Message-ID: <4f2fffb3-5fb6-b5ea-a951-a7910f2439b8@c-s.fr>
Date:   Thu, 19 Dec 2019 11:05:00 +0100
From:   Christophe Leroy <christophe.leroy@....fr>
To:     Daniel Axtens <dja@...ens.net>, linux-kernel@...r.kernel.org,
        linux-mm@...ck.org, linuxppc-dev@...ts.ozlabs.org,
        kasan-dev@...glegroups.com, aneesh.kumar@...ux.ibm.com,
        bsingharora@...il.com
Cc:     Michael Ellerman <mpe@...erman.id.au>
Subject: Re: [PATCH v4 4/4] powerpc: Book3S 64-bit "heavyweight" KASAN support



Le 19/12/2019 à 10:50, Daniel Axtens a écrit :
> Christophe Leroy <christophe.leroy@....fr> writes:
> 
>> On 12/19/2019 12:36 AM, Daniel Axtens wrote:
>>> KASAN support on Book3S is a bit tricky to get right:
>>>
>>>    - It would be good to support inline instrumentation so as to be able to
>>>      catch stack issues that cannot be caught with outline mode.
>>>
>>>    - Inline instrumentation requires a fixed offset.
>>>
>>>    - Book3S runs code in real mode after booting. Most notably a lot of KVM
>>>      runs in real mode, and it would be good to be able to instrument it.
>>>
>>>    - Because code runs in real mode after boot, the offset has to point to
>>>      valid memory both in and out of real mode.
>>>
>>>       [ppc64 mm note: The kernel installs a linear mapping at effective
>>>       address c000... onward. This is a one-to-one mapping with physical
>>>       memory from 0000... onward. Because of how memory accesses work on
>>>       powerpc 64-bit Book3S, a kernel pointer in the linear map accesses the
>>>       same memory both with translations on (accessing as an 'effective
>>>       address'), and with translations off (accessing as a 'real
>>>       address'). This works in both guests and the hypervisor. For more
>>>       details, see s5.7 of Book III of version 3 of the ISA, in particular
>>>       the Storage Control Overview, s5.7.3, and s5.7.5 - noting that this
>>>       KASAN implementation currently only supports Radix.]
>>>
>>> One approach is just to give up on inline instrumentation. This way all
>>> checks can be delayed until after everything set is up correctly, and the
>>> address-to-shadow calculations can be overridden. However, the features and
>>> speed boost provided by inline instrumentation are worth trying to do
>>> better.
>>>
>>> If _at compile time_ it is known how much contiguous physical memory a
>>> system has, the top 1/8th of the first block of physical memory can be set
>>> aside for the shadow. This is a big hammer and comes with 3 big
>>> consequences:
>>>
>>>    - there's no nice way to handle physically discontiguous memory, so only
>>>      the first physical memory block can be used.
>>>
>>>    - kernels will simply fail to boot on machines with less memory than
>>>      specified when compiling.
>>>
>>>    - kernels running on machines with more memory than specified when
>>>      compiling will simply ignore the extra memory.
>>>
>>> Implement and document KASAN this way. The current implementation is Radix
>>> only.
>>>
>>> Despite the limitations, it can still find bugs,
>>> e.g. http://patchwork.ozlabs.org/patch/1103775/
>>>
>>> At the moment, this physical memory limit must be set _even for outline
>>> mode_. This may be changed in a later series - a different implementation
>>> could be added for outline mode that dynamically allocates shadow at a
>>> fixed offset. For example, see https://patchwork.ozlabs.org/patch/795211/
>>>
>>> Suggested-by: Michael Ellerman <mpe@...erman.id.au>
>>> Cc: Balbir Singh <bsingharora@...il.com> # ppc64 out-of-line radix version
>>> Cc: Christophe Leroy <christophe.leroy@....fr> # ppc32 version
>>> Signed-off-by: Daniel Axtens <dja@...ens.net>
>>>
>>> ---
>>> Changes since v3:
>>>    - Address further feedback from Christophe.
>>>    - Drop changes to stack walking, it looks like the issue I observed is
>>>      related to that particular stack, not stack-walking generally.
>>>
>>> Changes since v2:
>>>
>>>    - Address feedback from Christophe around cleanups and docs.
>>>    - Address feedback from Balbir: at this point I don't have a good solution
