Re: [ovs-dev] The Patch netdev-dpdk: add TSO support for vhost-user ports

[Michael Qiu]

[+cc DPDK jianfeng]


For the gap, is it been fixed by below commit?

commit c59faf3fe8fe6dbd99e8cd9d703839728264c5ae
Author: Jianfeng Tan 
Date:   Mon Aug 1 03:56:54 2016 +

net/i40e: support TSO on tunneling packet


I just tested your vlan patch, the same thing happens as vxlan tunnel TSO.

BTW, the workaround is in DPDK or in OVS? Any more details?



2016/10/18 0:00, Kavanagh, Mark B :

OK, Thanks for your info.

BTW, you mentioned the gap, is it the tunneling tso can't work correctly for 
multi-segments
mbuf, and after some packets, guest is unable to xmit packets(dpdk driver 
failed to work any
more)?

At the time, the gap was that the tunnel header was not replicated across all 
of the NIC-fragmented frames; that has been resolved now in DPDK.

However, while testing with TSO with single-tagged VLAN frames, I see something 
similar to what you describe. I currently have a workaround, but want to do a 
bit more digging before I report out on it.

Thanks,
Mark


Thanks,
Michael

2016年10月14日，下午5:15，Kavanagh, Mark B  ：



Hi, Mark


OK, once it's ready， pls let me know and I'm glade to help to test it.


BTW， you mentioned the gap, is that the TSO and CSUM of tunnel offload?
I'm familiar with it, and I hope I could do something on it if possible.

Yes, that's what it was.

Commits have recently been introduced to DPDK that introduce additional flags 
for tunnel

offloading, but I haven't had a chance to test them with in my TSO+VxLAN 
codebase.

Thanks,
Mark



2016/10/11 22:34, Kavanagh, Mark B :

Hi,  all


This patch is very important for users want to improve the performance
of the large packets.


But you know, in data center, lots of networks using vxlan, so if it
supports vxlan, then it will be very useful.


Would you guys has a plan to support it? I would like to help test it,
or work together on it?

Hi Michael,

Some work has already been done to enable TSO over VxLAN, but it was blocked 
until

recently, on account of gaps in the DPKD i40e PMD's support for tunnel offload.

We believe that those gaps have now been addressed; our current focus is on 
ironing out

the

issues in TSO support over flat and VLAN networks - once that's done, we plan 
to resume

work

on VxLAN support (and GRE, IPinIP).

Hope this helps,
Mark


--
Thanks,
Michael

--
Thanks,
Michael


--
Thanks,
Michael


___
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev

Re: [ovs-dev] The Patch netdev-dpdk: add TSO support for vhost-user ports

[Kavanagh, Mark B]

>
>OK, Thanks for your info.
>
>BTW, you mentioned the gap, is it the tunneling tso can't work correctly for 
>multi-segments
>mbuf, and after some packets, guest is unable to xmit packets(dpdk driver 
>failed to work any
>more)?

At the time, the gap was that the tunnel header was not replicated across all 
of the NIC-fragmented frames; that has been resolved now in DPDK.

However, while testing with TSO with single-tagged VLAN frames, I see something 
similar to what you describe. I currently have a workaround, but want to do a 
bit more digging before I report out on it.

Thanks,
Mark

>
>Thanks,
>Michael
>
>2016年10月14日，下午5:15，Kavanagh, Mark B  ：
>
>>>
>>>
>>> Hi, Mark
>>>
>>>
>>> OK, once it's ready， pls let me know and I'm glade to help to test it.
>>>
>>>
>>> BTW， you mentioned the gap, is that the TSO and CSUM of tunnel offload?
>>> I'm familiar with it, and I hope I could do something on it if possible.
>>
>> Yes, that's what it was.
>>
>> Commits have recently been introduced to DPDK that introduce additional 
>> flags for tunnel
>offloading, but I haven't had a chance to test them with in my TSO+VxLAN 
>codebase.
>>
>> Thanks,
>> Mark
>>
>>>
>>>
>>> 2016/10/11 22:34, Kavanagh, Mark B :
> Hi,  all
>
>
> This patch is very important for users want to improve the performance
> of the large packets.
>
>
> But you know, in data center, lots of networks using vxlan, so if it
> supports vxlan, then it will be very useful.
>
>
> Would you guys has a plan to support it? I would like to help test it,
> or work together on it?
 Hi Michael,

 Some work has already been done to enable TSO over VxLAN, but it was 
 blocked until
>>> recently, on account of gaps in the DPKD i40e PMD's support for tunnel 
>>> offload.

 We believe that those gaps have now been addressed; our current focus is 
 on ironing out
>the
>>> issues in TSO support over flat and VLAN networks - once that's done, we 
>>> plan to resume
>work
>>> on VxLAN support (and GRE, IPinIP).

 Hope this helps,
 Mark

>
> --
> Thanks,
> Michael
>>>
>>> --
>>> Thanks,
>>> Michael
>>
>

___
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev

Re: [ovs-dev] The Patch netdev-dpdk: add TSO support for vhost-user ports

[网易]

OK, Thanks for your info.

