Re: [RFC v4 0/3] vhost: introduce mdev based hardware backend

["Michael S. Tsirkin" <mst@redhat.com>]

On Wed, Sep 18, 2019 at 01:51:21PM +0800, Jason Wang wrote:
> 
> On 2019/9/17 下午6:58, Tiwei Bie wrote:
> > On Tue, Sep 17, 2019 at 11:32:03AM +0800, Jason Wang wrote:
> > > On 2019/9/17 上午9:02, Tiwei Bie wrote:
> > > > This RFC is to demonstrate below ideas,
> > > > 
> > > > a) Build vhost-mdev on top of the same abstraction defined in
> > > >      the virtio-mdev series [1];
> > > > 
> > > > b) Introduce /dev/vhost-mdev to do vhost ioctls and support
> > > >      setting mdev device as backend;
> > > > 
> > > > Now the userspace API looks like this:
> > > > 
> > > > - Userspace generates a compatible mdev device;
> > > > 
> > > > - Userspace opens this mdev device with VFIO API (including
> > > >     doing IOMMU programming for this mdev device with VFIO's
> > > >     container/group based interface);
> > > > 
> > > > - Userspace opens /dev/vhost-mdev and gets vhost fd;
> > > > 
> > > > - Userspace uses vhost ioctls to setup vhost (userspace should
> > > >     do VHOST_MDEV_SET_BACKEND ioctl with VFIO group fd and device
> > > >     fd first before doing other vhost ioctls);
> > > > 
> > > > Only compile test has been done for this series for now.
> > > 
> > > Have a hard thought on the architecture:
> > Thanks a lot! Do appreciate it!
> > 
> > > 1) Create a vhost char device and pass vfio mdev device fd to it as a
> > > backend and translate vhost-mdev ioctl to virtio mdev transport (e.g
> > > read/write). DMA was done through the VFIO DMA mapping on the container that
> > > is attached.
> > Yeah, that's what we are doing in this series.
> > 
> > > We have two more choices:
> > > 
> > > 2) Use vfio-mdev but do not create vhost-mdev device, instead, just
> > > implement vhost ioctl on vfio_device_ops, and translate them into
> > > virtio-mdev transport or just pass ioctl to parent.
> > Yeah. Instead of introducing /dev/vhost-mdev char device, do
> > vhost ioctls on VFIO device fd directly. That's what we did
> > in RFC v3.
> > 
> > > 3) Don't use vfio-mdev, create a new vhost-mdev driver, during probe still
> > > try to add dev to vfio group and talk to parent with device specific ops
> > If my understanding is correct, this means we need to introduce
> > a new VFIO device driver to replace the existing vfio-mdev driver
> > in our case. Below is a quick draft just to show my understanding:
> > 
> > #include <linux/init.h>
> > #include <linux/module.h>
> > #include <linux/device.h>
> > #include <linux/kernel.h>
> > #include <linux/slab.h>
> > #include <linux/vfio.h>
> > #include <linux/mdev.h>
> > 
> > #include "mdev_private.h"
> > 
> > /* XXX: we need a proper way to include below vhost header. */
> > #include "../../vhost/vhost.h"
> > 
> > static int vfio_vhost_mdev_open(void *device_data)
> > {
> > 	if (!try_module_get(THIS_MODULE))
> > 		return -ENODEV;
> > 
> > 	/* ... */
> > 	vhost_dev_init(...);
> > 
> > 	return 0;
> > }
> > 
> > static void vfio_vhost_mdev_release(void *device_data)
> > {
> > 	/* ... */
> > 	module_put(THIS_MODULE);
> > }
> > 
> > static long vfio_vhost_mdev_unlocked_ioctl(void *device_data,
> > 					   unsigned int cmd, unsigned long arg)
> > {
> > 	struct mdev_device *mdev = device_data;
> > 	struct mdev_parent *parent = mdev->parent;
> > 
> > 	/*
> > 	 * Use vhost ioctls.
> > 	 *
> > 	 * We will have a different parent_ops design.
> > 	 * And potentially, we can share the same parent_ops
> > 	 * with virtio_mdev.
> > 	 */
> > 	switch (cmd) {
> > 	case VHOST_GET_FEATURES:
> > 		parent->ops->get_features(mdev, ...);
> > 		break;
> > 	/* ... */
> > 	}
> > 
> > 	return 0;
> > }
> > 
> > static ssize_t vfio_vhost_mdev_read(void *device_data, char __user *buf,
> > 				    size_t count, loff_t *ppos)
> > {
> > 	/* ... */
> > 	return 0;
> > }
> > 
> > static ssize_t vfio_vhost_mdev_write(void *device_data, const char __user *buf,
> > 				     size_t count, loff_t *ppos)
> > {
> > 	/* ... */
> > 	return 0;
> > }
> > 
> > static int vfio_vhost_mdev_mmap(void *device_data, struct vm_area_struct *vma)
> > {
> > 	/* ... */
> > 	return 0;
> > }
> > 
> > static const struct vfio_device_ops vfio_vhost_mdev_dev_ops = {
> > 	.name		= "vfio-vhost-mdev",
> > 	.open		= vfio_vhost_mdev_open,
> > 	.release	= vfio_vhost_mdev_release,
> > 	.ioctl		= vfio_vhost_mdev_unlocked_ioctl,
> > 	.read		= vfio_vhost_mdev_read,
> > 	.write		= vfio_vhost_mdev_write,
> > 	.mmap		= vfio_vhost_mdev_mmap,
> > };
> > 
> > static int vfio_vhost_mdev_probe(struct device *dev)
> > {
> > 	struct mdev_device *mdev = to_mdev_device(dev);
> > 
> > 	/* ... */
> > 	return vfio_add_group_dev(dev, &vfio_vhost_mdev_dev_ops, mdev);
> > }
> > 
> > static void vfio_vhost_mdev_remove(struct device *dev)
> > {
> > 	/* ... */
> > 	vfio_del_group_dev(dev);
> > }
> > 
> > static struct mdev_driver vfio_vhost_mdev_driver = {
> > 	.name	= "vfio_vhost_mdev",
> > 	.probe	= vfio_vhost_mdev_probe,
> > 	.remove	= vfio_vhost_mdev_remove,
> > };
> > 
> > static int __init vfio_vhost_mdev_init(void)
> > {
> > 	return mdev_register_driver(&vfio_vhost_mdev_driver, THIS_MODULE);
> > }
> > module_init(vfio_vhost_mdev_init)
> > 
> > static void __exit vfio_vhost_mdev_exit(void)
> > {
> > 	mdev_unregister_driver(&vfio_vhost_mdev_driver);
> > }
> > module_exit(vfio_vhost_mdev_exit)
> 
> 
> Yes, something like this basically.
> 
> 
> > > So I have some questions:
> > > 
> > > 1) Compared to method 2, what's the advantage of creating a new vhost char
> > > device? I guess it's for keep the API compatibility?
> > One benefit is that we can avoid doing vhost ioctls on
> > VFIO device fd.
> 
> 
> Yes, but any benefit from doing this?

