* Re: [PATCH V2 RESEND 2/2] vDPA/ifcvf: implement doorbell mapping for ifcvf
[not found] ` <5dbdc6a5-1510-9411-6b85-d947d091089c@intel.com>
@ 2021-06-01 3:31 ` Jason Wang
0 siblings, 0 replies; 2+ messages in thread
From: Jason Wang @ 2021-06-01 3:31 UTC (permalink / raw)
To: Zhu, Lingshan, mst; +Cc: netdev, kvm, virtualization
在 2021/6/1 上午11:27, Zhu, Lingshan 写道:
>
>
> On 5/31/2021 3:56 PM, Jason Wang wrote:
>>
>> 在 2021/5/31 下午3:33, Zhu Lingshan 写道:
>>> This commit implements doorbell mapping feature for ifcvf.
>>> This feature maps the notify page to userspace, to eliminate
>>> vmexit when kick a vq.
>>>
>>> Signed-off-by: Zhu Lingshan <lingshan.zhu@intel.com>
>>> ---
>>> drivers/vdpa/ifcvf/ifcvf_main.c | 17 +++++++++++++++++
>>> 1 file changed, 17 insertions(+)
>>>
>>> diff --git a/drivers/vdpa/ifcvf/ifcvf_main.c
>>> b/drivers/vdpa/ifcvf/ifcvf_main.c
>>> index ab0ab5cf0f6e..effb0e549135 100644
>>> --- a/drivers/vdpa/ifcvf/ifcvf_main.c
>>> +++ b/drivers/vdpa/ifcvf/ifcvf_main.c
>>> @@ -413,6 +413,22 @@ static int ifcvf_vdpa_get_vq_irq(struct
>>> vdpa_device *vdpa_dev,
>>> return vf->vring[qid].irq;
>>> }
>>> +static struct vdpa_notification_area
>>> ifcvf_get_vq_notification(struct vdpa_device *vdpa_dev,
>>> + u16 idx)
>>> +{
>>> + struct ifcvf_adapter *adapter = vdpa_to_adapter(vdpa_dev);
>>> + struct ifcvf_hw *vf = vdpa_to_vf(vdpa_dev);
>>> + struct pci_dev *pdev = adapter->pdev;
>>> + struct vdpa_notification_area area;
>>> +
>>> + area.addr = vf->vring[idx].notify_pa;
>>> + area.size = PAGE_SIZE;
>>> + if (area.addr % PAGE_SIZE)
>>> + IFCVF_DBG(pdev, "vq %u doorbell address is not PAGE_SIZE
>>> aligned\n", idx);
>>
>>
>> Let's leave the decision to upper layer by: (see
>> vp_vdpa_get_vq_notification)
>>
>> area.addr = notify_pa;
>> area.size = notify_offset_multiplier;
>>
>> Thanks
>
> Hi Jason,
>
> notify_offset_multiplier can be zero, means vqs share the same
> doorbell address, distinguished by qid.
> and in vdpa.c:
>
> if (vma->vm_end - vma->vm_start != notify.size)
> return -ENOTSUPP;
>
> so a zero size would cause this feature failure.
> mmap should work on at least a page, so if we really want "area.size =
> notify_offset_multiplier;"
> I think we should add some code in vdpa.c, like:
>
> if(!notify.size)
> notify.size = PAGE_SIZE;
>
> sounds good?
It's the responsibility of the driver to report a correct one. So I
think it's better to tweak it as:
area.size = notify_offset_multiplier ?: PAGE_SIZE;
Thanks
>
> Thanks
> Zhu Lingshan
>>
>>
>>> +
>>> + return area;
>>> +}
>>> +
>>> /*
>>> * IFCVF currently does't have on-chip IOMMU, so not
>>> * implemented set_map()/dma_map()/dma_unmap()
>>> @@ -440,6 +456,7 @@ static const struct vdpa_config_ops ifc_vdpa_ops
>>> = {
>>> .get_config = ifcvf_vdpa_get_config,
>>> .set_config = ifcvf_vdpa_set_config,
>>> .set_config_cb = ifcvf_vdpa_set_config_cb,
>>> + .get_vq_notification = ifcvf_get_vq_notification,
>>> };
>>> static int ifcvf_probe(struct pci_dev *pdev, const struct
>>> pci_device_id *id)
>>
>
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
^ permalink raw reply [flat|nested] 2+ messages in thread