Date:   Tue, 26 Oct 2021 13:33:20 +0200
From:   Alexandre Ghiti <alexandre.ghiti@...onical.com>
To:     Dmitry Vyukov <dvyukov@...gle.com>
Cc:     Alexandre ghiti <alex@...ti.fr>,
        Nathan Chancellor <nathan@...nel.org>,
        Palmer Dabbelt <palmer@...belt.com>, elver@...gle.com,
        akpm@...ux-foundation.org, ryabinin.a.a@...il.com,
        glider@...gle.com, andreyknvl@...il.com, ndesaulniers@...gle.com,
        Arnd Bergmann <arnd@...db.de>, kasan-dev@...glegroups.com,
        linux-kernel@...r.kernel.org, llvm@...ts.linux.dev,
        linux-riscv@...ts.infradead.org,
        Paul Walmsley <paul.walmsley@...ive.com>,
        aou@...s.berkeley.edu, linux-mm@...ck.org
Subject: Re: [PATCH] kasan: Always respect CONFIG_KASAN_STACK

On Tue, Oct 26, 2021 at 6:48 AM Dmitry Vyukov <dvyukov@...gle.com> wrote:
>
> On Tue, 26 Oct 2021 at 06:39, Alexandre Ghiti
> <alexandre.ghiti@...onical.com> wrote:
> >
> > Hi,
> >
> > On Fri, Oct 15, 2021 at 3:08 PM Alexandre ghiti <alex@...ti.fr> wrote:
> > >
> > > On 10/14/21 8:31 PM, Alex Ghiti wrote:
> > > > Hi Nathan,
> > > >
> > > > Le 14/10/2021 à 18:55, Nathan Chancellor a écrit :
> > > >> On Fri, Oct 08, 2021 at 11:46:55AM -0700, Palmer Dabbelt wrote:
> > > >>> On Thu, 23 Sep 2021 07:59:46 PDT (-0700), nathan@...nel.org wrote:
> > > >>>> On Thu, Sep 23, 2021 at 12:07:17PM +0200, Marco Elver wrote:
> > > >>>>> On Wed, 22 Sept 2021 at 22:55, Nathan Chancellor
> > > >>>>> <nathan@...nel.org> wrote:
> > > >>>>>> Currently, the asan-stack parameter is only passed along if
> > > >>>>>> CFLAGS_KASAN_SHADOW is not empty, which requires
> > > >>>>>> KASAN_SHADOW_OFFSET to
> > > >>>>>> be defined in Kconfig so that the value can be checked. In RISC-V's
> > > >>>>>> case, KASAN_SHADOW_OFFSET is not defined in Kconfig, which means
> > > >>>>>> that
> > > >>>>>> asan-stack does not get disabled with clang even when
> > > >>>>>> CONFIG_KASAN_STACK
> > > >>>>>> is disabled, resulting in large stack warnings with allmodconfig:
> > > >>>>>>
> > > >>>>>> drivers/video/fbdev/omap2/omapfb/displays/panel-lgphilips-lb035q02.c:117:12:
> > > >>>>>>
> > > >>>>>> error: stack frame size (14400) exceeds limit (2048) in function
> > > >>>>>> 'lb035q02_connect' [-Werror,-Wframe-larger-than]
> > > >>>>>> static int lb035q02_connect(struct omap_dss_device *dssdev)
> > > >>>>>>             ^
> > > >>>>>> 1 error generated.
> > > >>>>>>
> > > >>>>>> Ensure that the value of CONFIG_KASAN_STACK is always passed
> > > >>>>>> along to
> > > >>>>>> the compiler so that these warnings do not happen when
> > > >>>>>> CONFIG_KASAN_STACK is disabled.
> > > >>>>>>
> > > >>>>>> Link: https://github.com/ClangBuiltLinux/linux/issues/1453
> > > >>>>>> References: 6baec880d7a5 ("kasan: turn off asan-stack for clang-8
> > > >>>>>> and earlier")
> > > >>>>>> Signed-off-by: Nathan Chancellor <nathan@...nel.org>
> > > >>>>>
> > > >>>>> Reviewed-by: Marco Elver <elver@...gle.com>
> > > >>>>
> > > >>>> Thanks!
> > > >>>>
> > > >>>>> [ Which tree are you planning to take it through? ]
> > > >>>>
> > > >>>> Gah, I was intending for it to go through -mm, then I cc'd neither
> > > >>>> Andrew nor linux-mm... :/ Andrew, do you want me to resend or can you
> > > >>>> grab it from LKML?
> > > >>>
> > > >>> Acked-by: Palmer Dabbelt <palmerdabbelt@...gle.com>
> > > >>>
> > > >>> (assuming you still want it through somewhere else)
> > > >>
> > > >> Thanks, it is now in mainline as commit 19532869feb9 ("kasan: always
> > > >> respect CONFIG_KASAN_STACK").
> > > >>
> > > >>>>> Note, arch/riscv/include/asm/kasan.h mentions KASAN_SHADOW_OFFSET in
> > > >>>>> comment (copied from arm64). Did RISC-V just forget to copy over the
> > > >>>>> Kconfig option?
> > > >>>>
> > > >>>> I do see it defined in that file as well but you are right that
> > > >>>> they did
> > > >>>> not copy the Kconfig logic, even though it was present in the tree
> > > >>>> when
> > > >>>> RISC-V KASAN was implemented. Perhaps they should so that they get
> > > >>>> access to the other flags in the "else" branch?
> > > >>>
> > > >>> Ya, looks like we just screwed this up.  I'm seeing some warnings like
> > > >>>
> > > >>>     cc1: warning: ‘-fsanitize=kernel-address’ with stack protection
> > > >>> is not supported without ‘-fasan-shadow-offset=’ for this target
> > > >>
> > > >> Hmmm, I thought I did a GCC build with this change but I must not have
> > > >> :/
> > > >>
> > > >>> which is how I ended up here, I'm assuming that's what you're
> > > >>> talking about
> > > >>> here?  LMK if you were planning on sending along a fix or if you
> > > >>> want me to
> > > >>> go figure it out.
> > > >>
> > > >> I took a look at moving the logic into Kconfig like arm64 before sending
> > > >> this change and I did not really understand it well enough to do so. I
> > > >> think it would be best if you were able to do that so that nothing gets
> > > >> messed up.
> > > >>
> > > >
> > > > I'll do it tomorrow, I'm the last one who touched kasan on riscv :)
> > > >
> > >
> > > Adding KASAN_SHADOW_OFFSET config makes kasan kernel fails to boot.
> > > It receives a *write* fault at the beginning of a memblock_alloc
> > > function while populating the kernel shadow memory: the trap address is
> > > in the kasan shadow virtual address range and this corresponds to a
> > > kernel address in init_stack. The question is: how do I populate the
> > > stack shadow mapping without using memblock API? It's weird, I don't
> > > find anything on other architectures.
> >
> > @kasan: Any idea what we are doing wrong in riscv to encounter the
> > above situation?
>
> Hi Alex, Palmer,
>
> The patch changes the definition of the KASAN_SHADOW_OFFSET const.
> Does it's value change as a result or not? Have you tried to print it
> before/after?
> If value does not change, then this is more mysterious. If it changes,
> then there lots of possible explanations (points to unmapped region,
> overlaps with something), but we need to know values before/after to

