Merge branch 'dpaa_eth-rss'

Madalin Bucur says:

====================
Add RSS to DPAA 1.x Ethernet driver

This patch set introduces Receive Side Scaling for the DPAA Ethernet
driver. Documentation is updated with details related to the new
feature and limitations that apply.
Added also a small fix.

v2: removed a C++ style comment
v3: move struct fman to header file to avoid exporting a function
v4: addressed compilation issues introduced in v3
====================

Signed-off-by: David S. Miller <[email protected]>

Documentation/networking/dpaa.txt

@@ -13,6 +13,7 @@ Contents
 	- Configuring DPAA Ethernet in your kernel
 	- DPAA Ethernet Frame Processing
 	- DPAA Ethernet Features
+	- DPAA IRQ Affinity and Receive Side Scaling
 	- Debugging
 
 DPAA Ethernet Overview
@@ -147,7 +148,10 @@ gradually.
 
 The driver has Rx and Tx checksum offloading for UDP and TCP. Currently the Rx
 checksum offload feature is enabled by default and cannot be controlled through
-ethtool.
+ethtool. Also, rx-flow-hash and rx-hashing was added. The addition of RSS
+provides a big performance boost for the forwarding scenarios, allowing
+different traffic flows received by one interface to be processed by different
+CPUs in parallel.
 
 The driver has support for multiple prioritized Tx traffic classes. Priorities
 range from 0 (lowest) to 3 (highest). These are mapped to HW workqueues with
@@ -166,6 +170,68 @@ classes as follows:
 tc qdisc add dev <int> root handle 1: \
 	 mqprio num_tc 4 map 0 0 0 0 1 1 1 1 2 2 2 2 3 3 3 3 hw 1
 
+DPAA IRQ Affinity and Receive Side Scaling
+==========================================
+
+Traffic coming on the DPAA Rx queues or on the DPAA Tx confirmation
+queues is seen by the CPU as ingress traffic on a certain portal.
+The DPAA QMan portal interrupts are affined each to a certain CPU.
+The same portal interrupt services all the QMan portal consumers.
+
+By default the DPAA Ethernet driver enables RSS, making use of the
+DPAA FMan Parser and Keygen blocks to distribute traffic on 128
+hardware frame queues using a hash on IP v4/v6 source and destination
+and L4 source and destination ports, in present in the received frame.
+When RSS is disabled, all traffic received by a certain interface is
+received on the default Rx frame queue. The default DPAA Rx frame
+queues are configured to put the received traffic into a pool channel
+that allows any available CPU portal to dequeue the ingress traffic.
+The default frame queues have the HOLDACTIVE option set, ensuring that
+traffic bursts from a certain queue are serviced by the same CPU.
+This ensures a very low rate of frame reordering. A drawback of this
+is that only one CPU at a time can service the traffic received by a
+certain interface when RSS is not enabled.
+
+To implement RSS, the DPAA Ethernet driver allocates an extra set of
+128 Rx frame queues that are configured to dedicated channels, in a
+round-robin manner. The mapping of the frame queues to CPUs is now
+hardcoded, there is no indirection table to move traffic for a certain
+FQ (hash result) to another CPU. The ingress traffic arriving on one
+of these frame queues will arrive at the same portal and will always
+be processed by the same CPU. This ensures intra-flow order preservation
+and workload distribution for multiple traffic flows.
+
+RSS can be turned off for a certain interface using ethtool, i.e.
+
+	# ethtool -N fm1-mac9 rx-flow-hash tcp4 ""
+
+To turn it back on, one needs to set rx-flow-hash for tcp4/6 or udp4/6:
+
+	# ethtool -N fm1-mac9 rx-flow-hash udp4 sfdn
+
+There is no independent control for individual protocols, any command
+run for one of tcp4|udp4|ah4|esp4|sctp4|tcp6|udp6|ah6|esp6|sctp6 is
+going to control the rx-flow-hashing for all protocols on that interface.
+
+Besides using the FMan Keygen computed hash for spreading traffic on the
+128 Rx FQs, the DPAA Ethernet driver also sets the skb hash value when
+the NETIF_F_RXHASH feature is on (active by default). This can be turned
+on or off through ethtool, i.e.:
+
+	# ethtool -K fm1-mac9 rx-hashing off
+	# ethtool -k fm1-mac9 | grep hash
+	receive-hashing: off
+	# ethtool -K fm1-mac9 rx-hashing on
+	Actual changes:
+	receive-hashing: on
+	# ethtool -k fm1-mac9 | grep hash
+	receive-hashing: on
+
+Please note that Rx hashing depends upon the rx-flow-hashing being on
+for that interface - turning off rx-flow-hashing will also disable the
+rx-hashing (without ethtool reporting it as off as that depends on the
+NETIF_F_RXHASH feature flag).
+
 Debugging
 =========
 

