Home

Table of Contents

1. Motivation
2. mlxsw
   1. Features by Version
   2. Known Issues
3. Reporting Issues

Motivation

switchdev is an infrastructure in the Linux kernel which facilitates the offloading of the kernel's forwarding plane to capable ASICs.

switchdev allows users and developers to utilize current ASICs by using a standardized and well-known API exposed by the Linux kernel instead of relying on proprietary APIs implemented in binary user space blobs.

By using the Linux kernel to configure the hardware, users can use the same familiar tools to configure both their servers and switches. The sole difference would be the performance gained by offloading the kernel's forwarding plane to the switch's ASIC.

switch$ ip link add name br0 type bridge
switch$ ip link set dev br0 type bridge vlan_filtering 1
switch$ ip link set dev sw1p1 master br0
switch$ ip link set dev sw1p2 master br0
switch$ ip link set dev br0 up

hostA$ iperf –s –i1
hostB$ iperf -c hostA -i1 -P 8
------------------------------------------------------------
Client connecting to 192.168.1.1, TCP port 5001
TCP window size: 85.0 KByte (default)
------------------------------------------------------------
...
[SUM]  1.0- 2.0 sec  10.5 GBytes  90.6 Gbits/sec

For more information about switchdev please refer to the kernel's Switchdev documentation.

mlxsw

Mellanox Technologies is the first hardware vendor to use the switchdev API to offload the kernel's forwarding plane to a real ASIC. Mellanox's current switchdev-based solution is focused on the 100Gb/s Spectrum ASIC switches (SN2000 Series).

This is achieved by using an upstream driver in the Linux kernel. A user can simply buy a switch, install Linux on it like any other server and benefit from the underlying hardware.

Features by Version

Kernel Version
4.3	SwitchX-2 driver submission. Slow path only. Not in active development
4.4	Spectrum driver submission. VLAN-aware bridge offload
4.5	LAG and VLAN-unaware bridges
4.6	devlink infrastructure and port splitter
4.7	Quality of Service: DCB and shared buffers
4.8	IPv4 unicast router, port mirroring, extended ethtool statistics
4.9	Extended ethtool support, HW stats query via iproute
4.10	SwitchIB support, I2C support, CPU policer
4.11	tc-flower offload, enhanced L3 offload, packet sampling
4.12	tc-vlan offload, VRFs, ACL activity and stats dumping, OVS offload
4.13	Match on TCP flags, firmware flashing, trap action, port module info
4.14	IPv6 unicast router, match on IP TTL and TOS, tc-multichain offload, GRE tunnels
4.15	IPv4 multicast router, IPv4 non-equal-cost multi-path, multi-path hash policy, RED queueing discipline

Known Issues

The following is a list of known issues that will be resolved in future releases:

1. Order of operations. Network configuration should be done in a bottom-up fashion. Teardown should be done top-down.
2. Splitting front panel ports at the lower row to four fails (Fixed in 4.7-rc3).
3. ACL key size. It's not possible to specify rules that match on L2 fields and a complete IPv6 header.
4. ECMP size. 32 nexthops are currently supported (Added in 4.15).
5. L2 multicast (Fixed in 4.15):
   1. When multicast snooping is disabled:
      1. MDB entries are still in use.
      2. mcast_flood flag is ignored.
   2. When multicast snooping is enabled:
      1. Packets that hit an MDB entry will not be sent to ports that are makred as multicast router ports unless they are also part of the MDB entry.

Reporting Issues

To report issues in the Wiki please send an email to: mlxsw-wiki [at] mellanox [dot] com