Merge branch 'ibmvnic-affinity-hints'

Nick Child says:

====================
ibmvnic: Introduce affinity hint support

This is a patchset to do 3 things to improve ibmvnic performance:
    1. Assign affinity hints to ibmvnic queue irq's
    2. Update affinity hints on cpu hotplug events
    3. Introduce transmit packet steering (XPS)

NOTE: If irqbalance is running, you need to stop it from overriding
  our affinity hints. To do this you can do one of:
   - systemctl stop irqbalance
   - ban the ibmvnic module irqs
      - you must have the latest irqbalance v9.2, the banmod argument was broken before this
      - in /etc/sysconfig/irqbalance -> IRQBALANCE_ARGS="--banmod=ibmvnic"
      - systemctl restart irqbalance
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/ibm/ibmvnic.c

@@ -68,6 +68,7 @@
 #include <linux/workqueue.h>
 #include <linux/if_vlan.h>
 #include <linux/utsname.h>
+#include <linux/cpu.h>
 
 #include "ibmvnic.h"
 
@@ -171,6 +172,193 @@ static int send_version_xchg(struct ibmvnic_adapter *adapter)
 	return ibmvnic_send_crq(adapter, &crq);
 }
 
+static void ibmvnic_clean_queue_affinity(struct ibmvnic_adapter *adapter,
+					 struct ibmvnic_sub_crq_queue *queue)
+{
+	if (!(queue && queue->irq))
+		return;
+
+	cpumask_clear(queue->affinity_mask);
+
+	if (irq_set_affinity_and_hint(queue->irq, NULL))
+		netdev_warn(adapter->netdev,
+			    "%s: Clear affinity failed, queue addr = %p, IRQ = %d\n",
+			    __func__, queue, queue->irq);
+}
+
+static void ibmvnic_clean_affinity(struct ibmvnic_adapter *adapter)
+{
+	struct ibmvnic_sub_crq_queue **rxqs;
+	struct ibmvnic_sub_crq_queue **txqs;
+	int num_rxqs, num_txqs;
+	int rc, i;
+
+	rc = 0;
+	rxqs = adapter->rx_scrq;
+	txqs = adapter->tx_scrq;
+	num_txqs = adapter->num_active_tx_scrqs;
+	num_rxqs = adapter->num_active_rx_scrqs;
+
+	netdev_dbg(adapter->netdev, "%s: Cleaning irq affinity hints", __func__);
+	if (txqs) {
+		for (i = 0; i < num_txqs; i++)
+			ibmvnic_clean_queue_affinity(adapter, txqs[i]);
+	}
+	if (rxqs) {
+		for (i = 0; i < num_rxqs; i++)
+			ibmvnic_clean_queue_affinity(adapter, rxqs[i]);
+	}
+}
+
+static int ibmvnic_set_queue_affinity(struct ibmvnic_sub_crq_queue *queue,
+				      unsigned int *cpu, int *stragglers,
+				      int stride)
+{
+	cpumask_var_t mask;
+	int i;
+	int rc = 0;
+
+	if (!(queue && queue->irq))
+		return rc;
+
+	/* cpumask_var_t is either a pointer or array, allocation works here */
+	if (!zalloc_cpumask_var(&mask, GFP_KERNEL))
+		return -ENOMEM;
+
+	/* while we have extra cpu give one extra to this irq */
+	if (*stragglers) {
+		stride++;
+		(*stragglers)--;
+	}
+	/* atomic write is safer than writing bit by bit directly */
+	for (i = 0; i < stride; i++) {
+		cpumask_set_cpu(*cpu, mask);
+		*cpu = cpumask_next_wrap(*cpu, cpu_online_mask,
+					 nr_cpu_ids, false);
+	}
+	/* set queue affinity mask */
+	cpumask_copy(queue->affinity_mask, mask);
+	rc = irq_set_affinity_and_hint(queue->irq, queue->affinity_mask);
+	free_cpumask_var(mask);
+
+	return rc;
+}
+
+/* assumes cpu read lock is held */
+static void ibmvnic_set_affinity(struct ibmvnic_adapter *adapter)
+{
+	struct ibmvnic_sub_crq_queue **rxqs = adapter->rx_scrq;
+	struct ibmvnic_sub_crq_queue **txqs = adapter->tx_scrq;
+	struct ibmvnic_sub_crq_queue *queue;
