openwrt/target/linux/mvebu/patches-5.4/700-mvneta-tx-queue-workaround.patch
Tomasz Maciej Nowak cbdd2b62e4 mvebu: limit mvneta tx queue workaround to 32 bit SoC
This patch has been carried since introduction throughout every kernel
major bump and no one has tested if the later kernels improved the
situation. The Armada 3720 SoC can only process GbE interrupts on Core 0
and this is already limited in all stable kernels, so ditch this
workaround for 64 bit SoCs.

Ref: https://git.kernel.org/torvalds/c/cf9bf871280d

Signed-off-by: Tomasz Maciej Nowak <tmn505@gmail.com>
2021-07-25 13:52:38 +02:00

39 lines
1.3 KiB
Diff

The hardware queue scheduling is apparently configured with fixed
priorities, which creates a nasty fairness issue where traffic from one
CPU can starve traffic from all other CPUs.
Work around this issue by forcing all tx packets to go through one CPU,
until this issue is fixed properly.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
---
--- a/drivers/net/ethernet/marvell/mvneta.c
+++ b/drivers/net/ethernet/marvell/mvneta.c
@@ -4691,6 +4691,16 @@ static int mvneta_ethtool_set_eee(struct
return phylink_ethtool_set_eee(pp->phylink, eee);
}
+#ifndef CONFIG_ARM64
+static u16 mvneta_select_queue(struct net_device *dev, struct sk_buff *skb,
+ struct net_device *sb_dev)
+{
+ /* XXX: hardware queue scheduling is broken,
+ * use only one queue until it is fixed */
+ return 0;
+}
+#endif
+
static const struct net_device_ops mvneta_netdev_ops = {
.ndo_open = mvneta_open,
.ndo_stop = mvneta_stop,
@@ -4701,6 +4711,9 @@ static const struct net_device_ops mvnet
.ndo_fix_features = mvneta_fix_features,
.ndo_get_stats64 = mvneta_get_stats64,
.ndo_do_ioctl = mvneta_ioctl,
+#ifndef CONFIG_ARM64
+ .ndo_select_queue = mvneta_select_queue,
+#endif
.ndo_bpf = mvneta_xdp,
.ndo_xdp_xmit = mvneta_xdp_xmit,
};