netifd: rewrite packet steering script

The new script uses a different strategy compared to the previous one.
Instead of trying to split flows by hash and spread them to all CPUs,
use RPS to redirect packets to a single core only.
Try to spread NAPI thread and RPS target CPUs across available CPUs
and try to ensure that the NAPI thread is on a different CPU than the
RPS target. This significantly reduces cycles wasted on the scheduler.

Signed-off-by: Felix Fietkau <nbd@nbd.name>
This commit is contained in:
Felix Fietkau 2024-04-12 21:51:50 +02:00
parent c2f52e42b1
commit a205a5734e
4 changed files with 207 additions and 72 deletions

View File

@ -21,7 +21,7 @@ include $(INCLUDE_DIR)/cmake.mk
define Package/netifd
SECTION:=base
CATEGORY:=Base system
DEPENDS:=+libuci +libnl-tiny +libubus +ubus +ubusd +jshn +libubox +libudebug
DEPENDS:=+libuci +libnl-tiny +libubus +ubus +ubusd +jshn +libubox +libudebug +ucode +ucode-mod-fs
TITLE:=OpenWrt Network Interface Configuration Daemon
endef

View File

@ -14,5 +14,10 @@ service_triggers() {
}
reload_service() {
/usr/libexec/network/packet-steering.sh
packet_steering="$(uci get "network.@globals[0].packet_steering")"
if [ -e "/usr/libexec/platform/packet-steering.sh" ]; then
/usr/libexec/platform/packet-steering.sh "$packet_steering"
else
/usr/libexec/network/packet-steering.uc "$packet_steering"
fi
}

View File

@ -1,70 +0,0 @@
#!/bin/sh
NPROCS="$(grep -c "^processor.*:" /proc/cpuinfo)"
[ "$NPROCS" -gt 1 ] || exit
PROC_MASK="$(( (1 << $NPROCS) - 1 ))"
find_irq_cpu() {
local dev="$1"
local match="$(grep -m 1 "$dev\$" /proc/interrupts)"
local cpu=0
[ -n "$match" ] && {
set -- $match
shift
for cur in $(seq 1 $NPROCS); do
[ "$1" -gt 0 ] && {
cpu=$(($cur - 1))
break
}
shift
done
}
echo "$cpu"
}
set_hex_val() {
local file="$1"
local val="$2"
val="$(printf %x "$val")"
[ -n "$DEBUG" ] && echo "$file = $val"
echo "$val" > "$file"
}
packet_steering="$(uci get "network.@globals[0].packet_steering")"
[ "$packet_steering" != 1 ] && exit 0
exec 512>/var/lock/smp_tune.lock
flock 512 || exit 1
[ -e "/usr/libexec/platform/packet-steering.sh" ] && {
/usr/libexec/platform/packet-steering.sh
exit 0
}
for dev in /sys/class/net/*; do
[ -d "$dev" ] || continue
# ignore virtual interfaces
[ -n "$(ls "${dev}/" | grep '^lower_')" ] && continue
[ -d "${dev}/device" ] || continue
device="$(readlink "${dev}/device")"
device="$(basename "$device")"
irq_cpu="$(find_irq_cpu "$device")"
irq_cpu_mask="$((1 << $irq_cpu))"
for q in ${dev}/queues/tx-*; do
set_hex_val "$q/xps_cpus" "$PROC_MASK"
done
# ignore dsa slave ports for RPS
subsys="$(readlink "${dev}/device/subsystem")"
subsys="$(basename "$subsys")"
[ "$subsys" = "mdio_bus" ] && continue
for q in ${dev}/queues/rx-*; do
set_hex_val "$q/rps_cpus" "$PROC_MASK"
done
done

