Skip to content
Permalink
main
Switch branches/tags
Go to file
Signed-off-by: Elijah Newren <newren@gmail.com>
2 contributors

Users who have contributed to this file

@newren @RagingCactus
executable file 615 lines (562 sloc) 28.5 KB
#!/usr/bin/env python3
"""This is a bug compatible-ish[1] reimplementation of filter-branch, which
happens to be faster. The goal is _only_ to show filter-repo's flexibility
in re-implementing other types of history rewriting commands. It is not
meant for actual end-user use, because filter-branch (and thus
filter-lamely) is an abomination of user interfaces:
* it is difficult to learn, except for a few exceedingly trivial rewrites
* it is difficult to use; even for expert users like me I often have to
spend significant time to craft the filters to do what is needed
* it is painfully slow to use: the slow execution (even if filter-lamely
is several times faster than filter-branch it will still be far slower
than filter-repo) is doubly problematic because users have to retry
their commands often to see if they've crafted the right filters, so
the real execution time is much worse than what benchmarks typically
show. (Benchmarks don't include how long it took to come up with the
right command.)
* it provides really bad feedback: broken filters often modify history
incorrectly rather than providing errors; even when errors are printed,
it takes forever before the errors are shown, the errors are lost in
a sea of output, and no context about which commits were involved are
saved.
* users cannot share commands they come up with very well, because BSD vs.
GNU userland differences will result in errors -- causing the above
problems to be repeated and/or resulting in silent corruption of repos
* the usability defaults are atrocious...
* partial history rewrites
* backup to refs/original/
* no automatic post-run cleanup
* not pruning empty commits
* not rewriting commit hashes in commit messages
* ...and the atrocious defaults combine for even worse effects:
* users mix up old and new history, push both, things get merged, and
then they have even more of a mess with banned objects still floating
around
* since users can run arbitrary commands in the filters, relying on
the local repo to keep a backup of itself seems suspect
* refs/original/ doesn't correctly back up tags (it dereferences them),
so it isn't a safe mechanism for recovery even if all goes well
* even if the backups in refs/original/ were good, many users don't know
how to restore using that mechanism. So they clone before filtering
and just nuke the clone if the filtering goes poorly.
* --tag-name-filter writes out new tags but leaves the old ones around,
making claims like "just clone the repo to get rid of the old
history" a farce. It also makes it hard to extricate old vs. new
bits of history, as if the default to partial history rewrites wasn't
bad enough
* since filtering can result in lots of empty commits, filter-branch at
least provides an option to nuke all empty commits, but naturally
that includes the empty commits that were intentionally added to the
original reposository as opposed to just commits that become empty
due to filtering. And, for good measure, filter-branch's --prune-empty
actually still misses some commits that become empty.
* it's extremely difficult in filter-branch to rewrite commit hashes in
commit messages sanely. It requires using undocumented capabilities
and even then is going to be extremely painful and slow. As long as
--commit-filter isn't used, I could do it in filter-lamely with just
a one-line change, but the point was demonstrating compatibility with
a horrible tool, not showing how we can make it ever so slightly less
awful.
[1] Replacing git-filter-branch with this script will still pass all the
git-v2.22.0 regression tests. However, I know those tests aren't
thorough enough and that I did break backward compatibility in some
cases. But, assuming people are crazy enough to want filter-branch to
continue to exist, I assert that filter-lamely would be a better
filter-branch due to its improved speed. I won't maintain or improve
filter-lamely though, because the only proper thing to do with
filter-branch is attempt to rewrite our collective history so that
people are unaware of its existence. People should use filter-repo
instead.
Intentional differences from git-filter-branch:
* (Perf) --tree-filter and --index-filter only operate on files that have
changed since the previous commit, which significantly reduces the amount
of work needed. This requires special efforts to correctly handle deletes
when the filters attempt to rename files, but provides significant perf
improvements. There is a vanishingly small chance that someone out there
is depending on rewriting all files in every commit and does so
differently depending on topology of commits instead of contents of files
and is thus adversely affected by this change. I doubt it, though.
* I vastly simplified the map() function to just ignore writing out the
mapping; I've never seen anyone explicity use it, and filter-repo
handles remapping to ancestors without it. I dare you to find anyone
that was reading the $workdir/../map/ directory and using it in their
filtering.
