[RFC,00/19] mm/gup: remove FOLL_FORCE usage from drivers (reliable R/O long-term pinning)

Message

David Hildenbrand Nov. 7, 2022, 4:17 p.m. UTC

For now, we did not support reliable R/O long-term pinning in COW mappings.
That means, if we would trigger R/O long-term pinning in MAP_PRIVATE
mapping, we could end up pinning the (R/O-mapped) shared zeropage or a
pagecache page.

The next write access would trigger a write fault and replace the pinned
page by an exclusive anonymous page in the process page table; whatever the
process would write to that private page copy would not be visible by the
owner of the previous page pin: for example, RDMA could read stale data.
The end result is essentially an unexpected and hard-to-debug memory
corruption.

Some drivers tried working around that limitation by using
"FOLL_FORCE|FOLL_WRITE|FOLL_LONGTERM" for R/O long-term pinning for now.
FOLL_WRITE would trigger a write fault, if required, and break COW before
pinning the page. FOLL_FORCE is required because the VMA might lack write
permissions, and drivers wanted to make that working as well, just like
one would expect (no write access, but still triggering a write access to
break COW).

However, that is not a practical solution, because
(1) Drivers that don't stick to that undocumented and debatable pattern
    would still run into that issue. For example, VFIO only uses
    FOLL_LONGTERM for R/O long-term pinning.
(2) Using FOLL_WRITE just to work around a COW mapping + page pinning
    limitation is unintuitive. FOLL_WRITE would, for example, mark the
    page softdirty or trigger uffd-wp, even though, there actually isn't
    going to be any write access.
(3) The purpose of FOLL_FORCE is debug access, not access without lack of
    VMA permissions by arbitrarty drivers.

So instead, make R/O long-term pinning work as expected, by breaking COW
in a COW mapping early, such that we can remove any FOLL_FORCE usage from
drivers. More details in patch #8.

Patches #1--#3 add COW tests for non-anonymous pages.
Patches #4--#7 prepare core MM for extended FAULT_FLAG_UNSHARE support in
COW mappings.
Patch #8 implements reliable R/O long-term pinning in COW mappings
Patches #9--#19 remove any FOLL_FORCE usage from drivers.

I'm refraining from CCing all driver maintainers on the whole patch set.

Cc: Linus Torvalds <[e-mail]>
Cc: Andrew Morton <[e-mail]>
Cc: Jason Gunthorpe <[e-mail]>
Cc: John Hubbard <[e-mail]>
Cc: Peter Xu <[e-mail]>
Cc: Greg Kroah-Hartman <[e-mail]>
Cc: Andrea Arcangeli <[e-mail]>
Cc: Hugh Dickins <[e-mail]>
Cc: Nadav Amit <[e-mail]>
Cc: Vlastimil Babka <[e-mail]>
Cc: Matthew Wilcox (Oracle) <[e-mail]>
Cc: Mike Kravetz <[e-mail]>
Cc: Muchun Song <[e-mail]>
Cc: Shuah Khan <[e-mail]
Cc: Lucas Stach <[e-mail]>
Cc: David Airlie <[e-mail]>
Cc: Oded Gabbay <[e-mail]>
Cc: Arnd Bergmann <[e-mail]>

David Hildenbrand (19):
  selftests/vm: anon_cow: prepare for non-anonymous COW tests
  selftests/vm: cow: basic COW tests for non-anonymous pages
  selftests/vm: cow: R/O long-term pinning reliability tests for
    non-anon pages
  mm: add early FAULT_FLAG_UNSHARE consistency checks
  mm: add early FAULT_FLAG_WRITE consistency checks
  mm: rework handling in do_wp_page() based on private vs. shared
    mappings
  mm: don't call vm_ops->huge_fault() in wp_huge_pmd()/wp_huge_pud() for
    private mappings
  mm: extend FAULT_FLAG_UNSHARE support to anything in a COW mapping
  mm/gup: reliable R/O long-term pinning in COW mappings
  RDMA/umem: remove FOLL_FORCE usage
  RDMA/usnic: remove FOLL_FORCE usage
  RDMA/siw: remove FOLL_FORCE usage
  media: videobuf-dma-sg: remove FOLL_FORCE usage
  drm/etnaviv: remove FOLL_FORCE usage
  media: pci/ivtv: remove FOLL_FORCE usage
  mm/frame-vector: remove FOLL_FORCE usage
  drm/exynos: remove FOLL_FORCE usage
  RDMA/hw/qib/qib_user_pages: remove FOLL_FORCE usage
  habanalabs: remove FOLL_FORCE usage

