Re: [RFC v2] vhost: introduce mdev based hardware vhost backend

[Jason Wang]

On 2019/7/4 äå3:02, Tiwei Bie wrote:
> On Thu, Jul 04, 2019 at 02:35:20PM +0800, Jason Wang wrote:
> > On 2019/7/4 äå2:21, Tiwei Bie wrote:
> > > On Thu, Jul 04, 2019 at 12:31:48PM +0800, Jason Wang wrote:
> > > > On 2019/7/3 äå9:08, Tiwei Bie wrote:
> > > > > On Wed, Jul 03, 2019 at 08:16:23PM +0800, Jason Wang wrote:
> > > > > > On 2019/7/3 äå7:52, Tiwei Bie wrote:
> > > > > > > On Wed, Jul 03, 2019 at 06:09:51PM +0800, Jason Wang wrote:
> > > > > > > > On 2019/7/3 äå5:13, Tiwei Bie wrote:
> > > > > > > > > Details about this can be found here:
> > > > > > > > >
> > > > > > > > > https://lwn.net/Articles/750770/
> > > > > > > > >
> > > > > > > > > What's new in this version
> > > > > > > > > ==========================
> > > > > > > > >
> > > > > > > > > A new VFIO device type is introduced - vfio-vhost. This addressed
> > > > > > > > > some comments from here:https://patchwork.ozlabs.org/cover/984763/
> > > > > > > > >
> > > > > > > > > Below is the updated device interface:
> > > > > > > > >
> > > > > > > > > Currently, there are two regions of this device: 1) CONFIG_REGION
> > > > > > > > > (VFIO_VHOST_CONFIG_REGION_INDEX), which can be used to setup the
> > > > > > > > > device; 2) NOTIFY_REGION (VFIO_VHOST_NOTIFY_REGION_INDEX), which
> > > > > > > > > can be used to notify the device.
> > > > > > > > >
> > > > > > > > > 1. CONFIG_REGION
> > > > > > > > >
> > > > > > > > > The region described by CONFIG_REGION is the main control interface.
> > > > > > > > > Messages will be written to or read from this region.
> > > > > > > > >
> > > > > > > > > The message type is determined by the `request` field in message
> > > > > > > > > header. The message size is encoded in the message header too.
> > > > > > > > > The message format looks like this:
> > > > > > > > >
> > > > > > > > > struct vhost_vfio_op {
> > > > > > > > > 	__u64 request;
> > > > > > > > > 	__u32 flags;
> > > > > > > > > 	/* Flag values: */
> > > > > > > > >       #define VHOST_VFIO_NEED_REPLY 0x1 /* Whether need reply */
> > > > > > > > > 	__u32 size;
> > > > > > > > > 	union {
> > > > > > > > > 		__u64 u64;
> > > > > > > > > 		struct vhost_vring_state state;
> > > > > > > > > 		struct vhost_vring_addr addr;
> > > > > > > > > 	} payload;
> > > > > > > > > };
> > > > > > > > >
> > > > > > > > > The existing vhost-kernel ioctl cmds are reused as the message
> > > > > > > > > requests in above structure.
> > > > > > > > Still a comments like V1. What's the advantage of inventing a new protocol?
> > > > > > > I'm trying to make it work in VFIO's way..
> > > > > > >
> > > > > > > > I believe either of the following should be better:
> > > > > > > >
> > > > > > > > - using vhost ioctl, we can start from SET_VRING_KICK/SET_VRING_CALL and
> > > > > > > > extend it with e.g notify region. The advantages is that all exist userspace
> > > > > > > > program could be reused without modification (or minimal modification). And
> > > > > > > > vhost API hides lots of details that is not necessary to be understood by
> > > > > > > > application (e.g in the case of container).
> > > > > > > Do you mean reusing vhost's ioctl on VFIO device fd directly,
> > > > > > > or introducing another mdev driver (i.e. vhost_mdev instead of
> > > > > > > using the existing vfio_mdev) for mdev device?
> > > > > > Can we simply add them into ioctl of mdev_parent_ops?
> > > > > Right, either way, these ioctls have to be and just need to be
> > > > > added in the ioctl of the mdev_parent_ops. But another thing we
> > > > > also need to consider is that which file descriptor the userspace
> > > > > will do the ioctl() on. So I'm wondering do you mean let the
> > > > > userspace do the ioctl() on the VFIO device fd of the mdev
> > > > > device?
> > > > Yes.
> > > Got it! I'm not sure what's Alex opinion on this. If we all
> > > agree with this, I can do it in this way.
> > >
> > > > Is there any other way btw?
> > > Just a quick thought.. Maybe totally a bad idea.
> >
> > It's not for sure :)
> Thanks!
>
> > >    I was thinking
> > > whether it would be odd to do non-VFIO's ioctls on VFIO's device
> > > fd. So I was wondering whether it's possible to allow binding
> > > another mdev driver (e.g. vhost_mdev) to the supported mdev
> > > devices. The new mdev driver, vhost_mdev, can provide similar
> > > ways to let userspace open the mdev device and do the vhost ioctls
> > > on it. To distinguish with the vfio_mdev compatible mdev devices,
> > > the device API of the new vhost_mdev compatible mdev devices
> > > might be e.g. "vhost-net" for net?
> > >
> > > So in VFIO case, the device will be for passthru directly. And
> > > in VHOST case, the device can be used to accelerate the existing
> > > virtualized devices.
> > >
> > > How do you think?
> >
> > If my understanding is correct, there will be no VFIO ioctl if we go for
> > vhost_mdev?
> Yeah, exactly. If we go for vhost_mdev, we may have some vhost nodes
> in /dev similar to what /dev/vfio/* does to handle the $UUID and open
> the device (e.g. similar to VFIO_GROUP_GET_DEVICE_FD in VFIO). And
> to setup the device, we can try to reuse the ioctls of the existing
> kernel vhost as much as possible.


Interesting, actually, I've considered something similar. I think there 
should be no issues other than DMA:

- Need to invent new API for DMA mapping other than SET_MEM_TABLE? 
(Which is too heavyweight).

- Need to consider a way to co-work with both on chip IOMMU (your 
proposal should be fine) and scalable IOV.

Thanks


> Thanks,
> Tiwei
>
> > Thanks
> >
> >
> > > Thanks,
> > > Tiwei
> > > > Thanks