[Intel-gfx] [PATCH V3 4/7] mdev: introduce device specific ops

[Alex Williamson]

On Wed, 16 Oct 2019 20:48:06 +0000
Parav Pandit <parav at mellanox.com> wrote:

> > From: Alex Williamson <alex.williamson at redhat.com>
> > On Wed, 16 Oct 2019 15:31:25 +0000
> > Parav Pandit <parav at mellanox.com> wrote:
> > > > From: Cornelia Huck <cohuck at redhat.com>
> > > > Parav Pandit <parav at mellanox.com> wrote:
> > > > > > From: Alex Williamson <alex.williamson at redhat.com>
> > > > > > On Tue, 15 Oct 2019 20:17:01 +0800 Jason Wang
> > > > > > <jasowang at redhat.com> wrote:
> > > > > >  
> > > > > > > On 2019/10/15 下午6:41, Cornelia Huck wrote:  
> > > > > > > > Apologies if that has already been discussed, but do we want
> > > > > > > > a
> > > > > > > > 1:1 relationship between id and ops, or can different
> > > > > > > > devices with the same id register different ops?  
> > > > > > >
> > > > > > >
> > > > > > > I think we have a N:1 mapping between id and ops, e.g we want
> > > > > > > both virtio-mdev and vhost-mdev use a single set of device ops.  
> > > > > >
> > > > > > The contents of the ops structure is essentially defined by the
> > > > > > id, which is why I was leaning towards them being defined together.
> > > > > > They are effectively interlocked, the id defines which mdev "endpoint"
> > > > > > driver is loaded and that driver requires mdev_get_dev_ops() to
> > > > > > return the structure required by the driver.  I wish there was a
> > > > > > way we could incorporate type checking here.  We toyed with the
> > > > > > idea of having the class in the same structure as the ops, but I
> > > > > > think this approach was chosen for simplicity.  We could still do  
> > something like:  
> > > > > >
> > > > > > int mdev_set_class_struct(struct device *dev, const struct
> > > > > > mdev_class_struct *class);
> > > > > >
> > > > > > struct mdev_class_struct {
> > > > > > 	u16	id;
> > > > > > 	union {
> > > > > > 		struct vfio_mdev_ops vfio_ops;
> > > > > > 		struct virtio_mdev_ops virtio_ops;
> > > > > > 	};
> > > > > > };
> > > > > >
> > > > > > Maybe even:
> > > > > >
> > > > > > struct vfio_mdev_ops *mdev_get_vfio_ops(struct mdev_device *mdev)  
> > {  
> > > > > > 	BUG_ON(mdev->class.id != MDEV_ID_VFIO);
> > > > > > 	return &mdev->class.vfio_ops;
> > > > > > }
> > > > > >
> > > > > > The match callback would of course just use the mdev->class.id value.
> > > > > > Functionally equivalent, but maybe better type characteristics.
> > > > > > Thanks,
> > > > > >
> > > > > > Alex  
> > > > >
> > > > > We have 3 use cases of mdev.
> > > > > 1. current mdev binding to vfio_mdev 2. mdev binding to virtio 3.
> > > > > mdev binding to mlx5_core without dev_ops
> > > > >
> > > > > Also
> > > > > (a) a given parent may serve multiple types of classes in future.
> > > > > (b) number of classes may not likely explode, they will be handful
> > > > > of them. (vfio_mdev, virtio)
> > > > >
> > > > > So, instead of making copies of this dev_ops pointer in each mdev,
> > > > > it is better  
> > > > to keep const multiple ops in their parent device.  
> > > > > Something like below,
> > > > >
> > > > > struct mdev_parent {
> > > > > 	[..]
> > > > > 	struct mdev_parent_ops *parent_ops; /* create, remove */
> > > > > 	struct vfio_mdev_ops *vfio_ops; /* read,write, ioctl etc */
> > > > > 	struct virtio_mdev_ops *virtio_ops; /* virtio ops */ };  
> > > >
> > > > That feels a bit odd. Why should the parent carry pointers to every
> > > > possible version of ops?
> > > >  
> > > How many are we expecting? I envisioned handful of them.
> > > It carries because parent is few, mdevs are several hundreds.
> > > It makes sense to keep few copies, instead of several hundred copies
> > > and it doesn't need to setup on every mdev creation.  
> > 
> > It does need setup on every mdev creation, it's just a matter of the scope, 'id
> > and ops' vs 'id only' vs 'ops with implicit id'.  The other argument is assuming a
> > space vs time trade-off that I'm having a hard time judging is necessarily the
> > correct approach.  We potentially have better data locality in the mdev device
> > structure vs the parent.  The caching of the ops structure itself is separate from
> > how we get to it.
> > We might have hundreds of pointers to those ops structure, but the space
> > trade-off might we worth it if they're on the same cacheline as the mdev
> > device itself vs the indirection via the parent.
> > 
> > I see a couple other drawbacks to the parent hosted ops pointers as well.  First,
> > it imposes that per parent there can only be one device ops structure per class
> > id, but who's to say that different types of mdev devices for a given parent all
> > make the same callbacks into the parent.   
> We should have driver who intent to use different device ops for each
> device with single parent that supports this claim.

Why?  Are we not allowed to identify restrictions implied by a given
proposal if we don't yet have a user?  I can't subscribe to that.

>  For instance, for a vfio-mdev we
> > already support the concept of an iommu backing device which makes
> > the type1 iommu code behave a little differently.  Those
> > differences might be sufficient that the parent driver would
> > register a different device ops structure for an iommu backed mdev
> > vs a non-iommu backed device.    
> I am not sure if this is really worth it.
> Which driver should I look which has if-else conditions sprinkled in
> these callbacks for different iommu types? If majority code is same,
> adding few branches looks ok vs creating new ops all together. So I
> need to educate myself first with the driver which desires this. Any
> pointers?

While the iommu backed vfio-mdevs is real, the example that a parent
driver might choose to register different device ops based on that is
theoretical.  Parent drivers don't have that option today, but as we're
making the device ops more modular and have stumbled onto this benefit
of per device ops, perhaps it might be useful.  The "is it worth it"
question can also be asked of the claimed benefits of a set of shared
devices ops per parent.
 
> > The other
> > drawback is that it implies a binary difference in all mdev parent
> > drivers to add any new device ids.  I know we don't guarantee
> > binary compatibility, but it's rather ugly.
> >   
> Yeah, we don't support and there is no requirement for binary
> compatibility.
> 
> > Overall, I guess I tend to prefer Connie's proposal, the class id
> > and structure are tied together and the parent driver is only
> > responsible for one of them, the class id is hidden away in
> > mdev-core and the mdev driver itself. 
> I am fine with Cornelia's approach.
> It comes with small cost of additional symbols and it is probably ok.
> I just find it over engineered given handful of dev ops types.

If the device ops types are limited, then so are the additional
symbols.  Those symbols also add a degree of explicitness to the
interface (ie. register this device as vfio-mdev with this set of
vfio-mdev-ops, versus register this device as vfio-mdev... which uses
the vfio-mdev-ops over in the parent ops structure).  I don't really
see what's over-engineered about former.  I like to think of Rusty's
old interface guidelines, particularly the one about making it
difficult to use incorrectly for these sorts of interfaces.  Thanks,

Alex