Linux Virtualization - Sep 2019 - [RFC v4 0/3] vhost: introduce mdev based hardware backend

On Thu, Sep 19, 2019 at 09:08:11PM +0800, Jason Wang
wrote:
> On 2019/9/18 ??10:32, Michael S. Tsirkin wrote:
> > > > > 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.
> 
> Ok, if we go this way, it could be as simple as provide some callback to
> vhost, then vhost can just forward the ioctl through parent_ops.
> 
> > 
> > > > > 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.
> 
> Just to clarify, a new type of mdev driver but provides dummy
> vfio_device_ops for VFIO to make container DMA ioctl work.
I see. Thanks! IIUC, you mean we can provide a very tiny
VFIO device driver in drivers/vhost/mdev.c, e.g.:

static int vfio_vhost_mdev_open(void *device_data)
{
	if (!try_module_get(THIS_MODULE))
		return -ENODEV;
	return 0;
}

static void vfio_vhost_mdev_release(void *device_data)
{
	module_put(THIS_MODULE);
}

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,
};

static int vhost_mdev_probe(struct device *dev)
{
	struct mdev_device *mdev = to_mdev_device(dev);

	... Check the mdev device_id proposed in ...
	... https://lkml.org/lkml/2019/9/12/151 ...

	return vfio_add_group_dev(dev, &vfio_vhost_mdev_dev_ops, mdev);
}

static void vhost_mdev_remove(struct device *dev)
{
	vfio_del_group_dev(dev);
}

static struct mdev_driver vhost_mdev_driver = {
	.name	= "vhost_mdev",
	.probe	= vhost_mdev_probe,
	.remove	= vhost_mdev_remove,
};

So we can bind above mdev driver to the virtio-mdev compatible
mdev devices when we want to use vhost-mdev.

After binding above driver to the mdev device, we can setup IOMMU
via VFIO and get VFIO device fd of this mdev device, and pass it
to vhost fd (/dev/vhost-mdev) with a SET_BACKEND ioctl.

Thanks,
Tiwei
> 
> Thanks
> 
> 
> > > Yes, it is.
> > > 
> > > Thanks
> > > 
> > >

On 2019/9/19 ??11:45, Tiwei Bie wrote:
> On Thu, Sep 19, 2019 at 09:08:11PM +0800, Jason Wang wrote:
>> On 2019/9/18 ??10:32, Michael S. Tsirkin wrote:
>>>>>> 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.
>> Ok, if we go this way, it could be as simple as provide some callback
to
>> vhost, then vhost can just forward the ioctl through parent_ops.
>>
>>>>>> 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.
>> Just to clarify, a new type of mdev driver but provides dummy
>> vfio_device_ops for VFIO to make container DMA ioctl work.
> I see. Thanks! IIUC, you mean we can provide a very tiny
> VFIO device driver in drivers/vhost/mdev.c, e.g.:
>
> static int vfio_vhost_mdev_open(void *device_data)
> {
> 	if (!try_module_get(THIS_MODULE))
> 		return -ENODEV;
> 	return 0;
> }
>
> static void vfio_vhost_mdev_release(void *device_data)
> {
> 	module_put(THIS_MODULE);
> }
>
> 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,
> };
>
> static int vhost_mdev_probe(struct device *dev)
> {
> 	struct mdev_device *mdev = to_mdev_device(dev);
>
> 	... Check the mdev device_id proposed in ...
> 	... https://lkml.org/lkml/2019/9/12/151 ...
>
> 	return vfio_add_group_dev(dev, &vfio_vhost_mdev_dev_ops, mdev);
> }
>
> static void vhost_mdev_remove(struct device *dev)
> {
> 	vfio_del_group_dev(dev);
> }
>
> static struct mdev_driver vhost_mdev_driver = {
> 	.name	= "vhost_mdev",
> 	.probe	= vhost_mdev_probe,
> 	.remove	= vhost_mdev_remove,
> };
>
> So we can bind above mdev driver to the virtio-mdev compatible
> mdev devices when we want to use vhost-mdev.
>
> After binding above driver to the mdev device, we can setup IOMMU
> via VFIO and get VFIO device fd of this mdev device, and pass it
> to vhost fd (/dev/vhost-mdev) with a SET_BACKEND ioctl.
>
> Thanks,
> Tiwei

Yes, something like this.

