Re: [PATCH] vhost: notify guest to fill buffer when there is no buffer

[Linhaifeng <haifeng.lin-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>]

On 2015/1/29 21:00, Xie, Huawei wrote:
> 
> 
>> -----Original Message-----
>> From: Linhaifeng [mailto:haifeng.lin-hv44wF8Li93QT0dZR+AlfA@public.gmane.org]
>> Sent: Thursday, January 29, 2015 8:39 PM
>> To: Xie, Huawei; dev-VfR2kkLFssw@public.gmane.org
>> Subject: Re: [dpdk-dev] [PATCH] vhost: notify guest to fill buffer when there is
>> no buffer
>>
>>
>>
>> On 2015/1/29 18:39, Xie, Huawei wrote:
>>
>>>> -		if (count == 0)
>>>> +		/* If there is no buffers we should notify guest to fill.
>>>> +		* This is need when guest use virtio_net driver(not pmd).
>>>> +		*/
>>>> +		if (count == 0) {
>>>> +			if (!(vq->avail->flags &
>>>> VRING_AVAIL_F_NO_INTERRUPT))
>>>> +				eventfd_write((int)vq->kickfd, 1);
>>>>  			return 0;
>>>> +		}
>>>
>>> Haifeng:
>>> Is it the root cause and is it protocol required?
>>> Could you give a detailed description for that scenario?
>>>
>>
>> I use mz to send data from one VM1 to VM2.The two VM use virtio-net driver.
>> VM1 execute follow script:
>> for((i=0;i<999999999;i++));
>> do
>> mz eth0 -t udp -A 1.1.1.1 -B 1.1.1.2 -a 00:00:00:00:00:01 -b 00:00:00:00:00:02 -c
>> 10000000 -p 512
>> sleep 4
>> done
>>
>> VM2 execute follow command to watch:
>> watch -d ifconfig
>>
>> After many hours VM2 stop to receive data.
>>
>> Could you test it ?
> 
> 
> We could try next week after I send the whole patch. 
> How many hours? Is it reproducible at your side? I inject packets through packet generator to guest for more than ten hours, haven't met issues. 

About three hours.
What kind of driver you used in guest?virtio-net-pmd or virtio-net?


> As I said in another mail sent  to you, could you dump the status of vring if you still have the spot?

How to dump the status of vring in guest?

> Could you please also reply to that mail?
>

Which mail?


> For the patch, if we have no root cause, I prefer not to apply it, so that we don't send more interrupts than needed to guest to affect performance.

I found that if we add this notify the performance is better(growth of 100kpps when use 64byte UDP packets)

> People could temporarily apply this patch as a work around.
> 
> Or anyone 
> 

OK.I'm also not sure about this bug.I think i should do something to found the real reason.

> 
>> --
>> Regards,
>> Haifeng
> 
> 
> 

-- 
Regards,
Haifeng