[dpdk-dev] [PATCH v4 08/12] vhost: support multiple queues

[Ouyang Changchun <changchun.ouyang@intel.com>]

Sample vhost leverage the VMDq+RSS in HW to receive packets and distribute them
into different queue in the pool according to 5 tuples.

And enable multiple queues mode in vhost/virtio layer.

HW queue numbers in pool exactly same with the queue number in virtio device,
e.g. rxq = 4, the queue number is 4, it means 4 HW queues in each VMDq pool,
and 4 queues in each virtio device/port, one maps to each.

=========================================
==================|   |==================|
       vport0     |   |      vport1      |
---  ---  ---  ---|   |---  ---  ---  ---|
q0 | q1 | q2 | q3 |   |q0 | q1 | q2 | q3 |
/\= =/\= =/\= =/\=|   |/\= =/\= =/\= =/\=|
||   ||   ||   ||      ||   ||   ||   ||
||   ||   ||   ||      ||   ||   ||   ||
||= =||= =||= =||=|   =||== ||== ||== ||=|
q0 | q1 | q2 | q3 |   |q0 | q1 | q2 | q3 |

------------------|   |------------------|
     VMDq pool0   |   |    VMDq pool1    |
==================|   |==================|

In RX side, it firstly polls each queue of the pool and gets the packets from
it and enqueue them into its corresponding queue in virtio device/port.
In TX side, it dequeue packets from each queue of virtio device/port and send
to either physical port or another virtio device according to its destination
MAC address.

Signed-off-by: Changchun Ouyang <changchun.ouyang@intel.com>
---
Changes in v4:
  - address comments and refine var name
  - support FVL nic
  - fix check patch issue

Changes in v2:
  - check queue num per pool in VMDq and queue pair number per vhost device
  - remove the unnecessary calling q_num_set api
  - fix checkpatch errors

 examples/vhost/main.c | 190 ++++++++++++++++++++++++++++++++------------------
 1 file changed, 124 insertions(+), 66 deletions(-)

diff --git a/examples/vhost/main.c b/examples/vhost/main.c
index 5b811af..683a300 100644
--- a/examples/vhost/main.c
+++ b/examples/vhost/main.c
@@ -368,6 +368,37 @@ validate_num_devices(uint32_t max_nb_devices)
 	return 0;
 }
 
+static int
+get_dev_nb_for_82599(struct rte_eth_dev_info dev_info)
+{
+	int dev_nb = -1;
+	switch (rxq) {
+	case 1:
+	case 2:
+		/*
+		 * for 82599, dev_info.max_vmdq_pools always 64 dispite rx mode.
+		 */
+		dev_nb = (int)dev_info.max_vmdq_pools;
+		break;
+	case 4:
+		dev_nb = (int)dev_info.max_vmdq_pools / 2;
+		break;
+	default:
+		RTE_LOG(ERR, VHOST_CONFIG, "rxq invalid for VMDq.\n");
+	}
+	return dev_nb;
+}
+
+static int
+get_dev_nb_for_fvl(struct rte_eth_dev_info dev_info)
+{
+	/*
+	 * for FVL, dev_info.max_vmdq_pools is calculated according to
+	 * the configured value: CONFIG_RTE_LIBRTE_I40E_QUEUE_NUM_PER_VM.
+	 */
+	return (int)dev_info.max_vmdq_pools;
+}
+
 /*
  * Initialises a given port using global settings and with the rx buffers
  * coming from the mbuf_pool passed as parameter
@@ -412,17 +443,14 @@ port_init(uint8_t port)
 	}
 
 	/* Configure the virtio devices num based on VMDQ limits */
-	switch (rxq) {
-	case 1:
-	case 2:
-		num_devices = dev_info.max_vmdq_pools;
-		break;
-	case 4:
-		num_devices = dev_info.max_vmdq_pools / 2;
-		break;
-	default:
-		RTE_LOG(ERR, VHOST_CONFIG, "rxq invalid for VMDq.\n");
-		return -1;
+	if (dev_info.max_vmdq_pools == ETH_64_POOLS) {
+		num_devices = (uint32_t)get_dev_nb_for_82599(dev_info);
+		if (num_devices == (uint32_t)-1)
+			return -1;
+	} else {
+		num_devices = (uint32_t)get_dev_nb_for_fvl(dev_info);
+		if (num_devices == (uint32_t)-1)
+			return -1;
 	}
 
 	if (zero_copy) {
@@ -1001,8 +1029,9 @@ link_vmdq(struct vhost_dev *vdev, struct rte_mbuf *m)
 
