[dpdk-dev,v1] net/i40e: set no drop for traffic class

  Hi Jingjing,

Yes PRTDCB_TC2PFC is used to control pfc for each TC however we have noticed other advantages of using the register.
By setting the register explicitly by doing the "I40E_WRITE_REG(hw, 0x1c0980, 0xff);" it allows for packets to be temporarily stored on the NICs RX SRAM until there is space for them on SW descriptor ring versus dropping them when the SW ring becomes full. This also allows for larger burst handling. It also means SW doesn't have to be as quick to empty the DRAM based descriptor rings, allowing more processing on cores. 

I have tested using the ETH_DCB_PFC_SUPPORT flag in rte_eth_conf.dcb_capability_en and rte_eth_dcb_rx_conf.nb_tcs. 
This results in the NIC's RX SRAM not being used and if there is no space on SW descriptor ring for packet it is dropped. 
The advantages of using the PRTDCB_TC2PFC explicitly is that there will be no packet loss and descriptor rings do not need to be modified (can be left at 128 for rx and 512 for tx as default settings for apps).  Enabling via this register allows Burst handling to be within the NIC Rx Buffer and SW rings combined. 
At the moment for tests the rx and tx descriptor rings have to be increased to 2048 to eliminate packet loss.

Ideally it would be an optional setting as using it may increase the max latency.


Regards,
Rory/Nemanja

-----Original Message-----
From: Wu, Jingjing 
Sent: Monday, December 5, 2016 8:45 AM
To: Sexton, Rory <rory.sexton@intel.com>
Cc: dev@dpdk.org; Marjanovic, Nemanja <nemanja.marjanovic@intel.com>; Mcnamara, John <john.mcnamara@intel.com>
Subject: RE: [PATCH v1] net/i40e: set no drop for traffic class



-----Original Message-----
From: Sexton, Rory
Sent: Sunday, December 4, 2016 9:55 PM
To: Wu, Jingjing <jingjing.wu@intel.com>
Cc: dev@dpdk.org; Marjanovic, Nemanja <nemanja.marjanovic@intel.com>; Mcnamara, John <john.mcnamara@intel.com>; Sexton, Rory <rory.sexton@intel.com>
Subject: [PATCH v1] net/i40e: set no drop for traffic class

From: John McNamara <john.mcnamara@intel.com>

The default traffic class in i40e is set to drop versus on ixgbe it isset to no drop. This means when packets build up in the RX SRAM on the NIC, they are dropped, and they do this when the SW descriptor rings fill up.

This patch changes this behaviour and our testing shows there are no drops as a result.

Signed-off-by: Rory Sexton <rory.sexton@intel.com>
Signed-off-by: Nemanja Marjanovic <nemanja.marjanovic@intel.com>
---
 drivers/net/i40e/i40e_ethdev.c |  1 +
 drivers/net/i40e/i40e_rxtx.c   | 12 ++++++++++++
 drivers/net/i40e/i40e_rxtx.h   |  1 +
 lib/librte_ether/rte_ethdev.h  | 24 ++++++++++++++++++++++++
 4 files changed, 38 insertions(+)


Why do we need such a new API?


Thanks
Jingjing

--------------------------------------------------------------
Intel Research and Development Ireland Limited
Registered in Ireland
Registered Office: Collinstown Industrial Park, Leixlip, County Kildare
Registered Number: 308263


This e-mail and any attachments may contain confidential material for the sole
use of the intended recipient(s). Any review or distribution by others is
strictly prohibited. If you are not the intended recipient, please contact the
sender and delete all copies.
  

> -----Original Message-----
> From: Sexton, Rory
> Sent: Friday, December 9, 2016 10:03 PM
> To: Wu, Jingjing <jingjing.wu@intel.com>
> Cc: dev@dpdk.org; Marjanovic, Nemanja <nemanja.marjanovic@intel.com>;
> Mcnamara, John <john.mcnamara@intel.com>
> Subject: RE: [PATCH v1] net/i40e: set no drop for traffic class
> 
> Hi Jingjing,
> 
> Yes PRTDCB_TC2PFC is used to control pfc for each TC however we have
> noticed other advantages of using the register.
> By setting the register explicitly by doing the "I40E_WRITE_REG(hw, 0x1c0980,
> 0xff);" it allows for packets to be temporarily stored on the NICs RX SRAM
> until there is space for them on SW descriptor ring versus dropping them
> when the SW ring becomes full. This also allows for larger burst handling. It
> also means SW doesn't have to be as quick to empty the DRAM based
> descriptor rings, allowing more processing on cores.
> 
> I have tested using the ETH_DCB_PFC_SUPPORT flag in
> rte_eth_conf.dcb_capability_en and rte_eth_dcb_rx_conf.nb_tcs.
> This results in the NIC's RX SRAM not being used and if there is no space on
> SW descriptor ring for packet it is dropped.

Besides ETH_DCB_PFC_SUPPORT, ETH_MQ_RX_DCB_FLAG is also required in
dev_conf.rxmode.mq_mode. After doing that, you will find register PRTDCB_TC2PFC
is also changed.
 
If you don't want to enable DCB, why not just implement that function "i40e_priority_flow_ctrl_set"?
You can change the register in this function without define a new API.


> The advantages of using the PRTDCB_TC2PFC explicitly is that there will be no
> packet loss and descriptor rings do not need to be modified (can be left at
> 128 for rx and 512 for tx as default settings for apps).  Enabling via this
> register allows Burst handling to be within the NIC Rx Buffer and SW rings
> combined.
> At the moment for tests the rx and tx descriptor rings have to be increased
> to 2048 to eliminate packet loss.
> 
> Ideally it would be an optional setting as using it may increase the max
> latency.
> 


Thanks for the clarification.

Thanks
Jingjing