Thanks

>> Thanks
>>
>>
>>>> Yes, it is.
>>>>
>>>> Thanks
>>>>
>>>>

On 2019/9/19 ??11:45, Tiwei Bie wrote:
> On Thu, Sep 19, 2019 at 09:08:11PM +0800, Jason Wang wrote:
>> On 2019/9/18 ??10:32, Michael S. Tsirkin wrote:
>>>>>> 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.
>> Ok, if we go this way, it could be as simple as provide some callback
to
>> vhost, then vhost can just forward the ioctl through parent_ops.
>>
>>>>>> 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.
>> Just to clarify, a new type of mdev driver but provides dummy
>> vfio_device_ops for VFIO to make container DMA ioctl work.
> I see. Thanks! IIUC, you mean we can provide a very tiny
> VFIO device driver in drivers/vhost/mdev.c, e.g.:
>
> static int vfio_vhost_mdev_open(void *device_data)
> {
> 	if (!try_module_get(THIS_MODULE))
> 		return -ENODEV;
> 	return 0;
> }
>
> static void vfio_vhost_mdev_release(void *device_data)
> {
> 	module_put(THIS_MODULE);
> }
>
> 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,
> };
>
> static int vhost_mdev_probe(struct device *dev)
> {
> 	struct mdev_device *mdev = to_mdev_device(dev);
>
> 	... Check the mdev device_id proposed in ...
> 	... https://lkml.org/lkml/2019/9/12/151 ...

To clarify, this should be done through the id_table fields in 
vhost_mdev_driver, and it should claim it supports virtio-mdev device only:


static struct mdev_class_id id_table[] = {
 ??? { MDEV_ID_VIRTIO },
 ??? { 0 },
};


static struct mdev_driver vhost_mdev_driver = {
 ??? ...
 ??? .id_table = id_table,
}

>
> 	return vfio_add_group_dev(dev, &vfio_vhost_mdev_dev_ops, mdev);

And in vfio_vhost_mdev_ops, all its need is to just implement vhost-net 
ioctl and translate them to virtio-mdev transport (e.g device_ops I 
proposed or ioctls other whatever other method) API. And it could have a 
dummy ops implementation for the other device_ops.

> }
>
> static void vhost_mdev_remove(struct device *dev)
> {
> 	vfio_del_group_dev(dev);
> }
>
> static struct mdev_driver vhost_mdev_driver = {
> 	.name	= "vhost_mdev",
> 	.probe	= vhost_mdev_probe,
> 	.remove	= vhost_mdev_remove,
> };
>
> So we can bind above mdev driver to the virtio-mdev compatible
> mdev devices when we want to use vhost-mdev.
>
> After binding above driver to the mdev device, we can setup IOMMU
> via VFIO and get VFIO device fd of this mdev device, and pass it
> to vhost fd (/dev/vhost-mdev) with a SET_BACKEND ioctl.

Then what vhost-mdev char device did is just forwarding ioctl back to 
this vfio device fd which seems a overkill. It's simpler that just do 
ioctl on the device ops directly.

