Messages in this thread | | | Subject | Re: [PATCH net-next V2] virtio_net: ethtool tx napi configuration | From | Jason Wang <> | Date | Fri, 14 Sep 2018 16:24:15 +0800 |
| |
On 2018年09月13日 23:20, Willem de Bruijn wrote: > On Thu, Sep 13, 2018 at 1:40 AM Jason Wang <jasowang@redhat.com> wrote: >> Implement ethtool .set_coalesce (-C) and .get_coalesce (-c) handlers. >> Interrupt moderation is currently not supported, so these accept and >> display the default settings of 0 usec and 1 frame. >> >> Toggle tx napi through a bit in tx-frames. So as to not interfere >> with possible future interrupt moderation, value 1 means tx napi while >> value 0 means not. >> >> To properly synchronize with the data path, tx napi is disabled and >> tx lock is held when changing the value of napi weight. And two more >> places that can access tx napi weight: >> >> - speculative tx polling in rx napi, we can leave it as is since it >> not a must for correctness. >> - skb_xmit_done(), one more check of napi weight is added before >> trying to enable tx to avoid tx to be disabled forever if napi is >> disabled after skb_xmit_done() but before the napi >> >> Link: https://patchwork.ozlabs.org/patch/948149/ >> Suggested-by: Jason Wang <jasowang@redhat.com> >> Signed-off-by: Willem de Bruijn <willemb@google.com> >> Signed-off-by: Jason Wang <jasowang@redhat.com> >> --- >> Changes from V1: >> - try to synchronize with datapath to allow changing mode when >> interface is up. >> - use tx-frames 0 as to disable tx napi while tx-frames 1 to enable tx napi >> --- >> drivers/net/virtio_net.c | 64 +++++++++++++++++++++++++++++++++++++++- >> 1 file changed, 63 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c >> index 765920905226..6e70864f5899 100644 >> --- a/drivers/net/virtio_net.c >> +++ b/drivers/net/virtio_net.c >> @@ -66,6 +66,8 @@ DECLARE_EWMA(pkt_len, 0, 64) >> >> #define VIRTNET_DRIVER_VERSION "1.0.0" >> >> +static const u32 ethtool_coalesce_napi_mask = (1UL << 10); >> + > This is no longer needed
Yes, will remove this.
> >> static const unsigned long guest_offloads[] = { >> VIRTIO_NET_F_GUEST_TSO4, >> VIRTIO_NET_F_GUEST_TSO6, >> @@ -1444,7 +1446,10 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget) >> >> virtqueue_napi_complete(napi, sq->vq, 0); >> >> - if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) >> + /* Check napi.weight to avoid tx stall since it could be set >> + * to zero by ethtool after skb_xmit_done(). >> + */ >> + if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS || !sq->napi.weight) >> netif_tx_wake_queue(txq); > I see. This assumes that the napi handler will always be called on > conversion from napi to no-napi mode. > > That is safe to assume because if it isn't called (and will not call > netif_tx_wake_queue) that implies that napi was not scheduled, and > thus the tx interrupt was not suppressed and thus there was no tx > completion work to be scheduled?
If it isn't called it means skb_xmit_done() wakeup tx directly instead of schedule tx. This could be a little bit early since there may be still lots of pending tx packets. But it doesn't harm, start_xmit() can handle this by re enable a delayed tx interrupt and disable TX.
But there's a bug, look like I need remove the check of (!sq->napi.weight) in the beginning of the function.
> >> return 0; >> @@ -2181,6 +2186,61 @@ static int virtnet_get_link_ksettings(struct net_device *dev, >> return 0; >> } >> >> +static int virtnet_set_coalesce(struct net_device *dev, >> + struct ethtool_coalesce *ec) >> +{ >> + struct ethtool_coalesce ec_default = { >> + .cmd = ETHTOOL_SCOALESCE, >> + .rx_max_coalesced_frames = 1, >> + }; >> + struct virtnet_info *vi = netdev_priv(dev); >> + int i, napi_weight; >> + >> + if (ec->tx_max_coalesced_frames > 1) >> + return -EINVAL; >> + >> + ec_default.tx_max_coalesced_frames = ec->tx_max_coalesced_frames; >> + napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0; >> + >> + /* disallow changes to fields not explicitly tested above */ >> + if (memcmp(ec, &ec_default, sizeof(ec_default))) >> + return -EINVAL; >> + >> + if (napi_weight ^ vi->sq[0].napi.weight) { >> + for (i = 0; i < vi->max_queue_pairs; i++) { >> + struct netdev_queue *txq = >> + netdev_get_tx_queue(vi->dev, i); >> + >> + virtnet_napi_tx_disable(&vi->sq[i].napi); >> + __netif_tx_lock_bh(txq); >> + vi->sq[i].napi.weight = napi_weight; >> + __netif_tx_unlock_bh(txq); >> + virtnet_napi_tx_enable(vi, vi->sq[i].vq, >> + &vi->sq[i].napi); >> + } >> + } >> + >> + return 0; >> +} >> + >> +static int virtnet_get_coalesce(struct net_device *dev, >> + struct ethtool_coalesce *ec) >> +{ >> + struct ethtool_coalesce ec_default = { >> + .cmd = ETHTOOL_GCOALESCE, >> + .rx_max_coalesced_frames = 1, >> + .tx_max_coalesced_frames = 0, > no need to explicitly initialize to 0 (unless you did this for > documentation purposes, which is fine).
Yes.
Thanks
>> + }; >> + struct virtnet_info *vi = netdev_priv(dev); >> + >> + memcpy(ec, &ec_default, sizeof(ec_default)); >> + >> + if (vi->sq[0].napi.weight) >> + ec->tx_max_coalesced_frames = 1; >> + >> + return 0; >> +} >> + >> static void virtnet_init_settings(struct net_device *dev) >> { >> struct virtnet_info *vi = netdev_priv(dev); >> @@ -2219,6 +2279,8 @@ static const struct ethtool_ops virtnet_ethtool_ops = { >> .get_ts_info = ethtool_op_get_ts_info, >> .get_link_ksettings = virtnet_get_link_ksettings, >> .set_link_ksettings = virtnet_set_link_ksettings, >> + .set_coalesce = virtnet_set_coalesce, >> + .get_coalesce = virtnet_get_coalesce, >> }; >> >> static void virtnet_freeze_down(struct virtio_device *vdev) >> -- >> 2.17.1 >>
| |