Re: possible deadlock in sidtab_sid2str_put

[Ondrej Mosnacek]

On Tue, Jan 28, 2020 at 6:29 PM Paul Moore <paul@xxxxxxxxxxxxxx> wrote:
> On Tue, Jan 28, 2020 at 11:31 AM Ondrej Mosnacek <omosnace@xxxxxxxxxx> wrote:
> > ... the current rawhide 5.5.0-1 kernel seems to have some bug
> > that prevents it from booting on anything with more than one core.
> > I'll see if I can work around it somehow...
>
> I'm not sure how you are building your kernels, but gcc v10 is causing
> a lot of problems, I would suggest compiling with an earlier gcc for
> the near future until things get sorted (I'm doing the kernel-secnext
> builds on stable Fedora, not Rawhide, for now).

Right, thanks, I was using the Rawhide buildroot to build the test
kernel (derived from Rawhide dist-git source + selinux-next + the
patch). Fortunately the Rawhide kernel can be also built against the
f31 target without any additional hacks, so I managed to build a an
upstream-based kernel with GCC 9 and it didn't have the
crash-on-multi-core issue. Regardless, I wasn't able to reproduce the
syzbot crash locally, so I had to ask syzbot to test the patch from my
git tree [1] and it passed.

Nonetheless, I checked to see how the sidtab string cache + IRQ-safe
locking (assuming mostly cache hits) compares to the non-cache
situation with a category-free label
(unconfined_u:unconfined_r:unconfined_t:s0) and the cache (3.2% impact
when mostly hits) is still faster than the non-cache version (5.5%
impact best case, 65% impact worst case).

I intend to incorporate all this information into the log message and
then post the patch.

[1] https://groups.google.com/d/msg/syzkaller-bugs/1UwATFnIiW8/kOpRrjyNAAAJ

--
Ondrej Mosnacek <omosnace at redhat dot com>
Software Engineer, Security Technologies
Red Hat, Inc.