* When git-replace was introduced, --parent-filter became obsolete and
deprecated IMO. As such, I didn't bother reimplementing. If I were
to reimplement it, I'd just do an extra loop over commits and invoke
git-replace based on the --parent-filter output or something similar
to that.
* I took a bit of liberty in the implementation of --state-branch; I
still pass the regression tests, but I kind of violated the spirit of
the option. I may actually circle back and fix this, if I add such
a similarly named option to filter-repo.
"""
"""
Please see the
***** API BACKWARD COMPATIBILITY CAVEAT *****
near the top of git-filter-repo.
"""
import argparse
import datetime
import os
import shutil
import subprocess
import sys
try:
import git_filter_repo as fr
except ImportError:
raise SystemExit("Error: Couldn't find git_filter_repo.py. Did you forget to make a symlink to git-filter-repo named git_filter_repo.py or did you forget to put the latter in your PYTHONPATH?")
subproc = fr.subproc
class UserInterfaceNightmare:
def __init__(self):
args = UserInterfaceNightmare.parse_args()
# Fix up args.refs
if not args.refs:
args.refs = ["HEAD"]
elif args.refs[0] == '--':
args.refs = args.refs[1:]
# Make sure args.d is an absolute path
if not args.d.startswith(b'/'):
args.d = os.path.abspath(args.d)
# Save the args
self.args = args
self._orig_refs = {}
self._special_delete_mode = b'deadbeefdeadbeefdeadbeefdeadbeefdeadbeef'
self._commit_filter_functions = b'''
EMPTY_TREE=$(git hash-object -t tree /dev/null)
# if you run 'skip_commit "$@"' in a commit filter, it will print
# the (mapped) parents, effectively skipping the commit.
skip_commit()
{
shift;
while [ -n "$1" ];
do
shift;
echo "$1";
shift;
done;
}
# map is lame; just fake it.
map()
{
echo "$1"
}
# if you run 'git_commit_non_empty_tree "$@"' in a commit filter,
# it will skip commits that leave the tree untouched, commit the other.
git_commit_non_empty_tree()
{
if test $# = 3 && test "$1" = $(git rev-parse "$3^{tree}"); then
echo "$3"
elif test $# = 1 && test "$1" = $EMPTY_TREE; then
:
else
git commit-tree "$@"
fi
}
'''
@staticmethod
def parse_args():
parser = argparse.ArgumentParser(
description='Mimic filter-branch functionality, for those who '
'lamely have not upgraded their scripts to filter-repo')
parser.add_argument('--setup', metavar='<command>',
help=("Common commands to be included before every other filter"))
parser.add_argument('--subdirectory-filter', metavar='<command>',
help=("Only include paths under the given directory and rewrite "
"that directory to be the new project root."))
parser.add_argument('--env-filter', metavar='<command>',
help=("Modify the name/email/date of either author or committer"))
parser.add_argument('--tree-filter', metavar='<command>',
help=("Command to rewrite the tree and its contents. The working "
"directory will be set to the root of the checked out tree. "
"New files are auto-added, disappeared, etc."))
parser.add_argument('--index-filter', metavar='<command>',
help=("Command to rewrite the index. Similar to the tree filter, "
"but there are no working tree files which makes it "
"faster. Commonly used with `git rm --cached "
"--ignore-unmatch` and `git update-index --index-info`"))
parser.add_argument('--parent-filter', metavar='<command>',
help=("Bail with an error; deprecated years ago"))
parser.add_argument('--remap-to-ancestor', action='store_true',
# Does nothing, this option is always on. Only exists
# because filter-branch once allowed it to be off and
# so some tests pass this option.
help=argparse.SUPPRESS)
parser.add_argument('--msg-filter', metavar='<command>',
help=("Command to run for modifying commit and tag messages which "
"are received on standard input; standard output will be used "
"as the new message."))
parser.add_argument('--commit-filter', metavar='<command>',
help=("A command to perform the commit. It will be called with "
"arguments of the form \"<TREE_ID> [(-p <PARENT_COMMIT_ID>)...]"
