On Wed, 6 May 2020 09:09:52 +0200 Bartosz Golaszewski wrote:
> > > +}
> >
> > Why do you clean the TX ring from a work rather than from the NAPI
> > context?
>
> So this was unclear to me, that's why I went with a workqueue. The
> budget argument in napi poll is for RX. Should I put some cap on the
> number of TX descriptors processed in napi context?
The prevailing wisdom is to not count the TX cleanup as work at all.
I think the best practice is to first clean up all the TX you can,
and then do at must @budget of RX.
Perhaps one day we will come up with a good way of capping TX, but
today not counting it towards budget is the safe choice.
> > > +static int mtk_mac_receive_packet(struct mtk_mac_priv *priv)
> > > +{
> > > + struct net_device *ndev = mtk_mac_get_netdev(priv);
> > > + struct mtk_mac_ring *ring = &priv->rx_ring;
> > > + struct device *dev = mtk_mac_get_dev(priv);
> > > + struct mtk_mac_ring_desc_data desc_data;
> > > + struct sk_buff *new_skb;
> > > + int ret;
> > > +
> > > + mtk_mac_lock(priv);
> > > + ret = mtk_mac_ring_pop_tail(ring, &desc_data);
> > > + mtk_mac_unlock(priv);
> > > + if (ret)
> > > + return -1;
> > > +
> > > + mtk_mac_dma_unmap_rx(priv, &desc_data);
> > > +
> > > + if ((desc_data.flags & MTK_MAC_DESC_BIT_RX_CRCE) ||
> > > + (desc_data.flags & MTK_MAC_DESC_BIT_RX_OSIZE)) {
> > > + /* Error packet -> drop and reuse skb. */
> > > + new_skb = desc_data.skb;
> > > + goto map_skb;
> > > + }
> > > +
> > > + new_skb = mtk_mac_alloc_skb(ndev);
> > > + if (!new_skb) {
> > > + netdev_err(ndev, "out of memory for skb\n");
> >
> > No need for printing, kernel will complain loudly about oom.
> >
> > > + ndev->stats.rx_dropped++;
> > > + new_skb = desc_data.skb;
> > > + goto map_skb;
> > > + }
> > > +
> > > + skb_put(desc_data.skb, desc_data.len);
> > > + desc_data.skb->ip_summed = CHECKSUM_NONE;
> > > + desc_data.skb->protocol = eth_type_trans(desc_data.skb, ndev);
> > > + desc_data.skb->dev = ndev;
> > > + netif_receive_skb(desc_data.skb);
> > > +
> > > +map_skb:
> > > + desc_data.dma_addr = mtk_mac_dma_map_rx(priv, new_skb);
> > > + if (dma_mapping_error(dev, desc_data.dma_addr)) {
> > > + dev_kfree_skb(new_skb);
> > > + netdev_err(ndev, "DMA mapping error of RX descriptor\n");
> > > + return -ENOMEM;
> >
> > In this case nothing will ever replenish the RX ring right? If we hit
> > this condition 128 times the ring will be empty?
>
> Indeed. What should I do if this fails though?
I think if you move things around it should work:
skb = pop_tail();
if (!skb)
return;
new_skb = alloc();
if (!new_skb)
goto reuse;
dma_map(new_skb);
if (error)
goto reuse;
dma_unmap(skb);
if (do_packet_processing())
free(skb);
else
receive(skb);
put_on_ring(new_skb);
