[dpdk-dev] [PATCH 1/2] virtio: fix resuming traffic with rx vector path

Olivier Matz olivier.matz at 6wind.com
Mon Feb 12 14:18:27 CET 2018


On Fri, Feb 09, 2018 at 03:26:53PM +0100, Maxime Coquelin wrote:
> This patch fixes traffic resuming issue seen when using
> Rx vector path.
> 
> Fixes: efc83a1e7fc3 ("net/virtio: fix queue setup consistency")
> 
> Signed-off-by: Tiwei Bie <tiwei.bie at intel.com>
> Signed-off-by: Maxime Coquelin <maxime.coquelin at redhat.com>

Proposed change for a more detailed commit log (thanks Tiwei for
the explanation):

Since commit efc83a1e7fc3 ("net/virtio: fix queue setup consistency"),
when resuming a virtio port, the rx rings are refilled with new mbufs
until they are full (vq->vq_free_cnt == 0). This is done without
ensuring that the descriptor index remains a multiple of
RTE_VIRTIO_VPMD_RX_REARM_THRESH, which is a prerequisite when using the
vector mode. This can cause an out of bound access in the rx ring.

This commit changes the vector refill method from
virtqueue_enqueue_recv_refill_simple() to virtio_rxq_rearm_vec(), which
properly checks that the refill is done by batch of
RTE_VIRTIO_VPMD_RX_REARM_THRESH.

> ---
>  drivers/net/virtio/virtio_rxtx.c        | 34 ++++++++++++++++++---------------
>  drivers/net/virtio/virtio_rxtx_simple.c |  2 +-
>  drivers/net/virtio/virtio_rxtx_simple.h |  2 +-
>  3 files changed, 21 insertions(+), 17 deletions(-)
> 
> diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c
> index 854af399e..505283edd 100644
> --- a/drivers/net/virtio/virtio_rxtx.c
> +++ b/drivers/net/virtio/virtio_rxtx.c
> @@ -30,6 +30,7 @@
>  #include "virtio_pci.h"
>  #include "virtqueue.h"
>  #include "virtio_rxtx.h"
> +#include "virtio_rxtx_simple.h"
>  
>  #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
>  #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
> @@ -446,25 +447,28 @@ virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
>  			&rxvq->fake_mbuf;
>  	}
>  
> -	while (!virtqueue_full(vq)) {
> -		m = rte_mbuf_raw_alloc(rxvq->mpool);
> -		if (m == NULL)
> -			break;
> +	if (hw->use_simple_rx) {
> +		while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
> +			virtio_rxq_rearm_vec(rxvq);
> +			nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
> +		}
> +	} else {
> +		while (!virtqueue_full(vq)) {
> +			m = rte_mbuf_raw_alloc(rxvq->mpool);
> +			if (m == NULL)
> +				break;
>  
> -		/* Enqueue allocated buffers */
> -		if (hw->use_simple_rx)
> -			error = virtqueue_enqueue_recv_refill_simple(vq, m);
> -		else

It looks that virtqueue_enqueue_recv_refill_simple() is not used anymore.
We may want to remove it.

It means this patch also fixes the issue related to the patch I've just
submitted [1]. To ease backport process, I think the patch [1] is still
relevant because it fixes an older commit.

[1] http://dpdk.org/dev/patchwork/patch/35130/


> +			/* Enqueue allocated buffers */
>  			error = virtqueue_enqueue_recv_refill(vq, m);
> -
> -		if (error) {
> -			rte_pktmbuf_free(m);
> -			break;
> +			if (error) {
> +				rte_pktmbuf_free(m);
> +				break;
> +			}
> +			nbufs++;
>  		}
> -		nbufs++;
> -	}
>  
> -	vq_update_avail_idx(vq);
> +		vq_update_avail_idx(vq);
> +	}
>  
>  	PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
>  
>  	virtio_rxq_vec_setup(rxvq);

Another thing seen by Tiwei: virtio_rxq_vec_setup() should be called
before virtio_rxq_rearm_vec() because it initializes the mbuf
initializer.

Thanks,
Olivier


More information about the dev mailing list