Re: [PATCH 1/1] RDMA/odp: convert to use HMM for ODP

[Jerome Glisse]

On Thu, Feb 21, 2019 at 03:59:37PM -0700, Jason Gunthorpe wrote:
> On Wed, Feb 20, 2019 at 05:29:24PM -0500, Jerome Glisse wrote:
> > > > 
> > > > Yes it is safe, the hmm struct has its own refcount and mirror holds a
> > > > reference on it, the mm struct itself has a reference on the mm
> > > > struct.
> > > 
> > > The issue here is that that hmm_mirror_unregister() must be a strong
> > > fence that guarentees no callback is running or will run after
> > > return. mmu_notifier_unregister did not provide that.
> > > 
> > > I think I saw locking in hmm that was doing this..
> > 
> > So pattern is:
> >     hmm_mirror_register(mirror);
> > 
> >     // Safe for driver to call within HMM with mirror no matter what
> > 
> >     hmm_mirror_unregister(mirror)
> > 
> >     // Driver must no stop calling within HMM, it would be a use after
> >     // free scenario
> 
> This statement is the opposite direction
> 
> I want to know that HMM doesn't allow any driver callbacks to be
> running after unregister - because I am going to kfree mirror and
> other memory touched by the driver callbacks.

Sorry i miss-understood your question. Yes after hmm_mirror_unregister()
you will no longer get a callback from HMM ie it is safe for you to free
any data structure associated with HMM.

Cheers,
Jérôme