Re: [PATCH v2 2/4] seccomp: Refactor notification handler to prepare for new semantics

[Kees Cook]

On Fri, May 28, 2021 at 05:27:39PM +0200, Rodrigo Campos wrote:
> On Thu, May 27, 2021 at 8:42 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> >
> > On Thu, May 27, 2021 at 01:51:13PM +0200, Rodrigo Campos wrote:
> > >
> > > Kees, as I mentioned in the linked thread, this issue is present in
> > > 5.9+ kernels. Should we add the cc to stable for this patch? Or should
> > > we cc to stable the one linked, that just fixes the issue without
> > > semantic changes to userspace?
> >
> > It sounds like the problem is with Go, using addfd, on 5.9-5.13 kernels,
> > yes?
> 
> Yes.
> 
> > Would the semantic change be a problem there? (i.e. it sounds like
> > the semantic change was fine for the 5.14+ kernels, so I'm assuming it's
> > fine for earlier ones too.)
> 
> No, I don't think it will cause any problem.
> 
> > > Just to be clear, the other patch that fixes the problem without
> > > userspace visible changes is this:
> > > https://lore.kernel.org/lkml/20210413160151.3301-1-rodrigo@xxxxxxxxxx/
> >
> > I'd prefer to use the now-in-next fix if we can. Is it possible to build
> > a test case that triggers the race so we can have some certainty that
> > any fix in -stable covers it appropriately?
> 
> I've verified that Sargun's patch also solves the problem in mainline.
> I have now also verified that it applies cleany and fixes the issue
> for linux-stable/5.10.y and linux-stable/5.12.y too (without the patch
> I see the problem, with the patch I don't see it).  5.11 is already
> EOL, so I didn't try it (probably will work as well).

Great! Thanks for doing that testing.

> The test case that I have is quite a complicated one, though. I'm
> using the PR we opened to runc to add support for seccomp notify[1]
> and a seccomp agent slightly modified from the example in the PR with
> some cgo to use addfd, and need to run it for several thousand
> iterations, as the kernel needs to be interrupted in a specific line
> and some kernel locks to be acquired in a specific order for this to
> trigger. If you think it is important, I can try to cleanup the code
> and share it, but the issue is basically what I explained here:
> https://lore.kernel.org/lkml/20210413160151.3301-2-rodrigo@xxxxxxxxxx/

Okay; yeah, sounds like that'll be hard to port sanely. :)

> Can we cc this patch to stable, then? :)

Yup, sounds good to me. I will adjust the tags.

Thanks!

-Kees

> Best,
> Rodrigo
> 
> [1]: https://github.com/opencontainers/runc/pull/2682

-- 
Kees Cook