Re: [PATCH v4] mm: x86: instrument __get/__put_kernel_nofault

[Andrey Konovalov <andreyknvl@gmail.com>]

On Sun, Sep 22, 2024 at 11:26 AM Sabyrzhan Tasbolatov
<snovitoll@gmail.com> wrote:
>
> On Sun, Sep 22, 2024 at 1:49 AM Andrey Konovalov <andreyknvl@gmail.com> wrote:
> >
> > I tried running the tests with this patch applied, but unfortunately
> > the added test fails on arm64, most likely due to missing annotations
> > in arm64 asm code.
>
> Thanks for testing it on arm64. I've checked other arch and found out
> that only s390, x86 are using <linux/instrumented.h> header with
> KASAN and friends in annotations. <linux/kasan-checks.h> is in arm64 and x86.
>
> While the current [PATCH v4] has x86 only instrumentations for
> __get/put_kernel_nofault, I think, we can take as an example copy_from_user
> solution here:
>
> https://elixir.bootlin.com/linux/v6.11-rc7/source/include/linux/uaccess.h#L162-L164
>
> , which should be a generic instrumentation of __get/put_kernel_nofault
> for all arch. I can try to make a separate PATCH with this solution.

_inline_copy_from_user appears to only be called for non-kernel
variants of copy_from_user, so you would need something different.

> > We need to either mark the added test as x86-only via
> > KASAN_TEST_NEEDS_CONFIG_ON or add annotations for arm64.
> >
> > With annotations for arm64, the test might still fail for other
> > architectures, but I think that's fine: hopefully relevant people will
> > add annotations in time. But I consider both x86 and arm64 important,
> > so we should keep the tests working there.
> >
> > If you decide to add annotations for arm64, please also test both
> > KASAN_SW_TAGS and KASAN_HW_TAGS modes.
>
> Please suggest if the solution above to make a generic instrumentation of
> __get/put_kernel_nofault is suitable.

I think the approach you have taken with adding instrument_read/write
into arch code is fine, we just need to do this for all arches.

An alternative would be common wrapper macros that calls
__get/put_kernel_nofault + instrument_read/write.

> Otherwise, for this patch as you've suggested, we can add
> KASAN_TEST_NEEDS_CONFIG_ON(test, CONFIG_X86);
> to make sure that kunit test is for x86 only and I can add arm64 kasan-checks
> with SW, HW tags in separate "mm, arm64" PATCH.

Sounds good too.