find duplicate files and other space waste efficiently




rmlint finds space waste and other broken things on your filesystem.

Types of waste include: * Duplicate files and directories. * Nonstripped Binaries (Binaries with debug symbols). * Broken links. * Empty files and directories. * Files with broken user or group id.

rmlint will not delete any files. It does however produce executable output (for example a shell script) to help you delete the files if you want to.

In order to find the lint, rmlint is given one or more directories to traverse. If no directories or files were given, the current working directory is assumed. By default, rmlint will ignore hidden files and will not follow symlinks (see traversal options below). rmlint will first find "other lint" and then search the remaining files for duplicates.

Duplicate sets will be displayed as an original and one or more duplicates. You can set criteria for how rmlint chooses using the -S option (by default it chooses the first-named path on the command line, or if that is equal then the oldest file based on mtime). You can also specify that certain paths only contain originals by naming the path after the special path separator //.

Examples are given at the end of this manual.


General Options

-T --types="list" (default: defaults):
Configure the types of lint rmlint will look for. The list string is a

comma-separated list of lint types or lint groups (other separators like semicolon or space also work).

One of the following groups can be specified at the beginning of the list:

  • all: Enables all lint types.
  • defaults: Enables all lint types, but nonstripped.
  • minimal: defaults minus emptyfiles and emptydirs.
  • minimaldirs: defaults minus emptyfiles, emptydirs and duplicates, but with duplicatedirs.
  • none: Disable all lint types [default].
Any of the following lint types can be added individually, or deselected by

prefixing with a -:

  • badids, bi: Find bad UID, GID or files with both.
  • badlinks, bl: Find bad symlinks pointing nowhere.
  • emptydirs, ed: Find empty directories.
  • emptyfiles, ef: Find empty files.
  • nonstripped, ns: Find nonstripped binaries.
  • duplicates, df: Find duplicate files.
  • duplicatedirs, dd: Find duplicate directories.

WARNING: It is good practice to enclose the description in quotes. In obscure cases argument parsing might fail in weird ways.

-o --output=spec / -O --add-output=spec (default: -o sh:rmlint.sh -o pretty:stdout -o summary:stdout):
Configure the way rmlint outputs it's results. A spec is in the

form format:file or just format. A file might either be an arbitrary path or stdout or stderr. If file is omitted, stdout is assumed.

If -o is specified, rmlint's defaults are overwritten. With --O the

defaults are preserved. Either -o or -O may be specified multiple times to get multiple outputs, including multiple outputs of the same format.

For a list of formatters and their options, refer to the Formatters section below.

-c --config=spec[=value] (default: none):

Configure a format. This option can be used to fine-tune the behaviour of the existing formatters. See the Formatters section for details on the available keys.

If the value is omitted it is set to a true value.

-z --perms[=[rwx]] (default: no check):

Only look into file if it is readable, writable or executable by the current user. Which one of the can be given as argument as one of "rwx".

If no argument is given, "rw" is assumed. Note that r does basically nothing user-visible since rmlint will ignore unreadable files anyways. It's just there for the sake of completeness.

By default this check is not done.

-a --algorithm=name (default: sha1):
Choose the algorithm to use for finding duplicate files. The algorithm can be

either paranoid (byte-by-byte file comparison) or use one of several file hash algorithms to identify duplicates. The following well-known algorithms are available:

spooky, city, murmur, xxhash, md5, sha1, sha256, sha512, farmhash.

The above are all 128 bit except sha1 (160 bit), sha256 and sha512. There are also

some compound variations of the above functions:

  • bastard: 256bit, combining city, and murmur.
  • city256, city512, murmur256, murmur512: Use multiple 128-bit hashes with different seeds.
  • spooky32, spooky64: Faster version of spooky with less bits.
-p --paranoid / -P --less-paranoid (default):

Increase or decrease the paranoia of rmlint's duplicate algorithm.

  • -p is equivalent to --algorithm=sha512
  • -pp is equivalent to --algorithm=paranoid
  • -P is equivalent to --algorithm bastard
  • -PP is equivalent to --algorithm spooky
-v --loud / -V --quiet:

Increase or decrease the verbosity. You can pass these options several times. This only affects rmlint's logging on stderr, but not the outputs defined with -o. Passing either option more than three times has no effect.

-g --progress / -G --no-progress (default):

Convenience shortcut for -o progressbar -o summary -o sh:rmlint.sh -VVV.

Note: This flag clears all previous outputs. Specify any additional outputs

after this flag!

