Re: [PATCH net-next v2 03/10] net: sparx5: add hostmode with phylink support

From: Steen Hegelund
Date: Mon May 31 2021 - 10:56:00 EST


Hi Jakub,

Thanks for your review.

On Sun, 2021-05-30 at 14:15 -0700, Jakub Kicinski wrote:
> EXTERNAL EMAIL: Do not click links or open attachments unless you know the content is safe
>
> On Fri, 28 May 2021 14:34:12 +0200 Steen Hegelund wrote:
> > This patch adds netdevs and phylink support for the ports in the switch.
> > It also adds register based injection and extraction for these ports.
> >
> > Frame DMA support for injection and extraction will be added in a later
> > series.
>
> > +struct net_device *sparx5_create_netdev(struct sparx5 *sparx5, u32 portno)
> > +{
> > +     struct sparx5_port *spx5_port;
> > +     struct net_device *ndev;
> > +     u64 val;
> > +
> > +     ndev = devm_alloc_etherdev(sparx5->dev, sizeof(struct sparx5_port));
> > +     if (!ndev)
> > +             return ERR_PTR(-ENOMEM);
> > +
> > +     SET_NETDEV_DEV(ndev, sparx5->dev);
> > +     spx5_port = netdev_priv(ndev);
> > +     spx5_port->ndev = ndev;
> > +     spx5_port->sparx5 = sparx5;
> > +     spx5_port->portno = portno;
> > +     sparx5_set_port_ifh(spx5_port->ifh, portno);
> > +
> > +     ndev->netdev_ops = &sparx5_port_netdev_ops;
> > +     ndev->features |= NETIF_F_LLTX; /* software tx */
>
> Is your transmission method really lockless? How does
> simultaneous Tx from two CPUs work?

Hmm, no that is a mistake. I can also see that NETIF_F_LLTX is not recommended, so I will remove
that.
>
> > +     val = ether_addr_to_u64(sparx5->base_mac) + portno + 1;
> > +     u64_to_ether_addr(val, ndev->dev_addr);
> > +
> > +     return ndev;
> > +}
>
> > +static void sparx5_xtr_grp(struct sparx5 *sparx5, u8 grp, bool byte_swap)
> > +{
> > +     bool eof_flag = false, pruned_flag = false, abort_flag = false;
> > +     struct net_device *netdev;
> > +     struct sparx5_port *port;
> > +     struct frame_info fi;
> > +     int i, byte_cnt = 0;
> > +     struct sk_buff *skb;
> > +     u32 ifh[IFH_LEN];
> > +     u32 *rxbuf;
> > +
> > +     /* Get IFH */
> > +     for (i = 0; i < IFH_LEN; i++)
> > +             ifh[i] = spx5_rd(sparx5, QS_XTR_RD(grp));
> > +
> > +     /* Decode IFH (whats needed) */
> > +     sparx5_ifh_parse(ifh, &fi);
> > +
> > +     /* Map to port netdev */
> > +     port = fi.src_port < SPX5_PORTS ?
> > +             sparx5->ports[fi.src_port] : NULL;
> > +     if (!port || !port->ndev) {
> > +             dev_err(sparx5->dev, "Data on inactive port %d\n", fi.src_port);
> > +             sparx5_xtr_flush(sparx5, grp);
> > +             return;
>
> You should probably increment appropriate counter for each error
> condition.

At this first check I do not have the netdev, so it will not be possible to update any counters, but
below I can use rx_dropped. Is that what you mean?

>
> > +     }
> > +
> > +     /* Have netdev, get skb */
> > +     netdev = port->ndev;
> > +     skb = netdev_alloc_skb(netdev, netdev->mtu + ETH_HLEN);
> > +     if (!skb) {
> > +             sparx5_xtr_flush(sparx5, grp);
> > +             dev_err(sparx5->dev, "No skb allocated\n");
> > +             return;
> > +     }
> > +     rxbuf = (u32 *)skb->data;
> > +
> > +     /* Now, pull frame data */
> > +     while (!eof_flag) {
> > +             u32 val = spx5_rd(sparx5, QS_XTR_RD(grp));
> > +             u32 cmp = val;
> > +
> > +             if (byte_swap)
> > +                     cmp = ntohl((__force __be32)val);
> > +
> > +             switch (cmp) {
> > +             case XTR_NOT_READY:
> > +                     break;
> > +             case XTR_ABORT:
> > +                     /* No accompanying data */
> > +                     abort_flag = true;
> > +                     eof_flag = true;
> > +                     break;
> > +             case XTR_EOF_0:
> > +             case XTR_EOF_1:
> > +             case XTR_EOF_2:
> > +             case XTR_EOF_3:
> > +                     /* This assumes STATUS_WORD_POS == 1, Status
> > +                      * just after last data
> > +                      */
> > +                     byte_cnt -= (4 - XTR_VALID_BYTES(val));
> > +                     eof_flag = true;
> > +                     break;
> > +             case XTR_PRUNED:
> > +                     /* But get the last 4 bytes as well */
> > +                     eof_flag = true;
> > +                     pruned_flag = true;
> > +                     fallthrough;
> > +             case XTR_ESCAPE:
> > +                     *rxbuf = spx5_rd(sparx5, QS_XTR_RD(grp));
> > +                     byte_cnt += 4;
> > +                     rxbuf++;
> > +                     break;
> > +             default:
> > +                     *rxbuf = val;
> > +                     byte_cnt += 4;
> > +                     rxbuf++;
> > +             }
> > +     }
> > +
> > +     if (abort_flag || pruned_flag || !eof_flag) {
> > +             netdev_err(netdev, "Discarded frame: abort:%d pruned:%d eof:%d\n",
> > +                        abort_flag, pruned_flag, eof_flag);
> > +             kfree_skb(skb);
> > +             return;
> > +     }
> > +
> > +#if defined(CONFIG_DEBUG_KERNEL) /* TODO: Remove before upstreaming */
> > +     if (!netif_oper_up(netdev)) {
> > +             netdev_err(netdev, "Discarded frame: Interface not up\n");
> > +             kfree_skb(skb);
> > +             return;
> > +     }
> > +#endif
> > +
> > +     /* Finish up skb */
> > +     skb_put(skb, byte_cnt - ETH_FCS_LEN);
> > +     eth_skb_pad(skb);
> > +     skb->protocol = eth_type_trans(skb, netdev);
> > +     netif_rx(skb);
> > +     netdev->stats.rx_bytes += skb->len;
> > +     netdev->stats.rx_packets++;
>
> Does the Rx really need to happen in an interrupt context?
> Did you consider using NAPI or a tasklet?