BTW, you mentioned the gap, is it the tunneling tso can't work correctly for 
multi-segments mbuf, and after some packets, guest is unable to xmit 
packets(dpdk driver failed to work any more)?

Thanks,
Michael

2016年10月14日，下午5:15，Kavanagh, Mark B  ：

>> 
>> 
>> Hi, Mark
>> 
>> 
>> OK, once it's ready， pls let me know and I'm glade to help to test it.
>> 
>> 
>> BTW， you mentioned the gap, is that the TSO and CSUM of tunnel offload?
>> I'm familiar with it, and I hope I could do something on it if possible.
> 
> Yes, that's what it was. 
> 
> Commits have recently been introduced to DPDK that introduce additional flags 
> for tunnel offloading, but I haven't had a chance to test them with in my 
> TSO+VxLAN codebase.
> 
> Thanks,
> Mark
> 
>> 
>> 
>> 2016/10/11 22:34, Kavanagh, Mark B :
 Hi,  all
 
 
 This patch is very important for users want to improve the performance
 of the large packets.
 
 
 But you know, in data center, lots of networks using vxlan, so if it
 supports vxlan, then it will be very useful.
 
 
 Would you guys has a plan to support it? I would like to help test it,
 or work together on it?
>>> Hi Michael,
>>> 
>>> Some work has already been done to enable TSO over VxLAN, but it was 
>>> blocked until
>> recently, on account of gaps in the DPKD i40e PMD's support for tunnel 
>> offload.
>>> 
>>> We believe that those gaps have now been addressed; our current focus is on 
>>> ironing out the
>> issues in TSO support over flat and VLAN networks - once that's done, we 
>> plan to resume work
>> on VxLAN support (and GRE, IPinIP).
>>> 
>>> Hope this helps,
>>> Mark
>>> 
 
 --
 Thanks,
 Michael
>> 
>> --
>> Thanks,
>> Michael
> 


___
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev

Re: [ovs-dev] The Patch netdev-dpdk: add TSO support for vhost-user ports

[Kavanagh, Mark B]

>
>Hi, Mark
>
>
>I have reviewed your RFC patch, I found one place could have issue, see
>below:
>
>
>+if (v > ETHER_MAX_LEN) {
>+b->mbuf.data_len =
>+(uint16_t)(b->mbuf.buf_len - b->mbuf.data_off);
>+} else {
>+b->mbuf.data_len = (uint16_t)v;
>+}
>+
>
>Here, what will happen if v > ETHER_MAX_LEN but v < (b->mbuf.buf_len -
>b->mbuf.data_off) ?
>
>The driver will be get wrong data lenth in my option.
>
>probably, below is right :
>if (v > (b->mbuf.buf_len - b->mbuf.data_off))
> b->mbuf.data_len = (uint16_t)(b->mbuf.buf_len - b->mbuf.data_off);

Yes, you’re right - I have the same code in my local repo :)

>
>
>Thanks,
>Michael
>
>在 2016/10/11 22:34, Kavanagh, Mark B 写道:
>>> Hi,  all
>>>
>>>
>>> This patch is very important for users want to improve the performance
>>> of the large packets.
>>>
>>>
>>> But you know, in data center, lots of networks using vxlan, so if it
>>> supports vxlan, then it will be very useful.
>>>
>>>
>>> Would you guys has a plan to support it? I would like to help test it,
>>> or work together on it?
>> Hi Michael,
>>
>> Some work has already been done to enable TSO over VxLAN, but it was blocked 
>> until
>recently, on account of gaps in the DPKD i40e PMD's support for tunnel offload.
>>
>> We believe that those gaps have now been addressed; our current focus is on 
>> ironing out the
>issues in TSO support over flat and VLAN networks - once that's done, we plan 
>to resume work
>on VxLAN support (and GRE, IPinIP).
>>
>> Hope this helps,
>> Mark
>>
>>>
>>> --
>>> Thanks,
>>> Michael
>>>
>
>--
>Thanks,
>Michael
>

___
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev

Re: [ovs-dev] The Patch netdev-dpdk: add TSO support for vhost-user ports

[Kavanagh, Mark B]

>
>
>Hi, Mark
>
>
>OK, once it's ready， pls let me know and I'm glade to help to test it.
>
>
>BTW， you mentioned the gap, is that the TSO and CSUM of tunnel offload?
>I'm familiar with it, and I hope I could do something on it if possible.

Yes, that's what it was. 

Commits have recently been introduced to DPDK that introduce additional flags 
for tunnel offloading, but I haven't had a chance to test them with in my 
TSO+VxLAN codebase.

