All notable changes to this project will be documented in this file.

The format follows keepachangelog.com. Please stick to it.

• An up-to-date COPR package for Fedora (thanks eclipseo).
• Add --xattr as --xattr-write, --xattr-read & --write-unfinished to enable easy checksum caching for the next run.
• Unique option for json and csv formatters (-c json:unique)
• New -k option for rmlint.sh to keep directory timestamps like before deletion.

• Warn when using -j without -D.
• The exist status is now EXIT_FAILURE when rmlint was interrupted.
• Slightly changed progressbar ETA algorithm to deliver more stable results.
• docs: added a few scripting examples to underline the possibility to use 3rd-party tools.
• --dedupe options now checks if the files are already reflinks to reduce disk thrashing.
• Made sure that -T dd is the same as specifying -D.

Nothing was deprecated.

Nothing was removed.

• Dry-run was not honoured when a user-defined command was used.
• Fix crash when compiled with -msse4.2
• Relative symbolic links were compared wrong when using -f (See issue #333)
• gui: The size options wer enot correctly passed to rmlint (only Megabytes)
• xattr: timestamps where often compared wrongly.
• Fiemap (and thus reflink detection) code was improved.
• --replay did not replay duplicate directories correctly.
• --replay did not honour --hidden and --size correctly.
• Various build issues.
• gui: fix "render from selected" feature.
• gui: generated script now correctly removes itself.
• gui: fix display of paths with ampersands in them.

Mostly a bugfix release with smaller functional changes.

• Support for reflinks on XFS (48c2e5b1)
• Option "--no-backup" to overwrite output files as before (see "Changed")
• New build system option called "gettext" that eases the maintenance of translations. (thanks Michael Witten)
• gui: Add config option to choose how to handle the found duplicates

• -pp is now silently permitted, but is the same as -p
• Old output files get backed up now to a new timestamped name. This was done to not to overwrite the results of long runs. You can use --no-backup to disable this behaviour.
• Several internal cleanups and potential bug fixes (thanks to Michael Witten)
• Change the default optimization level for a build to -O2.

Nothing was deprecated.

Nothing was removed.

• Several small fixes in the documentation.
• Compiler warnings that may happen with clang and on 32 bit.
• "-g" now does not overwrite other formatters anymore.
• Make the progressbar less flickering.
• Startup error when the permissions to a location where bad.
• Endless loop that happens on btrfs due to the fiemap optimization.
• Fix a bug when doing "rmlint --replay x.json" without an explicit path.
• Fix -f that did not really follow symbolic links.
• gui: locations are now stored persistently and survive restarts.
• scons should work now with both python2 and python3.
• extensive memory allocation with slow CPUs.
• Do not use --remove-destination of cp, but use "rm + ln" to support non-GNU systems.

• New checksum types metro and highway
• New option --keep-hardlinked
• --dedupe option can deduplicate twins on any reflick-capable filesystems
• --dedupe-readonly option can dedupe files on read-only btrfs snapshots

• Checksum types for -P... options (see sahib#261)

• Option --btrfs-clone (use --dedupe)
• Paranoia option -pp (use -p)

• Checksum types bastard, spooky, city & farmhash
• Multihash output option

• Fix scons 3 compatibility issue (sahib#258)
• Fix compile error on systems with no FIEMAP (sahib#252)
• Fix handling of bad uids/gids in python output formatter (sahib#239)
• Fix escaping of dirnames in rmlint.sh test for new emptydirs (sahib#241)
• Fix another quoting issue (sahib#272)

• Fix build-error on macOS due to wrong usage of #ifdef.
• Fix a bug in treemerge that sometimes counted hardlinks more than once.
• Fix copyright year notice all over the source code.
• Fix --equal to allow hidden files, files counting as other lint and empty files.
• Fix --equal to allow path doubles so passing symlinks to the same file still yields a 0. Due to those fixes the paranoia mode of rmlint.sh should run smoother (again) now.
• Use a more sophisticated strategy to find the path to the own binary and fallback to $(which rmlint).
• Workaround the bash bug that did not allow huge functions by bringing back the old approach without a function.

• It is now possible to treat empty files as duplicates when specifying "--size 0".
• Use the cmp builtin to compare regular in the shell script files again, since it's faster.
• Abort early in --equal if an mismatch has been encountered.
• Test the shell script more thoroughly to avoid release troubles.

• Add a new -c switch to rmlint.sh that cleans up empty directories while running the script.