drivers/net/ethernet/freescale/dpaa/dpaa_eth.c

@@ -158,7 +158,7 @@ MODULE_PARM_DESC(tx_timeout, "The Tx timeout in ms");
 #define DPAA_RX_PRIV_DATA_SIZE	(u16)(DPAA_TX_PRIV_DATA_SIZE + \
 					dpaa_rx_extra_headroom)
 
-#define DPAA_ETH_RX_QUEUES	128
+#define DPAA_ETH_PCD_RXQ_NUM	128
 
 #define DPAA_ENQUEUE_RETRIES	100000
 
@@ -169,6 +169,7 @@ struct fm_port_fqs {
 	struct dpaa_fq *tx_errq;
 	struct dpaa_fq *rx_defq;
 	struct dpaa_fq *rx_errq;
+	struct dpaa_fq *rx_pcdq;
 };
 
 /* All the dpa bps in use at any moment */
@@ -235,7 +236,7 @@ static int dpaa_netdev_init(struct net_device *net_dev,
 	net_dev->max_mtu = dpaa_get_max_mtu();
 
 	net_dev->hw_features |= (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM |
-				 NETIF_F_LLTX);
+				 NETIF_F_LLTX | NETIF_F_RXHASH);
 
 	net_dev->hw_features |= NETIF_F_SG | NETIF_F_HIGHDMA;
 	/* The kernels enables GSO automatically, if we declare NETIF_F_SG.
@@ -628,6 +629,7 @@ static inline void dpaa_assign_wq(struct dpaa_fq *fq, int idx)
 		fq->wq = 5;
 		break;
 	case FQ_TYPE_RX_DEFAULT:
+	case FQ_TYPE_RX_PCD:
 		fq->wq = 6;
 		break;
 	case FQ_TYPE_TX:
@@ -688,6 +690,7 @@ static int dpaa_alloc_all_fqs(struct device *dev, struct list_head *list,
 			      struct fm_port_fqs *port_fqs)
 {
 	struct dpaa_fq *dpaa_fq;
+	u32 fq_base, fq_base_aligned, i;
 
 	dpaa_fq = dpaa_fq_alloc(dev, 0, 1, list, FQ_TYPE_RX_ERROR);
 	if (!dpaa_fq)
@@ -701,6 +704,26 @@ static int dpaa_alloc_all_fqs(struct device *dev, struct list_head *list,
 
 	port_fqs->rx_defq = &dpaa_fq[0];
 
+	/* the PCD FQIDs range needs to be aligned for correct operation */
+	if (qman_alloc_fqid_range(&fq_base, 2 * DPAA_ETH_PCD_RXQ_NUM))
+		goto fq_alloc_failed;
+
+	fq_base_aligned = ALIGN(fq_base, DPAA_ETH_PCD_RXQ_NUM);
+
+	for (i = fq_base; i < fq_base_aligned; i++)
+		qman_release_fqid(i);
+
+	for (i = fq_base_aligned + DPAA_ETH_PCD_RXQ_NUM;
+	     i < (fq_base + 2 * DPAA_ETH_PCD_RXQ_NUM); i++)
+		qman_release_fqid(i);
+
+	dpaa_fq = dpaa_fq_alloc(dev, fq_base_aligned, DPAA_ETH_PCD_RXQ_NUM,
+				list, FQ_TYPE_RX_PCD);
+	if (!dpaa_fq)
+		goto fq_alloc_failed;
+
+	port_fqs->rx_pcdq = &dpaa_fq[0];
+
 	if (!dpaa_fq_alloc(dev, 0, DPAA_ETH_TXQ_NUM, list, FQ_TYPE_TX_CONF_MQ))
 		goto fq_alloc_failed;
 
