On Mon, Apr 23, 2018 at 9:56 AM, Björn Töpel <bjorn.to...@gmail.com> wrote:
> From: Magnus Karlsson <magnus.karls...@intel.com>
>
> Here, Tx support is added. The user fills the Tx queue with frames to
> be sent by the kernel, and let's the kernel know using the sendmsg
> syscall.
>
> Signed-off-by: Magnus Karlsson <magnus.karls...@intel.com>
> +static int xsk_xmit_skb(struct sk_buff *skb)
This is basically packet_direct_xmit. Might be better to just move that
to net/core/dev.c and use in both AF_PACKET and AF_XDP.
Also, (eventually) AF_XDP may also want to support the regular path
through dev_queue_xmit to go through traffic shaping.
> +{
> + struct net_device *dev = skb->dev;
> + struct sk_buff *orig_skb = skb;
> + struct netdev_queue *txq;
> + int ret = NETDEV_TX_BUSY;
> + bool again = false;
> +
> + if (unlikely(!netif_running(dev) || !netif_carrier_ok(dev)))
> + goto drop;
> +
> + skb = validate_xmit_skb_list(skb, dev, &again);
> + if (skb != orig_skb)
> + return NET_XMIT_DROP;
Need to free generated segment list on error, see packet_direct_xmit.
> +
> + txq = skb_get_tx_queue(dev, skb);
> +
> + local_bh_disable();
> +
> + HARD_TX_LOCK(dev, txq, smp_processor_id());
> + if (!netif_xmit_frozen_or_drv_stopped(txq))
> + ret = netdev_start_xmit(skb, dev, txq, false);
> + HARD_TX_UNLOCK(dev, txq);
> +
> + local_bh_enable();
> +
> + if (!dev_xmit_complete(ret))
> + goto out_err;
> +
> + return ret;
> +drop:
> + atomic_long_inc(&dev->tx_dropped);
> +out_err:
> + return NET_XMIT_DROP;
> +}
> +static int xsk_generic_xmit(struct sock *sk, struct msghdr *m,
> + size_t total_len)
> +{
> + bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
> + u32 max_batch = TX_BATCH_SIZE;
> + struct xdp_sock *xs = xdp_sk(sk);
> + bool sent_frame = false;
> + struct xdp_desc desc;
> + struct sk_buff *skb;
> + int err = 0;
> +
> + if (unlikely(!xs->tx))
> + return -ENOBUFS;
> + if (need_wait)
> + return -EOPNOTSUPP;
> +
> + mutex_lock(&xs->mutex);
> +
> + while (xskq_peek_desc(xs->tx, &desc)) {
It is possible to pass a chain of skbs to validate_xmit_skb_list and
eventually pass this chain to xsk_xmit_skb, amortizing the cost of
taking the txq lock. Fine to ignore for this patch set.
> + char *buffer;
> + u32 id, len;
> +
> + if (max_batch-- == 0) {
> + err = -EAGAIN;
> + goto out;
> + }
> +
> + if (xskq_reserve_id(xs->umem->cq)) {
> + err = -EAGAIN;
> + goto out;
> + }
> +
> + len = desc.len;
> + if (unlikely(len > xs->dev->mtu)) {
> + err = -EMSGSIZE;
> + goto out;
> + }
> +
> + skb = sock_alloc_send_skb(sk, len, !need_wait, &err);
> + if (unlikely(!skb)) {
> + err = -EAGAIN;
> + goto out;
> + }
> +
> + skb_put(skb, len);
> + id = desc.idx;
> + buffer = xdp_umem_get_data(xs->umem, id) + desc.offset;
> + err = skb_store_bits(skb, 0, buffer, len);
> + if (unlikely(err))
> + goto out_store;
As xsk_destruct_skb delays notification until consume_skb is called, this
copy can be avoided by linking the xdp buffer into the skb frags array,
analogous to tpacket_snd.
You probably don't care much about the copy slow path, and this can be
implemented later, so also no need to do in this patchset.
static inline struct xdp_desc *xskq_peek_desc(struct xsk_queue *q,
+ struct xdp_desc *desc)
+{
+ struct xdp_rxtx_ring *ring;
+
+ if (q->cons_tail == q->cons_head) {
+ WRITE_ONCE(q->ring->consumer, q->cons_tail);
+ q->cons_head = q->cons_tail + xskq_nb_avail(q, RX_BATCH_SIZE);
+
+ /* Order consumer and data */
+ smp_rmb();
+
+ return xskq_validate_desc(q, desc);
+ }
+
+ ring = (struct xdp_rxtx_ring *)q->ring;
+ *desc = ring->desc[q->cons_tail & q->ring_mask];
+ return desc;
This only validates descriptors if taking the branch.
