Re: [PATCH v2 0/4] mm: fix races between huge and non-huge faults in PFN mappings
From: David Hildenbrand (Arm)
Date: Wed Nov 18 2026 - 04:48:08 EST
On 11/17/26 19:31, Lorenzo Stoakes (ARM) wrote:
> A driver that implements both ->fault and ->huge_fault must anticipate
> concurrent use of both.
>
> This can happen, for example, if a user calls prctl(PR_SET_THP_DISABLE)
> concurrent with faulting, as that only holds the mmap write lock and thus
> does not exclude faulting threads stabilised by VMA lock.
>
> A number of places in the code that implements huge PFN support fail to
> take this into account correctly, which this series aims to address.
>
> Generally the shape of the issue is that faulting the same empty page table
> entry can result in the ->fault and ->huge_fault handlers racing and the
> loser finding something they didn't expect.
>
> The fault code must tolerate this, but fails to in 3 places:
>
> 1. walk_to_pmd() asserts, without the PMD lock, that the PMD entry is not
> huge, so a VM_BUG_ON() triggers. [ vfio-pci, nvgrace-gpu ]
>
> 2. insert_pfn() treats a failure to map the PTE table as an allocation
> failure and reports OOM, causing kernel accesses to the range to
> spuriously fail with -EFAULT. [ vfio-pci, nvgrace-gpu ]
>
> 3. insert_pmd() and insert_pud() assume a populated entry is huge and
> compare its PFN on write fault, warning when that of a page table does
> not match. [ mshv_vtl, drm_gem_shmem, panthor ]
>
> This was originally reported and fixed by Eliav Farber in [0], this series
> reworks patches 1-3 of that one.
>
> Patch 4 of that series (x86 PAT encoding of huge PFN mappings) is dropped -
> it is unrelated to the above, no in-tree driver is affected by it, and it
> needs a different approach which is better pursued separately with the x86
> maintainers.
>
> All three issues were reproduced on x86-64 under QEMU using vfio-pci in
> no-IOMMU mode on an ivshmem BAR, with 6 threads faulting the same address
> in each PMD or PUD aligned range of the mapping while another toggles
> PR_SET_THP_DISABLE (note that vfio-pci had to be modified to pass the write
> flag to reproduce the third issue).
>
> Without this series the BUG is hit immediately, with only patches 1 and 2
> applied tens of thousands of -EFAULTs are observed within seconds, and with
> patches 1-3 (and the vfio-pci modification) both warnings fire immediately.
>
> With the series applied 30 second runs at each of PMD and PUD size complete
> with none of the above, with or without the vfio-pci modification.
>
> [0]: https://lore.kernel.org/all/20261112051440.16140-1-farbere@xxxxxxxxxx/
>
Lorenzo,
all of these are not super urgent and can wait for the next merge window, right?
--
Cheers,
David