 	/* Enable stripping of the vlan tag as we handle routing. */
 	if (vlan_strip)
-		rte_eth_dev_set_vlan_strip_on_queue(ports[0],
-			(uint16_t)vdev->vmdq_rx_q, 1);
+		for (i = 0; i < (int)rxq; i++)
+			rte_eth_dev_set_vlan_strip_on_queue(ports[0],
+				(uint16_t)(vdev->vmdq_rx_q + i), 1);
 
 	/* Set device as ready for RX. */
 	vdev->ready = DEVICE_RX;
@@ -1017,7 +1046,7 @@ link_vmdq(struct vhost_dev *vdev, struct rte_mbuf *m)
 static inline void
 unlink_vmdq(struct vhost_dev *vdev)
 {
-	unsigned i = 0;
+	unsigned i = 0, j = 0;
 	unsigned rx_count;
 	struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
 
@@ -1030,15 +1059,19 @@ unlink_vmdq(struct vhost_dev *vdev)
 		vdev->vlan_tag = 0;
 
 		/*Clear out the receive buffers*/
-		rx_count = rte_eth_rx_burst(ports[0],
-					(uint16_t)vdev->vmdq_rx_q, pkts_burst, MAX_PKT_BURST);
+		for (i = 0; i < rxq; i++) {
+			rx_count = rte_eth_rx_burst(ports[0],
+					(uint16_t)vdev->vmdq_rx_q + i,
+					pkts_burst, MAX_PKT_BURST);
 
-		while (rx_count) {
-			for (i = 0; i < rx_count; i++)
-				rte_pktmbuf_free(pkts_burst[i]);
+			while (rx_count) {
+				for (j = 0; j < rx_count; j++)
+					rte_pktmbuf_free(pkts_burst[j]);
 
-			rx_count = rte_eth_rx_burst(ports[0],
-					(uint16_t)vdev->vmdq_rx_q, pkts_burst, MAX_PKT_BURST);
+				rx_count = rte_eth_rx_burst(ports[0],
+					(uint16_t)vdev->vmdq_rx_q + i,
+					pkts_burst, MAX_PKT_BURST);
+			}
 		}
 
 		vdev->ready = DEVICE_MAC_LEARNING;
@@ -1050,7 +1083,7 @@ unlink_vmdq(struct vhost_dev *vdev)
  * the packet on that devices RX queue. If not then return.
  */
 static inline int __attribute__((always_inline))
-virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
+virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m, uint32_t q_idx)
 {
 	struct virtio_net_data_ll *dev_ll;
 	struct ether_hdr *pkt_hdr;
@@ -1065,7 +1098,7 @@ virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
 
 	while (dev_ll != NULL) {
 		if ((dev_ll->vdev->ready == DEVICE_RX) && ether_addr_cmp(&(pkt_hdr->d_addr),
-				          &dev_ll->vdev->mac_address)) {
+					&dev_ll->vdev->mac_address)) {
 
 			/* Drop the packet if the TX packet is destined for the TX device. */
 			if (dev_ll->vdev->dev->device_fh == dev->device_fh) {
@@ -1083,7 +1116,9 @@ virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
 				LOG_DEBUG(VHOST_DATA, "(%"PRIu64") Device is marked for removal\n", tdev->device_fh);
 			} else {
 				/*send the packet to the local virtio device*/
-				ret = rte_vhost_enqueue_burst(tdev, VIRTIO_RXQ, &m, 1);
+				ret = rte_vhost_enqueue_burst(tdev,
+					VIRTIO_RXQ + q_idx * VIRTIO_QNUM,
+					&m, 1);
 				if (enable_stats) {
 					rte_atomic64_add(
 					&dev_statistics[tdev->device_fh].rx_total_atomic,
@@ -1160,7 +1195,8 @@ find_local_dest(struct virtio_net *dev, struct rte_mbuf *m,
  * or the physical port.
  */
 static inline void __attribute__((always_inline))
-virtio_tx_route(struct vhost_dev *vdev, struct rte_mbuf *m, uint16_t vlan_tag)
+virtio_tx_route(struct vhost_dev *vdev, struct rte_mbuf *m,
+		uint16_t vlan_tag, uint32_t q_idx)
 {
 	struct mbuf_table *tx_q;
 	struct rte_mbuf **m_table;
@@ -1170,7 +1206,8 @@ virtio_tx_route(struct vhost_dev *vdev, struct rte_mbuf *m, uint16_t vlan_tag)
 	struct ether_hdr *nh;
 
