diff options
| -rw-r--r-- | Documentation/networking/dpaa.txt | 68 |
1 files changed, 67 insertions, 1 deletions
diff --git a/Documentation/networking/dpaa.txt b/Documentation/networking/dpaa.txt index 76e016d4d344..f88194f71c54 100644 --- a/Documentation/networking/dpaa.txt +++ b/Documentation/networking/dpaa.txt @@ -13,6 +13,7 @@ Contents - Configuring DPAA Ethernet in your kernel - DPAA Ethernet Frame Processing - DPAA Ethernet Features + - DPAA IRQ Affinity and Receive Side Scaling - Debugging DPAA Ethernet Overview @@ -147,7 +148,10 @@ gradually. The driver has Rx and Tx checksum offloading for UDP and TCP. Currently the Rx checksum offload feature is enabled by default and cannot be controlled through -ethtool. +ethtool. Also, rx-flow-hash and rx-hashing was added. The addition of RSS +provides a big performance boost for the forwarding scenarios, allowing +different traffic flows received by one interface to be processed by different +CPUs in parallel. The driver has support for multiple prioritized Tx traffic classes. Priorities range from 0 (lowest) to 3 (highest). These are mapped to HW workqueues with @@ -166,6 +170,68 @@ classes as follows: tc qdisc add dev <int> root handle 1: \ mqprio num_tc 4 map 0 0 0 0 1 1 1 1 2 2 2 2 3 3 3 3 hw 1 +DPAA IRQ Affinity and Receive Side Scaling +========================================== + +Traffic coming on the DPAA Rx queues or on the DPAA Tx confirmation +queues is seen by the CPU as ingress traffic on a certain portal. +The DPAA QMan portal interrupts are affined each to a certain CPU. +The same portal interrupt services all the QMan portal consumers. + +By default the DPAA Ethernet driver enables RSS, making use of the +DPAA FMan Parser and Keygen blocks to distribute traffic on 128 +hardware frame queues using a hash on IP v4/v6 source and destination +and L4 source and destination ports, in present in the received frame. +When RSS is disabled, all traffic received by a certain interface is +received on the default Rx frame queue. The default DPAA Rx frame +queues are configured to put the received traffic into a pool channel +that allows any available CPU portal to dequeue the ingress traffic. +The default frame queues have the HOLDACTIVE option set, ensuring that +traffic bursts from a certain queue are serviced by the same CPU. +This ensures a very low rate of frame reordering. A drawback of this +is that only one CPU at a time can service the traffic received by a +certain interface when RSS is not enabled. + +To implement RSS, the DPAA Ethernet driver allocates an extra set of +128 Rx frame queues that are configured to dedicated channels, in a +round-robin manner. The mapping of the frame queues to CPUs is now +hardcoded, there is no indirection table to move traffic for a certain +FQ (hash result) to another CPU. The ingress traffic arriving on one +of these frame queues will arrive at the same portal and will always +be processed by the same CPU. This ensures intra-flow order preservation +and workload distribution for multiple traffic flows. + +RSS can be turned off for a certain interface using ethtool, i.e. + + # ethtool -N fm1-mac9 rx-flow-hash tcp4 "" + +To turn it back on, one needs to set rx-flow-hash for tcp4/6 or udp4/6: + + # ethtool -N fm1-mac9 rx-flow-hash udp4 sfdn + +There is no independent control for individual protocols, any command +run for one of tcp4|udp4|ah4|esp4|sctp4|tcp6|udp6|ah6|esp6|sctp6 is +going to control the rx-flow-hashing for all protocols on that interface. + +Besides using the FMan Keygen computed hash for spreading traffic on the +128 Rx FQs, the DPAA Ethernet driver also sets the skb hash value when +the NETIF_F_RXHASH feature is on (active by default). This can be turned +on or off through ethtool, i.e.: + + # ethtool -K fm1-mac9 rx-hashing off + # ethtool -k fm1-mac9 | grep hash + receive-hashing: off + # ethtool -K fm1-mac9 rx-hashing on + Actual changes: + receive-hashing: on + # ethtool -k fm1-mac9 | grep hash + receive-hashing: on + +Please note that Rx hashing depends upon the rx-flow-hashing being on +for that interface - turning off rx-flow-hashing will also disable the +rx-hashing (without ethtool reporting it as off as that depends on the +NETIF_F_RXHASH feature flag). + Debugging ========= |