Re: [PATCH] fsck: check skiplist for object in fsck_blob()

From: Jeff King <peff@peff.net>
Date: Jul 3, 2018, 14:34 UTC
Message-ID: <20180703143416.GA23556@sigill.intra.peff.net>
In-Reply-To: <9162ed69-d245-8b2f-0dcc-3b345264b029@ramsayjones.plus.com>

On Fri, Jun 29, 2018 at 02:10:59AM +0100, Ramsay Jones wrote:

Show 13 quoted lines

> On 28/06/18 23:03, Jeff King wrote:
> > On Thu, Jun 28, 2018 at 07:53:27PM +0100, Ramsay Jones wrote:
> [snip]
> > Yes, it can go in quickly. But I'd prefer not to keep it in the long
> > term if it's literally doing nothing.
> 
> Hmm, I don't think you can say its doing nothing!
> 
>     "Yeah, this solution seems sensible. Given that we would
>      never report any error for that blob, there is no point
>      in even looking at it."
> 
> ... is no less true, with or without additional patches! ;-)

True that we don't even bother doing the parsing with your patch. But I think I talked myself out of that part being a significant savings elsewhere.

I guess it would be OK to leave it in. It just feels like it would be vestigial after the rest of the patches.

Show 8 quoted lines

> > I have some patches which I think solve your problem. They apply on
> > v2.18.0, but not on v2.17.1 (because they rely on Dscho's increased
> > passing of config_options in v2.18). Is that good enough?
> 
> Heh, I was also writing patches to address this tonight (but
> I was also watching the football, so I was somewhat behind you).
> My patches were not too dissimilar to yours, except I was aiming
> to allow even do_config_from_file() etc., to suppress errors.

I think this should work via do_config_from_file(). The thing it really misses is that git_config_with_options() will not respect it, but the handling of options there is already a bug (well, I don't think there's anything triggerable either before or after my patches, but it feels like a bug waiting to happen).

> Your patches were cleaner and more focused than mine. (Instead of
> turning die_on_error into an enum, I added an additional 'quiet'
> flag. When pushing the stack (eg. for include files), I had to
> copy the quiet flag from the parent struct, etc, ... ;-) ).

Yes, I think that's what you have to do pre-v2.18, where we don't pass the options struct around.

> > Yes, it would include any syntax error. I also have a slight worry about
> > that, but nobody seems to have screamed _yet_. :)
> 
> Hmm, I don't think we can ignore this. :(

I'm not sure. This has been running on every push to GitHub for the past 6 weeks, and this is the first report. It's hard to say what that means, and technically speaking of course this _is_ a regression.

There's a nearby thread of interest, too, which I cc'd you on:

  https://public-inbox.org/git/20180703070650.b3drk5a6kb4k4tnp@glandium.org/

Show 9 quoted lines

> > Here are the patches I came up with.
> 
> Yes, I applied these locally and tested them. All OK here.
> 
> So, FWIW, Ack!
> 
> [I still think my original patch, with the 'to_be_skipped'
> function name changed to 'object_on_skiplist', should be
> the first patch of the series!]

Thanks. If we're going to do any loosening, I think we may want to address that _first_, so it can go directly on top of the patches in v2.17.1 (because it's a bigger issue than the stray message, IMHO).

-Peff