On Wed, Oct 1, 2014 at 5:36 AM, David Verbeiren
<david.verbei...@intel.com> wrote:
> DPDK rings don't need one TX queue per PMD thread and don't support multiple
> queues (set_multiq function is undefined). To fix operation with DPDK rings,
> this patch ignores EOPNOTSUPP error on netdev_set_multiq() and ignores, for
> DPDK ring netdevs, the provided queue id (= PMD thread core id) in
> netdev_dpdk_send().
>
Thanks for the fix.
> Signed-off-by: David Verbeiren <david.verbei...@intel.com>
> ---
> lib/dpif-netdev.c | 4 ++--
> lib/netdev-dpdk.c | 11 +++++++++++
> 2 files changed, 13 insertions(+), 2 deletions(-)
>
> diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
> index 6b8201b..f81d638 100644
> --- a/lib/dpif-netdev.c
> +++ b/lib/dpif-netdev.c
> @@ -786,7 +786,7 @@ do_add_port(struct dp_netdev *dp, const char *devname,
> const char *type,
> /* There can only be ovs_numa_get_n_cores() pmd threads,
> * so creates a txq for each. */
> error = netdev_set_multiq(netdev, n_cores, dp->n_dpdk_rxqs);
> - if (error) {
> + if (error && (error != EOPNOTSUPP)) {
> VLOG_ERR("%s, cannot set multiq", devname);
> return errno;
> }
> @@ -1950,7 +1950,7 @@ dpif_netdev_pmd_set(struct dpif *dpif, unsigned int
> n_rxqs, const char *cmask)
> /* Sets the new rx queue config. */
> err = netdev_set_multiq(port->netdev, ovs_numa_get_n_cores(),
> n_rxqs);
> - if (err) {
> + if (err && (err != EOPNOTSUPP)) {
It is ok as a fix. But in future we should improve it by adding
max-queue conf for each netdev implementation. This can be used by
dpif-netdev to set valid number of queues.
> VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
> " %u", netdev_get_name(port->netdev),
> n_rxqs);
> diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c
> index 9c93768..c72fe31 100644
> --- a/lib/netdev-dpdk.c
> +++ b/lib/netdev-dpdk.c
> @@ -212,6 +212,7 @@ struct netdev_rxq_dpdk {
> static bool thread_is_pmd(void);
>
> static int netdev_dpdk_construct(struct netdev *);
> +static int netdev_dpdk_ring_construct(struct netdev *);
>
> static bool
> is_dpdk_class(const struct netdev_class *class)
> @@ -219,6 +220,12 @@ is_dpdk_class(const struct netdev_class *class)
> return class->construct == netdev_dpdk_construct;
> }
>
> +static bool
> +is_dpdk_ring_class(const struct netdev_class *class)
> +{
> + return class->construct == netdev_dpdk_ring_construct;
> +}
> +
> /* XXX: use dpdk malloc for entire OVS. infact huge page shld be used
> * for all other sengments data, bss and text. */
>
> @@ -853,6 +860,10 @@ netdev_dpdk_send(struct netdev *netdev, int qid, struct
> dpif_packet **pkts,
> int ret;
> int i;
>
> + if (is_dpdk_ring_class(netdev->netdev_class)) {
> + qid = 0; /* DPDK Rings have a single TX queue */
> + }
> +
> if (!may_steal || pkts[0]->ofpbuf.source != OFPBUF_DPDK) {
> dpdk_do_tx_copy(netdev, qid, pkts, cnt);
>
We can avoid this check by by defining separate netdev-send() for
dpdk-ring device. in ring-send() we can always ignore qid parameter.
Can you send updated patch?
> --
> 1.8.3.2
>
> _______________________________________________
> dev mailing list
> dev@openvswitch.org
> http://openvswitch.org/mailman/listinfo/dev
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
