Re: [dpdk-dev] [RFC] Add GRO support in DPDK

["Wiles, Keith" <keith.wiles@intel.com>]

> On Jan 24, 2017, at 2:04 PM, Stephen Hemminger <stephen@networkplumber.org> wrote:
> 
> On Tue, 24 Jan 2017 20:09:07 +0000
> "Wiles, Keith" <keith.wiles@intel.com> wrote:
> 
>>> On Jan 24, 2017, at 12:45 PM, Ananyev, Konstantin <konstantin.ananyev@intel.com> wrote:
>>> 
>>> 
>>> 
>>>> -----Original Message-----
>>>> From: Wiles, Keith
>>>> Sent: Tuesday, January 24, 2017 2:49 PM
>>>> To: Ananyev, Konstantin <konstantin.ananyev@intel.com>
>>>> Cc: Stephen Hemminger <stephen@networkplumber.org>; Hu, Jiayu <jiayu.hu@intel.com>; dev@dpdk.org; Kinsella, Ray
>>>> <ray.kinsella@intel.com>; Gilmore, Walter E <walter.e.gilmore@intel.com>; Venkatesan, Venky <venky.venkatesan@intel.com>;
>>>> yuanhan.liu@linux.intel.com
>>>> Subject: Re: [dpdk-dev] [RFC] Add GRO support in DPDK
>>>> 
>>>> 
>>>>> On Jan 24, 2017, at 3:33 AM, Ananyev, Konstantin <konstantin.ananyev@intel.com> wrote:
>>>>> 
>>>>> 
>>>>> 
>>>>>> -----Original Message-----
>>>>>> From: Wiles, Keith
>>>>>> Sent: Tuesday, January 24, 2017 5:26 AM
>>>>>> To: Ananyev, Konstantin <konstantin.ananyev@intel.com>
>>>>>> Cc: Stephen Hemminger <stephen@networkplumber.org>; Hu, Jiayu <jiayu.hu@intel.com>; dev@dpdk.org; Kinsella, Ray
>>>>>> <ray.kinsella@intel.com>; Gilmore, Walter E <walter.e.gilmore@intel.com>; Venkatesan, Venky <venky.venkatesan@intel.com>;
>>>>>> yuanhan.liu@linux.intel.com
>>>>>> Subject: Re: [dpdk-dev] [RFC] Add GRO support in DPDK
>>>>>> 
>>>>>> 
>>>>>>> On Jan 23, 2017, at 6:43 PM, Ananyev, Konstantin <konstantin.ananyev@intel.com> wrote:
>>>>>>> 
>>>>>>> 
>>>>>>> 
>>>>>>>> -----Original Message-----
>>>>>>>> From: Wiles, Keith
>>>>>>>> Sent: Monday, January 23, 2017 9:53 PM
>>>>>>>> To: Stephen Hemminger <stephen@networkplumber.org>
>>>>>>>> Cc: Hu, Jiayu <jiayu.hu@intel.com>; dev@dpdk.org; Kinsella, Ray <ray.kinsella@intel.com>; Ananyev, Konstantin
>>>>>>>> <konstantin.ananyev@intel.com>; Gilmore, Walter E <walter.e.gilmore@intel.com>; Venkatesan, Venky  
>>>>>> <venky.venkatesan@intel.com>;  
>>>>>>>> yuanhan.liu@linux.intel.com
>>>>>>>> Subject: Re: [dpdk-dev] [RFC] Add GRO support in DPDK
>>>>>>>> 
>>>>>>>> 
>>>>>>>>> On Jan 23, 2017, at 10:15 AM, Stephen Hemminger <stephen@networkplumber.org> wrote:
>>>>>>>>> 
>>>>>>>>> On Mon, 23 Jan 2017 21:03:12 +0800
>>>>>>>>> Jiayu Hu <jiayu.hu@intel.com> wrote:
>>>>>>>>> 
>>>>>>>>>> With the support of hardware segmentation techniques in DPDK, the
>>>>>>>>>> networking stack overheads of send-side of applications, which directly
>>>>>>>>>> leverage DPDK, have been greatly reduced. But for receive-side, numbers of
>>>>>>>>>> segmented packets seriously burden the networking stack of applications.
