work around git bug

Work around git bug that runs smudge/clean filters at the top of the
repository while passing them a relative GIT_WORK_TREE that may point
outside of the repository, by using GIT_PREFIX to get back to the
subdirectory where a relative GIT_WORK_TREE is valid.

git devs have been informed of the bug and may fix it, which could conveivably
break this fix, but as it is, this works back to git 1.7.6.

This commit was sponsored by Jochen Bartl on Patreon.
This commit is contained in:
Joey Hess 2018-07-17 14:25:23 -04:00
parent 5b7c84c868
commit e50ed4ba48
No known key found for this signature in database
GPG key ID: DB12DB0FF05F8F38
4 changed files with 103 additions and 4 deletions

View file

@ -19,6 +19,10 @@ git-annex (6.20180627) UNRELEASED; urgency=medium
* unused --from: Allow specifiying a repository by uuid or description.
* Send User-Agent and any configured annex.http-headers when downloading
with http, fixes reversion introduced when switching to http-client.
* Work around git bug that runs smudge/clean filters at the top of the
repository while passing them a relative GIT_WORK_TREE that may point
outside of the repository, by using GIT_PREFIX to get back to the
subdirectory where a relative GIT_WORK_TREE is valid.
-- Joey Hess <id@joeyh.name> Fri, 22 Jun 2018 10:36:22 -0400

View file

@ -25,12 +25,20 @@ import Utility.Env.Set
- directory if necessary to ensure it is within the repository's work
- tree. While not needed for git commands, this is useful for anything
- else that looks for files in the worktree.
-
- Also works around a git bug when running some hooks. It
- runs the hooks in the top of the repository, but if GIT_WORK_TREE
- was relative, it then points to the wrong directory. In this situation
- GIT_PREFIX contains the directory that GIT_WORK_TREE (and GIT_DIR)
- are relative to.
-}
get :: IO Repo
get = do
gd <- pathenv "GIT_DIR"
prefix <- getpathenv "GIT_PREFIX"
gd <- pathenv "GIT_DIR" prefix
r <- configure gd =<< fromCwd
wt <- maybe (worktree $ location r) Just <$> pathenv "GIT_WORK_TREE"
wt <- maybe (worktree $ location r) Just
<$> pathenv "GIT_WORK_TREE" prefix
case wt of
Nothing -> return r
Just d -> do
@ -39,13 +47,18 @@ get = do
setCurrentDirectory d
return $ addworktree wt r
where
pathenv s = do
getpathenv s = do
v <- getEnv s
case v of
Just d -> do
unsetEnv s
Just <$> absPath d
return (Just d)
Nothing -> return Nothing
pathenv s Nothing = getpathenv s
pathenv s (Just prefix) = getpathenv s >>= \case
Nothing -> return Nothing
Just d -> Just <$> absPath (prefix </> d)
configure Nothing (Just r) = Git.Config.read r
configure (Just d) _ = do

View file

@ -134,3 +134,5 @@ ok
### Have you had any luck using git-annex before? (Sometimes we get tired of reading bug reports all day and a lil' positive end note does wonders)
I've been happily using git-annex on my laptop and server for almost a year; it's saved me a great deal of time and effort.
Thanks for all your work!
> [[fixed|done]] finally... --[[Joey]]

View file

@ -0,0 +1,80 @@
[[!comment format=mdwn
username="joey"
subject="""comment 12"""
date="2018-07-17T15:52:47Z"
content="""
Reproduced this today. ext4, git 2.18.0, git-annex 6.20180627-gbd6799ebe.
In an adjusted unlocked v6 repository.
Running git annex add in 1/2/3/4/5/6/ failed, I then cd'd down to the top
of the repository, and it worked there.
The bug seems to be intermittent; I was then not able to reproduce it for
a while, and now can again.
joey@darkstar:~/tmp/t2/1/2/3/4/5/6#master(unlocked)>git annex add meow --debug
[2018-07-17 12:15:35.536935635] read: git ["--git-dir=../../../../../../.git","--work-tree=../../../../../..","--literal-pathspecs","ls-files","--others","--exclude-standard","-z","--","meow"]
[2018-07-17 12:15:35.542108221] read: git ["--git-dir=../../../../../../.git","--work-tree=../../../../../..","--literal-pathspecs","ls-files","--modified","-z","--","meow"]
git-annex: 1/2/3/4/5/6/meow: getFileStatus: does not exist (No such file or directory)
git-annex: smudge: 1 failed
error: external filter 'git-annex smudge --clean %f' failed 1
error: external filter 'git-annex smudge --clean %f' failed
This is git ls-files running the smudge filter which then fails:
joey@darkstar:~/tmp/t2/1/2/3/4/5/6#master(unlocked)>GIT_TRACE=1 git
--git-dir=../../../../../../.git --work-tree=../../../../../..
--literal-pathspecs ls-files --modified
12:19:49.691639 git.c:415 trace: built-in: git ls-files --modified
12:19:49.692166 run-command.c:637 trace: run_command: 'git-annex smudge --clean '\''1/2/3/4/5/6/meow'\'''
12:19:49.700739 git.c:415 trace: built-in: git config --null --list
12:19:49.704813 git.c:415 trace: built-in: git check-attr -z --stdin annex.backend annex.numcopies annex.largefiles --
12:19:49.705020 git.c:415 trace: built-in: git version
12:19:49.707887 git.c:415 trace: built-in: git cat-file --batch
12:19:49.707880 git.c:415 trace: built-in: git cat-file '--batch-check=%(objectname) %(objecttype) %(objectsize)'
git-annex: 1/2/3/4/5/6/meow: getFileStatus: does not exist (No such file or directory)
git-annex: smudge: 1 failed
Note that the smudge filter is being passed the path from the top of the
repo to the file, despite being in the same directory as the file.
Same thing happens during a successful add though, so why is it
failing to process the filename here? Maybe git is normally running
the smudge filter with cwd in the top of the repo, but when this occurs
it's elsewhere? No, I checked and the smudge filter is always being run
at the top of the repository.
Huh! It's somehow caused by the way that the path to the work-tree is
specified.
joey@darkstar:~/tmp/t2/1/2/3/4/5/6#master(unlocked)>GIT_TRACE=1 git --work-tree=/home/joey/tmp/t2 --literal-pathspecs ls-files --modified12:32:34.639710 git.c:415 trace: built-in: git ls-files --modified
12:32:34.640571 run-command.c:637 trace: run_command: 'git-annex smudge --clean '\''1/2/3/4/5/6/meow'\'''
# succeeds no error
That's two different ways to specifiy the same path, one succeeds and one
fails.
Hmm.. when git runs the smudge filter, pwd is the top of the repository
and `GIT_WORK_TREE=../../../../../..`
So the work tree in this case points outside the repository,
indeed it points to `/`. I guess this is why the smudge filter is
failing, since it's not finding the file in that wrong worktree.
This is absolutely a bug in git, I posted to the git mailing list
in message-id 20180717165834.GA5615@kitenet.net about it.
But why does this only happen sometimes? Well, running `git add`
in a subdirectory does not pass `GIT_WORK_TREE`, because none
was specified. So, the problem does not occur then.
When git-annex runs a git command such as ls-files --modified and passes
a relative work tree, the problem also sometimes doesn't occur.
Sometimes git doesn't need to run the smudge filter
(touching the file makes it need to).
I think it may have to do with the number of subdirectories somehow as
well.
Since git 1.7.6, `GIT_PREFIX` is set when it runs these filters,
and it contains the path to the subdirectory of the working tree that
it was originally run in. So, git-annex can work around the
problem using that. Done.
"""]]