On Wed, Feb 15, 2023 at 09:17:31PM -0800, Suren Baghdasaryan wrote:

Decisions about whether VMAs can be merged, split or expanded must be
made while VMAs are protected from the changes which can affect that
decision. For example, merge_vma uses vma->anon_vma in its decision

Did you mean vma_merge()?

whether the VMA can be merged. Meanwhile, page fault handler changes
vma->anon_vma during COW operation.
Write-lock all VMAs which might be affected by a merge or split operation
before making decision how such operations should be performed.

It doesn't make sense (to me) to update vma->anon_vma during COW fault.

AFAIK children's vma->anon_vma is allocated during fork() and
page->anon_vma is updated on COW to reduce rmap walking because it's now
unshared, no?

As patch 26 just falls back to mmap_lock if no anon_vma is set,
I think we can assume nothing updates vma->anon_vma (and its interval
tree) if we are holding mmap_lock in write mode.

Or am I missing something?

--
Regards,
Hyeonggon