"\" and the log message on stdin. The commit id is expected "
"on stdout. The simplest commit filter would be 'git "
"commit-tree $@'"))
parser.add_argument('--tag-name-filter', metavar='<command>',
help=("This filter is rewriting tag names. It will be called "
"with tag names on stdin and expect a new tag name on stdout."))
parser.add_argument('--prune-empty', action='store_true',
help=("Prune empty commits, even commits that were intentionally "
"added as empty commits in the original repository and really "
"shouldn't be removed."))
parser.add_argument('--original', metavar='<namespace>', type=os.fsencode,
default=b'refs/original/',
help=("Alter misguided backup strategy to store refs under "
"<namespace> instead of refs/original/"))
parser.add_argument('-d', metavar='<directory>', default='.git-rewrite',
type=os.fsencode,
help=("Alter the temporary directory used for rewriting"))
parser.add_argument('--force', '-f', action='store_true',
help=("Run even if there is an existing temporary directory or "
"an existing backup (e.g. under refs/original/)"))
parser.add_argument('--state-branch', metavar='<branch>',
help=("Do nothing; filter-lamely is enough faster than "
"filter-branch that it doesn't need incrementalism."))
parser.add_argument('refs', metavar='rev-list options',
nargs=argparse.REMAINDER,
help=("Arguments for git rev-list. All positive refs included by "
"these options are rewritten. Sane people specify things like "
"--all, though that annoyingly requires prefacing with --"))
args = parser.parse_args()
# Make setup apply to all the other shell filters
if args.setup:
if args.env_filter:
args.env_filter = args.setup + "\n" + args.env_filter
if args.tree_filter:
args.tree_filter = args.setup + "\n" + args.tree_filter
if args.index_filter:
args.index_filter = args.setup + "\n" + args.index_filter
if args.msg_filter:
args.msg_filter = args.setup + "\n" + args.msg_filter
if args.commit_filter:
args.commit_filter = args.setup + "\n" + args.commit_filter
if args.tag_name_filter:
args.tag_name_filter = args.setup + "\n" + args.tag_name_filter
return args
@staticmethod
def _get_dereferenced_refs():
# [BUG-COMPAT] We could leave out the --dereference and the '^{}' handling
# and fix a nasty bug from filter-branch. But, as stated elsewhere, the
# goal is not to provide sane behavior, but to match what filter-branch
# does.
cur_refs = {}
cmd = 'git show-ref --head --dereference'
output = subproc.check_output(cmd.split())
for line in output.splitlines():
objhash, refname = line.split()
if refname.endswith(b'^{}'):
refname = refname[0:-3]
cur_refs[refname] = objhash
return cur_refs
def _get_and_check_orig_refs(self):
self._orig_refs = self._get_dereferenced_refs()
if any(ref.startswith(self.args.original) for ref in self._orig_refs):
if self.args.force:
cmds = b''.join([b"delete %s\n" % r
for r in sorted(self._orig_refs)
if r.startswith(self.args.original)])
subproc.check_output('git update-ref --no-deref --stdin'.split(),
input = cmds)
else:
raise SystemExit("Error: {} already exists. Force overwriting with -f"
.format(fr.decode(self.args.original)))
def _write_original_refs(self):
new_refs = self._get_dereferenced_refs()
exported_refs, imported_refs = self.filter.get_exported_and_imported_refs()
overwritten = imported_refs & exported_refs
cmds = b''.join([b"update %s%s %s\n" % (self.args.original, r,
self._orig_refs[r])
for r in sorted(overwritten)
if r not in new_refs or self._orig_refs[r] != new_refs[r]])
subproc.check_output('git update-ref --no-deref --stdin'.split(),
input = cmds)
def _setup(self):
if self.args.force and os.path.exists(self.args.d):
shutil.rmtree(self.args.d)
if os.path.exists(self.args.d):
raise SystemExit("Error: {} already exists; use --force to bypass."
