[dpdk-dev] [PATCH] virtio: fix segfault when transmit pkts

[Stephen Hemminger]

On Thu, 21 Apr 2016 12:36:10 +0000
Jianfeng Tan <jianfeng.tan at intel.com> wrote:

> Issue: when using virtio nic to transmit pkts, it causes segment fault.
> 
> How to reproduce:
> a. start testpmd with vhost.
> $testpmd -c 0x3 -n 4 --socket-mem 1024,0 --no-pci \
>   --vdev 'eth_vhost0,iface=/tmp/sock0,queues=1' -- -i --nb-cores=1
> b. start a qemu with a virtio nic connected with the vhost-user port.
> $qemu -smp cores=2,sockets=1 -cpu host -enable-kvm vm-0.img -vnc :5 -m 4G \
>   -object memory-backend-file,id=mem,size=4096M,mem-path=<path>,share=on \
>   -numa node,memdev=mem -mem-prealloc \
>   -chardev socket,id=char1,path=$sock_vhost \
>   -netdev type=vhost-user,id=net1,chardev=char1 \
>   -device virtio-net-pci,netdev=net1,mac=00:01:02:03:04:05
> c. enable testpmd on the host.
> testpmd> set fwd io
> testpmd> start
> d. start testpmd in VM.
> $testpmd -c 0x3 -n 4 -m 1024 -- -i --disable-hw-vlan-filter --txqflags=0xf01
> testpmd> set fwd txonly
> testpmd> start
> 
> How to fix: this bug is because inside virtqueue_enqueue_xmit(), the flag of
> desc has been updated inside the do {} while (); and after the loop, all descs
> could have run out, so idx is VQ_RING_DESC_CHAIN_END (32768), use this idx to
> reference the start_dp array will lead to segment fault.
> 
> Signed-off-by: Jianfeng Tan <jianfeng.tan at intel.com>
> ---
>  drivers/net/virtio/virtio_rxtx.c | 2 --
>  1 file changed, 2 deletions(-)
> 
> diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c
> index ef21d8e..432aeab 100644
> --- a/drivers/net/virtio/virtio_rxtx.c
> +++ b/drivers/net/virtio/virtio_rxtx.c
> @@ -271,8 +271,6 @@ virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie,
>  		idx = start_dp[idx].next;
>  	} while ((cookie = cookie->next) != NULL);
>  
> -	start_dp[idx].flags &= ~VRING_DESC_F_NEXT;
> -
>  	if (use_indirect)
>  		idx = txvq->vq_ring.desc[head_idx].next;
>  

At this point in the code idx is the index past the current set of ring
descriptors. So yes this is a real bug.

I think the description meta-data needs work to explain it better.