Git fails to detect subcommand when hook is symlinked to a builtin

Git fails to detect subcommand when hook is symlinked to a builtin

[Todd Zullinger]

[-- Attachment #1: Type: text/plain, Size: 1648 bytes --]

I'm not sure if this is a bug or more of a "don't do that" sort of
thing.  At fedorahosted.org, the git repositories are mounted on a
filesystem with noexec.  As such, we have all hooks symlinked
elsewhere.  For the post-update hook we used to link directly to
git-update-server-info.

This worked until we upgraded the system to git-1.7.4.1 recently.  A
trace may help illustrate:

With hooks/post-update as a symlink to
/usr/bin/git-update-server-info:

trace: run_command: 'hooks/post-update' 'refs/heads/master'
trace: run_command: 'gc' '--auto' '--quiet'
remote: trace: exec: 'git-refs/heads/master'
remote: trace: run_command: 'git-refs/heads/master'
remote: git: 'refs/heads/master' is not a git command. See 'git --help'.

With hooks/post-update as a symlink to
/usr/share/git-core/templates/hooks/post-update.sample:

trace: run_command: 'hooks/post-update' 'refs/heads/master'
remote: setup: git_dir: .
remote: setup: worktree: (null)
remote: setup: cwd: /home/fedora/tmz/tmp/iwhd.git
remote: setup: prefix: (null)
remote: trace: built-in: git 'update-server-info'

Is there any reasonable way to handle this case?  Perhaps if the
symlink were dereferenced (if no subcommand were found) and then
re-checked for a subcommand?  I admit that I did not delve too deeply
into the code, but from the little I did I wasn't sure that any
changes here wouldn't lead to worse troubles.

-- 
Todd        OpenPGP -> KeyID: 0xBEAF0CE3 | URL: www.pobox.com/~tmz/pgp
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Everyone needs to believe in something. I believe I'll have another
beer.


[-- Attachment #2: Type: application/pgp-signature, Size: 542 bytes --]

Re: Git fails to detect subcommand when hook is symlinked to a builtin

[Junio C Hamano]

Todd Zullinger <tmz@pobox.com> writes:

> I'm not sure if this is a bug or more of a "don't do that" sort of
> thing.  At fedorahosted.org, the git repositories are mounted on a
> filesystem with noexec.  As such, we have all hooks symlinked
> elsewhere.  For the post-update hook we used to link directly to
> git-update-server-info.
>
> This worked until we upgraded the system to git-1.7.4.1 recently.

I have a suspicion that we may end up saying "don't do that", but let's
ponder a bit with some technical background first.  The "git" wrapper and
builtin commands are all the same binary, and as a consequence, their
main() function needs to decide what to do based on argv[0].  If argv[0]
is "git", we know argv[1] is the name of the subcommand, and if argv[0] is
"git-foo", it might be the internal call from the "git" wrapper that was
invoked with argv[] = { "git", "foo", ... }, but we don't expect that
argv[0] to be things like "post-update".

But it is curious why the resulting behaviour changed recently.  The
symptom hints that the link target (git-update-server-info in your case)
used to be given as argv[0] (which still wouldn't have been kosher---you
are not supposed to use dashed form since 1.5.4 or so without adding
$GIT_EXEC_PATH at the beginning of your $PATH, and that is why I said the
first sentence in this reply), but we recently changed something and now
we are getting "hooks/post-update" in argv[0].

I suspect the recent change to run hooks via run_command() interface has
something to do with it.  From what version did you update to 1.7.4.1?

Re: Git fails to detect subcommand when hook is symlinked to a builtin

[Todd Zullinger]

[-- Attachment #1: Type: text/plain, Size: 2647 bytes --]

Junio C Hamano wrote:
> I have a suspicion that we may end up saying "don't do that", but
> let's ponder a bit with some technical background first.  The "git"
> wrapper and builtin commands are all the same binary, and as a
> consequence, their main() function needs to decide what to do based
> on argv[0].  If argv[0] is "git", we know argv[1] is the name of the
> subcommand, and if argv[0] is "git-foo", it might be the internal
> call from the "git" wrapper that was invoked with argv[] = { "git",
> "foo", ... }, but we don't expect that argv[0] to be things like
> "post-update".

I can certainly understand that. :)

> But it is curious why the resulting behaviour changed recently.  The
> symptom hints that the link target (git-update-server-info in your
> case) used to be given as argv[0] (which still wouldn't have been
> kosher---you are not supposed to use dashed form since 1.5.4 or so
> without adding $GIT_EXEC_PATH at the beginning of your $PATH, and
> that is why I said the first sentence in this reply), but we
> recently changed something and now we are getting
> "hooks/post-update" in argv[0].
>
> I suspect the recent change to run hooks via run_command() interface
> has something to do with it.  From what version did you update to
> 1.7.4.1?