+	int num_rxqs = adapter->num_active_rx_scrqs;
+	int num_txqs = adapter->num_active_tx_scrqs;
+	int total_queues, stride, stragglers, i;
+	unsigned int num_cpu, cpu;
+	int rc = 0;
+
+	netdev_dbg(adapter->netdev, "%s: Setting irq affinity hints", __func__);
+	if (!(adapter->rx_scrq && adapter->tx_scrq)) {
+		netdev_warn(adapter->netdev,
+			    "%s: Set affinity failed, queues not allocated\n",
+			    __func__);
+		return;
+	}
+
+	total_queues = num_rxqs + num_txqs;
+	num_cpu = num_online_cpus();
+	/* number of cpu's assigned per irq */
+	stride = max_t(int, num_cpu / total_queues, 1);
+	/* number of leftover cpu's */
+	stragglers = num_cpu >= total_queues ? num_cpu % total_queues : 0;
+	/* next available cpu to assign irq to */
+	cpu = cpumask_next(-1, cpu_online_mask);
+
+	for (i = 0; i < num_txqs; i++) {
+		queue = txqs[i];
+		rc = ibmvnic_set_queue_affinity(queue, &cpu, &stragglers,
+						stride);
+		if (rc)
+			goto out;
+
+		if (!queue)
+			continue;
+
+		rc = __netif_set_xps_queue(adapter->netdev,
+					   cpumask_bits(queue->affinity_mask),
+					   i, XPS_CPUS);
+		if (rc)
+			netdev_warn(adapter->netdev, "%s: Set XPS on queue %d failed, rc = %d.\n",
+				    __func__, i, rc);
+	}
+
+	for (i = 0; i < num_rxqs; i++) {
+		queue = rxqs[i];
+		rc = ibmvnic_set_queue_affinity(queue, &cpu, &stragglers,
+						stride);
+		if (rc)
+			goto out;
+	}
+
+out:
+	if (rc) {
+		netdev_warn(adapter->netdev,
+			    "%s: Set affinity failed, queue addr = %p, IRQ = %d, rc = %d.\n",
+			    __func__, queue, queue->irq, rc);
+		ibmvnic_clean_affinity(adapter);
+	}
+}
+
+static int ibmvnic_cpu_online(unsigned int cpu, struct hlist_node *node)
+{
+	struct ibmvnic_adapter *adapter;
+
+	adapter = hlist_entry_safe(node, struct ibmvnic_adapter, node);
+	ibmvnic_set_affinity(adapter);
+	return 0;
+}
+
+static int ibmvnic_cpu_dead(unsigned int cpu, struct hlist_node *node)
+{
+	struct ibmvnic_adapter *adapter;
+
+	adapter = hlist_entry_safe(node, struct ibmvnic_adapter, node_dead);
+	ibmvnic_set_affinity(adapter);
+	return 0;
+}
+
+static int ibmvnic_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
+{
+	struct ibmvnic_adapter *adapter;
+
+	adapter = hlist_entry_safe(node, struct ibmvnic_adapter, node);
+	ibmvnic_clean_affinity(adapter);
+	return 0;
+}
+
+static enum cpuhp_state ibmvnic_online;
+
+static int ibmvnic_cpu_notif_add(struct ibmvnic_adapter *adapter)
+{
+	int ret;
+
+	ret = cpuhp_state_add_instance_nocalls(ibmvnic_online, &adapter->node);
+	if (ret)
+		return ret;
+	ret = cpuhp_state_add_instance_nocalls(CPUHP_IBMVNIC_DEAD,
+					       &adapter->node_dead);
+	if (!ret)
+		return ret;
+	cpuhp_state_remove_instance_nocalls(ibmvnic_online, &adapter->node);
+	return ret;
+}
+
+static void ibmvnic_cpu_notif_remove(struct ibmvnic_adapter *adapter)
+{
+	cpuhp_state_remove_instance_nocalls(ibmvnic_online, &adapter->node);
+	cpuhp_state_remove_instance_nocalls(CPUHP_IBMVNIC_DEAD,
+					    &adapter->node_dead);
+}
+
 static long h_reg_sub_crq(unsigned long unit_address, unsigned long token,
 			  unsigned long length, unsigned long *number,
 			  unsigned long *irq)
@@ -3626,6 +3814,8 @@ static int reset_sub_crq_queues(struct ibmvnic_adapter *adapter)
 	if (!adapter->tx_scrq || !adapter->rx_scrq)
 		return -EINVAL;
 
