[dpdk-dev] [PATCH v3 0/7] virtio_user as an alternative exception path

[Bruce Richardson]

On Wed, Jan 04, 2017 at 03:59:19AM +0000, Jianfeng Tan wrote:
> v3:
>   - Drop the patch to postpone driver ok sending patch, superseded it
>     with a bug fix to disable all virtqueues and re-init the device.
>     (you might wonder why not just send reset owner msg. Under my test,
>      it causes spinlock deadlock problem when killing the program).
>   - Avoid compiling error on 32-bit system for pointer convert.
>   - Fix a bug in patch "abstract virtio user backend ops", vhostfd is
>     not properly assigned.
>   - Fix a "MQ cannot be used" bug in v2, which is related to strip
>     some feature bits that vhost kernel does not recognize.
>   - Update release note.
> 
> v2: (Lots of them are from yuanhan's comment)
>   - Add offloding feature.
>   - Add multiqueue support.
>   - Add a new patch to postpone the sending of driver ok notification.
>   - Put fix patch ahead of the whole patch series.
>   - Split original 0001 patch into 0003 and 0004 patches.
>   - Remove the original vhost_internal design, just add those into
>     struct virtio_user_dev for simplicity.
>   - Reword "control" to "send_request".
>   - Reword "host_features" to "device_features". 
> 
> In v16.07, we upstreamed a virtual device, virtio_user (with vhost-user
> as the backend). The path to go with a vhost-kernel backend has been
> dropped for bad performance comparing to vhost-user and code simplicity.
> 
> But after a second thought, virtio_user + vhost-kernel is a good 
> candidate as an exceptional path, such as KNI, which exchanges packets
> with kernel networking stack.
>   - maintenance: vhost-net (kernel) is upstreamed and extensively used 
>     kernel module. We don't need any out-of-tree module like KNI.
>   - performance: as with KNI, this solution would use one or more
>     kthreads to send/receive packets from user space DPDK applications,
>     which has little impact on user space polling thread (except that
>     it might enter into kernel space to wake up those kthreads if
>     necessary).
>   - features: vhost-net is born to be a networking solution, which has
>     lots of networking related featuers, like multi queue, tso, multi-seg
>     mbuf, etc.
> 
> Signed-off-by: Jianfeng Tan <jianfeng.tan at intel.com>
> 
Sounds great. However, I think we'll need a how-to doc for this to help
people get it up and running as a KNI replacement for packets to/from
the kernel. Any plans to draw up such a doc? It would be good to include
it in this patchset.

/Bruce