Virtualization - Dec 2020 - [RFC PATCH 5/8] vhost: allow userspace to bind vqs to CPUs

On 2020/12/8 ??2:31, Mike Christie wrote:
> On 12/6/20 10:27 PM, Jason Wang wrote:
>>
>> On 2020/12/5 ??12:32, Mike Christie wrote:
>>> On 12/4/20 2:09 AM, Jason Wang wrote:
>>>>
>>>> On 2020/12/4 ??3:56, Mike Christie wrote:
>>>>> +static long vhost_vring_set_cpu(struct vhost_dev *d,
struct
>>>>> vhost_virtqueue *vq,
>>>>> +??????????????? void __user *argp)
>>>>> +{
>>>>> +??? struct vhost_vring_state s;
>>>>> +??? int ret = 0;
>>>>> +
>>>>> +??? if (vq->private_data)
>>>>> +??????? return -EBUSY;
>>>>> +
>>>>> +??? if (copy_from_user(&s, argp, sizeof s))
>>>>> +??????? return -EFAULT;
>>>>> +
>>>>> +??? if (s.num == -1) {
>>>>> +??????? vq->cpu = s.num;
>>>>> +??????? return 0;
>>>>> +??? }
>>>>> +
>>>>> +??? if (s.num >= nr_cpu_ids)
>>>>> +??????? return -EINVAL;
>>>>> +
>>>>> +??? if (!d->ops || !d->ops->get_workqueue)
>>>>> +??????? return -EINVAL;
>>>>> +
>>>>> +??? if (!d->wq)
>>>>> +??????? d->wq = d->ops->get_workqueue();
>>>>> +??? if (!d->wq)
>>>>> +??????? return -EINVAL;
>>>>> +
>>>>> +??? vq->cpu = s.num;
>>>>> +??? return ret;
>>>>> +}
>>>>
>>>>
>>>> So one question here. Who is in charge of doing this set_cpu?
Note
>>>> that?sched_setaffinity(2) requires CAP_SYS_NICE to work, so I 
>>>> wonder whether or not it's legal for unprivileged Qemu to
do this.
>>>
>>>
>>> I was having qemu do it when it's setting up the vqs since it
had
>>> the info there already.
>>>
>>> Is it normally the tool that makes calls into qemu that does the 
>>> operations that require CAP_SYS_NICE? 
>>
>>
>> My understanding is that it only matter scheduling. And this patch 
>> wants to change the affinity which should check that capability.
>>
>>
>>> If so, then I see the interface needs to be changed.
>>
>>
>> Actually, if I read this patch correctly it requires e.g qemu to make 
>> the decision instead of the management layer. This may bring some 
>> troubles to for e.g the libvirt emulatorpin[1] implementation.
>>
>
> Let me make sure I understood you.
>
> I thought qemu would just have a new property, and users would pass 
> that in like they do for the number of queues setting. Then qemu would 
> pass that to the kernel. The primary user I have to support at work 
> does not use libvirt based tools so I thought that was a common point 
> that would work for everyone.

I think we need talk with libvirt guys to see if it works for them. My 
understanding is the scheduling should be the charge of them not qemu.

>
> For my work use requirement, your emulatorpin and CAP_SYS_NICE comment 
> then that means we want an interface that something other than qemu 
> can use right? So the tools would call directly into the kernel and 
> not go through qemu right?

Yes, usually qemu runs without any privilege. So could it be e.g a sysfs 
interface or other?

Thanks

>
>