Re: [PATCH net-next V2] virtio_net: ethtool tx napi configuration

On 2018年09月13日 23:20, Willem de Bruijn wrote:

On Thu, Sep 13, 2018 at 1:40 AM Jason Wang [off-list ref] wrote:

quoted

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 <redacted>
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Jason Wang <redacted>
---
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.

quoted

  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.

quoted

         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

quoted

+       };
+       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

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization