qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [PATCH 5/7] vhost_net: move vhost_net_set_vq_index ahea


From: Yuanhan Liu
Subject: Re: [Qemu-devel] [PATCH 5/7] vhost_net: move vhost_net_set_vq_index ahead at vhost_net_init
Date: Mon, 14 Sep 2015 13:49:20 +0800
User-agent: Mutt/1.5.21 (2010-09-15)

On Thu, Sep 10, 2015 at 02:54:02PM +0800, Jason Wang wrote:
> 
> 
> On 09/10/2015 02:18 PM, Yuanhan Liu wrote:
> > On Thu, Sep 10, 2015 at 01:52:30PM +0800, Jason Wang wrote:
> >>
> >> On 09/10/2015 01:17 PM, Yuanhan Liu wrote:
> >>> On Thu, Sep 10, 2015 at 12:46:00PM +0800, Jason Wang wrote:
> >>>>>
> >>>>> On 09/10/2015 11:57 AM, Yuanhan Liu wrote:
> >>>>>>> On Thu, Sep 10, 2015 at 11:14:27AM +0800, Jason Wang wrote:
> >>>>>>>>> On 09/08/2015 03:38 PM, Yuanhan Liu wrote:
> >>>>>>>>>>> So that we could use the `vq_index' as well in the vhost_net_init
> >>>>>>>>>>> stage, which is required when adding vhost-user multiple-queue 
> >>>>>>>>>>> support,
> >>>>>>>>>>> where we need the vq_index to indicate which queue pair we are 
> >>>>>>>>>>> gonna
> >>>>>>>>>>> initiate.
> >>>>>>>>>>>
> >>>>>>>>>>> vhost-user has no multiple queue support yet, hence no 
> >>>>>>>>>>> queue_index set
> >>>>>>>>>>> before. Here is a quick set to 0 at net_vhost_user_init() stage, 
> >>>>>>>>>>> and it
> >>>>>>>>>>> will be set properly soon in the next patch.
> >>>>>>>>>>>
> >>>>>>>>>>> Signed-off-by: Yuanhan Liu <address@hidden>
> >>>>>>>>>>> ---
> >>>>>>>>>>>  hw/net/vhost_net.c | 16 +++++++---------
> >>>>>>>>>>>  net/vhost-user.c   |  1 +
> >>>>>>>>>>>  2 files changed, 8 insertions(+), 9 deletions(-)
> >>>>>>>>>>>
> >>>>>>>>>>> diff --git a/hw/net/vhost_net.c b/hw/net/vhost_net.c
> >>>>>>>>>>> index f9441e9..141b557 100644
> >>>>>>>>>>> --- a/hw/net/vhost_net.c
> >>>>>>>>>>> +++ b/hw/net/vhost_net.c
> >>>>>>>>>>> @@ -138,6 +138,11 @@ static int vhost_net_get_fd(NetClientState 
> >>>>>>>>>>> *backend)
> >>>>>>>>>>>      }
> >>>>>>>>>>>  }
> >>>>>>>>>>>  
> >>>>>>>>>>> +static void vhost_net_set_vq_index(struct vhost_net *net, int 
> >>>>>>>>>>> vq_index)
> >>>>>>>>>>> +{
> >>>>>>>>>>> +    net->dev.vq_index = vq_index;
> >>>>>>>>>>> +}
> >>>>>>>>>>> +
> >>>>>>>>>>>  struct vhost_net *vhost_net_init(VhostNetOptions *options)
> >>>>>>>>>>>  {
> >>>>>>>>>>>      int r;
> >>>>>>>>>>> @@ -167,6 +172,8 @@ struct vhost_net 
> >>>>>>>>>>> *vhost_net_init(VhostNetOptions *options)
> >>>>>>>>>>>      }
> >>>>>>>>>>>      net->nc = options->net_backend;
> >>>>>>>>>>>  
> >>>>>>>>>>> +    vhost_net_set_vq_index(net, net->nc->queue_index * 2);
> >>>>>>>>>>> +
> >>>>>>>>> This breaks vhost kernel multiqueue since queue_index was not
> >>>>>>>>> initialized at this time.
> >>>>>>> Right, thanks for pointing it out.
> >>>>>>>
> >>>>>>>>> We do this in set_netdev() instead of setting
> >>>>>>>>> it in each kind of netdev.
> >>>>>>> Can we move it to net_init_tap() for setting the right queue_index
> >>>>>>> for each nc?
> >>>>>>>
> >>>>>>> Or, can we call vhost_net_set_vq_index twice, one at 
> >>>>>>> vhost_net_init(for
> >>>>>>> vhost-user mq support), another one at vhost_net_start(for vhost 
> >>>>>>> kernel
> >>>>>>> mq support)?
> >>>>>>>
> >>>>>>> Or, do you have better ideas?
> >>>>> I think setting queue_index in net_init_tap() looks ok.
> >>> Good to know.
> >>>
> >>>>> But a question
> >>>>> is that why need we do this at so early stage? ( Even before its peers
> >>>>> is connected.)
> >>> For vhost-user multiple queues support, we will invoke vhost_net_init()
> >>> N times for each queue pair, and hence we need to distinguish which
> >>> queue it is while sending messages like VHOST_SET_VRING_CALL for
> >>> initializing corresponding queue pair.
> >>>
> >>> Does that make sense to you?
> >>>
> >> Not sure. Since current codes works for vhost-kernel. (vhost_net_init()
> >> was also called N times). We don't want to break existed vhost-kernel
> >> API when developing multiqueue. For each virtqueue TX/RX pair, we have
> >> one vhost net device and it has no knowledge for the others (which was
> >> hide by qemu). So VHOST_SET_VRING_CALL works without any change here.
> >>
> >> For the case here, since you still have multiple instances of vhost_net
> >> structure. Maybe the vhost-user backend can distinguish form this?
> > Yeah, I guess that's the difference between vhost-user and vhost-kernel.
> > Vhost-kernel opens a char device(/dev/vhost-net) for each vhost_dev,
> > hence it's distinguishable. But for vhost-user, all vhost_dev share one
> > char device(a socket) for communication, hence, it's not distinguishable.
> 
> How about using individual socket in this case? This seems can also
> minimize the changes of backend.
> 
> >
> > I was thinking maybe we could export vhost_net_set_vq_index() and invoke
> > it at net/vhost-user.c, so that we break nothing, and in the meantime,
> > it keeps the logic inside vhost-user.
> >
> > What do you think?
> >
> >     --yliu
> >
> 
> Sounds work. Then I believe you will need to set queue_index in
> vhost_user initialization code?

