| Message ID | 20250218023908.1755-1-jasowang@redhat.com (mailing list archive) |
|---|---|
| State | New |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | [net-next] virtio-net: tweak for better TX performance in NAPI mode | expand |
On Tue, Feb 18, 2025 at 10:39:08AM +0800, Jason Wang wrote: > There are several issues existed in start_xmit(): > > - Transmitted packets need to be freed before sending a packet, this > introduces delay and increases the average packets transmit > time. This also increase the time that spent in holding the TX lock. > - Notification is enabled after free_old_xmit_skbs() which will > introduce unnecessary interrupts if TX notification happens on the > same CPU that is doing the transmission now (actually, virtio-net > driver are optimized for this case). > > So this patch tries to avoid those issues by not cleaning transmitted > packets in start_xmit() when TX NAPI is enabled and disable > notifications even more aggressively. Notification will be since the > beginning of the start_xmit(). But we can't enable delayed > notification after TX is stopped as we will lose the > notifications. Instead, the delayed notification needs is enabled > after the virtqueue is kicked for best performance. > > Performance numbers: > > 1) single queue 2 vcpus guest with pktgen_sample03_burst_single_flow.sh > (burst 256) + testpmd (rxonly) on the host: > > - When pinning TX IRQ to pktgen VCPU: split virtqueue PPS were > increased 55% from 6.89 Mpps to 10.7 Mpps and 32% TX interrupts were > eliminated. Packed virtqueue PPS were increased 50% from 7.09 Mpps to > 10.7 Mpps, 99% TX interrupts were eliminated. > > - When pinning TX IRQ to VCPU other than pktgen: split virtqueue PPS > were increased 96% from 5.29 Mpps to 10.4 Mpps and 45% TX interrupts > were eliminated; Packed virtqueue PPS were increased 78% from 6.12 > Mpps to 10.9 Mpps and 99% TX interrupts were eliminated. > > 2) single queue 1 vcpu guest + vhost-net/TAP on the host: single > session netperf from guest to host shows 82% improvement from > 31Gb/s to 58Gb/s, %stddev were reduced from 34.5% to 1.9% and 88% > of TX interrupts were eliminated. > > Signed-off-by: Jason Wang <jasowang@redhat.com> can you pls also test perf with tx irq disabled mode, to be sure? > --- > drivers/net/virtio_net.c | 45 ++++++++++++++++++++++++++++------------ > 1 file changed, 32 insertions(+), 13 deletions(-) > > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c > index 7646ddd9bef7..ac26a6201c44 100644 > --- a/drivers/net/virtio_net.c > +++ b/drivers/net/virtio_net.c > @@ -1088,11 +1088,10 @@ static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q) > return false; > } > > -static void check_sq_full_and_disable(struct virtnet_info *vi, > - struct net_device *dev, > - struct send_queue *sq) > +static bool tx_may_stop(struct virtnet_info *vi, > + struct net_device *dev, > + struct send_queue *sq) > { > - bool use_napi = sq->napi.weight; > int qnum; > > qnum = sq - vi->sq; > @@ -1114,6 +1113,25 @@ static void check_sq_full_and_disable(struct virtnet_info *vi, > u64_stats_update_begin(&sq->stats.syncp); > u64_stats_inc(&sq->stats.stop); > u64_stats_update_end(&sq->stats.syncp); > + > + return true; > + } > + > + return false; > +} > + > +static void check_sq_full_and_disable(struct virtnet_info *vi, > + struct net_device *dev, > + struct send_queue *sq) > +{ > + bool use_napi = sq->napi.weight; > + int qnum; > + > + qnum = sq - vi->sq; > + > + if (tx_may_stop(vi, dev, sq)) { > + struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum); > + > if (use_napi) { > if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) > virtqueue_napi_schedule(&sq->napi, sq->vq); > @@ -3253,15 +3271,10 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) > bool use_napi = sq->napi.weight; > bool kick; > > - /* Free up any pending old buffers before queueing new ones. */ > - do { > - if (use_napi) > - virtqueue_disable_cb(sq->vq); > - > + if (!use_napi) > free_old_xmit(sq, txq, false); > - > - } while (use_napi && !xmit_more && > - unlikely(!virtqueue_enable_cb_delayed(sq->vq))); > + else > + virtqueue_disable_cb(sq->vq); > > /* timestamp packet in software */ > skb_tx_timestamp(skb); > @@ -3287,7 +3300,10 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) > nf_reset_ct(skb); > } > > - check_sq_full_and_disable(vi, dev, sq); > + if (use_napi) > + tx_may_stop(vi, dev, sq); > + else > + check_sq_full_and_disable(vi, dev,sq); > > kick = use_napi ? __netdev_tx_sent_queue(txq, skb->len, xmit_more) : > !xmit_more || netif_xmit_stopped(txq); > @@ -3299,6 +3315,9 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) > } > } > > + if (use_napi && kick && unlikely(!virtqueue_enable_cb_delayed(sq->vq))) > + virtqueue_napi_schedule(&sq->napi, sq->vq); > + > return NETDEV_TX_OK; > } > > -- > 2.34.1
