Re: [PATCH 2/3] mm: drop VMA lock before waiting for migration

[Suren Baghdasaryan <surenb@google.com>]

On Tue, May 2, 2023 at 7:28 AM Matthew Wilcox <willy@infradead.org> wrote:
>
> On Mon, May 01, 2023 at 10:50:24AM -0700, Suren Baghdasaryan wrote:
> > migration_entry_wait does not need VMA lock, therefore it can be dropped
> > before waiting. Introduce VM_FAULT_VMA_UNLOCKED to indicate that VMA
> > lock was dropped while in handle_mm_fault().
> > Note that once VMA lock is dropped, the VMA reference can't be used as
> > there are no guarantees it was not freed.
>
> How about we introduce:
>
> void vmf_end_read(struct vm_fault *vmf)
> {
>         if (!vmf->vma)
>                 return;
>         vma_end_read(vmf->vma);
>         vmf->vma = NULL;
> }
>
> Now we don't need a new flag, and calling vmf_end_read() is idempotent.
>
> Oh, argh, we create the vmf too late.  We really need to hoist the
> creation of vm_fault to the callers of handle_mm_fault().

Yeah, unfortunately vmf does not propagate all the way up to
do_user_addr_fault which needs to know that we dropped the lock.

>