A deadlock may occur in kni kernel thread while netif_receive_skb is called

A deadlock may occur in kni kernel thread while netif_receive_skb is called

[Tyrone Lau]

Hi, all. I have found recently the Linux kernel will complain occasionally
a dead lock, while I use the kernel module rte_kni provided in DPDK. After
reviewing the dpdk source code and googling,
I found that the deadlock occurred because netif_receive_skb is invoked in
a non-softirq context. The erroneous source code is listed as below (in
lib/librte_eal/linuxapp/kni/kni_net.c:kni_net_rx_normal):

*	/* Transfer received packets to netif */
	for (i = 0; i < num; i++) {
		kva = (void *)va[i] - kni->mbuf_va + kni->mbuf_kva;
		len = kva->data_len;
		data_kva = kva->data - kni->mbuf_va + kni->mbuf_kva;

		skb = dev_alloc_skb(len + 2);
		if (!skb) {
			KNI_ERR("Out of mem, dropping pkts\n");
			/* Update statistics */
			kni->stats.rx_dropped++;
		}
		else {
			/* Align IP on 16B boundary */
			skb_reserve(skb, 2);
			memcpy(skb_put(skb, len), data_kva, len);
			skb->dev = dev;
			skb->protocol = eth_type_trans(skb, dev);
			skb->ip_summed = CHECKSUM_UNNECESSARY;

			/* Call netif interface */
			netif_receive_skb(skb);

			/* Update statistics */
			kni->stats.rx_bytes += len;
			kni->stats.rx_packets++;
		}
	}*

The similar bug is reported and fixed in dpdk extension memnic. See

http://comments.gmane.org/gmane.comp.networking.dpdk.devel/3151

To fix this bug, we should call local_bh_disable/local_bh_enable
around netif_receive_skb to disable and re-enable soft-irq.
Best Regards

Re: A deadlock may occur in kni kernel thread while netif_receive_skb is called

[Stephen Hemminger]

On Thu, 12 Jun 2014 22:46:14 +0800
Tyrone Lau <tyronelau-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org> wrote:

> Hi, all. I have found recently the Linux kernel will complain occasionally
> a dead lock, while I use the kernel module rte_kni provided in DPDK. After
> reviewing the dpdk source code and googling,
> I found that the deadlock occurred because netif_receive_skb is invoked in
> a non-softirq context. The erroneous source code is listed as below (in
> lib/librte_eal/linuxapp/kni/kni_net.c:kni_net_rx_normal):
> 
> *	/* Transfer received packets to netif */
> 	for (i = 0; i < num; i++) {
> 		kva = (void *)va[i] - kni->mbuf_va + kni->mbuf_kva;
> 		len = kva->data_len;
> 		data_kva = kva->data - kni->mbuf_va + kni->mbuf_kva;
> 
> 		skb = dev_alloc_skb(len + 2);
> 		if (!skb) {
> 			KNI_ERR("Out of mem, dropping pkts\n");
> 			/* Update statistics */
> 			kni->stats.rx_dropped++;
> 		}
> 		else {
> 			/* Align IP on 16B boundary */
> 			skb_reserve(skb, 2);
> 			memcpy(skb_put(skb, len), data_kva, len);
> 			skb->dev = dev;
> 			skb->protocol = eth_type_trans(skb, dev);
> 			skb->ip_summed = CHECKSUM_UNNECESSARY;
> 
> 			/* Call netif interface */
> 			netif_receive_skb(skb);
> 
> 			/* Update statistics */
> 			kni->stats.rx_bytes += len;
> 			kni->stats.rx_packets++;
> 		}
> 	}*
> 
> The similar bug is reported and fixed in dpdk extension memnic. See
> 
> http://comments.gmane.org/gmane.comp.networking.dpdk.devel/3151
> 
> To fix this bug, we should call local_bh_disable/local_bh_enable
> around netif_receive_skb to disable and re-enable soft-irq.
> Best Regards

Probably better to call netif_rx instead, because that will
handle the case of overrun.

Other comments, this code should be using per-cpu stats.
it should use netdev_alloc_skb_ip_align rather than doing align itself.

Even better yet would be bursting packets into the receive handler.