Nah, it will not work, as vhost_net_set_vq_index() needs a vhost_net as
it's parameter: you can't do that before vhost_net_init(), but it's
useless to do that after vhost_net_init(). My bad for not being aware it
in the first time.

How about following then?

Thanks.

        --yliu

---
diff --git a/hw/net/vhost_net.c b/hw/net/vhost_net.c
index e2985bc..7197c3f 100644
--- a/hw/net/vhost_net.c
+++ b/hw/net/vhost_net.c
@@ -148,6 +148,7 @@ struct vhost_net *vhost_net_init(VhostNetOptions *options)
         fprintf(stderr, "vhost-net requires net backend to be setup\n");
         goto fail;
     }
+    net->nc = options->net_backend;

     net->dev.max_queues = 1;

@@ -164,8 +165,10 @@ struct vhost_net *vhost_net_init(VhostNetOptions *options)
         net->dev.backend_features = 0;
         net->dev.protocol_features = 0;
         net->backend = -1;
+
+        /* vhost-user needs vq_index to initiate a specific queue pair */
+        net->dev.vq_index = net->nc->queue_index * 2;
     }
-    net->nc = options->net_backend;

     net->dev.nvqs = 2;
     net->dev.vqs = net->vqs;



reply via email to

[Prev in Thread] Current Thread [Next in Thread]