We need to have a list of references to rewrite (which we pass along to
fast-export), but we accepted arbitrary rev-list args. This could
backfire pretty badly if a user tried the wrong but somewhat
straightforward
git filter-repo --invert-paths --path foo bar
instead of the expected
git filter-repo --invert-paths --path foo --path bar
because the passing of 'bar' as a rev-list arg means that fast-export
happily notices that some kind of rev-list was specified but not a
meaningful one so it gives an empty output...and filter-repo interprets
an empty output as "This is a history with no commits" and promptly
deletes everything.
Partial history rewrites aren't yet properly supported anyway (I would
need to stop doing the disconnect-of-origin-remote, the deleting of
references that were filtered-away or otherwise didn't show up, and
the post-rewrite gc+prune). When I add such support, I'll revisit
how these arguments can be specified.
Signed-off-by: Elijah Newren <newren@gmail.com>
Commit.dump() showed up in a profile. Reorganize the code slightly to
build up much of the string into one big chunk before calling
file_.write(); this shaves a few percent off the total runtime. (Where
total runtime is again measured in terms of the
cat fast-export.original | git filter-repo --stdin --dry-run ...
trick mentioned a few commits back.) Trying to make a [c]StringIO
object in order to build more of the string up into a single place
to reduce the number of file_.write() calls was apparently
counter-productive, so only the header before the parents is combined
into a single string.
Signed-off-by: Elijah Newren <newren@gmail.com>
We repeatedly hit the same filenames over and over as we traverse
history, but our expressions for renaming or filtering within the
newname() function are based solely on the filename and thus will always
give the same answer. So record any answer we get and just use it
whenever we hit the same filename again.
If the filtering expressions contain only a single short pathname, this
has no measurable effect, but for several paths (e.g. listing all
builtin/*.c files individually in git.git) it can add up to a few
percent of overall runtime.
Signed-off-by: Elijah Newren <newren@gmail.com>
Repeatedly using non-compiled regexes is rather wasteful of resources.
Pre-compile these and use the cached versions.
I ran
git filter-repo --invert-paths --path configure.ac --dry-run
and then for timing ran
cat .git/filter-repo/fast-export.original | time git filter-repo \
--invert-paths --path configure.ac --dry-run --stdin
on the git.git repository (with tags of blobs and tags of tags deleted).
Comparing the timings before and after this change, I see about a 13%
overall speedup just from caching the regexes.
Signed-off-by: Elijah Newren <newren@gmail.com>
Python wants filenames with underscores instead of hyphens and with a
.py extension. We really want the main file named git-filter-repo, but
we can add a git_filter_repo.py symlink. Doing so dramatically
simplifies the steps needed to import it as a library in external python
scripts.
Signed-off-by: Elijah Newren <newren@gmail.com>
Most filtering operations are not interested in the time that commits
were authored or committed, or when tags were tagged. As such,
translating the string representation of the date into a datetime object
is wasted effort, and causes us to waste more time later as we have to
translate it back into a string.
Instead, provide string_to_date() and date_to_string() functions so that
callers can perform the translation if wanted, and let the normal case
be fast.
Provides a small but noticable speedup when just filtering based on
paths; about a 3.5% improvement in execution time for writing the new
history.
Signed-off-by: Elijah Newren <newren@gmail.com>
We have to ask fast-import for the new names of commits, but doing so
immediately upon dumping out the commit related information requires
context switches and waiting for fast-import to parse and handle more
information. We don't need to know the new name of the commit until we
run across a subsequent commit that referenced it in the commit message
by its old ID.
So, speed things up dramatically by waiting until we need the new name
of the commit message (or the fast-import output pipe we are
communicating with should be getting kind of full) before blocking on
reading new commit hashes.
Signed-off-by: Elijah Newren <newren@gmail.com>
Treat fast_import_pipes more like the other parameters to
FastExportFilter.run(), both for consistency, and because it will allow
us to more easily defer doing blocking reads for new commit names until
we actually need to know the new commit hashes corresponding to old
commit ids.
