Blob callbacks, either implicit (via e.g. --replace-text) or explicit,
can modify blobs in ways that make them match other blobs, which in turn
can result in some commits becoming empty. We need to detect such cases
and ensure we prune these empty commits when --prune-empty=auto.
Reported-by: John Gietzen <john@gietzen.us>
Signed-off-by: Elijah Newren <newren@gmail.com>
You can call bytes.replace() or re.sub(), but you can't call
bytes.sub(). Oops. Fix the example in the documentation.
Reported-by: John Gietzen <john@gietzen.us>
Signed-off-by: Elijah Newren <newren@gmail.com>
Some projects have a strict --no-ff merging policy. With the default
behavior of --prune-degenerate, we can prune merge commits in a way that
transforms the history into a fast-forward merge. Consider this
example:
* There are two independent commits or branches, named B & C, which
are both built on top of A so that history look like this diagram:
A
\ \
\ B
\
-C
* Someone runs the following sequence of commands:
* git checkout A
* git merge --no-ff B
* git merge --no-ff C
* This will result in a history that looks like:
A---AB---AC
\ \ / /
\ B /
\ /
-C-
* Later, someone comes along and runs filter-repo, specifying to
remove the only path(s) that were modified by B. That would
naturally remove commit B and the no-longer-necessary merge
commit AB. For someone using a strict no-ff policy, the desired
history is
A---AC
\ /
C
However, the default handling for --prune-degenerate would
notice that AC merely merges C into its own ancestor A, whereas
the original AC merged C into something separate (namely, AB).
So, it would say that AC has become degenerate and prune it,
leaving the simple history of
A
\
C
For projects not using a strict no-ff policy, this simpler history
is probably better, but for folks that want a strict no-ff policy,
it is unfortunate.
Provide a --no-ff option to tweak the --prune-degenerate behavior so
that it ignores the first parent being an ancestor of another parent
(leaving the first parent unpruned even if it is or becomes degenerate
in this fashion).
Signed-off-by: Elijah Newren <newren@gmail.com>
Prior to this commit, git-filter-repo could only be used with either the
--dry-run flag or the --debug flag, not both. When run in debug mode,
git-filter-repo expected to be able to read from the output stream,
which obviously isn't created when doing a dry run, so it stack traced
when it tried to use the non-existent output stream. This commit fixes
that bug with an equally simple sanity check for the existence of the
output stream when run in debug mode.
Signed-off-by: Karl Lenz <xorangekiller@gmail.com>
The mailmap format parsed by the "git shortlog" command allows for
matching mailmap entries with no email address. This is admittedly an
edge case, because most Git commits will have an email address
associated with them as well as a name, but technically the address
isn't required, and "git shortlog" accommodates that in its mailmap
format. This commit teaches git-filter-repo to do the same thing.
Signed-off-by: Karl Lenz <xorangekiller@gmail.com>
It's hard to be exhaustive, but if users try something like:
--path-rename foo/bar/baz:.
or
--path ../other-dir
then bad things happen. In the first case, filter-repo will try to
ask fast-import to create a directory named '.' and move everything
from foo/bar/baz/ into it but of course '.' is a reserved directory
name so we can't create it. In the second case, they are probably
running from a subdirectory, but filter-repo doesn't work from a
subdirectory. I hard-coded the assumption that everything was in the
toplevel directory and all paths were relative from there pretty
early on. So, if the user tries to use any of these components
anywhere, just throw an early error.
Signed-off-by: Elijah Newren <newren@gmail.com>
In commit f2729153 (filter-repo: workaround Windows' insistence that cwd
not be a bytestring, 2019-10-19), filter-repo was made to use a special
SubprocessWrapper class instead of the normal subprocess calls, due to
what appears to be in bugs in the python implementation on Windows not
working with arguments being bytestrings. Add the same workarounds to
bfg-ish and filter-lamely.
