On Thu, Jun 12, 2014 at 12:47 PM, Junio C Hamano <gits...@pobox.com> wrote:
> David Turner <dtur...@twopensource.com> writes:
>> This issue bit us again recently.
>> In talking with some colleagues, I realized that the previous version
>> of this patch, in addition to being potentially slow, was incomplete.
>> Specifically, it didn't handle the case of refs/heads/case/one vs
>> refs/heads/CASE/two; these are case clones even though they strcasecmp
>> different.
> Good catch to realize that two refs that share leading paths that
> are the same except for cases are also problematic, but that makes
> this feature even less about "case clones", doesn't it?
> Also it somehow feels that the patch attempts to solve the issue at
> a wrong level.
>  On a platform that cannot represent two refs like
> these (e.g. trying to create "FOO" when "foo" already exists, or
> trying to create "a/c" when "A/b" already exists---ending up with
> "A/c" instead, which is not what the user wanted to create), would
> it be more sensible to fail the ref creation without touching the
> users of ref API such as receive-pack?  That way, you would also
> catch other uses of refs that are not supported on your system,
> e.g. "git branch a/c" when there already is a branch called "A/b",
> no?

It gets even more hairy :
If the server has A/a and a/b and you clone it it becomes A/a and A/b
locally. Then you push back to the server and you end up with three
refs on the server:  A/a A/b and a/b.
Or what we end up with locally could either be a/a  and a/b or A/a and
A/b depending on which order the server sends the refs back.
Since the ordering I think is not formally defined, is it?, this could
mean that we would end up with something but it would be difficult to
deteministically decide what the outcome would be.
As the refs handling is pretty complex as is I think we want to avoid
adding undeterministic residuals after a clone/fetch/* has failed.

That then IMHO means that we should wait with implementing something
until we have finished the ref API rewrite. If nothing else, having
atomic transactions would mean that
when things fail due to case collissions we would have a deterministic
outcome : transaction failed so nothing was created locally.

But I think this is the wrong level to solve this issue at.
receive-pack.c is only one out of many place where I think this would
come into effect.
You also have reflogs, remotes and all sorts of other things that
would have to be fixed up.

If we want to add this kind check-fails, we should wait until after we
have ref transactions because then we will have one single path,
transaction_update_sha1(), through which every ref creation/update
will have to pass. and thus then we have one single place where such a
check-fail could cover all cases of ref creations.

> CC'ing those who are more active in the ref API area recently than I
> am for their inputs.
To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Reply via email to