[v3,0/3] arm64: improve efficiency of setting tags for user pages

Message

Peter Collingbourne May 12, 2021, 8:09 p.m. UTC

Currently we can end up touching PROT_MTE user pages twice on fault
and once on unmap. On fault, with KASAN disabled we first clear data
and then set tags to 0, and with KASAN enabled we simultaneously
clear data and set tags to the KASAN random tag, and then set tags
again to 0. On unmap, we poison the page by setting tags, but this
is less likely to find a bug than poisoning kernel pages.

This patch series fixes these inefficiencies by only touching the pages
once on fault using the DC GZVA instruction to clear both data and
tags, and providing the option to avoid poisoning user pages on free.

Peter Collingbourne (3):
  kasan: use separate (un)poison implementation for integrated init
  arm64: mte: handle tags zeroing at page allocation time
  kasan: allow freed user page poisoning to be disabled with HW tags

 arch/arm64/include/asm/mte.h   |  4 ++
 arch/arm64/include/asm/page.h  |  9 ++++-
 arch/arm64/lib/mte.S           | 20 ++++++++++
 arch/arm64/mm/fault.c          | 25 +++++++++++++
 arch/arm64/mm/proc.S           | 10 +++--
 include/linux/gfp.h            | 18 +++++++--
 include/linux/highmem.h        |  8 ++++
 include/linux/kasan.h          | 66 ++++++++++++++++++++-------------
 include/linux/page-flags.h     |  9 +++++
 include/trace/events/mmflags.h |  9 ++++-
 mm/kasan/common.c              |  4 +-
 mm/kasan/hw_tags.c             | 31 ++++++++++++++++
 mm/mempool.c                   |  6 ++-
 mm/page_alloc.c                | 67 ++++++++++++++++++++--------------
 14 files changed, 221 insertions(+), 65 deletions(-)

Comments

Peter Collingbourne May 25, 2021, 7:03 p.m. UTC | #1

On Wed, May 12, 2021 at 1:09 PM Peter Collingbourne <[e-mail]> wrote:
>
> Currently we can end up touching PROT_MTE user pages twice on fault
> and once on unmap. On fault, with KASAN disabled we first clear data
> and then set tags to 0, and with KASAN enabled we simultaneously
> clear data and set tags to the KASAN random tag, and then set tags
> again to 0. On unmap, we poison the page by setting tags, but this
> is less likely to find a bug than poisoning kernel pages.
>
> This patch series fixes these inefficiencies by only touching the pages
> once on fault using the DC GZVA instruction to clear both data and
> tags, and providing the option to avoid poisoning user pages on free.
>
> Peter Collingbourne (3):
>   kasan: use separate (un)poison implementation for integrated init
>   arm64: mte: handle tags zeroing at page allocation time
>   kasan: allow freed user page poisoning to be disabled with HW tags

Thanks Catalin for reviewing patch 2. Could someone on the KASAN side
please take a look at patches 1 and 3?

I imagine that we'll want this series to go in via the mm tree once it's ready.

Peter