Signed-off-by: Elijah Newren <newren@gmail.com>
git fast-import, in an attempt to be friendly, allows the same file to
be specified multiple times within a commit and just takes the last
version of the file listed. It determines whether files are the same
via fspathncmp, which is defined differently depending on the setting of
core.ignorecase. Unfortunately, this means that if someone is trying to
do filtering of history and using a broken (case-insensitive) filesystem
and the history they are filtering has some paths that differed in name
only, then fast-import will delete whichever of the "colliding" files is
listed first.
Avoid these problems by just turning off core.ignorecase while
fast-import is running. This will prevent silently modifying the repo
in an unexpected way. Users on such filesystems may have difficulty
checking out commits with files which differ in case only, but that is
a separate problem for them to deal with after rewriting history.
Signed-off-by: Elijah Newren <newren@gmail.com>
Clean up the PyPI dist packages, remove unnecessary files, and
streamline the release process:
* Avoid adding extra unnecessary files to the repo; setup.py is code
and can copy the necessary files into place.
* Make sure README.md is included so we don't get an UNKNOWN
Description field.
* Add a long_description_content_type to avoid parsing errors on the
README.md file and rejecting the upload.
* Define the license and platform fields so they don't show up as
UNKNOWN either.
* Remove unnecessary pyproject.toml. This makes sense for most python
projects, but since I already have a Makefile with installation
rules (because I'm trying to be more compatible with git.git just in
case we ever get merged into it), the pyproject.toml file is
somewhat duplicative. Sure, the Makefile won't specify the exact
versions needed but...meh.
* Split the release target of the Makefile into github_release and
pypi_release substeps, to allow them to be run semi-independently.
Make the pypi_release run a few more steps for me.
Signed-off-by: Elijah Newren <newren@gmail.com>
In commit 509a624 (filter-repo: fix issue with pruning of empty commits,
2019-10-03), it was noted that when the first parent is pruned away,
then we need to generate a corrected list of file changes relative to
the new first parent. Unfortunately, we did not apply our set of file
filters to that new list of file changes, causing us to possibly
introduce many unwanted files from the second parent into the history.
The testcase added at the time was rather lax and totally missed this
problem (which possibly exacerbated the original bug being fixed rather
than helping). Tighten the testcase, and fix the error by filtering the
generated list of file changes.
Signed-off-by: Elijah Newren <newren@gmail.com>
RepoFilter._tweak_commit() was a bit unwieldy, and we have a reason for
wanting to re-use the file filtering logic in it, so break that out into
a separate function.
Signed-off-by: Elijah Newren <newren@gmail.com>
The analyze mode will handle type changes (e.g. normal file to symlink)
in combination with adds and modifies, but the similar logic below
didn't allow for type changes in combination with renames. Fix the
oversight.
Signed-off-by: Elijah Newren <newren@gmail.com>
Make it clearer that absolute paths should not be used for pathnames
within a git repository. Also, fix the comment about how the
insert-beginning script could be implemented as a one-liner; the
example commented-out code should have used bytestrings.
Signed-off-by: Elijah Newren <newren@gmail.com>
The prerequisites and installation docs were not quite detailed enough,
and no code of conduct or contribution guidelines were included. Flesh
out the docs to cover these issues.
Signed-off-by: Elijah Newren <newren@gmail.com>
Some of the systems I ran on had a 'python3-coverage' and some had a
'coverage3' program. More were of the latter name, but more
importantly, the upstream tarball only creates the latter name;
apparently the former was just added by some distros. So, switch to the
more official name of the program.
Signed-off-by: Elijah Newren <newren@gmail.com>
- correct paths to including missing "Documentation/" prefix
- use fully specified "origin/docs" branch in case the "docs" branch is
not checked out locally
Signed-off-by: Benoit Fouletier <bennews@free.fr>
It appears that in addition to Windows requiring cwd be a string (and
not a bytestring), it also requires the command line arguments to be
unicode strings. This appears to be a python-on-Windows issue at the
surface (attempts to quote things that assumes the arguments are all
strings), but whether it's solely a python-on-Windows issue or there is
also a deeper Windows issue, we can workaround this brain-damage by
extending the SubprocessWrapper slightly. As with the cwd changes, only
apply this on Windows and not elsewhere because there are perfectly
legitimate reasons to pass non-unicode parameters (e.g. filenames that
are not valid unicode).
Signed-off-by: Elijah Newren <newren@gmail.com>
Unfortunately, it appears that Windows does not allow the 'cwd' argument
of various subprocess calls to be a bytestring. That may be functional
on Windows since Windows-related filesystems are allowed to require that
all file and directory names be valid unicode, but not all platforms
enforce such restrictions. As such, I certainly cannot change
cwd=directory
to
cwd=decode(directory)
because that could break on other platforms (and perhaps even on Windows
if someone is trying to read a non-native filesystem). Instead, create
a SubprocessWrapper class that will always call decode on the cwd
argument before passing along to the real subprocess class. Use these
wrappers on Windows, and do not use them elsewhere.
Signed-off-by: Elijah Newren <newren@gmail.com>
During the python3 transition, StringIO was renamed to io -- but the
import wasn't moved to preserve appropriate sorting.
Signed-off-by: Elijah Newren <newren@gmail.com>
Note that this isn't a version *number* or even the more generalized
version string that folks are used to seeing, but a version hash (or
leading portion thereof).
A few import points:
* These version hashes are not strictly monotonically increasing
values. Like I said, these aren't version numbers. If that
bothers you, read on...
* This scheme has incredibly nice semantics satisfying a pair of
properties that most version schemes would assume are mutually
incompatible:
This scheme works even if the user doesn't have a clone of
filter-repo and doesn't require any build step to inject the
version into the program; it works even if people just download
git-filter-repo.py off GitHub without any of the other sources.
And:
This scheme means that a user is running precisely version X of
the code, with the version not easily faked or misrepresented
when third parties edit the code.
Given the wonderful semantics provided by satisfying this pair of
properties that all other versioning schemes seem to miss out on, I
think I should name this scheme. How about "Semantic Versioning"?
(Hehe...)
* The version hash is super easy to use; I just go to my own clone of
filter-repo and run either:
git show $VERSION_HASH
or
git describe $VERSION_HASH
* A human consumable version might suggest to folks that this software
is something they might frequently use and upgrade. This program
should only be used in exceptional cases (because rewriting history
is not for the faint of heart).
* A human consumable version (i.e. a version number or even the
more relaxed version strings in more common use) might suggest to
folks that they can rely on strict backward compatibility. It's
nice to subtly undercut any such assumption.
* Despite all that, I will make releases (downloadable tarballs with
real version numbers in the tarball name; I'm just going to re-use
whatever version git is released with at the time). But those
version numbers won't be used by the --version option; instead the
version hash will.
Signed-off-by: Elijah Newren <newren@gmail.com>
A few changes:
* Include notes about git-2.24.0 changes
* Make it clearer that messing with the first parent could have
negative side-effects if the file_changes aren't also updated.
* Fix wrapping of a line that was too long.
Also, update the README.md:
* Note the upstream improvements made in (not yet released) git-2.24.0
Signed-off-by: Elijah Newren <newren@gmail.com>
Partial history rewrites were possible before with the (previously
hidden) --refs flag, but the defaults were wrong. That could be worked
around with the --source or --target flags, but that disabled --no-data
for fast-export and thus slowed things down, and also would require
overridding --replace-refs. And the defaults for --source and --target
may diverge further from what is wanted/needed for partial history
rewrites in the future.
So, add --partial as a first-class supported option with scary
documentation about how it permits mixing new and old history. Make
--refs imply that flag. Make the behavioral similarities (in regards to
which steps are skipped) between --source, --target, and --partial more
clear. Add relevant documentation to round it out.
Signed-off-by: Elijah Newren <newren@gmail.com>
Elijah Newren
Karl Lenz
Julian Berman
Andreas Schneider
Matthisk Heimensen
Benoit Fouletier