unique and separate file.

If you want to write to a read-only filesystem you should look at overlayfs. You can always include the overlayfs mount into a mergerfs pool.

Why can't I see my files / directories?

It's almost always a permissions issue. Unlike mhddfs and unionfs-fuse, which runs as root and attempts to access content as such, mergerfs always changes its credentials to that of the caller. This means that if the user does not have access to a file or directory than neither will mergerfs. However, because mergerfs is creating a union of paths it may be able to read some files and directories on one filesystem but not another resulting in an incomplete set.

Whenever you run into a split permission issue (seeing some but not all files) try using mergerfs.fsck tool to check for and fix the mismatch. If you aren't seeing anything at all be sure that the basic permissions are correct. The user and group values are correct and that directories have their executable bit set. A common mistake by users new to Linux is to chmod -R 644 when they should have chmod -R u=rwX,go=rX.

If using a network filesystem such as NFS or SMB (Samba) be sure to pay close attention to anything regarding permissioning and users. Root squashing and user translation for instance has bitten a few mergerfs users. Some of these also affect the use of mergerfs from container platforms such as Docker.

Why use FUSE? Why not a kernel based solution?

As with any solutions to a problem there are advantages and disadvantages to each one.

A FUSE based solution has all the downsides of FUSE:

  • Higher IO latency due to the trips in and out of kernel space
  • Higher general overhead due to trips in and out of kernel space
  • Double caching when using page caching
  • Misc limitations due to FUSE's design

But FUSE also has a lot of upsides:

  • Easier to offer a cross platform solution
  • Easier forward and backward compatibility
  • Easier updates for users
  • Easier and faster release cadence
  • Allows more flexibility in design and features
  • Overall easier to write, secure, and maintain
  • Much lower barrier to entry (getting code into the kernel takes a lot of time and effort initially)

FUSE was chosen because of all the advantages listed above. The negatives of FUSE do not outweigh the positives.

Is my OS's libfuse needed for mergerfs to work?

No. Normally mount.fuse is needed to get mergerfs (or any FUSE filesystem to mount using the mount command but in vendoring the libfuse library the mount.fuse app has been renamed to mount.mergerfs meaning the filesystem type in fstab can simply be mergerfs. That said there should be no harm in having it installed and continuing to using fuse.mergerfs as the type in /etc/fstab.

If mergerfs doesn't work as a type it could be due to how the mount.mergerfs tool was installed. Must be in /sbin/ with proper permissions.

Why was splice support removed?

After a lot of testing over the years splicing always appeared to be at best provide equivalent performance and in cases worse performance. Splice is not supported on other platforms forcing a traditional read/write fallback to be provided. The splice code was removed to simplify the codebase.

What should mergerfs NOT be used for?

  • databases: Even if the database stored data in separate files (mergerfs wouldn't offer much otherwise) the higher latency of the indirection will kill performance. If it is a lightly used SQLITE database then it may be fine but you'll need to test.
  • VM images: For the same reasons as databases. VM images are accessed very aggressively and mergerfs will introduce too much latency (if it works at all).
  • As replacement for RAID: mergerfs is just for pooling branches. If you need that kind of device performance aggregation or high availability you should stick with RAID.

Can filesystems be written to directly? Outside of mergerfs while pooled?

Yes, however it's not recommended to use the same file from within the pool and from without at the same time (particularly writing). Especially if using caching of any kind (cache.files, cache.entry, cache.attr, cache.negative_entry, cache.symlinks, cache.readdir, etc.) as there could be a conflict between cached data and not.

Why do I get an "out of space" / "no space left on device" / ENOSPC error even though there appears to be lots of space available?

First make sure you've read the sections above about policies, path preservation, branch filtering, and the options minfreespace, moveonenospc, statfs, and statfs_ignore.

mergerfs is simply presenting a union of the content within multiple branches. The reported free space is an aggregate of space available within the pool (behavior modified by statfs and statfs_ignore). It does not represent a contiguous space. In the same way that read-only filesystems, those with quotas, or reserved space report the full theoretical space available.

Due to path preservation, branch tagging, read-only status, and minfreespace settings it is perfectly valid that ENOSPC / "out of space" / "no space left on device" be returned. It is doing what was asked of it: filtering possible branches due to those settings. Only one error can be returned and if one of the reasons for filtering a branch was minfreespace then it will be returned as such. moveonenospc is only relevant to writing a file which is too large for the filesystem it's currently on.

