Re: [RFC PATCH 00/10] vfio: Device memory DMA mapping improvements

From: Jason Gunthorpe <jgg@nvidia.com>
To: Alex Williamson <alex.williamson@redhat.com>
Cc: <cohuck@redhat.com>, <kvm@vger.kernel.org>,
	<linux-kernel@vger.kernel.org>, <peterx@redhat.com>
Subject: Re: [RFC PATCH 00/10] vfio: Device memory DMA mapping improvements
Date: Mon, 22 Feb 2021 14:00:31 -0400	[thread overview]
Message-ID: <20210222180031.GR4247@nvidia.com> (raw)
In-Reply-To: <161401167013.16443.8389863523766611711.stgit@gimli.home>

On Mon, Feb 22, 2021 at 09:50:22AM -0700, Alex Williamson wrote:
> This is a re-implementation of [1] following suggestions and code from
> Jason Gunthorpe.  This is lightly tested but seems functional and
> throws no lockdep warnings.  In this series we tremendously simplify
> zapping of vmas mapping device memory using unmap_mapping_range(), we
> create a protocol for looking up a vfio_device from a vma and provide
> an interface to get a reference from that vma, using that device
> reference, the caller can register a notifier for the device to
> trigger on events such as device release.  This notifier is only
> enabled here for vfio-pci, but both the vma policy and the notifier
> trigger should be trivial to add to any vfio bus driver after RFC.
> 
> Does this look more like the direction we should go?

Yep, it seems pretty good already, see my remarks in each patch

For security only vfio_device's that have been enabled for P2P, set
the vm_pgoff to the pfn, and trigger invalidation, should be used with
this mechanism.

I'd add some global opt-in so vfio_device_get_from_vma() will refuse
to return vfio_device's that don't declare they have support.

Add a flags member to vfio_device_ops would get it done fairly cleanly

Jason