Thanks

>
> Thanks,
> Tiwei
>
>> Thanks
>>
>>
>>>> Yes, it is.
>>>>
>>>> Thanks
>>>>
>>>>

On Fri, Sep 20, 2019 at 09:30:58AM +0800, Jason Wang
wrote:
> On 2019/9/19 ??11:45, Tiwei Bie wrote:
> > On Thu, Sep 19, 2019 at 09:08:11PM +0800, Jason Wang wrote:
> > > On 2019/9/18 ??10:32, Michael S. Tsirkin wrote:
> > > > > > > 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.
> > > Ok, if we go this way, it could be as simple as provide some
callback to
> > > vhost, then vhost can just forward the ioctl through parent_ops.
> > > 
> > > > > > > 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.
> > > Just to clarify, a new type of mdev driver but provides dummy
> > > vfio_device_ops for VFIO to make container DMA ioctl work.
> > I see. Thanks! IIUC, you mean we can provide a very tiny
> > VFIO device driver in drivers/vhost/mdev.c, e.g.:
> > 
> > static int vfio_vhost_mdev_open(void *device_data)
> > {
> > 	if (!try_module_get(THIS_MODULE))
> > 		return -ENODEV;
> > 	return 0;
> > }
> > 
> > static void vfio_vhost_mdev_release(void *device_data)
> > {
> > 	module_put(THIS_MODULE);
> > }
> > 
> > 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,
> > };
> > 
> > static int vhost_mdev_probe(struct device *dev)
> > {
> > 	struct mdev_device *mdev = to_mdev_device(dev);
> > 
> > 	... Check the mdev device_id proposed in ...
> > 	... https://lkml.org/lkml/2019/9/12/151 ...
> 
> 
> To clarify, this should be done through the id_table fields in
> vhost_mdev_driver, and it should claim it supports virtio-mdev device only:
> 
> 
> static struct mdev_class_id id_table[] = {
> ??? { MDEV_ID_VIRTIO },
> ??? { 0 },
> };
> 
> 
> static struct mdev_driver vhost_mdev_driver = {
> ??? ...
> ??? .id_table = id_table,
> }
In this way, both of virtio-mdev and vhost-mdev will try to
take this device. We may want a way to let vhost-mdev take this
device only when users explicitly ask it to do it. Or maybe we
can have a different MDEV_ID for vhost-mdev but share the device
ops with virtio-mdev.
> 
> 
> > 
> > 	return vfio_add_group_dev(dev, &vfio_vhost_mdev_dev_ops, mdev);
> 
> 
> And in vfio_vhost_mdev_ops, all its need is to just implement vhost-net
> ioctl and translate them to virtio-mdev transport (e.g device_ops I
proposed
> or ioctls other whatever other method) API.
I see, so my previous understanding is basically correct:

https://lkml.org/lkml/2019/9/17/332

I.e. we won't have a separate vhost fd and we will do all vhost
ioctls on the VFIO device fd backed by this new VFIO driver.
> And it could have a dummy ops
> implementation for the other device_ops.
> 
> 
> > }
> > 
> > static void vhost_mdev_remove(struct device *dev)
> > {
> > 	vfio_del_group_dev(dev);
> > }
> > 
> > static struct mdev_driver vhost_mdev_driver = {
> > 	.name	= "vhost_mdev",
> > 	.probe	= vhost_mdev_probe,
> > 	.remove	= vhost_mdev_remove,
> > };
> > 
> > So we can bind above mdev driver to the virtio-mdev compatible
> > mdev devices when we want to use vhost-mdev.
> > 
> > After binding above driver to the mdev device, we can setup IOMMU
> > via VFIO and get VFIO device fd of this mdev device, and pass it
> > to vhost fd (/dev/vhost-mdev) with a SET_BACKEND ioctl.
> 
> 
> Then what vhost-mdev char device did is just forwarding ioctl back to this
> vfio device fd which seems a overkill. It's simpler that just do ioctl
on
> the device ops directly.
Yes.

Thanks,
Tiwei

> 
> Thanks
> 
> 
> > 
> > Thanks,
> > Tiwei
> > 
> > > Thanks
> > > 
> > > 
> > > > > Yes, it is.
> > > > > 
> > > > > Thanks
> > > > > 
> > > > >