 	/*check if destination is local VM*/
-	if ((vm2vm_mode == VM2VM_SOFTWARE) && (virtio_tx_local(vdev, m) == 0)) {
+	if ((vm2vm_mode == VM2VM_SOFTWARE) &&
+		(virtio_tx_local(vdev, m, q_idx) == 0)) {
 		rte_pktmbuf_free(m);
 		return;
 	}
@@ -1334,49 +1371,60 @@ switch_worker(__attribute__((unused)) void *arg)
 			}
 			if (likely(vdev->ready == DEVICE_RX)) {
 				/*Handle guest RX*/
-				rx_count = rte_eth_rx_burst(ports[0],
-					vdev->vmdq_rx_q, pkts_burst, MAX_PKT_BURST);
+				for (i = 0; i < rxq; i++) {
+					rx_count = rte_eth_rx_burst(ports[0],
+						vdev->vmdq_rx_q + i, pkts_burst, MAX_PKT_BURST);
 
-				if (rx_count) {
-					/*
-					* Retry is enabled and the queue is full then we wait and retry to avoid packet loss
-					* Here MAX_PKT_BURST must be less than virtio queue size
-					*/
-					if (enable_retry && unlikely(rx_count > rte_vring_available_entries(dev, VIRTIO_RXQ))) {
-						for (retry = 0; retry < burst_rx_retry_num; retry++) {
-							rte_delay_us(burst_rx_delay_time);
-							if (rx_count <= rte_vring_available_entries(dev, VIRTIO_RXQ))
-								break;
+					if (rx_count) {
+						/*
+						* Retry is enabled and the queue is full then we wait and retry to avoid packet loss
+						* Here MAX_PKT_BURST must be less than virtio queue size
+						*/
+						if (enable_retry && unlikely(rx_count > rte_vring_available_entries(dev,
+											VIRTIO_RXQ + i * VIRTIO_QNUM))) {
+							for (retry = 0; retry < burst_rx_retry_num; retry++) {
+								rte_delay_us(burst_rx_delay_time);
+								if (rx_count <= rte_vring_available_entries(dev,
+											VIRTIO_RXQ + i * VIRTIO_QNUM))
+									break;
+							}
+						}
+						ret_count = rte_vhost_enqueue_burst(dev, VIRTIO_RXQ + i * VIRTIO_QNUM,
+											pkts_burst, rx_count);
+						if (enable_stats) {
+							rte_atomic64_add(
+							&dev_statistics[dev_ll->vdev->dev->device_fh].rx_total_atomic,
+							rx_count);
+							rte_atomic64_add(
+							&dev_statistics[dev_ll->vdev->dev->device_fh].rx_atomic, ret_count);
+						}
+						while (likely(rx_count)) {
+							rx_count--;
+							rte_pktmbuf_free(pkts_burst[rx_count]);
 						}
 					}
-					ret_count = rte_vhost_enqueue_burst(dev, VIRTIO_RXQ, pkts_burst, rx_count);
-					if (enable_stats) {
-						rte_atomic64_add(
-						&dev_statistics[dev_ll->vdev->dev->device_fh].rx_total_atomic,
-						rx_count);
-						rte_atomic64_add(
-						&dev_statistics[dev_ll->vdev->dev->device_fh].rx_atomic, ret_count);
-					}
-					while (likely(rx_count)) {
-						rx_count--;
-						rte_pktmbuf_free(pkts_burst[rx_count]);
-					}
-
 				}
 			}
 
