Re: SECURITY PROBLEM: Any user can crash the kernel with TCP ZEROCOPY

[Eric Dumazet <edumazet@google.com>]

On Mon, Jan 22, 2024 at 6:12 PM Matthew Wilcox <willy@infradead.org> wrote:
>
> On Mon, Jan 22, 2024 at 05:30:18PM +0100, Eric Dumazet wrote:
> > On Mon, Jan 22, 2024 at 5:04 PM Matthew Wilcox <willy@infradead.org> wrote:
> > > I'm disappointed to have no reaction from netdev so far.  Let's see if a
> > > more exciting subject line evinces some interest.
> >
> > Hmm, perhaps some of us were enjoying their weekend ?
>
> I am all in favour of people taking time off!  However the report came
> in on Friday at 9am UTC so it had been more than a work day for anyone
> anywhere in the world without response.
>
> > I don't really know what changed recently, all I know is that TCP zero
> > copy is for real network traffic.
> >
> > Real trafic uses order-0 pages, 4K at a time.
> >
> > If can_map_frag() needs to add another safety check, let's add it.
>
> So it's your opinion that people don't actually use sendfile() from
> a local file, and we can make this fail to zerocopy?

Certainly we do not do that at Google.
I am not sure if anybody else would have used this.



 That's good
> because I had a slew of questions about what expectations we had around
> cache coherency between pages mapped this way and write()/mmap() of
> the original file.  If we can just disallow this, we don't need to
> have a discussion about it.
>
> > syzbot is usually quite good at bisections, was a bug origin found ?
>
> I have the impression that Huawei run syzkaller themselves without
> syzbot.  I suspect this bug has been there for a good long time.
> Wonder why nobody's found it before; it doesn't seem complicated for a
> fuzzer to stumble into.

I is strange syzbot (The Google fuzzer) have not found this yet, I
suspect it might be caused
by a recent change somewhere ?

A repro would definitely help, I could start a bisection.