Re: [PATCH v2 04/12] refs: implement simple transactions for the packed-refs file

On Thu, Jun 20, 2013 at 08:03:43PM +0200, Michael Haggerty wrote:

quoted

I noticed that recently, too. I have a patch series about 90% complete
that abstracts the tempfile handling (the ultimate goal of which is to
optionally clean up tmp_* files in the objects/ directory). It refactors
the lockfile cleanup, and it would not be too hard to have a committed
or rolled-back lockfile actually remove itself from the "to clean at
exit" list.

Which would make it perfectly safe to have a lockfile as an automatic
variable as long as you commit or rollback before leaving the function.

Cool, then I won't work on that.  You might also have to make the
lockfile list into a doubly-linked-list to avoid having to do a linear
scan to find the entry to delete, unless the total number of entries is
known to remain small.

Yes, I noticed that potential issue, but I don't think it is worth
worrying about. We typically only take one lock at a time, or a handful
of tempfiles (e.g., one object at a time, or two files for diff).

And once it's abstracted out, it would be easy to handle later.

The part I am a little stuck on is plugging it into
pack-objects/index-pack. Their output handling is a little convoluted
because they may be writing to stdout, to a tempfile, to a named file,
or even appending to an existing file in the case of index-pack
--fix-thin. I don't think it's unmanageable, but I need to spend some
more time on the refactoring.

Please CC me on the patch series when it is done.

Will do.

-Peff