.format(self.args.d))
self._get_and_check_orig_refs()
os.makedirs(self.args.d)
self.index_file = os.path.join(self.args.d, b'temp_index')
self.tmp_tree = os.path.join(self.args.d, b't')
os.makedirs(self.tmp_tree)
# Hack (stupid regression tests depending on implementation details
# instead of verifying user-visible and intended functionality...)
if self.args.d.endswith(b'/dfoo'):
with open(os.path.join(self.args.d, b'backup-refs'), 'w') as f:
f.write('drepo\n')
# End hack
def _cleanup(self):
shutil.rmtree(self.args.d)
def _check_for_unsupported_args(self):
if self.args.parent_filter:
raise SystemExit("Error: --parent-filter was deprecated years ago with git-replace(1). Use it instead.")
def get_extended_refs(self):
if not self.args.tag_name_filter:
return self.args.refs
if '--all' in self.args.refs or '--tags' in self.args.refs:
# No need to follow tags pointing at refs we are exporting if we are
# already exporting all tags; besides, if we do so fast export will
# buggily export such tags multiple times, and fast-import will scream
# "error: multiple updates for ref 'refs/tags/$WHATEVER' not allowed"
return self.args.refs
# filter-branch treats --tag-name-filter as an implicit "follow-tags"-ish
# behavior. So, we need to determine which tags point to commits we are
# rewriting.
output = subproc.check_output(['git', 'rev-list'] + self.args.refs)
all_commits = set(output.splitlines())
cmd = 'git show-ref --tags --dereference'.split()
output = subproc.check_output(cmd)
# In ideal world, follow_tags would be a list of tags which point at one
# of the commits in all_commits. But since filter-branch is insane and
# we need to match its insanity, we instead store the tags as the values
# of a dict, with the keys being the new name for the given tags. The
# reason for this is due to problems with multiple tags mapping to the
# same name and filter-branch not wanting to error out on this obviously
# broken condition, as noted below.
follow_tags = {}
for line in output.splitlines():
objhash, refname = line.split()
if refname.endswith(b'^{}'):
refname = refname[0:-3]
refname = fr.decode(refname)
if refname in self.args.refs:
# Don't specify the same tag multiple times, or fast export will
# buggily export it multiple times, and fast-import will scream that
# "error: multiple updates for ref 'refs/tags/$WHATEVER' not allowed"
continue
if objhash in all_commits:
newname = self.tag_rename(refname.encode())
# [BUG-COMPAT] What if multiple tags map to the same newname, you ask?
# Well, a sane program would detect that and give the user an error.
# fast-import does precisely that. We could do it too, but providing
# sane behavior goes against the core principle of filter-lamely:
#
# dispense with sane behavior; do what filter-branch does instead
#
# And filter-branch has a testcase that relies on no error being
# shown to the user with only an update corresponding to the tag
# which was originally alphabetically last being performed. We rely
# on show-ref printing tags in alphabetical order to match that lame
# functionality from filter-branch.
follow_tags[newname] = refname
return self.args.refs + list(follow_tags.values())
def _populate_full_index(self, commit):
subproc.check_call(['git', 'read-tree', commit])
def _populate_index(self, file_changes):
subproc.check_call('git read-tree --empty'.split())
# [BUG-COMPAT??] filter-branch tests are weird, and filter-branch itself
# manually sets GIT_ALLOW_NULL_SHA1, so to pass the same tests we need to
# as well.
os.environ['GIT_ALLOW_NULL_SHA1'] = '1'
p = subproc.Popen('git update-index -z --index-info'.split(),
stdin = subprocess.PIPE)
for change in file_changes:
if change.type == b'D':
# We need to write something out to the index for the delete in
# case they are renaming all files (e.g. moving into a subdirectory);
# they need to be able to rename what is deleted so it actually deletes
# the right thing.
p.stdin.write(b'160000 %s\t%s\x00'
% (self._special_delete_mode, change.filename))
else:
p.stdin.write(b'%s %s\t%s\x00' %
(change.mode, change.blob_id, change.filename))
p.stdin.close()
if p.wait() != 0:
raise SystemExit("Failed to setup index for tree or index filter")
del os.environ['GIT_ALLOW_NULL_SHA1']
def _update_file_changes_from_index(self, commit):
new_changes = {}
output = subproc.check_output('git ls-files -sz'.split())
for line in output.split(b'\x00'):
if not line:
continue
mode_thru_stage, filename = line.split(b'\t', 1)
mode, objid, stage = mode_thru_stage.split(b' ')
if mode == b'160000' and objid == self._special_delete_mode:
new_changes[filename] = fr.FileChange(b'D', filename)
elif set(objid) == set(b'0'):