@@ -870,13 +893,14 @@ static void dpaa_fq_setup(struct dpaa_priv *priv,
 			  const struct dpaa_fq_cbs *fq_cbs,
 			  struct fman_port *tx_port)
 {
-	int egress_cnt = 0, conf_cnt = 0, num_portals = 0, cpu;
+	int egress_cnt = 0, conf_cnt = 0, num_portals = 0, portal_cnt = 0, cpu;
 	const cpumask_t *affine_cpus = qman_affine_cpus();
-	u16 portals[NR_CPUS];
+	u16 channels[NR_CPUS];
 	struct dpaa_fq *fq;
 
 	for_each_cpu(cpu, affine_cpus)
-		portals[num_portals++] = qman_affine_channel(cpu);
+		channels[num_portals++] = qman_affine_channel(cpu);
+
 	if (num_portals == 0)
 		dev_err(priv->net_dev->dev.parent,
 			"No Qman software (affine) channels found");
@@ -890,6 +914,12 @@ static void dpaa_fq_setup(struct dpaa_priv *priv,
 		case FQ_TYPE_RX_ERROR:
 			dpaa_setup_ingress(priv, fq, &fq_cbs->rx_errq);
 			break;
+		case FQ_TYPE_RX_PCD:
+			if (!num_portals)
+				continue;
+			dpaa_setup_ingress(priv, fq, &fq_cbs->rx_defq);
+			fq->channel = channels[portal_cnt++ % num_portals];
+			break;
 		case FQ_TYPE_TX:
 			dpaa_setup_egress(priv, fq, tx_port,
 					  &fq_cbs->egress_ern);
@@ -1039,7 +1069,8 @@ static int dpaa_fq_init(struct dpaa_fq *dpaa_fq, bool td_enable)
 		/* Put all the ingress queues in our "ingress CGR". */
 		if (priv->use_ingress_cgr &&
 		    (dpaa_fq->fq_type == FQ_TYPE_RX_DEFAULT ||
-		     dpaa_fq->fq_type == FQ_TYPE_RX_ERROR)) {
+		     dpaa_fq->fq_type == FQ_TYPE_RX_ERROR ||
+		     dpaa_fq->fq_type == FQ_TYPE_RX_PCD)) {
 			initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_CGID);
 			initfq.fqd.fq_ctrl |= cpu_to_be16(QM_FQCTRL_CGE);
 			initfq.fqd.cgid = (u8)priv->ingress_cgr.cgrid;
@@ -1170,7 +1201,7 @@ static int dpaa_eth_init_tx_port(struct fman_port *port, struct dpaa_fq *errq,
 
 static int dpaa_eth_init_rx_port(struct fman_port *port, struct dpaa_bp **bps,
 				 size_t count, struct dpaa_fq *errq,
-				 struct dpaa_fq *defq,
+				 struct dpaa_fq *defq, struct dpaa_fq *pcdq,
 				 struct dpaa_buffer_layout *buf_layout)
 {
 	struct fman_buffer_prefix_content buf_prefix_content;
@@ -1190,6 +1221,10 @@ static int dpaa_eth_init_rx_port(struct fman_port *port, struct dpaa_bp **bps,
 	rx_p = &params.specific_params.rx_params;
 	rx_p->err_fqid = errq->fqid;
 	rx_p->dflt_fqid = defq->fqid;
+	if (pcdq) {
+		rx_p->pcd_base_fqid = pcdq->fqid;
+		rx_p->pcd_fqs_count = DPAA_ETH_PCD_RXQ_NUM;
+	}
 
 	count = min(ARRAY_SIZE(rx_p->ext_buf_pools.ext_buf_pool), count);
 	rx_p->ext_buf_pools.num_of_pools_used = (u8)count;
@@ -1234,7 +1269,8 @@ static int dpaa_eth_init_ports(struct mac_device *mac_dev,
 		return err;
 
 	err = dpaa_eth_init_rx_port(rxport, bps, count, port_fqs->rx_errq,
-				    port_fqs->rx_defq, &buf_layout[RX]);
+				    port_fqs->rx_defq, port_fqs->rx_pcdq,
+				    &buf_layout[RX]);
 
 	return err;
 }
@@ -2201,12 +2237,13 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
 	dma_addr_t addr = qm_fd_addr(fd);
 	enum qm_fd_format fd_format;
 	struct net_device *net_dev;
-	u32 fd_status;
+	u32 fd_status, hash_offset;
 	struct dpaa_bp *dpaa_bp;
 	struct dpaa_priv *priv;
 	unsigned int skb_len;
 	struct sk_buff *skb;
 	int *count_ptr;
+	void *vaddr;
 
 	fd_status = be32_to_cpu(fd->status);
 	fd_format = qm_fd_get_format(fd);
@@ -2252,7 +2289,8 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
 	dma_unmap_single(dpaa_bp->dev, addr, dpaa_bp->size, DMA_FROM_DEVICE);
 