View File

@ -0,0 +1,200 @@
#!/usr/bin/env ucode
'use strict';
import { glob, basename, dirname, readlink, readfile, realpath, writefile, error, open } from "fs";
let napi_weight = 1.0;
let cpu_thread_weight = 0.75;
let rx_weight = 0.75;
let eth_bias = 2.0;
let debug = 0, do_nothing = 0;
let disable;
let cpus;
for (let arg in ARGV) {
switch (arg) {
case "-d":
debug++;
break;
case "-n":
do_nothing++;
break;
case '0':
disable = true;
break;
}
}
function task_name(pid)
{
let stat = open(`/proc/${pid}/status`, "r");
let line = stat.read("line");
stat.close();
return trim(split(line, "\t", 2)[1]);
}
function set_task_cpu(pid, cpu) {
if (disable)
cpu = join(",", map(cpus, (cpu) => cpu.id));
if (debug || do_nothing)
warn(`taskset -p -c ${cpu} ${task_name(pid)}\n`);
if (!do_nothing)
system(`taskset -p -c ${cpu} ${pid}`);
}
function set_netdev_cpu(dev, cpu) {
let queues = glob(`/sys/class/net/${dev}/queues/rx-*/rps_cpus`);
let val = sprintf("%x", (1 << int(cpu)));
if (disable)
val = 0;
for (let queue in queues) {
if (debug || do_nothing)
warn(`echo ${val} > ${queue}\n`);
if (!do_nothing)
writefile(queue, `${val}`);
}
}
function task_device_match(name, device)
{
let napi_match = match(name, /napi\/([^-+])-\d+/);
if (!napi_match)
napi_match = match(name, /mt76-tx (phy\d+)/);
if (napi_match &&
(index(device.phy, napi_match[1]) >= 0 ||
index(device.netdev, napi_match[1]) >= 0))
return true;
if (device.driver == "mtk_soc_eth" && match(name, /napi\/mtk_eth-/))
return true;
return false;
}
cpus = map(glob("/sys/bus/cpu/devices/*"), (path) => {
return {
id: int(match(path, /.*cpu(\d+)/)[1]),
core: int(trim(readfile(`${path}/topology/core_id`))),
load: 0.0,
};
});
cpus = slice(cpus, 0, 64);
if (length(cpus) < 2)
exit(0);
function cpu_add_weight(cpu_id, weight)
{
let cpu = cpus[cpu_id];
cpu.load += weight;
for (let sibling in cpus) {
if (sibling == cpu || sibling.core != cpu.core)
continue;
sibling.load += weight * cpu_thread_weight;
}
}
function get_next_cpu(weight, prev_cpu)
{
if (disable)
return 0;
let sort_cpus = sort(slice(cpus), (a, b) => a.load - b.load);
let idx = 0;
if (prev_cpu != null && sort_cpus[idx].id == prev_cpu)
idx++;
let cpu = sort_cpus[idx].id;
cpu_add_weight(cpu, weight);
return cpu;
}
let phys_devs = {};
let netdev_phys = {};
let netdevs = map(glob("/sys/class/net/*"), (dev) => basename(dev));
for (let dev in netdevs) {
let pdev_path = realpath(`/sys/class/net/${dev}/device`);
if (!pdev_path)
continue;
if (length(glob(`/sys/class/net/${dev}/lower_*`)) > 0)
continue;
let pdev = phys_devs[pdev_path];
if (!pdev) {
pdev = phys_devs[pdev_path] = {
path: pdev_path,
driver: basename(readlink(`${pdev_path}/driver`)),
netdev: [],
phy: [],
tasks: [],
};
}
let phyidx = trim(readfile(`/sys/class/net/${dev}/phy80211/index`));
if (phyidx != null) {
let phy = `phy${phyidx}`;
if (index(pdev.phy, phy) < 0)
push(pdev.phy, phy);
}
push(pdev.netdev, dev);
netdev_phys[dev] = pdev;
}
for (let path in glob("/proc/*/exe")) {
readlink(path);
if (error() != "No such file or directory")
continue;
let pid = basename(dirname(path));
let name = task_name(pid);
for (let devname in phys_devs) {
let dev = phys_devs[devname];
if (!task_device_match(name, dev))
continue;
push(dev.tasks, pid);
break;
}
}
function assign_dev_cpu(dev) {
if (length(dev.tasks) > 0) {
let cpu = dev.napi_cpu = get_next_cpu(napi_weight);
for (let task in dev.tasks)
set_task_cpu(task, cpu);
}
if (length(dev.netdev) > 0) {
let cpu = dev.rx_cpu = get_next_cpu(rx_weight, dev.napi_cpu);
for (let netdev in dev.netdev)
set_netdev_cpu(netdev, cpu);
}
}
// Assign ethernet devices first
for (let devname in phys_devs) {
let dev = phys_devs[devname];
if (!length(dev.phy))
assign_dev_cpu(dev);
}
// Add bias to avoid assigning other tasks to CPUs with ethernet NAPI
for (let devname in phys_devs) {
let dev = phys_devs[devname];
if (!length(dev.tasks) || dev.napi_cpu == null)
continue;
cpu_add_weight(dev.napi_cpu, eth_bias);
}
// Assign WLAN devices
for (let devname in phys_devs) {
let dev = phys_devs[devname];
if (length(dev.phy) > 0)
assign_dev_cpu(dev);
}
if (debug > 1)
warn(sprintf("devices: %.J\ncpus: %.J\n", phys_devs, cpus));