Re: [PATCH v3 0/8] Fix GIT_CEILING_DIRECTORIES that contain symlinks

[Michael Haggerty <mhagger@alum.mit.edu>]

On 11/13/2012 09:50 PM, David Aguilar wrote:
> On Mon, Nov 12, 2012 at 9:47 AM, Junio C Hamano <gitster@pobox.com> wrote:
>> Michael Haggerty <mhagger@alum.mit.edu> writes:
>>
>>> The log message of the original commit (0454dd93bf) described the
>>> following scenario: a /home partition under which user home directories
>>> are automounted, and setting GIT_CEILING_DIRECTORIES=/home to avoid
>>> hitting /home/.git, /home/.git/objects, and /home/objects (which would
>>> attempt to automount those directories).  I believe that this scenario
>>> would not be slowed down by my patches.
>>>
>>> How do you use GIT_CEILING_DIRECTORIES that the proposed changes cause a
>>> slowdown?
>>
>> Yeah, I was also wondering about that.
>>
>> David?
> 
> I double-checked our configuration and all the parent directories
> of those listed in GIT_CEILING_DIRECTORIES are local,
> so our particular configuration would not have a performance hit.
> 
> We do have multiple directories listed there.  Some of them share
> a parent directory.  I'm assuming the implementation is simple and
> does not try and avoid repeating the check when the parent dir is
> the same across multiple entries.
> 
> In any case, it won't be a problem in practice based on my
> reading of the current code.

OK, so we're back to the following status: some people (including me)
are nervous that this change could cause a performance regression,
though it seems that the most sensible ways of using the
GIT_CEILING_DIRECTORIES feature would not be affected.

In favor: Currently, if a directory containing a symlink is added to
GIT_CEILING_DIRECTORIES, then GIT_CEILING_DIRECTORIES will not work, git
has no way of recognizing that there is a problem, and the only symptom
observable by the user is that the hoped-for performance improvement
from using GIT_CEILING_DIRECTORIES will not materialize (or will
disappear after a filesystem reorg) [1].

Against: The change will cause a performance regression if a
slow-to-stat directory is listed in GIT_CEILING_DIRECTORIES.  The
slowdown will occur whenever git is run outside of a true git-managed
project, most nastily in the case of using __git_ps1 in a shell prompt.

I don't have a preference either way about whether these patches should
be merged.

Michael

[1] It is also conceivable that GIT_CEILING_DIRECTORIES is being used to
*hide* an enclosing git project rather than to inform git that there are
no enclosing projects, in which case the enclosing project would *not*
be hidden.  This is in fact the mechanism by which the problem causes
failures in our test suite.  But I don't expect that this is a common
real-world scenario, and anyway such a failure would be obvious to the
user and quickly fixed.

-- 
Michael Haggerty
mhagger@alum.mit.edu
http://softwareswirl.blogspot.com/