 	/* prefetch the first 64 bytes of the frame or the SGT start */
-	prefetch(phys_to_virt(addr) + qm_fd_get_offset(fd));
+	vaddr = phys_to_virt(addr);
+	prefetch(vaddr + qm_fd_get_offset(fd));
 
 	fd_format = qm_fd_get_format(fd);
 	/* The only FD types that we may receive are contig and S/G */
@@ -2273,6 +2311,18 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
 
 	skb->protocol = eth_type_trans(skb, net_dev);
 
+	if (net_dev->features & NETIF_F_RXHASH && priv->keygen_in_use &&
+	    !fman_port_get_hash_result_offset(priv->mac_dev->port[RX],
+					      &hash_offset)) {
+		enum pkt_hash_types type;
+
+		/* if L4 exists, it was used in the hash generation */
+		type = be32_to_cpu(fd->status) & FM_FD_STAT_L4CV ?
+			PKT_HASH_TYPE_L4 : PKT_HASH_TYPE_L3;
+		skb_set_hash(skb, be32_to_cpu(*(u32 *)(vaddr + hash_offset)),
+			     type);
+	}
+
 	skb_len = skb->len;
 
 	if (unlikely(netif_receive_skb(skb) == NET_RX_DROP))
@@ -2511,6 +2561,9 @@ static struct dpaa_bp *dpaa_bp_alloc(struct device *dev)
 
 	dpaa_bp->bpid = FSL_DPAA_BPID_INV;
 	dpaa_bp->percpu_count = devm_alloc_percpu(dev, *dpaa_bp->percpu_count);
+	if (!dpaa_bp->percpu_count)
+		return ERR_PTR(-ENOMEM);
+
 	dpaa_bp->config_count = FSL_DPAA_ETH_MAX_BUF_COUNT;
 
 	dpaa_bp->seed_cb = dpaa_bp_seed;
@@ -2738,6 +2791,9 @@ static int dpaa_eth_probe(struct platform_device *pdev)
 	if (err)
 		goto init_ports_failed;
 
+	/* Rx traffic distribution based on keygen hashing defaults to on */
+	priv->keygen_in_use = true;
+
 	priv->percpu_priv = devm_alloc_percpu(dev, *priv->percpu_priv);
 	if (!priv->percpu_priv) {
 		dev_err(dev, "devm_alloc_percpu() failed\n");

drivers/net/ethernet/freescale/dpaa/dpaa_eth.h

@@ -52,6 +52,7 @@
 enum dpaa_fq_type {
 	FQ_TYPE_RX_DEFAULT = 1, /* Rx Default FQs */
 	FQ_TYPE_RX_ERROR,	/* Rx Error FQs */
+	FQ_TYPE_RX_PCD,		/* Rx Parse Classify Distribute FQs */
 	FQ_TYPE_TX,		/* "Real" Tx FQs */
 	FQ_TYPE_TX_CONFIRM,	/* Tx default Conf FQ (actually an Rx FQ) */
 	FQ_TYPE_TX_CONF_MQ,	/* Tx conf FQs (one for each Tx FQ) */
@@ -158,6 +159,7 @@ struct dpaa_priv {
 	struct list_head dpaa_fq_list;
 
 	u8 num_tc;
+	bool keygen_in_use;
 	u32 msg_enable;	/* net_device message level */
 
 	struct {

drivers/net/ethernet/freescale/dpaa/dpaa_eth_sysfs.c

@@ -71,6 +71,9 @@ static ssize_t dpaa_eth_show_fqids(struct device *dev,
 		case FQ_TYPE_RX_ERROR:
 			str = "Rx error";
 			break;
+		case FQ_TYPE_RX_PCD:
+			str = "Rx PCD";
+			break;
 		case FQ_TYPE_TX_CONFIRM:
 			str = "Tx default confirmation";
 			break;