Re: [RFC PATCH] mm/rmap: do not call mmu_notifier_invalidate_page() v3

[Jerome Glisse]

On Tue, Aug 29, 2017 at 12:58:45PM -0700, Linus Torvalds wrote:
> On Tue, Aug 29, 2017 at 12:16 PM, Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > And then you can check if something actually happened by catching the
> > *ATOMIC* call to mmu_notifier_invalidate_page(), setting a flag, and
> > then doing something blocking at mmu_notifier_invalidate_range_end()
> > time.
> >
> > Maybe.
> 
> Note that now I have looked more at the users, I think we actually
> just want to get rid of mmu_notifier_invalidate_page() entirely in
> favor of just calling mmu_notifier_invalidate_range_start()/end().
> 
> Nobody seems to want an atomic version of
> mmu_notifier_invalidate_page(), they are perfectly happy just getting
> those range_start/end() call instead.
> 
> HOWEVER.
> 
> There do seem to be places (eg powernv/npu-dma.c, iommu/amd_iommu_v2.c
> and ommu/intel-svm.c) that want to get the "invalidate_page()" or
> "invalidate_range()" calls, but do *not* catch the begin/end() ones.
> The "range" calls were for atomic cases, and the "page" call was for
> the few places that weren't (but should have been). They seem to do
> the same things.
> 
> So just switching from mmu_notifier_invalidate_page() to the
> "invalidate_range_start()/end()" pair instead could break those cases.
> 
> But the mmu_notifier_invalidate_range() call has always been atomic,
> afaik.  It's called from the ptep_clear_flush_notify(), which is
> called while holdin gthe ptl lock as far as I can tell.
> 
> So to handle the powernv/npu-dma.c, iommu/amd_iommu_v2.c and
> ommu/intel-svm.c correctly, _and_ get he KVM case right, we probably
> need to:
> 
>  - replace the existing mmu_notifier_invalidate_page() call with
> mmu_notifier_invalidate_range(), and make sure it's inside the locked
> region (ie fs/dax.c too - actually move it inside the lock)
> 
>  - surround the locked region with those
> mmu_notifier_invalidate_range_start()/end() calls.
> 
>  - get rid of mmu_notifier_invalidate_page() entirely, it had bad
> semantics anyway.
> 
> and from all I can tell that should work for everybody.
> 
> But maybe I'm missing something.
> 

So ignore what i just sent, i will rework it toward that direction.
I believe you are right.

Jérôme