It is also possible that the filesystem selected has run out of inodes. Use df -i to list the total and available inodes per filesystem.

If you don't care about path preservation then simply change the create policy to one which isn't. mfs is probably what most are looking for. The reason it's not default is because it was originally set to epmfs and changing it now would change people's setup. Such a setting change will likely occur in mergerfs 3.

Why does the total available space in mergerfs not equal outside?

Are you using ext2/3/4? With reserve for root? mergerfs uses available space for statfs calculations. If you've reserved space for root then it won't show up.

You can remove the reserve by running: tune2fs -m 0 <device>

I notice massive slowdowns of writes when enabling cache.files.

When file caching is enabled in any form (cache.files!=off) it will issue getxattr requests for security.capability prior to every single write. This will usually result in a performance degradation, especially when using a network filesystem (such as NFS or SMB.) Unfortunately at this moment the kernel is not caching the response.

To work around this situation mergerfs offers a few solutions.

  1. Set security_capability=false. It will short circuit any call and return ENOATTR. This still means though that mergerfs will receive the request before every write but at least it doesn't get passed through to the underlying filesystem.
  2. Set xattr=noattr. Same as above but applies to all calls to getxattr. Not just security.capability. This will not be cached by the kernel either but mergerfs' runtime config system will still function.
  3. Set xattr=nosys. Results in mergerfs returning ENOSYS which will be cached by the kernel. No future xattr calls will be forwarded to mergerfs. The downside is that also means the xattr based config and query functionality won't work either.
  4. Disable file caching. If you aren't using applications which use mmap it's probably simpler to just disable it all together. The kernel won't send the requests when caching is disabled.

It's mentioned that there are some security issues with mhddfs. What are they? How does mergerfs address them?

mhddfs manages running as root by calling getuid() and if it returns 0 then it will chown the file. Not only is that a race condition but it doesn't handle other situations. Rather than attempting to simulate POSIX ACL behavior the proper way to manage this is to use seteuid and setegid, in effect becoming the user making the original call, and perform the action as them. This is what mergerfs does and why mergerfs should always run as root.

In Linux setreuid syscalls apply only to the thread. GLIBC hides this away by using realtime signals to inform all threads to change credentials. Taking after Samba, mergerfs uses syscall(SYS_setreuid,...) to set the callers credentials for that thread only. Jumping back to root as necessary should escalated privileges be needed (for instance: to clone paths between filesystems).

For non-Linux systems mergerfs uses a read-write lock and changes credentials only when necessary. If multiple threads are to be user X then only the first one will need to change the processes credentials. So long as the other threads need to be user X they will take a readlock allowing multiple threads to share the credentials. Once a request comes in to run as user Y that thread will attempt a write lock and change to Y's credentials when it can. If the ability to give writers priority is supported then that flag will be used so threads trying to change credentials don't starve. This isn't the best solution but should work reasonably well assuming there are few users.

mergerfs versus X

mhddfs

mhddfs had not been maintained for some time and has some known stability and security issues. mergerfs provides a superset of mhddfs' features and should offer the same or better performance.

Below is an example of mhddfs and mergerfs setup to work similarly.

mhddfs -o mlimit=4G,allow_other /mnt/drive1,/mnt/drive2 /mnt/pool

mergerfs -o minfreespace=4G,category.create=ff /mnt/drive1:/mnt/drive2 /mnt/pool

aufs

aufs is mostly abandoned and no longer available in most Linux distros.

While aufs can offer better peak performance mergerfs provides more configurability and is generally easier to use. mergerfs however does not offer the overlay / copy-on-write (CoW) features which aufs has.

unionfs-fuse

unionfs-fuse is more like aufs than mergerfs in that it offers overlay / copy-on-write (CoW) features. If you're just looking to create a union of filesystems and want flexibility in file/directory placement then mergerfs offers that whereas unionfs is more for overlaying read/write filesystems over read-only ones.

overlayfs

overlayfs is similar to aufs and unionfs-fuse in that it also is primarily used to layer a read/write filesystem over one or more read-only filesystems. It does not have the ability to spread files/directories across numerous filesystems.

