Re: [dpdk-dev] [PATCH 2/3] net/vdpa_virtio_pci: introduce vdpa sample driver

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

Hi Xiao,

On 02/08/2018 03:23 AM, Wang, Xiao W wrote:
> Hi Maxime,
> 
>> -----Original Message-----
>> From: Maxime Coquelin [mailto:maxime.coquelin@redhat.com]
>> Sent: Tuesday, February 6, 2018 10:24 PM
>> To: Wang, Xiao W <xiao.w.wang@intel.com>; dev@dpdk.org
>> Cc: Tan, Jianfeng <jianfeng.tan@intel.com>; Bie, Tiwei <tiwei.bie@intel.com>;
>> yliu@fridaylinux.org; Liang, Cunming <cunming.liang@intel.com>; Daly, Dan
>> <dan.daly@intel.com>; Wang, Zhihong <zhihong.wang@intel.com>
>> Subject: Re: [PATCH 2/3] net/vdpa_virtio_pci: introduce vdpa sample driver
>>
>> Hi Xiao,
>>
>> On 02/04/2018 03:55 PM, Xiao Wang wrote:
>>> This driver is a reference sample of making vDPA device driver based
>>> on vhost lib, this driver uses a standard virtio-net PCI device as
>>> vDPA device, it can serve as a backend for a virtio-net pci device
>>> in nested VM.
>>>
>>> The key driver ops implemented are:
>>>
>>> * vdpa_virtio_eng_init
>>> Mapping virtio pci device with VFIO into userspace, and read device
>>> capability and intialize internal data.
>>>
>>> * vdpa_virtio_eng_uninit
>>> Release the mapped device.
>>>
>>> * vdpa_virtio_info_query
>>> Device capability reporting, e.g. queue number, features.
>>>
>>> * vdpa_virtio_dev_config
>>> With the guest virtio information provideed by vhost lib, this
>>> function configures device and IOMMU to set up vhost datapath,
>>> which includes: Rx/Tx vring, VFIO interrupt, kick relay.
>>>
>>> * vdpa_virtio_dev_close
>>> Unset the stuff that are configured previously by dev_conf.
>>>
>>> This driver requires the virtio device supports VIRTIO_F_IOMMU_PLATFORM
>>> , because the buffer address written in desc is IOVA.
>>>
>>> Because vDPA driver needs to set up MSI-X vector to interrupt the guest,
>>> only vfio-pci is supported currently.
>>>
>>> Signed-off-by: Xiao Wang<xiao.w.wang@intel.com>
>>> ---
>>>    config/common_base                                 |    6 +
>>>    config/common_linuxapp                             |    1 +
>>>    drivers/net/Makefile                               |    1 +
>>>    drivers/net/vdpa_virtio_pci/Makefile               |   31 +
>>>    .../net/vdpa_virtio_pci/rte_eth_vdpa_virtio_pci.c  | 1527
>> ++++++++++++++++++++
>>>    .../rte_vdpa_virtio_pci_version.map                |    4 +
>>>    mk/rte.app.mk                                      |    1 +
>>>    7 files changed, 1571 insertions(+)
>>>    create mode 100644 drivers/net/vdpa_virtio_pci/Makefile
>>>    create mode 100644 drivers/net/vdpa_virtio_pci/rte_eth_vdpa_virtio_pci.c
>>>    create mode 100644
>> drivers/net/vdpa_virtio_pci/rte_vdpa_virtio_pci_version.map
>>
>> Is there a specific constraint that makes you expose PCI functions and
>> duplicate a lot of vfio code into the driver?
> 
> The existing vfio code doesn't fit VDPA well, this vDPA driver needs to program IOMMU for a vDPA device with a VM's memory table.
> While the eal/vfio uses a struct vfio_cfg to takes all regular devices and add them to a single vfio_container, and program IOMMU with DPDK process's memory table.
> 
> This driver doing PCI VFIO initialization itself can avoid affecting the global vfio_cfg structure.

Ok, I get it.
So I think what you have to do is to extend eal/vfio for this case.
Or at least, have a vdpa layer to perform this, else every offload
driver will have to duplicate the code.

>>
>> Wouldn't it be better (if possible) to use RTE_PMD_REGISTER_PCI() & co.
>> to benefit from all the existing infrastructure?
> 
> RTE_PMD_REGISTER_PCI() & co will make this driver as PCI driver (physical device), then this will conflict with the virtio_pmd.
> So I make vDPA device driver as a vdev driver.

Yes, but it is a PCI device, not a virtual device. You have to extend
the EAL to support this new class of devices/drivers. Think of it as in
kernel when a NIC device can be either binded to its NIC driver, VFIO or
UIO.

If I look at patch 3, you have to set --no-pci, or at least I think to
blacklist the Virtio device.

I wonder if real vDPA cards will support either vDPA mode or or behave
like a regular NIC, like the Virtio case in your example.
If this is the case, maybe the vDPA code for a NIC could be in the same
driver as the "NIC" mode.
A new struct rte_pci_driver driver flag could be introduced to specify
that the driver supports vDPA.
Then, in EAL arguments, if a vhost vdev specifies it wants Virtio device
at PCI addr 00:01:00 as offload, the PCI layer could probe this device
in "vdpa" mode.

Also, I don't know if this will be possible with real vDPA cards, but we
could have the application doing packet switching between vhost-user
vdev and the Virtio device. And at some point, at runtime, switch into
vDPA mode. This use-case would be much easier to implement if vDPA
relied on existing PCI layer.

I may be not very clear, don't hesitate to ask questions.
But generally, I think vDPA has to fit in existing DPDK architecture,
and not try to live outside of it.

Thanks,
Maxime
>>
>> Maxime
> 
> Thanks for the comments,
> Xiao
>