I should have mentioned it, but I failed to do so.  Sorry about that.

This was moving from 1.5.5.6 to 1.7.4.1. The packages for EPEL (extra
stuff for RHEL/CentOS) lagged behind for quite some time.  The
fedorahosted.org system was using 1.5.5.6 for a long time with the
symlinked post-update hook in place.  But it's not necessarily
anything that changed recently in git.  It could have been changes all
the way back in 1.6.0, which seems like an eternity in git time.

We already fixed up the repositories on fedorahosted for this, but I
figured I should ask, in case this change was not intended and might
catch other folks doing git hosting.

(The only other minor change that caught us was that some of the repos
did not have core.bare = true and pushes to these oddly configured
repositories failed.  Clearly, that was a misconfiguration from the
time the repos were setup.  Otherwise, for moving from 1.5.5.6 to
1.7.4.1, things went quite smoothly.  That's thanks in no small part
to the great attention to detail you and the other folks on this list
have.  Thank you!)

-- 
Todd        OpenPGP -> KeyID: 0xBEAF0CE3 | URL: www.pobox.com/~tmz/pgp
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The United States is a nation of laws, badly written and randomly
enforced.
    -- Frank Zappa


[-- Attachment #2: Type: application/pgp-signature, Size: 542 bytes --]

Re: Git fails to detect subcommand when hook is symlinked to a builtin

[Junio C Hamano]

Todd Zullinger <tmz@pobox.com> writes:

>> I suspect the recent change to run hooks via run_command() interface
>> has something to do with it.  From what version did you update to
>> 1.7.4.1?
>
> I should have mentioned it, but I failed to do so.  Sorry about that.
>
> This was moving from 1.5.5.6 to 1.7.4.1.

Ok, that explains it.  update-server-info wasn't a built-in, so didn't
have to look at its name to decide what to do.  It didn't have to spawn
other git commands from GIT_EXEC_PATH so running the command directly
without setting the environment variable happened to work.

But such a usage (this doesn't have anything to do with the symlinking)
was deprecated way before 1.5.4 [*1*], so the answer is still "don't do
that" ;-)

Thanks for a clarification.


[Reference]

*1* http://thread.gmane.org/gmane.comp.version-control.git/93511

Re: Git fails to detect subcommand when hook is symlinked to a builtin

[Jonathan Nieder]

Junio C Hamano wrote:

> Ok, that explains it.  update-server-info wasn't a built-in, so didn't
> have to look at its name to decide what to do.  It didn't have to spawn
> other git commands from GIT_EXEC_PATH so running the command directly
> without setting the environment variable happened to work.

Nice detective work.

> But such a usage (this doesn't have anything to do with the symlinking)
> was deprecated way before 1.5.4 [*1*], so the answer is still "don't do
> that" ;-)
[...]
> *1* http://thread.gmane.org/gmane.comp.version-control.git/93511

After rereading some of that thread (oops), I feel a small clarification
is in order.  Yes, the dashed named git-update-server-info is not the
name that that command would preferred to be called by, but if I
understand correctly, this is not what you are talking about having
being deprecated and removed.  Instead, the rule is:

 If you set your $PATH to include $(git --exec-path), then you may still
 run traditional git-<command>s using that dashed name.  Otherwise, you
 may be lucky and find it working (which is probably a bug :)) but
 please, just say no.[1]

Rationale: in that thread, the change to $PATH was agreed on as a
long-term workaround for preserving scripts and muscle memory.

Regards,
Jonathan

[1] In practice: the git wrapper adds GIT_EXEC_PATH to the $PATH and
traditional commands rely on finding it there.  For example, when
git-<command> is a shell script, it will tend to begin by reading in
some useful functions by running ". git-sh-setup", which relies on
git-sh-setup lying on the $PATH.

Re: Git fails to detect subcommand when hook is symlinked to a builtin

[Jonathan Nieder]

Jonathan Nieder wrote:

> Instead, the rule is:

Sorry, that came out a muddle.  Teemu Likonen explains the rule[1]
better.  And Linus explains the rationale[2].  The main reason I found
this interesting enough to chime in is that I was considering
proposing changing the beginning of git scripts to

 . "$(git --exec-path)/git-sh-setup"

to set a good example, which would be a regression in git's ability
to complain when used in a fragile way.

So, Todd: thanks for a very useful reminder.

[1] http://thread.gmane.org/gmane.comp.version-control.git/94378
[2] http://thread.gmane.org/gmane.comp.version-control.git/93825