Re: [PATCH v3] To make GIT-VERSION-FILE, search for git more widely

From: Junio C Hamano <gitster@pobox.com>
Date: May 13, 2009, 19:32 UTC
Message-ID: <7vy6t03ik0.fsf@alter.siamese.dyndns.org>
In-Reply-To: <op.utva2vxh1e62zd@balu>

"Matthias Andree" <matthias.andree@gmx.de> writes:

Show 14 quoted lines

> Am 09.05.2009, 18:55 Uhr, schrieb Junio C Hamano <gitster@pobox.com>:
>
>> "Matthias Andree" <matthias.andree@gmx.de> writes:
>>
>>>> Fine then.  Or you could just append "." to the $PATH ;-)
>>>
>>> "." in the super user's PATH? Cool stuff, and so innovative.
>>
>> I didn't mean to suggest PATH=$PATH:. *in the user's environment* ;-).
>> You do that inside GIT-VERSION-FILE, which is essentially the same thing
>> as running ./git$X from there.
>
> No, it is not -- the scope of the GIT variable is much narrower than
> doing  PATH=$PATH:. in the script.

If you cannot trust the top of your build directory you build git in (that is why PATH=$PATH:. while running GIT-VERSION-FILE may be scary) and anticipate a malicious third-party can somehow put random things (like "test" or "cat", perhaps) there, you are already lost, don't you think?

That is where my "essentially the same" came from.

> For any approach taken, we'll have to touch both the shell and the
> Makefile, unless we want to manually redo things in the
> GIT-VERSION-GEN  script that were already done automatically or
> programmatically in  Makefile.

I actually like the rationale you mentioned in the thread (perhaps in the original proposed commit message as well) that we should ask the freshly built git to describe the version if available, falling back to whichever git of random vintage found on the original $PATH.

If it weren't for $X [*1*], my preference would have been (as I said in the discussion) to run ./git if available locally.

But I think your "deal with details like $X to figure out the name of the freshly built git binary is in the Makefile, and pass it via GIT variable to GIT-VERSION-GEN" is a sensible approach. I do not remember if your patch gave precedence to an installed git on the original $PATH or a freshly built one, though---the precedent probably does not matter in practice, and favoring the one found on $PATH over freshly built one does have an advantage if we were to support cross compilation (I have a suspicion that the current setup does not).

Thanks.

[Footnote]

*1* ... and perhaps VPATH as you mentioned earlier, but I do not know if our current Makefile is set up to allow a layout that separates build products from the source material, as I've never attempted to build git in a setting where VPATH is involved.