Re: [dpdk-dev] [PATCH 0/3] add ifcvf driver

[Maxime Coquelin <maxime.coquelin@redhat.com>]

On 03/26/2018 11:05 AM, Wang, Xiao W wrote:
> Hi Maxime,
> 
>> -----Original Message-----
>> From: Maxime Coquelin [mailto:maxime.coquelin@redhat.com]
>> Sent: Sunday, March 25, 2018 5:51 PM
>> To: Wang, Xiao W <xiao.w.wang@intel.com>; dev@dpdk.org
>> Cc: Wang, Zhihong <zhihong.wang@intel.com>; yliu@fridaylinux.org; Liang,
>> Cunming <cunming.liang@intel.com>; Xu, Rosen <rosen.xu@intel.com>; Chen,
>> Junjie J <junjie.j.chen@intel.com>; Daly, Dan <dan.daly@intel.com>
>> Subject: Re: [PATCH 0/3] add ifcvf driver
>>
>>
>>
>> On 03/23/2018 11:27 AM, Wang, Xiao W wrote:
>>> Hi Maxime,
>>>
>>>> -----Original Message-----
>>>> From: Maxime Coquelin [mailto:maxime.coquelin@redhat.com]
>>>> Sent: Thursday, March 22, 2018 4:48 AM
>>>> To: Wang, Xiao W <xiao.w.wang@intel.com>; dev@dpdk.org
>>>> Cc: Wang, Zhihong <zhihong.wang@intel.com>; yliu@fridaylinux.org; Liang,
>>>> Cunming <cunming.liang@intel.com>; Xu, Rosen <rosen.xu@intel.com>;
>> Chen,
>>>> Junjie J <junjie.j.chen@intel.com>; Daly, Dan <dan.daly@intel.com>
>>>> Subject: Re: [PATCH 0/3] add ifcvf driver
>>>>
>>>> Hi Xiao,
>>>>
>>>> On 03/15/2018 05:49 PM, Wang, Xiao W wrote:
>>>>> Hi Maxime,
>>>>>
>>>>>> -----Original Message-----
>>>>>> From: Maxime Coquelin [mailto:maxime.coquelin@redhat.com]
>>>>>> Sent: Sunday, March 11, 2018 2:24 AM
>>>>>> To: Wang, Xiao W <xiao.w.wang@intel.com>; dev@dpdk.org
>>>>>> Cc: Wang, Zhihong <zhihong.wang@intel.com>; yliu@fridaylinux.org;
>> Liang,
>>>>>> Cunming <cunming.liang@intel.com>; Xu, Rosen <rosen.xu@intel.com>;
>>>> Chen,
>>>>>> Junjie J <junjie.j.chen@intel.com>; Daly, Dan <dan.daly@intel.com>
>>>>>> Subject: Re: [PATCH 0/3] add ifcvf driver
>>>>>>
>>>>>> Hi Xiao,
>>>>>>
>>>>>> On 03/10/2018 12:08 AM, Xiao Wang wrote:
>>>>>>> This patch set has dependency on
>>>>>> http://dpdk.org/dev/patchwork/patch/35635/
>>>>>>> (vhost: support selective datapath);
>>>>>>>
>>>>>>> ifc VF is compatible with virtio vring operations, this driver implements
>>>>>>> vDPA driver ops which configures ifc VF to be a vhost data path
>> accelerator.
>>>>>>>
>>>>>>> ifcvf driver uses vdev as a control domain to manage ifc VFs that belong
>>>>>>> to it. It registers vDPA device ops to vhost lib to enable these VFs to be
>>>>>>> used as vhost data path accelerator.
>>>>>>>
>>>>>>> Live migration feature is supported by ifc VF and this driver enables
>>>>>>> it based on vhost lib.
>>>>>>>
>>>>>>> vDPA needs to create different containers for different devices, thus this
>>>>>>> patch set adds APIs in eal/vfio to support multiple container.
>>>>>> Thanks for this! That will avoind having to duplicate these functions
>>>>>> for every new offload driver.
>>>>>>
>>>>>>
>>>>>>>
>>>>>>> Junjie Chen (1):
>>>>>>>       eal/vfio: add support for multiple container
>>>>>>>
>>>>>>> Xiao Wang (2):
>>>>>>>       bus/pci: expose sysfs parsing API
>>>>>>
>>>>>> Still, I'm not convinced the offload device should be a virtual device.
>>>>>> It is a real PCI device, why not having a new device type for offload
>>>>>> devices, and let the device to be probed automatically by the existing
>>>>>> device model?
>>>>>
>>>>> IFC VFs are generated from SRIOV, with the PF driven by kernel driver.
>>>>> In DPDK we need to have something to represent PF, to register itself as
>>>>> a vDPA engine, so a virtual device is used for this purpose.
>>>> I went through the code, and something is not clear to me.
>>>>
>>>> Why do we need to have a representation of the PF in DPDK?
>>>> Why cannot we just bind at VF level?
>>>
>>> 1. With the vdev representation we could use it to talk to PF kernel driver to
>> do flow configuration, we can implement
>>> flow API on the vdev in future for this purpose. Using a vdev allows
>> introducing this kind of control plane thing.
>>>
>>> 2. When port representor is ready, we would integrate it into ifcvf driver,
>> then each VF will have a
>>> Representor port. For now we don’t have port representor, so this patch set
>> manages VF resource internally.
>>
>> Ok, we may need to have a vdev to represent the PF, but we need to be
>> able to bind at VF level anyway.
> 
> Device management on VF level is feasible, according to the previous port-representor patch.
> A tuple of (PF_addr, VF_index) can identify a certain VF, we have vport_mask
> and device addr to describe a PF, and we can specify a VF index to create a representor port,
> so , the VF port creation will be on-demand at VF level.
> 
> +struct port_rep_parameters {
> +	uint64_t vport_mask;
> +	struct {
> +		char bus[RTE_DEV_NAME_MAX_LEN];
> +		char device[RTE_DEV_NAME_MAX_LEN];
> +	} parent;
> +};
> 
> +int
> +rte_representor_port_register(char *pf_addr_str,
> +		uint32_t vport_id, uint16_t *port_id)

