Re: [RFC PATCH] vhost: remove lockless enqueue to virt ring

Re: [RFC PATCH] vhost: remove lockless enqueue to virt ring

[Xie, Huawei]

If no objections, would remove the internal cmpset enqueue, should get
some improved performance.

No idea why commit message is lost:
DPDK doesn't support receiving/transmitting packets from/to the same
port/queue.
Vhost pmd wraps vhost device as normal dpdk port. dpdkvhost based
applications/switch
 also treate virtio port and physical port both as ports, and have
their  own
lock implmentation when to enqueue packets to the same port. This patch
removes
the internal lockless vhost enqueue implmentation.
On 11/5/2015 5:38 PM, Huawei Xie wrote:
> Signed-off-by: Huawei Xie <huawei.xie@intel.com>

Re: [RFC PATCH] vhost: remove lockless enqueue to virt ring

[Tetsuya Mukawa]

On 2015/11/05 19:21, Xie, Huawei wrote:
> If no objections, would remove the internal cmpset enqueue, should get
> some improved performance.
>
> No idea why commit message is lost:
> DPDK doesn't support receiving/transmitting packets from/to the same
> port/queue.
> Vhost pmd wraps vhost device as normal dpdk port. dpdkvhost based
> applications/switch
>  also treate virtio port and physical port both as ports, and have
> their  own
> lock implmentation when to enqueue packets to the same port. This patch
> removes
> the internal lockless vhost enqueue implmentation.
> On 11/5/2015 5:38 PM, Huawei Xie wrote:
>> Signed-off-by: Huawei Xie <huawei.xie@intel.com>

Hi Xie,

I think it's a good optimization.
But I  guess the patch will changes behavior of vhost library API.
For example, if someone has already implemented an application based on
the assumption that vhost library has internal lock, I doubt their
application will face an issue with this patch.

If it's correct, we might need to follow below steps.
http://dpdk.org/doc/guides/contributing/versioning.html

Thanks,
Tetsuya

Re: [RFC PATCH] vhost: remove lockless enqueue to virt ring

[Xie, Huawei]

On 11/6/2015 9:43 AM, Tetsuya Mukawa wrote:
> On 2015/11/05 19:21, Xie, Huawei wrote:
>> If no objections, would remove the internal cmpset enqueue, should get
>> some improved performance.
>>
>> No idea why commit message is lost:
>> DPDK doesn't support receiving/transmitting packets from/to the same
>> port/queue.
>> Vhost pmd wraps vhost device as normal dpdk port. dpdkvhost based
>> applications/switch
>>  also treate virtio port and physical port both as ports, and have
>> their  own
>> lock implmentation when to enqueue packets to the same port. This patch
>> removes
>> the internal lockless vhost enqueue implmentation.
>> On 11/5/2015 5:38 PM, Huawei Xie wrote:
>>> Signed-off-by: Huawei Xie <huawei.xie@intel.com>
> Hi Xie,
>
> I think it's a good optimization.
> But I  guess the patch will changes behavior of vhost library API.
> For example, if someone has already implemented an application based on
> the assumption that vhost library has internal lock, I doubt their
> application will face an issue with this patch.
Yes, i understand. If we have an improper implementation, we fix it as
early as possible. I prefer to remove the internal lock currently. The
last thing i want to do is maintain two APIs.
Would follow the ABI process.
>
> If it's correct, we might need to follow below steps.
> http://dpdk.org/doc/guides/contributing/versioning.html
>
> Thanks,
> Tetsuya
>

Re: [RFC PATCH] vhost: remove lockless enqueue to virt ring

[Changchun Ouyang]

> From: huawei.xie@intel.com
> To: huawei.xie@intel.com; dev@dpdk.org
> Date: Thu, 5 Nov 2015 10:21:41 +0000
> CC: haifeng.lin@intel.com; ann.zhuangyanying@huawei.com
> Subject: Re: [dpdk-dev] [RFC PATCH] vhost: remove lockless enqueue to virt	ring
> 
> If no objections, would remove the internal cmpset enqueue, should get
> some improved performance.
I think it is good.The vhost sample also need some changes for this modification.do we have the percentage of gaining when removing the lock?

thanks and regards,Changchun
> 
> No idea why commit message is lost:
> DPDK doesn't support receiving/transmitting packets from/to the same
> port/queue.
> Vhost pmd wraps vhost device as normal dpdk port. dpdkvhost based
> applications/switch
>  also treate virtio port and physical port both as ports, and have
> their  own
> lock implmentation when to enqueue packets to the same port. This patch
> removes
> the internal lockless vhost enqueue implmentation.
> On 11/5/2015 5:38 PM, Huawei Xie wrote:
> > Signed-off-by: Huawei Xie <huawei.xie@intel.com>
> 
 		 	   		  

Re: [RFC PATCH] vhost: remove lockless enqueue to virt ring

[Xie, Huawei]

On 11/6/2015 9:48 AM, Changchun Ouyang wrote:
>
>
> > From: huawei.xie@intel.com
> > To: huawei.xie@intel.com; dev@dpdk.org
> > Date: Thu, 5 Nov 2015 10:21:41 +0000
> > CC: haifeng.lin@intel.com; ann.zhuangyanying@huawei.com
> > Subject: Re: [dpdk-dev] [RFC PATCH] vhost: remove lockless enqueue
> to virt ring
> >
> > If no objections, would remove the internal cmpset enqueue, should get
> > some improved performance.
>
> I think it is good.
> The vhost sample also need some changes for this modification.
> do we have the percentage of gaining when removing the lock?
The atomic cmpset is a costly operation.
We plan to implement vhost/virtio performance test example, which will
cover the most common scenarios.
/huawei
>
> thanks and regards,
> Changchun
>