mbox series

[00/35] kasan: add hardware tag-based mode for arm64

Message ID cover.1597425745.git.andreyknvl@google.com (mailing list archive)
Headers show
Series kasan: add hardware tag-based mode for arm64 | expand

Message

Andrey Konovalov Aug. 14, 2020, 5:26 p.m. UTC
This patchset adds a new hardware tag-based mode to KASAN [1]. The new mode
is similar to the existing software tag-based KASAN, but relies on arm64
Memory Tagging Extension (MTE) [2] to perform memory and pointer tagging
(instead of shadow memory and compiler instrumentation).

This patchset is available here:

https://github.com/xairy/linux/tree/up-kasan-mte-v1

and has also been uploaded to the Linux kernel Gerrit instance:

https://linux-review.googlesource.com/c/linux/kernel/git/torvalds/linux/+/2700

This patchset is based on the v7 of the user MTE patches [3], along with
some KASAN patches cherry-picked from the current mainline.

This patchset consists of three parts:

1. Rework KASAN code to allow easier integration of the hardware tag-based
   mode.
2. Introduce core in-kernel MTE routines.
3. Combine the two parts together and introduce the new mode.

For testing in QEMU hardware tag-based KASAN requires:

1. QEMU built from master [4] with these patches [5] on top.
2. GCC version 10.

[1] https://www.kernel.org/doc/html/latest/dev-tools/kasan.html
[2] https://community.arm.com/developer/ip-products/processors/b/processors-ip-blog/posts/enhancing-memory-safety
[3] git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux for-next/mte
[4] https://github.com/qemu/qemu
[5] https://lists.gnu.org/archive/html/qemu-devel/2020-08/msg02677.html

====== Overview

The underlying ideas of the approach used by hardware tag-based KASAN are:

1. By relying on the Top Byte Ignore (TBI) arm64 CPU feature, pointer tags
   are stored in the top byte of each kernel pointer.

2. With the Memory Tagging Extension (MTE) arm64 CPU feature, memory tags
   for kernel memory allocations are stored in a dedicated memory not
   accessible via normal instuctions.

3. On each memory allocation, a random tag is generated, embedded it into
   the returned pointer, and the corresponding memory is tagged with the
   same tag value.

4. With MTE the CPU performs a check on each memory access to make sure
   that the pointer tag matches the memory tag.

5. On a tag mismatch the CPU generates a tag fault, and a KASAN report is
   printed.

Same as other KASAN modes, hardware tag-based KASAN is intended as a
debugging feature at this point.

====== Rationale

There are two main reasons for this new hardware tag-based mode:

1. Previously implemented software tag-based KASAN is being successfully
   used on dogfood testing devices due to its low memory overhead (as
   initially planned). The new hardware mode keeps the same low memory
   overhead, and is expected to have significantly lower performance
   impact, due to the tag checks being performed by the hardware.
   Therefore the new mode can be used as a better alternative in dogfood
   testing for hardware that supports MTE.

2. The new mode lays the groundwork for the planned in-kernel MTE-based
   memory corruption mitigation to be used in production.

====== Technical details

From the implementation perspective, hardware tag-based KASAN is almost
identical to the software mode. The key difference is using MTE for
assigning and checking tags.

Compared to the software mode, the hardware mode uses 4 bits per tag, as
dictated by MTE. Pointer tags are stored in bits [56:60), the top 4 bits
have the normal value 0xF. Having less distict tags increases the
probablity of false negatives (from ~1/256 to ~1/16) in certain cases.

Only synchronous exceptions are set up and used by hardware tag-based KASAN.

====== Benchmarks

Note: all measurements have been performed with software emulation of Memory
Tagging Extension, performance numbers for hardware tag-based KASAN on the
actual hardware are expected to be better.

