net: dpaa2-eth: AF_XDP TX zero copy support

Add support in dpaa2-eth for packet processing on the Tx path using
AF_XDP zero copy mode.

The newly added dpaa2_xsk_tx() function will handle enqueuing AF_XDP Tx
packets into the appropriate queue and update any necessary statistics.

On a more detailed note, the dpaa2_xsk_tx_build_fd() function handles
creating a Scatter-Gather frame descriptor with only one data buffer.
This is needed because otherwise we would need to impose a headroom in
the Tx buffer to store our software annotation structures.
This tactic is already used on the normal data path of the dpaa2-eth
driver, thus we are reusing the dpaa2_eth_sgt_get/dpaa2_eth_sgt_recycle
functions in order to allocate and recycle the Scatter-Gather table
buffers.

In case we have reached the maximum number of Tx XSK packets to be sent
in a NAPI cycle, we'll exit the dpaa2_eth_poll() and hope to be
rescheduled again.

On the XSK Tx confirmation path, we are just unmapping the SGT buffer
and recycle it for further use.
Signed-off-by: Robert-Ionut Alexa <robert-ionut.alexa@nxp.com>
Signed-off-by: Ioana Ciornei <ioana.ciornei@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c

@@ -858,7 +858,7 @@ static void dpaa2_eth_enable_tx_tstamp(struct dpaa2_eth_priv *priv,
 	}
 }
 
-static void *dpaa2_eth_sgt_get(struct dpaa2_eth_priv *priv)
+void *dpaa2_eth_sgt_get(struct dpaa2_eth_priv *priv)
 {
 	struct dpaa2_eth_sgt_cache *sgt_cache;
 	void *sgt_buf = NULL;
@@ -880,7 +880,7 @@ static void *dpaa2_eth_sgt_get(struct dpaa2_eth_priv *priv)
 	return sgt_buf;
 }
 
-static void dpaa2_eth_sgt_recycle(struct dpaa2_eth_priv *priv, void *sgt_buf)
+void dpaa2_eth_sgt_recycle(struct dpaa2_eth_priv *priv, void *sgt_buf)
 {
 	struct dpaa2_eth_sgt_cache *sgt_cache;
 
@@ -1115,9 +1115,10 @@ static int dpaa2_eth_build_single_fd(struct dpaa2_eth_priv *priv,
  * This can be called either from dpaa2_eth_tx_conf() or on the error path of
  * dpaa2_eth_tx().
  */
-static void dpaa2_eth_free_tx_fd(struct dpaa2_eth_priv *priv,
-				 struct dpaa2_eth_fq *fq,
-				 const struct dpaa2_fd *fd, bool in_napi)
+void dpaa2_eth_free_tx_fd(struct dpaa2_eth_priv *priv,
+			  struct dpaa2_eth_channel *ch,
+			  struct dpaa2_eth_fq *fq,
+			  const struct dpaa2_fd *fd, bool in_napi)
 {
 	struct device *dev = priv->net_dev->dev.parent;
 	dma_addr_t fd_addr, sg_addr;
@@ -1184,6 +1185,10 @@ static void dpaa2_eth_free_tx_fd(struct dpaa2_eth_priv *priv,
 
 			if (!swa->tso.is_last_fd)
 				should_free_skb = 0;
+		} else if (swa->type == DPAA2_ETH_SWA_XSK) {
+			/* Unmap the SGT Buffer */
+			dma_unmap_single(dev, fd_addr, swa->xsk.sgt_size,
+					 DMA_BIDIRECTIONAL);
 		} else {
 			skb = swa->single.skb;
 
@@ -1201,6 +1206,12 @@ static void dpaa2_eth_free_tx_fd(struct dpaa2_eth_priv *priv,
 		return;
 	}
 