This register base injection and extraction is just preliminary. I have the next series waiting
with support for Frame DMA'ing and there I use NAPI, so if possible I would like to leave this as it
is, since it only a stopgap.

>
> > +}
> > +
> > +static int sparx5_inject(struct sparx5 *sparx5,
> > +                      u32 *ifh,
> > +                      struct sk_buff *skb)
> > +{
> > +     int grp = INJ_QUEUE;
> > +     u32 val, w, count;
> > +     u8 *buf;
> > +
> > +     val = spx5_rd(sparx5, QS_INJ_STATUS);
> > +     if (!(QS_INJ_STATUS_FIFO_RDY_GET(val) & BIT(grp))) {
> > +             pr_err("Injection: Queue not ready: 0x%lx\n",
> > +                    QS_INJ_STATUS_FIFO_RDY_GET(val));
>
> non-rate-limited errors on the datapath are a bad idea
>
> > +             return -EBUSY;

I will change that to ratelimited prints.

>
> What do you expect to happen at this point? Kernel can retry sending
> for ever, is there a way for the driver to find out that the fifo is
> no longer busy to stop/start the software queuing appropriately?

Hmm. I am not too familiar with the netdev queuing, but would this be a way forward?

1) In sparx5_inject: After injecting a frame then test for HW queue readiness and watermark levels,
and if there is a problem then call netif_queue_stop

2) Add an implementation of ndo_tx_timeout where the HW queue and Watermark level is checked and if
all is OK, then do a netif_wake_queue.

3) But if the HW queue and/or Watermark level is still not OK - then probably something went
seriously wrong, or the wait was to short. Will the ndo_tx_timeout be called again or is this a
one-off?

If the ndo_tx_timeout call is a one-off the driver would need to reset the HW queue system or even
deeper down...

>
> > +     }
> > +
> > +     if (QS_INJ_STATUS_WMARK_REACHED_GET(val) & BIT(grp)) {
> > +             pr_err("Injection: Watermark reached: 0x%lx\n",
> > +                    QS_INJ_STATUS_WMARK_REACHED_GET(val));
> > +             return -EBUSY;
>
> ditto

Yes.

>
> > +     }
> > +
> > +     /* Indicate SOF */
> > +     spx5_wr(QS_INJ_CTRL_SOF_SET(1) |
> > +             QS_INJ_CTRL_GAP_SIZE_SET(1),
> > +             sparx5, QS_INJ_CTRL(grp));
> > +
> > +     // Write the IFH to the chip.
>
> Why the mix of comment styles?

A mistake - I will update that.

>
> > +     for (w = 0; w < IFH_LEN; w++)
> > +             spx5_wr(ifh[w], sparx5, QS_INJ_WR(grp));
> > +
> > +     /* Write words, round up */
> > +     count = ((skb->len + 3) / 4);
>
> DIV_ROUND_UP()

I will use that instead.

>
> > +     buf = skb->data;
> > +     for (w = 0; w < count; w++, buf += 4) {
> > +             val = get_unaligned((const u32 *)buf);
> > +             spx5_wr(val, sparx5, QS_INJ_WR(grp));
> > +     }
> > +
> > +     /* Add padding */
> > +     while (w < (60 / 4)) {
> > +             spx5_wr(0, sparx5, QS_INJ_WR(grp));
> > +             w++;
> > +     }
> > +
> > +     /* Indicate EOF and valid bytes in last word */
> > +     spx5_wr(QS_INJ_CTRL_GAP_SIZE_SET(1) |
> > +             QS_INJ_CTRL_VLD_BYTES_SET(skb->len < 60 ? 0 : skb->len % 4) |
> > +             QS_INJ_CTRL_EOF_SET(1),
> > +             sparx5, QS_INJ_CTRL(grp));
> > +
> > +     /* Add dummy CRC */
> > +     spx5_wr(0, sparx5, QS_INJ_WR(grp));
> > +     w++;
> > +
> > +     return NETDEV_TX_OK;
> > +}