It does seem a bit more modular, but it's certainly not a big deal.

> > 
> > > 2) For method 2, is there any easy way for user/admin to distinguish e.g
> > > ordinary vfio-mdev for vhost from ordinary vfio-mdev?
> > I think device-api could be a choice.
> 
> 
> Ok.
> 
> 
> > 
> > > I saw you introduce
> > > ops matching helper but it's not friendly to management.
> > The ops matching helper is just to check whether a given
> > vfio-device is based on a mdev device.
> > 
> > > 3) A drawback of 1) and 2) is that it must follow vfio_device_ops that
> > > assumes the parameter comes from userspace, it prevents support kernel
> > > virtio drivers.
> > > 
> > > 4) So comes the idea of method 3, since it register a new vhost-mdev driver,
> > > we can use device specific ops instead of VFIO ones, then we can have a
> > > common API between vDPA parent and vhost-mdev/virtio-mdev drivers.
> > As the above draft shows, this requires introducing a new
> > VFIO device driver. I think Alex's opinion matters here.
> 
> 
> Yes, it is.
> 
> Thanks
> 
> 
> > Thanks,
> > Tiwei
> > 
> > > What's your thoughts?
> > > 
> > > Thanks
> > > 
> > > 
> > > > RFCv3: https://patchwork.kernel.org/patch/11117785/
> > > > 
> > > > [1] https://lkml.org/lkml/2019/9/10/135
> > > > 
> > > > Tiwei Bie (3):
> > > >     vfio: support getting vfio device from device fd
> > > >     vfio: support checking vfio driver by device ops
> > > >     vhost: introduce mdev based hardware backend
> > > > 
> > > >    drivers/vfio/mdev/vfio_mdev.c    |   3 +-
> > > >    drivers/vfio/vfio.c              |  32 +++
> > > >    drivers/vhost/Kconfig            |   9 +
> > > >    drivers/vhost/Makefile           |   3 +
> > > >    drivers/vhost/mdev.c             | 462 +++++++++++++++++++++++++++++++
> > > >    drivers/vhost/vhost.c            |  39 ++-
> > > >    drivers/vhost/vhost.h            |   6 +
> > > >    include/linux/vfio.h             |  11 +
> > > >    include/uapi/linux/vhost.h       |  10 +
> > > >    include/uapi/linux/vhost_types.h |   5 +
> > > >    10 files changed, 573 insertions(+), 7 deletions(-)
> > > >    create mode 100644 drivers/vhost/mdev.c
> > > >