>>>      for the issues you identify around the limitations of the inline implementation
>>>      but I think that it's worth trying to get the stack instrumentation support.
>>>      I'm happy to have an alternative and more flexible outline mode - I had
>>>      envisoned this would be called 'lightweight' mode as it imposes fewer restrictions.
>>>      I've linked to your implementation. I think it's best to add it in a follow-up series.
>>>    - Made the default PHYS_MEM_SIZE_FOR_KASAN value 1024MB. I think most people have
>>>      guests with at least that much memory in the Radix 64s case so it's a much
>>>      saner default - it means that if you just turn on KASAN without reading the
>>>      docs you're much more likely to have a bootable kernel, which you will never
>>>      have if the value is set to zero! I'm happy to bikeshed the value if we want.
>>>
>>> Changes since v1:
>>>    - Landed kasan vmalloc support upstream
>>>    - Lots of feedback from Christophe.
>>>
>>> Changes since the rfc:
>>>
>>>    - Boots real and virtual hardware, kvm works.
>>>
>>>    - disabled reporting when we're checking the stack for exception
>>>      frames. The behaviour isn't wrong, just incompatible with KASAN.
>>>
>>>    - Documentation!
>>>
>>>    - Dropped old module stuff in favour of KASAN_VMALLOC.
>>>
>>> The bugs with ftrace and kuap were due to kernel bloat pushing
>>> prom_init calls to be done via the plt. Because we did not have
>>> a relocatable kernel, and they are done very early, this caused
>>> everything to explode. Compile with CONFIG_RELOCATABLE!
>>> ---
>>>    Documentation/dev-tools/kasan.rst            |   8 +-
>>>    Documentation/powerpc/kasan.txt              | 112 ++++++++++++++++++-
>>>    arch/powerpc/Kconfig                         |   2 +
>>>    arch/powerpc/Kconfig.debug                   |  21 ++++
>>>    arch/powerpc/Makefile                        |  11 ++
>>>    arch/powerpc/include/asm/book3s/64/hash.h    |   4 +
>>>    arch/powerpc/include/asm/book3s/64/pgtable.h |   7 ++
>>>    arch/powerpc/include/asm/book3s/64/radix.h   |   5 +
>>>    arch/powerpc/include/asm/kasan.h             |  21 +++-
>>>    arch/powerpc/kernel/prom.c                   |  61 +++++++++-
>>>    arch/powerpc/mm/kasan/Makefile               |   1 +
>>>    arch/powerpc/mm/kasan/init_book3s_64.c       |  70 ++++++++++++
>>>    arch/powerpc/platforms/Kconfig.cputype       |   1 +
>>>    13 files changed, 316 insertions(+), 8 deletions(-)
>>>    create mode 100644 arch/powerpc/mm/kasan/init_book3s_64.c
>>>
>>> diff --git a/arch/powerpc/include/asm/kasan.h b/arch/powerpc/include/asm/kasan.h
>>> index 296e51c2f066..f18268cbdc33 100644
>>> --- a/arch/powerpc/include/asm/kasan.h
>>> +++ b/arch/powerpc/include/asm/kasan.h
>>> @@ -2,6 +2,9 @@
>>>    #ifndef __ASM_KASAN_H
>>>    #define __ASM_KASAN_H
>>>    
>>> +#include <asm/page.h>
>>> +#include <asm/pgtable.h>
>>
>> What do you need asm/pgtable.h for ?
>>
>> Build failure due to circular inclusion of asm/pgtable.h:
> 
> I see there's a lot of ppc32 stuff, I clearly need to bite the bullet
> and get a ppc32 toolchain so I can squash these without chewing up any
> more of your time. I'll sort that out and send a new spin.
> 

I'm using a powerpc64 toolchain to build both ppc32 and ppc64 kernels 
(from https://mirrors.edge.kernel.org/pub/tools/crosstool/ )


Another thing, did you test PTDUMP stuff with KASAN ? It looks like 
KASAN address markers don't depend on PPC32, but are only initialised by 
populate_markers() for PPC32.

Regarding kasan.h, I think we should be able to end up with something 
where the definition of KASAN_SHADOW_OFFSET should only depend on the 
existence of CONFIG_KASAN_SHADOW_OFFSET, and where only 
KASAN_SHADOW_SIZE should depend on the target (ie PPC32 or BOOK3S64)
Everything else should be common. KASAN_END should be START+SIZE.

It looks like what you have called KASAN_SHADOW_SIZE is not similar to 
what is called KASAN_SHADOW_SIZE for PPC32, as yours only covers the 
SHADOW_SIZE for linear mem while PPC32 one covers the full space.

Christophe

Powered by blists - more mailing lists