IIUC, even with this using port representor, we'll still have the
problem of having the VFs probed first as Virtio driver, right?

In my opinion, the IFCVF devices in offload mode are to be managed 
differently than having a way to represent on host side VFs assigned to
a VM.

In offload case, you have a real device to deal with, else we
wouldn't have to bind it with VFIO.

Maybe we could have a real device probed as proposed yesterday [0], and 
this device gets registered to the port representor for the PF?

Thanks,
Maxime
> Besides, IFCVF supports live migration, vDPA exerts IFCVF device better than QEMU (this patch has enabled LM feature).
> vDPA is the main usage model for IFCVF, and one DPDK application taking control of all the VF resource
> management is a straightforward usage model.
> 
> Best Regards,
> Xiao
> 
>>
>> Else, how do you support passing two VFs of the same PF to different
>> DPDK applications?
>> Or have some VFs managed by Kernel or QEMU and some by the DPDK
>> application? My feeling is that current implementation is creating an
>> artificial constraint.
>>
>> Isn't there a possibility to have the virtual representation for the PF
>> to be probed separately? Or created automatically when the first VF of a
>> PF is probed (and later VFs attach to the PF rep when probed)?
>>
>> Doing this, we could use the generic device probing.

[0]:
>> For IFCVF, to specify we want it to be probed as an offload device
>> instead of a virtio device, we could have a new EAL parameter to specify
>> for a given device if we want it to be probed as an offload device (for
>> example --offload=00:01.1).
>> Offload drivers would register by passing a flag that specifies they are
>> an offload driver.
>> This new argument would be optional and only used to force the device to
>> being probed in offload mode. For devices that have their own device
>> IDs for offload mode, then it would be automatic.
>>
>> Regards,
>> Maxime
>>> BRs,
>>> Xiao
>>>