[PATCH] mvneta: fix "napi poll" infinite loop for openwrt

[Kabuli Chana]

Why is this kernel 5.15 based?


On 2024-09-11 06:18, webmaster at jbsky.fr wrote:
> From: Julien Blais <webmaster at jbsky.fr>
>
> This patch is submitted to the kernel.
>
> For your information, this fixes the problem identified during a test with
>   iperf, the packet sending stall and above all the use of a 100% CPU
>   because of the softIRQ.
>
> Before:
> root at wrt1900acv1:~# cat /proc/interrupts
>             CPU0       CPU1
>   24:   18652404   15171396      MPIC   5 Level     armada_370_xp_per_cpu_tick
>   26:     292136         78      MPIC  31 Level     mv64xxx_i2c
>   27:         26          0      MPIC  41 Level     ttyS0
>   33:    1024678          0      MPIC  45 Level     ehci_hcd:usb1
>   34:   23007175          0      MPIC   8 Level     eth0
>
> After:
> root at OpenWrt:/# cat /proc/interrupts
>             CPU0       CPU1
>   24:     194183     161947      MPIC   5 Level     armada_370_xp_per_cpu_tick
>   25:          0          0      MPIC   3 Level     arm-pmu
>   26:        246          0      MPIC  31 Level     mv64xxx_i2c
>   27:       5488          0      MPIC  41 Level     ttyS0
>   33:          0          0      MPIC  45 Level     ehci_hcd:usb1
>   34:    2015130    6330694      MPIC   8 Level     eth0
>
> Signed-off-by: Julien Blais <webmaster at jbsky.fr>
> ---
>   .../700-mvneta-tx-queue-workaround.patch      | 38 -------------------
>   ...1-mvneta-fix-napi-poll-infinite-loop.patch | 38 +++++++++++++++++++
>   2 files changed, 38 insertions(+), 38 deletions(-)
>   delete mode 100644 target/linux/mvebu/patches-5.15/700-mvneta-tx-queue-workaround.patch
>   create mode 100644 target/linux/mvebu/patches-5.15/701-mvneta-fix-napi-poll-infinite-loop.patch
>
> diff --git a/target/linux/mvebu/patches-5.15/700-mvneta-tx-queue-workaround.patch b/target/linux/mvebu/patches-5.15/700-mvneta-tx-queue-workaround.patch
> deleted file mode 100644
> index 32e8ef4b7d..0000000000
> --- a/target/linux/mvebu/patches-5.15/700-mvneta-tx-queue-workaround.patch
> +++ /dev/null
> @@ -1,38 +0,0 @@
> -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 at nbd.name>
> ----
> ---- a/drivers/net/ethernet/marvell/mvneta.c
> -+++ b/drivers/net/ethernet/marvell/mvneta.c
> -@@ -5006,6 +5006,16 @@ static int mvneta_setup_tc(struct net_de
> - 	}
> - }
> -
> -+#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,
> -@@ -5016,6 +5026,9 @@ static const struct net_device_ops mvnet
> - 	.ndo_fix_features    = mvneta_fix_features,
> - 	.ndo_get_stats64     = mvneta_get_stats64,
> - 	.ndo_eth_ioctl        = mvneta_ioctl,
> -+#ifndef CONFIG_ARM64
> -+	.ndo_select_queue    = mvneta_select_queue,
> -+#endif
> - 	.ndo_bpf	     = mvneta_xdp,
> - 	.ndo_xdp_xmit        = mvneta_xdp_xmit,
> - 	.ndo_setup_tc	     = mvneta_setup_tc,
> diff --git a/target/linux/mvebu/patches-5.15/701-mvneta-fix-napi-poll-infinite-loop.patch b/target/linux/mvebu/patches-5.15/701-mvneta-fix-napi-poll-infinite-loop.patch
> new file mode 100644
> index 0000000000..0fa191ea14
> --- /dev/null
> +++ b/target/linux/mvebu/patches-5.15/701-mvneta-fix-napi-poll-infinite-loop.patch
> @@ -0,0 +1,38 @@
> +From bbeea07de50e925df3877f63a24aee1d35828a02 Mon Sep 17 00:00:00 2001
> +From: Julien Blais <webmaster at jbsky.fr>
> +Date: Wed, 11 Sep 2024 13:27:09 +0200
> +Subject: [PATCH v2] mvneta: fix "napi poll" infinite loop
> +
> +In percpu mode, when there's a network load, one of the cpus can be
> +solicited without having anything to process.
> +If 0 is returned to napi poll, napi will ignore the next requests,
> +causing an infinite loop with ISR handling.
> +
> +Without this change, patches hang around fixing the queue at 0 and
> +the interrupt remains stuck on the 1st CPU.
> +The percpu conf is useless in this case, so we might as well remove it.
> +
> +Signed-off-by: Julien Blais <webmaster at jbsky.fr>
> +---
> + drivers/net/ethernet/marvell/mvneta.c | 5 ++++-
> + 1 file changed, 4 insertions(+), 1 deletion(-)
> +
> +diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c
> +index 3f124268b..b6e89b888 100644
> +--- a/drivers/net/ethernet/marvell/mvneta.c
> ++++ b/drivers/net/ethernet/marvell/mvneta.c
> +@@ -3186,7 +3186,10 @@ static int mvneta_poll(struct napi_struct *napi, int budget)
> +
> + 	if (rx_done < budget) {
> + 		cause_rx_tx = 0;
> +-		napi_complete_done(napi, rx_done);
> ++		if (rx_done)
> ++			napi_complete_done(napi, rx_done);
> ++		else
> ++			napi_complete(napi);
> +
> + 		if (pp->neta_armada3700) {
> + 			unsigned long flags;
> +--
> +2.39.2
> +