# [BUG-COMPAT??] Despite filter-branch setting GIT_ALLOW_NULL_SHA1
# before calling read-tree, it expects errors to be thrown if any null
# shas remain. Crazy filter-branch.
raise SystemExit("Error: file {} has broken id {}"
.format(fr.decode(filename), fr.decode(objid)))
else:
new_changes[filename] = fr.FileChange(b'M', filename, objid, mode)
commit.file_changes = list(new_changes.values())
def _env_variables(self, commit):
# Define GIT_COMMIT and GIT_{AUTHOR,COMMITTER}_{NAME,EMAIL,DATE}
envvars = b''
envvars += b'export GIT_COMMIT="%s"\n' % commit.original_id
envvars += b'export GIT_AUTHOR_NAME="%s"\n' % commit.author_name
envvars += b'export GIT_AUTHOR_EMAIL="%s"\n' % commit.author_email
envvars += b'export GIT_AUTHOR_DATE="@%s"\n' % commit.author_date
envvars += b'export GIT_COMMITTER_NAME="%s"\n' % commit.committer_name
envvars += b'export GIT_COMMITTER_EMAIL="%s"\n' % commit.committer_email
envvars += b'export GIT_COMMITTER_DATE="@%s"\n' % commit.committer_date
return envvars
def fixup_commit(self, commit, metadata):
if self.args.msg_filter:
commit.message = subproc.check_output(self.args.msg_filter, shell=True,
input = commit.message)
if self.args.env_filter and not self.args.commit_filter:
envvars = self._env_variables(commit)
echo_results = b'''
echo "${GIT_AUTHOR_NAME}"
echo "${GIT_AUTHOR_EMAIL}"
echo "${GIT_AUTHOR_DATE}"
echo "${GIT_COMMITTER_NAME}"
echo "${GIT_COMMITTER_EMAIL}"
echo "${GIT_COMMITTER_DATE}"
'''
shell_snippet = envvars + self.args.env_filter.encode() + echo_results
output = subproc.check_output(['/bin/sh', '-c', shell_snippet]).strip()
last = output.splitlines()[-6:]
commit.author_name = last[0]
commit.author_email = last[1]
assert(last[2][0:1] == b'@')
commit.author_date = last[2][1:]
commit.committer_name = last[3]
commit.committer_email = last[4]
assert(last[5][0:1] == b'@')
commit.committer_date = last[5][1:]
if not (self.args.tree_filter or self.args.index_filter or
self.args.commit_filter):
return
# os.environ needs its arguments to be strings because it will call
# .encode on them. So lame when we already know the necessary bytes,
# but whatever...just call fr.decode() and be done with it.
os.environ['GIT_INDEX_FILE'] = fr.decode(self.index_file)
os.environ['GIT_WORK_TREE'] = fr.decode(self.tmp_tree)
if self.args.tree_filter or self.args.index_filter:
full_tree = False
deletion_changes = [x for x in commit.file_changes if x.type == b'D']
if len(commit.parents) >= 1 and not isinstance(commit.parents[0], int):