>>>>>>>>>> Generic Receive Offload (GRO) is a widely used method to solve the
>>>>>>>>>> receive-side issue, which gains performance by reducing the amount of
>>>>>>>>>> packets processed by the networking stack. But currently, DPDK doesn't
>>>>>>>>>> support GRO. Therefore, we propose to add GRO support in DPDK, and this
>>>>>>>>>> RFC is used to explain the basic DPDK GRO design.
>>>>>>>>>> 
>>>>>>>>>> DPDK GRO is a SW-based packets assembly library, which provides GRO
>>>>>>>>>> abilities for numbers of protocols. In DPDK GRO, packets are merged
>>>>>>>>>> before returning to applications and after receiving from drivers.
>>>>>>>>>> 
>>>>>>>>>> In DPDK, GRO is a capability of NIC drivers. That support GRO or not and
>>>>>>>>>> what GRO types are supported are up to NIC drivers. Different drivers may
>>>>>>>>>> support different GRO types. By default, drivers enable all supported GRO
>>>>>>>>>> types. For applications, they can inquire the supported GRO types by
>>>>>>>>>> each driver, and can control what GRO types are applied. For example,
>>>>>>>>>> ixgbe supports TCP and UDP GRO, but the application just needs TCP GRO.
>>>>>>>>>> The application can disable ixgbe UDP GRO.
>>>>>>>>>> 
>>>>>>>>>> To support GRO, a driver should provide a way to tell applications what
>>>>>>>>>> GRO types are supported, and provides a GRO function, which is in charge
>>>>>>>>>> of assembling packets. Since different drivers may support different GRO
>>>>>>>>>> types, their GRO functions may be different. For applications, they don't
>>>>>>>>>> need extra operations to enable GRO. But if there are some GRO types that
>>>>>>>>>> are not needed, applications can use an API, like
>>>>>>>>>> rte_eth_gro_disable_protocols, to disable them. Besides, they can
>>>>>>>>>> re-enable the disabled ones.
>>>>>>>>>> 
>>>>>>>>>> The GRO function processes numbers of packets at a time. In each
>>>>>>>>>> invocation, what GRO types are applied depends on applications, and the
>>>>>>>>>> amount of packets to merge depends on the networking status and
>>>>>>>>>> applications. Specifically, applications determine the maximum number of
>>>>>>>>>> packets to be processed by the GRO function, but how many packets are
>>>>>>>>>> actually processed depends on if there are available packets to receive.
>>>>>>>>>> For example, the receive-side application asks the GRO function to
>>>>>>>>>> process 64 packets, but the sender only sends 40 packets. At this time,
>>>>>>>>>> the GRO function returns after processing 40 packets. To reassemble the
>>>>>>>>>> given packets, the GRO function performs an "assembly procedure" on each
>>>>>>>>>> packet. We use an example to demonstrate this procedure. Supposing the
>>>>>>>>>> GRO function is going to process packetX, it will do the following two
>>>>>>>>>> things:
>>>>>>>>>> 	a. Find a L4 assembly function according to the packet type of
>>>>>>>>>> 	packetX. A L4 assembly function is in charge of merging packets of a
>>>>>>>>>> 	specific type. For example, TCPv4 assembly function merges packets
>>>>>>>>>> 	whose L3 IPv4 and L4 is TCP. Each L4 assembly function has a packet
>>>>>>>>>> 	array, which keeps the packets that are unable to assemble.
>>>>>>>>>> 	Initially, the packet array is empty;
>>>>>>>>>> 	b. The L4 assembly function traverses own packet array to find a
>>>>>>>>>> 	mergeable packet (comparing Ethernet, IP and L4 header fields). If
>>>>>>>>>> 	finds, merges it and packetX via chaining them together; if doesn't,
>>>>>>>>>> 	allocates a new array element to store packetX and updates element
>>>>>>>>>> 	number of the array.
>>>>>>>>>> After performing the assembly procedure to all packets, the GRO function
>>>>>>>>>> combines the results of all packet arrays, and returns these packets to
>>>>>>>>>> applications.
