[RFC 0/1] Avoid pagewalk hugepage-split race with VFIO DMA set

[Max Boone <mboone@akamai.com>]

A kernel BUG can be triggered when /proc/$PID/numa_maps reads are
hammered by one process, while said $PID is setting up DMA for large
1G-aligned memory mapped BARs.

The 1G-aligned memory mapped BARs get set up as PUD-order PFNMAPs. When
the generic page walker (mm/pagewalk.c) gets to the PUD table entry of
the memory mapped BAR in the walk_pmd_range function, it tries to
split it by:

1. deleting the PUD entry by calling split_huge_pud
2. checking whether `pud_none` is true to go to `again`

		if (walk->vma)
			split_huge_pud(walk->vma, pud, addr);
		...
		if (pud_none(*pud))
			goto again;

3. if has_install is set, it calls __pmd_alloc and further descends
into walk_pmd_range

again:
		next = pud_addr_end(addr, end);
		if (pud_none(*pud)) {
			if (has_install)
				err = __pmd_alloc(walk->mm, pud, addr);

When VFIO is setting up DMA, the PUD entry can get reinstalled between the
split_huge_pud call and the pud_none check to goto again. In such
case the walk continues to the PMD-level and an illegal read happens.

As a mitigation, I propose to skip splitting the PMD and PUD entries
that are marked as special in the walker, which are mappings that do not
wish to be associated with a "struct page". The only occurences of these
entries I found were the vfio pci and nvgrace pfnmap mappings, which do
not behave like regular memory.

For a reproduction, the `vfio-mmap-bar.py` script repeatedly DMA-maps a
1G-aligned BAR and can be used to reproduce this bug:
- https://github.com/akamaxb/repro-vfio-page-walk-race.git

Run the `vfio-mmap-bar.py` script with the device you want to passthrough,
and in the mean time, cat the `/proc/$PID/numa_maps` of that process repeatedly
in a while loop. This caused the `numa_maps` read to crash on an illegal read,
when testing it against a 128GB-sized 2nd BAR of a NVIDIA Blackwell 6000 GPU.

Signed-off-by: Max Boone <mboone@akamai.com>
Signed-off-by: Max Tottenham <mtottenh@akamai.com>

Max Boone (1):
  mm/pagewalk: don't split device-backed huge pfnmaps

 mm/pagewalk.c | 24 ++++++++++++++++++++----
 1 file changed, 20 insertions(+), 4 deletions(-)

-- 
2.34.1