[dpdk-dev,1/2] virtio: fix resuming traffic with rx vector path
Checks
Commit Message
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@intel.com>
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
---
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(-)
Comments
> -----Original Message-----
> From: Maxime Coquelin [mailto:maxime.coquelin@redhat.com]
> Sent: Friday, February 9, 2018 10:27 PM
> To: Bie, Tiwei <tiwei.bie@intel.com>; yliu@fridaylinux.org; Yigit, Ferruh
> <ferruh.yigit@intel.com>; victork@redhat.com
> Cc: dev@dpdk.org; stable@dpdk.org; Wang, Zhihong
> <zhihong.wang@intel.com>; Xu, Qian Q <qian.q.xu@intel.com>; Yao, Lei A
> <lei.a.yao@intel.com>; Maxime Coquelin <maxime.coquelin@redhat.com>
> Subject: [PATCH 1/2] virtio: fix resuming traffic with rx vector path
>
> 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@intel.com>
> Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Lei Yao <lei.a.yao@intel.com>
This patch has been tested by regression test suite. It can fix the traffic resume
issue with vector path. No performance drop during PVP test:
Following test are also checked and passed:
Vhost/virtio multi queue
Virtio-user
Virtio-user as exception path
Vhost/virtio reconnect
My server info:
OS: Ubuntu 16.04
Kernel: 4.4.0-110
CPU: Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz
BR
Lei
> ---
> 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
> + /* 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);
>
> diff --git a/drivers/net/virtio/virtio_rxtx_simple.c
> b/drivers/net/virtio/virtio_rxtx_simple.c
> index 7247a0822..0a79d1d5b 100644
> --- a/drivers/net/virtio/virtio_rxtx_simple.c
> +++ b/drivers/net/virtio/virtio_rxtx_simple.c
> @@ -77,7 +77,7 @@ virtio_xmit_pkts_simple(void *tx_queue, struct
> rte_mbuf **tx_pkts,
> rte_compiler_barrier();
>
> if (nb_used >= VIRTIO_TX_FREE_THRESH)
> - virtio_xmit_cleanup(vq);
> + virtio_xmit_cleanup_simple(vq);
>
> nb_commit = nb_pkts = RTE_MIN((vq->vq_free_cnt >> 1), nb_pkts);
> desc_idx = (uint16_t)(vq->vq_avail_idx & desc_idx_max);
> diff --git a/drivers/net/virtio/virtio_rxtx_simple.h
> b/drivers/net/virtio/virtio_rxtx_simple.h
> index 2d8e6b14a..303904d64 100644
> --- a/drivers/net/virtio/virtio_rxtx_simple.h
> +++ b/drivers/net/virtio/virtio_rxtx_simple.h
> @@ -60,7 +60,7 @@ virtio_rxq_rearm_vec(struct virtnet_rx *rxvq)
> #define VIRTIO_TX_FREE_NR 32
> /* TODO: vq->tx_free_cnt could mean num of free slots so we could avoid
> shift */
> static inline void
> -virtio_xmit_cleanup(struct virtqueue *vq)
> +virtio_xmit_cleanup_simple(struct virtqueue *vq)
> {
> uint16_t i, desc_idx;
> uint32_t nb_free = 0;
> --
> 2.14.3
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@intel.com>
> Signed-off-by: Maxime Coquelin <maxime.coquelin@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
Just realize that I forgot to CC dev@dpdk.org.
> -----Original Message-----
> From: Tan, Jianfeng [mailto:jianfeng.tan@intel.com]
> Sent: Monday, February 12, 2018 8:51 PM
> To: Maxime Coquelin
> Subject: Re: [dpdk-dev] [PATCH 1/2] virtio: fix resuming traffic with rx vector
> path
>
>
>
> On 2/9/2018 10:26 PM, Maxime Coquelin wrote:
> > This patch fixes traffic resuming issue seen when using
> > Rx vector path.
>
> A nit: AFAIK, this issue happens after start/stop/start? "traffic
> resuming" seems not accurate.
>
> >
> > Fixes: efc83a1e7fc3 ("net/virtio: fix queue setup consistency")
> >
> > Signed-off-by: Tiwei Bie <tiwei.bie@intel.com>
> > Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
>
> Besides that,
>
> Reviewed-by: Jianfeng Tan <jianfeng.tan@intel.com>
>
> Thanks for this timely fix!
>
> Thanks,
> Jianfeng
@@ -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
+ /* 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);
@@ -77,7 +77,7 @@ virtio_xmit_pkts_simple(void *tx_queue, struct rte_mbuf **tx_pkts,
rte_compiler_barrier();
if (nb_used >= VIRTIO_TX_FREE_THRESH)
- virtio_xmit_cleanup(vq);
+ virtio_xmit_cleanup_simple(vq);
nb_commit = nb_pkts = RTE_MIN((vq->vq_free_cnt >> 1), nb_pkts);
desc_idx = (uint16_t)(vq->vq_avail_idx & desc_idx_max);
@@ -60,7 +60,7 @@ virtio_rxq_rearm_vec(struct virtnet_rx *rxvq)
#define VIRTIO_TX_FREE_NR 32
/* TODO: vq->tx_free_cnt could mean num of free slots so we could avoid shift */
static inline void
-virtio_xmit_cleanup(struct virtqueue *vq)
+virtio_xmit_cleanup_simple(struct virtqueue *vq)
{
uint16_t i, desc_idx;
uint32_t nb_free = 0;