[dpdk-dev] SRIOV hot unplug

[Declan Doherty]

On 07/19/2016 12:18 PM, Eli Britstein wrote:
>
>
>> -----Original Message-----
>> From: Tetsuya Mukawa [mailto:mukawa at igel.co.jp]
>> Sent: Tuesday, 19 July, 2016 5:49 AM
>> To: Eli Britstein; dev at dpdk.org
>> Cc: Iremonger, Bernard
>> Subject: Re: [dpdk-dev] SRIOV hot unplug
>>
>> Hi Eli,
>>
>> On 2016/07/18 17:47, Eli Britstein wrote:
>>> Hi Bernard,
>>>
>>> Thank you for your answer.
>>> However, to do this, I have to have some communication protocol to the
>> VM's application in order for it to do this sequence and acknowledge that it is
>> now safe to proceed with detaching the device.
>>> This implies some kind of integration from the host side, which I would like
>> to avoid.
>>
>> I guess you should have some kind of communication channel to notice the
>> hotpluging events from host to VM.
> [Eli Britstein]
> In order just to notice the hotplugging events inside the VM, I can use add some udev action in the VM, in /etc/udev/rules.d/XXX
> However, those are asynchronous events. The host proceeds with unplugging without waiting for the VM to acknowledge it.
>
>>
>>> Do you think might there be any other way for the application to handle
>> such event in a smooth way?
>>
>> So far, I guess having one more virtio-net device will be easiest way.
> [Eli Britstein]
> Could you please elaborate your meaning? How do you mean to use this extra virtio-net device?
>
> To clarify: I would like to have my bond device automatically set the vNIC as its primary, and close/remove the VF before it is unplugged.
>
> Thanks,
> Eli

The only that I can think of which would allow bonding to support this 
is if you could force a link down notification to the VF. The bonding 
driver should automatically then fail over to the vNIC, but again this 
is asynchronous event so I'm not sure how you could confirm this action 
from the host other than by monitoring traffic?

Also I don't know if the DPDK environment has been tested to support the 
removal of dynamic hot-plugging you are suggesting via the udev action. 
There are no hooks that I'm aware of which would detect that the PCI 
device is no longer available and then clean the resources away in DPDK.

>
<snip>
>

Also could you please drop the footer/disclaimer on replies to the 
mailing list (see http://dpdk.org/ml)