Re: git name-rev looks at refs/notes, refs/svn/map: stack overflow

From: SZEDER Gábor <szeder.dev@gmail.com>
Date: Nov 11, 2019, 14:09 UTC
Message-ID: <20191111140956.GJ4348@szeder.dev>
In-Reply-To: <20191111042149.GE6379@sigill.intra.peff.net>

On Sun, Nov 10, 2019 at 11:21:49PM -0500, Jeff King wrote:

> On Sat, Nov 09, 2019 at 12:31:31PM +0100, Sebastiaan Dammann wrote:

Show 17 quoted lines

> All of that's obviously a workaround, though. The real issue is the
> stack exhaustion.
> 
> > I hope to hear your view on this. Is this an (confirmed) issue with
> > git? Are there beside the workaround I mentioned, any other
> > workarounds?
> 
> Yes, this is well known. It's covered in the test suite (unfortunately
> still failing, of course) since 31625b34c0 (t6120: test describe and
> name-rev with deep repos, 2017-09-07).
> 
> There was a proposed fix recently in:
> 
>   https://public-inbox.org/git/20190919214712.7348-1-szeder.dev@gmail.com/
> 
> but it doesn't seem to have been picked up. I'm not sure what the
> current status is.

It's getting along, being polished, clarified and fine-tuned ever so slowly.

E.g. it turned out that the performance penalty from eliminating the recursion is basically entirely caused by using a 'commit-list' to store interesting parents (the overhead of a malloc() on each insert and a free() on each pop). Switching to a LIFO 'prio-queue' doesn't cause any measurable slowdown.