Re: What's cooking in git.git (Feb 2011, #07; Mon, 28)

From: Junio C Hamano <gitster@pobox.com>
Date: Mar 1, 2011, 17:50 UTC
Message-ID: <7voc5uenmy.fsf@alter.siamese.dyndns.org>
In-Reply-To: <AANLkTiksROE_dQTUn3Y_1Op6NDpGzEsSJ70JNYyYrufS@mail.gmail.com>

Erik Faye-Lund <kusmabite@gmail.com> writes:

Show 5 quoted lines

>> Erik, please check the made-up commit log message and sign it off.
>> Will merge to 'next' after that.
>
> Do you have any hint on how I can create a repo with more than 2048
> packs so I can test that the patch works?

A stupid and brute force way is to prepare an empty bare repository, add something like the following to its config:

    [transfer]
            unpacklimit = 1

    [gc]
            auto = 0
            autopacklimit = 0

and push into it from a random existing repository starting from a reasonably old commit, walking to a newer commit. You can use something like (totally untested):

        git rev-list --first-parent master |
        head -n 2000 |
        tac |
        while read commit
        do
		git push ../victim.git $commit:refs/heads/master
	done

Show 5 quoted lines

> There's also the added complication that _open_osfhandle also seems to
> be limited by this 2048 limit (see the third comment here:
> http://bugs.mysql.com/bug.php?id=24509). We use this function for
> pipes and sockets as well. This alone might be a reason to corner-case
> test the resulting binary on Windows properly.

As long as the use of osfhandles for non filedescriptors is reasonably bounded, it may be enough to make the 25 slop configurable per platform, no?