• Do not chmod +x files that exist already (previous versions tried to +x e.g. /dev/null)
• Support anonymous pipes in the sh formatter (Supported now: rmlint -o sh:>(cat))
• Documentation should look a bit cleaner now.
• Fixed a crash when a file was clamped to a size of 0.
• Fixed a rounding error bug that can happen while sorting mtimes.
• Fixed the slightly broken --mtime-window option (was too permissive in rare cases)
• Force permissions (0755 for binary, 0644 for manpage) when installing files in the filesystem.
• Fix wrong ordering when using -D without -k or -K.
• Fix a space wasting bug that happened with -D and -c sh:link.
• Fix a bug in treemerge that sometimes lead to directories not being detected (should be also slightly faster now)
• Fix inconsistent handling of duplicate directories when using hardlinks, symlinks and reflinks.
• Abort early if an invalid path has been given via commandline or stdin.
• Do not silence errors while commandline parsing when using -g.

• Added new checksums: blake2{s,sp,b,bp} and sha3{-256,-384,-512}. sha3 is an alias to sha3-256.
• Support for GNU Hurd (well, at least it compiles :)
• New "-o stats" formatter that prints some statistics about the run.
• Progressbar now additionally shows an ETA time (how long rmlint will probably run)
• New option --honour-dir-layout (-j) that makes --merge-directories (-D) also check if the directory layout is the same.
• A new --equal option that works like the cmp utility, but supports directories.
• Add a progress percentage to rmlint.sh for big runs.

• Changed the default hash function to blake2b.
• The default formatter (pretty) now also outputs the time used for the run.
• Updated existing translations.

• Update translation template.
• Various fix to make rmlint ready to be packaged in debian.

• Output of rmlint.sh uses a colored prefix now.

• Make --replay truly merge different sets of duplicates.
• Call exit(1) when getting a fatal signal (somehow was missing)
• scons test now executes only the sane part of the testsuite.
• Be more friendly when no manpage was found (and show --help)
• Handle readonly btrfs subvolumes well. See also: sahib#195
• Various build errors fixed for old/rare systems.
• Various fixes in the gui, mostly related to old GTK versions.

• New option --mtime-window: Only consider files as duplicates that share a mtime in a certain time-window. See also: sahib#197
• New sortcriteria O (maximize outside hardlinks) and H (maximize total hardlinks) See also: sahib#196
• Proper installation instructions for macOS.

• Re-Design --replay to accept // like the normal commandline does.
• New default sortcriteria is pOMa to maximize the chance of deleting the most bytes from the storage.

• Makefile related compile troubles on BSD for the GUI>
• Check for accessat and ATR_FDCWD explicitly and try to work around them when missing.
• Clearup installation instructions

• Fix symbolic link emitting in sh script (sometimes files were omitted from rmlint.sh)
• Fix compile stop on BSD systems in utilities.c (thanks f99aq8ove)
• Fix some compiler warnings and typos.

• Add basic spanish translation.
• Add basic compile support on cygwin.

• Fix bad size stats using --replay with hardlinks.
• Fix unicode emission of json formatter.
• Fix broken gui/SConscript (was looking for python4...)

• Add unique formatter to print unique files. (See sahib#161)

• Removed --with-metadata-cache and --cache since both introduced lots of code without giving many benefits. In total about 1000 lines of code were removed in this process.
• For limiting memory usage there is just --limit-mem now.

• btrfs: Do not try to clone files on different subvolumes.
• gui: Tie down version for gobject-introspection.
• freebsd: Fix very bad performance due to mounttable and some compile issues.

• -S now understands two new letters to match via regular expressions:

  • r: Select the path as original that matches the pattern given in <PATTERN> after this letter.
  • x: Same as r but match only basename.

• rmlint should compile on Mac OSX now.
• Bugfix: Broken chown calls in sh script (thanks Shukrat Mukimov)
• Bugfix: memory corruption when specifying -T dd alone.
• Bugfix: Make -D and -k / -K play together nicely (thanks phiresky).
• Smaller compile time troubles fixed.
• Progressbar uses timeout-based redraws which leads to much smoother drawing and less cpu footprint.
• pretty formatter (default) produces now valid escaped commands. It is still intented for visual output only. That's why a note for this was added.

• A fully working graphical user interface which is installed as a python module by default (can be disabled via compile option ie scons --without-gui). It can be started via rmlint --gui.
• Support for automatic deduplication on btrfs using BTRFS_IOC_FILE_EXTENT_SAME. The Shellscript now will contain calls to rmlint --btrfs $source $dest for duplicates on btrfs filesystems if the user specified -c sh:clone.
• Benchmark suite that will track the performance of rmlint from release to release. This helps developers detect any speed regressions or improvements and is a tool to help develop and validate optimization strategies.
• Shell/Python-script now does more sanity checks before removing and can be told to re-compare files byte-by-byte before removing them (-p option when running the .sh file).
• Add a new --hash option so rmlint can be used as a very fast file hashing utility, eg rmlint --hash works like sha1sum, or rmlint --hash -d md5 works like md5sum. Also does sha256, sha512, murmur{128}, spooky{32,64,128}, city{128}.
• --sort-by learned new keys: l (path length) and d (path depth).
• New --unmatched-basename option only finds twins with differing basenames.
• Smaller performance and memory optimisations in shredder.

