On Wed, 2014-07-09 at 18:21 +0200, Joerg Roedel wrote:
> On Tue, Jul 08, 2014 at 05:59:58PM -0400, j.glisse@xxxxxxxxx wrote:
> > +int mmput_register_notifier(struct notifier_block *nb)
> > +{
> > + return blocking_notifier_chain_register(&mmput_notifier, nb);
> > +}
> > +EXPORT_SYMBOL_GPL(mmput_register_notifier);
> > +
> > +int mmput_unregister_notifier(struct notifier_block *nb)
> > +{
> > + return blocking_notifier_chain_unregister(&mmput_notifier, nb);
> > +}
> > +EXPORT_SYMBOL_GPL(mmput_unregister_notifier);
>
> I am still not convinced that this is required. For core code that
> needs
> to hook into mmput (like aio or uprobes) it really improves code
> readability if their teardown functions are called explicitly in
> mmput.
>
> And drivers that deal with the mm can use the already existing
> mmu_notifers. That works at least for the AMD-IOMMUv2 and KFD
> drivers.
>
> Maybe HMM is different here, but then you should explain why and how
> it
> is different and why you can't add an explicit teardown function for
> HMM.
>
>
> Joerg
>
>
Joerg,
It's true I'm using the callback from AMD-IOMMUv2 when it is called
from the release notifier, but I only use that to destroy queues that
are related to the pasid that is being unbound.
KFD driver still needs either this patch, or an explicit call to
kfd_process_exit from mmput to release kfd_process object, which can't
be released in the callback. I went for the explicit call but Andrew
Morton said that this begs for converting into notifier chain. Jerome
and I followed his advice and hence this patch.
Oded��.n������g��� �a��� �&ޖ)���)��h���&������梷���� �Ǟ�m������)������^�����������v���O��zf������