# When a commit's parent is a commit hash rather than an integer,
# it means that we are doing a partial history rewrite with an
# excluded revision range. In such a case, the first non-excluded
# commit (i.e. this commit) won't be building on a bunch of history
# that was filtered, so we filter the entire tree for that commit
# rather than just the files it modified relative to its parent.
full_tree = True
self._populate_full_index(commit.parents[0])
else:
self._populate_index(commit.file_changes)
if self.args.tree_filter:
# Make sure self.tmp_tree is a new clean directory and we're in it
if os.path.exists(self.tmp_tree):
shutil.rmtree(self.tmp_tree)
os.makedirs(self.tmp_tree)
# Put the files there
subproc.check_call('git checkout-index --all'.split())
# Call the tree filter
subproc.call(self.args.tree_filter, shell=True, cwd=self.tmp_tree)
# Add the files, then move out of the directory
subproc.check_call('git add -A'.split())
if self.args.index_filter:
subproc.call(self.args.index_filter, shell=True, cwd=self.tmp_tree)
self._update_file_changes_from_index(commit)
if full_tree:
commit.file_changes.insert(0, fr.FileChange(b'DELETEALL'))
elif deletion_changes and self.args.tree_filter:
orig_deletions = set(x.filename for x in deletion_changes)
# Populate tmp_tree with all the deleted files, each containing its
# original name
shutil.rmtree(self.tmp_tree)
os.makedirs(self.tmp_tree)
for change in deletion_changes:
dirname, basename = os.path.split(change.filename)
realdir = os.path.join(self.tmp_tree, dirname)
if not os.path.exists(realdir):
os.makedirs(realdir)
with open(os.path.join(realdir, basename), 'bw') as f:
f.write(change.filename)
# Call the tree filter
subproc.call(self.args.tree_filter, shell=True, cwd=self.tmp_tree)
# Get the updated file deletions
updated_deletion_paths = set()
for dirname, subdirs, files in os.walk(self.tmp_tree):
for basename in files:
filename = os.path.join(dirname, basename)
with open(filename, 'br') as f:
orig_name = f.read()
if orig_name in orig_deletions:
updated_deletion_paths.add(filename[len(self.tmp_tree)+1:])
# ...and finally add them to the list
commit.file_changes += [fr.FileChange(b'D', filename)
for filename in updated_deletion_paths]
if self.args.commit_filter:
# Define author and committer info for commit_filter
envvars = self._env_variables(commit)
if self.args.env_filter:
envvars += self.args.env_filter.encode() + b'\n'
# Get tree and parents we need to pass
cmd = b'git rev-parse %s^{tree}' % commit.original_id
tree = subproc.check_output(cmd.split()).strip()
parent_pairs = zip(['-p']*len(commit.parents), commit.parents)
# Define the command to run
combined_shell_snippet = (self._commit_filter_functions + envvars +
self.args.commit_filter.encode())
cmd = ['/bin/sh', '-c', combined_shell_snippet, "git commit-tree", tree]
cmd += [item for pair in parent_pairs for item in pair]
# Run it and get the new commit
new_commit = subproc.check_output(cmd, input = commit.message).strip()
commit.skip(new_commit)
reset = fr.Reset(commit.branch, new_commit)
self.filter.insert(reset)
del os.environ['GIT_WORK_TREE']
del os.environ['GIT_INDEX_FILE']
def tag_rename(self, refname):
if not self.args.tag_name_filter or not refname.startswith(b'refs/tags/'):
return refname
newname = subproc.check_output(self.args.tag_name_filter, shell=True,
input=refname[10:]).strip()
return b'refs/tags/' + newname
def deref_tags(self, tag, metadata):
'''[BUG-COMPAT] fast-export and fast-import nicely and naturally handle tag
objects. Trying to break this and destroy the correct handling of tags
requires extra work. In particular, De-referencing tags and thus
forcing all tags to be lightweight is something that would only be done
by someone who was insane, or someone who was trying to mimic
filter-branch's functionality. But then, perhaps I repeat myself.
Anyway, let's mimic yet another insanity of filter-branch here...
'''
if self.args.tag_name_filter:
return
tag.skip()
reset = fr.Reset(tag.ref, tag.from_ref)
self.filter.insert(reset, direct_insertion = False)
def muck_stuff_up(self):
self._check_for_unsupported_args()
self._setup()
extra_args = []
if self.args.subdirectory_filter:
extra_args = ['--subdirectory-filter', self.args.subdirectory_filter]
self.args.prune_empty = True
fr_args = fr.FilteringOptions.parse_args(['--preserve-commit-hashes',
'--preserve-commit-encoding',
'--partial',
'--force'] + extra_args)
fr_args.prune_empty = 'always' if self.args.prune_empty else 'never'
fr_args.refs = self.get_extended_refs()
self.filter = fr.RepoFilter(fr_args,
commit_callback=self.fixup_commit,
refname_callback=self.tag_rename,
tag_callback=self.deref_tags)
self.filter.run()
self._write_original_refs()
self._cleanup()
overrides = ('GIT_TEST_DISALLOW_ABBREVIATED_OPTIONS',
'I_PROMISE_TO_UPGRADE_TO_FILTER_REPO')
if not any(x in os.environ for x in overrides) and sys.argv[1:] != ['--help']:
print("""
WARNING: While filter-lamely is a better filter-branch than filter-branch,
it is vastly inferior to filter-repo. Please use filter-repo
instead. (You can squelch this warning and five second pause with
export {}=1 )""".format(overrides[-1]))
import time
time.sleep(5)
filter_branch = UserInterfaceNightmare()
filter_branch.muck_stuff_up()