-D --merge-directories (default: disabled):

Makes rmlint use a special mode where all found duplicates are collected and checked if whole directory trees are duplicates. Use with caution: You always should make sure that the investigated directory is not modified during rmlint's or it's removal scripts run.

Output is deferred until all duplicates were found. Duplicate directories are printed first, followed by any remaining duplicate files.

--rank-by applies for directories too, but 'p' or 'P' (path index) has no defined (i.e. useful) meaning. Sorting takes only place when the number of preferred files in the directory differs.


  • This option enables --partial-hidden and -@ (--see-symlinks) for convenience. If this is not desired, you should change this after specifying -D.
  • This feature might not deliver perfect result in corner cases.
  • This feature might add some runtime.
-y --sort-by=order (default: none):

During output, sort the found duplicate groups by criteria described by order. order is a string that may consist of one or more of the following letters:

  • s: Sort by size of group.
  • a: Sort alphabetically by the basename of the original.
  • m: Sort by mtime of the original.
  • p: Sort by path-index of the original.
  • o: Sort by natural found order (might be different on each run).
  • n: Sort by number of files in the group.

The letter may also be written uppercase (similiar to -S / --rank-by) to reverse the sorting. Note that rmlint has to hold back all results to the end of the run before sorting and printing.


Start the optional graphical frontend to rmlint called Shredder.

This will only work when Shredder and it's dependencies were installed. See also: http://rmlint.readthedocs.org/en/latest/gui.html

The gui has it's own set of options, see --gui --help for a list. These

should be placed at the end, ie rmlint --gui [options]

Make rmlint work as a multi-threaded file hash utility, similar to the

popular md5sum or sha1sum utilities, but faster.

A set of paths given on the commandline or from stdin is hashed using one

of the available hash algorithms. Use rmlint --hash -h to see options.

-w --with-color (default) / -W --no-with-color:

Use color escapes for pretty output or disable them. If you pipe rmlints output to a file -W is assumed automatically.

-h --help / -H --show-man:

Show a shorter reference help text (-h) or this full man page (-H).


Print the version of rmlint. Includes git revision and compile time features.

Traversal Options

-s --size=range (default: all):

Only consider files in a certain size range. The format of range is min-max, where both ends can be specified as a number with an optional multiplier. The available multipliers are:

  • C (1^1), W (2^1), B (512^1), K (1000^1), KB (1024^1), M (1000^2), MB (1024^2), G (1000^3), GB (1024^3),
  • T (1000^4), TB (1024^4), P (1000^5), PB (1024^5), E (1000^6), EB (1024^6)

The size format is about the same as dd(1) uses. Example: "100KB-2M".

It's also possible to specify only one size. In this case the size is interpreted as "bigger than this size". If you want to to filter for files up to this size you can add a - in front (-s -1M).

-d --max-depth=depth (default: INF):

Only recurse up to this depth. A depth of 1 would disable recursion and is equivalent to a directory listing.

-l --hardlinked (default) / -L --no-hardlinked:

Whether to report hardlinked files as duplicates.

-f --followlinks / -F --no-followlinks / -@ --see-symlinks (default):

-f will always follow symbolic links. If file system loops occur rmlint will detect this. If -F is specified, symbolic links will be ignored completely, if -@ is specified, rmlint will see symlinks and treats them like small files with the path to their target in them. The latter is the default behaviour, since it is a sensible default for --merge-directories.

-x --no-crossdev / -X --crossdev (default):

Stay always on the same device (-x), or allow crossing mountpoints (-X)?

-r --hidden / -R --no-hidden (default) / --partial-hidden:

Also traverse hidden directories? This is often not a good idea, since directories like .git/ would be investigated. With --partial-hidden hidden files and folders are only considered if

they're inside duplicate directories (see --merge-directories).

-b --match-basename:

Only consider those files as dupes that have the same basename. See also man 1 basename. The comparison of the basenames is case-insensitive.

-B --unmatched-basename:

Only consider those files as dupes that do not share the same basename. See also man 1 basename. The comparison of the basenames is case-insensitive.

-e --match-with-extension / -E --no-match-with-extension (default):

Only consider those files as dupes that have the same file extension. For example two photos would only match if they are a .png. The extension is compared case insensitive, so .PNG is the same as .png.

-i --match-without-extension / -I --no-match-without-extension (default):

Only consider those files as dupes that have the same basename minus the file extension. For example: banana.png and banana.jpeg would be considered, while apple.png and peach.png won't. The comparison is also case-insensitive.

