On 2025-05-19 11:02 p.m., Krishna Kumar wrote:
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 has
(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
add empty line.
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 */
- **rps_flow_cnt=512**
- Ratio of packets on good vs bad queues: 214 vs 822,392
- Avoid updating wrong aRFS filter: 0 vs 310,826
- CPU: User: 216 vs 183, System: 1441 vs 1171, Softirq: 1245 vs 920
Total: 29.02 22.74
- aRFS Add: 6,078,551 vs 6,126,286 Update: 533,973 vs 59
Skip: 82,219,629 vs 77,088,191, Del: 6,078,409 vs 6,126,139
- **rps_flow_cnt=1024**
- Ratio of packets on good vs bad queues: 854 vs 1,003,176
- Avoid updating wrong aRFS filter: 0 vs 50,363
- CPU: User: 220 vs 206, System: 1460 vs 1322 Softirq: 1216 vs 1027
Total: 28.96 vs 25.55
- aRFS Add: 7,000,757 vs 7,499,586 Update: 504,371 vs 33
Skip: 27,455,269 vs 21,752,043, Del: 7,000,610 vs 7,499,438
- **rps_flow_cnt=2048**
- Ratio of packets on good vs bad queues: 1,173,756 vs 981,892
- Avoid updating wrong aRFS filter: 0 vs 30,145
- CPU: User: 216 vs 206, System: 1447 vs 1320, Softirq: 1238 vs 961
Total: 29.01 vs 24.87
- aRFS Add: 7,226,598 vs 6,960,991, Update: 521,264 vs 32
Skip: 7,236,716 vs 4,584,043, Del: 722,6430 vs 696,0798
Are these numbers with the patch applied? Can we get a w/o and with patch?
A table might be better to visualize, also may be drop the
"rps_flow_cnt=1024*" case. I think it is enough to show min and max ones.
Also, please add instructions on how to get these values, so that
validation team may be able to replicate.
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).
Signed-off-by: Krishna Kumar <[email protected]>
---
drivers/net/ethernet/intel/ice/ice_arfs.c | 45 +++++++++++++++++++++++
1 file changed, 45 insertions(+)
diff --git a/drivers/net/ethernet/intel/ice/ice_arfs.c
b/drivers/net/ethernet/intel/ice/ice_arfs.c
index 2bc5c7f59844..b36bd189bd64 100644
--- a/drivers/net/ethernet/intel/ice/ice_arfs.c
+++ b/drivers/net/ethernet/intel/ice/ice_arfs.c
@@ -377,6 +377,47 @@ ice_arfs_is_perfect_flow_set(struct ice_hw *hw, __be16
l3_proto, u8 l4_proto)
return false;
}
+/**
+ * ice_arfs_cmp - Check if aRFS filter matches this flow.
+ * @fltr_info: filter info of the saved ARFS entry.
+ * @fk: flow dissector keys.
+ * n_proto: One of htons(IPv4) or htons(IPv6).
+ * ip_proto: One of IPPROTO_TCP or IPPROTO_UDP.
+ *
+ * Since this function assumes limited values for n_proto and ip_proto, it
+ * is meant to be called only from ice_rx_flow_steer().
+ */
+static bool
+ice_arfs_cmp(const struct ice_fdir_fltr *fltr_info, const struct flow_keys *fk,
+ __be16 n_proto, u8 ip_proto)
+{
+ /*
+ * Determine if the filter is for IPv4 or IPv6 based on flow_type,
+ * which is one of ICE_FLTR_PTYPE_NONF_IPV{4,6}_{TCP,UDP}.
+ */
+ bool is_v4 = fltr_info->flow_type == ICE_FLTR_PTYPE_NONF_IPV4_TCP ||
+ fltr_info->flow_type == ICE_FLTR_PTYPE_NONF_IPV4_UDP;
+
+ /* Following checks are arranged in the quickest and most discriminative
+ * fields first for early failure.
+ */
+ if (is_v4)
+ return n_proto == htons(ETH_P_IP) &&
+ fltr_info->ip.v4.src_port == fk->ports.src &&
+ fltr_info->ip.v4.dst_port == fk->ports.dst &&
+ fltr_info->ip.v4.src_ip == fk->addrs.v4addrs.src &&
+ fltr_info->ip.v4.dst_ip == fk->addrs.v4addrs.dst &&
+ fltr_info->ip.v4.proto == ip_proto;
+
+ return fltr_info->ip.v6.src_port == fk->ports.src &&
+ fltr_info->ip.v6.dst_port == fk->ports.dst &&
+ fltr_info->ip.v6.proto == ip_proto &&
+ !memcmp(&fltr_info->ip.v6.src_ip, &fk->addrs.v6addrs.src,
+ sizeof(struct in6_addr)) &&
+ !memcmp(&fltr_info->ip.v6.dst_ip, &fk->addrs.v6addrs.dst,
+ sizeof(struct in6_addr));
+}
+
/**
* ice_rx_flow_steer - steer the Rx flow to where application is being run
* @netdev: ptr to the netdev being adjusted
@@ -448,6 +489,10 @@ ice_rx_flow_steer(struct net_device *netdev, const struct
sk_buff *skb,
continue;
fltr_info = &arfs_entry->fltr_info;
+
+ if (!ice_arfs_cmp(fltr_info, &fk, n_proto, ip_proto))
+ continue;
+
ret = fltr_info->fltr_id;
if (fltr_info->q_index == rxq_idx ||
This seems similar to a patch I tried up-streaming before:
https://lore.kernel.org/netdev/[email protected]/
not sure why I did not pursue further. If that is correct, then
obviously I have no objection to this patch.
The exact flow match will reduce (but not completely eliminate) the
chance that a packet may land on wrong queue (since there is always a
chance of hash collisions in aRFS).
Thank you.
