net: ice: Perform accurate aRFS flow match

This patch fixes an issue seen in a large-scale deployment under heavy
incoming pkts where the aRFS flow wrongly matches a flow and reprograms the
NIC with wrong settings. That mis-steering causes RX-path latency spikes
and noisy neighbor effects when many connections collide on the same
hash (some of our production servers have 20-30K connections).

set_rps_cpu() calls ndo_rx_flow_steer() with flow_id that is calculated by
hashing the skb sized by the per rx-queue table size. This results in
multiple connections (even across different rx-queues) getting the same
hash value. The driver steer function modifies the wrong flow to use this
rx-queue, e.g.: Flow#1 is first added:
    Flow#1:  <ip1, port1, ip2, port2>, Hash 'h', q#10

Later when a new flow needs to be added:
	    Flow#2:  <ip3, port3, ip4, port4>, Hash 'h', q#20

The driver finds the hash 'h' from Flow#1 and updates it to use q#20. This
results in both flows getting un-optimized - packets for Flow#1 goes to
q#20, and then reprogrammed back to q#10 later and so on; and Flow #2
programming is never done as Flow#1 is matched first for all misses. Many
flows may wrongly share the same hash and reprogram rules of the original
flow each with their own q#.

Tested on two 144-core servers with 16K netperf sessions for 180s. Netperf
clients are pinned to cores 0-71 sequentially (so that wrong packets on q#s
72-143 can be measured). IRQs are set 1:1 for queues -> CPUs, enable XPS,
enable aRFS (global value is 144 * rps_flow_cnt).

Test notes about results from ice_rx_flow_steer():
---------------------------------------------------
1. "Skip:" counter increments here:
    if (fltr_info->q_index == rxq_idx ||
	arfs_entry->fltr_state != ICE_ARFS_ACTIVE)
	    goto out;
2. "Add:" counter increments here:
    ret = arfs_entry->fltr_info.fltr_id;
    INIT_HLIST_NODE(&arfs_entry->list_entry);
3. "Update:" counter increments here:
    /* update the queue to forward to on an already existing flow */

Runtime comparison: original code vs with the patch for different
rps_flow_cnt values.

+-------------------------------+--------------+--------------+
| rps_flow_cnt                  |      512     |    2048      |
+-------------------------------+--------------+--------------+
| Ratio of Pkts on Good:Bad q's | 214 vs 822K  | 1.1M vs 980K |
| Avoid wrong aRFS programming  | 0 vs 310K    | 0 vs 30K     |
| CPU User                      | 216 vs 183   | 216 vs 206   |
| CPU System                    | 1441 vs 1171 | 1447 vs 1320 |
| CPU Softirq                   | 1245 vs 920  | 1238 vs 961  |
| CPU Total                     | 29 vs 22.7   | 29 vs 24.9   |
| aRFS Update                   | 533K vs 59   | 521K vs 32   |
| aRFS Skip                     | 82M vs 77M   | 7.2M vs 4.5M |
+-------------------------------+--------------+--------------+

A separate TCP_STREAM and TCP_RR with 1,4,8,16,64,128,256,512 connections
showed no performance degradation.

Some points on the patch/aRFS behavior:
1. Enabling full tuple matching ensures flows are always correctly matched,
   even with smaller hash sizes.
2. 5-6% drop in CPU utilization as the packets arrive at the correct CPUs
   and fewer calls to driver for programming on misses.
3. Larger hash tables reduces mis-steering due to more unique flow hashes,
   but still has clashes. However, with larger per-device rps_flow_cnt, old
   flows take more time to expire and new aRFS flows cannot be added if h/w
   limits are reached (rps_may_expire_flow() succeeds when 10*rps_flow_cnt
   pkts have been processed by this cpu that are not part of the flow).

Fixes: 28bf26724fdb0 ("ice: Implement aRFS")
Signed-off-by: Krishna Kumar <krikku@gmail.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Tested-by: Rinitha S <sx.rinitha@intel.com> (A Contingent worker at Intel)
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>

0 files changed, 0 insertions, 0 deletions