+	if (swa->type == DPAA2_ETH_SWA_XSK) {
+		ch->xsk_tx_pkts_sent++;
+		dpaa2_eth_sgt_recycle(priv, buffer_start);
+		return;
+	}
+
 	if (swa->type != DPAA2_ETH_SWA_XDP && in_napi) {
 		fq->dq_frames++;
 		fq->dq_bytes += fd_len;
@@ -1375,7 +1386,7 @@ static int dpaa2_eth_build_gso_fd(struct dpaa2_eth_priv *priv,
 err_sgt_get:
 	/* Free all the other FDs that were already fully created */
 	for (i = 0; i < index; i++)
-		dpaa2_eth_free_tx_fd(priv, NULL, &fd_start[i], false);
+		dpaa2_eth_free_tx_fd(priv, NULL, NULL, &fd_start[i], false);
 
 	return err;
 }
@@ -1491,7 +1502,7 @@ static netdev_tx_t __dpaa2_eth_tx(struct sk_buff *skb,
 	if (unlikely(err < 0)) {
 		percpu_stats->tx_errors++;
 		/* Clean up everything, including freeing the skb */
-		dpaa2_eth_free_tx_fd(priv, fq, fd, false);
+		dpaa2_eth_free_tx_fd(priv, NULL, fq, fd, false);
 		netdev_tx_completed_queue(nq, 1, fd_len);
 	} else {
 		percpu_stats->tx_packets += total_enqueued;
@@ -1584,7 +1595,7 @@ static void dpaa2_eth_tx_conf(struct dpaa2_eth_priv *priv,
 
 	/* Check frame errors in the FD field */
 	fd_errors = dpaa2_fd_get_ctrl(fd) & DPAA2_FD_TX_ERR_MASK;
-	dpaa2_eth_free_tx_fd(priv, fq, fd, true);
+	dpaa2_eth_free_tx_fd(priv, ch, fq, fd, true);
 
 	if (likely(!fd_errors))
 		return;
@@ -1929,6 +1940,7 @@ static int dpaa2_eth_poll(struct napi_struct *napi, int budget)
 	struct dpaa2_eth_fq *fq, *txc_fq = NULL;
 	struct netdev_queue *nq;
 	int store_cleaned, work_done;
+	bool work_done_zc = false;
 	struct list_head rx_list;
 	int retries = 0;
 	u16 flowid;
@@ -1941,6 +1953,15 @@ static int dpaa2_eth_poll(struct napi_struct *napi, int budget)
 	INIT_LIST_HEAD(&rx_list);
 	ch->rx_list = &rx_list;
 
+	if (ch->xsk_zc) {
+		work_done_zc = dpaa2_xsk_tx(priv, ch);
+		/* If we reached the XSK Tx per NAPI threshold, we're done */
+		if (work_done_zc) {
+			work_done = budget;
+			goto out;
+		}
+	}
+
 	do {
 		err = dpaa2_eth_pull_channel(ch);
 		if (unlikely(err))
@@ -1993,6 +2014,11 @@ static int dpaa2_eth_poll(struct napi_struct *napi, int budget)
 out:
 	netif_receive_skb_list(ch->rx_list);
 
+	if (ch->xsk_tx_pkts_sent) {
+		xsk_tx_completed(ch->xsk_pool, ch->xsk_tx_pkts_sent);
+		ch->xsk_tx_pkts_sent = 0;
+	}
+
 	if (txc_fq && txc_fq->dq_frames) {
 		nq = netdev_get_tx_queue(priv->net_dev, txc_fq->flowid);
 		netdev_tx_completed_queue(nq, txc_fq->dq_frames,
@@ -2989,7 +3015,11 @@ static void dpaa2_eth_cdan_cb(struct dpaa2_io_notification_ctx *ctx)
 	/* Update NAPI statistics */
 	ch->stats.cdan++;
 
-	napi_schedule(&ch->napi);
+	/* NAPI can also be scheduled from the AF_XDP Tx path. Mark a missed
+	 * so that it can be rescheduled again.
+	 */
+	if (!napi_if_scheduled_mark_missed(&ch->napi))
+		napi_schedule(&ch->napi);
 }
 
