Day 1 / Track 1 / Talk 4 Workshop: Hardware Offload Workshop Chair: Roopa Prabhu and Or Gerlitz Report by: Anjali Singhai
During this session there were discussions relating to hardware offload starting with Switch ASIC offload updates from Roopa Prabhu, new mechanism to debug hardware datapath from Ido Schimmel, Doorbell Overflow from Ariel Elior, Qos offload for NIC eSwitch model from Pieter Jansen and Simno Horman and Scalable NIC hardware offloads from Or Gerlitz.
The discussion started with the general switch ASIC offload and the many devlink updates, such as health Monitor, which monitors device health, and is used to pass information from the device to upper layer. There was a discussion about the need to have more hardware counter visibility for upper layers in the stack. Right now the hardware has lots of stat counters, programmable ones but they are not tied well into the different layers in the stack.
The discussion then shifted to packet drop visibility in the Control plane which is very important. The proposed solutions are: 1. The addition of a mechanism to allow ASICs to report dropped packets to user space. 2. Metadata can be attached to the packet. 3. Drop reason, ingress port, egress port, timestamp etc 4. Drop reasons should be standardized and correspond to kernel drops ( ex: Ingress VLAN filter) 5. Mechanism should allow the user to filter noisy drops, sample and truncate. 6. Filtering could be based on stages in the pipeline. 7. Devlink packet trap set DEV (all | group…) enable/disable 8. Show status and supported metadata 9. Monitor dropped packets 10. An eBPF filter could be attached to the netlink socket
Discussion ensued with Jiri mentioning that iptables enables tracing and looks like this infrastructure is missing in route susbsystem, tc subsystem, and then map that to HW Tom then added that if all that drop packets are received, how will this scale? how do you weed one particular drop?
The discussion continued with talks about policers being configured between ASIC and CPU, to limit the number of packets. The point made was not eliminate stats but augment it more.
The next talk was about Doorbell overflow recovery. The topic of discussion was the discovery and recovery for RDMA queues. Possible solutions were fast dequeuing, CPU stall and drop message detection and recovery procedures.
This talk was followed with Qos Ingress Rate limiting and OVS offload with TC. The focus was on ingress rate limiting and policing. The rate limited was done with TC offload by adding a matchall type cls with police action and introducing reserved priorities. OVS should install Tc filters with priority offset, reserve higher priority for rate limiting. A possible issue with ovs-tc offload is when going from software to hardware, tc police is in software and filters are offloaded, this could break semanting. Possible solutions include reverting to original semantics of policing with offload isn't supported and ovs forcing tc filters in software only.
Rony raised the question of why were priorities chosen vs chains. The answer was that recirculation is a good use case for chains.
This was followed by a small test demo.
Finally the last talk was about Scalable NIC HW offload. The talk begun with discussing the large amount of scaling hardware offloads. 1. Scale without using SRIOV 2. Multiple dynamic instances deployment at faster speed than VFs 3. NIC HW has very well defined vport based virtualization mode 4. One PCI device split into multiple smaller sub devices 5. Each sub device comes with own devices, vport, namespace resource 6. Leverage mature switchdev mode and OVS eco-system 7. Applicable for SmartNIC use case. 8. Using rich vendor agnostic devlink iproute2 tool.
The question that the presentors raised was how to achieve an Mdev software model view. A couple of points provided were: 1. Mlx5 mdev devices 2. Adding control plane knob to add /query remove mdev devices 3. Mentioned vDPA from Intel 4. Create 3 devices, netdev, RDMA device and representor netdev. 5. In HW mdev is attached to a vport 6. Map it to a container…cannot be mapped to a VM since single instance of driver.
The talk was concluded with reasons it's been implemented that way, as the devlink tool and bus model fits requirements such as providing vendor agnostic solution and multi-port subdevice creation.
Site: https://www.netdevconf.info/0x13/session.html?workshop-hardware-offload