netifd: improve packet steering on ipq40xx (and possibly others)

Some platforms a single ethernet device for all ports with multiple rx rings
and NAPI threading enabled. In this case, the steering script was limiting
performance by keeping all NAPI threads assigned to the same CPU.
Fix this by assigning each rx queue and the corresponding NAPI task separately.
Additionally, if the number of rx queues is at least as big as the number of
CPUs, skip weight based assignment and distribute the load across all CPUs
directly.

Fixes: https://github.com/openwrt/openwrt/issues/17611
Signed-off-by: Felix Fietkau <nbd@nbd.name>

1 files changed, 47 insertions, 4 deletions

diff --git a/package/network/config/netifd/files/usr/libexec/network/packet-steering.uc b/package/network/config/netifd/files/usr/libexec/network/packet-steering.uc
index b2229e5d79..34df46d18e 100755
--- a/package/network/config/netifd/files/usr/libexec/network/packet-steering.uc
+++ b/package/network/config/netifd/files/usr/libexec/network/packet-steering.uc
@@ -65,8 +65,9 @@ function cpu_mask(cpu)
 	return sprintf("%x", mask);
 }
 
-function set_netdev_cpu(dev, cpu) {
-	let queues = glob(`/sys/class/net/${dev}/queues/rx-*/rps_cpus`);
+function set_netdev_cpu(dev, cpu, rx_queue) {
+	rx_queue ??= "rx-*";
+	let queues = glob(`/sys/class/net/${dev}/queues/${rx_queue}/rps_cpus`);
 	let val = cpu_mask(cpu);
 	if (disable)
 		val = 0;
@@ -76,7 +77,7 @@ function set_netdev_cpu(dev, cpu) {
 		if (!do_nothing)
 			writefile(queue, `${val}`);
 	}
-	queues = glob(`/sys/class/net/${dev}/queues/rx-*/rps_flow_cnt`);
+	queues = glob(`/sys/class/net/${dev}/queues/${rx_queue}/rps_flow_cnt`);
 	for (let queue in queues) {
 		if (debug || do_nothing)
 			warn(`echo ${local_flows} > ${queue}\n`);
@@ -160,6 +161,9 @@ for (let dev in netdevs) {
 			netdev: [],
 			phy: [],
 			tasks: [],
+			rx_tasks: [],
+			rx_queues: map(glob(`/sys/class/net/${dev}/queues/rx-*/rps_cpus`),
+			               (v) => basename(dirname(v))),
 		};
 	}
 
@@ -187,11 +191,51 @@ for (let path in glob("/proc/*/exe")) {
 			continue;
 
 		push(dev.tasks, pid);
+
+		let napi_match = match(name, /napi\/([^-]*)-(\d+)/);
+		if (napi_match && napi_match[2] > 0)
+			push(dev.rx_tasks, pid);
 		break;
 	}
 }
 
+function assign_dev_queues_cpu(dev) {
+	let num = length(dev.rx_queues);
+	if (num < length(dev.rx_tasks))
+		num = length(dev.rx_tasks);
+
+	for (let i = 0; i < num; i++) {
+		let cpu;
+
+		let task = dev.rx_tasks[i];
+		if (num >= length(cpus))
+			cpu = i % length(cpus);
+		else if (task)
+			cpu = get_next_cpu(napi_weight);
+		else
+			cpu = -1;
+		set_task_cpu(task, cpu);
+
+		let rxq = dev.rx_queues[i];
+		if (!rxq)
+			continue;
+
+		if (num >= length(cpus))
+			cpu = (i + 1) % length(cpus);
+		else if (all_cpus)
+			cpu = -1;
+		else
+			cpu = get_next_cpu(napi_weight, cpu);
+		for (let netdev in dev.netdev)
+			set_netdev_cpu(netdev, cpu, rxq);
+	}
+}
+
 function assign_dev_cpu(dev) {
+	if (length(dev.rx_queues) > 1 &&
+		length(dev.rx_tasks) > 1)
+		return assign_dev_queues_cpu(dev);
+
 	if (length(dev.tasks) > 0) {
 		let cpu = dev.napi_cpu = get_next_cpu(napi_weight);
 		for (let task in dev.tasks)
@@ -204,7 +248,6 @@ function assign_dev_cpu(dev) {
 			cpu = -1;
 		else
 			cpu = get_next_cpu(rx_weight, dev.napi_cpu);
-		dev.rx_cpu = cpu;
 		for (let netdev in dev.netdev)
 			set_netdev_cpu(netdev, cpu);
 	}