 /* Allocate and configure a DPCON object */

drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h

@@ -53,6 +53,12 @@
  */
 #define DPAA2_ETH_TXCONF_PER_NAPI	256
 
+/* Maximum number of Tx frames to be processed in a single NAPI
+ * call when AF_XDP is running. Bind it to DPAA2_ETH_TXCONF_PER_NAPI
+ * to maximize the throughput.
+ */
+#define DPAA2_ETH_TX_ZC_PER_NAPI	DPAA2_ETH_TXCONF_PER_NAPI
+
 /* Buffer qouta per channel. We want to keep in check number of ingress frames
  * in flight: for small sized frames, congestion group taildrop may kick in
  * first; for large sizes, Rx FQ taildrop threshold will ensure only a
@@ -154,6 +160,7 @@ struct dpaa2_eth_swa {
 		} xdp;
 		struct {
 			struct xdp_buff *xdp_buff;
+			int sgt_size;
 		} xsk;
 		struct {
 			struct sk_buff *skb;
@@ -495,6 +502,7 @@ struct dpaa2_eth_channel {
 	int recycled_bufs_cnt;
 
 	bool xsk_zc;
+	int xsk_tx_pkts_sent;
 	struct xsk_buff_pool *xsk_pool;
 	struct dpaa2_eth_bp *bp;
 };
@@ -531,7 +539,7 @@ struct dpaa2_eth_trap_data {
 
 #define DPAA2_ETH_DEFAULT_COPYBREAK	512
 
-#define DPAA2_ETH_ENQUEUE_MAX_FDS	200
+#define DPAA2_ETH_ENQUEUE_MAX_FDS	256
 struct dpaa2_eth_fds {
 	struct dpaa2_fd array[DPAA2_ETH_ENQUEUE_MAX_FDS];
 };
@@ -836,4 +844,16 @@ void dpaa2_eth_xdp_enqueue(struct dpaa2_eth_priv *priv,
 int dpaa2_xsk_wakeup(struct net_device *dev, u32 qid, u32 flags);
 int dpaa2_xsk_setup_pool(struct net_device *dev, struct xsk_buff_pool *pool, u16 qid);
 
+void dpaa2_eth_free_tx_fd(struct dpaa2_eth_priv *priv,
+			  struct dpaa2_eth_channel *ch,
+			  struct dpaa2_eth_fq *fq,
+			  const struct dpaa2_fd *fd, bool in_napi);
+bool dpaa2_xsk_tx(struct dpaa2_eth_priv *priv,
+		  struct dpaa2_eth_channel *ch);
+
+/* SGT (Scatter-Gather Table) cache management */
+void *dpaa2_eth_sgt_get(struct dpaa2_eth_priv *priv);
+
+void dpaa2_eth_sgt_recycle(struct dpaa2_eth_priv *priv, void *sgt_buf);
+
 #endif	/* __DPAA2_H */

drivers/net/ethernet/freescale/dpaa2/dpaa2-xsk.c

@@ -196,6 +196,7 @@ static int dpaa2_xsk_disable_pool(struct net_device *dev, u16 qid)
 
 	ch->xsk_zc = false;
 	ch->xsk_pool = NULL;
+	ch->xsk_tx_pkts_sent = 0;
 	ch->bp = priv->bp[DPAA2_ETH_DEFAULT_BP_IDX];
 
 	dpaa2_eth_setup_consume_func(priv, ch, DPAA2_RX_FQ, dpaa2_eth_rx);
@@ -325,3 +326,125 @@ int dpaa2_xsk_wakeup(struct net_device *dev, u32 qid, u32 flags)
 
