Re: [PATCH v11 0/6] KASAN for powerpc64 radix

From: Christophe Leroy
Date: Mon Mar 22 2021 - 10:33:48 EST

Le 19/03/2021 à 15:40, Daniel Axtens a écrit :
Building on the work of Christophe, Aneesh and Balbir, I've ported
KASAN to 64-bit Book3S kernels running on the Radix MMU.

v11 applies to next-20210317. I had hoped to have it apply to
powerpc/next but once again there are changes in the kasan core that
clash. Also, thanks to mpe for fixing a build break with KASAN off.

I'm not sure how best to progress this towards actually being merged
when it has impacts across subsystems. I'd appreciate any input. Maybe
the first four patches could go in via the kasan tree, that should
make things easier for powerpc in a future cycle?

v10 rebases on top of next-20210125, fixing things up to work on top
of the latest changes, and fixing some review comments from
Christophe. I have tested host and guest with 64k pages for this spin.

There is now only 1 failing KUnit test: kasan_global_oob - gcc puts
the ASAN init code in a section called '.init_array'. Powerpc64 module
loading code goes through and _renames_ any section beginning with
'.init' to begin with '_init' in order to avoid some complexities
around our 24-bit indirect jumps. This means it renames '.init_array'
to '_init_array', and the generic module loading code then fails to
recognise the section as a constructor and thus doesn't run it. This
hack dates back to 2003 and so I'm not going to try to unpick it in
this series. (I suspect this may have previously worked if the code
ended up in .ctors rather than .init_array but I don't keep my old
binaries around so I have no real way of checking.)

(The previously failing stack tests are now skipped due to more
accurate configuration settings.)

Details from v9: This is a significant reworking of the previous
versions. Instead of the previous approach which supported inline
instrumentation, this series provides only outline instrumentation.

To get around the problem of accessing the shadow region inside code we run
with translations off (in 'real mode'), we we restrict checking to when
translations are enabled. This is done via a new hook in the kasan core and
by excluding larger quantites of arch code from instrumentation. The upside
is that we no longer require that you be able to specify the amount of
physically contiguous memory on the system at compile time. Hopefully this
is a better trade-off. More details in patch 6.

kexec works. Both 64k and 4k pages work. Running as a KVM host works, but
nothing in arch/powerpc/kvm is instrumented. It's also potentially a bit
fragile - if any real mode code paths call out to instrumented code, things
will go boom.

In the discussion we had long time ago, , I challenged you on why it was not possible to implement things the same way as other architectures, in extenso with an early mapping.

Your first answer was that too many things were done in real mode at startup. After some discussion you said that finally there was not that much things at startup but the issue was KVM.

Now you say that instrumentation on KVM is fully disabled.

So my question is, if KVM is not a problem anymore, why not go the standard way with an early shadow ? Then you could also support inline instrumentation.