从 /proc/softirqs 到网卡多队列：Linux 软中断失衡与 CPU 亲和性调优实践

$ mpstat -P ALL 1 1
Linux 5.15.0-89-generic (node-01) 	03/30/2026 	_x86_64_	(32 CPU)

14:32:01 CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
14:32:02 all    3.12    0.00    4.50    0.00    0.15    3.20    0.00    0.00   89.03
14:32:02   0    0.00    0.00    2.00    0.00    1.00   96.00    0.00    0.00    1.00
14:32:02   1    4.04    0.00    5.05    0.00    0.00    0.00    0.00    0.00   90.91
14:32:02   2    3.00    0.00    4.00    0.00    0.00    0.00    0.00    0.00   93.00
...

  42.15%  [kernel]          [k] napi_gro_receive
  18.30%  [kernel]          [k] __netif_receive_skb_core
  12.45%  [kernel]          [k] ip_rcv
   8.12%  [ixgbe]           [k] ixgbe_poll

$ sudo ethtool -l eth0
Channel parameters for eth0:
Pre-set maximums:
RX:		0
TX:		0
Other:		0
Combined:	8    # <-- 硬件最多支持 8 组组合队列
Current hardware settings:
RX:		0
TX:		0
Other:		0
Combined:	1    # <-- 坑点：当前居然只开了 1 个队列！

$ sudo ethtool -L eth0 combined 8

$ sudo ethtool -x eth0
RX flow hash indirection table for eth0 with 8 RX ring(s):
    0:      0     1     2     3     4     5     6     7
    8:      0     1     2     3     4     5     6     7
...

$ sudo systemctl stop irqbalance
$ sudo systemctl disable irqbalance

$ grep eth0 /proc/interrupts
 128:   14520391          0          0   PCI-MSI 524288-edge   eth0-TxRx-0
 129:          0   14389201          0   PCI-MSI 524289-edge   eth0-TxRx-1
 130:          0          0   14412093   PCI-MSI 524290-edge   eth0-TxRx-2
...

#!/usr/bin/env bash
IFACE="eth0"
# 获取该网卡所有的中断号
IRQS=$(grep -E "${IFACE}" /proc/interrupts | awk -F ':' '{print $1}' | tr -d ' ')

# 避开 CPU 0（留给系统进程），从 CPU 1 开始绑定
CPU=1

for IRQ in ${IRQS}; do
    # 计算十六进制掩码 (1 << CPU)
    MASK=$(printf "%x" $((1 << CPU)))
    echo "Binding IRQ ${IRQ} to CPU ${CPU} (mask: 0x${MASK})"
    echo "${MASK}" > /proc/irq/${IRQ}/smp_affinity
    CPU=$((CPU + 1))
done

$ cat /sys/class/net/eth0/queues/rx-0/rps_cpus
00000000

$ echo "f" | sudo tee /sys/class/net/eth0/queues/rx-0/rps_cpus

# 全局最大流表数
sudo sysctl -w net.core.rps_sock_flow_entries=32768
# 单队列流表数
echo 4096 | sudo tee /sys/class/net/eth0/queues/rx-0/rps_flow_cnt

$ mpstat -P ALL 1 1
14:45:10 CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
14:45:11 all    4.50    0.00    5.10    0.00    0.20   12.50    0.00    0.00   77.70
14:45:11   0    1.00    0.00    1.00    0.00    0.00    0.00    0.00    0.00   98.00
14:45:11   1    2.00    0.00    4.00    0.00    0.20   12.10    0.00    0.00   81.70
14:45:11   2    2.00    0.00    3.90    0.00    0.20   12.30    0.00    0.00   81.60
14:45:11   3    1.90    0.00    4.10    0.00    0.20   12.40    0.00    0.00   81.40
...

$ sudo perf record -e irq:softirq_entry -a sleep 2
[ perf record: Woken up 8 times to write data ]
[ perf record: Captured and wrote 1.842 MB perf.data (14210 samples) ]

$ sudo perf report --sort cpu --stdio
# Overhead  CPU
# ........  ...
    12.65%  001
    12.58%  002
    12.54%  003
    12.49%  004
    12.48%  005
    12.46%  006
    12.42%  007
    12.38%  008