• -g now checks if there is already a sh and json formatter before it adds one.
• -PP now defaults to xxhash as hashing algorithm.
• -o / --output learned to guess the formatter you want to use from the file ending. For example -o /tmp/test.json will work like -o json:/tmp/test.json.
• JSON output contains rmlint version and revision now.
• --replay learned to merge several json files.
• Internal refactoring (credits go to Daniel) of the scheduler and hashing library. The duplicate finding process has be split in separate modules.

• Compiles on Mac OSX now. See also: sahib#139
• Fix a crash that happened with -e.
• Protect other lint than duplicates by -k or -K.
• chown in sh script fixed (was chmod by accident).

• --replay: Re-output a previously written json file. Allow filtering by using all other standard options (like size or directory filtering).
• --sort-by: Similar to -S, but sorts groups of files. So showing the group with the biggest size sucker is as easy as -y s.

• -S's long options is --rank-by now (prior --sortcriteria).
• -o can guess the formatter from the filename if given.
• Remove some optimisations that gave no visible effect.
• Simplified FIEMAP optimisation to reduce initial delay and reduce memory overhead
• Improved hashing strategy for large disks (do repeated smaller sweeps across the disk instead of incrementally hashing every file on the disk)

• Incorrect handling of -W, --no-with-color option
• Handling of $PKG_CONFIG in SConstruct
• Failure to build manpage
• Various BSD compatibility issues
• Nonstandard header sequence in modules using fts
• Removed some unnecessary warnings

• Issue with excessive memory usage and processing delays with very large file counts (>5M files)
• Problems and crashes on 32bit with large files and normal files.
• Bug in memory manager for "paranoid" file comparison method which could lead to OOM error in some cases and infinite looping in others.
• Fixed bug which prevented option --max-paranoid-mem working.
• Note: much kudos to our user "vvs-" who provided many useful testcases and was prepared to re-run a 10-hour duplicate search after each effort to fix the underlying issues.
• Handling of json formatter on invalid utf8, which fixed --cache in return.
• Bug during file traversal when encountering symlinks to empty folders

• More aggressive test suite, leading to higher coverage rates (90% of lines, almost 100% functions at least). Let's not speak of branch coverage for now. 😄
• A primitive benchmark suite.
• A GUI sketch that can be shipped along rmlint.

• Most internal filesystems like proc are ignored now.
• Improved progressbar
• Memory footprint reduced to enable larger filesets to be processed. See discussion at sahib#109. Improvements include a Pat(h)ricia-Trie used as data structure to efficiently map file paths with much less memory consumption. Also the file preprocessing strategy (eg to find path doubles) has been improved to avoid having several large hashtables active at the same time.
• Improved threading strategy which increases speed of duplicate matching. As before, the threading strategy uses just one thread per physical disk to enable fast reading without disk thrash. The improved algorithm now increases the number of cpu threads used to hash the data as it is read in. Also an improved mutex strategy reduces the wait time before the hash results can be processed. Note the new threading strategy is particularly effective on the "paranoid" (byte-by-byte) file comparison method (option -pp), which is now almost as fast as the default (SHA1 hash) method.
• The optimisation in 2.1.0 which detects existing reflinks has been reverted for now due to conflicts between shredder and treemerge.

• performance regression: When having many pairs of duplicates, the core got slower very fast due to linear lookups. Fixed.
• performance regression: No SSDs were detected due to two bugs.
• commandline aborts also on non-fatal option misuses.
• Some statistic counts were updated wrong sometimes.
• Fixes in treemerge to respect directories tagges as originals.
• Ignore "evil" fs types like bindfs, nullfs completely.
• Fix race in file tree traversal.
• Various smaller bugfixes.

• --with-metadata-cache makes rmlint less memory hungry by storing its paths in a sqlite3 database and selecting them when needed.
• --without-fiemap disables the fiemap optimization when focus is on memory footprint.
• --perms can check if a file should be readable/writable or executable.
• Json output is enabled by default and is written to rmlint.json.
• --partial-hidden does only see hidden files in duplicate directories.
• --cache/--write-unfinished can be used to speedup re-runs drastically.
• Checksums can be stored in the xattr of files with --xattr-read/write/clear.
• New progressbar output inspired by journalctl --verfiy.
• Better support for reflink-capable filesystems (e.g. btrfs):
  • detect existing reflinks using fiemap data (significant speedup)
  • support replacing files by a reflink if the filesystem supports it.

• Optional dependency for sqlite3 for --with-metadata-cache.
• --hardlinked is enabled by default.
• Support -n (dry-run) for rmlint.sh; require user input on ask.
• Default digest is now sha1 instead of spooky.
• updated .pot template with help strings.
• updated german translation accordingly.
• -T supports arguments like df,dd properly now.
• New --help text that shows a short reference only.
• sahib made his 1000th commit on rmlint with this text and wonders where all the time has gone and why he isn't rich yet.

Initial release of the rewrite.