So I debugged a bit more what happened here, and actually the culprit
is the call to kasan_populate_early_shadow at the beginning of
kasan_init which write-protects the access to kasan_early_shadow_page
and hence the write fault later when using memblock. I don't see the
point of this call anyway since we populate swapper_pg_dir in
kasan_early_init and then we write-protect the access to
kasan_early_shadow_page at the end of kasan_init.

But that may not be ideal, so I'm open to a better suggestion than
just removing the call to kasan_populate_early_shadow.

Sorry I did not dig further before asking and thanks for your time,

Alex

> answer this.
>
>
> > Thanks,
> >
> > Alex
> >
> > >
> > > And just a short note: I have realized this will break with the sv48
> > > patchset as we decide at runtime the address space width and the kasan
> > > shadow start address is different between sv39 and sv48. I will have to
> > > do like x86 and move the kasan shadow start at the end of the address
> > > space so that it is the same for both sv39 and sv48.
> > >
> > > Thanks,
> > >
> > > Alex
> > >
> > >
> > > > Thanks,
> > > >
> > > > Alex
> > > >
> > > >> Cheers,
> > > >> Nathan
> > > >>
> > > >> _______________________________________________
> > > >> linux-riscv mailing list
> > > >> linux-riscv@...ts.infradead.org
> > > >> http://lists.infradead.org/mailman/listinfo/linux-riscv
> > > >>
> > > >
> > > > _______________________________________________
> > > > linux-riscv mailing list
> > > > linux-riscv@...ts.infradead.org
> > > > http://lists.infradead.org/mailman/listinfo/linux-riscv
> > >
> > > _______________________________________________
> > > linux-riscv mailing list
> > > linux-riscv@...ts.infradead.org
> > > http://lists.infradead.org/mailman/listinfo/linux-riscv

Powered by blists - more mailing lists