Re: [PATCH] setup: translate symlinks in filename when using absolute paths

2010/12/29 Junio C Hamano [off-list ref]:

Carlo Marcelo Arenas Belon [off-list ref] writes:

quoted

otherwise, comparison to validate against work tree will fail when
the path includes a symlink and the name passed is not canonical.

Signed-off-by: Carlo Marcelo Arenas Belon <redacted>

I take that "path" and "name passed" refer to the same thing (i.e. "path"
parameter) in the above.

I think you are trying to handle the case where:

 - you give "/home/carenas/one" from the command line;
 - $PWD is "/home/carenas"; and
 - "/home/carenas" is a symlink to "/net/host/home/carenas"

and the scan-from-the-beginning-of-string check done between
"/home/carenas/one" and the return value of get_git_work_tree() which
presumably is "/net/host/home/carenas" disagrees.  I wonder if a more
correct solution might be to help get_git_work_tree() to match the notion
of where the repository and its worktree are to the idea of where the user
thinks they are, i.e. not "/net/host/home/carenas" but "/home/carenas", a
bit better?

I tend to agree. Will cause less surprises (such as this one).

That would involve tweaking make_absolute_path() I guess?

Hm.. can we avoid converting work_tree to absolute path unless people
explicitly set it (via --work-tree and GIT_WORK_TREE)? Basically
worktree will be relative to cwd. Usually it's just ".". When people
run commands outside worktree, it's the relative "cwd/to/worktree".
I'm wondering if we can just avoid the use of make_absolute_path()
completely in get_git_work_tree()..

Note that your patch is the right thing to do either case, i.e. with or
without such a change to make_absolute_path(), as the function is used to
set up the return value from get_git_work_tree().  Anything we compare
with it should have passed make_absolute_path() at least once.

Yes, I think that should that be done inside normalize_path_copy(),
not prefix_path().

quoted

 setup.c |   11 +++++++----
 1 files changed, 7 insertions(+), 4 deletions(-)

Also Carlo, tests should be good for illustration and regression
purposes. I know you described in detail in another mail. But mails
tend to get lost.
-- 
Duy