Re: [PATCH -tip] kasan: Emit different calls for instrumentable memintrinsics

[Marco Elver]

On Fri, 10 Feb 2023 at 22:37, Andrey Konovalov <andreyknvl@xxxxxxxxx> wrote:
>
> On Fri, Feb 10, 2023 at 7:41 PM Marco Elver <elver@xxxxxxxxxx> wrote:
> >
> > On Fri, 10 Feb 2023 at 17:13, Andrey Konovalov <andreyknvl@xxxxxxxxx> wrote:
> > [...]
> > > > Probably the same should be done for SW_TAGS, because arm64 will be
> > > > GENERIC_ENTRY at one point or another as well.
> > >
> > > Yes, makes sense. I'll file a bug for this once I fully understand the
> > > consequences of these changes.
> > >
> > > > KASAN + GCC on x86 will have no mem*() instrumentation after
> > > > 69d4c0d32186, which is sad, so somebody ought to teach it the same
> > > > param as above.
> > >
> > > Hm, with that patch we would have no KASAN checking within normal mem*
> > > functions (not the ones embedded by the compiler) on GENERIC_ENTRY
> > > arches even with Clang, right?
> >
> > Yes, that's the point - normal mem*() functions cannot be instrumented
> > with GENERIC_ENTRY within noinstr functions, because the compiler
> > sometimes decides to transform normal assignments into
> > memcpy()/memset(). And if mem*() were instrumented (as it was before
> > 69d4c0d32186), that'd break things for these architectures.
> >
> > But since most code is normally instrumented, with the right compiler
> > support (which the patch here enables), we just turn mem*() in
> > instrumented functions into __asan_mem*(), and get the instrumentation
> > as before. 69d4c0d32186 already added those __asan functions. The fact
> > that KASAN used to override mem*() is just the wrong choice in a world
> > where compilers decide to inline or outline these. From an
> > instrumentation point of view at the compiler level, we need to treat
> > them like any other instrumentable instruction (loads, stores,
> > atomics, etc.): transform each instrumentable instruction into
> > something that does the right checks. Only then can we be sure that we
> > don't accidentally instrument something that shouldn't be (noinstr
> > functions), because instead of relying on the compiler, we forced
> > instrumentation on every mem*().
>
> I meant to ask whether the normal mem* calls from instrumented
> functions will also be transformed to __asan_mem*() by the compiler.
> But following the godbolt link you shared, I see that this is true.
>
> Thank you for the explanation!
>
> So the overall negative impact of these changes is that we don't get
> KASAN checking in both normal mem* calls and the ones formed by
> transforming assignments for GENERIC_ENTRY architectures with GCC and
> with older Clang. This is not great. I wonder if we then need to print
> some kind of warning when the kernel is built with these compilers.

Since these changes are already in -tip, and by judging from [1],
there really is no other way. As-is, KASAN on x86 is already broken
per [1] (though we got lucky thus far).

Printing a warning wouldn't hurt, but I think nobody would notice the
warning, and if somebody notices, they wouldn't care. Sooner or later,
we just need to make sure that test robots (syzbot, etc.) have new
compilers.

> If these changes move forward, AFAIU, we can also drop these custom
> mem* definitions for non-instrumented files for x86:
>
> https://elixir.bootlin.com/linux/latest/source/arch/x86/include/asm/string_64.h#L88

Yes, I think so.

[1] https://lore.kernel.org/all/20230112194314.845371875@xxxxxxxxxxxxx/

Last but not least are you ok with this patch? This patch ought to be
applied to the same tree as 69d4c0d32186 anyway, so this patch lives
or dies by that change.