Signed-off-by: Elijah Newren <newren@gmail.com>
Once we rewrite history, our history will be unrelated to our original
upstream. As such, migrate refs/remotes/origin/* to refs/heads/*
(our sanity check already verified that if both names exist they are
equal; if the user used --force then just delete the remote tracking
branch and leave the local branch as is), and then delete the 'origin'
remote.
This has a few advantages:
* People expect to work with refs/heads/*, not refs/remotes/origin/*,
and will be more likely to write filters based on those.
* People will probably need to push the new history somewhere when
they are done, and it's easier if we have it in refs/heads/* than
if it's under refs/remotes/origin/*.
* It encourages people to use good hygiene and not mix old and new
histories. If users really want, they can push the repo (or parts
thereof) back over their original history, but they should have to
take extra steps to do it instead of just a 'git push'.
Signed-off-by: Elijah Newren <newren@gmail.com>
Apparently, despite the git-fast-import.txt documentation, tagger is
optional for both fast-export and fast-import. I suspect this is
because there are several (old) tags in the linux.git repository that
have no tagger, so if we want to test rewriting linux.git history then
we need to make it optional for filter-repo too.
Signed-off-by: Elijah Newren <newren@gmail.com>
Make it easy for users to search and replace text throughout the
repository history. Instead of inventing some new syntax, reuse the
same syntax used by BFG repo filter's --replace-text option, namely,
a file with one expression per line of the form
[regex:|glob:|literal:]$MATCH_EXPR[==>$REPLACEMENT_EXPR]
Where "$MATCH_EXPR" is by default considered to be literal text, but
could be a regex or a glob if the appropriate prefix is used. Also,
$REPLACEMENT_EXPR defaults to '***REMOVED***' if not specified. If
you want a literal '==>' to be part of your $MATCH_EXPR, then you
must also manually specify a replacement expression instead of taking
the default. Some examples:
sup3rs3kr3t
(replaces 'sup3rs3kr3t' with '***REMOVED***')
HeWhoShallNotBeNamed==>Voldemort
(replaces 'HeWhoShallNotBeNamed' with 'Voldemort')
very==>
(replaces 'very' with the empty string)
regex:(\d{2})/(\d{2})/(\d{4})==>\2/\1/\3
(replaces '05/17/2012' with '17/05/2012', and vice-versa)
The format for regex is as from
re.sub(<pattern>, <repl>, <string>) from
https://docs.python.org/2/library/re.html
The <string> comes from file contents of the repo, and you specify
the <pattern> and <repl>.
glob:Copy*t==>Cartel
(replaces 'Copyright' or 'Copyleft' or 'Copy my st' with 'Cartel')
Signed-off-by: Elijah Newren <newren@gmail.com>
When non-merge commits have files in the _files_tweaked set (they were
modified by a blob or commit callback), they may become empty. However,
new_1st_parent is more accurately named
new_1st_parent_if_would_become_non_merge; it will always be None for
non-merge commits. So we need to get the correct parent.
Signed-off-by: Elijah Newren <newren@gmail.com>
When we only have an output and no input of our own, filter.run() seems
weird to call, especially since it'll only be closing a handle and waiting
for fast-import to finish. Add a finish() synonym for such a case to make
external code callers more legible.
Signed-off-by: Elijah Newren <newren@gmail.com>
This will allow exporting from one repo into a different repo, and
combined with chained RepoFilter instances from commit 81016821a1
(filter-repo: allow chaining of RepoFilter instances, 2019-01-07), will
even allow things like splicing separate repositories together.
Signed-off-by: Elijah Newren <newren@gmail.com>
We do not want to kill fast-import processes unused; it's better
to abort before those processes are created when we know we need to.