Thanks,
Mark

>
>
>2016/10/11 22:34, Kavanagh, Mark B :
>>> Hi,  all
>>>
>>>
>>> This patch is very important for users want to improve the performance
>>> of the large packets.
>>>
>>>
>>> But you know, in data center, lots of networks using vxlan, so if it
>>> supports vxlan, then it will be very useful.
>>>
>>>
>>> Would you guys has a plan to support it? I would like to help test it,
>>> or work together on it?
>> Hi Michael,
>>
>> Some work has already been done to enable TSO over VxLAN, but it was blocked 
>> until
>recently, on account of gaps in the DPKD i40e PMD's support for tunnel offload.
>>
>> We believe that those gaps have now been addressed; our current focus is on 
>> ironing out the
>issues in TSO support over flat and VLAN networks - once that's done, we plan 
>to resume work
>on VxLAN support (and GRE, IPinIP).
>>
>> Hope this helps,
>> Mark
>>
>>>
>>> --
>>> Thanks,
>>> Michael
>>>
>
>--
>Thanks,
>Michael
>

___
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev

Re: [ovs-dev] The Patch netdev-dpdk: add TSO support for vhost-user ports

[Michael Qiu]

Hi, Mark


I have reviewed your RFC patch, I found one place could have issue, see 
below:



+if (v > ETHER_MAX_LEN) {
+b->mbuf.data_len =
+(uint16_t)(b->mbuf.buf_len - b->mbuf.data_off);
+} else {
+b->mbuf.data_len = (uint16_t)v;
+}
+

Here, what will happen if v > ETHER_MAX_LEN but v < (b->mbuf.buf_len - 
b->mbuf.data_off) ?


The driver will be get wrong data lenth in my option.

probably, below is right :
if (v > (b->mbuf.buf_len - b->mbuf.data_off))
b->mbuf.data_len = (uint16_t)(b->mbuf.buf_len - b->mbuf.data_off);


Thanks,
Michael

在 2016/10/11 22:34, Kavanagh, Mark B 写道:

Hi,  all


This patch is very important for users want to improve the performance
of the large packets.


But you know, in data center, lots of networks using vxlan, so if it
supports vxlan, then it will be very useful.


Would you guys has a plan to support it? I would like to help test it,
or work together on it?

Hi Michael,

Some work has already been done to enable TSO over VxLAN, but it was blocked 
until recently, on account of gaps in the DPKD i40e PMD's support for tunnel 
offload.

We believe that those gaps have now been addressed; our current focus is on 
ironing out the issues in TSO support over flat and VLAN networks - once that's 
done, we plan to resume work on VxLAN support (and GRE, IPinIP).

Hope this helps,
Mark
 


--
Thanks,
Michael



--
Thanks,
Michael


___
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev

Re: [ovs-dev] The Patch netdev-dpdk: add TSO support for vhost-user ports

[Michael Qiu]

Hi, Mark


OK, once it's ready， pls let me know and I'm glade to help to test it.


BTW， you mentioned the gap, is that the TSO and CSUM of tunnel offload? 
I'm familiar with it, and I hope I could do something on it if possible.



2016/10/11 22:34, Kavanagh, Mark B :

Hi,  all


This patch is very important for users want to improve the performance
of the large packets.


But you know, in data center, lots of networks using vxlan, so if it
supports vxlan, then it will be very useful.


Would you guys has a plan to support it? I would like to help test it,
or work together on it?

Hi Michael,

Some work has already been done to enable TSO over VxLAN, but it was blocked 
until recently, on account of gaps in the DPKD i40e PMD's support for tunnel 
offload.

We believe that those gaps have now been addressed; our current focus is on 
ironing out the issues in TSO support over flat and VLAN networks - once that's 
done, we plan to resume work on VxLAN support (and GRE, IPinIP).

Hope this helps,
Mark
 


--
Thanks,
Michael



--
Thanks,
Michael


___
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev

Re: [ovs-dev] The Patch netdev-dpdk: add TSO support for vhost-user ports

[Kavanagh, Mark B]

>
>Hi,  all
>
>
>This patch is very important for users want to improve the performance
>of the large packets.
>
>
>But you know, in data center, lots of networks using vxlan, so if it
>supports vxlan, then it will be very useful.
>
>
>Would you guys has a plan to support it? I would like to help test it,
>or work together on it?

Hi Michael,

Some work has already been done to enable TSO over VxLAN, but it was blocked 
until recently, on account of gaps in the DPKD i40e PMD's support for tunnel 
offload.

We believe that those gaps have now been addressed; our current focus is on 
ironing out the issues in TSO support over flat and VLAN networks - once that's 
done, we plan to resume work on VxLAN support (and GRE, IPinIP).

Hope this helps,
Mark

>
> 
>--
>Thanks,
>Michael
>

___
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev

[ovs-dev] The Patch netdev-dpdk: add TSO support for vhost-user ports

[Michael Qiu]

Hi,  all


This patch is very important for users want to improve the performance 
of the large packets.



But you know, in data center, lots of networks using vxlan, so if it 
supports vxlan, then it will be very useful.



Would you guys has a plan to support it? I would like to help test it, 
or work together on it?



--
Thanks,
Michael


___
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev