[Intel-gfx] [PATCH 08/10] vfio-iommufd: Support iommufd for emulated VFIO devices

[Tian, Kevin]

> From: Jason Gunthorpe <jgg at nvidia.com>
> Sent: Tuesday, November 1, 2022 8:49 PM
> > > ---
> > >  drivers/gpu/drm/i915/gvt/kvmgt.c  |   3 +
> > >  drivers/s390/cio/vfio_ccw_ops.c   |   3 +
> > >  drivers/s390/crypto/vfio_ap_ops.c |   3 +
> > >  drivers/vfio/container.c          | 108 ++++++-----------------------
> > >  drivers/vfio/iommufd.c            |  57 ++++++++++++++++
> > >  drivers/vfio/vfio.h               |  10 ++-
> > >  drivers/vfio/vfio_main.c          | 110 +++++++++++++++++++++++++++++-
> > >  include/linux/vfio.h              |  14 ++++
> > >  8 files changed, 217 insertions(+), 91 deletions(-)
> >
> > mtty, mdpy and mbochs?
> 
> They don't call rw or pin_pages, so they don't need to do
> anything:
> 
> 
> 	/*
> 	 * If the driver doesn't provide this op then it means the device does
> 	 * not do DMA at all. So nothing to do.
> 	 */
> 	if (!vdev->ops->bind_iommufd)
> 		return 0;
> 

OK, I see the point of this check here.

btw It'd be good to document in vfio_device_ops that driver must provide
this op if the device does DMA.

> > > +
> > > +		/* VFIO historically tries to auto-detect a kthread */
> > > +		if (!current->mm)
> > > +			flags |= IOMMUFD_ACCESS_RW_KTHREAD;
> >
> > Can you elaborate why this cannot be put in iommufd as the default
> > policy similar to what vfio container does?
> 
> Snooping in kernel structs to try to guess the calling execution
> context is bad design. The caller should know its own context and it
> should declare positively what it is. Someday this should be lifted
> out of VFIO as well and into the drivers.
> 

with the last sentence it makes more sense. otherwise I didn't see
why putting the guess in vfio makes real difference from doing it
in iommufd as there is no vfio specific state referenced for making
this decision.