>>>>>>>>>> 
>>>>>>>>>> There are lots of ways to implement the above design in DPDK. One of the
>>>>>>>>>> ways is:
>>>>>>>>>> 	a. Drivers tell applications what GRO types are supported via
>>>>>>>>>> 	dev->dev_ops->dev_infos_get;
>>>>>>>>>> 	b. When initialize, drivers register own GRO function as a RX
>>>>>>>>>> 	callback, which is invoked inside rte_eth_rx_burst. The name of the
>>>>>>>>>> 	GRO function should be like xxx_gro_receive (e.g. ixgbe_gro_receive).
>>>>>>>>>> 	Currently, the RX callback can only process the packets returned by
>>>>>>>>>> 	dev->rx_pkt_burst each time, and the maximum packet number
>>>>>>>>>> 	dev->rx_pkt_burst returns is determined by each driver, which can't
>>>>>>>>>> 	be interfered by applications. Therefore, to implement the above GRO
>>>>>>>>>> 	design, we have to modify current RX implementation to make driver
>>>>>>>>>> 	return packets as many as possible until the packet number meets the
>>>>>>>>>> 	demand of applications or there are not available packets to receive.
>>>>>>>>>> 	This modification is also proposed in patch:
>>>>>>>>>> 	http://dpdk.org/ml/archives/dev/2017-January/055887.html;
>>>>>>>>>> 	c. The GRO types to apply and the maximum number of packets to merge
>>>>>>>>>> 	are passed by resetting RX callback parameters. It can be achieved by
>>>>>>>>>> 	invoking rte_eth_rx_callback;
>>>>>>>>>> 	d. Simply, we can just store packet addresses into the packet array.
>>>>>>>>>> 	To check one element, we need to fetch the packet via its address.
>>>>>>>>>> 	However, this simple design is not efficient enough. Since whenever
>>>>>>>>>> 	checking one packet, one pointer dereference is generated. And a
>>>>>>>>>> 	pointer dereference always causes a cache line miss. A better way is
>>>>>>>>>> 	to store some rules in each array element. The rules must be the
>>>>>>>>>> 	prerequisites of merging two packets, like the sequence number of TCP
>>>>>>>>>> 	packets. We first compare the rules, then retrieve the packet if the
>>>>>>>>>> 	rules match. If storing the rules causes the packet array structure
>>>>>>>>>> 	is cache-unfriendly, we can store a fixed-length signature of the
>>>>>>>>>> 	rules instead. For example, the signature can be calculated by
>>>>>>>>>> 	performing XOR operation on IP addresses. Both design can avoid
>>>>>>>>>> 	unnecessary pointer dereferences.  
>>>>>>>>> 
>>>>>>>>> 
>>>>>>>>> Since DPDK does burst mode already, GRO is a lot less relevant.
>>>>>>>>> GRO in Linux was invented because there is no burst mode in the receive API.
>>>>>>>>> 
>>>>>>>>> If you look at VPP in FD.io you will see they already do aggregration and
>>>>>>>>> steering at the higher level in the stack.
>>>>>>>>> 
>>>>>>>>> The point of GRO is that it is generic, no driver changes are necessary.
>>>>>>>>> Your proposal would add a lot of overhead, and cause drivers to have to
>>>>>>>>> be aware of higher level flows.  
>>>>>>>> 
>>>>>>>> NACK
>>>>>>>> 
>>>>>>>> The design is not super clear to me here and we need to understand the impact to DPDK, performance and the  application. I would  
>>>> like  
>>>>>> to  
>>>>>>>> have a clean transparent design to the application and as little impact on performance as possible.
>>>>>>>> 
>>>>>>>> Let discuss this as I am not sure my previous concerns were addressed in this RFC.
>>>>>>>> 
>>>>>>> 
>>>>>>> I would agree that design looks overcomplicated and strange:
>>>>>>> If GRO can (and supposed to be) done fully in SW, why do we need to modify PMDs at all,
>>>>>>> why it can't be just a standalone DPDK library that user can use on his/her convenience?
>>>>>>> I'd suggest to start with some simple and most widespread case (TCP?) and try to implement
>>>>>>> a library for it first: something similar to what we have for ip reassembly.  