Signed-off-by: Elijah Newren <newren@gmail.com>
Allow each instance to be just input or just output so that we can splice
repos together or split one into multiple different repos.
Signed-off-by: Elijah Newren <newren@gmail.com>
If we are using --stdin, it should be okay to import into a bare repo,
but the checks were enforcing that we were in a clone with a packfile.
Relax the check to work within a bare repo as well.
Signed-off-by: Elijah Newren <newren@gmail.com>
If we have blob callbacks, we cannot pass --no-data to fast-export. Also,
with blob callbacks, any file the callback modifies could match the
modification done to the file by a subsequent commit, possibly making the
later commit empty. As such, we keep a record of all filenames modified
(by blob or commit callbacks), and then check all these filenames for all
subsequent commits to see if it causes empty commits. In particular, if
files other than these are modified in a non-merge commit, we know that
the commit will not become empty so we can bypass the empty-pruning
checks.
Signed-off-by: Elijah Newren <newren@gmail.com>
If a commit was a non-merge commit previously, then since we do not do
any kind of blob modifications (or funny parent grafting), there is no
way for a filemodify instruction to introduce the same version of the
file that already existed in the parent, as such the only check we need
to do to determine whether a commit becomes empty is whether
file_changes is empty. Subsequent more expensive checks can be skipped.
Signed-off-by: Elijah Newren <newren@gmail.com>
Split a lot of the logic out into separate functions, and avoid
flattening parents when the original commit history itself had
redundant parents (such as --no-ff merges).
Signed-off-by: Elijah Newren <newren@gmail.com>
commits may not have any parents at all. As such,
parse_optional_parent_ref() is used expecting that it will sometimes
return None.
Now, when commits are skipped, we have a scheme to translate anyone that
depends on such commits to instead depend on the nearest ancestor of
such commits. If the entire ancestry of a commit was skipped along with
a comit, then that commit will be translated to None, which is
indistinguishable from there having been no parent to begin with.
Sometimes our scheme needs to distinguish between a commit that started
with no parents and one which ended up with no parents, so we need a way
to tell these apart.
Also, not knowing the original parent makes it hard for us to
determine if the original had the same weird topology that the current
commit does. For example, it is possible for a merge commit to have
one parent be the ancestor of another (particularly when --no-ff is
passed to git merge), or even for a merge commit to have the same
commit used as both parents (if you use low-level commands to create
a crazy commit). There are cases where the pruning of some commits
could cause either of these situations to arise, and it's useful to be
able to distinguish between intentionally "weird" history and history
that has been made weird due to other pruning, because the latter we
may have reason to do additional pruning on.
Signed-off-by: Elijah Newren <newren@gmail.com>
Oh, boy, timezone +051800 exists in the wild. Is that 0518 hours and 00
minutes? Or 05 hours and 1800 minutes? Or 051 hours and 800 minutes?
Attempt to do something sane with these broken commits that fast-import
barfs on. Also, fix an old bug in the handling of ahead-of-UTC timezones.
Signed-off-by: Elijah Newren <newren@gmail.com>
Apparently, the default for subprocess stdout is unbuffered; switching
it to buffered yields a huge 40% speedup. Doing this also exposes the
need to add fi_input.flush() calls, highlighting another performance
issue. We may be able to have fewer such calls with some refactoring,
but that is a bigger separate change. Just having them highlighted to
remind about them as a performance issue is good for now.
Signed-off-by: Elijah Newren <newren@gmail.com>
As suggested by Peff, use rev-list & diff-tree to get the information we
need, instead of relying on fast-export (with some out-of-tree patches)
to get that information.
Signed-off-by: Elijah Newren <newren@gmail.com>
If a tag points at a commit whose changes are all filtered out and thus
becomes empty and gets pruned, and all of its ancestors are likewise
pruned, then there is no need for the tag; just nuke it.
Signed-off-by: Elijah Newren <newren@gmail.com>