RAID0, JBOD, drive concatenation, striping

With simple JBOD / drive concatenation / stripping / RAID0 a single drive failure will result in full pool failure. mergerfs performs a similar function without the possibility of catastrophic failure and the difficulties in recovery. Drives may fail but all other filesystems and their data will continue to be accessible.

The main practical difference with mergerfs is the fact you don't actually have contiguous space as large as if you used those other technologies. Meaning you can't create a 2TB file on a pool of 2 1TB filesystems.

When combined with something like SnapRaid and/or an offsite backup solution you can have the flexibility of JBOD without the single point of failure.

UnRAID

UnRAID is a full OS and its storage layer, as I understand, is proprietary and closed source. Users who have experience with both have often said they prefer the flexibility offered by mergerfs and for some the fact it is open source is important.

There are a number of UnRAID users who use mergerfs as well though I'm not entirely familiar with the use case.

For semi-static data mergerfs + SnapRaid provides a similar solution.

ZFS

mergerfs is very different from ZFS. mergerfs is intended to provide flexible pooling of arbitrary filesystems (local or remote), of arbitrary sizes, and arbitrary filesystems. For write once, read many usecases such as bulk media storage. Where data integrity and backup is managed in other ways. In those usecases ZFS can introduce a number of costs and limitations as described here, here, and here.

StableBit's DrivePool

DrivePool works only on Windows so not as common an alternative as other Linux solutions. If you want to use Windows then DrivePool is a good option. Functionally the two projects work a bit differently. DrivePool always writes to the filesystem with the most free space and later rebalances. mergerfs does not offer rebalance but chooses a branch at file/directory create time. DrivePool's rebalancing can be done differently in any directory and has file pattern matching to further customize the behavior. mergerfs, not having rebalancing does not have these features, but similar features are planned for mergerfs v3. DrivePool has builtin file duplication which mergerfs does not natively support (but can be done via an external script.)

There are a lot of misc differences between the two projects but most features in DrivePool can be replicated with external tools in combination with mergerfs.

Additionally DrivePool is a closed source commercial product vs mergerfs a ISC licensed OSS project.

SUPPORT

Filesystems are complex and difficult to debug. mergerfs, while being just a proxy of sorts, can be difficult to debug given the large number of possible settings it can have itself and the number of environments it can run in. When reporting on a suspected issue please include as much of the below information as possible otherwise it will be difficult or impossible to diagnose. Also please read the above documentation as it provides details on many previously encountered questions/issues.

Please make sure you are using the latest release or have tried it in comparison. Old versions, which are often included in distros like Debian and Ubuntu, are not ever going to be updated and the issue you are encountering may have been addressed already.

For commercial support or feature requests please contact me directly.

Information to include in bug reports

  • Information about the broader problem along with any attempted solutions.
  • Solution already ruled out and why.
  • Version of mergerfs: mergerfs --version
  • mergerfs settings / arguments: from fstab, systemd unit, command line, OMV plugin, etc.
  • Version of the OS: uname -a and lsb_release -a
  • List of branches, their filesystem types, sizes (before and after issue): df -h
  • All information about the relevant paths and files: permissions, ownership, etc.
  • All information about the client app making the requests: version, uid/gid
  • Runtime environment:
    • Is mergerfs running within a container?
    • Are the client apps using mergerfs running in a container?
  • A strace of the app having problems:
    • strace -fvTtt -s 256 -o /tmp/app.strace.txt <cmd>
  • A strace of mergerfs while the program is trying to do whatever it is failing to do:
    • strace -fvTtt -s 256 -p <mergerfsPID> -o /tmp/mergerfs.strace.txt
  • Precise directions on replicating the issue. Do not leave anything out.
  • Try to recreate the problem in the simplest way using standard programs: ln, mv, cp, ls, dd, etc.

Contact / Issue submission

Donations

https://github.com/trapexit/support

Development and support of a project like mergerfs requires a significant amount of time and effort. The software is released under the very liberal ISC license and is therefore free to use for personal or commercial uses.

If you are a personal user and find mergerfs and its support valuable and would like to support the project financially it would be very much appreciated.

If you are using mergerfs commercially please consider sponsoring the project to ensure it continues to be maintained and receive updates. If custom features are needed feel free to contact me directly.

LINKS