On Tue, Feb 18, 2025 at 4:49 PM Michael S. Tsirkin <mst@redhat.com> wrote: > > On Tue, Feb 18, 2025 at 10:39:08AM +0800, Jason Wang wrote: > > There are several issues existed in start_xmit(): > > > > - Transmitted packets need to be freed before sending a packet, this > > introduces delay and increases the average packets transmit > > time. This also increase the time that spent in holding the TX lock. > > - Notification is enabled after free_old_xmit_skbs() which will > > introduce unnecessary interrupts if TX notification happens on the > > same CPU that is doing the transmission now (actually, virtio-net > > driver are optimized for this case). > > > > So this patch tries to avoid those issues by not cleaning transmitted > > packets in start_xmit() when TX NAPI is enabled and disable > > notifications even more aggressively. Notification will be since the > > beginning of the start_xmit(). But we can't enable delayed > > notification after TX is stopped as we will lose the > > notifications. Instead, the delayed notification needs is enabled > > after the virtqueue is kicked for best performance. > > > > Performance numbers: > > > > 1) single queue 2 vcpus guest with pktgen_sample03_burst_single_flow.sh > > (burst 256) + testpmd (rxonly) on the host: > > > > - When pinning TX IRQ to pktgen VCPU: split virtqueue PPS were > > increased 55% from 6.89 Mpps to 10.7 Mpps and 32% TX interrupts were > > eliminated. Packed virtqueue PPS were increased 50% from 7.09 Mpps to > > 10.7 Mpps, 99% TX interrupts were eliminated. > > > > - When pinning TX IRQ to VCPU other than pktgen: split virtqueue PPS > > were increased 96% from 5.29 Mpps to 10.4 Mpps and 45% TX interrupts > > were eliminated; Packed virtqueue PPS were increased 78% from 6.12 > > Mpps to 10.9 Mpps and 99% TX interrupts were eliminated. > > > > 2) single queue 1 vcpu guest + vhost-net/TAP on the host: single > > session netperf from guest to host shows 82% improvement from > > 31Gb/s to 58Gb/s, %stddev were reduced from 34.5% to 1.9% and 88% > > of TX interrupts were eliminated. > > > > Signed-off-by: Jason Wang <jasowang@redhat.com> > > can you pls also test perf with tx irq disabled mode, to be sure? Yes, performance doesn't change in orphan mode (both PPS and throughput). This matches the patch that doesn't not change any logic for orphan mode. Thanks > > > --- > > drivers/net/virtio_net.c | 45 ++++++++++++++++++++++++++++------------ > > 1 file changed, 32 insertions(+), 13 deletions(-) > > > > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c > > index 7646ddd9bef7..ac26a6201c44 100644 > > --- a/drivers/net/virtio_net.c > > +++ b/drivers/net/virtio_net.c > > @@ -1088,11 +1088,10 @@ static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q) > > return false; > > } > > > > -static void check_sq_full_and_disable(struct virtnet_info *vi, > > - struct net_device *dev, > > - struct send_queue *sq) > > +static bool tx_may_stop(struct virtnet_info *vi, > > + struct net_device *dev, > > + struct send_queue *sq) > > { > > - bool use_napi = sq->napi.weight; > > int qnum; > > > > qnum = sq - vi->sq; > > @@ -1114,6 +1113,25 @@ static void check_sq_full_and_disable(struct virtnet_info *vi, > > u64_stats_update_begin(&sq->stats.syncp); > > u64_stats_inc(&sq->stats.stop); > > u64_stats_update_end(&sq->stats.syncp); > > + > > + return true; > > + } > > + > > + return false; > > +} > > + > > +static void check_sq_full_and_disable(struct virtnet_info *vi, > > + struct net_device *dev, > > + struct send_queue *sq) > > +{ > > + bool use_napi = sq->napi.weight; > > + int qnum; > > + > > + qnum = sq - vi->sq; > > + > > + if (tx_may_stop(vi, dev, sq)) { > > + struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum); > > + > > if (use_napi) { > > if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) > > virtqueue_napi_schedule(&sq->napi, sq->vq); > > @@ -3253,15 +3271,10 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) > > bool use_napi = sq->napi.weight; > > bool kick; > > > > - /* Free up any pending old buffers before queueing new ones. */ > > - do { > > - if (use_napi) > > - virtqueue_disable_cb(sq->vq); > > - > > + if (!use_napi) > > free_old_xmit(sq, txq, false); > > - > > - } while (use_napi && !xmit_more && > > - unlikely(!virtqueue_enable_cb_delayed(sq->vq))); > > + else > > + virtqueue_disable_cb(sq->vq); > > > > /* timestamp packet in software */ > > skb_tx_timestamp(skb); > > @@ -3287,7 +3300,10 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) > > nf_reset_ct(skb); > > } > > > > - check_sq_full_and_disable(vi, dev, sq); > > + if (use_napi) > > + tx_may_stop(vi, dev, sq); > > + else > > + check_sq_full_and_disable(vi, dev,sq); > > > > kick = use_napi ? __netdev_tx_sent_queue(txq, skb->len, xmit_more) : > > !xmit_more || netif_xmit_stopped(txq); > > @@ -3299,6 +3315,9 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) > > } > > } > > > > + if (use_napi && kick && unlikely(!virtqueue_enable_cb_delayed(sq->vq))) > > + virtqueue_napi_schedule(&sq->napi, sq->vq); > > + > > return NETDEV_TX_OK; > > } > > > > -- > > 2.34.1 >
diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c index 7646ddd9bef7..ac26a6201c44 100644 --- a/drivers/net/virtio_net.c +++ b/drivers/net/virtio_net.c @@ -1088,11 +1088,10 @@ static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q) return false; } -static void check_sq_full_and_disable(struct virtnet_info *vi, - struct net_device *dev, - struct send_queue *sq) +static bool tx_may_stop(struct virtnet_info *vi, + struct net_device *dev, + struct send_queue *sq) { - bool use_napi = sq->napi.weight; int qnum; qnum = sq - vi->sq; @@ -1114,6 +1113,25 @@ static void check_sq_full_and_disable(struct virtnet_info *vi, u64_stats_update_begin(&sq->stats.syncp); u64_stats_inc(&sq->stats.stop); u64_stats_update_end(&sq->stats.syncp); + + return true; + } + + return false; +} + +static void check_sq_full_and_disable(struct virtnet_info *vi, + struct net_device *dev, + struct send_queue *sq) +{ + bool use_napi = sq->napi.weight; + int qnum; + + qnum = sq - vi->sq; + + if (tx_may_stop(vi, dev, sq)) { + struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum); + if (use_napi) { if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) virtqueue_napi_schedule(&sq->napi, sq->vq); @@ -3253,15 +3271,10 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) bool use_napi = sq->napi.weight; bool kick; - /* Free up any pending old buffers before queueing new ones. */ - do { - if (use_napi) - virtqueue_disable_cb(sq->vq); - + if (!use_napi) free_old_xmit(sq, txq, false); - - } while (use_napi && !xmit_more && - unlikely(!virtqueue_enable_cb_delayed(sq->vq))); + else + virtqueue_disable_cb(sq->vq); /* timestamp packet in software */ skb_tx_timestamp(skb); @@ -3287,7 +3300,10 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) nf_reset_ct(skb); } - check_sq_full_and_disable(vi, dev, sq); + if (use_napi) + tx_may_stop(vi, dev, sq); + else + check_sq_full_and_disable(vi, dev,sq); kick = use_napi ? __netdev_tx_sent_queue(txq, skb->len, xmit_more) : !xmit_more || netif_xmit_stopped(txq); @@ -3299,6 +3315,9 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) } } + if (use_napi && kick && unlikely(!virtqueue_enable_cb_delayed(sq->vq))) + virtqueue_napi_schedule(&sq->napi, sq->vq); + return NETDEV_TX_OK; }
There are several issues existed in start_xmit(): - Transmitted packets need to be freed before sending a packet, this introduces delay and increases the average packets transmit time. This also increase the time that spent in holding the TX lock. - Notification is enabled after free_old_xmit_skbs() which will introduce unnecessary interrupts if TX notification happens on the same CPU that is doing the transmission now (actually, virtio-net driver are optimized for this case). So this patch tries to avoid those issues by not cleaning transmitted packets in start_xmit() when TX NAPI is enabled and disable notifications even more aggressively. Notification will be since the beginning of the start_xmit(). But we can't enable delayed notification after TX is stopped as we will lose the notifications. Instead, the delayed notification needs is enabled after the virtqueue is kicked for best performance. Performance numbers: 1) single queue 2 vcpus guest with pktgen_sample03_burst_single_flow.sh (burst 256) + testpmd (rxonly) on the host: - When pinning TX IRQ to pktgen VCPU: split virtqueue PPS were increased 55% from 6.89 Mpps to 10.7 Mpps and 32% TX interrupts were eliminated. Packed virtqueue PPS were increased 50% from 7.09 Mpps to 10.7 Mpps, 99% TX interrupts were eliminated. - When pinning TX IRQ to VCPU other than pktgen: split virtqueue PPS were increased 96% from 5.29 Mpps to 10.4 Mpps and 45% TX interrupts were eliminated; Packed virtqueue PPS were increased 78% from 6.12 Mpps to 10.9 Mpps and 99% TX interrupts were eliminated. 2) single queue 1 vcpu guest + vhost-net/TAP on the host: single session netperf from guest to host shows 82% improvement from 31Gb/s to 58Gb/s, %stddev were reduced from 34.5% to 1.9% and 88% of TX interrupts were eliminated. Signed-off-by: Jason Wang <jasowang@redhat.com> --- drivers/net/virtio_net.c | 45 ++++++++++++++++++++++++++++------------ 1 file changed, 32 insertions(+), 13 deletions(-)