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

From: Christophe Leroy
Date: Tue Mar 23 2021 - 09:28:56 EST

Le 23/03/2021 à 02:21, Daniel Axtens a écrit :
Hi Christophe,

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.

Fair enough, I've had some trouble both understanding the problem myself
and clearly articulating it. Let me try again.

We need translations on to access the shadow area.

We reach setup_64.c::early_setup() with translations off. At this point
we don't know what MMU we're running under, or our CPU features.

What do you need to know ? Whether it is Hash or Radix, or more/different details ?

IIUC, today we only support KASAN on Radix. Would it make sense to say that a kernel built with KASAN can only run on processors having Radix capacility ? Then select CONFIG_PPC_RADIX_MMU_DEFAULT when KASAN is set, and accept that the kernel crashes if Radix is not available ?

To determine our MMU and CPU features, early_setup() calls functions
(dt_cpu_ftrs_init, early_init_devtree) that call out to generic code
like of_scan_flat_dt. We need to do this before we turn on translations
because we can't set up the MMU until we know what MMU we have.

So this puts us in a bind:

- We can't set up an early shadow until we have translations on, which
requires that the MMU is set up.

- We can't set up an MMU until we call out to generic code for FDT

So there will be calls to generic FDT parsing code that happen before the
early shadow is set up.

I see some logic in kernel/prom_init.c for detecting MMU. Can we get the information from there in order to setup the MMU ?

The setup code also prints a bunch of information about the platform
with printk() while translations are off, so it wouldn't even be enough
to disable instrumentation for bits of the generic DT code on ppc64.

I'm sure the printk() stuff can be avoided or delayed without much problems, I guess the main problem is the DT code, isn't it ?

As far as I can see the code only use udbg_printf() before MMU is on, and this could be simply skipped when KASAN is selected, I see no situation where you need early printk together with KASAN.

Does that make sense? If you can figure out how to 'square the circle'
here I'm all ears.

Yes it is a lot more clear now, thanks you. Gave a few ideas above, does it help ?

Other notes:

- There's a comment about printk() being 'safe' in early_setup(), that
refers to having a valid PACA, it doesn't mean that it's safe in any
other sense.

- KVM does indeed also run stuff with translations off but we can catch
all of that by disabling instrumentation on the real-mode handlers:
it doesn't seem to leak out to generic code. So you are right that
KVM is no longer an issue.