Boot time [1]:
* 2.8 sec for clean kernel
* 5.7 sec for hardware tag-based KASAN
* 11.8 sec for software tag-based KASAN
* 11.6 sec for generic KASAN

Slab memory usage after boot [2]:
* 7.0 kb for clean kernel
* 9.7 kb for hardware tag-based KASAN
* 9.7 kb for software tag-based KASAN
* 41.3 kb for generic KASAN

Measurements have been performed with:
* defconfig-based configs
* Manually built QEMU master with these patches on top [3]
* QEMU arguments: -machine virt,mte=on -cpu max
* CONFIG_KASAN_STACK_ENABLE disabled
* CONFIG_KASAN_INLINE enabled
* clang-10 as the compiler and gcc-10 as the assembler
    
[1] Time before the ext4 driver is initialized.
[2] Measured as `cat /proc/meminfo | grep Slab`.
[3] https://lists.gnu.org/archive/html/qemu-devel/2020-08/msg02677.html

====== Notes

The cover letter for software tag-based KASAN patchset can be found here:

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=0116523cfffa62aeb5aa3b85ce7419f3dae0c1b8

Andrey Konovalov (29):
  kasan: KASAN_VMALLOC depends on KASAN_GENERIC
  kasan: group vmalloc code
  kasan: shadow declarations only for software modes
  kasan: rename (un)poison_shadow to (un)poison_memory
  kasan: rename KASAN_SHADOW_* to KASAN_GRANULE_*
  kasan: only build init.c for software modes
  kasan: split out shadow.c from common.c
  kasan: rename generic/tags_report.c files
  kasan: don't duplicate config dependencies
  kasan: hide invalid free check implementation
  kasan: decode stack frame only with KASAN_STACK_ENABLE
  kasan, arm64: only init shadow for software modes
  kasan, arm64: only use kasan_depth for software modes
  kasan: rename addr_has_shadow to addr_has_metadata
  kasan: rename print_shadow_for_address to print_memory_metadata
  kasan: kasan_non_canonical_hook only for software modes
  kasan: rename SHADOW layout macros to META
  kasan: separate metadata_fetch_row for each mode
  kasan: don't allow SW_TAGS with ARM64_MTE
  kasan: introduce CONFIG_KASAN_HW_TAGS
  kasan, arm64: align allocations for HW_TAGS
  kasan: define KASAN_GRANULE_SIZE for HW_TAGS
  kasan, x86, s390: update undef CONFIG_KASAN
  kasan, arm64: expand CONFIG_KASAN checks
  kasan, arm64: implement HW_TAGS runtime
  kasan, arm64: print report from tag fault handler
  kasan, slub: reset tags when accessing metadata
  kasan, arm64: enable CONFIG_KASAN_HW_TAGS
  kasan: add documentation for hardware tag-based mode

Vincenzo Frascino (6):
  arm64: mte: Add in-kernel MTE helpers
  arm64: mte: Add in-kernel tag fault handler
  arm64: mte: Enable in-kernel MTE
  arm64: mte: Convert gcr_user into an exclude mask
  arm64: mte: Switch GCR_EL1 in kernel entry and exit
  kasan, arm64: Enable TBI EL1