-n --newer-than-stamp=<timestamp_filename> / -N --newer-than=<iso8601_timestamp_or_unix_timestamp>:

Only consider files (and their size siblings for duplicates) newer than a certain modification time (mtime). The age barrier may be given as seconds since the epoch or as ISO8601-Timestamp like 2014-09-08T00:12:32+0200.

-n expects a file from which it can read the timestamp. After rmlint run, the file will be updated with the current timestamp. If the file does not initially exist, no filtering is done but the stampfile is still written.

-N in contrast takes the timestamp directly and will not write anything.

Note that rmlint will find duplicates newer than timestamp, even if the original is

older. If you want only find duplicates where both original and duplicate are newer

than timestamp you can use find(1):

  • find -mtime -1 | rmlint - # find all files younger than a day

Note: you can make rmlint write out a compatible timestamp with:

  • -O stamp:stdout  # Write a seconds-since-epoch timestamp to stdout on finish.
  • -O stamp:stdout -c stamp:iso8601 # Same, but write as ISO8601.

Original Detection Options

-k --keep-all-tagged / -K --keep-all-untagged:

Don't delete any duplicates that are in tagged paths (-k) or that are in non-tagged paths (-K). (Tagged paths are those that were named after //).

-m --must-match-tagged / -M --must-match-untagged:

Only look for duplicates of which at least one is in one of the tagged paths. (Paths that were named after //).

-S --rank-by=criteria (default: pm):

Sort the files in a group of duplicates by one or more criteria.

  • m: keep lowest mtime (oldest) M: keep highest mtime (newest)
  • a: keep first alphabetically A: keep last alphabetically
  • p: keep first named path P: keep last named path
  • d: keep path with lowest depth D: keep path with highest depth
  • l: keep path with shortest basename L: keep path with longest basename
  • r: keep paths matching regex R: keep path not matching regex
  • r: keep paths matching regex R: keep path not matching regex
  • x: keep basenames matching regex X: keep basenames not matching regex

Alphabetical sort will only use the basename of the file and ignore it's case. One can have multiple criteria, e.g.: -S am will choose first alphabetically; if tied then by mtime. Note: original path criteria (specified using //) will always take first priority over -S options.

For more fine grained control, it is possible to give a regular expression to sort by. This can be useful when you know a common fact that identifies original paths (like a path component being src).

To use the regular expression you simply enclose it in the criteria string by adding <REGULAR_EXPRESSIOn> after specifying r or x. Example: -S 'r<.*\.bak$>' makes all files that have a .bak suffix original files.

Warning: When using r or x, try to make your regex to be as specific as possible! Good practice includes adding a $ anchor at the end of the regex.

Tip: l is useful for files like file.mp3 vs file.1.mp3 or file.mp3.bak.


--replay [path.json]:

Read an existing json file and re-output it. This is very useful if you want to reformat, refilter or resort the output you got from an previous run. Usage is simple: Just pass --replay on the second run, with other changed to the new formatters or filters. You can also merge several previous runs by using --replay more than once, in this case it will merge all files given and output them as one big run.

If you want to view only the duplicates of certain subdirectories, just pass them on the commandline as usual.

If path.json is not given then ./rmlint.json is used as default.

By design, some options will not have any effect. Those are:

  • --followlinks
  • --algorithm
  • --paranoid
  • --clamp-low
  • --hardlinked
  • --write-unfinished
  • ... and all other caching options below.
--xattr-read / --xattr-write / --xattr-clear:

Read or write cached checksums from the extended file attributes. This feature can be used to speed up consecutive runs.

This is an slighter securer alternative to --cache, but the same notes as in --cache apply.

NOTE: Many tools do not support extended file attributes properly, resulting in a loss of the information when copying the file or editing it. Also, this is a linux specific feature that works not on all filesystems and only if you have write permissions to the file.

-C --cache file.json:

Read checksums from a json file. This json file is the same that is outputted via -o json, but you can also enrich the json with the checksums of sieved out files via --write-unfinished.

Usage example:

$ rmlint large_cluster/ -O json:cache.json -U   # first run.
$ rmlint large_cluster/ -C cache.json           # second run.

CAUTION: This is a potentially unsafe feature. The cache file might be changed accidentally, potentially causing rmlint to report false positives. As a security feature the mtime of each cached file is checked against the mtime of the time the checksum was created.

NOTE: The speedup you may experience may vary wildly. In some cases the parsing of the json file might take longer than the actual hashing. Also, the cached json file will not be of use when doing many modifications between the runs, i.e. causing an update of mtime on most files. This feature is mostly intended for large datasets in order to prevent the re-hashing of large files.

