Re: Making split commit graphs pick up new options (namely --changed-paths)

From: Junio C Hamano <gitster@pobox.com>
Date: Jun 16, 2021, 01:45 UTC
Message-ID: <xmqqsg1itwf4.fsf@gitster.g>
In-Reply-To: <87fsxjyz01.fsf@evledraar.gmail.com>

Ævar Arnfjörð Bjarmason <avarab@gmail.com> writes:

> ... indeed unlinking a file you just created (but hold a FD
> to) is a common pattern for getting a temporary file that you don't need
> to unlink on atexit(). It's just not used in e.g. Git's codebase because
> of portability concerns.

Do we open one r/w, write into it and later read back from it, without wanting the file to persist after we exit? That is more suitable for say Editor's "swp" file but I do not think of an instance of our use of temporary file that would benefit from that pattern. IOW, it's not used in our codebase because we have no need for the pattern, not due to portability concerns.

But that distinction is not the primary point in this topic.

> So not a big deal at all, but I wonder if there should be a warning
> there on !Windows, if you can't unlink a file on a POSIX system that
> suggests e.g. a persistent permission problem that won't be going away
> if you ignore it.

I do think we should catch failure from unlink and probably should error out, unless we anticipate that somebody else might be removing the same file and we got ENOENT.