 	return 0;
 }
+
+static int dpaa2_xsk_tx_build_fd(struct dpaa2_eth_priv *priv,
+				 struct dpaa2_eth_channel *ch,
+				 struct dpaa2_fd *fd,
+				 struct xdp_desc *xdp_desc)
+{
+	struct device *dev = priv->net_dev->dev.parent;
+	struct dpaa2_sg_entry *sgt;
+	struct dpaa2_eth_swa *swa;
+	void *sgt_buf = NULL;
+	dma_addr_t sgt_addr;
+	int sgt_buf_size;
+	dma_addr_t addr;
+	int err = 0;
+
+	/* Prepare the HW SGT structure */
+	sgt_buf_size = priv->tx_data_offset + sizeof(struct dpaa2_sg_entry);
+	sgt_buf = dpaa2_eth_sgt_get(priv);
+	if (unlikely(!sgt_buf))
+		return -ENOMEM;
+	sgt = (struct dpaa2_sg_entry *)(sgt_buf + priv->tx_data_offset);
+
+	/* Get the address of the XSK Tx buffer */
+	addr = xsk_buff_raw_get_dma(ch->xsk_pool, xdp_desc->addr);
+	xsk_buff_raw_dma_sync_for_device(ch->xsk_pool, addr, xdp_desc->len);
+
+	/* Fill in the HW SGT structure */
+	dpaa2_sg_set_addr(sgt, addr);
+	dpaa2_sg_set_len(sgt, xdp_desc->len);
+	dpaa2_sg_set_final(sgt, true);
+
+	/* Store the necessary info in the SGT buffer */
+	swa = (struct dpaa2_eth_swa *)sgt_buf;
+	swa->type = DPAA2_ETH_SWA_XSK;
+	swa->xsk.sgt_size = sgt_buf_size;
+
+	/* Separately map the SGT buffer */
+	sgt_addr = dma_map_single(dev, sgt_buf, sgt_buf_size, DMA_BIDIRECTIONAL);
+	if (unlikely(dma_mapping_error(dev, sgt_addr))) {
+		err = -ENOMEM;
+		goto sgt_map_failed;
+	}
+
+	/* Initialize FD fields */
+	memset(fd, 0, sizeof(struct dpaa2_fd));
+	dpaa2_fd_set_offset(fd, priv->tx_data_offset);
+	dpaa2_fd_set_format(fd, dpaa2_fd_sg);
+	dpaa2_fd_set_addr(fd, sgt_addr);
+	dpaa2_fd_set_len(fd, xdp_desc->len);
+	dpaa2_fd_set_ctrl(fd, FD_CTRL_PTA);
+
+	return 0;
+
+sgt_map_failed:
+	dpaa2_eth_sgt_recycle(priv, sgt_buf);
+
+	return err;
+}
+
+bool dpaa2_xsk_tx(struct dpaa2_eth_priv *priv,
+		  struct dpaa2_eth_channel *ch)
+{
+	struct xdp_desc *xdp_descs = ch->xsk_pool->tx_descs;
+	struct dpaa2_eth_drv_stats *percpu_extras;
+	struct rtnl_link_stats64 *percpu_stats;
+	int budget = DPAA2_ETH_TX_ZC_PER_NAPI;
+	int total_enqueued, enqueued;
+	int retries, max_retries;
+	struct dpaa2_eth_fq *fq;
+	struct dpaa2_fd *fds;
+	int batch, i, err;
+
+	percpu_stats = this_cpu_ptr(priv->percpu_stats);
+	percpu_extras = this_cpu_ptr(priv->percpu_extras);
+	fds = (this_cpu_ptr(priv->fd))->array;
+
+	/* Use the FQ with the same idx as the affine CPU */
+	fq = &priv->fq[ch->nctx.desired_cpu];
+
+	batch = xsk_tx_peek_release_desc_batch(ch->xsk_pool, budget);
+	if (!batch)
+		return false;
+
+	/* Create a FD for each XSK frame to be sent */
+	for (i = 0; i < batch; i++) {
+		err = dpaa2_xsk_tx_build_fd(priv, ch, &fds[i], &xdp_descs[i]);
+		if (err) {
+			batch = i;
+			break;
+		}
+	}
+
+	/* Enqueue all the created FDs */
+	max_retries = batch * DPAA2_ETH_ENQUEUE_RETRIES;
+	total_enqueued = 0;
+	enqueued = 0;
+	retries = 0;
+	while (total_enqueued < batch && retries < max_retries) {
+		err = priv->enqueue(priv, fq, &fds[total_enqueued], 0,
+				    batch - total_enqueued, &enqueued);
+		if (err == -EBUSY) {
+			retries++;
+			continue;
+		}
+
+		total_enqueued += enqueued;
+	}
+	percpu_extras->tx_portal_busy += retries;
+
+	/* Update statistics */
+	percpu_stats->tx_packets += total_enqueued;
+	for (i = 0; i < total_enqueued; i++)
+		percpu_stats->tx_bytes += dpaa2_fd_get_len(&fds[i]);
+	for (i = total_enqueued; i < batch; i++) {
+		dpaa2_eth_free_tx_fd(priv, ch, fq, &fds[i], false);
+		percpu_stats->tx_errors++;
+	}
+
+	xsk_tx_release(ch->xsk_pool);
+
+	return total_enqueued == budget ? true : false;
+}