-U --write-unfinished:

Include files in output that have not been hashed fully (i.e. files that do not appear to have a duplicate). This is mainly useful in conjunction with --cache. When re-running rmlint on a large dataset this can greatly speed up a re-run in some cases.

This option also applies for --xattr-write.

Rarely used, miscellaneous options

-t --threads=N (default: 16):

The number of threads to use during file tree traversal and hashing. rmlint probably knows better than you how to set the value.

-u --max-paranoid-mem=size:

Apply a maximum number of bytes to use for --paranoid. The size-description has the same format as for --size.

-q --clamp-low=[fac.tor|percent%|offset] (default: 0) / -Q --clamp-top=[fac.tor|percent%|offset] (default: 1.0):

The argument can be either passed as factor (a number with a . in it), a percent value (suffixed by %) or as absolute number or size spec, like in --size.

Only look at the content of files in the range of from low to (including) high. This means, if the range is less than -q 0% to -Q 100%, than only partial duplicates are searched. If the file size is less than the clamp limits, the file is ignored during traversing. Be careful when using this function, you can easily get dangerous results for small files.

This is useful in a few cases where a file consists of a constant sized header or footer. With this option you can just compare the data in between. Also it might be useful for approximate comparison where it suffices when the file is the same in the middle part.

--with-fiemap (default) / --without-fiemap:

Enable or disable reading the file extents on rotational disk in order to optimize disk access patterns.

--with-metadata-cache / --without-metadata-cache (default):

Swap certain file metadata attributes onto disk in order to save memory. This can help to save memory for very big datasets (several million files) where storing the paths alone can eat up several GB RAM. Enabling swapping will cause slowdowns in exchange.

Sometimes the memory savings may be small since rmlint already compresses

paths but storing them in a special tree structure.

This feature may not play nice with some other options, causing heavy load and long computations:

  • The --match-* family of options.
  • --cache might use more memory and takes longer.

Some of those restrictions might be removed in future rmlint versions.

The metadata cache will be stored in $XDG_CACHE_HOME/rmlint/$pid. If the cache cannot be created, rmlint warns you and falls back to normal uncached mode.


  • csv: Output all found lint as comma-separated-value list.

    Available options:

    • no_header: Do not write a first line describing the column headers.
  • sh: Output all found lint as shell script This formatter is activated

    as default.

    Available options:

    • cmd: Specify a user defined command to run on duplicates. The command can be any valid /bin/sh-expression. The duplicate path and original path can be accessed via "$1" and "$2". The command will be written to the user_command function in the sh-file produced by rmlint.

    • handler Define a comma separated list of handlers to try on duplicate files in that given order until one handler succeeds. Handlers are just the name of a way of getting rid of the file and can be any of the following:

      • clone: btrfs only. Try to clone both files with the BTRFS_IOC_FILE_EXTENT_SAME ioctl(3p). This will physically delete duplicate extents. Needs at least kernel 4.2.
      • reflink: Try to reflink the duplicate file to the original. See also --reflink in man 1 cp. Fails if the filesystem does not support it.
      • hardlink: Replace the duplicate file with a hardlink to the original file. Fails if both files are not on the same partition.
      • symlink: Tries to replace the duplicate file with a symbolic link to the original. Never fails.
      • remove: Remove the file using rm -rf. (-r for duplicate dirs). Never fails.
      • usercmd: Use the provided user defined command (-c sh:cmd=something). Never fails.

      Default is remove.

    • link: Shortcut for -c sh:clone,reflink,hardlink,symlink.

    • hardlink: Shortcut for -c sh:hardlink,symlink.

    • symlink: Shortcut for -c sh:symlink.

  • json: Print a JSON-formatted dump of all found reports. Outputs all finds as a json document. The document is a list of dictionaries, where the first and last element is the header and the footer respectively, everything between are data-dictionaries.

    Available options:

    • no_header=[true|false]: Print the header with metadata.
    • no_footer=[true|false]: Print the footer with statistics.
    • oneline=[true|false]: Print one json document per line.
  • py: Outputs a python script and a JSON document, just like the json formatter. The JSON document is written to .rmlint.json, executing the script will make it read from there. This formatter is mostly intented for complex use-cases where the lint needs special handling. Therefore the python script can be modified to do things standard rmlint is not able to do easily.

  • stamp:

    Outputs a timestamp of the time rmlint was run.