 			if (likely(!vdev->remove)) {
 				/* Handle guest TX*/
-				tx_count = rte_vhost_dequeue_burst(dev, VIRTIO_TXQ, mbuf_pool, pkts_burst, MAX_PKT_BURST);
-				/* If this is the first received packet we need to learn the MAC and setup VMDQ */
-				if (unlikely(vdev->ready == DEVICE_MAC_LEARNING) && tx_count) {
-					if (vdev->remove || (link_vmdq(vdev, pkts_burst[0]) == -1)) {
-						while (tx_count)
-							rte_pktmbuf_free(pkts_burst[--tx_count]);
+				for (i = 0; i < rxq; i++) {
+					tx_count = rte_vhost_dequeue_burst(dev, VIRTIO_TXQ + i * VIRTIO_QNUM,
+							mbuf_pool, pkts_burst, MAX_PKT_BURST);
+					/*
+					 * If this is the first received packet we need to learn
+					 * the MAC and setup VMDQ
+					 */
+					if (unlikely(vdev->ready == DEVICE_MAC_LEARNING) && tx_count) {
+						if (vdev->remove || (link_vmdq(vdev, pkts_burst[0]) == -1)) {
+							while (tx_count)
+								rte_pktmbuf_free(pkts_burst[--tx_count]);
+						}
 					}
+					while (tx_count)
+						virtio_tx_route(vdev, pkts_burst[--tx_count],
+								(uint16_t)dev->device_fh, i);
 				}
-				while (tx_count)
-					virtio_tx_route(vdev, pkts_burst[--tx_count], (uint16_t)dev->device_fh);
 			}
 
 			/*move to the next device in the list*/
@@ -2634,6 +2682,14 @@ new_device (struct virtio_net *dev)
 	uint32_t device_num_min = num_devices;
 	struct vhost_dev *vdev;
 	uint32_t regionidx;
+	uint32_t i;
+
+	if ((rxq > 1) && (dev->virt_qp_nb != rxq)) {
+		RTE_LOG(ERR, VHOST_DATA, "(%"PRIu64") queue num in VMDq pool:"
+			"%d != queue pair num in vhost dev:%d\n",
+			dev->device_fh, rxq, dev->virt_qp_nb);
+		return -1;
+	}
 
 	vdev = rte_zmalloc("vhost device", sizeof(*vdev), RTE_CACHE_LINE_SIZE);
 	if (vdev == NULL) {
@@ -2680,12 +2736,12 @@ new_device (struct virtio_net *dev)
 		}
 	}
 
-
 	/* Add device to main ll */
 	ll_dev = get_data_ll_free_entry(&ll_root_free);
 	if (ll_dev == NULL) {
-		RTE_LOG(INFO, VHOST_DATA, "(%"PRIu64") No free entry found in linked list. Device limit "
-			"of %d devices per core has been reached\n",
+		RTE_LOG(INFO, VHOST_DATA,
+			"(%"PRIu64") No free entry found in linked list."
+			"Device limit of %d devices per core has been reached\n",
 			dev->device_fh, num_devices);
 		if (vdev->regions_hpa)
 			rte_free(vdev->regions_hpa);
@@ -2694,8 +2750,7 @@ new_device (struct virtio_net *dev)
 	}
 	ll_dev->vdev = vdev;
 	add_data_ll_entry(&ll_root_used, ll_dev);
-	vdev->vmdq_rx_q
-		= dev->device_fh * queues_per_pool + vmdq_queue_base;
+	vdev->vmdq_rx_q	= dev->device_fh * rxq + vmdq_queue_base;
 
 	if (zero_copy) {
 		uint32_t index = vdev->vmdq_rx_q;
@@ -2801,8 +2856,11 @@ new_device (struct virtio_net *dev)
 	memset(&dev_statistics[dev->device_fh], 0, sizeof(struct device_statistics));
 
 	/* Disable notifications. */
-	rte_vhost_enable_guest_notification(dev, VIRTIO_RXQ, 0);
-	rte_vhost_enable_guest_notification(dev, VIRTIO_TXQ, 0);
+	for (i = 0; i < rxq; i++) {
+		rte_vhost_enable_guest_notification(dev, i * VIRTIO_QNUM + VIRTIO_RXQ, 0);
+		rte_vhost_enable_guest_notification(dev, i * VIRTIO_QNUM + VIRTIO_TXQ, 0);
+	}
+
 	lcore_info[vdev->coreid].lcore_ll->device_num++;
 	dev->flags |= VIRTIO_DEV_RUNNING;
 
-- 
1.8.4.2