Re: [PATCH 1/1] vfio: remove VFIO_GROUP_NOTIFY_SET_KVM

[Matthew Rosato]

On 5/18/22 11:12 AM, Jason Gunthorpe wrote:
> On Wed, May 18, 2022 at 10:37:48AM -0400, Matthew Rosato wrote:
> > On 5/17/22 2:08 PM, Matthew Rosato wrote:
> > > Rather than relying on a notifier for associating the KVM with
> > > the group, let's assume that the association has already been
> > > made prior to device_open.  The first time a device is opened
> > > associate the group KVM with the device.
> > >
> > > Suggested-by: Jason Gunthorpe <jgg@xxxxxxxxxx>
> > > Signed-off-by: Matthew Rosato <mjrosato@xxxxxxxxxxxxx>
> >
> > ...
> >
> > > diff --git a/drivers/vfio/vfio.c b/drivers/vfio/vfio.c
> > > index cfcff7764403..c5d421eda275 100644
> > > +++ b/drivers/vfio/vfio.c
> > > @@ -10,6 +10,7 @@
> > >     * Author: Tom Lyon, pugs@xxxxxxxxx
> > >     */
> > > +#include "linux/kvm_host.h"
> > >    #include <linux/cdev.h>
> > >    #include <linux/compat.h>
> > >    #include <linux/device.h>
> > > @@ -1083,6 +1084,13 @@ static struct file *vfio_device_open(struct vfio_device *device)
> > >    	mutex_lock(&device->dev_set->lock);
> > >    	device->open_count++;
> > > +	down_write(&device->group->group_rwsem);
> > > +	if (device->open_count == 1 && device->group->kvm) {
> > > +		device->kvm = device->group->kvm;
> > > +		kvm_get_kvm(device->kvm);
> >
> > Did some more compile testing, since vfio has no hard kvm dependency,
> > kvm_get_kvm and kvm_put_kvm are an issue if KVM is a module while vfio is
> > built-in...
>
> Ugh, my other plan was to have the driver itself capture the kvm, ie
> we lock the group_rwsem to keep the group->kvm valid and then pass the
> kvm to open_device in some way, then the driver can kvm_get_kvm() it

Hrm... If we did that we would have to re-evaluate some other usage of 
the rwsem e.g. if driver open_device calls vfio_register_iommu_notifier 
it will try to get the rwsem but it's already locked.

> Alternatively, I don't know why kvm_get_kvm() is an exported symbol
> when it is just calling refcount_inc() - inlining it would be an
> improvement I think.

I think that would work for kvm_get_kvm, but kvm_put_kvm (which we also 
need) calls kvm_destroy_kvm after the refcount_dec and that can't be inlined