 drivers/gpu/drm/etnaviv/etnaviv_gem.c         |   8 +-
 drivers/gpu/drm/exynos/exynos_drm_g2d.c       |   2 +-
 drivers/infiniband/core/umem.c                |   8 +-
 drivers/infiniband/hw/qib/qib_user_pages.c    |   2 +-
 drivers/infiniband/hw/usnic/usnic_uiom.c      |   9 +-
 drivers/infiniband/sw/siw/siw_mem.c           |   9 +-
 drivers/media/common/videobuf2/frame_vector.c |   2 +-
 drivers/media/pci/ivtv/ivtv-udma.c            |   2 +-
 drivers/media/pci/ivtv/ivtv-yuv.c             |   5 +-
 drivers/media/v4l2-core/videobuf-dma-sg.c     |  14 +-
 drivers/misc/habanalabs/common/memory.c       |   3 +-
 include/linux/mm.h                            |  27 +-
 include/linux/mm_types.h                      |   8 +-
 mm/gup.c                                      |  10 +-
 mm/huge_memory.c                              |   5 +-
 mm/hugetlb.c                                  |  12 +-
 mm/memory.c                                   |  97 +++--
 tools/testing/selftests/vm/.gitignore         |   2 +-
 tools/testing/selftests/vm/Makefile           |  10 +-
 tools/testing/selftests/vm/check_config.sh    |   4 +-
 .../selftests/vm/{anon_cow.c => cow.c}        | 387 +++++++++++++++++-
 tools/testing/selftests/vm/run_vmtests.sh     |   8 +-
 22 files changed, 516 insertions(+), 118 deletions(-)
 rename tools/testing/selftests/vm/{anon_cow.c => cow.c} (74%)

Comments

Linus Torvalds Nov. 7, 2022, 5:27 p.m. UTC | #1

On Mon, Nov 7, 2022 at 8:18 AM David Hildenbrand <[e-mail]> wrote:
>
> So instead, make R/O long-term pinning work as expected, by breaking COW
> in a COW mapping early, such that we can remove any FOLL_FORCE usage from
> drivers.

Nothing makes me unhappy from a quick scan through these patches.

And I'd really love to just have this long saga ended, and FOLL_FORCE
finally relegated to purely ptrace accesses.

So an enthusiastic Ack from me.

                   Linus

David Hildenbrand Nov. 8, 2022, 9:29 a.m. UTC | #2

On 07.11.22 18:27, Linus Torvalds wrote:
> On Mon, Nov 7, 2022 at 8:18 AM David Hildenbrand <[e-mail]> wrote:
>>
>> So instead, make R/O long-term pinning work as expected, by breaking COW
>> in a COW mapping early, such that we can remove any FOLL_FORCE usage from
>> drivers.
> 
> Nothing makes me unhappy from a quick scan through these patches.
> 
> And I'd really love to just have this long saga ended, and FOLL_FORCE
> finally relegated to purely ptrace accesses.
> 
> So an enthusiastic Ack from me.

Thanks Linus! My hope is that we can remove it from all drivers and not 
have to leave it in for some corner cases; so far it looks promising.

David Hildenbrand Nov. 14, 2022, 8:07 a.m. UTC | #3

On 14.11.22 07:03, Christoph Hellwig wrote:
> On Mon, Nov 07, 2022 at 09:27:23AM -0800, Linus Torvalds wrote:
>> And I'd really love to just have this long saga ended, and FOLL_FORCE
>> finally relegated to purely ptrace accesses.
> 
> At that point we should also rename it to FOLL_PTRACE to make that
> very clear, and also break anything in-flight accidentally readding it,
> which I'd otherwise expect to happen.

Good idea; I'll include a patch in v1.