 Documentation/dev-tools/kasan.rst         |  75 ++-
 arch/arm64/Kconfig                        |   3 +-
 arch/arm64/Makefile                       |   2 +-
 arch/arm64/include/asm/assembler.h        |   2 +-
 arch/arm64/include/asm/cache.h            |   3 +
 arch/arm64/include/asm/esr.h              |   1 +
 arch/arm64/include/asm/kasan.h            |   8 +-
 arch/arm64/include/asm/memory.h           |   6 +-
 arch/arm64/include/asm/mte.h              |  54 ++-
 arch/arm64/include/asm/mte_asm.h          |  10 +
 arch/arm64/include/asm/processor.h        |   2 +-
 arch/arm64/include/asm/string.h           |   5 +-
 arch/arm64/kernel/asm-offsets.c           |   3 +
 arch/arm64/kernel/cpufeature.c            |  11 +-
 arch/arm64/kernel/entry.S                 |  28 ++
 arch/arm64/kernel/head.S                  |   2 +-
 arch/arm64/kernel/image-vars.h            |   2 +-
 arch/arm64/kernel/mte.c                   |  87 +++-
 arch/arm64/kernel/setup.c                 |   1 -
 arch/arm64/lib/mte.S                      |  41 ++
 arch/arm64/mm/dump.c                      |   6 +-
 arch/arm64/mm/fault.c                     |  59 ++-
 arch/arm64/mm/kasan_init.c                |  22 +-
 arch/arm64/mm/proc.S                      |   2 +-
 arch/s390/boot/string.c                   |   1 +
 arch/x86/boot/compressed/misc.h           |   1 +
 include/linux/kasan-checks.h              |   2 +-
 include/linux/kasan.h                     | 110 +++--
 include/linux/mm.h                        |   2 +-
 include/linux/moduleloader.h              |   3 +-
 include/linux/page-flags-layout.h         |   2 +-
 include/linux/sched.h                     |   2 +-
 include/linux/string.h                    |   2 +-
 init/init_task.c                          |   2 +-
 kernel/fork.c                             |   4 +-
 lib/Kconfig.kasan                         |  60 ++-
 lib/test_kasan.c                          |   2 +-
 mm/kasan/Makefile                         |  21 +-
 mm/kasan/common.c                         | 554 ++--------------------
 mm/kasan/generic.c                        |  33 +-
 mm/kasan/generic_report.c                 | 165 -------
 mm/kasan/init.c                           |  10 +-
 mm/kasan/kasan.h                          |  45 +-
 mm/kasan/mte.c                            |  76 +++
 mm/kasan/report.c                         | 254 ++--------
 mm/kasan/report_generic.c                 | 331 +++++++++++++
 mm/kasan/report_mte.c                     |  47 ++
 mm/kasan/{tags_report.c => report_tags.c} |   7 +-
 mm/kasan/shadow.c                         | 509 ++++++++++++++++++++
 mm/kasan/tags.c                           |  14 +-
 mm/page_poison.c                          |   2 +-
 mm/ptdump.c                               |  13 +-
 mm/slab_common.c                          |   2 +-
 mm/slub.c                                 |  25 +-
 scripts/Makefile.lib                      |   2 +
 55 files changed, 1653 insertions(+), 1085 deletions(-)
 create mode 100644 arch/arm64/include/asm/mte_asm.h
 delete mode 100644 mm/kasan/generic_report.c
 create mode 100644 mm/kasan/mte.c
 create mode 100644 mm/kasan/report_generic.c
 create mode 100644 mm/kasan/report_mte.c
 rename mm/kasan/{tags_report.c => report_tags.c} (94%)
 create mode 100644 mm/kasan/shadow.c

Comments

Derrick McKee Sept. 14, 2020, 7:06 p.m. UTC | #1
> This patchset adds a new hardware tag-based mode to KASAN [1]. The new mode
> is similar to the existing software tag-based KASAN, but relies on arm64
> Memory Tagging Extension (MTE) [2] to perform memory and pointer tagging
> (instead of shadow memory and compiler instrumentation).

What does one need to do to enable MTE in the kernel?  I am trying to 
color some kernel module data, but when I call mte_get_mem_tag after 
calling mte_set_mem_tag_range, I get an invalid value.  I have assured 
that mte_assign_mem_tag_range has been called in mte_set_mem_tag_range, 
so the initial sanity checks pass.  Thanks.
Andrey Konovalov Sept. 14, 2020, 8:04 p.m. UTC | #2
On Mon, Sep 14, 2020 at 9:06 PM Derrick McKee <derrick.mckee@gmail.com> wrote:
>
> > This patchset adds a new hardware tag-based mode to KASAN [1]. The new mode
> > is similar to the existing software tag-based KASAN, but relies on arm64
> > Memory Tagging Extension (MTE) [2] to perform memory and pointer tagging
> > (instead of shadow memory and compiler instrumentation).
>
> What does one need to do to enable MTE in the kernel?  I am trying to
> color some kernel module data, but when I call mte_get_mem_tag after
> calling mte_set_mem_tag_range, I get an invalid value.  I have assured
> that mte_assign_mem_tag_range has been called in mte_set_mem_tag_range,
> so the initial sanity checks pass.  Thanks.