    Available options:

    • iso8601=[true|false]: Write an ISO8601 formatted timestamps or seconds since epoch?
  • progressbar: Shows a progressbar. This is meant for use with stdout or stderr [default].

    See also: -g (--progress) for a convenience shortcut option.

    Available options:

    • update_interval=number: Number of milliseconds to wait between updates. Higher values use less resources (default 50).
    • ascii: Do not attempt to use unicode characters, which might not be supported by some terminals.
    • fancy: Use a more fancy style for the progressbar.
  • pretty: Shows all found items in realtime nicely colored. This formatter is activated as default.

  • summary: Shows counts of files and their respective size after the run. Also list all written output files.

  • fdupes: Prints an output similar to the popular duplicate finder fdupes(1). At first a progressbar is printed on stderr. Afterwards the found files are printed on stdout; each set of duplicates gets printed as a block separated by newlines. Originals are highlighted in green. At the bottom a summary is printed on stderr. This is mostly useful for scripts that were set up for parsing fdupes output. We recommend the json formatter for every other scripting purpose.

    Available options:

    • omitfirst: Same as the -f / --omitfirst option in fdupes(1). Omits the first line of each set of duplicates (i.e. the original file.
    • sameline: Same as the -1 / --sameline option in fdupes(1). Does not print newlines between files, only a space. Newlines are printed only between sets of duplicates.


This is a collection of common usecases and other tricks:

  • Check the current working directory for duplicates.

    $ rmlint

  • Show a progressbar:

    $ rmlint -g

  • Quick re-run on large datasets using different ranking criteria on second run:

    $ rmlint large_dir/ # First run; writes rmlint.json

    $ rmlint --replay rmlint.json large_dir -S MaD

  • Search only for duplicates and duplicate directories

    $ rmlint -T "df,dd" .

  • Compare files byte-by-byte in current directory:

    $ rmlint -pp .

  • Find duplicates with same basename (excluding extension):

    $ rmlint -e

  • Do more complex traversal using find(1).

    $ find /usr/lib -iname '*.so' -type f | rmlint - # find all duplicate .so files

    $ find ~/pics -iname '*.png' | ./rmlint - # compare png files only

  • Limit file size range to investigate:

    $ rmlint -s 2GB    # Find everything >= 2GB

    $ rmlint -s 0-2GB  # Find everything <  2GB

  • Only find writable and executable files:

    $ rmlint --perms wx

  • Reflink on btrfs, else try to hardlink duplicates to original. If that does not work, replace duplicate with a symbolic link:

    $ rmlint -c sh:link

  • Inject user-defined command into shell script output:

    $ ./rmlint -o sh -c sh:cmd='echo "original:" "$2" "is the same as" "$1"'

  • Use data as master directory. Find only duplicates in backup that are also in data. Do not delete any files in data:

    $ rmlint backup // data --keep-all-tagged --must-match-tagged


  1. False Positives: Depending on the options you use, there is a very slight risk of false positives (files that are erroneously detected as duplicate). The default hash function (SHA1) is pretty safe but in theory it is possible for two files to have then same hash. This happens about once in 2 ** 80 files, so it is very very unlikely. If you're concerned just use the --paranoid (-pp) option. This will compare all the files byte-by-byte and is not much slower than SHA1.
  2. File modification during or after rmlint run: It is possible that a file that rmlint recognized as duplicate is modified afterwards, resulting in a different file. If you use the rmlint-generated shell script to delete the duplicates, you can run it with the -p option to do a full re-check of the duplicate against the original before it deletes the file.


  • find(1)
  • rm(1)
  • cp(1)

Extended documentation and an in-depth tutorial can be found at:


If you found a bug, have a feature requests or want to say something nice, please visit https://github.com/sahib/rmlint/issues.

Please make sure to describe your problem in detail. Always include the version of rmlint (--version). If you experienced a crash, please include at least one of the following information with a debug build of rmlint:

  • gdb --ex run -ex bt --args rmlint -vvv [your_options]
  • valgrind --leak-check=no rmlint -vvv [your_options]

You can build a debug build of rmlint like this:

  • git clone git@github.com:sahib/rmlint.git
  • cd rmlint
  • scons DEBUG=1
  • sudo scons install  # Optional


rmlint is licensed under the terms of the GPLv3.

See the COPYRIGHT file that came with the source for more information.


rmlint was written by:

Also see the http://rmlint.rtfd.org for other people that helped us.

If you consider a donation you can use Flattr or buy us a beer if we meet: