Re: [PATCH] Wrap inflateInit to retry allocation after releasing pack memory

On Thu, 2009-01-08 at 12:22 -0800, Linus Torvalds wrote:

On Thu, 8 Jan 2009, R. Tyler Ballance wrote:

quoted

quoted

Tyler - does this make the corruption errors go away, and be replaced by 
hard failures with "out of memory" reporting?

Yeah, looks like it:

Well, I was hoping that you'd have a confirmation from your own huge repo, 
but I do suspect it's all the same thing, so I guess this counts as 
confirmation too.

I never got a real solid "consistent" reproduction case with our
repository, just a lot of users that experienced the issue. I think the
Linux repro case is a far better example, and yeah, it's sorta
confirmation (waiting for operations here to deploy the patched 1.6.1 to
dev machines).

quoted

quoted

This patch is potentially pretty noisy, on purpose. I didn't remove the 
reporting from places that already do so - some of them have stricter 
errors than this.

I'm assuming this patch is going to be reworked, if so, I'll back it out
of our internal 1.6.1 build and anxiously await The Real Deal(tm)

Oh, it shouldn't be any noisier under _normal_ load - it's more that 
certain real corruption cases will now report the error twice. That said, 
the new errors should actually be more informative than the old ones, so 
even that isn't necessarily all bad.

Junio - I think we should apply this, and likely to the stable branch too. 
Add the re-trying the inflateInit() after shrinking pack windows on top of 
it.

I really appreciate this guys, this is one of the longer threads I've
participated (spanning over a month) and I'm glad you guys were finally
able to track the issue down.

From now moving forward, I'll try to get a reproduction case with the
kernel tree or something equally big since I know it's frustrating to
play the game of telephone with a proprietary code base ("try this? what
does that do? okay, then this?").

Linus, I'll have a chance to look at your comments on my "variable
packed git window size" patch this weekend, and I'll follow-up in the
appropriate thread.


I'm relatively certain that after this witch hunt, I can get Slide to
cover a round of beers at LinuxWorld or the nearest GitTogether ;)


Cheers
-- 
-R. Tyler Ballance
Slide, Inc.