Re: binderfs interferes with syzkaller?

[Dmitry Vyukov]

On Sun, Jan 26, 2020 at 10:51 AM Christian Brauner
<christian.brauner@xxxxxxxxxx> wrote:
>
> On Sat, Jan 25, 2020 at 07:13:03PM +0100, Dmitry Vyukov wrote:
> > On Sat, Jan 25, 2020 at 6:49 PM Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> > >
> > > Hi binder maintainers,
> > >
> > > It seems that something has happened and now syzbot has 0 coverage in
> > > drivers/android/binder.c:
> > > https://storage.googleapis.com/syzkaller/cover/ci-upstream-kasan-gce-root.html
> > > It covered at least something there before as it found some bugs in binder code.
> > > I _suspect_ it may be related to introduction binderfs, but it's
> > > purely based on the fact that binderfs changed lots of things there.
> > > And I see it claims to be backward compatible.
> > >
> > > syzkaller strategy to reach binder devices is to use
> > > CONFIG_ANDROID_BINDER_DEVICES to create a bunch of binderN devices (to
> > > give each test process a private one):
> > > https://github.com/google/syzkaller/blob/master/dashboard/config/upstream-kasan.config#L5671
> > >
> > > Then it knows how to open these /dev/binderN devices:
> > > https://github.com/google/syzkaller/blob/master/sys/linux/dev_binder.txt#L22
> > > and do stuff with them.
> > >
> > > Did these devices disappear or something?
> >
> > Oh, I see, it's backwards compatible if it's not enabled, right?
>
> It's backwards compatible.
> Let me give a little more detail. The legacy binder interface would
> create all devices listed in the module parameter
> CONFIG_ANDROID_BINDER_DEVICES. These devices were created using
> misc_register(&binder_device->miscdev);
> in the host's devtmpfs mount.
> If binderfs is enabled these devices are now created in the binderfs
> instance instead.
> For full backwards compatibility with old Android versions symlinks (or
> bind mounts) can be provided in the host's devtmpfs. This is what I
> recommended a few months ago.
>
> > And we enabled it because, well, enabling things generally gives more
> > coverage. I guess I will disable CONFIG_ANDROID_BINDERFS for now to
>
> I would at least try to test both somehow. It's likely that binderfs
> will replace legacy binder devices so if we could keep testing it
> somehow that would be great.
>
> > restore coverage in the binder itself.
>
> I'm not completely sure what you mean here. The binder IPC codepaths are
> nearly the same. The difference is mostly in how the device information
> is cast out before actual binder-ipc operations take place.
> In any case, testing both would obviously be preferred but binderfs
> strikes me as more worthy of testing.

I've reverted CONFIG_ANDROID_BINDERFS to "is not set" for now as the
easiest option. We've got 35% coverage of binder.c again.

By "restore coverage" I meant that even if the IPC codepaths are
nearly the same tests knew to open /dev/binderN and these disappeared,
so no test could open any binder device and get any coverage in binder
code.

If anybody wants to try to test both, you are very welcome ;)
If you are hinting that I should do it, a single person can't test all
2000 subsystems in kernel, and testing each of them requires quite
some domain expertise. This really requires involvement of developers
of the code. I am sure there is lots to improve in the IPC codepaths
as well.