+	ibmvnic_clean_affinity(adapter);
+
 	for (i = 0; i < adapter->req_tx_queues; i++) {
 		netdev_dbg(adapter->netdev, "Re-setting tx_scrq[%d]\n", i);
 		rc = reset_one_sub_crq_queue(adapter, adapter->tx_scrq[i]);
@@ -3675,6 +3865,7 @@ static void release_sub_crq_queue(struct ibmvnic_adapter *adapter,
 	dma_unmap_single(dev, scrq->msg_token, 4 * PAGE_SIZE,
 			 DMA_BIDIRECTIONAL);
 	free_pages((unsigned long)scrq->msgs, 2);
+	free_cpumask_var(scrq->affinity_mask);
 	kfree(scrq);
 }
 
@@ -3695,6 +3886,8 @@ static struct ibmvnic_sub_crq_queue *init_sub_crq_queue(struct ibmvnic_adapter
 		dev_warn(dev, "Couldn't allocate crq queue messages page\n");
 		goto zero_page_failed;
 	}
+	if (!zalloc_cpumask_var(&scrq->affinity_mask, GFP_KERNEL))
+		goto cpumask_alloc_failed;
 
 	scrq->msg_token = dma_map_single(dev, scrq->msgs, 4 * PAGE_SIZE,
 					 DMA_BIDIRECTIONAL);
@@ -3747,6 +3940,8 @@ static struct ibmvnic_sub_crq_queue *init_sub_crq_queue(struct ibmvnic_adapter
 	dma_unmap_single(dev, scrq->msg_token, 4 * PAGE_SIZE,
 			 DMA_BIDIRECTIONAL);
 map_failed:
+	free_cpumask_var(scrq->affinity_mask);
+cpumask_alloc_failed:
 	free_pages((unsigned long)scrq->msgs, 2);
 zero_page_failed:
 	kfree(scrq);
@@ -3758,6 +3953,7 @@ static void release_sub_crqs(struct ibmvnic_adapter *adapter, bool do_h_free)
 {
 	int i;
 
+	ibmvnic_clean_affinity(adapter);
 	if (adapter->tx_scrq) {
 		for (i = 0; i < adapter->num_active_tx_scrqs; i++) {
 			if (!adapter->tx_scrq[i])
@@ -4035,6 +4231,11 @@ static int init_sub_crq_irqs(struct ibmvnic_adapter *adapter)
 			goto req_rx_irq_failed;
 		}
 	}
+
+	cpus_read_lock();
+	ibmvnic_set_affinity(adapter);
+	cpus_read_unlock();
+
 	return rc;
 
 req_rx_irq_failed:
@@ -6152,10 +6353,19 @@ static int ibmvnic_probe(struct vio_dev *dev, const struct vio_device_id *id)
 	}
 	dev_info(&dev->dev, "ibmvnic registered\n");
 
+	rc = ibmvnic_cpu_notif_add(adapter);
+	if (rc) {
+		netdev_err(netdev, "Registering cpu notifier failed\n");
+		goto cpu_notif_add_failed;
+	}
+
 	complete(&adapter->probe_done);
 
 	return 0;
 
+cpu_notif_add_failed:
+	unregister_netdev(netdev);
+
 ibmvnic_register_fail:
 	device_remove_file(&dev->dev, &dev_attr_failover);
 
@@ -6206,6 +6416,8 @@ static void ibmvnic_remove(struct vio_dev *dev)
 
 	spin_unlock_irqrestore(&adapter->state_lock, flags);
 
+	ibmvnic_cpu_notif_remove(adapter);
+
 	flush_work(&adapter->ibmvnic_reset);
 	flush_delayed_work(&adapter->ibmvnic_delayed_reset);
 
@@ -6336,15 +6548,40 @@ static struct vio_driver ibmvnic_driver = {
 /* module functions */
 static int __init ibmvnic_module_init(void)
 {
+	int ret;
+
+	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "net/ibmvnic:online",
+				      ibmvnic_cpu_online,
+				      ibmvnic_cpu_down_prep);
+	if (ret < 0)
+		goto out;
+	ibmvnic_online = ret;
+	ret = cpuhp_setup_state_multi(CPUHP_IBMVNIC_DEAD, "net/ibmvnic:dead",
+				      NULL, ibmvnic_cpu_dead);
+	if (ret)
+		goto err_dead;
+
+	ret = vio_register_driver(&ibmvnic_driver);
+	if (ret)
+		goto err_vio_register;
+
 	pr_info("%s: %s %s\n", ibmvnic_driver_name, ibmvnic_driver_string,
 		IBMVNIC_DRIVER_VERSION);
 
-	return vio_register_driver(&ibmvnic_driver);
+	return 0;
+err_vio_register:
+	cpuhp_remove_multi_state(CPUHP_IBMVNIC_DEAD);
+err_dead:
+	cpuhp_remove_multi_state(ibmvnic_online);
+out:
+	return ret;
 }
 
 static void __exit ibmvnic_module_exit(void)
 {
 	vio_unregister_driver(&ibmvnic_driver);
+	cpuhp_remove_multi_state(CPUHP_IBMVNIC_DEAD);
+	cpuhp_remove_multi_state(ibmvnic_online);
 }
 
 module_init(ibmvnic_module_init);

drivers/net/ethernet/ibm/ibmvnic.h

@@ -825,6 +825,7 @@ struct ibmvnic_sub_crq_queue {
 	atomic_t used;
 	char name[32];
 	u64 handle;
+	cpumask_var_t affinity_mask;
 } ____cacheline_aligned;
 
 struct ibmvnic_long_term_buff {
@@ -983,6 +984,10 @@ struct ibmvnic_adapter {
 	int reset_done_rc;
 	bool wait_for_reset;
 
+	/* CPU hotplug instances for online & dead */
+	struct hlist_node node;
+	struct hlist_node node_dead;
+
 	/* partner capabilities */
 	u64 min_tx_queues;
 	u64 min_rx_queues;

include/linux/cpuhotplug.h

@@ -69,6 +69,7 @@ enum cpuhp_state {
 	CPUHP_X86_APB_DEAD,
 	CPUHP_X86_MCE_DEAD,
 	CPUHP_VIRT_NET_DEAD,
+	CPUHP_IBMVNIC_DEAD,
 	CPUHP_SLUB_DEAD,
 	CPUHP_DEBUG_OBJ_DEAD,
 	CPUHP_MM_WRITEBACK_DEAD,