Re: [syzbot] KASAN: use-after-free Read in rdma_close

[asmadeus]

Christian Schoenebeck wrote on Wed, Sep 28, 2022 at 02:57:07PM +0200:
> OK, maybe it's just me, but ask yourself Leon, if you were the only guy left 
> (i.e. Dominique) still actively taking care for 9p, would those exactly be 
> motivating phrases for your efforts? Just saying.

I didn't plan on replying (happy to disagree), but I'm actually grateful
for Leon to have taken the time to look here: Thank you!
While I probably would also have spotted the error (the change is
fresh), it saved me time even if you account for some bikeshedding.

(Not particularly happy with the amount of time I can allocate to 9p nor
the maintainance work I'm doing by the way, but I guess it's better than
leaving it completely unmaintained)

> From technical perspective, yes, destruction in reverse order is usually the 
> better way to go. Whether I would carve that in stone, without any exception, 
> probably not.

I think it's a tradeoff really.
Unrolling in place is great, don't get me wrong, but it's also easy to
miss things when adding code later on -- we actually just did that and
got another kasan report which made me factor things in to future-proof
the code.

Having a single place of truth that knows how to "untangle" and properly
free a struct, making sure it is noop for parts of the struct that
haven't been initialized yet, is less of a burden for me to think about.


... Just happened to be wrong about the "making sure it's noop" part
because I didn't check properly and my mental model had close functions
noop on NULL clnt->priv, like free functions...
(Uh, actually it is for RDMA, so the "problem" was that it left
clnt->trans set after later errors -- but conversely virtio's close
doesn't check so also had the problem and we really must ensure we don't
close something not open)

Anyway, I've sent a couple of patch (even fixing up the order to match
in create/destroy), I'll consider this closed.

-- 
Dominique