Hi Derrick,

One needs to use the latest QEMU master and provide "-machine
virt,mte=on -cpu max" arguments to it. I'll add this information into
the cover letter for v2.

Thanks!
Derrick McKee Sept. 14, 2020, 9:36 p.m. UTC | #3
On 9/14/20 4:04 PM, Andrey Konovalov wrote:
> One needs to use the latest QEMU master and provide "-machine
> virt,mte=on -cpu max" arguments to it. I'll add this information into
> the cover letter for v2.

I am using the latest QEMU as of Sep. 8, and starting the system with 
virt,mte=on, and I have also checked that the appropriate bits are set 
in sctlr_el1 and ID_AA64PFR1_EL1. Finally, when I print out the tag 
immediately before and after mte_assign_mem_tag_range I get the same result.
Derrick McKee Sept. 14, 2020, 10:41 p.m. UTC | #4
On 9/14/20 5:36 PM, Derrick McKee wrote:
> On 9/14/20 4:04 PM, Andrey Konovalov wrote:
>> One needs to use the latest QEMU master and provide "-machine
>> virt,mte=on -cpu max" arguments to it. I'll add this information into
>> the cover letter for v2.
> 
> I am using the latest QEMU as of Sep. 8, and starting the system with 
> virt,mte=on, and I have also checked that the appropriate bits are set 
> in sctlr_el1 and ID_AA64PFR1_EL1. Finally, when I print out the tag 
> immediately before and after mte_assign_mem_tag_range I get the same 
> result.

I should also mention that CONFIG_KASAN_HW_TAGS=n in my configuration, 
but I have added my own macro to enable my code and added it to the 
conditional macro in memory.h:224.  I have also ensured that 
mte_init_tags is called from setup_arch.  But still the result of the 
ldg instruction to get the address tag is always 0, despite the desired 
tag of 2.  Perhaps I am missing some other needed setup or configuration?
Andrey Konovalov Sept. 14, 2020, 10:50 p.m. UTC | #5
On Tue, Sep 15, 2020 at 12:41 AM Derrick McKee <derrick.mckee@gmail.com> wrote:
>
> On 9/14/20 5:36 PM, Derrick McKee wrote:
> > On 9/14/20 4:04 PM, Andrey Konovalov wrote:
> >> One needs to use the latest QEMU master and provide "-machine
> >> virt,mte=on -cpu max" arguments to it. I'll add this information into
> >> the cover letter for v2.
> >
> > I am using the latest QEMU as of Sep. 8, and starting the system with
> > virt,mte=on, and I have also checked that the appropriate bits are set
> > in sctlr_el1 and ID_AA64PFR1_EL1. Finally, when I print out the tag
> > immediately before and after mte_assign_mem_tag_range I get the same
> > result.
>
> I should also mention that CONFIG_KASAN_HW_TAGS=n in my configuration,
> but I have added my own macro to enable my code and added it to the
> conditional macro in memory.h:224.  I have also ensured that
> mte_init_tags is called from setup_arch.  But still the result of the
> ldg instruction to get the address tag is always 0, despite the desired
> tag of 2.  Perhaps I am missing some other needed setup or configuration?

Hm, this should work I guess.

Could you try CONFIG_KASAN_HW_TAGS=y and check if that works?
Derrick McKee Sept. 15, 2020, 9:46 a.m. UTC | #6
On 9/14/20 6:50 PM, Andrey Konovalov wrote:
> Hm, this should work I guess.
> 
> Could you try CONFIG_KASAN_HW_TAGS=y and check if that works?

This seems to work, as the tags change appropriately. So I guess I need 
to figure out what I am doing differently.  Thanks for the help.
Derrick McKee Sept. 15, 2020, 5:51 p.m. UTC | #7
On 9/15/20 5:46 AM, Derrick McKee wrote:
> This seems to work, as the tags change appropriately. So I guess I need 
> to figure out what I am doing differently.  Thanks for the help.

The problem was that the kernel uses the PAGE_KERNEL flag when 
allocating memory for modules, and the solution was to change that to 
PAGE_KERNEL_TAGGED.  It might be worth adding this new flag to the 
documentation somewhere.