>>>>>> 
>>>>>> The reason this should not be a library the application calls is to allow for a transparent design for HW and SW support of this feature.  
>>>> Using  
>>>>>> the SW version the application should not need to understand (other then performance) that GRO is being done for this port.
>>>>>> 
>>>>> 
>>>>> Why is that?
>>>>> Let say we have ip reassembly library that is called explicitly by the application.
>>>>> I think for L4 grouping we can do the same.
>>>>> After all it is a pure SW feature, so to me it makes sense to allow application to decide
>>>>> when/where to call it.
>>>>> Again it would allow people to develop/use it without any modifications in current PMDs.  
>>>> 
>>>> I guess I did not make it clear, we need to support HW and this SW version transparently just as we handle other features in HW/SW under a
>>>> generic API for DPDK.  
>>> 
>>> Ok, I probably wasn't very clear too.
>>> What I meant:
>>> Let's try to implement GRO (in SW) as a standalone DPDK library,
>>> with clean & simple interface and see how fast and useful it would be.
>>> We can refer to it as step 1.
>>> When (if) we'll have step 1 in place, then we can start thinking
>>> about adding combined HW/SW solution for it (step 2).
>>> I think at that stage it would be much clearer:
>>> is there any point in it  at all,
>>> and if yes, how it should be done:
>>> -changes at rte_ethedev or on PMD layers or both
>>> - would changes at rte_ethdev API be needed and if yes what particular, etc.
>>> 
>>> From my perspective, without step 1 in place,  there is no much point in approaching step 2.  
>> 
>> Currently I believe they have a SW library version of the code, but I think we need to look at the design in that form. At this time the current design or code is not what I would expect needs to be done for the transparent version. To many interactions with the application and a separate Rx/Tx functions were being used (If I remember correctly)
>> 
>>> 
>>> BTW, any particular HW you have in mind?
>>> Currently, as I can see LRO (HW) is supported only by ixgbe and probably by viritual PMDs (virtio/vmxent3).
>>> Though even for ixgbe there are plenty of limitations: SRIOV mode should be off, HW CRC stropping should be off, etc.
>>> So my guess, right now step 1 is much more useful and feasible.
>>> 
>>>> 
>>>>> 
>>>>>> As I was told the Linux kernel hides this features and make it transparent.  
>>>>> 
>>>>> Yes, but DPDK does a lot things in a different way.
>>>>> So it doesn't look like a compelling reason for me :)  
>>>> 
>>>> Just looking at different options here and it is a compelling reason to me as it enforces the design can be transparent to the application.
>>>> Having the application in a NFV deciding on hw or sw or both is not a good place to put that logic IMO.  
>>> 
>>> Actually could you provide an example of linux NIC driver, that uses HW offloads (and which) to implement GRO?
>>> I presume some might use HW generated hashes, but apart from that, when HW performs actual packet grouping?
>>> From what I've seen Intel ones rely SW implementation for that.
>>> But I am not a linux/GRO expert, so feel free to correct me here.
>>> Konstantin   
>> 
>> Regards,
>> Keith
>> 
> 
> Linux uses a push (rather than DPDK pull) model for packet receiving.
> The Linux driver pushes packets into GRO by calling napi_gro_receive.
> 
> Since DPDK is pull model the API would be simpler.
> it could be as simple as:
>  nb = rte_eth_rx_burst(port, rx_pkts, N);
>  nb = rte_rx_gro(port, rx_pkts, gro_pkts, nb);
> 
> I agree with others, look at ip reassembly library as example.
> Also, GRO does not make sense for applications which already do the same vector flow
> processing like VPP which is one reason it should be optional.

I agree it should be option, but I worry about making it an example. I would like to see the GRO to be more transparent to the application and supported as a generic feature for DPDK. Maybe the application needs to request the support or it is a config option. The problem with config options is they are hard to test and testing becomes complexed.

Can we not figure out a way to add the feature inline instead of the application needing to call these APIs? It would be nice to have IP fragmentation also a optional feature to the rx/tx ethdev call. It would take it out of the example zone and move it into DPDK as a real feature. Today we expect the application to chain all of these little bits outside of DPDK into something useful, can we help fix that problem?

> 

Regards,
Keith