threads / discuss / 31731

Re: Super long branch names corrupt `.git/config`

Subject: Re: Super long branch names corrupt `.git/config`

## tl;dr

6 messages between Oct 4, 2012 and Oct 5, 2012.

replies: 5people: 4as markdown or json

Ben Olive· Oct 4, 2012, 17:15 UTC · lore

Super long branch names corrupt `.git/config`

My `.git/config` can be corrupted if I try to get a local branch with an extremely long name to track a remote branch.

Here is a (contrived) example to reproduce the issue:
    $ cd /tmp
    $ mkdir buggyrepo otherrepo
    $ cd buggyrepo/ && git init && cd -
    $ cd otherrepo/ && git init && cd -
    $ cd buggyrepo/
    $ echo foo > MYFILE
    $ git add MYFILE && git commit -m "Initial"
    $ git remote add otherrepo ../otherrepo/
    $ git checkout -b `ruby -e "puts 'a'*200"`
    $ git push -u otherrepo `ruby -e "puts 'a'*200"`
    fatal: bad config file line 11 in .git/config
    $ git status
    fatal: bad config file line 11 in .git/config
Workaround available: Just delete the offending entry.
I tested this with 1.7.12.2
--Ben Olive
Jeff King· Oct 4, 2012, 18:29 UTC · re: Ben Olive · lore
On Thu, Oct 04, 2012 at 01:15:25PM -0400, Ben Olive wrote:
Show 23 quoted lines
> My `.git/config` can be corrupted if I try to get a local branch with
> an extremely long name to track a remote branch.
> 
> Here is a (contrived) example to reproduce the issue:
> 
> 
>     $ cd /tmp
>     $ mkdir buggyrepo otherrepo
>     $ cd buggyrepo/ && git init && cd -
>     $ cd otherrepo/ && git init && cd -
>     $ cd buggyrepo/
>     $ echo foo > MYFILE
>     $ git add MYFILE && git commit -m "Initial"
>     $ git remote add otherrepo ../otherrepo/
>     $ git checkout -b `ruby -e "puts 'a'*200"`
>     $ git push -u otherrepo `ruby -e "puts 'a'*200"`
>     fatal: bad config file line 11 in .git/config
>     $ git status
>     fatal: bad config file line 11 in .git/config
> 
> Workaround available: Just delete the offending entry.
> 
> I tested this with 1.7.12.2

This is fixed by Ben Walton's 0971e99 (Remove the hard coded length limit on variable names in config files, 2012-09-30). Not sure yet whether it will make it into v1.8.0 or not.

However, note that the tracking config is only one limit to branch name length. On many filesystems, you are limited to 256 bytes (or some other number) per path component, and writing to "refs/heads/aaa{256}.lock" will fail.

-Peff
Junio C Hamano· Oct 4, 2012, 19:28 UTC · re: Ben Olive · lore
Ben Olive <sionide21@gmail.com> writes:
> My `.git/config` can be corrupted if I try to get a local branch with
> an extremely long name to track a remote branch.
>
> Here is a (contrived) example to reproduce the issue:
Don't do that, then ;-)
I think we have a change that is already cooking.

Ben (Walton), want to add a test for your 0971e99 (Remove the hard coded length limit on variable names in config files, 2012-09-30) before it hits 'next', perhaps?

Jeff King· Oct 5, 2012, 00:29 UTC · re: Junio C Hamano · lore
On Thu, Oct 04, 2012 at 12:28:39PM -0700, Junio C Hamano wrote:
Show 14 quoted lines
> Ben Olive <sionide21@gmail.com> writes:
> 
> > My `.git/config` can be corrupted if I try to get a local branch with
> > an extremely long name to track a remote branch.
> >
> > Here is a (contrived) example to reproduce the issue:
> 
> Don't do that, then ;-)
> 
> I think we have a change that is already cooking.
> 
> Ben (Walton), want to add a test for your 0971e99 (Remove the hard
> coded length limit on variable names in config files, 2012-09-30)
> before it hits 'next', perhaps?
I came up with this test earlier today when responding in this thread:
diff --git a/t/t7201-co.sh b/t/t7201-co.sh
index be9672e..a32ecd9 100755
--- a/t/t7201-co.sh
+++ b/t/t7201-co.sh
@@ -454,6 +454,11 @@ test_expect_success \
     'checkout with --track, but without -b, fails with too short tracked name' '
     test_must_fail git checkout --track renamer'
 
+test_expect_success 'checkout tracking with long branch name' '
+	z200=$_z40$_z40$_z40$_z40$_z40
+	git checkout --track -b foo-$z200
+'
+
 setup_conflicting_index () {
 	rm -f .git/index &&
 	O=$(echo original | git hash-object -w --stdin) &&

but it feels a little fake. Why 200? Because that will test the config
limit, but will not overflow the NAME_MAX limit (at least not on
Linux! No clue on other platforms) when we try to create
refs/heads/foo-$z200.

So while this is a nice concrete user-visible impact of Ben's patch, I
think it is a little flaky to be testing. We should probably have a
straight config test in t1303, though.

-Peff
Andrew Ardill· Oct 5, 2012, 00:36 UTC · re: Jeff King · lore
On 5 October 2012 10:29, Jeff King <peff@peff.net> wrote:
Show 6 quoted lines
>...
>
>but it feels a little fake. Why 200? Because that will test the config
>limit, but will not overflow the NAME_MAX limit (at least not on
>Linux! No clue on other platforms) when we try to create
>refs/heads/foo-$z200.

I can't test this particular case right now, but I recently had an issue with Windows Server 2008 due to a long filename, that essentially meant I couldn't move, change owner or change permissions on the given file. Unless someone has more info I can test a bit later. Is the idea that we shouldn't allow filenames that will cause issues with the underlying OS (or other people's OS) or something else?

Regards,
Andrew Ardill
Jeff King· Oct 5, 2012, 15:38 UTC · re: Andrew Ardill · lore
On Fri, Oct 05, 2012 at 10:36:52AM +1000, Andrew Ardill wrote:
Show 15 quoted lines
> On 5 October 2012 10:29, Jeff King <peff@peff.net> wrote:
> >...
> >
> >but it feels a little fake. Why 200? Because that will test the config
> >limit, but will not overflow the NAME_MAX limit (at least not on
> >Linux! No clue on other platforms) when we try to create
> >refs/heads/foo-$z200.
> 
> I can't test this particular case right now, but I recently had an
> issue with Windows Server 2008 due to a long filename, that
> essentially meant I couldn't move, change owner or change permissions
> on the given file. Unless someone has more info I can test a bit
> later. Is the idea that we shouldn't allow filenames that will cause
> issues with the underlying OS (or other people's OS) or something
> else?

I don't think it's that we shouldn't allow such filenames. It's only that the test is flaky, because making the branch name long enough to test the relaxed config code means that we may run afoul of filesystem limitations on creating the ref itself.

It's a separate issue whether we should restrict the length of branch names in order to protect against filesystem limits. I tend to think not, as we handle the filesystem error just fine. The only reason to do so would be to protect people on multi-system projects (e.g., you make a long branch name on Linux that cannot be fetched to a Windows system. Or something; I did not check the limits for those systems). But I have never heard of that happening in practice, so I think we can ignore it for now.

-Peff

← back to recent threads