Re: [PATCH v2 2/2] mm/madvise: allow guard page install/remove under VMA lock

[Chris Mason <clm@meta.com>]

On Mon, 10 Nov 2025 17:22:58 +0000 Lorenzo Stoakes <lorenzo.stoakes@oracle.com> wrote:

> We only need to keep the page table stable so we can perform this operation
> under the VMA lock. PTE installation is stabilised via the PTE lock.
> 
> One caveat is that, if we prepare vma->anon_vma we must hold the mmap read
> lock. We can account for this by adapting the VMA locking logic to
> explicitly check for this case and prevent a VMA lock from being acquired
> should it be the case.
> 
> This check is safe, as while we might be raced on anon_vma installation,
> this would simply make the check conservative, there's no way for us to see
> an anon_vma and then for it to be cleared, as doing so requires the
> mmap/VMA write lock.
> 
> We abstract the VMA lock validity logic to is_vma_lock_sufficient() for
> this purpose, and add prepares_anon_vma() to abstract the anon_vma logic.
> 
> In order to do this we need to have a way of installing page tables
> explicitly for an identified VMA, so we export walk_page_range_vma() in an
> unsafe variant - walk_page_range_vma_unsafe() and use this should the VMA
> read lock be taken.
> 
> We additionally update the comments in madvise_guard_install() to more
> accurately reflect the cases in which the logic may be reattempted,
> specifically THP huge pages being present.
> 

Hi Lorenzo, one more flagged by AI:

> diff --git a/mm/madvise.c b/mm/madvise.c
> index 7ed5bedb8f8ef..2a165e9beb5bb 100644

[ ... ]

> @@ -1173,8 +1180,13 @@ static long madvise_guard_install(struct madvise_behavior *madv_behavior)
>  		unsigned long nr_pages = 0;
>
>  		/* Returns < 0 on error, == 0 if success, > 0 if zap needed. */
> -		err = walk_page_range_mm_unsafe(vma->vm_mm, range->start,
> -				range->end, &guard_install_walk_ops, &nr_pages);
> +		if (madv_behavior->lock_mode == MADVISE_VMA_READ_LOCK)
> +			err = walk_page_range_vma_unsafe(madv_behavior->vma,
> +					range->start, range->end, &walk_ops,
> +					&nr_pages);
> +		else
> +			err = walk_page_range_mm_unsafe(vma->vm_mm, range->start,
> +					range->end, &walk_ops, &nr_pages);
>  		if (err < 0)
>  			return err;
>
> @@ -1195,8 +1207,7 @@ static long madvise_guard_install(struct madvise_behavior *madv_behavior)
>  	}

When err > 0 (zap needed), zap_page_range_single() is called below
regardless of lock_mode.  If lock_mode == MADVISE_VMA_READ_LOCK (only VMA
lock is held), and the zap range contains a THP PUD mapping that requires
splitting, the call chain:

  madvise_guard_install()
    -> zap_page_range_single()
      -> zap_page_range_single_batched()
        -> unmap_single_vma()
          -> unmap_page_range()
            -> zap_p4d_range()
              -> zap_pud_range()

will hit mmap_assert_locked() in zap_pud_range() (mm/memory.c) when
pud_trans_huge() is true and partial PUD splitting is required.

The assertion will trigger a WARN_ON because only the VMA lock is held,
not the mmap lock.

Is there an additional check needed before calling zap_page_range_single()
when in VMA lock mode, or should this path fall back to mmap lock when THP
is possible?

> @@ -1715,9 +1763,9 @@ static enum madvise_lock_mode get_lock_mode(struct madvise_behavior *madv_behavi
>  	case MADV_POPULATE_READ:
>  	case MADV_POPULATE_WRITE:
>  	case MADV_COLLAPSE:
> +		return MADVISE_MMAP_READ_LOCK;
>  	case MADV_GUARD_INSTALL:
>  	case MADV_GUARD_REMOVE:
> -		return MADVISE_MMAP_READ_LOCK;
>  	case MADV_DONTNEED:
>  	case MADV_DONTNEED_LOCKED:
>  	case MADV_FREE:

This change moves MADV_GUARD_INSTALL to use MADVISE_VMA_READ_LOCK, but
zap_page_range_single() called later in madvise_guard_install() may
require the mmap lock for THP PUD splitting as noted above.